From b317cbf558f0ec8985eb23d96a9f9610f7a15387 Mon Sep 17 00:00:00 2001
From: dongzhihong <dzhwinter@gmail.com>
Date: Mon, 4 Sep 2017 23:21:50 -0700
Subject: [PATCH 01/94] fix typo, rewrite graph

---
 .../images/multigpu_allreduce.graffle         | Bin 0 -> 5294 bytes
 .../framework/images/multigpu_allreduce.png   | Bin 0 -> 114947 bytes
 paddle/framework/multigpu.md                  |  60 ++++++++++++++++++
 3 files changed, 60 insertions(+)
 create mode 100644 paddle/framework/images/multigpu_allreduce.graffle
 create mode 100644 paddle/framework/images/multigpu_allreduce.png
 create mode 100644 paddle/framework/multigpu.md

diff --git a/paddle/framework/images/multigpu_allreduce.graffle b/paddle/framework/images/multigpu_allreduce.graffle
new file mode 100644
index 0000000000000000000000000000000000000000..2659baf8ada1c7ed09cda326fbb6e56dacb8fdcb
GIT binary patch
literal 5294
zcmV;f6jAFRiwFP!000030PS6CbK6F;{T%-aUVquDyAm|-N7nI{EZMPh9b0S3_Qs{E
zTQCVpSVIH@LYAFW{`>ZT6iK}J7GG+)%A|xIU;y-)KHWV%^W>l3uA|1+AnAs2`=`g$
zBaa(FyBW8__T^8H-=DqMH;@1M^wE?5IevR|_S?zxMkflpY2)Pmv)8YV8jtr64h}n=
zC^$GcK09ulync0h)_`*y96W#XxbgTZO*=mx9NgaCdcKqLn(?(0>>ivXaVJR9@2}xx
z`_RB^rLD(sncmrR-GRTh!e;vP(GO3)1mB+?Hq-EH@Y??#B(K`7;M<b}_g6R=w$tD;
zNS=}>2X|lZn!vTgrk}zoKlIK#Xyhk}@4o!-q?;yi%TJ+)UVPmSFO&GD<GqD%KPUdh
zMHHwf2YrzdX{xNpICP(Br4$<xd2(=f>RuOa()h6TpPOzvxOAG_1P6D=pZU!%?vm{m
zT&{B!HV0z;d+&=miSFI8K{x*Vlk7FfUwe%|sg!iDp*8>OuUvm!gQLjrmfN1jQ8>J4
zp~<V`q5R!LWuwoxCw7{CkAk7(U6*?Y;gs>0;n9)9z2nD;e>?2fuw!rDj&=tg=-^4*
zeH(u1J;HAADUPqFP>3VH{nhWjO~T7?s0w$_GJBMxZ*th;H2fnd$qyfSsZku=T(=LS
z@UmU%6fJiulRoYEO*q9sJ_-EjDF@}tzTAnzFHW0}{j}5{MyTE=W}le`V?R><Bj=63
z6Zk3h&YdJqitS$pQP^&L^j#nF4UWTZC-T3aHvOn{RRC(ElQw>8WM94&o4yF6;Ou*+
z)XWE16lF7AJdT?;*Fifi;&t%&Sn2uob-3$>lnQ+3m9hY&5hPK;rDq8fiZE-85qz(q
ztaRYAfOIMiWt1CXtrYOyrOG3%3Ti{FUHP4$3T!2`4ewMcn*&@G;!2q6Uiq!~b`(@Y
zfojy%!EOY1HFz$c`zTJ@LGsRTg*W+2WIAjHx|~1z=vUX~Eclkz-(mjf@co@9=><La
z+ubnxui3ezOtSN4+>Cv9ND}er&-078ou;3o^9yo*(GKCWdF3bFASDe@|F1y`*5XZY
z6a3>ZkZs0MoPh8@&Xb_Uq&&Y&f}ribe2#7c_{^mK_dNL4@xdUqhVsqsC1~GylREej
z5IgVqN$WiQCKZAIDP`G*GVX&hk$o8AJ~+=rxXna_$V8y;q}SSgD4K~V4I=KmVbb=~
zo5YXKe{Z@!q~YaN8niF`CZL==dX(L4m$Sxs^MW)Eqv&1Gx@iV~O&VoqfDWT{-C!{Y
z76i+(!SXDL{Z`ZOrb`5i<tlGxGUyP`P+~N6(`dptTO*P*A_<YS4Q5AEwqejw2pxpZ
zHW(qHoz*eMBZ6C^sFs9sy*}s&!4WzLoo%o>z}!4ywd7V4N~zRx!;mBpNr<Ga@Pu3h
zox}ZDfQSMNIXcHFct=ujqX1BnQst1u;0PznL(YpZEvQRh;<x9g71GftHts~<1>Gzv
zJS8=(LgnkQJqDDcCmKCd!BX~wPvUNvjs#2x@FAWMFjh%LjnbNehU5X|#NWkf7FHgO
z)Q(=06#8C3-bN2D^PC%pmQ@ZdM|DF=@9};F5SP)He|xoqPtEY)fLA{_z#DnhzKHAJ
zxRa=O<AWDgP^=6iBN?OtHeE7f0L9{czYP4=%!7jz<DptJI8~AFlNkOq`s@ZNl@T46
zRC|(WAU;G{Cal(b4NX;%3>6;8AcH2j@U-UANJ=bMLO6oNJ#DlCa%BZ)M$@8t4EwS9
zqmO%uhe?v_z~KBZ!YB&v4v$`_Zykb5zxh4a@WPL}S?>6UQ4>-awrg^YLsQ|WX^^xl
zTQE}D<jfDF+S?%0kE3v!Syj6187}1a8tE|t;}kIjv`!P@D2H}`(^hjq9k0*GJEL1q
zIoQ%D2c3~}7|EnoQgXv609rGo97s7fDK3DNV_B4A%=kQZaH8{44&kEyf@qIe@EaMZ
z3YT2kI*u_9LqU%5usDXKc8MHAY2X;t>7|h1_bcfp5^>^5ZHS<TTCKPhJHj!@vN(pc
zvxYScHOzo*u+ms=6jBVN7@HI@-7t!A1aBmXqp3{nKIP0af>D4LGe(2AwAa|z)T7|6
zbIQ2^mt8PZ>(^H_pVYc!bGJDRBlJ9+NZR#;VXTY7P|ZP13dK0HSr}%UInlG=k1$Ab
zC%>C^kGyF^V>1+3k!J;`0!9o_R{}~bB=;mxCZ-IxR+wI5m6L7fP0}olH|be<lZA0u
z0NM=~mJ@-o86-}dM4UdJ|IJd2&&)s@^t`lTp8U%w%d^rWio?Kc706lea|NZIWL9Vx
zs)>bv0wXJ97{Zo?N$RtU7!6%00V~z~5VqPg%xWvC5|qd+I0_&nR0~e16ogAmz}_E(
z*NR7uc)TAx%DM5#C2>v+Axse~7<WdIFpn84C}+fK&bUB4?j|1DvfbgLF0hg=fIDo!
z9X8~?i(4elD@Jc``I-2Zf4$5~WFbp*h^rDES{-OA%Uax-$`?2alV(&AL0SvO5CyA5
zLF?l(F~Dhwf^b!$AgjXwLku7WCdd7j2m#P^&lm-)$qEC;Kv9=UV4emzNTq~S)Y=`%
zB+#X05(s7zFq43pga@2SAoGSjrkX-96gMt&Kr}Vhf}5+2n+OMN-P3{^B9vCr5QdX7
zcPMpGKO+W1l?zQ85_3-)vR3EfOWe3nC4i6$q8VV4Xrb%_nl!Z2u}Ps2fcFDHGcN#*
zB+g6G76_J50605K?hXOCWdM{*Ob$2)xAgKEUtLf~;3{$Gv2A)bz{)bifm<jtoYjFw
zVI&k$u-a^Zp=FY>&WxClhyldFO0xl+SdUW9B{keCDjiu868@$wCyE+jtkFA?4iHOQ
z^g&Qe2Vgn?(*X}R9U$gjqk#+KW*}%|t&mIt|8TJh>v01w)Vd{h;E=3~Y`F27Idu>&
z89=#l<2BVDWy-(+M}f~<inAbhC_BJ0I640#8Vg~vqGeX*K|EsiV2c+(K>$tzKt6YR
zjw$K97DlvNz?Eu4dw3jm<^@AM?j|0^GAjuzymxTTgi=frskq})PCpFrYALKGL|MHX
z3lj0A71r}v3+t8PR63kW38*6i*I7pY7Z*L;O)@~|Dl6pkSL#dZF$?31g`wJjcJ=bT
zo-h)4f+UK-!P@C|^H2#S18aAv=L&UK&^}h8eJjiMaNH3KE!n0|6A+wxR%Qi*oNGlC
zfJRB}SxdDQ6ozeh4R$y7A>ndq{8-G$k15y~4liB_w^tu`#t`{2@?+eGWC<&FG9^6e
z*61Xj;Uhn+HekEoZ!)RHPOZyH*j%(-%pJqC%yS_YA+;tPh&r2Tdla{pi0rK&B5^fN
zT#a+mYMjdq8Zyv5?dBE>!3_fk(QR|+mLHLt!Y^)Z1<>Ce2^#XHks>xDDbnDWSuGfG
z8>ImAv`C7O6k*VCDWu41XFZkm1Ux~hB{)9}{JdI5)RW(z&|sGc(HN64cD!WsCpATn
z(I`(Vuq&E#@USUYwV}&77!;BtBu6DpH_KeA+%2cWnP$R*_2w`lyPaz#^?)zIkRg?j
zR9e0x<Y;LNjPO|(7$H^xz9mrxKADgTLq^DsaFdrMu_H5U#HiQeAq9+tnvWlm^#qKJ
zWIc98F<yj}XtnVo!K5dRb{owrFczS|juN@TE9N$xCziqyx*L>;*=128W@e;BjJUaF
zZgVZdtu&M)BSJ=m@uDSdduKFQTIw7-Fv8Lpu)h0JdYUm~DA7c@)dXs%SyeTk)Ox4V
zYx7N!XJ$Ry-tGPrdE`;sw+o);%CpAJkpUHEpoA{%radl*RYDMO2`pv1F&Bv~Z7vd<
zbuN;&;M!^K*eWr^a->m6qi`<LPR`XIGdzzSoWufKa+%6WJBtw)oRI&(XoaOqq!o&}
z{kNF~PlPH?VK{TVLV2KUR$5^dxGk3LNO)#x87nqx#>z0prKFNlt-z^p;bY_#$SW{m
zg}kC>NST3GfHqYocaT>gub4q^<x*$D=xm3z(lRkCt1zS%*ONrUx7syT87tRQI}8-+
zsaJ;qffzsxtagVrvEl{iiNs2Bgv7ET@o)cY7$h^}qDHZQPU4$RRSr6e{7yG$<sFkz
zVh>;<6#(jEq|P%aX<WWrlYOvSFD!rZN>6N}FIMf1Rr+Ijd{y~mRUOLl+i{Vzs?+Bi
zkF>2O>_rbP|0YV&U#5A<OPeZ0oO8#jN4cs;Q5xwALN%jZ{diY+WHn=6<y9(Ov}v5Q
zS~soesEu=?YTc-!5|lI|?@Z06aB7^w<-V74^aT@J*v;DX_U^4;vSkwV>y7OsAXGcM
z1zI_lSZ<E~R=n-J<*evaQ39>&qE6Qgus~Bb&B!tV!E8ZOMy;x`Qd8MGG-cao$|T`)
zn=-`g5-`gb#B7Gdy#{3hG5cUJOAxc`#H?6`%ZP}5NDiaCM~BgfHF%2>y+!mE7xNZ>
zo7-EQj0wj1T7oF58W!6%K-Xn5Dsvf#rA@%nuQw6KUKZno6KM4W;Opi1ag!1W_3j?>
z!fo{AD0*^>qIQwwpIH#K^((ec4Q|BkHsY2cZr6!hyNXN)q84)<Q+b@3v$>WyOciIv
zI1KuZv-pnp6o%=8Vm*PFD(ixQ7zES3gK3dC%vd;ant2-QnH*NdchSrsG8U*tSQaj*
zRO>?pM^qpx7Ar`&CsYW%*XWgH3Ri8QiCq~e#N5nlDV#`dFemEXhTU$JmhYg3GttxV
zpHJR5jv9&I{?cIMlzsTLPa?k=TnG6;R^5Ub`C6(%X?ta9nJ`tGl>X``mtniE@+Q-$
z{9ZlTS0>AUs}SREvE1Bx1ci=<-P4=Tmkw+r9T=?LhjmHYImt}3`9qwVE*9bxl)2EQ
z9>4EOvBAWRS}^drjnmJxzL&sfe=&%X%B+0ajYjtri8$6St=rt`1<c)|br6(HW4&g!
zk7(|;(A*h;+4*&1dC~d^qkD7QjK>hoh-RE4w^lUEB{-spWyEr(OzQp|QBzd0HIzkZ
zg^nmXqMKM2l|h7Yz5ZB@#+f;YD=d-PN+M|%n6$0UQryB>avD>6l1L`E(J{m#jD_*M
zNQQF3A`GTPuXak!if42VcN5RJ6xBNMY{-gJFo<SM!pyO}N2WBJ5bQCyVk$bN8(DkR
z!^kE?dkn*94+3j<>agH8dSzJapY&$1`UFw3i8Wc8YzROE0t5lFf#q3>?llyza$uyZ
z+tNgOnk%ZAQbZa9hgO9ciM3l;yOpoqI?QUf$YzvoDSfY{TU4s5;w|Q}92~Kd>!HO}
zC7(a8Di__BBijd<>G6wI)ovDy(bH{=o=S5s(Nn83Zcf*+HnK#273nw@X|3CmnFTD_
zf`uKbw+kYjY{)9D9U8JNG-OI*Qh%M0U9fsXYuTGY#}Tp*2C_J>Zmp12%kUQwu@A{x
z)c5Bt8omZsQKPGfuHq(^QDu-|d@Vb4OM`Q15K34%wUvxeEif}%+NC&g6H*BEt{!s2
zZS>->nqpzhE|B$83szGY$$HgMbPCL(C$}@G#igs(iduUA84koOrZ}dWM$0ojSx5Xu
z;d-j*G;U&LQy&v+3cx6i0T{&6?i4c_%6i32YnczGusj2Sv5ECdN}Bb7LLw*-6q{JH
zB<x;8n&FO8baj|eSe1lTN%^XzXT2Sz=w_593A6egrPNY(lp40{mj{Kd(OBR>c?U;N
zJAM-$_u!=cc;2up{d@mUoH+3#r%^-Ue9@C%E?puXEf+0)u8=EBT}~yHOeNKYyR0az
z>L!<;M}<eqxpH@1dPoCJ8P${FOUNGXP1p*$ujA&IU=aDNw6i19AHr_<c{m8(iU0Qe
zv?%}b#J>y%x;W@h&Y|ro_%#R~fgwnPq$EyD;{O)}owN9~>5uec5UuM4^!)EZGQ9FY
zO$)l<ZoK@!3j`yReB)nNyqfGXjSA&CfnU?IeEu2TnUpAO2Hokp@gWRuC!V$rXVBaD
zCap)>VLQC`)8TDacO(;S0t1YH4co_IH!Y&l6kwUjo(3?&`e_)q3*y8h1^FoJ=JPg;
zW97cI;<VJ6cHpP6%Qw6Z+L!57=>m-Oewa0caj+YO3uf(S=wNyqcfzUU^I(H6eyKxG
z8En7$m+c^(MZOE@IlT&kw4nF{RQwvZ!V56duGhySZ~E?}j8MH#>^_kVqJE_MM`jv-
zcTcW#t?Wc2SnVG^j@y@?uEML>4MX`RZ{jqV{$}1LAsEiw;~k6#tTY~gG$_l|CwHUL
zGk8N^x>LwHAlF0Y&FQl^P2=mH*&S-ah+XWT?;?u*G!N+hcR}*t9S&B#E4+-8@Q=9d
zN6_!ET~N95g_?28jpY>=dUlxp(R%qy^xH>Df`31YKD~MwABulH``4?}=U0bUZ{HJ1
zKY#osYF-|}@2$7*>8GpDFF!<waQx`GJUe5*{Ql|V8*=#S<MYGQ-$CZZx9Ilh=keWt
z9bBY~JRUcmj644c5d0*U@vAs}H#=T&_CFP6UN$%=O_(vnC~z}ue$&v3i#oM<7kmXH
z6IAHxHoTX2j_%&)i#dxs_i_6Cn3}zax9z}Tx`qq93;fpGcJzHR*x$vshf#Rho`Qma
zXn}9_De|zEzaK+^@%SuE%j43K-~Q@%bL64d;u>6$tQ*dE=uxi2W*UB-hH+($ol*AN
z&0xm1@7v+;H^I=?7<2dUyTS8z8Ybf&`-f`sS024ThV52-TR+Ag-6RP-iR?+1$!*?`
z${IZHJC^_7d&DE$YG@Dp^9zzuS~rY4xP=U|wSy>gkvPq0KJ3tcwfSC;dOCN<-`zw(
zQY!$j=d(ZQ@W-$<%4r7AkyP9{K<+g1KfeqdwhO$3^8$_O+sk*sWp3meroDtGnc+lQ
zVy)0H2<<iYDJR);5kh+IpjE~wsmxx363#uV7^l`MX@u4l&+JniUzhljWY+y?%8k|T
zI)#q9Mj!m>rt~7q4sJcRewgZdBSM_yet5OtTOOB&pPdMR`1H|#0RRexa|LMu02u=-
AUH||9

literal 0
HcmV?d00001

diff --git a/paddle/framework/images/multigpu_allreduce.png b/paddle/framework/images/multigpu_allreduce.png
new file mode 100644
index 0000000000000000000000000000000000000000..7741bb37f47e10f13f92c732cbbe9d31f7762be4
GIT binary patch
literal 114947
zcmeEu1zS~XyY2)g(%s$NEh*jICEeZK(%q#rDAL^uB1lRpjS5H#DlH&!-sxK3xA*?e
zI)C6?dtKL3oMVhR-g@f3?}r3Ubp;GmQd9^8f}x}+s||rXz=A*!5Ru`*CzD>sRp2i;
zZ*2u>NW&EQKKKWUhoX@;1cHVK`yUQc_>36bFz={q=wqm+Dq`*K%4uoiZe`0E;OYVH
zhCsvuM8J=(wmz2B0j@4?-Xa0ww14gp0YAe&=AxzkbBm9&IIW?YCbf*amo2pbCqE}Q
ztpqAHHMN+Rjh%?Lto+}PgMW$BI{5f_h;VWF`}=eH^KrU+*>mv-3k!2`^K$X>a)3KH
zyaU~QECV>)yy^b>l7D@VtgW}Tm!pS|qq`e5?0YS(+<kq-X=!0E`uBf-jnl``?!VsT
z=Kc4yzy!HqSGagMxw-!R-QZC%*ry^g?yeqQw%*?0`y~X#{@nRLKKrlt{B^&kgS(GA
zSOqUfYb7@yTQBgik0oq$61;yu{r|q>Kfg=E%h49h`p?rme?R@-Kl}TBF)rA||Jy?R
zwaR}!1<NdfD#rEiEt5cP`pWwW0+EC$$x7)4z#SI(m*UG^cwkWH8>l)DkrUx_W9t&~
zDd9iH;h}a`FD*w-bR@vnM2FH@z@J)4!{M#d=D%%-LgCqkJU~zSa<%;Bg3Rohko)r&
zy}OfCGaqyw`~G|x^y%@h&pqvDKl|@rXH%glO2Xr&M?mSR;r{!NH&PZP!vB0pk($1=
zomqAx@;{&XbNhg{aS}Q7AHES>wS~43<}@EE{?iK;C7m;@|8XdarO6Qs<-CNHO8<D-
zfx`>^p9h+aGh?j#ZjkQ3&+EVLtd~puA2a^<l}KKA<D_xJ0Q!HtT-rX$^FNpBUk_@^
zOFC<jpr!rCO8$MhjZpZ(KVKO~cHm&O-pKgJ<&0RM9r?dD=g(J4lEBr=*|Lje{^R8t
z<1OO<dskr(GSii|S0~3P{PPxp%|rO-EC0KB|8ySyck}+qnE%)E{>hmCf3UpBAliSq
z6>5EG*!|bjFOMY1InDa=!>t+027!aAQnJ{u$}_Vou)XwO2G@>Q)bO1+fbCG91NUF`
zI!RCS&ymTLg{;oF_upO|H1>xby+PW2<~{Y){k5?D6!{dr5|V9B1awZN)a+G|$?Y)-
zugz%CLH&rqp~AkN(4#^`zi_vo0dNm@zTA9t=79ODZ;}x2@N-N^y_|39LDgR#EV2(5
zSyc{=h%@BwY8OlVY=&>@VTE~r7>9A&KvKz~U!5u#h-50X$*1pV5G_sOdXMp`3vcN4
zS&{!|qbFMD%umREPd~c9*@&Ym4*t=74XGK~ZUj#ZLA&DIU!5<ky7hi<vV6N>WM8C;
zG%)v|Q(=unkm)ZkQJoFZu#rxDeSi16+k3ae?NiWrHcRmLc7rMgaTah(=^pgCdeO?O
z;9muN_HJ#n^5+MO{z}GG;<(okyXoT4%Z<3N2ut46;^%O><KtQ2ty;)XiHNZ*rV{_p
zfe#TnzJ&drv8ZgXW)`=q_<Oo1(0tvS%dctX;kW0WBNU4*wiBn*kNQJU`c;76kH?FY
zdh!4VIdFobv#x0nb>om$9iGsLFcc#91M)O40_r3Kp@KwyZ|plk#gUIh7_fcUzF}|l
zSMZd4gS@_awf~~z_3g!K;OWly-k+}#oktQGKq%abgIMwoNqc?e#nGseY$m)tnc$L5
zHuIU4BQo`f+$hh{KB<J#W8=ZERj5C@^^vB2du~5AWpXx5WL`Sme}9<|$Itjgl?N6a
z>9Zjd9rq_AyNPTKhTiW5Di))bXm<!6-5k9Y{n^+07E@G~D}qBA%%++EPG2`8CXS4#
zv(fmmA#Ds_D8w*HS=M<zI|7=iVv#0(k<V=j|Eq!8l>f~_`J!2{>#tA#BgQ`CEHApQ
zg}Ujq$6#Y0iiV`g>pQo)1lPAO1f9<tJmq)9ZMHxjs7xAmxO}AQOu=)rDX1X1oM#@^
zz5o1a;Z;=^h2=pz!$48uZvr7sH}m#a9r)u12+T2N`U)tO2@t;=q0a$yJ`j3eh86pg
z(1kX~5653ee*OFs{-Vh5q+UeLTK6$*s`xRGG+at%1SCxQSNxxmDfw}dq}I6jIGG;f
zSaZ2~!wdr)ua03t6Q6G%UuspzzY01xZgbG0)_(E;V-&G<PMJ->WyxcN!oP1LmMrSj
zFxw+P#>yNVu`OAwelK89RGP6X`sShM1O}!VDe9CVM{g$KyD2oD3c4;n_OHJ)Hk)kd
z`Bw97;?nsV*<%qkkE2qfx5oiWqUu=p@dJ`3QL<9s#qF;zTtg;MG-tE$TVHuy+yZ8A
zTG%-}dd&ZFq5Si!;NEK?90YY5*jm!3g3Z$q{obadA>PiM8F%-ryv?mb3wJ%G<jq2}
zwE`o7_s*mxBJHfJeUYy|GeWOg#xeBwY&wmMy4bPobE*vu0m_i;{TE=!^Mf&H!DoB3
zt@hK>Tub)w-eU}TM+T=0ChqghPZ8-gX`%2`l4bYT%1}G)kA1tv!3&T5ALpJnHRG=$
zrjqb6rr$&&QCc`+VMeTfXm?%&H>8&Q=s9Sp?^t2g5pAH52npCBNevWfdLvx>5liey
z@)_LK3epn`CrZNY71Vy(klm@GUf%;j&rR727mVrP^dR^$1nP_h+RELkA@tIhRL@zN
zVMc5SRs1ymgwg7)+NBF;?m`C{Pb8UpUxv_m*zGkU+DtxVE$sLAI)z$h`Ch;8f~zjA
zBz$(dH!rXGRMl779Ceyi^Ee;hT}_AMw#eqsjZlS8O6(ofcYgjH{PPyru}*$Yz5<2s
zN87x|9b=47v9qTxm<t4&3^M~UEoBRsz6AdiC;Y%8FfO+5xsQExp{Hi9aKrpk9ygqo
zKXU)WixQ7Pl(omrPs2~fm^wcPu|0hj+v(=^^C9Fi<n+A|nOm1-i9~2ic3jQf^<qCJ
zF}KCt%~c)lV6UXEBe6%JEkiqp-)@n=S-?gN@e6f{E1P7UvgO*2b3=()D6S(1ZDm$J
z9$y4w&=yM{2zHSV@hoB|SBZHja5N%m-8iQnYofMT#w>%e^9nG$7n%J~VE@pL{B!89
zPm|(K?Pa)63)9GiQnSBT`rCx#Wp>%?PduQ3g@6nN%7jq+p?I9boZp*CT$G+eIFqQv
z(Ur#i-#b>DXvMGhDpJUVBz~=VzjgNH;tL!Y9cJv^P{I`P*^A&KuhrjCN>hoB5NfCq
zG3M3FP&x?wQTq_$_c?G^QcC})7uW68Zpp;HNq^XFMVds=#o;pXmx){5eFOc=7P|t;
z*>wZAiF~gGL+g~u!L_j~%^fg=9kB}v>D+jCQho<b&AQOv7tzQ(Pd*2HO<^er<QRDy
ziy>3#af@TWq+0aQSz2c{sw6&=)jz>pN0qwIpX2yjI)2PmmiM{4wup-GA2Z32Y;t)T
zBfXW%BVdLb%#Fhvem1Mnm4~2}@T2?J=inz_fq%u>f9u(5_dbe`p4$peLa=b%S@8a(
zT*Cs#I^azdB3J(n?<~ffJI0OnL%cy~L?e3)EEbJ=MAxIzv4pv+!wlQ7koX1+*0uxN
zbXaz?=`E&pF`v`N>h<J%;6LZ_keMFO6P+&lU{iB@@6D7I^fV>2ojQ9=ZZ_67zV2~d
z?O0o7yE8sUzqp-&C%93r*6Owr(`@Md(eoDT(KABxp#3TZ%D}DHzkhtCoa*UVXVjqI
zUw`vtB$Yx1f?pe}ltt8`UWwb>FeGXF`p6W-0x{$${mYl{`h2&yj-5r2>8Fi{#1{jm
ze9$G@nH2`aJ;&jWYoUqi5lomR<w<1_M}`O`<S$;O6was~7n}C)!thgjm`Xx2UtJl#
zxOnmCmZk07f{%8E*(M6@`*ir7_oaZC1>%s`1jmTMx1m>)y%Ky`M+TV4rD(4=BUi)k
zZwnq<aG;{qy}DR=(-w(OW+^{!)N02wMspx&SD{&ago_39$=SScCqtG+SRCwcD2S;!
zOge@n-@>huLJv}fcl6EO#N7q<z=TDM;pg#@!>l|#o2Caq98}q^Grzk&XbhZ5KSj=i
z*dmK=77WHk&~55H$#Tg{B}m|PZs@kN!?Q8pWYfSlMH8#x?7umgG#BT@fln|^c-~$u
zD{sGFZ4@u=2_N*3Gk`tjpz0t#n?<=>HDFSk^VONQw24t)ghoZ`d2ly$Q+{AXZ7qo2
z_T*^%?(BmGQpP>8?b2ZKGyDP(Uh#GMt7Kj^zC%y>I!(i52kNP4b2iF#al3A#?ll>F
zG)9iLsY2d`(&qW8t+A}P6MQD<8g8$zcDlL}IJ~~LrY9i5QVDo?Hjo6c*4CXIwoX4V
zR}vhGW9YFg^v;dQ_>zLPK=tE7N?UmHwg3~-P``&j9ka$aEtWMF_+zb`K&Tfz6f(ip
zFZPZi`L5o$eLTb$3tjSK*8T$vB1BX8lJL7gyW2O%;diH0+Rm+(@IRt41eZ#!;HE2!
zC28yBUj8H?U}u3}D@E&M!c3}K#vy+p6`5WfxKrSAJkuW@esYn~F!&u8b!gcsNsb!*
zm^l~H@zf1I{QmbBp6&F<@V)d;zQ24?Y~?GtBOReUzx%yJ4<VrGDn}1%D|20SICG0~
zkZBScVKaY;KFX%sjBc;kF|Jyw^Flp84q2h5O*H_w1P<Oja66|`f}vI`;4{LX>?Z;u
z0r{~7hw5@0@)@|!bo9<&>-6$fNAx>k7rizyQUS3yecK@rXjluP)Q%|18!3D=E*AfR
z#<kT1%E9W7AOkpBRQ?3%8*AOXtjtm$uPagUu>0BhC5*j}&XZC-sR<WLIT1h{ncv)U
zQs^QFzbCxKiwd9$Azj;+>6DLvnDaXxHFWP3S}gN9KNY5C%#Fqm5M!3<L7mgS`R*Q0
z7nXBAz5R?*ic#$3J)bqrd&-1gicdzoIYT7h;k5C>P?LBPxW3*kR;ylHea`icZMz9V
z;d+zK1~XMJC2TfXGbgu|rGzf1iN>%>KsOV?EI7<*s}Vkn>`K*!v3y%f3>-KpnU~lt
zL(&hPW8~zbYiyDYj7jYr9wq-dFZ2+&7Sk*a2N2T8ty}D=nn^oxt#66SWQc9K7miS9
zg*w-~CRB?LTje#GmJgzS1s;mJ^uC+QDX`^D3!L8NOi@c@54Z<Gs9NxJ(Q8C3bV}ra
zogA_)`)VswC4hCqao~91#hPiUDMOeVN93be+swN+_px}%cZ!)a<E)W2DC;J6;vRba
ztvdzl8et4j6eM2tglUQ2$&EKC>)M!&q~byR-myo{9^4-5l6E*ZV@JdTl-XBoNm&I|
zNADgv&e5Y{+bTa~6R?)H`bnnjMbk>O!!34{pWD=l$rM<Oqpe~h_V&o2l{u(=ULUPi
z4wC5>Z7!ZJIfrs&Ab%r><DJ1GJ&y1!{igjoIVm|%mDK9@6-d%w=ntkR70Ov&k(Ny2
z8InXoROOCHJ}Tp)<voP^VX)TK6`J!fPI>-UgBeS%Psvb~GCHI4YwiV$m|Z!M0J(lp
zIW(^r)QG;Lh&n?N-F#JuX5^#aOe{~e0I$QxQyfxVZ(7hgE%DUiew{wJq<J7EM2X$c
zx<pNX0LLjB;qrXZ91GQ@d6-a&?M<&Z)0lVkX@r=9W&1X}{`;pwSm7tspP*f}ah+#d
zw%zhUk8l<7U<dEXqpMWWAJlwuiw%#Bk~@CB{%z;=R!(vfM}u(<kvL*$Kf;iP(wd$J
zBas5+lP+14Jj)t^$%ssRd>cGkD12APY?Lm2%u+n}pn3yI!68Z$h1W&QlguFu%e}_$
zdU`ob+ANnWhRj-eJkzkEBwg|x8s96novmYt82bt704v6R>e*O--M3Ah{q`kW!;9f$
z2ogkMMR@|-Z@e7N6fGbyK@PRmHZEpNmx9uNf6n&jt6z*{6g1V($hj>bMn@8$RA_?e
zA&yw~&|YkB;SsJ=hg?mA4Cy5d40IB2c->5=St{=7A%z_pbaOm%`jCbkD-V8$Mz96t
zeAhAhWqhN}U@)i`(}Z`3$tT2?1~G6RVxsafy(z~hk1IHeH`pAcp5eMo;C(0`W&PJQ
z|El=XB~frKY$Y)lLhPXPt(!T;%w^vuqWe#BDf1H_%CKw0yvfJJh`=PLn2AB$(MOT+
z-xFjka9%Qxst$EH0B;~Vg+W-P$11+8F&24!^cJgiJlSD>#z@PWQH)YLngzF&R5!_!
zHBQ00iZ)RDKT(%NfS+pNY^=6H@6se`E`}`CNR(1xd37!uQV9a)!_u0fS!DiCAWJ|-
zqkW9XQfm}%qBde<H%NCGA5tx)ELKQ2FQ;duTl;dR<Ue6bYBx~057#SXzt}>?x6&YQ
z*~-*H)=^i)bH<sKiH47{A4Cw8uqN<HBIH;q&N`H%pJ((oyL28#kTyw>qs^nNU0t$|
zvk7A_k+@j0R!TEyEKV|EgFj9FpBPDGG>&1<^}{G)rzZoGT4SA=567&Lbuz5vV0r2N
z?a%P$J{v9qqbi=Wbh5@DcBCC&Y2LG*vl})=Y%dg$q}Jpi6^v~u=xTkkt$s}to&&0M
z!<?~V2QfEVveiSc;V8{J;?+*~pM0E~KRSc~8X*L~#=4A`BD){Z?^R1PY!GOm;kjE=
z2%S@o=oP_v@P^})WJKFVOAtevr01_86g=J5FZT6I&5=|3Ca!w5D(vikZ~MRDS5HO&
zi%-3(aQUko2PIx2B$%N8_m5s#1yBZbMnj-f&m7;x-Np)i3Gv<(KUi!uXGW<rfS<>u
zrlJ0N`UFej^2^ap@#Z~Q8C!-@`oL!orVmceCklU;;PogD95n*GT>r*j5}<~{upC((
zpfEo{3<Q~(jr$0w9}4rwJswP*ytS~KYFPv2=vPq2!9W3}gN{|IlAZTY@iYr$KjjGH
zKwn>OCOLk1&La&oyK>!J1Sq2l?}G<2_#9C**B>+<SssLdkuR4;@q&1P8Cz{)%T!)L
z&>{&vgTqIzo8L*2+zf#jC!NfnFVOwV49HP#Zce8wuk6i7U+TQDqO{Tu9l}xIcU|%M
znj+M|!$P2Q3^LqT*WYS2B(ApJ3w=pGc4&Nc@eNdyAblj`cfeL~&0&dza$m!$Cl{6L
z;}NRXPf!`3kJVar*i7!mDJ$vkvo7geRFkKzQ-x!1q~gfEM$6(Tzj0Oxl($NvBJUw6
zN-;e*iK+YzD7B&=$1nZ<Y)LvDT&w9WGN*dptCMlIw{ckZB??wi+{ffTGg5V2wMh|m
zM!ld^<QBuJ_*kWq_d#9M6;$B0Dlujz<4{5lV}8HGB_8W{7BMy<lwp?}?+XP?><ZR|
z*v<L^c*vf92Fa)YnkQ21EWoTl)uy%VZjxbQ?orD>;j6cYtn2eLs%HJ;c8DC6Bq#>G
zj^MCIUSO4SQf##A;Z*M-IIMQKx`slMr(cwWi3CwW{Li^3+84@edX`TA3|?~%7$HN?
z?~<TzFIBi+owYBT9jtb8l@a#2Y#UJ6MW~L(VUR6UDrJF~@Vhbm);a8;v%?3^f@cui
z84QpOJSCq&6sALv!20sw>e<9B?<Fu56N?9(C6n-}3vc5Wf}%%&m2hH$q+V{`;-j%H
z#<R7sIiz$lp~l2mNhjDgmy>X_IL3HV*+6X9nrwdqkfsYPWNR2DmTMV>J_}7XDI(($
z2t23K?=H8@3#~jJ$g~0E0?}>x+Xq$PpQM9-ZReSTP0vcW{ju2i#WIef1EdB7>KW!?
zSKA$+hB<7R808nU8jAnk`Q^2TtC6<|GPfP){nu6icNO_A7{z?n^_hAmEE9*x(Dx(x
zkwna;H*o9kNB<0W<*pK|v}1^!iXk%l3erX>CmevgKzPTD%fFsXi@?mbam7u5we$IA
zB5w9AQ~onPGJ8w0eFD#hO)5GlfB#f-bhFNyB64~~<me<Q4iLX)g*RJd!ezfnC_=`G
zdZY9}Gx;g{)cS`OTNPfZx{ej6j8O+VM`~d7p06%JS!wR^1`*V6Tv~GC!55WC6<Qm=
zzJyZ=xsl_@BT)uW!{^&QBt2Aev7M__dQlw2s`R0mBU4<Qmq2kZ6(vWr7ZeH4!Mv_q
z4V<+E#M%k<5SV_KoWa>lk7v%}DOO=cjD>O6dNCNE(Xl?~u)+msM(-%X_{r#_`>W}?
zyI}~vYAczLn`zF>>F%_>5#xt*Y6RxrK43vPx9-F)*Pmb<RHGK|dQ%VB92VFYTYf?+
zX5gLbaQwh%cuR&pyY8poZ2Wl+J_%QJA|39`{QjCq$+bKE3>zfzi7t#rTHMq0Q6NW^
z_0eJ80-h@1uz4i2V2|qlQh_21FUEJ_TLkr5Bmz?A{2s_2f39_V*Eh6#d_xHVs9Ul|
zCaPuQGC;!;%YSCG=roaJkS5T8618?m<pDTs>3??51$Njl7$)#DB#lEMW$g|uI~HTt
z*jFd>2If4BnGAj|RSYWR%h;a$IC)ZbNTw|%p%+F|b6y+*npo{xOaeOLQ|*CeP-;nh
z@d+hvBe&M~g6zKnE|%?=%bA6EO-HwNrS)_b+jWag^vv{7p#d`%aQ32~N3*w=Ra=%o
zEjwgyTAz=nsMz-){G&~A!+cB(b27SOl-Q%YOKX~Mom2F<)Wo@3v%UWArZdAiRB+e`
zme?%jnI8R{?}op#nG-yJN%`}e@BVwR*AOH@M)j!i18t3|#W~J;I)uzE(*&aDIS@|T
zLAoQ8M*zbZp7T7}8bgjCM$S?o`ICpObv`a7e3ZaAKus;qR7QBXv{Ipfh3nTcrUG)Y
z16=9p_K%X{ALWZas@}GNz-m`@Bu6vQdwBQzr^XM%n_RIM_opTI>6?}S5j6E%w)M`6
zL+gH~T%CyeGWQ`}VQJYK>ly1G+!@)^wrF30Ds)I2F^Vyf6N{G`<pg^m-bF4?0WMJ;
zLWJ0CN8k1u7&a_GnFeA*c(5O2VApjGF7-fbXeGyWCkyEvm^_|%^p!^^(UN_i6M=oh
zdO@|a#dIEN<Z`QhD;+DD1NRGN{VwK7U=i+6N%Ux`O6jwxg>q@ll;EPZL&dtLS$VPz
z_-)L~85t~!puIAiu(67@zor*ikwf@uY$wB`r<N8zo*|`<=quxBPwSf}NLjdS{j54P
z^SpT#D5T}Wr0?DF78pUGErcYM@{Biq!8_<r6ey~$rpu(p(xxH9)n{KXz@>?Jkw?Nf
zaTAC``@%P*O`ce`2gGC?ESj%$`I(9k_`av#`%b5>A>`=I43cg`bH{KF9Esl_vd3HW
z;jQ_92{m!+W*Ccw;a~vBwN7AHRJpIszIKd>xo%VJZTk^CcE5BP&Hyzm{akCnwHb<h
zYdiTYU^4;nlYb)s-1((s_!TPnQy@9<+bkgM2!oFL^@Avb%iy`@IWnx#Q<aXooSx2P
zfj+ELDUE0Goe%94h2MHLof*&A{v{FZF;R2?@_v$0?hHF299$>_N3tW~8Hz_Iy8sn1
z$7c#EC!974=w>F<qCdqtes9aTXxWG~NlQ3nFZ^1v#-dNCb0ky7js$35WU@3scCLYC
z28O>695!;*drnhWp;agk8LD>0$OQxdEu-Y#`H$Y8sX|e03pheWq28mh(XzzN&2AS4
zM!bm7e66$ZtNs9gqR%z6j3L;3i4#Rbx!<XfMp^%cxj0nUNjpX&;%nL?iP)Y4gm}vr
zvMkzr2o@^Uf32ujGC?bcV3OnKzQ7miMP`B91Q$45iOg*(2eLnCP)_jW9+0Il!F|@h
zzv%2g9vttmg*kbZpfj&IW;Hs9759it<(_Sm99Tj!_-XN9<iY7#2Dn6ug4;=k4^Z6u
zj^6=r7quXXIaW<U2{M5utkBr=jdkHG-WyWe5ACjK$4lsWlkD=u^Mpm)r|ObalAkk-
zEh$m5>##szqEygQt+l#GTUbdi#z1)Q_@ao=s87D?p#ud;dUVVqnOJnG4&|i4mP+Kw
zXTp4n_zFRSxom_6{!q!}^92*Ee_02;W(K5}YvYQVNApHaF7haj!jD)qnwBrqKzD?1
zpKc&EKsGDmd1sfqsZo@i;zC99y)_Ji?HjVf9Tf$88>rwE6G*Q)E&*t)KypRTy2SUR
zB7TRLfGIg*m3-HgA0JQ8uCntc!o@9*FiV{!m-zl>_mL8@tRI$y+!uDnhzR1#=nSMc
z*o}rZE(U-kQH+?ha*-H(t`3}%#jv?`Ff781CU4ovC%IB>t8fdFzWpFMpgviV$a^mL
z5jer<8DoE?iJ?tT8<642#;(MA(fu`??vw%>(_s@9vS|07{pT*hI3@3`;QeYFdRSq0
z76t(2F-T4J6{E-P`C^IBoRWt4xo(|c5=xsaJ(P~~Cx})Pou3x~Iku&^4sun|TK(k_
z?vaBee-d6g;S(Yas0rsoIB6Y#Y@Yo7+5f1cJF!MOVgQdxUG@_G`x{fe-r~Fn3CM`#
zBuaJ{Y{w=??0;-FKNhDEgDRK;(xZD(dz~!<10_xDWE8&B`rMOVlGo9%lwONGq3SF1
z7i|?9Vos}E)z@IVj<Ypi(rao%k*}@IJOQyluj@>&5Z0zxeUtcD9IoY)@4-R}zaGzI
znDMKvH;wTrN7WW9uU$Nh7|p?&Y$a8m<41v^)XT*X8w@vQA)a<61fA`@p|ly!dv1)9
zo_;6?Sj=vdce)cSu8%TxPM56HY|S^+1efWjdR)OVJ)8kWxe2w$vu#~Ll*(er!Am3z
zQXhb{zq1*0!W_sTznX=A?A5iVVx=AtM~c}*96N3|jj^ExC)d(Hl!;r71`djE*=~Cb
zV<{G7&^QTId@@n<9_!J~%<Ir#$_pa%5Ko-K#?Y)o=O<iAZJ3H?ziVI720{V-!C9WT
zGg%<RyML*YmJUjLPIx-Mn<_nNl*&v93S(e(v-5^LnQOKKVg?n*3}Omyef(e^9k=6L
zWu943!W4(Ugr_+eg(j0{CY#45EXg*_%PAZXt2+k{Q+c)N9^`C2e&5Ia)W?)NqAZiX
zr!RMqUwv3$_|{4v$y6K=Wh$4&q``T8VXrc-_eqF08JoBn=XH<Q9n6Er4lFfWAI#VK
z1LBkD`L&|69GYJlDadCk>sJhEP_Sv$Fm<Ihv&b$9H>F~5=;WI+8hRIERmGU0U^!w+
zor(hKOR=aUUbZu*Jz*u(v-C3p1hiowC7PoX-B3P{H6fRoYJ8Y0XN$`qa(e5Nf;m7$
z#CdWXmD6$mu4u*+WTjW}s>%V74mPTU#Q?;v3FO#5bG}kxZN$P}Hurb8?=iEYKBMbW
z3-eDSU-YTO$p`;@<)6i<rpZe2ty)eDWe*|VzMol75_pz}xNU?n%yS;xt}Fb~iJ1q@
zec;q7q>#oCI_{zSR!c%2&4FqqXp(2-5ZDDlnCd{nW54C?uCJ-q)^vI4^rfBnJaW(n
zSd<~_kw}8>pTz`3%Fsy544n8`ZRJLL;w=%K?Fy+&F_jB6dDTKm*&d(44b_O^Kc|Ir
zi@1isvl}O|Vy0CVZ<MJUAAsOx`@MZpy&BUQFr2QPpJJun5M@VSSe#pBUGFZMhf_TC
zus9#yykQ%xfKMDUQnbyIEJO~uzx{PsvefDRg`Xvc)4n*6F^I#oi`zCRr&Q$dhqBjR
z^p7X{df{d2fWMMYqkOCfH?3023C!bTOz|RNVT444Dma<Ej`Y8~{C|FSk?RK~2!nBo
z_^&m<=oRl}I&k2>ri}z|za0lqGl<4YOIpW^mCq~-cELSQXsfgE3vzVe6Bs22*bl5m
zDRO~A3RR=W`K$*3?QC~j1yCHrHiub|8HNJ(ok4MjU2eEkiIEY1vZZgBKAi!a38SbM
z!)9<uN9h~enX}}h(qI)XJSVRxy<yyaLMq{2l>EKu#=Rr;%nJ>wpxmS8O13Ac%uymr
z2C`QS*x~Sp!4aOh_>Yc>1dvc>(Y~*((jfU9cSo<Wv1@;Js;%Fo(|%1a#gdy+Vey)^
zjS>Y4CjSt3Qx$A~_vcw-T}(yV{!H@gAv_J6$Zb$4<trlQ;*)a3VvxPN)7fj^;tEq?
zg*%Z62N2Ymqc(w|wi~__MYi!t+ZO<viQ!pr4ZZ_37dSxE=kF>+&ZI?slH(pKXQ06e
z?yV!T=dopZyhmYr77?YCu1kU^1Kw)YQ=^MGkEGeK+fJB+kU83T7}2p)=<Ii}z>zU$
zz<5RzN8a*|)S2GXzH_{d<=oqa&11#!#fGvK!l6%-B+f@E79nDFg}53y7ZlZe?#RBA
zW`f2aQ5i4XZ!V908%IbJrPqOdhk=6uQlekwZIn_K)v;04BvK}VTC`yD*@_BG25O;@
z?Q3;sEn(`h-^!3FK<!Qe{lr#+S^W@jaGz(&WH7jD2(@{`K*A*|^vUOgpz8{~%-Uh4
zEJe0-`(dJqG6-4%kyKz=0DrBTMOR511%pL^xk!IVd=$ud2W>jrSfed@d?WW>O<Gq*
z3Zt-l_OwPBm|r!15|gzwSc&0O#qW3x<u#7t-e9B|iZQmQEIR4)C(%UROH7ELf&eE8
z#1>qLiP<`ERhso{O8tNkyP}OCvF-fi_f;yr5!KodYo&~?@SUchTusIR1*qx)fbr3g
zeHKv@TL7;<YH%cOK;zNv*#H6-Fnf(OW+h?xYsfg<M(sj*qAIQ*W4(@N&juh*{Ay34
zF-A}~VW=c{r~N?*YCruQRJobn6;y!<i?WFH*edWY{xc1adkqtFHOdSRPe1BtY`4lE
zBt{PaS1twLYxllTZq9g?RO$^3E&+lLu0p8bGg|oQq;Cj{TH((%2|^(RqI%9+a*RqM
zr)%ER9EMG$DtVkrxFXI*;N;Hm9I&wyi^5kZq}dUDUP{PeibUoj1x~G)wUrb_(E<hD
zFlnX3wRh+xf0*SDejtD0cgpoDA})bSy$B#No(X4GJnNC+->wj5J-@6~@+@vlFPDhG
zmTSR0gVS9x2hL_VGe65v@&m<mf?up8lifpo>@%S0*&*PoUr_z{2`CxtCi5c;33Pu&
zL+Le{;9>ke5>-eOP?9XZ%%wL?PIX9pt%R*lGJ$;L^LEbQLj08iDgTGnPA%u>Skz9N
z9R_4zby=&b+N-tM#~vy@aA0RStj<LD^v=ZnJ~L~&DEx-)P7V;l8ow{T8<=_yZzgFI
zqiEKHR0kYYf6JOqeUxm@NZ`d{<6{9IEqIPRgQGSpbYA#e^$k$p<Vqisy^1Yle%kUw
zr)Ur6$cwu<77~FIQ}%S^sKHm^Rp))a2t+<G^#IA{4m|_xaP0WB6A=_ZM9s+o6eLGj
zKJcE(Z^^RSXkEU}44y$DB2;O~uSxc!yp9lyhN<}n+o!G_Mw>%ZQtJTlx&av7_=YrB
zh9o2KR0aHEsWjm49KXbca&R=hZhd^Ili;ivTj!gOOEKGqNf6-ZiQL3=29O;XK1KNA
zs@hmiO53>`&QG_`^@T1nx2SC@X=KNzUVn49?rj(~?dAF%5D2wk#GvZd{}Dp-T7--X
zIv=&56oeKvNGjl@JEPCbX6`$$2jbGOU>6Kg0`x6k_-1}+$iwhe*UlW#lDx(nnx-lQ
z*4-*WHWnkW2g)4>S>;_|U4oUJwW<s=JzHcqw~-_5C>(ivWLfx-*JZ7^a)67od#18M
zg9;+o8uKl_Up*^9j-<^(`NrG0*2*&&0jBE@YF_}3Jn}vP;FRcTNMtIC7EedDqzjs^
ztIgNeXE%=lM7Gv_3KHKNP~gTr3WWtEMk%YuR*&CiU5R_d@8sD#E)Xu8r$|K1kz4qF
zG^Uc&g8<M(j`7DX5r^>^fJK4mO5{h>g3)XJdgE7Tb1Hi#yM=Ys+VBjS6WGt~ASpf;
z%Z85FmwOdyf-4Rc$Ue8Aeoofg8)GU#&&!4px<H%-bOS{|O0<LwElmI^T(s=sY@01x
zAPx5NWgYYPcpN%pI;;5DN0wsOK=hO#hC01<J3v$DU&&x-X|q5Z3SoHI^R2-PHl^z0
zUm#6lU9?TO5(&O=p(WKfw2T>2d{$e!7yo?ksWYOP0Xe)2;R5gkn{5=XwbcnNvmK7k
zxQfDXgABO<A9KhS`_c8)L*<$0I}$WP_R?VxIWrYYoEu*Ww!EL6#Sg8}TF^g{rj9xe
zoYynJDVbQZ@UUZ)`r`2xgA3@}67ZNLPDbhWPIjKU^9sSxw_Fpq0-|}=`XL%)7DA1M
zL`fS#{ZL5zHzG77C@onDLFC995|9d8<uQDn1XUy2wH-DaDlpY%TqR2<dda*hQN&<H
zhyX<i4F(o{IuZQjI|F0Ksy<ykYX(^-Q~yuCvSonpy($Sg8GCehoCZ`uIVs<r`?$>!
zXLOylUR%iPXq1gG$uw(}NVu-F0cDKpL@tLYD0cwv98-g1ICKf)g^+un6JJ!oAv}og
zC)^yMF-ESb(zk~SE?~a|PACS+2shp$gz+O7DDwmjey`PVVjdeRGG7ECV(??qTcF&7
z3D!XR)ySDczis?c)<{=9abTP!@)FOm%VPs%ALIqk)C&29Z_hp)|I9dw@);z0?4xC)
z_90#8nAvf+sy-@py0QNbw!J?AeZWnPLNQ1@1V5(f<)v=@1xV}7KrtK74fCgEI#3Z=
znM=wo)zTTk1=TkIyPG?DNCqE6SuzGok;m<BhMHBuvG5)qZvJh}h4(e9GGDBoE~_p6
z_i!oWc)6F&_)0odTO;C8?U(R-ab<zPEE5=N-xd2St=ltOO`*{xd!?$SKFW{ex65YO
z9za^+8>~?4i~Lshqw&-OKrR79OC11liDLltqzelE$9KkB&LBJ#1Ux5&N2Dl8xB%m*
zv0ECtegRygP~oSTujy=<z!xsumU*b)#=C@YE+1+Sj}z=-b1dS7<w#PX$Wl#6ri!^y
zW03N%e-S}7U!6d^&NMc$WG}&p1ZxB84r~&Tr;{tUjI-AAJAGunAti^3mY(m;R;*n~
zhj6qlzQX1;v0p=m4d!JFaOA_AW6CJ_J|T=Nu;VocmlcsE*1dct1TO^!le)-j_MA?0
zf$>C+y5-nNYrT9c(K}xM5P4b=B$%ioWR3FNpNVR(+9e+e6L4WK4eFW990P9Lnf3~>
zDtP1l0mqx0uMvf1-@ZPgwueW&t$nG>3Vf^8u}u|tTmtVq0sxHnBa{Zlm`m2=rU?vY
z9<4)t{NUx1EkfS08a$KXk9?w=VXYR?|8lbJb#rd})L<FQ^+Go(TYRdd)X1*is_RHt
z2vC!1R!;*B1>Y%Q%r2d%N4$Kj$S?-MftnJPWFnFtr5@YfrE1$BY$495QcsSX4C{PD
z;vkF%7)uL{=H79^nyKgVOe-M4Bu_5EmNm*gZ1VAcf+5I^WBZ}ZPHKe7d1h4pV}JNP
zu-ujy`0ln;fJBy02E3i}ijQR$8!eEOa;+0q(JMt0Xd*;**{I?8^TtxMjZzbDDU@|v
z)-yulf)k|f&4#YcL{P%9T;E#Fv~`)atKw7I%dGE~gfl`OKFt))M)4lTpb*KPP#T&_
z`VFvQ5`BV0q7OBCY!^v1D9mQd()Z1!?5uI{0YD=25vZ34b^wWBG7^OA8ECVrPah>u
z8;$<Xena-v@jEgltM?yVXUpX(DJ)`qe|~<2wrH9nAS&xDYZFO|0?Zafq?&h`EX<{m
z0-i<0DOeT&)xSCba5FGP`cWhc7p=_)r}sqpopFe0gt9E5INrl$j{Hi6<y!?mI+m_U
z(;Omwk}Vq^W~X>{{)CTgH%1wXlPt6Un&ULyu`OqGx1-xP_apvPDVN~pSO5L6uNF1U
zvE&+1ZS>t>TaNS}rd@pU|560txq|oYF3Xvr2%`Q!o&~mGkHCbI81>EXNMowkL{#D=
zwSIgX9z1yb%5clut#jrp9+f1hb&lX*xIcxcA51WuB<eXD7VpF{9;iuhH5k><P^a5T
zyOYRv*798W{M_a-GVF|t7??h;4yuPL$3cF&{_HzdPuNQ$3i_ZX#DQb0Q^6>629j#d
zY7tm+s7ETCjJl$nl5?83cUhN_iReVHxwjlp1^-uW%Z59Hq?w`Q@qARHv}WieB;}7a
zJU=h5d$wq-tClu&`U2Ra>~3O`eK<=F&<RC!B>*m|@F=AVCD%U?q&!B#Pk$BGH<d#7
zczIC?+7j}0QwssDmzuzxm|+USd*-S?18p}5@1n-A@Q&-@1xm(i5J_>V9VBFBb<W-^
zs4d{&z@dwPaAX@vngsSC&LQ%8#YrW1;9G5@`sOZ>$qJuC|0})v1kC07Y@bD^M%z~t
z&z9|WEDLwE7^GH*Mn)~nplr}heEj#j0ABx8)ZYyr`aUim2u0tlJxwdDqoajCBYLG@
z`p&A$JXxkqGSs!bO0^WI)2gt0BU2rW$0>?5Kkc`J-~m{;hh-=BEP|-z&t>>kObKga
zA%e#kLTl-tylPmoQUR4h-$tdfRsZDk%N&?pN{^gs8z!N?vXGE)XweuVpej%?#Wmd?
zA|&86cd7FlE@K?;f%8j93!I-8h5Ni(49-lS0a+(1h-XSKHa!orUAAbV{XY3Fs@;Ah
zLe%0jgcM^CPAHa=!ZYAz0A{a{2Ya`KxP%>j5fo)<pz#SIGrv*J=+>mp)YHb0zk``a
z<`he6%w0}fRlAc=@zC%)(c8<RAd~EDGsHpXcHDRbIKkk-TfHuUde0f0?=#InrsRn&
zPvWGqo-Tvgy`5{~6s*@W{srjbKlS*k<W4G+nSf2SS`|Qj0LvyZ5$fe~tbm^b%q0x9
zgp1md83IZ0b3ndcRga03(_z2CP8kd10Xjpcg*qc28&@UxDk+Pr?Yu7_Xa508H2m6N
zAQiS`9lR;BqR^>ZfuR<FUZ`BcNpZqsR0S?xAyHIoz8nU<D6e$nAb|H0l!2Uil%B+1
zaM=bA%3*paAHs~?CG}Ch)IFq9JChV8Td|La99_7<liN@dz8h!n^DEJ;`LzbMA-%67
zX(OBI$XWy~v~cChV{bDz&of~)rjeHav?4VUn4DL{4ujSoJLvdRX?e#6OlglojB900
z6ZZRXu|Zye;?u&aT-pQ$#81i9>25!unPXg^E4Eh<t&~}3taL7mVeie?E>>$*oM>vX
z*8&2RAN8F!e~dCK7`)Q=?zwamP(p?+80%&zSxvzJ8bfVdI=6<Xa**=}wJ#E+Kwdr%
zySuXAk(SG%&WzPM0x-t>IVp+@8H+Id`M3Fl4=<)cDy{QUIH-(5<hgb)uGY8eid0%Q
zr;Q5;2d#Xzl@LEZdpWvbkp!60+~NzC4~hUZz?_GH2AQuZJqF+a>}d+5-kH%2RdxAf
zFoAHnBS&Lvz%fcwSd;T&LtQKLYv!122qH<z^Da($>jO?lc(5;PfLjCPejYEXL$sMg
z9`7q<wIYy>7X!u^AlOPJo6Ipz%=r@YAFJbm<>BMJR<nm82g1sr-3?A@ZV|>L(&K`1
zCN7KKhr?D*)JOJ%v6=m`A-2W2`?q>IdOeyt5yCD4X?REaCZIp3@SfxW@>i7tuYbKy
zvk1WdcR+i63G^!;Dn-AQ;@4CuvG^vSfW_t6M;@-u2Da~I_{SQW8XNA^4C8Y0yFmJP
z1$c)$Ao6&j=dSzvV%6OrROxN3)n>p9BJXa01FiEbuGg1V<k4eE=m5Ie%gI}*tL%cp
zl)9j+<Lg$I24@5Dl`Dm>ChGi&NqxH#1S~H310E)DG+Upv5!%ive&%5dU25J0>O|Gz
zJtM_rGl#fTqwID?a+D^Cq$2;36fz=(jT0KYnzGV~nUgnjBJ`RJk$MJ?eP>FgRLy~g
z^cO1{!!A<W!~M!EMdN6*u?%CSTDk~J#N9U_5vXO$e)dHZ>Z{2rOv~mt2!lue%<US&
zD{AVj#h9{1(FkG(pacxBrO%!ecio&!6eLzypNPmO+$)p<m<%S=fI(vYc>7mWRc<}|
zjB_HD3Jd}nlBOu>3hMxT=>>`GS5^y-t)9(9Ho%zY6^jQiY?3CudpZCU9{`fS#A}?z
zB?#MH$vGF305+PMow?j}Np-N_ThZ^6%IUyz*wlmW&5Vx-@~42R@AudktYp=US#$6K
zuI&8#YG>jX@gj8it!E;8rBCFOi8IhWo#_~qDFW6b59(Ug0fJfwrB6qLqrD86D+0ny
zpuaNXTpZaU-D0ZQ16WNyzoWKGw01f$GXVDh-$Gm-SY2DKR|=n4s;*VlbbN+4gOq+J
zq23aE4H^9W!5F_>@tf4wcPW4<HaDxts*xn{vSyX$YOCG!*I<f;G!|FLrzkJ)gjPS%
zX2~a}w<Bh87{glv$`B-U;u#RWwqo6>g-@`^A0B}t#2E5*lrD|_*l5h(B?xlF_)fi@
zK2E(s895DgKe$TOEv?efD-^==W6h@;U<-Q+^>U8H)1XZS?)5m3O|WB(-;})tiBT#z
zV5BL^Mj0GHx(Ik?B_sE8Oyx3TmS1%}D%4i+buNzqoH`}E`z(I-y@9u4>W#x}hpWlS
zU=yJ=sGURi%HmG+`p@z6AQV9W_@WFwn??CVh0%;dNibQhw-%`@QV&{2ie9T%Kfi|T
zJ*<~Qp?IYd{|fK|XTX5~;K*##CA0tRVDTr6D+yvSLb2td<F*-12Rz;e01sG2bGGPU
zfT3~=?l@<pEH1SXailPC$%@a;0KGFxrTBUfl`Wk#Hj6k7ombh^U7z7-O7y2{SOn+e
z9$1TzUQWVpAXMHZS|KJtxmx!n%!B8ADkQBsgPX}QPSWNrhVh9PP5I)%eyC>!ytb%4
zrCO*yS}_p=E6$B6I^qH-Wm@2Pjo$&GAK(pLFtOq#La*uBeonvo^*L!0tVqESi2a=a
z_)YmW>)NsWj=DmXcTD`<r}-Y=gH#I9JD?F#nuXT_MOFU<XD@Y#wvGPOM{LD5taYF=
z2U38-GQ*d&*FeUhC&HcS55ReSEBiK}_hFQBXUfr*59?536gUAOg1IA{24gnlld{z!
zy*xA<K_cuXVDNky(z_N21NLUIsSokrj^ikh9|I>L|G|pFBoJs96fe>ya~@1SkdkM=
ze>0M*{Uzj7jpFI)NjV<1Q!++Amhe{k{oSvAIilzCaM&%DZ)5xX9|K*m0AWW2^hBL{
zRstfnb`3d^px@NDO!y5(IqWLA!#NEi<Vs)QDV~NH&o5Rh4~DEG06zEvWwcTj<Ngx6
z@B>Y28u+JZ>yb390FM@K#%5-s*_kRzLY+JS?FH#Nh-RSccq?)Mh%jL;6p%DOfN4y~
z_{P-1>)?nArdJ^s)H>ROsLU|x0Mw|zC(Q4x(zOIiagXu#gFRD1Kn)0DP*OJ*)^Whs
zZ_BN_AS^^KB@EQ2@55|AG_;TxSQD&V)_W4THJZU^H@TisqFJ9F0@5@W0pE=nBK0DX
z90wLNE0|6vx~5`j8@cIY)Yi5OO7_g{EcWm9Cs?lmnxa+t%qi@z&G)zj@X^rgXnKUi
zuKmNOGJ@p78e+%L<vQoSMCuB_5hm_ZOUcJkg>^~E3*aVs!yU>3zd-vO-Ne$xo0x`p
zvmO*EHk%)G@pm_LQCs$5%~a|Qsa1!dRLfrTG)dSJJ27I+nmNLdRu7Pa74)<4pO%<6
zfab=1wtN^IGy8NchN7ZL9eOUXkMH;Q=o6>S34d_uDwZFOvqzrctn)<D#D++MWqXX?
zF2ia=*Ac%ZyPQM>C8>XX{f#&u)*N-n)n>DX%-u^}o@>E#%^%xaNrMacZE363L^APh
z7AX)<l$$=_ReFuV4kd#_wjIMaOwq#uGADpl!NTr5;xY3X@Ikbv0MvR)QVTi_LoZnD
zm71jFkN2v?6yL1!b@T(qLcsGS|1?lIFJuDk6NiY!DK}=FG<+cKK;bq(BU35G3#n5<
z0w$<o3&oE6*Y5&oUuY+#m!hnc?s%^DbfH5eDtVOpEjguimXv&Z6;bVp7#C28lSXMP
z;U581dPqSqT8bVjBSz(Pyide{5(=3_wo;U#l0+9}P|oNwc_0C~fHG6S6a>s8hO({{
zAn71Bx&WqKS+^lQr6%+VzaXQJwt{;!-rsHK8v$AXqY)H<P=SJcn|kL!Qlyf&1*2)<
zvzQP!vS?eqluKoi`Q6s?86JWDB}%L$09`ll8MjJcDy1X6Oy{r<3?LbnR!hfGF~{H?
zC$T2y91@kr!YxX?`UxT)M}s0v2=hcS!Fv^-sHk26mX+{7b7c^Mui8Vn6n88cP9h2i
z(apzW4w`2C8DQWg1st){)npB}!BnAVKw_41RIw273552^8h*d3?MzZM6Y3r45>*OA
zT=HE0L_$EMUXn%RWl~M_6~n-j9e^;|dgXj^s(tCYF&N_xIw0uM#5VaFNR{<&qm}YP
zi<ZJNk{{{iu<Qk{U|gKZW{L@(fnE#umN7;km|NdwN}D&<8-fPJV^6#_J_K4C{>|ra
zcN|FJu<80W`2P}j$6y3v;~q6omRHZe0c1h68s3xNXk2}owP<Ks#E+mbh01($SY}K{
zN~D5b7!{sx_V|Q1HTJ{kz4xHJVu+hK4M2gXmR}d1Flb+!e{0gyEH|UL8>dO3%KeZD
zmh~?=#k+vaN%(6rUyG8>lb3tlyT#ToN!?aqU9=s_EQW&61&~gF4%TOrg$NJON1ul@
zkE4@6#HLwZu>%kxM=p8?8yt#_OvGaxm>i=6fvJB9u&{yl<2yBd*<~EYu=D8%BoOpm
z^AB>4vgt=1-U}?!52NEHi!He5fzCdf!LY)<{N`V5$2B;OR;9wERgOUGQ{tfI_sVr!
zvJ|*Yn_K%Tdi6~mbz#Kcc9=OIZ*f<u&M_%OfMSSRkK$k`WBSWFp!{UwVPe1wkdQ1k
zOLK7%!p*rHciD%cD?MP+C^lv1fh{f2TfL;V+M0k87>U0loJTze$FvQcS-uD-Up@Yq
z4Aie8xC<TK)T8mioxm#Oh&oTqLyi#(Is(FPL6yD9NJfiOl@@?ZBD9>%L<%qFlfses
zz&FffgCsHraTQ~lNDz72eS6!fBFq@f-3G(B+x+x*4mF*=YwK8{l>8MUnWt-V0g&`G
zzdHg6eGpKyqJ9!lo56lR>kxTI-la!sqO11^{6SE_J{=C~52j}x@H;0!G0KfJc~a9u
zRKa0v3epB3D*0w+*-eVfJ801cS{>FvYPlVD;@pCq{vt}j8^8>Pf;d%Mf2oE7-Sc|E
z&Li@?8g7~c!qQIw+y}kLO#6b?43zr^>E8fOtuVNPDxZ$CT4gD@JwRQtGMA)xC9q)l
z6?DP`0rf5rmQFn8l4zkpr{&p|6^l+zvrwi5NRDEJ0zxc7aga^LnnOhj3?!1aIj&H5
zdj4nr^gr6=m%nL4o_qudb1YIkH60_d@5^!LBt;8Qdmeo0fAo8xWgz|b1V;J43shA*
zWy2-V=E(T~#h4{enez!VX%>VDvn~pOd8!qO5cJ-F_RSQ;IA0+|SD1XN!y#h;!&o7<
zuTwF7_GsfREY<<LEQ)#efMmPK!-fF&^%0mHAICODgb|0(s<e<A9>1(Nzw!rU%_YK9
zw<K1d7_s|F6J3+hfZRd>=XWY^VcEc{^QjpnQ0>r|`R9?&DovtZ(j|lkJ@2zt)rpZk
zPA7|x^He2guRE6{mQA!A(kZ+azw0?V6X;;R`=t9ST8uFj2uL2??rXU8j@8Wdyx=c?
z@U2YnTjfr9^bf0o6_?ic=)5(r5AcrJ95wTkaEHca;_8?K1`Q_-iN^jVw;MkwcC3;Q
z{Bs}(7_5_OXtUm40Re;+$RuefcL9{lBqxa}pc{_jtIc=*0o7H{{GoY2u8j{MWM}t=
z67Wv-UYvveU}qcvn}=f~jWnn&2^|FRHCDs_y}|@O`v^cF+`$~!T@FYOa`EM7G&lKc
zlA!t2)6-qPEfrlcv(du=cLt_flMRO&v9?HA76jl@KPJDM291wBrJeoXd|x<1E#hfa
z&lzsp@xkvPQC8#PGH!!)xLF$n&$VOczx`$~C(qsgo-sz)mRTvVVbj>Dl+vkUiuMGB
zk+^Q>Zuil>kt~mkh&fd8>*xgf20D5Z$Q>|1vp(pnG~yJlB#!Bh{#%52%3zXm4zTA3
z84Kl3>suFD5)o@%fqMK<?XL*P^>Qi*sl6dl63EH+X^2Y+Sz?&!FH{Q#S(gER;{mc1
zcAsSJO3+dhXYd{Damk?@Onedcl=0zxJYDJlciKc=Xws=kce|Xt1(2uN*!Kf;<T7t|
zkTWMk;Qej%Bh=4L0I2pbHn+{uqjg@1wNi<|QFOscdm429O^%MfEU(Ue(`^~(XG%O{
zEK;3Q52ZhVFe?!%^0Fd9Wjk{8Flal+z-&Ja3#pH07ANE(^@H3ILZho@yyo&a#2V-~
zP9R*4N#}?Y0ZAP|<u#-gM+e3w-<Lh$N+5%KM)>A3WT(Ks>5_YCX*d}%t<D4`)}d)9
z<9O`Z7RZ(>&?^cJTNmn0kYw;Xy@!-NYlq6mMTv3c*L-fCR=nt|B`SY<B%)Es?+B=P
zSreD0HxJgHp<KDFw6S3JlTt0qrcOT<`U%rQfJWadZ@Q}iv*7P?ZH!3-u#Qz^$JUow
z{d?Y1!uEsFDD19`Ore5#hKy6t(bhMYeP2^W<Be8*7au?Q&G*Zx1HMtqMW4|t@Q58q
zX5InpBY2c1j$5|7@166hHfW1zlzhVh@H(?}#+SHm8GZU@USkX{Wh<TT=H<eu9Jinu
zQq*OMn^)8W&&EEI)P1vi?1|YEBrK3))p|B&%<yz7_pxF^^NX9&1cb(6U4vT2rEqvo
z<H3StL+Aj4UKI)_1zpl=msv;*>Hwqrh6M4H5i84g4+quZy*<_STWNgbROAm9<x@H>
zpvSsd?hW)SpuD$pKBCCG{I2GVvJ|rlX?dZ*Epe5XA{#i5!miLYLR>!kG;Z7x{IUe{
z^vOn_f`xKb>cB?}Na_BS`uXPYD1XSApWKROuIQ$IQRv!g)<Zz_q(tLGy@k|@A4muw
zjR5@;xg%+rB>GHh6z1+%7`&{qCGz^aGFal#892MsfE<oF4H+gdm<;{^`(*`Po?p4S
zHWLQDl+nXKKW!S6#RHk_-j4~XfIdv`-6`8At<q#Kz0>QjRpMo626Cw@fTD@mMpwMH
z@_sy&#Vs<<IH`t27yJ>a6Ayq({HPk=oeW9sKdlx^`zQ^G_QbuJi0Z6^t3Ccdr`o!I
zTywu=O2;)|I?r-C27<E`-iHN?w;(IWH_4Zh_x(9uzIX(dDZ{mp+j{T`NeMu(>wc4B
zuW7mnj$t#z7A#rUX~7!zX9e^kh;e6(@!zVmR4Y~re6mgxKk_(omm*`8-QLoo9o+-a
z3{x|Y5jWawwC}q}kUu^=lfK{u<W{jBjq*1HW~ko5Bv)j>=Xn|DS>CJpg_TS>2puMw
z5N=#Jz!z|KCnZ8ee{|6mT#Xefc~u+lOaI%gfWULdtR^>_tdTEdXLfC*C;J0yPWz)W
z4%$c}<ZZ%M3p%}iWvdpVmbFW@!nayjhS}kjS@$mo^D=i_fM7nia+@JtAD*5IbY|FI
z3MHTlx-6v!sLr|nI#|S74QdK@!6neLgEiD#(Ml$Yv`l5(Gnf;|>uc2N*Bf%b5kc9q
zk0mhb6{L^!f?5$x8LPieG*9RknDCWj4-|d^SFD*zT}9Z<_aqyGJ%Ne6<f}zSNIo2X
z5?zU~_fAuJF_se5AbQf|(O2wuR1ZjSp3T)7H1<)qWNzY8YngIO1m0L9(U%~9&B7v)
zGywFkMf`~egDj>yW$*(;5=oMh|HIvzKSKGw?c<CoGZHg)vW$HR*|XMI#$J}}WG`#7
zS6b|{XU$&rWXm3@2-%XNjUrhRB_S0mzUMu9KcDCMe4fAH`%ArEF>~Mdbzj$co#$~L
z$8oX6oS$*#rAZq~m8@-z9}D5bu-RtB`V3~JsE$}k9-B97$Y&54?v`|r*ReCgxduXg
zjdzW@tUB}hOzm^@%wn_xNKO9wc$=8J7Au7Ud8~m8Nhj(`bzV5Xw{<F#)>8lriD`}5
z`;AF@(X02XzMC?kF9F+mKrya>ErIpWjiaI4JTZlL#|o+UanY+DDGlL`q%FC6*uzkF
z?lgV*7oa5d3La3AzitzM@@(sX3^gGg|NCu^u?A@{iDIw*&wZE}k$Ufmml3@BR^tHC
zu+olz3PCd?e9-r44euQ>YjD=mul&-N$lg8bal>yyEmrf-@2Q;|?myg~=00w`#x;@c
zSPeaxxYwj<S!_>$eNnSW<26u;%1S>85=s%d_JdMO?_G<(5bjy~Jzzhpyxf<r6Gq_g
zN{FH78Sc=#Ng;6U${a)A%8dg`R0oFW!69pNIy=P%aZ9nZzaWUaekQwnHY(O3E*s6h
zA0Q)0N#y%07?k7Gn7iFL`2f21B*Gu@hUialDT;BS&&LTj&i#ND(dU?QPxEdJ3h$^y
zPn8>%jC$oz=xNQ;_wa1BqjTi+<)7nyhT3~(yqgiV4ugIe&L&uMU$oojCpjx#8ub0H
za)Arap=5%7-ZsGXzaa@=tc-O=G?4w_6z?{0l7rD>ONzrqF;*u>L18k%cEG!iGAA*u
zy-w0Xms#ilj9f#JYtD{QSXeNF^2%9;IByUqY0am%?WJCWOgKY}Im4Q)3ZzjFD;uot
zp!1sFh&kYc=@8lqyDXgTSZ9-u?S<_o&;2BmJ)*q44xpebcF>uj;s&ha&zl=w?36#b
zgiF(D3tjq>KejGio#Vv2$^jcPA`$aMxF13Hbd#3=C`Z;$G@mcn{G^4P1sKu+^ah(*
zf90&&>?(Ts?qm^4cr{~iu04#;2_^lj?+StTgNE`so?=qy1)@{rxY#VE3^}2^?}IfC
zwNrx`U2+Vh8Z^7w@1;gGd+K(&n;z4#w=raWU{7`@)O(>&ucS_JIAC@dG@j(JJ;_S|
zqY^|yrk*ka<wLpwvpu3eBk;mA36z#F$u6TO7$ug~n0PjogFdeVltR4*FMwAy`C3$S
z<{Z=+M`vRBfd|vmC3l_j%}@U)ed$D6Xv{Nan7+~tt6vAJn_Aemq6A-CDIVH0bTyYX
zG3a(FlMew1jHlb>WG-CeR7%^g`6c3sSBWsLn>SwfbX%>bz<J9%@rFguI#&9wS^$RA
z3L>3kDutp?Cl-SFYj4W)ES>fR`<)MSiU6+i`Ap6OU*pEPAW)^11sGJMLDheKtk`(E
zkNaYx?DU%`Zb&y3?#tcvdXr2jd0ASa>+T@ZLLL<@O+vqG6ZBgTdyLK@l>KrO+f+~@
z;shyaHT$+#i$*>WE|>F}@!#~Wd%gjoUkwyM9U%4-;eBTzit`n;5ncD^4$PNw()hde
zqdQ)a`{<QR8}e<`r4^2i*<c=m$OVB`odq-SU8cSy`ITEUI7X;Dmjjpl12l(zznj{g
zNAnx_u%sgw)`S=AJSKV@ujggrd9R@^KL5(r;v=C9by?2FLuZ}x66=B^62~OZ0Go!;
zk7+cCl6aS~D*!E2L<d9W3$4Qo&=(c#UjY11EE(6bJI{iE=3F?8sVVe_6(II$X51gW
z^%LxeczA}XsSz~2O(t$zKV<~kJ%BG<xPz4m0i>R>Cb<h)Td)>A1v=4@c6Dg>lc3L<
zyD!yWq3z59|1N<0*Df1iHuN&_|1D`cumkC~>KeSdNUC8iF?E3?yzBaq>@%F}%lj7s
zA`|)4efW2^c&;czDSNQMK)76oLN^oCwMomgh=-DZG18{`O!3U%ATN*`7({M}okGDJ
z1&tj><}JL2PtR^T49dMcNxn7xa0<pmkhL6Lf369?1i-UBjZIYv3`@8fbh%OneI(++
zPBA`WQhWBeBOoh6{jq``-i5fOcJ^~|nMUL<5imZCf|=%mbJY?mL0po2GA}ffN0uoG
znk~geMwZ<+38SxWdZr&mq1mC+{0$~Ix9n+JuNSm3gFH4~WCy|UYzB-F8)V5VBw)b6
z-ofF@^x}%CII`BQad0zx1QN`2VDSb8{yB!SJ=6SZC`b0!xEoJIVlB98Ooi^86e14}
zAv}<A42pUCORJ9oK2IZ*ARGSVU}XXFTucc1KQo*}l&>KnQWjT6LjOW_<1k;p`$VhX
zX&L=s2eAj_yLZ(I!}b=cM`C20Pb2^<io-R7OTmEbX21{KB)_Fhf&S8F5$uD~L}>ys
zWu#;K#~mPbzaf;ct;g`eT|<atLT5Xs48+=9@y<G(V#3sOMicp0Cuq)y(7wrdQ`s!?
zk*hl7f&hFRcn^9hBLIKE1cIK%npK-QaxIuB&g6=>9lQv@@FW#(7kGFHMJe&S{*dFy
z*nn&;3gns5BC07w3pCtuyIe$K>pQ^m-WqCKn>417pGY8w!mn99#RO^#VZ-bcSMv;H
z8W8nhNQ_4(9SlA;v%F%+ek@q`()5XVBP~GFhV7g4dE=5#3H;R2<gWmm45>%ZVqqxK
zrsnhwQ`f%z&5s!zuS^l*alFdmYn2`ngZze7ViLe5>kjtdQVt~ayMu6cFMFoNuc{j4
z6{H~x;_GP3WUy_E!2wwJdNRfy$ZG~IBRygN11m?aCKhJ4x>vVyDb%&WLI##3?30r)
zxcg=FgU(jvlW*k@52X+rWMQjNPb!~Z`2ZT@DFXjh<UYgakl8!&FoEd5`<64{En^~0
zfGXAc4&LM1u!S0OL57V;;`jNd4(F%$@Fi}rxy}_L>oVmZ7a{sO^v#N}wxN1B81zKh
zVZ!eqoU7c~_xIC`Z3Pi49o@nxtkfgN5*X~g`#H2yMu6{N(<Nm_9@Yk6a_R=}Fl&c^
zRAbt+Mo`Z&L81i{kr)iRt5klS+<^)Kop605N{C3ndLENZ!tl2kv;(|sU~r)sF)&Es
zgT-w6GH26-<YjV2(Ix?`QPflL<99B&_`Io|gmUK~JNY?Mi1)BB?>O8flz>4S;*E3i
zUZ8Zp$mC=VD39T1a3|+>UhAvWkea(rIbmH%VT2O6*AOuIdIw@7XtOz`m{`0N-<?u=
zLeHzMWm^moPoG%Xd*FMNRvq6!zEKajA^m&5Wi#H`WKec%V&|^GD}rxQ&Pfz9%wBoZ
zvHglr@~5a-h<sbIqR1;c88`xj(yO%W5l?yuCD}jIoso&%4<eQ@aYEVs*w@7w=juK)
zXLSNI@Q6z<t$!$otJy?yUC{1#Bz8j1|18+UP?lCEKCx{-^5K7~i`+uxW){JUMke`G
znDokDn;@X*yM7XFB2SPxEm*+OtoCes*^$ar$$XZy8$B<<L&Q%}lhPoXq;|gZj>BUj
zVQcuB>s2eAZW8yS<1_Sa7(_^@X>_EkHdt4AiSik{A$M~dnW4;+R$hWlEU!3nSP>OJ
z1kR=LMxfRVx14$R5P0QnuV&5vMWRw1#IzIThU^O7R$%D`(^tPw^cFHh;~`Nnb4^>8
zh0r+J;mN9k0)p=N`r-=!^Y6izASC6z?2l`&GVQw6t{sDifkb4gn3Mp}03!C@tbM;l
zwn`u%(}WMjb?r2?fl7EXiZd{<W!y5l2`XYJdQL)~mZz>=_Hy$xvZWPj!JKXV5V4Q@
z;!-^*sEdKu0jqOC(fAA`FwAgYrNsxlEHu1Ko+gU(hHJp&kNrAIr@TOjOr!~aqfDOo
zgc!80<2q8(bzLKrAb@WSgIF4Q*j*9yTp=UPUfauL^rG=5qaotQxD4cAHrI8ee6wOF
zAv+bUA6IpFDCnXXkpVcY%ocP88(aJN$fJkTfv1`oJGhw+JNmYjO*-1254V>cSW|12
zRkbqdCdZi`?2b(uyy(3-i5Qs~KpzEu<ToI%&ngY(m*9oc&v~zl&4O+X3S<x=BMkNc
zp7t}?;5}1JG=7`Z?>@{$gg3$8zW_pi;r<FpJ`qs==Np?{mkVbSJW%yLeupk7Gg<Y%
zq!LeJy%7;wRxr5Sr4K9sb<FPmdB2lh>mmPE<9%#OCm@#Gm*xz)Ktpst+FY6ZlM3Xf
z#6|>uQrD2;zAEoA;-`#AJPuMdnXvajh$Qb494giY<kt$!cta`Zn%+mWOCQG|luGrx
zW0kKUc~*)D3=YGqlk%9V=8C<<ZSn90tT%KAAhdsZU31`}#mx8pGVu@hd7F^85I&*j
zrFx9{f*_`bjc#eI9)H&-fIP}6hv19WNDT`Bj){e^iiDxXcwsjZy_Mf}@jF0)up7a?
zFD5z%+EEzQ6k<F^>JgiwFErHN*aaK|#yd=Elc~yhNIm?Kxl{P0nWe0d(VVv#mLxDg
zZUM~v21eVn-TLwFY>kAZ16biL$MO{RvY#?uMeI0<YY1V%{-%Uz{p(`E3meB!()>6y
z`|5i`1Xq3lyO1=au^;*KMAQYLPBQAUKQa*!HmDwU4BDMB%&;$;qz;FwfFs1SmmGpA
z=U{c3{92y0z3X!s(2?u`k#A8oyY`<m+Jt;ZIj`5=L)p@Pj7ArB@67;4qXEr)*V!gv
z<~9EyfA}C4>RgB(@S1^EG6wX0z+XAj-F7Dq2_1p$O*cBC<UiNHyZ1BwsHud364Ze?
zGjk6%kePdi;Rd;#y$8+?83bJ8R>R2L1Lg$)IEy2NYshWuM|L`FuVo>!WzIZUOFs8*
z%Vk&?Aj2(+w8gCpAWg9TTh_#21>MZ_<H~V_65v6_3KZ^TR}$L2?H`b5fOnnJh5@Ld
z9srYo%{60-+fp4C^kbwVDGPu&S#~Sq1j+ze&g(a!8LIfMjNP>f&C#*@UiBd=WXpS5
z09uc1_$3*^4W1Bc1=Zm}`0iPIneg4y;}tZ<A`@11wTIRIJTFEj7YK8L(W-*v^qfH#
z58)mfJ=vhD0e%xkW(#_?rltAFfFOGuB<Sy_Y!aN`tH%q6b6~u|brg29s!@coFI2*D
zezTDr6$EZ}f)xXlV$T?mloBvy=}ns3h}(BKgCGpZFc&DQ$UZ0i@vx@Xp^*>ZWZcpl
zVC@Zjmf80&ZDKHR1P%|hsiCn9EP$_IYIr8a#QCD>BjnwQm(de$M8XE3`kMr5OfSex
z)wVZgL4qpK-V0FRddJS+vo%#KsDIPG7=zS8-dVtpSBXdj7DSBJKy4j)+JWyj50O1#
zjMGCR{+OjmEc&w{Y_E*R6~n)d0`tOzrJUW>%&pi0Nf?MIKjBVv2^8T_5A2pWmST~K
zEo06J0PrHj`7+s!bD8EA(}oi&+MCGReTmQ&-^w2RzzkYT%$rI<V<zNg!O8|~9wx{W
zQcRz9tb#mtT80n`-<5r>@8iBrD`Nvjmxu!_P+5tYp{$a28;EnAa~IhpmcRioq<A|2
z*g<k-&yE#HguNV0AXVR$sR2`eFS}s%cP3Er!W{MIgB1ragf24iJ#>rmE8}KKAk^V)
z5y?oL1spUQ?5L2agfPlxp5KDS^L*bRAv^*HNwxI5vOVhBKmdciRb9mPHu*k#;)sfm
z+~40<@_I2NV6)r+9;2cG&rT#GY}kuI_J~6*0Shyi=|L9cf)DmFLBTHw0pZvU)ts-R
zh{B{x;q+JX@DswF8DW<NB0($EqeyrM+8wKH@x2ykfd2)Al<7f5Y-szrZx5lJ_^R3T
z4zl^Aknw;q11N4VQ{{p<JWT5`$r{T$_7K%!+LX!kyRUqB^m}JD+{3psCwAj45T)T|
z^jzzNRP-+tWhuYzkrG&tj{YBXb%EB#6%GaOe)&I7l3AMD$3~GmOG9$v`Vb>8lluO5
zwdi6YUr78OKn)BV?wMDW=35bouv#zg@bJ{_r3N3?M2`Be<he`ChZsxfcm~}lwbMgv
z435Ie{jt$OV9+5H_{Pd*fAVaUwKZ}#TFByL%E<Bturp2dk5IT7;~2;-#wEXksN}dq
z$_U`X%|OJ7deFBD9j@*>+`DUifME<Igr|q|+aOd3sP!*^D8|O`G4tkK*LKs#g0RjI
z+S&{yur!=MC<uqoFe<`y>n4;yv;4sZ>LXM!%C5C_(A%2=>2d0{_!^|H0och1X5aDX
zQfhX@e+Ei9sxO0#Q#@kG%rB&a*ZGpt{;MPq4OQ%~UneSKqSUn^H+ctc&AA?KdKD2#
z0EmFugotWiKDHwS`4j|>{(MJ7U=4Djiw)V;fQj@Ebqp}rdLXNKt^n4?eg2vz4TNjD
zT8h?>fH@3WA_i=oZHsqK1O!Z@Bfy)nEdR8THJi4ek(oin;x-UX&;A1~ApE@3IIIkt
zzwNyg9$%J*t+HL|@$)Kl$e~~{@OZ0t>1#Oo7TP_Q=kO(*#h_s@TgKJ=K3}&NW~dBI
zcbX(1D?^pgfp$LT2=sZvH~G{2FqyQ(N7S|90%!a0K?yTm4^DMNe?;2o1XVz1{v>Ey
zA%;%Fhzhpwweig9><d&cfqI_vYuY>cxa;}IGG2X&6tZIi;rPr(WxA34kqb@(46&n~
zG;<Met$=lZne7Wz#g(Wxe!v13+rn5<V3}@qO80SUri{R2w?P^ye7HaM)*dA;=)^pS
zzEE<)HFGS)(R+AM70<D|<*&j-X(%$kO2>sb*xJW^wm6k~NBc?>kyS2aE>%ajH71#K
z{kB=21HjqCs^5n~|9)`*orCPVvX3mTb2y1t(<SfQ7@v3Ms;JzlMs&B8oE||zC(O;y
zdOtg4e}4lUADk<CEV=^s32g@c`UdSZqDB<7uu}z<<8p@!RI+#nJM@_Bb7r-7F)wZQ
zp$1?CB2o(eh;uf3yEj&6r&k~^>pSB(XTsg&;>3|VUw^!+Qi3wV6YhGjk^Tk!p`3At
zW!!$N)pLVM3GXa_F6iGXJ@>=^vYK-b#wP@<B=y_AFK)zt$~;OPXn0i_`^p)pPl_FS
zj>=HWahmzYqLQ3A-;X9qM8v>x6d!gN6~Q0{W;<eT5I5Tu&%EsSCr@#J*$O73K&SGa
z;`p%#sTWA|n;BdslR&)$c12(S8*{xIR>t`ZBt}TZ@KarzZiFv5=a_H?g$b!N(^cLF
zax=zCjUUc{vi;nkkZhY{uw=KSuPp2U&D?caw_$;ptms~#lF)ebShp<TYA_&jBMIF|
zg#w+BJ_6iYM1TP`25`m%lQciVGIho|c8K{8LKaiamZbCcG|0jG8)F`(ws^WQLkQ_~
z>_L}6kji09M4H=(MG&xh+YWLTe+=6&2M^Zgoi=Jbp-&Mw<5g#MO&(lOUW)NM<1dF_
z33`TS1+|PL677NK2@YNLC*Q)3_)S@#>m#n_$P3%a&>EN8R^_yHe{2SZ^7<^o&lD_P
zQm5vUgWeT`iNdhwUPjsm2S(SPhe6fO#JAwCb=Tv?-YGq$0j~e+`iS*}Id4rSa(#0)
z#(?*BesPPHrC3C4-KaGK(jv#jLnh6xGK0scuDbh7+}B5y8GcAQ{GL$Sp@k_F-MSvz
zuxgp!JcMQs`ux%Zn8=7c0`YtRB1?3SHjuprncnl}>UFyS>wrk*l@u!usUi5OYvex2
z^IWv4`B+u?K;#+!3d}oAnXVo^dyKxTt1HK|E7G^^C=9*#J;CWJ1dr+oGe_JAI<5Bs
zjLrmJE+ia!=s26*@#o_qYkOv<fY{eGS<x<}FkQI@yYH>Mh2k8KV!M6==MOY|T-Sv%
zH)QGN=U8v;E9P~WuvqxA_2bJY@6<!JE22IH;#UxNvp0n`AS6O&2UwpnR4n24x8LYE
z(O}nj&k4wfG)5sf9ME}^(JhgemfJqTHuPY{FrL2Oz9-j$dqMQIv~^{D<Rb|4)F9jk
zmvO{|2Q<PbfFV6jn=8mv#78%R+jB0c!H)XwdhGkK?iRIv{%```vaUp*yDG`ZT)kFt
zxdlQnVuW@HeNbKxBSRBh2asR(A3&T`o(wc<Hoc6A-_j>IpvTH9wG2r(P0(bpDt{SZ
z4TpuU+*k>T1}A8hfj>@rTv<zY`n))(ZV>nTp<XZX>&YlKBfo}n&8$Fo$op*F>UtE{
zDL-Jx;E`o6Q_zj*|9+e4jz|Xd@|l-e)5EUlr^bToc-M&jr6cc59|F}z-g6#TgHPSF
zoAus6Tb)zznP`=&!|;O8?-RJa`ODlu9n3^KABVR*q*B4uQ}vP>CkDa6#FZPWx6;&7
zmw?BDx9^Vf9Rzqw#m$M(vI*}|(dw{t$hte{Ebs=klqCL5Q1T|O5%v))ZrizUvQV9>
zcohMO1$H1FA;0)KUWl~sBcNXpQNtkiO=yF`&_pZqj#$8B<_>y&b5um4t|%#G<haUL
z&@~4-PY~)M8_Tl2@f9K`W$}CD0;9x25Fn~9>YkN`{1q5v_dz@DDn0uJxP2*mf4o#j
zsv}W6;90h3^X`fq_Icbf_*4<#3qf&8JsRn~N$j#Qcij{c!qL-Ppj!Qysca}egOix1
zEdGpqy{QJ6c;xN>?sV1I45HdKDYi1L!>`~60>=9270MMSw>ZST7DAYu?a}Ywpot$%
zA9k!sKTb;`5H{+4R`p_@e6?0VIvzCNVcI*2(WEstTpWSd4JcFA&u>#l3#V}sPfdzm
zldoo2s1+Qj??(?lh62Vj$d2zx(aKO`(dukAre!DnIWdjtBlLh&e!M)PH-qzhwhCbH
zde^=~XgAJcUUsqxGC4&dk7+Bule%vFbl&4*l^gvf%_2uG7YQ)vGEg#vrroW8tP=#N
zGJW9-)jD4$+GN{{1oaq^LmS9*11VL>-ysO5g*Itd5m$pH4B))V&_5BnfmHfDuq8rS
z5Iyc5&1_*+=co?G^bJ+rrn?{9%{`Fs032doBHrbWNIE)A133>uWqTEoYo_L0j@Dzn
z3#vl_(v{s=@P9sp06f4PcAWuyj-Z3e+g=@5bv}eND*k0PyLWdU>4d0msf+%PfByG1
zL{%Wg#OoXt*@a&JYg~4J97_$&#nlDsGV(Y5w=0wXa}}|L<NjO7^nX7Ma(Mzb;-?8f
zk;cDK`X8U}iG*(4`SnQC|LcPPxnt->Ub|vUvs)Ma&j-TMLyu%;gkGQbe|+P=FPc$D
z%B}zZJ}i+JyM)+rgK-NK9%2gqK=WQZ0qDcQZL{SefKd(uf!<&EuUPmcyw$Gqopc6F
zC>UM*(~l|Kw)zfahh8J#qI*BSIamNauxAhgjF(Ft$XEk(L4X*MAt;jYuf7#XIgbLy
z{RZIPvadFv2@0FSjUicX1A-75=)i{|n)r8306F>NFl_c^dQd8O{JhxEbo?HoK5OXB
z8%1C;<gKRghXYG{6@olajzZ>tAWayE$%vWbP&#}c9U7dKWv~3%Fw}#d+AT<GuYJ!E
zR=Z9VMLGwA4{j<>GJZn1(yh|vo&T9|a~g#}yxS;7_Oqdc&5%w?YVxTc13)x73LN2S
z%h%|j#ZkdOpTp5>x*q`20}<Fbw56vi&8rZZnB7yLy0$_NTYoy{6VIg;_1A?#NC!0l
z_R}BBlPZ6dL$@zg7lZSb_W52aY-u(`V5W0EV&(?<$ODS?73u$HqyV)jpQ=`I`C+~C
z4q&!lVEN`J)3^X^*cBMJ7&nAzju4~C_d@yrfP!!{F8wM{kea;58vqZCqB~`Q6w*7$
zsR-qUSfd8ouZ59xo}kP*KM#V=%iurNeFyw!Ltq;U1pyP{Osy*oc@0QCqAIV;AWY||
z<nD;rub-zP9Gge)Gw5!8p1R*vCj*tyYqavYWy8{581IOK(EE9h@c#gu)u}nShQhu~
z>VzWUGPJpWf%_R8OyFw8bL7%S7$wye@VXkZ0nxkQ%-b&aRxnrlJfg!R;WjZO{;eLZ
zp8^BIu(Uo_nSTVi{1{zJF?Rm*RB_SW>9>r!6Q8HvRxL2Y=5hpEH+MD2ZF)UVH%XID
zNg(fQ*_vonke=i1GB|cdi7yO}DPZq%Jo|7(QPoihvF4$!*Fc8R2FRq)BZwxKY711p
zv^`No&^iOcBXe(tFQzJVc9d~2l`FibX8Y-J38ZzlN1$^(@0A<a2OzcwfMW-PQiY{X
zy6QTS84XyT<`oa<!fpJ91gNVa7aX(eK$(0Pjv1NhiM#Z;RgY;@=E=rSd%83S1EknU
zTi#X_y{XR=_AF0ve?43O%;zb#zz9BZb;7Bt(5&ZA55dz(W9pU@<!ijfN#bA20bB!D
zRr5p9qK=iHX79A&t>BQCwiUpyv4<S)pI?PzKf3QcQiql_n7%j`PXlK3066ilTWOz8
z+t3v{7=*)BO&30He94BZzAoONzh5CZ7vr_}c}3AXduH_L-L1PcBsDl~NJ}Ry$PI`V
zE?dV+wadau0JP5lw!%pxck$=a$`}=PE};|FrcMUV$;T-g@ukYTE3y&C+6YMukfAdn
zyZZrPInQ-^y8bg-8FuGKr#XsO3ImbWVZp=X#`4?iAf=E04vKZ}+HDECbG9`^LS|Ly
zLH9*0OeqXH`wCPGvq<q%_PQA97=SPV65%bF+=rm_Xwf|w5v!0BjH*`e_8bRmdyAaQ
zP|u6UtHP&1ti2c?au*H_f*FP=7^^sH!6El8(w#iPNBt8{bTO&l3!2DD*x^gtL(_{-
zlHYrp%%9@)7q2=->RZY@v8Db$&Ng%WhITT*D;78F_nv`vIRIS}VhRJt&B2!^07y?r
z+3BMNm9W!e87Orou@ivpcwf*KD}H5RTLViku?c{&Hs|K-6p`=1U$OL*;4=ya*!dzb
zCuPC`4qXJ2NrhFUn*bb=0^PEy$E%i7nTS8yD$jXRO3Hy+c49*3q)?VX%HXRGgUW@&
zumLKgqE4=0-(3V~yAB=g5GTNwWU|r>r4o3T0>1StZiBEqyHkH|KHSo9XJ?1tI>m<-
zwKa?9(;K`VFCO=^RK!$lB*IjL$wW)V#HOO6bWfZ>rrIe35&c+<_1dP|!};F@Q$bl5
zo=vC(%;(GA`NbpuYbYnc{>~@hY309p+w5gIdPw=M@8;r2$<^N(b{JGNp$_7z%VX`%
zAZLcvL;g46^oPKuu~db3&0bf}_cL<?TPWq`sVCQi?tOXx0G|I{ffol%O}ZFTt5L0W
zBOK<W{8Ecy@MQiQi)yQBv-HqCl<jyNSJ>pN7tOh!ixlxWs0Yt)Gf=+y{R=j!bWEXG
zYSu$jyzhfh?BIuti#im7M>+6B9&o~ZsV#Fy5ali;WXgxZpgb^+PyGOfC}SQpQYX-$
zV_gbz*(Ew8fByY+sKX6(DV@e>Q)Y^~FgijeK9FMBNQXL%gF)qpn`InFXImb}&nGp9
zSR;s6FpA0)uG+ER#!0w*I{B~{Ov)776Q@TgS3YfPQFq?_{$hl(@Xi(>O8S{%Bwl5T
zPb|t5zwq*`&bFm{-MT~Y-yFL44w?GbtZz35R!2#w6LB~+nv(q2I;nGnvXw^nGM0I8
zNC*C#ASM>YY(|k>QIVoe6WT)Hruk-X^PGipb^BGYdzxl;=KGET_9N3~|K}_IebW$$
zSW22tKG~F!@at^ThW?jrT-~$ZZ+O7J@BjH-bhsjZ?9q>ig!FjAWlgEnKXi&8C}u_I
zyH1S+N#9Cjr3^9pu`Q81)8*MSeAa1Bcy$qXbqT}}fe%8z$H=_;efXQUDmQJa!!`T4
z>-uQ67r<>>YZJ9FG)e4$YS3q_E?lwMRWbftv}Iht>xXGSH3XBUsv~_YoJM1YA{~4#
zgVO4}VeB=ddCu$RlaG5`Xxf6j@9A=Muq2+`{f)n{)Q8h*-Ko7Fp&3|4agH2%=;v~V
zb8cD~UG($Vu*}nYcE|7E!;}B&j}JRvd@7nm^S}J;-+=3ip!g*87drR7iwGkSV2b=N
zf4L$9Y3-D%@!#Lmf4>e6%@7~f^S^z~6&&I7AFw>p@0o$}{UbD&D^1JdZ{8|R{?El6
z<xtMHuInq<)&F^AUjFBI96Hhj4Y<Kj8zT2Fgcc19W8fz})PcGbj`rCEM*r1`9690L
z+b(3G45ZNA>45$cAfa&D>3frRZF&#ln)*ve7;OIM5EN^FfVVdRhnj&B)_}#Sc5xZ*
zlJ9^w_;9NPF5rT`jXvDHLt-l?mX2^#m1L8_<Fl6WpWEIYssc<PWIajskTr0bj6kv&
zRbcW~VpH_r4Jo0Dsrt``%|&MYzaNdpe8`gmZ|a%H3GETQS|l)vEKoToY45=zBQB0b
zUW20w`QuUl>y^zh;tmh*n^#$Ue3G>T;{Ytk$WF{p9P>$F$^bZpFafG(EVA*)x+1V6
z&+yHU#lz(N^MS|+VlmzdYH{c$BgeWy$L2L`R5s0?-ggjU`+6;6e^p$BpZ_;zi5e92
z(O?{6amOAxj!k%K9AaL*geUCoAm;o)wTVA>QiZ=nk<pp_0bMz`ZxP>(1m=zVs**4i
z16qY=ryjnzWwZk^|HQbP>;vGWAu#amXn%?SOu2=)PB5v%)2>?HHZ4OoOr#T(NrRZH
zFOXWRgjY2(|Gf|k^m9tX0_C8wIwz>|;5@Sz2$#^_z|;SV`g9T<k>zJ~>nWD_GbaCn
z9*N%<xUF9SLAw1Tp8245!*u3B%e1qK&33i_H7O=>R`&uufId<m|53aiKuZ8W9L4g)
z#3DX02<jW=$dC>XMsx4Z|8P4@Ar;~8IHEBEdUO$$RpaSnP#JKU;I9=@b;g)!k=mX=
z5odi@Y-m`j1)(Nh`vXk;uEYhHYXOyr7lNR{b-NaCYk1Q5D^{NWwdni^E09(g-cV&%
zgIfUui{+slbvaJI#K?8>lp^-Sl{>TD0r7N4J*VUZFS#5pc(aU*-xyL^F3j?ug9rNH
zfI)1o06uiSnbd?C{{t%ehe2N~oSE_U3&2%$!){xi+2sANaSx{mui^nN60lfqfY++~
z4KU4z1D;eC)N_sh{Z@)FG7}jUL$ASN;yDiSQ&*MjyaphGDnKnLn<=oiLfY=`cn%x?
z9!`)b0|7WVNtKa4lj2`gDa4};u8XTrtAYomjO<FY@_e?79N%$W@(d;7`u9<eP|oeK
zt-C7VV<L6Gw1tpDM333JXq(kLlZIlUxR)1K1rzCO`YcNL@LE9JlRw^yjb~7ljygqt
zFM-+=E><%ZW;HfHz#>r*_#+ShtQpX{kjzTIr(n;JqsxXm1s4%k$3e4-6&M!(<2kV8
zlxTYFUeX7ue~CR~V%$^ukdVTJ24wappnb*H1Qqd*xL{)51+TPnX$K&5AWB}3=(Q97
z*+XLS;f5|&N~9h|&e3M_6fGj8aNr^vvE1}9$1O!94EWn{I1dv^*HSy2t|SV2`rV62
zr}|ZuC4yT(rFZc6(%kUgX*=?F2H=PTiD{8&H`qCP*;sTu0^C{k|IO3}eu8q?QmdAO
z*l6*B^mT4DWdIJ$EAhX**D#KEuZ507SoEO#s&N}`WGs>*xpquKNbC6iTPt4AK0JOD
z$Kx{QyO<`*A$emomwc<%kW~;ZtTLG6n{?S<j;0a{51}@4kAY`n7#)Jb!F(|)Z-oA&
zrh|X942*LfS9m0ya_mAToN3TfFtRg!79dH)5ST$RDMBaF=_)->vh)&&#VSYHEXv)P
zN<`1tB!*$EB)BI;tQfh7OcnReY+0Vr%kW~AYr~<_Vi99Pu>QCPGtL{7KG0gUhR_6}
z*Mr5y0>}4E16>dsj>LV4=Ob~&7?1E18%cFk26??~KSvD?)!4PkET(5k*!Egh`%K;&
zTpDqWw^6Q%s;R*FqN`1sYpvTM7`YOkQzBw`s1*gX=#;}C%+*2TE~oy)zD2$-jWVRh
zF22#?>N?#bg5yzCYu-taq6ilRjsg+!!KKYLAUOPlj(q*iGt<WKiVTTiJqa+2WWz^s
z?#Lm7$T1Tdngi}A7c0KJ8*ow{uqpQ6lo)2NP#(-Ycq{v2PH{YC*0P}v5nWw=tq`?E
zqAsLxy|r)0Q*2f1gn|e%?t4&R4E6=I-r+CUg<=&8j^!)*fllR3gbc3+GA+V)G%3_6
za)yzE`l3)fmZv}MKsBy#I~2?2CVa8TKU!M+xSPO2wRNKqL-BceP8pIWm#xRp6*MR1
zL670DZa<Fh;JJu}Ca(nuF@_`6B|<H`TY9Gk8rt9&L4P<CIG=BL1gO2BsEn=X?@DF4
zj;8#M@0|Yw{RYq;aR2->ACR6-)!N<3c$|4`?2|xO)Bz5T?m;Gbrx{EV2X9zd=3ilD
ziX`{=7$zNRijEPvv*uMcdwNntbAx>1&X`zO(TZeP`R%1mQxHJT3K6Lt5S)OO1}&B<
z2(dJ}>7tbsu8SJDI<{~ysyy@l{wA2yIK5I{?P2^nn{k41rVh9gnB!k(C7F`t4##AQ
zatw+X@MG0Gnd?Axn^&fv*^ud_L%jkkH=eHZV(bW|uL9C9tcAh}<(wg5`?~#ML_56g
z2yx_Xvl)vFDAKZ(w(9)><e=ZD*T$~eN;-d7$mqs>D1}A&^~)<{D$4lPJ1x-2VRZn`
z@I2d7D%K-gYg$Ir#ed2vv4qIX>gfLfe<-ZNuwuMkjKRdjB40Y@@!}E`?`j~X-H1m{
zny{{VY%g*`?k6t}MXLdmVDEklt<rcp0-xh$g}ke|`lX~J^{;Ak84pVGjz|Bco*DJY
zcjj}4Tx6lD?;1}i=)IEJq>F>UzWt0%6|LkBSTw=a(Ni90-G4yAl`9}k!3QuMC?n<@
z;zk710<Y`L@DtGZj{)<>521HLXmlN5$n34laosqSD*Bn{;B8S%Y?yp~05OolZ~x_R
z&q}WIM<hqTe_@S4&<;4?r&Dq8`|RO2C!k*c1^`a%kL;?5V;{ZQ5$%%m@>q`<)BP1w
z<glE}KM|225FLV`x)C9n1kT+$mQ#GiI89lZKWzUKh?|=ejs>@3h9Q(W_ary5nrL)&
z_-#`PSMirGL3?L)CpI^CE;x>4dMBp+Iam%jfiBN2(aE;Sy=at8AuzbuD<w$p?qkZd
zBXDrSr39_59Hr%kN;hW>Y+WJMc?eL@<Uf;67uI=rjm0(jd?g))Ur}xtTaGJQOmBnf
zl=%+y$qELlCxvhIhtC}l9py82W=KD@SLLU`T*9OM;YqR89o%QBzfS|jqWeY#*u5<^
z<{dse9IbPC&t2DeF$(K@u}qYyEF_w<B1s5d)E@_jPM(|`eg=`@avXyorccU$PPd1r
zNXpBL|8{TPr4Fs0d9~+!5^KFWciWeuwWx7{#_$J|wAz;hy3S9n3vHF{qkIE#^8uXA
zRWS=%YmeOzi}~T!!6*_^$vp@7n!Z3V?De8QQ9VW?5qM9?B`*>#)1a^#2q*TC!$Z~6
zEj_ZMNoNd4rIL7UCg5|4(YZczi$r?VL*Pmv=Y}8$f=r2PUFq>>FiEA;<d?&sr1`?=
z!VFW<rm?9p)h8EH^v(X`1yDjoRC>gX98ElAqfT=Ya;reFEN{eGkl0f32dS@7OS1;8
zN`KKPC7n)!L)9LeZ-dn%@0@TB$4*f9bslj^$U!qNNn7`&Uskw&3XJ_KAOiX4K#B!L
zY%>^ESRd>&hq*5~SLUGJ8G^?35Ke6XIdrTwpoq7F`JlxgC=(&-P!%&toV;hINi)lb
z_1gDvsdE&@K|6^9X3jD!zW9-|%95O=4zLC3YQn_z2+RI)B9P*AQkIb?^JC`~y(F`J
zP=mzpi_hNkqn=ja7_IFYIBwwUGteFaujAGALqATeLkw5X#B&iY?Gr3B30%>C5<yJG
z;68GceT3gid2D<E^aP4v6B&O=|8g|Y7gCrvpeoQXPrh*p{Aa+<#?mOpLd}`>Nne6%
z0VD*Hy><6I6q~yAWAW&;7z*VuSfQ!1`+A#SQPx$eC{sj8F6z08zYvlj^aA~(CNfJ?
z+<5HG?J*YiqvhFWaWAeCp2IYpDcHcAKbrfS`mZnLAz!qEuiGC?bRR`n;PS*KXibEz
zX;CFehy;qAazZ@kYi6Sn*fXU^HxHeVyU#8mfklPmJ6S0r2Z7$8T+DYv_=(lW!Az>d
zOn<3>n4($!ZN2f#I|Ttd7)c>+opm*@H<9?>y0=9_Xe=rY5+aJvM)~(Sfoi8r(cANi
zGLb`GOj<J|-Q*0F^IWNw4_b@poq$4>p9=a}GPmzZrc5;uo&Z#Hxl_*k6{JuS)=|hY
zLa^4oNZBzxQo(>#^S)r0yb2sg;;r!CMKOUPDRXKhHSVEVmSVoi(-*nz`uQOpOJSx|
z^^F_;2}MEsxwsPRB8@%2ZXIErD|A%yUz}!=gY$Bav&N0KD)-g_fcF71z_JeMY85Mg
zBr6@n5vhp^2QED~9HDg7p;Anyb6lNyERZwq*VmKdQ1{}-;j`Rs=`g+QfP0JX5jlCU
zsX>XK>$Vn@vSK3^ldKmP07sk23;<$8<#Kz)wl0fiCd^hRA<fDk=-NgS@A}bSp|H36
z&GyzmLwHi-ZqGm%o^>*K^?Tt<a?;wVk5)7T8W2|n5pXto2#Wo5=0=PL;;vk3N?Lrp
z&Odsw2bH`OAuQYnLbWeo-AO!0x0_xZZ}iuoIRQgpOh8H{9P#nB1UjPhhx&6_iYD$$
zu?P7-z<d$x5h~`3n&RjUB+lwb8MyPeuOK83$hk!{dt;NK8k>ZJkn@MHOW4LJa~$>%
z@ETA$%_gm_Q3|PDg^nYDx)0!#>Hb>N2G;?kLPa4*xczY5Sj68k7qGsxgkzqz3SWpi
zH4*X(L-PsdGaYsW)EvW(PP7K%40z_v94S*BoURZmnJ(<deXIH%d1omSHG<;0SY_lt
zz2Hft4WuYX_vp@hG4s9Qlc2#HCgpbd#U*wjo)a@{KX3QEeHq)j!ymn&Wmg<RVv3l<
z9C*3Yo<6aI^?(McVOJ;v1GCWsq(uR7EJGw@pa9p^26yjDAZDO-VurJl^LnTTD~b&$
z@wSL_F0}O1qaT3172BWB8rJAQ0Q5gXt#cGS(dzzq!v8=cz8)nzd{!CdSgHrHVnqXD
zh5drNuTdZc2dzfFKx~u9@%560Ji9JLi@=6!Hx<qm650yUrXXP&VAH2W(d}_npBGm8
zP}12GtDu(~0K+&fs|bfZI$NfjWbgk4UgF;9Rq`Ms(~vN>14JLYO0(Ng6wfC1b`0L5
zdyCLirI{WJ?w{otm46S~MRi-L;)gD@7s2`)*Qt@{l*A^(Yl?eL12vDkJg1~|F_mwT
z?pm~Ng7<;H&@7q)agR@-;+Up)>8++%Od;FncB-&c6R7E=V*Cvl{KJd3<=@zyfD%&U
zF5e+rC|#jlVVbIkq3bFq=k+AM&ck#i%rmV@@_MlrbNtd4os>_%UWOKjOK;+%v?&Sc
zzkE;GSaR^ie9|%Um#w0&s1wy-E*f8JJ9n*^YNc1$<ecDO^O^2rgYQ4Ts+!gRY;*tH
z_h&a_F7ZkD^M`402b0g3FSsr*plFwte_0__+$uGdWjausE!np{<uw@{s>4zqVTU%h
zpl?H*1C7N7csc<!&;w53E2pRFtphw*(EYHX-%vYnXi42pTiG;!BZzhwwi>F=K&ET6
zh&_|!de_&i#sT2uE9K$1qZ+ZoOsoDFRwqVV#&>*-I!_Z$+vOm3BN1T?wf(E#;Vq=S
zp<12fzqt&9uCt@nqDg6RAI?lI&pE#rL&MAa0QeM%()h~L&dpa=AG0db$5)7mB$_%v
zhMgAE32hk9SCu4Zv?jZb**erd0dg`KSNd68tOELui?E{Qv;dD(1(veV^3F-Fu$a!h
zf_f-tG)Y|P2g$~D+A2M`K+6Rx_W<u7CR-j)1oEz|J^rBl3p{ja*=3>>J>RBr&^R5m
zto5FHsOHZnsVF6G-B1d={^GR#du%-rs9;GutdoUbJCX1(ug%+^bH+I8f?4Is8vIhB
z?-j@bhmZXBPLsm5a=45EkGxCI+zI`cJQEk<!WJl;d00m%dyb%U8S5hbJ_TTU<z~aE
z=%kYUPn`%yO9Kqa1^$F;U^n;)eZ%Z^(DP*If?e#b*NS8KV^9|f&K%b%o~a)I?Gpgf
zQ1Og|;|jjM7+w=Zaz*I!{Sba<Op|(aaK>?AY*AYwzysW*F8O3Es#@QJwocvYryLXh
zr#NmsrA}*jQ)OxNiPqf%{b_3)EUf%L4o(6>&?e??KAfE3YoJ(=`hw$@tiO9~4F7w|
z(Uk+#dfsO-s>o4Bz59}59z3aBmKIr{82iItBlUI?n?=_^<kRdk?~1gZ^eNlZ^TC_X
zp3W~R^+wee<Sk3$iwt)~YUA+c9g4XOrx|ok+@dNVk&OGLhl<jFp_9Hq4j2kYv*6rB
z#R4<#CMeR@PXe>%Nrq_TZA+kNWpJrvei!4SK`=KQGipP0wmZ6suab_bCE<eE<9KG1
zq!e#^r@P_p&MM()h!4uWh2jLM7@*&Keh`}5MO|S{1T2G>nrBl5W_~4I5hw)%?&8W}
z6Pgm|Xtjx|BedN2t4m$(`2(M+)ik?0XkFY@I0R-OBihVpEy5HzDT(1%vXrUF1XMoF
z=U1KWP$V-ZChgmY$Ux0VxP;7Add5Ux^R#<R${g&mjgS^^C;~aE<3My5_Ty2LJ@D(_
z3U$9lgk8p1l~bLEeeDPkhMr!<a&>~8g9N3t9FNdsKch#g55Fl+tp`t_QroAlUeohS
zDq<6aMja5=S5uAMp8=7XHP>pLc1*A6aBqn#fPChaW};0eFW8gI&{Dpzt`^91dhWLD
zrO;r7Y2pTSx0mt!iB7;v^AXUB!fSok+8YCBg^>P~H*@gwYUAptAkI~A>cAgb?;bwv
zVSIW+TCg<;2=rr@ms?A2Yjh%|Rd)plT$kZq<$WBX;%tYG^R&@bHN3uy-Dl8r3eUFU
zBFHcZgv}vYYnuVuLT2+l46{+~KEQK__MD!hjiXoe(PX3&uGsR+>SeQ#<BZl6lRyAA
zYLvxi3e0M+=n^ng)*bCuC9ttYY`zEEWzDM{KRPn?Gg*%vQ1EsSNoJreF;>Kd+$PGI
z9odGq9dd-T6+_X%8}OVU*ombQE9T48BwR*7+8WFlevB_55S(|a5nCeL8$-5(`aNtG
zWH5z*X@yQn7@*g0B<eqOK~qGccR-Um7)o%u5=ZL&0DSi;Zc~4&K@KAi`&0fs(ut|I
zqvxfT0&>R5;t<_$Z@GFT_W`C-oElRxpLa>1kd-?lg6VdOqi<Z<@<{6$$4@9TUf_Hv
zW?74vM&L+%-+f92ooNq<-XCp+&EP!2@d0!@JFajq>4YRS4uE=8G0#_Q+FOH2%cwc7
zO_M6e%|ux*K@*={{sSu=W`2unl;&j(`fnSsdF38&fsrQ+g8FYrMR}y1dw}&>2d^K@
zBHvOzY{}l|gk*AZ=sMZ2AXQm@(30!P4Q(f2LYx4z0JG$3w(d$T+bsqFFyzqHZzu`3
zr|nXDn<H*6P+)fSKeQnQJ)<z{S?Y0hEfP)2-G}>9S;nP$pwircgFBGYa=dfssH$T=
z{RwC=jwY2rO4sBwiLjq~Z@9x&vX50#U@)>qn0>O7H3Y%BO)$9L<RLqSJ$HvWkF4(p
z-~x$wvS*`hBf2gOW{Aprz3vIT3#d*-0k+-`PO*%7m(V0cUS-s<1FldAf3pR&^MLY&
z=k%zVxOr;9nlF%N?CECdGR-{zV?70k@cOqY9!C8h>6~+~vi?AP5z)k;;*HRIo8X)X
z<dBM&>AY&>fRchkAAy?%gew&EMZw?Off7Q?bL0=Wqmfh^V8UxZ67Q|!*GcZt`T}4w
z;(EQh#<E|BNzot=Ne~54*b&M^E)^CoBIMi&=+YFL<3!qFApjB6!`Pw;qR&G32`!}e
z8*IK=A6|QJL%DvMqDEcU^%sQQ(}6|m+%z4_kTyk5(A<)JSqIRX2G0{8TP@Nwm4w9Z
z9Y#?2Z^G7`A#N>GHZ*K7T8w{y(B+F7vpi@I4k&9<TXQhMjJt_6{hmbFx~(lQP;5l~
zg}f{0oTjMW{@6#iwNTDc5tVXj`=D0-sL`9~A7KkBjzw6`GLaJmGiR}nT6^B!3-U_&
zq*2ulg9#~XapGjF-)zFpAE(Z<>6Di4x_X-MyaY^Cd?%RuN0;8146dm8{BDejUX|G6
z#hI_ki+8bVum|;N<Wl}du3ripoI3&0S<}OgW4Fn16s<*_!sBz>D2|T6yuZyNGE6t&
z<p)l^L&$l*SRXa;9<ZG~*f_5CDL+8UHVOTZGl_xH6MSp{hdukYX;DwOKr2JhzljSD
zj8UpIs3#{<5yWf^q$h{8zSk?Z1_@bZ-(h%`*D+)pu5afu2DL$)v^Qa)v;lrsD;~CB
znsCcNI6+GI8A~i&odzv6-1l{#PeI%~XpdL6AD@^1Zm}TPBfy70&>kH20328mQs;4T
z(8m}paOp*oM*xFr!?l8z6Kwse5!Ha@o`X*=t}m#UjzUvnuJ2^7^b(L=2)%WiyrM=B
zgoq1Bcjv4wi8|$Q;F_Zq269|+0rR`F=)1j3r=U%5JbAZk8)Kus5MS{Ze_iq<m`VoF
z#_As3gy%|Ms)lWTCa~N$zw96+%sBKmbvPvCxc%x(>X+pDvT$!o)r3q%f7Wd4F}qmp
zSm_;u--;!J_2KsyDAp*BamS$$h$Or+UP1Z_uiYsqI?V%Fym(^c>$Y+f&e$xN)jjJ?
zjHQ^CuxUc7YhPvl!~1VXYkt!u&nqNLkEEDk@K&V<S>n-5iK-XC&OAayORaThv2|~$
z>wSj!O<aP9gj~MfNQ1W!ht{$&!-UkU!7lMFP2BrKXL!9yVV#TrUxv<zFggHIy{sk?
zLQS?favEe5iZmPk!WzT~2%uH>HY=E%p5o9Ed{RN5&1hBQ`dHW1y*@K8bF^0i@n$fR
zUpnRog^mvr;eB?md+(`UkuVN=r8MX)mQ1YAvEU19g@VG@enUxSQe`1AbAj~;1<)#N
zAV(M5?Jh{dQNCc-SrPE{=lnrVIS+O6{#k)1{5=Kb0V6Ynl2J&EoQ@mx-ygDdrp=3s
z%RqLj`Q0n#8kz9k(M6tAK<lkJ+tDk3AT5K9u8QlF2}2+>L;(^+@fZG49tQ{&WYhY|
zjmRM$A8reXU4c*oM~-Lzu}Q!6?@_x*6td5N6o#AtG_Bxe=LP4H^oF>DVWH&5Y>w%R
z_E1Ae#{@~I?t}O7oCo%6DeYpa=b^<k(eE)#-ycex3(}-WU>lto&XE=Nn3n>q^&UX3
z(4s=lXQt-J04so!YA^yALJ+BM$WbR*ato+Of|=wk+xQwqxP<Q3L(f-YKTdvm?E<wr
zXf|S0-b3-f19CdwW5rDn6eO)1<{+}aPO>S|A&-_Fj76vVUan36=o08YfK(??|2bvK
zEqCzo>bbe64C=in;{tEQ)7HlDxOc^4R|#yCB5VibJ)@Eg>-wrICZSCMDdHb6boepw
z6d}5T9;g5idbFzZ&A%Wog0*jS-R0(jo;-Q-XYhvkmcm{+K<iV0m9a^JBS11mPxMO1
z2vSkW@Z*_Jp#%kCkrqloP3#EO)LYYvL@7&Z#I{?BQd-Qs5{w;>`YEJ_u?=dgD-Rxk
zeRc4_JvQ!3VB8DV*`lwf*tT958U#PH!Zv%nh$)aecvdprS2<77k1zSOont~2)mxxx
z<y19f?1PyAQKct=Lhk?^eq*Tm@ch8@qH}N9egX`sqtXBy81TEasJo)}lJ6YF3!Nde
zR>ii$Z#vX<&|vjdcr@z-Gg7~X@FpUx1rvKFwM8@ixD_hW_Lk<#%^k<?3n1uZm{$4D
zB8ar4j$Jc!d#3uG_H81JNqQ`7$JX^}yDk4fA`xY)Sdwy@7O|Q^+&o}^Pp5Mc0AvyF
z5{H55(&}48MM&#~LsG5)JJ8Xn=WTM6t`<r1h=0*hST>`fiQ8FSa9gzD_~6A}W6u3W
zb48bcFb19a#+UPO%1$6qkS%KMa;l}_<d!@TFG5E5cu}00uZ;!CcZizktzX|l5nd!D
zgB+vfVFKQcP@34nOoaD(570(eM?(0pi4GU(4Hzx&<0}OX9*L5DG<qA8k`>7~4_BhP
z{ihvbfI&IkUmb}YCTJ3{pBteR2-tqt*L=P>N4osDuh)HB5=z6K?PU#At_8sGLAoqX
zU1*f@2W%q8Pei@a^&%0^EJmK!&<C{&n-L%Cz-pmf#^6w**=xon+WQlo@Hu6dr+`;?
zu8H5Oe<tc7Fx3Qcd`?Qtl>V3D3>)F&SQZ+)j6)7bnj`G?p`aP39A-x;{hV^;Z84Q1
z^bO?L#d-O)(J32eq0QaLwbK_&%|)Tj3rAfcEn-f7J?nOschJLdx9r~<g^d&%M<yzR
zGlT^oP!X(Yd$kk`p59P9a7_H_erNmG3X$ZIlq2Lj?q{S#+4|NZ_1yPM?}z$Uv<S9)
z*xBOcVm`JzG`Q+T>7SMnGrJl`q2FNkjky*GZtSP2N3Robdtu5Gavs$r(8f-Grf4}k
z#l3fIG-e#>NaLQRZh!v<>+C+10G|D6x(RZe8G&gC@PQZ;KwZ6fm|^obNW&1hEQr5;
z1Av+8jVjeufw#{X7bUez19*f%df$!XAdc&G<c>&zP9hu&pcfsPNbBGsBjTcLR%@s9
zO##Q8>aYnkztviyb2U03ewW;UFmh6%4R|(1zfVY_DUacREOm}Z3+RfBNf4o%i3@h)
zz-$S7ImJgcNSUJcILW^BY6xN&%O&ls6-~puqUiO!@DV~mU_i`-ULG%%Ca)_rB{N1q
ztZvWvQ{TeiBqz%no9IMZ!A71!1doD57|y2YHad0NMH&U-?f@wMxo&E0p8@tO@Kv0p
zJGye?cEh>Q@Vi!b1rjx3w7B(j?9`RGMWQ&>Xz%aKo%ce(h5sv@T=07d0DQ)&5I}DM
zYlf!d&e6~wbaS$clL!yGKjYl$(NNVTySc%FETb=LiR4^~2;}5`5ySo)#<s6&r4ya3
zlOX}zdkRcwPhaV+16Ah*@L;wN{TuXkj4|RRUsy^#^f=ZBI?hj3eYFc=M8^|hPX!3l
zIlljED8FMoog=lO<@T|*<qpCItYpIeBEO7J;o|w+KsA8(dr5|$rT(Z;g#7YpcZ;}n
zy{!Q`TEr{?5lMkE*W=@wY&b?rZH3L~6zA82)6Y&A%-P^DQ50z*Lu{K9Ob5$F_|ffW
zG`JCt8*u(vLT9CFZ=Y;e*m=uQkAd_~5;<Gkkax)m^nQ!qnNq#36xa9lWPdAQq?5Ij
z7ro0J4ty2oN((Sy*nV}@dR{5jI_^<Dpa!2cGlen)T+OM!A`K~c3jj^LKySwO6RA7L
zU17Izy*=;k59r(62)+^3Zqk)#bw_4^Dql;j%dH^g6C4z~ul^i1^a{v=Ji3llDf?yQ
z3!d~&kYmqSautwmmH##buxny=qWfz^E6o>%xTc;kL%WXs?0_6(jPul0BmML!iW?la
zKS}p(iR)2d?nH!1W2`vG__CzE!(Co9d7QXE;@2lx$8}*M5Up6?U-7}=1v`FALX%=~
z2Xt}cwYI%kPmbFVX;lxvVV(ImD`FLtqL2t;g(=_5AMiUZVU7@XzTYG-Q^Z7r$HatV
z8+cn(S|vBpcW)@1;h2}@7-5nYb$<zmHo?(xS9#mqZ2v&Ea}v1eAj`C^^?@_AEHAwn
zQ%JQIvg8sq!l2NBx=*FU%}oG0Waw(I6UDGabpmU-#e3Y@!PhI%$>eN4L@KM|)bHFS
zZ@^zsw#QFBZWM?Yuz)0c6v{+rah>^;Nj8v^B$e1<S`NT2faY?5KS)Ebx3<1e^e`TK
z(RP)-3m}vDmr7WCf+Nds0<Fpbf7tn0BRUlS4psNwzkr6rNv&BsI;G{Am=E`5&n??s
zmDeb4eFI0z8Hy4CI?<qJ@19@#_@o3gO~o2ATQAo?!j+^vbx4D5HZUgTt!B3gA$sgx
zn(H~BRP1Ttzvh)pA@u4dklUfE?&?tNo@tu-gP(JTt*}ch{Aj}>*WYJTuuD9Cc3z3W
zNRIvwyULfp2_uxW9CESGy|1?ovasX2T_*lcbP@I)Dn(M^c7~%r%91lNDf4;QFtONH
z>Mg14&vE4<ERjcs@-(nlI<!u<!5IpMyoyJaH)D?gDF}V8^K9&81s#lVgiGP;7&<O}
zS`(Be4y*D!(sWz~jkt(G0qDnh>gdzvQxjN~h%gj-tQmKVUDoJ$<sU!Kpyb>4Exft5
zmCClC9mT9GmH1nHbm)@Yn3};&+ym80aBMsgJ8Pju0!{d9Q0I@W>!=a`{{D3oI5oyq
z`EYIn>E~lgVC0t}JaPz*5GF#p2nd=nHoewo>`W285FU#6_w$2Hkx?ts(R57Pyw+~K
zf!`t3YSEbTqTJquR%SRqS>^z*svWr4>aKMH6LsXmH{>J~I5B5cXia~HsauJ$jdw<1
zZ*G`6*b{dNX4z@P6abXtKph+vGe_vm7jVWOIPdj3Nsg-%B#(Tfl$=moK_5!{-ATyn
z{%Qw2!YMx4vt4N?;Kso5NR(#=*la3b^GgpohP-C)m+p|{JG3_)b=W&+zPY0|CK|1s
zMO;1|_fBx;i4aHERcZc4RG*M`xwf4l;XmP`v_Q($11Um6{sLrThoU5o6PEIUPHUsf
zzgE47`4ESg_tmd8d{|>p%85DVroV6vPf7PmFV_F*@l!xQN!Z@hRPFIx%9ZeaBnf)Y
zZ%o>S!?jFB-4KoEm(OE<#;sti^?-b)SX~}NZ*V|GWxZk#M|Zgs!}_NvEF+kiQKGf2
z=})(AsC^%j1&X$Y&#)sNU1VCO&#5~W!tncAk&_5gg+NaFf?Yr6eJXA3_NTK@$I{Bp
zKB1cfM#}?07tbq*M4KMPp|Y7(H=psOwFA`1TQ{plQ@X{sjCEFbHAK!V$J3r?m##sI
zNGO~Ni5X0$Ic28BcArow=Opak6;}jNluB$P`ARZZ-ag96!4p9g`Nx<5Rc+v7BAhJe
z;iAxOXcLS^ttxL=LH&!3WO{!jYI`^<mq082AvSY_rj8o&ae)M^pCewd@&5Ts1^YSp
z8+8Sje*=WXd5og`k#6Yo`qk)D&kLF;E(L4`92rcW>mRqPcW^cl+^fLjH$M%N<uV*>
zv65rld*nM!2xm@-ZbHT*?m~Xf=h-;S36!dfuKCaL4Qi*RYmYpEz%bneP4MU!!-M_p
zSJhdR-*?%jvkZ}t14scOarKn$8~Ugr@&--WUPkghIHKokOfu`Sk0Pk#o0*uPnH%}a
z9*d*$FhB47NdU(d=X^`qyYfv&rF*YmKGu6!gncPrD(3Nxe(4z4o~^li>(o`AR$o)3
z50b0OznCyf8QJf#{||c~m_Jd-pB6QAyqkP;pF*|!kyMqGIOaC0x5t;Ped({BF!kH?
zQ2)H46?Qi!P*?FA;hN8ADV8n~+TcGnt99giLeWFcLXNz#6<NfT9;LxcjA}Oh-&kai
zm6F;D6G%1!kmz-%Hru&zg=wyUCriEKU#NIeSW<=ZI`<`CIHW<%KyW?v`5mO4eOeog
zH#9O59XSdpyB0uT?Gm5;120OukMjK%u*DlXH1=OrlaJsIwvPQ8(SndGI{-dPmZ6ZC
z1U}2~cKpIY^`pz`zsruCY#lA$S&7O1ag}9{w32r=sU_4CUs8zIIR0%1&+l3MU5kJ4
zb4*M6{=;hpz607t@tv_xeOjODd!6R()Jc6D9{QBew&E2d<vfF%Qf3v+)M>ht`LFo@
z_KrLOflI<*JrtGQSLxl1*jcvaNRe^|vw`74PtW8$mAq(2H{cVe&Pq$8;%&$}UR5NJ
znB^}Mo~99<gmW#ldI%7s?-A5@Ful+C=55tf)GGAg|KsVb!=hTlElx9pz|bHyLnGZS
z-5{y7bc2)#h)6RuLrF>_G1AiFkPZk)NeYsRf=DW1+;7f(?)~>XM<4ghp1t31y=(mz
z)ZX(f{&O2pB^aCIK7?kiI-bS~Q+>T@dOOD*`Wk9Lnf0xL-p$Vej=bX0!d2d<`vD=-
zr%fwbuPL-xQ%!23i*cUH&ZmNR@$S<T#qtrwIAJ&yWnC~-%;o&dRm-SSwJ9{g2IWe>
z@_Cq)u`oPH)huS5Y3#9JdrMU1*NKC00Db1eVZ`$x8Vir3mEGraQO>u*^M<e<g*)qN
z-?>ApINRfTH#-Z!JLe#n1A70G0eZ$3ljs{g8Jt2rK(~?m=5+c_j`>vV*Iil$)VISd
zzvmMpexA^eP^l<gsuA?t4?HJ6Tdz5PYc-^XMFhKW^zTMc+g;CMMI<g#-0LdMJs3#o
zjNxW|U@TV;uGHNQR@DyHesO>i>ybA4h-L@_{?7=jCHtVCJbNm^2tqHW8F2T#Y@St1
zk1c@~VN^1R>r=gGm89?cYti5mJKo1v63#V)s;s6KLz*&y_p~ZcJJ|3d7u+zV%PDD$
zEx#+w)RWsW&IBKHQuZ{1ao8#|!}M*y4Jv?4Mcf}MO0gu~Uo*HO<9Gb?b-ElMH^6e0
z#OI!l*$FBIXBKSmn8Z!{EBpf9slVY6qcvRt=Td?I|6uV-b5CSwMYLA)?FohD0=b1J
z2dqfuR;^Q!m$Cz_7YwiJY063UH&py8P!zSqB>3g%Q9;Jl0q?f=TI%T&&t20klAZ{?
zh~Jnb2*tbi9&90-fY;Ah`yyMRa_r4Pbuq4M#)$a1R`HC>vk4D*>C(PpTu4^E34Eh3
zDIY4Qx-FYt^E%1HnNe!EDK4E1+$_3X?Y`uAsG)^nA*VMnY!zUSJX{Hb*D=4Xz|OWn
z^9=o1*0I~>VeCM*K+``WThfy1P#C>K)+uPnBi!Svm_IEM^q1EOf?H2V^%LxY$r(j)
z+|}(#CpN_7`OIokV<KOEjk!Y?=?Dh;k!DB}L2iWO<UE*|0Es5f@?NnpSP?T^QDhUz
zJyvj(HE(s(xPKtPnPwBGxrvU(ms6g9Y4eOuS)R8N`-TAG=~h)ZD5{u)&LvIKiCHa$
z-V~#vKt_;OLJ%ghco815ml=#=G2AgD^zvc8NM(&*UmZUmr8yg-+ev!`l1!id;@A>N
zK+$uxq;YU06p}v*P5Je8E-^Ef@z>^csmyUVn~TL4^AY}hQn=7~{Op_Ymc~bL^KuvU
zny4=?ILV!c3~j}<YtZ=VWw_3_r3un;@Kiaak~pGhGjW_fLjQg^eDjowl&n;X-pS|9
zzb9{-?+dD+6i}+^B$CRP1P*3H_()dd?k4}Y8CEC3!343gRGt{Qm|zY)4qBFN%+fIj
z-G%dq&lZ~*+(irtL-?Ef{My~)QR0PN^R8(U-)3HY6rjRgAC4zu2)r29x#0=DkPtSI
zkQ7C@BKIT^!2_XUBz0)V+{NTmHhgVDpC>nE{8Fu&eid1AAWyoaAV*iLFrh$=pM4@-
zNiF{TY#cu&cYe9NQvhz|KV~xi_3mxypICS_^N%=~Ovs--dL$SHFxDCm2FhdB;vel!
z|Cm+ibu8TV#w;fcrN-pTmIsyhtL{`4H@39lONvj|@kX&N=fLfETs3Oc;cxsuY2@hv
z<(qk7uS&C#A}Z}hj59DZGvmDgk9i#0R_|TW!Odb;!B)FbNe^<>2S0<>G-so|;yxs?
ziD|kF6*tu6xNG9zyC|n{HSH`l*eAtU5iKxk3~IT>^*Y@pJ-y{(@u*sJD{0iJ=#)~Q
zri#rYx>*?Jvja>WY41EZc!LB^A(SVwY16bfK8q8L;v~NIH1aa&QiVKn6Dr=?VM&?8
zl~(N&LPU$tsM$wBNyn8s^!n~$nn6?dW3Yp)JpRlWYf$oAB*G6V14Ttk&>ingC(}@G
zfry=2ivfWi7Zkz+&XrmQk4;xAW~!xDs&*$1tGF}>LS1Fw!1P_cU}0S0w+0`>EgV&B
zs<P$iYPDS19HqVlU>~^?=R65I-IA`7co9xQ2pbKwtkb5WCGEDJ<|2zR4BzRjm^uPQ
zG7%lf2A^4k-Iz%;f+ua<c>62mKu&Bn!F3K4f02wwE%WF`Os9?Lr2e&|xsCXPZ$sfz
z`mij|?uHi~Mcih8l3*^_0WV$zZs8SASI)pLwpm>i^xEfa?X>S5R(UzyCTE&Cr@s4c
z9M5UTP;_6w$g#>0{Uenm`;kJ}ZG<rW4>x+^`R`#eS?$Eo_5gAXHI2|Z#;`>4)ER6L
z$VeWRHBf(Oq=}?=CC=N;prwm24iI??<g*hBfxH!Y><2$9IGEE;mlqnC-^myr;aWQR
zH}@>EIVzN}Xw`g_?!I$qj4{VUOtY9OKn77i)j=Rma7qx5yFL@dUnc!ir`_jw>jZ+6
z=IsoksIr#_Yo5tJziNuT(A`UCB?$?`Ji0)f9A+{xpX<X?PyhR(X9d&dG+{T!+vqtR
zdfNAZ90Jp-Q`QdWBBR3BrK)+~zS)*?pH9!tNZEzQZM&M%SD1U};pli$28!n6icrRW
zOqaQ(v??_-?LqSvT>Uf>2^DQJE~l&eU_1s95Jm+cf0Y0OxZ~?2;ALbS;H%P1r@(>s
z{rlcP9Dlh+&8iaF<*FQiEb`6WWgNv=@W>*snVknmT32*e!;I@7U)w<oEs#8Ejm)Jp
zM{C5k4DE5DTEOmku@=Um(TgaMS$&Iek&#SkVYiz`a%Q+px|WXj{Ln1O`{jSCY7y^1
z_r|p&bw+cQbXZB+HhrD^1j3QA3hl77{L%Jc=v@zRpq1CE>pCZGvZs4OwgawQ=M3JR
zF(S<Cipg+Ko;V%Fd(aKAXv^LkENO_G4rH#fVtzp;U(s78r1NASDr#*Vr%mUsX22<B
zl~=NiCzeUzn8|Sc8%9q#!)Jxw7Ed@&!XXKk)29;S8qId(sk>O*aM;s`$MRDpZgKSy
zl#^H1yx_a>lEj7q?_+Zq*{&h2C-SWi<vlRn;ZWAD*=X(Iuy_S{a2N<E*VpO!J%gSz
z7~G4wNgcpFtH}86lC)F1^&PYrf$7_nwsG(_c-$U<q0z~+b}&;Eo~<}wgwJW#M5T~U
zIgxfOV&DCV$D7#Zt#n}ptDZb+11y8c(%qpK)2MSxeECnA%n&Y>9F!!Z1-8X|y)r#V
zDr-wqLi(PtjH3zm53WYK(^yLvEKfCb)RHn1Or|OCC|<#nB)o?>G(L!HeUM)0;CY`3
z@3USC)<~wE+rhI-MHEI==fh{F5SZpDH-V+yyARFx_Sn(T;93xDRzKvNC{GMsExu)`
zh>fIP9(6kdkgDPowTnuSesC|x**Vahbj~V_*EK*YCG`Pf7T&=JxzHfqi0<H*$4Mlt
zaHQH+Q}^ZmVkTL)68T7xCo0CKtJ%LcmHT5d`yrrEMf(l|g>dslt2Qf=zB+cubiLu8
zi0xm|Y!g&(rNUM%%~OVMahEEynokeZcnwOGOJoNQ+Qe2#-lfXfYRUZry1;Gb@-oJ|
zAN4pVUH7QoABp8G;Zj3LU}bn>+FfN|r<oZ$fc$=k)@BKz@wj`bN<qHb8zQ=8I5C3A
zNkl}I&%$GNTMj+Da1B|RF|j%kkIl#``xoXI7h=6bs`nJz*K*}^Xa<>*hQL`<zfp<O
zJ}W4NgI|#K>IZJ-;bx9u<gH;OVL$e_lfQnzr<P=cy)iN;XK~Sx#5&+CJ%XmLT?;bQ
zgG7v@zGB$PjxAl5m63r>Oo4}AsAh;XLU7s|oEZKUZiaepX~jhwq|SuwG2?Nlf1whH
z)93gtVXZMy#$#A$z9IN~kumNUetl-W8W$`Lxk43Dk5)1^o)u*pnnKX!eP<Z>cBaA4
zXXf>7+jjtozb!b$mK@(f5gVK=Pnd@5;t{E6CIO#KNt%Vjw^Tcz0;&kEp{T;9=ToE2
z-S+KRowl6I`(Y#LRgSY@!Xwx`TvV?s9xBFndG+?@+5ErZ4Q%}>=hiRezw%5OWcior
zTgiDSOd?zeJR|P<@JkvI<4g)>alQFmYubgYBW$V0|7AZm%_a54ws9Av1v>a#-rEl0
zt2IySd44o|0g8W5{Me1wxRqv18pG_8{Dz8Dj=uKIkRPjJR3^xEdwe@AMH+j)38Wv(
z290Dz4De#O5ZrMB%tqquG*7R9dF1@kVebQw`Xzk&3apo8R4w+t+bnR<Jh#1_zg_V~
zP!#cS6QhFD>yC)#T3a6;?9QgggQ4HfqKZBCdPOy1Ks}dsMop}f#Z;QjF*%YK(LL4h
zj8HXz%A@g2BMrR4WS6FIeK$SJSZ7N@vgerRC0VkBfUuepH8r??lPXQNW(MK^u07On
z9(>z3^#iwMds9#mo)(Bx&MKpPQPo@ZPB5bPPQhasU+1q%8Ac>XgS?}@CsB-#ko13^
z<!=^$feQOKzjr2whY#yA#oE|f!Y0YefxX)8$mcHv^8~DtfZx9^bD<(K1nd|4Mu8kL
zSOs};Xqjp@8|o4^f`N477nsfYjoo3ut&y)#iunzLJR);`6UC?lw=Jf4;B?@@m51Yg
z`^BtUb7m0XS5~VA08{o_2rcJ<wQ&t6>URKtaJS`We?+YZXy5J@*t42ir={^+ICUX<
z&2vl`wgvi&NtCqT#tJwwB5Mh34*IYI5Lk7^`@<UKeL=9{hGo!0916}?ULVxN^$Xug
z_P75@8o}x2K#bG~#gjrnE^6^e5glx<dLI@t(3vl(+&Il^{0=U3#Xh9@TeIbv3Du<h
z^pY`C76<u>(?p}D<ipt}ju4plD3FoN+45orcjvW*V`~v8_u6x+o4boTw0EFy%GBaX
zjpI|yXulV&^JMn{pTR^sw#7-to`YnNLu6FOM8+UXD?@b_ZW`>otPFr+G;blTFlE6p
z6hF9Nt>hmf(LN6;GK*6J$Hm4JK|?83;a&mRMC{4n18l!Hqe^@yjXqiEFh|+JTCdCl
zc#5tuvxfCfv?Xj7K7(6@=iN?$c8nWV@IcBjSDjS<lWL`w{4_eIH;5Mk%p1kqm{|I|
zu>z}}0QU^Cm(LPOG-s<dmoS;Vq^C5wm2~YQ;88=XtjaFftSn$SfkDDgMHUG=YCZ?>
zc1ZGPf+Jpi`D3BjC19lZeKDHKR7pf-7QSid@<)^5rEEve(+KckahU!o+AGzmV`9Z4
z(oMl+t1PVUl_7du5sC3?;Q^_RNoc1yF%<C&*&TXFy!N)qC$&xummxqzhliu@%6~gN
z^ZW-^t?KbN^M)np`+-GTi-$G!@H9UXI|j7&Rcq3NoqwjHo91jVYt#@e$)cDwua^5=
zHYt(e0s2+-z;<sV_q`pMyDDJ}m(p3@2T9CjeBT`)4p6SxRa9%DR*vq7H<i`aHBlmg
zD=#wgr?YCEYhvO*ygUs?GY5+}KM4=AfDi0)fn0&%iU}ax*#?5KT^PRBPeA|3z7*cf
zE3V@cYtgm7!`!-rPgiZ48wRTJ#tuwRdEBTuxW(WO$?XsYT9k=@ToSYauQ%bM?}nFY
zn&V~W*9YYL(R4QQirta_{yfT%FZ?v9M|hje6uI^F!I|a7+X8>hW|XfV=f6~r&)&fG
z#`d5>Cq+@k?G*iy1CGlVX8x`(e#s~phhVrFEEP6d_57_j<8tMuYjQR&23zK7y@o2X
zVcWopEr25W&>1Uqv&In%Iw@<kYWf&^r#bhQu?*Bg3zI11G%d9VBP!rCNyNsK5guf(
zt*9TX`d3xh9w;Dm<HYlJ^EH=IzoMH9C*rWCOw7X7gC-!3=@Z3aNv~&NPPVZIU$O1D
zEfQ)pWJqwEvV1M__{hN%?NC$u*QAle5wj!6Rf^<MCm|E1=7U!zjW)A+v{esPTb~Nb
zv7PmHVo@N2UG~8<nl6xsWJ3K>bJ6(3*60nO_{NN9mRL;7AgfKc=ywZoE_GADdz)xj
z)Q%WEtl~HI_aE57PYl8M7W+)^r{|J>{+T2{(kwJ!Bt6}s@u%zJD>5qFl97<~%~1Vw
z4Nqtr%PNy3Z=}PA7Zg2&JF+ISgPCioEbgH?FriV58GGInQ#El`=>-rQaZoFQ?|7|t
z2n8krH{$iq%($0<+r(NxyQI!VOgxkz=x@@NVS8ZCcZYF^Cv+>Y;(8TVHWh7z>}x*d
zDSHksaUJEz^%z9v9xsaR3ghLlDTQb1E1zBL_Q0aXMDI)fIk!*_*q($3LrpmT^d<Qf
z0Vi`grVpcTy|afCJf&JZ=C(jRgDcxirI0W(9mA2&l`yC_j+d*o@U_q?6wLb8oU96f
zErahXn#CuiR=5gwB`Tt>MulB^nO}htz#$kdd3^w?S)o=qa=$>2NNo9C+X%W@5q1*|
zO@dsF9QwTb!LYN7r+M2{j^%lRR$Kp|MPA}*VqaYjJ)b*?&;&tv=(ycB!N^@;r<3Fd
z*%J{O>`KU;_F!+HS51U29YYGfFb`#+=473h%#9?PzDyL*n4dpeMksT$2*|QRfd{5i
z>&lo;p1anI<ZGIc#h%dHC64Edpfl@dB+VSKewr|I)ZISKq)(zimqjtwIhUL>C&^Lg
zfj{Q@+S?WejR4qF50>RHs^1_%N1f&eZn`kVhD}ajSsu0@_*ayx;fb8Vg#^So_7AE>
zXH&Wowq%Nmjxq*<-CxbpwoBj*4&ZOK1+q8dN7QDp8{worRUniBtuanUOyA9``WIvt
z>_W}H*+lsu5(&@k^fOvG%HFis0{Jj+*CviNf1Jf{F3nLwPESa6FRFGAJ4Omg{=ts>
zTsW;6P4fW*o(*9AQO8Yh8&u07kE;r6>rb*`qV)vX!s5X)>{B>2%5&=_aB*Wz9t#jw
zS<oJ2-|m4-a^5U;TiWUK@z6Q3^~qvM!vuF9*5vXUS)YNV9=5VH;@@nP=pJuNo$=g&
zPNXD=T4Ce$vL&w`d|LNlVfk62;@y>1@p^(XGqiNDh-C5pN4_m?^(Q}M72!*6k6-rh
zH;3==>HGW9D9H?|ahQDf5EjBx)zX!ITE0tNb4c6`0A`yhgSH{Y4os&TXkR1JG#AWb
zeSyaeRh|R4EjtfdnO&@bIVhJQb8%X*^^b8%#_idPd6}7OelSLF(3t;sx)4q4m3sWs
zr#+RPfZjm!`EQR}Tom$Yb?ayQ+hBM4(-ZF~o@;X!0FZWDzXl)$J)r^$5zS;W6UR8_
zlR}nJNEK*(tzy8|FwoPS$71YS2-(_HKn3%)1RUdEPISo^4lCsUMA<Q8BH8F(G@;e{
zpu%e*Tk)Uy0A4CVIL#j!VX0SYDnXtzz~JNFejv4IIYf)HawacInVb)tNCb@}rPfYw
zAA)@oV&N@?3)nz~_~(p|fc0VQ2Nka*sYUEksiFe-Ay0X(96|7rjgK%3dqd*u(mdF#
z^D3Q=j&^vWLgluFxF~1(a`T!zg=JJs1TLPzPL;(i#4I1Y2a2+{i5dRvqag<u<=KkL
z2ebM6wYQMCHk>%?Jr8e@C|HIS97g(2cT?bgJ9d$l^fPBeQ#Y}0Taf+!;ee~<z)MNB
z18t+4wzB@OVE0u!Plw;P;V6E4u>-2_q6vnsLi<vW53x(7*RJ6(7`3Pu&;6wlhWkG9
zA_ZqO|B!zy+`3u0^KWOTkUnb%dBqnx>0k)JHR13??%^nAN*>a~^e!_2VaI#a#GY$_
z0IdW%-jh*o9)yBI2ohhu<8_xNK!`kr-yRS)tJYMQ=6iGFD%K%?xNU-R7aOR;^`jTD
zH&Et`4V};wAmWdBL-ITjbf|o;kImR^b{s-<>sXe8w_#<=gvH$O2{`$Of&I}&l-W9{
zcyNQ0)k$*x9MmARF!d_WUCbvbb6^7XED3V-q8K*jN0>XxmsKufS@g}Ef;(ysz#XV^
zGWDg#vc;+UTbxb0N(6{=0|IH?{Wtx;WEw%_EKUvJv2PA?t7I=gTHZcO|5M9zrlCiq
z#~QOjZOAcezvWfT56}v|TSF;fCd;j^+i=da$T3p$S6&ti#I7modxD`dm|um#4|Im9
z44hXF?Fy<-K=fWKGVMjjqe+7G^EdU|J5wE*Wf;KsNO4tZS>(P0(e;!qD)VVDE{|~~
zcXLgw63!8^=@8tvt0ZetQwh86JWAf0z%;3IcWc>p4Y;eph%gzxcs2-_3C@P%K<a;>
zW%&Uatf1esI|~n_KLp$b@@^S8(HhoX3(a$^E!C1B6MJseW#!bMwjkv5Mz=5b?%T7{
z;<C*1iXbf$-9oy#<><e!ycbW#Y?~c${l3vRZx7%p7#&u)9oIT92wlsHlRcH1_qruC
zc$JXx2WVrm&+DSq0z?oK#t;2zwWrIvyhqkxx^Fv~jQ|LtvCCU(t$brM0^ut8*8_NP
zP3P)j>@h7V=sl3vH<`XCXou}lcM;+4QRH2RfBNs(jC}gv{p0}dCxr-t@!=}Y*!+Ou
zYuE*-1&2%EQ*=iP+%!)d;t?KY96bbEbRy*U(s}H8&1D~85`IN;Sy;DMKC-C%QL%>}
zPC&6}1tGN*neU+%94Cd`?*U}-eV3mAIk$oLI1vQrmw5$Z?nmrH@{&YXDNLX&<sPoh
z*x2+V$(MZ`ee5*{ngYC&TdW}tqze;Z7xW4rxAkQ<boQ<Z)S1BBvip><)45NvOJy=>
zMw4L~U^4<b)_%Y9$VywnC*JmRyhxy(kZt`#lS&4=48r=y;`);QB9m%)<Z)PT3OzRm
zkE`>K-9e5LaGZzUA;9D7;)#A`EHgM*L3|&#N^>bC{zDnhcP!1S*{z`#l4-Rh#teUw
z_Z01$NAKvOIgjr|a-FFDuQ2qcJz#)oXeXHWZ$;cQMpO4+0pQet+Xe6lEd^*RyrqR<
z6Y`=)Ee5%X5b_|=UP{eS4X5|II*NZjiTtA-q$t8UxcGe=E=3sKOtrC*CKO=#^6N6i
z{Z6jX3M-LdhY+!QEwRp(gKxr(I?BwSfPA23SY<!yRhiW}Qx^zs@$ML=KJfs7-)pKk
zeH=C7-zndN1cFTlKg90nvrK-ZRuC2?ZnH_rC@TYeD&VO}{CC`H){67b@=;~$kley;
z`kMKj1EsJN7yB6yM%)xh7Lr_u%5N=Fdm0%G=NPJL1OGMd%~%tmQt1kK0*g&`jw#>1
z;Vv%a6dU`&#=4vEwqtS>g(h>5HM*|3j1F&@4E^N8f~bt}yeOndbU$zDQFMLOkpca!
z(h?Z>y}g@89Es%3gph9m3!NU7u0rVX23`PTF}99s;Y|`72VXC@JdUgJtv$$?td<Pe
z+~1wPmulzGuM*Y9NCJ7i4<^`DbLqQ8&Uz!Ho>1?{UP-@Ry#i8#(su5>3JsD6e+~P{
z%%0f>k_7Yy$nVcqlZCx9*nv|`0)NlIF>xf;UACvnDM0fnXP2Zq7CKMFO?~KJFp#6P
z7kjmsDDqc6<e>I;om#meqz`P?y4bdPooLJo`p_JXW(OfZhxzg$;KRnae5#=yJcZ&{
z0J|n|v$V#FFalRm2Wwxbf3*4@Fsf&azrB)~!6%UF)F3V+a(8|0&=mE)<ot6Y7(6+)
ziwyNjK*C}<+ywdn*Cc8Te{^%@%T33Z?FbHKjqsmIF##WR#MQ!~`<?lRT~rVMfxhO!
zUvI}mHu6>jRMBBok;5XLu_cni{%q|_f%RakBprWfbW`E7ju($at`1Mx#&?H5JY9{?
zTr6jSofjwhHEGDe&IPCHcMu|HB@I579&9%lMB}!eQds}v8C59RvjaY7O4Tj`2SY1h
zCQNbuB<qh9%bQ2bQpKg%r@ovlmZO1MrV`YCWhT}=>A%e2^#&=y)0@2hQO5W^SN0BR
zp$5Z3rk@@PDmj(C)5?Z~wVXAUnUVYFw+|1QO+OnK#^g4GNN~y@LBv~vSdAL<r24a5
ze*)P$>MBtZ4s^}hEWAqr9DlP!bDGD(^4mS=Eoc4$KSjPCkRbzf8D-}W$cxW`j-bvj
z9|b(O8hDvWSabl`_J<&Ry;f;ch5&r;6}JeVBbMl>_Vb>G{7?3vaU)L?>OrloAXzYf
zlr=>soy+*H#M!1w`U6u{dD7W+Inx{J)ECXR=Vr@g@?dP`xl7nmKwcq7RoATb4NDp)
z4A&DN!{XHfclA;p!s;++o{5^grQkHuxFz~r0gKk8I@>-p50mA^VhpaU%-}D1h8KCc
zSh@`XZu7O_ydwsD|3iFnn)u=YrqtXT)^Xevb%%p7!s{Ursdygrl?s{y1L96F(&Rg6
zp&a@ae9k0wU`KtT=HMY!rP)l<mN5IH)aSqq5gt|R(YY3K`Y$H{3yeAf>6n|^9mE4u
zR<%IJXL~5v760~@c0X5N5<Phyp`b&wj>9?du34k?k%HhKz2YAHRrD{|(gVjxbjZ`i
zk1!r3LN($q_dBPj$r7Q2!}o!2D?P8sIM?%13DmNtKfUiYt1=}g%K(xmw`~WDVQtrj
z<P0sptpQUdvg_+5ZDkT6eAD%uOT@N5^JONd_($n&ue4Wx06;@mC!Zmh4FqZ`@6>{8
zdTG^`u;UL9T5)$E<S&WtE1fp~9}6H>c6bm~es;}!Gxr@Jd+OqR`3I%(&Z?(h`~?2^
zkj7TqZ0e=_^}YR}kpT{$L7%%+RmM~h2Hap(jbb6-59*q9^ZgFamX51AZl^l1`<8e%
zZcn&Asx8G?{rlGB|2U&Z3S4Ay7t_Fe8Fl`$S<D$MS0T$D8YBOT>ug`0QYjcX_fmOm
z6W6=8h~$|0?<&(sBaBSKdyBWO1}H>7;JRF%>|!j`ANdR)73kC+7zjqv*qLwl0EUe%
zMtfdljHy{jPmKbOqp^1c(wnQ$%yimnOEWFWTc$-NAP~;(v}N$Nmc(Hr(~~+yDYw~n
zkzOx_N|=l5ii2Oy|KXHW`$7&)S>*SXS0E(Ro=}NO&BQ_cIRntB)^(|tBXCgQ{?=dU
z?3;Rl#$JA^q|V})DqwH7M)5C+WKK4Nh@}rdOWfDjz(Ec~Inxe~Ar#?BJnA(2^y&P#
zJMV*gUWg&UV<h3nU#&6^(s6leqV472o*EZVlS5cJMt?t`f2}=0<1nFtOh7V3`*7Zg
zs>KW(czSH87YL($9GACb#01Mu{Ir8{!qi`Ui3K45`}wK*D-XP629c>tLNhC+w{R*8
z%ttzR|8x@?G#Yu5HBS@C(Q0hX)w1=QypR83CXHiP^h$Q~Zq0PD%zq36zei!W8&_ld
zk_e&~((3v4q9?g9zLil%E9A*I;eWvxduz|^@`E<Dyx&Jzie#*_T8QI5#|Jdne@6ha
z!tSC2+nw^~=MMx+w6DdK@$Q@4Dg}NWyGGk?b5H03pAHjCK<to$S_AN;@_2bdpL_=?
z1TO0i2T8GzbPx5I-9<3Z;~O((P%?=RmlOkTgj;>rHs%WvyH;EWVqNFK4|8?vIC2Je
z$zOHN!j}FZ%;lc)X}CFRyq9&<1bh&cFE(Ql|0>v~B}v3NhO0@+ycI3J9|fy^j`z3S
zP1U=pJyM9%FEvbaHq;b9TLGd_r(jq*9qRk~5Hp47uKa2nI``fqPryO|HtM94Vrmh;
zj@f(q2>3@yNKS~0L>vAgAo27ZRBO6Yw`5t&z#B_fM=Ea%9AJVBSf#vZndI8;rdMP-
z%L!`F1RnU-GwDfJHs;_yg}abqOJd<%ziiY=5^e+7NfI+v0(EpHVP(=D7ip4E-bgB*
zkuIg;5gOd8SWac{3g^Ut5vU({!~(tC3<%ju;>Av`Q?hu1o)(<cmvTSKMjklLOT5AS
z)Fd|M(r~kI$W6JrRhP?y<7P!S$y|w=2iP*Z1sg*83LY4S|8F0}!@AQi%^c$?YHx@$
zQ9*GwROmYkS1y3aE|Xk3uAj+=-m9RCbDd*4E@RInBVVYucvqP&)wx8G6+A}~!YGxu
zR_hxeH9#uuxNs<ALQ@-0E(K38FTOql>J#*42~b#$MAW{|_nO?diO{l;$Vo5nZQ~pc
zjz5OaGq)1S?A(}zOQS5SZZYclqzGoxg9-ZY^kxr$Ie{3S1OWaHhk-_Rd?$`Lh@VI*
zr#&@}P7<dOJc6-8#hqoW9XwKRYj6q0Sh+V^i-rh4x3-f#65!Mg#xYTU4J?R!1mm7A
z`;{c<qv`aKO>9C^&H4;{k}XlU0l%&m7q~8afQbkkloTqNns}n3w}+6@O|5_14Xr_d
zP#%RjN_#)<+ZErUjT3Je5?Yl?@=WBVwYv4NoUP$}%MueWEuMEWVZs%Mt{$X{mzXu!
zc)qr$Wy|SoE5R`T-19w?6vu4;8l;&`U@4)-30uyp42pILoE&Vh$bjl~Q-9x<c0?Ng
z<5#s}22#P=tx-mJ?%Y+Hal(OHS#Y6bo$<_3<_Y-01^D{s#?!9HZbAte5Q&A-zIxjf
z96pat(}I|^dB~eKfe3rCAlu2P{k8+4x?>27v8Qzy6g0Q`4eU?WWzux|DP@MiIMGCW
z9<G7}bg1v(ap15(;c24b)Q`>e0=GR7WO~5KXBB{-A02+2S0p+1!uwd8v!4A?+7sF#
z{gx#tRp!0p+6NdO9}PvAo+XRt62(WTDa_OpssIalZ{$G1fIC4r7ZTU^J%67@^wUe4
z5yC%_fLUqMj_NW@hbNy>*RUnCxLJe;458wLO%>l}Etk-!Yl!{mn>s^KBoEbZHGU|q
z%n?JJ_rFiGhH*hnVt$jP(_}kmt9c_5jrB8c-FO_T?B!LZDVl;dNfU3P8cvRoChaEK
zI7;+2@%mBsX&Rid@K_Js5Q(*;&89Tq$#UkQi==mqVKp%#`NG<MVQ4U-nCA`h?|6@c
zOW+9JorvJ?3PLaZ!nZ#VG;>`Uyrh%jiSK;>&N7tKQ>fVkgiDWK6bB}oO+9J`FKoXO
zMk!p;{iTMU;hmmlL+iK&&1kvU*cu9`3Rux4iE=OQSsNP>XcOgNzRw&FQlPN3#KDik
z#orm+l<VRTMj=0#A8z!37}g5KF}p1~g{&?nCG>UIOD0)DPHsEe7!_JxiwEJ@C@3yI
zkuC6IyWHd$ml~J(akaVr>LRt?htI?gF66XvvDl&;(C_NwcsaT`tr739E#&P@&fV+M
zH5I;CHpoXr!7qAUEcqhJTzwk=(F%-nQ_F6h;J1*{2{fxZv2aRJ1-UTMze#F&uk;@Z
zV3Rt^OYVI4MrS~$)DuZu4^~0TC0u$YcCwNbe@h0)+4(#{Sp`RF3A?m!Ha-8}=Sjbs
zN_Fq*uoBU}_J9dpgwc;R-h@V7Emd*Qsl^R<Ogt%G%a4&Qvou+TiWEeVK?yInZx&14
zW#Jr=l@|zZq<b5*t6%33v@xaCts{tMQf#h-9v+z;i%s|<_18lu`uf+8A11fYY?9x6
zEuZ@_Q`TA@tnAN-XrT+4!&fLzxo9+@)}(!IFv_GyVL}f?&DCw%yfFtDH-}lZ9-S@S
zN3H1(ZvfHi@Q!dXCChcTIi=LWAlQP&CEt?A7dvPw>Gm2}s@|%5|Ejeb?P?W_BTQ>z
zyHSV^WYB!)>FZTNqO#}kc8kiS%g^P?voLDDRLl!6zZ&@HAHS%$;$MGb0=vG`doV`J
z*x0rq55;A%XIU!&m#-G=-Q}Rsc{51NWsxnOscN;ql}wq<q*z7KX=H$T8zJH)cx>xP
zpO-sui<O%mcSN6nY1;nDxy6pgK?UZ0+T|HANnQi%^{t^#Zw~dDbR8cK3mZ;W3=CqT
zOAT!J6q~sM_r3&;V&H0~M|MHs8u2PWs*Xv?&|NFA3_V(kZ4KmGy;^Xrz?!0+mVwCe
z;yIr~UI3Sbj9r^tK-iZn5I}af;?OL#Vca6TNlirIED7F_RGZ^X{3BRNbVtVEZ|R_s
z@uZPG4R{Z%z=Ys9W-ImkPp6f<T4l2>^5WNCst<erDop8;#6^84m@us7T6`7Xp=nZV
z5;UMK2j5}jj2c8S<8~0^A=d{(?Nw2-Myi&WUN?T7l<BpW#L*3m2~60ohxU^ZVXs$C
zUI;lKdj{XS#t+K?!LA7{cNQj%T<{6qo&3x(yJ*O~$c`JqR+;|%a}4@(jr;rOfOFOW
z^d>VzJoLpA7cGqMnitX@jdCho&{@{?Ew)Ye%5umPxzD!7BOY(gkDtXk^)Ix7|Dtg~
zir9RZ1@cl*Cx;37=@{70#oZ?I0R?HxBJH8X<9GBz)?ncZgibS^-()T_e%%rgkgTPl
zu6USzm(ecxmB0pb@0%t6q6Bg82l>BmkF<=qI|cspb$eNt#Ipf7v)eS;xoT~s0f*SN
zg8GX~z_s7<U+fBScCw580F*cY^~qe(7c9h?TEvn9-*=%qbW=q4Oj#_VkJH8Q_JD_1
zEAI2WocprWO%B3Ycn=!#5h>ox4ZosO@N~JsnCM(fS@k<by*NfaWy?@KxQ%cxlu7*o
z`-cW?R2SpxF2@CnmBgu6l6-(8;IN3r+t))E074~+j<JZg^HHlm1)Cz{DqL!8#(~Do
zahV%klY)q&u!1>B=!Mm{BMw@+W@#dkB6bVfgF;+kw*s{WW%w>q6ZFH}yk~B(1X^~m
zPlN?ix?DzpN4eSBrkJI})1tfx8mG)*aKx0zo-)>B@kFjYaN<*;n}x*k$w#T-%zw8c
z3I+*Z(%M;78k{1~*h?sQYeVLnFycmHw%guo|4|KN^tVYFEcAp$IU;n0LV(^7r;!oU
zY}(cqChAwz4w<6!z8JY)bPlJBO!*GQ^&oxl0z<!K&@0W)c%!vx1V5c&X}JL=WLg=;
z{liUzc@PeTbmFCO2K*TG&0gqp9x)kU(x!LSGILeDm+yK^QgYwZ7%;(by1BCUfnrya
zt9={<_7SzuXM+>&*#;24F!M<Vtjw2&OL5D5J5!#lBPZG9`Fo`%mwbDFKs3c^9+F(V
z!eDy4U)D<z!crdVU7Mws5U>cOQAU$*6{}ve{V{QR%8u?S?*ax7WoCG3XDq;>D|l&i
zR{Yx+U>q&xzNh-5#T$>xzw+S`(_bxIj`-l{a-8nm3PxGwen7IN=Oq+3YpDgx`U3k*
z7P<LR@F23*I8iQLkzK9Q-1=jKd0esCc2H2Unq7U8LPHY(w>qby>?XD3(qL@ch)v)v
zWcGymd3gOitn1ST*b`ccmh#R6)?4X^TPIryR)kvq<=gx>3K2NnWApB<8H9Y`N-cZt
zPV*DxA8?2n-kZL6V29>W|N8r;f*2un6cfLI$4?@{4(H)*f)9-r4SC=jKEQ|lcapco
z6OHDbWe-gBFPdDjJ}AUVHzLh~y6!5SsdXHa+tNCIWidt8Ro8t|b02sbd+`zW`Lr&r
z&Cq&Sv|(mmZX4SBt0}u-@ZXMTBM*K(PkP4oBP}Z8jwruD=&;tUyC`%sgn3>kg-U+O
zu_C6OFFE>__+1vSJj<=_K@@8}-F6;=fO>-K&iS1dM%<-f+7}%s7`fuPu9qLHPY8?l
zRs$^2w6|Y<oc6A%rinotUp>h(w3qHsAqZ*zYj{SY6Z`=PZ>$~PY}_PU=v60C{odzF
zPMG_3rrhU$wB|Wyi{m14z=Gn(?bI?b1{>5@(2&*`8nObQ3F^+RR>>{H%cNm>(sFg4
zYk~lWC)&Rg`$p5Y{lVrB9>U$c5B>MsRhcfFB*7elQG0OrNo?(1(cv4fKf#?xMs`Qi
z$|DP(Y47#vgA9Kx4pA_Wq4e;WSuo%8=*4n$spK63@BEeWsk`V{fqoK#{n!9nz(tPp
z51&6pjMsqz|F5EKqY9}`hd_NO_+gE}Di|J^fgO})Dz(TB(s15UGS3y^z9Y5LC(ArS
z@G1J|p(YnBZg|X(35<1GIZ0K(9)pC}AH}j(inKp^#p6v*4Zmlv{uc?dr@p{4-uS?}
zieOsy5Cy5JIMh_V4*jPEwXVik1~h+W5iv_NPzArIcvnKS9zzTceF(vklZ2NHt*B1d
zr)(qnF(+v&l<`FUa?3D;GrE9e8mqv>2~sf0K565BWSTyzW4UfeGnZLH0_8jL8$b{*
z{|07e84_@34I!y6Ic04)!wx{iF^n7#boD+b9&f*W4+`VZlA)zg^7Pvaom}~~N5}B6
zcEjFJVBW1*MLbbO6yl_61-Qt8Y5qxw3wUs_LcZvtA|p=j>%3axjA|9w{Nb#fUu26P
z{awB>p6hDT0ycbNz>lA%phhg^=MfD0<V%;4R#8={j4j$?{_8DUXyexmGCqT-_lh~N
zDN9d*0tgkP^uccs+4hyW1_Eh_#-I06rjS+=n9pmTY^iRbIFavxm+KFeXs7IKI`+Hj
z+^#zn*np0XO`f-l5TCmTaKa>kH!|-g(0%*<(|rlJajd>!<8r}!(*S})xQ|;!F6hzq
zvssoy+wJ90|6j1ghpeEi&UOEc!rIO2Oof;Ca6~z!WZ3S5-OHt-)elWsKqaUMxmX+p
z-?LLcThN5qc=61~Mb133|NUhz@RP%#Lvebl8Q|C>w^|9Jl+`~Ug4z<ZC2(KO4<Z`B
zLhO2;-kwEZAPrV0<C^XaH5Mq@113jl!H-p}s%1I^vV+S{ObMgGx*A9wC>=h7rp&l<
zm&UE>Vffdc{fd3Hmc+>!0IeWrRpsF|n>XJ2lbII7>dAxuEK_ENl4s@vfXN0gY0Zn@
z!ij+~zs*7*m^4~c0uW_8z{17z5;p5XeQWi#23kSO!UN9~8sbFk?CVl-DPg7eHtnYN
zfjhx&oq<(|p+Gy}ic?F^ypf(#^MumFMfxu)T-A$mSy3!Z2dGXI|F8&oLs}f-9~GTr
zW7C~)AMxCw!Vh&JW4L-?uObZUTcZ8Pwqv565-gP>|6-R{f58rw2j%#9tFR(pGlIjb
zNYe8Jl$`=ZkkGUlk;Uj1s~3~uj+5}RQ3pqGQLQ_<;ZY(hZmG{iv|Wv0Z@#!I>1XP(
zFdAp}elAG*_e@1NT_LEB0omSc=J*>>{6?idso~q2ZgKg!rC@u4cw(=qhcgS$8h8;B
z9Yk4(BbW+K2D#Rar&5i7mJpjak{=NBdM_x&@mCEk6ns|m;H;oG>jeJIzc*XVtaCCY
z1fq@B_;CIvenb@_jEr@7G=9G&J0Mz;_ReoLO`%~AC~IABMN!4N$YeZsMUg0k!&P*o
z%%9Qk3-R80P9ED_$NLv%Y619uVB-zOs9VF*ue_42R=|j&_zRiR9T!l6UMVd_j6vEx
zE1FY>V7<OUzLg(aFyYbZtSMNI5BC?tiR9REH7uG*!Q=kYU(ORYPze22c=kssiyXD@
z>svzb75)GTu(?lT&t4U`vZ9IyT>f>TmIJ&;LV2o;cXr7PaR$Zbe<SQ9JEvjB2X^3a
z0|M^b#N#?38;#GPrfdmJShrcMB91f=x_Z?rE3#zOnq1dh)16608S}qhV!Z7&m$-<6
zkWz+CQPTrk&!is)gECD*-Vm_C>|TuE7wZvw7Wpq{$PFY-|13T$j8WNyd<4~~ASKZh
zv?JUn_0l?u-x4qkitd_gYLgcSLXsN)FxspG0uCrSF_CJ!$8Xbt(jUWGO8sULp@vnK
zzw%TWfL5SXLvVMrhfUgtku*fHpXMB>@4vv0SmdwOV{M=00$uQTV)090kUVMw19;!b
zaJ$=D$b8&L3d5J6-+)_@BWx%C3Yt2CsZTvO^eQ)81nf44T%BzQn0dlzo#ewOr52m#
zUsxdi@S45>RC~H@kpxOF<b_08#EZ=5vZMlkl{VvOW(Tu|jY2F5IQqG=I6MzxJh!Qy
z6`m*du7MAZ(u>W7^Y{ti9w-V}`jZ6_9T*EbRO!%XmN(O%iZfWvosou0Af-lkW9rh+
z7@T!Fhr7sr7S^#aX}@8Qaio&|GoK?r3PUHl0Y%H_Nsh_Xt&HuL&N=s7UH~5<n@usi
zf$1J@jl-mD>nTruHK->Z?wRT(^ksJ$O5-ZIa}&^8;A7l5J-<c}`E&QY2+bOsdJpA|
zGHQ(XvyEb!sag=B5r02vXl2Kz?7nknNUHz*)Z!_amPr^>#g~*L9KHbodf>7C(}jyI
zns%~iyt}0YMy@Stf@+v=2IE+4_rz1+1r&cE$Ia#C@DyQlw~_R79;Z7@l;+E^Bt2MY
zOZ%i0@0K*nTC#ZtCOaBHkB-F#*HGZpp!4hEqz3xquSSK&ncufA?giVjq5;>p?~5=T
zJ^DT1^wO+-AhutkSz=2!MdtbIjRvuUu9};ROi@HKgd(wh)is;&**#_(IqU~EXru&A
z=A%efgac#Tq)LX4t11UA!P0RVy)zAso_C=no&vTLp~?QWWVtpjli>LhdgxV5zj;Ws
z-z8Ybs!@#+mE@e(+0h)o{X=qnumY5p?L5Zdfn(>hy&0eem8eQG=n5ogdaS+6z4feJ
zrbWl4qA^NS9VGpwetjCFB%p8w+l!dq-)#4O(jpBI>?_s!3@$8KS4fkzeLLXqIY8f4
z;qvl0F^0*8eFU379Rpwh%)Uf@<rbung#w7mU(eLckE3hZQ7|=^9c|Bt^Z&RWvxe<w
ze#BxRvJL@I4e-fEa=lz<V;`&lO2^cZ>$-SP-1PPyPVqw?M$S>)cS{zuo;GnW%V=EQ
z$sR^fr<O4jInKF1nQQpabiWkmx+ZAT@ad=iCt}`r8q@~~xm<y(@6IimUEcCndO3WX
zus1UCNZa5+W_in{2@BQnG_9znx~UFLhZM;IS~jiNVlBl3w7GTp`R#yy&YPcO2+4f5
z+JUnN5LKBX$`5q6Q)E(f*ewSu-;BJAJ<ad_Nel6Hgr%tlJ>8OvDI^WBFOxyy(ydn5
z#vXp^K+q$8Vk_CKJG^69saZs9yF?i@u5xHPm2K}4rFtWRuu9<9A?MAUOq_JabrfBJ
zVKvWvCnRq(k|I)cN#8B&n_UbsWhw8(+rEZ^hYZig0M#Dzr3h2+*X;`>Ht#5+PfUxG
zid$|Tv*g*#SX<g}dXpYJ;pCRYS5PQUWE7@<Ib8AlKWKv}P1vJBQLzvn>K}=pLAZ|U
z^K^}eS!7q9{1rzOU%~Fou=y)1u{bCySyX2%d~r){&?Ymx8c-FDhKb2!f%YWCf@GqR
za&x1O0-Vatj{^hbO4Ec4woy%q*(oa;qgb_7cB1e-%Ucri8d*}FQ!r@|IT^lR|BGef
zlw?tnM6r^Ggf6T5v}M*P#(br+PumWLrM*zn-x2vubRqNd<6f!EdMGE%wWWV-#|fO>
zYH>w9(8p8{lnv7Sao?fnHAxbA+9|ATwES{keXOLCAhbkh7slpmOU3$$N#=JSn@nvR
zw$IU|4RCfWYz)NSU>;Ga+IoWS5TEAMy2nF|OV2XAD_%%OxfO8NT}iL)18*D73`=P8
zE@kW;^*hjF0&FUpay_j-ek8K29ZyT**047MC*coQ@m60w%r`8PfD`OQJgw7n6<5TL
zq2h2yGg08-ph<dP#gv4fL?IgfejHfT(yTjt#_U;Cv%Xph$|i0sMtL?9CX$ZEobmJ3
zX8q=IGjbc@uE0mo`mdXDu_8l|E3*2w;9`l%8No?CVLmXZ3CrxJPIP>npG&AV4?QE^
z=4>qlM}Gm6@hgYT`23`}qF>Iyaf*fEEOz;?1bYl1Z)Q%3uscn(md9AHpTd=9ioHNW
z{mVOdmRL>dRQmm{?R5M^xc7=VuJ%h%U|^8#;8~8ZUr*T2?KvqvO|TTS?)7^nbN7XX
z=oarN9;UCA2qGC1NeOo(U2vGLO#dj(jJvKpsWcPC&=xlP%70XX*$d&GLc+Njj+-Ap
z1_nih6|Rwr8?r=K$Wsd&!#5&>u!{ev$;<3yy!UFJTD{5SCKbbw@fK^V+~f(~mSxG?
zB8(~jbB}JV+nKzDa~RO0;C&SwI1>e33V6zM6!RKB8m=+_D)twc>G@C5S!2wkk#|QX
zPWJTtM<A~Q+5<vATLU-e(|E`gK#=^Rgf{f-Cw<Pfns}Nb;C+=tuSz}RPNLnEB`x)2
zfEHEvRln5U|7tglyC+(iD_)bN!dtEB576yBZvbBVw(O+*Ec`FG7)!y*+&9K!np3zF
z0kf(X2ln#ul;F9OYEi$8o_azv!QmvEvADn+q3WToD2PU=N+zla5GMiOl0!=diO}f{
z{Kzfu!kfdy21h68eS%f^AcDT&P9`gBt|3DvlqD)n)`m8y?YJU&2ol2*a``&Rzz(GG
ztf@}0R!0toeC2mgb>Qdw=<Z7R0ppCm5(Q}$!Y7(5gn9h=WOUul>i7w)7|ic^a*bxE
zfg5W9Dnnjld%&Lh<P17T$zGze*HX``X59`}=AVPD*B4Z>O^3#tW0P*6Jaug5x$FR8
zy2>2U!*Ko`sOwBaKPDnUU9oT$|4~oPU*uC*Tg_)~f){A%<az4CIVObSQZ&CM*;$Uc
zG#ReF5@(fWSu!WNvaQ=dk)F`7d&1?Nmk!HP9TgrPVn!}Zs%jb+5<E9&QV;?9RO~_J
z8$K$=nt^9dT$+3E+m)cyktT@FA<k^r1xH=LAW)|&Z2~bo8`-t+lH&TtPA=rf$zz&=
z*C5_0RTHL@;-MFe(|rVZbHyAL?H6@*H56o={26jBm27wCoM~TEN#eg<1?;D5AC^ht
zhKl#&k@b#E!JtR(x{k@I1nl<nqK5{!hovo~<A<Z#Bq}P(R%Gv<Yv`mfl1ES^HG8yc
zh(6RCfPja*mt$a~N#|C<h}#NJ-m-UAw0MJCZW(1?LovKrbc=RiRoQi%Rz2*??;ucw
zcvCOpZmT14P3wA=nm5%3p`@m@28*E+BUEll;|O+h%W*(s>$I#OmCN8|`|V*Zd}4|!
zDXtm59?^SamAgyq-5``jGWQs`IPVn8&hUzX2L(I<`F`Bmf-{sDHvsy(n~e%Pe|DDw
ziWZ8lM3Y*5JY+r-Sft)x8@h<6wDU1c0h6-_SfO-tNVTwhl}nBW0UgG09&RR?u{DJH
zH};5yQU7>_fNDJ61u5q*61n+LBj)G(4tw?yYw31s(}$KHvf)B_%xX`IilP*KJ;rE<
zlm%akn%qv}wgObO!F;QshoCKF+1}mtOUJ5TtW^hDTvtEXg0(J)Rc9XL=2vB@PSD1@
z{u&2e8sp@C@u)5Eo(#c-Wg+-W<s%9qH+_obaeZL`OVYHur>#e6gxbTgjfSR=g2J7k
z#}VKQnjz>1HA#L+PzDyROq%jLQ|Bh{#zz2pX}oVTTz^$9(D=x%(YRc#rt_Vl%$mj^
z>AlD*Au?e{M6y58uzlHoxfS8(1q@;O!0eq#Qe&oukQw~4jV@}4dA<uyIVwsYdmlYe
z8au5aZBU<}e7#?_#!aH+e$+k)e;PPECnPTN0DgSld=Iz?HIZlD7Ukfzjz|``i^@-w
z7Pgv&7vBR*6sBj2;={VJ{ZJwLkSJJ(r;{1izaf_ujd2N2o+hq}aDtx&ou>K-uK>V~
ztIo?0O{=1;ZXgU8cLbKrx5Fk>h4?UsymrXK#rh_0yTd??iqEJe69bn~J<0Dzq%qdP
znBMMz^)rfX2pku0a%dA{IHZ#Ug7F?s0)Pi4k3g*>zLJis+TF`vevp;Wy>nf)3E<Y}
zZP-c<qNp(~22poL74aK9Fc<}C1dXSLPfFlqI|w{T^2y^cSa3CCyM~)+Ufd6v8s3)@
zo5cVSYLv|!@Of{axDFE{Rj8g3*v%Q&ZFPS(3!XCB10Z$3^*%YZ0+haYpUN>0{HyaW
zFOhLy&c%<Cu>=8%8P!Pe^LhBl#wiGN;TTJcRq0Ct0`_q;4yUkL((8>=^f&briaDsm
zoutzd&`Y1<5{I&BMs0qx;@b73!%$9AD7>8RDh2<0Pn~jV><lF4(vsQ5W=BQuH|<sl
z6HIjQC1czTYDAor=kaTJ3xB{Jt0h$&_I6Y&wQUBH?=|RRbv(amr*PpifC0$9iE?<x
zYkE8$gBO*YBB(qP^S4IW9oJ6gPmk;-{0?WxwPc0_;xjvq^k@6c2k=Uf*$P{;3Mx;i
zFOMxTx^K{<1(*m?c@Rh98_YX@e(o|8J{eDPj<TT%TRw-46;abGwQ~C{w~q8$FZ6;N
z^Xm8IY1?`rDM`Z}y;9KG%^3Sci^<(m+K%nqxo@Nhf@CDT2@}QB&2#0l4fEa(RRnj@
zPvQLcSc<fmQ7g!Ggj!;pY3@3_30F2!Vc&{n{&uymrsv}QkLlG?d=MnKRO&3|sC3vt
zybubk<{zoZS@B>BRCp1BY1#Ax8^F4Nvs?QWPpqL!e7RD#t{+tIBwv3hSD^T!^Y1ba
zT8h8~hvMv3jEnt2lS<tqWF2qGRuc~#WpW37O&o9RNGX^>xB6`KGh0ax9N<ajlL!Iy
zfXJV0^KU*EdP5}Ze91XQwZEiZ3FXKj3AF+GgXd24EWFULkOHTPNuvOmL*S8$^eG3I
zx(HM^a8d8M=e0#OTRWM0#Uc7Yg?Qf@-%AOmV>31XJ>YkDU2t)am&HG(I$_YzgruhG
zd>b#?V<db7&@TlBU(LjQmPGR&5j9fzvSje;TMu6Z^x${-JGb2`qR+f>9H{aT*PUI<
zKMsgdBSK<_P9QU8IeT@7W}(gf4jC$tu?BFGs7*Sbj=o~H!hp|pI#~e5BaeLm@d@E&
zf|knT57QxyYK%LoF2T<38t>Ar5<BBY#BgS$U*mTtW`@D&!Q7h0E~d8-Fks2}gZu$Y
zt8y)hO4Sk&vaw~nDoWIwbe}HOe(p#2oLVszi<!T@&e#wEIS0X9EP3GR<}BRW2tg6G
z+X#fLy`9V#8$U`rtuz{2rmKn{W$w9QSS@c$!T4`6VJleYZt{H(w^l&hn-1xXkGtvz
zFe)9Lj9w-XJ8Dl(9obZF7>hsUIT@H;fKubK3vU8E*MsPlE-t;4d>7y@8p;uFc-_fk
zCVGo0Vk)VZ;MnR1Pe&$KPK1$t<fSVO)WLm1m*1GcOi-@nizIT{pb&5c43b?>lHAs3
zy9r|MD4m-l77KLBg=`5Xa_OV-M`w%8le!ynX>JXxP{%PjzGFQM9`BV3srJE#EG5=(
z{m@c-nwd|lgMe%SxMBC;c1M=I^%tCqxbrc}h~UpGotRvD`M;M^h;^v?LtA}pEfUp#
z{$Fx!zFGLnofwJ@a9A5%5PzUgpJ6inogS|e5QI}Xl@??!UkV`uhtnDMb@0B?G5&le
zqAVQev4;7ejz@qpKE8{k9=}@2Bhbu}K}M@PRM}7q*|YG_wTL5{%)6w8%V*(k_<5a-
z!5khCd5Nw1Hvr#G5a)tIq~MfXKLM<HDx=D&*#eBg%97)@7Q?(q79UvVg>Y?d^&;9K
z=g6tTo%ZNhbG)=l%9*3e=UVl@$)UKF)tVbodQ^`Vu`KL(dqnfj8m)^8ITu|;!$j|+
zb4sE#l5eW8BHX-5U;?@Z{E=U^H3#(aPj@sj<M>=aFL3;?ltmJ#F6jg9A$rDM;K7#!
zeZk=6d7SkYcEgZvl~RZF*MDJaOaix11sRBYZ1`*mr9R43Xnrkd%ts^qgFtiCHfQ61
zX!`1)D7*LXWm#ZJm!+GfL1HOEy1ToiJ0t{YSyH;CK{^Dbm6VhcL6HWL7LZ3$1V#KE
zzVEy<&iIcyD|_GPT-PT<WU!4wf2DTQWYl!|??)A9$ge!{t_~tQ$xy1q&=qsYWV$|U
zqJEc2=<!vr&9`o*oZ%{?s1M*gXd92&+5r>eh*v0N?gKbSU);^?i<;x|Bz{@|i*?U|
znA7?r=|0AcYKM<gM8hfkdW0ZejV?np!J2fBTl((S%U+)ig{c%=ZO>!XiDY5;@b`ol
zdhBpVmi}=c5IeLO>TsT#g5CXmQo>#1*3k>(FZhGA=X4GN^rW}VlayGCN4BYn!44qT
zQjLh*q&2uLx_jxi;jqKJP1NeP-7?vY@xMyxC}tp>bc=m_?Rcnnc1w@m2}w?dXiM9~
zKSh`Z0Ium$V2^nzhV7fSQQBjFX_oSE(DCJy*}quC8t%_<<xL~RjC*UZObkn}i+Al9
zbH}f{UAv&=8-g*uJD`>_U_AuXO0ij~TC@SM|MlD0tl{OUQdaa}ce;3ZtXQogp;mxZ
zvOIp4naqSv%s>Zx%QfK@j`S)BK8&I{1%YkAWi?Ryld{Aui0X?>muTK(;k_`9;G)jL
zKtck>VYNKZ*hNR4z4v?kK7E|>wud~DveXixN)50xvR7nS8vtE9|C~V@aBTeet>RT#
zS+eRFK&olflae158{%fXV>qcoi#*g{E^T@#zcoyn(Hg$Yha<b6{R$=b2Ew0lI)&mP
zm=1KD_k=A)jM|$fl|DBlj=BS^SgLSeV<l-#`$Gg>vcJ!nN6}S(66P3Aedd8mt2x%T
zEmR77nEALOEOk%Yu`atqxl@HxX@%tt;ndpLDrn<ks^}Gdpa^l4OV%{YpmHT}Fz-%e
z;~gQejrxU1m{AL6tSBSk<j(GB_zyhdo{X-{WSkz3yU}BYm+p=^EBb@mc!JO`R-j-m
z?0imGxmJKzvx(C-<IXt;r8-Z&gh#F+UqrrU#up`{G9Ne67*alZ?@2x0_HtXYLFPLw
z6|Spq@BR7teY9w*OKpFMmz(jpm*qdrwdaqcinq?|@2JKaMqMTwjsnu6@g=r!4C>`m
zwXMx#5GpeRg*P=VqWGeyOiBPnULJU`*j&*vmzlbks<SU~RHa*OPYZ3(99Cg>h5p9g
zby@xr|D@bygf=z{wPG*h*zB3aYX8A2N7vxVDBBwxSt{7$cM*|88?$wWtl`{n{b8bk
z=_<fgQ41*j=4dfWY~xzkQ$_KW3z*^|G($MN86aeTz!*M_6QAwEf+?lfCxx=IOfg5l
zZ$P-=Z^bULp75i|-b<#iS_ddOeW=VD`P509amt$&B+K5D0RCNeTGk}K@?a;ou%Hux
z-gASVY4J*>vJ?kn-LAIyjptIqLVp;0I{s?mJPV*7^&2x>sw)e(%TH^F+(|fg8`k-3
zXcL+J4Mfv%7HfcnW&-`9Q{H?IQG5&1`4~IdNCDIdM&vqO4*B>`I%i=?Q`}0&W6w;=
z`|A#_a#70mA4I&CMP~ekMRchAl>*h_<}aQ+JlkL;;AV;iRp8J$e}C9IRtmQ0f#J0e
ze%I6SaR2akD?*Pg0X?BFhds)fnPI0HYn}c@vtww~B+`owYwUiP|9{KR$Y}`v7m=V<
z7+<M;KBlCAm^N1tV}R&4BauZ@4y$J!vE+=LHX>L-URy06E_Xkvm8zQYgHo*6K%NN7
z#{VMG;}rn5wf_=rO3NZUMKY{5UcvA@4Q&7t&PfXBbaZ=eNZq!SEYACeA_dX3jQ#?#
zMH}HHFQ2=o%O?`38Msv)K1F{yIJTI|6RZH{Y0z=^T}Sf2ZTekx^Fe}oyBrZpd)USB
zbDpp-F7Hz<4Bl*Tr8C;fOQbe=k7u&u+@sjs`d`>cXN6Tfucl!>e36sSFKGKb%9r3P
z?Q7aQ(&u{dC!%hNxNbQk`-8?>PWB##>dBUH2xk?^R~hs-)($XTG-22c=*keFz99J(
zUH&Za{Ed-N*OxafFEkY+Y+%n<eaBoH(+uyO2n6|pSft*bU-(4Q{J?(~A#KCIX-Q0t
z#%vf~aNOos4Oe7Q-Y<5q#YX67jD~TY&I6FKiBf)9Qw}*puUy&07A2ah$x8)@#69_r
zau)n(B(WE7BAcZ-&Npqxb@+A66C6dpOpn;S5_m6kBjn0#JtR&l<LF7{vN*8j|I3fA
z!p-;xk-k8yARsgyy=Kl*r~jI7A`Xj#*NL6}{9AeHbeDa+po)#%gD{~*;1dTqyOM^{
zYc)FkGU5z&&6t*FRgQ#CN!dnVXV&O8(w#msJ`4L6DDnK!l*`n8;YZHF5jVC|MWH|b
zbglgYzjfyPavCMooCEQ`%;nRCVgkPeLmL1mna{ALIMlMzXd}Bh{vj)0npuS?qlt29
zQ!2)<9^@0`&yAIR@@*5U8P7?4q(*(Q1dmoD`Pujol<VldRK~qo6kpe@zQleFi4rll
zUZ+yCJK09^Z~K+l1kLTG1OK(0@70e%v#gVvdw-tH6mT^R36s`9ox<^5wFiF4rP2*F
z9jfBTIT5po5M?sVE3yV;gt=Su2q}c`oECN6qZgp)=z!$Jgsk1UVuyxC;}QcMW}-_x
zq!%WdpeC|q@mQ%tYCPJuP~J%4&BHH?(9VCF+3j2d(pzoqW7Nw|OptSl5~njYG?mP6
zf*iqyUNe{(B{2QTkz&Ym&KrK8_Hel4cgx-(9Hisa%9Eq>lrxfhKYw3C1`prE{LgH@
z8Z6m{r8RCK-aU_ZpKUXRMwmmCQIV$}Dv7G~i8+%n?|$r6LJO<vT2D!8qhv^l3%mHl
z1kFN09q(zRYT{L3A8klEm9+-Y5I@HDoKi_+)b6U_!C`|i8A4_6<OuV&tF#u~e}p_7
z{a|Yys~8=GNY-eGTYUfF><3fRKM-U}#SguAqJ+;f3A%)4$Q^)pASU~0v6=P5lb+!F
zp<h8>x#GU8+YFn@Vlo_m5O#bvF+AK%@1lfDe82!x>y4y}qu2$g;v-caN-?2iA4a)H
z&RfF#&BweZym>}%cGBXJ^!7ooYQGLM)m<KPFa0>iqxFnVPrraoMx0JFl^cx5GC#g2
zczsfEa3d_z{8%UA{r3<M?oh<|&i)@PZxV3sOT&9ZS_5I@Q^+{|>dgE0?s=-bAwB3k
z!*L>^EKkR)iX4JatKD2NJaNf7;>q?5#a8`%*$A0N$_&ncB3fwrpaM?&qi;HAe(5#-
z%sNk^l7onD8xzFq&BKW9B99mkz4A<MToZ!fBLpmct|O1g{G=7^g@0-~`eFzm4lH7M
zb*UsousneUlYl4702c@b2~=0bry=HV9zMu%x@Q!#!70jc*tSG+qQtxqnQOdy&~sv+
zw)+^MY^DkJTR+(Int|U>J)nMMoCQ;p55vj9d2-5&qg5+dgLuw9N1;j3nxM--CVU6F
z_~YbJy1D1$!AH+$cZaJIm$HsE#RNXdOTi$C+T<A};xp@mLp-B{9f3#ITC&$XSBbTP
zFN}}vT^s(pPKr*6jrt};_Mvfd(dI-zMRA!7I`T%U0{>L>Ndmv8Mhf5a=F0crWbKsk
zS%;Y0HYE6HwkrC4`_=z<cqEmc{W=J+Bb<&LYsEEZhcWR#w&YzE26SVKDLs95X|^Cn
zC+|t4!$FGY`XdKcin;#SH1awp3*YnRH@mqbjSO^3zfBm896aJbi`Tv(C?PvYF!-Co
z2S>H^%ypp6M-bu@r9u70@5#tt4fJPU`4jickqk#O6U~>8b+T&3{hU~>vmJ6uv%hY>
zp_ZhZ7xAq7XO=Z(|9R-G@QfqVQemoaJVc;2?@rT>yj}_^q8ocf&l7QtqPV~*DZBRV
zH$-yy*;@uR$M9wywgwS0OIqxPgGo@lSGd;Vw_@vh_f=Q|_c3TY(%3V3E)#D5#N!4D
z%*WVK;)@k6goB-9Y_xq!Tcdj!HY}P|4a$PwGO`+;zcNqKqUq2*JjLoykMbHrBb%X!
zcnt`v6A1UEMmK*L8vykNCzYTyfw}_wE~HP~<PN15?;o=iZC%WUo>(%;a7;_c@j+xD
z-XoNHBMjAYXzM-k0(#WfQ`0otzm%LCKVrVJm#hqHYCctv_u=}FH6kHf@*7qnUh2NB
z^k}z3!vXK1)T1Ty4QyMUgo7$#!|~wiBM@eAy+ZY^ODr`snNEGaoH}djErCJcYf`gR
z+f@IlkNr_eb<|0?e<-ZenknC<tJK`2$bFkCUaCC(ETltbHmpnf{jksS?-5zGxf&&U
zc=K-uFxA<=_T6OjHqZNp?;8U0as^R^ineqdWAV?HrLvZ>=r29LM>AXuS`tqSUpp|X
zhvj^}!<QJBsic^<M`3Cup9fJHpRHCg#kmuSut<s$OPCq6h6Oon+`((_1wKu>n@SLm
z9v8?bhL=K{R7~V6*Yln<jqK+$CfNy#Y}<Jud%nc%(KadChma&0$eK|z;zU)8;FEe>
z1|0cytR&!21{!1>{0kM5-9fSDo2PxfXGSSL&D5vFOB%<C`NG@CX!cJ~WW6`_8@3z{
z|ER^Vhrm3R30qEJjo8|ct=--D?JY&dQBBv`(z?G$;hq1TqOq|vHWAm<m_mFXHP7Fp
zTxyh%4LT*7#gbgjr~aDXDyxMNL-#Ayn7d<<Gw04t*G~1U)A#2!N*=-fG|@V*$Q0v(
z5OI>SKk=$6D33rR(@F`X7kPjI6B%>AKrEN+aSvE}mJB#u^G{8|KSwkm_xj(AVU7k>
z2p$%r$S|edYM?gTg?q<diWbuX?wp(nf;nzgqqYCd#(rE@oYjMW?h;^gNG81V0Dd2b
zhH8i(%Z$go!|d(h0b1?vrK_=61FqgTdkU7!UcknsdxtbRdZj;1BRUvVB+eX2WstN_
zWA>tvp-ff^nW?(ZdTkAdU4h<_nM2O4+}0O`n!g3ZKH9KB7iZ7|nc14I*l4><=H{$Q
zYCoC$&as^Nw9fSR`!TP;)w*Bw3k%m?{QKWp_U&S{(s77x8{<W;Cl<<z@@{|luP=DQ
z<TPZ%fAr4TR_Zg|<<e22f%p^eOw|!&=0-7x&f27$G<jNnCBy0gjJqQ}>3UbG5Q`@=
z*FN#dv7XOdug?X`$aq88kXbb;cyTCdF6(rQ-$QkHC}vn#3gm*+U^c<?XFKI%{OQlc
zBid?={~e%TFpXK^FQiGh9jOYWzwCF8ZydDloh?MDkzq~nWymW&=CQgzJeGy+3n+S?
zbqQ^w-pBVNQ_LpI`cQ_E^i&+hhr(*#gOphs>7m4p_=f_)wb3ffjOW`oNXh25O+Jm}
zY5n0FL`;Hjale$1i5kt&uwn;6{?ltsoL_khLrSh_)l-37vh}CEIfP|gBNo}rv}k8<
z37S|r0Mz7D8MIVPao&t6B{Rt~z7^!%yR{zWFYYT@AT6|*xf;}C1e$x0{lVZ`a0tO5
zp{?ZVE=nksacm|SrDES_X(jHRK~)1${O&3?*xm^X)@M$W`-_xU-1=23iM<q|WS)E1
zMCw-@rD;ZXehB82A?b%ZepLRmh&3h53ae+2@>)?mkw)xyIE!$@T;3n-*Z(wWj7^MK
z9yimFGGzI2p74hO!Ikg9aVl-2JY;HXc=Jk`K;2|CQ@7orj@(<nr}KRfhlSKYOUbvI
z(Rr@#!NvFxs8V--xK1_Bifvd`wlZsjjC<FHa@8_X<ExHsFVlP%u8X?72KCoz&jv_>
z#ZQ-!5K;5FgomWw`ZK?fagVLV%Th%4Ch*mngVMyQN<}_yXY@05zTKW^?i?<6L}Oa`
z)~PbjgSwbypdWh%{!sk;)A1ESFzlCbm9~><j67qy%j(OnWnw?286|8kcQtg-7V^^_
z{*4*MQWD=+J4g8(^=_frLkWN2E*>99{-6u6W+G#i(`Q_@l3H<%dxjV>A&OlL7hnH|
z8%y@s(m8#92;2DviIR=NB>%l@I8t47cDr`aRL^>{3>VmN36q?#Y566nF1)c*k()0^
zPs?N@c$c#)?Y^nzx#yLQy^u`u0Y$RSTW56Rzh6$VsX&~y!ZbnuyZY*+>et(|tYxvb
zW7>~Fml0WylGS-#Zql>nzeWW|49K++ep2i?vSqWUD_ReKn#Wq6QWRv2a?M_FF%S&H
zvT#=b0a*zjvDPG^$Bs%k=1B(K#5VyH=E`{+k(r!W5n;dHeB$!9CNQ=ohbO<^vypJw
z#C}hI54!U8V4sXtqMAt|riokbW)u-qbS!W$U^m0eYLa-87nfs_c#&ZI!Qi!NZ%Dvi
z{gs%tBaIkt0KrHq1D%rvM?(T!calC|?o*?~E6X=iT%Qq!Sk-X}6^;<$_?6)%#HX1i
zdVLefUF4%kJFf44v1>%FmaWy)<Le4^La?wHa}xPZ2bDT_v6Jxgr^G36iKssU38nSn
z1Im7?AyQe*83NMJYuP#1@<Py{upn(I>xG0wr>%j<gB~qaylD-}_@xeX`wSkWg3CQ+
z{ey*aw&4_Ln3^<ec1ikmgXBg-K}#Kn9W)d#^4h#_E`J!_|B6shQ1a>Uhp>~E4a#rx
zIPkB;pog<>JeQ(kap;OZ#~)k7@BHz8|5Y}EI-F>QNs)|JB|+mmcg(Qm;G@VYRs-j}
z>Ajd=h~7}8!OTl)XS_-2AbT<~%?(S1{*FeqHJuHM2S_B$k&j?psnFft9@D*X8(ly~
z@K~64YS&aEUIBAyZl<5me_HBIL`!n0B;K4rn1qzvVj73pg|2@u_cIp+tHE84U7E+)
zYt58)4gRlITkBd#z;Q<Bt;M)a`^SRtSL|q{b{BRee4y5n!Hq9-bn@?tI;({;25`+k
zL80_4jv=W~@x}CW;|=8Np)UmN!r3aYhyvn`9%czQaLfROfUxcXGZyr)d3Q2jaaAws
z9(RPYIKLYQQKU9{N5{Vmfq+KJV`j)2@E6a6vpBKtp+oLh!e3Kg3?Hv6@O=A-;P%%<
zp8{whO&B6c`i|XTZ(v_@T{q_br+xj|2v<^!i6;n_uN)dylB+JyYl5w_%~|*;1Lw0y
z_yhRM*s!Oj=!4Cwua-d}>w&6fCTJ~X{nqLNSAEQ0g(+eWY!u}B<Y}7e*54`2TgM`m
zDLqKrNTgn~_Qeb&uD%%)me9dK@|hjasuWd9N*^T$HkruVKPydni-amL6TuP8*mX3)
z#-OW&8ue6&!o0{oFEFUF^mX*R3mq}mUP22AKQ7z$v+BC|5JGSV>wt{0qBw=WPbsku
zg8JXnSo~D17q4cC{wf>}D~)A06;vB`-*>RDdI~Y^%5EQ<ATGW4kktp+P&dD@zZmNa
zZY6&>^fbEaK2ox~K^dJ4sw9M3zR594_MT>X3V~i|IR(zD7!0B*3bI28Vjc%3R)NG6
zDD0SpXlEyOP~O-9q+zv2s4;O8#3OXtu%RBJprDfEih9QM5deo{7%lDYs7LVrMCib{
ze#uLCi3i+iomZk*GHWOE7E77~8CHWtZU$Y7>nSVaIg76<M*?b^8<c;nH|S@|VU!co
zS-6%wtu8aE*WOmgP=WB=@Q-eV59aF(pRRV*Vpz{CzvDHi!ip`tk4AHhiYrbT8hxhl
zb${wv`G{{8(J4q1VxiWZV4xyq6#0Q8tu|C}q6%&3BXv?4-H3|y_^&JG3zS`e`~IL|
zwDFf$5510_s{R9=0PC>kUY7uB9vGK9?oSGN_p%v_2?O75ASe8>4iUz~Z{<DuK}3bo
z_wfcn(K5C9B`uUV>y*R=1hXWe0I`M2324mmNfl)~3A)q*4mX`4v&OmPc8OsQ#E=Z>
z-xyss0gDS&P`9O2On@~U9(m7eSqk)zHcUpbh{8(WF#p9OGVET>splf>b(v<A+E6lH
zoGDW*25Iwm4nZ!4Pi;g1oh^)copZA!k%@xGr~~MmZ-Ew2a|koDsmr0jV&U?bd-3dN
zkD?1OR$w#0H0u@r;?-GBu$({$(5T`DW%;fL*yV&9jFEOGtVe;Mj*YwwKDOFXg$qL6
zs<nIv@ai2!-c%cH!LbBbm?iapw<_$@pZ3B3Q}3&{{k87@FreuZVa?;)?3++j3l(>0
z1lWx9va#{luH-vD*Ur)DXerc-{(|h7TNyGvm)`FsC-(7D`bVOqUm@f)k$u;6s*i^)
z$7;TO_;PQBh`B#1FH*ml21$iTWiI(X<<GQeogffR5{>lJ@lexYay;-HD`Hgp3%*Wf
z7&Ji+c*;Rx*!F<M&k}_{!1YtH$GF;DlBqEmM24mClreal^j<mu;|nAM`6CY`A0L|a
zJ&~?F1t|ggVO?S_eBpKLZwF3q<}JOezCPoV)_Ztcss^MZ8GOAs9a;UiFA7(GV589)
z4C*7X9uliLM>cb*1Ya;&1BoYwgl<6mB7>!3Pk@C=7l9tXn*=J}{rCQnPOxKH#mXr8
z;Zuq-4>dJ4fLcnG+@<O=Kgk<Z;sWSlJ7BzH(!)T@-E{--0gc^M0!>*W+v%d!5;w_+
z{4OmOcWlV`_2A7zmm{}xoCnzBX^v`@pHz<p{(=%O(?)CUlvqeTkB2^n)AU6!6ibbO
ztM%$8<o{;@&{rTS<~gzbzNgTMby50gCX010&A&6$1V6LuSO|$_klkPQ*1k%1lG$3-
zU4iB=+4G`Q@r4?0yudIwbP1$stj}>ab#fXqSXC9q0zl2?qIgW#DGBrv4Uz?(@>PSP
z2cNTlpgRr}Qlm!Gb8&QieOEzY6gYuEQ)GQ$=~)3A_&;x+<Z|jVHsU*1NU7G?l#Hvm
zHTcK&KP6#^^5_rDwEV<>?8T*T)olL`5Ee(kU?4Tn=K25tOJI`Ec*J<gtz#n}#(P<N
zVpQ|>*)zse`ukA}sd|=IKN|WeUfgTY4@wDI?J(g>;$v01Y@mqg*{5+?_@MHRe~$MJ
z;bhR?Fu<XKb|L<`owju6SXUq~2e9McTfTJG>rO~GuuUl^l7Hgh{C?<RJe{cSI^A#b
z7F7BHRet_&9#1}>q*IU&^(~kty{06@9b{n<Qt}?Rq|;;*6CR!Kf364K8+n={>H=d1
z2?YmrQ_l=Cc>Qz-ya!kuL9q+v04sT5{0f(kjovC16KP3#zngQPf>CSrNK@f69KY?c
zPYve6>@OB&a}V_4LD*RNiB@?}K|$@B|7UGS>uz+;Q5<tqd^xeaD8<!72_ddyV63>U
z<RT&{EfbD{SdW9Ysa6Ght)qp;ulX(|2_Ia6zWO#SuA>g#Arbha2Y|ES<8!ip3=Y%V
z%^+4e23J$ldkb=$co~U!2s^I&7w{7g1+w#m)ki*TM*$yL?S|t?Z=G<zl+JRL8vAY}
zbPZ46d?CM$Qxt#lg<jtQ|9|gXoE(sm1`i*tyb-Vssj4HPQ3+yw*mE)>^5@eImK<iA
zRd<I52V!7TE7#6TKXTi+M6R6o;0H)Mp#SzZogStBlx>@dHP;6A*l;SnlKQ>H_O3R?
zMYGxn$IIyRDX<>nuE~@?8j6PJj{^?6d=ZNW8@e)jyjPN4O;abR6_i2X#wn`p&I^-@
z?+4NfX)fa+rSg(}uQy-fJlLdoEI*y!R+RxNRu{PO@21RgEW+3M*}d5pB{aGl;61mL
zG6*wh)g^MC_wJ3rbr4fWz0Vc+^|tn*#w9SVBZTH(J&+W``!HtKLg@s%NGF*vxdF{C
z?5zyZq`R3-6y;rjwFX0@W-%dkaBUdrOd)gGvQf)z{aWEYFXZc)Tm?CBkW<4EWQ<!4
zI-T)y;0MipNv!JK;&6MXPAuu1I2*jN4m(k9y?a@6>_<0W3)+7cUKl4O>hHcOksgaD
z9By@mS)+fb?>OK-$#LL;k84hz74hno%Jnpc`7$1;P_Etl{qsk`Vj-po)JKtLfo@TM
zuVF1>HV^&k^gCZVK;Q>!q1u6N*?paRR$g{z*oGACz;zGibC9N@+LRmngQTex-DWZy
zCl2Dxc-xH0z{wEP--p1Q9@;$rIS>G=MSKOC8Gu)_OA3{+d1`^=g_K~_lS79z>3cJU
z#HQ!KgDpP}U)H-H9~ti~v=Vv-`QbA55)kf(+9zz!{)6jPQ~+DU>%e*v=~PG@_VoZG
zl@P;8sRwiRF3>e<z9tKXW-3oc|L4Sax?Tx0WQ%HE`j(O6RI4gPtF!<L%+1m5pf5Ns
z7+zS*K=$oV({mDasLx`J`EPLYBdlOQz`V2ftrk|0op|dTly2Z3G4Am*Y|V%N8;ha_
zh(V+-*pLp|?9L#n$c<X{l+7m&jP80XT_#4K!f_NO?RyHIp=Ygp-%*ssvbQDL07@@L
z5!Wo7MH;_giw_`wOuD~HL!h=dBfa`SIUbaVx8DP<$^vcO?+T0XXRHZ4ESW{JJy8wf
zMdxiF9}}8(_-p_ZEeM3bejG?2bHha*C7r?f5Rsih0v-qPI0z0I{dY@-nk2arso>IX
zMzpP*0fRnue67J}3nc@Tuj0orM4rQM(<aIXAlUPeX)pgEAE9zvUJE|DpTmXJu%03D
zhaF$hqDSz4Np*LW{KwV&sg&}P@TU-`DX7hP`btUAw4=Ea^bjAeTcv596!wPY%3%HW
zB>+;h!|RfpzCM?d;J5>bT`Q$vFi^x#A1<|G4d;^WSAHbm^+P`R$p?$VKyY`!ahI4}
z$;?G&#jxIG4Z;Ay(Dbr?OcQ>!U)>%L6A_x40)-owGRem*tD1na=mn@*vdjinjA#zD
zVE(C3-wt4ICUOOs-i>a*rKS}OhXL_k^7AzER$z2TNZ{pT#$J3@oePxSdqb3kTLMY`
zIsc76y#xa$2x6UH5(Wvg(SZd%@Lc>nzfuWmP=KW?E<N%XwgDBe8{GYtdvv0Qi0s}Y
zkY*$=sv#*ssKZ1)He^#qJ2a55@5TZ<R9OdJ+uLT!4eeT^y38-(^|F$rr4!qrr6a3K
ziQNuMvXoTZ*(@pb{x0?P3~s<DfL-2UkZ`m}>L3$$W|rJ}-7U~5cc?NucVlZjy0*~y
z$i=;tx{?oI!cT`ofO7vf_d=|ue#RAjhZ$4$5!=Ag0M^2*p2<wO*oBBZhaC?ZytEMA
zbA7qB{yoRyFibl92C+6K?g2b4;1?_V+k+SNaPB*F3!?D!*f*fl5A>9wk;HdLKiBAT
z&OEa+Z#@B)r6C&e?MY#JITB+a%p(ai5_JYU4oD&y@2H@T2x9BWs68+P=ZqFgu7HPI
z?eao3jj<ObK7+Z$MFW}QK?+Pll9?|QwVK~uhzy~XoR`2<sJf%)n#SLE?WcpZa~!7J
zFtaDzV7;+L?!lkAf9YxufX&Vj!qL9qR(AuvNZ#*&AqMj_e&rHGBZPqaa;YjlYJ%|5
zb}bRGEQI#lXl8Sf9@QTgarVSZ;0|9|nBEU6!894)ZWq75yvpHb>5qO4+HyPW%U{(5
z5me#X^>LI79Ms!JfSxb1aaGoI8FSH!mWa+Gy|Bp(Fgqkx7CV@FiQE1)5XDuNYB{xk
zHJF}z*i+jJ);JQB+zdZE>DV7;-e->K>1YiJSasnSMnGP~eB_|8HUb#~hnUJjJ4PPR
z+zSX!QfiM+AQ>5ax1*1UO(dCm>w&~HgQZEyxEsHFM1uy4C1LSd#<kiWix?fz_sf+<
z``cJ&4G91gF+EV5r5?Zlkr=Jm!=kj|$2LX=ae`$nZf|7#`=zivH9=>u3Q<e|?0U^$
zo~3WP^v?EZZ89IWs<w$az?c7kbrQt83V$~pkogQifQ#YV(n<q&AW*q=405z$FCfdO
zOGdbu9Q>92zJmBXC$pq$1JO8iy>#e>t%Mc<F(>U=;!KS$6K9CYb8vjH#X{cG&b61m
zsRt!GPGjnZ#s{hf5a%9WaOWo=ExP>cl-1Gbxd!tryrPUaXR82E+lAPS{5y6YWp>Ed
z`eE@*3q+S+&*vA%R;qATj+H8JRTqQd8yv)iKw4cHqURvXf?@v+q=anlL^|cbaS%-q
z7sPR08(%17NdF&53u^nZe<zO<qG^5`D{I@C5bQjf+0NB^>2Z!`=}(2x(J1+<Up*kM
zZ2DL@&obz=BsbI&hd&-aS3~SFroA;j+AQ#5gqib(zKvNe(cv$cBoKZuFwj9|)j;ed
zH+O7tEr?wr_;|d<7a4E_kOuqTS{tq5-DpN1m_N0f>qxA)>^d3y;Mj3gUT^RfgrZ6G
zICa|z#cs(4g&~L;B(gj3y_8w(@DBv*V)g1Wmo!9ZuCZ@xlBWq+DasnkTA-Ep`Wife
zLYgytU~@<Q6xQFMy1K{rX4PK*m#s)vI~VmC7vFUZBslnTTr%JQCW=j}qMo2rO8H4q
zY0{6PQu*2v#JSo3Bxcl;vF9s?$b+V-`xnY{)SfQsrSI|HhwbaaW9YG1glt`gbcP4V
znAk-?TIY7U!=}ei)F%gTZG@n@bY;->@9&KXa~oKy>9QOOjPA~wsi^wF%fZm%;F&Ev
z#_^XI;u=4aMtho${35JU_DEszdoB)I!z>qjKsrrCcaMmb)DGh)8&o9pl@(H8kFY!k
z{bL5DBT`wH%;*vbqsm89eLBJBw^)Bn=F2hKwX-K!&>lk4qO_ON<`#JV4g)(jtUYi-
z(kJM=!6hxl2aefo$wUjkl;`q-&vuIeYaOF{WxLB(4qmuaR8zor6RBQM-Ufz@Wl&E&
zD6zk&cd*C~gZuAPMNN8yQ6Fd1Jxqo<Y&Gr{r*JN~)f7A56IMt>vt6<8MK61S`a+(B
zplSk%V-hF}&&8Bg2M7Kgs>_;E1y=v~+OSDQwo>VzvxW_PPt1+2h9)-U%vMWD=|$bM
z-(8%-%wQ9R9~}8kcYD-c=rPM}-kl;As(R`1{_~Gy;VJY><v8pdJjOaN*yq7>EWGrz
z!Jqn8qH!;$-&znvLRjUho`GGkuq%do0K37t)`%5(0hXFSpigfjx-~PHI03cfuK$VY
zI04~*4@4RD$*BgrDNxjIZBRuHv2yB9f=&ngw6`;}a+q1?Cut2Ra(6Zs+<K8X&&Vte
zD44RVU75X^On&17(nU=qNS0Z&wcIa#dCY>maw(f%MiKHHSG4TqVHqpMkV0B}JfG-4
z+HH~!yM#nXeAYnh7jRmo*S}5$e-snb-LnEwan_UuE#Qq<heBh>V|~bWp-nV2Q7{pn
z)GV$q<;sLmHBbUbNNB9W<{P2cDYrVhO79EaLf}Icl|7JygDZO#fG5Ziei2V4$csi$
z1Y)1iI^%+t!ph>*<><XFSu`;sF`zR~iV`WDc^3gkP-=de2P2~yRk5Q8FK=bZz;(y~
z4iz;O(oxBt*3Gr(L?3l$GZOgV$Og-U=4&kZV&&8j$NIkK>62;`?nrtg(3#%S9Au=w
zm^w<fFAFw3jZQ_T2SDne(jk?sKqiBfR4Y%!6;x|=)2B-Qr1G)%-G#d|A#`yaaPnSy
zb;~nlqmuY9IpyYA(!S;+1=isRVoh$1!KC=MH?=%viaxP}KL(!6@^WimECF~#5Ou{x
zOkB?1plnyqYwMdcrv-;mV14&^phpUCvNqylwx!I)C6o*pP%FNdy(6f)3{6I1&41BL
zS-@|gn0VwHQohB<nr=;cx_-WQ@YMs(Yu&TRb-gJ%5XGkpmQQBaiY}|BIgz$4cEUk(
zsPj;%XLf1}r7~7|nnJ%9>^9RW`y;4!bo5;Zx;QF+D{6};(QRWL1n)^#5x^EbW&yMN
z%w1Ci{Zps>?u+MZU~dGWb0aHhUEB#HNT|F5#9mZEC#o{TUyzi!A#UAdGXWqp(uq$B
zzTL;Pm4lUYyF$5yXiYGco^c0sa~mETokghx<;64YFElc7E5n>Z<=Mv9Nns%1TaCIx
zH3t6~bx?N#-CwYOV>br4VkhO9n<qgdLN(=?m@!|U1?j{y@=my;$&@XUz{C}^6l+An
zL^>SOIf4!bIn?dq4DKAT7z`j=No|5QVkznB^KB68bALhhpWAAOp!X@8$!!6vIeP?X
z_+}OwEy?^@-(Qq3>Oooc)&2z&HNzLvn8~+bP3}~DN^-j!gF7!_1zeR@K%D#~SBce`
zGV)7z;OO0@S2ND}1+-6|frjC`?E=E=7|E^8w5OOxu;2e6`qr)wUhn1POD#tFciDTJ
z@;`(-r8)QmxRS`Z!QDij<k!s3ei#eQ$@s7NVLHw_)kQmqWO)p(_H;_+7wdHnEIQA$
z*Xw99&{%S%p}&4)y!(V*nqg!s*l1LTWG5`W+XZytTUb%D*s^%W^;@Fw{$PP_xiGxj
zwfA}OZa%1rY?5T*d8zS?2kdPup2SBM;#u-xV$rV&SDPJFzgruxsG2@KL39A8azfp3
z4%IDF1k~Ms6Ystfzm1YF)g<5<v&e?Jc5O$%4;P!Gk0}qfUQ`<!{2DG(eEJs9p0%1|
ze;7nu^{8DLh~5dN0JHw2YRszsSydXi)GIL<-X?C$?un+FTquZ(0}J6h&%hZy+Ge|c
z_3yqZ24qKD5jO=QZkfwFT?8<n=m8mF+kS{=Qr^=scebh%AiVE?6H|5tSa5O;I~9xt
z|FTHPF)o*0DZ~xaGZ>_IuUm<r^Y$I<epvbO=g`|RG{vpGyUb8VI`$-4yGUH{MOuNV
zheEbuSitO6=@X(yP#?W-Mj`lEWRb_zZ?&H(hWY@D&(K>J<-gz|wbKd)Bj8JsiwAVp
zoM7Qsd3K@cwF$)c9*YCx?i)6ns$6a7pu=V0DRL@guFj4cQx8w)bV5NAW3*#Sq(c=%
z7-u_KD>HU#fl%D8(8SONhzhv;(_v7_z#qd#O5&5il#xE{<o^2(lvlztp(P6IC3ajT
z+DB|tO3H6Le&B631zs^JRH1(XE{D9Mn9^vRemqaAHX-8&^7pcmRZa_j-h_Y3gsy>O
z&*Z<0Y|bPtmCc0tMjH^KcGUW&0RT4Ur=mMD0cAD_%oLFd;R_EMcB^Za*V`hb;$R@X
zozmcDAB5#4`g^r09rH}tz=`QE+g<q{cpfx1_!K65D>_9So8qp(R^}iSSOhAz4%5|U
z@8om|w&~mI6Y2SDz+pI86BSC&h6xg#XILM=K(wQd)F~LB3m&R6c0U2&qSL>tvplP}
zq)&TVK?I+k&7c@k^J5z+2w-IbRksAWII~7z{4P@?9N@k!eO_j$k0xT4d$kAr=j{N~
zAxi#lJN@!?jdo#N``EaA&)S;?W$Wjt#_`+04E=_`z$#35CBRu6oIqu%^gNq_`kv<^
z*a5S3{xSAlfOj^s%2G5=T1FeG3$ILuIF%(*@)ex!bbTy@AG<%TXANIkludu+rD?iP
z=wN!sFVajNslaj}Yhv+28$-Izh<SZIaksHjZ3?W8mTj)`;s-NTn%(~U$j@1{wp>w_
z7C<F`8TkvOoT_P*5$b!B`w`!FD`-&0>aH&~eu|_3tNIUs8|@&A8Mr_8ayC&K*Xl(Q
zFoQJ%j)n-kObxG73sVX7TWE^j;t`leQlO@itR>HXOE65D-mJVE@cTh<<x?;UHGqU*
zAS7bU7;UvL3Ya%vGAIi*3Bu3efQlC`ty+U^9#}nX+TAIRNIm1g`LyyjJk&Mw^PyF4
zC>EA|;y4?C8H=f6nl<Cn&n)TC6BdpO|B+umxj6bdt@#A(`(2O!qYd+7_>A~H0v?MW
z*@ij^VuQdwtC~u${-rI5`1;A3&0PTzgC@^@^^L$q_C7=tLu5lcgBOzm0}wW7tda8=
zjZp>z#9fhCbdAw#pKq&J@AW3=js8$9892gz(gFEMMIU<D%t&Rhxdb||Q9{J{9~Z%=
z1p77`nX!x=(fLH^wPNuSiLUqtEN?$=b28CxYfq^+!&3FKFeUtV^Ih|#YjF>nbeaX4
z>hhI;SMU3ev1m`hX9PF9R+Al%dCRuwf)$yR1)c**;Nc?Hkuh-BT?<tW&-0r#VR}<-
zU^zt!{WvhLsHxN@kQL9LT!oT~f=qLCv!VJcK^kg@iosrb?;r3=7Z1jgJW%-b3!jW5
z{_8R(5IEn#(H{VJ5BP2fvo!@NDxBamM=dwmDS8sf7%SPkBtbIl5mfvj_kAVDoNhfg
zmW}KC6;a7R?-u~BeQ5^B+A+uZhmXgEwmb)`ox<$MaMm_coGLBP*9l7>C8Pj4A!XkC
zLg~Oy{s*WKd<q+BneBdjs3C6dGMTqw_6uyjm-vwYtSk<X0LU!yzozF(xee9|7;z_}
z6M#(lwkUu`?BL7e4G>pu+Cn8U`vedLv)C*T2g`2$&Uw8JyL#BULm4s7>sG9S`Q>HC
zJ7{Gv;TF^TS#~8fk-2KbdiW3N`|d#h4en@8IYL@t7LSa;hni^`gLkngJ_Zfdh?lD>
zpFcALM9@?YFL9I^^mVjp)bBLe8*rggg?)>zvIXhkpeDMt@u3L*U6`i|hMM4Dp<hbj
z5)yH_igcV^mU`Xwh!iW3NHC9r>xt51OM(LZ`|l;aOKv7!>R|m>$-m*GJE^3OqD&r9
zx`QLiB128KT}@fJWAGtPC}W%?1ca9xk_#-=7kTsn00Bg27eAW1KSjKeTLOTS_gYRW
zQvgo+?}OOaZDK6v?~`hX4*kIYmSZhtEZfIUK~T!X+weavUi%feXWyn>mfFkT)?`mS
zm;DEz)5POj`UH+8MrzjLh<_A0Ep<i|Mh4hcNdT#Qp?~1e83(CmS$JV4SNH+f0c@bJ
z++6<1f$u_EKp9|+?Y%B{9>@26OA%YMou86ClDd<~nIRMp(MR2Le&MV~N?a@h1&zVy
z0BY3=B4w;T4}EWY=M<K={@%h-p)8sjzCD?L29~6!#rz|I63>v(ri*Nl@4BMtyFA39
zBjxg%FMPJYPc-gX@@~b7R{3&LKY*uz>wBd__0Hnmx&S|*NL*Rz3ppc)4#PV}_V7pf
zo{%Q;hZ00q#th`ngBjJ<)W4h`&mAey|2jq__#mH0RI;v+jVR>=s?b4$t6fz=0Rz`!
z4%mClTb<3_4nN#WpB#KUy!*9qvy=B(gEBw%tBt`_%)3PvTb7?kAfkQ=KrU!zRvmZH
z6J<1TTp{tHP>k~@u2>6GZ0@82vnb50AxL?9chp6WuMkY<)$?huPD!Dh#r_6ha<(Gt
z0ExSWV&1!-)|cZiNVaXL)Lx6rasOO&NRt;+T>A<XtG8MHQ?>#l{ibCKv7Q2^(d(!t
z*H0`qs!&{Sm7WID4dq*7Em2!;tQPYMz{JbSFj@Tag~ws+9{Ru&UJwGsN@`Xf(}2!`
z*vSN0S)9&snkPWqX4vtuQ?piZQj+tNJ9u`irlrte+<tm8)rp#IMS!g>Xb!|zTLUo5
zK<CyBrc*3&m6X3GsrFk3nF(Lbruiql!6QQ~0Q7}q@32sh{mBGXM>ppl0$v`h-u$Mo
z=tJxT-h%nboH*zr0E|dJ;0KrbY6>92GSMu`i9uFs3ExOz2Mdk5>P%dsQY^YH_OL~=
zbYAkd+;%RUkNyb=HBu}t6rkW8*vURnLp(l#D2E#WN3cm+-{Pp$rFjo--|mt_CnP5P
zfjNdb33O{q#H4jmK-c6kK_ehh>3<Lr5}_*Bl^*Dr#WtV_5r(~?S0VUXH}V~(OlS4F
z@qP?al58<|Otj*hZ!P1tg+XYsLN)E-Y%M9J^iy2pH#Z6xiMVp~uZ3(U6ASmBdZU=!
zjLF1|!99R8+u1b&hXcU^?kN%s;rCw}OjN)RzxF-<lJ~RM6u^BahZ`Z<@=Uc{&EIQ$
z%E)r%QpP!k466jQjI$1jqK^AOGQOjLQ7CF%02Z&c3O;W<DJz)Tw*|)8Uk1R|IW4*$
zzx;VI)h_j;W6O4Lh<|?-p>rN~d2Dx$2X=)lvf}QqoOVherGFeavUK~&1%jxod?Nn{
zxeh0@%IR>~t<z8ePZ>5=1E0}r8USY<&4N*7HhUZ}L29+KNJvz~xFi<$Kq1j03)t5W
zD$lz>tvBGk81DI~KiFv>!K&qzi@~v;2eD?M#Fpc%utR0cVJjf)R#s|mUovvs>K6>n
zuNX64fICErK7)rl813Wu(Stbh?vVoMNzm4Cp1$`M0NG&5)57tX*5DeTSHPSPQQS4r
z8(jyxeE;v@Nx=|7fA~K7F=6N@5F{Ss^+k1JO9S=mHP~@wGMC)++quM3n!JCzEGW;O
z+y(&^W5~bSx|r)lc^ofVKD-H3jDHf(S5EF^U>f{Fun{Yrbqyp8m1OeZ*kvc@fB{_z
z4#zrT?suQ$$criz=sAuGoLD)DnxIo|*-%E?PXcwm;o|R4I0_t}z~fcGJuLF725|{s
zJe$z}L@%)x&xVZN^$lZ&wEO|5AEKdyzvThD#lRa5z~QK6V6lFG%yu0r%G6wv1<%Z$
z9C*m<q2KgP4F5mdI_BgmP>`f<RZ;zl86rG*d-4IYxl^d%Z*)^lO@TAGi4KpJzcA9>
z^SlJQMCW6g-B(?-eCu#I8wMZZs<l2+W|7Pu2$E=H@QhYWc+LfJ0nqK6H)9Dh`p|MP
z0wq|8Uoi9)g92pWbD3=>rQgrPv{-fe-JfZf_=yi?qa~wM9VfMp1FWkB0Fo6J#2&83
zzkD5bbt<E;VE-E*3z_ot4wFuQYCsvj4RhUKNNA`f4>8|epk#1*r{ofYd4uRrAAP;p
zx1k(q9*im)Jbb6~-tyPeN>b=JtIys?6qGo#Psj8UiRHNp_&OG?XOPfM%ieJ-ohc|X
zgj1G(K#R;w367hriN73lT;K2G7mBb<OB#;SRXzsUYlgF)gE9Ev{~=nt!(HVEtcxL5
zK?&bHU}ysyK+b?exa}2AT+k2W@Bevu;s?^f{()fbs+4ftQZWBO(Fq2qOL&{dLh64q
zKR<=l@Kh!^@&cY<B`|Z8+-%%c(C(A^R}_ZW;<4`64;Jpwvz#E1#bi>eX9sQ!QTJK3
zmt<70Sb|#RBB20#4+J8mxXZI?Sx)VO?GMQ3o=D)<d6xRP>s}Dtr&~c;EaK)G@RQDf
z|3qVhBv18(djo~8)9xZc$F>d1dA!nh%ebk3cun{x_P&P=iy!AZuxW&ZlS5JR{gnX#
z-+bF=M51Mzfh9?TXHZa=T<xcS({+d4iyS&@nJiI}HlSz=cGDipCqNYdtOxagyZA*>
zxhW*o4U*c6bB7*-SlYSPDT#W}jmS8BR+*(k08uW*i~QNH24Ky9Ai6C(X79WDdHXRC
zJg|5KErM|@j_9O?B|TWXSmN2RmH=Ts=@PkRHtC{hEkcbJE62hhSXk{Kzin8R!O3w9
zuX+Rxznw{&h9-=%2S1c*0=171F&0me0x!W3*R|4Nc(@%8v<N^UZCd$4juRcYMg-Ww
zOE^&v)<NZ%jy(4Rq%AD|Zgw44|JirKIb0HYapy^DD=O=p94#-#0(On;TmC#ld%irm
z|3`v+bJ^)Pe0{bU99XK%Z&z0aTY*i{eucAtqIN6Vb&}uz`_JnN@UG2&OrJ@I>Kl72
z^h-b=eVb}uGvkYn>>p){m>ks9Sm+A)4u<ULloi#Gle}n(pJN%UYuQ(T>H$Y{=kJ61
zFH~t7-#`UgW!M~unBBbDqy6+AzOGSunw`s@?TJz?1(b<TFRAo_6aBT>{tSbZ6Q3WR
zi+2h%i4FGCi0j<1>CGGFvp5kTq>#BKEn-^uNu(q{(pY$VduzVzx!9~*i;-)ABM5R2
z+1-x<VL{$8r7=PUYm+$|MBd;Y1m9v5m4t7rRyZLJ(PEc6nGPR{Yh<0D7la;7o*!aK
z5$aTPC~rmeKbn^2q>sVml9v8pq{{?^aH!$ftp^bdUVN6X&#Qy-pB?N~ZjuY#trwJl
zDrPHl#2>N!f`(+GqT#kBfIP7Pfb{1KJ22-S>xF?tjkQ23R<`YVa;oFEwfmFtuZYo5
z*!cke^wx;z&~K1@dJS5!4wu^rM{K0GpkW^OYiQpP`zxm}ONhmMmiAtG{StWB!F>Dp
zGna!hcT;Th9{BUal}=UR5?*;qt_fGHa>=$V6TSwV@PA<v;Szy-n(U-rvnxoHg-N=X
z)!fO?#0aswgA!X*Ff?^IP}{Dxd6LkI?15X#^cm#L+TM87)Q)#Qkx$6(QzXI%fn`7D
zduLECf0sI4@0b=7D*WMzRR_HvIJP9ljJBb(v<l!1N7*v>lT0sEv$T~)(V?0t1OKCb
zR;`u4Id#}**36<?A~ltt*rejM*t7wbHo&2s!JEKrVhBzIL9;&4>Ghz?|D>UAjNw?6
z;2m+Z5dj`*-3A6ugp&J@cGvHm6%Z`ip?(}YS7{2^vU_AuIm-*dn6Q#LDxH>)WpZmr
z;6^y|o#JlUOAzk`@J_UB`-~nu2!iz@Bvg(dff;rv4>>3YPv$iU{~3D?`gqWFdito9
zoP9AL{x=0`<zR>Z37S}*E=SjJdPb8s7l`|^xREZ_%;Z{0!yu6WWC5M4Jdi}S*pDpf
zvJp1`H9@{}z`-&gGoyIfrq%ZWY}@!L!}=LdOd}8F=gZ}E>az)ipybkZR-LCHEw8s>
z8&I!YL-GN10)*U4F>(;Z?WMx}cE?K?(xnz#iGVjHq{lRZb656a(P0Uy(99VX4j!Fn
zUhhTd(kq+5d<M9!>LuOD_dtd-K}PnX4kQTya+G|Ro1Ok0jAxP)IIJ_krjBN1{(_C^
z2lM0=7}g_>2k@2-6hzoHe*mnc!0Y`vo%lC=HJf_^=gj}_=LnD7Mg)OB^^}O_6wGiO
zf%!=<d#xN6m8nItZ@+yL2!Yy(^@qr>S!wh!+@%k)N|l^7<evSp<m2I&1&c?zm18}t
zn+K3?IKTcd<B0|e%*EjuN5*7M>>{O(RWL#`G)7M0F=EX;s+G;k@C%I`H6Z|-c#zMS
zszdM_mKktsz6I)wyL_MBh757B=Rt7)a~;lsw}2<Hoo;L+<bV!SJ^&CL0884$<|?)a
zR=oQ_WPyiwqD!t($6F8%<Ywx9pb9rr#aK~TR|dZc1E|wNqs??Pe6XQ{Dmh2#vt!d*
z6(O}f#OaStsrGA!&NG#aw}jgO>-}ItB?_m50;whADimbJbpeqcyn^|Z#S^afwFz~r
z6dwD-Etm6h^GCyf$IfkFwhy;o@|-*I#s*8iWS|LqACRroRd4Kwlh+s2K7c`Er~r3B
z^!9iBBkTJyUXc%;tSeS0Sn=tVVoc?7={MvcjH@7kDO<p%#9rO!t3t?&pctRYu(llZ
zayGBI&~MKsnLEIG>6F702?kKQjkc49Gjh6}y9PpD3ne~jz~3JVx`Tn3gN=+&`>aUO
zSN*leOA@duJJF}nl2ST;KP%Aq)CQt8eVz}Q#(o&v|1{T!7~_TZF!b+y?Th&5`L0x)
zqgnady2UGB0Xy$yO8v|)SS7lvk3XY*%;xM_Tbr2=cquhuz?$%h<7EQ>r5N;>;hx(J
zjPYZ{^VGiyiJo|O7?b}Ne{nINU;6sAtfe1tlUiN@-6w_-3e5aagrK$00FJKI2~4&b
z<T*YEP@He!K~gEJ55Oq#9=xltZThj;yafnO!oOz#6B(aNtP|Q{trhG8gHPVq>PMNU
z(&TZf7;u-J7d@J~0wJR_``XRQ`=wOL9>G7IEJ;v=p=@jU0L#~wBE7ak8|!~SQ%h(N
zqSE=GtN)muTyB6Jk;`43@dq$bJ6PYQi^wHw69r6qf4-l9jfwZ<eSM7^tjzL&wRjC0
z@t<$x0(z2Q&3BZ+$aV*3YnV)zNSH)^u6WPMnj!RjFXQECrf7y=&u!rn;6sa5qCX%`
z_oV0Dism9}5zJqbO+SDQ4`Js!XKWg)5y(xa16$8zu5|%b97@(8shb*sr1nUT!CASh
zazYvTZzz5vb}=eX#?{GZhMeC@e7Dkeh#dx^VP7R(z0`>V%1G+<xy>r_pSNaoC=k7A
zT<#7g#I;dEnK5soD>Y3#32ZXWb@9i7aA3*pUoPAML5*HWr4>!2DB*e##q+9jw*Y-<
zYta1{OcS&ZWjbB}MWPY~pLvT?U7k$d%WNfH{c@PhG?;d0C+2=Q{$s({zXQr!S}8JS
zRe)x}lN20Qc-~ak&AH{)NW3NMa$I?Zb@0NyvqG|^@E$<vOo8n&K2cOl*-BY{Cul+9
z93$MIJW>i|2a<T;M$|kksWYt6kwOK?6v2V$(gVzpDhZ1~@P^mjr=`5nLsr=y2{};$
zG{mh=ZGXs-rghTU(My%uB*4=cmi~BuCe{jzGb%4ia33gk9T+Rx*WKFe-w~82-cme&
zi)Fe~AxoX$^6|4bGl;lNz4a~flnitgW)S7qlwDh=d)R?G=#jbAeL|st&R~BR$A;r-
zjvh>$kia4O`$Qf5)xrBm+ojEssP~ga@;x8H6Iyd6cnZn1opi^yzRyBuDVy3xx$$^J
z5xY~XQceGXsOv|}VP|dDZ)!i7EeFe|xE5;EsQdvT7B<$JZu4_{&Orl*qT9oEEAr8j
zQ?C9@WR<ve^z4-CtN+gBAk3jJ0qImdZ@JzGczKLrvR70(OL?{O!1vfC_V{9)v9eCg
zhsZ$u7p^Y#+~-9fB@ll3Uhn;Bo7khY6GXQBB|d>>P&hR_{Wz-lwyq%ZTk=BksG=><
z1a|uEB>wWkOb?mKcXAPl-1pl?&u*I{rWs-12&ctYJopMnxPOnHT?0jtGyn;187z!l
zu6rE{^LZ-lXUf4qPh`c^_X-4@6Y90x7WqLsQ->-|P0G)hT49`grO7V3*9<UKK`o#;
zFnR)YK|)%&25bj2o;NQEwbGR|hV-(;%_LmiL(Q1fD!354MHtwlOds&YTE{WlKZj4}
zIoO$Qes}~XX2Cw~Gd6wCzUx!M;@TFu0<{A_6e0ogy>nBM6$^jEH_v0OIOL$t$Xo_Y
z21KDW^3qENR7;^x^N#+(C{Gpu_-dbY8(JEws_zYyjK^NB!v>Jq3<8Dz`>>|x$j=Zr
zeE>Qn>x*xlE~mDST@JJ~WwCdGJJV<6*rLeEJc$>&ljfWzbqdFMhL0Mc{C+2a-Dw$)
zUdUdg2Kgg0x<$`D%{K$~^@M1zyKno2>_I?Ex)r4P-tdskw^$4y-4&R*VLopfta}Td
zQl+}Ggd=<O5tguAAA@2}*K5$c3G~KfU+KJVrAzif!NrTME_bGvfX<G+L(e9b3~O_p
zT#<aLE_a@rxxf7PAfbR>(s(l3&&x}!Qq(>Pu?28!lrVo0R?j@A&OFO^v*6wx0Iy#C
zr|%wV`Sp8?hD;%Xxiy(REk69B9Y#ZHsWev^6u#1(e@_+lEAJTZhJ7&J56woBp6>aC
zE)HAaYV+O4IQj7XJOJqdV7!m)=eh@sJP+ahTL*yFez4_sk6h-U91lAZ4jZDkGeLjN
z?oZRbH`J9j?t2pMUSyd1!KcTLlNL0b&Z?P&=oLa(ZDGEnMtYI>w~HGp9fIX5!Weoe
zh#W9mA-nEyUwB0<E#*yv-=%SJcgRACrWP=2>#n*Nr%mkuFJcy+I!Xc9USj+zNxX)^
z5EME)VV7TV!|;9&Z1ocqJ$`aHv}Xn#n`<H8`GnhKoD0cyT6Sqo1x19KR^n2rynP5B
zT%98kr9LgXvAK&hst}W)t<geXbvC(b1n6zQ`#z{z<=0I|US!jPUO4^8FW5{qy{zUF
z8g^E-;ChSp10Xrj?BJ&;7DR6R=AcMzT7pF%)TcA|Gp@vo5+=MgFobttAXQSJ=3*cT
z|5NZu+<O2#Bndkg2N4^H>uUcWQSTj3_5Z(*a~vEa&atx6acr_x2pxNG8Ih403PmC6
znAv+PGDEh=4p|L}6dDpL6=f%h-~H+J{(OJ8TmN*sy<VN?dfdl#zpm@fGh-9zHTd{W
z-)E<eh5f_%nEWrQ<JT63Q@u-LL`BU?by7CcB&G?ZQ}LufIzK)8L>eahH^TcKoNMv6
zIko|?h+m7ct&ii`_ym2dBQ$?3hwEn|KH8?me7(lXf0GKg_`a=4{n$2#g`T>t28p&C
zq#Lskho_}&JNX}skcl?@w2uGs;?A=l+H-C;j|8s9PG-WQtj7U*m%^v-LWlmPfk7>a
zOit{4D%~Gt{N^1Z>2jEb@?-vtyI1&5hO?@u=9hMJ!U9*Ac2%Vi<L^E>yK(w%%vjIt
z8jL+cVZg<osnJDdt0CJoqN6Q?p8YArQ3JjeK(j)oNOKO3Zmw?@FA#`>zkutE;VfRu
zOC2ULo+xOJX-J@)RPe&B^^m;<Tv2Xw{IeQf31bI)E<c%C=4laDU_Xnw5I`wwt!LD2
zgFbA1ILz3qlTveZea*4U?|IhS(1FMSZJ}vB)l61%$K1H{eZ}c&_)v6XfD{O()_^~)
zdlTP8JzO*5e8eTbXa8D}%R9NALypmcU(N&5f<^Sp2d=}+y_|dh)Nec>9+o4dA6-T%
zT{ByT3A~U+L9qKoJozDvNj=*@gurLwzrZ7xXXI+mi+YZiK3Vznm|5<P6~tEPUWh-P
zMhldu$<4QZLOIG%Q>oJ-`H!uW)PpQGgMSO4nJNE{U{%p4z|&76dqI|C=4+Auw@;t8
ze0O56%A!(D)WdV~Z2;?931$0TT=1Vz&$Kj$22E|Jnn-uLx2ZW#PU#kXOC8*m63%{c
zv_aFtU?`7xLojHY2WNbKDv4NWR_1HSTtpmT!z92$3l1O5(r;^~AXDjQ*nY1I|CHVk
z4DP?2Xvj0pKQXcs<tc*hiy-4D=}cKMKa~7*T#g6jD}XiY_uK^lmgShWlzmfA;p9)T
z74s!O%HPZHe#2<cu{Nr@3RuI?HQa>#*-8>k{miHvs$}kDAtAes7>5cVu$8mT82=-L
zSSkd)3j35a`5ll&<~ZoBjqPAqW0Oz0W-wKw`TnWPh>4bHMZ)P_m9eDBz?RZaZ8gek
zw(zlYocS(jT(A_bw7_#|s-7HmBtg~1999GKje+ql+2khZuyJ*Ujb9tsOsCk3TLFXQ
zST%gIb$3e*zk&hg{;fmiv#&G4V{v!Uy`Uo+F;jr$cv~-Hxv2lzE_a73KeB(5)E8tO
zv;7x(wfy5=gHH`uJGm7u2imwj5d{4N={EUZUS0oq^xLx<cn(YB#=vk-QJF71%di$c
zJm4<)sm)30?Ow0{ZTr+6YD<N!t=zjO1+T%HXYWu12t|)1xX5wUX!3V|lIXF_1TT2>
z{C6>4tQ#7?cqa3W)JL*KqeSr>bPzV3pC_41fFFC|H)~q#Rz@D0l3lA9CU2=|Fx4;-
zy%PWBjrkRRr|>Yiq~08D^LUeWHdU|0E6Ez-%t?275EqgRlIF*|qKk%IGlBDoK!|P=
z;d*dUs#(a0y6H}P4CiC-f37k9SBj5k{lcbH4qr2OE9v3qMbJ}3`v4D1q?7fsIj+a^
zx9j|evRWoSBAZmwr!)b5vrkDD|6mcsuaA%R!7VZRrbc;j|I2T-NIg_^nd&`oDHrom
zdVB5s6K@E`y_GTF>S*+B&hn&B;mo@24hzPU(U4$oaAsnQ&Ko{4J3otkfxEGeg@HGz
z0!%x7J}4Y@t8I1h;cq^CFz<(i*bXx;^dHy+V7U1!p>pEqv&n}K9z}kHiL$QH9XTLz
z2_Lz2h|3Z<dRq#6W~$?~<VSh&?xbYFXT=poEo`d4P&LV#rMolqhRiRgx6hqod}{27
zGgZ2JQ7`7Z#k30Ji^~~0CLd#sRA|#pLKm!zS9B^YvKpreQZS8&iK^p|QP^(SlBhmT
z?5ImIcyM^4|H4VI6I5I3!Y8<Z%21p5f;4OBR}mDCP$Fn=QTmuUdqz4V&4=`(aa|#7
zz4ReNJ3k|X6z0&hK%L4GD2^?Hf{c28d{+6i{2i)P)~mDStBtqQMncgAA#1t<of^Hh
zAFNpLJ2`@=rPc~Fi_XARL}Y<*M`GcPUqCJPK2twpt06Abp>|qHzS)#eB(kG*XF~fc
z7L*<?z5}M;s?EyuUg+lc79=)?A}r%PHxqi>=of{jPgm#G=f4}zS6rn;fh-DN!tr9l
ztF9|6e6h1(Tg{v18z)MJ)ivsDM_qN!qJbi&exC5**c~4K8Yog)@bF;8rGl%{O>E<p
zjx#d}>R?RP1l?cU2H;W8T*Pkx7CHXSc9zLexKjRi-c#c@4xbidA`{c&`KFqNPYY^$
z{K91i`b1%v@nbOxC>D7!&^8RKrtGNo&|!u-W}|HCt*C6`mj=wcXu=-h>c#4H?EJbr
z3iD3-AQByLOrN=yrtkOzV=`9-EgG2N+e5!vWU~CB<K!EsAKCX^@wh2)J;iUPnIZoq
z!lRFSx7nwmLj}ZwzGxCq$@2Za{h&fY*%&dW_%F4eHCrWG;y3Z4_f))n&}Y`{q1V1Y
z@x<1T@czK|{nqaU?gwBzY4VNnC;r!}DZ^0}#Jlx646uaIyYW3>4FfIusr{#tPY&uh
z6;0sX6dK*`-;zLeYDqYBbEm8B5aPrnLg_#J`UH(w32hqkWVCVepalZZQdd$S2g4&>
z5l?qmVrenv`+IH7F3?uyhmwD%_As5nY}18-bzn6d{=GCC__*OVP$akw@*IuIS#k6V
zDHGHOpfLOvFHqaMA?^R@7OYmQwte^$x~hPi9*K)slm?xeLn3R=BV=E%Ml2J<q0&pC
zaYV?PMhc%<3Tb~OO~EYqnmaym&KwN}(8*wjp7QmF`5%FngTgcg1N_F}3z=rQ8a)!@
z1b?iRouH7+>xL_zpCTql5XJAk4Xa(}Mp?(h25JFoSs)>EkJ1Q*MUuYa@aD$2?1I9w
z=a}Ix)t^^kkBkveKdHf79DZmtft&LTk{|Kt^SpR3(_<F?MwrqtQsSFpxzd+qxf;YL
z9=RGH!y1rXz+mqy|FWTd<0uO0sP-dYJm{Z#dC$18;DU%;_bu%Icr?M#C_C{Z1|A@O
z+=P4ui^HR(lo@qJOyQHmqKdRZ)&BA$3e{tY7Jk<dwC)z?`R?s~#dQu8Y?Ve9u>e=*
zn{1mvE&RtcxVtO4w6X*09#9OJ1P}bR{2E1uNpaYRj%PLIm}NCPt`)f5+W;ciLka`d
zIh%-?2KGQM6-k=JrZf+G$r6Hg6MF#A6VA;zb?ws4`@G*?spFw)1_c)67|Y-fi=Gd0
z+Yn0M^O(*5p$!96W3t-jl@*~wXbJXvSGgBn_x<b1|Ge6QrgdXnAHLfTh=^bd`Wc8|
zIq*-ywb8K%&x)PluReHRb{Ou%5&>m3SYLK>55s=%IjO<D#R8WHw|IheG@JiP+wSWC
z>S-_&0PMH#e7OQu&oCyLx}Bn{QV7^IA6m9vq&fSw@Fv{O3381{+$Lns%MC)eebqe1
zQsxc{nvd|dQh${jpb&8_1v1C|2=E~c?QlN7DEtKGOtpKnd-D@g;1OamO_hs}_W%6?
zRJI9<N!aWO+*t>jFda0|^y-8P0>d|Uos&#YsPjuHvZ3Ig-=-k1`LLHGq$`4sBt62)
z!~7FGZ8clt9`!RcJutZ-<ptIR?Q~()7?Va%z`jcv?p%F)0nAwyVH-qK%NPeQmN5*R
z(qpFxxp8x!O4bi9IhhYGc}PfS4k6rHp~2sZ)-b}MEc?A0z8`}q26^A?`Mh@GLku`!
zC~x$n3dg5~BvRlWZo$o6>e_sh!pFDiOsGeYq^|*H|9!Y2iKN|%?gAAWNx0f;IAQC0
zN5FDB#T<kM`0}GI@}5ErF}p5uktPh1CQ>XxT**31eSZv~Vx6&vpyo55P4MV|5PWrQ
z0TK5xW@U{J-ava!R5;zd2O7bH;~|p>M?SZJJQ4X7c*~b~rQBa|%fL3H%XB`4iA~pk
zUoxu|PSl<jEQ~FPZp|^VsyQ1K74rzjXjo$n*HOiQ?~%EsB8TMFgD(J8ORcJH;^0}M
zImum%DT9N{ZX=%#5u<@Ofr$can9;obvIx7)Mcx{yoyd6x?kM^Y7uX|6Sd@tak~@Ci
zO4549eI!q_Y#oT8v}reT<2D}0Xw1S*nZ%Cj5dDuK*NLG)#UVZd`i5}A;qUU_p|Cnk
ztA?i~AFqQY@j6aE6+r-HF>hX+J$;^X5l~g6K>LPY0i+m_%>6I#ARh&ee1Pqb-y8m>
z!??xqvS*M)7c`&nr~h}^{c5u?*a(3bAo2?_%uEYqb2P`XHCU7FhW>;CGvRmU3&2yt
zfO-H6uB<!`o@pd;3Qnl7x2#0y5!G3+IUG6;`;Qg0gRSnt5AsJKhZf|1`OjHBxwOzT
zVD&;L+35?%*0#mF_x;h?w<}P`oJr<U?vo!HpO!(vB$xF5PGQ7nu?ds^-dcDvytUp!
z=L?G{RH;<OP9~ka7lt>5oqQ$^V2m3;y+%MfKmAZHH}vdf!QbDZkdCbT3gG3b`2^YU
zL@i)_Ac4j4hw!o)k*vKIDiH)qgo@JZWA{JlA#QW1>{)ik1P(rNUaSSW-<D+}H`5;e
z9!MDm!zAVwlLsEAbr=>61-#P{-~=v?l^~nX0aaZI=GG7|OFx7GPF+@h<ik051gCW@
z2Y}{oHc|tE6&q>ApYj*C8jG1`WHr<u!7f@!fLK6h1WTN5K%51mh6m8cebUs(J5Xgu
z61mYVwk*`_cMiU-sYtRg`p}(?hDzB0=ER;)k8i;Yd<;51X6p@v2Mwu4hE>zKbaej~
zS$Q!GBM>#mj)STUYtd2HU!d<9+Bdh@rwj8y-vBl$q?v~oUPYfOEF{A1p9F3@)Y~A2
zd}hINC`;wP-|sn%8U9eprHVZL*<&bn8DvDX#9p*?ByJ8)-)Ye3LANYzNrEsR%|s<~
z2aSTCp$-{^V`n_aStu6@?5a>gWxqdb7?c_}_ICehFd+jU`2GoL9?+4K%VxIIR!}h`
zXC;z`+$O$H?%nu73d3U1dw)OfgC`?J_O^}ton|nUjfq1QWo7tshXEdPN+sHv0x_S_
z7_F4cbS9&2a{lgP2a>-U+lK%yU50cVURnA83?Pt>(H(0dcy_!sZH<ib2`GyA>K6Ey
zNR_YCTEPHZ<QWr~f&C9P;z0p<!xMOUa7>T;&Gkd>9=2d%6u!eqg@Sq1Dzu@JZ3|4r
zEKj&-qs_i~twKty=MV#wwq?G*%ZU&KnUeV-&B{YiUUGyXJg!Wb;x-+7eGq0CkgPo3
zW2P3BoerlV4^1vrH3M6(S>}4Lr~6;iJ~H4n3Jq;RU(^B9f)^`_-un4Le|#Etr|NG_
zLkY6{Lq65S&`v93<V{QpuJRazUPBeI!hZngT*RS!*ERcG3&bwMx7h|5+;xa=VYl6Y
zrhzJos-}U>gP6|7X;Gc`Hk~IRWBqrCA(C`Pla$Eq^KmxuX$3gXc#t@S8FJa4lyE!!
z7QkU-Z{BbUukP>7QA&U$nTvEQ&)vT-lhlKZKB4V%wWFtjQjkQ(QUi=kL5_PMfN^?h
zrC6RoMmhg!!|>0V!MT{{b)9Rp0`Kl-wgF8C3Z3HsV8Z`@`(+NW$8Y=g(d05c_nP1P
zdU`)sF+kM59gt|FJh>-cKonzqI0VcsJ06S22X(-0BC&u>$i(oTF&NsY0{-~oaofmz
zZqAC|{9p}C?_dV`EL|vhY!X1EXSKCpk-xNFWm!X8gZ~QH`iS&G-j~3XOS9~^$~I$#
z3Oqh`cLnVMS+B~zOOg9FV2PhWJD#wP2%d@`xIVitos<55UI1(}sgSfRn;IBolE?~X
zzysO+|86~~1Dpu>mQT_@)f@81I(Fm&N5?<dcl|AFatlAF0qnFm;NpYV2Y5Nxr_LtJ
zxN~=i#5xQXlkGxz6a?l?4p&Bwv+Mc3)6g@@`wqUfqdfGs<1mvaJnDOQ4F;=gXTngF
zch~*^K$GZGbW>0(wu`Ou;%*|lbhykqjMq+q*M%{ZpA7ZJnp^Xg&N26;ORnqj)9@jA
zLWi6@2xUe#@=gE2S0fx^2F^Iw1FA+Q;TMnKHcl#@{|XclE0eIfRBaw7b2T$%W4}wG
zCqj^ZCg43xs&pYKiKgoHoBDk0{vdxE+#hD5bPTuvZ{*4VA|5C2oByciR%ovHh2j2}
zbks{r(hR|>On|?jl1WP%VTI>ASoml9C9lynKOcp9_PvqPKcF{1^8aS0<_3Vf+hLZl
zCF;H}Fh^pU=cFn@w*1O8y1l%snv>~7<;V#AXd>_vp<9?D3!ClF_vz-i#hu~q(!@k^
zbKCl~`<Rt_fYEPmsdtvUr*LRy363Q{8TT)Imn)lYwbd#b->E!frFf9-S8%8`bf90O
zPU+g@(I>mnlER@4xyk;D(o6FFSXeSOyhME+D>BC(ogk}~pRBSIlZ@z8qj~0)Zb;=d
zJo+>bg4r<fSc;ewpqFg`VT1GH_HXD!BI-#LpxIKwMY-I_>{HF#DVHji&}9y*H?*1B
zFkIiltbFd=3sOCGNoVeGMcdCjsnh>9m@@{oEGoSsJnudy^hEcLTtC+Il(Xbs+r&QX
zk3my^vXBjq*nc-AvYPsjH*6*nlk2I5ec_8RN!<w#okIA}$C&t)2-w_c-AQmRp9dxj
zu$4<!Q*Vk{R1jB^{zy+p$ZfjHw{(MG*{yND{S6N#V)<srHKt8*;mq$b7u|^soX(ZD
z4k#UeOjJg&mpZP0iJu>0xd84Z{x8~<LwEe`Q2n<M>fhj76|XnPy8Ad*=HeySct%i0
zY7hN*y-v|z(@49_`sv6<?BOup!Q`W(>EZV=OUG~Ne8L-snzN7H(SP~OZajHDV$V$7
z5r`aiF_jpmLm_tHCr!a&f9xvP!X0IBz_|T2=OzfHB9iGwKXJI<4qH!o_Yn3S4W1c%
z1m&Y!$z$b<Y|6cH8cR&6dU8Ft-T<#~cnc|TdmJ!`{YPtEgUDZj@4S2~bEDM7uwLB7
z#s82f`ZkeFjUanYx$5RzM#+ms=ksWtDR*1A46zc9Cmu^l7&?%-ZbLCf%JTr^GRsbu
zgsI}2SK1OhJJ+9N7YF2grx5*@Cwo5tEARLzCyJ{J!?<;4EruzAn~NR%030u0YVtlo
z0DSccQjb$94bCw4=^*n*6*b&nXv&;FN^_I3XM#S!vv%1|lSoB{S2rq_6J~aLn(*zT
zln^dnFY+5`S2v+$;rwXkoSl6M6EiZXI2=ck8EZU${u}+tIpFvvgf&5%^4<li=lmd!
zX&1N!X9w##du4C+FF}p3F-M@K=W+isT4>yAVo?+aqqNmC;MCDJpLqHPl{MB$ZdR$b
z#yMH}&5|LtS`RI%>X;*?z~tK=J34Y3mVs!mlE1Ea`vqib^GHWbK{lSapnjq$ji3Dv
z=6DQ-S?;BkE-#q~>+R%D{n>!|WX|TL40&Cl!GEOb;U%n)tn}=lHuvSgQvFQP0S>m@
z3T121q&~mDa{7wXg7*5<7)76`^H#EwH{SV1Svr!D!yNG`ubGKS3<YEl;=j>IH4icq
zDO4K6`VaLuxO~5qA%t%<h&!xAUd~pa3`J#TrIW+?3Dx-l#<zK(21SeNmGBv8PDuAM
z=%J7Jp%;>JDX@~-c<#?eF~Wmcx>>n@GX2OBanw^q^}$+uxQqQkgrCK4VRkmmMt+8E
zn2nIlh1JIXVVYZ8>}F>;e|CNO#bK2GumNw+v(wz+xtt$%T6Y(wJJ0&mWzk@Jgbb#!
z6?Z`Re_96TThH-^a6Y^>>#k1|^;NsbS6oNyy0*?<b=WS9Jx#b0=j+pC<HqRAHYer&
ztz=KvO<-*C;C>Bwv5v;Xa(uI^oojI!icciT+tMCbk*F9)Gk)VW#43g@kCS`FFojX~
zdFoglPQoQgRfcmZ_JPyi^Gwe9AJa}7Z(!dLWcN7ue)L7r(L9!Yw@3sQMGx>w|8)af
zH*JTmDPq&>^+6*w>xUlx`{mOQ9B=GZUhkCPa+}YqrU5P0eb<-i>zGs6dNuEy^szTT
zad~Jewgy`IP?jgI-nT)m;Th85@`Nek5Q%*Slpp32HL%fS(8~0bT+_)(21Kv@M*OBc
z=94rjFh9qCQfrcUP+$GRI|ieKZ7=9)e)fA4@?Nc6sK@S4i4u7v{u9mJb^+BAD4m7m
zA1`T8Um)C_DlR2g0>9Y0kNgH{GH!MTk+`taOBPedE|c1BZ5$1;Nntj%j5whs%0cZb
zZPt!D>s6JH^JF=GB19`Ey<X1r9z`uI)N<i;V8#ubif@`?tX0W;u6iDiu6%D$baL+0
zR0ecwn?Tu@i;i$gx-9E`QQ^tM{-+K5<?Rko+1YeYhcmmoYf3faP<XQWd|xc;8@@;V
z)v0d^Si%zEherbZFScOVIoua;Ig<2)0!x)Tx{sT+7&yr_pe2tQ3=d_=`~nLV&U+W@
z$UXLk$%o9gWT1>%7<m|_97jKg>#5qrLnfb=AAWS>(wz!yavQ;t!o8N5V4X~C#lM0?
zllx*&b%jKdzR{4pU+g~qX>7BiV@#?v%@87Gd)cce?9L@zM;Y)&_#xHey3VdFjtGsX
zT-`!*F7~{A{dCC2k4eT!f{T_$L3$cUish_Zr^p?3ik1J8F}-Y!Z7MI~=xR83SCpCy
z^LQ5>Taxq8;C4XDF;wMicEPrY{FKlwzGNk}5ywksPs~M%NGnF)^j2q2XchmHEvzt?
zDL;EWwp9OpX>q>p)g-ldJd)S|F}3^MEQ~y;>W8U3W3wWQl!=%hp#D2qHc}kv^f8Je
z+1Hz!y)|1}=+n6KQO#No4RXLC2+KiP4PP9^`}$(;_#sYB@hkg)GD}pQ6jE8EANbWr
zl~&{f;t_9Kop^~QE*jJ9)M1+EukS$%+QLUBBp2_l6y2P0i;g6Ocl{9R%lm4+?lfvN
zFrRXIW1yuhz4|<!S5ghwEppv+hh2HxFLJNbmzN=h+pg%HbcS#FNjiI7Ss0m_L$|X|
z-x1LV)n<|U)EANO-X7GCxzFE~JUU%6pYMA8Vu<U62+8`zcFZ*9S*73Yw~xLp=D&?;
z7d6Ms&V1bzF)y*@&gM#EM~_<{YcYP5>nT3;IA!R3XPw=t(Z?U`*w1mvrBJlhPpJE~
z1A<x|Pd^<WSoLR;EQ>{o-oPj#p*yD{=7~Vqi@+R3<>F`mO7Kc1Zf&mF+P}Ab&%Rbx
zQDu@8L2oytUnQ<dM*B>aR61y3grw8lget{<`NKJ9J0|s4nz4TG>$B`+&3yTn;KRT*
zrf$+}045B>^npJ-VtwAN&|*i!Zsp6#8LX84USll}qRq(f$`mn=R-#$w+m)JlDR;Wu
zHzluIgnnU7V9{0L+nhM+Q2u_aD#E&MGQVN5n|ri|+qdqpnB{Xl{pKeuDBq(dsnT%g
z94Eb8$>Wbl+t}M=aD-<s0Q8-r`9}T(RoTo@R&;{1-uTJ`lKsT?a8S#K!G=9nR4lG%
zxv({?n-wL9Dh*h-cCm6V@LMNsru6%Y<~F7ZBr#obJ*houUCB#qyy@a9*Tg&{r<W~p
z@kXgpS<uoV`KPaMCl%uA6_IfTB^t_QmzNk5QrhV>dTR?qFVm+eWLRmfe%{{bY(o&w
zY2~pJ`5GIeqKd~aokLz1sZ-DCib!dyB|91Qy}p$g>yyl_aNhofDvNV8t%*iB^DpaM
zkJXg1Al?)Vf(t%gO0Uo?f27~mNEUAXmhnC|>pJ_%Aa;$a^VNNW9P}m8;-?$K{_6~{
z@)GsDK|Wq2Me3TWRT1A(KV$CwIq<lMRor#bG1)~o7tzL;f;>%Lytye%8W2GKDDLiE
z^U|rQ^TNN##O##fGKU9SxVZ|*z(1;sqkwE(fYO-EOtX1xXaXBEsMu`ppNYk=qO8kp
zjym02tA69=8n!Pcz58I3^Knkx-S>oN^nA*{^s!7qmWKPZT4xV!aJcAhTTzOoF%LZT
zYRSy<<U=<SNZP9ao(_5&eu_yqR#EVx5PIPaEbGOWMH6RW{qHeeO%uU%_7`=mREEhz
z>55^|J1hqcau}9CI^+Se8?Yyg9lsNTdPR6__0rs=H`b#BY4NMdRH&#BXv<BsEap09
zn@}IKJH+BiPeE`h1N!&f6NTgzqIB|p&Z%pO98rq$(Dw!u?<f3@RmfC3-5HSXQsC#r
zPx^)JF*<Z~1v&<x@>5dRzglyC82Y2;;qY-N@pw2|S@KAnZJ>>@(a%;CkKKW9;9%N2
zD}&nJTFIErk=~EG{16Zl3yzXmvr?&6-<avJYbb37yRwBSQ+6SER}r@mgec$q!h$@&
z=Q^nYAvqS*_|0s&2RX;1MKpECe*w3~eS$suLd?lxKnn7XM{_W%<pPen`kZtlJ_4w^
zeoP&M3SV!hS823MP@xQZ6F6kU-Noky@~_iBU55|SghJ<`VM5zvWhBPnj^bu5tcw6d
zqq1jFjsH)nZaRA{AU*7{D<$oPYgt(OPy}>EK(!tt^j!{_8nx)Cm>tupXrL9)_t`#w
z)IY9vGBv>Bt%)YJ`53a_hJflGt-k0l=!o%1qBv1xHspuK<lT+=_29xQ?xP~F$rCsq
z+dNCpKOX&u)`u*MPyQ4AJ{gVIf(<*x){n{L?>SlpN7~pu;?5!K&|BY|XX;E25Xr9a
zn3IXqtuxJg-7J`M?SRE}*1RHy=fJ=9rYLWvY&03j_fB2=@)6JujZ{8)A?DjdfY+$p
z{`cZgQNe%62moN@<OMY=7jjy7i*MB*$`m#wlwTCS--G=lZr7T^g8r`L_RrEC727km
z%08@oUmSH*(Ld(8T<DWOPNg4X-NT-WT2F0^#F50^9UJkwVGk&6gZ1CXzA}#{AFe)z
z6`^?)56Lw((c3qQJwWW->><CiA9ap&`M!rH7}Nu*b3Z7Z#xug&nE(O%^n!O^TFU9k
z`;?5=vwrh=zv;`#=C*bAz2T`=@E8m0Kj$kySXR#U<ap1k&gFn0PfX;+O2*D^;cc+4
zm=_v`@6Gd}2C0sAxr%32VYAf}<ahH!>^`qvY!zG{+;uxLPFQ(%j^&c->%(2c#&TKy
z4R+dw?JBfk-q;nD3#KP%HA0TmHIjLuFL-=Or=sOwm!I=IQXWsHY73NJgNNA$4vdV1
zvJoHHiUe*!*I?^(m^7;^=Oi{UIe`B6m;PAh*m!O!N78b6G2l&AFk0tDSyUZNqaumc
zICh20{3o6UzBMmQHN4U31BR9^Fb6A!31R|PCz6;_tr@#9aqR`)OIUsBdHE)wU@sUy
z4`UEE<r%AUloSM!L~iIPUcU{ONBQ49RrdWnAa1kt1oEt@Kq3Eu-Tt4`u>}X2_TYqN
z<46-eP#LXKg#udt<XHUgniq6oE;uN>ZTcrvCNS6@-3Fe;>*|T&84qd5!&|_kfPNs^
z@L${HA%81u7cfxsgU|5ye7yWje2C|I_4UVIrFZdUO~`s<jP8%=rmxaNr5(Is77~I|
z3rwCrxS2Wb{1<n~eF!~TVg-)<f3qbWRAwGRhrXnV|1gr<HC}O&7m8TSv36MM$V&XQ
zdSNX!P{mNI#G0q~IiL-tF809T3PrbSY_gzIO2i$yZ@(4$rcbGS;`5PLzqFPAgQTH3
z;~)<ailj^AQEmrsgHktj!4W?t4#yLvst3<?DxTz7JRN^8m<|aTuw0kJSD>sQ9s?Wh
zX3O(0h3ox!@XF|l-<1C?#0gKbW2w>q`q*ZXHfcq%L#p*hZ3oOF+z<H7|NB9847XC?
zOW|`qd&p)l^Msr=myC*gZW8-dX7|)*@4x$hPqh=X3pfvqp8!TN2;@!qa+NQIPn)5I
z5?Sm6DEQLPcfEVF+Slr@1&E{P@GC%CEASuJm3{=W5W!=L|BEa3L`N`&0!J-~@~7B}
z_jkXZo{FG<hc)s@syzRNl2;%C3yoy8rj0Z*$KfJx^nv-nm8T;2`uEPdFB4}0!28hX
zst;qIbtv~w-@LJRd=F(1uP*Zz^kWN#USj~nWUAyR9Qhr!v!aY@H8ZzY`eafFcrolQ
z6gyMJf4%FM52eV#I0ADNHNkf;jUr05{?H$>+e#r<1QPe7;K%pF)sIFJQRsR$(@hao
ziqK3IREG1=B@`>ef(cBrZu&jKj#CDS9~F72;xe4Facku)BfwH1oe)QVmm&|+Yk~WF
zOt}s;yg)rsnt5$B5JX@Q{m4Xhw0&iF>m$(6j%iHAV#b8fKCpqXUcvg8s`&qft=4F=
zF?CRNMJ7moi3vQl9seiOh?ZOu8DC|)Yc~ebg~3p2$QDIu%K4m`WxU*A3P^68l@0(R
zhv!{T3-SpugfqPa4A|1A9H;ldUZ9g>HN($^AJi(2I~U$QsLZ^N47uP8!=MyA;vFAd
zdQ&|Om?a|G6`QO|{BNd$A=kzst3hF~cAmi$T<!q*Qvd^K5g2>JH~8n;PyRM6Mi2$2
z3&<iyX-3T|Z9z_Q`Be9C|D}<q9vXUx$%ty7O}h(i7(_&YFYRxEAr2ICx^JhheSnZ6
z33PbSi26XZ0g^L(Cgx4$e*eR@naI!A9tBC#ux#p(w;jwku|ZlxYIrha7ERL4fEkhp
z1}x^g0XHb`jp@@xaQr1w{}Hs|>zf-z-Rw9m{t`;FQz8OrCOi@U2?#<aKxYMR%RQhE
zL#Saj)dWq5$k_)T;rBN(|L=~uaNEGf%9c2%czHpmK%?JsXeoDj5vtk)7yw^^5oQ*z
zkQM8Kxg?HVIarP=1C;S8X;hOk^n&1JfR`<UlDbs;^GN>W$M64agL6uq26lcc{8kBt
z2?oK=pYc-Sh7a0N0Gw+rKr@GW?y~QbxT*W!v`8j_jJ(EeGx&Q(7Z_KxUHN+Ad&Zec
zvw;muH*38^l4vDl?9Y>fj)jkxcnn<tj8<kNdGZL{1C>$@Kzo*fe-{i&i$nQi(8}Ea
zZT?`p(bqKfyRb+*qgekZXEPhB0SoiVFwGl;_3yxB?p1md_J6nj4!8X!-$&=-iK<Di
z7T()Q(HAR#UJ8a_&M<Dt)Nnfmdl%S|r=Jj%!KQ}K!6?D}CbDbe`|D>WjKx0yc;`HU
z`1)=D0}MO_V*i+g$Hv#3?SiRM%htUYeepd|Y&4#Y2W?mnWA3AsIl{>QUiB4>iDslm
z6MPsQz*7Q1v*$zxXkeblmZV0J3bETATFr$E%HM_2D!_e$iP(7dd*QKV$<0@(C%kvc
zBQmvX553C}D3d%<f^5QquoFF60UtL{u1qYBAKl8U`)b}#-K_LfaPRZY|D;dxq`Hkl
zzn{b03beW<rDwDXKCuU^{AkopR;)`d(!^JmsgMa%+B>i;*w~foq>=<;-(CLoC-lri
zO7vmHqD#t*Baq|UrLv-RNW|If4&e_GRg%F7o;{ex({(>SoOtx$&R`{hd<g($nTmS9
zKU{NwjuN8KlA(eC)Xg+t+1`{F_p%ENb!Awi7I;$fkjxgAfG3kYzV$SLzD6~QSdS#M
zsdcgCB~?=1Pt}@1S4Lv)k1{Bf-~pGB&ceA7sm)MO?NnKV;6{_a+Mba%`Vf;M(GHt_
zKE_G*{`Y`;USXc!R%im7wN(xN!A_`0SG!*aK}T<fIQ(Z6y^3;DCB?nSFfs{gcK@31
z8;x)apc;RYEm`Ou_57|QOoXdI5f8TinPr8&HkhfGCOFLt6L(c9&c_rReJ}aHPe=@d
zy`%S5#%^MjbF$Lh9kB5K{SdcWkAMW#ty1GA+wYr=CsK2=-A;G6avPp0lyA;e1Shbd
zS4PWc&wb`Z<*diZUcM;rK#xoafq3v0ki-yhBSM7cY0WPVo6no`776T(-rQRs=1az|
z@569c!?Vc2fuB<`DYmCKf*SGRgILR@!N_Zq`rvO|knEz%FmCq38o-r#noot@9)aO9
zoM<3kj!g6<(lB(a{%@GCjl(oI$DE;|z5tY>BT-X)cOKskhRhSF!5+&CM@|VXym1w(
ze+cF>_7q}Yrz2sDnIot+QV`w0vFm&_sr<w$b{p6`hsDoo$ng1v@ZZQ1J>v-(Y?bpz
zCpJb6(Rl*DaQ@2$;Jib322e~fFunAk|3W2@weZeG6E3Q=i1zNNr^-B?a)7y~a_|N|
z*}!y{yp@_I4oA{WFyjSY9h)GuMh$F5l}y1UKW~82lt<$(RHNTgTvFx#^iVK9hs|&h
ziQof4W=qro89Lnuc^7^gc=wM%5V-~fk3NC@33Fq_8mK$PK`od(4Df5*yQ$8gD|6hi
z6zdw;KjUBLNIO6HyVkb^1gcp%6uQ{#7Wy5QWg_}w3~>S|9qCp~%f`2<PV@Z}W8veg
zKc(nVQKOk5Tp;Vtu+W;Mwv_sUmu&FfSDX`nG}f0HX=2r@=Dx&yOUP_IldU`r7X!{<
zAD|JA<1qiNd_>y+qC5`o<p!Nybm}6o0AMA=H<(|ff|3*XD3<pv6RQW2PVx4z&W{PK
z$9SnOloKUPpPK*b9Wpp4fbNVX>rVxD@z`A3JiwsfE(;P3!Jev1q_ec~F!`6WFUYR*
zjs*G?xwZWSXXm<9lj|-aoVVGK!a~!0BOn;R36VC2jQ0dj##60Su+n%4gX<=%lJ6XC
zJYe(DYM^<dk1M(scwn6{tApn6(_YjGR6gzz=cRkV7+G2B*+ZCJ-(KB=$#ir+FXfRo
zr@3>Kzk@O=V-P-aggy}dkK#ODSFvVXl@=HUAQXgQE~0TXMtEb+7bId;)|twb$n%jv
zGQDxE?hN`!&$skV2F!Dg2OYEFbT2JH<C<$OlE#991*`g1U5_~Dn;#o(z?==-Yx3p+
zbN{1%LpR?Boh5NjFAKYiV~?Iqwg%o6RCkQY0Wc+o@3$`(jDjilObHq?V)wLuA44J1
z2|st;*n1zW%zP^wu*32wI3^=!X<uc~KKko76s2f85UECenF{@j2lLnGFEDrZL11VE
zCbNs|porg{HIx#Q|56VkKu*Jmd*^14Gg;jhH&5_-mUTmEie{!0@`KvZ{H1q!)B&kF
z9IB+@SN}ib<s{{71EvnWHKdy^_`1e+QmKxDH-(kFz#N9KdT-WaM?HRLvL6i<L%-K1
z;d^tYU3AvqW}bas+Dq8>(|E;Ef8mv2apu2@aMs*J6Z-SDk5@V{9nCRZ#SkWe!5P_w
zxlh}W0o;kfKo+022bU>aI}bfm<{@j;a&jfak}7~EB%>lyNrSI#o$#d_Ju^{pwmT#!
za3iz4<zP*M*5;E4ieiq6Y8nx<9U_ao)1FR-&GYYZIrF3P^9d{|4Xu9(Gl_&`#MlGI
z_|I=jUkaAWQthZ?xkPa;4bU=8b-!UnpDa<?HSzOLufQiYq*MnybTQOrDwl!&Np~0V
zo$EWTHJ{c>b%vOYmB{0TCJI*9d_twC<#+03h_YHg7vY=dpt^rgeZ3LHN(S>?!JC&W
z_PslR^4x-r>I|z}e#={UYI5RhtQOf@IMJ@I8VTB<m*TUCG>{Wpel5t4sn71w4}r<;
zyZd!l9sZLqE1@wAagB$ZBE^9Z@NF<d^s&v2kD<R2JIGauNkE-gP;7U|bvJXT2c51`
z4a45F>x}7ozi#fyT$l9pH|F()=s4A_V5FzYnxsoBaU7;wa8uUt)sHAv)`)wJreBo2
ztQ48gq9xiE(!g#_q38x!?9Mprwj^@qRh;bm=;#>_DHk$d4hAbH=R8Qy7wiG&8QOX|
zZ}lbb1}wZZBmJ(}Oa^bD*^<5E9kI9t_5`#+i9aaNfz@u7daZL$v&5f0%U6H5@!a#U
zLc3`T<L`h4-LFZYUEKg^>4fLR@H1~$gxw%e0R$@w23T$gOMxBVjH|s+EiIskAP(Hz
zJ683o*~=<k{Pi&ljEiSwq#7NaU`r9S60gZ_lb(X9LgLZW1y(?02+LqaJt5RP4>8=3
zlzPR{Qk0Wh`edte3(=@gD^gTnE6d{0z)u_>lZokwW<yzG9q*i(RuLl#yekK*fk-B`
zhpSMa2wt!8S_ZO{@ug~M?h|}q0k1xjOz{DT6~N_NFIh;Nro(u@Z*72-4RoXRa48dV
zV00$gwgU{Mp^iOD-W$h@46JlwoG)fn4%5jBaKMPa8Pj-qG}r7L7)t_$ryEw9`*I9h
zxbziMK(0?ygIr%aYv#;5djho-`}g-R#&yN9lb4d0p9X3eA!X*j#?a0HA|0iBmDt&G
z8BD_*EZ4P~0r;uXzsO51f|x7gin0sgmwUBXRScY@NoXX0flHU{s8SQ4X9>zsoLLDB
z%di(=JEzkqj1t}<gdb?|;aVilKEO2LE=uP?S#awF2jkxnLaPym{2rDh?h5&OHCSjZ
zsnGYJzlE+eMFsU#L;tAyCLVoTlW6DnV($J-6lotPHM3RoqrtHk<$1~)=3HT_WR35@
z7_ApAckQR$l7_}nlT(=*tdhd5H}`MQPNb!UyuENw^XV&aGY6lRPx9a1J@i=!IzkuL
z!$*97^yh;o=|V#j51_-59{o{@lrTLKu=S>qPIr?J&8-2y%MkvftQLfa*G1K@>k^)%
z1)+5we>xKcq@Yf<cWW)3;D8N?9IQP0QyE~23cQIK@737=#Y<{dc&bMai5zm>HPgub
zLzJwh(5^Zs*+PKM`HHxw`CkO^!uiJibhVy)7%Wkn*Uh@3A$`#H9aLJJqZSM1H3!t&
zJ$Pq?n<Lq!ha^#Radf-ye?wM*G!MeX8^z+HG2c=SUr*40WsF6Ehi=P2;anehan3D@
zmbK3;r8Tkp5EV*;$&`hL54WlTa7v#Jn(J-ly@EWEWhvIzwdbw_g!F_K;NQlW$}Vc$
z)PoD@rbW-Ue}jdBg$!>~Bh=zN8J2_t5(^P7<O9_K_F{}DCZ0%MhzLX_)zdbB7_Y=$
z-bO9^+*J;(VyGXc{{~COy98dUg@7+%*9uOuZz6>jz-NCxPMww7>fGQ!3Yl8Z1#}Eo
zAAkSHtebl=yyp>i#B_zzD0%W<V29nm5%r3>Xb);v9p<o-B4hm~x@<{p$0~<;&doau
z5)@^h{sSoSN6;0@yTCqLYgPCu@2)JiatLP4?|f-g1t@AHE2&MMq$p~Xt@tlYHXpD9
z{dazj(5@l-4@+HuHP?8fOg2n|RJMjR3?7B}MWSd<xEWvy?&L!!B6VZ<=b*Xzlx2Hq
zEMK2i=Q|UeAxJ-G{y~P5$?|u<*7Ng6IWCGMjxp&teDc`yJaB3->P#lPE>Am>_AMPp
zVEi@DM{`7tv@^skWW>J{aacSi5?M@2-VZ`^7OzzCa>8vfN?I0Qoxp$}IJ7XcRxROr
zNA+5(kEj}z{G`b1j5F+DdZ?;K;&xga*2_A^VJ;FLn^3s$KL>{%w!=-i01Uu&DJAtt
z^JE*((Lno9e|7oniH{r{XCvIu<W~bCQ;-Jj+t|%NpATcA@9=2;`%6ilhWNPe?oZC-
zQqP$D+uwf`j*UCUtIw-)RKF!o&jV+sXY=;3+Sy|l%x%&vS;W`wdsqpldNe;sEz>!Q
z)16KeH_O`ir`-MG;h*PUU%da({p-&4U&luO%!84r+m~a>SAM=%7!B&OcLx)sgg7Y4
zK`zk~@S^bS6+xrhF^n3Xli#bzaE`D4RKI33egX~j`>XFRqF={Zc>wQ-LB%>bdv!%v
znI9d~SpXi@fr+atMm%+xvUn0EhRB9l{HjEkaz<kXvx4nmkqF$hOnikWa)&{GQUgv9
ziAqf8;`E0hTiSpm0U_m#1tmFG-wFT;zG7M?mI?*X48mlTbs38u;9lR7ax~*&V!|_-
zd?_zScpK8k4J2WWjoq9(F6lRPi^o8I2M7q0r5?2=eg(}Fx?%cYEO4afLnE+&SCd4~
zp*4u;K2>6Y$EQW@=01HCM1{{?!b!aLZdIKj#&@A%1tbQbW#Jj(jAfSY-|f|To1+h*
zwSL~Q*4}Cii9F|Ck%+K-!^3}RcizXS%YoDT{n~cXQv=ibye$p?O12{ONyC=gLdD13
zi)etdo%iUb^FP58gu@eHaEY`(#z(<9xqQfG5yhH%9M#!54GwQW4V^zb2fP&A=-Bhk
z5P>-4;ca7d#AzK0t2Dnc^<3ktTmR!9Ke4%oe+haCo<k6-ofy>v8hwx3b3OP4$~dRk
z`z)M#t}nof#4ZaFezQnS3bYBVK%smjV8XPmCPI%<F$@;PFx?a_@x=L^fv%Blf%nU3
zphhFy!#AL2@oM?J{8qU~aKw1OYf036x(C>q?M;ql;Cb*#bchuLt-GyU%oKTq#p*;)
zQ~#Z_;)YqIiDDPssGmv+*Fg7aIQt24PWvdo%h6ylL~sOVA0l{E<hZ&t<+rgoW)?s0
zlylz;91CvWFFte010$8c$^ShFg>QYT$D`MyeG8>ccF_a8W3OJss4Fy)ly5=+?)?Z%
zD0J<)s(MTe3FDGZnqzZKZ+S{fMNTPBt99D7JPW`XJ#5$UCo-lnP`fZS0^Tl8S5hj%
zYU%+7LXWoR=tA<+s!2xAWWt~ny$Ow$jl_iLPhjpgZ8bC7Aza`Xt03c2g5W5`<rph8
z-6^cNZI;IBwy?4=8t&*S7#=YhknK0Z%K}P#c8~8dr(OonF@WKK8mXYNW~5_gYO-Uo
z$wH1B<|VJxREVrA4%H*JXsRsdR9n!Uix!p@IUMCeBozS}Xm7o6zrQJhs8dkwIAVue
zjuogitId&kE*u8^0IQ_EQi*1(TQ@8Y>|{YUY)qEu(WA-xUAepKRr!=mw8WlwVcg#(
zuoXSm#GgVl9b@Yq6(Hc~42usT`1*SujxE>`Xw*5YF<n`&ha<;f!$?diXIWk*nK#R7
zIs%%m(e_2$tDM%Pdr!~==mYS@%TleKHPm8wM#ak{Eno{(F%+emQ%jw0X|o}>$AB4^
zJ&Qi_ls61oycuR2sXu~P#ecwWBs$<3et`Nc{gH|LP9&L%3{-0bEgnUjzBB{76xp)m
zzjkSXnT~69UMd<yXHD5jbv>3aGTdx_U#O&gR0Dh8GlNf4^Q)%q$ctTwg%|soCfkVN
zG`V+<^uXwQ&;CzB?7M6rEp9NFfdVG`9<tFWYTZV!0`yWr?NU$DgN_G(&a|p^cw~Qn
zQ$Y9m{a4mWafc^hAC{A&|Ihc+_q|^|p;k!4TVk}3Ckc6V{E6PATRPdfXnhgfkLj2G
z60isetYEkSN{~;lrW%V{40Y=RJSw0IZTuaAMjMY=F$Nve^UC=22PirxX=V4IfKV>g
zz+vIZlKar;h_7YcyG*A<K0)AMqwGHX%8Trk&lhJ%Ji!HxZ2*|cmNMshE{qly3hMy$
zy930L+xOzWElgmR-nXUGUEf;z18l;(FPksCqLLh#Xk1&Vf}-J5bAMxKPB9!A_4bxq
z-m=-OT_>(@d9xIQ1eN)orUs#V7;?3;W2|td@MQJJNM5F!)7Iyc@y8OL1bU{j&7Q)a
zxr4=^;@QOGtVY!t6!>R&4?;2t6R1!0Ons=7X*)}U{gN)eGGan8j(40W?sQ8rtsJ(9
z#x%Fq>WDw}nq2s{3%YGbt^UFBK)V?CIZ#4`GVNSjof`m@EFhLi!1@dK6TI>A_!9+Q
z=;_MAH75W{cIW}L_;-L`$lRIOv$hCDC6P(yE%qjo?SaVMc$s+~+m&{l4Gd~b?-Ijq
zKhTVdBkrI9<963-Z8n+)1AC^?gx`pd$RqjBcoYHtqzJ;1x`3Ma41aOr>)A+%dLA+*
z`X;CiR+`+{UDd8xm<#%K_mNbYHa|*|d;sTa?LohO$ZpicT}e`+j^pQa6ZP14D5in|
zo2sz2mz@05Mj-8|yATL@0!j6M0YqkwiW8A*7*9N*Va?|6G<P?MYwws9w*|O@);}*i
z$94Yb6&)>kiu^~_Y@$QoJs9Yl&!RQ(M06<|Wkr+488~9zSp0Wt-uvhD-@>G%Q4Nlj
zU9y{Es{B&`64eJBd$SF)SyOjc8RtHqP{j4U4ElQKH#kS)GzNVxF&x`XP>VEt3Tv;<
z0R=wB-%vhp+HqgYww~)P3{R9kH$XF%p!{WYb@y4Hs(KAZnf#}ztlRT?7Q)XmEiN#x
zVryww$_Dq%4eSIF`7SZc@VuP98#ZgAHF_ve^1=1T^1QSVI)?o`N86Y5%LK;v@+~^z
zdK`|=qd+*>JLLYR+6L<GM1_T0$&qs99b-*eN!(Aa2cEXj4e(`RKYb$v7_v-ty&8;{
ztzv(VD1Xi{zdbFgiAHIYhV8cVUPJu|NWF(T^n&zu%oa}nosLs*8CN;q9!6r&O>R(y
ztrqxo1pLUd9ZCW1_#R-e02FQL^5413czA+v`Xu>e6QF1V^1^@G3UGB?Su9|w)&I>K
zCUG!f@h|g(KE|1AF<2*-fW@O0Ri|bMRBdnmKm_(Pv?@koNDPrk!{6cIMa;kFc~F7a
z)7%|S;8A|NS@m8|(;_yrCALomWW<x-MaFa;0k&cmH1I8_CPZQMfJ;qi@J5-w8@5c_
z{Kjy4cLlt}*g){XgVxnt9Txs@3*B<*181Ejx=dyH!K=9=XAYuP3~?B<XlOO8XfXnz
z7NZb$3ItF9?b8PFb$QJoC#4^!9hmK4c_od~#?tQ;EwXkbTikZA&n1EIX814Ev%pV2
z0Sd37*!_H87z4B;IW}CAl=CCu4`i#4&2Krs)>>x>BPYR00N*TJt$hD#ED2Mp;|M(g
z;6BSBwuqZ_-<HYpc6NP0f0(IY$qz?s>^zZXz|=GJ%0vB)!+SCsk26rb0PrlU`{lZ}
zaF&6kYEV3JkOm~XoL?<7Ox_syiqlKVQg!~|Jt#j&7>z_o?G>A?Z^cjqUosTD7y)T7
zM&C4szp+GcKi>KMtBxb7aEG5gU>cSxLS>&MK*@E&<CxN|6a61y|J?dR$Y$F&s7eH?
z$}sYndZxm$=xWWUFJsxi9&<cuEJ|SfpmPMDd2qtVko#&6t;-t1&X!3tnjuC(5COQa
z=p5zKB%K+R)1<ddJ_>aV*rw50-dS!&1{F54&kBk|8EG8l{J}@9LyO_OV&bL^v~4ki
zG7wf_j|+BT=5cXmCdMU^EF#zBhti<++n3O@aEyAWE6vpZoB{RokjhMk&HXf5CM25U
zFpBWSs{T@a{gDd=D$B2JwD%~Kw#wLo9?R9Y>4?(WO-jFGw7av+I+!r9XZm-B$>UY|
z2&}48!+W~7KyXKBYt#(fF%}arJ^_cqk3mUh8a?}HQ1RP<F92+NWJi$L>e(kZ0JZ+-
z5cJvvDW^1g9zrPj)8b{(P|ZZ-_k7>Fs(JVv42kc8aOwd-5jNuXQJ=%fjD&A-U)GV(
zv&;>`uK-vo^SV&RRS!Ty*d%xRmRQ7<@8Hdjz5R5jp?sUy9*_kH##d{&rK_CBA?-g?
zn+4+VAWY5cg$VdDP%EB=c}dQz4O4W8<}7GH9&fzf{IgL1m;rHn8LY3{(a$_6AF$XI
z^&1#PQ$w9>vR_%bqW()$t1zt=>{|sB+gBr|lYa}4aS;>x(n{#&V+Yml*bp(0v7~|i
z9Qb61f5n0fMpAeL0f$vv5~q*c&j;<0obcvvTjl+oACh+7ARc2fZ-DrNb4@5R14}b_
zNkiNLH+BKE<Ox~n1$kg4sNAN~nV#&yZ6nT$ZGXJ-mix<CV6q1~fuC-*&2uBS2J6K*
z{`d8kszdSh?d7XxUS27vZ=v-kuu6aId3f{`-<JCLiQ|t|9RX*Cb=S|}5<!jgS$>;w
z96%n)QM^#@QmIs-cI`kKrX(3KoKNcD|56XbGM9V)jQYLQk$L&_;M6UBV>eLiNG+k?
zVyiyG#BD&&<=JnaYD2fHmbG@V9x;8koq>LNoxloE3Ek-gW%Hs&h(@*-99d{k=BAM?
zLl{{?3(O)Mz#<DAAbW|Ak0+Q42UVGp6Qp2+zyp1?nvORd<R6Zoi$#(kk_UNFX7Ms;
z%W0}uvNdcA*_YS;071b|Bx+v63RS3v7sR*MS~tuko%)GS`2z{Nu62%fa$5s*EDT+e
z&BFQrBJ}UB9#`jAv+j+?I2X&lkf-<<=WY4;T_Z*!K0ZBDDB2sr-VSA)b&+<u<tVlo
zktOQ#fVVs6^_EY_A!!Nc?3l}d(aIg(1ENd6%O_FqW*~+e&4F|%ao{EBsE&iqaqiB$
zb@!w9CHCxLq6|cOLBS*AX3uGMgi+J{J<CbbOPreVp0~k`9QL3NyM<b09Yk&th93_>
zfSmhC96Ai<qfWLP&@Sb?VQR<xM$epG|Eb8doUFz7&6Qx*1)PnkYn{NDpmD*>{LrF^
zl{_08ECX?!MX`J%2e<mA6PQ-r6}gj{JvSpqDqRE58KtbLKNxs^4>D<mnvNO>7QTS9
z+$ck8Ftc35_2Ygq;I((_cdo)RurkhT%3k^a$?`eOWqGcM3u@*Z8Q}k3f^oS;_7>|S
z0x`_Go?bUY`iWjhEh|h0GQHZFG~p1#AG{Ep<Im_0+D#m}U`4<9`K)p9yEYd;@8Woo
zzj;{uPKMT*!1z-k+rH1S7?*e&^|@FN&d-bE7MRY>H@4eo<+WkHW4;AXH_6-Fpi6<^
zKG6I0^jZEl8BkC(ZhsyvM2ykiQ+0Ly5{La+05`4*_VjQ4gb%kMA=KnOPw*lyee<!(
zVYx6CE!>Q!WqkD!T}`&Qm|xh>w3C|eS(D}GV(5$T{Ey7J=JE$nTdn0tQoFBPylESK
znWHMYk@X-BV#({E`&^4FU84~tk1PT{Ktn;&Yf=>5VHKF1qKy-CQ>=wNgs|v@@7EC5
zQy6Yr6A(|*u!^gtqOsCC*6;p8JI+o+A4bs}^9WjWZm$P_?{(R0TB@Zap8P8i?XQ%~
zC0{CaL8zR)IRF0cLd}cMg_q0>$N6VD!N!Vuwb-uAgj*t|Gri{$hWd$flER>eI=lW1
zU=JSZGzK>2!ur`rzDH9238=xJ3SP|od<r#ni`n4tI^plBo0c<9_P65TnO{4S<sSvr
z1~Lm|Iu$+K3XXgRJBoJ}rOCU`gGS~Ru<_>oy1Pef!LJqg3CCZ%bjR7pZW2y`?rB~3
zgK}(jq$ICHgL`4<n`YB~rpt!L3b@CIYR@=kXOL?x_jPKu>Ge<-%Dg#pMLA3zqLYg1
z>dR=m?1$I7^#1SM&fbJ`yZceQ%87~wPqX^{l$-0NtPcv{ml*4}4y7`%ZzUrJ#D35p
zxETjkc3k?Pi269V;Z3Dyo<)scflqb4{2!T6)T-1^q31ly;_Bk<AhW9L_^CL#@_kG8
zN@MaSbH0~L8H}{BT*J+li>e_X>=<1o{J_U1_nDo-tyrRciQRzLU&tqs{CY(7NoH%k
z%#qEM|M%g(L^Ck>FP{m*>D(gwMG>^^toj*7&jSmuD$Rv<k6^mF(rc6Vgh-M^a7=@d
z6rpmM@lMoh@EmX6HA%lcNuI1zT;uGWxt=F4MyliXyE&(6c{wx1K+tBZQF=xl=2s<6
zG&wPI_6r79dM-3~ZwSU*j@oosbK#Wp=;<h17@f@hUyQ>tpc=(#wNF>ER$_aqLZE!E
zwX$!&c?DXoLS|yViQPNzEnL{S>zm(?!{E-FjkCWIk*8KI7Y5Ee(Y8F#alMeteOZ=!
z5tB4irk|Z7iU9T1HEih>$)<v{P8s(>fv1&Q`c~o~qGD~}%@E#gdbVOB(@Cd4gvg8p
z0~ibh7le|tH6msbk1Ut_MnxsD4><%;J(j;k8>`OtY*Vr-hTSpBdbhYgtY?G|=~#{B
zf{YCxz>DwbH#7jy@Tw#uW_p>$M{p(?Vu}AdSU8<@sH^12zh8!;I0f6o@fD~PNLrYE
z8}WDmOaiRI$6ZM*_Kp8y5l35tdEAflU$+UL+BoDra%-jN_-z`xJ`JXY0M&y}M!Wo3
z^P13Y?2bjaz>{5On+q_ohQ<IaWKLS=uYq=~C2zJpg1=v0$eyWh5%49`YHR6)mXws%
z6|f<`+NBX|DtK4e%JgSw^8qrZfj30O*=7h5&%x74;X(m+)9{dzd1Vm@D(fdkmln35
zD60kwt*FEH#FfS+n@-VS7+dL>0*j<jF`&vSC|dAncx4k?T9X2>nl$X1kfMlAmb&=M
zknM$;Ty+_d1nXltb9a!65vKY>r@rSEd>+=zv1=%J3&SzfcK-_r^H3;VhGBNA<@16V
zp|?=7#_GKR=LqO?U_2kkjX6E^-FxhT*n`KhvwLu1@xfDd4yW(EtP{4vaWi^0{+CUV
zf-2~lc^eeUF0o`r$}?I0#U}vw&V3l<CUml|t^iC3U;@a*<76LoUVwd_f>DW#+070y
zE(Ow@vcn<eQ@%1*v|jvxy~A;A&lMy1y#SdBI49V4UV(K#bDnf;ze2{F;Hg6nu9goP
z(>x5M_%DzchXiAKqisII(z-O4=iW=P!(ky(vwh?jVR!;*CE=#sE)(_*?<fs|q++6~
zZ2|7h9&)*)*%qh!mCv*m_D9YeSjGJ>E5?xS5Vg43mMA?OkQ5`M+Ve+Hfw^q^v@yLL
zqK8|!Rdt@+AA;L?pYVU0w8w{BDQ?|;_}gb{VD@Pw&lnJU;R)VLv5k@JO}h`q%oVJV
zN_Hk6?R^f?w^2+~LTFoXPGqQ)9?GAL(^b@TS7>CSt98<i?BFOz^z$9CDH)blo&#H(
zjC_)SDU_RtH4s%48EqiSCM(8qm9tpb3Oc0!J|QAHjUV#3u8E={a=p7cWRH(}i`7ru
zy}auh)Ce-#fL}Apf7<6Bxv><3wHe7~8qw5t?0_a$uu$FqMgP5o*B@aG@eI@yM<%N8
z=eWHy{cKrhpO~01jm*N00G%3U)0So1e)e9`fh8kqSPwj;bHSpQ03~&Ny-5d*(gY21
zQmhWQ$?^WUyw8lczR^kYvcBykka3oo<+JGVOYM3G*_w!^4~q!2IK)OL`kZSFS;b!a
zrFX9SZ5+uMgVh&SEwwMSkh%35tfwhYs?V!y{C|DD2RxPiA3uyFWSnDVk7Mt>GUJ%Z
z%1*M0WEZ83?7degBT2G~QnGhOL`Jf*SE!UC&*y5~_wWAypXa$>>eXH8T<1F9>-+tD
z-s>X?1o~&#7PwdJMIlQ(G#~h^BqEFf@IcRM7ApmB^3&^;!|~G8)@z3wV?bg^rs#}6
zuJx#P|7KsF!jHqY{J_K8CH!Nf`=35Og63!}49!n$!^zm_32ie+2Pz#4bqUv|%7asH
zjAyx&KYxWj@;OB2W;rhcauj_{(SOI@$akl6CenHN*?5odpz6dVA1GUOxmP<Uw9=e1
zLDpCS$z2jxeVrK6FnAc07~bRvl&UtDS71*yUn-Twp!o1$dZoRwpF&W_+AtVqY*B3n
zugnn{|5)SQQ$}F<n=<BN62&vdJ0WGuOlOzgwzS66q4+Hf@Fu~E?V$B#4;H|+=~q!}
z{#Dxa4Ms4XyGmE&6Y=+UZKCh}z;~3*ygBu;O=4}JM4j1z7WTv@N&aC*BehZ074B>v
z(@E;2_Bw@vPh51&{09YTxBjei{9M=;?|!|zDvW2t6=V#D4|U6A%jbz*2p4DjjGE~f
z63!%pDCV_7QaUjX&nZhWn6A*av|N&lqifSGv%DRD0Jlp8COo@td>v~vXHl)ZwH|qt
z#_?#mTawXyG-q3svudcI26OdKifUrxRG!7{;euW$j+{F33oOkgp0NKq5LofF04z=A
z;K}8JE$|2))aL`I9yIe(7?;2efcl`ul<v;)xAmQ`bSQo0(m#JqdD4X0^ebAtC-cP!
zb}m3cRwzx#g4u)aJ-<gx7Zypo(NUF*vqej~o#9I@C@bedY83XjTI_*MO0Dke$Q3vW
za40jC0u|qgV}v@N1pEJhc_n)w26gHCK`Ul;g9GoVVsdDJS6?MaSjtyy&a_4I)kuD8
zA?CnrsBmDkzXL`s;1*e8<*ce1F%*swBN*2C1;GTrrUdBsyO7C|6wWQcFt=k7bk|<^
zwf__vkv=6T?3#VtdF8DOc*l{s7;|%hffGy1cr@u;9LH+aL>HdAy-ROmWK6T5QswGS
zfUX)aRO67KX_@}Y@7I&-!ahjoXs&Okt;Z!B+22PO-v}Uq!Fy&98#$_;CU|ZaVb6tL
zJYh3c)COc6K}<oxC?kXKrca4&qQrxn<(qxcFZr>Umhuerqb|@PBEi%%IRcCp&<29u
zs9HZULaj3O2Y3T>)^w-^EvTG$ROtQ9OqL(p+SbEEgygh<U+qUE<W)*q#Zv1l%UEUR
z>4!jwgQ<KsYwCt^AxSVmIn@^C^7JL4V(ih!G*<qtH4I}~1`}Muj^P8o+(~?v*}h01
znY}5yy3D?JVgni10htHslWJKfAHvp!V93Q@rVT&YgK*JTD3P-EcuEx?BOdB>{bOso
zhp%9Svn#O2nYZ^U-qt-p<KBigFZFh{u}e4+C@nKye=8456*~VNXicwwAY`~FJg7u&
zf#lX992~_!E$k^H3#x18ZaUHMKop7iU!|a740Y<b%bphzb5$*j?LO>4a($64uPX&=
zXF;Z%0~7xiFRfR4nS*H;rci}~+!5)#_7&zL*{jVmSgC!x*dmZ1Kut-L^M|)_);c?5
z{09nBw#g2zGJogzv+OY_S)>#iU2yS`hhNMkX(=KQ|FhX}hO$pt_B6s)Ij`^I`?Iyg
zr(M31Sw;B?!)3s-fxm^5Eow)gn2hA^09uvs+(xPNpq_$O(uK-V%rDDE&*-t^nFsC!
ztl>uP>(el60ETD%s5czhyUDIRzXJQ-*l6V;R4VZ3%vtC)94}w_yHb$#@<5IL2xcn~
zAKErGD%$<C9%|@RA3~&?=28(6@wB$#z1p%|S$16~FFD#a$c$w>63b?LTh^PaL%6N#
zCghEAP@4x(tXEplQ7|~VmM_Uey<fh9)@d8VKiV9L9V4|l{{?j4(BF%M@Ptgw!_(gB
ztQv3>rD|RpE35+h+_zUR67Nc%{eobazrMN@;i|@+$|-JW2Zf!Fv^@R94s-SUlW5!r
zG#ywI66-}i9Y+xR^2!==ZYmTA1j3x7F%%oc0eH0q&b+U(>x3_t<_B$)w(+`|BKaVq
z5{=qc++7|#;tJ_Ks#y(?>I*(NR1%_=D?5zTgqRZ_`!IbeskS&sOiBP&*10SKwIJxW
z47jxhgKbSU>goA2l>b~GJzSr;>Z1@m<tCXzEy11IhlKgsgf`}v>I~n*y$8dH6Se@*
z3?wZHRM5^aK1Vm8e2E|ZA`0pN3db($cDJYR7qd6ByBOa|Dg;D<sNx5d5vZuwn2VAa
z^ia`dpXQBaWcDPyr$$8O7A|m6@mZfi-=BRAPJJyP8I*pg)%1|!TGhIbRkBGtL{u4Z
z?FvYAoIHzU>K31~xn#TQHUWzjfx6tM<Xe_mDXI)5XC2QD2#J~MW{HIfUzD%m(W8ag
zMtS!?-SQK2wXn$#S3py6U<P4lFk9w}B#pAj^(3*!{Q)A__L|vAgY$3FegUPwA8=;a
z%IDd19H{;9z(!{MvIsz=qYa$0ZaDc@dNFcs!73sy<JCyEuD1od@Nr}i-9{<shq6yQ
z&ngXp6a@g1Lg0k0-U?IsZ$K_9gb0_PsRQ5vUT0bRmkN;sa&dLCTSp|lIg3?;uCq&1
z-t@5)Z_cl_?@lg{_zm0szM641sEg!sjI~7wtt+=+9wdHYxnlVOu#o49G|_yT9PbeU
z4*aiOL(hfu!X3{?<oc3xLP($&&XlHFPV@sKXcG*Y6XPf@Xcf_phTF}Z(!bv}TI>yz
zRm#gkhM}nT5$j&8fn}S0@Kv3PJH7Xn<BE;o8dU8hu5Uu7gzE8-&@f<#_VI5`HI_F1
zIeSBcV5>p-6lBlP1%r?T99OR&uNP#aPm8@%y@*&M+g_?4Yl`Q(t7-NIuy0*Emk9Y`
z5T06`D$0kny{)px{n=3O8)z_5SDZ&uTuvu{aHaVe+Ix?Z13>gU57nSBae6P^(E<5x
zW<+HAnMNr8iPrYIpZlNmCfQQRZ{*f7ckB0eW#F@&gsxaTwXyn4FwU*<(7Wma@n4xz
z(XPbSzEOWG;(t5-t9bbF&7t$qbeTAllfNj1JzA!i?HInPph{ztS`w|Q$umZKR@(wf
zKSH}FP*5r`mqY&EJYOoJFYZtE1-jLYw*8@%?hw&fs0x;YNoqYRaW>HlvE4xk7ws{1
za@VlnA#goW3xA7=IJzLhUr_XdzG9(L+kpQX6vyfeVu6id&;w94`{+6X8kG#{U;pPh
zP(gTd=jVo`+u=c?!kFQB>2cnS!2J(k#cSg1yX-da#8);$^O%{=%79$qGb$V>-N=4b
zuLu0o$-vJ*bL}-^mU=>Ee+CD7oJPx$W~)sY>9qa9Zs8-;p1EE<yqXcAN(rZ2l$82z
zgtU+_%51@WNzU?5$f@Oc`DO#-2T3JU;Q(Jdg=#>aY0m*&bfNJlv|7z49YFC!exY>)
z^Jim21(%?qR;Xb%y-u#I>@tXp!|CbQ5laX2U}msB>ELj%y>S+{X{bOrJuhzRoqFdo
zSP=G-+3)lLD7fKTtGTmpVCZBi4^l$Jk<&Y%fTk)yZo7W)GBTA+=Uqzx<TZ_2dXsi&
zg9^%Keeb@8VL8ziGg=$KwUibZee${SJub<w6s5#Ug*I}eyE?0CF#8vAXYo`1y+mB}
z7By^d4Xb3vTwn1;u4IbXsBEsG=jl-;U8RcE=sf4pA4EOLI#7iM#9D-kpXei++u2Rz
zE+M^w(z*PIxqGA^n)_;fiI@lPA(<;6ENWASc1IvJGZzCPB%YSXAD>Ovrc=1)11Ja3
zGK*MGU<@Hhtv{=r>fdW1WC_02I?7R_i9f`qO!5I{vE5HmO|94%EIhW0-HmIK0FzBQ
zs3lk{k+K@L@p5B_=If(AO5D{R46Z{=;C>l@z%h6Z>2lUn7c6>^gHxy$WQzFJB3fz^
zyTB1+?|Zn_P$+Q2Mn{n4>91={hloOcec{K$)7s>l#g}7__w5l0v%Jh>2Yjik_II1!
zDm;irScozkDsLoZN^1|gQM;kNZ_w@AN-gMo)gu<ov~3Y_D8KCNRu{bzLquKfgbcde
z@B|I|DN*~!UyqU{2)H@w4c=TNoXOt<3t8u@Lm=4LbbRXzwf{K;QNfWYSSv~KCgN(W
z=kWvt)OzE6ytlu8fi$R)$8lvh)Gd=CH|mx7S43|C{9XTN=VI&8Hzfvz>ra`yMWVTe
zM7Y$Pa(zv9`4e@cpn+2H;8>TQb+@GM_7O{1uUVczCKbBkbwDi^4sL(j|GYXb8iw+|
zlO%!uG8v-}@-Ura+`6UAbTU!a()dq1^p~OT1QnyR4v~OR$D|in8)(_bMo@U<h!PSo
z8Smb`*`JYj!=HFgdUo#RXMvOmF-0}PAR;Iy&+;o7@phj?-Mn2neJPV_Rw6?5xMF2M
zBcd+n@P2)f2DgxOe3bVb7?NVB6=4RZ=F7JicbSy1JJq8aKmL@KfY#5bLLC(_##@>!
zOt?Pv^0rxBn|wzJ){4ld8)%xDnE35-iqML@=YgsOw)sNxwZsq$>5>q{(YE?=?2$D=
z%h1Qr#v7o70=Q-sKET~_b-LathsLB<XCcY^S-csc%4dN%NfrLOt8#4~#IuX1uZ`J6
z_}}*C@=_hnly=WUM-krgg+Z4TFS0cu4K0gDsLcN-V2NX-<2CtEu+fm);<K(lhR(Yq
zH~Ij)8}eoN!$SYA^DZ)^)R`?a48aEZBk8AA`PMX@4{uJYR?~3HXlCMBje<ohLPQ#n
zu@0FqBKey-W3McM%_vXNkspMxm}$sM0QSQ=5@g>D%UE*xt)PMX-`mzWY2O_F2t51<
zAzJLqNDwEIW9l<p0I2=m{iVHRAfw+XeerzM4fTrSKwcj$yP;&dYa|vR@rR}rd28L;
zI?6b%iu*P|q?!O`R7lG99y9T_#Gv(y>upwkrr%`V?|My!72d=#Y!M{}8#Iz?R-7$8
z5_(c<+copO?2+-%+k__M24>`hx5<8JkM%A^v5MQaMf&B&C=7xQA>{m#6ywE1<5HvT
zMup?M`iN{Y-IE7fe7@r*pB2L+pE+YWg`Kj(pX1w&MxUWJqmaC#yb_6xWE#pPm+|{%
z(jjgox0q@7T46B!^%e5W1l`}4^!w+(IN**x+U(^uNl|?d+N!5Lb(-l1&Y8OxWs(##
zqhJ=JDC0(Sl!U#zQE#>RnQV!um8(Cx-zuhKbxa;5C{DAbl1rV~)t6bej8a451~DEm
zunOt(`ux`jkB7;dzPRF6R^VloHOv>J`x!V|jCau$|KV~>0ByLbz@?>$Es(bT6?MVW
zgnY45;!_E+Q5vN==@VI(2E|3)IN2v5iK$g9o9e@yf>~Dr!PX%Yf$`U-LBtUMgpd=J
z832Z#-I8--H0@`t5+f=D*(fMY)L;|rm*uit@)&i2K~JC|N<r^<S=o#dBLkU}qu1>{
z<{3c>r_kP@SMt%%FzG;I1EK8ye%S}ZE+?RDf<A@u0PO|^$(Rsr*i`K@_!{eAM3i}E
z6d^0pkAWWL*02rH+DXtcgm$(DCT{P8u);iKFhAn?SW17Ec!b9w61)Z)xhVhl5DRHq
zfQ_0gXl6=IG^C|5AkA3`Cuhlf_nm`>Y2X1I`~1yBC+)d|o(eZVUgzDtaYqC9v8nO@
zJk_Xc@Kmczx4T&95#7BlB3HvT##Sq6E2Ue+Z22&9PDdsWdNTQeO0OHu07p6Y=VYx}
ze0#nRWe4jsm^K_~QLX;sWw~EY4bTlB^Z~t@Swny<Ze!oYU2m`IK!9$`Auwk<%H#*^
z-RC?rybQD2!Z3kIwH`G<bYSp2svQ?_KP8VD<?ypWQGAfBgDG!->BL?&`rFt3Vhew=
z&FtijE2W)1gJm_vhdR5%e#A^7bflK>j@;sWnfw7%g$UeZTJlC9L*}duB=~AA3kSlz
zhF&jL%Km{G)R8S`5JxZIo}C|iwZ9%AUi}1cj5+b(=KB|j1d*ob)BTZi_(x2=Ph-#R
zgCnubVV&rzG5AnvDxDsP(a?Mj{OqX|>pgPn58wt1EI=1}g8@0skymuMCSocgh^f^h
zy4##n?CcigGn+@!r;rAuW==R1CkiY?9~z}=6MKn<@?(Mq;lz%>S0%ZN=(QM7I^B_^
z^H1+}?hT!X3Kg1^7r9@N0TRl&Fcnro^!5)(a^i@o(a@)@W^Du=7PeN5co0G}^)v+v
zNs$THKrX9aj&D(krXk(<XW2z6Qvvf8ts?v|j+cT~?ux)e?W3cc9pjhkS3wb@Wt|PX
zRW(hG-ei2vmnVj1)N3$Wi%o*0B|WVqmx^%aQlwO<?l8<U!#iI{{oTV;I{lLCo8q`C
zKEBF8RVkHuNY<v4(&v3L@$ht_N@eag-&5AbRu>1X_X}t$s+d(IQ>3Wcbd7TVsUWee
zdI+TmvbcCgT-%1w;=X5j=*(hhQp)(DUooDpjZqds(iCb%F7XgR71lD=_Oymnzn#!C
z&SDsLOVFdRPxUz0x%ESqsHY$jgFF>h);9`S&~UfK3B1nnYk!24(Ehc=6=HedAKMsO
zLVIfNLjG{qzs&Qm74?t$qz?5BxKXoux^rpJGFGE#!%2j=!}y6#6anRLtJbm$voq0e
zvW<rh#<4k;%k;Gn+E|r*VlV@c@^e6M1ipP+N}ZIcVA;icT=mKH&%V;Y{;Oqd(E?o)
z0Rmwr;2ISeC0++~=l}{3a~hFXYEF-3HvJ&UbvAs)g6G1?oL83QyuXyFe>bfV{3l!%
z_;LybXKsRPCW)Ewm3CJiP}QiCG#sJ)X=kpud)>Veo~rEDl@-tLSin|F^z<3fis51R
zpy@hR8l5N&(GQ5U{m>9CqL3PmQN-?Do?F0r<dl&P0xpn1Cr&W^6clU?fQQ5K>9-h^
z<t21i93JyiyA!a}4KQfb|M3JGu(Q4c1&tr}-4CF$QDw(SF`p@0GRZn-BeO&p!Wq7O
zo4FX`c?#V4wz{u=Y6SE_&UP@=*>i58bOu1>6f&yN{-iH3XDERF?gqp#O#6Kx4K><-
zj~H~2eGri8@k1IvGfZ=yNEX(!kPQ%zJlB3uVeaTp`UtB0<sW~2dihBEDRKTmcpGO>
zcL!7;%0=^ln4a9!Ld!rzCO1y7vPc<r3J}c6+Dz(6!1}ZjsSY(JbaAPg>#A0f*-^=(
zE|K*5bW8;h_NmNf%RV8!izQJi!eFGi$2A&GlDZMvMD|&9<W)%3`q4UZnsJJ78=`k_
zC4D(m!`XQI==O9v)8`S*T~!AOtR(h8gv|IDQfzE-!wLO!{lrL*IG#M=?#dgd*_JSL
zMl^{&Ydv(UI>}jrOS`zDOI~SC?7f18ApHcn8Ax3Fn{SrF)_3{-KKB{b&q@M6x5)=0
zmZ3l52I0CWnmHaO>^6f4!kjujXnI%WM#Dgh$B%pYht8_djds$IyG3fMf<Q|$&|Eg2
zQ2(YYp%~B83k`&RnV6$Pbw5ceHU*^_b5>X{$#UytHeJI58T9e58~x|5nfISjHByy$
zrK;Nw>{6t1B_aMwl&qP(@wq#wnHni?YWd0l!#g>qg%FbKcQZI_@0>dSV`!_ve@V%5
zCpLX<pcg>>IL=pzY^rzTrdcQ#@zFLCSi6&ht!yVj_<=_O%ZscFcoyOUqu?+4@NvS;
zzx;#^FSwAD3FsQd;c(mU2K{N~-sQ1u6$x4tuTdkYGhQXj63Fe06u?!hi@!9gac{P)
zrQcGN@6nm^+!N=03_`4*b}x9&JSL_Q_%?HkocUXQVqWkHjSD)-6RtGueJ7;PZlCdX
zdl+6qc{+qG`r4elk`<&Uoo@gD2y~p8TV@Su(PpLKb@JTK)(6Etv<^tvFJVPj1(=H3
zX_^iVlRV4q_ca8KEOnIVm$VBOaXA(f6&KTk!b@f=*s|YwmJA_VThkJ3YJ2s!j9&T>
zq4v@u@_oFCM!(>ImCkLVD$u)+*u4w}e7#_f*2Tm$gZtqVpbH4j`Q|5xI9-6$7VU$u
z6WkV!QMT_Oo3ufh0?HWBvz{T7WT63b0Gu{{T!~=C0)i87`V#Lb@jS<re*Jfj6rm+;
zInEE3TC9PUV{dBisN&HVT=@n+=qL2Tai;Gg?Ivth&{Dndr=$EH=Qw1Rt<&l&uIHFI
zzspE!K)!cca0gXfy~F+sQO0DlxTe22xqa(|__S`FP^j{ED6Y>o$YV~-sY+ZN;t0vC
z2hr;{6f$n>iX1zs!IFpvnZer~Iq*;?y^A~@=bh!-4N5sAS#_<@{VGfgL4Dk#5SIuK
zb$88EX`i>GJJRP>9;-FlX^DP~iesx_qd$UWp+q$aC8$wH7I*$&Akp;CV>3IqES;@o
zcTtSAv&D3=fJj}lpZsFvH971(nuP$(hj>|>K^e!~<;BjazxFytXeR$aQVK1`s-0{P
zOB#keL|+PJV9g!`6f7>_F54<TtvGhO72`op?$UOBs>b!iqQzUz%&B@<1aCVTH=(7?
zU2T7evmH)+LWuFBbJx&<<;wbN{Qk85l(giFa7T4p;0KqsCk5-oiF{gMZht}mKl*F*
zXcgMOqc}?<o$HOqWmdq}iFk4XQj@*mTk^;a)av^8N=%Mon;P*^Vo2QOSSk$BeN~eV
zl0WY=rnd5USb>g~$gt;|(~Ss)IF+R>jsH6QZs?2jqxN9yrnR44)zr>Exb4g1DwjEI
z{&4hDanb=S%c@<ByXE<{AnH8VYdn<5#hREdX)j&W?k2+}HQ;uFAC&jnP$Gj3@KHn@
zxcBC3-k5)d7Ye8231KVC(^lV0*CerLuKFSObI&F5p%`Z7_%Gu~JNg|mYvO?+r#<Nf
zBYtce>t|u5i(EL~k=YA3&JvzIfPnC^dzc>O!Zz$0M7VxxsF#r~z;Aa}El~rZ*jrxJ
z(++oNjnV$?^v8w*lNT)aSZD%2%I5MPDi$2dvchK(TF$m=UqB^?Tb5tNH+N67?H;BT
zpYUVD52262P;=jst7A&t^;T0>9oC&{WxIxoFWxOXXw(ZMGBSWFw`pYZ@RF*z>Y_V6
zUw@vj?Kre!P&-}1763`PMd=5VA7F-MNc&7~27kLMUVp=u$fp^{qv^b7i!+x*9^=zq
zu#bI@Ky)p~Sf9@3rLVj-h=9E24QSgwpkc5MsFtNtqlS2qWG@i%LP6``;ciJkRBoqg
zW?41QVQ^zMTEPK2l7KxgF79k8x#sW><+?8-H)GiZb%pw+)TI`(UCU>af~?h(IA<P)
zyGEXT!M?!Ob3WHMwpZ|ct(q?Q7lH}VGeila?2%YawYEAe8#0D$u_pLjIIqeLnyliv
z8gS?E-7NQ;h~JBeU!f)-$$<ThR#-Y`20@(=R!8pC-?|uTcl<RQx6aqn;(DYU6DNqn
zh706Ry_)QX1zVd(V({SiD{7*i0UCgg5_w}i2Qf#cV7Gpy!RU$Quqan22t}-wV&~7&
zn74A@3q1U}O>A3dsbl`(uO1kLv%x_HH)OAy4>siJcx4vAr}(iOm1Z%I|BwnDi68tS
zUXDzUW#1JuXdx+9iVp#c*qkuGQ^k<h?p3FY^>r?J{~JT-rqF*P%*|UbjM%}Po*R0p
zz~>zGK8201ri;jg7lR_g!}7;s<wm-kDOvkl0EYS?nrK9o5c_M3MRr+?@}#m#mKjUU
z5!|34HJ$+E9~ejf^G1JB23vBtR+uW?EFe5ILV<jnsOdPf>;7Nl8o3@+0`y~k_Nl*h
zGczafU|W8Cj}#bA&ep9$dposcP>Jo`D=3Zlw}*qOIy3~nP2Z=Go(<ud^4Udf+X(l&
z0N@jSG#LKv9SrlMDgd=-M0vYQ;(xyJTUgLArtx#=>aTCaSaVmx`T|f5dSbL&38bLy
zs{Rs||K&a4<m=*rLV~XSVejwi*^pxk%B~{6dqV$^BY!&Ygr??|^W_iKSjX_q_9)0t
zxXQtS7(>PD1p|N(4-K808E7?N-p8Up8XVT|kz{oL+Md^dPqx^WK~LFfbq500rGVNl
zO651*%#UEURdTfUHU{nwN;L&Yu>Mh_wF!ZT>(FHB?ARk@@~<N7&(A14MC5deyOd~!
zVi~p_!1cQqpm{-HQ`m>0$zBIFEMnYRam;<cc3&rJ2Tp7Xa+w$HrB5!QfGPfocrQ&J
zeFqK*sXn`@E|Ol-{RyacVstq~%)gCogyFTxb(Ty2b~e8{Yb;{e>bmlh{wE6_j3ajk
z2=_}CD=!eFFkHEqYlMb!V?GB5gKkY0mHk(Le5}l{w`ucr{*l2zm~h}rx_+eK4bgw!
zktf~`Hs%|S{-^W5lBbI3TMNMdddK&zIiwk^PQ8R!0w#@<KcL8dU{Kfrb~AsN#R{ln
zYmY5NYaq!ru+f9&`a+d$J5+rGC&4E?33g@5ggjvXV8iOwmX-s6on+IU_J+|zlMN(H
z{Gvz7Pe3*DV0;&#*`0AE5S{~L3zD=B!CRmwxA6o-Mj1Q?bHMRg0>-j}m@%+0Hs-h!
zBtlH)JZPL{tpvUes=zeA0Z=<_1I-LW%W_m+Y*n{Oth`n|{K(l1KroWE)UihdQ&j+u
zvnY(uK~yu)Qj4vcWle3S$$~XUeH6N|%V{^eIWuAzWrqSRLH-LL5N!&?Thy(VK73no
z!96z|+fJwiv}~fo{WG^6@40e^;M-1eEF3a4i=yswaeq8Kog?|`;rn%ut8Tzc0e)cU
zT(KW`HmLaY?><MKo3*CPyQoKG>{mz?WQV|m5aur?I9~5Bp4@jv`hDmnAL!u43Olsi
zz&r#^!X99u@qD4__F=>rdi2@v(7(nrNXK9Ngyarzb$mlXSbi#zk_(wR+ky^+;<tPT
zLE=@gS#|FLS_|sloA2D7g5;u&>WsPOk$?b5i}(AErJVq>2~?GdiGHbVW1*#n3RL+0
z+$ahYgUlS@3d5QB12;m0nqcTItY~cw%B*FWX_>!v)|txYk(omou%4DNh4-SawDEn(
zq(^TMk`B+0HwFgUvKbH;l4U%FN8%9Rk42QPwTL+3qoKV9AdbQl);+orA&f9b_6mY4
z!=+@40}pwzL0I!)>=yK%$h2YFO><di?reOIl=<SGj<Lm#$l?Bt_}k}ZoM#TN1J*!L
zK2ho+hQI!^u{&o#7Ixigy0QK5(=ba7F`$wYonjo*sx`fjl=iTF-42o;Y<b`4a^4Ub
zK+bG`Lbv@gp)Ubj_=Adtm0LCa04N+nV+P2M^ITgLGQ$A~*3M5XOLk_#4c;Kj5-U_E
znWNhVN%k`+24zF5KXov;!<m?98bf>8+NI5yYXFP_EGO8`J_Z#XJHeOD_qVOJiY%{;
zI~0jL0A(-B6-MCJ8J;)64H?LX1{5O;lL(T#Wmk?6kcd3K5h_T`8FuUqAV93lgN!M4
zKf}XuT7#y9&p($CbeV2YIN}IpUgj3ZW!96$M7Ci)-Fq<JoW?U|?Aa<HAY98e4ux$$
zyo-2Jdq7$&3XvNNG*}&7$ZMzkZUpoN<%zaIa0Mg7KoEM;CW5I$Lkvs(9Si^fW^`M3
z*pd-N`ScbhCVCVX$_5P6nB$3FPTo0;Ti4w3Zmrv1968Jwm6{|vQ|MQ5>*W><Lx?7=
z_J8nUsiXArpB-d|gY_K-iZE<E&Rd$y{mT3wm7?5B!Q@ZrkA!4NeLrr9W~s_-%IMTo
z3uv1<MlNYfUIDY4<k(60V~sw;@}9HoSi{aU=(Z+@XjOA6%?^g=hII#kpLzK#&`oiB
zT^E@32200ms^iWaGen=HN<+qIXxSqs7JFZ?4|;Kl);-{hv7c>sT)3N?h_+_632a7*
zZM0BDdaTzsD_V<h2Pse6t3GSP!xK#%bnRf9SU{u^GapH{2qRegZIr=EK&d3g{|OTN
z`X}@%TTPE#>;}eGeaw$QDugS9m*86JkE;g9X`2GrWwKeNnWE`~Qx5_^)UDx(NFH9$
zSp^g}jDf46BVd&FWwG%5Aw!L5cZJ*cpN*57GwJb%!U-&hlJqD`r}t#A|EdV@&S6cL
zv&LVzwur_dq;!Xb78gTolV>Nj@0>X;=36C9*FWDZu>y^o3=EOq=o4qr;zpk)=IX|G
znyryFB$LwSu;p(HcS}0pD>#vrz%`$Ez|+CBsG{2x7Qw5q9aWOQ{go}LFq3I^@Yatb
zDFLhDA8vP_2~^bzqC5U69h{)dYRPefAr@^oa@KaDtIc0}FKOz2%Z5&t5hE}d$#jbp
z?Gisu$cIHoCOBq}AF~H}UQ|xnGXb$S`~murMSB9iTXe26incaiuNsh68mR==Xh>@3
z)WwJA;e7XV`Prx`TfX@mb-M5HDvXVuG~GK1S*ZlaWhC_N%sNLkem3T_M2M1pZaX|d
zly@mgylZ-RKTy=osv)4xSQ;DHA(gF&+@?xr?j?MxM4_egcZoUllyO&*o!YZ@u&*eE
zav6#6O2o0~>Ijor&c_#J>EpiW4it&RSVvf=k5Sr)VtZKyZipy$8Xmb$9GgHW7~6al
z_rm(RKr1mj3HouX7@6%IExDf#X56#&s)_7*=Dc4)!Bk+vMajWYs*yMLlv14GHkr;e
z<Old%M^BKG;)b#nIFE1KKuH)-RFzhSW5wds1Y0hUP!biOMY}v5>*u2R*H73gJ7FG(
zKPkhF5^K+Va>sF68eJjUrc4_xOPWS3t{~O##6mrCa~7AH=STMhISa1dKC+Cwj_LLt
zf;Xcpd-6J|N5=+Ol3EmL7<Udyr&)C-F_-Ikza1izU=|mn9JU)F6cH+n*il)>L6W^g
zE`stxHdbZjlUCV7s60l$)P;LgA&Ms->=EOa1TWXu%S$!v<CfNzW|QDvS3a2+ky3~6
zeKXx-R>$g!*sBk0xWbg1Tfq!t-4bv8<1fYJ=%iBk;>pkFx1+ebh(fw$Pw~f{N|e=h
zrln#Lh>dh=^TG2x)`}^@<a9++waOBzm&rO<iDaL?@^R__e6X1}=H0IN-EvdhIYG%S
z|Dh|GHfypSCYuBsVV_qGT1QYsS+<hV(+nnki%gSKJW=F)oCO!pake$=HV3oKTlF&j
z;Vk>!n04jXu7tk!RybnGSvVQrud}D4deCHP-~+N*7cD6}4mI8R-uPY(@#Ayo*8%pS
zw?DS9MKHEhJ|+JE$plFu{ET#7spem&)zao(g+Z9A33?)m_0O$tNwG-}Is6EdL_)>6
zGCr8}p3m;{qkMHJ?V>8-aCl*AllJgJK?_?0d&Xm((tPg2>5{kt!rF}K8zzblt#@hV
z&ZzKgS@*t|3E>&GpQf0Z%?k3NU2g<Nqqi&h0rx4T-lmS8Ow&hYB(0C1+%Q<T7bDt~
zHH-;)ceS+63BTcB_grzCBfIlaE|oEpG4`WooTBHQdd@u4E^dA{RVhHfR)}6DJY=;C
z#Io9y305ktLd{k?H+jax<pcgrfu!I}$Hzj{RaJs#*>9RhCB{8IG8-&Q_ca1ENfZ&Z
zaZw~$cI-6P)y-#%&8vwInR+8VAchD}m?V0}K}L;L(?IKzYUaFt`a-Sah@~7ZXHx||
zJ?ZK9@75A*$Ud$i^B7NeR<WF^90*fy<G3rC@o6f-)ajWx>Sm0&=&9<hQ+CT@CUmA|
zO>Ez{N)rd2cghS335GZbv_V!&c`GHYUq@azUQ98IE9-(jf%skuLt|%eD(zNep0k#H
z*998OR->LC@l$TTJZ<>(JBPl$qi4nsGj8*Jp!{;joj4)cgi&bkFx-#lxs)J$5V3>L
zq^oBY!{<XiMLQbn;~o2Z_3HNn&tK#hEkibhG6jaoE8AmbK|cN)iVxI4k@QJbGf1*a
zKc=sP1o!b>fR%F7Ugex|n2IMwPw0brQbfBR_II<(lh|5y-^KjI2a<}4p)Z?+W7T>H
zZ%Tk0%qqyPrb~{dg+z3kt5+nZ$P9uLTaIVG_{2=`AqEH_ks5WPxn&>>;Mq<#1my}R
zsko+4=V!DXX*CmM@)9Ht-v^Wx;s;qKa5MBHAok$ym2FLzg^Ju6il+I9n@;;?1+gPE
z1PtB@K)iNpYdIbhjm%&pR26U^$iz-IPY_Z{Jz-E!?GE-KP}(ar#W+BIV<J(4Y-lr6
z#+FpLje3!>guvvLUN)T+kLb4Jw`^ijP3y~0zqH{*7wRc|q56=L(w&$Q2X(St&h3Yv
zO+g{u_Or2LcGT8k?+`=n>CQ2)<md_c4fgoi-pQX2Ys9cgPS*xTNgq%Rrc~gl$UnZ(
z(dZk$eYkKsVu5fDeF62vXszUmiaJgfv%wEIG-q-X$6*dcw1X(rUr$q6oXb_tIoX9i
zu6oOX_wt5a@M$~6OD|P1hxk7mxopu(ZJ!R+QlIVzZXB&jxEipN1~w5egC8wAO?g*n
z03Y|}T;2H2+dI4?y!UgGP1$dpz1MZL=tGvG8G}-RV_dSe@wu8lYTSmgES_v`sQ`~0
zS#g=|*N&;#s`Bg>KZ;QA2w$-;85dcTL{-q?3T$Z9r5k46$;6h|>>6|uJvw9FyT6ga
zQFn(;Kr6lc%{so<yAwK(lQcHIi2jJ>C!qc0w8YlIcVyrTOepEvqmikYV`7iUS=l!9
zifJE?K4~4@epS$lQ$bIoR=KF@_y~EQUY&0ge?nO+5sdvQQ3xcUw#9Tzam<l94&ZJu
z1m_Jp=O#R$zebwD>Pn$Brro)md_1clF@zwR*qv$aZFGaXq0YK&9{N$CePmBJDUGy_
z{aJpdZNeLO66EQ3JTBh*R2G%QWFV-Sf6daBNxRa{-6-UZ56^LYimyjiBPsn3FI5md
zm7I5vjn|qN?vGmJp?^+C6sc_)cC_k%&q_E+Gn0e@mVnW^=%d1%?&Hilx9Q%I$Xnr=
zNA$YlJmE8ZecS9+%SrpU(@htWsfGp*9;tRrM>NX&&8)`o=lBvv-6`16V}Hx`L)TVd
zSKMy5E|KBH(~3-=8==ZFo2wEW+gl>iSEerS5J&f6`R3L8oF!U=800?Q^ULl2B+k2d
zC2qD}wr6JkXYWBelZj`0c}hjNtUXjuiY)#cq2twGN<i#Oe7SwGBi=zF4AYqX@LNoA
zg0Hw-{oRGW^t67`JEP?RZDiT%S?-i9o=gkpE9g3&SDkx?_awzHGCLe)`VfF!$H2Bo
zbSz7bRBtCfT+HdLdx@2WOB;$uZU^}+@~LMW$c7HS+Q`t*xGCnRsyuVln?BJ0oHA)&
zT<*h4gGw^diAVWew~qBRe6jDpZAym5BQLzjY*+HwmwIz|@j<t?=X))Jm`^I@)rj0a
zBwPyJC60SZL8QXND*O<SU)9`7YG2>{M@~SmXy=yVYkz89od}OWk%za>aeD~fGdnf>
z?Cr!7K}0U01KlUmHHwW$<vc=>k*Z0z^%G*{wAQ^g>rsw(s(PBd721eMyc&-7Km4{g
zpD2EwcaUQy?NInE;r$mvnrF8T7h;;?a`%o{b;z%Zp0ieL{Y2-@uy?5Ll6r0V(Dib-
zBT*MrN&r*DouRZ^bxwVAU+*c&@-iCvD`o@*Rm7SnA}>c@;-pR#YA?SUt)-}U%2&O-
zutu@se(4e=9z9<<6PrM}nOAa2gGc-3F@fPzwQA)rQ)EOvr?X67=h(TG%$vE)9J*G=
zRS9CyT7)s4?D{*p30>wLPe{E9(v`&bDpk8^FTL7x?rrwq5m!{?NvpT-sBNha;Zn`R
znDIAhZZNh>Qk>=-iaO`CFQjEpiaWK!HvWOQhmDa>uqsb_?Au{-R|nowJwefBe0ee6
z?VpBKei5m84tPZ0@%S|*d|sdvjitrKqE^ypsC{DH&=lmYBu{-(hziKlom2beoUA+<
z<YdV5HRpO=L$8y0(o4}VQcFbmrS)-Msy7t#G>((RA7~XQvx$|FiGIxC`Z+U%FeUU|
zMK>xdkC?-bU5m@a?$eXjuxi0*4FN+?$*UK}s!j>oupDFj^yO5H^j)1^YZz$Lg0twy
zLBWdWezievvb8GDI%wOW`C25?=FslLuF1)<sNE)mSM)76VerfJIN+4th42vAB#F{5
zFs4VcdEiS2lymOfaxma|hmmW#yZ$CB>S44E+0xeY+<XTqMT?HxT2o_mcJfso7Z%8R
z6=h|lZDqFbRj&00%^#eyeTPZ0>1CCU6Jwt=PxpP7FnUEO>MGBv=P!36PM$XY*IPtr
zAA+NID1Voui|)9a)DI$cqL#C<5jS5La1R`e@yxCSok>jcGE)y4mmCPO?7x&1SI8JG
zx;h_4I?XsKXpX`O?iO3M55x80oD|oR5O<qb`ONFZ&UGHpGg<6{Y|pLY2W;Y~#qQk+
z7p2$unbJeA%Fk4)7b}$I`-*J}UJ^ZlSrF+hYkRos&Z8ZxEZ)@1Mj%3PrdLR&jU!Z6
zklP}{Lx-In9Zxj&EXn!CjYP3Qo^jfFXPu7tZp(HJFZRle?$;cv6@>O#I3E&^n#J*F
z@zN*QJ#Hlr_9U(?rJ_tE<4i?~=j_4&yd<LgoOFrtel}csTvM@!Flk|CLe=P41G=of
z`77kd-0XM}@kU4aq|tdFGm@{*4nGO1DpkAO%N0S-NS#(bxwfhO2Erh^ZX^f!IL30m
zF{4b3H%@rA?e)Sj@v%0&SL1=j<rNvJ<d)af1on8oQA;%5{`EmV*h0Gxj8=K(?u5OZ
zyK;?M+xrzU8No{qJgWbytiW)YaN){>F8j2m^P@RUt(LJy_9M;EHP^KPWaiX2Uxi=w
zgk$f}V);o5SuoXpU=}m&gRUA&g1WU(+=Up`asasStCevcP(6sm{U{LAEi05BDf2cT
ztZ@t%nU9>rd`0__ZBRdmENuHMvE=d8>EHsT^;nTL<#(jsJ_tNmyRcrlSqfb;Uf&Fz
z29th>2m_l%TwT>UIzx<laZXftm?$qxUALW^J#@z+=%$M|0OmwgzA6bd01Sm*RVvZ+
zcj2wQn4`yG-oe4{50V^tzD&|qr&RVSjCD)+=D^)*FUJ+POY(gh92{$Bj_|%M`B2<v
zJHBhvdyT%Rl0ur!i>hY%V}bVe0!D}e%5A!dB#TtG${v${Wh1`<Fw7}v;V^@J&6kSs
zhVD(d6Uo2b_t*4{*&i2fctkx-I{j8Kf|@d&u_b}m;s*LiTfpZdG^=-?{o0@g$!}yl
z3zwDU6Oe7*v?%%Ba!z2&XcA*u#4$&HH0N%GkCrM$pevMyQ#kt9-V$=q77%M@j9TFo
zQ~Rv23N3mEN%6%}vbuI1Enmb9VJt4S<t)jcTi3Vo6zevQ{@#*S&r+#>Y#4L%vJNBT
z=(TINHFfmU^;&fEp4B$L^xVt;Sif;X&x?{qPvly1)b|!@yB@P=s8jeCWcgJhTz?Mg
zkaV8UZ8+`k$FX_UN%>^|(o(`*5fkN;Ay?32?G{c!l$J+^vO;ojS2U-AlZ#h^KYAiq
zKNXuHdJ->!XOKCu-#)E(n#4w{*g2O{mG$Y3SrSdBQt-6HK3?3i8{{Jz@ogD?tG{_%
zP!D&-u(idwL~Sv>oM*Iu%HV%Cgg<z(G64|>p{7%s_Ih3BmNyZe7xXD_YaBcAiyl)%
z?b~j^iTPFojwly=%Mad3)SgLvCS}Xn+_)=amgGsa?ukxHthsYE?Q86RVu$_@7#d_Y
z;M$J1t$5h?9~k)`ko>=N!#}_M_ZJ3;$1wV3xzztk#{PW`<=+;T-@#J<{`OyO88!((
zaGu?3MH4#9-)4lcx%qi`gl1x?b=7_R2q!4|ixd3M_x+<ePdkDqkWvSZP)4ft4;V1s
zU#Pa%Vlfr5Fn)vb-F+Hfk`iV3N0jnMPj?%*ZX~d*2R9g|q9dWMvOHs*YIyhpisjZ%
zN{iW6*WLgBUl3w&a%4Ph$%uD7lOVAM$s9=Fr+rj<G4kY(bE?DL9Ll_2<vv%^bNF-r
zFisV}c*EbD3}T{T<9-byO#tMDt?FgwH*}u+BY7^$P063BPl`Uq2lc@>m&gtrqDc)|
z(~~cB`qOl|WOW8nzAd9DU$MW-V;B<`oJc9Nztsxz=xe|pZgEiU6mt#Avx3lwoYh)u
z&L=mS6yel+5KbNUeTMhF#~<b`SL~Bpso)(3{Lj186da-@R0;+$IbvD_YG+I!w6SLU
zTx75@u!(@9dIq&Ao!0PD2Fgl#6xG)}^GZ1h-lR&J_q`M?va=G{+$~9Lu1+f09AQy-
zl%Nph7C_#Ge%o{KRy(>%z^n&=J>J~lsbfr3-)kwOf;eaWoHVz%7++WNB^%8T=#GMJ
z;>%K(PQHnZDTGy%Zd8Clgx8X};dHF`LL{2KW%Nl<t){9(@dvm;LCb|HK~cNZVcbqm
z?>^k)nJ|?XlhuOzWBU~f@>+zFo8<`7Mv?|0@B9vcuC6dJ2HbYv`i>u{i(eEtQ#>ni
z!7%C0F<2zwC3iZVc6?L3@8zcD@NY;2R9CRMd?X|_kR#pbWWVctL02r`D#S>xfL&gf
z&^hvC)paqf!#Us(Q(sfe;BKcMoO3=`5`Hu%DBQV8PJiT32~YD5GU+<h8^kl<RE!fi
zK~g5-oOW12ehkDHD{%i=$aDiuy0Jg_Ov`hW9%iIi@J^WOR(A3msSay({KG2#eZnrH
z6TR={@V}u~&T2G6^e&$Y1>`w1oiBv?5bC82U5+ZaaWC%mD-!r1Boychj?E4%_D&4&
z8u>xu9#{qlE&C31jD3GIZ;@<4_g5kK?pvMg=hp%+&G$e(Cc<D<=fz2`F#9d8Pz}PR
zvVjN%PD}do<TDrpOWQU3LvtB_eb&EQi<^Sw*sbQR8c`j=A=H_`NUw)&3Ex0%tE==-
zgjDcN`s>M&i-fo<kwFiFe2Sf-cEhf8bd6Si?H)COPmX<vuHd%^w&O~AUXYzjMZ!ND
zPCj>2VfcQa>h!DN%Xh<_@4nDW8=%K+;4MoHaoqq%hd#k$lX9iwVuK9`v>~3m>%>`T
z!cS^&J>j|xXaUBXU4B`z|MRrZG!pVmWHjeHUpSN$*#CS=Rf1C0$R+7c$FtSnEA{XF
zg{%09e9+Om3ylVQ7bJ&JUjibv1;1CIi^>1F!he3Di<jw+@G3z*u&!bpm7hb>Ll<tJ
ztjK-}5X>7W<5btfSho@Euk*4=1D|SPLL)c76t=K$X+Z^++{;<7v30ty)e&hlAsz1Z
z)U{ixZ>1!EL&(4PssH_%l?g<yL;3ez|J5P>=Z`PJmKOF%V}$wNGUhk6`iEeRfc@2H
zdBx)Y{f`M5fTy)tGpPChc=;hjsm65e`TuqL|Gv{L>R_rFyJob3gcJSWU%@^g;M9)D
zEB_m1^54G{409SRsA`#wum9(B1J9cB%-vgLJF1dKzf@_Oi2j)B?h3`9Z|@?+?c)~<
zG0p-XhJA9-Wy|g5XdQVm#1pY4j=MI+S<^M}zUcXH+6WsN^yOUu*e;QQk1B$uXW>4E
zXa&iQ08^~}{?R*Ak#|Ys7T}y}e*VyjrpgQXM<RoKO{Mpa8Vk;4Xz2|xlM*@nb|j-t
zH|?7(+W%~=FYsFE0fsc;QHtWPV{d>^z^rlxQk_4U-wQBex`}-qm07!7zFBBxU<_|=
zj&i0P6!pGqxH5?@XuTn1c2Zu^hp*M~zfZOzwIG($%$s|4TKpOER+swtfoJ^!p5Ucz
zm)u`0|2UTj4AKS^iZ#;gY|N$6rs;)KT!BCLK~^bnuC{D*_2(x4h%P>MCXh(yjv`Y{
z<n@f^q>&|Sy=y@i*qNvI9EKD|Df4>%EK02)USz1tw{{zE=05#s$%<zntAK-~^iCa#
zr%5d=^*q!Jv!r!X5K!wFc*Wy-USwyutYz0LTY-WFN>KX!i0Ye1E|G+xX*gHEz%+Ss
zpe3K*V`(|B82JdvlU34%!z5$R74<^Tmj+oGN|ZC5I}q~B1a?#<(cX}1F4}{E!~Fq>
z=>ppK0DNP%<5;G<ZKqcdCFCm;-J6?CBamCb`V`+|7)F>*jaAt}ywU*>tw3ciMY6}w
z=sNTXd7V>eOzV$(|4ho;bm<Jv$eZ);-ES0qk5q0W%6#!1$^8P=6L6d(W#-}+8Uhwq
zzOr7n2`*I4l?J0JnD~9J4Of6;8}tLpN(RAe8pdYGa~pts9sxN?(SAqdfqFkK^bBV@
z_)?DqVp;)ixd9ws5Xg85JkmgZvcN`(aF>u0y1K}3QW^_FM0=vGuJhoAfo&AtUG8fX
z<Rim*?+1F?YX7Gs!~u%}C5Mw^@V4qbKd4_`1hi)uNNf=Qsa?~z$iD_FJM=bA)d-lp
zS`HXh3xJiP7P8Fz5fC(jx#kMmT@d+L0zD5WBe2<rU>I4}IyrgE>=Cld2kTIZoV^Ra
zQcK7kLjtVb7`_JUoNyVq<A1bP4Ph{e(b?#;F`yYl!lQKfLcY0lPW{sShu*^ZbZn8@
ztdkK4Mz929x#xs`yp37*ka*K-vPB<s{ToaZ2WGxwBbT=q7Dr}n;-(F6&#w>%L)M_4
zf%h*UBaH`Z7Ud-_Y*<7n?SJ}!yvPmB`mA+_q%S-P%^+tOV@S*L{Lf0nH0iM`rE4(x
z8UZCW82jzzlbZsMyv1JH_3>vKiPX%DlGbQ`iq^3+{RxXwi$Z2*7}Fhq@uio(a`}TF
zT_NRm##6kvPt@G;Fp?C*X6P+)T>f6XkJ?eso758}Q`r?K?%cdpN<;U0(x%TX<AuoW
zz~zTrzjyLr_9k*}rNn)!Y?gk@kvn(9t+1cg!gcK(PrN11S){I8rAmXeJ(PIO%y|hK
zbsfQV_k1;8X57uGLpjrNVTLOA2Ws2>roVy*!5`uOc|ka#0j%HSuSjbx>}#;(uW5+?
zNZE^YSQn+m76OLq6ImpsiYYIxi!{%;XZs_uGq`BYJ2=u>M!r8gZd(*x2=)5Ci*6AG
aAClWSZV)j(d6|O)|7oe|V#`&mL;fGq_pbl|

literal 0
HcmV?d00001

diff --git a/paddle/framework/multigpu.md b/paddle/framework/multigpu.md
new file mode 100644
index 0000000000..61ff1ba204
--- /dev/null
+++ b/paddle/framework/multigpu.md
@@ -0,0 +1,60 @@
+# Design Doc: Multi-GPU support in Operation Graph
+
+## Abstract
+
+This Design Doc refers to the multi-GPU feature in  paddle.  We propose an approach to support multi-GPU both on a single machine and multiple machines. Every device only run sub-graphs which our framework issued. We use `Broadcast`, `Allreduce` operators to join different device sub-graph to the whole graph.
+
+
+
+## Motivation
+
+Paddle supports training with multiple CPUs and GPUs, refer to different physical devices. We need to support multi-GPU training in parallel for acceleration, in detail, there are two aspects. 
+
+- GPU Data Parallelism 
+
+  Suppose to we have `n`GPUs, every GPU has `1/n`part of training data, and store a complete model in GPU memory.  
+
+- GPU Model Parallelism
+
+  every GPU have part of a complete model in GPU memory.
+
+At the beginning of training, the framework needs to issue the same sub-graph to every GPU in Data Parallelism, or different sub-graph in Model Parallelism.
+
+During training, we need the operations of peer to peer copy between different GPUs, aggregating gradients/parameters from GPUs, and broadcasting parameters to GPUs. Every GPU only need to run the sub-graph with correct place information.
+
+Besides, it needs interfaces to synchronize model update with each other, and issue/merge model from different GPU Cards. 
+
+## Implementation
+
+As mentioned above, we summarise that several kinds of operators are needed. Currently, we need to issue parameters to different GPUs,  named it with Broadcast operator.  And also synchronize parameters between GPUs, called it with AllReduce. 
+
+### Graph Converter
+
+To be compatible with parameter server design doc, the graph converter converts the user defined operation graph into sub-graphs to be executed on different devices.
+
+1. The user-defined operator graph will be partitioned into sub-graph. 
+
+2. Control operators between GPUs will be inserted into the graph.
+
+   *Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, Send, Recv in multiple machines*
+
+   <img src="images/multigpu_before_convert.png" width="300"/>
+
+After convert, the graph as shows
+
+<img src="images/multigpu_allreduce.png" width="1000"/>
+
+Operators are added to the sub-graphs. Every GPU assigned a role of `rank0`, `rank1` etc. 
+
+- **Broadcast**. Broadcast operator distribute initialized parameter to all the GPUs from the GPU who owns it. e.g. from`rank0` GPU.
+- **Allreduce**. Allreduce operator synchronizes parameters/gradients between GPUs. AllReduce implemented in the Ring-Based  communicating method, avoid of the bottle neck in a single GPU.
+
+These two operators need the Multi-GPU context support.
+
+Need to notice that Allreduce operator force GPUs synchronized at that point. Every device only need runs sub-graph in a loop style forever, the whole training process in asynchronous or synchronous mode depends on the Allreduce point in the graph.
+
+### Benefits
+
+- can easily move the optimize sub-graph to parameter server,  multi-GPU feature can be  compatible with distributed support design.
+- easily plug-in with NCCL2 library.
+- GPU Model parallelism becomes easier to implement. we only need to replace different GPU's sub-graph with different part of the whole graph.

From dbaaa4979b7cef401a28491906b4954acc9f9b3b Mon Sep 17 00:00:00 2001
From: dongzhihong <dzhwinter@gmail.com>
Date: Mon, 4 Sep 2017 23:33:25 -0700
Subject: [PATCH 02/94] fix typo, rewrite graph

---
 paddle/framework/multigpu.md | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/paddle/framework/multigpu.md b/paddle/framework/multigpu.md
index 61ff1ba204..c8501725f5 100644
--- a/paddle/framework/multigpu.md
+++ b/paddle/framework/multigpu.md
@@ -53,6 +53,10 @@ These two operators need the Multi-GPU context support.
 
 Need to notice that Allreduce operator force GPUs synchronized at that point. Every device only need runs sub-graph in a loop style forever, the whole training process in asynchronous or synchronous mode depends on the Allreduce point in the graph.
 
+For the simplest implement, when each GPU compute the gradient of `W`, followed with a `AllReduce` operator, accumulate the `dW` to full batch of data, then run the optimize process individually and apply the gradient to its `W`.
+
+In fact, in the way of every GPU optimized full batch of data, wasted (n-1) GPU compute resources. We will enhance it in the next stage.
+
 ### Benefits
 
 - can easily move the optimize sub-graph to parameter server,  multi-GPU feature can be  compatible with distributed support design.

From c11718550b959aeeba28d0f7cd35d6a9f2f872ed Mon Sep 17 00:00:00 2001
From: dongzhihong <dzhwinter@gmail.com>
Date: Tue, 5 Sep 2017 09:57:16 -0700
Subject: [PATCH 03/94] rewrite graph

---
 .../images/multigpu_allreduce.graffle         | Bin 5294 -> 5704 bytes
 .../framework/images/multigpu_allreduce.png   | Bin 114947 -> 118413 bytes
 2 files changed, 0 insertions(+), 0 deletions(-)

diff --git a/paddle/framework/images/multigpu_allreduce.graffle b/paddle/framework/images/multigpu_allreduce.graffle
index 2659baf8ada1c7ed09cda326fbb6e56dacb8fdcb..588f79feb9a3f996b6d6208e556d61f0a9602aa5 100644
GIT binary patch
literal 5704
zcmV-O7PsjiiwFP!000030PS6CbJR$({XG1NzI@q;yI8LGW0qmV7-nED11vVMLpb7w
z=$36Qa?7KZdCUa=eY4c<Zav*ExnE9*z>QTURhRN)p3JPQdic+;7g6VPkPO1O|I>Zy
zk^7yX-;H}=|LmvxZ;yZ8Gxz`b=-$KsdGhM;`1jXOJC{*7NIS3JK7R51uycQJfB)d}
zG79$hpBz8wyngZg=(q#-*x!Hp@_y(3d757SxWE79%NNggt-Nl0;TjJ1UnlWpkfh&U
zz|HocgV#%Y_u(<ayBCH7zwL$H^wGT^9)1eGJv!*7;pgCm|1C(K_j|$Dhx_iga53zs
z!C8<zA`kbkKVEl%XNO%sg<HNG-g(@~PZHmK_~GFoP2ej(f)RT0ML#@C;;T#V6@2<7
z@lQ{qKt0?aHOaN6%6g2$@R?Rhv0Rgf``5P~4&f?|4|@N(8l>Y#r^!{Ye|`P2-~Hqs
z+3&&QF3-d6xY_8tkD9p_gX?SKVf^(MdC(!>A9VhrQql(<t@-!wh4H!$hmk+1_C1QD
zaPrVnm*-C=?XM@5<>z<U>?r*f1(TN7Lmpm)TgIOzS96z#*PkT*m&vdu1AF-@KOC6Q
z@l0HQntT{e;UM@B#}_jQ#G&8+><?Zg;aNBlh3mP@rZWE~Cq0hBKZ8p9$&^<*#nIJ8
z{~!v_`jtV^>Y%dLN0)vVZZU431b+00gYab^uA9PtoHRf2)5>@lp?Z&)J!U$L{Yd$b
zoOk|C;7?_A?j~_k?*A-^!hYwy@5WH<@FW~uM*g>>t{+vN3P9~#rk$TU*@v&?u0MxS
zaQyAE(#;20l-p*q_$2OLT?GBKjMwq>vC907WO(3IN)5h?LRkV*4oOr{=4nESRaOZ~
z8N6J8w#O8g+Hz*8<-$0uG4`OtNfBsuNSi?I+`kO!uy#F3>2dHit@iu<=}}GVDE?Ay
z`X(4;GFon(1M*}NrQroE51YOS`!}F9f1~`JidtdR_;ox8)7<NGzt7|t6ecm&5+fKT
z+)6|FSi-M6zlqaK$n$cOe<mr7eZTmIi#n<DJqG7~FF%3#2K9QWNipT0T(9xkxp(;s
zmSm+ghT_$Ue4ixgDLwJ~gE0H2dwE8gWGCIY8~g5(B;wv*C#P{gO+Q8_r{v_cAHvV>
zxt|PzlypEnd=65urZ0o5;2+<iZ8whM1RDS2Bnf&<%9FDs2>R~B$LK16pINK_Jqf;E
z`e3kollI-g8KCo|OPzWm(Cp;WPkJZmSE&g63)stkDC2%GCbA!fxF4KnO}NdP2$3~`
zv6EqM_e0UFiPF%--8W47etMPo(a9fO_lq<<J5PiDsow=HNbcRszU+Xr&Pn%_bPm5a
zSfAnaz6TIj(N}zXIg3^7+@KyD@EVoFA-s_!j$X$g^ZCAG1($ord(>!;fF9FKC?cgX
zpsDtx^fY6}P@;(fb0zurlB|s_T21$9y(fI<T0g;npZ8DWW)XEQYD9D_VWrV!pw1cK
zu`+SCV5~dJFX~y~_vY4SQrxUt8nh{^(WW!KFe(i51ZYB~xB^`$fKRaI*CV*_+$%5=
ztQ1!QnoGq!ZL|WzWd&zO(<!fKqt)Z%;Lk8f3OB!I%Rl5_>_L(g200!B@Qo<AzMLP-
z$ixL_e)n6U<EbAFiVtvbEZLq0`Zlq5ewqeJzqSV>wOx+=Flzi7WOic!x0zeam2DF9
zM%C5Kn=A!C$Z#b&F^mv!ha7LBp5#Ju5KCz_AJQhe^|Xn$Mw`f-w22x;m@+Vx+8CH^
zjkF1A(+1I|$4Hy*E^QKeA=*S1rA^lG0&5}$coQ`q%(l@WzeW?rS<~yXjzv+;wIos~
zXlOGUg}e!Q(=^_sTjWiYS>sWwEFs!}?l2s96O$fgoWn0bZwo@`Zp53kZHqW*GbeF!
zQwd%Xu?iHl1p0);35gTNn_3}GR(leZmIY(Xh}2SY$*h!3hIen+es4;LouWp{EJBUS
zT5gVr5|_lYf)nr-i4|Z=9&`j@9y3-@&ImZ;jEhDl%Q-miL|dYWvP@V#ome4LLZ(y=
zq_Q@dlF-7F8U#=iFyK~#R4PF|E3NyVR7@L?>D^#Tx;3Vx=447nSPLGNAd160GlN74
zi4qc}X5FyOti&$FO4Q<cE}(vfS<u!LG&hmVu@Xl$s0OY=r69lroAX}Nvt--FLXcrB
z&oFqKIV>$b!{J{gl%gP{E*8U$XSvW?0h1O2{FU8EVrkhL!w~Z_49P%=z^gNiN?;ib
zDF#vuOk(XO#aJjOR8}c-<b;4><k!Y<<0uC4G!;ZMP&h;jWgBxs3uKg#a@;Y>AykW$
zLs;&CoPkQC7G#%b7fF@c<Cz;yIQ+}_u27EFmK$q^lmjWp4pEK+q#SdwgM}B`3d%iW
zl(rxzLV+%@jvau5XQZXVf;e)iO%pqi)Oy?keK3k~%c)h49cXY53H(kot1R3G*}?X+
zg9Qw^A=V?*RbN>}HK9fi*Kr6Xjkbg;P8GE~V#r%t0ISK|>o_E_S}@}5I8iQwiG~~v
zImQlf9mh(-8ET<$#==WnLEF_3mRU0<MhK6)LIEX=6r2f$>o_pvu}H{6wz`gkOV1Lw
zN}gE|PT{gFHj8ms4jf7v%4in<kb=53+3iuKdU&e3WvLr2_J?uO50W>2FT5%~BGY!o
zB^$^D*5uMh(VL)m)eXK++s9&_#o{Kj(4?%w*M}X0*g@>H&Z3jp?{)n_+9Gx=S4FEs
zHw;M{k%UOv1}mi9DlSc|A%ZGunA$X(q!3Anq;0TCLOUa5j7J2wL{SZXJ=g1lju0H7
zgV5OqBLqawBUVdpHKCNsY~jQvAxR*T5J_8MgpBkUf$<PSq>-8kx9VBD-vQI(f;y0~
z-N;-j)>h_HQCxzAnM+)PgsYybK&{f~nbL%cF21fQtEAGvh%99cM1@N(ZSz7VT`cd;
z>bq5$JK;j6yPWf*ZSD?8xfO0d0xC2iAX@|^PDt(nEw@Y=Zmlpww$916a|~(P;uuoT
z$uWebS_{oxA-v^8Ajd$Cu}K``{mI{|3uKXJ+@MSJ;<eWE?-HEf0me#?C=QC*Drjfj
z{xOt#l3AgJ3;x4jfehgWhOljg)M9RPf7fam7uyzGu5G)7UEe1B)*-xQ8yg@VXTqag
z)23>O$F0UAYug<z%L1#&0=UBl++jm*d&PvrWEFaQ?MIY*{zaRMUxlpDA+9QPXmy~e
zT`S@4RDHl<m~^9>2+~?GhA3Di3VQF`!~mxi3c^)|f~*b$3^9Nhm>&0Q5dt9Tt|Sgv
zlNH9gYHOEDV4emzNTq~S)Y=`%B(zp%EeU24Fq43pggcx`Sp1lsSq{p{6&<_G0nyZ0
z3vRA9ZqXR9bx#Xwh)`NdLl`bD*rC*c&={(`m7}qQUDGy*yLU9!PRAw}*pUzcI1>QP
zdN%Gr0B#uo+uR?+Ik=UV&-m(sIsz+^(PP{6Y=D(jh67)r%y3o*8ikQiM8Rsa0ftsd
z#yT@%LLvqb11rr2aAG}5IhWLMtEhBjNl5secDwRYBaAy7Za2~aqP5~hK`|YG=>SXz
z+}(75Sh&`Z3*%NGXk)FAOalLKu?g#O125FNQgOJ{xM{4@*UX87aLEA5jnmgudz2{y
z3XV#*6FaVg+@b6M$Kd4RQ*<MQ$%@K_7ePE?c7KZtc@coK0FW=7o?}WnuZ0mUgyd8k
zI>ck%nHLQ4xSMzsZB`Okc<<nv38k1MQgO$pM0TQ$B2rjOh_ZS&Ht*rB71r~43+t8P
zR63kW38*6iPen)n7dP+OO)@|iDl6oRSL#dZF$?;|x?{vZx`z2)PZ$Y2K@vsaV&ihV
zMW_Ulfweo-3x&FCXdkQ4zLjNrIBv6ymTWVY3EcL{SWJoRs4Go8!*hRFW6Wf4)MZ+a
zojM(hu!St=!iRk`%RJ}h2&pyUU@O`5ee6o*{Bl^rFYyp8+z@ZtrjBj;?cl6?+AZ%B
zf*S@L&23-pb^wr?!hc+OGsy3b`0cHo`^_017m9v6`t5jJ=$7mQyjUoKEuL2pLq?-K
zt-!8m&aI%7tH$WT0t^bt5t5^dq#NEMIdVH%aHg5CV7)nLWROZFsRw)sREAVSQfc{)
zkfYXW{Q0~!{=_Q4w<OBIPbRVx=#U*DJHo@%+h9j#-iXn#$6X2-3AGqMBI^kl8Oet1
ziekJ7DbZ@<MS@9B8f_(LR4^7Gz>X5R+6P9R0=5*c(A}U!%(g{|n3<CjG2)hY7{Msv
zRvOBY5g{YOcu^~iXpb5VmX<ol4verg2CVO%l%8hH7)mrzuAsMZr9)l!X}xzUy|&mY
zRkrXR7K<Dcb(dDDB9B@Pk7_ftq`C5}Q9?VhW*~$v?WR30h*d%ma0x6Oo^{E#o<_;m
zXq1>Y?PkHX)7-IDVu<BPqmV{z5{){<<+5u&>yn8jPQ@{`i+1pQOQaQZ$gQO<(h9|>
zM>VtHiBQEUs58ealn26Qr4?3z+p;TXXtUNbR&3skm0^raNhPIPfm7ka$H*&?S75>l
zc}0UtnS)n=G}R_|kXInDm_u%5t1Dr2cF0j>r-qdk`A~~nx}o7)?Y4%DRZAKj21+H3
ztHXdm3?K$pd&rSk@t)5_VlgK|qHRe0+y5K}Nn7Qdza;V1WnB(Biu}t#&?`D7d14P>
zA~gUSVWi13s7PFKSkry5MlY;*^IA`AsxMaWjn(>NMSRuyWOV~7@Y|arXLYwPcFy&!
zC+uYpt@tJ?(O;%{#Y>whL|kyk>a|?mR8%xwsZoQrYu0zADQnPqwNI&a(PnYd8r`(A
zqjr-M)#yf*g`gr4MQ3V0fm1gLT;X#iM_)3rr2}Qn@c7K;T@$83zuDMM146yCTcVXW
z63fkz-<r2ww460#Dz`xEx~$VR2Q1K(&7xVRAeb#_%BWQhR%#}Dho)@%OqnEnVN-^f
zZ2_}<Nz7(Qe9)mxAZG6jW(i_;otPDExQvL{yW}v+TXYzWScA7H(OX1saXD}C_l3R1
z>6l=guO*10>S3{63c4wiQJc#^ENudoe!Ge=_Ocu&oI<Oo0AIJ`$4yHlG`oAq3%Aje
zqv**ki`r$9e`ZP4Ht*OvGq@4A+lX6+xLqf1?J6=Ih+52b%;a%q&gNR;Fm;?2<1pwu
z&f`1YQW$0giuD9y>Z}U}Vh~KX4yI+|FgL=9)6COY&*HE;zKdoCk+DQILR+|?Qmqda
z98rO&Sgs)9mQW$|gU+xlQ@CmaP3+1*Ar@v{t#Bf_!JKIDB^(Uuq<jY@oQa`?|ML26
z=dhFb{ZAcsldw-_`*q}ZgNvXT$ZFa!qgYE-Ds8VVEfc2hoWS4w<SgtrRo-M0RXnRF
z`^aSVcQwtp+bp-RA3<rL;o#`%<Cz0nZUE!G`>-x)I~SR0ws?qB)8#^(f-)Dn)Hk2I
zQtmMIpcV{#ZsYPZt#2jpIob@OA~GvqcB9cfMIw&1OY1gwdI@v4Y#jt8vskaW?IW7I
zEi`vVV0M0;SYEb1!sy`|H{&rxGol&S$gLI4vIR#Jv5Z*Ggh}0=BWlVbwuZ7Ot<Vug
zM|2a*qB4juuGb%{(YP`PafKyPTS+9%5{tIAX~iv^C#Ny9Cy8Wo8y!O|!dM#5%Va2*
zEW%()3~Q&<ym&_Ea5wRcTT!hO&xWix1%qhDB+LTadt^eh3BewNJEo#jx{<Y4LyT-f
zw8t=v_8_o!rw$9g#;^=){fpiVR-YhBHnAp4lMMlgK!6}XHn2QP(FYyH>l_&A>b5kI
zp5}^brWBFJz@=3oMq=$2)@~JRw+^z}EwUM<TT0(*=@yl$u6T=itN=%><hpBdRmm5R
ztIB1M<;czfX8PvMsu~Xq#^~ubMo*==mFTHe88@fv*c(}*zlwAmi?r5l$;=X#Y{|k7
zH9Q58PB&z=)(#EX78)`oF{!^!$SzsEp|u>YpyLSHI|Er<SGQKksy6&ZMC@Ji7WM6U
zi-xbkRn+JzqN})xWmFj?7+=c?-O}J%8iW#7PHiP4R7<SPmbMiqZb}NF+0{c%xQ$*M
zR#PmE*(I`mYRPH}BiXPziq3#p^yGF1wYYWFT2V`HKf{5T#T3U(%V>EfC+mp6DBMpK
zoyJY9Y#L!=O#v9iF#v;D+MQx1L)oyHX)Wu)6qaWoFgCG%NlCLlP)Gy?f?^YEmV|xK
zk!EtD6kQ!A6jmi+RZ_7k>GAMHDY_YDNy4oDL@CwEiBgk(qw=7zmyZRGg?D`Q=+f`P
zbdPU3iswy+GWzcSi4)iS&`H#!ak1#hua+(mk5-G8KGw9WwmO@sRWehn0es7v!m2@X
z_A%c$SI$-V)|E*bOG;i&CNCkI+^et`3|_?DPr*3yS!-u=t>1-%@Z&@YUN`^!=~21;
zv)BGvFqXyfcnS{fP{1#s;UO4;G)OAVX{Gu91;OQUeAM-GqZmi)h5<eQTaZkiJQmZE
zEVwUTo$wOD$XdSiFKS**_LxqM@Vtiqrq%YvIl49}QP>Rzvkl{27<`#}+a{dBXydE2
z8EFUo@WM|gU$cH7S<@~s!1%YY|0EowWmK9HEVH&p0cco14dZ^PIq^uTeUuIJX&>}h
zbu7I&tqi6g_-X9&4KITJS$bZ103*Xc%o@Tt*p1Qyv;K1oF#Qr=hBLJ<f(?fFsR=z*
zu>Izr^@DVt_5(oA(RmQ0CBYw~;<vaLo`RWnqkfWm)7Lj;gz7zF_lWFJ^`n$O65IK^
zo4LxfvK!^FI=cKM?w@@)56@$#hKgNY#%VD7mwA<hU^olYJJtuRq7R@o2+PbjcUtK&
zyrEBn8QMA^H$&#-(c?Hx<BOr$ok&8?E{^VZ8pVEE1oY^$AVu&_l-2MF&*CKfGw%Bl
zj63X?M6UXvZrpRayyiiV57Iw-&wh=5e@{v9@5j-H=g;B;@vq1KdVciu{NVi6TO#Sl
z_rFHnvqSiQ@6}uS;r!#XchLb{KYS{WkJ+z(e0cwo96W#j^x)_ZX!G;e=*!_RHxK_s
zaGEYNJx-t8bpO{t@RLHuujclH?0U`J|5B8BnQ~B?Fk^^O;AYt3p`kSobz<=*_zXrS
zsFBrecrR}p-M!5ha~xmZ#_fwbHG2`S`hmlA0~dG`_`O&C=-YI#zlpybMB!O~1_}aA
zOMGiYktenMqZkV4<Kr-`>ZL=!|JffD$iuM51-K&FFr4o&q+AExH2gdZ<H|ZaqwM#)
z!JK{H_QOA}f{Cwj!?%As2%h%SFu7^k-_?`9s`h>#_IvS{W*vKYl_W3|*-Tc+ZPAa)
zIy@aYmj6GT;vBb{*u&BKf+SDta_1H>SFvofBq;d5F7jYX`=N~9ZgJpwXXpC*o2w{D
z8XLeH`sh!){66gEY0Y>lNzIoB)E!0s$7g|qcZr&CUScvMlldk%D=4q=JnT^^9(48?
z$@Yn%MiGx|Eu^tjDap)(4yA^AlrsvRH>X5eT{BG|;`pM%k|eY4kEYyMtwF_=#^yS`
u^P{WE>#Mr8&4~J8rsw4vaZ>oh>%HBoHk$mri~xR*?)?`7r1L9NkO2S$Mingp

literal 5294
zcmV;f6jAFRiwFP!000030PS6CbK6F;{T%-aUVquDyAm|-N7nI{EZMPh9b0S3_Qs{E
zTQCVpSVIH@LYAFW{`>ZT6iK}J7GG+)%A|xIU;y-)KHWV%^W>l3uA|1+AnAs2`=`g$
zBaa(FyBW8__T^8H-=DqMH;@1M^wE?5IevR|_S?zxMkflpY2)Pmv)8YV8jtr64h}n=
zC^$GcK09ulync0h)_`*y96W#XxbgTZO*=mx9NgaCdcKqLn(?(0>>ivXaVJR9@2}xx
z`_RB^rLD(sncmrR-GRTh!e;vP(GO3)1mB+?Hq-EH@Y??#B(K`7;M<b}_g6R=w$tD;
zNS=}>2X|lZn!vTgrk}zoKlIK#Xyhk}@4o!-q?;yi%TJ+)UVPmSFO&GD<GqD%KPUdh
zMHHwf2YrzdX{xNpICP(Br4$<xd2(=f>RuOa()h6TpPOzvxOAG_1P6D=pZU!%?vm{m
zT&{B!HV0z;d+&=miSFI8K{x*Vlk7FfUwe%|sg!iDp*8>OuUvm!gQLjrmfN1jQ8>J4
zp~<V`q5R!LWuwoxCw7{CkAk7(U6*?Y;gs>0;n9)9z2nD;e>?2fuw!rDj&=tg=-^4*
zeH(u1J;HAADUPqFP>3VH{nhWjO~T7?s0w$_GJBMxZ*th;H2fnd$qyfSsZku=T(=LS
z@UmU%6fJiulRoYEO*q9sJ_-EjDF@}tzTAnzFHW0}{j}5{MyTE=W}le`V?R><Bj=63
z6Zk3h&YdJqitS$pQP^&L^j#nF4UWTZC-T3aHvOn{RRC(ElQw>8WM94&o4yF6;Ou*+
z)XWE16lF7AJdT?;*Fifi;&t%&Sn2uob-3$>lnQ+3m9hY&5hPK;rDq8fiZE-85qz(q
ztaRYAfOIMiWt1CXtrYOyrOG3%3Ti{FUHP4$3T!2`4ewMcn*&@G;!2q6Uiq!~b`(@Y
zfojy%!EOY1HFz$c`zTJ@LGsRTg*W+2WIAjHx|~1z=vUX~Eclkz-(mjf@co@9=><La
z+ubnxui3ezOtSN4+>Cv9ND}er&-078ou;3o^9yo*(GKCWdF3bFASDe@|F1y`*5XZY
z6a3>ZkZs0MoPh8@&Xb_Uq&&Y&f}ribe2#7c_{^mK_dNL4@xdUqhVsqsC1~GylREej
z5IgVqN$WiQCKZAIDP`G*GVX&hk$o8AJ~+=rxXna_$V8y;q}SSgD4K~V4I=KmVbb=~
zo5YXKe{Z@!q~YaN8niF`CZL==dX(L4m$Sxs^MW)Eqv&1Gx@iV~O&VoqfDWT{-C!{Y
z76i+(!SXDL{Z`ZOrb`5i<tlGxGUyP`P+~N6(`dptTO*P*A_<YS4Q5AEwqejw2pxpZ
zHW(qHoz*eMBZ6C^sFs9sy*}s&!4WzLoo%o>z}!4ywd7V4N~zRx!;mBpNr<Ga@Pu3h
zox}ZDfQSMNIXcHFct=ujqX1BnQst1u;0PznL(YpZEvQRh;<x9g71GftHts~<1>Gzv
zJS8=(LgnkQJqDDcCmKCd!BX~wPvUNvjs#2x@FAWMFjh%LjnbNehU5X|#NWkf7FHgO
z)Q(=06#8C3-bN2D^PC%pmQ@ZdM|DF=@9};F5SP)He|xoqPtEY)fLA{_z#DnhzKHAJ
zxRa=O<AWDgP^=6iBN?OtHeE7f0L9{czYP4=%!7jz<DptJI8~AFlNkOq`s@ZNl@T46
zRC|(WAU;G{Cal(b4NX;%3>6;8AcH2j@U-UANJ=bMLO6oNJ#DlCa%BZ)M$@8t4EwS9
zqmO%uhe?v_z~KBZ!YB&v4v$`_Zykb5zxh4a@WPL}S?>6UQ4>-awrg^YLsQ|WX^^xl
zTQE}D<jfDF+S?%0kE3v!Syj6187}1a8tE|t;}kIjv`!P@D2H}`(^hjq9k0*GJEL1q
zIoQ%D2c3~}7|EnoQgXv609rGo97s7fDK3DNV_B4A%=kQZaH8{44&kEyf@qIe@EaMZ
z3YT2kI*u_9LqU%5usDXKc8MHAY2X;t>7|h1_bcfp5^>^5ZHS<TTCKPhJHj!@vN(pc
zvxYScHOzo*u+ms=6jBVN7@HI@-7t!A1aBmXqp3{nKIP0af>D4LGe(2AwAa|z)T7|6
zbIQ2^mt8PZ>(^H_pVYc!bGJDRBlJ9+NZR#;VXTY7P|ZP13dK0HSr}%UInlG=k1$Ab
zC%>C^kGyF^V>1+3k!J;`0!9o_R{}~bB=;mxCZ-IxR+wI5m6L7fP0}olH|be<lZA0u
z0NM=~mJ@-o86-}dM4UdJ|IJd2&&)s@^t`lTp8U%w%d^rWio?Kc706lea|NZIWL9Vx
zs)>bv0wXJ97{Zo?N$RtU7!6%00V~z~5VqPg%xWvC5|qd+I0_&nR0~e16ogAmz}_E(
z*NR7uc)TAx%DM5#C2>v+Axse~7<WdIFpn84C}+fK&bUB4?j|1DvfbgLF0hg=fIDo!
z9X8~?i(4elD@Jc``I-2Zf4$5~WFbp*h^rDES{-OA%Uax-$`?2alV(&AL0SvO5CyA5
zLF?l(F~Dhwf^b!$AgjXwLku7WCdd7j2m#P^&lm-)$qEC;Kv9=UV4emzNTq~S)Y=`%
zB+#X05(s7zFq43pga@2SAoGSjrkX-96gMt&Kr}Vhf}5+2n+OMN-P3{^B9vCr5QdX7
zcPMpGKO+W1l?zQ85_3-)vR3EfOWe3nC4i6$q8VV4Xrb%_nl!Z2u}Ps2fcFDHGcN#*
zB+g6G76_J50605K?hXOCWdM{*Ob$2)xAgKEUtLf~;3{$Gv2A)bz{)bifm<jtoYjFw
zVI&k$u-a^Zp=FY>&WxClhyldFO0xl+SdUW9B{keCDjiu868@$wCyE+jtkFA?4iHOQ
z^g&Qe2Vgn?(*X}R9U$gjqk#+KW*}%|t&mIt|8TJh>v01w)Vd{h;E=3~Y`F27Idu>&
z89=#l<2BVDWy-(+M}f~<inAbhC_BJ0I640#8Vg~vqGeX*K|EsiV2c+(K>$tzKt6YR
zjw$K97DlvNz?Eu4dw3jm<^@AM?j|0^GAjuzymxTTgi=frskq})PCpFrYALKGL|MHX
z3lj0A71r}v3+t8PR63kW38*6i*I7pY7Z*L;O)@~|Dl6pkSL#dZF$?31g`wJjcJ=bT
zo-h)4f+UK-!P@C|^H2#S18aAv=L&UK&^}h8eJjiMaNH3KE!n0|6A+wxR%Qi*oNGlC
zfJRB}SxdDQ6ozeh4R$y7A>ndq{8-G$k15y~4liB_w^tu`#t`{2@?+eGWC<&FG9^6e
z*61Xj;Uhn+HekEoZ!)RHPOZyH*j%(-%pJqC%yS_YA+;tPh&r2Tdla{pi0rK&B5^fN
zT#a+mYMjdq8Zyv5?dBE>!3_fk(QR|+mLHLt!Y^)Z1<>Ce2^#XHks>xDDbnDWSuGfG
z8>ImAv`C7O6k*VCDWu41XFZkm1Ux~hB{)9}{JdI5)RW(z&|sGc(HN64cD!WsCpATn
z(I`(Vuq&E#@USUYwV}&77!;BtBu6DpH_KeA+%2cWnP$R*_2w`lyPaz#^?)zIkRg?j
zR9e0x<Y;LNjPO|(7$H^xz9mrxKADgTLq^DsaFdrMu_H5U#HiQeAq9+tnvWlm^#qKJ
zWIc98F<yj}XtnVo!K5dRb{owrFczS|juN@TE9N$xCziqyx*L>;*=128W@e;BjJUaF
zZgVZdtu&M)BSJ=m@uDSdduKFQTIw7-Fv8Lpu)h0JdYUm~DA7c@)dXs%SyeTk)Ox4V
zYx7N!XJ$Ry-tGPrdE`;sw+o);%CpAJkpUHEpoA{%radl*RYDMO2`pv1F&Bv~Z7vd<
zbuN;&;M!^K*eWr^a->m6qi`<LPR`XIGdzzSoWufKa+%6WJBtw)oRI&(XoaOqq!o&}
z{kNF~PlPH?VK{TVLV2KUR$5^dxGk3LNO)#x87nqx#>z0prKFNlt-z^p;bY_#$SW{m
zg}kC>NST3GfHqYocaT>gub4q^<x*$D=xm3z(lRkCt1zS%*ONrUx7syT87tRQI}8-+
zsaJ;qffzsxtagVrvEl{iiNs2Bgv7ET@o)cY7$h^}qDHZQPU4$RRSr6e{7yG$<sFkz
zVh>;<6#(jEq|P%aX<WWrlYOvSFD!rZN>6N}FIMf1Rr+Ijd{y~mRUOLl+i{Vzs?+Bi
zkF>2O>_rbP|0YV&U#5A<OPeZ0oO8#jN4cs;Q5xwALN%jZ{diY+WHn=6<y9(Ov}v5Q
zS~soesEu=?YTc-!5|lI|?@Z06aB7^w<-V74^aT@J*v;DX_U^4;vSkwV>y7OsAXGcM
z1zI_lSZ<E~R=n-J<*evaQ39>&qE6Qgus~Bb&B!tV!E8ZOMy;x`Qd8MGG-cao$|T`)
zn=-`g5-`gb#B7Gdy#{3hG5cUJOAxc`#H?6`%ZP}5NDiaCM~BgfHF%2>y+!mE7xNZ>
zo7-EQj0wj1T7oF58W!6%K-Xn5Dsvf#rA@%nuQw6KUKZno6KM4W;Opi1ag!1W_3j?>
z!fo{AD0*^>qIQwwpIH#K^((ec4Q|BkHsY2cZr6!hyNXN)q84)<Q+b@3v$>WyOciIv
zI1KuZv-pnp6o%=8Vm*PFD(ixQ7zES3gK3dC%vd;ant2-QnH*NdchSrsG8U*tSQaj*
zRO>?pM^qpx7Ar`&CsYW%*XWgH3Ri8QiCq~e#N5nlDV#`dFemEXhTU$JmhYg3GttxV
zpHJR5jv9&I{?cIMlzsTLPa?k=TnG6;R^5Ub`C6(%X?ta9nJ`tGl>X``mtniE@+Q-$
z{9ZlTS0>AUs}SREvE1Bx1ci=<-P4=Tmkw+r9T=?LhjmHYImt}3`9qwVE*9bxl)2EQ
z9>4EOvBAWRS}^drjnmJxzL&sfe=&%X%B+0ajYjtri8$6St=rt`1<c)|br6(HW4&g!
zk7(|;(A*h;+4*&1dC~d^qkD7QjK>hoh-RE4w^lUEB{-spWyEr(OzQp|QBzd0HIzkZ
zg^nmXqMKM2l|h7Yz5ZB@#+f;YD=d-PN+M|%n6$0UQryB>avD>6l1L`E(J{m#jD_*M
zNQQF3A`GTPuXak!if42VcN5RJ6xBNMY{-gJFo<SM!pyO}N2WBJ5bQCyVk$bN8(DkR
z!^kE?dkn*94+3j<>agH8dSzJapY&$1`UFw3i8Wc8YzROE0t5lFf#q3>?llyza$uyZ
z+tNgOnk%ZAQbZa9hgO9ciM3l;yOpoqI?QUf$YzvoDSfY{TU4s5;w|Q}92~Kd>!HO}
zC7(a8Di__BBijd<>G6wI)ovDy(bH{=o=S5s(Nn83Zcf*+HnK#273nw@X|3CmnFTD_
zf`uKbw+kYjY{)9D9U8JNG-OI*Qh%M0U9fsXYuTGY#}Tp*2C_J>Zmp12%kUQwu@A{x
z)c5Bt8omZsQKPGfuHq(^QDu-|d@Vb4OM`Q15K34%wUvxeEif}%+NC&g6H*BEt{!s2
zZS>->nqpzhE|B$83szGY$$HgMbPCL(C$}@G#igs(iduUA84koOrZ}dWM$0ojSx5Xu
z;d-j*G;U&LQy&v+3cx6i0T{&6?i4c_%6i32YnczGusj2Sv5ECdN}Bb7LLw*-6q{JH
zB<x;8n&FO8baj|eSe1lTN%^XzXT2Sz=w_593A6egrPNY(lp40{mj{Kd(OBR>c?U;N
zJAM-$_u!=cc;2up{d@mUoH+3#r%^-Ue9@C%E?puXEf+0)u8=EBT}~yHOeNKYyR0az
z>L!<;M}<eqxpH@1dPoCJ8P${FOUNGXP1p*$ujA&IU=aDNw6i19AHr_<c{m8(iU0Qe
zv?%}b#J>y%x;W@h&Y|ro_%#R~fgwnPq$EyD;{O)}owN9~>5uec5UuM4^!)EZGQ9FY
zO$)l<ZoK@!3j`yReB)nNyqfGXjSA&CfnU?IeEu2TnUpAO2Hokp@gWRuC!V$rXVBaD
zCap)>VLQC`)8TDacO(;S0t1YH4co_IH!Y&l6kwUjo(3?&`e_)q3*y8h1^FoJ=JPg;
zW97cI;<VJ6cHpP6%Qw6Z+L!57=>m-Oewa0caj+YO3uf(S=wNyqcfzUU^I(H6eyKxG
z8En7$m+c^(MZOE@IlT&kw4nF{RQwvZ!V56duGhySZ~E?}j8MH#>^_kVqJE_MM`jv-
zcTcW#t?Wc2SnVG^j@y@?uEML>4MX`RZ{jqV{$}1LAsEiw;~k6#tTY~gG$_l|CwHUL
zGk8N^x>LwHAlF0Y&FQl^P2=mH*&S-ah+XWT?;?u*G!N+hcR}*t9S&B#E4+-8@Q=9d
zN6_!ET~N95g_?28jpY>=dUlxp(R%qy^xH>Df`31YKD~MwABulH``4?}=U0bUZ{HJ1
zKY#osYF-|}@2$7*>8GpDFF!<waQx`GJUe5*{Ql|V8*=#S<MYGQ-$CZZx9Ilh=keWt
z9bBY~JRUcmj644c5d0*U@vAs}H#=T&_CFP6UN$%=O_(vnC~z}ue$&v3i#oM<7kmXH
z6IAHxHoTX2j_%&)i#dxs_i_6Cn3}zax9z}Tx`qq93;fpGcJzHR*x$vshf#Rho`Qma
zXn}9_De|zEzaK+^@%SuE%j43K-~Q@%bL64d;u>6$tQ*dE=uxi2W*UB-hH+($ol*AN
z&0xm1@7v+;H^I=?7<2dUyTS8z8Ybf&`-f`sS024ThV52-TR+Ag-6RP-iR?+1$!*?`
z${IZHJC^_7d&DE$YG@Dp^9zzuS~rY4xP=U|wSy>gkvPq0KJ3tcwfSC;dOCN<-`zw(
zQY!$j=d(ZQ@W-$<%4r7AkyP9{K<+g1KfeqdwhO$3^8$_O+sk*sWp3meroDtGnc+lQ
zVy)0H2<<iYDJR);5kh+IpjE~wsmxx363#uV7^l`MX@u4l&+JniUzhljWY+y?%8k|T
zI)#q9Mj!m>rt~7q4sJcRewgZdBSM_yet5OtTOOB&pPdMR`1H|#0RRexa|LMu02u=-
AUH||9

diff --git a/paddle/framework/images/multigpu_allreduce.png b/paddle/framework/images/multigpu_allreduce.png
index 7741bb37f47e10f13f92c732cbbe9d31f7762be4..52d3eacd55834880d09f4efbee804b5c567a57fd 100644
GIT binary patch
delta 87925
zcmZU*cRZEh|37{X2gitWY%-6Xk-f85LPWBXJ)+EHcFg10GkcFh#W6A>*(2F9GO|M|
zJ4C*h-tW)n_xnD+|HL1+bME`PU)S^X9IvaeP@=e8BHVQ^^sc$iuZs(|nqQpeA1m*q
zCYLxB_A!1jpyqo%W^=R9sg(N7<d{lPGA(Da(d3=_kCa>`aMQh$8;a|{b9;YSuHW$b
zD&{yO<MsZctZ6t$=G!~VZ=W)~D&HEu)GK-5{dj3_{EfNPZ@wl3j6~V$_x#h|=(ynT
zU7X7gRu#<aU7e@UmCTRoAI~%HCi2T~O;xcf-bxqx;XPG)EB(i8gWQz3sLP}PS&VzW
z#e4rJZ_M77MRBG!OIHb+*MBPKz97N#;b4=6H1|$X|J)vjQi~Dz7aDTmRsR0<`z!i7
zQt2-0dQd59(f@)nwVm&GJVg7MI4YQ&RYCr3(kGWFNLEx-DjJR#s>FzA;>G0Ag>;VC
zGeltVpvxvN9W(~2VuCJC27a0PX(NQGLyQO@0tNV5_Aut)KjWj&0!&x@@XXyJLie99
zFz24%2O`)KS@Gh@A(_bK03RgmrRtSM@d)x@#RgMv$XS!(D0JuBnYTA&<JGf2KN^y`
zZr#C>w>3qOquYYi1|z0v%iCvgmR*0QaM0RR3;y?LL1upITL=UU231!2_ag}D?hUEL
zyYoOLLUjM~cWg9?H33a9$F*xENgeru*vO9NblN&~r55yC=B!nOL+NifPt0v;n*Z-3
z{`*N@NQ84zlP#U#+Vc5Tv|#JlJ4M*`{@=sih6&35=VE0xJ(!RC`^DuFl-5^-4(A7a
za_Khdk74en-uFN8t*8et7`{ixylBKRR_~kBpWF<T?PCA4KdgBF*VmRpox{I_&Bz2I
z)zztDx~5V}XpeU%=1Sn>Xc|N66}Rt`Z@y2Vt9bMT8RD1kFWp_vd0SQTUjOEb)c1ny
zyBT%t`-|!3`GI@HEL6)YDKe4vcb|&vW~e-7&K64UsUomV|IZbQ7iYD#CH!BY=E5Q2
z5#q264N9o^6{-^kBp&z$WmRhW-{=1S|ENK+<vE#JC#j^k`!Ew}#Zb!bbkwKZ_Mu0>
z@l3(`7#N$B=>J=HZ~-+sBpk)&@t=#>*x;aKg8%!26^50)_+Isavidmv!LyUim-*L>
zUvGTlE9dy1tLWo`r)e2(1gf<BUw^682S4~Z->eozNyHksvk>s>{eV#OcH>I?pP$XY
z3^bXa9E@syI9bnTf66^bjQ-CTmZ?LSA!8SAAI^OSvm}1K|DE?`NOw!OM4dY>ZRFbj
zeBr;lLrCDE|DJqH6EtuA^HZ({?xp(_eSSzcWA*A@)k6KF*|LE2BZB{wR)@NhH{@iA
zZ>;FdJ>}*9*4w|;+Ey#D?TvRv$Nm5xgN?TITcg>I9edn=`@d^Bh(s_eJ;Uc@ZTWxy
zRzMB%H9Uz!^XI$om6cX-1B3VehYrJeard$!$VhMfTfyo0;0?&$RqG_$wa9FFuK()^
ztZ4866g(O&>L%tv=j;qheT`+7qYXiBvtN66yd3jVFZbQJo}6j-8{vOTtt5&7x3aJz
zFjmx3m_kGlqQ-%H__^Y$TT8Xy{`!@xlCST+GG6=me8!U1Z#8M-{-58!@|~Q*{r>wP
z#fUal>yB_@aMAnaX#T#Z^*%q$Tw@p|9vW2`@u82Eo^{=y8tUU63OD)p7nE_}&Pka5
z+7KvQPap*lB8uGj{A#LpSm6(Vo|iSl4*m%@jvALRSMQLnnf=c`9-#^Wcfv-$Torh-
z*YNe(pObI-`PXRtf4rR<iE+PodpY9Tf9@ybHM{^nL)88w&bm|b@f>=(HRunhlf;&^
ziPf{vj;qYU0^6n@0|L~uMLIeqYT)yuH5P~DehFB%_&$7Z(;c1U>f0SnvwBreFx}<9
zkEpC91*5%gIalXw_}WJ1)~9?$^g@eY#>W@Y(QL<U_!K$vel?F%NpJkOwwn-2L`tW%
z9P!f2)#zRpRU(}K;g-<-chW^D7G<K~E9KZ|&%$(5p?dYq46P_o+thJww-wLB8Ad^U
zt{ct!eapR-%3E#MpA5i`_8d3yzgV`cw3wjyH$&W+(tlkyTk^@qNQhXNlk_q@(Wc?$
zN`)jMm572xMqiuOj=&1=rFXl~RAO@A?Q(N3aG?bEdNhBM;Dxq*Kpda{{>r15)2DFN
z8Ci3}ljmy#Af@XX%4Rg%t+k)>tm@02>i;f1J{0es;@M7Tu|}Ni;e_F+yp+?3RV1^|
zNSJt`%73n6yA4w#O5}0>%R=Rxq?l8Z#+K*{YyErr9UaNy@AN4{1XkDjEiVA9vwSFR
zVIi01r|MpoX7?qPsXnn?ePYOeSB{Y#=QN6<9`aiM#4d*a;MGdJoF2z=IQ4&DZOdh$
zTH*)-w3cZJI&VEB#UoY=elOvm2CB0&b|X$|p{=9#+4*UiT@n>?!~f6G4-@pW=rZ0A
zDpeEOfB5VO6UJ~oRxj0lUiogTLTpHSW#IaUE@5Leby<qc5$=E6h!r*h6@NU}=sHuA
zE#<bJBN}kz>Pd}lDr4OUokY$S6<}`$DYNmxU1#fml)ubVVZG&!3YC-J>*EOc+2qK$
zFX1wI_6vQHZOp4k07d`z*6ERb;+W@#YO$T{_d(GIvmr?E%Aq3C*^9flDa(L}u^RDY
zQYhHXDjFXNLRL{bpYJ_b@|N`3p02q2ikx1!;PL#snqj@4u85(3;3G@~P5s0xqvbwX
zV}C2w+9Opm(x=eqzb``$uL!>{{?<V$N)R2*V!Nq~5GD#CAf?3(b0<K2{v4+>2OW&A
zj|ioLVhFwgU)+eEw^3Q0a3}e2L}tm9b1Z_nk}i~4w6x6khWPSc6KSBMP`e0q(?YG7
zRAoVCV!TsSaMjH5?h5h7&bNH@g7>(to$Rb_!i@v2d%RR&rDu`S$BSW6mVW4I>&Tg*
zjf-RR56{k(iI`Jidc})K+i4_+h80=B>E_E<U|IMWyPdSNX*t;NjAMwTNG8>Y5Cj#E
zv;TfJqFnnzpeqCm;Y96p(>?J1KKTR75uYLBwVueXYW&9RN#_+QY#DEqp|US6%v)u#
zi|TeD0C(>qUKO)N>Sr9=ZRoKVBAEBvH3Qd-%9Ku5xo)NH&;TIM6Acj0EK1Ykx8}bF
zV}iOEy>XeW1hB+_@P^#*Kad91-d`F<zF3@~@bGJEki+(b%DcYP*E*xGao9Ruet|o@
z@aY7|Plt31&wfw?r<bJZ{<{MJN>zq0<Ii15Bf+v+;o)(jEpud_R1l|Plvc!s<;M$@
z!w`(x8dUV1lrsmbDYq}T$cd{*rP61E7di~xa|6_O$r`;KW${o}5<xJUk?XP$0yeq(
z$M+1Qc(Q7glj31eLB#Rl)C17mZ{`dMD|c<W0_<)6)!Iy(N{sxsjm}u{;COxBf7^*z
z*{IRSA4KIY(11SPJipZ@EjL5@oF9~nEQhk=;nwFQME9$uL_Wr?X>Hphjzzf1O|Qp_
zuP#m!HXwL$HQZ1UyJ!byiE6=*Br((^c(f>-&yO#a_nVi!*K>3fz#4&e7+u!akwlvY
zgYF~S)&qAI+rQ;LWjQ)q_ex2@Z|rPK@q4C<KHD4Mt?cKT1Eu_Q!SAI}`5h9i6ut_G
z-KGixM<2qio8|hYZgWJH!v3@DTB_mpN~$5%GH$6eva>vBE}|o8y$2_%uF7EV%*hUK
z#G3^0-weR9-t)%>B+q|~ZxrfFAtDPI;;wps^hILILRBettn^mE)IIlmFPw}M)zN0D
zG_?36Tnv%WFlwo&4`+L%GlM2EF*N*`@tKmFDL%80r0muDsqYfL_2z_6N4EhWX}tLR
zXQe6FD8jdghTmAtw?rV;gSu=PPp499nE>&eSVGQN9d+$vv{(;LI`_uuMzLe=6WRyW
zUPCfGE^ntv?W21XU@F1NaBT4B0(5&c)@T%}275j1u&-|zRoBN;|K+ZiO*a>3Gwwpk
z;!B*N4(yHPd&Z!{l?26SM>CEIOfoeNgIQa|NqiM7UYiC890gR54m3bJX0K$>)fznj
zwyoDhjdCS>r)e&8u)&aPUhDd;2V=L`Y5{^i@4;o-9bZr4+hXM}=Sz_ZW<a0q_Wp-|
z?ypEq8HN#FF>CVVs%G)Ji!09&T)D}UtM-XGJ0m)j&N3Qw{p#Uo-`_M;g7pM%99ha6
z9e-BGXE;HjeMi#~hO?f*S>xwleZ4<1()8rJaqM)`5lUO@SZYXd(e6oC>U*#=|FRkw
zrpyUdYi)gYr}MvFi5A%g;2$gETI(=q#Mp@$((}GjBeYpBtu)4y9~kPEq@^g<d*Ap_
zM~4?m>kYZydb|v(#dQpdIM!GQ{e-;z8OkPPM1|l;iB;kv&M$hsbxB^uGT;7LrItwy
z<WM(*Y3tbRbzAl3g$If>vg43Rhf^y*tQ6#rrfsLGen<l<wOyMn`RHSWW)MMxUGz1G
z={qbnKwNbaH{Lk6g%$>x(LcXFRL!28M6y1U-f6=p?iGGG!JC2&QPl_J6d*D3Dyr_g
zZBJ~w3x(n9X?1lrc0=*_O|~EsJYp5Alf5;b9Ctj2`Yy4txJNo;!g1_obIvbwF%emh
zS2vX?@LHR^zFE4r6THlSCYy3Klo|F@MwBT<KJ+U7&$-MLvrr~<{p@3oRh~@7izl6F
ztO7#MuiZ@!6fx~h36e~+zrrY9X8B&$GZR#!PNrz7MjWFlfA3vhLvf+i(g(_I{xC!r
zoBi?8tlLbZS0Vcc&=G-H`U*CwY7w=dWs8X_$qI_tg|v0MZS0MC!>#M@4_;NZzWcop
zV0(}qPfE0T4}nI&#x5HyBb&oHHFL1(%36w<dNqCE9gfeTD7Nb5P?UpmkO2Nos!rpp
z>`0A2<1cjiEn6Dv1IzP^o@`AhKSwgN1(I@=8hGz$4}>Zu!4YeWrEAmGwn^Rket^=I
z-6_>hx?NtS|He$B2JZrZSlP$X6bfe9IzVN%`bW?enz7a=zDciZL9e6l+<LJzLc-6k
zfzV>a+klugJk}i&pr0}(YODCU0fE0O;Yq62$*v2gAC}*bCATE6R<eHZ(#G8GEW@T%
zS%)=LDdjq0%+qC##dD{htOuP0Ep5oa_S{_#Y8HIRSfOepM}mB^od`U30kAEw5@wDD
z=%mhf%Y3JM1J^ZWFYKc+QpKfQ%51`T$9PuyGLPq^HI2IV?HFF^7pg2`iqHj~Yaja#
za-RVR#Du@HA4uz3!L~a7@E7i&H<p_SXH#Yq0Yp^kJSruf<Jy=hQCmi%Z18BYCXWNw
zQ?%ldQfv<OKwK~!`^Z44t?u>Ev(!Meb<-Wp;pX_ur@t1qKi3tV161dT*%`mBIA3PK
z@diGgNZ<2#^0ld*Z08#8(QN&OxVz#+q8>$Z%geIfFuyq!qmo(I;V@FaOHx~d3MLX7
zoq2X~{`9wyhxkC6;P<~1-$I$F1vG_j<b-nTWVCIf3>feA(jS37G3QeC{s1k7c|vqc
z=u*ZofEDj^2&2g2lZ9+ML3&4_0v3&loMLi6Go<%PeZByGw&Yp^2-5e;=J;TGzy&jE
zWH^}1)s#xquZ<ed(*<nw#9du^?ggaStDntff3;{;pVJDOS#hgY9KtGV)!7d|{82nV
z7zb~mLm}Ye{I_HN4K&zV_kCu%hr{k@{v$H{$VGh*(7B57P5>_I?{sj;8Qstos7`!x
zu!@KZ$9}%b^Lq955`E|1E~DbJv!-u(e~woY@;!POjP<mvwZUhQr~!dSfLC_xF$(1+
zlt21O$wLyy>|qZJ1Rxtet!5OyW8RrT?qm_=PW%XMN2l=pb?vjWT}}Hw`q5E<KVROu
zM2^n@7Ot<k!X0NPDAP;E(W8Z>1`OvAz75__!<;AS%w283)`9p`ZR2~HQ<O)hcw_l(
zG-pf#a}kP(ACZfTHLqe+UOMX_z0SriyZ=dyK_J#TN*z^%Ql!$@g5&k8wT(h?d30$0
zql-9tol)ra``-)nv|Nr}d5j2r=3qXkNb-qUMBiJj7!{ad@e6kgRPhdw%u&*qplj8V
z5~!aSTI>qQ^fF6i1e;6;;o-$k&fjPiGo5^e+@1VzwQj>-fX4W5mA~YSw)Er)cNYRQ
zmjRC8YELTMFJ0l@4Hm7~xTDmCo&>A>7l4L?^{<&)9>Ndr^l72Yz-np&y)Z#$onGKW
zmXurT<9XTFZL#~=@eq3FJ7A43NqI1xEaMy%b1-2Tf`=F-7)^D~d&-^eUS;_nPS{mB
z5%nS}sx=R(gbdvV7#AF@3N=kUxk-M^D=bo^o4iiOB2S;+{tSJ7ifqn`?Qm_RkXc1=
zVChi*n65qs(*D=&{oy2d4WRm|9L4|MoqRG9|IT|ew2)WGnSLs<a<{ViC&`yk<RP=|
zbaA|FG=$cM%%v-lI4LyP7>*@ZdbQ*pg=7vURZqUIW@b0Lq4;OJ_WaN9r}*9NEH)fB
zO4ZdM-TLUS%#hgF%9cIt*uuHwa0{z4p!=uNh3;pKyy+Xb)PF!$8QD>WhT-@wg#tjh
zow!b?2<gXqx|BsUCt9&I?n+((tio1p0Q9O?pWf2>pe2S;cL*=jinZ2dH7<OB&`$v3
zcf$7PeDMaPW%uU9{V2~zv+gN14M1?C8<XGB{`zya|B^jlD`pN}kQkY3aDIAt|E2vE
zV)ArO$9!E*Qe(Rkb?ICrhGR~sMB<;ea8(JutpPMWJjZo8J(T8eyxdQxudU~EO<uaJ
z)UnpRaaT1*9%n3PJng2rhsAUrh|ru8gft?A&4_|Gd0BdUnV(u%Y(&gncSNUTiaB~p
zvdhCAsq}c4*APM--1Xe`>fG^Jbc2nefD5?5pj;M7E8p(62KrSIEjMc_%%304Cj5<Z
zjl@gybq2<z#4Z_vT1Ca5f=L=-i!vJQdU(DZbKRYV+Q2}czO)r+UCt{ZK#@<D8H9Dg
zF{1Hh8$fU_Ioa`^C!z&=aO?58`a*CgY;ZBY??z#y&fG<_EaXba742kL7d+a;b1SB<
zucMm#hip{O`lrm5>8;{}@wA(g=c_U8Pkx3FvCxa!!96DGxFo1Mj9;6KzR<n8s%GJC
zWm4y4kixC2#&wmXehdrlVmb~IBB&A1bTd%OYi3ots}R0lhBj(8<Cq2;b<S(}bLIVP
zPBK_!5B*U;Prmn^yY$K*5SMS=AH^mcJl(7A4XLIGS%+;~&7ClXiVR9G2II#TTg(`-
zDSw)LW8NFfpqh}6{wov0^ZW^OT`}vk;}8#dz$s4Bt%#)N-0`7#VUk7Yb70g08nX|a
zJq{anbTi;u^^XEZl!-QP*B4Tw(5c<+Bh}+^uIpo%D_4)_c*p%}ysI93cm|kz<Jw1Y
zy=)3jd&OTZ``$yrX&Ua8`Jj&q8ou~+<M&+3t5D;$?M8S*_&9^i8$kCXF~BBZX1nsi
zBeClU)3$#pRa+LFrkh`Hv4n+^2M;Qap-GD8bKN)DNI6BlEx6<}KL0Q_Qz9GU29+>H
zzu8m)ZghAkS;!-1Ig8%lFs7K#n%hJCiR_4}5iEkKlS$U^_DEwMu+D%cBKAltQqlBl
z#&#lR<LE_-0OQZ_zwf_U`Y}HDqu+D#e;XT10zFIJT~aYa-X$WL?UqMZSTuQ-X=q7S
z;bz<JSLZ7A?;ecbe&9W-PTy$^tR#1;z5jTJO~^Kx-+<{znHj;nj&J7pt!LVa(=sdl
zEKHLl<N)p@A%&&!n8^JmfXVmybS7o}X=|w71LunvuY$)<x!fQ0o2URydOqsMVh*27
z8r|82LQvO8!1GVgUAhEq8*Q|iosq;E!+AI6>mFy)EDN^N)6M|i@)ozMvJ}3TUlxkG
z7X4Je<f(F-0!&~WDi5?!e$Y|odS^KKB4C;+!Sq=aDUVm9>a-e=d{~xccw$IiL_vY^
zasYk%8IT+bKph`z@z#C&@b{EdSHPceyJQ4?b;SUf2nj~7O{zzH*8bzKB}YR@X0s$*
zD8GJTguw6D0aXAPf=f5D2M0PMD<!G}UgU?u?|jrS&;4xNEG`Mb0iO;5AY&8{AjYWZ
z(SrxeQ4YfylymRQ8y;^fJe?agVqg6G=U0OD1dut$r{AltN(p~OUz~5CM~in~cILrL
zMTDNn{isXvuz<J!6O>L7ko8P~H+F<0(G_@69v$~D(aQ3@Kc}!D?rjHaFFox`F{AUw
zTef8UjPV`p&uK5$5Fg6_1QPDMLhZ8@LFX@Ad7K98(kE+tKA(Bb80RTWVizs24Frw&
z4NYQ<0WTB$y%Cy?@fz!;QmtxQXB3DUp&16x-5yHN1*I`a`pBNxdik-&ux+?&l8o%P
z94|dUka27MEv2v3Vy&P5vn4`<V*ql{i#sKn+q6d%<13lRrbLo&YblX2YqXOycGdUA
zdz?`$nBTr}NDV#F{YKQuYk`g`k8*tmq$7GPllf@I$p)Cp?bkn?Q3mj;sZVQc>frwC
z3z@>@LG$K}s^w^W!o4RQc(+V{sp3^7Dq3xQUh7sDz2qI=QV#&hJ4$W-%u2r&-rtx`
zIG}Y*aUOr+A-QDKF%uq4&yA^xqeeN)^Wxc<Q&WrRey6g5ywZ~^p(!&bMxS~C?ffwM
z5dC+5{^{3{qa;lh8PC-|0FZt^zPz1pe%gQ^_(+7V1FQp^Gu+DYOOu!5T)-oV&zI%r
z*0h!u&n#ELWGaHW8&!XBX$d>SO{(yf?sPV>#kmlG90l?Bd7D>fBt>smRGU|cA;KUD
zZ}rVDl@fJHvR$g%X!+e<R<uRHkGGP-`xF%P-#RBgKIeb3;xtoZ4^pE`Ayv(jCHM5_
z*~xy?p4Io88d5tEEZ~8)sPV$9ImFU++uw*qQUrAZ;$#`P+x^LWDXqXgPu|ZH;FZYy
zjiqoZeK{(d6<~XiH_$oj{^NKQz&Bik>Mb;=5up>My!Q%NVO>;%6xwD};9E&Cg5jm&
zQee4TiCBEC@)d%F3)PrX{O$$y18|KJ4QDGjR<lx)^WGQ%edzWLhOy#*Ndd9)BA_xX
zU~>a0=Op!^8Tda~B*){%CBiz6?ijj+P}5jz%{a!0s7dh1U*EdiYiUvML?a1?#$H_7
zu~AAY-jJ1^I8`Ej;98-ed?le&h_KrJdk=!Rpduv7ck&K3lw4z4c10G2+TYBa-L15k
zftZCQ_Wkq&qG&@UOrHALJn(X9n`rbL%d7(qHbNBA=IgQ19Ykhkm*i2IjR3m3_jJxf
zGm4Vy&d2Q*PC>>|bPORm6TNIeW)Cj;wU~Eesq$qpMS+6G34kT|6;B>~d-~Uhi#$K6
z)Z~d^`W$UeHVxh~ZH^`cNL_^NKJoH29(F}h0RzlZtnZ2n*NZ!PMg>6Z5&8+$y}Okr
z;JnZx^Klw%x|XwHR>Ct$2KNX|syOFAE>i}VM*}{twKel7@=HZ)+zrsI)LMGlWvD4d
zFkIi&-#+}@HxyOHKokWuOgjHPq12vI<Ujtgr}{4hNjpv5fZQ2Pbdi8#>$I*6B^~U2
zz4)8HI2rWk=&eMnX!gA9lP#-yM&dvjMlQSw;2pj)e)FZ#W5w|Hjpw``t4?z0&5{&P
zqjZ~pp9%{RMjo%FKftcrdboer!3jU$%9BE8HsC#P?7=0-?_udX2}AX+UFHWRfZ-J&
zeAS8c;r@HBA>ixFlprgCbFGV$nNGkuh5ULDdr!<Q!cs*CMOj_#LNq-K-VMkbR=R3H
z=n%e;B+h%e254Y(l7=*c&?N`hIAE&-Z=|LLpr+cjTq!rpXLF?totF-a4Oup8@?I(^
zl<<Ec493VC;-|_(fr$v#is?|f^F&V1-Q|;|&$Kns=OF&=gaVrIO6(U_>)qvU(4f16
zu?Hztt*6)ehZrSYdG#DgdQIQ`3lGtYXm}tHOiu73_;RsKYj`8~50ND8YPZAN74)5_
zz1k0fU8`gPo|cVbRh=l2$Z1cD8ayp8iP7D<s9n;fYy5XBN{VVt{wNuRe;%GFH&lHY
z8YJnU27|i+;tbHzU6w=YN_m9pFD=G@RS3!}46;n=Do>$sY^9hokt}2kIzx)UYrrw8
z4GM6!p68Tpp~2*!Nx7KJKieOa$`D|FcDxjhr@j(0ZVE`}5s2S#U($U*Nnd8Ip(tVd
zkJM!HQN45r$}U?}KN*`Vsy_Vvl46n~pknHnd$(BQ%CaAw0>rIlx99miOVfJSnZazS
z>lTgfN82-HUO$f3`I*F>l`o+oqzuV+ZV%`;CC!{H)fh;0tWtVXxa@p^lhxq%b)n)e
z{}0}nWZ<Y0#xcoIaul4&&HynEq>grctxIJSX0e>}+OwT@q6E30jb~>6o6mKqL&S!G
zu)9}fd1KBY&gG5caGnXpV7}tlQYJ%=z+noxZQ+#Lsodx+34Z5wVc;0I98Q`8jM)K9
zmXI;^$@j^^*DP{A+n46fuLYnwv&A>?+PR+pN3A3PVySi7>ba(Uc>nw43J@l;uG7ME
zyJ^=S{GBv+^Uy7-TOZ86ZdSK3ZJY2p&aCjC)(yFQt!>JL%xwUW637>~YllgEv`3(c
zFP$dJ3{w84GNOTFbH~*#^oE1_4}kN}KtwsLc+qnEu_aI`AjCKWu*$uN{IgAG>ATdj
zKa{dI&EtRPKOk!lXJgs~{X;a&+%YuPxY(rH1|%S!vDW=<!Jo9$)H9WCUdB*P|C*<;
zRd7A^HI7f8p=ZFN(Nl`uTN}7I7)RgE(dA^r)EFej?_Gl59!3a5^c94!z|YD$l)|)1
zY$VT%8LI${l#p{aUa+7B@r88<jA^j3jh2|G4v<o@d-i<)2$m88O9xGs1g3A4Z{Tin
zq3W+nk!@f>hWprUj+KJMsEj<E_{u--0aF3!Qts~IiB|raKyC$Wq(#@z-cHMaAJ5H{
zD+!>Fkce5-qe$*t9E`9*FLF`wq8+mjXd_R@({(}Z*3&Kixr09MBS7g32Tk5Sq=}+e
zGfJi5=lf6n6Tvuf5n?3HJvgu<w{)fEy$1vvHiR`)!v6J<kV|PxN+%0JIWwU__DRI#
z4mnbL{X;88&1iL*gYJ}rZ47nZyK!N73|=O?aEH)5eRMH<s(dPFvJYw<bRX`M(h4;G
zH+`7Q0JQ|7uwp=+mc7$Rk}UoOK?Ur1L8k}h<_$j?9CHfUACYKv5c&Xz7Z`ml0E+<C
zQYR#&+(0vVeG>5wntjy3u5y`rQYW(C-$;P?$w*uO1Xz~%@dJ1l4R%{2=!xW;g&U9@
z4?1Qg(QhAuEMj=F7;Lv-4L!|wIoVbtl;^*c8@>c2@&ZukmEHuN0{pzTGV6TbWnuyO
zL3)G2OHmIC>EZ|`;2r8X%U0qP#xRT6pkaUEpDXTm($glHe)$~)Xqu*6XoP^?2B3<D
zc<hvEhDf%lUQi5MsaAz;AbHM7{P%)D1u`pD2Gegqd>@TMbDm6^W|5^OB02E1&ZdBt
zS&6@>od=<)K_7c0(;m=TU0Qc0E4%AK#`1-xrd9mm#73&2-e`OcS{nWig!(sW!$gpJ
zj$m*`jcUL;EQ1aZaa-|{|D@`BYQb%^IsD|Z8wK+DQu`wE$Ub9bLJTU-`s8OiuK`|R
za%L>aXKPe&vcTtni0@#CAaW+8jtd?h0-CIrgZp<Bz725cfGUfG*4J4}0vAlU%d6z#
z42)%<6@LN0hp~op40VNwE4Sq7*P!2-j(MB24Q{7#YJZ=tJ-=7kxEu|X-V$hA+}RMm
zv4>NYxmujz&i|m1Ov7rd<Yva;<#r|9M}&I=X%1B4H6}Z|W9<}WhVepxaJ~&8W!STi
zAl`oSRHDh{1|FS-dpjWKKWh5XPk>6C&-Y}9JB0#z--2%ge2tW-3Vena8xS?Kl1|s%
zqITlPLEyGF`5qsjZcGFK*#6#^!~wj&m|j!h_)EG@2PZmxOyw)|LB;(G_SoQMN^Axp
zf_;xM_-|dvngMFr+3{wcLxL_r7TB<$`#rmMi#Y~4jRrm`u+O)SIrrRvl=hjir;9nE
znbf&-*oyFlJr;V@Yk^ysMB6Q)dUkY6hscH@QF7rI{^V!#W}im@#-|;JbLG;HC;5$E
zJB`1%@@M<c#R&&@*%KyG)e|<Q7cDUtCo9(9Q1^be1e&EgUUDQOnE&sex*FChN8w*8
z&^~e2WCNZ?O`i6=0l?bc7sI><@vtQjl~P{=4dsr)+1|q0?B<?K-QP?>@3!8XZ93qF
zQ(sCUH6r~k>ZI8VjW%lLHu8+hvuw|KwDJxxyR;ZpR(OZ;O{H(_c2PY5oYhW4zy_@Y
zbD0|)c}=;Qa**Oozu+}|188jY5ZGne#zKlzSjaej!md%zXe;Gk`L}A71R+Xrw({QC
zq+#7^n`luL+${I%($?o!->0hFcBRNyvVe@)t9pZs0nzjKjxY*^KI#sZdJwWG+(tI7
z^DqYJ#GWk3ED3wvP4WDW<h}n+{KZ26&&v;jGVJrHGN02k0kQrPtzhJ=a#O|l$;i(y
zZ@<6L&1UgiMd~IdV9C58Il8n$7FuZ!hq%G?ixCU62tV#B*C*`ZwG{cbn-d&?lf-_1
zL5n~Hr0;xtnde`@y<J;DrP*&f)?KfaqTH{i0UE>pko>`jDz+c%5CjAp3UA}<2lR_F
zcpvTu(1yROWzZkV^HaHrq4E-UhHc{yO_L!AO9H4Z0w;}A7qXqk>wV#|(&Jma5`W)H
zZm)3@?k>Bg-~Qo+80x?r114{PvdH=I&OWedyM&;tt=4+6E#xFB+9^f#c(FaK8liJV
zWBrJWu3s6ln^-I<2gv6klBRU&`dE>;lKAOO_wVIs!<RM+#YkObN~p$ruWv(|Odhc0
zfcoxBwf1((vU&D#^V(Lu@9y&Pi|R+HTAfPSJ9!8gd?PW-7VlLNkj-4-Ppu6R!tjtN
zDxSwPHQj(7(svduNdKrC$Ek6AyP;9#m>=YWT?wTu1=Ij+x*+T$ZNnb+U>SVIx5jVK
zW(=B52v;c(t;1zo!9UR%Zw?Qqpje))7hV*s;lQS6F888D(Y6a5DbxETYeIGr`Ajg4
zzeJcTrvTXjZKk+L(9-2cw)r^%vP13?uI5qo#?Cn7PH?bqQ3SatX7DmVibg277fP|*
z3S*idE!#gwtqSWo!pm=nqnQZqVBZVg^y!UfHosirb-2}?^BHi@AVfcl5fjkkA7WJ&
z0w&${*ObQ`*$dSdXUw2woW-Vd?5gdz<yY4(e4uA;3x?vu2OTnBxh6qH<a_3a1<7i5
zR6u^2faBjae6VE_1#xc8gp`G10}3d@%JZC}ZA3OYTliAZi@$G8JbzXB?jv>ResM~|
z21xrg+=<RG(~e%@?kH)(hmST!3cC<o!muP9{te#ef~p08`U9dz*lTb!ZeiY&i{pUv
zDh1ys9ZeP4UoH8_`qKDY#z&X#$)$Y5t&6|P1;Y%_Yp-2U`m|rrMlMbQEu04&bp4J-
z8iP@pZ>~9F{s&Hy=LQt|KnDVoQ+ndsUEm7UTqiG!p}YR<(yG1w;L@G$VEYR?+w{_2
zuf7c=Vj~2N^z`rm?FFG0W_4#;1IYIF8jc4^CY$=u0CB>YMw&}DewAI`B3|LtX|FAf
z|K3oFN8VL@5V5Bu<S~n`_NQ=+L36yvZUPmoHhUiswQ~)xc7*f*!QToCPDe_{c0Sf+
zFm@@M1v#Sl8n}NYWg(9n?KqSZO~dcd^hh1~4H2$(Tj9iR{@s%#s$5`KMc}%L!*j<|
z`3y;%)$$~mq4J~PmKBXHX)P>KJma`c9pZBlR9m4eu|efb`KKHiuXaT=Hqs>Sel3B}
zG*|8BYHkO7V&*x@+oaYZu6c!hnx})qjthviWn31cg21EgZS+VntrJ-`=H^H~X!7XA
z5F!M<{f&&J<==n{4Q7boH_3#lBf<f>^J269esfz9gDIQ?6au(>w;g#I{53@zxAH%L
zTuWpm_PSi+<%nHTjL`B-0k!tI65PsoXF?IJ_4DCGSx8&_Z{EbMm_{BfL6XS(oeokp
z;YIr{bluVABLSZa-{6jxeVTTZXD+GI!du@*F-4Qm#CqTT^+6!92O@%%*UB%$dlGMm
zt0|HsQD{7Y9Kty>m#h-v%N+Px>hT;_nZ%7%1^w-oE`pHOz3ZbB!8AT-=b_!<(l~6g
zsil1D5Mf-BX$j@;pf7dK@!XJcs}e=@#`JNA=!Tvdq_ycxhSgrvbJewS`g(O4C4P2-
z^#0`}0|+gb#0q1B)@cr;MKzOX`ND1}X-XG}up?f;e0vfrwrh7OI-FR;K1)9O(za<G
z$!skB<8^FH0KHw|ieQU`bGrph)@NH9gv2jsy*^n8g?s8EaOP%R^_00VBJ^D~^=)-{
zP3KJ-8g+dU{yS!$P`D_oi;o}Rv2aw6pM;mW>L0a-Rx?v}fnmw*L&UZw+fhum(_k24
z)tc;T-T0?d?7qHc?Z>7DX%=B@&SYD@hwDfIRNpqO1LpJnxXG$BdY!5XKhq>huHZ*k
z9>H&SX)c=deA<%u+b~}jEp%k~KKb4Te^m=5QIGcXYbSw|#f-~}cn%^|IE327EjF!X
zCi~VX7Qxl0F`3+BMHVC7C_zd<)t36qDr?Ex*nPoVi3tv*VZB;0S}D4V1idYEh4{pX
z8qq`GPMzDL&iHP#YikI{4MD8EBgvVC)xjN8V<I4K=LehPGw26}n9e*#Wr}8jE+6Gt
z&_lj{_5UuY$q=0OGn&8B3ax(dynyw0!TFthAq(eGw6E%?Q`MNz3?|7>s9D9gSc4Hb
z`^(=KSfs?LUlG#&iRx@W=IV)Fqg}DSx>c8h)SN@9L@i>mxNo`bgF}+w#8wo<E2=i<
z<J|~1u}GnzEp*bel@0rm0_8F$Fz=&a(*-|DwO+byD6Um2u1DuIHkUTDUDLakxSdAt
z-Cls}?D0-|&#2wu%f~~-%V#tVS0;>Jll#>|^^HRq4Mh3L5JvrKw9DAlrRN$CoD{A~
zx?-WXOyYQ&m?afA@RC~D4^{nWN3XUYu$2ovv|cCVMpKr%*-$UmTJFZ<aB8ECn~y3?
zKQpHuo3Tr;%SREjhX&L<wid?Pu<>H#JHk{kWQEAABIT^XQ@&L)9mwB;AKwl0@8PeZ
zMCru2%5sugdR4BHN@opys{A&X9f=pr-sX__T2m{eHeO{tERzd|8G%+wrFkD5sI_yv
zF?W#QL-Qe^l5xSG9z&-XTtkpoX{HPQNRJf(dn2Y;Nb*BYu?VY|o4QAP#ZRHl^?r-(
zG}?A%1N;n_0a>8MBh)bWf?S7N-N%exO-;gO^C6zmx7P(P&6Xkr?zUMtOHsh@pG<6~
zaY$J;sVuOb-fB<1KgKf^yR39w2~2Vv%<4v`NBod*bPFX{i!>=nL*9SuG>XSBxGFu(
zP#r~f;9KyqOk>n`pTJQZ>5e#(7+JX^Lj$cF?2FnVF$H1GLqD~t@CTQ{W9Iv5EOOww
ztVD@cjJ$?^TpiXDIEEr5nGJQHq0%Lk{n%rPmL}{I+`?eVheXQBeBc34A|l_SbK86(
z#JPsw!P$a3sngjx*7&z*8{`j2DOK4^Zvd);B+}#&(|%4H#lz?X`_tKylACEe-@AY;
zXLk|57>lNw3qc#al0{bFVC1XZAGd=26NM2ZuAf4ynOh|PxQeg#ZJVx}JfEKOG=neb
z1oOBbzBYTi1rQ(aNw{K@SQh$h!15!5{*j45v1=EDu1|mL3)}tv0fu0<a^QarWn&qV
zsKnz}oW5jZB{2r7XWAQ$1#UK};kRac*xF=Wb>~n={2T24N82~$8Bxg<9s;Kzejy@f
z!uDQyXs!@LNKkKeW63;roM8a|f<dd*ePGOPxyZaHjQEDwH#7=^Ko^)nt$`Y`2IFfU
zlB=a{WGuyDU;=}0V=7DN*l3CFa!i*jW=C7UpKekI3>f<#FGWT@>E0~5G&p_<J15pQ
zP+?TLTr>{4D@Df|)lI06zE35rsDevPmWsVZPZz@|egdmeuVU){Q)Xt7+-VDBN<>@(
z5`PKN|56wo9ZVm=(H4nGxd0l(6lLOCm@-AXsp_52;z0gYdJ?Ad3im3R_lX|`>aR`T
zLuJZYIG#;ezk6iEG4s<o<v!)uKRf5c*~b<XH5FbJ%hWmqm&1{p57tbk<d7BmHSC6M
z`{PK$>Lu>V3=YFZsOCs;HS1M8O-M*VQkEJ*2~DCEEFvy{@$mOoMR-KSZ4OeA`mYzQ
zw$s;X*(RAM*?M7&`GI)2@}-cz_`baz>cb6k5nAg`dVLM5b|z(61Q&uN-otL&H<9G5
z`r&@XPH*xcY3EYN&S7e?g}oAkuEkbO55wVx*<%_k#BCW^;xd6YC73$+83Wx521fcW
zbdA;lgW5Z*mW*(tMI4|xehDGPc=fGuj1nEMr~c>)Bx?E1F~WS$e<ae%paK@?=DJK8
zsVNWvjC%n!Ys_K}v46|2N`J03?R6thLFxD=#`0nYn&oKQK;tTGSk!O-+0pE|@pLnH
zUU+a%V%4?NT0>bUXQWC0E1TyXy{T#R43v>!V}eR4!Brc1eNPlM&Ytw3(G7L0R4c~w
zii`Ic>K;*B4`F-t2E_A9h#`P{2+7P_Px>YIA#&OoAzMaqOo7y&rk|)yC5etLwS){*
zjrGu7!;iJ7V5n%mdh$Kb?YmWK?E?5NWGHJ6v5xy9{1GQKH5JWcs9iPN>Ou5LWGV3-
ziaI5ebcEy{n1S!K7f(!)dF~fZ<d>u26iw*GfNvYzxggFX!9B+oE>}yd=iwROy$Q)?
zTcGVy%5-AC-UcUFV=?Di;I}gY(c9(d`lzp0tnv%JzjO)>ecABIw#;Z>Pp58TVh>>3
z6qs=1aV5iHRL}YEalg2ME`Kz-F~<04%+ulN9ZD2li)hZ*#8T0>iSm^1X_jszRR|Xd
zj`in|7!`{q;Ij6m<3^~jtou``mRMwoGC~!qX}bhbR!r49!UEAArA02}Wr<J`e?4M_
zN!YrQ3`Nq$_e9Kw22YaO4rb+#`cV$Ob7ze+AiyyDHrEtRg^eDx4X4IMWfW`I-w;8f
z`e9l9fl9)+Bt#wVU)2B?GEa30l<|P-ZH#Up&UVRbn7O_<S5_ADKfNPS&5DWL@F#v{
zZx0pO<IZ?O^HO+>d+}*ccz<(l3#%Gk%u}%#{H>#wt2e)d!M{ioqS#R_MzJlbggs7|
z{uECxqTlN7)0gNXt(Y48fj!M=5{0tH^vhuzKO&l2k8&^9-kH`~`wPh}d6`e<WH(yD
z$o+-3zoLm!OX%REYb()JBJ#?-iL!Rn8Em`Vd{K&=YHEJLdR;dxyh1z9x-QtG52<Ts
z^>Sqk^EL66MvRYsimU%QD9RY;^zMy^HOsHc1*-WyMfZ`#ot7ec{`0b((|R_G5((E#
zW01o4%?`kj31zWSKF5t@0&EB;eBANN&P;a|NvRMIt0-0N?pUN`I^~|M?OK=SB=O@P
zX}wRBEv^Y<vk_k*@iFm8D)jr%2hX@yHWEGYu*?odLpRyj_h$$M>9_^8BwD{%E9x<n
zCYmT1zHo8$IjrOCaTGb5;+$xH$1&ZZn5eyR?5E~yx27qmAlaN5o<)3=KA$**Ijm^9
z9p3bPB)Mi))IM$PO*uIY<}_U<5Qw7tU7xal*s9fB&FY>`5ZdTV(yUppxotoHyGUms
zpz)-jLIoH7Rn)v^qv#RAGRsvVO*cY6SR%Uj2YgoO_3xR6CHwi8xAXNU3QMZHb|`wq
z!*`|6Yn)%$`;)0v7!5^0P839rRqCRx;yH5omxk^tI>}ra6NnopU>?0cLR->$2&Wem
z>?n>FV}Fs9B``4L8bYrXD%=+n^cN8m>69w4=)ur=sum1;m&1jnr(Ee5PBA)jaTi8=
z7j?}0e}9unLunbRk1-fGOvOxwzafP_o^FZQ<BqQDv#pQh2aGS6XyT_Jhc|@$t+Y6T
zn}+qX5m&<i_LTGmfBTiFMoMQEvctfLS&o@CxSpEbf_t_gPU6LlFVd^ToVNsVd@OX%
z20mm-n<&FJbpl5bf?1b$1sv}>k9W}EFph?1dal6zTg2S+U=ZWA81BB^y%L(t?Pnv|
zOAu@hHoSKlv6~X~Xv3L+sWM70;v1fR;#49U$xh&TZ-PXY8Ebt!-EH$O0+lPwV34AA
zNiukSRA!_!b#Zl7WKgeCbr5Q`7i*iRc*4Wvu||@q<;+bzIHq+#PU6tD)?4@IE2E-<
zuZ8Z6y+f{#sJ1*Bq7lfOvha!vJf0KD5_2SXBo-}b;(B)<ae4}-vpD??c-_FZk?)aG
zb4Hp|!vc%&ca&zqEMzC%s4&)}u&_RPpD)Yk(FZveDy4Qo3lnLh9?Hkv6i<P$>y!lX
zK<kd-OmVzc%vifMn(x<WF_92Dqn~$jew5i`BI-ko1%v1q-}o=7LaMx=BhzJA?QoAP
zkAT>^JHy&p!Z^cUr|?i46A%74@$^Df`A#5TiL}@_!-Gk>n3h6csJdAB9)7;JA7w*&
zHOc8ZUADL8OYbXCjh3h;`lLMsQrO5$4JSW&<BC*4xhlRN8WH8@_RxA_%Ac{*-+i!>
zIybVu`A(##iZq8unR$ay(~-RI%^|_i5}fzO2O_0+k%p2dZxEWa$Nx&hzqT08^*Ct*
zBCjNQiwMHO^a6ysJ(|KkziKVYgv{!&g1caf!Sk+`gN(={VuzxpC7PKTwi+cDi6q){
z7tBg&6#6QXP*g-?w$KG0OqQ!!h7bzb8FoU7gk?#4*o4&XJg`M9P?V`ukR*kRYdlol
zDaZS*q<pi}j!cfj*rfge?A16`X@O38c|rq$CV6bFbO0~y7g@3~zAb8T!o<RpU7bO!
z)^%>}7%Y}(MxzYIo=k@Su+?AoYUjHDsO7U0do(8b_2;H*8;0@K^a~OCF>Sm{v>IIn
ztKV7OMhq{g!m1WZ(=D~n6!sLU_qat5nEcg8lW9)$OVq*9iD<!yo=r$Wrh4+~6w_T2
zDM&xTE#9@jAY4-Ht~V129lNB;w4vi0(&|BM7U~U)CF2A1`IChoo#o9XJhC4ps`y7A
z`J$=d3b|5Q><VxOY$p57hC$lz-N2_B5MgnwG)caG49kR7hwH0%-dwj`zJ2M`DzK8h
z{{pN++^)kz>#wl8g$-_|VLHB^=T`ACd9`h9lMx*F(Cs{S!es9ND$;LyX7F$b>#bVP
zqQYMt)Df<YA=9zdq5;?8@p`(B;UcbR<1=LGE`f;p282}WMC8pI(^_sJXHyw%751p%
zk&T}0Iw9VnvPSr!mR7jRKiq7yY^>ArhL|o%7=SnKD$pOdvMp4zS9+IyC?Qey*_rQ~
zvginU`Iu2Ij)u6>-}16tYNDnc`$tflId!vx2R0b5Jk7?oc7+5H>;{qtZBO>t!#YCz
zdbr-vOh$X{2?yJ6kO9>Kmr%tl=5%JJ)^VQ^<ajJ8sGN7UB6>GhlGbQAI9AE+^mLoK
z6z{__wy;O=y-2n>!Ps*=8@4ES580Ne>f$65yqxnKZPIp?A)1n>c5m`T?W0XsBC^6i
zB_mjcW;EzK*KIO8*^AIc;K0jjMS-JCfT2!_>2M0eK~@*tYE%FS84j74DJrnmYDJW;
z)%(w=Os|2wfvzLs^;R99Y_Pc1AP^R@Aa=v=7jPCDKOQ^cp~YVfWuwgoxft8_<td|m
z$hZMPP{cIBSDUfpMwNUe!I-)YG$!Ua#^!o4Th^0;>O!<Ar!xI&cgFn~r>&9)%G)ZF
zvHdALujB;?A_mG6EC>ydro;N**(E<#CGqB=o+*QIhe8s%m%Lx3-s^Jvffy{)c_97z
zR2<zuu;104xRO#bYMWSxrJgsb2j>IEx<o=<u@tvNU7|%?E0dQ0++s7&#pnE1l;f!*
zIxO=I%@oe<zmB(G(6VC=B32$k{iPpWV|XkgaNdBU?f&dO8_HBd&@I`FL0k_awl>B$
z`0n)VyX0ne@r31po9MXLu)zk8nq*w?{f5WP-=y~nB0d*<^Z|p|K`U`m>((ks-p)u#
z(bdY1Gqp7*f*+)NY?KkVM~FyYCumcNd7?}D&IX=@yeapaixSi&Sp^n^^M>UB8%TU9
z6<e!qdn=$_sg}M}rsz?1{&>MPR@Q$Qe!+dpbLXK(WU%rWYL7GVP*Rq+=rO~xbz!0<
zldTB~M{iB`y`7J|i9ls()OVs`L=ZWH19OL17TojrcXB~r(A(9$`b!wDU36?j709`g
zZxR*%k$(XOD&;W(Zq7*rvcdA438qQn_lXvT)jOWVPDYJpKGz&1^}ux>B{KcU8fTF8
z>oe64VY^B3-d^I0=QH+S<6w3q)F<7^8Tm@`EaekF@58O{s;id=5cAJX^dHVFEqw^)
zLiSLb&=k&+eNnMEyc{h-=ddIcn!SDaX#SmqkZAOFgK8wXUF)qzF2`%yI<Co+DHwqx
zCG~P6drMUoHPPmYmZ(ga*3DRW21>AloWO;e$=ar?qOFo$!7VJy$ks%`?(}M>zg2e{
z`Rz*JP0ts{uQ%ZCv=4_GmH3`sy+QlegWCcr;F3Mtr^tal0f#TqslC~}V<?Km8F^tT
ztQJI@<+G?8dp#911S%ph0u^q2O_^gXZEa6p$Hyhz>MkG=@ddYG*LR#&7*Fy_`S6F;
z0LFP5jL*=QxlHo^UZiqeBbm@udyd?*j-mG+e1cWx)}B}bF}kLj`Nm-KG-)UY&6>?u
zWgS#%wLq`HM+fwKitx{N+;i@oSJdh!(E;<X?p+sUyUZ90hpAQ=?X+1v``BZ1ky$2K
zyWEIHt*>e;gd)T+G4;a)PhGeww<G&WtT=<VNpO-9l0jD-bEuRp$Sm|Z%dvmA;OqNC
z(5?|Zhrp3Usg{{L?GoewaXO4gh(1}PGs0}Y4sAt82|TB4;GKf?-8SQtg-EesHf15)
zgq~hU7neRi3%PRB`twuI*G9q(w<q{8kBEJ0Lh8A;db@8Yz$Q97c=eB&jSo4sV=~sJ
zhYDnN@n0!C84+abjp{&RwjEZbJ5@=XMKtnGaoZ`B$(y+9$H1DS`a;cq4@|Wr%IRc@
zqx*RhjE!wYZGXlWmnwJA3B<Bx?BAw#D|%k+N@S>^6@y5pn*U1PW)1U+UX93=f*$IC
zhjh5YF$1O+hnwU1o+1mtdAB%=y~5t>Q&*rmn3JN-s;a%Ht8TP~F<89KZ@S*I)Xqty
z>g(z(fbS9&c#$vbNY<tSdpaayS_CbCp@p?`icr(up)C`y(VP0zae0z}adR5XEzmlJ
zB>{>vHV5oMQG`xr3U?)?oh-2HuiRoAR0vq^ic+Ih8k+xrMw7i=UVfGM>;#LthAQJM
zs>7WXkrwt|)G|fFNr9z<x=5HdN^y?8$>Vi<pE@8jLsht(SA>+}yDs|Ewu-j5suJ26
z*(32fT;oqcDY<NHTfAnU{1knfU2YFU=L^p)zZ))-uL~+Ze9}1JThede#KL3KE&a<C
zj3V@iE6h~<ZEw-|4~R%&za6%P;J^RcCBCO-mP)9#H7tMGsD*jiG`36}-)q>!hua@)
zjGF#pyr%`r2g9<AnrI&um?Ai>s#ZMP8Wg}X{P4IBm@~M0I~NXI24;5N5EUltj4GV|
z;;O2tvH&MptO(}vP=IcAZ~rQzlAio-7jR+`U!k3wpGdOO*;Y|P!BMGn<N}_zL(hO3
zt=M@n(Eyo2P>^mlr?J>7gXc5=V{U>24RAT(=Q49I&{sPhJX6g=N1z6y;IvJunbYYL
z5L{?M>{uU^uY`_If;=_d%%)u;lB9?VK`r_{JL9=>60LniHy_!xf3!S<djU+uUS{I1
zaU&rd^hK+nqh(0b{&mLa4h>*1oWtGR{i9f&^zV<&{FY19S-}}EC=pZJ@9S6KVb9Q%
z*(4Iov8$T<;4GYZofDYX0r72YPrLP3;yMgJ-<N?iNYrn>mQ45a?l_dC=)~{%PL9w=
z^;28T<gYa2Yhd3bN4>-ZV3eQZ*T;(Pf%EI&Y(>MT=_@`aEZ>Ine|-{{i?xmkbYVfC
zd4!>^(e}q1LgV21o9$slws_n7NOWz5$UQ%(-<3hsEXCX7V-7*>AgcHIl+H;+3d*J6
z<?TSm8_h3s`~fFrt9GN+{satT8{^5nYc4+zK}hP9&kcE?MJW^nf*ql37P|gq+QOPn
zqGtC_CJLp`bD;$E#MFh+AE~%?%FNzgOB0x;F<%rRodAxBB1b40-9vO@LJ?h2=H9v0
zRc7@S|N4Gc&Oh804mbMuZ_^00X%$F~qF3%1^25Z_X>&oosxe^c6(^z7o+Z(PYns*(
z><#q&IoDLq<zw;t=X|zP(6$md0N_fbS#BW@0u+-gC21yJ=UcC;P}dm69LU^?SYN1_
zt%lrTbH3!viQ>Do=)8&}VrF{yBoWM-J08oL<<VL_jO(*z+DD=GkD=_QV@2Q;(qy?|
zFNM(QR0wi<SCz*vaJp1v%HSRfzg`2<0#4RZ`r!6N=g$`lyqNJOR{C9?=E~%Qxh&;+
zUeO+ZZO4mk(_>-l<8WXvsIA)o2D|1%XI<C%i72RHiw%03{fjkO_D8cxm!ml_OY+Ho
z;6O!%(`rJe6$gaMw6~?`CeL4vKd(M{%4rLux#?(xoG&K_lbC?+#V=ACLxE$JY|e9n
zDye(y0iff@7Pn}H5(fu>b0oB;gCis+W(cY;`pfV8|AJ)@zP?AbEqd3i4nJ;hwm=b%
zxrvq^IMe?Nk^<zS!%|1Ys+cIY*P_HGvEz64A|Da6w|=GUu?hKA$=HmT@WaiKLVU&n
z3kh9;4`9|RUC^A6Tzc&XI4i<klc+6bb6v+spR1JC%r7nzLXrl!gH47lG;F)jBk~N4
z`hc_Y)=;?)6(QK=$Q|de%l{;UKMnm#uW=3Dbva`KPT!v%eh)yeRJNWzRAHYB+#g(`
zEV$_sKD~0Q_n8ruKZ*;#lwA2y6gR^8x`)Kj<#{=2g_aDu=R)^og+x_;rN34J({naK
z`t{P`L!nQ@0;_KR#;B3Ci`V~Hcy4-4@nn^1b;|c)z?9D7{fk$!;FLBgF*>>hWM(FW
z)B*SR!V^<5wZtFmYwQQidaR6nw6;`Ez-gA61&MbEepu>nFf_{Hk}(=<$CWqu84R?6
z;eqtS%FT$&altIRzC`wgSX>3ircavpK&kqwz#JG~r1x}FBYQUs#@NE`>9<~veo~2b
zq<x$=U6J^Nbk>d1trY!P<eZ0x*Y5pB&i^6lJRGTR-#=~~9Qz#O*z4GPL}uo}5wc}F
zW*OO}$gv%JCNncD37JvZn~;o>>@BlGp8DN>-#_3vPks7)?)$pl*Zcjtj7&3J`vZcC
zvA{(fDs??S$bU<;A6(y+lG#yZskECDII5g7)?-iE{N;9H{reTC55NbKa?ickaFHB&
z!<3D+A1*EM!^-y|diwZP>?hlY-^_gH+}~<+_g=t8?_W(lKHDy-Fd=%E=EM*&)l02X
zTm0JMEq1>1(TxS&c(IT1twXX#<(lkH>-N_-F*RdyM?V!ugp}mY5T`5aaTt_cIaDr}
z$!F-N$wILL>rG$D0g0f>!XPP}=fCpzS0|%9*7mJlg}=G0r%T!1y<~pEph?1a(|JQ8
zFHw4Rs?D*P%EL>(V;8upxZZH6gefu1D+c^n#4acs!LL@Wwd?hw7lLpJ%TP_>sK9ZV
zxwbO54WgC*BwWOe<=qg6^@Jw%u80NoS8UL<2Zc##EeQ#X+!nnfh2m{&Ya87}?#~m&
z-VT`?!IRz`Y}#vOdiJr)F+L>kl`Y2%q#qPvY@owMe>U^I`PTK%QuR<WQ~1mC_m9ip
zVu5!HFgqyo;rLb7Z_2b&`E&{EV&1|Lh+F4moT5E;BDKUaZd>vE&%<S`H8sP>dQzrf
zJ8!NCz1}oIxrIyFVY>4wte10rUyp!CpI<?${Dgk1Ba$Hd=ri!DGvDmJwg#mrevz+$
zs@)sQiSSKusZl~<a-tTV2=s8%gkbNeukQQ{yQqW(%|kGQTmRtIg3#y~JQ8ou_+EGY
zk3r{edAn)HAD0lmNo-GGV`ZWL``N#d=*2Fux%7^({4|>+8L#=+bqge=NFUKUaWGHe
z_bEeGXu>D)|2h6CAhXT#`N?@fqkW^bFNJ)pdw+J7Tj2g|?E?7X53rs4Sy+)SmDz7q
z`#vc@Q)2{W<U<Vdq~q^4X(l2-lSG~@!&$LkF<Hb>2py8xsWmDiUhzKhQ2SUJwDjm+
z(&>Z7j|RQDddc(4n4j6^Ufl*c$YPE3fdYh&kO)X4925B1H_MDAlgm)_@QuCY{bgG$
z#Hp0oKXUKhkl2SSKWXGuBG#peE$TYnK#Q8DvQuhPWhJhBW*e0I)4P-eIszm?RVJV@
zsY|6y=>vll{TaOzLykCXSwxFK{qQkh#FRHSNHLlSnWGaezMANEzJB!-0pDSI-APET
zG?|a3J?xp`eP*M+=v6QQX257WQQCGW>J&g@e6AwC17jHp_VFZcq)*o6k@^n|q**OK
zTG7T;!iD;>`0vk^#1BBk-!Z5$yiZGT`2LSr-Qaore4i<wxCDBhuA}glBpDw8c-~SA
zt<}*-p?%>3ck94~N4~#IuH`;k>qa8ZZnF$~3x;E5Ck=n~&+yd7MW^tqx-Bim=Y^O3
z@4dTlh$6;P4xPJZ8gBuPM6;b~pS3C-<|N#=?bn5tg|H|RN1gusNh!ugCL<}$uJ2N`
zLQxF$O=UC3>F7V-6Y{o=ARCwudo{=L=mCa5nY@5Rw54Eo;Z3D}Nyn#wJ7K%PAb+PO
z{G&O`qlD*WFfucW7e5C{!yR8o+Dgd(MH^m&T(EHjww=imqL~st;rpFYvmfG*@9;kc
zp2U3(m>9l=iTO+O@zr@*DjVk{UKt50$Vw@LlgFn@18oR`8{hryfLcbU_usYQFM^`0
zR|BN1oHOBd?-?w(*=c<fXCAWYrZR3fPVuPQiM$iurXytK6FoknJA2Vc?#BzT9OI5E
z<k(foZq9T;vr?faysY?lY9NAo1kif7)+XqW3Hhx89tvJhKXyLRJR<WZppDa37+*tB
zXo#s~6bbMBm(TdeFoBBm7Z)@g&v}*O$(a#2IyF2^zE)VUZhkC!E&51}Gz0|XX&OHu
zryKs|m;~KU%z)jbDe5Jq<%2w-)T{k~^<C3icPx+S0EP(t^6%8+=T9+8F&i~SBgCJX
z<a{_AgmB}|u<;0VjEpfq1)~J#q=~U5OaKxA?rMu#ave?Ln+`Xd`zkz4Ir~}!?B93l
z>uszO-5l=f`GBAh=zzbUYnN-Ty#mG^K4|zi?@}raJ_S5BEiMt}wD`;n0WRz`Qp6V>
z$bbkqZqcX`l*XSWAqY~fvuHpX8HN&gl{Del7<E`dpz{YUfyPTN(MbIqnXIsX&Q3+@
zGlVG3j)Ru`f0aouuPxQUkUJ#Tcncadp$^gnzp9+NMV0L2Pk2}Kn=thv7<_UR!AoH9
z`$Axs;EaFMVOF_JH|`h<EXgOuQB;k8IyR{!cLZSya{u1O#M;Rv32^@X;_*7LX)(G+
zZ}f2Lg)YO@ALC47zlrYlVuK>(6`D<o+;85FZMcoOJis{W)&E$@s4oH|6@SV9XvH|j
zGln&NrG@FPNA1BiTSmR7v-lL;HE#^6M6(cpqL8=l16*)fqy|umSHy*95x+9gJ$l$z
z*nfXL^QS%81=q=SF{8Dq(bmv>i9P%O=)De&S}Rny!)ax?bC^vDZiL)UQ!e+@o)%U&
ztjvxMdjiIy-oU6l6}>mfa(9tan;RT6J;YOjx4oa4*r?3oS15#w*uJG}ci&D)W>lsn
zeOegb?_HYcZo^PYh$FwiUGmBkwqP}NhSVV8P`39Mn=Fb1)9Iyj9wRtcP4_p~QECmn
zf*nQ~d<s@6eaD7Mn(VzV>Td9!m}nFS4=+jGM0@OPo*%)?Yuk<%P|D2uWIct`;58K4
zwRl?2cJLvTfqM(}?XG`qGMs}kT0bW_op?t<9_ll1j*uYwL)6F=1a8Az(U>TScc7N-
z0A1b)!a1i|^?h1knBU)2tBd{hgWHt8=Oe<eIbAUMS?*Nv`{KK3%KfDDdY{NfjP$~)
z@I-6=&6J^UK7nd613!iettgN|?g01{sN3_FGfh;)`id`>$aLudL=(KX&q9wt{f_%$
zX~v=71^n+L7epqR@IDDF7%cZ6h#d|FhNQ`+b56cBq9?jlp9^V!KsTwAuLf8B_yLe`
zh@J?2xJ0W1)D)a5q4ir3izuf(=jSfq4)z1(>E;}U-Js<8Pj8Vz?LU&r1olS4MZmCM
z54^MGpmxzx8(GG&ur9%^o&bxek#6F)2Z-fyO*Img#WyhcqgZG}KJ3m^6B;t6QZq^Q
z6&BD7nPM{M`fRY*NU5Y76Gy$BJAp$Jxc0Igo8V6$?`MlxN^uQ1q7#kHVwQ%su@8X>
zmz4B`Bwym?KO8g$!3q1OBO@%wYSQuW@D5rgdsWhDz(<em)?;ZV`8Vb%rD;y`BHB}a
z`wwcEZr?1c%fiPZoyouif)dH|+ggJ(Rs~qL5}0AJnT`GQEZ6>X;FIb2!adeKp>(SE
zkQIKh;u|y~Sh=6M+OI~~;u)XGHy}h<Wz?H#GQfj9Y+IuW#m#SD%LG`sd*xE{<(JhP
zHF+fg6yFNjxGIE`7)oZ-WZJmA(+~{~{YcwxLjXvKXo4c<kK?H3z)fviV_FTc6v&AV
zV6BD|a-n=e99WMt@KkTJy1`d^VMGdR=!zK7SV|(@;A0B@pCcRmk7ss!BJMKjlwt{S
z%)5X?pvh%;hufM%A1^j8NQ4S}^}2M^r9jb6B)W=Wuv4X!0P(D!dvg!BC%c)(gRft6
zj+CZHPk&ZEm@Wm@J2GLZ(6u*yEM5-#a8-v)Hg-X$o(<iXvuOxls=Ks-B0&SBF<p6%
z5hcp#3>t!+VSp=Om8tQnZR2Nr*XBU$`s9nh7A3C9Tb7_Uj;y7mnh`W=S1WNGd@#bV
zd7vAu%q$p`LomyF>8(+5dWIM$`!i|tQM_IOfP#z6-9cHbP9~qI{rk{?BqrOApm&$M
zFF_!rFpr<hPCuOG>2yfOS3c36*-gNgJKdD8?NC7Bx%6-ukS20devMF67eWb%I`a4Z
zSRfr+L58<uzNDnJviUzD*z5=_Z7co`Tz@<X+4QZ$Mkrsqo*9kw-rC#ZUO7)EH50Zh
zq&oZ^aRS)Q!ns`bDK%dUv&H61|12ruhNa)_M#9lN$lGZu-M)@~p2qhnk74elelHxT
zKWghIH$h$Day3Sq^pO_nAGNq-@mc@Yi{3uC^^?V_+L*Vv_Z!7A>@%aeJ5Uny*hK}W
zk$UyqBZ4cPQHT%m)j5NHq6*;_nAAEG$Z>Q*Y901T<;IM@cV`}6Hmaz)xpbiX=+`4A
za&MkjM2LgTI||i6xxFL*y&J@uI|3bNAYHXiuXkxO_j+!+!L1mGHYup`z%vmwDuwi9
z*_3&x$-`Nh43VwcgXD_!AV2tq)W5^0#Xjev9MQ<a-RUtCv6ph|mxW><)(2l)tei{?
zyZUpr#dVo3DJ#GTN?=rA`fld}NDTmRbKP}jJO~$<6Zqb_7Z)-I-tEQ`RA77)E@I7o
zkgxwDyCG(SB~F8uzK5;3_LXGTe#VNzo^2Fa8KYcz*S`OB0f8}PG9V7!Q#X@}tQ(^4
z8_5<OC64z^no%7~cd@wQ8&Lj@!ji!Lz@l~WYBpw`$u16jn?0L?6bII097rsjLq!6_
zdK_#yC+AFP!xJJygTm&U^sUOzL->|Hgz;H&1u`0*mButJZ&~tmc%krioiZa$p>Ue7
zQontJoXA$g9W`CFJDmx{P;~35-afMyzDaaz+f|e8=Byu|iob{)k5FXUz#<!<C|~2B
zc(e?<{M5|P=g!Nxh<<#kOx6Bx(p&IvMckba+EF91(-&P1oCD_b`qcZ;VyhimI&?<C
zQcU**<WZJBRm+~KCKh~AF<IeyfL-Tn=EI_P_usUxc)LowkFw9cf$^=Gdh&cZU!+jo
zHbr0khoH?G0D(-XRGu5j#u7YpnZj=~MEOQ|5RBT=iB>jO$MfNpM6o4mpjO_d?^L=G
z7-I*jNHN(7O#%mn3A>fS|FmNERSp{YO)zZ}k$WhJPq1&mL+}8|re3W?!oPWq^{3f}
z(7C30{TUHE4@UVC9OLLoJd(KIS2gtc$5>jJ6Tc3~vVhQu!<|*JAwc0tr)@|?s>p7B
z0x^ZhAY1Zu<<S23i=^^;1;=L<u2UF?^)7=7Q95=D8u_zRc_(K$m*9Dw{H;>;;)m10
zF+>=fgEVbkvN+$;At&CMH}^hzzrP@E@C1Fei2<F&s9Z*Fe%(SrTc`guji6B^4LL4K
zaU4lWN*E`Zhwb-v#QAK@WjS7JHzzE7{|p@Xt6ceCgNNV>%sKFI!3Upzs%*2%SQZQm
zjrRPQ<yvhH8P~MQ!<!8v*sL(l0N3@`QZP(wUoK;X`z3)>%sd(ID`5oM+d=6yr#WjS
zr^veP8@h7M?H}_RPCn~ZFa~kEB=5>dTbbLsidf>TRBfOOKdhBQ*tnM6NTDbl39gfP
z>FN*L+2c#zOT|-8HPSZlSHt*!dKdB=sFE|U?Y|55_251!OL4Eacr~NNG3E^h9_9Px
zH}8mP%IHWK$gMyE@`vb*Ur?Y^<o>P2S1flB&#@Z9{W+T&IgMOVVT0G;GC-*SbO>zR
zU76YOebIawk8A~p8uM4TxsJutzO`*k9C4u?sn-y(!7B<+uWxwh-7tN@B^7YVe{eYY
zM>mKliXZ;3HMhBfFE9K$mujGAPBt#o*M113DC)vZz?8}EwFS^XTp*=6xHnl;!b07}
z8}1Xx+5bE0rB+@exIS5nvAVvdL<5Qg(BcpO{bGUVnQRrf%z1{>p%b?WxP7!oy0UTG
zclXQFWiYHMCCX_h;#9t5RH6y8lvCq~;x!kFcTY@Tvp1|do~_1O-gV<pn42!<Tt<f5
z_5wsowU)&vp#GX;JdGEZhI{0FmO*lDhDSR}KPSfJ8D72^l12gR9D^Gd*r=XHHV&@d
zHaOROA7IHr+M2fhYKV#Sd-3uOX0eUjSDLauLk<gJ5?fQZ(l$;(*NzU18^lEX_rU5z
zAE!nm3nE;43L1jG<(fes;!E(%(A*TIfG3A<*m{p?5;&5Cif@zX<9UW1_cPSjhJUZ>
z;JnFtC$Hpp4-=MF4gaO=_r)kGh5LOtxSL{?g*1MfBWLgJ@rljuwg=Nlxx2M142Slq
zJ7(viE!vP`!OlXYHRi^_mvW5s(0c^6oE_%g>)0lAGC_7{KHLqOh~iu%6G`cNN8YM$
zb(%u2O>~cS3Ueb*Xsx|RB++2|>UPvMxOa#*-3HW@*qg7E5{f_&R-~p<x^Ttpu5JGg
zU4EIsV?7UA(1MV)`&(%@T=5X{UEVWQCVPwpaRVlqNMz-&n3+`Z6lQ7y?d@gen(_>X
z#h-9(R<*3R+bwxbuQ^LH4CWVB{R-;;=%0~l27CrgKem>JrliXYfVjOkY&IFPpN9uo
zyw8RWjs0MRhvu@y2&j6!5Fk!TJp{hnUY&?BQ;`}rg@9rP9&M^-@Y<vy(mOv^-kiAU
zmLB{Maq|5BxkyOG_Alq^tJOYK0owdlpGF-4f)qX=a$dX=U!i3c&61b*E=1_i$i|5$
zM#FPkG-NQLOmQ&rPes0$BbPbTM|xL!#%^p(N~QFDB*8kLNzt$}q8LN#>a=5}R>_EI
zjOvrNRkLq+h@VgK;?#S;j}DFJT3%JWz(O~l0#Bti?En-m@Xy^q9led`$Pwv7+68%z
z#tyOdJh<uM(cnQKryz*d%6j^Hl6+y{hL2J73xBj`MRr9Pc_Gn1%2ofkoa?_17Tlw~
zwOS@yN!t5MAMVtV{mbla%cH2ulT%Ll&ZNdVPw<vh8^!{%2Nb=?`-a@2>f4yOc(UnI
z0phwL3*RcsV8QdplAqJ_6ltF~Bq}@?9{W;aCFvE9=IquxG5trE2oKlS_H`FdE$Ed`
z{yJhSWM3cuo-K|@@7QFN@|~*{R)Gsj@!~uG!n%tSISQk__Z?Lfj=HNKfH&pv^1X`|
ztp|K&Y`4SwJ9q5W#^|%V-@Q?{kmEW-+K+%h{<jzU1$@NiLJB3AFEcY-1Oj81O-Nnt
zrK}=!Z5d3#xzPYdGW)O(JLIt_M|F)pbG%E?!T$53{iaFDGYmQ=Tm*aO`dJp|PNcCo
z`A1aBykM<AdFT+Z0H0iJJYyrE>+|n7R2d)T4p;(t_%1a(!na@w!h=sqwH~CK$T-3S
zWvf3!*`<ooe>YcL?ciN!vuSiDE%Mthe1)ZTvcIGG|F=6DPs+EzjMgzs4zyc9{gy+W
ztv;v$Zdjbrst5Sp!@giH{jZ=}-F_qfwOq4rp@&wStf*H{)hQK^OeYA1Kh!rXx&rwz
zAK%71WqZ`j=RrY(ylbcNHyt2Eoj|y{&f6;BoucU4L;^-=*N(0!HJmSDQb=<kAxdT|
z_z@uQ=sj0jSTW7KGtJ0oKUXw!+tsqxaI9vbp%2O#kxM#R@7Pb4iJj)q1yfQ|Z^!-K
zbl3&AoHh5O{plCvk0=xacTt;hAVU=jl0%S+?mtNwpR0!VyxpjnREFWQ2VedJP(a3S
zTKnYcGZ2h)0=F)3BxUYQ3u^gkkEHQ2`i0f_YD7H)WYYO`%kc6f#s+OkpQVP?uF2!o
zAM*`sAmX+ai%yk(F6`M;`%x*2Yr?`!C>IKY7I|L*Y^77C?k5gx_sH0A0`<$13@N`U
zJ3wVA7af&pL}cBNzK*tlt2Xlp)Khk#>3sr?qzM;!s=ql|qT$_faI=;qO{%_l<xB5s
zIBRY*)t}4jhT{Gk*U$VvS>h&v5dPM=jIb+dW-&MvzzNn*o!>F6fM^Wv%ZfeAz5FKJ
z#p6ZaK|!p)mcM<fh1mbTPs>UP9EyTgi0GDMa?h_SfK01?a&32`yXn28)GSfzp`iZk
z$M;leq*RZ4cBkLeLJU2VFqMI{O*Me5(;^hv4=^90Z`kSAvpHd61gtCj%GCsO0G^N!
z*t%j~%q`i`CC+uLlYPjYw|k{~>J4$@*eMX5l4hk-;gOWjEe&&Lwjv=?oYo^2<MMXw
z?B!7i4?g}__Rt&;qM?N2g~M;)93xg?YQrFpsBNWDKFnj@5)vQ&FsWt|QeL{%M>X(e
z2z7d|F-sXBU;4td44;o=xy(TV3nd%~9BBXZSQE)YIB=~F2se0!f12+eN$<wq5pLS&
z4|A}1z=OJqtAhGJd1AbUkbCA+ohZvQ2*UI`3T5;|iNDv#8N@`)YO)@(wF^U|ad>}j
zjBU)mw&!rnw~`_9n19|@Az(Fy5R()!0r~MOAn)`%FfW>|4~}EhqHWL7eVQo>{l>S;
zq(OV-pxwpK;*WVII4XAj*byt1ULY&^YehG5d7f&R+qeX9O#rAYq}pg6Cb|5j*IMJs
z#90exc_ANmbw0mkRu175MKXoDGsyieChh8}0?KdohJyXtm(*m&rQ-7-pGMbj8ae6;
zo`r5=(KnIql7U~F+}1w86<=r`KR5cX%pRjpQ25pekIeJ-*Fkt;F*Z#Pcb>{HPg(!!
z0C&z$!LPikdhwd$@i<5}UN>)KDGrQ09goh`x6}dMKnl)T&RnCL_%j=XRU_5r$JW0!
zX3+57Hm$SCeL5ob-X<vL#kCdNpQAm;dMLif7c`Q5lNFI>Uk(&%$*JP@TgID{PX_P-
zINJe;UUw$)vD=im$c|D%Lp;AFTg;!-p$<SvVC_JExB&t{O+V|JaQ7?w{y$?Lrsu`r
z?4m|@9Y$;IYb0N6wjg#;dMYlB4;?i;<qn!!RF6|_0FUnQ8X>?T!2KD3{%xQam~S9z
z>EnFdP0Ph6_>{oy`W;B$&s6=+=cR1?9ove0{fOQnlW|OQR{9-*I!U7amzk+YJ1?FU
z%<ELgY}%c^=-cMgx|Ey&2ipj_cqdGTvVH;B_NqeCe-TBPLcD<X1Yim6>FvhexnYzh
z^|OE<^NHzLN2S0IdThB+>kbCQpigG4as4F(;%q_j|3J08sLu?49Z$Engq7M}1s7*e
zR@aEd94O~Bx>E)pmM--LZ$00}TkTizIK=5o-f&xJ+yGk5bgku$FmP67-hG}kTl+O7
z&j*AAHvF#u0f7xn>8g`4B3kAywcf_$58eXE4bMvyv-W)gpYbx-rzyaMfBr)GmHNI)
z`sFu({Aa{Uh8iaJVT<rL6l2c%j0}$t!A!gWk<Am-*o(wF=1qs_I5Ft4#I?=Tsw^IW
z05*x?DJM(3xHc-446)c#fUihL&M25c<SbCz-}%3%x1{Hs?Awm1;r_3maavUJN8~$K
zA?@8A4lv6_Xr=!mkE&;KW0(%MfgY7C0M@BBhYJLIG_n@F$GAbiumhM6qln5O_W`qT
zL5mwH`<<k&)%?ZO4RxnpYF|kfe7JSCKE)ywo}&FoU;KC?iQGJUYSo>!2Z-wFpUMPh
zF3U4HMNlG0_M(5X=~K=tju$5jnx!=t_Mfj)rv$`h6fM*oImkSoCF{#K0}MsQwotnI
zODvC%P$In}T;2Sh@q6cC8rQ$qzxQZVkjsehkH%u7pP57}t2}u`u3TzmgAYhgI$E!b
z97gwdpP(~MD4Y_)Q&ZtTTmc0_`|aQ}+2r8g{4~6!4^0L|8#W2PQHUvqEun~~Y^Oz=
zL?drGrdl3~?5@}N*r7ZA@hafy(nxV*I~LYWP(E_y-^B7o+B&D_N7{J@Y+*sWQ`#KK
z>HA>VnWTX(f0X9{yXmxU^!APj!G5q|VncibDgOKp2otBW-|}=E$!ExZ?76)1r6bSx
z;gl#fhcc@gBTInsMCG(WAIES0{Wl&TfCBT4CQ3{$tT^)=;7BM-+tJ37BI?*Zy9&MB
z@^^roNE=8b!VG}U#qIy@gmL-^a34b#H-5J@R^pr^EfsVE(|rL_Fy}6dIKZiWZN#}u
zhttwZCMgvt{b=7}=ivif{_)V+Ga3X*0S-(m{z+@0z1Du{biuf!PRg1TcCi`nQaS_C
zKU^mYNOeFn%JH0}&Quti$sYPCMd<_`D}I_i#T(78O|o2jqnlCkj@3Y2Df%nYlc<o>
zdcb!p6LaT-1T<fHFW>GJfRRJyWedr@FGGIr##n|}vC)We3bIH#qO95lRY+e2f?Vd~
zp+kBJbrHN(anuaFwj?TM9QWm!E7%Y5x_Y8TGdxN<myZ3oln(24J*!-4sm6N9eNU(K
z2L$@d%-y0M<tLYzFQoPYA=A%S$C1+{*+<*n$0y55u64JF)U7ShPLqp8!70TR>IDqj
z`p&Ar>Nr>Yr&Y9+sPr8xaEu)P$n%rTcqqT0wjBKT>f%vr@xI#XY~0>RxG$w|3h<?h
zyNC#E;~+lRz9V3e-xrp-ph;AUNDerP>b8^0G(DtI3O-kIP~Nm1ukC2lV+V0;<?kr-
zz|jYZa>b3`n&dorYPrT@^#`O#vHJr{p{M%gPKX<cC~g@Sih`5ulX%TrU%?3AmjmC)
zk7x(SeMcVU>9PBX2vOu%iT+fDBdg$!0yPB2`v$9&ms4|7_>DxQ8Jt+q^3ci#9Zc@m
z*yJ%c5r2-suk@I!i{X}XDjgWpBajH(H7!t-%PApwyBtY=^wY=BD1YcWR}>*9vh>jg
z*XCQ$g-IyGNPX{vrH{Q9tle4ossi{{7q{xROzNY$sl+?@)*nGWadk2e?!b0<P8h-n
zu~#gP`3<t4!N^^G#PLa{&M;87j0iE4;cHw=p*hSe1lg9N2{cst_t_d-k`{$L+O@e{
z$?C)TWk@7b-HE?B-k0!SD=QexzhBAE9Hs5~DuR%NSvnyC-Tx>2flIM=Qk@<ZfruP#
z)J2+sy>G2{;&GPiO9NN@Qu!mgQ4ZoFZ0vKS9O*SgkiAp;Nuq_9z+OEfrIcIb7vT$f
zTqLpEjbSKe-k?Wt;$=`c9(QnR#fC*gR5P-@i;N~el1n3&&hitE49$uT2%icuPZ4`}
zHVr^DZy2=GoF5~|C%DdKFwCCaC6NsPXiQea1KOGo_Mn&xfCd|+2cW)aV%^t7e1Q#O
z_iO7Ruv3_fj=!^eA1WW;TYqb8gn(-odKq;Lx6}uq_@@FcPwJ1<$51`ve<WQ-vaUs5
z`?VgoPbEa!UaATolyD8^Rw#tVVzV6YCVs+Z>XE#*;W(F#zd|o>>JxqzDe@pEO9!<^
zotiIA*6757vjRqiMhy14ymO;ZI*+e{syI;V-<h!~US0OM_25g>q++CC?bYhjn3#zy
z>HDk%5Sp;f+<WsyOxll_yRQeVN%MI)p5%pEl_{0eLHDAcb!sv5N3?&7uKFMm6U>Lb
z8~p@s@Xn_DQ>9?5NS}i5h14;@YimIL_HkUM`O7z;2$BF;93A%8XU0zOD6H1NJih|M
z<!RMkq7pBuegEGWPLoQpja?Dn0zNFL97UVOTs8zV+R*2@BR`hWcB9YL$&C!3<BrgJ
z{iq4cz3f4@8Kk&?lvjE$-PAoM*EL3w3`|aBGxDXSIew*m=uW&BhRMFZcOiVzL*?(e
z6SIeb(lHit=gp5sbw@d==E~2KtNa=Ij*tPg5c}wgLaAbB7hP@t(6azG8G_Cyd4r;U
zupH$l3inRNWO&>|)jc^@9q_f(JwM?phP=Z?SQ4M;-Ou$#x5mGA^MY|mXeXu2k+x_?
z!$qZvI29FyXn6;5-6NVj?u5pc4Vsd&EaJ!a(-*N^FUJq{op*09mfC#cHe5tj!h=N0
zJqIU*@8(xzl?{Za3*5vkLBaP=0=AF*tSdVadqiQu0Wh6qjh_TAd!&f`OOOmfZ#Sz_
z_6H2cRRe^c>jn(sQ@-Ra<2ES4!(6VE(L!&T2g-V--(|m4U8}5Z#JWv^13c%G+av3z
zQ&=V=?Rx({xOSHn>R@ki|GKcgO;jj2mfAwarcmrZ6nYJ#;>{&^?Xk<jZk8(}BhDp9
z)bLhEv_!rzVkn9YJYZD=9TyN9{(M8-3Q{B2v)t=~=&YVm4|UD?H$#?fOZ$6e#>Ak1
zF2@Bzn(m}^`wW;_A;*#oWD~>?Gc1p`c*9+&?fShLsOY^qf+U{=o|S9z`*5Ht@`1A!
zPxi0gX0O0`Z_d}wQH9_0hm65q7#pX&xF%c$*o{GMpt8Z}M)RWIKwl=%g0Z1(r7z{N
zg?_5hnffFJtM}bkRvT}8r8#O7l!gCPYpG^{pzf%_IBZV|1oA0Zb8+fePZJjtG=-cZ
zw-8xPvXhxXbQ7+-Yby<!sHF=rsiaHlqSAYT;M-yj+*AIu-|PB?FMGkm@nf#pn}k!g
zL!nBjNRpi9QHud+L=BhZ|JnHV)5p>jJ@uTAsIF}-pJawmWFS_aCw-PqkJ5*5G;*;c
z^mvr1^M=tqGq?X}c+Oy7+pAb<&?J|V`#)8(O8F12R6+VF1lX<b<VbY$OSZW!<88cY
zH2++Gz|ScbsghwN7kypNu7`n4ex`W<LWe@|tC2<>h@KqErVoQ<Gd6Nt{Gz46FtZ*x
z)8|*dB|r>{RoCs@tZgx#hups+u}79Hz4Sl&o{|+<%q11NeKd9adu(Jvi}i7Ab$SMC
zQZ)YwGQr`olixS_I1N~=r26Zrb%ph24AdE}Ys_fLH?bZC=moobE?WRSA8pw}9^D3V
zh#hZcC11eoCG?DkIbF%RmjZgA`|If$pE@ktU$PK{is(0FF0XLNUgIaea&r|-Lf!+W
zhF3hCft4qXw!~oKr_tz0Kd=xFbdEpP{i<F{c48|AAjuW2AcMH5o+-PE#GZhX2(HXQ
zvSdSlCUR{_+I3i6-qv|8Yi7ab7`AQ8e;|gxFBerwkGRO>X4gv5G&iw~kWCNPn-><b
zZqal35cB(Ur2z*@6HweoHCUdBxWOl0qBxRt*)7k}9iqkFsKgkyV(W}C@{E5DQV&QD
z*Po(HQ@fV(*-EVnLVjS|mjcf|B$iXZ5blPb@qf-&w+ILR!aBZ(qN-isx57`~^%C9(
z^+QE?Dme}RXGqu_TEL{%LQ>nRsh3Kzn9=8H@Wg!j1c5V*C08nI|2F&*QyR&9Q(!#0
z_T=yztm$}iFEcu)l_Kztc|ZLYtKJJaZYH?oG6>W-($K6HW4mP2wSZ}1PH)uO3KdmK
z=wqZwpgTf$qTqY!vb6}ZV6@!H*Dky^nV*N>zi*7<$dM78Vx(^?N{svc$1ArB%3$3;
zK{JwI$hSAmEN<8tkG{0Hh{vZB=lvUUudyUSBOtBlW{TOjZ`t##twTgv`V<Wd$;uV#
zN+MmjQB$m=__<R3Wm>{?0A$`|HB{1Vx*W;00Q_ho9&L{iI*+V~PBh;1-Vv7m!4Lie
zuNvw#jC3SMIh%L^A)iL05)*t+3z}3AEl5m?>d9+5w<`=@jqkhV;&3lzB`Vv@JaSKx
z<^^cq#B0OKJV=T}807&&M$UA-VNSg(dM6+NyB_mdDq?p&9&jOZS*O__oCJ0SQ8*NB
z`VvoFIrR)I3Sa3vpJ`fUdEBN$ik6N?@8RX#WfYx#c{lYl4c9)989Kkew<#+&koTqV
z@j!}qDDylr=_%E&IsqbJ@ZHhJzsGKTEKq#tHEVr`%i*Wajz3V9o$5adtWk_sjFyFA
zlUjpmxoFBT+z>6!;l`wBVebd5sr<TDeM>~A?ZE8HKcHidnFHSCN1i=o`*%MMl5T5+
z8fw^LZPD?b2P$QV4j-zBA8%sqV1F|i^|Wpf9M^}lM3!i9I+Lh}r@ru=Mlu58D0K;|
zj4h`t{awBgI{wes5GLOb3IT&Ftf(E8B8gR6crF!nv5l#G1&2k+W=*k%e2_6(YeHeK
zmY9-xy;ZkcZ4FGQK1V#Wk><o>RtONd5d@$>VZYdCuG3JYnK>I=5J2MM&wRJRByt6S
zFUtp7ZiU0RBAQ5z-`i8%Z}=32`<QTXX&*R+)G^K|<GV|!L4h1+Y$oBRFdM)Djf1|<
z5%i!I)0YOcv+qJ4lGpk%RsjAHS*GqL69mB6jxz6Wa$?&-!|@-^s4<0npeOiV@LJkW
zHJ2D@4?vq7gS`*LyXnx^=MxN#_S7?lA4isxN)(o`5G?7QzH>Dlz;0fVKSNCU0bZ!c
zLonpl=8sb%!v5<VNBSkl@NINPp=33FbX``H>R1kMFSg&vUs83O`}9SvGqUBYD9_Aq
z06)Ul$%JkRZ~<YmJ4@{4yH-9E31OztiTDo4vE?COOBzpB=pDN!Kg}|Yw&<`j?x+Np
zlX%yqgtyWCw+`v?BEjOQ09<Mw+RdH=GWHY^!8I^lU>e{4@+qxwpDIl>lf&q%VN=j!
zvZ!_`OY`QXL74i=Rn<|jR$!T+V{d(WHmyJc<E=FaTQ0k?)uve|txrt#jyra@EgRsl
z7`WcYgSAnN-$5>}sR+V`?b4vW75&O(DO-fStoW-aL<<aA<h38ydY@QZBR6nt;<M?H
z8DWWiar;mk=+xWvmvIoNMak-`E1)E}H;a-c@OUE$<GRqTJXW1@dP0saF#Bqs2R7<2
zG{=ZVyd}3)Ct(4ueP&o-rH4X?E#M6J3cvpVU=3G+R^f{q>Db(+hTef>u_}8Wqg4x^
z19VUSL%17!m-}a)GfMM<h#}gPAi0cs>Zt6w{9rkc2qY;RflnZ$wrKm(v^r8<Tttt|
zjYNu*ID>4;451#rdM|1K@{==l9B<SW7`(7RIY@OM@hUmp;`Hqy<>w^Jm2gmB`ENB)
z9hK4Zb*i4oY@iNo2m>M|;PW|kuKS`jO!|KjX7oIv)wEidY;@2`+y-v~yIg)S`5N0g
zm8Up4-AMGPqFTBn18>{&k9eP#-mQ6QtycqpqvO5uu&;;u)=Li`WlHXF@B>1Xr2R_}
zvNSit)^X81vxV|~i2XZLQ+WhJF4yZFS+;?HraJTC4m<W1_HXD!@-F|)do0M&O?>im
zZkGWj$m(gZ4Lcm-s1}~NlEieQ{`bH(8w<%zt%fwU>!2veuctvAe2q$L?Ne~Twn6z)
zasQ!&c7JCB*r>)sO!jg?=}Y&wePFNPSrj?uv!Dl#m<I#WsX^21YA`M4wkCDOFsY!b
zM;DjaWwq3FI9|;b`7vy@1{7bge?)9N`DyE&@w-zgb=rCtOMvupX2+QCU!GdW{ZCYb
z`1(~;vrDIzqK8v5Sc-Dd$5Jq<t4b@yO4I@NX9<;!L7@WuS{5)!1BY7VEAJJahL#)K
z(4-?Vo;&s{l)IAzCTt8&&h9yVCq_Y7B*lOxqqc5<==RDrhi)tlxX*UC^eMC+EcrFJ
zv7=g3y(sxk`S6N#*zf(={1zEWB1yw%&2=-E>{9A2`Qo7tjIm>6f;x7B9WLB`6}^X2
z-Tu=2vM42~D(ZFPDccA&rHhif9S{!DR&<Gb1yVO(l+{9yUPwrL88It*dSEBzd{OJ3
zMM1jV1`01cpp&BEy5*hzxRI}ipQ9w4XFVYc*g?m;iN!@Gl#$dFAz*odIbz_7X2V$u
z_c*(6K1lq{7>OMuTDvPJUKPcA&R4IGRF(~D7AWF=48n-KUPg6rim8^NgUC0sQon)f
zucQ@n?`<sJP&nhWdZhRymTsUih+fAciM|VhPZ>Xre#Bl@LdGwc)mD2TP<c9(5usO5
zo1QCyG+23=Zt0~>n#s`z4e%f4eRec;_aj}&`+5<FA&YfFAf;rC)J%NYFdenuG<KGS
zn+e!hK^#!zgE$2`-B$wh(Go*ZLjugLug90(h1{roBW8Rlp_@R04SkLNlkvN<5VIxs
z-)^tbN*1T`+KewCl6vfQQ8=LYxTf3WLSs@>n(zojM1s(X%P~*yfiGse5ah)wJ)6t9
z-<FxYoe0^<tj>wW+oNePBFzi`8-9@+BwJdvo!g+b;yl!nL!J9Zw9ce*0>Fg0ca5IB
z3C*~M%UAu`V8q7$BZ%u2G^s#c8pBF)Op!&pKrfK6!=3%_Os>nNjcLNYw_#Hi*~#q;
z#M9dcBF`-bPUGkCc#(c6O=uEME{a0zje0q65`$}MUq_c-l5`4V4JMAUNxO3wH<EWj
zmt5vAgWRy>I}W=YObfWWC}}N(HR^O^`uu1WYY*PK;38jiygJKZ&2tcg>{W4_=J|Ci
z*!NE^nb4kG%0^mcqlw0!9w`n>!cp3lPq@O0%J&3<9|3Jkr_v9gA`zu>8<ryfWJs;x
zv(XMJ5#n6O;nG)Wsx;qqhcfjIEP5Z<DBjdbWId9U=Ill`xL(V$Zf}{GE6Ei~JFq;_
zB4#SB7f0}ztkfPbu#It7P#`!8ABbT7DHzrU`PRF>bnw#q`}+deVo`-Xrd8(Ai8GOn
zPbiX!sl{}gXyIM7VyGwmycSJ(L76U~_@dKGmi0lFY_KUFZG}qJ<AEF(xCDH{htUgj
zb?+z*+)A}ZulSy5(uU976#JTl^=40oJdWOQ1d*%Y%9E|S0}Ij>Tl%WP=(|BnsM?S&
z%5;u<B#z~%I8yGI8sc%vN`AY4xF%nu=;c3doTLOZdw_UM!HPx!Wtv(78QR$G1uOEu
zk-XLB0-!(JH3LBRcOZ-x^ePGJS~$pzWelh%bv-{fiFy6UbK$g}$_#r`>>>~%f*yqb
z618d)_<$%4CUfK@?U8<%l^FTZ?UQDgvRisDb0tHP$@y&kXGZ1FVBE2c8^B%w3Pq2Z
zUVEC=BlsGEpi8H^x0QPA0vfx)FdzZXgi`v)kkL;Da%a8;rria6C$lVt$lA<4?>nNB
zmkNCh-TYLnCkuOP*ew04#D9ZJxg<-bubu#ob|%GCcqxr$Je0|J*a9i&gflG7YDY6W
zXFKoS%mt#Pz^PJdyi|<z(9FB6-B-W;zyAIa@`LHO{5@)qgZ~Ot=07}{wcC62KBTC5
z?s4p4FZDfOeiD35+xD}Z2YBZqWErl^8Bf}t)cr@(Z$-D2f>pdHDn$$o`kWP|avwE%
zUMv51W>TtI|4SXMp35D_nk`|G+13{jF@>*Ay`7-fofDID%Kc3VU5ZJH5p3Ba84yor
zxeLnuD`AMck<IA93F@4_cyGf+ZI|g6YUxkP<F!7kv+w)~Xv&;_c;Eb)xgn|)({`Z<
zr4#$w13AYUEVE*e2<nW%K2-9fq(?0{8}ixKG<qA!a|w>uH+8(Jac~WuTFn+Y>XOB4
z25Kv_ytG1!`46|o7ptr^v4J;xI5RZ<(`|jrxqR0nJwoLs)p>^C->$c5oca|E=S0Ri
zO6}3)?FGLX>|_+mt*I#m#8hk<J(RE>Z+`H<itkatcGCI^8O}ZnpQYQUT5p3>j6Tl7
zSZu1Q_i`>cC`%hN3|G3Ez;8e4x2<5xB^Pu0k-su;y1YtYL9@HSaiR<P2=X<*_vWZS
z!tfvay3?fJA84+{`8>)X>6VJBiU#Y*GI-Ped_XlI`XBKXwjh`^I`#>}?X)d#rpo<R
zFMEx_GWC8*)C_kk0V6Fse6VRKb>`V`86(~K{~lSwmd-)5IZ4a{qn}hSEVC#WUV?TM
z+~BE=Da=yog-L|PFUnA81)Jnd>-4dJ6{+q@v8LukLF$Gp(0huO0B*X(40-W(Ihni}
zQ5ctT^=gVFVi<-m#lk!IB5Eq)Nn7`qZ7D2sEwM!bfL><=Xjjw0NJ8~Uuat5o>Svh8
zK~aaQ=b(;yL{CMzswN>M$>Jz3`>_Mtb$JL-k^P!o$marT(T~7U{fW*Ato)450(%M+
z$HQhi3!1$F-nZlo3tYB24!7Vm(WJ}wM*EVT6n>$InM~Or=ywqeu}qsFc#93t7UGkY
zO7dY>^><xU6nlT9XReoH+Pa~EBBhd4x;x{&sv)H8aXVWvR5i*4Ml=09{NnZBOHQGu
zHc-cV@ZPuwLSuu+ai<Hdo^t>&`wC|a$r;O&N$iN9ic=&y<rUUt)M|ga$n8b>8=lC{
zLVRlON79}t;(TGK8zUq3`x&;7u%K{;c=L<lXNxLT)4tgrrQxVc;2Hchuf$D)fOCE$
z6siJNsk{TV=hG2#7o<|FGGJcKuu|)KvLZY5j+-pJnX>gMwWuU@w}sj|!+5Z_dYyLF
zv%4Qqd?ZzP5_~7&xw}IV5qQ5xn=K2Aeu3x(a9#BsW*uyR>{EXeb}TYM;;P(>MEF6r
zMkjnI5c4L%x{yu$*M0Um;HIVNANZsz;RH7KWbT?X^C#jxccr1rL7WfYp!d;84k0_X
zG`5v|;Ujxo?d0ex6>-3>LXz9>%HR20=fO=~dXvdaMg&+CNJwRod8A8yi+j)rjwgUH
zLOotZN0C#%WTjohg5aRUe@8FZg-Zvj%)19Bk6=gQbqsuDOQVbT!(0X~6R?uZ)6a*C
zydP=Pu6tb3*ji*U>nc0lX7HH$;C%*2q{9SGccQKH=ra^QkUZx{1g#6K46;aI@VIUm
zaVW}@64XZHeynNUFYPvnd5ZA#*<@P7Gd&)KN9n%`2=x-dmiI8X9r0_rx4)9Rw6H$a
zU7JgxL_QIwu4$0&bK9TOCi+PF6IK0pfF>Gd86+dNDB=WVbum<0p4gxyn=AGn34Kdk
zEKahj34?eTZNK9N(dH4GYD}+7vSVltHt8v)KXc2>hgzMvH`@;1nA{tw5A;)Rp~5MJ
zHVHi+!Aza7kiH*XQgkd3j`E)OU5Z)MB%QntNdA_XWM1c7=c4{RJj*G+@BDf>;h(-5
zQi@XvWB&d${6rV)+Vk-AKN>UjTF+=NP)m#Mwr+|aU%Qd~Qwnikqy5wK?_bd*bCe-J
zpek-RAT-~l7C8S67`JjB>GF4cMvWIs=cy5!j$^`xtC|up3XU>L4~ECT;wJ%ndwMbz
zKuw-RHHLS!nsY0ssbSwq=H@N@SdDKI3NYG{{q9-9L#j>n3_DVF=P%t$bkziZ?0*UN
zC3Y_)J;U`&Y^qb3rHgLpsYB`;8cR6kva(A)phgN$3HmQT9$@Y!K8$;`6fAf;;o=fQ
z9d3=)F%^H%_EhhD@jU?%+0`h+)?3+>dr2fZwVFm=vU@@TES6dr`AGR`c@08Zu?j~K
z=GwdS>`tNy$mD`cdx%8f8}tT=YXM(9t_V51@bAIT!w(YU>W$%Ch(m24@RzpuuJUw<
zjGacyfvsECIceiu^^~;EeR>a`>m1-wg=S-a{-p|WxZC@O;vXZLrl_AqOs00#rvF#G
zh5Ed;o5qx3#(gO5wKag0wo8C%OZ`H~5DJ;JqO1BNJW>&kr-)FNt4j7*ap}rdTk;~a
zeH2Lgw;X5!^>Zh;@djLYHCSOG@~gfy3C94cZsZ+-rENK@iLgjq<vfgU7BFT1ggv?6
zhA(o7H(%|@wW^^<Hl-7r4yE}%`Fvyk3FN4z@i*7wtTy^za*O&`+ZG79k8$?MY-lfS
zmwWarl36STW$<gK3p~S!(FR1rLBB#*Qr}t1LKQH57X;X=aU`LYm8(NN)q4ZtuXr7Y
z*0LRF4JTRRFcr<)_P#EUpPWnLVP$2zxEgji$~e0}M|;CuFd=qgY}2BJnRyec6kBP?
z_xp(3=ui3GXU`0^m;A?`c5Xp`Io#BnS1u$5M=m!r$MKMW^AD=M1XJzO<eobtMZU1*
zXX;rT!Hp<a9kxmrcl(BL?%RnIOpy_u|6Ww`J+*m3N6>j6Cn{yP{sTJDfn{QNm%g($
zfs;zmgy4!B=BK}p2Y0NLbKRXY%+x2(N_<?c^x%v1N1|A<=mLU7M?=G>i^#;}TW$lU
z!+rf{8^hZP)HPGJUKW(a6a7Lco)BQ(;reb`@BC`@p$)@vaJ8D$JlHB+8l_`B5SQ?9
zHu#f6oa6I3P56uT_2258IIO0EN)X4s!PAWIbg`D-3Z_49QC~zO!$m1+MczxPvxJjo
zLgEW~`qNHHBTGbIarkpR`$TFx#r17k{trH}qukzqXx$-22OK8Epo#t0v*>V`SR(y0
za9@QpB_xrqJ`s4Faa2>z40A@0+eZ2wo6+>45cFIf>4K{~gm1pA>0=8_x757G5)(!q
z&VReOIO^q+LNuC5X?+@uzEi)gwxxE&z|I(pDYj5^-=XHeb-mSn7K-8E(kvV@f7M*C
z8@%Kt;iZ0ZKLDpfyvg06P{}c5Lps;m&^uIT%4r=jKKZy7_MVcPjnkMx9&A&>;hD-f
zPo@>#CZmx&Sy#nWW3$3NSm5R*!UQZ0IT~1v3Icht#el|-oUfb`QFihWW+>?ksyZq9
z^3QjOS09k}MLl0kSpPD1P(i8ljIh!vvBw7!@*()x@x*anPs*w1Olz9VSp%ZjcPxEE
zPkCR<X;Ir|Xd^)V?j7y6=;6arl?Xbq&(3DtO*EO5)qm#$EU>>TEthFCIWBs~T5{4H
z{&G`&hpL9#gdd#%gBL0xZsWP>;LxcbpEAJ}4giI;IPD4*0xcvnV+Nm;*9Z)MMQEIY
z|BI`S*t(y>$^kCo<Lbt7f2Q}7Bs}quc+r#N|C&pg@oSFarG~8o=&g24e%~A`&?`xM
z$ehqsk9o}b3452)1HMAB^SjKC+R&)pATdcTh4;pDP4m=-_hu_fCNKIB{cqt{3IXwq
zypk%}2tT`rkfG6tyr>5v9+*aJ38>bR7m_6>Pr{*C#974mSG!<ww^sbltr+s3Z=+{a
z_S|W5Sps{jB1gt{H}k)}P!Kn#@N{qC0@x4&j=tXas#q%irgTr6;lI2*5zDt$ZO&JY
zOYl`Bl-Dm$@_Fb8#DrHZ(e~@I!ygUwi2}0{G==j2IrH+@BvO~CW+h-GZyU4loJwdJ
z#gSXH%L*%R@n-U|n<|(4^gbk9VdX*f(E@!ZDNN7dXc%^edMln8^JrQQBl4UhlT~`(
z`r-zH01NWhH$QOh%G5i$J|ozGEhb;bdwFw;SY|m2zAmbHakyHZK7(GsA<iOGZ7%D9
z{(pgS2H8uu>aSjV5#u-5Ub*k@m@K>nHimExL178hE4O&vmJvL9GDC*ZJ9qS?n`S#V
z=YxGbqZt}dR36!1ij0^nO!jUsybP{3qEW<35zA+Hb?zkHwXP-hvxC*+@WS+mQk~3Z
z$ZW^0_6_r@Q}Mh!OKoxNjvq7}Tx=>EekqEiE*#P5eg!HhcPEtWK?hxQsAv<U-1eK$
z=Z^p;KmV9fi%>XnN2*m<Ut0L#jmIpd(P)pJqk-{F4Q{)wP@J7ZPxE^!+)_uuf7Dn4
zJ^y*A=l9Q__?u*M!isIXD#9;Pxr*Pd+fOgHrOrP&ANRi}DdFJNWsZiU#;_KV!<g0Y
zK$7~aJ@D_Mh$KpZDCb6x`8&pT+z}6Q2WvIT;S2BD_R4a@eMzwHP3|=qW$Qcjw!qzY
zvq%vT4yTU~ahTzR?&06@==skSt%HTq>isn(!4K+Ddl_Frb<Tdf$2oF}qp}Cw<p!h<
zCuX<kv=&Mnj(OZ--jNs;jZIJf666(1SM*GP=?PZDVh<j_552Gmt}mz;_zGxF=y>Gc
z=J}=nT=kxI#;gpw9IXG{PDmY+n*J~EeuG2=Z4?`s*S-bidgnhF=TC|Xc_h~P)qc5!
zX=Z9{5&fs>qm|x;{c~#AmNXKE@91bwg^1&9jP7H&{>^dMvE3ZGBthF*)oSk4ThLMA
zqM42~VOp-P+l9!KI~Eo==3a)8N`M9ewFP;J?)kHg!`IYW9KFBdE>b5Y_V;|r$f9?F
zMjQeUyDP$0veh7|#x#VbG5^J0Qs1XCK=Kug%q$)RpQgjQ`OTx`wuT4dd~Vlj`f_Hx
zhC~!|bzM<BB#gK*M(wlY;7i3!F#9tGO=UiF9q0_&Mea%Uxsy_o1R<xBon(%Vlo}f7
z-nDoMshei=IK~_wWnZ}6M|_@m8uE+QFl?HLLC$joA=C3Nfe?!ULcM3GPo{O}?Ysy)
zD$fEv2+~4Z<5|0Ht>)}ubyENfegY0Y)9SOLzAqz|^q}m3`k}C}8&n+nSDH8OG-<mW
z5UHLCnhYojlK8?f%!QIp#O$MT4~+j_uazs$-Ig+`tgW!Q-pYi47oS1=2~zMwp9l^5
zT9^**xXxo8(U4e@XQlXHQ1J3oujPIbW*s!6a4NSG60Vvz!B6wYS36WoZzB(@tLle_
ztDs%M>O9P@i&ftfss)5Zl;w1q5jIoKB$cnpOPXgg`2JbdYU)G`AEwlc*d@-sMg|WS
zvegBqm$qc(-8XVb&s^d@0{eK>;hM*77CX&LiC9K!g4N&k2)JFJhhh_bpU(RGnCIzh
za!^b;e|~DhbKfuJSM&plIpv%Ux7$hD4Z%K@s|Y+g&9uDqyD=J14b3I#EjwRbr}+Pn
zWi+n%GzG_(jiV0%Sb~*o`(Q?zo=x8GniYS&mC0Bx4XqxaYWxhn)`j%BJaZc|BgDQ8
z!_W_!LOxWA?lK4%hTCY~l$uxe+>zic<T2Z!(Zjj$CGsL(#v2#MH>0b+a`jxXL|kue
z|HfN$&Hl=ZROjJWW)&GPIr-`$ip;W=$-o)P94Oog+SCuS$<VJ#iK~<yt9s_CGHnO`
z=+c^aH6a=4xw+MF4c<?a_YL*2GJYnWRf%Xw1zALXlolI#2-)xSU@|gwAmV=J;?S5W
z$yCtMN2;g3s*lV0=H65$;gU^A4KH3K*=b@Y<J1U=#52-Fylk<CQ+vzqz%vpb5DAx2
zLYCSJGZGG_dF!D>{Z`KZ>7-?gadq)%qTyMdH}8as#sB@#3$r`B?mty4A3U$9VQ3HB
z`O24*3RRHl__={^wpA56h?tN+K(AMHvW2eC^A@sbK1P6TYc6m`AXu14Co0ht#Tbdj
ze1#0sdMlhW8@J8Y=~{>yo@6|g^`_iAV+~67ddFe+AK3OW0k#{s{zqeVo7u|Php0qS
zbG*t*GtCN?MbD-W!B(~Do7tZlXj2z&hTq8A`8H*AH6#Dy;Cji#)1B(6MX_0=%B`2m
zQ!!s0uD3!`>wPi(*3+wFi8vjA)fu=^^eB$s@Pm;@ID1eG?($DOYK%Dq#x#NdiaSz1
z>Ak;R2KE6tS!GmqJyT0Z39+fgE=8V^NSql4yZ>fY3dW|@vy}=d(Q&U#0^PHR#JtgY
zI9gPL<$hBcl-@e!Kr<Y61Nj3i>Z8!S1fHjcCx3!u_5>DQHlHC9@Zec=CDgapwtc~R
zm*Oaldtd+^HsFprlX$V5%%jp8!LiMc6?k~6cseLG(+17!*FQ3x7c$f~QK;DZqGCig
za<&8u_G6finb{dPL!xg40%3Rf5@RwM;VixX5Nvlhj(Z_PB)U!mw%(vA@f&Dg#9D8g
zfzUDUzr+~t|D)-<<Ei}n|DEHgW5hZ3b{u=lO7`A+lTF##Lg_k$jO?wFy|POoMWQ5?
zY$_`aMMWy9^n0Dp{rx>2_n)6T&T(Du_v`hXAB>M@>k+zN1@cQ}=*cPjW%Fr1b4lCO
zA6q)JF_|(Mt4siyV1{Z|zWx3C+6q*t3e?z*o<md`XD82!ajq<;?bGYc@**dtLwi<J
zYwadRnFyWw#;=_NHV>iA(oWO88DsAug<-bc&zw~FN2ha+BZ|E+H_}BYV!zSGvDOHm
z&v%vAbpL-Pp9oeKu~EOs>3d-N)F8{^in8_#^&X#sE!!F%{jO%?-%ePZ$}U9Pbde<O
z8gWaK=ot{1-#z~~-UxYXS|l#Cz0nyq^;hlHF{dTRHrJ*!Vj&k9gZR0*P?M%MmT_NN
zW?_&R^}_!&TTout^9t-cCd^9I;CGsmX3O1h{jBLzOU!yt@`Y%-b7f8OVV8&AMvyzW
z8gnSl9VPqLJS-A~hX2v!(Hz1j*oV19=;)B4G@UkX88IZ|l(JOCZaW&SCkpa6Y2@Hf
zaxR&W{nG-(k`mNtLk@+BU>?n}QGO0t`iS=MJe$Unahc68^hXMVZ%-)Z{xM3`>tZ|9
z{1;k`4pWIMvgXgg8TgneyhVjtCQ=xqQhaLlQoeK1^?nDNtQ&2Q(8J##%YBFtS=B%N
zM2uW$(473F2dwT+S(sT1JG##X>T+9xXUta$-3ZAdPfkDWe!0G?6~+!Vz`5sG(^Sj4
zQt5eU_#J`uQDZhvjiE=KV~60c`>c`BWHt0OXVnwp7T4LkoUJl4-!wH;xWEMs%AYG2
zbbeMq$C$c1;tXwC4tTk_3ZUW<-}ar_zi}HSei6>6F&c3@g07T{#-HV^0_hFY*MOjN
zOZYwmQR!F<St*xg&<O#YD)8R*>F?@7yFI6`)^D0(44a`sX0`jlbzSS)!+hr^u4<oT
zmcn^v4+RZ>de|N4EXlz?>D@+HH(v^mQ5Uq^G-*E7>-OeQEnj||*i}US9IyZ*nYL~g
zNd)4T-{ib8rQkhC+V~Z{aKYW<{I9nxj;{QfS4Qv1+=v$xPE@Bn@e8OejXz&?KXTr<
zIVd&uzWBsxYYyDI7@EV+%j6AD@~amfq5T5>gC~VzZ>s1l=$^LF^Hd$JGZijp!kS(?
ziiaKz22pMOuh<bDZ$xS5H5>EK5yRw$(KF=>?YNnirlVJM&|zcMzT>6apqYJMB6Cf!
zrtVS|+@Xz^NTnIvcUAjWzGiqt&)Ad0AipiX=v3IcGDyr&+MBVf-7w$Q+k~*8R<}E|
z1yv`l81RD<$It!<K4o-iLp^?1)T)v~sfI^22B;WAXWJJ@hu4E!r(e_fgY<6I6D}%V
zbmQj4D??WH19_Gdl=N)1tk$~aVSME<q3JhJ_%e=d+<9l|vuoYw`WfL7i1;`y_Fd1`
zv>LeU>9pN&+ZHe)4oRha!|+vJelzCuU-8?AY>Goa)#7$4^&zw%ZiG+coYR@C&7x~`
z0{+MiUP$CPd#7kAN{_TIP`7g_YS_xK?D6p-uXE6krfr?PPzP#I<!3C8n6hmYL2x5C
z`H#=KjjN~Su<7xUd8PncUb_~3F<lWh;Vg>B^R>~rOAD3Sl(E0j-SQ&xWs)KpJS0v+
z$2l|3d!DYjE?KJzonE&e9BH<yN=ja=Pe*E9rLnPTgTKEFd|CY<r|{j(%WF`PXq8MP
z>rUOJAuR1A9CY-;=pNI@C#dT`YtP1$OSsr%-hJl!!6sGN=zh1)xti1Gm@z*j7{Abo
zm`hem*jM29fxkg?h+g|4#-KdM(Zk`c#XEq<5O45Yz`tXpGf32dW+r!bbtGOPTH15H
zYDd2^>UQJi_M1mdzRQi)0-0M|A%|Xf3)U>qHbM7QPddfX)zkDMKKjJ3K%rZI<;q?f
zZC|_A1ej~-jjL-<Yfqb(D?IMU9f<Y6N>%&_G5z50YVa=n)!KEqqDQMx|H03sZ8H;%
z`Y&H1?5OzT(&H-%ppZ5V`Ci2OzUs8^+7?8@uuk>~uP0M8RhEvI$u|TE#4F}HW%lc6
zguOX{Y%WRSLT=4*3nkYO?QfK}EZyL>hq-6*di%YMyOLp+B&w9E?rNoBh|=+{YW-`@
z&iS1B9{M*V>cMI>7>B7;i#(6x;O9l#r6Hi`ggP*ucI%tSm#>0eKKsy=_*AAWOQvTf
z$tqOI^#$2JDkR{q`<8G>&TaXfeKXJCU%K9C2<x+BGFuPFq5`7|A*H11S>1_~_`xue
zrjzsbBpY4D`Qk@V^<&)SmM%hY9%c@fnI%$%W@zdY*)K;|I7epqDx|eFxqh|`@$pab
z;vK3FMeh9)e)v;cb1MWL263ijQ8|r0=C4xP!GBL~ZN?xZEgt#HU8Gg;u`=TA*v*SG
zBRv^X#}qH8e`SzL>kOO`-_wH~GG?(OmL!lKHeG8FK5$e<u`$F=qWQzUq(q*h4o^+y
z-JxA!X>f?qX#^_MjFfA0W<(|nDv!q;MLH|~IQ9LWVWRkU?DLVBBc04Yr{X$Ue{Gx=
zZ>9ATu*()~PBV(oh@~nES8lpF5kq^J#sB545Rdo`p*yNSbP;Bn`xT>HbGj#a!7BkY
zYOE-UP6}piYJ8=q&XCgNpaiDt)Ki+*-(jD+R!G5oh}b-RlQqxI=oL<Z5A&`RIbdng
zhw|FPpAN^@#|tOpq3Xyukhz5G#ykGh@<9KMOLrirr1ICyWa5*Jpw+8_-ML1-6=n^O
z@(&bsNOFGnWa3ACEc^YsB#9b7FqnqPI;Ttp2=ii%{&(;fKC9>3pYIS20fS4GDq0gQ
z`mcth-TKU~Y4B(%+`IZL?Q;m?nO6k?Ok^LQ#ZIq$vuQhg@1E2sx-qC-6cKoA6BMwH
zM~AtQtJ0xjX!emGlg|0R04X9W>(Y%Ap?6vCF~Ac1lQE_M8l>10hA6Q?l2i(<46El&
z$Ap>1Yy>amh&dnqTGPm<Q`BO!*u4zOvGW=>;~A_{CLg$c(TI*i>uYl}j}-;v$p-C8
z9D{7Y3iY5<s2ih4JQ@yvHs$XwH~qbME{+yoILt6R!bF&@9(^g|w`oXm9fw-C`k}_V
z6EN4tqec}=yTxa@%Io|)W^+C-u#o1-heUto(FY!@aamp&HP#<X3l$g$x%_0DKpIW1
z7KC*b8;T4)J+{n!yXxG=QM6et^l_LJGZrONZITRDO1-31Krp&($~CPIu(V3tPHlkH
zn(<0$qMC!^=dZVVw4jC84(7$4H=`xl1d8h_Kkh-hk=5l9LI?*C0D`l25By*C9A#(_
zs4;1v5g=7rd2PJHWl5Um3yz|1X_U&Sqk<79=n$mjX^rM4o><%Zn$e(WQ@ZEp+m)G<
zb1sMyZSc?;@TRmy>{c6WbIFB=vp#WRMu#qIA_wFxEhDw)4OFl(ZkV*SEX2a?aV1gr
znc4&RULd;QjaE(b(RJiCgkx5yW^7-)IK-9F%uVjebj0G^wP_gd{mm0V44*+~ES#L~
z#>6)UMA86HBr^XCW}f_JU9<4{U&@F#JMl&^-1>I(DL+0y5Pkhoyn#V5K3k|5$GLsD
zK+CYv4(ORTjglT>X%xp!!|>4^G<V8T&36<Rn-+TsW;l8O34()|j&|LO)D6d4$oE!s
zxCpr|<}K|>kx0c?;zxGe`yjWI|KiZWO;PDDoJ+jpyrsGhZkj$}#;@p~T^U53XYrb|
zRZE$?(BZ-Kg6&7qz(2vLlyeH`jRJdl$=@YUs7D?J4}kQ>YVo+kOe&>AN%K)r?<eDz
zB`fl_jyigcW~#}MTM!4CFh+}Yg6KHD<hp!1Mx24MB~ob-VkMw0InJWz$mSM>=f7#3
z!Uh)wjVfW|ROv-XIs6Ee)g;4gHH_Un8s|!Xa<;^9Zp}b7q+7-yqy1IRj)<ZsSe)^2
zGvtHnu*J=LIz3u`$e1}VZLAXc%RN?(+u8?13AWbA&H2s>Y~^$%9HRR$SJY(ajmZil
zcC1_016lEteianOLePXufD@wYo|@P!(PPP>yKty+Jc-fJycW92o{}Fou<Z~Bc7B$T
znN`JY^54nk8x*I5fhYIz=3sa7ha2@r#~!h}<-K~u9O{%A)x@b=Is?fqM|p%(Qkk+A
zsv+R!X?jH+x#ZJ`233zJHV0{kI#%mZkL4EBi#PEZqdr*pB%San>I|K%rMA`q2jopr
z>O=K#_HHN1PflFAe!6c%SVygcugo2nc{lFmCB^69F)t7v=)NtE35n9lpmR~+U4&?I
z>m4F&$l8?rq<!!7^N>H1#wlR11Ce+&N&KK5S;IBUZ44r@{%e@uw%n~Cv6xl=tMLuv
zDF)LlhfIsr#GeGoSt7FR@xnqM<CNzqa6wdBY~XqoHEaBoi(Zc;t-sT(EI&u*{>@do
zCuf$<aR*r|9y-4LzLHAYhWiP*3K*R|2pr<dVKLVpS)sZPrtcqK=%N*(&IxvWz6)81
zuQY7;kmNhGZ>S8lF6O14a!vb2sbNjLL?@HBaCz)-BdIk}J%3CmBe)8S(}JJ~%{5_P
zF-Dq9j~rvEWRs$Chw8^Ge6et4eZ4Vcq%Rs&*+OCcMMg7`Tr>N3O!@J63<M_?+`~1P
ze{ej%gYh&GI9)h@-HIv!fzP$P>}_pz9D|HLd0@HJ7ai2dktS<mqJi#C<0X0*UnfYd
zEyI}0cs=whX-8z1NP}A+FDvkoYJTXcdI95huH%)=FDx2mNJ^udU+7>udUU5M`4pJQ
zDYMgOKk8OLS}Z4c=f@juE5qn{E;{x;emy5Wx1@X_-)FR0_FhN3yrua0N~6==$eL)c
z4DZjB=#t-W#?mqfT6&taK19||Lq7HRn~?Lxo{psy{H+~6c21u{VT5q&1N0c@__oM%
zCG+e(WEy;e-4_oONEQ3=MRUYu{mXBO#ilfhv;0ZlI=e+oKau3jdj13@vw86kF&wg0
z&J1YGyp&&DwYu+Om-P$M)-u>+5R2HtjN2x@<VF29ZvIvXSH%)gvY><a0=L6>!Rai?
zH_9m#iEY8+qx}yrqAYt?AEwE(9|Zmg?(Ze%6W%Dp+a&WFmsBzY-hK+DsCsI)RL89j
zqxi)g-n;h*Q_HaJHB{;dM5Q&J8`l-AJi@1OrB}bic|@^Al0JFl2*P>a>p`k>uuKI-
zrdHE}S&M$)5?h`)4x77|5qZHZ;M3yHEy#|)79b_Z-O>NmC`6!?wVpwlq(C!g7=9v<
z?!;j$Sm2UK<yfnLI@p7zg@&<h^~ps0E-a2(HEd<@+WCEI4P}Y4xm@=AR#w)(sB4fh
z38^X;F;VS{w(dx|71hVsSRv9SE?QJ|DZJ&JpH3YA;)pQ{dG~XaTc^r^=-pxBGy3Ct
zSK9_z$ewYi>vxA-wWb<MuloLPR8tvC(cfu<7X9;mm#Z^RVOORqg>M!qMm2gneE?3b
z_MOLtSEyr>)_G=L>!-C?v@cBj5a)Zd1V&79`IH?KlCLFI{F_{o8<>=?mPGOZudgU6
zQ+8m<oN=O2*`j(ELK?LwHa@(l@v7Rn-sKg)@Bv+AHAhY}pB?zaaX1lzo}#}jsasI~
z5+`9j2e)Rfvw3Zzre7Z~s@iV`heENkgznx0rmt}16}a;_Q%=*MU2fxp@@8*aCeQIA
zs3C7>`e$YGkNSiKYrMSpztg>&0~<Rh8>6qDH)nwVl^2b<hqnj8Uu@GM%4={NLc#p)
zS{L@e2}n_)6!E#&E|>olqH4m@i+kKtTX5=c77c#%DSh`ESyP_UcJq!eOf6w?>?y*N
zYzMG&!7gJya82twK_M-%?{#Vy!d{bA`J3Y7=c=*!6@zT<0IQ~W(bOQ_mmjD}epg+M
z<^(Mx37xEU)RnNO)TU6p=qRif>`yZ~o5_P$!8nR;^0pX0&7btW*6P%WX394R@E`u-
zDP?o3YY<%I9M|P;j;IPfB^U5AqCQ2st#?Nlh+%q)4tZ3FvDnjMOGl8p`@<G)vNjD_
z?XgeuZc+H8Eljo_4i@Qfm@j&-vm&e4e||)q=dQQ8K(0G}C)eV6L~Ay7917M&|5Vp}
z-5<N6iJZqSHcB_N)65F_6;Q28qHfH}q)#zEbIxyhqhR$>qLWoDT2-%<Kn#OB3ooPq
zmq_Lq8%$>4UYXcIxN7wu+B;>hq!O4QI2R9@cRrq)-Hkn$pVITHIkh--I(@P}CY0BS
zL22?+baL{GoSzgO?Xzk`EqA3;cqw9TGg{sRP%Pto`TOIfn`F<|dQD2j?()%WZuhsa
zAOE}-DDIyekiVD5s~(?gqi1!ChL}#5L9l}Q`XP1}1TSpmy|N{IdiQeZz3scRN{3hz
zS=tK8mA~k6xW9UBi4E8AgI~tVxvCFmnsAQmRn1nr_C|Qc@E7R{BsmoY(G?uiuCvd)
zty@m6Vb;fV;Fn`whOly(%8%el`Tb^Pa@fbScfm#l@o?TEmfnhqWqJ9Kto8&~=3|+T
z{24pu2c(U`#_h(BzYPq3vhd$v*r{iga#??R`8Mp3!y#FK1Hpjw*_0>{p9#_FhrNq}
zOD0dnyyhI=FI?_i;u;tuH}Wo)%*wpz6gr5-?v*QrG}nCHcp^Dly?c}{ia|L;?1hlz
zr2q}gbXEUah6Lb1ew}>9waIs_kU_bd>a$0>D5AVPRKe^I!v;Z$#$LP0hiI`iCTwLQ
zgEzo^#U=jOdU=u6@TV0che`s5E(l-n3+7g{$LJ3k0GF^|Kk<iS@3?|@U|OBgm2cvA
zwARvE-#J2&B<E3Y%;z%>hI(}{(*yz*mfX7Pc0+<g&wNC`*$I0y7+IL?b5wK?iO`@&
z1IjtfThem(4!k78cHU1ir*3@qFu$U_4PG!fT5o^Os1+V?%Cvc4-nP$%hLk?x{*?V5
zn%SE+zv!<5ya3+<T|sXDGKkm?C(hot;+~p^^rCh=|C#x}`I_Xxn+h|GzKr${bngmX
z*2u{7`7MB;LcXrt2YIkbPk!oT8Wp95i=v{5ZL`e3eJeBhr{|N-9X~BQ?+0?#xZupv
zi=%;pb$M&Gd&RN}fYuX6SJFODN95_()a2Z+*3Z0T^TaqSU%-%ll44;<FuKS+i}Euy
zE-24E=eo{hm$aOjevaBsnvVAdfu^Dp;JRaGlL*W;+J(C3qUYlV4gx~a9Uj(<cQGh)
zJT`mv#1vQIz4=p|oOfvFf3`%$J~(pB(~LvGKey=NVz-y-Hr!w)0Rz*gUomL@p-@$J
zmilA*2b)rS@87}MJ*y<=b`?dsgfJG=POapg7T(;xQbAkvQA)H_Mcg}%w1<D=zhDD=
zRgh-LzmL{BH*3?PuW_eJy6MK5FYdz3Ltda#-JHzT;7m2XC}dr$Goj*XM1wvf@azFx
zG8~(xdCH?4Fv3l6qKwwO8eh0j;WhZt_1&Ukzp=yff;D;l#U039POHJ+@E0gq+!9+>
zl;cPjSp(N#YWK-sJ8z%nw~PH0WBZ6CQAC}Rb+K}LBDz|gIj>Q`B(Vi^hs_OEsH<&-
zpH-#L@CQTxIi!nlK}{gNd(`sz4b#5O<$-HMCz1=_7(@mbAxFMJ{GtBI&pn`Uapuo!
z{2~)w<6q<GiGNm14xzQywSYHMO93OA!|1>$;UaOo;n;t_sv=C+NZ?u?Yy{tH+iV5^
zgg{*P+H>{+I_Hn`Bj**}OD<nax3eG&js^bN+YWQqGFOCzh>V-KE0ayOA|em-HJFGL
zGVG5l3dJqv<pL=q^Z`Ha1wm5d^2f7`%Rc~2>0sTsuCt$?*)5ttcVXtioCjjw`Gd_p
zmVag7O4Mg0zGa@cU)!Pu>+-+t9He$?U~R7Ac;unt(^-LZ59Adg0%gRSWgCNv##KsA
zgFv`0TaB15h3)8mFyj@K!dAZWXL{^q=Xt{W#mzC+1+_23{A}1OR<JPcxT@a4kaE+$
z=eXT9)+{Fu%2R~I!677?FAoV<hk0X3VbduHEgOKdO3fnnwu;?Ms*00D+ou5je53pV
zcu?F!zQcNVx9e$@1K=$Dvn_k4mP$xm91@Iydma<&Q>f*;nC>Gb4#Hg)EdT@84DbGH
zB{w0<DjPf<v9%3~WlFwc`t*c=4f$vc!bZVnY?Ajd4Bl^{eBo}!E5mpsw^RRJcTm2n
zI*zIVxCpm_^9e5x54GHa5S^EIzBfMrs9J-crO&bJagT)n=|<T|2&|Tw%wk>VentGi
zOQu3U=?nx^_~&2cb=$X08-#`gTT@;6v%>YIW|N?yDD^uRwt^c_9H81mqVF5F==Z%6
z`|?&TVjX+=4><pjTQEN?nbx#odVDRfDC0$@Z$NG>khnq>`X^YJu%9Q!FMCeg4Wt^-
z6dzF*v)R&WW5S%v)bb56Ty7T6_`4Na*#Edlq9hSBl5S*V$x(Frb)U6v)@z?Jb<t@4
zlM-ZDM$)-!UYcd5#a90Df^@nRkB&#!eJZy3$)c!gJc2Gw9#s(3x}eny_mig}vHP&*
z6(&uNEKh3_I}>~F9g8Z0OCu|XoG+;?`viBq%_pDgrO{S_KTl(%_Ne{s$l&KKO^Sm}
z+#E-Jb;H5_o~`cfn-F8|^z6#<H<G0^tdjEphyHt^!cGB{2iMYdP?7YdoauXmJE7|L
zj2oF%*@zLGqeW}zknHIth9Cv&myw9I5DMLYd7L%CB@-(f2elR9S*1KKnx0z+5a#3@
z`X$OV<KDoxXR7quCklG5Dh4?;*mrayP7%MqfAaIWI9bGiBe{9OxyqiNR&*T#!Ih}^
zMBmD@H2NF7E!*8XqnDxWYGiYQW^bhgNf8uHp}l$s^D_o+hpVyP-*+%+-J`vJ<C^LZ
zWJpboa0M0`?dCY}2I!Q{C#Nq=sDSt&_1JYoITdBeb$Squr};=mhHcv25k3;Kv;Mfr
zU22XReNSfvUbK_15#S^^54=1!gvXq8NmT05Qd5$F##iK0_n39Ip={tN;yFkHTlRzU
zj_lh*VFv1z)1p@xX&UCT(a~Lg&zFFbPmVXfpv6CUo{Ke;P*P~~FXnlL`fnH}Ej=uX
ze==*EWZg~si=R5XFE>9!8OouCg3y5Yr&|CZe5%?B(F~@>D*Iiu4A)y0->CKmBE!h2
z_>i_F>g9oi_TQJG14sgRtm%JwUcGaXd6R_XmA&reul3W`O2DL5Rh~aGE|ARd_5OQS
zks61DXuTmPtf0cMmZ>!ezRd(X<2xdFZ=gqVMo-_ERBky^bYprUPNn=r>s)ry%i&X&
zrRhVwbA&&}ySW|$YZt<PV?X{{FR%>mWJR$B^kVdFgyoz+v?2XFCK+o&XNw}Y(r=Gl
z)BFd)wd=#U|2{#)8pl`5YcQ(NzJ0L3KN!5l*Qz$-3mvpQZ<#W*iXZiazoCb9rlZc_
z$X7SnEf;GKu{ww%u<8GRZ7;MVULvdrqiI<sJ^7TOu{nQB6a3W>bMoTpi<6RbxIw$Z
zjtu>BB*4F3dg9cLK?pFs7-w+j0u(@XiUm&Art9WQ6s4yQyoZJTP21~;CsOnq60ZX@
zD6ILb%9_gq@lKsplLHXU->B(Hiw1}967=8b#Nv5{PpIrRGa0I**qx?!dwQrdoibnP
zofzl@IrTJb9rGKzOm6n+URz>zjaE7vNZ8^7VGA|Cp<+(Ou*QcSv{)m#a$3R@Tp&^F
zNa)8}(Cnjrm?W8y6H4l@5R2g8O*S9>5Kz*M&8GW9WOloyA7QC-zXT%CrQVCHG_4@@
z(uV^XWd^&F72-aYnFQVvVYobYVT~#|%=RQT=;*QYMX9IgwM%Uv=BUB^GIFZo#b=6q
z=lK$D@ieo)+&#Gp#J4aq#4vPVTaKYweKM``rw7X9{UGhxF*N0$fO@#ir@8h9numzR
zhg}4~IitJxvswx!h}Q|Zyl69rV0_%(Hiy-ddN_zSzKb{pkp=;gw}JaWIf>EJ25zS)
zn(CC8B{=nAI{t<PV%3{obWTa*7<@7-3ppXpJW783FO<sHK;hMu`|g(#+aXGU^R)ne
zMj8Vl6AjUVm2S#=?R+m=^|kb3t=rx2ic+?Uxk6~9h5|Ev%D=`H==8al2BSbvZ6_gD
zn*15qQ!$MY7{`g6r_t>iJ_l**Ha!!)rj19(k%|I;#rLUUZGP_TP{zXWLzQ*xSr$>9
z7M(#LIkp1m1)&rCZ`Xlk2z=ewZD1j*FFS2VmycaDJr6Q)ZL6Vy?aR)8SpUhdikGLd
z0gIKsMvCEhR!w}jQ=ioO8eBTpz$MFdx(9W)gYC@LV$lM{BYR+QzO6rIvNt-8aNsas
zIp&L6m~q25tqS;*J0Ad3f8*|?L#s<szJ^uc{L9v$yT?CLhsAkY@+rV3mQ%G$z&hzu
zn4|$9?DAMoJ{}wvU@komLVSnAdzPw!pVv_JrLO^5_@lY-M36$&&UujMHMWcJENFc+
zZ#p@`Pme<%_Z5Sx_%;OhuMa(;km8&FzEHIK3!OtVRH$*BS_HdT6n?(2U<;$q#ke>(
z{^K!khbKvO@+u!ef@!{FRIz4PYAFv080$lqf4V`#LFWe{iQbz(h0xPSnQ0>by>>qT
zJvZ-o6<c?FqtD=};q+%bTajQ1Fxv!H=!`$}U8772?G}A{b!ud?m4Ks46fn$|l{w*Y
z)(woJu|LA@8f-|s5Icoz?x(x-?y66hO3D?AXO*?Q^YzO;it3fIXUxCrut*EUv6k-|
zgb81y<H(nAl+GGvwJy*h&3uIQsi1GT>LGc|p-Wgy3DHp_hljx9ql%$@-B*@sz-|rK
z_!$T@Hy&i>>W76}rJ-cEMH0Q9%y_7+C6-v~j`a+w>Us*t!gofl0G=a~)kwP3<mWtG
z;=fD^6k)Gqn%d}ByYKnqNm4si7kN!gE|Ww50fLrX$?r4iewgjR8fNIs-5=`=ktKmK
z*w2?lYq01Fg4Pd?=e;7jPbxnXNRPQ5AjhgCKGT(i6~e{09bq<$%Sn8o+pdUd+e~Te
z``DbkSodnsJm>`snB?A5ai<2!R@X)`bS)q77=^C2fcbj;Rr`02T3}+=qFGdFQy-%8
z<!(i@q-63lwJ}G8LJ}+Mo3mUNncbs@aCNknIL{gV>1{|Tziy$GA)>5|sEFQIjNL5Y
zL|5~9XNu~VP1~C&p5mf7?hUv>QNB2pmyv-g973;pBfrOgtE+5sB?7y1Pmi=vsxM1*
zOmlCAG8)D|?UROi!Lj&D?9gwJ=Xm=u6<gYtDvezFvkLW3EZ%Z$HrQO-e~`2_$S_#W
zS1*1bbcp*@wIN@E9U^vIpS=}N9-TXzbJ&j(#iYJ+$FlcmJvD53FPJ1K6ls$ca_waS
zr+UgmZygJi0^*J~rSg)TR@rWdR|%q$AF5D4>j0XGCdQVHrus7VvoYpxkBzj!rO$1n
z=k6&hXE~Ngs1!p_I0lLLrd{Pk!$>Y7L&E63J8IZlW{co*KFYg(hm3ERfnr7+*sgCc
zEj=v$6BF}gDFFb3g?T?;$6H@Sv7%3|9FJVriV#-D;yMhRxP=Vo>A?8Y?$qg>+!p$(
z2Y(_YkEow?;e!81PP8B=1G_AV#Pdmj@t&4V%J&~XZbhjaITDa^dvxi9@LiZj2bX2n
zJ0)k2A^oE%Wph^>FhgSF?zOyd_^Xnhh(6;F+gUe8AGz|$Bi?}q_2(wYw+LR%n0cn<
zkM!XYUD(_kq~wq5k4EyY4H9)^JzX=Vg#We&Ki7*aEKp5Is5|@~=Q-3sjD_P_H1G$Q
zh+ctm;ZV0)ogEFc=*p{}$r}Vx-af3gJcvqEW3;I3@V6G1&`=-Ko|lBZc8AvTtOSFY
zlXpZ25!kcOje_&;#+rmCkp!>`A&Z4?HNL;#-gy1kg5O*)^g67M7W8ZG%Yrpqkjn+-
zOyl1_XNe0?qhka}M>rX7yuNFKh!Y2;J7c%kz~c1!R>%oxr?h9@mMOz7jRXJk2!skn
zMH9G9a?c0&g<M6q$vJewrct9rulWwI3`)_p-b)lY5_VC58bUSR{nZB-y$<=qVO*20
zBrPHOKrCAe%FsGky}ImNEruW5H7~PhSh#cZ2J7)p?AHG}anw{ODV%Ho55y5SrwrET
z-lLY3Qy5y+GWiA<6EZYJd&<Gq{5TG?t3!&<jHi~I1Ydz#P%tf6xG$1_WeArL+r`CG
zCu+)Cc}`KQ2YfloQnI>`u%2glh5n_nVX$0YJE#YTQfM2d(u;zOxz>H^>K0PZ6flUQ
z&7;=`Suv-9PKNu*Js%f>vQ;rgCF-Qm<SD)yK>p1k?q{V^p19q?$h8CaUQoKRn-=_5
zx@QI5VA$CSCr%gMY{LW>hC|i6*u<#|emB!TCZN0Yg7blxQPDG@@h=?@7(WU<_TLy3
zog>4rHCU@vW^ZeDgH;jlt;_eC(&(wp2vvjDonSkSdL2ENC-&xpXh2L`7cJ(F7Qp3G
zh@F-6PY(1jOGeYE-%o{v=(IezQRDajI&D_LIs}Q!L(yKHcSW9#PQR;*gUTai_4vp6
zLRDJP@yuMmL5`H=892+LWT#P1&EpvkOo*C4kEe|H%xRuO#jgwdZY(V=6~jn%5_}=z
zXBi-I3XWQfYZ@Uku9KGuo~mdB5Z|u5mK#>#7^r756?V7j1FmQaLJUuwEA1y4`Q50y
zzV~6f`~u5&ID)n>$9>-d&k-ct6a^=hN=@2qoP)=yu)W!-vs#MBMC6_}g(EQ>xo2(2
z9F9Vn{VU{+^dk6R*)><=1>&c>@wV)vDJCUiMs-6%-cmNJ(mjY2rDQ&CTHd^odvi-k
zROJ%6bTN3LE06l5n^SB|rTk6uCpdpKu9ucP9=K!AuI4XcP%rTW=4U@N>U2l&p36qu
zxRVNQqs3i38L<780C^^Jb4&f@BtNsP=O<xv<(bGzc7aMoz#C<!zWuOf<vxKcwy@Fl
z94|fEgt&=1Qlmc;v|g*T3L0i}hx_&T8vcT+p<|&-t{|)Pm7&x5{VhMk)(VBkq>5%T
zb}SmVq2q}cE@C|^By8)=V9lD!OUz}B6@_?;H%U3ic#ik7^7G%W?;-71*H?;p*$#*F
zZj>2o8}$mBW6~N(ZWVJ`L8;L_4De9COhA7BxKA(erAHyuLssT}XJQ-H=0w8Ga;MMX
zcYBnF>3aeSJJD?OLrw}{N;Y$GSZtU|y^K}m<wFl+g{!$mqLaQpZ<uO`oRhO>d|_}`
zieE1iv+&TF$D@7f&z-!TCmh;?(imCxzaVA{bWwsX?F7*4G%<QE1)YjM$MUdH8tK#*
zmW&`^mD4DWuJ;`Fsx*tFGZLLYw6bJ&gO5%6;w!Q-@}2@qqC12tI1fVr&MPXxlURU$
za2QGAAdO{^Ej}=R$H1s%#9q~)4rn)LjL6ed0C&Pvay_!1*kaA46ihPbHlang06RV1
z=B7$FOV((BilIo61*uDKc0G+pWpH^n-JEv<43&mB0e?ge`BN(lw&qHk9m*uP#;B(9
zHP*KP+Ur-=mJxDU(d_0%Pr8J368o2dhPDoWCf4;KT{!^)j3#x=S2M%onf~4nvh?E6
zze&NbW&edDK%#k{2i2Gf&LrOyE~t<)6SuPqLovROK(40?#{b!V)DC-4qk8PwAtfXH
zb%5qdZellWr5lBH8*>Ja9DDT@BwNig4Mw8aVVxC-I!OQpOWIoc3}IKTiF!Pv%I}J5
z>bJ`iM%*mVj#NcI_m-O6|B?PxTaQ#NB4kp?4qVpQ7>L6z%5mG>wssy>IQiiF)$?^#
z3MY}=Csc_{A{JWbo}9L=#B!oFl<@gq?ZL<mK(BCr1M^wpydU6Y)F<i}NP;Ojng=~F
z_67W4`o06X58Dru13^mY1_bzpI1|R~YBbAw0RU%D0fWiK<6b9_$nurH|2LgdNAs8G
zi=6?a3I4U^!Q+-%`&~%W#095JyB=EN1wt-28s_Z;SeapLxcus*^eqRdV0V)RB)Vkg
zG8sfl((1y>Qd((cBq$#46X6rOe&?o$cKSFo-y5xwx3`z#Y4bCaRq}7$g3BiP>1431
z*AQmSc*>Xek`Uz;y`EQOxzNO+%_f5T5%5B47?5_kk&n+~ZB9Rb)R+<2%um5U|I6R(
ziv#8kL_4ubx%k$dR0)ZFBD<gzJTH6I;!i>=pil{jzA8f03bqD+H#ve(ai(VdK7<j#
zEYMyb8)KaE=_~*pG2ey%V$rjBf@4aS!r3QRRnN3}3{?0GFeW=c67-i!uMclndJCX7
z0<o2w8vX2A&i9!73&Q-*#4vB%(dvuYDz_?`I5f=Bw8=!nVxMhmKw4&dg&r1L4u8ax
zY>kV{i*SsnB2-eL$*my$*O&>WqOGcts{I^r#%k>kcTvwA<$~T(CfJ@ce>LnMhSTmJ
zZ=HnK`)601>iL7Rqr&Yr<A*Z$k)K}}Jnni&{59pCXLWk}Tf?_}bcXwQglO78+PV*B
z<vsx05ir<@7Gr)&)i!u-V>M>o>}d5V3@!sypIQ;V#-7v<gk`+q3w02SnBB0W_VqGX
z4w3w%He&EAqtFaa%2#h5a)n-j0}XKh$9kqyY9+z0HjlD-j4f}tAz~X7k(knl+l_cH
z?OPfz?K0|t>dJV}eNQ~tpdwQQy~mC!wXB-_>7X5RIP4c$O;N?p{7CRt+Op|Tv6upB
zVX;R%!BVTMCaA<-e74#*wDZ(f=L)Dr%CVr0dJ&jf6Tdce#>iPZR;xA5K>rp5b~*=?
zo#(2Lp8cuaD2e1k_EF&VhMZFJR;5Qld<Yj5Fh8Z&=KuWqMo!zvWHXWf#~%jEUT>ux
zfQ*gIZ|t3;*jpf{3+!SHy8m7QD;d#U-E8&?8#pfG+$H$&_QsXkY*flBz7k@=p@_$}
zJkIO!L)JQhUf1iw=->qBV%^ap&UE!}yBo+<x=7MB+IsP3j%G3NQgUI^xX2Z4A-7+@
z>orjO=K%{5fIy;w#=M{Rw%^ob6v}vLq;;lOW$%Zmh=fz%^0G#XmkB2XClx0*q!xn6
z%?q1COSc9lUap>uGQ<xa)u(b$VuwpIGK#SAiV2hEe(KXa{K1lWP2cH_Z6oKQs5XU`
zEyl%;Rd-b2w^KuS@d0HD#m1>%MUYRMB<llBimtmO%4LRG$tN_Dvp4kI|NQy(oP>Mi
z<Rd6K?|wAH4RgRCkjmyw>%CuDbl<o@#ImCC7!{W^zY*{aM$@5heP!{4WR$y<QeES6
zjUj{ZS2!c?A7e99&_jY4@PyNq&<n#$<4K=&`$DbhsbCt9d}mAQQ70<QrQX*+0#7<+
z%Ffxeq;gb<x+RHoD)_HtGd#p-*HFSj_TiL}2)Sm^)b0g5kFU%|%q;W*uJWSEy+|6*
z0fT#Oa_qD3KO(g?Lyr27QkJ~hEKo)Z>Un|^bW?h-11U~=lzbt%#E4RpKkUz5__)yd
z$<b=*j_~f&uge>q9@q`rGuvfHx9Z9$KXCD-GqK@4Os;S{_j4Q&9}ZKa<9ucqGJQs>
zuqMX9_Ba1`#EyAfMEj-^j!@AE#~Fm=iX@au`qVpHzKcora(LvRPZ0oj&tsM_V%lRq
z<QVu|4tjD{UA`*0^hJqT>qD76!#|6I7#>ttB#aCj!yUB;AjK|TQ>Iou<A0P#N!nI(
zX$4-Ti&snjk5nR8XJ;HSrkUvSOlA@P#4J1gHy;>0%(r_57apQhCFQ@-ucaIS6Y^A3
zmw`b2u!Ys{34U>YKfJLU!QPhV<s3b!qfhgx8I*Ru&&lBel-U&HdV-5qtO6^_Y5Pf}
zM2g?Gn+@~n9n<W704QH{TEWrj8y~J;b2?(9%BOOBiV{U?=_XgDx`f^v0Mn30|H!;B
zR*V?9C~?Y!&$~R;1k)*xIvyqC8R;pb&9aq@)_G3dO37b>z*R5>jpxB8wHz=$$INx<
zr+;gnvf#lxK}<mGeT!X?oBpbhcjp<mFCf>K0u<$7F3z@|*qIbz&F(LEYTTdhQ9Gen
z9s7)JDg4%)i&B>DZqEBS7_LJ5w++6%d+KK8kvc((x)Sql_Ly13Hx#|93_^pxuUBbZ
zW6**0UYy)m!>8-F(ddendEb(cyy%h{dqYdr{B5r161{{AEY370D&2|H`9=Zkfoq3N
zKh@^H>(b%WPn3OMNGL5xehFaxQ4IdKUM#p*?l7u8hlp>3ze-OF^g6E|k4&|A)h?M^
zzkbU#+w$@qPgBRqbOJ7e;#=?V#Z)xo;2VT6*Z%|*X!gaSy>D3`xEokm>U<=gy@reY
zXX%OOYcx)ms^06oTVoWep+)){a$`83VH(eZLA}Qfo;NY0tknlGyQ0DBr6-Z9v&1)t
zpA1_Re*hYZdz{@p^Ne$gK5OD+4M!1PtkDchealH*kX3BTfOOV8Jqw2X)o;VU5E$ki
zO6Y}Z?hl-3bx|A!r^un00*WrBTny@jt>paMet2EL^{N$!T{-RC=}jdg!~H^Ue|W(f
zcLGxq7FeU_{vED(^D;TAWiF4akg7BMwd<;K>{r5<1#*DItrB?<=_o6w^V96f6F#0W
zBpH(%nJZZfIK0*uU!6v<0mH2P4^Ni38Ju#JrHn1;@tD1AVmgBV05Meq>LyZOKS)S7
zr}?p=eT$F2#dkjMEB?yro!MB<d_D3k+tfSJbzY`}!2vyXYx>k-6l4C=&<kN&q8=^m
zi-blQmHHgA0BZ7!UOW(m)MjbXzvX;KwX(GU>VJcI9w9Y;A6d`5@5p(bPvZSTCmo+g
zKWLYxl^eJ(uf{&GPyni}AMFs|as9~gO_-5xUO%?+h!t~*@VU{=*i*;OX}y+!JB;#I
zI0Uddkv;P46+TBuS**I~SAuW2?mHmYqF<N@C?-qx4__Gu7qTDrG0*#-4_U7^5{b7H
zs`DGYPzWmUt?X0lf7t=Mn+rEw?-x^wPc)VcvD}8*T=|<CftQ&r<lb{S9#xR*SWL}k
z3k{=j3y7!2S}8l*uIdLr$CFr=_O|1t)3GZoXgF1TIDZE2j)uqC4r;N*EnSx{P3)Mn
zO5lF@Q(3N-ThsJBhme3#oP=B7E-2@)b9L47hyl@cxpoMz#N}so@gA{9%)z}^SHR@<
zHr_KUWVn1JRCt*N-OR7G!ln13+4$tBu2|0F0+Aa>wMU&8-QjLBdsZ6lt!T8WrWiMW
zqa*0ygR4VtH(Zl0a5(CzsP`aSNTzv@=B-QqY-o5*PV(#9y#PCdm*cg`99HbgGjDT*
zW6aH5S?EW3qq|k<EfCvXO1THyu~ML)OXejFd?iS*nAtkQu_Jegc*}5VeYGOI|5SaJ
zuH>G0kW?23*6RLkP|ks8_S8rVy)aZYfpJQ4Cht)$rgAMaxf@z!or5*kj%EHSo)@|$
z0Wf3or@~j58`?0jKJ_VKz#^xsz0;pu7z+!>kxuD5t*@;Jcr>L`Q1_d?1M=72R;{+$
zmGDr~Oeb7WAwr;Z_03J=kJq7Ea0_%`KLrJITl-?7j(SV&|D4jlbm&kCOk?YyB_=cp
z22);{7WiQAGnsd(DsJD{KmtfgCO!BQz4?c}oA2<jn+hd&{#XdIs5gPL{$xO;)R=dg
zB^oFQyS{JHAEl)92gzy33kwxICW)j~XHt|@S^t4OeGrcNCMFNCAH8Y@d%&X&72s#-
zobu4B;SYbzu*uw?UibZ$yO{FMzH48uYGRjE$86_pox)~oO@3O}upoWDD(U7OX;&;?
zb)gp^dIY4=9oAH%ts$FIO>H1;tgl_XKCw3-&CF7}ZE~ozD!q5UJBTeMW8v%@Z=#d#
zGQ&;{RbM4O60;ab)pF?*8VzH&MAgo;YK=SH18q`Lx9Dt-*x`?lKIVR_4`8vyKc=NT
zM3*+@^q&<C);7Y=uVsD}UQ}iacR)5^h49$oZ>JJ|f&XFo>fO$kd+WC=(P&`A1nR`$
zT|C+6W}j=~i^EJBdcxt~A2j(zq?<SF&|qOwO*x5ENvw3o_IIJUeTLn{bX)f)@pVJm
zC~bjVlOs6u_Wb$k0k76%Kib8-#>c8EOOEH+3r*ka2IE<#I)zun1sOaH-Sm)C5?E$&
zJM5eSp2jXXanK)T_s+JV^EpC!SsBi$1P9Yk?>kw&TvF<@7tReOQ_sp27AI+i;Rrs3
zN2oPtRvIPCAL>xFVRSM|v!DXw8-Rol1K-X30)+LlnF5<Z$xI2E%?ZaTgU<(u?W1!#
z4)!oo!DDKERYlPPIs3~5H$j^?bKUU<YBwX9-jCwkxj$W44^b$=t7rYu5RYOPmr~_n
z|2*%&wdw+p|7ifl{3-eP4!?F8t}&aXJ6Ri>mJoOOS3)(<%h=-D`*AHSy+z>Vt2TK}
zHkBl)t@FbF1%>9hr0r|_kP|Op4M)Bg^mi_u_31W7eyGh7cdGd_1mf<eg__(t93X?;
z<4uGoqtkQeZh5umIgF_G1e80}kAbnKD>Zw0z91Be&@JOQ11FT#`&#Z@F>u<?Nufyj
z)`8nz7(g_=qQ=MV?msC#>eKQA4CNEdD0F7!M~1iFk>>dDLp_JOX(_2xn6k9wS}J_E
z_mCMoUo|cN+n)E6EUv0hNrgR2S+@McFM4hDRJ=IG9`^W@2OmIemnKlU_1ItV-m3JV
zLlMyG%5M&>N{Ewe-W|u5XJkJnQfRkhScPcIpD)__WFA9LD+#7w4-iL(p)=tCeygz4
zwI9LhwV}y~N<zbwV-Eyou`$3c8wgx8MPH0U`FC{8^z*_VbG*iThTntz5$30!_Ow;H
z!LwVJTdvB|Wm(b56-+2>N_ZS__y^%{&bG`OP4B#IJ!0Ps|8*fre|e%HVp9QaXWak3
z;_MY(p__>ROV7ZZYxJk?bYmY;nIinQ?+uv2VmpZPm(vZ+03|vhR|0XAzR%x*%}t&M
zHWNsk__g75=A{h%noq|yzzR{O6ygFS%++E&2M-GY$-kGexm1^Ydf?l#>k+W$J1bgG
zjF4L8+pZ<%i<NR?(b>eWuxqJ7{QT$N;UgR?A40y@e$zUUnRcjwSTE9ZM^m)jcoX|M
zhie{hOfejCvIddL|20SB$c=C4X{Y{?{{3iUrgy@pXK{RyBdbQ%F%s<ad^Mu~Ih0_!
z!}<|M&%olAF*k7GCKa_?<*VYEl8X$!6RAsZ3Im!d2WneUp5H&ec#!*Kk84wE?;hbq
z2XibX{Co{zsbdKv@H=U|v&dnGuW(m-IR7|uq?7o!CV*RoOFwzHfE~k%AMu&1PIfXB
zd)aXC2KD{w#N+G8<tqib9_sf6o_&{5|CBbk{<r{&cP_MQLe6Q%lP?yWoP^Pj5~~Fr
zz=pu0EgUR-gpo*^Y6FjD^1Mi}*bxkBnV^-Wxo7<C{qryP4A$tv_#zt&>27vo;}JHR
zi^s{o!<%`vn9Wk+5JsDhYieR8Lm<1nD?xDH_I>unS!sbn$7ASUS>*}-!)NPYZ2a({
zNLDG^bd7(cA$((gz9FiP3jH-E{AVXiv9y~DZKA%>9FD-FmD%78Yew$DPDVJ2!#P0+
zQ}&d%g^&8OF(oC+!L{QimNN+<5h&15o$v^JaAaYW<pj=8>)gDvsLdDRd9ImzM{ahf
zxmfgl>1)-r^v}3A`_lA@0>0yki=b<Gz!LVD%)suXX_FA<PVeOlV@^xNLEg)k-t|Fh
zH&{5C^~mA3cZ&=S9FL&fdaSXMNenYQ0Vlr~eH9z|L<lBPTg!1F9c#qb<9}@FF^L3$
zLGi-mH4p{7X2K0~;P_(1Pu}@fktJhbvwUdLrI6WX(h1j*O(7HtgI(cyYQfn|n_JIS
zHpOoa`^rK9EQ=tX<BGuwWEeoSa5A<9REo0RH&4Ij)RAsDTy#K%62a|l-CouJ%1G4u
z>Ic_-GZd;5VHKj{q<R)XW+sK;-~TFs)=j0`7Vp3sjd0-J_wIvl2m}wGG``1j=e3*@
znk7nz8gpqAj$<}yBpKU!*25etqk-mtc_s#+MJiuB{)!HX(NS{bbnn~Z>ur5BG&Ar%
z(tql(y<j|0P6+!j0lY;PQJ8zM(6HynA`(<{PNbn<i?koLxFyOxLailudXC(#h7dYn
zNd*XW`nuzDzK?rmw67riaZyMKG+=(us~T%CXgFk()>IO#lO?1R=-fUhCi#Dz8wWV{
zZTLy$Gq@G=<p#@ZWa9S=$8mqy1dDbLpUHJRS9lFMSSa`?A9EjVnvFqN7pUr`{u%{a
zvtI%PNi%#+Z{GNqEufo2ICizzFnHU#pG4D44YebIDxIeD{T_TqFP_L_<_)BsyD$gu
zc7<?lJ-bJv1}+-%#(`t=D0Jhv0Wx(nlF-9YA}vq>@CJOhWDXK?aJ{bGeODe7mFK!1
zg-Qh38OsvF$95>fA6u;?-S>fbQKjUxt(KN|=8)EPKrZGv{_huAG2b0#vII;|c(eac
z6Bb?Wk|}u0DDS=D@K*+O>|oma=g>;HFHY^~y3gLXtP4o(T<Zz=-?6b%cQpGheZZjg
zXmvcYSV|dorM<-?(7vs|$3esbLw`NeP9t5KVH7jP#`jZbF3L8K-YVBm%Bh2-wj8CB
z7eob5WiG$&p)_Wp4x<G>t*LZam&d@QP}+u_*LO)i7-DI7qnmp?@*N!WURI@Qab4Lv
znd5m_PpNX1Xz^{fuseF^CCJ_vf(F|6<KuJeeAOMH7nn{A-_ZU{R{S7vy8b4&n|k)p
zU2r*I1{uEgAP?pMq;OhJtIwR2vWc@quKR^;m_&r%v!>gKVul%&{{j#q<ujjMXYZFc
z6TB}Hwa_kru;@M?4fn+F!JbWd{p?tVkM1?)aPP5_!(VmG-6m$gOm9pzJj&r?V3RSZ
zhN0;?oL1KmSeO9b!A{*+O27Hv57!dFC|sQ%ajAVFFI*i+qBs}?nS`y~2});uy+=6G
z;wgYz0n#1xT%RXQcEWPjlSPs8yjIp$a-wXvayR39>y_Kt6sJ=`rBs33V<`EsKKJ0$
zBMCV$b_4T?Ku63yI28bty$QlIM^_{~B|5FP?!P5(k)JS2+&W!`mBh~HP~e3M&BpSD
zGXJ`n=Ij#&b^d|osG+}4(Bnho)Pazg_7IsW>$@}vFqv0MSe%tx>H}rzFB&IK7uTz1
zDEIlC0{x;t46Zb6(*02zWIY49M;qv#K$yw1xsR{8ih59^5wP2>aKT4@Stnc`TA^jF
zM~$ge#Ghgk`kXwclY8k0ggx8LSWKO{Ft-;cdjQYi{HG<r$9XSF=>{fg!7G^TC^UQb
zBalNa?d#`kuzMS=iPG>m3X{@gi5Z%J9;oBC!TX~SRWHi;|HBB#Ne<eBd|2?m<!OGd
z@#iP6xTGUkyH*#_^5304TGY?NH4#-uyC(<H-K%3(a%pW?%!1wS`)dfKU&KIIY2T7s
z!uk)Uk?-UJy4bItj1x=HF9qVZj3%RNkE{e);744>PfRndJgD`*9NH2k^4}W0gLREv
zO(D~{`rYYH#s^4X?+_)Xa&o0DPPT(i3>%v`2p28z1chO~%28ke(fv5_w!{M*GM_;2
zCbKFuNifGKytdYx+0QqX=D#&s!fj^#y$FI8eq-rCr3ay4pK~MxgfmI?k$0Rj^Y6a+
zFL(okGR5wGn)?D0Q@ENJnrs^+-3VFmB_LWyFytP2MNGIA%~~y(p<g&r;^=mM4*a_d
z16g5<w}ef9Y&bd-T!r_quzq4Z9?FJRdB`f|+%+uyy^aqi`i|^Dpao*V;AG=4vM)3_
zDS;VwS!daiYy0o15+~-$`#l49q}G&d<U2E#x3slpzEf$m(z$0esOu2pGU}nRJHWWX
zqE7#Ss#$UTRiXUFtWpzuSPx|X41?1dS_OUXCs94%d~%JRpOgaJz5#voXRF(47WCgG
zUjyx_#SI8D<Xy8iB%Z6IN0;xg8c*L<$;mY$YZWx~NL#bu(4suv#t=PaRR_#Yt}_$n
z1S$2{qqBJk#0qGQv3xo{rW7ZWb%N1T#|`oVHU~5SC>N+m8>=8!KIa;jzZ%lmR~|Xo
z*ZXHNbAyaQ*(>Ck!YPeG3E^NNHBUrS72<?A_yw7yIBwRq{`bj^iDVfp46EEw+wk~P
zZLa_pr|{xtfURL)u|>22h4hlw10=Ob+Ksb5@_a|4c*w8Cpmz1&n>RZ8EiWr&8C3(*
ztve}GPrS<K_xq$Kh-Pgq)lVV=lQsYS^NHG(aK$>hMHj_<Gzrd&xow6aKi5}rm7=OC
zp7oa3vQ4!`dK1Jyk^B5?pX#+v;|YJ8;Zm~;N-BTf_o>bFu7EnyK1?$KkgWPiwk9{7
z6~G`sJRBV$%VF@}htt;Lz`#*oc7i|2KU)j7S8^mpFC+1u05<oMaM25HZ3h(=%%tuL
znZ1WBVsVES?6UlVpN0~P7e1599-cMt1R)5gqD+)%>sf2N^O;cro;osnmy}9!%1lCH
z9_f1+;`U&j4mV#8dvGLsSNdvTHbtjCbE`58Wbfkr1rET!XPCp=n$<&WG{nNA^7hut
z(<cB|8Gc{-JHeF;4yJ)G6IB-sWuJ7!6JU@^1+m(<$O&E7Je&W0ia{3gBS|wZm$r;2
zq*Q??;{2S=t|=bHi{wD406BwM%w49$*?{C04tqDa9m+&3d7rQuKN#evD3R}VoX$&Z
ze_vp!@ffU71N@^!5^Pk@{<PNkm!Q!T(J<~@N;>oF+xyVjmW5w%M0|UCtYISOKcx)$
zcylEG@Hth9c(Mvv91HBqI+#7I^*OJNp%CgI7;oShD+gl_+|nkZO6TIQg;)i_Z5^B*
zXB6}K)rQDeWHsB*FUx!^m9#Q{jv2qz{P*iQFr9yw|NZrtJ{2KHz~YW%<rVHAPBYN%
zIXnO`ilAd&i<z-0-KabvFffZk)2o^3ib}Lekd$7m8xHh+(q~N8$@dN7h9~4!fs4P-
zq`m(Yd4mPP^Le1)7@hR!<|?%>q}5uv2jVBVFtRM_5F7!xNzFmYbM%Dh*?)lhja-Zd
z#0y^Cd(e%wb_{~w(g)$~XuJ)PGXfRo&Z9wP^|C+i8$jfAs&{1US?e$;4`RpJpQoFH
z+~POt?jh3_g*Hb}P3iTEJoq^BsT_*+938y&<yb58#NmuJ*rR_#7YM{9pdCmcv5+7X
zrd2ASfxG59%6!k&ci`8%%hXk5$pYDI9*Kdy+8#_eL2bPcS&gw;rRryY!1mY!D;g;A
z=B_)2ewe7zkS|(97e*1WACgFj@1>6s4S)C>hFO2T(a2!^&vlQ%Cc-xnIazC5dJ&4-
ze!-~;aYXiJ`l!CY0gWz@tear2pZoU4s-^DPUONWNsB5bAsI+s_K;o|^Xlc91z*hkV
z<u^H+()GkU9YyE=d*VCN1_3#PZ#ty2)8^^0OX2Po;cX%^Cg%0M<2bddXJEHr#rge!
zu`WTWi?08}TnYG6)icuBKjDeK3ydUaSxHkWxw2`K+S)77RLL#tsRiF}x{Y9{sA!<b
z<Tsd={f-(?(BFEq7Xo6L187Jb39XsF4XBiFl}^eg;|Bb3g*zI7!CmnrV{ill^53mw
zgLQ>a+asI`MjsjVIl(Hh!KF>jq9pQM?b{_|9EI~cz_g6`D^x%C2yA>l_;Ku(I$XhL
zXxPSt{`Y;+Bm%mZ7c~ogW?j!`+uTgA7*$TQoh;BHRN?uXmRwJp`BQW2x&gmoQ)>iH
z9C~hLie5Uk*{iC^nHXorlNWyjzuh))5dSWjETE964NAhFVGU}57j<Oc<QdtsO{nBV
zpZ|_=1<cX@x`X|Z5Q!boY15een@U|ILdp_-rc&|gbN3;wvy1%I#O(hI@cWemia(}u
z-`-vYJ6e><_!Z@F9jP)sPVK=4tk#P~5Y#myEuW^F-q^C}OP$44da7Ny_ha|1x6G(n
z#7TH3M_J-OHZ{okJyQcBc_$y6l^7uk+llm8^{!Q6G#8xe?kbJ$<WQNqlhN+Vyg6{U
zC~Ib4`tPqtdV@_984OP=;MRgjF!T?{ZzoRoQoUDd_wyctsh)i0!7GpWf`8yc5C}qB
zK1RG7f2`6Dv(lj_<RL&g)pd$H6eQ)!54{3-U>AVdY5W^5>=n!zI0?J!4k1vx^Y9mW
zyUCmc5#)kEUgT+|V#??E!KpitE%bZ@hl<iBt@#&zp44}lSr-f~$$ww_w^_J!w1)Hl
z@$}x|RR90~_;EUPNU}$hV{h3r^VoYuB6|}Jk?I^Hd+)soAtIZQy&_vGD<Mj#DDk_W
z*ZcGPURQs0xt!zqe2n}3w(dwiiCX)Cf()E-&FbN=_qWZlpp5t%!CS=$d0}OS#b)X|
z`NK?oHC1H0hcM53tL1yU|9XBXdaF-A@I43Z*>}bSZgkpuHUX+Rnm=?O+abhe&jjka
zyj>XD+W3Qi|7|x{%{1m!^jj$4OAT~XXW<_5n{!C@#y{q%yHXs(Qq>z(g-cQgp*G=x
zaWgaL;As`^{-HL%`z&k+hK?e<t>L+l(tf!mQWurn2Z)sxam4jokCU;a8b4cm{PxXv
z48F%rdf%0GWdo~Ssyh2}DJX0mvcS75>d!(w0h-DlWHDe_s$oCDly#MEoUXH1>5jc^
zX%`W%f0=M$S2S)68v>nEaKDOWk%^9O4GrY^BqfLX(pcuuyQY7-H=)y*?fdr%JQUEj
z1rM*j&AnyrZbG?5Fm@T2QTIXsH>ACujjq`E&?4dJkVa${lnnO)RJ(A3)Ut?lh6C-$
zNFl}bw%?H*F*3aM<}j3(BI3$&w|P*f6kZ+|`^p>??WBl48TH#Ad18L@q#%MD`u^Xd
zr$R21>i=+zS1+CkwSV%40Ss9SvG=5}Zb^1iD&Fz@6(ko%>rJlQ%Tzc5Ldb(!xE2^6
zM(rJRI`{qwkq1A07=QSwusD$E{RCi$(c>UBnmw5KBBf#QQaBge-g6u+#k)y~r&OO4
zM5)m8{AV)SU9Mfe)}FPUOE&*Qf%>(=$o-k%!#??AqT(7Vlg;@&;kX8LwYAqDCOb4A
z$H#Ie>4zCc|DdxWrRXYi+$TM0s4-mDn+`=iq`}xzoK}>46U}%vP$Yjy=$OqmOy9eZ
z-TX=8pC~)eK7uN|+8U?0qIXP4-tUxI%N@8}R3`YQW}ja_PvPZC^Ymrn^v1OIn->fU
z1ME#_)*Z|Ec<SAi$c$6Q;`m24s^D0J>WtGeK5%4|p)a=V4qOB4S$_Tip*5ny&OMcw
zPkRp%_ulK0r%F?XK%tfu?75<}=EL|cjO{Z|m&`hiAQGqC3Nolx?#74?tGi+U_b@&*
zq1uMIEo~zeSbGR_7MiGys37!@LFH5p2kTbY6v>y9<KM6M==zS{178+>L1I~ImGKKp
z>}}v4M$SmhLdM6LOF1r&ZuXw$rT)c%Nj{*zas%9mO7DjV9ozcT4ybKm_~N-)Gu(M|
z!Jqt^k@C2)JVk<OQ&HgzYB`Rp#+rF>R*L`HpLN#hktYSx95%TQQ$U<K<5h*v^Rm|)
z9yQs#A<5%7-`_1F>UX7__52RqFAmw+M@pxB(z2qTH(o+L2u&!gL4c_-Ywj86X+A1B
zrt2hxu7r-#n&<AmPJUoau+A_JveNZQgIUg`m2(8gpAS@|8XBbhI|}tlI5ymV2|PiJ
zf8EZ5i;syB-$*|wkG!wUsjYpIme@#k7Zeb4c|m3{BfW(E5veb7*86#0%b+kw?|WL5
z8^^#85g1Sl`q=85C%Z#}j%u=zfE#tT+<z<go3Wl{h4<wY`V!9eLvJqJRgNsw^$*`S
z+`e*as?;Ee^Wy;}Ycw&H>Z#E^1%g{i_4@7$W@{><*QWH<rR}aG)>Sa=J<3*Dj72Ur
zG<tzXPj;z^5rd~k2e;cN)1{%s!q9Ehv3xwa|MHlN1NY+FP*xJxCQBWPyEQ1qb3#GQ
zKd*9uQ{?I*t~Xb5bw}MOD@Vkdx}F4J>q0-eV<!Rfe-pAk`}pd?(toVy&h(ew8qk2M
zbw>_a9ly2ER0}ug@w8}K4&vwcGH`@8sa0r2NjtkA<y9U+%v)Z8sC1Wcu)!Y8ZetIO
zkwo6<8zu@%Wu=U{!>K|)oGmjWl%4VE4kDWB1ZNxCwaN5{Wj<_AW#OrKik)Hy8Pd?a
zgG|%2o+XC)9?kOFSaTHd)yk%p39MuM1PyBT=nc#Iz?jC_LA@NIlp=we=U2bd%)Sh7
zuX(Qg_Adc}cTmpr>3ZY-s9&X{)6pz4j8sV}uw6&W+?z`G_7Op|Qv0XtMEi&;-AMmD
zE`6j`uERVmHM1=C52x&u;pdU}?^z1Z0OyCuJ#X*I*k?jBY~}Z8i?o`7E7kw5(7~L8
z&cekQS~YrJP|=YRsR(gX4~u&(Dwx_$F5+(oK2o=hA@_^&Vlfg=7`N6dIy=OW$oT#!
z<;brfw3;4t?2mgjqAH+iu3-vss!zh5BK*Os4n&*3fkS>e#W_r2T9qLFUs>TCfbjd&
z`iPDy*_ok>q|E6&cWmg?oR@Qr%8!Z5LEESz>{hlw`++TuE48RT!Cp{NXf5HiM3DHc
zyeB?G&p1ccjZb{6&{RhlfOx>;{?zf;+I%O7DVSSUgMuh8p9*WhMW@}HX#JeQCL9Ml
zlPeM-G6hCjS<d+f=dVoAD9mZz<>QGJ`LaSUEJw69d}WQ4nTbgcF~v`M9s;X4JDD6g
zBQ9FY<vk(hR*C4hvm^(_Zp_qc3)nYF+Y)oJYyUX9!aA;)^6=41bu1|VIVb4fA6|{R
zn;x&(OLQ|TsS1am&e(JeKL0d28-zW>8fF=#6GYdkPN;>WUxcGqt5bHLF<TQ8wa%4I
zG4?X(<V!5em?go8qK7}0dwRWuh7CMJiupMYQ_Q5MKN<h#1=46jLd)xD))bW}9*95(
z3%8Rp2P%-EcitIu%7rf9gxXYB@yr*2&Ww=Uj=D@qoJ5?P+yy2lw%yDNBzDfz#%A4K
zmGMEU2m?T_^J}axVT+r<s&hV)i+V)pea%MTNuUHPEn|{;5C<zoaT3kL{4Ss*%~h$z
zpPc;r&59@ste~#>x}>6FTu&2RDs#1NpBwIWY0TJb9PU>Liff&#^?tgJ9)etFE%Y~o
zLnlsIj>s|8CI=xt<E|WaWcyZ*s3b1`w0m~m3sSklp&EcN+Ksy9X^2Gct?Tb;){g{J
z)^<AnAo@aIh0P(sy!HE}mnO6X2{5+Ac$V;A6{pk{R#;-Rl|>%fH7S@-&{KvSl2WS~
zejE;1B?P_%OuCZOki1pym;MV8dp*hU`9n%sLhK>EJFy<<AgWmQ6KZy~R_fF4iR#VW
zI)Ktf6^*6zYOvrL!kXzcEHTl<M9Eav?Pq_6l{53OTU(PBr=*Z|&0<d3d;oJ@UY)tw
z?TQw!MXH$SQ>T6lud&^HSW(z77Q}_PVKl!dU^m)%Blszm%nBG(YUr0vg3jT@DkM0M
zS1(0;dP>Zu+~7Jz_MZM-!I`hBN|W&;;fnv|)Qnvs(jE{J{o6hqe@>Q&G*?tO&WnW2
zy~OsEXROr~<lEU1q{AQE_wO`{7}fh#*8wf*z3xSy(H9?GnA2GY=8gQ)Tg7+Zu{@rj
z()i>S&TMykvF>PN;boE=ISo!a*dr%mD8y$nLy>&{v)=Ujfup(Ag<G9(d#)#?T|Mr0
z5B@wUsk{B@-oB29%DER9a>cTYX#wQ`|J?1wp&$;Ki|5#`qKj(S{A~ueNj0D963e*f
zJvn2#xcthwIK9xrw<(jRu<+E}l$6^9bja_gGNb=B?)VyQZNa#@_(XDCtw4_J-EZu~
z<fjs&%dy*JkFOdmnQ;V+`lhRTlzi8qwO_XV^X@wF7q(RM4^r+yV<JCe!ZOCyCnifC
zEGsm5f&4i>!}cHa&@zeQ-g$5Vixh;0)jfn!+u_6bh8V3Kj|)~r%ThaDi6J4RIr~=a
zU7~sccWU2%t!jypCEdg69oUrSyBw!?DZ2hC5sDj1%}bRkK69JOdG=_D=<NN^@k4Jo
z%dYxP-+y(Dqxjiq^wPCFOyS^p6;t{Sv4Bg73ja#QLi|{mHig=Rx$h2h973mLogfwy
zeY@<p8o$j-T7qb!Mm~es@v7{Q=Ze{}4(Ocg7~5O!Z6s6|3tm{n{rX*ORggr3j+-8F
zi#N%p$>!WjJdEr1;z@rQ#~wrN_&GSWbPRuw^JNyA)<lu<w*&K>e4TIxH$o=b?M2)4
zQKEG6T20~U;M}3+Q37wF*nvYY{$@A317OXu5x;gJNb79QlFGOoY22vsyDTq<U#xSK
zep*1G9Q(s0k{-t&##!7!ODY#4MSA^;I7ho2)C+!6r){E`tmyL4bQ=HcYUs;wu_{V*
zR2?>d;X*6pn#0)(^PXYkB+sntq+@v9E40nM9U31Nn5}x_Gp_k)=A@U;J*!j5b#_<z
zLVpbwWi4GhYY58f04zRf7C^=LE+|t~Qx)^dxz}(kQ7ZbZG;~TKlHKVMGNei@T`QWd
z)kTs|ZI_p7(jsJ0w?gO%xxrsh@nXXM!5^<p$r9ziJp5?}kuPnv1*t)3>-1jJO5y9~
z(#ga^bc#)45WK~6)+h*(Nwq;vQNYw+rSY6)i(rCbgT@O%*`w(zq&RtP+Q9fO1(7T+
zVpeAHd!bjb8!8vTx-O;e&{O@#`P#kd;U5;H#In~9K~2>?$v$bVYt8$ZV6k9cL?VM<
z35Q<bz9B%)sf<R}0G?vNXTRFYOi2S=_bj&i=th7b%(2dYYF9DIsOlqWG?>D#^^aKj
zoEYyJF1+^*jF+IgV{mIVR1r9r`$d~QX-3@p6A-YJFl>@+uxiLawES@KRIl;b7S4+u
zx@X8*MqoU64JFl8O%eeccwBbF(R`Mhb!<LA^(wBJiVys*JEONj2z){jrBZdwsmdK-
zun~>2RpgxLU>0Vv@h(KXJaIrbn#+?PvqA1=;3^hkkaLzqG)r<1M@k+bunO~hKC}5m
zcQ>j>+EX$++CV1Tz|DL$3-OP9-iuV-EUHMVpO~L<7DdJ+0bo<1X(DYRcc8ZvqBF3|
zVb<i3mZWeTP9ti158Pjweff(rw7BQy^9*{NgJ%VrfWg93gG@clnE5q;*FQj}9vfOR
z9N$eZ6^^Ug-&RARGxEZ|LsmhO$!O`-i|Iku(~N7pvTuP%3ZkK~ay67yj~1VJ`fz_;
z(Zz=KZI#zjmve+=P#cZF0|%371EFER1*Efioz|W3An{C(M$1K+XQDrWYP|fc_z^o=
z1%oKQ0TW;k7$|&asI9MRZ7(#@j2^4|iOMNc<F3B$1jVl_53Zx9_Y$#z8{Dg>VuDHq
z(I}<>z-t(IjKjKw6-gBDT}K7%*<Ov~#oQ`lL_Y$1mTP>8pww3E#sg~Z{FhFFB`=gE
z0+%J(xZ~;<^-uCYNIxDDLZl}XEA@STp!oPVnDAFevorTpKWMH<)}?Rv>3(%<>2SgM
zn7w}T4`>6ZrG-bH7&m4$Q-h3jWguUr?)IS24+TThpqaDjLErF4Z*j+olGonJyG@o-
zmCn!%#xRMH7_9&ZHxjgZe(~FaX{y@NQYTH`9ct&Xl0QZJ9%Fe#FG!B>AJ%9%Nzxuc
z30`9m!twgqM1F47*{zCi&ob9v3gB8PK{U0d7(}wsn4#&~e3!|LGo-fkJAimM8a(BI
zj^woWD$v`<^mM1av%EnhlOF;%x^EzU*~MAIhWhzS#9bA@zK5%XPkvrKNjlX+$?P1-
zEcl@k3v4E*+4;R?Xal-9LPPrm9u?){neq240KPFR4v*x!QCGTs0|#TBe+TnDHC=l3
z7=2$FIQUt}C7cEH@V*MQMuHMv=c!xPHw{t_xOfS~oS9mD*!HNIaD)sR(40WJ7#X{{
zYY{v2a};Iw*6*T@?D=;OD-EsMpeT7^*^;t}9n?B+*&*0|Pyi5N+y-@j#}DOma%?39
z>*@n@vuJW0NWHE7!-_TV_q&X{`&D?(UXm8AN4NhWq2Q2xN1au?sMB-V70aV!z^2qk
zrvL7Lu(&1V*hAaYS&Zdnk$59g6gn@@Ob<Ylq&U^!v5Q2u^`KmuYPfyt<%G^hF8+Ig
z-^kvvuN#1<U(?zH0_^H2ErN~D2CVeO8m;3VgU=FbRqb2i>YqW($~bqfDfkiQEx|b>
zhWhpGga5yuGl^pWoo&ED{sEGGD78L=j>r%Z&L7dOGWy@%%Hu<NJ-i8NkzaDJ2%^;7
z+ze!UlG=DNY0RJsMA5?o*Fkc%c74LFBt-R^Hi55;{=;Ylb2WgaIH2!pPx2Fz(jsV^
zxTPbAq~>8exL>7<cO&F-TvDV=55glHFc3uSz^4dgsa+SR3q#74C^h4_G*K5rF652d
zJW6!&cK^fl@re7RVPBrQMi>@J>a%!9u4Yd@v}JHJpuB81J}!=8J3AuIRb2R*5>jjb
z0FlOZg4+jh)eWTql(iM$!CC@MDy->PV`3`i)=_~nNc>~_Y`29p`s0lxQeQ}yBE=Q9
zcLUl0i7D4@zm}du4A4&ASPP@JuvcO^{a@1nq1W^{4xiEz2(y|4x8bs?`TMvK>Kj?-
zd)EPSD|?}3ObV0l=;)=m(JG3II3O-=)j3l^Jb8PfQJ{h*dM6OrF^Z{{XWHorhA?&;
zCEqM@YN)hRbrIw_5yS*H4&_%~DM2(>U<U+h-Cgac;vo;7j&lVFYDTu~)tUK8C5rzL
zUpV*(`geH2aCRF8o1i1*{eEOh+zzn~<Y^#5-_YB#<;RirC-S+@;}LTbsKhrcuFALG
zVaZW&80<pS<dgkzHBdmodGHn7IHjyRxRc+ws~LZnL<gKFgKso?v3#eV+CK!K*_cLS
zb)1&S^?os*)WuBz>3MIpKY}*F@SBPNElO4m@2g$%XDqH6IA^(^lScW0E^TFK6O|ZZ
z9D$61FA8YhG*rZoAs8g>&oU_^geo!h^)gj*`qa>;H<*At1O=RW!KP85>hArAX#JRM
zp;1^qa)4q0*(zYwD(Nk;dg{$(we(;6e@ypXeR-J?m~fH{>4q4k;>lLOyTvNhHlkP%
zE)fq*$MmSS!f_W?Q$3EO87Cw;!Gm<_4x`GzFx#F%k|0|r0lFi$W5C&p!(<mFE(x9u
z7fM#mof298|8<7Jgt(VQo~v?1+AG9lw9cQ}+sXpXBDJ}o%HI(|HohuDe0DlZH+HuC
zhLW_#pP=c03L4fNnkJ!lf5j$H7L3^SaC0Mmrf3V?SibUU;>%=IivB2}e+$r_noe3{
z#lXi8l*QUW;%OrT@YOjDr+Gse4Sby4^|SXE%s2QYJ*tHfmR|c)H;fio065RPrT`&M
zU78iHZG~fhO2T#m@eOEw*A&<A18gguvZ@{vk!_?|?*$T8GY)1EvY#mQZG*v%!KVcN
zn;9L<N{EM(aG#~z(_d%UqBb`R&G|8Lb-)<{Bv}%cC(DdB*RhKFgR;+AW7P`;k-#ZN
zT@mnxrH5={dHkh}mEk80wVrK%++rZG199Nv*S%-h`r7LOBAN+F40xxq2+uf6)=${v
zozmZQ@FMBo3`Ju4K*)%r#P5c<avIN&#S9)#Bo*sI5hO$hDLp*y8wIa87nM#v@QkZN
zX>m_&5pGn++ftqjk}0F>fn_-uVNiDBKP>@6KW*8feh6ccyWSJuNP(-P-gzAJG9Zfz
zbrgYUeGRl0&$u%DPDoG?i&wEpv3<?s`NFCk%Q62PmwobZh2>ehVw%<1FYhal4`a?u
z7!za|0YWIV^<7dKZk%f<;ww{u8n7RX98%K-pbmbTJ^5L0iT~Ya-SDhHg=)$?3U(hF
z$oZ>sSoQ?MY~W}}W$ShDnF14=9@n{->>nL}-`;~a`{g)KxPaN3Sz&`Rjk%0jSD%GD
z9cRmvaGJFmeQ$Hp1t>R7Y#PZgtFYz&-bxW3QFW^zg3FoGr*gUmYs*j8k_qiN1nUPP
zwAqEMAOyL1H7}l;_`*8WX9VPVHmaY<6sw=PT58!ME?x*H{QJq=DNoQ4*t{yMqd_fW
zgh15S_(|<Bp}Cfh*OXQXSHLyaUxsJFKd^@dmi8UAb613>u$KQFr$<o;PsC%pKA;c3
zs2Sa(ftO$nFd@XKGm1Dy<)c+*;1eX=RgE=mZFE91zDZ~7?qOJ%Lc1^SMai5djEj&t
zHNe%bwg}L9<m3;?Fx)Xw%Sjc1zm77z1F#q!XB@xzK>V!%LY5Yxe7r=}@#Y_7ZG-ja
zhT0OSP*yBy(uo5QEFb~>GWHVb>^U2iB6#Z<9DN^w@drj?_w2>Vpu6%+{180VPNYx!
zgPv$G9ZR+YT$B+J&u%L$0>Zm2EkEVY&~q>Bx!j`zui1wG9wXuI!o{HV_dH?i1G{bB
z>RaQp8~IAT#R7s`z$5>fvcpk@m}!%%4Q8mqF#~)@?K!V}j!9rUo~rksZV<-v=)Yqh
zu6DzF0cA-LU=X2HjOkGc*gS$Lv!BnRp3jqdSarZkU>oWO!rw@%f;T2zR$X#>1c&P@
zR@^fhxA;K4q_Cok$fMSOeH9fK2f0a&VX=a}(LioK@nThyO3VYQeP05@J&B0+O*KF8
zgd@Q=_$5A3RD+7>&IsJnnae$Q(qUyB2_k_r8-@svm1q5PCsTUIgkVoaW(D*D822Vy
z?BBivej0p#UyGZ}BKMnb;V_k_(fSo%ATZr<a{API<yj3$Ww0Pa?(TCS<1Jd0HcX*<
zv4uX@gEl^?jR53t&C#@bU1&+J3XDP+tJrMP(5sOf{GV#fP#N*@H}EW<fl$}mzSE=$
zfAXXnJ|WN8LLTV!KLruNJ}hI;ufKqi_3jIe<G3mc%XQANX^d5`%4iGzC7E(w8Ev-4
zwkpANneF2`GiXx9Vm-TMd116Ep!7gBnako2@D={hXoo-SiS6`;dd7XjRGO5Jpkh4T
zy36(v6cs|`C|O5?iLtJXfn0G+HIPwlH<R2WdSfnWQT2k3MbGpWX)dbVez^>Xk&v>{
zav=~v(ilSM!7qkXOP}<ZuE^msmJZL)uV$>iLM`W->VL;1uM+FGyBdBk?%|yaHyQUq
zsF-}%Pn-pAV9>-gKluIxMhS_^00N2w+xl#W9seJwd9TXX$Z6H;HOYd=*6!T4z|us~
zBN!=O2@Sy}c(ZWU<N9je(?Vw7CgV9%*Z`z2c;W!6B`$l?PA6cWesqGgl>X+L-|!1U
zH^MjtQ$b<-EOMd<;j3Jd|B9%Z)wxu=E@&!dzLh_D1gIIQ1<yO?vFQT95lQ-7f=bq(
z_uFyW!eUq8L!q^FL)jv`AqnHrOCb7;BXMsWRn|MP43@NRd{9$#D<YnMh842tz^9+r
z0fs-w1gn8|2Sf#zDfn>qk9)^0BVSZCt)2gQ=KL<25NZfQ6(|r`>(R9<P}z2Gu{{^$
zeLD+?_*F={ME<D|P5B?1>KlU(wq<K229c02=uIKHBPf`k4(CD%Qrx965yxTuRb5H`
zUbu<Fg}olqZrq~OrwjndhL_8PFTQ%yke+k8!Fv>v7Bb=-1vpotd@(*h!F&`wVFl8D
z3>C9HL^9l$RhJE>b)2uUQv-_A&*0l!?ddn&V37{c<J;wd0eIa?_<KADKY`TuFlFY8
zk9Nep{1ZWZeb%pkP2GZ7To&|&1iX-tC#%jClKdCB0f+MQ$~xnG9eDe&B^@r1W`~kJ
z=rUk}8Y!@3{p&i{yC&v54y2GnDR-Ks;Xioga4j(p1z9rDEtgJ|J?RdxY}S&mXC|CT
z1PgS;hwC@Jd+cK6LPo-FoyEmY06ppZpw4+bP`d!K#Mef%-={1U(d4pFh6ba*?WvEd
zm_IhhA;R9kGEDvX>)rn(CFiIpo$z!!5RLm9E?LWE)l(k)U3PEkBQyn$f4$GM1daVT
z?c++>PUskRn2<Ca+B@m0`I|@ozT^6I>wLE7;q>={@KW6I*zM0|lwq-m0f+mWlL2ZZ
zZ^j@3E4VhYD4#%eb`(H!;1x1OPWnqFtoyrK`&m9DTp`>>YbWnz_VCY3>Q7#aQW=cP
zma`?lTg<oTVIhq!A<us%I;P;rZpX?#rEEy%B2+}{KsGL$M_s0$KoTMR_i@g&+v_h6
zMw{L~#*H@pmh#_jd;KN$??<QYwwdN>r^?!2Rx{02ZvS|TH@|N4a^8f>W*$no)j6Ow
zsU-P?v3(e3B*vPbh@XUQLKGW8+ZmWy$&=7DyXUs+FA&gxq^wUgBL}4sXsM(+W!mUC
zXo~Hmo=UqzIQpb0%^~mGC~BdaWhLs|G~Cq2EW7jd9TH!`YG6zg@xp}DuDn+4dIwJ;
zg0f?GA7pOnlHQnWg1Yy8`F?Fqo%y!{NC8BQZWeu2K-rQCc@i|xH5=hZ`@iT?-?l&g
zTYwq)0GE^Z3*!QG7#?k=-eFE|txR5#%%3zy7azq<B0e*aW=jd4-^W2XlCu$f879pz
zkwm4$%>SZ~2h~|+m02LO8{FrpD>VT{)06$7lBCw2lPmJov4N_Yqm#o6(^0Noj&Z{U
zgcFO@EW_oWdhpUZ%VBk@{&BpHOO_0EEY>|^4`_9@kHvOW>fs7o;#IPe=oq*-xNbIj
zX*O*vBWOAX`84BkHawBipg7xgF}`{GR?QgndDNMuv{aNtR6A7gd{xBg_auLD$7t@4
z&*=LXNNuY$aB%UBLM3|6Ns#4ZXj=Z^@1Uc}z-E2cO?H#`iLtWFFmXy|ecrHbsOKvk
z?<h8{eTb|u7V2qs`b-<1AH>qUNZ{GNc*n7wPErfTHTD{#`b5hj5`yVNhM;FUgbmbU
zWz_b#DaM#70-l~Oz8dc?<=PbIM;qIN*zy$W0ok-0klt<Zf>BG`yJn_7o@!5A6)j7t
zez9wxE3-?$*3w%+c13CAz?k6oozsU^KPY|SVC?q#nZ0VzrgCeWQ=yJRR%=dfo)>2T
zN3N#t5THu_><Z#T4oBG6exRRns?|?sN0Op<m^t5x#AY5H*4l^u1+2ryz5({zvP$#M
zC{_w8KXI@>De7b!!JHzN!B@2pk`t>=`J)$E!W2bWieGY54}OIf(!SqZGLz20Je|2i
ze_)=?igQ5H!CncGxi{Qhiu1htnBBq_*)87tH}hHREm7j|=ZOb=**Hk62YZ2yvI%_^
zIw$VnVdmT2gEvb-CXcM%hK(fsu8B!2huMy@-x^o8D#ZTX1=VOt#yk4CIn@MB)Ug)c
zyB=&NjWoL%m)$YiShK*z`y$_0ZVTs)NkVCraVgs)##FV!s6G!Ty;c3D)BrLO2xGf+
zrV?CkGQ1!Ssqe=!4-Vc%f`Z>|<6W^i7VRaAe#@c-wTrq%5<aaCOS%}mL;Lk=>uld2
zd?@A9a(9scb&q)2ad%Jd0eT-~sc!ZC$@<4tzk^_gkxjQsV>BqC8grT~m|k8FTu}%f
zh4tIeK_ZOHB|1g&2r!0#iVPk)4LtTn0dI2^4_J{mVBD8C42e-qfh7#fk&=jkbZ1oR
z@t=Fz8<xXH=HUt@r1k^N*CwX)k1>?S$OyXB{1ko|Z$(nfG`NuL4vXyjDgo5|Vs;us
zsg(x;?J74)TP=f#FR0n;J(Xs`(8M#~bUn68)zf%vjbFJ;tU(OZuhu|n1(b>cDp1oJ
z1Xz3sH?`Ef`rG<n**y&OElhusQ4o1tq{*r`)=^;=K|#<YB-s(v?NBorHe(t5br<2T
z?5z)*HS~*L1@)LeEG7eLs#h7MkGR6Ud^j9$TJ27@5za4WWD_+rF6^8P$?C&&0ZO9h
zXmtyFle_ASRAGAf9{Um>Sw}A&GP4TF%UN^bCbYQ~4rRdFD~}go`DrwleYNq~LkxUo
z1WhY)nxM(?OAR^5fUl8I&-I<HHtW=%Y*??^lrt_#Es*vrs>2lyw!v3$ixjef__75Q
zu}P|?xMp@4)Y{uzz11*iryZV2CS#D8ID`KNbU>8XooRC?hdNpZ0b~_TTz&ujqCtI>
zuo5DfEZ^tAp6!8yD^b9E&z5cAzY^;E152uvP{yI!TM*$38t<(PMeTHe+nX><Ev@<7
zUhKRM2|7rZ#(WzMQLt+ysm#)ddEyor*{QTBpyC-UndRZtKyFz|!u$c4&~{hO$Nm-!
zzXJ7f&T3n;HcF-A&VOkW{x0SMMcp(3X2^H+J;DYNgfrGx0WUYddTyEM!BB*KME}C&
z{2gjrDr)xeFe=uC?2lTkF6Nn>?4tnr2wnAuMU~L}CpVHqFDD9zEC4+<m}_Ns`&myk
zrg0ZYt+8f#3Ph|-G8edLtD>XX(p8xGk?zqFu*X3N`s{v#=dJ{i=^s*t4hp(!l_7Q5
zYaWu}{74L?DDoh!yYd};_bdKeHOK${t|U|bMza6OM*IkY@JCi#(k0h@5bosOd#5qE
z4vHq3{V|;KKP$0;@AO1}R6+2ik10xGS&J}~E>anO6KOzs>+XoyWwN{itO80{1%Z{f
zDu&Fmdm+2m>&0mr;nsn-?aFop@>8junZHM!;%}1OF^hCW^@UF$9IBPa^0Q+u;*CUf
zyWig07DKjM_yIVL0uq&~u5ThgQCS=DaO~kwY|Tp{7eHj9hFF0dhL4~)GVZ1`83zgL
z<31H+A3MIs97NqKIRM9I`w>OUWy-&%C$=l&!V)&gOHFGqCB7qeN^aD7D&{_0tsak2
zuXC9w0i|%Vn002+kR`oK<Q=CXUL|!r>4g9#C9@x6rc7e;n(1&+I%Xhn<6LN=ELU!*
z(Y(RJVAQ5H^Ii`@rvMhk6h3pM7mXnM_bc}D?jddY4ISA0IYsI<O=~#q8uMQPg!3&K
zRQ_R^RVx2GA>NXLJdc5Z{LbUQB&Z{9lNwjuOeZb*u034Dzgz#l^qv+eM5+gJXBBs8
zR18x#e#13?1#_6?<<0bgW85*s1QF7mo||q1>!MwTb_@p|@0Mo6x-GPVAvJL6FwJL>
z`6p&AT=>_ecIP}2lp`hFBg^)DiC~~?DYcIpxyAjzz3oENQ5{w-_HtkQLZ0V-3ce2e
zargR$I5kY89_l=O<z|UC(}kyr+Bhaw{YEdF)&Lm1w)T)zdmfT+B3_;UYf%$^zntda
z&(E*tn1q8by}Zew&Eyz>@DB4e8dg%Q!BW0w{9ZqMulOJa;b@!o;;*hdbU^UQESw5A
zl|B;D$_uQ1=$M&Gt?!>1;an0?LVP-VLQ?e>9Y1D24|yRznT}t{7~F<?aZ@OMN7g8i
zZ9&4gM0C}CNLR?cJ1XUv%f~=$WID&P1A^keqnna4_QskCD8fV_OrD!sEt~h@zVP%Z
zA<z52H&%>jsVz`}D)>|dC7S7S6JGUXLrW#B!hC_gSOZTcEs;xPLKkg!<g9rs90#b&
zKE-zqL~<aN5_TSrFqR#F!a>h<7pI{oWTiFl+ASp>s^`HsVz}uQ5Q;2yeof;G7lfYF
z=OmEQ9kyW-<T;UM%(@w(8)<eiHXh8++>!*4m_;?DB(HXs<alg?mLruAs)*lEk21g+
zNp^!)hfVtvsqz7++dtRz5=Pp=nWAf|^2`c8-vjj4duc4=9t?t3!%>*n!}VX}RK6fJ
ztiSy!Cx7_vNJ^#O-8NN*Sv_m0#k`-gAs@mGx}0NZD}+wsG6+0yfJ&NN^}g{LSc0Jn
zfJoSzt&WWaT!D*nwa+whIGvsOeZlbVE13uT!l)!23sqy1*8P=-rtp(+f#WFA1Idjc
z?j<qb!e`lC&Ximb(L@O5L;dGe7uJ}5`(rDUU=LUqdA^ym{|Q74eun52+n{L>MmHiq
zfwv?$FT!6lToEG#_yxEiB-O<XEn#1(Fb|`CfnpnRV(SvwwGe9ZTM`vUshH%3r0$b?
zRzyh&0MRc<jnl+kQGZpnpXL9(je(tNlcip>4Xy@sX0|H_X;mN7HoSbu;!+EAyEn*-
z^d)U7px3r~qVx>>(JSi=W({nYO?M?WHMM|`8sG%48e1IGLrlFb2!j#L@>=$Zj5(CP
zkQ62Exv;J=&SOkCF;qG%k*#&Lg;zUyA;JC|&U4g!lpmKeWaN#pmI?r=X9bWK^U<@r
zj)SGXCgsL9yO6Hfl%%<_&?GWLlUoV^1owxu)gD6PEda*RMi+i*!Wp2xD`1)O0O{_$
z3wJoRV_J+iESWEIZ(6_Lr4!DY>NE1+Y^;Wo$nUMblzk{dCq#>z>g(arG4Ahoz*?Ea
zTxsVqNDF&4M6#_&0663htKU%6nkx-(5H=MijxAt_;t_~>`(!K<8fN{{t@O9t88kjK
z5!GRZ`F5k)FA<CU)3vjAA_@>HW-_EBDgmHbMUls-Br0&y{#mBb-NHQJfAS@^-W8uY
zZ{JE=YjiKr@EB5TaKL=!-|Cnq2s%^Cxj+`GJMK!D$$hPyj>lMGr0=%h=42Uw1?R6l
z@`XOAi09?A;+Gr)@FVmV^b}_z?p%t*{&(ao-a-;F&N^7Ee)`5a02}`KStd{W;0m0u
zj-Ex5lw7>>H-cd&kd$ns&Hpz&w_&%&Oh6<G=D)377!Er~tNgK}-3o_ojl{j0u)5kC
zZ4d9mQ4}M+Sqfz~ya(>>NB8We(TmdaFp3Ny)|VgXQdHcFE2hFKIveWq;MkmCnl+Zu
zRGOp55{+LcoX0yd0f^R`q5A&s+4{SYcyyz79xxz2->&&U<_`^JD&_o(R@`pH(0z)>
zKI^ysfNK}R&zJibAA}e>!j+@n`E!`0jXVX-MMi4S{m~5KR6@>Tp0*FM5hG4v9@74G
z2MgjVG&%6xxLrHj)?xiK<B6&IW>Nh0DMV(0F_lS%|F;*ALe)d{T*H^l?lTZkg(>GL
z1R#SeV(u<O{|kq@Tq)hC^w)|;Cy{LRi=Tma!kij)hK_9bp0`au4k!aq3|=g`qIS`{
z8-^AV(wwte;l^|_`2ZnZ!N16`r~b0{%(I%Ymj{Pq#&$Y5c?@p7(I7<pkT1G>9fD|~
z=aM~As0{5iR6b2uPxv7}{nI{k^kx1ycrIZbCu({6`AY7QzfRcy?KwciQ7PT=;aVpy
zYaEcq-@wEigKF+F?UhChjI-GQ?0`N;9~9X;F3+;*f7^UFX{%9LpLq5bV4Q9rf5Eo>
z1Y@j~n9A}UP@jgJvRih$C%%eew29iJ6r9OA3i}*RzpGoY37>qBgzJDmcb161YKO>-
zB+kQNA;komRs~;DT3%DWs!W6mu!cbN*qj^B@>@&yRku0m%FzrS`l+8_O&eX`zZL;L
zKaywtKnTzZS}izADGR&@nra@fr#NN(pEQ)#dRbt+u?1`yj6x>5sWrbb&uPsrf4&#+
z4vHET1ssZx4PPtu{GpVGqVhrk$j)zY`Rrl_>bd+(n+@R{qeYGN6N1FNpKn%{qvL0<
zg>!AK*WP)^*Fz^W91cLvHXufiH5XwoITgy4MeyBjhmOLTmv`uWKktuYpfy;OLl!j%
zP#mW|9g7uWI=3yb;tt`%J9v>IILBehFIjg9zB7VaJDCMx*HVY|wlEDRyeDgJBHPc<
z<7M?TxHs6Bg^7p1g7XLdxOp}MvX^Er87#=CPwyLIMytO(R^WZ+>JWiA?t)V0>RnRg
z7dy&~#x%q*7xi_0C&I@uTchAjwu2DdPHs><Rmy}PJ_}cel@hBWJPRdF{Pjnsq!pF2
z7WGR((=Q#gPBMn=6Zr*{w%Aigs7*|F(0sSrvlf4MK2LQ`l0xpM!*O=$?(kIzEwuvr
z@(#bcV&olc(V71aS!$#t1%V$RNL$|+`CQXsPA2qh>q&%a`5*ytxnGJYY2N4xuFG<W
zX~ZRlFhWdjE&GKuCD8*?QK-PdhV^Z&Yyqjg2yUNq?W8Ck+a(qaIXi?SEs>|bi4Nf`
zzlhv&QyeDy#qAiI1b$tq4exVd7D*GMB;GF?gbeDj^x_dJyCDiNaE(HGo%!OUR&J}}
z8_J7OBBPQJ#a>Cp+0KlEbNuo()2%MblY9M!v7k3*duHzwxB=4_vIC#0@qGLO0B>NQ
zxo>$zhGQV&FC!52KY+6RVp#xa+yDcMhrw>ey<+sN2A;DY(k}ruQR)zuVNWvl;}inV
zDL5Ic@uGga_P_K;#oi#Iw%^fzo*!(6WySajoUB7Qt9dB?$+1sO2C_&gHKe~J)@LnI
z_0W??6x+_9dhKo8R$r&HdFLYZqGg|oWNtE7Lo5q`1q4wrIFD9FzQ-w3c(WKeP}w3S
z(a?{1L4O8PP%Y9R*>S_J;z(=<4|4Y5iqBvQEob&ULx?gofAZw#)b+{){jSw)1WvPq
zFp1(H8PI*xaI&(AB-L|i=tH~X&WTS5V@?84)2_RZY0Af&J@*0%t_(eimp}N<-7KjI
z*{L{iryIWU58mAb<R3uA<nVmhJ#_-oXxtf?zU`X*y(H#p4CX*!#)QEw9Hpq94)JWE
ze^c$IV|phJ^C&oaafg8Rc74!4elADEHzr>)9*-0%+_~kZVN2x|L>S89cg1eX21F^J
za3^x*!Qi~N_TPm<swuv0w>l8Pq60oSF#rVNlE{XmX<&#JU-RL#*7`t0XccY1KKckY
zAmGxnc3?o?W}*N$Q@Yi)<{>rI>|05838FK&=c+qv1cL`SgX;C`TO+##<8Kk_Gb?aW
zfnnuyKWl6&urGT#9zJ)*hUI0lBs>??ldd*vCiFap#PJ}Kvt4iNpsbut1Z?gr?5{Tp
zv7MhnstFvW%6*4x@VnjLgq$UaOhE3$ZZ2c#cq-t$m}?OyiG<bN5uDh@hvzM@mRm6d
z;~vaU-XRd*kfbTZ*Cp5v6)IO097?JM*A2^VJ|NvG>8v7#Qs3zZwTZQ6`e|lQgX1Kg
zibyPT94gEuw)Sc{#x^+{TsyN&-WU5b$_*(y)bLWXV2uFo;li2VeYm%$8FCa+v+tYl
zKNmb&!`VuGN$PrZ`943#LdD0d`_twXtArdV0OzGFZ+l3K;MgRVe^hm}u(nm%4<rDy
zQv6EzFIH$+zD7(r*LGiJcBNWR%284h*<INMPHrpNozCx3KdnA)h3oXb<^0xjnM1fm
z5~5=-5$QV#-fWEf6!<6V4O_?QSLYrfZz`JLbWlhQyqX@kOUx%CDlQ4QG>X55GGOa<
zfSc<o_ocrv2ny$|8$zHA*~hz}W2LCSX1gpnzvZ89<DUysO$ZJ}Y_hl$X`B6}xFiiH
zap+>_ldQauX9Mq^ZxDJquv`;zjDft8kLmON$!A6}@uz)^o`@wK;lEKcxQ4TAMC}^(
zRHU7YpmPuN{b9?%bL({c_^`aPr8BI)XW2gIZ(u);d$SkP%Jx??x_)ex@_uSQ6V1bb
z6EK{wbJ>GnfUG-%+@&cHn)i$McIhKqH~Y#CYJdwAP+h9{sW1+mjP4uH3jV}S9*3UQ
ziraMiXfTn(K>4NnyD0R8IR7)6$F0<$Sv_5k-Lke7R=OMIk9<B|VfjVO?ZfS}L^u3^
z)q(pz>&KRQUPuRL&-t<-LG&hzl3;$mon|Z(Rh1KjcAv9<CARnAxjEDGxwBq?y{A3R
zlqJN!*&2)7qnX<Tzi1>6y08F*6sqj=C--4503y$f3Q@(e$y|uco8*&Sq;P-c3)r0s
zvkb~CWZ&N`XbZo4o7)WYtdeK<b|ou&YCTs%NX#2(qj|g|!)||o5i~|8fX|(AV|9rP
z9KTSl;tNf`kTP=|zrhni-DW+>%lOI{OtCwp=fMhJ{$%mEyxsBG^A&?;0{|I&uHQ-{
zv~i<FW07cYI*YhiW2}n0m^`XnBX>#gM}kpA36sU+V*9WTk+-rz(H0@X*X!X*yoSNI
zq{kK|mtx|0pszDzFZl{$?H-sW7L6&?61dfV1lL6@sSADGd`Q_f)vpgA55k5miXEAG
z<!RZk<5D+#gVe;ZJdI^snr+(qEU3Q@Ytm0Eye7E9AyPiAoGLTiAX1E+0_u+Q7PpFh
zweb4pozvd_k=KhO@wd>Fg5L=Pmy>^i8MaersU*|W@Mff$q~;F>_%=sVRteJ%NT`vx
zJMWEzIO|S>n!pyzz;8IY`|550#TKW6aT6FMsboUh-p&?xJqZZH@vz8W_3?lsbT9Bm
zXVzwMI!@PqHz8(3zTrBA^a!S#`QL^!o3w(S`|>=-XX+4)>6aw>Z>%Ouj|vl&MZ6@x
z*%OY>M(Ii}q3V>}^C=ygM~65=wt)kxHgeaXOYMLbe(r$pTbU5&<%?^B&%a^tn(I)_
z#m=W^wvmV7p6Eqgkev+fe}Dp3Rt;~jlq?UL77!c6mG{4=KLSJ|oO`8`JSWN1@mTx#
zaHn0}i*Vds-RoN&q45hecJ>`Uu8Cw@JI^!2ES9?o=p8dMfCc|^4A8qS)696WWo*)8
zx><QnMvHhAQ~ue$V0Ka;Vg!cD<G(NhoQk6H#PjSr?1uM*CGf;a{`9C>{qzyR!p*zj
zm0S9*u=Anx-s4xfh)j0NI|J;q2sLh^?Nr$~YU;w&<OT&({wYnDwOpQ>*{?N+oL8&n
z<T(n0IV0P8X1HPflk5&f^}JU-0)<d3odUc-w3${OkWapp5IcJ@bqO1ZqbFoDZ;fgk
z`EyVavcOE;N$FD&;wbM4<s2@(y0`~;{p<1|CIYihc;sTlsochANzPJ4_h_y~xP7J9
z&hY}b+rE9RM~NY{HH33}p<GLU4-Sy;{E^w`zZIj@zC$*HvFACt{kTsL_~+oK;HvT5
z#QHud4=0PlImL6n{>s&w|1W}d8$vDV++^!6DbO(l+$4p{wgfOy*`P2#R{=usJ1mlw
z-)e?cbmu-2FV#JL`|46L{~%pFK(|yoGyz^D2Wj%SYw`0xUJteV5ziOEZ~I#oX^%A>
zIT8s-zx5k*g_7V?5f^lNfP5AP6g^F@nk){PMJs=~!o0ROeCPg{;%Tb1*U*p{PsiRT
zjT(j^Q$kqADWgj!DsV8+pctPYR-Ypb=rDTaYJ$08UqJhr8<|5X8ODD@j$3kDi7KtZ
z%QpN_cen0&#RaG-Un=1FpEC&D?AMhr=NkHeVI+r?f)p58sGa9$wAAt8NC)Sc)I}~%
zBW01v`4htNFT`dV$5U4>f;Qv+ri0_LWSj7w{Qj77_Wo=TY^(4Y0%ieRB}VqGB&NF!
ztZP7*Va~rb(v@PT*Rm%QM4bGYd7T$*n}l`%TO*sQMQL1^_8|aX!A9m62h{q$QO`n2
zN~Lyj$up;OD4GQqvtHe(O^p`zeKlkk-6d82g(|}ndbFJsMByJOl}F;L8nGYipK8MY
zsV6GH!8y??68awP+ihV<!C9$GaDDtpCV&k3zEU}So*KitSD;Pa9LqoDo$+B%MCj1{
zL-WsqNP&n-3Xb*lhgs+1E9ZvYAj>V>1Y2{6!pRUj@%ltJZI(X=eIu6HOEiZtN2AK5
zm&OB#4;A*W?;F0?^j)@iX|Hidf#x^3t29lLc;zhPcmjsYBG7JKUmN%G3H|gc3>WJ4
z6fj}A-$tmh!xVN=(7dG{^)w5!YK?XA5H@Byhl1p}2a%FLJNiDz)+Sz1jKt~6Y>SpK
z`39#WRQ&`~J<`&i>~(gA`bI<1440kX$J(=o6ndbUt8Dl<4}~4G3D24Poo066dNrel
zWR_U|Gw<tMV8U}=5>m=euRAAseG#fD5piCHk$e3tq4tV*=vV-}XBMuy>Rc>e__)_L
zhTnYrnAa@%YiV*m8V6sM5qVJA(q4bwy8azKaih117!_KW2y$~A3gm>w(g1o`q8qq)
zCJzqIrShPE*z*oGSEVQvhgT1+8G=GeAt6!Hgbs*^=mq4XGk2ww((`_He5e4bmZ+*M
z(&SA7QVWx$lGc?CwGCJJthJ<I=cVc%NkEV+ll0ATPqPp#vxYyVMr3NBD$4>?Q1@E*
zAa*0x+ZGP*ER{=yX9?JW`XXq4e6BWRy<9GS2vP{f<m=_eWzaQCSvumHWyiN}vubb=
zp0O0QFU)U%X*5nEuNDPatf7BMSeg2vyZBoC$&s()_X@{Ppv7`04bT{enj1H}7G>gD
z7J*5xIewedXx2@6J<Re>F;7G=&F?gY8<a9DQ6f9%{b6EHRokx{N{S;{Qi+sUKLffR
zeywm!oH)_Pja0!~Ils0&KZNy0Q~>rzulJdi@I7i!pt`@1xo%6=d3shOF;D(rrdI5-
z06_@W!1U3QkKN52x?G%FsYx6x7HL!%dr?O^X~)vTX-Md}jJ^40?*$aBMGKHgLF(zP
z`vHW?Tp?axY`io31;J3yMwOjiD;<W}fz!;E)(9AjeHCAMZ1&VlBMKx7YEdlOA9tMg
zf6m`nnCHIAhyq=KKhy%tPriEGhmgobNF_+CVZVcWEq@zpxB=f6{&7<=|GM}-(^g`z
z%M5B}zrNn`jNcwUUl=7kZ)%#G->Ip?ricc2!+8e~3aEIu$OJtME{dK`Ac)6t$B!s3
zonh{3v_4Tn$X_;=`rz>HhZt+1n1Fk?tkS831KaGU2=B3ULu#ksjyhxlTfq1Bzq+{2
zXNmU3L$WMFpv%IH6-r1*|K=O=f`TiLrC^MJ)xsBL=3tO!-7u+LAmh;`)O8D(v6ct7
zla5h7z-9~i+45?d8C%Ve`uIV$38NXNDg$4p#MWWWl`8|`rwuvqCQ6V~VARepDxsYt
z%Y;3bb1mi==|R;JUW=2T2Ulwv<EaJ$0TljXd`=yiMgyIp45$n!6}H3fn(Nw>7xNf$
z)}$a5*@fQRgufNgSzhDP#qca&Udqybz(K-htr&$y!Mh1Jb54IDVBc_D3!$DI{ZKRB
znX91C1jvA1C->Ti(uk~rxlRNf^{yoMHEI0?4DO0piV9u87*$6f&%OJ)#a+%%Z@?!&
z3IXn9uTwCiLwU<D_Nf7n)W-Y*X+M%zfR&FGJ$T`F615RVq3D!+m2`=*n!f<Ymh@YO
zUnsLiul(!f80W?O0oU5YYANQLM8-D|UE*<$OM}4cr4hzocu8tLiXL^$?MS`#WB}*M
zZ!YivL@r6{&O@bl{6zYgDq*7=Y_uQnc*F~4Xea>(AEqE00N<#}ymk#7`eCx;aD;;(
znEJ*6$U396XKK*)VTdX}5Kftg2%j^l2irFnD8@_6w@e)T@;OcA<}nq%1*Z5m>VYKx
zS({L*O|8xAw=EQlBNcF(sCr1c-Y6a59w>}tQP^P~7B^QP5k71B*h{W<vnJsIuza7h
z^Y5V10f1L$Nckx6tTeb*BNr?C-HAe$N*jcWZb*+;)KvG>b8Yq#0@uFe0Z^+}{M9W1
zwWG(vW0SxTy805X{WaqW;G#d@)7*e{ULbF=U2-o{0H+(&5lDQ7@$%JG#hPyiJ0<aY
z9G*5iiY7(rP8_p5dQ|<CL`v_hr4sUolReI(N_4X|na@}yz+afwyY1KXbmq2?%77LH
z{YPGp_@~jB_*NDn$RMmg_q$|^GPN1NBB-ACV1@0Z5NRoQOaC+tG1hVl^@s4O%dX`M
zciM25XLwFL?`61cziY%Fv=AaNR#ZPEP!lj{rip|k&+(ozYC;hYy^aVV)gVx3oAMX&
zmXo1wJ$~y1LpebS_eSa)6vBfl=7UE6l?nD1Nu=d2Uk29rp;XwjxZT#{`P#)P;a4;P
zPruJEevRrz5uxIRiUg-F<8<zJQhmSj0nfYuvVy$&nsusnKuSm(8zIWY88%?5)2)A3
zp3(=I6kJzf-*l5CB{=%vcwegUU${Y2?E+|rw;uO67G6{nTEx?oPmJa>AsV7PCB|Rz
z7YzuphOCCxV5CaJzrVk0c?e_69Bxr3!(knFnFu#WU)(qHv!!~Z6)GKba^}%}QqHW!
zhsOi<&cBpwUI&&&LfYrsH4(las>OeopZ`+X!-{D1yHm#1b6rSc9kf;&jg%2+2MtiT
z`n1g>WLo`T6xX%$4MyNnS@KCa?iYOaK^d@G&U0maqD#OzBI&o7&x|DqB8Rd8Cv1h=
z>#w_Z;*n)~%S)J@Tp{ixfcp$gTHGb%Oa4KMFWtq9gO7nz+^gzh{=Sq*E8#?|3K1MG
zP*f0#nEv6*-TfhgPZO8HV@Q8kTEph*NKi;TXeW>{Hq2Ri$47?{=M&zEtI*+aAPlMY
zHz<psw(dLS43bK@hq(RHn@zg}rPZQlHg!N9X?2h!$va>lWht(@I2Byj(^aU)`Z=W;
zQOuq|1W#^TG`|7eXSQbq#PE@t!aU|}CVA$iNzuMb4=%R%UM`PV>>ziB$&-A+g^UkD
zlRDeF47#)LP{e>DPq9dLO(myRQyBqIok*tvm59K-wh6hg#Yhl051)5@Zc2b+4Ad@R
zT{^l0p31S-ROL$Dbu>N%p<jgN_;qE(bH@VbA1n(S;hZ_@7u_yP^<DO_I0(aApy2HZ
zPJaUU!KN)zT>ut6d5cO@^6OoVAfP^76ts3zfK?2Q=1XWbOAY^DdE{%tYlUm|-JSWW
zIYUYdAq)_#l@ff#W4wU4(+qM07Jy3p_^OoeL(tT{1@4<Ke0_ug9Fxg1i8zPOJ11aN
zLav3H^*&Iz8HttsT2agFg0ioGTK@|~zpkn$H}-gGn6>);fZAH0+D27{F;Yvg?H;hz
z^Ikkx8(OorKeoxd{f=v0U2O3un1Bcc`6VjdpD-a#O+yqu9ZXSwsxSexj6bDYE|R=-
zE&mTYMJ40Subg$=J~9;4GRBfHTnHkdvUGW`+-8>V-K??2uBAO;Vd8yOhpo|0qP9Zz
zGU?1?9{6{_eh&X$#aF*<_yGd;1*lauj6vga1}D&H4f%)_ag^K~-+pS|i^r3yuG@x8
zJ7N_s-aKzWogY~=%@wY%<C<4n*%nrAm#LnhMz{hA|CrAr-S5l;Y$fVhw+%GAyBpQO
z3Of<~`v#BqYsIKep3G_)chO%EL1zg%igwfIP6g((90b-CR;0_jB2rDv5iZ>veEN!m
zBAiHx1utdYXUN2ubMo}Avo!gEDjMC(7GNtv8p#m)kN<rr;dAc-xjJ~!;`5d~vfIxR
z;)sA5{m0p?LrvrQWijaFf?}?uK3CMv;w5qSG#%s(mNQnIz~@8mUU~!3VTCmdi?Axb
zV~Q&r@`i9Ot$##J!-VMx0-yi;hh~Oe9ywcDZa7aUWdQJNth+rE9nt|q#GHJjS)1QO
z;b!H#M7;L=h!E9*rlR8J@a(zzgSphGBw+FV9YDl?ZA_?<wP+XCi<|&8^`1Mei2DEk
z_3$~>Ssl2Ulb=*XR5&5j=FAH{GvV=m)&PL%K0aA|Q@=o{uUhikKw$wJqu;LHyyYJu
zZ0dGQsenq(s9nAUh|l3l-%Zn)Z(uzs(nu`7{$f<%J}8jY<_{14sPcR(Y)>U|<vTz>
zPVS9v{!c3L|DGQ;;p`&1m`!$x2-%0LzCQyMAwzS9)5|VM^60&~4ZYWhJ&Whsi0QKX
z^|!tDTM-$3kLDEjCXMt)>^NLLjnZ#iknk*$27yx_E6{RB{wsh`&ZgUfJL<SnsVbq6
z$M#e7O~DCbDCwrDOK-f|1aUzHon9be-{H~01TjLs$3qj@YeeZdhMq48p_8DqT}d#7
zQdxF2zRhB;lQ2mZWd6)NLQ4UO+>!wge+1L$?g+L1hG}VfGTU{4zA!z#)Hg86E;K_a
zbmnh8P|?4^_|ddG$a~7@#cTxOjsOawW<wJm^3E9|+V-uS6$TgfXaDPsd_n6(n^SM9
zgG<CPdl(p|QCsZ;lV>;3<pfi~P$STKX#tET*s}F9SYqxD^#}UD%ddct*EoQ#9B}4m
z-3d$fPEP81aHT+2-8N90icdBJ<(XTm*{igyN1wn{S@?K|%47JZ8-VU1HhW4OgR$on
z2;MHQ)DFl14_EzoNQ;B)ZJA3leXA(CFN${ospM=_cm}Vkf*o25;%Sviy0_m0f9%!*
zDGe~+%F)vA<^$s=E|HJsy{Ki3Q0RNnzcowloKtHjh<g1_U)u_S?Bmz~+ElLB?gPPR
zN1<tnVW%Z41o%5&NOQq?*o5TV<*bAAdZPuswr$OpxE7zJ(OZEYdXO|lqw&(5%Y?&}
z)<XjNIdFj$lngs=r1Qd6X5g6HIZT(ICC90dcXb0b;rtX%r<cMu*SekEW~-N=4Abtf
z7k#<AN<W|N6zv`<8}LF2%lsV$aljE>(%%ocLb;GI%7s{P)%o3=>tW+c%u4h(0J?wS
z3}|&fV$EMgmnU6$t@%I20u$R9NO+(dv#-W0I1zwGt4FMaET4|@r$P$trN3dHSV3U`
z=P9iVXw=^qY0xC8@i9Gy$0sG13yaawX(}#j)P0g&*K&(?iocA&oVR*|Y6pm=ri10k
z4};SokH-Fee)|F^r|sb!K|Y9=W0s#Hr>7jNU<&>rC3Z!0rL+gc%_q=r|0!-Rji8X*
z*B=r8_s%N82vvZt?x}q8+CY2vIGN%0vaYROH`_9Zpq!=%gUbl$HFS;{zlUoRm93*x
zfBs6(rYJOgq>P^v3yQzpJyMld8W8wTrTaYbg{*p<$NR4PQ0_xzr1%ArCVv150EiM_
zOZA@dJcRmsm9a)beV7py))zeQ24qZtamY(SGmxQs&A^EGo3Q8hx<?B$3B$~wK0WV|
zs`mN0*>_S}ft!p!n}6RqxQVR2=?zwozr~lyZmg!JBxJ#^G-rhd<~yt@D%^Y^C+kz;
zqFNVn!3OB_X$`o4NRDQ(xiGMK7|gHVD}?<bNNWB51O`vjap`xYzT@N+eIIJ0+$|iR
zew%TM@o&aiCSkTFa`&bPg%5R;pm)2t4M^q9x>ATVyXG93Brp&lWd>^4u^%>ut1`DK
z#4G)ObT~l=4Z7msIviHuKA8iJA#&@a{~L-=<pPbtEq(NG)c@DmTZcupzJH?(3_3$e
zmw<GOlqez6-Hk{Msf3gQ)=(ndDIwA#BHc<iBB1CHA{~lIiGU*SJ$rwD-}imbALpEX
z?Q36qZ!efNYd!0E?)y`BCqpeyA$;~R(tsLHX4j*=bv4r^l#Y&!i3X@Y3X(Cqh~iBs
zyQ(ZZxt)?d43qNwGEreU3P6g5mI-2RFlc9k5b`fLe`FUTxKQ^xY@C5N>HD<`-UAh;
zd07vtYarLO!cyDHRe!>~)fPUZvr0&VZ~;5PS|mt5B3&*=>m{bv{mIyxR?`gAq|)=n
z55V&b{qrW{(lwE$ZW_ciO4z0qx6u@tcFuze3(>avwWQNa!<uZMIO8nvqoNG<qPBQ_
z_aYtN!3~q2L#aZmTqp7EcnVPUr}6{3;1iypS6Ja|9xRFASE0j@M380;o+cyQ-gQ2A
zFvt2}P@rUKC_VnzPt-SdChOFrF*tqz2$DBaqx7N@%7Q?=3uUiyX`8S_d##5`-dr=B
zm=~gx1~&O1k<F){qsqlY+n&0`ae%sjn=`{3G#nAupCM(vo&SeBkw~DDQjnc~9M(De
z^rbd&&Nc73^ey@w6;e%g(~D5sC2_a^macFy3|G~hb$jOo%HH?%Nh-+eyzM$!Ijxl=
zcJzJv$AE-$(zXpb@yotKw}L1dIow5XdXbiQD?oIf7}n}H4O1#{@uD^m_>HO`Xjp~S
z;xReXXiN+_={35MQ)*V-*m%4dz;L4WyO4Yk9Cmr}RL_}*F5aaz3;-G3t*mS)H|}c%
zWqb^)ilAllpt4&1^ZU!m{v$ytQHiL2-q9-jeqN{>G;vY#?~N~>1F^wO(@s{7n(;}D
zW|Zwcd%*RwQ;R-}d&_1qyy-JVHbfFB!HzULqvyIr=6Fjx4U6*wfNml&{7{a{5#-<N
zg&Co6vSLWI_uTXjJA;U;#%}LKmZ5CdhuxcLvo^ob{x($wnwa10eIJW2nHFbxbD{tK
zEtjD;ORYi<-!qPVhzvNVE#ug(9N|yw65QXf{q_BSrINxx#Dj>P)(p|Q&y+PR%Ecv5
z<@Pku9GkC-7sF;ew>M|Y^qd#Vul`c>BjiRekAR2abzv<sfnWgHXow!!1~?v%v8X;I
z-x8zan;}xpieCuj^`y8y4BI5_FXj&liR;Lwh@1Ya!rsx@4y=yJ?NGxH%X@;S!W(nj
zrN+;EP)WZQ`cD!%!N#O4aWoM^$+GIVES94eCz_GNG|47UR(Gc%4iB0$J+V{UT-m^Q
z3Y*uueTv(6Qmci)%HaZoHT?Lo!8a2<)Tyb^l=Gst@<WI=VB`a&A1E3qeawh-PlycZ
zOfIn`UHX}IfKIHlX?u9Zs_p|sVT+HAIFLnW=pX+2c`>Q1Ry9%5kL3;_8CD;cF;}t+
zk%f%(e@2usRw{W0<-a0pWMQXn-wTDr#n4zfLgO3*(<SI5`mCeRvHu=S7q%QP(VvMH
zY34V1Sv***8+4jHOAk7S8d&6*;*KZW*auVJ--jp6#+F{Q!5-iCy|9uT#_Tv00$Pxe
zN``f(jGPwAET-b!5Nt51+{z0``sNRJyQA=&mZ~z1VDN#G)f+Go#7r=*?E3byo0R3Z
zHZahLIXiBMDhHuHtm%b=uNjr-2wjE_U(Zx%d*d_313NbXZBeSNFar9GR#&wn7UJ3Z
zh_~T^Q&b<qSw8{>4!wZ4e!ok;!Sg6_#fzkYJ0DeNcs^pk@xOqPY^~=UVOmL4y8=PW
z+TP}bHRl{9yYLmO#Rs5Nx-6b2$e9yio_K3uBoz?1pxotFtbF7B>N*~J8Fv?^+nFA=
zZqNh+VXg7J^4527bvZ`G!8`2EiyvH}6J=l|;d=J%tpuBASGeQ%J)k@5ZdMah$ruBP
z5w<b}2i!d#a|WKVoc%7pRg&tt+aC-89P><KuzHvYnA@d<?%g1QyEFx_rJQ87v1}1Z
z^PHKnG;ymyo%Av0S`I?8VdgyH=XVeBAyi)*pT6?hOCd;p4WxJY-^hxzrl)V`)4X1`
zK}M=vQBh~80)eR^;Cqu{W6!Z<ZXe#D#y3zK1Bk_|TJ)=3c^$}5x*N6|j!}Xv_iVai
z>sLO9qQm(ZO-o_PF(Ywp`G8C(s^u99&PnckI3G?RraT#+u?T>GVjU<_fY-#^yMFJj
zFOsUL1gSdPMm%|uaPrG0>041VP_oKFIsTdD%+cX_Q}pw`eb{Uk@+)Iv_>}ASy*wiy
z+W!8$@Ra02)gpBUgsQ`&1y*vZV*24@T;|9Kk1TOT+;oJNPAb6&X#k?2rB88S=G(8g
zp~gIvo!`(@8j_y7alt-Yz_yZdAWbcif3yi=+rh<%chIcH_Uw4w_^b38q?6F%E#lKw
zuF<MtLA@DcGCXO?5EBAOMT^TjHqOyZeV90iGtF$fLy{e1GEUE<tobP{k)VpmfD!Km
zz1)K=|9klwKr3%E++!HU2-EFsKZytu*ra{yY}FHU{p>Un9~Z>cOVc-y$zSPba>nCs
z<`~EkbfP#)RQ(drgrIikeDf1RPdo^*tK0dRkwz&cSCFW;;pDWw94Z*PC(<QXsx_(M
z0CPlb4LLDf<CM<$MelJlk*qeQg2PNQyg}ge!hBC__x{>g=<>M@wZoy)$5gV7_Tekf
zOag_k<+=tUOQ%D12ny6$>i)R-(@aYROxTO|6nBDz8k{jX$7P&%dSH;<s4n)>XId+U
zvqlC2c!`;m1~nN|KrzP0{o|U&4L(=WIO#Rl%Ad-&zS*-{UM)9yo&<W_wFF8#;vE>?
zeDU6<+IJQfYC`)LE0Cq%>lfnWvhRqp`fO@Tg*t^$!dB-x-hW($MXME=MDL-tDguC>
z+mRY7J0pZ(WJ)|kl|DX%eVf$c2<L^e2AMXRvSY+(z=(nLQw<1x;ZoxsEHe!my<^w*
zr$jjeqJ*Sa>U#eRlwN%}Zf&O$3$lv)8#q&-qiBD;U*1Mjwo!gk;vEQJsHT12=<TqP
zuuio7Y6SBYtf)_W#!XC0(GWo%zx4;Q2GUNYiS7~AsxD%CW0mby?r(~IFuzS3tUwK!
z64XB?THiq7ER#TIQk2A}Kxy$YiM1~UblcSwDSy7x+X(y%^@*zneJ9)I<!P2@t8_<(
zf4yYPW0`#6N8Z5^HEHOX-y{x#PTz&_qer#gZ`m8$-znyL^Z9(k3??w*vQF*r8*9&t
z0U0e={+o*QYdDhgzqNhs`1abN7kn(%WTpH3z6!esA!v8HaA6TXH&*k`9xLGo##tES
zO&p1ckz>x*C0YdNoGq3)w`QQI5@zRwEP=<5+3xmxwHN$>yrDf=8sNldLz1pq65=&D
z#X;^h+Wl<#M3fODf9-Azd_2GU4T}>9=U2tK&Gm<#Wk!VhzWEp`F+9Hd?rA(Zl@{l0
z>KMbQjBb|H&z?I{;px#?^iD6M^96g{J0_$1Lc`<v$Jg^A)&CU$!w|hwQ+$4^$(N6+
z&YPEya~!mZ;djK9L|CJOT#x~6OaAkg+k%%{kRLuyF25XCVdp|tMt;17mF99hP7~Os
z!F^65P(zLA&=485O|W0Hxe=Woxpt2oUAekIbVg1T@ECAXA*(~AO{2U?&T7c>V@8>v
zpu7dOJ8Zb`){n}I7Txw?4PUrci@L=%9D$>fSmFq)nN)o_HGlE9wI4Jppp2x}$_J^^
z{Es_;YhHx|3)`(Rcx}~O@m~Vu-+u;kw@33BVl4HXJ5}ej+ddgiake$&ExpSj;5~~h
z{nC8ft^ZPecRmRU8Dh`p^Um}$v_@+aFBrTLL`~X~9Zi$8na7?!KXOi27%*JS6Q5FT
zJ6Ir<RJkt?na-FS?~e`WSfM=BtWOl73C8ZB!e(BlQ2tE2wOM?sx;1$G_vdvHDvhk5
zUSur#hCiF{BE(P2z$7@rTqM_<;C+=L2khwgdK$aYK7<D-6FAK2GVc&}Lf;d|amJ$6
zL;8d1FplWCAfY?-&tfahFmD4Co@qiZx5qS8D*VUbp5dds3zY7^g>l3x*fqrknySxW
ziLv1~HRKLtbqOJevI}|j(z52%IV<-`1J1=V%7Zf#h1DjRQ8b5is7$-8%lhs&m>#3)
za}UHXCYY0_G+O(afqp7atfNyb__!y6uAQjog-znkrT`yyiB~y@%zH<hFCS%8BeU`*
z@AEYw_ru!3^9ql1NplSQk)DLI((#&{^GXKBxtZ2&*TpRcoYlu`;2AXSpm>5_GGT~)
zv}nda{8y}xrK=o2y>D)k%F1AYmQ8Z)ceQj^$#gl4w$-q?+sot0Y@6+LBk>YGqWMs1
ziVWR^=^>Gk;1^H9hPX$n_mW7H6l7Nz5-Ikkp30A8YGiza^-*G7@r(T`fAc{s+A|hi
zaY<%dXb>7tkSsLk`i6@;YH;tVhl{^cs*m|Pk{~k8>AFVaNI2Xw1!VaWbFobWl+RSA
zGCzMCwS;O+fE-QhdoF?XTj6Ekyd8t_={j&NJ;6*x$}>mLqkA!$qpFj#7xs^`jb!{c
z|14wEZS&)~<jffJr&hZa*C+Yf2`{pZ#t_((rZGI4YD}lTB>J0P4K|S@;KBOq(e9qC
zJao1b3gMBD$!Vwg0bMzRcGVHx`p!Z#0!lZX&Ie1lN0DY#m>#)}cOvxXJrbiuQ&VG+
z5Q;t2?o!}5JAkCDY&4ROa*_A0=_C!p-VSOk=}hk5w*eXCeGp?>ELWJloB{|Hl(h||
zUtdJ4A?Dp?M54SY&6EJ+bk=X^x56aKxZv9<7BIk8t0`<2C6-$IUx*!++ZOBU0ptcB
z&0J3v(P~6atJ&3cDvDOQk8_R#5sOa5Unef@AuLVcvg2qkmS7#2@la7Yy#d4ib?4Wj
zTjyZP-*a)gk|XIvO)l%He963Z`m0nBEg~x$a;Uj5JT&M^eesK4i_QYlf#J#icaEmS
z*m-<9(4B~<7v0PNs1Ep`I>iLH`c5cLBueoXr~wQbsNB~8@$vS=%Rio%<bJ(B5sbnq
z*lpmjRd@kvBB?uN?<pp{h-d!|QhWcAy~ATLLoL{iLz@nBY9>&5r=o76&Yb1fqJBV!
z&_}uIM09%2XSC1Oh1saOfSoOc|3<+Z3~zBnu*lS|mKE(p$KFL(WQ0b%jI?Xrpw$Bc
z!nD@q?Kb}l7L^L8R5X+1)(exJ-{r<)r%;y^{M4WtlW4}iYF-0|0-Q0@*-w*Gu4bjH
z-ZtdKqd!C?Y>vEM>?_yeL_z7>nf3XwjEJf-IA})O!HOgJ-i?i(&0=smF+CcCmRb+!
zUBJE)o0$32Dq)of#Tcou4<RLay9pb-bnJS?KnmL0YKe`d!xh?LH3v7<uH%_;sq796
zJNJ8q|J?tpG%`|yrBZyDrebZY6d_J)cEWj{AkNVA4P?zWAu&l<y56;mw{Vpl^lGKC
zTk&=jzF?IqNsT4FRK@WbB2I^A1|DsQrx7=Els~Qv_mzXyxgkMTtrPqq-*qgXN$5IL
zG?3m)SQ}>T^A+$7oS3vY9(F_0Os0cf{=r7gyRu-6^u1|?svU5eRck6ZC<%MW_Ju+1
zSrr1S;Wvndx(CbWjKk3$4Y$9UHE^DSyFUFsg_F>pubp`V{AWED!)Z2fHuw)w3T#+#
zebt|{0%QQD-`LV5W`9>+AZ7!+7CR|&F?9!J{P|ycN_{pgj+*9WKP_u?kSTJ>hIyK$
zYO|>7lcUTsKua$7y3sVerNJ1(=~0C$BLTQDW&DM?xb=lWUk6OXDbi#7u!yLabH<61
zuJSCu05?hg#i$05L&!uGDlDA<!3u86Tlpv$PxK35JFNRe8C^JE=@z=3%0?u!T>7so
za$&#`{deRnS2-wpS?_OksIwi+RX+sYbpCGmj>gQN&$Wu_vqHa&vYOPI((m$og=CM%
z7b@XtxADN!&VZ*)ihtTl^?^EeMY{#@@|ZSlVKuaYU6}dS^HMcGp#zh@rT2fD0+Mfn
zMS|)$ev|2B7UuEA+Y|SULffosVNksGY$0#(ng+qE*MLe_ud|VI8DatBy-rG;rUA#~
zK1~8q-LVHZ`X8!W41w9R6N)01L9F9(`Ftzt^3%poy8R!07w#y*agrMcwVGU?%5NQg
z|C)2of$;J{9yi8Ugm*Sa278+`C3K<fpiw%M|1O}up%&*SH2+VoK_4k9t*#$Nbp-W<
znu-hY)2O0#tu28Y(U=ZFRiu()9-wIWfS>sVB||Gya18vHrob~53P_a3C#VJU%LNZi
z++Q?P-`ASTx2V*j3O1m$3mq`{AZnZnRn&i}2ly8hXE0!^3*mRb14iv85b;yHO|Js^
z)W|hy@2rA<RC-<=f|e#gbq;jj+huF*=|Uccb6FO2wCU_E2wE&<FfBmaDTKn-{|J79
zxxIk9m50bH*zcV0t%D(0HsQ8O(f@UofcEg^HL4EOjHluY5b%+L%;_G8n<3Ih8nza|
zA9sF&{s$~7H6dpD6aOWE{{5=#bilk8DQSNKrXCxi5QMr9KMKb@sC`AXILm$l5Wxbg
zXy=&1fb@mn|5@m8Ch#Hx^NY}W4aA}CAko!t3|Q?lX#x)?<b&+s6(VCK)t?=Bhc9LP
zdH3X^P(hNA?-}%mGSlccM__7E9mEH@0JrcY1&D5PNU~eRmd3rccBeB0>kzytDzo5&
z?VE-5nJ56IO7sLhYv5Ai<>#t4#s9H%|IeKdwxa>_N#(S8opzK0lyvwI29aYZYCs$y
z1GZA&4qf0V#k*gPep&gB0rDjS7&$b_Bg~Q95pH;(8}C%XLd^?a-sKeIq=8TCxa)tQ
z_Ad|w2i1G9h0TJhV|5>ZVl1dZx&@CwWv-!OWm0Ouv06^?|9B#unqXh_s&Bb@d_n~o
z>%C2@x!<@a;0=d&5}0$nrNSxc7esitL5s7Q=YbD2?|@3VV#Q^0O5+|3@T8{*Tl&D~
zCbFKbDS*H@c-<L(t^pt_pf9P<Z2kzZs`%f3dmxu0ovorZ*nXfXUlVi0*pC5^2}x~n
zlRXgkW&mQpap_IqaKZRs3;tD5C8~O|*TH;2K;j_|Dc7|F&b_-ZTML^#2)Tx9uzmIx
zl)yuQm<Ipxw;g~wdSScPLfOS(PR23Z0bLojb8zt4fUw`2O9sv&)K&L!d?d?z(4&F%
zP4;VP0Vob?0(YIqf!vFxwugsu{Sd_X8A3lVF(N^B=Z`VO8%$t)%?aFFys%&G^%l0e
zTT%e~_r#hTQ2q-dty@Y*(psee6h&;R<f>UoV!9&$BN+RJN6pq#o9sFW@1~=D9FdI>
z4+*AeG2d4BV*r|-%3A_E3R6^8pccFxrlP8fuj~15P68-!d0mDwn35WgGO+$~p*YaY
z?^VtG4fUFHN;p<7?|pk#>&Y#Xn(^wG3P7$c;N~$e51w2icLg3!T!nqG#nxU(yxVee
z3LY`1?6j&;pxPX;M1jCwfI6~Qj)?;@ydnozU|xZW#|Pjx_AXz3eg!N@8<*vw4zE>O
zOUgD-Y>k#E`T=gfX>Bo&H-I1xcf)8m3G6$$Wc1Z3mTINcU2q)9USxp+5+Lo&?08n$
zdhkWV;I(gfTt2E;*o5dlD~#}XmCGMq`4=xh7v#3#5tC9h=~fPH|A|!U6eLMb23W;=
zZs}Dh0VRHv44ATSxN_3q4F-}Am{U2+CtRLBxC3;U?}aoK-l?&EY$BJ~#Un`SR4!4v
z()}Fl0*@>-Qj}hGEx&}e2g1EB^>7g7Wa_P@jQm)Jko|d~(=(lZ47M76#lU2=06;XH
z>NxU)ctY?>E65?>Q0(=3H=qas3VMVf4G@l}lgBqv?}#7z|IzsqL3m~jzWPXf(yE|u
z;mNHE25c5-w)67(7ufAU0FjZ7%$?Gltr+@;Pz&;JW2z10S$vR7hnXEeKKMqlmFMBH
zx?N|tqq2K@S2MsLg`bvIYTZxT!ci;1kI6Wa$(6VGb{|?byrkPo{cVQWA`lAq7Gp03
zoGjiY6B_-2@93E>4zlNR;wSZZU>iR73P{Sr_JN<bV|lF=b_mgn3nIyUcUT%xOk)6`
zhKiT!Kb1#na~1*aS19@%k2F>-9zdf$N{YAnWD5fQBUl1DS5axc2kk)!=lXa|w36n0
z&mMe)a-hm>ot2}rjD&tf*V;^;8+|B+SIRE~%t-c~$&J{MH8EXEq1x>0@BC64Rqr=`
z1Cot*t<IsK`$)QspVQIdQlDElocl4<$wMM78{&J$4KqI43n7=a{}h0^rWM8s@oGph
z<<ngQxDrU3^<j(QXG#m=fc|pK>!{!Md27r7dptVt%POCPU_{i@)JT}(@&}5OkpjN<
z+t^>)6N~}IKTl%a%}lm%q_*BYbU5}bRn|6>>prmObW=y{Y?|YuEL616+=xkf;%ZCN
z=O0jZ9v$N%=l109XC&`-KDLS-c$6u1#9*R&zXwBT%T^$`Aha(ZP)c<#0GenVlf8w#
zi7G@h(2Y0=U(S0oFD@*Q!7b}cAU9Qstr5+<0D~cqKR|Gn+nb&APsV<xETWUW#YCo5
zclpBtVV;e~QX`mN9dCmP%9ls1i7HjMdJE(0Gpg8*8cnq5&nL~iYfqwb`Hoqo?@L}b
z%C$>}d5tQk!^Bf&-Hte#hp@-rY={0H=23T6A4PD*DHspjr@Jc7w~Iab{bdB}9YXYw
zxil^N_yg*!exKU2$j+1-qRW4cKZwhSwxLFN`*llH*xiw%`8z~Q(n>vspD+dZQRiv&
z!)Lr8w(?K00cFir5PBTM_^k5M(!#!qu*PBAZi2x{!cqqO58wXb&LI;Ml$#Husg&%V
zeliyR<&`v@)EU>@`oLc&Sf|ENoAQ$%Adl%6r@qBNv7h}18+hEfUYwOZ<R#a~n$bX!
z%z0w_IEDgWlaEtaCC=vQe)l1<dZs8lq<gM@q3}G3RrHHD-Y7~#)Go{JJZsM3`FBD6
z=eq)LOErK1>^TsWiIy3!3X>yvh%f9RUF?q$rW8_I+!&$m$L82S{VfqTj&GhMC?~p*
z((>DRE)iZ2$QY+FIE9XFI}@4Rvw@0di#f?$;yN35i}%toIXA*tc-SgqN-TLv`Ie%U
z%ITrHYQ~>)cb(-+vn$VH=(Silg~z-)`?ASB;tnZIA~+4?pIGWTK6Vt>Tq^H&O61@T
z(0e7t@ZJzsYS%WFj<f|Ebu*%f+PAR3R4r2ClO+{O3h|+_yAYJUs|G|>go;j)0^wyS
zBGZTUf4mOR1?+PIp&WwWnBYwnkwF#QQcb1#WS^-+9iSi)REZJun38NQ*^VhTc=*gY
zI4t9roFyYAeomD;!5HD$_e$L8tkbnZf7?peZ;<YJ|G~!YoLsEWF5F%J(uURh+m<0!
zAS9W-oj$8Hl)nM-=WP`9tPu)n=&f(s7u;3vhDy~93AjHx^<DCb4PxCWgjc^d#cO@R
zfKowLnf6(KomrZN?%T1k%WijhANjjdKpw;usvlo2)7q_vZcC>z<BYfH^<3@AL{JCR
z10!PPqET3m?5C`Y>~Wd3HA_Rxgd?)8gS#VIO*(7|gN`92=<R-*-*+(14)252yO><E
zsVu|dh?EI>C=F{lzdtZ=&DV4c99*X!FroLL`io&iQN~8-#Yrrhnp@Kaa-H(yO_H%5
zT2IgsUQbR;S7pnJ_>{qs#~9TdBOu~@_kLrpLv1NG@zP3Ja8#>BA19?8p+)*uMxKov
zg#jY(mDu0@;^GUfTZFdver>)r-yjK$+Zwv~Y%Orm&UdgsUW$+Ju~V)hS}rd#C2W!^
z$|0O^r%&@M1T18V+7kv+8=pC*)j!8|s%-}9#HS8^n)bUv&U3n5C%>LOW)lzPOX~r`
z$g&aZ#W5u9{641fZ<~ss<5uCz<rBB&?7vY#b;#>z{91oo=BsBfB<Wf*WzR`8p843W
zLfUTR9lgl)UJPH299v>Qrf6qY)s;`v#}9+u{{b(*JRodcdKS&0XW;%(X9Rui23g3k
zrBYYkc1LGKfLg0g!AwVGgfxMkqe|2~GnQ6um}Ho+pov)jDQH@<96n65R6c}_h=GPr
zZm~a@F*61OEup%pN^{lNTo85S^)7^rF#XfRs4J*fe+$=W<Oq?&osUT`X|fn)pSeJ;
zBbp?3AqpAt=9E5tP2K4$nd|3d#WK&>h`Ppkm6lsIh&+hbTb?<9?bqQBymgjX$|;@Q
zR5s6tq)_h`Osn5<7(S=bK^I3RJ(RusG+%3H5o4p0ZyW}1e;879UV4t;0oQL1yTwqW
z@a>tymsJGqmZ6Vq=_~rwOjWd?N=0+rdgzJYhM2C_X3OO@rF&e-X>i;)ZTVfHrP~_(
zA-z|h5V5=n)Y}&O`vD#a)|r+6%8Bi!INbk7%7<~y*&aso1AG-H7k-=~=riN@xf1kj
zb|&fI!gv0YFWnJ<+g<0L)51mVo%)Xa_px-h%F!fHgg!8H%1Gq#l~ufo8&5wYRWDqJ
zVH#}lTPns4I4*o~1LxeWX-%I`D7Xi40#%1|$I+3L^>m5YM8fQ(fX5DlZ?%Nq+g-nf
zqmMlv^CEA*+K)R&rNQV%x)G0~$ozFxjV61lF^IwcgMv=YWO@`17x$b>;wBuAU-bj*
zM0yXN65A|L7C3M!ui*GeMJMOZ({xdGw__h$BJY%x3p$KwQI!iCAV~y9FlRf3<ej8F
z-Qt#HwgeE^+?#8zxUcd?56OJJu6Eq+$+S|-)%6;VRf`-5k!q;P(`9rgD?_*SWK)I`
z-ydCT-ZMR-<2}Rw0EY~lb=`3ckGx4uRD>5Ak3=rUM6UK&9`U{B-bEHkBwuaXmUAvv
zd91!Z&>c>87~Mr&Y^0ZqE}RgcLWv@D)MI~3uoYnO%CSLWore~8q)b;DoF&U*PUPG$
zSMuOb8TxKp$j=ev^b^Sp-l8>48B^P~m}k)Pj-hMSkxlIHTtVd2pIo?hwO38PG5(oP
zCqLOEilH&|AMD%RYk%7if2&^H>|mcW!#^w!MzuI<VJ7AuNb)IxpCUF{bO|CnnCrqG
zSi`bzzXt?p%1H)~s>7{^F_xnMsh}?~TbZtW$5B2zTV~rPuNGO3rtxf7saIPb*5)vF
zn{7fL1=%uYrBMBez)_ZS+tq1g-0UP1JIy{UnpJo?^Xl%fL`KVEc+j>gzLdRIfL_>u
zaIRqIsPjb@2TnG@Oxz6)M-f%0Cgri9eqvPMJX1k(MsACsnaES|R9ep19o4&%mx<Tg
z{gI-F&pxkTjWg<fPMAafa{(qutQ2zlw!b(P%>h-gZk~SrOve<8hRE;I9)B{?y5GYc
zb=ye6T_BsBIvw=vz533V_l~|r6^2ldmeBc<yyu>=*y_wXyQ|i-5%K)OEeUM4QwiS@
zA44+7J}s1FMNhL@_RPWQPpMEBWqIz%a(6=bs@E`k{9tdRoNW2QO?SP_+gpC^JE$%N
zCEMMzlAo5oB4T9*24acK#0?O>LDh_P7eWcNsn%o}5!pjArzN{6>LrGa7&%#P<hT*~
z20dXb+K0|#<VnIIIChDLu$VvC<X2oCebzzCz6z}VjEU+7qv{WFbK{alHQ_3M;c4Dr
z;8K8-iGP{(<V%0Z<z89&!1>!nz^!5k|CEZocMks8G|bd+XxK+Fb-nlAy{S;C2KA%g
zzA<J`8dS)qC{7$H0@}rJ)nDJGd<8|I?u5n{TB_zCzt;D^zjY~NX)g-CT)t4n^U@qC
z{=qcBq3bkXy^)_Dr{%czTY*`s90#My{k|RkeGwzk@a|7BYjlY&P3cAfvX>%fzOfXr
zx_D$(XB?`}CM|h~S8`i6i$rtJd02@Ks9b^JN7BW%V;uohkyli0ixzvX^z``oJS492
zw@36n$L(bswf0~Nu}Y_Ost#T*a~FPFZ#@xH?IWEC|Jq{HvgMf3csMHWg&=vN!QR@L
z)UQPlS?V?!Gb|k~-vBn35py;geBQ}bw+il%YG0BXWATn_3ACFS)=udj>vFFQT6*qL
z1?e{9ha0VzvQnuf&G_STS~4D2c6ONC#09>-GafkIJDi}4<<PMACS<Mh73@&N2%|my
zj!iI{1pUP2SJl%<MLv9CWKHh3y}mna>9-cHJmc&X7+H==?&y)^+|SU<^&_AU5cO~_
zByPEJ*Y~3^QQnIMp(Eb?YOl=&aq;f+k7&ooe#@2&m#4P12H0hU*FEtip}VT8{wDTv
z3FjGVhsaj(uh@*Ucvc~Q{3jhD^TdNOa%%sFHRQ{H!7*Dh%sw(WGhd*>xoMXuEh<=D
zF|0q`#763tvN6-SmD`)=p;Vkd6)~I;oF1sCamu4)0ZqY~#a=Rmiz$W1ab*J{dIJ;Y
zYwh3gGU8aR>^9acKbHugHC#3^hgYi=Cqpyl)6KE!Y)or3=+-%n&YKA8C&4)_m4;q=
z+w5~GOt`>}{7L`WVavc}T8=G?#d(gn;SuZ0Sy`K_h22*)PHAt%&McFySDz0Fm*-3n
zN4Q@ReHHPV{M!TJKdcwd&Uj@zh^){({n?Id(U@-v*>0QsJku0(PoM{#-}`z*jELn9
zXNuD>_ENT$AKpl0SGIULOJ!E0EkAUdo;uStZvBN{Yhy`cklx5E_I(kyr`b`u=257y
zWCF8?y~lwd`n1@3MmslgE;Pr2nXzi&OSiV*g*yJ5K&?gOci?Kp&FurFKmJ=6IuH%f
zgf*DT6=W_s%yzywe@aU(D4=#;GwR1#tbk5{!PV{X!LJ&na(UK!Cl^DXO%jS<Ir#!~
zlimBUbQxxu<|qewHJYK5OZ$EKv{d~ph=xwHLNV7>+WL@Rm&m6E8ZMu*6&}uW6l<sy
z45eVD`Fq`WuEOASVYDf)Wv)_7eno-x`E=t2OC<$s9#3VI_I2T0?y~a>m<9x?IJUo@
z-}4*4m{@T4>raf&ykd9lI}&n63}!Oi3*4hMx<`5@1A>o7D$j|T`;gk|%CqE|bZSMs
z>Sd>REWWrLRzZJ-_9IVR9o5<V%q#Sb8LB9jeBE%BO^Mu+8Kn;%W*04M+nLg88dGG3
z=mIns5gwlQ(f;&GhUaLu`U{u#(yOsPGO5dZ$Rs;r^j8foc6BKSc?F#Yr1q#GuS03T
zXzOB{r?}#Js9wUjrLS}3v_-Ht$pe;NjGFg50g*1E&%%nnQAO^a^Ib{QT51%-)RI5C
z+{+cMSS}x4Z6v8yj~9p#)oeU|LT31$OF>*JX=%|!s))!l#lujq8$2g8OfRtGH2x1A
ztQ;S@u3TuoO-6>cvcE~?WTnC>$ib2`+{|3ekGwBQKY8QKl+MG+h2C;X3Himb_#$fJ
zM~<U<#SPq^Dn9L^kKHVU5e0`^6JoPEK6a}Ny6+__#0h=fkCZ>yngs>?O<K!XWjwUJ
zs^M<sC|;=R@J(ATxzuRb+yxhpece;&WvN#}U8qiirTo-$etJp;xw1a=BJ~G%1O1Ik
z;zeC^Yo7YZM?7hf<^ngP8<#xVL&#jC17@~7smwJ#(PS>#hX)x8?kbaCuULG0Eu87=
z>pit%emz$*W7Z>duDbArv!)a)-_{sd>L@;WrM>cgdo9vCbe;0nYt%-2hYObcv}vuL
z+NXBO$melY_f(TR1+<ogUrmji*JN`<X(tt_+pN6gc1_8UU~Ek^dNw*w`<BYBNmF~Q
z+9E*J<_F9B?zId|tW0Lk6S5p3nsbG*kIrTiaj`vnLd`(cS%cqE)^7O!?XK%GCcBr$
zZyVN1Hg%bThT$m{E@q&3Pm~Nh6)Bxn(<H_DOsXm_ekRBV%21W0g$o>|7F%{b#C`TY
zp`wFaI(~+mc^)q8u`CZKGeWH6QFNVNZ;wb?3ZHGprzG%+-@{aIDI-r+h?R_rh;1ys
zi}-R$s<ZS&2ZtEpBmkZPK^gk^Cx{JlbPc7WR}x+QAT<L~(Z?byghldNsr~K#QQCwN
zJ5)Jd7BMk*)_6C5)++^h^ba8ZANV;z4uKi(vl$k$UNQ+|(3MY2mwr{27`c*>K2rKi
zZ#*yu?5}0MRaK9~Yy)|E6epyMLibRYoiT<D1wljTTQg}sp6{MVE}wj39}h+aC=t9=
zTsm-f>>tABd32hoOsIz5p-pJC^`5k;ch8)peA`6O;6@g>C9kHt$h~q`;quiAjwIpO
zp;&;g@35r^A5>NS1mP?$3HAB}AYPmw&N1SLv^L=i<9RZFX<|ST`&lGh7jO6Pcf{34
zeC8G<>CMX^la8)OX*>Fg6?-f_waN1=S#zXVKy?>ttVmgN5vJSMA?PTiXbcpZYpv9$
z@@vP|)zUdy5Hwgygf@CEO<U3pvt|FCj)zn6xRP32)=({Ha9wEsSw;x-@x%d-sq*nI
z9>$YuI)QM94WXN!+^jB=JAK$zfbozL;`RvX=u0e!(|A+rpMKnhGoPeNgtMn`Z|9&L
z>$F+N1HiWu2|{{*W~~K-B6S4#eT{IHVz+5*{rUM9*=HsE=h_lP42);;hi^yBEf4=l
zlc#CDyIiH#nWP>hGFwp}LHhoba6ixz-d5_7k!alLJsL=LvAkD_<BH@XPc(jZ!(m>n
z=loz~*#iBtN^jX-Evc102>}7=u_i<c$hMi*d08`J?YGUQf`|)=iWlgc{z8+($)4^|
zhoGV0;G~=j@&mc6;Z-wJ)%3Y$1U=du@AQk(V2Hp|Ec7H1c#A-O@Bhy~#{SPwA7m*&
zoH67+-NJtA@K{F>(l_{;KD<F~zz|DZBEmeRBJh7RAiD6WPRaFyJ3$J-N`p@Nt6~Q|
z!y(L?NaHu`#2~tOo;16<-+K|=`=6-M|HAf;$>5`yEPPol$L{q9yWRKddeeAl$axWJ
zcDpKzcCW}{5$5UHulZmD`X`j(zaKlv4UPw?8!v><pg_8W>5XAxqjmRA1RV{2{P}r$
z(d+rZUe`A?y9(13fwu}s-m^o3zl<&>ECaol3-n$Ap59xiyEz(1r!Qd11;6ei*5C#E
zEtKE)soQXr!m{2k4*0EQU4<c5XvaKt<w(`AmMoi}P%Q{rd+`I+|5~zspxozzDUs!@
zLe&EBnBrP^9@FN<sqYrsDM=Uc_k_HTWGB2GpnZAZUjQSGH+**%?Jh`%0w2U_l!bee
z(1+^MXf>&}W>;=<R_yPTe=c<7|La2GI^UtL-<foR(YY>=_*rGSrGp|R>1{6qP!Li?
z2qTCWG{R1}@<eW0lt%xxeL9->6pg>&pZ~8Hj6JL|OCq5LHH0773}E;Ii1@qvOQ8n<
zX(i`iK`4bxzj9)=v7}9(JG$?Ppd-smp$kH^^EVRo>u7Hq-X1gg9aGVeMzC1~)S3Ek
z%Mv|Sw@1UpcD=9tz@IfKEU9RDGmrmr{Y8V60NBbhdGo|LqAJhzA!+Ulk|nobOMA7p
zCQ|2&tIumGSd<W|5}Y4^<PQKvr0*YWCF{vNhsZfH{@@b>T|w>C=a(na%lcR3Uy788
z-1I$3aUY;bxOanFi1f4eghJXSgTeqA`e0$o(F{uTWT&F6QEPS_-F?9{GF?j$nd0_R
zjNr7T7O~#AMJPw+DZjPte-X?5zzba(eX7C3w8o1{;jqQ+lSqsOSkrA_D6yBRe11Kn
zh5fkOv(vTA{k2JT<g{ave*A=!>>%1P*#D7X7orb(JCU=Wd-$9h@gynd)7tv(|1zUn
zXAn*4*EVg+@Cv0ryH@Fi2oCtkPYxkydDRbzL4NAkOJ(y~ThBwcTzr)?HkO0iK{z1~
zB+`Apz<tK&-2d==!E!Y1#<zh^lJneIeK$ThsDJf1Y;p?-#DMiN36Xc?uBxYM@n)UX
z;mjJ4)!?Zm2%t2JNY&&(u?4FW47K_><Okb{XAX?-JZWI&|3Thu5P8l!Q`~6>n`>&5
z_5RK8i)GWFc$+i|#X3r^Ix@#a1S4;lb}PK^LCZhdjOmb$J>_KbrN#eMli^hU^>DJB
ze{E+dJ}{hhPHvXfK^vw8<~FS$?WugEm!B4LhGz@(hQGc~`@>r0fA#t8q<>S@Uc`kz
zz*Ibrc4F>SzZqR!Yx4vdp3<zEKHrR>rBy&u(->?CboG6nGY0MG=50tk#~%kYRCX5;
z#7}*$9Hi8$n*9nu^Zon3ITC~r!Z9FhQ0;DqT=mBim^mo*-$$6H&{+KIWccUcprb^5
zp;R#U*$s}r#a(YP8hSoWH$M`;SJIINp&YaDn-OotPC5b(8FFBT+nwPpo}7_M(*5VM
zNI|Nj3!J7qjd>vcYUoMrB;Z?vo(M87|2eJDlf1B#)JMLM`{(n>@jkgs_^nb6agH?q
zu-l}R!R`<P%|G33nw$O%cyMeG3a0}sT9%c1^}pt_|L%?lsqKoVS(nVmcpBiQu7N3*
z_DBR~GW`=Y@R4$~wev}F0z3TY0QG2lDvdzmKx7Qw)auSV7h{dTaGj)oGUim+Rgl>(
zzW6;I9t7pYbyGSpl<^_d%3yLBma9eL{$UMUX6g~RT|SoPT6Paw!&m1a94YnO*?mzT
z!IwN+X#JIXQ0-sG<ddL4i@V|{_BP(bVtvBa&63x6A3(AQC1MB#vs&=cqiJZT#LIFj
zqcmb3W2K&kFEk(vKR6pLTK1{buqzO;UsNA)X&th&*K}o8&9EPt{<E^F=IjTCuEb}_
zEhjC!9cj&ntS%tbw6{I2A)1z*aFQ24cn6$O`IO$p^pk1rn>*~rm3AG<XRg#jCMNU5
zcTXWrK0?kdTSCZ}Q?=^GIe9t$yn;)mpS#ihO;(>bwy9#vg!kXOSB1<V#H(ZZD6mg)
zHB4IGD|GmYD_{iO%NU|`aT0<sToOcYDfpZBU|zwa2#Go`p4E0o!Zc+H=_TNBH>T=d
zsQql01T_GmB2RniY#a6jlFvq&#&^XAe^=E5nY=LnAy1Jgv+bzrkh=y?RpWGm9xsCf
z=F*)F3pTOgkm_R5qu*HI(!n|l3aOt!_qA7}new{x{osA9c3*b%{iY!Wcjp#*t`tl#
zNuURSe6UqGP%3~+fs{O^=h>5w8WtwaWQbb4J)BKkcnG#7%cei7f#pO=uRZ1~F!>um
zdV@FNI*7GFbCl8WC6uxAzWV7*aJWiA@HuvG&$SKQNn@4v%2)F5e~P2Fz(415&h{?q
zb1o*T<peMcrV_+Z2NaIRy1Sojv9PD$C8K!lHi(qo0PQSbTzntC1F-FCQZFnt8R2n{
zQqQHf0DBHPZG1R3irf&EAA~-Di{NrrL6Pm8gp1<uckOR{v8yBS3nen*nI!OurC`hC
z{(PY>0dMi@;8k#kDL!Rlxegj8;C=;8J;r;sZ;JDp8w8G?m2$@u<oJ09P-%Eh)oR`v
z@7U}v%fa7j$0tFVgsCFdM^D@;q=9Ghn(lqT{{E5L%DYy=>3pC<(d1MzT*dTPPq2S>
zA!a<&<iPs|U-eBH*v&w9y>ez)s4`Ca&sLXq3V(2JL>r4N=~zc;ade#>Q;^twI36KJ
zo~TRKI0G2d*c+tlB5C>Mc=-ePiCqtwUX#)ZhJ~Yq3{UOC6i?B9*(t2;6g?t@D$U<j
zo{?#L9ySrlqsN}$>!idYsJu{Hu1ofT;BY(V-A%O?6JR5AMtk*%&biCGB)oJw=L4Fb
zogZ)Ie2+_QgMA<g>63GfJ5#?@#KF#mv2Td+3KiCp&a3$RH8jdp-@r+=osCd0gZQIX
zWOMstb~o3!1Lnj1Z{{9}h_@)WaBDSt`Tfy!?tfo{H(c%5(B#wH@3eR4_Ou?dxd^N5
z`YG!>Ix~zA>FtudVpd%S##2T2yhU_xXT)>WCBa($WtE_@-)+9BL^rwOSfOqGBm~<9
zJ}0TY^=K#E+*Y1tD^mE=mnqH?-n+qf4A%amu<;-?eAi#QJFpll_u)V^XRZrDL-uz?
zOkeWzKW}#+UpsrGSx%d@?)e>?FrR<bM3rIfCqy4Zs04_4SMC$QUz)0VDiun0A^!_q
CYgk4A

delta 84704
zcmZs@cRbbM`#*lpapV|rj!nj~NA}9z$Id1*d&@2(Iksc(J+d-NvJw$OR+3SnQg)PV
zLdNgb`~CU+9^c3J`;WuJ>p1tg?(2Gv>y}(h5EV(#An1>cxTSM(ae;RH&PvE*X`H~S
zIg*pW7k^mEZOA(jX~<1!pekcvQBzZiAJ3jG=cLW8g+rL{9?NyD{w@A~H#hkGR~i4+
zBGG4mn8p8$=KJ4%_CpH%JpK8($E`_0=@&l7>k}1mf3vI*uo!g1-^~t(H)=h1?a6lp
z{T}jLwgzlqXVN9UJoktT8lqyl${?Bc(CKZ?-Iu?<z509j`%AG0eZHs*X)LBquim`5
zO1D`0gDUxmb<wnGo!NqMCKeysfkHBd&n<Zn1s<+L67pfM#%sx;(|`YTY;|QKN_o?B
zy9`l2udsw8^I7-RAsh<yC`IOwwXzRtBF{KruAH2PSXKlA=7z9+=R0E;_lRmlPYF!~
z$2k-i(D8Q*9G?9U>~>k<QR>l$8&W|U!b#EDr~t^Oo;pD|J-Lj$%dL3&p7uie!?#Zb
zX(AQF5R%dmM>X8mECyLbDjq&XDGfRQ>O2A6{&B8K(DQI_Y64pN?Bw^0?^@Zs7#3;B
z4{B*F<PS=m#@@2>ic4PzYRjhY%BZ2=)BW@30d>**N{UD%91i{WvXAMRfVL57JRqbR
z8C3`0qj`Cwsf-~hH8p8!MA%j|6VcXfi*ah`&iTiCt{EzM+26W{>EsuT|NlSn-@i19
zQLMzQrdYRRCI0cowj62wI!89V|2xh6-BA1gdKexdiJEzR5Sf^nfPSDNlzvJoIS*Ol
zChK*YxGP+pL<0@cJ2>YrSnPH0AM>_b=Gs|D#%;V;xCnTK-HR3Z`1i(^nk*AZx=rfs
z<s2<I?Nsu$8@ug1raA^m-A^rFc+FrNA|#(XN+txx+=}=A)Rl3l%#r-PF3QW)ZZdW>
z%Es%#>CRw~4lY%1l_AGu?oGb~QTtubmm2il)JfL=dE)GJ@BLUty(@v|YdAUeB+`~%
z6Fx4cv+SaZ(14HQ!n2+6|FeSso&LX<p@i_Da;F#!5&ZI(nLhWZNi?`&to&MQJ;Yxk
z!!vK*qa}8{Tqkq8bme~!&|=9G!u$W{0W=zfO=S4LmoOyy=jpTVbFKbGo3P`R>W5*E
z4Id>utgcM?pQmiaVD^@d``d&&|9#7UUs%}2pV@S0aA-3Hj4R#Wdi6)3W20#)MP?I(
zuR4Qof4Vx;@yV}eabNSrxDeQXA6QBa4S;A|bRVx4zL)>IQ5^Wq;7hyatN50o$_a9d
z|8u70YV%+xzBW0%eepd^o7%2^eG?3m^_u6q0&$suHRRo`$*%tlQLznyBSkC7Vl1+l
zy>_$yXWajpWHBOOeUc*?+c+2aozY-ITkutMvB78lW1jykL8%n19B;@wd^sxj|9)=_
zV!lwwz8|;vCb1)uMHO@R*Qa8cK!MwC)FOO*gqJI5EA#K|(f`knHbL;&bak3MF<p6^
zi@g9#F_=Ub)q^-hQdx)YyqT)H`NgfexNgxj4|RD=&~%rx*z%=zGx#<>92Kcpjxu`{
za&ULrdK#=j?A2?ZYK^Kbp1XTKBU3o_*{}ce)nBi?47K{-RZCKakj2DPaG5^4eYX2n
zkn8&_$YPEB?qAz{^F^HIv3SR*Kx-Mc3P}q4fA7!`10(15IZC8D>Xa1#i!rQP>ixZI
zs#@>j?^d_btc&QYL8kjw_kTaHjq&6EpLkh|?-@wGh9cRyRcrXj()-l`2w%NyqF1{k
z*Viy@Pq=oJv;Vs%rtop7PO<denFhN;i90{BA8Mzf$mBI{YrFe}DJ>+EDpDWTwK-8I
z+%W)o>@tmE*nGa+dGY(@ulm}91gh(5O$*uAZ)JE(KDVm>-@uqdn!OBg+gp9tn6Sdv
z=(Es^PfTsgTgb~BIZw`^rIRm$xiJ#M^q<5NVTCBgL!Zss#W5bvIABxD@XVT>Z~obw
zV>Cdemg4oq8mnU3$KUeNd@0w_sZ<5IBSCE*gnw_+8VET5$D5sdV;tH!R;`}qV6F*a
zX72xcW_;1xgUnc+5?MMqFqT4@z|GNoB{2ryq^SLWVWl_(KDsB*b*xCr6*dlK2>j}c
z&1VF=OF^+ko=zRS#1;sqG;f1AJ^#Cv$$8WMVR8#%_H-##q3iB1Q+p~@<J!TrjSJ!V
ztvkmo|GV@NkccPD!OQ9CZ0aQelbWt2*Th^K`bcHglnV6!`&1g&WjInuVpGA9V%7y2
zQ9~AH^WflPAC^o(He}xwX7N4vNAbbWk8G|edJDgPh6no-i8O-ul4U8#vLXLX3#BNS
zb`GD?f~@y9SrO6g=hCZr4q~jj8vnvn5jjK*oq#UmTQ<b8dl?9IR<@$k6yY<L4r<1Y
z_Lx!J?fbvQ`DA}hKTFl3mO@3;zwrK^vQmx3C6G%BM?3x7n*cS4gV~~~@l%U~gzNQ^
zzgwodPIt9Q95BX}dufDs{^ge=B?zM;I3w4MYqr6{`03_@lP5*gZ}ovJK<O9mj)8uE
zWfwY)65e{}%i(gSUU@2qB<eLIh|U*3O72$)+jQLfeR2M$+bw3-2<$EPU*g1}B6J+g
zq%i%G43p3Q1!@^OjUE}*OvpqABK;OOzcyPt66t0j;8<ye)DVFE`t-DK-E%Zg_~VmX
z%wn$N?LSJc|J_&~yRu;Quh{q_6^D~DqTnu|91PG>tGoHT(rEoRT$|X?(%~Czsy4in
z<!e~s940u_A8FR&iu&uhi?Dr>ik$@+(v9@DaxECiOcW%geuke@RrhN3w#WPVH?QNF
z9cFyiGkE9)N+t{bjny2$1w~V{EUJVS4Z07L>F82H*ipgFJfVniQtbKh4-hkYRn2mT
zGb%QIF(P248yF*^NTqRW*^jibxvj+$7Yp8fu~!M0f^5kplOxrW?Bd9nQ-n8at4khF
z(wbJeQdaQXu}BI>nDH}x<u;>WRHUqVdFSMok!F?$jaWMpo`LJ3u?d~1P)`M6=Xtm>
zs}f6M4sz7<>D$-K%h`4D7tE34gb|lD0HpSb_;Q3>gwauQ1D<vvc7XP9Qd{t;RlCS~
zW-h<wz^yv3xtAjw6OIWM(odtG)*yZ0bq3Gt%{xF=a#S3L?*IJwYg~nlBbQV<93;F3
zqiWs!MkC^N)?dvcnC4#A$@01ejWV>Y0<5jj+HaVHt2kitJNDvL=<%-0!v}{&{$q^~
z?=0wpUD8nu?@z>nP!f%pM7oBatXG4BgH%=-c7-{Zu@;v2^b$^6`Com=KSos}J(7LB
zI`4jdc>*+dQ%7A9URRY`3Om6NltLV<i7(uFcT|lexG$F;AJ%<-@{2-GHe!m76DL`G
zvq;kS&s^YiqzH@hWoSfPs6dn;`D5SjZ^=c&F0i+pe_VKXpGnkW329xp?w9$+2=nQ>
zu&>pbH*=Dg2Z=L?U{n|_-A2RU4AARtW4~Pv6fRJVilqjh*wL{%6Ll4wt^X<KJ<9M;
z_15>7l5Z>qtQP6WAI(RRGD_oWhI-SfbKuZDRL|;Z*Wb-2x1KQ_p04_5X5zK=^-yGn
zgehZ>78AMahdI82BlF^0$P<t+i8L~K9zz`0m5~j!5zg-l?tJS!PE#SZ`zXEBPw{&x
z%ZOsJA&w~oq5FGDfHFnwMr<|@{Rp=<8=+DURYMmj<Bznmo3cIB3AVuzOd##KkDGw{
zNCdBB->UE2)q>e)@mqfC=!)zBTNcTOz1%ZeeePjN657f(&C``!n<1|O3#&Ra;i554
zy^_u^UL8DaT2(b7HILYodQ9Q<9h)!D{(NVCk6+2#FIsQ-h^Uv%$u6lvk&4FF?-fcg
zyhf*Hwvm_Iz`(RfIQGr@#)LYv+)|S0Q9|+EPv+)N-rVLk5<6r;!bp_CPF}%YHC3%l
zAVss<K9DGkE6}RM$TxnhFQB+Cz%m<iMzA>PRb<cV3gE@slfhJG>}tMC3axPYz2BdI
z5~lM!X7XP*KsJy;b!o0%lW=78&yeu?+~Nj+{%TYF1V=_dj`|`S8jf1=U)As0?j=Sy
zI}Vd)4>aUE9to!N{A9jf%abV;PKP~Oh{Ge5?2#PVTe>lA^yl0Ct;Z%$a0hu$B5$sE
z(xob#2ib4V^cz!lZX5EMHH1Gnyyo5M`_bgq-Hn-aZj3>|a`nyp@-Y1jX=%3bt8YQN
zJN{~0Tpc?G5|Dj=N(+sFa&NcqIh5-OTh+$hpC5KEa0La+_u?^z&tf~q@B3nt?E=3A
zyCm6Jwz!r}(n<u}EB8pdtNG#$G(-OU$RRXQ^(0?vv+1#my)t1#DP9mrE4QBAzLQBB
z-g9G@+A$^QF)75itRS2=-x1EMc|_@Ca&i3QVexR?99Q*F#PT(sNmhM(@=U=iGKcKT
ziLb9lB;)G3nY;=178(bc`bugfgLW1-nhS4ukHx4zz<=SGzzZ>diKB$3Q)7s{xsxxB
zmQsh}zdiV4Jva6q9Onn|<Q#}WA-`pfOk9MD{xH`=fadV8v9|mINoc@>%&2Y#gF&+x
ziKVQg9}vL^mFGdc2-<YXI=A+VPA1>cnG{N9vG@w!mlRXrwY;uK2J1TB`(fkr8o#<E
zgg5hABqMw+P0Kh_Hu4(YuTROD7Lg3t97`{jWv+bsi@SX}%zOd>z-5ODn-69_kjQbm
zN-tIePOM)(04IRaJvV)7VSatk1jL!Z@*^9io)9>)AMJFsTIfB~=3mCrO?BP$6i_B`
zO7Y4m`K@0Xs}L=*5_%AKZS+inDX1mWgN@q+fu1noEiJ%MX``fwMQW%h20+}x^A3yx
zAkN9TdhS$j?6)<Ojfnw<Pasan!#`*wVZTXYc_&(x_V@cdVTWn!z>hhC2S)oK?I~rW
z7}24`9FGhFwzb|yDyAclEA)?6z*A{&^n+974kcHzj1oe%9>ma#yQfhd0Z^@Ml2Q`7
zHeUAo%Zpoju~<_%dy=<<iH^mfkO~YmymXUn>D9ua;2E(H>2Poy(nW~}o_~ZkJeHA$
zL<+2HI`U0%@}mc)^c0>(<*M-M&rH|OP}3<?<#{8g;?UzD3>J%z5J3tBe+kZfpn^Vd
z8y}ZbM9d&0JYe(d@pc~>mpKWn;xZ%MER>K=#3nF&qSBiHl}32-SkOqUla8OC5Eg;z
zp@Bq=G&&4Qm$Q~|y)|1I$;P`ud4~7#v5896)_(JyFB0rlU@<wF)c55)rlU{;4WG+6
z;e@bwfFvM4^@=(f4ARG>Ap`v%p-J@OyuvD3nFe?8?3XLeyx^*eo{2D6mDAnBie)VE
z+ia0K?8W)n$$`VODyrY#!~XJ{M?VbtSqqMv2PQAJU+A9y>L|+;oj@F;bbw-Y#NV>m
zfRMChIgA|B=n{1~>zwm_CaWPf0A$jrpt3&kV}}4!cLc^)n3pNz|EZRzSuVAe@>^e^
z6!TqF)uoA_Y?B^nAOk@Hz|$jaw{V?|ncpR(8$l`(C`w!l9oLzF+N$G8rjXk1EWTmS
zpY<K=&$nrqD!Jj!<dO+u%nD#Uct5xG%a$f7HpW_2!2Nh9Oy_9suT|YFT+H@gPFH`t
z*->+@L9LPvLe`E`<?#W$z8S-^ukZ&8Dx&YwK;!#w*o(7Sm-pBLagSW<dpmojAO0b}
z9xb?}@mJqL#u~-*qa`wsnTAMB^eD9X92w=LR9RW<4+aRYOL<5RG=z3?>g4%+3;vPJ
ztZ?2krQ7VMOk@NWK_7j|HIRyLQWHrPh36Xsb0rO2H*nY4=0VJG^mCT8Mt`Rt2vA(o
ziq>{z!**=bxfJd6ktuMi4B%1#p_yxRP%j$G;kS&HroZ9F?lCOoOe?IWTnSK7jk@jk
z2Y+6zR}~J`8#Xx(U$Q)yuI!HLf#cCS>X`|<dFTd`YQZdGE!yd8$VUXx4-nL(gcIx4
z565inNVNILQB=qILWb%{4Nh1JY4|~>8Mg64++GMn&;XLNgIQM5yI^-kk^o2*ykBG0
zgNk*Um7fTulU&g28=uq*C^N?*?O|>I=MT8H3$~aeRjcw43}xgp;@XFcj?AwYgl@wA
zt2(_{a&juYR{+OsgH)8o9Wq>u>}?0bcSI{@!PR5Oa#ITX@kVOOwV^0WEI6YxH@=5v
z#=HvL`bK;DQ`sCI;5pm00?2~G&zg#cQ+M#g$-%EinCO_dG|aYA6JWh<Wsoy&ueti@
z1wiOYl#AGLE!?{M6>Ei7y$}8#a?m0>47mz8IZr~=AQ(gwsI*WRDZZo9DwouJMNdCY
zLbJdhEaZZyqz`sp(f(GZLEhCruL+nv;XD7<`7##KGZ*h}e7AaRTnnh>Qqn;8$V<}C
z27plsQ@-K2x<o%I{w?h9ZzW5i@-7FG`~4_&WRG%^T{5i*iy?BH2#^L>aRve5ay*|h
zjolcHM9*ty>r>U={ol)Zl}gU2cve=o;1H#>Kd~Kn;pql|774}rX_Na!l%0!DTNwcw
zru>3c&=T;!L~4G+bWH?lZ`D6ojA8A$L0Unz*d`^ei8Ebh6E^LEzWe=Pyj00yAnA3+
zcYfhB)_1hj`KWSM;aKqpW-(z(igF+z^;6peQ_JzT2e=FZIYyq}>AN}d?dQiQOIWR+
z7M)vr?;pm7vhw?}g)1}NQ#<%I9ax+r&Ms+{QS?9}ozE;ioq8cajW)%nBkj{aYSBOl
zm8Y}rxX0ns6mLUhdL;kgB;e-<xAW8EDl(#h28Y2E4%ZtxzM~zdyKgalb|-m6XOW6l
z2->H3Jpq)dOAaYgoC5p3zjK?1u^Y3^QiVvS=w?DL${jxhjU9z0#TQlz^>>NRR{6Js
zSEwSx$?J!9_P|DDOylj$u|3=b>#lyG+NuK8?rd?Js>%<#f*@jHc|}0Rlq8IL>}>x$
zZ08M)Bw0cYH+Pbu4Zx=vu|45XJok@}G4^m3I(6fHKpy<XM6$w$sId-a{)4*QG;lFh
zAk@j8ZufC+URX1R$M#&|(1h8;F?>=77z@V*v&?(Nst$kv`+FW3oVdMZFAUp#GcWz8
z2ak+|P9$2={c{FAk=^xM^`76m<os9#B!&3Qn<~wloy#+>;#<01GVTqv@E&S5wIp;`
zVY{av!=iq4u<>K8^;6VFsZSWdUt{urJu`%mZS)Q^TqQxhnkIJevwz{Z79YL_v1g{4
zfcDG73n&_*e1G^O2E>-a@koVI{G5y;;KD$kW0VUKlM4e723}(P>F#(+*Twn$fkxGM
z@>A`GD352aav<-Ou1$2G1F;rRhFQD5SX5)V19mV4WOd2y9=rptFZx933L}fQYcuO=
z68>&W1IeLYS?jXqFK>9xG&sMb|LW&NU;U0CqiOodExjKku5R!jC$pWwdwwe;bAV{K
z^SYXhr6l@jOBSb3(N~N5x#D8kzCkCquMuc^dLv}JK{6U3PKkZ>_VK1L_Znp8l<lby
z_VXNJE@_h@tID8BZ<P6**CYJ?c~92cq`t}NB~MO(D7L}3PX2<sJ|1P#Em=VBOs;NJ
zjaQ7p=nn~xmSz5dC;tMF3_Hw`i*(`P4aj<tC`r;dx4qE!Hj5{!_Ev>jau%aZ_8u=I
z(WQcgq=FssEcL~-us$ZGY)uR{euGV*GDuGIylwX)qQhP$B4Y{g!<TF2i9)eYf&4Vj
zx%?kN5}XYG&8IKExwq6HR<|=H&-SN=UHPQLzpC&DlBg-lPLMEl)>S&Z^c!RhY%|QO
zySvZl$Q1&@kOFS3w;O4A0CYm~KPf_n++P9lcpkK4Zw12`Q&Os7YX$#l9upD#B5`CC
zN#9=G?{J$Fx!!LXDa_YYCVls(?Ql#l;fjI*K6rksRAVbL`~kwO3NHv$Ir+1FqjzzH
zj6EN3+3-V!#h5wghA7<9z1R$qa{VYYo*c1~($oGwLXpSAnOxl<!|$Z)yS~Ss9nch*
z?UT%CmfaY@RyYFD%;d2#Pm6W%?SEXZk~O@y6(FDf7p+$y$}}KWTtr>QJ|wUu*==?G
z^<r0zMyc+pUWv8zYR8;bJ*Q`5QA0N1fs9wfpboyMe(CD!r+8N)%+Yo?l}T>q!Dd@U
zt#Z#Lwdt$*1;8H|thkpgU~L@+6>8W8y*N3PmV?#Ld!OC7mfDb!{{#h)K}SW+Erv--
zsZ8+6m!dnozO>$%dwEyW*22~!Mj_>rPS%|G<_^%BiZR5f@bX{ERRl28pAf{(0MAA*
zvEaU~)eb))u8h{_dWWb<o$ZP(<q)O$%ryy5*N_x)k_vrEWJs-|7}i++{P2sn4PbF*
z<Ym`Oz)Z_Ai8wsoa@K4W^n**?^C@BmHN)y3Fed%0&PZ1Z-?C)Xazv}2*7Tn&fkw0e
zFCiRA!qfbX^S!_};2#)PA-D&KENO);!BwCp4@#aBY(3?A7@84&TD2EtM_J$9|3Ouv
zd;0eMb?hkVUr({dOE{Ek6BaK=S<dQN5x{BYN^vQZ_VB?=!cN>W`e#vJr740Lx_EEZ
zhrXe@g~YJ#c~x@QlU(EJoNNPyAqZ{zinP?GJB(>VJtVPtIEhYFve1Wj!Bbh0ghFLj
zjVN7=i4xk#Pn3{XbwDT_Zc@!SN%Y|<{0!UvyY-9Mn-hhLW3UWLe*r*c7PjxddgE&Y
z2!Rs0`=6m*t$D<Dm*`Fm-XQw>)6;94uUZS-nSeKQV=llbSHRjhrH;1mv8v?>IZy(`
z;C0hbbnSvw+Q9S3hieeTh1PugC99_3g1t}IRZ_(erM>7PPN$v~>IVc6d&G=Ay`+4z
zJ);V;wnbz{PfVXqCRYHM*WxvIvR08gP~r;e3NeiU`$*JNF1ooMnh>xuXafXtAZ64_
z0{_JrgZ%zgpa5LvK2h%&%gEl~kt`l@k5BypFMvQ&Q%u<C+e^q5{Ux!5^WRw`0QGPi
z!nI`y<U{N=>ELlP*kb%D0msu{0C?c{QTG}aTm#Gf9jL7PTi=-}wEBhf13%`Tf<(iu
zf`BFH!LPKa#F5Sch2ujei`>7=l7@uKgV6%H2)q57;ND)xTY(M|X8F_49?vi`O(5*m
zUh@q!u&Jl?tNxmCNPV?4aBXLo`l>pXQc}C{5;d~J2q&OPj55@WiU5iy!Yk8=bQMX)
zIDsga8BQC6=6N5&hHk}tJEHZ;o&V%{4v4ff<f)Q|;~$WLode30n25SJ0g#INN>nZS
zp#cb0%FJ)QELAZJc>MhT8Db>F?-*3-EIxCQN26ofljW#auGsDQ*JaJQZY0--rBw*b
z=_x_jJ-jr>-bPxw>~5|>j-t;16w41(;nBQ`dtIvvvyYB^ES5nGNbHL8K$2girCx#9
zIylxMQD4!<_Pi^q_@7_i^+@}n{G<acC&01gNpo$!ONke!c0Jyi&|9t=nkrzzXhd{E
zkC7f$E~c_%?q1E)pI>J1$Yd4!Js65qSWpgTO}9VP7({Ip9ToVd(Tckpze%7{akHZT
zk9(j1S0zYe_Nvv3Cq+|TZ8wby*GT*@&!8!Iakhme^pdmffmkkGZ=RL=QRFXVnGEk@
zNG6B6_eK)}Wz+j}%ep34v1*p2Ukf7^d%7d_SiA8bwu)5DLjb@csJhoE*>;!HEbkfl
z`@-%~%LpwihZ#T`;)OlYYn4slS=9tkk%=O11wdj-x!zObz{eB8g~kSd6MGfMua+uB
zZJ)mh65lgj<LGB+MP2K-7yjztlcfSk*SdUvH=6a<7&W8*tpNfHYe%-79e#Nak_+Gm
za<}lWzrDY|rc^lz)c)l`>jL2oV!=-IK*KQ$j~>WIiffm~iMIxZAniuLuB{mYAYVRq
zlNU?y7Xuqkag6u$=U()lBP^8G5^By+IRe(7JA?g6a2g87U}BjYb)f7o22}YJJ-Yr+
zoIb}AL;#?UysAX06e>2Di=3{M<22+se0r`)<zRMUrg;-QhWn?1M|7@1b|E*J;$fHi
zMa1I-3E?mnD<{B6O#-Mrm~jaWC;1Y0EY^%0-VY?<APf8!EtllmN1E-%)!G=9EscM%
zbqU_1AdBRlvSiY0Xoc*|PZ|8Cpg3mmdCvDlG#ZpPIQX5B_e6%10>#n8Y%c8VU^<VF
z5hR;v7-lJA%6z)s`cg+~_?&z19~4ExBUy#+{xl|$jUS}hSMvBtu}R<Y_*vJ<NTzuv
zpE4j<22V`+7lUc!A<teNS^%WY`yY{A;C5%w=GXQb|KEqJ*BQjzl>QaJ8Wj4>37~um
zDv$vt0kW~n9L(f!q*Vwz1;7aup(l(^2IK+200eRwjIH!kdKL(b6{+%CDb3fPb!N=h
zvh#+443H|dns;iES^eKJ_F^ES@0;JtcYx}@AmL%<@#_Acfe_b)-WY+EJo<(Ew*k`6
z)X(C?=8JN<tqK95@g>k3f!A^hJSPJBEC9(X>X<~GM=!;lEkHVpbH&zRuQ{^+RZIO;
zNCdyeNtgRa-UI7jpBw;fj=Op-R^}TZ;uizgynSw#w}68vVBWM066&X9i!$|pleGe|
z)T4BTu?NUiiKO$RqVI_Uw8o)n)vn7!>6sT<^w?|S?$OCQ4TE(xbC0xh0UA4<b$NI&
z{O-xKu8>)$++rXn!)5Iso_+Oy4~~{zj>E$>P4Rz`mBkY7Qwa*vFo{<&p@>DY{hNer
z9-{%_SPncq5jGT+6O4lc9D!1PKtyc;efexylO%MDV0+pID+LwiHF>-bjsw<LB0ZdT
zL`_M0<Qk}6BiCNiGKKyq4lD`z$@A$u?a8N7?R)Rd2%o!6#TJWQ_ZCR-kuH?<WleFm
zvQ3P}`&@E=Gyh3b7V#zYLnHbVA)41Dyuazlb_ETOmjzDK9s&VQU8V^X2^CT)Rf68=
zEB}TLV*y@<ahqg&DRxVppaD38J`%5&>_GTvJw;DPhPNVcchXx`GIh;hQI^#z+Yc{n
z`+|SgR+BHt>``-Ex+phObzI)d?vZ>>0;`gV14YlgF$0*>VATT@6=9Z=6=}{Sf5C~g
z0bmSN(M(bYcCmks8xjG4RGUv39y()SAr8wN)&nK5)$9M9(xRb?5s>#QQvlZYBsS?h
zw-=0yBm#ABeL$b#YT0sG=vrAEu$h}btkUpX1gt;GHLbVGuM>Wo!(S-wVG6L*n{_@a
z)^_SkT%o7}Kl!tT-2@H~fL-*HRpIQOwtxv(g1`e?`x1zKj<)ECF|2_)BWC)=S3hFb
z%3HUCF;Ke6m~%v0ZWP5T_$a*E(++SJwW~<a)cNx%fg(;9I00q%KmRiLJg7WhxJuW@
zqtg-tQ@&06;VEE?#h~Bh(iO1lg~O!h36tm#cLtdUd#936RJ{M<82EQ);fs8u-Gr(z
zhna=GIP6lYeudiABs*08qX*vuzds0UVKW<AjP9D!mEl0L+DTDC{T_tw4s+=eQWIHa
z3EId%kF+9&0S8Gs+zKi@4&DA-jwcC3axpqS7B)3_+L1S~Xa^ZD2JaCtLey9Ym_skk
zO$Ihi^A74SK(KecHFPowsui$03}i+YvFNqfYhGw`k^xo8;&&x-*KYE~UA4cBtKm+W
zNR$6J<{=6w%+ja+l0of-ej6R!Rz+Ll)6ph~#jbSNG&yQUYdMSX8pp*$w3>{!sOs+@
zZqPXsOy;1GS3uyS^ItzrB*88GgtU5pXSy;o89RGvrXt^`pYLsf(}wK?h}(TALnBd)
zAuAx`QgHyLZ3tkq>o>@c|31pS{Sp+c?u+xkKtN3QgjH(DfC!+7j~4omiTxUJ>DxBX
zdepZM*fK??0sm`lqCzt!Dv89#O@!M)+PL0IYD)r1m2Puf=-stb&l&hseLj*{;0R<6
zYl(KO*Gxm%Un2oH^d=H??*uB609`}G;4Zq4k`dIrfIR-&Vv?_)UeW@&H_5jv<3hH;
z`DGNgSB)W}+LrVfFMWOKpp)M$mKOh@TPghS1i+zG2zK5E{MmOtrT3EUZauY#Nh0BB
zW-h&yf&l0|Sexk6bE<U^hRI;JI{g9XW_+l6e&<u&Ea<tg7w+yYhzEHK2tb@;U)o$H
zFa+qy@-}DO_ptbNMLxXAfxi!WUWRx~*Z=<Z3akbYz$w0m0R0r`y}&EGP{8)XpQkw3
zaWlh*vI2J$u(BIg%Ol0PdcSs){(%z3?y$2VZk?eLeany7df_BH^Av!!t~hyj9Xp2&
zG(0%j4_=6&xs5%$`+J5W@&jvlS|gD8_lLWKzCZbG6+!_|Py}NG5}<v;&~Gf8Z33w+
zfzGY-o9&z3=%YC1F!a?j->e@Id__#aDF=HTD#AwHsK6O1zUl008o#f3GAu^21?+er
z5qn@y_SfB3c10r)LUP-*dQJ?#uP?ke(q6VeBBCJ~A)~a%Un#FwakIfY?kF=|_FaXY
zols+!gz9UpJ0vbX(>EdjkP?suu9*(YhMi@r=lULHx<^=f<6wW@Rxt&s=q+)c7fcbL
z<--24X_XG$`Nuf(RjG9I_}7hqtp8NOW^-n-JDL#uzVv0lMgU1*3Vw_mQY7e^bp%z$
zv3Awd4`^S@6R2ibhe2<W*)x%0ydqV(UYFvuAKj<B#EpT3*sw`WF1KF!C(^!IMI6?i
ziDGnF8;<OvZQ@&_K=(|_OLF~0b32HIHT?c&JC~6$+;Qfx8?eI}1}AKso>pcwq=dfq
zhw$JDpC$zHEHYdeVa71Snk(!X;V{+WX4E<1J1Ee=82mK=E?MkXGjB6RhdSX`fvjH~
z?JUjKTMp#D)wNJ0k-c_>jH&2h4NgKT`ZAN4A*!yh`TxyGTt<fn4H*<>a~mi#8yL`^
zgU%v6)ryDk7bOyRu=J~<^b?fAJgy%)?*6(5`p#I|T`W%l{;<CL(WK*dt7W}c;N0C?
zp;I#w>E@ib7<u#%FnEB*JK+cu1EAcI_jWY!AZVj|Dm61(J!kE0d_0ou47`g#$}%fY
z-(#wn7JV-(+V86rKM6Vqz%df&mw3dpgzPDTCR+foflaQkLbP53B>~Q81HNEJKb@TX
zg7j0|oNA9F>jhdx!UrirRH7~QUo?r#Fk5&KPLCAEc7fM-<qSyDzdk<6y-=^L%0|q0
ziUw|4#fdAIw@sUhT-V`I1e%310bSqf?#kN=!~z~o$Wo)&&;(<$^mjpJ(xrgdw9hJi
z23T~=%$E$uK+qF_-^!NikpkiT_%Q5b9dOy+ZppsImc>)lvOOq_y}S_$P3w$j@4thJ
z@#cMy6uF)l*(sgOud~DnC<(n7Zrn4ceT140+2`=}>}xNYSg`UEc6N{}96U<NMx?D4
z5{C2Q+6YdB@_xowaZ^Xhj}|HuhGDx^t=kt9uIsQ!Do7vW<Uwx>Jg~<b_dr4&!4>IH
z`YD!_QH#U?rh+7t8ILlY6@df6WFK%YUj`?Ich8-^63Rq|!{9T|kvHi?^>iPf`numu
zsqHlNWIUI8bpBoeu`gBOXS!n3D{%CE5&8Ay?N=f9a0!ZpOQxzA7htP$*BDgH?5oQu
zp8fr!02)K|pA`9x@L&$#K<}<kJ^~=~C&-f^UxZ`|l~Cyb*hirnWotO2W>Rl8+r(y*
zZnmxuy)Skpv5iWw`&8r_i>y_O?5@Uc4`_;<2;Kt1%ELQfT6Z{iwH7J+q$t{17TK>9
zge!3s`SfO{moSoM@ffjALKy%@0alM%W@ia43FxkpDS#Z30pK8>IbXtw_5Rigz>;J8
z9SrIRGFnwncBday^~9nzfS&}T`E}77MW7qgLW6C$ZXGcn-wOPl#@(j*U2>`X*qbg|
zYpEzG=|@eR-#c9=(BqJJeyk$f@4ivM@AFz9|0dTKrCRA&J!M%RNS&*jL5RcnZG^cd
z=yZ~-tBNJlz9xf}*{{K;@2j#S-~SO!NWl~|4~46llR38?$?c7bHaovl_8PNA!OILE
zX)$QbgpmJDEwkgk1n>Y#PSI(_eoH5*KmXwk*d2-3(p%DH&^GZ3$ZtUs+%bj<NLU_~
z?D2c%j1Hj1u{12n5mlbBZW7unIqJNZ`@?`D38KRaKfR<@fJ~GsW>FgEKQb)0`T<J|
z<iGu;yyh;XlK=b?W27(^5<{a#jP62;Zc+l}v5A!uZC%C82+?S2=ZJyrNFSL2T1Xg0
z`At6hd@Q#BO~N;i%bp-=AOM(|J(1AY7RAxr85CN4m7c9xgJbhA&{r64LaJVCV6o$k
zJ26h<#VwFf|6_mok(A}3S*u1HdjpOu63o7<3t*2Qj7k02&etE1-y;fP$gj{y%zDjK
z%EmCBheQ~`WF0h24rh-D>{?P^y90xLp#_+}F;inBLmhES)Y9L2Ei<V{F{RqM0zf=)
zbQc~RT~X$qhC@3ZVLY!-XkvZzF~<G3qtP-&#s&s(s%F+lZ}?z^h{I>6i9I;9ebFt^
z7ve2?l~-~e+#I;UO}9}b6#J%RNH`YMQ*)+)1|^wyb*YkMcg3C*-A`PCMh&@LJ*6AG
za+Mw;?#!cO`y%Bbu|%D#e7a0pJXJg1XWb2RAF?<jL*HXJrJqNt;V-bU0UDB9=v1%C
zN<z{k;2s|w*Qty5W8tIroVF~unq0-SeyNdA3OS3REo#x5m6-O@M2rPHfM<Q_&#hb-
z!&SEX<0%?#&j2w=qfX~6>>*AG7E~r6M>{(G`~uM8pz0AtiE>twEqyVxcWta##vvi^
z0L4Q>4E=TjI*zeT-O5+v<isPH?wR9$N48$}^K}DpB}D|nKL>g)#?BMgt}7{QD*b&V
zZ{0Roez6^!cMwO7FD&JmhiS#Sqbf>K%DP(@C^q-<Jyo`mpRui(S8wbV`wXj<CG_Cl
zd9}UM@^EJ9QBQvpfqlnnS@V$sx=QXMmHmY1yECbE5%we7*}%lxSAV`Gjx+*t*B3|T
zqDM!4E{2H`(_RXQ;Cy!{|DC{nE7D=FcqJMVA{kE|n%O61>`A$PA`uzNG08~BQnN0g
z!THF;_SBglWNmqxwUo7SzCzu0+mBgo27H~iA^Kxb#uvN2C;{a9HA96g7t|}14q3*}
za%yKUtSxK$38xxV{uk92f>@g}x$s+6fytaw{4=UZhq25&Bx@VjFX}5%+*cy)huCB#
z9fy$gWjyqivvm)Z$9mxNWS_g_4}!(@1z7!+j3P(qhVS#WG=^8Kk8Xs#j>PAN(84^g
z42MKhO8(+@l+QDVc_0|KLmV{K_4FuZCto_5Vly*fDBCH-&<3?n@oQ?@MuUi}FH_n0
zUjj`->ZkqQTPgZI;XkI7G<82WGrUSh-*~y9zM{m#HIZ+620wD{dM>_VJ5_s<6dW4h
zi0)bmCAa=Nd8x3l6G8u0V%Atf4mXCO86RoOR01kAmjaa31JtWGhJ;^A=zi^?2_7rL
zN`;!w-1a}@TvH6d;k(mKfDnAzFsYmpRg5YeQcNXyoa9a^Ct>5GMgb+Tv%!-S$TmVL
zen~1qJWB%&kCi#<0XL7bLC`#$O-!b3JS{g;NoW~zfPd6TT*CxFm@3kOwJkxE@}?)@
zZ2Afl%?N(Y>6{Sr*qyD9m*GKuv1VTgJ#T3k?baniCV}NeCt}!D?&bSdr5i0wY4{oN
zO1Q62fJlY0TY6A=gj$2bsLX7>)Taj<jOp}EYI}lDC0{ch(L8@lQAVUaujo}?M_NNb
zh*#D#z(@P#wMX-NHI<an5~mcC_y_#eG4t=>L9kozfp7N=w3g|sA7#p5D~7u_9u+~I
zQs0P-s1;2(=8n2bOFZl>f+8{=8?erI#^<4=l$In@s$RvrI?#&uKZ{Nw1_7hl)^djg
zR@e8bpWnValBpb+ul*G)thMxm<(HTgsDeRXOIuf(UG;5=KT&kOkt}XZvMq>MCVs(p
zlT+ZOlWg9&sQ)>aJr)6dOocoI<|zrS5X430Hqq$3ur#+(RZ?Ww)AdL_sxUhM#osg_
zp24$2?cRL=Fh`v!INI!1kpS>jX--HpuSngOv6D1za#pq3;AKm;jK(bVMBquue|Tz{
zOC>MORdHD@fWEj`8FKLP))Z<g{vDOD`ApC?VI*mKD3+)KfisYVg}O=}r_u^XazwtN
zmJYs?F0cA$-EMU|+2LkS3sDDeykYJi^rB#F9W{GnG8!FQKoS^~_-B18E-i}o&%)n_
zH+I^Y98G>0zv0OvV&xx+nd}~EtPf;0E^}0=3ja~fMq=N0(@I#qss}Hn6zZTL0Z)Oz
zl-R`M(b#axG>C(1(D}#BZY+|B_@NrLy=V8ugTV$HKE*njI;9jsG5KS-tx+FdI%E3E
z0?*G0Mti>ASmDwn&Pb`q0JIjGl3{6RW(Q4m<nVFTWFeKKkS4YdZ{Zh@dizMYNI}~N
zr)1He6VJc$BB9^<V~A<|j`}sOx}lCl1$0Ekgs?cL^fggj!0RAk!rC6Y?CFF(COma~
z&j;6TcqN%X`%`F!PTy@4M{J*bVd3{vJ?s>0i*NMdVL~5@E&p+RD+AhO&x~&PJa%bL
z`Od(J$=gQL8Ia@#2J(e7%G^`+eq9zh7ZbQLzBMV+VOOx?F*F<7mlT;NS>|8XrL<gG
zRF7?J#uF16ujLA7n$2Rpz3ik~rNY|n^IbJZ3kbyYpVlfd^`zm+S0f!j8!bJo8l-#(
zWs}E>Q13z!BVUvCfS4<ZlJ}wi9A+}YJ^EujlQ71yujoltmJ0@g=O~}Z{%m>X$?f<^
z^DCcdReRN(qdV+PiT18Lngl+=EXEJmV+;4lv?(f?Tq7C;SUi_OM?~^lS2UMD?;e6|
zS2B5=^6IL{l>tcHD>q$toi-&zpfi8b;xa?x6jVZ~lOGo$GNF7s97QxD*Q(WdYpCIK
z>z(_+wX4`!rH#@lJ{1i0O22`six8#SSsqWIAYZI1^iporforj&5S)M&s_9_w8-6LD
zcqIP0a%FV0l3f)Z<Rsb6qV43)63iZQz4tz=iJhWlWu_EWjhYjao%j#m%j`RPz!|I2
zF6JZ#3$^FRS*rF(+Ar9dQ7&m=rWy2E@a*@C$V2N0M<e{0SSO}gu1+NDvc}I}9sHBH
zAtgK+>ohf9rb0LpKFg^YI*v_#qmguOO`(80!?Psye0aAs$NH)}6>0QBS3F-fFixkB
z{1#zF;}sJuM@x*3GFmNDR@UmbHY?rj6-ekN4#-X4j==1fZwme5)wQd<*|U{Im>DP&
ztbh}s-g2fU`0zXUMn($(s>PQ?RarHtmNqz!Bx&OEVu`LaOUCQRn^;{;I<*r)&PpmJ
zRj9tNAQtFgyQI81pQSI?ev+>@tA#2j$S>~&ZAhro!*}P|`(X{?LHLDVj0dP}9D*YC
z(RYU$gq0Q(jmWVYINUCz7<a#I6&`aOf1yJ8i^ur|(2Xz(saL{FpKi>#B^<6}gsU~z
zQWy#E1rG%t;dZf`X>{}l+AK+X|2BJ)FL?^Xa&@Gcn$7lQ%No!$Sk!9e)kD~ck$9TO
zs`o7m&51TtTGqPqj-1X*kI3!_a4g>)7>Q4<qxkxnKlN8+o3WXM5;;BPM#Cbj%3fGH
zO6Zw>8CC5^{b2;JK%H;;pJLE_Z1*ajm4BD^HFqTjn?edfbzb~_LJ)I2N6x4i*5)K1
zvm^M^#c`yLHb2lLr-7LhBB3d?qH4sxk*98}*2A0d&LTaV%D6{0s<Cg4y{-{Bw$n8%
zXw?o}{*5n#IL8}eiH*$G<LPXvj_;fvj=bE$<mdeH*;6u!v8C#EYE7EJd?D(Wldwwp
zMzV{*qEo-1TCL5tn!Wqpz65<-Wi=|A4)F`jDQ<`*-xH@Powxh6IOEzPV(9Jz`|TK4
zs@DwalDB$`pM;J3(N~(&R})K@ca-vLJotqYvapC&r*cu&VG}pcDV~K1r@`&gX#SqE
zP?Jq?oA)e<#2&;$2m_?Fk;3c)nRcADN7M6I+cnjg4xm+wUT5!wmyuLdRdZi`N@z&~
z``QppymFJ$Eq&0F>=r<L5VD#%OSLr!gS)RQT8CHuajlkDF6w|r=T_u3a$k-~S=yh+
zL@(5v-lB?dL!Ao}mbM-`_gj32C^T{`Y>i|ECd)TyS*I{n;fX|J_C&4ImslWf2fM%+
z_%$L$K*s!K4!O=t(EVfPV%srPuW?;u_N??q9~3EtB1~;$S>*crd3=gU(u$&(=skWZ
z`Wedwt1{7YeGOHBG_r5IG4V|RF-F=>q?|KeU=6pazY*Q!ol;=yW<v*C)k?hV#!}C2
z<y=8t2}ag>avRC`CEH1j{`obHC3ETf(DBlmf&yv+UwCwvZL};d=u6RcGudaLC-``~
z-9ZdoEEpqqRP@&l=s@ou0O?3$zg+*DD!3Yv<c*tTZRJKBsS?abwDL$p;_%DukW0!c
zcQ}3+iPe4%3zX#ykF;!S@Tp1S_?p1H8M40?ai<?>eGAo1>Wt~skD{=BhTopt60|xO
z$~36!sE}E#p}^9#i#S#qRekqX=F)j6Uo820qgnV1;oBrBD>bPLK<yRi%SvfYziP3)
zb6P{b-xkiAfs!K#SxRBT>TRX#<MgzxKu_1Bw4A}I-fy3&l#zbqfe>1}Ir;{d6^9Ei
zn`Dh#Y20*cgWhDMAHqgygx)tIuITzSMf)kr<KCk+*_OF%=`4y~y7)doM78VX$Zlu&
zC!%@y7*DpKX`2gKH^Z*?r{i~B9-+K%8{EqAmY2bi_HGi|gNUmIeVW^-q`AL*G>^f(
z3_ESv;N4{_stj{o9m~jkSFa$<JqaA&xCLMm7X=GJZrDY&=m?#piF<4GFto}KByY4f
z`c%|Hb+nZ8W`XfM-{~}M^dG#(X^)lJS(4K~BSRne$m#1(3eokA;iz(c)4cw<ugb%1
z<o<K%&g*o%CB8AHgqT*+sDK1%{A8%3YnZH&DC=Zm{3I*uqEai0R|&`77hdqJWNe_J
z{dNo32yA*8yT{#nqLVBKzSuyP53E93Z80ce?&FieYx^HA`sbfa*BG!Z{z<<mOS_R?
z_GA2dLQdUK94kM_6{em%V!8y7cYGP_-K#Z*ZBPvXGi9D1zoL>Ild6~W+YpVY-V$jK
zYup{R50q{P+mlB>pFG?3V%DAGkee{554Mi?>cbW)Zm)MvknAW8DUOn8zP!_FCRqRS
zCvVD*r2m_Y(AQi;>^P2SUV2>-W(w>HXuUW%wq5%O#CuWC&U}lR8%ic?zb!wZDN3vq
z@|Mbf@Co7WE(|F`I$UrO>~r7xd)qUqVJtm^lXXR|YaK!u7NCYpIMj`PBaW;vnxeBK
z!4^_sRoz0BbWqYgFvBn%4}kWPwJl6jGW0bTMa5q3a6TnY5akt6mLn$zD_~HGVQQce
z_&58Yn;*cAU(@zmlI%#V3w1NW?Xgwr<_RfXe{9ck<-L8t@X#HHb-7es!gQUS?al|`
zw1Mf=XSsO`ML$piJ`L|3NLWMqv6^&ImR4pg?}*Dl%d+#f=Q(2N0dR;wyLxTfr}D6S
zz~bv10D?ttC8B!MEXimx$T9fM5TYF~-)PEfSdRu;C0#BTdCGzaBzEWyr`2G!DkAl^
z-*9Emfw(ACPbxR0`QxuMnm?)5Eq4DDRL05D%-$cT-3l6R+s#&eaM&8xiE>Ok@0$78
zH%bb(PkWA6UUBZ=2QQWcEfUJ94@U9Y)27$JEjKzZArY{!VtiPhNZ$dC`B%;S??On~
zp1Xfk7SS$vE7sL=7XJp)j>f&I;*TLhfEzl~+p;|Y#!5CU_g=rq6O*qW=hA=6%74P#
zhdISG<{BX1NsrH9_31LuB)j&b^%$|clh4riSUd)RY-G%o6WP1+L0urK23O}&bAY_I
zGOa<Whw3g>jv>;INi?GTUHFYNvlBS<qb+~FV{LBSMicg;1qW^z+|bETOI4a=H4Jc=
zmACIlPz)lPSmgPJ>v+JF)e^P@BRubtrWGanjR+6R{=(KZBMa8@iASS$d%d4~&Suam
z#=M}=CJyRHmu}2;-0)^i)YPX}wOEcYW0_=~fL6jhT1b2BqWfiB)1~MucOPAdH<62d
z9~jo^h1~;<4@FBuQPieUykFd=>*v6Az}z@O%;Y1?%uref(H=#11yzgSP?QFxyb?2C
zgFNreQS4AU>z?2=erwgLFITG=Pa2?qkIU}!N6mmHgwvyeB)W<#NTZO2n~rA~nx~Sj
zS=dmpTxi3y!nKDQwRDWIE1HQznTnsjtd-(kI^p7CltX-z?BaXuqlN1TswJ1H5i#mZ
z2H%xms5<u5TX(RA^q5Nj+}wA&0G6pr%(WbBzbJJVXggNpWK7!J%L}{21x=lgFx0>-
zPm^`VOa?H9_fZl~3tMohxtKDE3--UJ{-Wa7;-TxZw#?$9gdfQ+x1aM-Z0I<qQxAv-
zWS_0eA7Sb~Z|jLXE3K)0MwSk`+rq*Q9h7RF;^Hn~IjXcqwkFYDqOQcgADN~6u<U*z
zvavru)HZ|rnr%beP2Yh;l6fq!kW*C4E!?PSp-0~|gGcqqFgy5aFmiaqaF5fO9L*uj
zYMao4kfE$I@QIE`J%M$^9(7)IH%zuWPFwdT`4vHBDJ|O`c5xP%DqZlsR|{W(*f4!@
z-Fx5cXfWRg(@=NEi|rx_z3TB5+vdyUU9OQRtLVJf^A*hs<qkG}+Zi*X8vud#Hvpet
zzDdi+Vj((PYP>3I{;0R{1EqUkc_zyeXm94NyYlfx6#um<J7B{pmDZa-`x>R!kbPbM
z2FgTjD30HLTzs14N<Ozi+~xgQoGbAwbMpI2K9%KG8!|Y}Xpx+iJPdZW?oUKR!D#el
zhvU$v6Ypn0L({$z4U87a{a|~mD4lh16Em+-RGES}9mFcbdZFuAG-rfuRn56UJ)zyu
zG9UZHbWRmQVj*&F){6Ln1}4|$M}3bh_V}cj_B&nx3t6hS?RLNf#u0Isn7=E0Dguwd
zqGulHxsfAj#U|MC8|jY>7pYeYAjg`CppIYZ7qTYk=~wg`JHG&i`e+{u_jISBO-eS=
z>S4U}n+E>Zx)S59s%LFS>*?VkX0-T3-F=TsRNYX`3}5KPxWa5dR+GNOUzRkG>`j|X
zVsHu4VBsGaGGfkoV5ls@DDes0$*@%}WL<Xu(#r1_H;m2X`nx<a;;!R7I_KLWrg0P*
zgM#~?$1mP&@yq&cJA!aSE&7%J{Q{Lt>d{T_Y}n&0Edi7KR%ge4AKc)&KP|i>0Lx;M
z3rW+K-#=P;$r3&+WFvNVU?T6k^o}(EWq>~GjQ1)8y}@NeokJ=X%WIH;F{Pqm=TGDl
zP{{@)Rwnk%cqBTHJ+@bQ1eUG#X}!Qa2yg%kHb$A_<xIRgLK)or$^~D50ir16q+8IY
zm9`F)0yJNQ*xwD!Oiq+JMCy;+pRZA6rdFX;lY-0!su*~-Y8G{lO#sXO5jJNDX;+rR
zXTErWDs&e~7G8JUbStqJr0_>@QgJvh*kNDA=UZ#Udm+{Y`Fq=9v&UVTTRteop=Ndv
z`9J9UGU}!z9cGMmMW6i`&8ILv$eqQ>b1?8qGNM4QNrl>pK9w{_jXU9bGQY_h|KJR|
zWIBp~>@bXC3iMGL(4%V=_R^^1Nqb7eX=@#d59s5i$aBEkoL+fYv7kcfK`$YeVKx|f
z3hslECwqa&3zmJdzy`X2`)|L#D9l!MOP_!ue538UdX*zG$(-;T(}aWuX#IezS2eTb
z=yw$j+C{B^NRHbIUkGr<f2WP~L(%l8jt~=2w{t0-gWCYsK}J6<!##0v_y^Y7sg3B;
z?v@%LGv=&VMl<G(FnD37lI=vLwD^_Q!fVzpu_eFQ*(D8Nhcu-=`{4W9?FsNSYB{K_
z{A<wCP$ib;u6T1gN;d;jv!dS8%Y+;h?zZ?8v_@aHa(qHCN=CA1-EMJBk9Syj=sdIB
z>FKJmYvq~2kUSL7zFC#crE9Sd?wv9E^gu%|Cc}l+uw0F``ctnp!o<Kuq|58ylEu%x
zzkhROU^pyR^k~Z{e>F;$7WA;UkZ{`OEB7LY%7ZOQS=Jfn`%k-mHH0j4Yx{Un$ld5u
zMjQNg72pp6Z3`)Khh}e+=4|6ud_jB6m`-yaZR^kr=hLH|D;Q(G49030JxYEG#4a;u
zk32J0_zAs#Hci^}IA7D5eqyS^Wci016MuOCx}K*n{zp56k~_rjA9}VVQNyWqFeRt1
zHPE_r?4zdD+X}!KJ#>R@$FMI<-m-bH(zNcI&gX{Wl?iPizA?H3F@DEy7M`w%dR!yx
zyVwt8H;3U!>qdO0X2oP9xxb>ULq28LE63a+ECoE@$yz52#OhrNN7b6V%iTX^%3c`1
zh9$v6D9&yOh(A|W^mm&8T{|u<8{*SueUx?P4kU$%?>_jA#{GEz@S%nM;9X!N;XVzL
zIs#+qo=?`u8{iB^eN=Rh7oWcLj>^kHYvCT}i5<Ae$Ha}FaT%3(niSsHg<AKA;}H?I
z4-}Eki@2jQx*ym6k|Z(=2n(``q%tEjIQZ$M1J<N|6fF_Zoc-!SY#fzk%F1s{=KZR<
zo({ESgM52=_d21BS#bVln9q1SDfH)#qlB23F;fruGsYzo;?pf#sG2Po8FCrbj2t(&
z*ayv8tj}9GTfKhvhx6RV3U45P3&q}a;{PRnfAi9bdgT<t!a^?O&S^9AhQ`KIS}=c~
z;v)Ub#JQc3BTpeSktnwXT*|A~NCBV#YJhf2UxUb|iEUD|XdS1$I{fT0U~rrLHM~b|
zy)rYQ&e;{8j~_>PhkNIn<1@<_q<gFiuLX=AVPwX+yRV)^+2&#Un@7o3E-x)Xy{H9m
z>&UZ31&v~OukeI+6PEbRM-Z$&j!Io=u<OIMYnVL*wXvdc)MToE7bpY4z;v1qH~Qw)
z@!Oe|xK@nquvZn|#L&GgdpeJ>)Rs=vnIu+<T4;!gK5$4+{<zX{ID@)@jc-}m7PQ$Z
z?vDMaGHbNQw*QQ_Oi>I4SB*VB)z#W`cm26h4{mlvE4MAgUhCyh%>0CC9L!BQtKr;N
z)w-g^82Op}CVKMrqI(e!xc%R;*q0b=Fx%v`#LBiGHcVvv+}-38fm5Q&8eh<jp+5xy
z-WvT#9fvQQ*7=Wi!Od{Bu;l92z<2O(2i=dCmd9Gtu%$!bA5kPaDK<`X=pi)SBnwMD
zpXirGI*~X##Z?Mq30k)D{j#ngZd6tbR&W>~X^N$Lr(wD{Yc&TtmH}jV7c#xy3$FD$
z+!O)UfVZ02$MgUh?UyT)tbWv+;9mofwBKO8GI!aoipsI8d8j5zEL>vVW^VsP`4hiq
zw`M0c+jMaMVNq$?LAk%0fo1{K)NI80bC2oW;s1}Pvy6)B{oel2L(k9+Gc<y9cXxM)
zbfdI@;Lr@+NJ=9}hcrkGDWV`EEn%SWl~xeM|IP1N>v_dmykIeB&biOt*S@aLwrh33
zdhmm$b$0|?*5GHQ+i{)qlE9Ue2r){0(d&-D=%0j)zd%BhdR8B;;xCAvH465l(tKCm
z<2}BC(4OpOFaS~g<{oeHjf$=92(+u%jR)|wn!K-%v7c&78QueNa^rWF{HS}^z{KSu
z$k`{yy@~wvzpgq9^}n0S0o+uw5x6s$YWCOyf6Nu)q9Tz6DT+hZ`-G3tEOCtM?gZV$
zPcSCO&-$-y@zy}&vJaB{zn^fJ+jLevGOs$U+#ALuG%{%fW@*SVST-&>z7cS@+~|II
z-{lu2m);ik<HQH3jlLCbzL@(F`_TL(p*1pNI8%j(YwK+w_>oxC0NVg-?SVQU_nRH&
z5C@{ASuj9)iG$VNlmlO|ForwhdRz6L;&nR@$n_|_37l7_T>*3l|E`Ve*B;qv%QytP
z4;(KNsAi?w|5BzBBQ67P?P75QNjF(ZbzCwKCPsm0C^K+yov-%>I*KA8!QLUjH0$Cy
z{L)BrbhMJ-K328*a!UM%a<2d6K$@-9t)UK@X}v7WfQ-yX4g2QPIJ&4W;y4jrC93`_
z3T3npd?T)=nP4`w6JcpIOfmE)&(eOx?E=_=mi;wlC#ev(`E4PCHod$=C`q7DKe>9S
z8p`RdwwB!APlErbM#+jH2N&0hNO8RAR@l~7nt;Dm)Atp!``sL&mDYkm4k5ypZL!W(
zqp!jZy2{NWK@!g@tZInpPr3CueGf=SaevlNjr33>fL+1jbRjAP*D3!6@&y@>eu(|7
z%k*ZQLiVl@L5FQhMtQk<3JoyNB>p>YN13)m{#iY$Y9EtcQlzO}{CprEcH&|`4|0Fo
zg2@76OHl>w#VV*LK}fc->JISRxL4CHc=Bbd;4l`R>z-CnBy|@NcZ!YuU~AKhr|6g*
zMW)UWXfs^jT276PAs&wWup}gTcToshEVQ4${3!ZG)R7*|ow8B{c<NriLm#08^XEfI
zc7SP42c|6>dc1}0Pgi0O-eXIzl33Vy`Z;AFu10tEpwm(sl3<8$@y;^!p2Lt*R0ka)
zbZ{TkwW($j7WmFO<3yft@5f$AzhAxtT7&#<o@J#P;lmsK0b<jqb^(O`{r)oh3pK=H
zFZDhnVQ+wwXXKdR2{4nP==YS_`6xf-ej)6Qg)icBQvCEQ9Lbg6i~aL4QSe44<e*Nm
zUZp}GIsgW8JuJK2PL!sF1H){NrUxOvFgyiNaJ{EoKGjl;p2Be|fk_j%R@!3)>41T#
zi+Lc_FIsgE*uOJoCjUsz<KT*Ss}YprySolLv_!ovJ^z|WieeCB+b!1DDFx|;m2hMD
zLoDN{X`G4e)o+ZBO`T{q1-0;BNiqH(v_w?G;rrbMKYL)o|LUsJM}JRF3vT7F`6~~H
zRX@Q9cE^^A-SuPXTn=~vMo|*+KMfdVFYCFn38m|?6>NP!^M<FZ@R$kbF0pb$k_U-G
zMm}G#EC2Tih5E-#6m%>x+Nn1>jMaWhX7i70Lbi18UJ#^>T)79=LEjqq2UDCsN%>_7
z=PwRR!QxWvQ=8tDNK?YCQgQ3PGZ1K>3|(e$dxJdQDI>RElo3wfAA5(iP(6%*$){j`
zd8hK%8adFgwzKAPQxd;|PRy9;yQedEr}A1s1~=s|iXT0R8>?1pmh|E*&kt9so}yZa
zkZrhjV*%MCyWiN6EtK0j9ahlkL1Q(411uQ@IzX@V*QS$NJRm7K2SS4?uT0c(BAXg^
zX3}k39=HWU(7s-4R4HSl@x(7V1$i7F4jgN!{IXP&`NbMIW8i5_F{-f>C<O|V@|Nf%
z)VZYZYMfoFm>;lERV1BVRnWhpNPXUFcW$~;E(3~LE(^T2LXt{pSbeL;k2vL6e7tZS
zKH^)%SHKsYlwd^t)G*g<ZT@ml8d2ObO&*_lds4mKz%UoR)y2mcELX|V-!in%^Yd<5
zEVx*$SNe;NQ{WyL<B3zp5%D*nKyj+s#Bq|<{~V3EyBPu%kLMcxqclu51rkLFqGZEi
zfn1t49QGttU_*^maqtkYR&OQjNLV;5^Eog@hey?UbZ>;5{>$~hr5qiB@Cl=47s1G!
zbsdoXS%Ud{;wSHD4si@5(U9ch@jFCoIh+Hpn$56o7C-o}SKWiYql9k2lpokmqC-$0
z*AZOucq#<n?sva?N1O=9!`ufht@Qk2qdd<``MyVo>8Sf&3rcg+QnVn~p=dW+g6O!?
zC!uY-wgLE$Cq3UUsj8CjklU{39D;T)GMh5lMb;&Dz0zL(1(XPFtpeI07R8&3D#JSP
zp_;5a5<dHBG7<a78A3^#@bAl?w&rM!>Y@Gv9i!t~uG^^=;=UEOty?6=y2^5#^}io2
zevdQy#KFH;(nCM8SWZ!}Vj6RX$dk>shsVgg<T%?`C6^BZ9=lX7yTr|&9eim9UJC_E
z3ABN6cz?;xpAj;l4_GdjCts$_RUi3ajtaHv4)pk+P~J1!?E|C~yD7~@!D;$dlz@&3
z84{wlcLXwyYw*l;sv0X(4Y4~W#ibxV&FZwHH(5vMu$Adaks_bhYWI;wCk1wwgX0ey
zue4wAoOtI_E@k<X|5jgu_)KR)Rby%<1pW665Q^G2#oLZR<<I$JsL0tj^<tRKXVfQE
zOzY#ca@1H?dlc`oVCI`v+iuhYA6}-o?}Nbk4MaDSF1E2oyp#CG)0|V(;*2}@LrX98
z*i-v(LSCG;YO_Es7vM1Va&S+Li>J)RtD2^{??1HB>96)Pp@EoBEJX9?q7$sm6tpmX
zwiHWv(LRnVJCefu<tGm`gCJq5&%eb2AHMzK+>1XR*vYhlbC-CgIdF^+>Nd*%;8B=j
z^IGgc5^nTT+6$h&=r<fwzEx2Mt0c*g@EcHQ-q|y~{Gdr8^I%<yEcw>rEWmc3?ZYsb
zWJdtPz?Y8>EQS@&&L8rZYF-H|VBa^sQwBUZ_nPf`%{<{tJX-Wj{;^}SDh<FE%H`z=
zkNgj~`#Eg79K?h}XwyAZW4?R@H8+kCT?RS52&S|I5ER^UK4R=pAfj^$YL3&ff$zan
zbLifx?{=Z@-+JWpne!ngoU~F*%;Ptw_E3+2)sm3#grHcc;U5|rPs0Y=&=$WV#bgT3
zd~Gf9{2kB^1nM!1dr{F#cUYuXW;#prYt9E8_`aamk*I3U#X=!nh=66A6Up)0R+TV8
z5#TYB7+?uG?c>yp;MM@?_uu`i?6V^VihqM1F~Kjl{X=$=xPRvCj4R`GKS;~rn&88E
zVg5SDbb<qQbp}KfIAbZ0362C+{#)!fzHb|$VJ@1RUeVvdjtPoCMlLe6<4b-fT|i1S
zT2<en)A32+&!hpB-_5)2K7hypfh<wfO#~)Pa@`Ndi7gKLKrofvo&utk#4do0psQ4J
zXPWK=kH4ZChd_+Ad$WyD$lYf)_oR;a*tLUD5M$LrV3gy*o$++puOb{idY3-FeH)ik
zu{i@Dr7cPk@WJY^fEln4_+>zmFIz?5!W9+0JC-HU(tgvaZv(PqGL2J5X>VtId*a(P
zA@TYlq1CB`PX(K7RBbTQcKVBL%k<n-*xt!_34hqMb)ePU1k6E3i*<c%J5JnZI|=#)
z=bmrr#n~2yHlVF6e9H;7PCcM}S2&qs$8~bBy^WLC<djTG(G*Ge>g<xS`8+mB3#8ZN
zB5BzM>flArk4Z_?Z3q)swL!n%2kPmyPEX*;TwjzUJods~ugl%cS(osHcS%e#1*S^A
zb=>%Xz~-SO3)8V;@?0ichnpZwJmIp4khca7WVBebc<)2v2L3D9rxg0sL^+Q4_X&XS
z8Fx1J=%*u-Pbq3yl9}Ah!~Mq^<L;WsO=ho@QmU$<gbxSi&d_AZV=s1^Ka^GF3Zu`5
z-lo|gIN-)H*M#YmInFsM-snUl-OM|rk3$u_ysFiOQihGwL|R~&<Om6(Uc#-TL|<dC
z!}?F}K;@3j{L>AcSU1sXLJ6LTGY{=78pjxBV*|o(%$*nddgF5W-XO7z{Wz!;iRS(s
z5rq0rR^a)A@ZAr%tsH;!nyAIO;=A9zwhCqU6lnDTDZu0BB>~B%bB|iV+5e!FP8>^U
zf4QL#^SQ58-zIKJJz6?8ww4U81g1hsLYyD>Y>W(WHSu$&{+mA@C2PE8#f`4tE%}eu
zO`$&iU;=u`5WLj~ayhGH$E<e4Da7?LsFcvbo+f%JJa*1|R540a+~yC%Z!>#XIQVwJ
z#OuPyHX}YGdH830^W{bA3m+cidq@GNt&5Ls+WteXK8}|Y+wauk{WJx<y-7IxU3%ui
zKbDX3;FIwRU6qJEk1|u;1sJG8qrBAeJ1018MAUq(%1%t|;;=v$dYV^BZExjK|I2T0
zQ$%@5pMN3k_U{&t6vOgh=C@kLqM>IcE=}>XqJ^HFFXEP0vXzyxO8Dl`@ZLO2`rT5d
zZK1`CPc`U)5V#1VnQmr;NBvo@W}{Y#!*tC?mTVNnNR?X|ufPQhqln>nm%EH5;ucKo
z<5DtwLCw^YfnRj%9Rj!JG*G=-{Mg1NX7a<B@rCKwgm2<E9$L{?zYh<M70+ywUwyB5
ze>h*>UJ<0=M~7~s4tb9wTaj|nY)qj}^-OPqUWd$>1}JB1yHxqp4hT**(;6LWJL+|f
zcOPB>0?<#xyCvjIS2<?n;s>K(^cI(VM+Qgupry3eYh<}*r~d8Bc2v!<t920ME|sy}
zR?%<(t@>+EU$06+r9Fqq9hh;?1D8LZMNx}o!d}<~HNb@Z_<7|Y-pyCWh%3YXgJ~+d
z=8i2HI2MyV(?%({04hP=Jq~K!jDf-~A6eq*tJj9wi4|CkOO)iC#zz=-(Slz5$99f1
z`FSIEm^o>%#&vP&QSa;{&&@xp9aK)eO}jh;mc1(wdA&3C>D5nNdTqyG%w7FAl_R55
z(PetJJaVlZ0ejy9C#H}pCP()I!`0%I4y%ueh=+St5g9sEWV>ovlR9;vRJ}FgI<0^X
z`uQ^+hkSnyQOSE9(*9xJ{(!WoyEU77kuBX09wP-lnKMc(<r4<BJ=4NFek;6vmeji`
z8#OR`V_;9I51Cy>2p}P*t1tFZ@8rGOrE<*k<2RaAG2Xw6QhLO&8vnzc)vw|B_%gmr
z-MGX!a70rY`I(MAYP697tBU|TYjZTzUb#`qK-p@l-;Gx*<=sYG;>6aJG2-sMKB`X!
zc>P|v`KS;Q9@Y$=O|?TSfBv3X4R<DbwLEZ<EZ%%&TllEY@R1cOf~6|`+1HriueI)P
zp8+;h1JE-}(ed!-kuDljz8hXy_QQ=%WlLHs+P)=r$zItG`GWUZHn>FM&3JLL+2{U+
zR`OmnkBAePVVIyz{93v6S)YzMpW2DI%@(u))WD8;TCnKj*E9k)V9p2RG*hi>Vi(B=
zy`mA&?B%hZc!YhA!58qbAV%}<n<oDz2l?EG1=ot>ZR74v0snm6n(C9dwg7N*mog_$
zrGv=-=dFaT`uy@mL-UT`#~y!Yr+cv<fYt-RESami{6&ztkGGVxw>`sM+9^V)JrgGL
z=;L%@>^<Ns)rkALDDA!?&d7$hfb1KFu4jp~aw7jw%X+$y(j_|Akyl@*sFuX2rtIje
z26f==h0?1&Wc|>f+1Nuj*yFflzM44qQj7<Lza2gvfeQ2md<hURlBns5sk$Gv`;oE8
z(XAoHr{^6g-5ggqhig;NaVRotgp;UF*o#i|L0iuP<&!MIdrO+50vus?0yIV?c`i~D
zbi>@d=Si6Yth!id!-B|NE+fFPZo088Y~_F|&X1sU%EW-yMKov5NQcQYYvZ94j}r9)
zG?qsuN(Hj`pEZ6ENMjP!Ehtk$itk4p<MH3wlKgQOO`4eFwm0a9nou(u{_#fgmRZ51
z5D~gc#z&(MX{MWMHR<RG6M9hG37w<%z8Jsy=p0V{B;`LimIu+p=TkJxdi@f-bfoPq
z<2dQG%PS2C0h6jI&O<jfhEXIO+KrvU?teJyo3k|FJZ?NfuSw&oVd^SpS>Sq1SbE>n
z2=JXCy&T#5FTGJ8Yx6XZ1A+Km_iQ03!P3qj@42Z@dgY#Qlm2qt3eV>`Pv-HH9Fl^)
zGUH30y}uyc;Is%$E?K2DQ5=%;l7li;#Cq3d>m>Mp1OgnL@sG+iuZ5wQI2{E?_mtQE
zqdz6*xvA#OK`9@!Jh341;~OwCmT+1sAGUd8!~Cj(QAhMQ8dz-cLD3bE-n>dWDTN`x
z^rYd&6ESV819Rd+`%EV3#R+gC*{Gc;l&wmwRjcp(wVirgx!rM4Sh<!{bCN<y>5sHN
zhmrRZS#hY*b!^2Za2GLn!XJ2e{rXuypb4NARAg-x-3QEf(tqxp>?By@Y4}y_@{$&z
zA-$;SMfdg$JRa~_t@<q9@!}U8unFs1zOy{IH_WB_{hCo$7>^=~o|n(#7d~E>GiDd}
zL$i59KKPaoa1j4%678^sqPZ7X0}}m;->ljk6hYDrh_d0XU*ykJx{gWgs2smDnPk=1
z)PGWOA3?=ltRudj)~9vo+hBwm<`<<=yYT+sEjbON|2_*fbK$(;O3&DRqyZ!7ih3{#
z$7tNKXdG^ZGAwGPz+{#kD`Pr&lB4g4STK3zTkZ6YHZs4U-pxnjQp|D~o*TB&VJ!zy
zJ?}c1k}i4Q)3jc5f>*q^=5IAjrTB9Fop)^wWz4YApU7-|dx<V3+>p*2{WC%>RL}>Y
z?J#%!uyqq_qfvg-IOKgU?=*G4-h}h~uf`(#LP=aiE>Ob{6;sQ77m#`{Dk+I-_4Szn
zj-t`<PP^ES{$&zIhNwc7>k8N3AyV@?abO~C*AEPvkRhC{`|y9)t_sxQgb8NQDV2v7
zk+F3aLKxD)Pe|wS@h{_P6;Dt_Pc`=jbV2Ys7J?rHbPgSC2BxY1@@a%~wP7+20fsl?
zKJ^z}tME?(h{G+=XI|twANu?)rn?H@`~N-qZetPbbZn)=9Xwy-h)UXrreNlwo=PD|
zN)*mLLF~DDci>2+EOLcQ069nVEY$dd$qk#~F|LtTJ3ER<38?0T+<uKr8)aGcM=!a&
zNhpw(_Nq5oAkOFu_^HhgZK`qKN&Reu)>i&hSH24UrvbOAnX>Y4{mLY0ny9A?9;<j)
zykQ;M7zj-W?vazImn4;tR?nv#1DR<j3F}6q*%vEq7$|#mA>q5*{yvVMj9x0TgZGh1
z`h-@_%LNMECR<5Da6^q{!7kaQY}vyOKu9i(1mHt;z9t^;PQERbeL5@_S_UUcS6u4m
zD5yI+Muv6j_kRLKY~>m{QW;$YN!9Rokp=~K4#j(3=@UI(oHSV#0`UhPLBB!PVI|8t
zsydafO;gx!vu$VCXpmOYXSDIHTrOhH%9F1U3WJtCyavf$-<;_8tf*+GoE&Oa3sp|n
z&y}~a4=vk#Zx;a`cMss6NCNLv?roq54*X9M#N~uo|F}&{1?{~9<Pp4mtZEW|kDgzx
zQf!)TO=CkhVB!m1ZL~hu{yTx_G^;ncyRrunVi%WWxev7YOF8R9btw=NuMD~PI03#L
zry-WWS(NZh$^809_I%?1_e?JFk|5z@aXQKwpt+D<s{(1dnqR@EZzN5Z!R@d(if;Z6
zz1N3QTtH7j8>~-e)ZJ-o%^Rf-=p7~a*Q=S8%e8Q&Mpq(D@S?$F6LipxT0D9!nQ;{^
z%{%Wf$nSmomHR4fiErisAb^BfnTs1`x@ffXH#04U*^>+BX{O{nIoJFLz`+b!)>ss|
z1BpQ}OlG6e^lEKN{^;`0!2HA2gjjH)xU)8>HY}@Q?tyIr4{;)J_I0Ve6t&iwOk>nM
zaL4^puV)>i&({g?$P^OuuO#MFJmEA*!J&&vSJmP?=0+y^gU0SgUW_0~11c>J{SQX%
z*xX|HV;$Qaem7VfI*p|Rg!j9qwR+Kh)4MTIDBR_;tee~U#to}*16QNt<DH_)0`&+s
zuVOLJlY@@8fXpKB&J^E#Vu#s_UjMVx-4%l_wxHrVcM|=hL}siqpV{bpwfy}BBCbSV
zDW=0Homq!Cpy~f*$|0#yMZ(S#vWEjX4DqCJoM!m}ag3t6TU<e28Iae}k@o63kOgG+
z$n%isK=L98e<~>Q^K6<=#hd>wqqj+u9};kTFUiO8R*x+eepT^cucR^U1_r<z#x_%%
z+)PovXoC$N$W7vUV-ecGNQ+DDdXo46e_6u2pxq>elGVSw9ktos2#a-*%y{P7NGKbQ
zRMHYRdrGq}z-{=9B(}Al`vzfR4lrwAItpr`I~a+VUdh(0pv;i_ot4rX=U+LzT2_Xh
zhIV>Zwx*6D`h5d^tJd45!lTofQ*P-goZs|klVfL8ZlO4`9`~1T*iTrb{MaV%E*z#Z
zNmKX^qV5PFS9$$K!JIdZHD^u4+8P!QDC3&~Z3ozo1oD;X48IWTLq<gwuhI9!y5Av;
z4(@?Q1;m|q31+lFOcsY$MZpTVcNCedBaYP2+B!8#t5U=@>Ks=b@4AzUGZue0O?lg^
zFLU4rW{K;!M7<l?c`6YcWFNYX#~lI&U8vrV5xl~E!cU+4%N=tAdBeXSpBBX^Z9~`n
z#z=wfz9sPU-2w3?n<!o@kZLcsFjLngDG7ikHUFiv-2~W?<u+$y@eWH~+pb34>04Lg
zg9UV`ezn!_d}UhDm&@1UT1@n@NchkZg~$z2o&#m}H+Yr>Z>%3{`Xm=>@1vHQZoxO8
z9z5y*rS!lFrqgXBWHD|$h4x$EH2@dn-n}RD5}rCf^&<6L->X9ZBfw`l<Z10fLEREY
z<s=jSM*L&z;&XHKUv86EfY437E0{p;m31Ln9`QW$nG_M<jr?{T<-%w-#vsHBmu-k6
zo6Yke#&Z|;wCFsse**=s9k~~a3;S^-h{DV9S@{tM;vX3CJ5+1YWLC7&oQlv|zds`i
z70nW#_!3i}en#u8)s5*P{#8`ZM6daZHO3Jp@pmy-W&$ys=mrEEpU7O}xjPxVP0qQN
zF3*7rjm5SE*+6fJRqOCZs{NF!pypYW0;X@SAJ3Q7Wh{*YRcgoxfEmbX&fLD=<LH9-
zmR@+K%`N>uWehTEjrMa4Vp?Hs=+KCpPih)D@hM+kyVEAU_=++|fwGs7E>&b%A;RGY
z&?ZM7>!Ow}b|^cEqp>Z@a1C7BRQOe<e(25IvWI6;z)2?ZP@0p&%K?S9wP+^#n$PZz
z5Tg8cj1r>(!%qpHl#(x{tx{Gjo&m{@dhnxT;n9so&;V%tzBs9cAKs`{s+~Rf(Zji9
zS6)2gI{CJEimgw#4|F1}ng_!BrRt@2)N{n1zh9{lIB2W5xkwgABtyv(JJ(!u@Sa*S
z*h=4CvQYzh&}vU)F{2&m;@&7_Xt^r0QGv7NxSYnBl2XUJNDN!{7V_R=KPXnAiAB$U
zzC1kkGG@puB>KT6nA@qqCh$vhPwVed9#8%yygFC~^2u&KUC_v}^V!}!P+v-w#b~wp
z64X65UgzC;+9}zl<x<%krLGE+-{QYNO_Sr2xdL%JrvIA7@)uQ>9-4I(aH45lm~M>(
zEfV&ffP&@#e_f5m&E-TFCKa|0=0jR~z&@69+4!B4pG*o4P#QO$shR5&8#z%36_?MN
zp23U%I36>H?Pspvq5-mh0;Cn-LyhNoxh}*$T-{D&ojY>f6zPk5x4Q=^3Fe|>pU{52
zY)<8A8`o4$>GE3YCn|y>wVVOp@x6QG`-TrK_sbwxwSn9Es82(Y!rq1rs-t)uuD}y!
zct`RJcg0JcT%K*jE6MmQyP#198LQ=4bLE*dji{FTxh{2w6tO}o7LC{v4Y`A1Gn<NY
zMgJjo#;-AW#6CNnzzzZsm-G=82im(Sk||oOR-;v~#$U&tq6&I{Q9*qj5oyYSs2%B;
zA|ikLa>*<#>a|L{*q@)e&@|{@0S9f9HeKE^%oI$5c0EG0YJYMmyA|v#@l(Cgfw$`0
z_c7<~+)PM1-DV?op?(e5eWxt$)-19oLd&{t*+1^Z5RjK~&rS|B6b92iod$&Nsc*$o
zFCO&z!U@b!UBxtsX>sClE3MO3T-zBN%lj>_(t~E5+>&?-i$w4Z!gMb&mCybc<QJk0
zdo(H}9KuC$nD`YWw_so2)q0pd>B*P*!xqI;_+=hr_R?B74&In7q%|G>aYtp;HZ!LN
zkO&Ph1jJH6QW9XwqNfyhbEAj?%)ssSkr5Jk)H^&{yQr4LoRn3yiCd9VYPRSt(<C8D
ztrQX0DJVVgo%G*s-rUkrO0!{vQOqPEp)0CB9htRqG2dbKX}dv)wCD1=p9Qb+FC?4R
z_sS$UL)j6oZ9~(aoj@m6hb81Od<=W2pqJ){^}3NpoiLHBlg!#y<3Zlb^(vSs9?EJs
zrzo~SQ#|&8uvkIwvGLrle&>72v=R2MrLB=zQigG*>Yd2pE|GWa8kSrHSTszSFCs<6
z<U9Tr?(#YvAGkXp^Gu=1U&vz(RSn@KxVPzB^3AlN`0*#@o!A<pcQD>)NWx!^lAVDU
z2;Z<wTy`M6d)lPu$*l?>!$puNM^hm%^`}gFTTP#YlSC#I{&oh~o6>B$e5UQ0l(WBE
z^GhXeeT?#K#Y-fbh&kissms3Rax-un=d8p*Q~7P0axiCwWUWf++JTQtKtcz4?}WvG
zz?NH>F-4-|<AOXql|}d&!7h7y5oqT5jA#BhY{wTQ#TEZ{2A)OKEi2_?kKbyL$4C}q
z=A0m_(`<W1jLqgLQhvU~3uK+Yy>@4cRi{X$+3(p+$4NwbubN?LHUS{JUXFuj1&;0u
zydh4{HzMzFmjgGwuIDqqJU18G;hw;r8feFdip4x3M>-NMIlQY%Ul(D(+EjQWKOaTg
z5w`HsZ$gy83ypG5A!OeU$0~@Q2Gt|l8p}Y*Evr;pz*FNkrfonBWfnQC&Cl$nvwXQo
zq1s~12%{Y{+F@>&eshAoV^um?JY~ZB%%fLBF_XLK{S*Mhd;bXvn2&-l`=hvyVg`{D
z;cAO7V}FC{ocDA%dzyjhiN*Nr$)1kiICRjz(?1mT)du(hQSs0}08a6n9Ny5kpY%1)
zdKQ(!=lv&_Mww#XolvtSTS7c?gsQRryKb4i-=8mOoPE&>9P#RemEI~Xe}QK1NeW=r
zyHamt7LYfb!c2v$?|;x8Q=TH7a2b@n*lsTXwmdTLjY`yS1C%E`6Z9E5bR~t}Xk`zS
zsvQ3?T3IYng^wT!_+1>@G6)4uNpYU+co#8Z2=tClhWByTkfUgtLc>gE<~)7cOgK|i
znv^Y7V8?M~@)$IRDdci6N$(zrdfC8EZmn``v;_*U;i|w4w(jnV_hHI;_>T|>K;T8H
zuj1wN77$bSwyNSJFi%ZgFOsOCTAfBnH~f{xyr%bn#WC^>{+^t*RB5m61-FV#Cz!8(
z4YJu>QpzzIn`w<rB5m~4vR&l32k5`mX6QcJ^Z$T+PCs@y`{b~1|86$#QD5y%77Ec(
z`;`;-`LIOtB1O@AdbHefG_Mu$S+1D`F_yhNd$oFbGCPTao!dyUjsObdo^WZ{<gg;%
zRq5d&Y~Vt#tfF=y%EdUJg7(*iu?ALb`6wBw2b?)^sP7>atCp&@DFdk;;!H6v5LG_C
z0Ilk@Su|$hNzYOf*%j&MJm}%cW6Huo5Fiw<4bw{T&<TR{9s%Z830r07MSXoO88JI=
zhICsMi^Y3qszI0-3TJW+0D`W3nBEXHRKA^g^1`tt=<2`jt9p8;QZNb5j~*N0#7LM+
z#ABj5L@O)HSEVe@)wELRNFvCRT0J_|go1TOpy0swatvrT?%pXJcU$GkU-8b47HM$H
zD<>alC_%Og?NE)ZDY(v1sfK;K4!r#Glu;+bqTP|8wtW+ot>O*a!V^=sQDf3~qJv8>
zs~sU2cN|C5c23I+Q#tgSIu$V*I0R(X;vDlleL|MgRbQ4_dqLWV@cm=p12im=n&%b<
z2L>`B>+Ot<IeRDpmOuQumj!lPl=hYZ*is{PsXB$~Ovqv;u-3f2(svO_>ExlE172AV
zeK6SRWfN~>`YxRu4Pqa3;ar^blG7V#)gP=8OA~(avi>#L+)Lul-$e5YP!WrZ13&xr
z(HrUaRNnox`jCSZz-CZE6&FXz`Fc!KjVbUq2^lLUaaseC)@XtCSTNXuuww7-`mJm2
zL9BHbaa_+3m_xNM3D;yE<P}tBE6-BJP!4{NgD+3Bb3T965nw5adtp_C%<<V_l>$3M
z0v|_!Z(<&I2yA<KkOZe?;>e_|Fq}I#emye|7(6orZ^CugR052S?ll`#sML19)|cE+
z%NgB4qEyh5qu+LiiyLB|eF1%plKguAqlfa-r?o^4s<Y&S`_&tqg!1l3ouf!p0Oq}b
zh~Puy@p-EyFq5fgoq3yAfY&uTna`rJAW`D3^#ZcQ5==bkpUR0~v}1?h0yH5}h%Qej
zQ;vUQE~{!YqMlqW9F^g?zY4oebkSb^n_;MLHTeahY1QO44S4=Dj=;P(i7|#1;UFCH
zJE2P#o7-5O4kIy2J`+~-v>XO62(O#7jBcH5>75?8zBb`4zT=W@Hcf&lHi=~aAnf2b
z0RKSF#aHKuBd_JEVsZKV0dzHV?_67D7Ff8s8+MWd$!bkXKs=gGNraROfk2TFMNoRG
z`J@EA=>$f2F&-HRtvN?)hx=Qu7Ru-ML*_91;=&74fG`<l`yM!dcTZd~cv(uYr?~gt
z8`bahel-o6Gu{Kp>Ia(x(kg|Gy50jy$6Uzo&R@9kjRtZr4ojyK_{ip!pMZxU_>rws
zAlTPoBq3a_D+Vyb$E^?)`(4wd!L8HbAF3y0@8J%HNvGqVr*l$V;u>vR8@GR0bA0in
zo+5ukCfoF`r>qsi-dC@X8aoe-xwK+%vHh&1b4|5VN{<j4ZK8`Kr9j7tdmaxW+j#vA
zt(TP{teuVR6m}U|JcFlmvX1AB_hc_zMy95@%VsN(8G|&~Tzbz@RqTR|`GjVZ2Dd}0
z&zZkGa$0b@oS|1@8KUT~tdtU8?YAEys{|J+?My3So^W3-JA&bXQI9sTBBU`N>PSG!
z@cHkrFAR8*GfB=-wv=Hj=ZNWI3L5!#&Ic>)<NY>E{owxlbA5T*u^B)_*kGtrmLfPr
z7aOTT?`|b=j|Fw^JTP7i&7#9&L<nJP=eu&)g?aCW%7I(y*WH4*x1?TSgLdHf54D<d
zrnGQ)6|P{Q#JUs9@Z-<EijIr-Kl+!;@qy5wGV!yRqcV(xND&;kkJn)&%-9H77<L4I
zS`N*~7MPMlzSMok7H;SfS*en$9|HS9lCM5g$daMHY5lv5gO{PFf<hr*DrbcM4ikx6
z;xW&(<*0}ROwhZ7%`$AS?h%nOq%nvuL_f6?V}k&TDvxLg2%SIq#j^P7YmqlptlpP|
zT}bm=>K}nz$t*lg0FK}?j9x$%=@*efTIkgZfx7_tM2;rq;8Gji*b7{>d+zxiQLQ#k
z!X{pE=mEnrg@Fy8CQ<rhQx(5G;DL5sa&eEB!a0VW(5k6JQ&Y8m%oOj@;k^Rzket-7
zcVvL3G%g_g(+T8-tQP(lQZ9A48xl7L(A5Ig4TW(xYT_k>^%S^f?~?ftTr#&u1&;uB
zCRl-YoOAq@4Fzh-G*)4qCI1mCx<itOOzjGFjWA^^01;LIHh3j@if!somny&Zqx(*+
z=}UyoCa*HK1VN6JKkqhfj@n*8+8Ch8qP{c(&1vs*=JTz?vTkd&=C*g$C66-q+@_eV
z6vYv|cjy7f&G|he&wt@IvS`b9A^q`je}(`z2umv`>l+U)u#b%841QxkJ-KwHggdy;
zYV#W5n(|AxeG|)C(JKO(aJ^*LlO(szg<jkkL-}(v^v6Q23IRLZ**uzPoQZ`Jv!vdJ
zJjy#5C5kwD$Jfk1gJ$~0Lu!0*pv#H19ETbzsLTQ?Ei`lou;qG3ce}FfZN5P&<Ibna
zBZ9s%bx-Bd$lP3tqc>4-)nGf_A+Ul4_CHG|i6+kia@8<~YzuTn6H6iwb!jq;-~C5}
zT?H_ysqFGgl9x>aSpk@I+I=nTAJlZeo(d}5jq});`k;!9+h}xbam#jnxtfowo-LUb
zt?E#1OCey-#6{gEf^ISEk+@s2fONyj@1_f4^MJ~T?z}*~0=R12IG09r3Z(QZ62P@#
z232zlg;RQ~%Z|Gmw2Ok-JYWVE!m+*6kM4w?XU!FT?u?E#!%nLrpFgU4rt#vM1ddfz
zqrMfT1AFxGmiilSk8b_ELFJ-E!a-fxFx&s=oE$%m@P{&@7%Tq`@P1qYhQ#l>+5@mB
z@Uxm*W*m>p4kju7cgjaXxHi#4%`qCfe&CK4^I7%-FQ(({Nd&1r^%}Vr(eM97WTz51
z1uBVw8-L4ZM<Df4rflmVzY!1R-9udU3A@~_a0MK*$G_A2>9d~q1$~bx%~D=_@~}2E
zXZSlSH&9Q6?t3rRjaZnYnh_LF??Jg<CEP5&cz#rwa+d+<P5mDLJ+dJO{|5_d&~IC1
zE)8Gb=SXuVRTPTZPkLd*33WjZPx}EJ%<^OBv-C8a{<2vaFP+;wC2(pGjxL@tItwzv
zZ@19VU@adyq$h_t15BV636(`hSZ8ei%I&7J9_vnVyW357sh%wVRm;b)0YjQ+e8{!S
zfzj#hjF3D$H5H<dlE05mgjoiIS<8js9@|PB$2WbmjF2FCWMw2+@#B|0yj;Q_8N__~
zkY0oh|2Ef_9FbX%_UM`NBCNmdRIn?zh2#9cfE9JYTj5|K4TqhkO()Q~yN2R(FspOr
z>_VgcnNpGQl69)Ydck(7%7i&q3Nr?ABR$YvkECW4p8Pss$i>i~0O}RUwT#rA(>(JG
zrQJ~I6EB!8mWbdAE9otML`=kj(JJtXUv}Z!d%btp?-jSQ!vUYHBAv`*wMN`i%4SNe
zZ6Hpaz0ZONADM>?sa{r8q-st9y_a4)B{i_r6hCX6`M5?%?1AxWd25sME(V^}5xFXW
zhf@5W+l)}^hTP3MnM3dqy?x}e<b%6{nsl;EDK9f6jUfl&7fq<YxthGL(*cYzHOTMO
zyX0~>CF%%Dcj}GEcz&esP^=d5AUmWgB7INar9QVuy;p-<Z4KE?JhwTu4mSC)Rt<{Y
zQ-!!FrRrK`(RvU$TlXh(@J|pq#QcV#k`}bWSgI<BxOsDX8vg?~tq+R_8zr~5%h%W`
z(+e+`{59jz&jiBQ8+KsHRK(c_QS~}Oe%)4Xhb;0lE*fpV1{v?CrUIx!-K-5YvkE^?
zcpO}S>EWyq|I11T%276QWR=JIki^yb{$$BqpVs#=eqNTNLF9kBn`I#}rMqXS26D}K
z)0m4?(@8KLZ)w62jYBjgYVGbE0no<+Hng>M86gluYw-*$KB;{4W4W!W{X=&8Ub?||
znZp{xS|<kVMqP{srw8^QoUiVy8|WvM?h_30IfykU1(!CT6n3XKzWIhGPbN9K@f2xs
zL&j0jW7~`MrtFctP-6_q$Xiqm7}n4Us{P?=H%sp1`M9Tv5U3QiB!KCLb3J5+P+ozu
z^(`Jc*Bu#E&SVV7W=GDk#lCKYc@pl%udpBA6{37Co5pS*tmgWrI%i_Omm%wfKPOa?
zb07))T}}q}6oJYxSI>yhW5K~QlYzHV)oK-K&gO=Fo#^c{xiFC{6w5%*Up>6MV5Z4{
zDbtnuib%)8jK=6MNk^U-gZHNQqjSFjOgeX|4xrQ$8JC>!7jlUc+Q66M92KJl5yw$t
z*O^KbAy=8)#m`z3)_OvGvT3BZoIR9c)Scdl`L2pB1c{0n&<3eJ(uP{Uc=F(Mo1KW4
zHU3U>5B7(kAlxl1R2q)>57TQu!oI}m$e_saHIWcoAoLq6;Y@O8XF2M|+h=a*_KZzh
zMEi1JO-c6!{kNJ&MNjm<th!z?x?25gN=*g!)>=c78DiW<K~Xa2ynfmfPsPH0UkoeM
zcenk`#n*YQa!sqdVAW6jBR&*O%YV^A(^WwFblQluWI!Pm=g2Tx(^br68A2w2%}iFr
zXka*S1NYojvpIV;7A-8qz!D@FU$PxZ)>P(|shmutW#U<LkSMhA<H%;NK)4E|SHYUB
zeFNG5J~MsQwjL#FaLgB@al-wDDdUUSaDSa{W756NlgZ+sER){)a5|e4Pl9@9_kZk%
z-YPp3nqSwn5W37Q5D@zLEXJSc6GImRx%`<C`dHjE8Q(L1;`^w%o~x6$sdlO@6vACY
z_DMnL8+#9U%eO{xnlM(t0DnETnXzyYB(0{%=o%_pS3kCv!tUdit@}^8w`7<~91Di}
z176MGz;8kl`MV(W1kSv>JGP{BLQ&hMs9#*Sa3M@p4vloFR~-(_IBPP3=VS@UbgeYX
ztGY@Q3ry<OtsF7p>AL*1P-f_!a!)0aFqVSUmp_@q)*A1dzRMQ$I_?RcszA1Pe8Ds5
zAoh0nrPWq=f?U?&lj>C|5Rm>?7+Zs%^$#NdOGpC-q3;>Aa+5px*LH(CC`E;Bk-A>}
zTWfN?!#Q14!@=oIoYXEj&qc+lreoHn#b{hXn#HLb*Pd77LhPE7YX(|PoqjXJw}<AZ
z5#Juklm*VY&q<30x`m-`9Og<M|Getk{0)Be=Kpq`gzJ6)+*G#8x5bizzl9$+0;h2y
z^RDVx`&!F=#hvLlIfe3UC=HUVR+^O^xwwD^kP|QbFjX<{-zoBZIzKs3i|*G7G**l3
zyyXGd6=Kw+j(@W(wWV8ofzt(v5wo`6qSbOd{!DTAbHKCvp&#}#K>}9E|MKUkRnGBq
ziK{0IMLdmTqVVU~u91Wu`XfJ;(iulu4>SoATuC{^NV1uiRM~^GP!V4Cd?G55UrtK;
zB$x!LdwL-Gap9Zfmz>y-WARCWe=yk{4H?80Ptp?GwFyz{k(-WnC{{L8>3*=WjNSWB
zH}@sai2QEn%PG25S60ZG%rn<hEg@QkKScS$je}OWRx}`Lr38<n%0FxTb27l?McA)M
zPCUxUtyiGL=%ZwYQhL24t)oCd=xu$qSg3`YdLP-5FoO0f;Q!9ClSVJfCC;MC*MKdH
zsPl_Em9g#*@qNaCThq6dmeNVXoR$#rYn~N52Z8M(p_8slP=#~2E$3R%5ll}Sg5xu%
zk-?(hSH*|N0bw=8RwU0C<$tNsQ6e2kaWMJD!9HF!HWZeM(rHXse*Nb3Cu{3Jz=ou~
zi!E`iMu?mR%TX4noPk9<F86S`js4A&fiUUEpTN=`)$fX)3mg{9sZhdE+@sSO45p3A
zT@9b~2Q#kLU^Hz4^;aOv1#0rqNQ$KTu_y()?TX%QJL0$CFEH!=l7XgRatdv(`-EYw
zzc}Ds`FVsQVDO2}%>0Eri#mDBT4^#J&lVyfe0^N><3?1hEyMuzdOsZG>`Pe2o&Mn#
z%z|0;3TAIiZzN)Rjv~RhHe1@st3Z=Kd_btcbecpY$H(QeDxc`F-Pfm*1X8l~q_Z!x
zR6C3d6;TQ;G+EriB@Eb^qbhhW1HT!Z24p@DVlzNJiAfD5VFp%1sRru^k~<WKEC;>?
z7B?PAVbBR8<SUPfK*|7l6(`YiT^Ii-B8W3mvY<Ylj0Cqp@S{cWlU2}&hXFd!W$8({
zb@zjN`K}UXaogPD%m<w-WXEc3AETd|um2c0cFOn~0z@jyB&Xe<oCR$lWu+b5Ffomc
zLOn0U#mmQga>9?NS10@&R>t{(T9>FJ$&i^+lpK5c=W$@BwNJ=sVA<kVOil7i&XKO9
z;JmUNE+ko>D(jil!q(^*-{fe|qeFW=#cRIH<T~LO=0{E*jsIPz#HPi^d=sI3(=xk!
z|5#8%b(Io(qFb(t@I?Gc(p?{&G=Z|V>enc6$o4AuZ9%LZ8k78VJ5&R{{T}{{nMh~i
zybc9}25uLw&C=(m2XW}1yUHFaBZl#%G(LV`>31P!$FJeBkpQ~573j=P{b4vhgR0)u
z#{aCX`|Hz*7G_4ZKNc)z&fe(vQgz)#&nSN+nfy%?fFjxlKJ*A#Pr#&R%0q`sU!y3g
znn#9nFM~*jmB=u$Y$QvSQ@!kZ3FkBG^$ugM8BW)2H*~U$OJY9t|EzN6oZgQOi7vRX
zt`w(>q9KBH1>~(?l#SBhuzs91BOll`g8CQUvx;l~VN(jO04LkfwmANrgN{&QHd&jm
zk)S%18WgRw`J>vg)qfe0#Cz1#joQ@NvnW%Dbb8`_Ljl5{I5ASoRqe#1y;B?vuhe!Y
z_p<IIb!!^cg}-IxG?q16r|8l57#^Hp4QIyqP6<)8VZ+cm5O!BUO{EiRdowixi15eN
zfFn;=g>wh;O3H$q#+U!fDox)oYTHOMl?)Tt9)9#A`Vm1OLc0YQ(SCq(csTO{`|#Dk
z^VY@TmDH!`3G21JRCV0ewi6X)Kc4^CQAxS7->{Mi(!V>%Pxd=E{@_253tX|@#&O_F
z`cXq_IvrMf2twu8Yqa0`B-0<KGHNeX(&fwz5t%&df?K6KqzBEt8;(h-r%S=VqPEXy
zqZ!ZCcz%<n^lh$mrTXMOO2mND3S$W0KX_&P_mHyIT8l;qv-7tHM46nr{C7AWS{HmH
z^bZGIQei}~sskg}6gqEJE@u^s>B46}migDHE$Lg)YiBm?i2V2D0?7&4YN`c$)RuP2
z1rUwt#aazZJo0FmO-hVp(!!KIZm9D%Il;?85RhZMsRjh<ghwa=NrE(nlu{CZrGeM*
z49fF7R>c{cT!$qQCxPVL$xb<j2aps6$fj8v>{wHeXkI%Z3rg{6>RIrC0(Pi**l$F5
zZV&aQe}Vq>o)wMMTh>=<{O|<ss15!W7OQ{4Vq1gh-*A-h?oQfZyakuAEI9HXJ(t}4
zx%>4i`tz<T%cL&KW3jycFNNrr{~f6D@eB7+*K|?Egnn8+e<pdfXdv5+YV^y`p6VtB
z>3*xO6@`sSS8K8L$5SkvxwT%qHn6{yUeb9Mh<nw_;I^hviVrZ&WEEFvO$~(iBQwit
z846#jU=vnKwqe0|9>t>p&<?GbaC;P<Sb{%CGNROl`nO=3uR|M7fW;z)p>bS)q|foo
zt7k7wkM$$|2PF#_Tf(|d$8h)d_k^4TyL<n<WI#)iN=$waDvd`^J9Zb#iqE>oYUtpH
zkk;P|4|B;ze4}qp)X3~XFx|?07~bmQoB1Gv<j1HgY4%7uv#e7Fn=icrO{!Y>LX9;0
zwLKIUbqT_|3kTf0Pdi=|>;4grcz2%zdwD@<Bs*8v14qa)o2@OUnl3=0uy-ndDWf;@
z_4l}D@TUD<@*4}^NoxOZ9p~p#A+@7$!%mi8Pkpcub~JZ}BY$rYh*Hs0Vy=wNIMy1n
zy_7N$;*SVt#x1pBl-6c(2X6XsZu$bffM*zOR4~ieiGfU`%XEm%6NPI(bZWd$p2zi>
za0MlQI0r?}^E8441Ralkrp=$RdIAI+ZbTa77u;kq$>;nfO$g!J_oNg0S}gzl5PHEn
zWk>iiL&oz^Q!uk>*bUt>YCpJG4AY{-ni0rSRt@2^lg3Qt;P`{lH{X_wzS&@ibaa{(
zYPRAP!ZezRmaFtYRBJy}9ofQ!P1-_uAQ)B`tHH){_W6cFwykqVKqvLBG3Evqm*ihM
zEGJ^2ML&j7?I9{myw=70UGQ;C%|l4@MDQu)R^nhjaRtwWO)eXQklRBU`gkRvz!cJ&
zv{%h>-%P2cvMRC+3G?sWrb`N!zpJ5aC~WlDcpA0i1ltBEhQq)+!Z|#O0=t7s+g%Nz
zF^$8DOQZVgy}TOlmr+d<al-x@4&1$CB-VRw3+ZK=OJ3uewd6tS$CN&M*Cg7PT;&;N
zjsY;Xv@!VyUjk@@7GaxeQETiz1L~V4XcC?HeJE12@k7BCPS-zOdUFdiWC*H_9?qN-
zz<qo-0+^coe;lPVv?xR7b}>7b>O|TWli7wZo$IL{8V~fo4&}0u8)<*`?fK*q&wf}b
zAp~2ke>l>h7H`cVq9#|Jy-~rdZ(F@)m87M)=d-V6p*zp7`hrI7u8h19vM{NWRSJl>
z^@pSf@Q20=zbO(z?4>Hw#P(1#gxYMO8B(<6V(&g@4YT$ReV%FS#Z<ZoMYRiT(Pp28
z_OU5ozZ)32qW*W)^9e?T2}ruk_>yf-amsp!-Jes>!f8%7M$}sAa_mQE__-JK8yf=o
z4E?I^15FuX{9~K98sW$t0s+8bU<|frrDRt!W?8qBTk}ZBgH2eF#4n>FrMAA|$5V#b
zx@CSJ!*RQzKq$sUrT)ERI#FA4dfSiC*Dm@XF-7-zM9Gdh^a7GJKXyB6C@obAy;aDC
z@vr7qJATtF@G7X7dLf(Q2bN~G?Vd7D|7n`z(15t<Md*SR(c0^?nxAeDwf5z@p0@$a
zhAb2f2t|XgN{fL_RL~|Z6gHyNNj$GQaOl9{#8|QwnOMMHnN|{Nj_}C+=x!n$fo0>R
z0_a6a@31yyv5#ET@T^lz`blqssjbxuwxhGTu}~4eyXScx+7p>OP(f2)@7<Si-@$p!
zB!Rv5>Bo0Uc9~jMmAF=3rJG4uT*;B3MDW)vE4x|JS$=#j)GX;T(e%C1Ys<m#;Jt=R
zNqZN1N&H}<iF9U0R~xRzB&gvmQ=!s)i*vJW_Z-iAm?>6mLQ<6rL=?S-X@$)%v@#i6
zK<-ckMmzHC|HXMOZntW$rH!sHG6=`QV#!YyI2l#z;m1iKES!^~#wVeB2i%7C2lr@(
zX~*D-x(h@o`1{S={A*<q?9hl%eL4G&Ny)CeBO#;S?KS)vjp~Hu&Wzugz2U;E0~N!g
z#Yzs5)Izvg^6a_KGOrtDw;PMv>$x1U9}`4hTfh2Hh=C3_6AKH=CSu-195*$p4;63`
zUP@vgEOz^>#Khw<mb^zF*`U8%J$(I15k(hCvc{^4qGZrW(%I*Y!`O}nM%S>LxD~wJ
zi~9{5e5^K_eL?3&Fe@MGL@BAeZL2ce)1tL$ux)dXf&$k?fM{B+*vrW&s(<@7LQY8(
zBFaDa)lvqn5_Ms1Wt=p8QvQ&Hf$Ts{sx65)g+g|hbsE?4@z#E2fVD7a8SZd>r4Py7
zY@<OrHU>4Xcht9&frE|lp&rX;hASK5-|>^t`h7Uj(2+V@W>10a$=Sba+Uz#!QDB05
z4z}nn;+eulN`JjAGvB7@8rvY^6wTGZMHP{553tF2g7XH9-bD?6uwh{zw0)f|R9!cU
zk>EwCOWpP4B8k=)`eG1N0fS*jD@QF*q8bU8&%$!Ju_T1TrK_P`biXi1TPl3t-obc-
zbcIfU_K+b8k%f<Qn(U2y*WEIVlb-)>yomCEM_Jr1K(8Lds43M}7PR7SaeT1xQwPaY
zvB-PSruc|NOQ9b-HJ@xl!?zx3T3HC`sT+6H7I_#)?NwRA_CTjWWlWW!n~AdDr?&2x
zidLubhIf+5b+NyS8%bX8o)VQYh=LTdxt!LhYCJ0sOnua9q3o1bo;E~*t-?kEg|Xq(
z(}$UZ^$>K3L=kH1lAwY|p)KWIvE#oCB-savZDay?9A2K*)}zCT!7JGSAfQSTRQ@K?
zVEqUkeob$4uG%0~i;Vpwii*T7kLR=$*1~vwcXp~tgjn|FzMPsNEthz}?gyfTI~$xY
zre1}0P(2t+jIEKTknL|&7fJ=-ev$TXN-VO2C)qwCU>B~E3U^Ih6s$D{@`V@>`9@~e
zfuR!^+1NxF7H7U79^wSUBkIf$Q&MEGK%oqbsS&KGsG97OZUJQ(2t?sH7F$PhZ501G
z%mA0?x3Y|{R4{qRk{b1j)l15Uk|`emU}%)Y)1*&zD{XB$f4RAOBKUb*qx#RSM&oRy
zDD~t_B+rVE-9<Lt=1^@MEg+sozVj@;w^VPMxZYP6#eQnLPtd4|TWa(0a5CSlwCaSp
z#cz(#@Z8AG4<)rz!l)uojD^^8#RZEwCMn)<Wz;<uo~aQs^^-fUj%`82d;ixLw}H(m
z$SXZ+8f(7M{J{4xQS%=_yKmvP4Y~)@@!|4#;h$40dsQsh%$WH11K-}yEr=)qVF&-=
zPZHWFW50)h@vBf*_^h1<Z;OVsh-i`Qc7HpzlAsRigMeH~uB)(n9T=HrGG{+`dv@F2
zusw<*8#BH!yJ!WO5ZcgwTe-Mkdnhzo!go~;ELv=wjX{dzmUpxL#Ue57U(avgAs%#p
z%Oba}X1=^op;`)v;p7JZQs7sI3TAY`W!vK3d6vvd&1cpF9L2X<4A|p_YjwXqpBjaQ
z<njLSBJX@a)g8nTaF}s*8&v-i)Z5M=k0C@th|)%N<-U73mBiaD(T*1EhmXKYDyj<T
z)Ot@1KHzTE>iLflG`NU$*P87@u>_H9vc`WpRK7Evyn_Cx{i?y?_g2t@;MRHK=OMQ|
zlE;X4THePf&@&n3;t+6LD)+oc)qP+z&{Js;{|)(KZD-0_CjWhqiqy|n?H`$%ag~Vc
zOzs`axmSD~`F2Z{n2WtN61L%(f@tGXdJ0-tI@`1TxggeM`y|0wvRH}$18*$@R+k?>
zQza}~f5A7(hKrq~1meM9W7FpmoAYNXS0E@;wa2pF_AFaxG?WrcC6NUc_8Bgb3B>y%
z*++Myf5?U$SiO28Uws0?+Qt!mlI{YL^;<(DCpSyB4{JW<3CJ5gxZO|%d>L5+gLpkT
z!?%wM&+wxbvr}AbKbfuYc-;?_3m0n8j~1JNKw(Pk3A`l=SQ<`5Scpt9?4$8n;F|sZ
zIw;x|_efE-Iz}0VIiarb*3!}f`lj5oJG6b)#|5KmJV0iA2?o${nGs;w*L5Sv&P?4@
z1LsyU$H}tY3NP7&@>d30-uUq8>(QGB?uVXdc=vFoGhDQ)=QWQ6|AH+_U}?L4T0EqI
z&)YbP+ww&iHkK9<Psim=I0T{u(N<G7W2&N{UgE|<1?;CWO7_wC>848dt}Kn4>VnI5
z9S<S14R!p>+3~8HlWeh0^Q&OnM(&b0ExK68lOIIQ#x8&XjJ*tR#~{Bki(OM?Di~}N
zT$YOKJ0Ziq5L#-oO%$jF%znSqe_&S^*usmL%zTPx=<mM{w)KFc2yDY^cvOB`#F4{=
z(%(6L%5C_6G<{`Ml<W8QFbp!JGjvOrbSTo@4N6LvgoGmP3@KfLG}3~Aw4x#{p|nVd
z(kdV-0wOBvyXXA=?^@@JXDLiP_r3SN_7xmWWFD1@I`<q)N42jsT#xU)Ma>@L(;Jd)
zH!Zg3CunBh;_@6;9AAN&uh`e_bs5kn=<u_yaU2U7JE}#B9M^ucytk5@%aP7@A!a_^
z)b8h(hF<#n=NimH(?Z{NUJ^|a<<UHDppWa?X7-#P*M2VcTI3PMSm+;IB&^X0;^^1!
z+cQ1ly@1pW0OcRThI7{Mba)hqjkJ>KruhXv?fP7tOxE+7?7h?t8EC+UFZd<wD;kip
z57s>&!<aK6Pu?4hiYORF_F6NUbEqi39&W#?2epbCb1_rlMRs|0S3N7=ELwE^WGA`{
z^nsA3!O(}}p$kb|5)5!Oeam<TTUyb3#iK{vvaM#QxhR(;Xmez`!A6?z{(?&0ju|?X
zn5ZDxKK~Y^aV}qfW#s1YRN~Ip1kUEf3Mw^O`kz<jr3Cjtlt73?q9iXXmySU=j6&8=
zo4SkP*ZHQE0?*Q<aW9}W6a2h7>(HI@Az!+10OZ0O5a5{#i!}X^MWww5g{t|bEgeas
zvV0<f@08v<_|PMO?49)XS^)3Y;Bg-3!9*-UL-s@S;EpDw$IF{o>_2)(=kvBRQ10eC
zHZxbU|BhaQDsVQ#2EQ&n5_bx(uA^qw4&}Mp^=(Mz_w;)rRs1H8$$MrJ1kT*CR<%Qf
zZHN$eL#LI0`3uO`**<n>vT^C%;@v#Mljn%KVKI?e#q`p4^MeunQH$;ne`D<71WeDT
zmX#~64aA}gMq#I1u|OnDj9ePt+bYeeVXl)j49z0HA|PwzErL=^>;<-ilHkQq&5F`(
zzen#9e0UXw?WPYQEDHAoHnJul_&%HVntuVkD#qu}w;-?ov;o}(c{>Cpdz%TR$DIE&
z+$q=ddR#B_#D8?xUe!MaT{cE)?!je+^Q7Y=)~yWgsH05k7nlA*SO;Gldo1=tb~Am&
z6Ieq-BWX}V!4y^-iJd9pEPrF!N{GNJI_F38a3)V(6$h6(6B<LyvC&}sI8hZoYoTK*
z9sgEWLWzB`X#CB?{E<WCH?6j_$M^PvSm$aXc4hwl#gt^T506TfMiMCo+q_T?5?}P*
zyPmpv$5j|TYA|+CEMi)w+SL>p$g!i%u>AMW@89aS^Kr%J{8;F2Lx$pYzd=JPULUiX
z%+v31o#5ZVQ>43NQhvesoV}m(0kH*r2Pn&-ZwG%1o#wpwFVxLt6782}6Xd{uN~HZF
zE!+@sy*(KG{;uQQ?|s3jTFeSiSpePUQ;;g<^)&=Je_9yV&0W_FrE@byR91(eWL6tR
zzcIa#7@g=LwG?rH_~JRz2&?y9BiuLM+|Ov!%1TfaJPfI)R!T=C5TEpMFiNq1EA!#Z
z`2Z9`gNL+X$ZV~#*w5~w`>Rzb3*MNP#gAEO?zK8nESmEW-E1S#0r{v2Vd$b-cG{1>
zo9|NVAp;if+5CbkfU!q?fevWvV=byMCmDR0YLptphAwsmS+o_P|KUbrm;h3UHi6-3
zr``Dgp+mD)tCzNZ)0M;9bm_?@%jqZ*4F#93!d!`2`@rY)<?(#&$&RoJjMF$}9mTCg
zR=6Pw2;mu%A7$Z4r@uqb%z&;PnoPRg1%<QF$mEmy0y>u`NtipkSfwkbL9Y0)ed-2<
zRcF8&$WRscdU3=zLfGQ|3ev?WWeHrx<YZB+tD@x+5G1s0pX}K0rKpxFhLt_#kZ@8p
zXE$R?tTlgSt7*;^s4*3Z$#?x_)y!}id>D7nY!&PjV6+KIo$%U&T#pn7O)aR#-v`PT
zUqkmQ=y`LgO}#CcNUJD~x`l9`Ksp{~E|rE(y4h&rkh^4PlWGH-l*mWiKh0MOM**l_
zyC<c8dwM)e!xDB7&TR~mVL(Kmgj{l>!93dSswr}jAezhHMNu($#Hsgi@g=8LaSGDf
zv#&mJ0M9mPqZ;c+4A4KfYdR89GEx&05T9_YoVv&T)&O=K&j4FXRbC^Clf8p1#Le9b
z=y?y@o)>xIXPyslb~CX^MxlTumwGpYt_>vP7<tkH{K)-RI<G^NwjNR5K2?fH{XX|m
zEVC33Ed`bFGir8u>*2n6QFu1KAO6%1ZBR#LYAjy!8FYkP%r&82yDe7PT};liui!GH
zCabR?Phot9Ze+l*oMoV|zzlbV8&b$uL9PmZnr&-FwUDX)E~;Kd0bBNB6G9hmS8MV)
z6DgEo<vgrY(l2~qI+-~YJPj!4X?A%x+mudPIS)>S_Ya?nKUUq<o_+n->EiJ6{Nrn$
z-fc`(qJSUY8Vm=<Ji+4O{Jr`aFNxEf`0^=Yb2oETtB77UCk1Zd2`2gwwbyUnXe+zr
z)wr?x>5lDgq*Byh%<_nw4@gSjKOVmBo%{7@q0_AMvgpkCM-Z+CQBe{!RA+}@-7^uG
z$+fp>`v#(+0cN?)F=;kc>We^}qmHzc^?>OITrNgCE17Viyj@wfJJxXD7;>o=3Jd8T
zAL(QO{{hT;p;PhHr%Cg{f=ovt`@V)@%kv|d0SQfyMQ9au-kW-5i1nNV8Phns4Kl1*
zyHKq0{B@$)Lg#wFd(sh%nw_CQO9VodzXE|b{?9>=g1?oxbPPrV;ZQvns}o~hP+Z%r
zr3A4+MAu(~96{{Y`n_U-u0;7$gFkV$Ru@7`@y!<rW5=&|5AFzY_r~6Uw9!tNiU;>X
z$*W17d-yBFcj}#R5aUEUs>741Y$IFQihEpa8aZ|Z%|voliK~^LoYNn3*;87It@|zy
z`ZM!(yK0}oh)2z(IwQu19r?{ElIxb4DWNZqswsNM0o#taDSmnf%W8cb{Vu+$$eBYJ
zqEBHrie>Vd22We~euuA7T~avntR4uRS=`N<7oU<n9nxne;!aw~%{tND;y%wx`RRCR
z!G+++t1k~|(wWP;vrK(>U~sXVzgnCzc*D`sJVCO&^~xjV>%EFZz6KDAs!bUej5=91
zpX2VCEIYk9T$_p$CDt*zWCLjQZ<sZ~kyHAURiE-JKnacp2^oUs-auv{X!rQri05f4
zCQF9|&$wRK3R(e=Gk5Eh6LZ-FiF&1oBPV%7aw-9qgXEceCT9e~FWrS}gf|}XlI3+r
z*`s=hMsXj}v$(jUV~+6X3WOdYiN^NH^*Sv*oP@;d2AkWY#aRg+_Q8<KgjtXNGjbSX
zy(`)_zHnd&-qR;@1tsxS+5%N0Wm+3GCD4<@om>QT=BjX0R~2qN-!CBh@p@AlR{+;P
zmOLSp|D-mtNXmlkGZ@FTf7w2rFMu$xAvl3LJx&Vq7|!kxY&-Tjl;G}7N3r6VH3Rj2
zUZ$#Qo+_H-?ss2Q9cWD;8%=zDkIHkzXk&D^Mg0B{r@&vfc23=7*FVr&V1iK4<RMjb
z!6HydD88f?e09jD_BKSLgTDe4;quFHtu5-Q1V;esI@1-ep?Eo!Ra(COk-eDwXJJ1P
z^v$w+>6`9ce0ID@6o!f&C!f<v=BLH&OtvFg7jIgZy{Iq4d_qjfqMan?p)YSJZ$(y}
z>uK->QnEl)-}-yCeN=CQ&f6`~M{ixsemKe8?hs@;5EMO$LxhDLk17Q3pk#@$+ImUG
zWen3AnqxuqRaz%@V7TS-or=R$*+sMhp#~{J7ml=EGx>UEmf>EKzKq;9LC3KXaZ5RQ
z4j2#ikDTF?0dvL8OxJav?_ATa{ze!{J*BG9lV6{jl3uS7>13+#t8;jJ7^HL6_aFC1
zl!VP}kaCPR9?9u{Ng3{H{u+d{Pus$x61O~E4F|c8wt%fn$$E~G2kVS?Qwc4WTH!$y
zx?t=MAz#WIS2m=0`<PRr6v3fAWo5=H+0sq)`_fzm-l%r=8xgXL0$ZHXsAOYHdfm0}
zy(3Q-ngV-MBENspP~);&PJ1fl{12_y`C$hiN?_+5r}NY418)VKyL7JR;7^wnpFuts
zbRBOX1-M^+d%<*Pfe(eg{=Pb9%qNm*FNgJNDhlVi(e$AtO<?}Yy%LXe((1_)yg&K2
zV&C{dlALf-Xbrji9yO9%_-Mkulb`G~<MF*Rb)M^qwUJ{Qyk$z?XDu2;Z`xe6A9%5@
zYW=nh%O>lc^ZaNZWsQ1KH0~z4|Dj9wzA2~b`q>F8sp>|bm#@C0N>4~MY9$cgA?2v^
zL%r(99TJgdX7qnIvzJY1ob0wGk4EsQR_DT?So9>0sgJn9qt=p#<_N}}-;kQN7TcB`
zMwP_1=xsKAW-keI{Tw*>^r-5DT~Vj6-B@Fc9^er$8v}cOvW)JTSylY3$IXleE;?^s
z?o;(L3BJ*{`H|3k*LZRIpSkpl4E8j-IGIotPT97Kpv*TTwscjCIbzBhh`Xm`%Nuvg
zdFTh!GddDQW&g2kQg=EhCA*RFgb?3HvuHm2u`l_hgo^c?J-B%}FqpT(xmZOa<LKf8
zXg?sEnVDlyGQ#P%1>aR@Q6P08>y3iqqBd_J&h)YB+u^rtfp8c>6)LN;1N|qI@V9`k
z&<u&m_-Z2VH$kw2OH$j6r-UNWZ2=ShXYvrIcgd34oYf{JMy6=yIS!3sFnBiyeWW#G
ziJOdwNL6XyNq8R#BNHQyo2CoP6|drNW?ZM&qd|^F+g}knch62VUR4*fXBK#vIi~x<
zn}*F2;-_0%LM_b}(s9GI+bS^o>Oa;va~aqY+MSB2>a@x<3c3}^rzCaTu9z{VPHRs+
zdw%5YL4jOYg`K{iY2BNd@?izCMgi41?u?ZJ8u3*$hRQ%lzdt3h{ZXxOxkf;I|Chd-
zvm!!9M~eWUNOJuYq@t?eYtVA87jX)_^V$%Nq9^*~bJ-M&LpM8EigG$J<ej2W2=3D@
zIhXTZQs)gam5XTZooU)USp)rxYk}bv8=^d!4%q$G!>yeaAGC-=*Mi{5x@=#Js0mD_
zoL-er?3-W9v~TcH^xGgk2FiSM(pwp{h$=GFd;L%Z|D&8KNbz$ETm_aLqnJID?akQs
zkxufLSln9-NR>xyQO>V~?iC|{*J(X}=lc+phF~x@w3P8gC~1fWsiux_kySU2smi)8
ziRElaSiQ8K1SpPD^3B4J7x3*BFpq9lYL!whOQy4Nyoc~)i#^MO812yf#53FTO=pC(
zQ0@_GyrZjF6!<giGF9rtk>xUlK22i%bAmgVByI}!B2LM-u}MMVA`H~daN<{TLyZ$T
zL|%AH&}!MHKob_X7;i~^20Ix3ct|1)JcK*s*uD8tad_C%VjV-*;u%=$3mh@4um6D4
z+m*MSlK%U=mk6PlHhegUVCIw>F3tX4-&<TD>%&mJ;&T5C6Z<=>xUp^+qH&KkZc!6P
zXQ+TlOHeuOfo=RQPm{-kA^P3Zkm0k74`w_H3R!OELip?14smHdyu!wM#w~mk%>Q2~
zKX#bGF@KYKY{<d(JZEcN?ThqdC4RAB!4x_nC{%PQK`nfIyYa|7S)W_3W|FKj9(91x
z$qlIVnG9O@SL<AH+{U>^t98s-5=5$H5kJ0UJ)b5n%d&KmY_hDQ;iD)!`vh3r8$>ZG
z#44m0>o;W4y<y@{RU=6WL-F0P4+Rhl*sQ=U+^C-`3^N_KFV$CDx!Y=y=VKpIylrvS
z`Q&hMNyqBeH%upJhm-0C?=TX!OTfhblYI7p9MGU~#AOEL!XvghNUtZGIP~s9OY9!Q
z&c^+ki{?KD%QbFw!^X7MVB|NujF%~s7dz#1$uv;WkLkp{H9M%zfFfUo$BQ%z;Xiyc
z&{d59yD}ISpZkW)n6k7u_v(ELOf#qbx`F#!D03zI&HE?hsDRi$Wof5ggl|gztr2hD
zYTP$q%J)8sEB^^gbUJ-!ZM?168=3IEQ-Y>th%5N6{!r6%Pc_LO6OP)`t5s9q2f9Zj
z=n0zZatmdp_;0C3#d4DOGYVyW)N?c<gJ*x1-K4yR3^4s@8BRt-8@<QWZIr7K5da#S
zfo@N&KZNvn!!T$I8)c3zvpNV|#etJ!?va7tI<Vz^7WziL*BsY%1ltcGmZR?$%|!U?
zI%g()JvvjsoePm1XB1zm6rnD|G5eUOD(ihM5W}5|E?GDND?#vjCJG7R@Pn3`a$_K*
z$Q*R{{&gBDBJZ2j8bj1boKPhZaDH7;F=#WfuRDjHv;>P(@0~}h;%H85q%~d4+?&Jd
zzEBPV3%J{<+WUpy|3*z`BbVU<y!81fM<B&edp&8c$q}r|zP3GT08qzfU-rFnaJeG}
zKS8NYao>lT&;Eq_;ie2$4!TrA<A$~#msXSSeP+bqmT$!icjtALr&TXOrl7wjs{SId
z^0ACtbHdMk+)Hs+sgPnwP}<GZojrZ$iDbK(lNqJyb$v97gy==3X6_()vd0Y1WfZvi
z87uBXS6z5jhvVr@Kq1}#{XEFG@5bKhY6BZOQ^)@EGzL@cSg_1I1No-pc_g!z(3O|#
zQuGNC5_#XS*T+(Fsy^5PRd@%$K9s4SH#2d^5APWjC3K98s&y?tYS41H%hfbWaIG+F
z_ygiyil5>FC1FX7cA9r{=$X#>F2H=4WBiY!=LpVjbhVvqf|9ZkjS0Fc72#f<%ph90
z|NhBT5qi)2Ry|MD;(|)%H9rHZZ3<Vb(?QYJYBcKHM=F<W?;GK<s~q@;Coj%6;i_~e
zV6L=l_fnJFnW;8-dVQPb)om6h!I&yrAk-g6{{UM`UHx(jGk>}ustZ>N8?=a?)|Xto
zMMDpx{1*V1-qVz@3r#%}XlA%rYZ^_?3G)Y90wL`=F}TVk%_!b$YamMiEbvWOAZ97l
z(iL=*4`MT0v_1s?x*S%8y9M1*130?^C5a<zxXq<Fc+O_gygcGkDA^r;q+HSBTL+l-
zL3ioc;Z6S)>ze=<)lzp<gje>fUHiNUA|jXMQC`3!OBmx?3=%RA>{umU*t*SsraQTL
z^mS#@;3nV!Pi}l>iS%QCh50oEkH(iA3*)5o{h(&mNoT|9y=xDpI=Q**cvf7-yxF%`
zZ!KhjuLsl29^KHv?srC=9jFod%XGq)BMe~x&}&f1E_3*8i*A&dK3s=X-|-TKkRiTr
zoroz$wus#p4o36!MTjKklA<D;T7Z`c<G2J()*E8n<C{_sHA)t#P2|>4`P=#HW691N
zTbj>sS#~<dMQKsLTF%;}oXGhwXEM(-*HkS1`}wkWkBDU-UL{m|!)9k;o7PR)d3#z#
zTT)l*klh8MuNOh(b|O_hI45S^jDMEyh`Lj(-b;cKjG0Jp2D@*u)M_rp7{nz1Q(ms#
zD!WHlJGITXGN1j1lU>puPkmW^`Ue>;KQ3|Q4L$@E=}-oO(cZzIEKp8X;B+DYuR7*U
zv$KXTx$;F#7ta(#mJ5ba46M7C?%1%d=Ed_0e)>sS+V|`oU~unR0D#-$KOCRhqi_=5
zdHYzJk0y$z-HN2z(Dz}|;*`81z!!=(&+cpP1gU+2sJL6Gq1M@_H+J>qa5kP}`D@lc
zU@$%=iw0P;Bnk&8F4e#0yPA0o4(fP0cgk;oUj??RLxOSG@!lHvO<T1x%Fo_}&B81(
z_tpOLzkgo)bw~cZ`uIHqZdBw-i8lU+pS4K8z4?nPanD|<EJY-9Ru4H0{>Hw18ghM2
zC{{p~f<>C!CoANtK?Y90|11&xROrAxxe6Vv!(40F9!=a4p^C9aK8&@B`IVva7wRix
z<j2@*Cvbm;h~T!St1@I~QNG%ECi0#6UPbj|1n&4Pjoa)S#fML>VTnR0CG+V8Z))DK
zBQG?&@KV9Q^zVyCW1`PhVr%XuW14zg45>q!r%7bFm6>QaYj6xz@6E>(q_Re-!@+&p
zf=+y~zSyS+5C$+lF1a>wVS;KcuN1JAmxk`z6M(pU9zVabNk!!GX-pT>X%_PNO}w45
zUB}2in3=rjj{4o|w_SPa;Nzs{Vn;>yy_^?!RsI3Wn!HEYxS)y7W35}=$i7BTDjrZq
z(!!!v0SBndJGRTNj}s6z-1GOXRg1<?xx#$<;ELyORrorT#I*p7?0DvR`bJ{U$F%dy
zPID6qUlmU$3uH+pBFwnXdEECf#Zr|hBOxjF5KvjeP%KByVZ=+P)BC9W)t9zz>gBOa
z=*_W$0~nufm52?6$mh}^n~!q9U3AGH5cg()-&oP}p=i`>Z;xz3Zt921CBuq0DZKz<
zffReGQs?x-*}C8$;5aTV^@JbLAqUZ&LtA9SqBpV0ViDxgRdIcJbI`atSzWIXBlV^c
z?|m326Sh+zEUQpg#7mj>4XNJAe9EZ0k{$&`@(xTyHf<g@S9ZtGWsdcC58^(o6s<oN
z$!*XQBYv>fpN@aN!0p8S{VVv4F9Iqf!O3In4H-a|eS4Mi<3;BMzMneZdL}+^Or28}
zW!(^}wfSM#Q&qGGn)aHxj0gKzq(I4abLgk-X*<E2wdlP6hw0TfiAU6%j*Pkw<y3{f
zFSur?o!3}i0Uj*DksUYTBtFz@RjwZID{d9L%GK;O&F!dzJms(5)qq{oA_!e6+w<ZL
zIhO!JFE7X67E>$Qi{$e$3;keQ7>*=LY0(<dN6v!JW0Hyr$>UMxn?RMb=$v|NP%Am6
zATaF>&yW42q6Bo|x4un0=9+E8z(5!J8mxyM0P!-%856`{#wGGUQwo+9bblGsWTW5a
zF#nTY6IP}Y2gV}pb0V%A*J&knhSy(Tg}pvZ=3-V4*utH~yP>7Bp$a_;hNdY1KDzV=
z6M!Bo$8u{WhuZ5VeZ-=6=9^6P&IrmXa+|cepcZH|Md;e|Is{3kt|wu1u|$GgU>3gU
z%11lSgz%XLpM}2wSnw)15Fl5kvi!@0Q(n63PJ@rSY=h@afoNs<xQr+C$YnKr=n%Lc
z{f8sbV02WUWQMri=Izr!NTVLHX_K$i4ShmsvD!aux)4XBK)aAPB3pSVx|~Ju)0bMP
z)XBIyTZ?5-x^?Q}qrW(HynI51#E<zL_e-|k-%W?{d0Au0S^Zl8N;^9@frx^9{?tAV
z^yU{D&0kcacUOAuzRUmq%nD$^Z@X*ZMrvnj1zSGd3n-_}Q%xJ?AF!yFynXTZE@jMK
z57?W#iCac+t%{=(4d22GPQ>y+GruW5%GVf-S{1N;dZQ8d{b-^?@k{51)7F64_FIhc
zVdU|i^9d;o5w~eeo~{TuYl7MOC@8s1dg;VN8SMk2e@l4{rt+v73p%edGlCe6n5aS2
z@*y*TUSDUScbUx@h3%1UtqKhy1~;leB-B?#xxh^bhOat&0;$fhEn`0ypm+Iw#}HAi
z2u?MQ#9<E1!y>5cMtM-XTKL^ISoAF|-6Nbex>Ggx*;^iPWZj2?qR5sdEEFaY;5Ow$
z6@B)rIyl}N>3d9`>ju~xIzK}apGkeeJ~nmy;Q;;FOQyrCFzfgJ3LBFQm%NVNmbgI?
zF%5pwaenV~UTo-d{dfoyuyQtT@rqf8pj=wB|1ZyZt=Vq`%bpTE-HMZ^jcJU}-h$oL
zzabilHxorG=-kb%!tP5p5oPi$gBMU~wlIlvPU;=#+cnW>qOY92-edgvF-78Cdodv|
zmcGTjVAX0}(~@u*1NOhg!mnu(b^d91#0rH5Wggta90S_t81Y#acTBW!Fktz-XAnPN
zcpXRq;rhm88$KukWZtj?_oSh@?dpXQr&XlvnU>Pq=<J-azN;cWX3ZbZlYMro<4moF
z=o>xz%9;n92<d*p9SEE1$5MrVVw>;XEA)V}wEOkVI}X-!zQ;gx^w?wm@ZbrH=qg&(
zkv)K_dbtOO<&??pLeNmI^&haDmwxSuIRY&F@1v0<Wiw<2G(}0aa!2evB@lfJa+=u|
zEZcTIzSaJ5(1n=}`R`;%UKUygol&eq0xWE80d-cNKY#UYV&lWepZm&Y>Mp;?h-lJo
zojzmSn;u+F=EzysA0825Crl-J7Pt`Z&ow=9arj4=-pt{L3q5OE(KcaR#r?a_jbGaR
zxK)Kkj`9R-O>rTqvU`lJj%kQ|-@=}c+b|aqv2OQl)ZTakNfR!hBGzX}Ypsbsm1;ot
zCUmd9Hy|hiW0#RK7-OQf2YzW5v%dXtWbj60Yu^i12Z8P}+&(nvqYp~L+!6LUIJl78
zACM%3e!2MRv!5?n2rKFzctlsHMV%^xjtD7{WX^SrZugl_|E&CdI`W=yRg#+sEEbkR
zvWMxcN4+GSr}b92@Km-pOjgEiy^#lZLcurbQmv^oR1aD2S=~lj#s}P?ZK}~ofa*hl
z?UZnQFsUM7RJb`#bJv#!2eP2t^czpkg`xwVs%xFc{XK!rrU$G}C5XX2Q!nXUUy;|O
zw@m1sQ-fB%h|<||A*L^WFRmv)dx;v9+beM8)en!NLvpG0Rs{nR+TCMGZRnImq(Dt-
zURak}6J+-H$!R`6I^?Wfs(fWepN10*u3Zc_fmHz7hI-37T=oz8s&y&Fr(p11k~eSz
z{9s?VO~}_n3LeMcLDg+za)eeHY4rCFT>wD;f!|tA+}0<(!;U?mOmO>zE<lHtK>4ke
zJ2T9vL_T-WP@JLwyGUo3gI$oV#ptpS-LW+flh)lSXdl#JcXt~htr?O)6B^a#fftrJ
z{o`dg1E38BP!S0itBa5i2%c!}MnyRbBZ~noX*Vp8a(mHv%95NoY>_19@+w418LJ6h
zrg1_gp3SKf>^=BI@%j`so1DiOlRcxuojZnp9EE)PIfV6lmeE{AYVVgMwU+Yzk7Q=N
zaU`g8<zHT2>`vnAvc2jD|74w)jOr9Vea3zE`w2VSmfu#nN)cuk{ndNrk=H&>bSzto
z#zyxJpTUjw8|cqJ3H}7VcWl~{PWZR{So-fHSv<=*KVdxtxAfy*JN54vGxR?~##L41
zYw&hm|GUL9{Sv*ZU$vi;$Cu;FrBen=mVjqir5$&+6YK36c13s5AW~2NG-L{n7xmpz
z@Gz+P(Id`-1R3CBC==Ekajjz1iV9QAw*LP928~@u>vhX<s;y`YdEPFc_ty|qsDE5p
zoK)fR*d2XJf2fM^O~x?F2euhTQIJqBJkg^y7UdEgS``xkf1<I&T_Qya;~IXgjhNnR
zlS%??ariuJ*%uDh>_5M8KO!tf2&zYD#zgHN9@c~w<nC-$t<y=Jt(TNXYUF6~Cw}Gq
zfeg>)ibXq>!oI~8;Okde&d~ktnMQ)+#c~KE5AWt29phei?e<vW11bq5>ab62a$`t#
z;1~Ffo;+uC+<ntQG32PUfsFLIaF68?)pf1RH}dD>UMczGmL4{O6dk(W8?OXiwS=1E
zTejdI3|6^UN6Gu;YkDR9<WWs^;(ifmE`a_QDIX;tB5J^g^_yLy;j+Dy>1Y4?+xKML
z`Lq4<8;mFkrYIosF1Pzqv&d{g$+F5t95}dK9JX>6*-jRfa-NRH1Yn3<MgmXws}_7<
z$~4_$!QmsM$8XwqvIW5fCO=}iiJWCohx?q%iL;k_a=wPUy)1^6tA!!tGgEGjVZr*u
zZj*Hjw@InuL_u=1w%<bY8Vq!>dOJWjL$hK5H-n^g4@5RyescX=L){4bo-FxusuoLf
zQm$1;6x<A@7p{F+J$+C}zTnLDWzVA;XCUsx=aa?gp(ceoLzK#2<-SN$uDbjc%EQ;d
z38B`lB)Mk*0mZ6x%$TA>5k&XHBUH(gI8nQ5h@A>}JWs%-XXGKItVq<EnsF`N**+g2
z_NY*|0%rX0kgaj+P3%1Z-&neqLb*WhE7*m5GkNw(C`2^iFp#n-j7Y}m#!qo)s^I<t
zutM|Pm4wKg)|$St-LPjIhH<jAL9?uzR?)j^a}}z_^*Q8HNIIoD`^Wp>a`&uZ6SlU3
z18Tvn;uJ!QajFOmp~7%>zxR_yJkgD>!l0XzGUJ-y3RXE<a9u>Iw+O^U!Q+$b_fm#6
zv#J?dHrR;aY8Fk#UILxx1ugCUI`I5~y(--YAwIU}DB&>y&~RtLBrn0k`HmRh3r*%v
zXyS2ueWZ&!>N0!=UjR-i^m}>8D)*60*YTYAA?N=-Np$okCKUeGElTEnXzTbx3Q`)M
z*}E=iG0Ef*o`6dTa@}tZ<IyX(GjMEiLS>h4muY(FtLFZ;6BQOyfr%v3-mRYJiVx5~
zJRXkE_(}k)b764CjU!cnxLC9EEp*oGP0<si97HqMYE^D$1w}*;Um}M=Jv3lM$2IH`
z_YBm~ZXj@+6@7JOz~U6~9K;LUH5TaWhRur8WK%l@KXQ=aG9Yy(IF&!`=U-QD_AUAM
zfRTZZf3jzuWoI|o_*}8_-_b!^>)`F_9jd|}MFKQ6-{d&ig6?mqWK6xI`O2+%xtfB>
z1>yeNxXkFG>*E9MtZs@;z=g*zG0LJ@k-*;Mc#8yww<kdQLsxR{bNixoSDLv96y%ZL
zLaKKf*S|F2KJR(7L(d=OvgrHzn?Est016?O_<Df+QdhI~UXHH5u=X+v8OI3A2FmVV
ziPs!1#Q8;EzPYMVlVmSyT86h$&0||rMR2?Yv!onx$5I!)fED%d`=N0GW0CE5B;MqR
z*hu~IZB{-Fa!>a=!qL!`vaYp%YdD}|eSg+m3g<Up8jt~I0TGZH3yct6T2Z5eVvRt(
zhdzzeFiE?!Ww2lw+&&!?>QA?WHD<uw0jv1&k?rZ%J(v*@WEXqy`;{KtKi}tNa{Mh?
zCl0NC1?t54jcN5WKTuT?PkO|%Iz}81a@*Uj1K?~Lpg{aE&EJ?Lc6=VW$A0e03=YLH
zg}a;nCn?#N^fX86pOSZ;Hgk(Bx5`_4VXJBA7yR^_1x}q)42h9+2<Rg*QT&EdX#q3F
zyMWjQ7VV^z@dN;xNqYHtrBm~lg_aFijZ*wM0Mumkb#k55d!Aa!9%zN~S8A?h?<>(I
z=-{ES9Ts1k_z5<lGuuWjIIZn6##EoM@9uWgToe(!%LRb_8!2K>>?JOmjY}}e>r-%L
ziXz|sJ9G1R4F*U-Z#|ALAmDa&Aozx;rW#QOkNLm4kVK46^c8t{aw<%V`!Y|-G7|SW
zFnV5MjC?t8nBCIpJ;}x(tqG+xwc6{FEq4d@h{LU{#^G$)tRP$$Axj5VsU_O6<Cy&|
zrMWA`uW_{)&UdL+<1qbUoC9B9w2Idc5n?q)t!8u9dO~i6rURH|Ght|$d{W~`md<K_
zV~GAYkhm7V5R<R$<sL9YCuT4Ap~`814+RFV4^n<M8YckRl4<qzdNtj5f)pMJ6gvIm
z5B*6AxON64C;m}vmBA%na>r~N6S9#|5>#r(yCZK{Z%fRlv&bgXmkM5v5q|LaL!rd6
zlljv>&|a`yRqnhG97j!hQJYrHx_ssQ#vDx%vkH{*B=o?uFJ6Cf`^QaYb{e8Z+UT=p
zwSo2Ei-lW@)SvOBC9lGPhIG-P;*^^DDh|%wb9)-zG?5)7s}(3M6O&~Pwd}Rj-iOZ1
zAL6AOw1&!nxS&7^_0eFrw9ew5u_9Nnaxof+oIcQ4Y9}oKeU9{LkD=CD7j0EXG~$~k
z>@5hax9tHp=C&~pH$NSwOToSuQJFWkXX5RN1Y+`IB)5SH?8;Huv5LFW?(&?xBAKuk
z`ctC(G##<{7%EYai?un9;9eV8?5hysCCx4?W}vdX%6hevYo|+@Kqf^ZVFAMTJb{<Q
z%SNI<`Gq_Q<)52+@K5_MeBCT-iROAaR;<=F1&`^ypOX8CY^O1AGKUM?tad8t?F?%-
zhBSyD8&>I>;V#R1P2qPR*WUic6ohi(=blbzW#SrE+o2#LCR#RWx$7d(Z|+)5xOX?A
zuPwUgnaxC3%Q?i(PUt-Ne5i`xjA(>SRn45;Y7^}B@fN9mF`3*IHClb2;urZ6OGrk_
zALGZPL&)xXdDaUZ7Ej4zf(razUT10Mi!n-)*|?6ApN1qzG!r^=FQ$YL9uWO8bv|`i
z!wE=+kAvPP|M0_44p^%_7NnH95VR>VyJ>})<UoC-n3P-c5iJ@L`ZauT0@NfW01*g#
zFT8V}cO4i0)kNg?w4J`L=*o#tl^C=D((g@s^te*CF=K|VV$ek<^-;Q|X6M*v)>=5*
zl+ad4h>Lx}bwrJ=S_WamjPKt@3d2lI{Q=Y4a@O)*-VxSkbSnihAByp)YuN$t<vJvA
zy4*!i=DRvut&d-WMq0AR=z!PEx94P^qNKJ}wNQ5_hzpa1`1E*PgNKN0Eim6_xg>n2
z&eBF1It9uIW;LZnbH>GpTlrsc|4<Cy<^ia6Ng?=-M(CLN137ZfuWisAF`GrM-g^Pn
ze3#}G;))rd5~*f#8~bn4J4TNCTAOe4J%JJ$Ftle|>~51Hf_$Ihk)gPcCdnn^>SOqH
zI*HHy4O(J8XMqX)PLxfG@A}%T2XC87v79_5WZ<}f?MkLSq6K$uaKQ0nJRX+z>YPHT
zfJY4u&kAorTn~2?O_v$Q4wV;RULxRi0zr$wWTstV^&^m=TtdT27TP>dPc8y&kFV3z
zF`kxaeU(mwZlW%4PKdL&;#WU~xM|90szi{V-}x$8mlVtfpx#{4V#P#Va{|V5+@Gw&
zp1lVo7WebBnRm3^%BNO+<suAcTPj~hV$@LwiWzIC`MN4JYUydgIUTMa`Fo^)`T7%s
zBzUPaGktHeCJ?LtTz@{5pcb{A571van_ds?_f;PhT|UPDmoY%xL8vz!K6#v>cYK_w
z`5h^{XF~s2UzlWlX<?#tvFB!#cd<qKctDr40FDJREoXHvg_{;3c$`py!<MGeWQ3uP
zD-%KFCBqT1D@oI5y+nJ`={^5|N=eb5ne2)F(b)l8P3Bry#I1Un9PKxE0I0+%J@GXS
zFu>%*x73OH<^xFN><j0D%D?O{w_xgjLEjamO2Tqr-nqUU{z+82UD>0E&bajh3(iVX
zMyh!!A)V1bfc)~$LmEt3K&wf7-U71@;!^0wa8aNhuWAhjqPx$&46V`r@wYQSx_RC-
zif!lzYNm!w#o#70ACGQWy=})1a25<Y#pp{UY1V%6)2BBtqGEUIGdX)XmJ($tlDf?;
z(4BZhwIT#h^retZhkZ+U3CIL4McShzl*(Mr#E4_rCe&wWprrr3^V7|Lj78|84{9?Z
zn<H7iRf)pF=2g0xhuLC_+<1i){7>C4-u*%?^Mi*m{q2_@`r3#b0&o)eP|42U$u=d`
z0OJIMAsu7$a_ov-cH+kndS2Wq?AF`IZ5krS%$E8ZcA9wdZjgtpF5p^R7bWewc-zMj
z!YyB{Q$9?Ty!@oI?qSozcOz*&7b2Yog7vS)ESA9p|DlCiJslx))+LMLrxNG)5><Rq
zw58GjZ%)}Yp7SyE%4$Vb-7JvV2|Kd;nW`yeqqc`CWr@?htNXAhiGsBlZ?5Jugq@~z
z+qkZd6l(R00COV_>@$F}MOZzpJ34u|&uv}dRvr2ZG-@<U#eQMd2>$fB(vHNIRFe4%
zKG^+U#2Wx0Wxmh6SLakj+e4_~n%Sj7ZSFmYO|cLFv}M2ZqJ~}AyQ6oEr~7m>8_yi<
zJ3R{=yYnXMdE9dyfkl0_TzU(qf@GKeirea(QK(j2ump&+_JNzNhvRG`9cdhQWp_*I
z{TEX1wj|R_<&?nx!3EeFbiyCru`<&3vHbm`akE{OQHJ~Wna@bY5cAJKuoti_4fmW)
zA)-Q?HZwer<$JICCzL&CS*FoN*n6gGbZ>2dR%Xr$Zdn*>RKGY+mMSe0SsZ<V9+kQg
zSEu__{0~DHzL$`~1wRbnn}zF7!_|aGK?jhIm_X#42Uy<z-ETl%b29t)zGa=du$TAz
zn_&9=N|GmK*zliG-r3oXEm}Ixjd7k7Z?f`774*utvxZJ21oI}&v}jox4i~E4<PSUM
zz?!%$q^s7MR|go;ma9HzK&L{&1diDsz5BzOCd^FoC1%j)${)o!Zo~Si*_J|+qI2WF
z;=P4X{jmt<%C5{k3#yD4Gcp{=06vWQp!W%2wR9rZ5)N&>cjv#H-Lu#YB>C}q=?Cl-
zogTzD+y^=`OcOXrPgWFb8>B_vR6}?oA|p>)(T=r1hpS;&Hu=LHc?owR>{awy`uq++
zkhIA#dYjn8g2$$aWZ!V1QS0psw{cT#;o4M{6Xogj`Ot?|YmXb1_U+)2XTG?@Z&JD&
zqqxa&b)lIE$)2j_W&x3~JkwyHi<tio8gOirQR~MR2D1gGicSD4nZFs09+e%bbMB!*
z8E}irV)b=S3@!;<hUx>w(zy8rh+R9HN@6AbZTGo5MybbPF1<g@I`Q~7%yC7sK0_}I
z>ATo21clmojPip-17A$^x3Fn&CgF{(CMWYHlh)7}Z(*qq9G@W?JSMmd+}GA6^iLw1
zf7re7gog9N(NRI!dHxVc`1Oq<K&GhiR5uw`T%#6m_nKJmrCe}$M{Vq!<-&NNIJYk5
z&Pl8wwv5w6^HE`#+q@hZtbDwfD?jM7la6u<Q6x;-6|k_&LV>x5sqnRw4|5jRc%5US
z;WhQ;>sX|i9x$l+q+QF<6s1is;~@B<<lwbd%7nL!o-K!Nqyg3Xu@Gy!pG1cM&X}~V
z<VhmSfbSpo#Ov28Wbb^%WY&#@Sa?+S@^TS?R)V|&J};pz`XCEzulEo4jd#@#Xn0f^
zB+}Qi`3%h0(k=f$OvBfeo`2wwn1Iu)sTiEvwj*>%<w2+30zGr92&w<~3GeJ2cST=a
z;&v06fsHN4?FFoPw>=|{KLtUChrHO_w1>KLJZJ3#E&uVAB{|&<#hnT^>+*kh;f%+F
zPB(wv4#xio15h5Hz-A2f6FmNnzJE=voBi@~{!x2R+zPCt^#n>}fDk6gev69r5^(;G
zF8noDn;|DU!^P=|&)|PqQCt3yLG3HDF+;2BWSP>4c4`rKd|Bc23lk@-nc@vU{lp#1
zMdi~IK{>jnD@n%6<hM<uHmywdbZaf|v@UW>z&0P2u}+^RAv*72n#PRk)22+rcE;Jk
zE9b#LP<{6qXQ~_M8g*1B@T*3FwP7g)QU&=f5`S}7?>JY&R3h0kqbGoAmZUPW4>UH+
zq>aL5hw9c<LeFd&7G~W0>AmvW=N;%>>2Ivod}}Sv9*;tmM(*qJb!qlJcxOe&`756v
zx!YN5ZrK(3jUeP85KxlfTfc(b>U*idZl@_K@KjwzQLe*`yIkm(cFC;HM+~S*+yb7P
zZ8w-#>rdzv^_8ZzMiJ=Cg&wAqcDRIzyDBOT1<gfEGez>>NRXhw;;dwznDuFTb|+@5
ztL0eTmu#B_DMjxdZ~L^VA%DwI9^Jgb{Z6EW<9Z`>Ee|>2vlW#Hzkl1*E=5t6mWE#!
zZMx6zIZ+G<rM`tw;PYy-vfp{GQu!CI)%@P@*8G9x%uY0wO6>KXYarvnX+zJ?xf$e0
znI;G%0Sx*Al1}a{$R|eBGJmP}QlLkeSL1D(txnmb>^`@6Mi$+JZ9dzqYaiJC^lssK
ze}WVUT+o+8vJIR*p-tEC!}JCw_YN@3mYaTl_w@Xn^QXLecTVgA7@l%|&o(FgNCHhi
zdFgw<rtT>&94D$j0T0BqqQLD>r;yM~#;(i%$r_~3Q%(^3q3Y+mkm3;bo<48*b-))+
z)y@_84#CX*){j)ScJP}t-9G(=^XoS=>NCm&Yt{%@dkLUUQhLFs1}1oge=o#G|7-1(
zwSc~J*SL7_NDSGfE#}zGc3bTick)@WC>-UxuWK-*Rgz~DJ{%pa9Egqp-qaK4%SUra
z)}~M}itTPC?!47Op8)w~V>shSRxiy(^f5&=_!Kt4CGa=f1|D&sSRnE6K8W%)>sCck
zd-z<?H-MyZJB9B-=b_~Fo?DQB)?oMO3(Q-=IIpxV3{sOq-7yVEpWj1>Y1K@kp{A<3
z8kMXHMb!Ahd%C;iU(*+0ksRX5nYM0$0z2mnusF~B_{riY-^2g9)l4)nCP%L1nipvH
zip_9e$5`3(3rM|gx%U1A!FLHR^m}ij9~`hDty3U<n9o`o*v)K{WCGD~_>Wjl3lnT!
z>D_AF7<%NH+Ls#;_hJluQc~~>hu<`r%Eoe$P%b%9E?k^GW+Il?jnP%n__VnSnX0Mk
z0!`IXuL8}L=oUf-Eco=wO|?8ZJcA?*TmRxF+6Nb=?wQ;zy&@#jeGBt{bewP)qn&vP
zMj{}${0_wnf};~8lyIl@gv{W%!ywCbKvX}qf<*RSLJ2OlgdKP!doR7Imp@=eQnAPo
z(pG?Svn~CY?*aT}GHlsi0lD;Z4KHX8Ov9i5zVtC3fzEXN2eZ((M3GAkOxRA~jUNNb
z-~jauDD4bF=9-xQlv#23c}!-<h>tXP*hw0$7l1#(f*ggAtGIfu?N=I{&`g8c3kuGq
z@ON9@kFdvZtG?wh-}p%$4Y<dAgUu@|?xzB~5bu27^KL}IZ*ShQ21~Ne!x;m3c%Q=E
z2)3c`fmD_cKN4OX1)bpP*^9gl|9xflXf{T%;Oa(9027gy(V*|T1a9c>R;gS2Esk(q
zt&TsEcK>t%AT?YL1~T0aPXXWkmuS;my)NNz5(GqQBdYqKTx<2?T^zijb3~eP*zZtC
zgIWZN1Ju2yGuQuXaaQ09%lcYlh$Luwi3vO46~Gx`*@yN2N*u_yZl=?h*H@I92_GT`
z-G-~%>iPHQD<HUS&|yMOCU9~c$)a^pK>KSFC=HD{JaJ7oJEy@bA<uw>pMOkD)cEdi
zzJQ((Dh}U+k%P@CSlViJCie_dx3!yIk?;X?geMfy4QSI=Z$Qecs7r3VxdP6t@{sk=
z_Hde+3v)Ue+|QF1xC@fk*D7%b|Gnr89(d8i0=nx2BCiz+Y`4gbVjZjV-l}sBq6rd3
zuDQSdubp`ark}9XHy*u@<<Tb8RbVq7!P~p~?E73M56^d3?q&kq^fB1{e@|~@8vdlV
zyHuG>9A0h+R?xcH32@)8vyOs%JVgnXg)eY$nV3fi`H@9K4JGtXaKYTMC9R(ZTv>Pe
z5y%HER#Uy6!fkwGf0LjlJbmZR%+O7g_iU{*?jDE||2rg7X>jxlN{P}4zXrbPt6UPE
z6KqltTogp%e>bfy;>T6<2ZCuji~M(!N)tNsO|2Sj<Kq*1cyw^-7!ytRN%`QNxPmGn
zDmjhtD?s~FzyEmlzmFTmLgZdS5*}V%On81&(VCp5u!4Xf8Li@}a?CL+QD>-58p}m^
zPYG14;LspuSuF~v?@Yawxb?E<c%fGH0gy|{vu_q89}XmHuEM)AO`6nIy)Hs{p=dHB
zE}7tDU|<9%9sY;h4)j`#<m#N{L^Dkg6t3&+UmFI17wvbyRTUSKEdXE%P1`o-9>B^3
zh2H-aA;N=!C3?>AGjNW#jS;&wQ~eT3>ZaBx@9qD+?m_icScOEwT_E(8;Om(k#o%Ot
zWoWU+Ax@&cK-Zc2BX<HImuMg_%r@FtdHwgX#;GpANfq>!7K>t6TLmY^VOdDqy`U3r
zbq{_cZ!F>Lg1)e5jNY478?_9+GZf-o0dRT=Vd!tqtiIWU#^z!Mhf=@X@XVqV61Kqj
zw<ULD-=Cc|{ofZCgUf(#uJ5kvl`SN)N}~2xE`^*Ank$ouXgL{xj+;R1CV+W?22rdY
zsMR3;A3M-r$2ENfm{VaRRXQe38<-=68^-CA;AeAH#FGcmnh-EYXfu6Qy8lQHvspmL
zPxot@@4x4q3uB2Iw4<9X$aR1}hmu6&K+j0~(6rsltO-r0d4bcJ9Y}^-)0KqmYXD&P
z!r3%Mt8Y`GfueV(DDKfZd<ZJKm!1QQ-h8|fWHC09@?Yfqj#?{dmZdc{dLTqr98ee#
zG(lq1O}KKwETSDoyfrP&!v7TEajI-6Iy*X2rjq}juLWEjBAPPl*I`SYbTDmN@7lmE
zSfNkDG)QZGNT66l-I1c#a;=*DcMDNd0qYAYcKMuNj%rf$<!9W#f{<wVpM~W=ifAv?
zP2jjf&3Sa?Rf59ZyQ=U3oCoGSwBew?d}&EXbw~MsFQ8WiJ@T%KRgLfktcu`}Oj3ls
z(k@f^E_?Di972`7USb%gB`@LO%1@v<B~*e!!tWuCoP^V7vd(f>CJKmJQQUQd*Xg67
z_qh4@<fkSLVMzq1fIHg@26iIp+$ZF<r)UW$CXS5ACZ%6yY39FrhC!qL{$BY9zLCUv
z$2OiN9XPPRO-x3X-^mGJiw4L<K{>&dm|%sYIaw!@>xv+4bKUW`2aRE)-6KGtf}lFb
z)K#^E3k6}E^u(C*+s<@m?;}V?LAu1#w8^`RuX~-{30B}jFKNJ~?EjD@4kQpjyn*Hi
zr&Shm+M$$>Skj#qEWw>Vg@i5EHyBdIA8*oB(2cT^Czx;he1i(v%zPFQ-G>{4&z!l#
zpm5SYBHMceYER~90w*jNHo$Enum8Jms0l@Srq^6OKJPZ1i9!^S#QkN+bEaGGi@1F~
zM2G#k&m@eyrTy{$ku|t5!l<J!?+j)ZpW%Ki8yqb{Q}X$fT$ZVky>`y{Tw*4+PJ~*&
z<vvi(KLMj%$g%sxJ@4{ExW%})S6g7=_f+%<bL;_B4`g|K0~y2wqPtobg>~QBbe)6p
z^}n+WH%eCmUXjS52y2s)U5fQi0Wnpy5v$#KF%OkD03;I<?nZE#TzWsgPm%z5Nn5U0
zce%b^D!rGmHHDGT-JX&R+Cw~ot`S(C{LJ@2XX89&C0~<@AlX<uH2S%3csX&bscWB{
zZ>cQzG0=UW^^par6aIbiR6bDCiv#+}3TVcBHvWE8`BxwxEbQ<EU~A(-nNc4got=I(
zT#UPL&7Q+@;J?P8SXH$YAz*5B&jkE+)BvfR7<fFsQINkExG~fSTR+%dzPv4vF+C4R
z)5`}Bz%qaL;QpmX@<z^&0L70>ZWewB&AhtGbgTM%l0d1~O82v{zd-)G@9<2Xtr$SZ
z_aIzPJs?P{D%#y&)=TE!Mj&wib0c6_&VmwlsjY=tDjiX}4DQUH|NqU0Jq0fWp3mng
zU+52C$2dJL0ItvV@PLCiklz+_SrfQ!$-wmo(-H8JZq8cGkn&`EDwO0nRDn2whA9jT
zr5vw~%QEQ)ENSW+7w&+|?j#4L-3+YsxqJGTa3Qdi-M<)(Bq`hf2_R0Ie>sj{JLwri
zo!?0slVpt40c_<Iz;eR`dQ$3U6Ro3-I@iQ{$;t?QUNRm+&Cn$x3PYMB#<vaHKN@4#
z;NdvE-U~WP8o`MkcpvBGT|NTO#L6^!Jxhnf*+Si1$t3V<)VWB)pcDKS2m38OC{e;$
z`h#ZvE8XwpRKS8_IbPS06@W?B93b|waLxUX*lvYlP1zh7oT4Dz-Ib&cSEB)l2AP51
z)r`LD@|2Euc~CrrqGN0ly4<(MihqDO-E)20T-^ikc@IXq=v~tO53qcq+h93WNl56G
zYW0{ZQ^V!wI-hbn4by5GFdkuySU^Ot4i@=$Gv8v%@pNmW!FzZdt(^?elRiHkxW3!B
z%GP^#czG92Coh8SE9WzJ2E7KWZz#?#rDZGm|Em8U7P*fK4Q$n|+Mho&dXgY67}Zjc
zaa~?=S8fmkky0bOq!AdQb*6+=Y3-s6obs>441zp1fn!7QrbJ;&&)NnEb)!X-h!Zn`
ztab>b4;H`UA26cCHsgswAEt<nce|O}uU7c0MxsJYk2bp5$Y$Zt=->#w_rB|In3}>>
z-F3#=$6pF1RQ?>&<_u|D*7?MGmyL1vCiIO*h_t<6sl4}i_8+9gph(v&anj+j|CSbn
zzGeO4ie8RG7kH~<CISfjR8Q_l&XK$>iKc;-d^R^bf2F$CYyckyc+gefvT$cDYgPBs
ze@ZUK%6xa1d)N(9X^+;8CvzMm1V1`^)@dE$D;IxEyXnmyVs)=QehTg6r`ft#rYfg{
z4=EeNbXUMf<@&@EC0x|6>-NY&Fd{kmHuOf>XHKNMpLuURMSP!Q9BHinD3HrXfigrh
zJBWUb{V<6!nrkTI%;fEudi1XBE!{OvqbLie=@Nsfm-aIm8?k@QHJpG9VxL%trlE?o
z2N!H&W(ScQteYiDV8l`UF&_t-tk?{S$u(xr;^>3SrAKDCO6{SGLp{(_dQ=W5`7tQ<
zC2Q`|Wa-QF-kJlZ;>Z!9?d^3$6MRTIn_7f^^X+i)?5$6dRzOC$o7;6NVN|gyLY-Us
zveJFrdQRoUma7X&cfr#RUdP!==25TRbTkEsYmcEn!xwr1#CgYF8jrruv!QszZ@WH-
zOSex3^1pMW!PNP_cfj(KRd6Q0u{nZ0aMh+>;P!B{hu;9FEP#}#pC7<s18#|dEE2<u
zwEen-#ig=3_BYIHt}Bvk2&P;RAb3CcY_Fwy>J+DjafOT^t@Df2?G*_DY>Ix|HYjGl
z!z{zHV(yxkcNLvDJ|sVqjKfP!x4O7&Q=VT3@^ET&8w@aS-Jq#13S(Y$gE!&oXx-$X
z<g4)^to1dT9Y%f*&kvJ#O*&02%adUP_w6D$hCJR#Q@MG{8sjXps`#MMHA87`*N9ZT
zmmGOt#EFD&{!Onv1(6Ni^8}yD-|n0an}kJAkM7wE(kZH&IE`~{*}P08XxT(Gt<02_
z9v4E0zBr@HU^RFngQa6PM@~;*=nsBzWEUeKEx9_R!xj`;WsocUoSC7ZR>>Nauw(Um
zD%YGhbq*G$iTj0Jk1`Z}IbCm7-zF63XeZ9h*R-nX)TM~_lx;GJjv-oAV&&H8L8@Ly
zx1Fmw0-U&3!(>K9q8f%$CFnGB>zo~3X4N*HpMFya5>%A1ekG5g)~sY7wLXeH5Ve_J
zK#URB;bePnoG3VSNAFI-&)h&n8kTgSyg6L^FV=4`jQdN)8(6uGFc4j27-8VfW5wuT
ze$uS(Z*?<KyVxFlo?t{8j2LqDzEok0DPVc`3)gwFx!7gv-*krz(<@-Oh^Z|B=K&Qd
zY(O*f9JfnM=bbYDHqsmRaj#g<G(9`Az2SI#H%W#2T5^DYn~lfm0ETr5&+W>;dLDe!
zTmQXdLoSSyiE0gJhr7LdMKzK~DBbsx0}w5diB9IcCi?{68iqm^wmqZF(|TiB)3_$A
zm|MD<6w7W>yYp`s8rPzgY!#D3%pAA$UuK?9{qYoJ1JX-dszAi(trb!I(+w=(>|YdZ
zvEO}O|2Jc;2BJoOe^~p3<x0B2<(nB^WApx-S@6S#{)o?%O|+Q|eOe{w{T(dM1r0AG
zHNwarMK|j1TMUX2@5rZaqcpf!Dm>F3$V)858X@gu$jVGXrtSPZH9;&sCyJ9puSw!V
zi}J6}YtO6x*K&R3bP7ib+FjAtdLK9!`BuF~px6Fqr6N%r=Nehr6F#*{Xs89`26i<`
zuW*+wR8$cug1K$e3a??dl!v`x95z~I*K$E52><xj;h9LAOdPG*rPy=Y!px)rlIBm4
zwm$cZu-<gC=OVT9Z#T=B-6MVg30iI}j<nbjGtuv6Xlpfu{|#^1gu1te<>%KG7II*c
z`wo%-nS@y9^dM<pzYC*}24A%Nt9jxG-JVVNVhpXPr<Ozq7Ria&C<?$Jw>f(?UMXx}
zz;N#Z&d81nzV7!B&BW}viOm<M9OB!p)72a3(57Y~(u9#cekdl2F82cjJh=E)=*kTE
z!mdQ>SUglpJ(?(*M22~g*<Im58kOnV=SM3Zta=NfZ64zoe2mX74TQ}k)&40>Gp#=h
zjaXzHewz&MJ8?4b#l&V$*92c@mLq`A^KsC?rJk&(Mj#EjVgm^+<bm#qG*%Tjp)yUI
z9R(N{hvvn|!%y?4=egaqHA`=2gP2*yx^scZNw-4jH-cuWGpVhnocUSH<uYMXR`kFL
zDycI6kzsLgrU<g`H532i*rLp+Bc2RJ^>L@Gw&&L4ge2t?aK0K$shy%f^8_!f=i=m6
zWs|B5-d0r<>D@?IU*ZtQ1fNx}@1{G=foyn`#WB4qv_+DJ{seNv^VQ=Oan398#2Ep;
zY)qYbk^*Zpu4l9!Fl!P4P{I8<>`u!EC(-`?#F9XQdsNEaKY*mg?w=P>-lu&2wf|Ih
zxj!gWeC>2o#CBQ9%<{5E$i`md?!kcikOv_klj5nE(2;YC0x!Vn{w^wj>(xe+3UV^E
zh-C5b`NQhl4KAEq;_ASWk?CG%bm#E&V>_U%sU|dYC&DGSske*Lb^kuQgO*yRD!!vq
zB4hwc(Q=K24?;V>UZewTE@k3p7ArT3+#~!V-DidH))U8xi|Chifw<x~J=<GFZxWvf
zTcB5$KYkaosI+6tW6fql%~*>(H0deu79AeQ9CqnyvR9b=e_XwJJXQPm2h4G*L*yJ&
zNOF#Oo@LH(%tPimAt4lsqSU6$^E{R^hGZ&ABxEc@rVN!NNuf}L#Itty-tX^up4Usg
zZhzdDefHUVeb#zU>-jqe))R+hlBgQ0M>PGO0N7d)OFI!8SiU(%k-;QEYiN8Ut~skT
zs)skUKQK#CrQqeCLM~+!w}#W1I=?o2?M78p)fvPEu^Tn17u8n9BYeh@!m^XvIDW|#
zH~uow@BZ+<vxrFimZY!V`TR07Q7c#e6=Xpi8q_ptHH09DZrZ>Fk7%Da%QX0b(8ydF
z8N=nGU#rX|K{RQ(Ptyg>Bb2FsaBoYDyplOq;+vGyEJ!=M$~%8r?8}TOhAwyau{tW~
zeqeshVmtfDeGcEMgTjbauf9QD4->}sh)J>}e0H{@UZ*KzWg~1IZPVDUXU+mro}~3!
zUT$yk)DdQE+$sH`<%c1TBOBd8_3t`sc9=2IgqEe-k3*Z8F?^V!fFCxl*6w+JKSI`$
z+I@wxYm<2s7|xwOsxxL&MpUasTu;k9V4RZC&lEd*t;o1IXc3vG{P1~lOd+OP5hXbG
zV4;h4eT6d#Nja86Yhz*LW%>ex(wsJ$E{?`_H9?%mR0a#>@7o&Zmp*@WDtI(sbKi`f
zpoEr2qO)<^Xk>h}Pa=nci(|h!)2Rp=6Y_P&?>5;UD@lVv#3USwA3k49E!8S{X3$Ve
zaov29{vLwQxWalgh?QLKQqk7QMq3ymdaO3|pRw^GQBB_)%<K6QAy3CERYlh|EzP|@
z1<DFq$6N^!k)C&R6{?NO%h4ionVSMNf#UZ^2zTz77mbg*2z;jyc2JH<@9wPUIFv^L
zp;k}X@+f}r?m13jrd2oCHG+@oRIGFKN7C^)W{gdV?Gfj@s}--^PKWLaOKv}0V}G6%
zbLTDTB`vqgcLO{_kd@IcjkX=#Pc~P*4eLF^DU2O2yz0|4Jh`#8q!67Iz*#%JjXp%B
z9#>WHqT=uO1_Af{tNam5SOhFW)H2~qW$mwG-fNWDm8uvTvBmUvd=AA!{yCio*n1#B
zF?#4$FlLza-1?QdNo%x6AqoMM7pv~wivnP7Ld8*6a2urRsO>H$Pg*LHb1`Vk@5tSz
zEEU=>?{_MBHJ&Y8Q68q^fP&q)U(pKb8poOfQeE@>oOwdN<6DgBjx56@0ak%M`>Kam
zPd&Q2aQ{)aEOdPcC);DDoDlW~u;ADlWGaeAZ^3i$mnxo?!fdQBr%h)`?uMU#1R#n9
zTj8`>k;G%)sEN=^7@Y$S#k1k2to-nG-H1d{q+Ihev+{tp>m~;TXIbpyTFaDqm?aw_
zsHHdb9V9yLBdif;qK*~-l8_@C!N#bO4TvkU(i?Iu_6A6R{TLcM6+YigEhj$_+lw)5
zjboF(?k+mhk$Z*q#SeHYO{jDq8O1e>m4)F9Zz---!qN-CJjz<;HF!3Q^io+X0S#h}
zUM_rcdo=@3dlkiJQS7b77<-3J>b^!qc+{5Zcm(JYXxr?2pS2HZ98C_em_$sp_L&c&
z`+7*2=7EZ{{(Me2j|3{Ye6ub;EH39x)c1#HhB=N1j#9?4Kev6EnkyTzN#jG2!7cxR
zc9(+MYu1*PYJFiW@oScL-r)vTkC+qaiuA{C&C_+qI@BmGaGFzy?*GB?(F@6&Io$|*
z>}-gFMkgR^77-$rNk>p1;1^W7H1{3=5ppuOJU`>jF2HnDZ~p10%g_4<JOP}Hk}~i9
z@|mo9VnJjzUAllNspPD{y%zkYsKetVChS*bw?9_ynCO<l71nN*d!m>livCeoWUlsX
zIv2f<b`O0aWHbJ=FD69P-KmyKFZAIa`^yGzqle8SD7z~V5xgMvvsgH5ljFU8!&w7_
zPfXMK*}d02{M7d?%*bF*P@h@Y<2;lW`h)>+vJd^lT}cV&p0`P9qtkvLbAHj5P|R%T
z?s~(GE#*8eV9?85@^^pvp&r?m;ijd4AWvM_*)sa3W`PX|(D=yT4G*2?BXUR;%~Cn%
zv;wmB*c0r2AG=t6hWpp^&UbFR9UdYrzdXToPJQ&?)9&Lk8U8g6Iz~@aX+pj6%c^Hg
z<!H#khpTESys&3H=2G|4@cfXU@jP4-OQCKLQeneKnT99n=}E<XK8;KD5MOi&mQaUE
zGM{E2!^bBE(Egfhk7kUH<&ba+St=>0YDTuoftn{`EXx0KSYc{51XhLfJaR7rC(Ums
z8(nMif!s}3pw9|`dW^&ChN&f0Xhr`Vxzx}73RYiwo<{&H>!<(JjY9>RmewcsP>}@V
zIbhTnoxCoO@xP-~{M7}3xoKJwWyW|QzkkPe`?rbcyixSrh4&1>QDzcCMtq=ITG@LQ
z(ET2f*kAYi_lLR?pdYtwA5k5_;al`-IOnfu#$Qi+L``|U9#Rr$I}(llG>p;tTjQUC
zt~K|vB~Q!y^DjlaIImP(dG1wohl}C?y6PCGw=nVGvt(CMBQexMj89^g!E=Fwk?qz$
z1&GpzgtTZ2=Q*|i&J;hXI`s%9_(d(QcYWDUhf0qUp|!<{Jb`78%xWK2&a5T}sv2n*
z+Hkh^0vu7~>IizUtJufnH6~0_q3Frxo0Gy{4ECtVeL6h+T}R~~{2JyqHp*cBu>J9z
zDo-F?p~y{>ug_1J%}K6E{qLO}M`A-f3npUkUfz#Z4%o8G=F3~07mI_9c+(}h=dSB-
zJ{&f_;5Y6ci3atFH!jxtUm0A>*9loxY?OGsP}%q*UD*AvQu^QDsEOlH4tynW!e<9v
zKV}}6wd$Hyde=>Cx7_Tm=JdTc|GTeExNX3K5Fo7pX)z2WCe2;sPU6-%P`?LVRRo~<
z;<q=gJJULssxJkIVrRIPK?0ZOKcpx547?|NB8vYO7)LcHNUMURFR1uRECsvUpO1~-
zpnZck_DCpmncG9;y@BtJf$C^0iE=-OC~JLyPq^?x@NWB#o%@m+G7UiAyV}zRz!m*~
zG2j?-ZAW$oV;HL`H3@FDStGAO0Bh1!bK?&G3g22*!8|rIcU1mhavQ*9*mo#wS<Z9i
z>35$!f`9o7+!4$O_uUk#aP?}sP2ro$!54Vrcf;Yg?_SqD5~ha1R<oF{393_FO;^RF
zo$5M=VMY$jngIWW_&vi9?FkY+BKS(xwL9r&<gy(-fLA_m>P4|%B`AaR>){hV-}3`x
zeS!Ky)E{-QpfUj(`<aRQz?0?e_4goZ6CsaB;|BS$KCrZ~TEXVKy6AuP>v06dpe8uT
z!s5i|q5_pRVmH%`X(+`}&I`$O*QpKK9|}}fmpx{WDSIy?(@=@wIB4eB%T5B~d_Cuk
zMvzai5xm=rAgnEV!FFs10uGwkR?_@jdB79nv~_0kVOjb;lp2Fq88}u*xOaMX?sdfk
zcwAAlu<%%U{6B9Sj#7t!E)@kn&gFn9gz#k|C$A|$Y_%Y;=I|hHUi!w<fXDG+Aou}Y
z-zZ71Rj$LwPAQ-4e%*hu?}Z0hAB{UvZ?k>kN)rmTk$3LNB*@P}zrO#A(xrD$Tf{+a
z55`m<C_5mZhDXKxL5bhLC^!S9OXU$TKXpqdcX>MiGl~yFw7WF8C{QLlL@y19Cl4GX
z<TnGbP|_OJribGGiy`fYu|}iF&-`Xqf;La#9y29DUMvHb8rKI<8kvAd78EjfL6;0Q
zhuKsMj3<J24?V8m`<edVXFf#O0AW|A*a^k+v$}cYcB`(%?4@~Vbpa+oK!#R|6JQiA
zl@@*qj3$9q<+9A)G;ro8rR*J(qUD1)1up4g=(LM;KK12}z5n{h{&H54bH~;%g<q;y
zf%70a_|u<@|LKEu;)S4H3$W;6dR=#Y;@9;4p_ruNC@Al{tzG___7vhX8ZLa6`<iB1
zX4di3%FRZf4iU#Hqq+gkENnD9!fGPeJP^t<8p~7OAUsqpA_F3{1bV*9;O3yq9fUFO
z8e{`>J~95BqIm~4Yo`?$d}FU;!PMY^*$maXM*8s;G}5iguS5U)@NWnkbGbgcXXVPr
z4%HKHCPbVq1<fiXm7M~HCY_A9DZ%mxR`d&q3rb_r;L5@&1LqU=^yk;nmnQTD3jox!
z%b`KUKS53gDF)G-CfB25@7p~Es;PecZhu>B3v?f~cClbYYoX76wmd`X`}ga<fHTnw
zBR_zLqY)HO0DZQ|rGe9?H@Yx6JmfZhqfxu=HZrU7RRF6BGAba_hBDs@3@(YU4JXTa
zZ<X9g*Qum?lg3*tE?0<d_=5TqF;EJRIA^w0G=T^En5Z}W(NEK?Na=EG@7g~utJsjL
zpSS(Kb#o}t=oJ=OYUh1m4Om{N)k##WO3c^dDl1l{5ZL2*l4;h~p-4A5gf|*_bN>70
zRm(?vum=_M&#BP&!I}R=A|paKL~S~=QQ+N;a`DS3`rVnr`Sf1wb+sdZKO1aSkS_#0
zEnQLn*Skw6Vcdjrv}mNj3*9s|<jp_GiFx%D0)VBM!)G~@a?pM<G>(fxTz36M9PNGe
zaJ6dmbQ}L6yo4;Ohva-yzu$S9UM+i}7<wnT$HgIb@cIZ=rYWd5sjfnqqfT9Ul9n>?
z2$v-G1a=p_j}dSE=N`8V<9cr@Jb(ze6*5m}6EvwS&7(mu@tdJ53exh2!=2SbV*1mJ
zO@iy(Kj->JpyC8*(t9$+Z@Y*0-f=>JPA&&WJ}fRsFTUMsiyMDsLeM%haz~xYC91&q
zYvF&-kuVN_OMg<zVPu7UtjydUp!0u!6Ng3%uh`zmqT>(jzpmBFC1+*29czBfVPtt*
zzAjr4BFch>jhF0Bd}7CB{fLr2e^&k^Ey^r{a^W)|jz};BB%wC<l)C<d<{!<8`Mg^L
z$j*;$?nL~LT_A?Zp7|$F@~|r=M7Olw*oWo>LfJh;rYG7aKl~jNBt7riO_)Bj3czNL
z)(Zi*XAlw%Z#HN-`$k&gsSh-+{3`~~A>iujqAaQRodIpq;qY<pThDJ^ej$k54J5Nh
z6XDHM#tQtr2x9>hb}2_H;m;Fcu%XQfoFPeQ@WAoYA1gvi<W}$-paME5YDbph_6z2@
zmLX*63Fp}g`$7{wN&_{Qf?)CE+!;{R!{7#JO#!IJ^kFo^#8K1}XCg>vHJOB3w=|Qb
zUW}gv##9Bl{F5uu&~%&f@jj*)LWo|R84+}PEPOCE)v)B3F<dVEb`AWkoa8&uqJBwo
zO_twmp`!1Fy?0QO;2}W?seJ)yy3K!&Vff{s>wgX&%vCtr+jw_J%#Wj4Ma@YkH7-+z
znigsk-i$W|U6|p3U0;_VvziMT&5}Iz@b_xlB8ai3WiZ$Rvq<b4Jkv<T*(kLUkcFgL
zGb|myx%U|NA7LhL?rJ4PkJ9^EX##oD?gwTcCul4t_Y)~P?|vpY^F-i%8PO3}vwUVQ
z>PuXDt!1Xl1Y8Y7mc4^9Ifl*rm&##D|FiN0E-yD2`XZ9&!OZpy7F&D)#v>W*qo8WB
zx@Q$%(TR?cH@kHgM(`d(MS9R*6f%6M`>juBc!C$(6iv~d47uph+4eaARKe%Wi#LMM
ztfzuaMYRVh=j{3^t`PeIee&HJzCj#yRkFzy*I@RWENG*lW&ShZGS?c^+$aj79B0}K
zZ6p~YJstt={lL2LD_a96<c&Nw)RJpE6j2GHvL@j8U@-i!6E26zB|Kw4cNb(Q%Zn{L
zs7mb1@Lk}wBXWs*4mUW@oY?a#D6K3CRXX1=07JnMg2!k%bJ|5o-T{D2!G-236hm#y
zb)3CFM6Eo-vMedgB@C3S$Tg8FOYGs6FR5z>aJ_5~8>g@De`NvI+H7;d6ebKThBYXE
z`i%W!-E-rgK+i##PR>Vw=r;|n>E+sDGsI5lXW+L9tP#_RHXz}Gj*mVu04Qd7h}*Iu
z*_dj_lth*izN`I91cSOW{M_}T@4mA(^DV2vcgtf=bs-6Tb~M9jG|ztfb%iV1cZ6x|
zJHHqEofG%l=Qk*TcEQqU43W6=tYFCBo;H#YmY=Hz*CD&njk_nNWf`n*ikioHz0A0#
zJWf5;1ZRcDk=(^MIr{>Vb=lNIx`+QyH*<8)bPcWwyWXXj$~Su7eoVQFikQSqiR76B
zfvf-eNA!TluXNTUSB0@}bwapWvnE`1SK;Hl{Is|ex8X+~W~)9k%vX^9r}R2=cDxC#
zOXc(BMqFcE)S&_>n4k&`dy2b9Q<Dbri_SnMpN5C$sZKZ1y;R|(sMU5pO^OaF1-+4!
zs$dy4SLNdo-~9uYBc*oD!9hsi&-9Y|zpEa!*LrSH6|j|7Pz#!EsWHhr@93u4_I{1Y
z`Zyr}F^(yz=J9XRR6HpWO&9?gKrbSB?y^LNdSeySIjR#W0IF$e`gJS%WC%&Gin_eG
z0FPCdauu`}!kF`WT|3&x++9Vtu53NN|6#RAw~O((GG#3PNFK7X>T^|cLVl}SN=;h(
z+dMq^Y<utB)%;NlwkX4oPcN^XFWvQS1T}R%KD;TkYH5;qTWM_MbF?<aB)r+DPm|+x
zz?#Kv5oRbOyfn(kgR9PLF$f0M_szYkizojvPAg+^2V!dJoWn#xPw=HPP3XDpwf9$l
zp_!G)ir5%nQ%o3^Ta<ch*%O`47x#9<PPi-dsruiM9jPngp8m&)zR<6aH!B$Ht1~C)
z(TK?cCk7wOCbr@k)$*#^POa&jvX`|Yqa9YPVKxP#;1u$&K_u|hDZTo5_MFnAZSS2t
zW8nmbb1(~smytQuNy~lN172%h10+MnTXWI71`l7HvBA^m^aI`?-=%m%?6Zi3gan$P
z_ysC#V1=8N{^OY!8KN&==4!tA+3R^w;fZNI{jY#oz0V2YnEeU(lbq*B_e<~7s6K+X
z6tr4Y2Vj3AEIW33J*3gEUX(|5L#+eZ5h)+8^RkW=9Tl;_xq6m`AsYMl^VR3Wj4?`a
zGvoyXJH#J3mS+u8h|n}<Ob@C0RM&xP;u6Db_4!%ZMLp|f>!?G1LXoQCQgH?wnQKIL
zP%5f1f(2uRce-Ucp(?D#`=$iK4#F5T9<4yf!guAq*AmF4j-RWL<dEZr{Cdr)M5=e7
z%>ZrSkHT3bWnw?h`|aZzIKn}=^y9jO2_?ux6YQS=SVh*vAE9iGA)+*wZq%u>X=UB}
zrFq#vB2b5`JwK3bb^_urK?>9i>(PDLx-DGW3aH@hPtkz0zi8U*6meP(vl#vR*LV6K
zii1bbB`&=PBpai>=bzfEO#y2A_qdlM!geK6xSAQ*CTu+gAXS-vzL&-gwQQ-=7>8gU
znc?~J0#G-N!Ib$O;=`l|lpg@d7N-LJnl*2?6l*@dX(ENnIPMK990C9UE=J7mA?^X;
ztYi+f36XMa^uPN^kBxDZckv;jr{VZsxy-+saNGCq^a~>^lB9}xK{hy|xyFUPsio%N
z*FSS_Dm-MDAThOmf1$>y72~O71MFC+Iz{ap2yJV{%3S*9ut+8!F;Pk<Gm8s6Ms}~!
zjHILlPoBA}^<o%;*da6KgZ!5_k9=l>4(|_b;Z}Ql<lDoZRQ|4!hcNp{4lER*Jxz-k
zEb*z`kLay&V>!t1iw|61D6Ryn;}s#zD|)1!lpw6`^ADCmpe}0Cc(YpH1j*fi+`$^B
z&9VSfOdx_AYSsM-#+u}e>&YH1A(U|PT`~>ZrQ5@73bX8~1be8}S)b8hH~+JMZk)Py
zFIA)EE)K7zLv*t#tx4^)e*?W2!aiUzYkvQ)q<06tHxkm9Swts}nTgrI{q`3eF>vgI
zSMpkcs8H0Gq=Q%D$gpuTKalRG6!g>8foEsj!fBY>%#t3*H`DFKkQq#wsJRI%WI$Zn
zbg_pf%kJjo2rfx5k5-<z0x(k#4M4@UQDsl}5yl^$Nj1&yMK*pxjDv*~@j)&0=$vU*
zq`y)aHMo*@d(pm?@;H}Tyj<e!jX+F7HBAke{|X)DZ8b7aTx8QOfQDk?_hs=|*T8d?
zQ2fOl9z*W!>oERV*c0$EMw5A;^@+}oBnpj|GuWs@Z9MJoGmxDqq6@08m?(7~piKOY
zfYj{XweVra{2l0Mbs0ko^N$-m*q<q`<5Yg~Bm3H|Suv{O5C1@!T!*oxD%+rAT79hW
zP2OEac$p4}>NmdB>bz9<#mn}Y^du>gi<kXp$Ljvd4gI&i_0g=}e>A0E<O(9Z+H1wq
zq2!R_RdkT>xXm@MCX{s@F$2)Sow7?VOfQOO2Ij608TRJ}a}8*8zcRq91ZR=uA2{!%
zGW^{ywR~GBA%usB;!GM39u?`$fwTzYrbLP>@-%%ZUs4Gq`tKjPsSj&}Y;`g68}n@4
zI4BwwhVEdcY=_i9Ca+}C64FiKJv2<dx`6=;1grp~b|ne9(pa(j*hffXul$(cXqqJ}
z1T(5@gm5^ocI#(|u$c>9kB+-NiwdG(z-eJQ_#hjQ1Gq5dgceE8%;PiEFjrJxT(XmU
z&&FnV!wpM$F(51no!!0+BAcHM;v#NwYW?}WhcX2Xq2Jyeo66odW%6ftclbI!MuccU
zL{yI$)W_(15X|&#Cl6}aiJURFO|fDUUA^aFEs*R{_b|Cw_Xt66B1P0J<L4ih=Ke>U
zy`THvE;N6?b>+Lrz~)DYJawBBNxbmwt-?UiQ%85mWr~Y|P8}Q>Edl+v?Jn>c-;AOs
zdyaiAryw|uZdQ!49Pfeg`t8LxXR)I(h_weOm=37gBxbHG3#jm5qnh#{T01a)W!ads
z3RfH(f{RjP!OeeG-k)SiUdAosTbRTG*R0})xzsp9VQ{$*aTM{&3@2g?y5a1!1=9m6
z?imb9I8_lD01*I+De3sV7(j1OZKSm8U_=MUk97$rvqKCFTnr|2NJ$B*{g9M36o<7u
z4m0kAgkNx{MZoR{APIw&{yr_PWi02_pQaB71N&Ou)q=cuB|+!}maK+tQ&-FL_%N?i
z*Jgl$>hS5QKmyUxt-LAt_#K#L0jvQWT3||T$|^(e_r}Uc+ar%)^6qV1ee(D?oa<S4
z^92Ruk(x(;Qf|GC(v*Rm`g@g6gp@ia42bnLe+t)c7=(0NZSWV!y601a96aY4a_SGC
z9O>W)<Zz;c(eWdY_FU3sJC9*bmc=wRO+fq{=uJP`&49XvFfiD=7R(!i-o6c-t|-mP
z+vQqc7+NlISN+)Zj~!Xt;d%v@3{Hd&%C<ATW<ag4tmC1F-@%CI9DR?8UH^1HgrYcP
zpssT!u`%8T-eu^Y4+o5xHr&6VPp=pXYi}43p+ZlBpCycxEVIP9fk1LCe26a~^CD{d
zy!ukMLvq4-zp0Pcda(ltn?1s|1X>Ypu}0wnkc>B!2%Ex9M6d;#YM!S4TXv#G86hJD
zu6p6$l;bXe+0<zI1DfjgUVfXc#-fYl)SW~`J(h}W7Z=C<e$Ee?#SS@V-SdLXgq!yY
zEYEr1By!hyz6N2q9>37%)Nj#=#OP3*^?;<?i)V3~3J*e55U9ni??F?Atvpd)jf)Dw
zxu%97Gc2c5>tE39rQ9P*qkZgjNd^Eo`uL}^JyNDo(9JN^0z5CezpzM<+0+9fjh=1H
z><`XKsR%KCDHRI+DFQQ?t=NdrH&6sWSg&JrKt;w;)<MUMNPGiOqNA*_`<3vb_8H`r
z4FMGaEPSMkkio=YNU>WB-&a2Z8;x6hMVwm?cn$(w2Wm?N)m39%GgFf-i!~-nLN}4J
zOiLlGs^F@=TD_L)QdWh<{u2=b(t-!WUDZPJK~=Q#<MwX*gBxnPc@<854uqv>-b%B|
zEU{jJP?!^##T}IkwH~`Q!)nAPCUjHCSfM^Gma_dcM{~70x3USME!_JQ=l&*+CBIjT
zO9{&$!qPgxOCszyDxgE*__v&F>#zmUxM`X^QC6)_pd{c!Ll~0mm|i8A*GX$R0r*aC
zm=pCXp)qOR;WGjK0X+B8dspp@v=6-8OJtDbwTE6AI#sRl#U{6u>EN4#ptH-I#vXn_
z3<VK6GR0^svB`&5`~#9nwUb;27QpM4`pm?AD~v)_3OYEFHm4%NK=J^q1jUm0pC^*M
zQ;nCZyj00*r?jNUH1(N6DG1ZKZCAw|!)sXEUK)NF|F~lM1pRVnqeEs`K2c{HA2>$o
zoh8{Z@Yb{a13&9F3wWAq45y%@$-G<Q4`G+#KWz0&!J8HIM4~4(NcQrlbnD8K&m8Zq
z@#^*78@5S^IXD7)zMP%>zrCHf=RMrBPa%cN3a5?UOgLtTdi2NKQYlVE7zh#;CSLiA
z!D=JOrQraqK|Zx&?_l`+Rkt<(tpc86kAH>Y(ZFd|fWt<$3?D!C4mwd}j7E9~ItrB{
zG64@am(qt?S9CSw?)m-7lp`cImOae}hrK9<edbQVDReonwgIqJ=_1B#&)I?7w*_<o
zCf)+|$<4blUuH*ei*FlJ_g`6G+yr^zomX{dhWCnhjMT0!mqQcrp{~95YE}WfIGRuD
zBfX_FnV-sCS@%Ym3c&cvcvp)|YVL-!Tg53_AYEXr;(ZvA0hzFIN#qiV>k0HsW|>yv
zvb=@IVPaWCW2^@>4=C_V5&u@_5L}>vnrHGO<#hXLYW!TP=yIP4)ex7{NI{cZl4)7D
zMFg(yaiy;43$L-+FWcZ5chYX}913)Za-RXO1#*>P>(a_h2Y_h-(G0vca~wTf$CrkB
z6o@eIl|YJ50JQQj8))-vfnt%dDZXWO9y(1m2B}Frf}Z{$60bea_!0j!MV1BPa!l{2
zh2DIq6&|Cug#|?1UHfsJ@dV`WGYrK2LIX{n$$#R)kl-hY4_>S@7-ZxV{L7iU<!4-Y
zZ}7O#>!1c$&vI>hMd$wPOwji`&m@X<crfCW9fZ?19<)E`90p9>mBq!X*uG6X*f;nU
z`m4(TV3pcCijQqJkV3XJT}h-I-h}GkfH*UT#|X;Q48_ZlZCL!BXYK?Y+7Yqluz>i|
z>fYNY4!In;psS6@Q{_IZU=iy2>T!U!&JIiFQo|Op>?wU<VF_=VH=gV3KH|L-29rRp
z)M~)n<&fy6m@NMS5J$}pr`AlvOy=b675bS^a*BkuS3#d|{epZ;0=d)Y+yRm8IE^r)
z7qHFi1PJV-{Eg&uCY*M)?W+$>0^OqgsRrh>I2B}WU}gJdo4V$GoC@VPQ)#!}Y9`XR
zV(mkasl`%XvzQ4np+E5>YUsm>V}x7g{MD#V3rlXH!pL8)7Rm9_MkO7NAK4n_QqPm<
z-^$nPit4jD^@f92v$f0pb%iao<M9f!k%?h4C5?j*v=cabt^^*l&<pToU_Ev%81$F%
z3{|aU`t#P&zxq@@rJ3KH5Yob8bV9lvHgZOxsRU%-Llfv>S_ei8=YQT2CAf}@?2`x4
zVuWyj1#KnIuQ6aD!@esCjO#l9*8-UIQ;+A?9r}YKq+>@Z#~uLK)*&yj*^o!5I+Vc#
zNnh<>yn(<wj3{07FZP28=hUJ3%ev7dJQrqOeSC_vx8e0B6k|Ul>jJcxM5(D!^E7&R
zspa1F{M(W^Qs3#0<5ZbkD}Sr6Wf7fTAKj)3{^hZ+f`fWa0B$kz8Tyv&BZcGiL8eWr
z@y3|G>9$H)`*Pse_A-Q=v4E+B6RW4Sf^-YKi^SHSd&gd7g)LTHdN@3D*z#}vic;&w
z%^O*FD_ZnH)x{}<DuLGur~sNE?x$W;==;)6&<V^uX>}oGkFAxT5)HaQB~#SCv&}V*
z^t$^ubh)5>mIF6gSM+YKFVF^0&?EP{R!G)D+$J1d@9QSnKO>bkX#(gA;mpfbCs3hs
z?_zWau1MDjh6O-Um%z{wGv>Y_mEnEr^h4T%40(%w1e)WgMp6tJTDpckG+&>5OF`{n
z3GED^)Y5u$S9An246W3IV%0jS!79xD*(%NCwV|&lt+@2wrUl~Pek24(j)#-JbnWq^
zFa{y{QefcA;g$5F_f0wQJ6a7Ao40;_uVPEMy~V>CFaet*q2o{B-E&CJLqs`JuKhi%
z*Zc7Z&bEzfdkcBXi*fR}YKGf`5fxf5UPZHh7I8XqT!=*fPWLcZ`rkKw5dC0Ju2yAq
z;irow>CMpOBB%r$7j%zsYlR?|12S5)R*6TTvw=vr{Z_Y@>QHLKR{CXL!Bu){CmDap
zcx%)?;G!76W(y--RHqbFSXfGfpPhPMke-fnji<PgZL*-;X*2l>h8ea24^8E%>TgqE
zyzWw+O0&I}Lc@R-b^=Zj9$WSA>Z9)uVZhk=xq;>`l`^tk%o6lmrn*5_h{j<|@(sPi
zttIBpxQ-pu-%|`8!zF#NiBE&e)4d)F{Q`*K_8s7evKRrz1ws_x2PK%1TXwM!BVY@`
z1hCVR6$Nc8rhBdd#{K7H(5MO8rjlD8K}p%H_p+#|U{K@neEWDs>);6>nD2l`3wa1&
zg{`P#_^0a>#sZNX=XJ&Ot+Io-mH}6l8oe!bS|89vSd@1&QuxM&uMjMczxiUTreuTY
z2+#x6<%{<@B+E|?K^yQ=V;U66oj|%*^OLv+!5?V{6ie3dPg88L*7Va~Q_p{nu5Hd%
zKR<v*&vaH-Z5gB!C=qyt@%SXoYdA$yBi&4j?-|)b{)^))z;*Mrtw84PaG&Ye?>v-a
zV!)nTzPk4Oq<V97urQcqQo#KV;<SU`qrq(>F3^Vp$?Em-6Nm5Rf=Nk6VC|Q^%I?;J
zxPv!1)fmicpb!yG$K{(r*cyCKT^s?|w!bwZ-#LSqTfvZ62By04REC}%_~2+fWW%Oc
z&Qg0Z6EK2cMEK@bSvS)c*;y^j_U|)XqzV1lmsi7eM4|+Ay#Ni6m?hu0JUXJpy{<VV
zC;MF82|#IB^ZgR85nMl?<TvPt00<Hv;Ns6-ERwjbQ`wP%D?}0j=lP%sX*AWqvAN#$
zr#EQb7nYMt3lZaz*R}(-4_A`fE!L~kOx!y3U0?q4sW5W8XjSPD?Ge>x-xTPV(*&{t
z)vH@gV5*)c2a_|s;LU}}Wo821n}m}l)dP8P5|X<hrm|J-xon)NKv20UB}oEyAUrhC
zxQ{fxW}|#3dm<V=6VYRlh%t+m!dgx2jiy+|*Hg@m`U6k}zaw7j5}v<Q{d!*PlS-SK
znFOV8TuGa7fLA@9p&8pCgIngnQ}H^1+&?!A?ySga@@Uw!M&M2rNcYQAy^ry>dj6&s
zCl(u<n$92LjY4#EX?CuX&XG=t@cfMoA=ih*=B&|mpI|yku~V5*=K;@^Ik*FQnReF?
zLf&<tIyasHk5YWcEAVW|g0*ty)|(&hNA8L3I079FI(<IA!$W4h)LR0W37(dv1j$8q
ztys^SkfRPeU%TC|T4ek!;vqO5E+__+9EW4Dp}?#<+y6vhlFE9`@C5e@J7xFdn<B$f
zqBi%J(<HN}p!*oV)C8gnYFFIUf;LsOxLx#RDX8&GiX}g@2rF|<Alhm!%^pi{LBjgV
zP6wVaPFmG`*wK3z4%Si)T@A2A%t7wmz=7n;j50T_$ofS=7GUJBY=tEdcDQIspL+*K
zC(;X?F6RYNKCP_79XwwPajuaRlXxFNsAe|Rw0dchJ^I0w%s?fidp%*$f|nV7xcK3H
z{)BC$StC&Lm9~qzWDG$}x4Ec`tJ7JYx^A6YY1)>7u}Z-kzP)&yYb>?qOtc64r}-fZ
zT+`ZX`wgthYB#rtZ{CYF$_6)det|;Y(fZ<;9nb4D=rC$GJ`LPPL)+f&efspfDC`9T
zTyll4rG5PyJid8xS54k>UY;k?)`^sL%Y-s%6Q(?^Vk-{oX|cpaeaC;I8Pj^poG3pN
zMVrsX^UVCvOl}8e{c#C;!rjv=ShJ5f&sH8$%lx+@;wh^Tm`zt>idLzGD8uqWP*9VX
zFd7oR|DY<6uGr)8*-1A19{hN0+}BHJKr2w=HY7AYHMI0%kx;axuFad@Fs-vv(}q&j
zMLmNFp2O?m@4HVOwX8If;*b93jqq1aJS1NvafZKywIKK2_U!%sPq)vR84dAFvqQMn
zz7?dvq1c2&EU78A<s5Eb&#453P7h61gDF5KJT$2dZOsJ?GSTxNZ2{z<HGjc(HvN+l
zW;~M7@Zb;9ZzaTP%Goh825$SM!x{eJ(0-sCB7;&y_hU#5G~80WH7`lo>;hJtVG#Fy
z^lNS&sDx};5ID$=x;Do+BGGFkC9qii(0iv6T@fZuJXzy@`|9gD({_gQM#mKh&*>^H
zoifuXwU^qOv>WtW_T83xefWY(s3w#tRrQru5e}J;E<M%%@AaKV;Pu^puT$>4mzs-u
z<*Sn0>2n$H6d=+u+V2N6&>-p+ztJJQK>Ig<IjF1=IrmNx^S<+^_g;PTjD1|o@WB2k
z`9r}Uz9R9Bzn61KR8#Z`Sh}hjzbTF_e_fZpP@A~MnCs<Q3`7|IP<Nfxyn66E2l~@u
zeh{mZ{n9}pGFr{C(4oWYHykcuJo;*yV~jTX>3wTS|9g60VGkVeU$P7$=tfd}N2r1}
zPN{za;<{sYxXhe?`!KHgP-<o34u6Qa8l<@akwq%&roR<F3UTRm+a{?u$0!qZ3+|ut
zPXCc3FC0?E;fKs<S)ET$GUT&eua%tA1lFtY0d-c?jN`1KwZ1F$oojqi=fl@duDY_z
zc(gPY&kl@b|6A<v2QWyjD2>jsa^^C8OL;JK)s@zcPml{RjooHc%QbO$<GoG@J#l62
zt1J+H-Yo3xwWuk!eCc-JRF9683)_|36z)sX9P_w@sbYi7EFl!wuei^WTAFCeH|?Bu
z4;+8m*+r9Le`_n=7QP(Pn>EkJxUkfaWAB2c`Yt1Y9D_~<mG-Rn37f>CyYs&6-7B^W
z2MTJmESba}SD$X#*u$(CddoQD&HQe+zA;yD<4W`)IKkl|=>Lj+O$`_hQ8f__`%5o4
zOfX3kj{o=lBIu?<t0hDE^*o%jN%$7F&!EOYkBI5X8_zo+r2w&k6nbUh=-2-9`D_g}
z<}nK{pEpP!8rbALvMVL_^Vrrr{m_{b40;i6DV>s+b*ub0@mm(xd3&}QZO;JR4FdxN
z+8njXT?M0Bea`fg8$9ju{EiH5^8iAbR@g|!)h8uAUWVZ0i%-eXrhIn<tWCdNt^13g
zk-3b(e{P?~r*;BvP|_g`;5#*s_AoBbgUscJ+`!`OIy7n(5bYIy&{1u<cG0#;=rWL9
zx~3pSxvdybZXFcCcjS861D2Er3ZRc7J3K(!r2W$S3g)`(&rD@&N(si<h-A&&Lc2&%
z1#~HW&B^=Jt)JyklQ#*Jn&}gN<V@U0=$+037yj6)H!q5R5*pcP{nwCR0TU3=`!O82
zV_jdp2OkPQd>%c$16LM%dA#c6vAeIT1gr@h^q#f<94%;%74*`)0orHRXbNMMsf_jl
zIY7y?9|gJbA8o73gBS#04jDL|9m8FCS$|N`D^oDKIUs1pDroRjPSU9;`AS*Sc<}&M
z-w0NiY+;gLKbTn{y}`ce0xU3^@!ZeyJ$UE=+;}*WE98R)wGM*jeh!YwF21PN2;29t
zlP|@!*LyL#J2Y5gx{dNI@GWr0#32sb4EWc?0csR<#f4Yc12ml3<$6x5AqLU=%xzi;
zYcf9?TF3l*NJfR+Qq$&OS={4s5<PFE)LS<B6c~#)jvc3!LG5?5k>xIw{!r#ze6Ih|
zqLJ-7O%-|P(J!Czj_DU+oP!`3hI@E7$v#TFHRT?JNtZIiv9dMxZ0A#ufvsY^GAaRs
zw?%4S!XtU)XpEksmb*eN!~RNVy|6~M64bxn2<xF?Ip+yT*-6U{;WdRG6b&od8%{wJ
zh_TIxa#~?85U^h97Wn5mQp2Y3z`3qxqG*I(@U||+^CRBE)gyP#Z=VjT1@~^i_bHXl
zCo|97m<k|uEyQ*LwemJ@ffpFzyRGTp-+ni4^gV1|o`NRh@JPkIEVnnNpRB4J<KyEd
zP%7ORFsyDC9ch-0mv0qMGNpxg>qGK&He?Eupwo`6HfaO~j?XYF$@*Y}3~}N7E+d!C
zwWbg+o14zOX?BcEpM+%3J<<Prc&lM)LvI9H(6Q_n>OC~bfWRX9%DL8=ipiLe<2by5
zfO?VRZ40T$QAo_*b5!%ACRrBv_n%N6_`kSHLJE3VKCocP)Fe|Nj~>^}Rf}E|W<OpD
zXH83c?<vYyQ(ChFsvg)0z5Dgs*ZT5Qe($#DBfGh!!sBDxKfd3Gv1vS<AWY-T%kj()
zhH4(96Q4nbPKEEAMw$As<p(~spTA(pd;$@_*&cHMFU4F|4O(@z2w3f!j`Em)Jn<r6
zP;>IF5a?x%`M-8e>ZQA9fl#svBE@7b1-KtdCt*l<Da_kZ=wF@BeSzv~`?*|!gcriV
zl1^94K$Vcrg%R-OxZvAOU)p-VN1CsWKf9xD0UpCCOTNRB7)#7SITud0QyJ}_7X&(0
zH^Kpw60KSZS)BJh0w_59SNuXyjXry`1*|2nF_!p2v(=$bz10s4NyYTA`uBF3g@ICS
z&eN<=9b06DM>s6B#x%e1=LlH8WlrvBP$~Gq$HpoAry%{tKMy;C!`^}ldFR)iui_YI
zz7R_|iJ05ow*8!3hgfvtPyBS}ka*Ss$d6tqq+^rf@tbmxf(;SvpXW<&Q5ZY-$j`4N
z{DB*$0eh{zmVS*tvpu3&ePc0d4~yHL%CpH9voXBw(H@#%qB@+lI}EinmT3ZWD<cKH
z&@Q=m<`+01wlWX7|DFz97(KvM(|81K@j)@9!a(0XpkCH0$Y5Ctw*VT3I%~Gu`(HP8
zy)<MDP|EoEEA?Ru>04mYT(SI`1(YzrNxm>z9SG$PDSlE+!xkRJy3|>dLb1$Bw~`sa
zEUKX7OJ@=OUkCXDs+Cs5uhEO>{6)c-bqrA*mL`dHJq!){4Xal+L1F6I_nTG9@hT5y
zk7`O-@cF)K5W-Y`aW-m?5vr5jXrtjFEotzO^tS-f7C9SL>gb`V8#xp~lG;0B_!p%K
zeZ>ec@s^ksRVwc?pq(pmC=%={{wipSfQqORRn~33!~U-e#sG;>&BWT;iVPh5Jda_?
z;N#g>uO<2E(b+c{EgT%_c1)UlJ&7=M1KetPJ><Q7fA!Xtl!ow+GKRW~E9r~zDHg7G
z2_;v9X<@UV4TM{cX{U?YucPF<SmQ&YwW>bg=qR=dgQmPZ_N)mbxrGq}w<{p$Odyi-
zKuS+-8m3iG7$4C*?dh8$9Ck1xf^w<WBq>s>I_)=@5pOOSGK<<VxeI8p2iaK7KX7qr
zf_Dkci~)Dt?<k1Xly!__Hc^vz%*r#l4nY#u$~|0ZOO}PSp#b;P+Sw|zmxf7k$L!Pj
z@_)^TS!^3JnCKb44-05?7mYcmYmrC_cS}xfg=_ER61sN;Viz*rYGo_lgCd3k&L!uq
zn|HWFC<zxC(PnplYK;&F25->xz=`eNOW0fO1=Mp^?q=2NhB5+WE{uR_xwV#_5!4{w
z%zU*`8J;F~Vhae-^Z$dJ(eGaXpTsYc@_dMbLKV>?t6SJHUA8@I7z3|GkwnlX1{RJm
z_s%=qd6BV~w8DvZp%f|gMYX-E7ODRL0_dBt({RpvvHH9%n1x}vRu~8<(V=WxWhRQd
z)+UpSIiQ;ui6#~5%37V+&81lA>WU2-C`es7@aGce7Q*xI!%xC1pxx2fCr@8x_>Y-n
zJynzvc>E#aP{ogIR~B){o3HL-gNSVj>E|wH)iC}b@tGc3!Y)(fNS@lOnu6v40j8Ds
z#Q9kHpC$%YSx+W6slaUK;}#FxjLLlWaz$A9EWU?r39uCGcaz!Z<~44$*)OR6WPGf8
z7kU?XgJ$fGH}9|f^55fvJ~U-o>{O&>1Pddb78C2Xxvv%>%^!yj+k6@l5`Iq3f|uIn
zt1GS;=B32AgxDNei6RqSauqJ{b&9vwT!SPu9%kEMhQ(?-HikoPUX`DfVGgMLLNII}
z$L?v3B9GG{&L`GDR1WjM#C3t|Q?u~acX?<A@4;)@mdA-}K*ig5*_d?asQDU-cK-e8
zR)x!&aIa*zsvi=5BFe$w;VIIkt%os`Wh|X!D_W;HAwxG15eq7q@qSY&5J6yd(>R{U
z@CSV2B1Z3OPIbYLXKPveTbBgmEJPw71X+^NE2``BgH&D+C#0F(3^BxDyTd1=;=797
zC|yWg@plb3k&|nWhZLtoV01kyBJrO@&&~tNt#inwt``63gc0NaE|48A&{p&Qb&Ohz
ze4(D`YW+Q2zCO;`*1N&%9o&9!rMQy|0PcVQC!_%c)*i4A$T+z2voN7C8X5xzw{GT+
zvya}*<$TNO=6EBk5*!V>jNi~#;G<uW&d8DoLq+HOTbJh16`Sy>o}rq#2$3_YkNZsf
z{;hD(1?T}QaqPEN-3vNO$9kXfYU0t;ku?^4>mo9pisonraEZOMH{$a3>?Az0C?e)F
zrQ9~p#ZY4=`@zk8KupToI9n=Q{ETv)z;RaC*OYhn$BjTKFFWV_2o(YY4`~o8gw4EA
z6kW94Ek9aU%HN=^?Wp_k*7U@1`d>)(pBJ4Is`@O^@HBHE20Uv{Pm2I(+SAOdaF!zf
zVlN3%Y7f<r@SLbc!+<Z?af|Oq_v|f|gAgglz~jTQArQ0xhEfQGvg2!Url2JVk%idf
zN}~oqCg2y%>;Kg?@<5EQVb1eW*$ccSnju%Xk1}5KcaUz)uXXK7sf_#w;r-`wCWz1}
zDc!Wxm%uqyFZUwS1K7e!)yhx6U)mSx5`=Vl-l0Ms_~)tS7bi0cSN-nOo5;?Hp-E#D
zQ@UC=W6CUGelTrKif1^fSHwCNacV~0<Zk;|$pzRFQ(19+1bS_Msh+tycy@FC{L*IP
z&***W_QhFr0h)PQui@)cVoew{%?wzjZ7gzQ>RI`-e;d!R5U4qL|AL%pwop)`fV=C}
zgT;azf`QZ<%`>PAv;C>|zLo^OJGwT*fRr1b@{Cj-0XeIkdQm>40WMcO=+A*Uz*Ofj
zbJYnnQN|$UqZiBfu-;-u9>Cmh-_wFl#r@sU&Q1uQv!Om<l4XJ5gC?JMH0*5uIR2KH
zN`LiM180v(e|IL9s0c%|blS68qfm+)6Jd9>MG}4;P9u2HI0Z!i*C+oUB5(<VO=u09
zg?Y@%gPy*z9LlwP32`6RT2$j%i&iqLrqyE(S1y+V#z5SuhtOoIa8^P-;H*#@f-0B$
z(*l61wPEeJ=V&}cJsze(<~c047@hPX={#!lgK#j`ecJS1;h{rduwoY9h>hIyNkX`w
zsFA(uQ?<US@MUP3wGT-l&%h7}_-oFdMd<o_O9xG^{O@hhKv{N=CuVdj5h0S|q>+T9
z6M~t@_D2M)e=WRy=g!W$3sp?BJm6$=G^JPhiI1SjuyFl)`~`ToQ-ITf+57ycP1-??
z?P&^_ex5muwmN=;Jyk0oPks6h4e71(F9dZX!;U4YdmcO1clCN3Er<LvY|-Sc{DB@}
zZs)Ewb9|INW-T5Za7v{XU}{!-pqNfR+d*IqYgIfA;wk!*&qqoBmN<N9DFK>?I!@~=
z^lECJgOn7!eu15FY_Q#Cn(UteCJL9U=1{SrCE@iu^X<6$8_&Ul@TZ)C27f>u4i{U?
zpL3POroi|o^|~~Am<JcI)C4Q7G!33Z_pKSeX$yu7sWFEObV#Q@2&$YHH}C40RZ?HH
zVRZ)d>)0m_|9pO|ueb7-qSQFKuxC6OULKk{oIyqW*}{zfT_Yu7nHdUQ^Vb8Ge7~>~
z7qcXsHNGtnX4&zj-Q}vagglS1-_$Qs452j*j<XAsKG;V+v$~$dUrKioZ+PxLXU}MV
z2qi*QYa$a{_6{P$g2SV~8M3a5q-EVAK^%&o!~O^68TxE0m;C_;0V-#X><3gMguL}<
zS2F$IMc|Hv-grL7Q>Tj^BC64Tghg|AQ#I47mIe#=t&`W|TV%k?Qwf3#u4>ej3l+W6
z@^tH!G5=%ywZ4bEkfFn1vY9C38MH^|pT)FKb}!I#RIDC^kJwrXJuR7aAehN_eb3f0
zsPIF{+nendWNgWXDog(UU=!e_Z3)O0x@n8qxKGQeEdSs%R_>DPotD=sWicpiQQ^#V
z0nNVB+k{c|)X?SzpaCv&pA3H;r;$XmT-qc~SDy1Y+Ys|5mYTWJ9bM?T<Od3ObxGF;
zzxEtR6glg5eDK;FZaRMx3~OC4cY&898U{2K>i_RZ#D_-V;dx0;u#(o|I)M=p(dtd`
zzp(Oa4I-+pKZvims_l>hVOB3~f1#2M;Qji4dXzYg4VOax%i^QM7bIf%h9vm3+;0U~
ztqUg^N5gET;mfml^n;HBbC17N;$q$WB)W{z3)2bhSp>ulxcq;9JtYCQ5WkV7h2b-W
zWCB_<!+re5<(zEt(N0IPk2+1}p&12Tq=zB3h*;;_^I)D}<sKi!W9Si7TnV_%h2G5%
zJ5hHc9{RoY>)OoE22B#`zj_9{m7vEo7d~bo*rSNQmRtSaJB#UqOr+#~)#~79sC1p@
z!mBeZl!bH?V}kZzeTrpPg<YSzPhX!|KYWb5T05rm``>;O!2M}W=&XXr1<O;!af?$=
zb8Q;h5#`QOvLm&B4^TVPu?Z`c3}IDyPY}($F8N~0^)wJ(>6sY9)Bg4Q_<g6nZA0I~
zo?Qh|6+k*);Y(bv)Mo3AKK(5DxrdnS-E6_kFg0`FTGGUSt!rGK1x@ap!R2x1$e`Q{
zeCIVsvX1)X5u$N70$`sfLt~CE1hr)gAYv6kI{-S4(@oUnLlehB?}Fu8d;pVq=dGAO
zU=UHRz>bLh-_zcG0BwFgRG3{ShXkQPtu^^hEM51mz16H`;g{FV!Z?nB9V}AfDBx$E
z^5N9V*9^yBIspGsnQEv!2+1|?A+iDZ5~nDTlpmTWBlOD4Ap>`}+820Phj)J?yWb()
zi+dhT3`HYX{YMG_<=?#fc{2s5?BQc;e)oHzgK-<k>tp3Mlg;poA_IVa&$=pap-10P
zjlxTH+Zi}mt0(?v3h9XNNRzKiO`orM<#Guh{)Ytor-hT}f={uJT*OF1$RVR?HYGVn
zN5d55R=l#_DNs#><>IX9O>O9Lxd(phkN19#=8|@4j|#jMt1<{mhU+J&QZ46pEz2!d
zo~i6#H$jz{8GZufk`tZRgd7ju_n)y;QdzAi+?iN;X7Ll+H3r$+YG0zrQHO_a(aQ&J
zSanJ}D$Qk`dZjWL@#-Q%pOt9*&qe+7_kV!lstLj7>1A0-6JOSvM=u(5GyZsFt)G!k
zR?Uirg_DvTOEJ+hu09sMwYEk%GSOcy{qA|Knn5<Sx}T`}o|r~2cW~WAe%>Kk3k^gh
z`NAG9bO!7TTEOiOSGI1vc($V8ykqu=5E5ei6Zl=yg`$f=BbB6J)(C46@6VIVpp^Sp
zA@(n7Lgr$`GKmoxEz;dhsGZ6~;-jxB_Q}YS>a{AS`bg@q!b>F7>G~v+aQQvRAQB$p
zc=1`mU@TfMwI65cFn3IftVW0oj8<Au4*M1O96tMwdBXN81k^~Wgw^wo&SNy_7V4Ck
zMOWWh4p7#qbTl6?{T^tR{6}U9rTqW<aSerfC!%IWP$&6<{)2&bT#Pk*s(v2A)*4{(
zm2Flujtd>tK-6+$#F=_Q5tI;NUC)uEm2J>w*j^vZk9;zo+Mg{Q={sochy;BC$z62N
zI|#3|E+9w0EoeQQl4QopVoG=OO9cIqyf<4s0*436pbZdirB-z14|ycs`e0FT{pxKU
z$_Lh#|F5?je;M9xjrSd%PI=UKawU}N_{NF#qAqgAMVt=zqGk-`^I%+49;iNlwG}Wc
zxBi>i3w9gGlWk~3xE{lT<tV#a?Z&5-fj8@bpa8`XsMl<ogB2*B^_^M2aK07+n70xF
zd{&~ZenTC%_scwQmfapsLeLo2=mF|Krca_fC;=Zl<~ywxVGf+eN7)8gVF$QR+$D?u
zrry6S;|_6pHE-fry15Y<;j+d;``zg}Y9kXiT93P}G&ftJ{0CY^6cVy79Tv%yH<yRV
zVZC{A1a3FXfpK!KWm?c&wi%&TVn9f{zW3_?qN4fw0}vZqn!#`H8c|InOVN+JqxM)T
z>k9^C`)%+*R-9g>zGMmRSC(q`GAS08pU6+YV@TYEQT6{GF*R(kF$@g@O1h&j*(j}~
zG$c?jt#3?^Ew7aMGK5OoMzN>T1*c_Ao-Roe`4qEjk)cm>UNTIW6fy`0cqCSn_71B3
zI)pduiK3f*RNS>WbOPE|m|hxh{X*AZsNI4cxGIvlzad?UC$>(<1Yxb^dIp*<a=jGI
zAcT7!*A*$GLpOVa@kaYxLYqbm3*FNHp1mkFCg96rG{o=43o@`OT@<;ezh~@Q=Y)6D
zR}j|dIpsjvs%5D={x;#}+CwuN<^@<8MwFp#%SbQ1#e|#oj*<&A9)UG&m=lG}TfTlr
zGrSS6;ga|oA)#sr;?kIfNN_ee6Tyd*_Y9IWs&B0bJ>r^s?s=Mgp%JEy8cq$_R5_-0
zV~bn=Yb(iI$5Ad4RCWm*lrGI-C4G<c2w5dCy;KOpAY=KyK2}Wv4RolN@D|t$jal1x
z`=dHKlMb9;yxj=YO$1ae+*2<+8qW4Z>ZzY78r?h<UePxO3D$7KrT3n<7uNp{d8Yjf
zNndD`fw^oy)E4%l?hYg-cmE%4{r9o@pM<3WzX7bOO%t1aI*gOGc-9D7G5&C2>Vrif
z8F0~RTZbK?gy9^^p+6G{`KH4>dp#se)}$O9%miTF9taWS_3Lux<SZ43Zo&PU58v<f
zl_UL^Du^*6-IKv0;Z|T(6&a&h1PJL5bR@Pc5-+vfAIN_TgxJ{+5z}@8ClznLbf6bR
z^{xN?a9zhf<g>#nsT3Gp19MI?C+?+wcix`;2n)V6okA@Cy$4@>aZhigpC-3cb=8yG
zc2HHRA3X*>G9tVf=COU{F-b=ue*#s`@0t-LRMKKds^racGoQ%5H!J7{0XIlwlivIO
z5$NHX0X>IDsNZf-K@jJoIx-%pbvt;q2LRLP|KE#f%8hWn0VPf#`OR-2!toXR=tw^q
z+p-RHeu&DJI23VcV`KWvbw71T18^~ZxuX-@2l?Qk(0<RH4P!F}`lpaXgLQ|!!1ho9
z%zjrP3gUg>N4hYJ?RTgV2(5?!V^8d|1lo{V^CYs#euYH9MDo1)LD;!!^44Ee1No1+
z{QC{&qr<7piL!`x-jJS7Xi3zHW&vDP{H8~chbT{eicmF)3i1?Sop0;2nBM{>sF%dF
zt21d#$<*3V^A+8Ql``fT#cslOxByZ~)rc<KlSkLdw8x4_q@(;3I?al7XK~I^eRiFB
zRdV%#tnR#s4X*^h8DafK^4hKz#nS8hx$iR${~Xm_*E}slmL>m@IAFOC?LL<I;l$oq
zJUIG78l#L`|1#|Up)H({8AEHr)daJwVM?~>=XFZS&*#(KxqC$mK~4(VH4x7Bw_YoU
zB6#lZHouYP&toDxEA#`A^DsH_gDhT>z=<TlO5s+}QP?skguGwZxY|6>=KH%if7e43
zX4Nhh`m<5GnjrR)MQjxka3<Hha8d$~PB!BLvq*b}8g^)_zKI^?xcP<4TY=v&tK~8V
z6xREHE%n=9w(U34w9u4!scGB+lvI>jH4ghFTEWKE(*8Dnx=!u_vvPH?+3lN%bs>cM
z23`vd-?Y~jegIQ9cs?1A+>Xnb8R!KPAfESS5>fL`{QDz}b6A424EdDe;B%rPs7Eji
z@QfYk25N?;z#??b1in(Z|K}bOF`<QYhk@cz9f@$+GVM>dy-*oPtV+~lc!eLu8=Wtm
zmqBl6lnA9(LxMLlrc?Z(qOIRSQs}-><*kD!{7tVrJ?i=7H~oNyMPy_820iCSQ&L{&
z7Zy)KvL9S(_`5F1%3U$KaQ0qADWk!4V$9_k<ztSJ=yakPKqAnFW-Z$^r^ncogR9DK
zCC3CD0Weaa>BP9I?kaFEb+EMTnkD;Hx^C-;S|ARF$JjsX7tB%e%q6MLWrReOeyAem
zyzwg?LW^9>XDDx*O*br_`cvaK(<Af!FHn!IQGj*t>|_mcp^wI=0ydX5a>f{wVq3vr
z@dMxo6cHWX0b!^oP~wt(kh?<JrZYy|0wtz1-WuS@;1`RT^2v5Ou*kq2S^9l3l8X!|
zPlEMR%pRIqo+*>2EuK`d&-zNdA070#km~W_y4#u<_JWHW@P~c!cm&1!JLoJ5H4CP#
z#vnGvt$4Q~n;gUE0n*3al4jRA=uGK14MbP*CAF*Ef8l3R))B9aM#YsI2c_Q|H;9F)
zZ9&Uz-mFYIh|FlpoEhS|p49|e+6_FqN9>{!H=XHc3Eu;zuUD18zn%Oh${_wic0dmZ
z@6Z(3g+iZ8u=fQ2eUm|2CL+woD^H_i)`5+x?|fyP-dL9d;T1lfSVd%~A|GMIw}>Tb
zF-Mhj{XRsi8<a=1v7bMhqj&0z6kS)F^;`k9Hlo{4e<td(61kY=Q?TwmOg3*w<~|=~
zDSPc#=l9`kl*?r02vS`7DX#Ef!={QD@32E$IfH<e#RuOZe#Np%lUJUTeCg>u+pkR3
zc^#ayd(E3Q)dUa1>n@IOF!0{`+HsF!C4%Ol7{_;y?x9Z(Ulw0s`_ubV(^JmCJvBz!
z;4eOXKL)PJgZZq$`tHC60m84Cu^K{9XGyk92H&vnIWB<WsReEZr=?tg|0j>$RI8zX
zv(#!2x#bxaFNLOZj;F!W;Fon7AR)Y9WbLAiA&YdjMTI~4_Wl{lJx*ooAH}|kxD(cx
zvBwoG<*cqBzUu1d|H%zN1iX~8Mr`km{U}NP0}p4-I?2c3#6k#jp4VjoM)Y#6tUv32
zKBM1rfRfIX-yQxi7PsRi2AkpAi2UQzo3DFy_9Z3^&dP)0jt&NnTjdL~WTQ)g=pEhk
zOu8pUn%@7{ozg*ggKo?yGvjroJ82<IPUre5e(VgwM{yx4=hIblocSM!HXko0y-+J`
zCoe$W?2qOqd4&ROhlggV7O2i%3u+pizRhY$2<l+pH-w~oa@ah=f_zuFCA_Ozu&clY
zUq@Ia@vCbAlM-b`evQyfv2J@Y={Xknz4^D;_fTlXZzwe!PFugArKUMzJoTJ-8J|$H
z{{Qs#-SJfR@!v>7*0D0PH)Ul+D96l*WF>orvWueYknFw5$jB|pDxzfX>@t(RLq;h@
z=lL9U-}iGrujlvck6tA@=en-%`F=m2_xd2k2lcuSNDTC#_H7uNz`7`#C@;D)3HD}r
zTaQ7j2IZ6sIv428O&TAVUjd^!16!vvQ-oTT33}@`MR%s{k0uLTE=XMxeNM<=VH^1o
zhvYRKV}Ci5mAo>n7Xl&E>(J57p<qxD=%J+&!v+O#iDMuThAP+24RJMN8Z5C<Ha07t
z=hpL|2|s5Cy^7zS2QOc`lw4!*T#3%>(2YnA5uH=LQmRr5>CPoHaTrV0IG*XJ!Oo#)
zUvn*RbY08zj_ek>R;8i?4v7F_)HF%V2v;atU8SW4&4C-RwpbZ(HCW(+KYd!kY$dpW
z1n-sjj7RT9M6V!pL}V$j(=nWu&X~qwQ#d9_?)=~7GSu{hpR8`Q_e)FYl5>xrAbs#Y
zSN{CliB4FxRas<uPkz6oI#o5G1*kC5cb2n|ykrDc@3(3!?r3h)5>=ugoSRep+GTo^
zX1>P=_SY7vb&Z*p$=spN7&&Zf&;giSf7!bx19cf%9;vzTBtLgjC7I3>KBYy45(dmg
z%TegE?K^W0n4S?YB7_h~bjAap^92w!?@^<J_I51%_#0&Cq`vgxG@n52X`J`W;g!C}
z1$fqB9}_qPD?3OG1vsfl2+%^2Xt|*dM;i9tCcvtYIk8v++L*}0-4?guaw<(IT}U%#
zt2}~_!Km>2;r@p))IXPc$T4in;hS@o>!bliq80?4-FRi2_O9!Hac<ljqKbi5&Br$J
zxBO=M1Oe>O&mWTmH_FqznP+RG60eFxG3`ChEeiR!zk^~sC=h<`l2;xR2acS|;S#lO
zA>Qu*AW-btK=AhuFs+a_7bK)1C9O`e|M|i14>0dJ>%Ro89{wPwC0`jVGQi%TV@BO5
zgn+hP<wMo_-%JS==wCbdp|W6XecJuIsMhD+gi0*m=kck3_>@15c}zp&%C(ZGD(s_#
z#(UJ1C!FQrc#NPG@PLs<II9lh4P(%Iz~GWaUlh2zKjK2&eXF|q0my8!E{2A)-8>t@
z;iZ5W&rcLG+Q<nxX|0IZTY6pr(h$nGa}d1!vr=Q@gplDH^j})rwm2sGurB-aI~4%|
z5c=o&ibz(dmtp?_9-wmp+8P|f3R^Pt;A<eV#hJub>@z=p`b#Tq8xC-4N}1QKMK4^a
zfMx!LvwIpo`vJTWt|aZGMXr;(===f<JT|-p67q*5>tR@Le2v-Czdg_I6^b*5b=-Qx
z^oxxECb7E$-1{|+T>!XMm?&MqJ48>jKA(ZdiE54)ll`r1#tEB`73+65Z1m?G89-o2
zgE#rwk=%DA|G6ZD%AH^v9q}FS($2NRK&*VG=3KDdW_u5tK%m0v<QoV^U{yQ&6Kd}#
zdU<VNZu5n~vj9&vcUeQk2QEklHhs`@Ta;V3LM7OD77WT0V3ekb$pY38cCKzsX*s~-
z$u?XW{0*KOuH$0Rue&Au0@x$V6X}%&?7Kax5W;+5@<N;j7TgQEa_cWZu9U*BHw)aL
z8Ne_rIKu}ThTt4`fLMx&j3@Oo>}5dHL1mca(+9e#E#RnO9$UtR*P9hBk}Lhn*v~wT
z02<@+pW60FU_uPQcBXmJ894C`G~D8=#%YrqNwNrdfK*4IL%W%DuahSwl0|mV&kS_M
z@Q@f1AZ4Rwwdm<^sfBB1FuIjk5h&eQyT7y_+CO&Y3na9j;9kHoH;Pd`;^q5{y_j+4
z?bDBIZntj$nFRoYzGLY=av4)_@pls>&&OWb;hEnhI(itL1tBKzad_a1h>MKX`bxz2
zXiI+#a`A!=Up%i(;|AE2pnKQ@*fd%&DB1P_DJK)^@(<{4qnV|nukYZJ4|v<YQ$g5&
zB8j3Cg$c)k7KHjYeF2H%ZE$pT?iK3gfok~PhZ`?Jwb4SWZK8g}&kv&ReSV-B#(|*%
zp=NBXPhtx)8dRizj+U^OkBZt*FEs;5#$Z+<WIYJe08@rxU2Ce*q%Fhj&YVw|9cite
z8QT>DUuqVS_c-iUi{RH(Ce%7H`TO~?dIYiN&?C*GNO=h#3@iX7i&Wm0q7uZ<2KVZK
zZ3>QAbL+&(X-;#eui!vvcrWP^+4qni*XvHJKaJdk&J;IbnPJ03){!rr&@ExUpsQ_k
zu`Tp{?|SsZtJ99t*t-Bm5S5G<xrr05{i^TG=#zyFxPozX|Hl-}iNo1mDM`+=jA~RF
z{e>&}VIO;lksoOKSnqUAA2)mbr0ofv)|<rML>$3SN*7k{SM~znfMB6%1N6snwj~TV
zhyuvj_8qn*8{>dFPmp+t=c$s-GH!v~yBZ2a*`SJ@HfC2iA!Cgq7;aiRwOrxt11AF8
z369IpL6pZu^mXIoLraZ(vs+_!`JzuiG0b+01*m!ER}FAW`f{K<<&=e43h|w?D@TaP
zM4u!6K_a9)501S9n27!4082v6ui#+3X3XU2Yp*0loTlo|9kGYfFLjgqCi~d}66*(D
zoqI6ao!&iSw0adV65i!XyS$d4p85QVU7%SOgXoV1dbE}fgw50ZFaW}X=0wW?*pYEl
zO>n%XRS0XFnmC&72N(<h?dU`-+pH-_!o0fkVq@J3@?`zq)0+^8-HhMHMy;uDdN$W=
zEe>H*Mx-W4wDWvQ@4wlEfezv16~0gAmfDJL{@F^Vx!FH(LS2TA#v7(j-7m}eStiD}
z6hQfs=}2Ij)Q{u(D7Ny{hLm=76~C6rW0YsCDNA7>l^hqDP$XjT1s3_NS=%S{JhM)7
ze4s`JkK)WgaOQ(fKd@3SUj~vYYOllMr0zh`s8vOj_A!0bS=uDrI1mk6oK?ow8~lnM
zyka#^1S4!`TJ0AeWyYc`*{zU9D)IFu+EBN(+C~LSi7kxMl&x}g3ju*x;(&7-$M^zH
zTrvF_A=e-dVehq40<!{*qPWlwL<;tf>y|Ye9l72KY_IC550O+LZ=e9toy4EF^^P+%
z_;JakvrDsvGX*3bAagaJ2t?0dEwokvuzkSHTi50{!tkcx^gNciQmnJ|#*bh16C2aX
z(b&^5Y&d!9QMPu^i2&bKQGso2Wrw30A%fc-0NTx`w+k%VCS$<oB(=)2k=l1mof7vh
zKh4-X-zd2PO`Xi{sLgth1e*pQ>LMv`C!xbkrK~=MlqR>eP)qQQxC6o56KOHL^RWl~
zZLEtIbQ&In2%OsrE6myY#u1m7$~rS}|L2hezt!Nmhn?DpvId6n_)ap|PfsE$@lgYT
zrY(4KmNsImjbFQ6)OEh6L!ZmSDLfF$dY>HS6g@^Ph(?9R*r$#ivjw?dSVmH{uy_k$
zAJfpHEs@}TMrRoXYpZX!^~lQ%E(BDnozcvwi4M-f`{CpCt6p8UWaAap#UAW!7?wR@
zw09O_U5Sp%NM5orZXZ$mRgdJbg@}=VX~CW#$#Myk=$LxHkBHqcuk)+9B8^7cq|y~|
zcdE>hZ;6mL_SDj~N2J`kO86`B4y|e1=(jYdGFeCn3MJ5}iV(vYp6AzPnWDbx^yQ02
zScX_8kJ4C)p}X0I{Y4ep^^e>ojf^1{iEKQIZ?U!}+)T<vhI-yCj!;-<YsmewGv=GA
zRgUG#G7<O&TBuw@UK(!hBDJj1moyU04=J>!Ai==vE^3^T96yL7*Kus!pGs1Xy1b|?
z7%d*9BGP0*MnjT|66<icubmAST03E_<iPnv;zcojn0RaIi){NTX;i6LixNY)EO`>C
z#5t*62R6E)do%cS2>;K{aY{CP-F@6z@;Ii`g9ZObNBYEFa<{g1Fh4aZ(6ek~$*0)0
zCOB`_3Vg@n=GrVShS_b^LvACK9FZfk_5);lSYD!%JPvjxrL*SggH%~8K8XvDX#*9`
zKG`EBEDTt#t(BLm+s7}eDoQ8AzpHdMD<q+Y(DPn0%>U6cza{>5jsyQR&BkT`^Ju5!
zuy3?WM21!(wKswMd`>GBUk6EGr|fy5sPnP1nvM*#Y{HSD4lP~;?#G%r^Eop*LTQ_2
ziB*eb?aW2fFW&k*c>wm?RR8%Kw|ws3P}n{}!zVx25x|f((F#La0u0cv$_Fe%sKd;f
zDVXR7kht%mNpcD&@*R(};S<=;G(ULAebQ=JwOHtVnr(N)no_?rvA3-`o_Ks3UdoTV
zT**{jD2gO799gZ26&4?dI`3L{bhnzs@mW;ApKZ{?&rKX5EKOxEDd!*?Audpeh0!C?
z<lB@=((Ky@5d3P08V_UtbF-TgtdawP@PeTzK@z+vbB5j5(tCVp-eRSllqK!37Lywc
zSc}{yjykTC=UPQMeAubNs9fTzlqr8h1-s@)4075R_%|)PKgtC1zqg&Do}Nj=crmQi
z0|(OG5&wkmykd7lTUV;lvtqL5=P&&A)@;Q|He~f90zcd?s&OE!JJ`Ki&|;5pIUeP`
zFlso;b<~(g?3zQDcC}_fWA$X2Fw>_z)G9HS-3CFkTNR6xDXu~tS2a7K?dCK`cuzPk
zAl3f)DZ27<k;`27Ou}NLo*$VBkY)T90_rB}5QeBQvNRhmddrH&%LOJCq*&JOP&Wuh
zA~;PJK5ZwX!mh5T;i8;6e+fxmsInh2lf&m}C}m<IzxeUPrx+`W&!2EZCoi^_vuT(0
zJy31oesm^fXEMgfp<06KUWAF*`HIc+Hp}9Mj7G){96vUTVh0?zi}mt|2DyneL9I)3
zKOw1COa64UxI!9ln#Cm|iM<5o`u6TbhRx6{M-AH!3wpC=gRU-#^EVK0{uaX8ZLGKV
zi1rvZ<)Ppl&DU&K(wKNdmQ#D!U?2WhQX=pyA_uCaD`pge=Y!nE+Ujeg?R&d*Yxj|B
z*M-E2aT`PNIp&F5TcgDoFW+^ACn}(A+EG@=oawj}(bGnT|NIePPnk(?Gqmj{qsdX@
zm%y(ns#yzLy)klg0$ru*y_ggG<cxws(3^(Skt$up_awo3W{$C`?2w~xA`_e9?G}y5
zH->P>rsJuvcTN(`MF43eTB%Ahy9^8ie8%JTLI1)-E}?#^{R_&5yn+=ngo$F`{{?_7
z1SGPJ<ELB-f#iipx3<)s7D_XxsT<})?m6rm=OU3q^hC^_F+ke3YiigZ6AMk@AXXN3
z?Mp>ZG>#L~NWEZIP3#QtB2wJTGvc&^kjYq*7}=mk)|5@DU@O)9E0V$!x4JpB65PUD
zj^DS6q-t1OhJvOAFFa58+*jJUgoMu6lqjg4Z89D{{bB^->y}^jZ8Ia*c6(S(I(Juu
zdtTAr9YQ2*3DZ5}zhEoH(QyuU`bNl~&<-S&;$4t`?%!7L9leEJxEQiPJd3iRdSURX
z@YV%YytI>gKj93W&Ws&{X%^vjVo;L3Ok{H`xp39NCg^$jFgL*)f17}dHVQ6plsU14
zzv_9dQA;g5Se3+=`^fsy@|fFx+erwRQNW6RH2)&aqf>o^`1fXO#<qvE1%?Fv%7{1O
z^1uAJ<7oa|nu0O2Vy=Buyycavl|6L$b)#wg>3mXtZW*!?GF$!jiRsGnT&6$ssXh?D
z<yyLuZ%Gzb%7`z#u2z$*pPHSDE~(trYbSZ8ZPLBJp2A&|%^|FjT=H&>(BlJgLhE^)
z+WJ?qpOHdD3_A`>9BqO}`o6-jm5x1nnbKKSu8@qCEq#xO*7xCOEyG)Hi<tAA(^acd
z%5OM6MA@TT;~gdxQ`}4vqwu5f9705A&Dl1|JxgKVhriAokTu|#8S{kc4tWZ@GqvKV
zX8Usd@wD97K%#I`SJv6#@H$t0L~Bhp3-v6|HngjgoL*YX_OcM`7O{VJj6Bn}+x5pg
z#bI%*dLrsMcg&nwHOp*V4FcbJ@gFCo{&rM3l*R|^QcCjj%)D!4w8s4D-mpb}rdNz4
zp_*n7j+P$?nxBqSPbH&<Wni!-_AD=>^Y}@vhm0S{<jo08Lb{#tUI^;<KSYe*Hl4K{
zo@%g&ryU$Pc&6Mo6;dznGrbxil;KSrmYut<%QejLQ^#6(SHfnuCYJfc%hFUY{~#rq
zja5nRtxZwsTa!1pNyB^4g7d09j*`t7X1ULgeKI?DBm@?3Ma|U8c1_R!>OM$jHFR$+
zNhl4LwS{_0fz5Y4XsqIEVJ!kNZ#P>P+oJ8xJ>aZQfBHS5AjVrluJ+MFPjXT(dG<(&
zUkgRLYMLtzn>*{mwNl2mSLIi$30@@lgr)~m89fCM*dA#KMMbh@NOiXp!n?Vc_BggI
zZ)siO$U}_RqM&NZfo#yg+w~MRwR_?|%1YBmJ(&>St4icOQJGK6^pG+NvGHd)9rurQ
z)qS<?eP~31BB0E>e$uA!xi{U$?&6b9P4|x)L=ig|N-A)2{=i@<REGq<%Q?|fH}haD
zfsnF^xzzq8lb;!W-D2&V3jMxx0$L$%i0IRYSNYsT9vh#3Up+j&L==)qY{$4m{)u`$
zR4I#Cbf|pd#@Y$-5(dPw+iES$KD)fD!Sh@T37JRT(cY)u_vT|It_ciqPbXndUnc(R
z^(pnso7jbjhN#TFBj#=Lt72Cz6`FS#J(>5gs!oY_ma)z^gY8K=pmqWrBTAdbeAO}W
z-Cw#d$(EN<xNk8g$}J~VKM{H}+=YiO_Ec-h?QjhR-Sgh6B}iVSLg`;cOEd&bf+egR
z!X?HY@r8A6tsBRL-=D8iDS4A1Bj!GpX4IcybE9zH*l8N;ToF|+!imx#j&SF?w5=1<
zVbb=3+><C-QDU!5xr4#w?Ve+IqZ_}3f&zb1t!-OXQ*9uxau%ntP=oq9OY0fxi#&s2
zR~`0GY1oqEPp%*wV{@ck94vw&<yq3B-?8x>Z3IiTMEN%f<;4ZIe(9I{gd}F!5s>^K
z5K@=)dX0*`A}t{vwvs$e=M{MaMNQdE_R=eXB$qPTF|k+9!Q7o$PKF{!eYV><=q`mj
zlN8e;on(klQV)+yt-iSXm8dg>ea*tfR*^C?;m^69zd-&KaYE3C(oU+hEK+V8E)8Bo
zo1GWU4=O~$)r9pygs++xsXQrS#deHk=j-_h=|@`KmN1{E32)Jdo0^@#^>&@yM00tT
zC1%U6@lGh~#^CPLj){rUu-yi|w@gj<V2aJu7yz1{dGJxNiW6g6U`Y<;a3hrVE8*F`
zZ>Puq0ram8kJjFWg*^?oqFCB|m6>BFrC{3jP-AkG(MG=9&0>L~TR~Pv+FE9l@WP#L
z%>2Q5>kpg>R^9B<QQ}+^Cdu9(Vn%MA3cJmJ{?(i9kh2%B{A*$&whhG7#mYb8?qEEA
zL+U4qDoNAj$dG%l_4xV@M)_w}Fxs(k9*D6jX6#H~pjod=T2vlOxY+7^82J>-gop_h
zUO=b#s_g@OFP;er4M~X`^B2AdcyRGv1CUJ?y&&6lzu*an1l3~a?zod;KVho$;M)@I
zGSz}}#W~*M8zL@ZCpZ^GyNg?%?z-}8Mk+}(baN1i5@~mzl4;=%k`>`I4RO=rVnR^S
zB%{@Fj{g3!;sg9+4D*g!ZPA@(t!f@zWhtHg+^eO;wrO~Cu}6)ggwh0<Vr-r_QwF$`
zRu$3G#8U7iQb}a&!hpXxlD~KuW21ency)OvBe5_oVtQQJ;8-21xVG^dgxTC{dmZvl
zOX;k^H7{ebZ`Ifx5oN^+r^lJ1C}djdqSDz<8=CJRlA^;u9^+*n$@9*bCN}y?O!bz>
zYx~&enoJ(`2c|byWTfJo`YVWR3B1FW=skUFF<$6An>iNqEE8AaZk}DadX269(ukCR
zB|CoQe+64$@{F;#^`yf#so~m4Mnkh%q=D^FBlOfYEdZ=JAU*Ttgyk=|_pqkRFA^vM
zX!nt{lMGKfD$R&$J_X?~L@1X43`SU?g!hE@Nhtnj;fPMzQ^}z+!#Uu@W4?a!$XU*B
zC?AS-x+kG|Eng&;++I2ySkPFG=3CPIz?I<#$iXLzwX%&O=%ERErWw^(F9nJ+bC^ce
zlwW1k=Tt=sGQxr%hzYROblTjoh2~il)o|S(08yNL_(Du2phrQs6${mUodl|`XXtX9
zv~hF!g0x3oFqOR7A(3m6)3Paeqwi6LhyBV87x^AFZthRoM+Alo=L&kP$9Ao{?=a<;
zQA;y=&{i&g&ehyn;5<bQRXF2#oM|FQS(hR5w?N=;r2yx7Xzw@&dKz8w2?ifex{}Bb
z@B6B|N9>QCu6sr|MLsnw5<*9l%+eGiVCs)L(&G2!2>t4#>cgFCfYLvuI#^a#P*}Ec
z!?f^6(^cV3g9%QfeC}DwqZyA%y)=}m5oahDC-E-b87Agt$R$P8Q%2133h2C6*iS8b
zVx$BkY1o}Rj+QLq2eK3tS@RU;%&uLsau@Hs68@tpsg|wm(y{lP_ik#ju#DWf^H5#u
zQnGH7PF8hQ;~V$AoX@rECv-h%=ygT!#E1Q8qO<8TuBJLqXdx?fA;kIDfEHQ%wamJU
zzCPR=w;hzu_AV{OJVHbbmCgp<LXEbXI$&tbjtr&+X5g==PXU(~tqA}0LZo&wGDYkx
zK?wiA$=F`or0yv)D~$rjOd4hOm;N(k>JCL<sfRtiplKK5B^L618NSrpI3}WtzoOsV
zbfr*bF}Z|)q<2#9|JoS-V9QEGB;0Dm>JCYoYc;8xo+Jbw(7_B>+P4$tKc^1cw_b+>
zbXW}zEGI&<InOvc_c%er;^lNc{FPBN$~XqsSO-P+%vt)@Ph5{9QuI3UX^(?G2R!28
z)}>E-QU*Hk=}G=!AX2vAucUf7n_~Ez5~<(cPa3N9+v}qDDn+*wpB_QS0m{w4dnbQf
zYx(>1DRgE|76Gxbcw$XOj}VUj6Fa2;{sWEwqd`wPLLi(_1IAGn+O;`Ou<Bb>*lMsD
ziJD${N9DcyGPp1yO#fd!3J!I7h(lc<{HzN<0A{#DW-MYP%Rkz#LO>|2P-6L_sDNX2
z&GmnOMF`T#ka4%BAl-G3gUB0PPC<+i!?U96p=W<yRes;ett8-4;&m%ILn!kPu~d3U
z5&qsEaPArozJ8Eu0^)w!ycUs}U)Or&i;II%YKZ?rcUJ5<At(;MJB7C4kxZz`8l8Qu
z)tjWlE2}j?<=r$w<t_gAQ+>}F6%b1!v%lF4TIoAL6mD|UZWr(l$g_jUh?3n>W7aD(
zo*YM_ci~8M{EsP~kDt%Yo2}T!H`Br;`W;>-^83!JiwrUlD+YkL9OrBV24_?_sJ?RR
zYG}X}U=9H!b;q<QUR3i?0%}TWgsP`;`mItNT;_!&&&LTG6qhB@nVV<OnOce9g#`M<
zt&j?*+X6^C&u42EF1D?s5bS&a$rH>DoIl1&`=g2`48t?y<AA7d^0M@o3C0`D_vws)
zTH@<chgOcEj1i<_ldqS8aYVrEWZlI`&xKGFSJTJ~OqIH_WWgNVBFu7L0w!#i?g5{J
z!-u)Y{NqOQ;<6fWhpfLrWnG0M-)6Wmnz-~r2v^SrMAaDv)_@=Hd(ZI$Rf+4u+66O`
z7Knaa_Ayu~!G+oF4%^-do{uw=GKAJ8{VGa1oIVqi>&cO?w{tymw9pavyAAP_D`2qK
zA$ANsTYgs@ZFd#u!o+^+Y5c9^gR72L3xkhlV1gYR<Sq?i6$#X{aWlDtJV7-BZpJ9#
z6J*7rj!D>3%449oSb;~tMxlcshMnl2f<`47iceD#Oa;b`bjsR=43yt%wEe>%{!K?+
zN5y(R&JcRXq?A@~j1#)-<O$33q&ivzc@gU-4BiaO^?w}I|CS7%h2%NL++#C+i{0aW
z0tP;i;Rh}PVzZtDtt;NYDYj4!%wMaJr1u@Ek^Ryy{KjMt)MKK|<~1HXl;>u?N9CzN
zNL4zpo<Lqn-<+t1F|`aGGe6ar3D;(PJ2m*I*^b?B+^iJS5*eh@MnXNFw#0k~-K~z|
zQ&Dn}cgg(|L)VG%S3)sQFkS@?VY?4*wRMb?ed`=Cgonp9$e8=vEZcS~8Og%fIF-iy
zLr~@LHRJ{FBg#{618zPFc6{_&H>r;azfPbyG0=G(j2@SWo*R}Z9v2^|!@(62nY#`=
zd4@vddUs>)$^iX8*64I-;Qr62e!8BRe>|lz$I$|NMtJ|rd1Xl&Wdo<U?6&GvPUOFI
zA3MP>+`|svU8vXFvp6$I_0=y_Q{+b}s(|vJxA@N=bO=&iapVcc3l>*|z0xZPi0HuY
zkrmxf06ux0>Pn(BHuA<0`tZDKP$Q&W7+1^8DS}Vj_ascN8Q*eRKiZ&1r(YE(lqP1x
zznl2!zVfit8MXiSfc$ec?5v=69nAU5`Cl#K-$&>nr0fEl+k<CnLnl+ja62<Tq)ouy
z>b1OLdiaHl@FL7Wyjp$ItNb5cVh|@rW4-gr@c;h~npDA2GkV8h9hD*qdpC&}C^I!f
zvP$R<|BAqb2UFUrlUMHk=k9|oO=<ekeTr@6GX{scGIgAqnD)^M^`DBVgBZU@NIdXL
z8u%-0<1shQwwA-S<i&AThfN9mPm?^A9ep43U;XBH(4m;GAN;^xi5n$(LB!}X+>1aB
zjNDL82y#;M<FgM&q95YMOu-q|`0A+xeYpqje+k8S8%ggSy&^K3qM<u@lAOfuxBr;x
zV#B`iqV2Ed+FXx?E+8~xo+T(83ZMIkg^kOmA^v&C<guRt>pgUTSZdXB$wr>J-W9mI
zS(@n*5XpOg!k3A4Lg^0TW-w(T`E%bYqJQtF`6?K;izhesYBYpW<jq}rgn$?Q3J&0<
zEvL*wYJQAY6sC3qKExhsd^y5pq+x2|JTLNVA7X%nuT~Xrtp3^%8qy&|Pa`B+*$S+c
zp?6an<A#<jb?;y-xK2*(*$ti>p~>p{v$ixa0+(tKkA?58R-Pw&E`^^dtp+&+oC`(S
zHDvCFRj_2UsHRVzv24Hq9nJ|ad0f|nLYvo8cCD-#C{y6?q(2TRzkB8sN*t7gcl#?0
zu%`qP@zrCtrfUkJ&v1#eiaKyKr0h9Uz1H=n$Fcc}G}G<dkOs{P##2SH?m(o1m*Jof
zyWa<yUBJ^GfHTZ`49)te<>CrX(D>F+=iUbE5X2R*za(^f4<k}1N6T#>X6XQkQ6L_d
zLb;-+b?kahd0ahLpVS-mv0BR0Xh|Dy=-suCuKxKyLX}!bQeXeTg@A!(2^h(tViO6A
zI={u0Z|pa%0`ioROlh!`f^*;N&ihiZXano7q^KA0N$rX;ZaMqm9*JK@T)58>d7#UW
zpQ^^23UKL?#Mx1R46Xx97bG>hfF&BZN;cT^a7-g)h^{UQ85Tu?oY0nJv*Q{VU0_Ru
z%ggMif*55u!F@m)TjBeXj1=)(6fVS0jKYQ0y3eWJTm%^AJ@C$;;8VGC$u#E<tnZ+Q
zc;!Rj;%YizQO*Vai0+hG>Q6t+5a%6d(BXnY#|$WSIO%|Y#)1=MS>x=)edA}iJv%^)
zM)dL{aFCio3K=fz`Udm20Id^Fec8T8t5k4&BSMNwM_rBp<sU9aN=q>CyHoq*p^!i5
zKHko;MLOel7N7%hW9Rs0PyZZ_Say?q*KD}Sgt_}2hL;1I-nO18&<zVDwJLVQN}%g7
zy+NNH)7AI!H7E=C(Z*pO5~ua^iKTrn9#H1rK(SXhZ<F-|$Dx?z^ds~U2DVp!)+1+w
zE|+4m8mqSfkVgZN-`~8rC;ZG){EbbIP^y7w<@5-7rTR{|mW|ObSeY8sGSlxlJN-Dl
znWS&sobzGKquEM%Nih6E#g%~b4E1|s)FvgqP&dJ6&9GMus<CGhxfI5;@7*t=XY8M_
z>ba5fTJ#~Z{FIjyIXox>xEd(=6l3?z)7g5>hO)CI%+Wiaf^}@|PYhFLEYj7i(k2Ok
zOh%$|`kExYs+P!_YmSP*N&JmTtenx<Y36d*Cn{ULMu%~PV0-ZWvT!=44!quDZ^^4n
zZ7b1~{q&@N#Oe83?2FRk3x0#Ov216g3TWhA)<hepUDJJW$uqF|X5&d~7<m6`duW<(
i5#;e((XvSrfTgstUngOC@g@Tgerc%api3@T2L2yeBoZ$G


From c8701bdaa79ddd9e04a61d675a39c9a675e07e05 Mon Sep 17 00:00:00 2001
From: dongzhihong <dzhwinter@gmail.com>
Date: Tue, 5 Sep 2017 10:08:56 -0700
Subject: [PATCH 04/94] rewrite graph

---
 .../images/multigpu_allreduce.graffle         | Bin 5704 -> 5489 bytes
 .../framework/images/multigpu_allreduce.png   | Bin 118413 -> 110982 bytes
 2 files changed, 0 insertions(+), 0 deletions(-)

diff --git a/paddle/framework/images/multigpu_allreduce.graffle b/paddle/framework/images/multigpu_allreduce.graffle
index 588f79feb9a3f996b6d6208e556d61f0a9602aa5..cb5bc420ceafe8ba4c87694d44ee4e5e4ad06779 100644
GIT binary patch
literal 5489
zcmV-%6^`m3iwFP!000030PS6CbK6F;{T%-aUVquDyAm|-N7nI{EIF}r9b0S3_Qs{E
zTQCVpSVM#YLYAGB|NHiU6iK|q7sQv^sfs9}2N(|ebf2D{o_X@mua{BlOOW)!xck#%
z>XFB-pxcf+VfXx}$L~&l-ZhW^`Sj6~|2cefaPs@n^VU@q_R`kTyJxRn9<(0s?(OYg
zT}8p(-r>n%>*&?X<C7NLV{h;I>&LCf7ioI+<KEuQ&5h?fDX$%0I>FxFQ4(JTN&4*-
z+-w&bc%8KK7#`EVd#*e1*G||@pFaBG$>-qP)BScDehFUr--6_2w-bDQvgiH^7sGBE
zoCnEM@?`J!<82docG&h)xaIr)orjJ5B=OycAD;Bm1itc9=%E*1cEj@|zP|F_z^7jl
z|LiOZ)RVn|$e1)$)?*yH&$LpCjfp(jyS;V43)gA9-}%pVFC9KQO|FBz+w0H#_GkCV
zZU-KBbrH6QVuSBK5OWf}+iSya{Ph>vX_4<ct-q+0bf={?|NcGKU)SIu@_XgB$8i*n
z9$IMf@^B=7JE&}YeEVX@>9;5tN#1t3e-UmOe;!>OyWGEinD{rNZjCzj`ptNEU_gf>
zar<fXp+AJZ;A0$LPN5J7e)o&tdy|Cc;Yby3$1)qr@i#eYaUA{`l;lT4UTPFa*O%S>
zC_L|$Iz`K!%A}94{5ITTD4zs=^pu10Wgl)u;TNaPhkjb>4<l6X60^%pi?JUm|B>_7
z-wFIH_0HWSPKxbc1X0*+eehi$@(m8d-c{s(J8t_?>8SwJ)>Yd2sg-^BT5S4r7zHQa
zu1d{(fJIR@)5XKMeSH~p(;{An!^cYFH_+jpGbt7L&MRdBNMlH%f-+ANO02R<P|D!t
z0<=A*xYU+2ODz}1VU4k!7AJY2RUvHzwG00$sKVOqAf+e4*R<U3_vgnI(ouX<7JVD^
zG94{Sj{$i!h|=&JmiwZw!tNbt9lz1|o$^*;Abb?}!gTEQxxZ(63<i@JYl#tz5^klT
ze5m2Kjo-#;rsU&#GyY6c==)y&4d-oA>3j4p{Lc6Qjtf-Vr8>hj{^V+l*UY_(U$CGn
zg+Ao3j`aI9Nzdr1-|dCj|JqmQlu35lj@z;CE=eLD{dIa4chmG!bb3Zk&$=P}Y+v|E
zFGxuX%)^%;1#kK~xDNjD9c0^a6el43kJBXRFey*ZlOX8251*pz0Dfjt|9cvIz4F0f
zcSiE<-Z`N2v`w9PA|Q5p<tLrf^s7_^{twv8ekkL9Feb7ehPWS`XCmBYB0^*$(09^r
z?S3eli6{*s?!IBt_0#Lbk52z+yFaAi`9&IZ&-^x6LGtKP_GNpVwNBe-q;>GU#{2Y#
z_XB{qjK2Ka8}nHC#x>@_0k76L9Kah%;^-&_ogZ&IG;m|@c$XUO5wK&L2}Ptd1}xRC
zl%8hH7)mrz;I1V9UeL9nN2_Q)sr88O+}bDT@XPL5T(6=|qC!Q78dm6C2I?^fJTxZG
z7YuDj@kPA|{LakQOp3d0ON}*U71ngB7e<9)o&Zaz6jxvi1@H-0|9S)$o_hsGf|ue-
zfVfoL(?%<BTvl*qG@bBz*4jKi3;qm)BzN;$w*15Ri``F>TqlQp0KO3gx0lB|GjMUi
zx!?YlYk1~Iz5D|l91FguhQ5vbou8&b(yeU4NM(}~Ka6U>2ASR&z-?w$b7kw)yw-Hp
z^Cn9n4l-OxP7EUi!Xd|-s3*CQ9Mn=;&HJ>8Zai(GjnO7DBW<Ea5vB|br8WjeTO(~k
z+O$Tr=^4_dhfAA;o{Kh-d1;e1Jja@d0p3K72cvB?=&#X)aaQ-btYcABb1jJ!3IuIN
zqmVZtZ<@rLbc4JJ^j3NToH-H73d^bEO-y=LJDpTO<%yMgJLa3TZHhQ)GwXa46_RoW
z<Vg`BrN;RtBu+Tr)C_U5+LK_kEI4CEq?VFPW~Hoiyt{q*-3bl0YetF9Lyd}7ZiX2p
z&LQ+B+7d;SWy0!Ns1fCocvf%%=8{;!INxar!aQcIa7GkDdBz1YC1gsIn38FdDG4n+
zsX=`;;YI^10;40SXQg$&NyW4Qt=tZ#q#I*OYDT7HgtZV+38FZRGc!n(kSHNhs<#ar
z%u4DL0C(CZHGiH9sGngLtThG8O(b)?!Vwj!fo)JJ2r$9=yw~I`*=Dg2WEcxG49T0!
zVQJ|Z4*yp|DGEvoyjOAKSuV6zz=ef?cx5}1SXws5FvP43LozTT@ahbs5?BU9ih&dZ
zlUUnHG3Lq%6-~+vIU!&e<7;EMaTJ4inhK&B7#yO7vb8y(IWkH}IUX405UN4SVWjY+
z;L>tNxsgU0m;Panh6qIsaqDR`+Y-vr*t!`#>$(}n40sJi4Cjn7O$1U7q#Q^&F#WTn
zl`(QI7Bu(D7$bzo-3mb^z`<}P*evNEBOyjJN?SmaP{5h&<Y^jA{~%K;lYiA{`iD!;
z61TXHSx{}^CTDCm8)LZ#T4|`GP{LX%s9Va_9961@r>a`bves&U5GUOrdFyw=>--}!
zYnRXTAQM;(6WEWUw?XH+9ekfu{yGD6n8`deDLR1FVaFhL5Ic>t=p^<#ZNHZ`h#kvS
z-s{jcLy|@$A(A%13u%}9)x;VisG^3cb@NFIk%UOv1fL|db3(>=L~u(K)exC;y*lU!
z!4WzLolS5;K;%4Pwd7V4N~zRx&5$GzNr<G4a6(3Uj6i>gA<{@qgj+17-QR%eaX}r(
zwj^_@SW}rxMRBndW-hnMN#kOvrqYCE8a-8-P|<nQgo0)Zur)9uOBn-I;gZYR)lrh_
zd`($aRHlmD30Frw<eVS1jTXmS;R+F{(1d_)5zsgxxd*h|GG(~6!t~iXC!5YOq-ly{
zNIfIR5SD5!G<WrLmJ@*-13AVzaf}bAf3p;M#`Kgx>iJ8wX5U<Jyl~t~k0=g?*(#8;
zu5g@EPckdCaFhS=pFoCi2SeDj`YSQByMJgU5cBOODpyuzx$V1)-aCXhtR52aI29h{
z{Cly?zMBf-aa-`nnhuAHy1+`h0O7EKaM+OhUNIptS%uMF^E1jF|FX%&uR@mS5LYES
zv^>yMu9a|iDnH;LOxjV!4AM$4hA3Dj3OXN}!~mxy3c^*1f-DaM3^9Nhm^|;-AOt|u
zT~QmbCMyg$14Ugbfq5F>Ae9nQQERs(lh9b5wIrBHz)S*W5*~0SVgB9ur`cbBtmxQf
z4v414S_pHsaf`-)t$SKfw>7#}(h!D=3$`eAAT&-@-ph#^!meodr0qLVW2@&T=h(;y
z0XP)^=iaw{o~<_!fEx=y+2l@Qw&LY8zPw2tftAP@v8{VHz{)bifv-?xILiZ#!kJJ+
z!E&<!hL%ajDl=k2A_foxOU(vwqCG2(5{eUP8OsV`r1cncK0tC{(n5*tNC$|<iWddN
zbO5FUFdgu4(*a`cT0<_3TY;dBwL&rp{KL&nSdSZcq1Kg(!=>6yW1YQrYGAe7DO&?~
zaoGXwIa@6x=hT8J$5oJ9lpWwWIXV9j-3ei`q;lbT5RaJM+u%Z81mH9P<a4Lzn35iK
z>-Dq%r$@D+eLRi_^MWBBw-b+|$w~qX?;S!jp%jxuD(?7{(+>l@S_*3kQC4rq<~_Wz
z!g@YyVZAb(N{3S^0d+*+VX7Ga;^sZuNe1X#WrckHN_|N^X2HH#Fx3XMtDo=ngpt4#
zBvAw|)-Jc3he{wBSb0D_*Hm`}?PDd{x3p{z$8DC;ldbzQ0U*VmcFSml;D&)Mblb$b
zN?WNZ{Nl<8sos*4fH`lBAF~<xu?8DtwP3{UZAp|nQ3v@k@?+%3cr4#i_S>9~va|W;
z?}Qs;L6oNz_!Z5$6~Mw(Elan|bMrAmYE3w>dbVvlZsJLN%a}YAhBr8QD46u5(N=<S
z1-}8x33f(Wh$k7%lC<E=Dcufo#B5XKh?!Z>EGBNw&<I8ex6)9K<Os<TlA}h*(JnO_
zyeMVbo$My1G2mHurSvpo#!#Y(a;wB^r_ZXIPinnYCmPR}`ebt#M$NPL_(Mv4B9B@M
zj}lGtDD94xG)icv)(n);%`s_@3u2XkAVEq?`v<77ji*tvF&ZUionx{BhL}6HN(`|a
zX%y0^b)r!Rn1fyM0V+%^u+_5ZZ{{lT&=jN<bJGe-8>AJAQIBeOS~aDLQ!r<aS11pZ
z%}Oh*g0N*<4rXJGZFR<G-RjIR#-*f^Qmr7UPy%@c@``mj^AUN)s(HoSl@?B)1{=R1
zuRvZgH?NRQPFH|BxqSjk<v&YXIYuq+6M&9yxqD?WR$eFTFi==0yF3gC!~kMoxjScz
zC2u4~B(4xfNHh(JfBRp;Aek8#HH!U95?^0c<)EX;zv=~@d|)z8>;X)q0zd<d)I|m*
zjmvjyatu}*h2?Kv8Hr7d#j2yR%6Ke~ud0}=szW(`d)Lfa)$Q|*$J$mC_F{yVe-owI
zU#58}N}FnmI2VppTe+&CC|SCKP>r>#w|9jhtFd{NPpJ&irU}w&!?a?cc2^Kp8%7nC
zprjG`U}`pnQ+Fv`?sF+eUvRO7Q;JOgOqTjmoJr8Hceay&P#x?RXyu*6a((o-;%(<W
zXGNci66jqQ4Z3E41-i0nEXxE0vjtrlwW`KTP37;<m2I9YlZ4Oh$`G>+V3se4*$jz0
zEy@I9_Q7D5AZAyIS<ytuh=_ei0i(RffYFE*M2ixmMT`~~ixz*MJ6fEK3C8(KCQ(#1
zEH=wP*JUy)a~X)Gb->bZ*AdRWEXD~Z(CP`m*UiLnlM)H_;U4nBO^oCyMskayc9G<t
zSrD~clKy5MY#P*JJhwHtWr*8V;?^!B(}AePT*p)qXBKR(WFDr9v*J7q#*VYZj`uVT
zGXTYECSt0*3r@r!nC=}+i_F8^2`5goNMkjN!>af$x*0^q0@Vmj;etxFI#h5(1)^fH
zf`ofQh0r^#ep#k))drf_rGY}s&Ab}nL~@NeQST<~^{TXd3pJdHzJ~vD^saT#O8oBU
z7Q0K?N27fd`R(8`$WLU|ZJ3d-r7D!RSC*CuQ+4>nZ+>zfcIzr{GL6b_OrL#Zvi!FS
zG43|Y&Fx1}=xEqGzW#LXz&6%_;of~%m$aFS%ru*Sic{0YrZ}^6I+@mYZ!%s!*mJ_J
z<T&xUiObKlzL$y5!DbL8m09|-8;#*95^=0uTD7~=3%I*Q>mVqZ#(T}|AJN@ypt~~y
zv-7LO@}l(-M)%ja8IK{F5zV+pZl!3J4FsZyWyEr(OzQpuQBzd06_iD3g@GsrqU%@|
zl|h7Yz52NtjVp5yS6CvokwnrguxMMGM#92bavD=dl1L^uF)+j;jD_*MNQQF3B8+Jd
zLY~J2Af7i7&$t!UD)DT{l2b5<W=z7&vAsv8H0v<gV{pe*3`*Cs_NtGOb(rlj3}<@~
zSld&F1z)3IhPC=dZw9MR5GCtalcmX;07M`_5Fl$<o~7tcOYy1zM!LE!O{Ax}qM9j1
zq%m-5S%{HXyM?t|`P!}htagj6N9mT*_gcC|rK&34Vjj!E5i7YK+Ptdd^UtfwMUUmk
z&H-ln?#-%d4-3ZG(@mT`mF8Y%Pp!(hIbFrx$P(jKq~lnmwQ5gh7Vu;X7IvuqDTs8k
zBdhdw=*TwEktvBu{Z&GC!RigIWq$=7N60=H$l|)Xl|oiE5icTQA5yfa?=M<3d<CJR
z#!wMM#dR#B${@k`N>1pO2G`Ocl(2GYBN?GuU}d(njRbKMQV8{-9&*A>jN-7GVqwfK
zko8jwR#O<s`qfc%3d~|8w>7B6t*cgwT6+H(4#X^`IHp=g%QHP$#r%uH{Zuh%T*u0$
z0VY;70i!rhz#x{kr<lo5)-PsS$$Bt_<rxT!b*x`f(yR^?5<!8WSjU<rVRu^6j82rI
z%fp1iswAvR%2y>l>z^n^*P|>+nB|`+r5ZU=YSeB}9u#)Q=K_byJG^>)<+owDhc_L}
z=Z(5D`0oFS6DNM)G-@QAFM9IJrAx%4<)Woe6>?>%^Qokgsib=FEh`GEddc~xapAFY
zuH3gS4bo6k#`R?M60*U)4m&~bRowm@3}-$o{p^_Z`>+>&8X3V`@!y{x7v*0Z`RBn<
z7l-}H1+;wyzXZVpa0F?Pl*DOC{QrXB>LfmH`(wQr&eruO^!#r@GJ5h*O$)l<zIb`S
z3j`yReC=OWyqfGWtqSEif?w0Je142>T}l+TgWh!AcpnBg6K`9GGw5x6oz^35zZ+iq
z>F8@#cO(;S0|ShI3%iG5FD;_d6kwUj9tU7y{WOfb1##k$f_#*9^LZESSh+8qI4yOi
z8~ADL@(r(o?s<AqdH^H+f6N-fIQWgi1GDxsbTGY%ufnP1^I(H6ey&4L8En7#=iMNk
zMZO2<Ilc&jw4nGyRQwiq!ZUEwuGfcSZ~FG8j8MHx>@JZls(&QrN5)!zcOzGNR(7K?
ztPU<8#@+Le7vV+h%uv3`>o^Uj|1xiq5FBT2c!&0YmFxjXgR)G0b7z&F!5jMAn?lwB
zxgIjFkDtY98ejI^?no2H{Nmt#XHo2@c|Z?73z7%#$XNBC@FGsaKjW?+LBGRpLFLL1
zYR4UC%PSuAY(M?8^WxX&_Yaf=|9%#IeEA~Y7yo+pub0QqFZM6qyd#o+`tWPiK0kop
zJ8$06j~Aa_ypQ(b`oVK~a>9Q7<Ku_dWdG%d=ljQhfXvTdqnm?Y?(Y7p;4EEac$_`C
z>;6YT@RMA|uj2N-?0UuB|5B8BnQ>5>Fk^^O5N6o?p`jHIb!zc8_ySHQsL<6-crR}q
z-M!Bja}r<O$L;erHG2_nx`D%V4HtMD_?<W1=-Xtlzm0G9qwu^t1qA`o0^b@?<WVjE
zU=9WB@kyAL?b3nY{o?m><e}f<5<-!z8!mR}Q?C7X8h)9Eab=C2Q}(;<V8*uZy5S$!
z!6??a<J-UM1<$)_nA|n&@2kmQ*?NBnyPf!^-o_qWCkc#1Hj-s>n~$Ti2G0k9<^T7G
zc#K<({NZ4IK{8J3#x~rNdZ)!TbrTN||GJq6Q`iq>@OJYZ9}jkJufM&Hf}~ae-q0t1
z(&Z0fXPnjyhmur$c|hH9<bQe*ICvMR3FieSGjN%2gY%s68jiy*m13v0%Sg6Q3^j^)
zTx%hXrAkR=c3PAg>QT-pMBbbdX?4XgeT?JF5=)ZIx__E-W3>hoSLmB-^xluIORujS
n*488HhpC=7Cd5hZAFqyf%hqW0^C|-PJ$>|FMQQRD-+BQ6USy`d

literal 5704
zcmV-O7PsjiiwFP!000030PS6CbJR$({XG1NzI@q;yI8LGW0qmV7-nED11vVMLpb7w
z=$36Qa?7KZdCUa=eY4c<Zav*ExnE9*z>QTURhRN)p3JPQdic+;7g6VPkPO1O|I>Zy
zk^7yX-;H}=|LmvxZ;yZ8Gxz`b=-$KsdGhM;`1jXOJC{*7NIS3JK7R51uycQJfB)d}
zG79$hpBz8wyngZg=(q#-*x!Hp@_y(3d757SxWE79%NNggt-Nl0;TjJ1UnlWpkfh&U
zz|HocgV#%Y_u(<ayBCH7zwL$H^wGT^9)1eGJv!*7;pgCm|1C(K_j|$Dhx_iga53zs
z!C8<zA`kbkKVEl%XNO%sg<HNG-g(@~PZHmK_~GFoP2ej(f)RT0ML#@C;;T#V6@2<7
z@lQ{qKt0?aHOaN6%6g2$@R?Rhv0Rgf``5P~4&f?|4|@N(8l>Y#r^!{Ye|`P2-~Hqs
z+3&&QF3-d6xY_8tkD9p_gX?SKVf^(MdC(!>A9VhrQql(<t@-!wh4H!$hmk+1_C1QD
zaPrVnm*-C=?XM@5<>z<U>?r*f1(TN7Lmpm)TgIOzS96z#*PkT*m&vdu1AF-@KOC6Q
z@l0HQntT{e;UM@B#}_jQ#G&8+><?Zg;aNBlh3mP@rZWE~Cq0hBKZ8p9$&^<*#nIJ8
z{~!v_`jtV^>Y%dLN0)vVZZU431b+00gYab^uA9PtoHRf2)5>@lp?Z&)J!U$L{Yd$b
zoOk|C;7?_A?j~_k?*A-^!hYwy@5WH<@FW~uM*g>>t{+vN3P9~#rk$TU*@v&?u0MxS
zaQyAE(#;20l-p*q_$2OLT?GBKjMwq>vC907WO(3IN)5h?LRkV*4oOr{=4nESRaOZ~
z8N6J8w#O8g+Hz*8<-$0uG4`OtNfBsuNSi?I+`kO!uy#F3>2dHit@iu<=}}GVDE?Ay
z`X(4;GFon(1M*}NrQroE51YOS`!}F9f1~`JidtdR_;ox8)7<NGzt7|t6ecm&5+fKT
z+)6|FSi-M6zlqaK$n$cOe<mr7eZTmIi#n<DJqG7~FF%3#2K9QWNipT0T(9xkxp(;s
zmSm+ghT_$Ue4ixgDLwJ~gE0H2dwE8gWGCIY8~g5(B;wv*C#P{gO+Q8_r{v_cAHvV>
zxt|PzlypEnd=65urZ0o5;2+<iZ8whM1RDS2Bnf&<%9FDs2>R~B$LK16pINK_Jqf;E
z`e3kollI-g8KCo|OPzWm(Cp;WPkJZmSE&g63)stkDC2%GCbA!fxF4KnO}NdP2$3~`
zv6EqM_e0UFiPF%--8W47etMPo(a9fO_lq<<J5PiDsow=HNbcRszU+Xr&Pn%_bPm5a
zSfAnaz6TIj(N}zXIg3^7+@KyD@EVoFA-s_!j$X$g^ZCAG1($ord(>!;fF9FKC?cgX
zpsDtx^fY6}P@;(fb0zurlB|s_T21$9y(fI<T0g;npZ8DWW)XEQYD9D_VWrV!pw1cK
zu`+SCV5~dJFX~y~_vY4SQrxUt8nh{^(WW!KFe(i51ZYB~xB^`$fKRaI*CV*_+$%5=
ztQ1!QnoGq!ZL|WzWd&zO(<!fKqt)Z%;Lk8f3OB!I%Rl5_>_L(g200!B@Qo<AzMLP-
z$ixL_e)n6U<EbAFiVtvbEZLq0`Zlq5ewqeJzqSV>wOx+=Flzi7WOic!x0zeam2DF9
zM%C5Kn=A!C$Z#b&F^mv!ha7LBp5#Ju5KCz_AJQhe^|Xn$Mw`f-w22x;m@+Vx+8CH^
zjkF1A(+1I|$4Hy*E^QKeA=*S1rA^lG0&5}$coQ`q%(l@WzeW?rS<~yXjzv+;wIos~
zXlOGUg}e!Q(=^_sTjWiYS>sWwEFs!}?l2s96O$fgoWn0bZwo@`Zp53kZHqW*GbeF!
zQwd%Xu?iHl1p0);35gTNn_3}GR(leZmIY(Xh}2SY$*h!3hIen+es4;LouWp{EJBUS
zT5gVr5|_lYf)nr-i4|Z=9&`j@9y3-@&ImZ;jEhDl%Q-miL|dYWvP@V#ome4LLZ(y=
zq_Q@dlF-7F8U#=iFyK~#R4PF|E3NyVR7@L?>D^#Tx;3Vx=447nSPLGNAd160GlN74
zi4qc}X5FyOti&$FO4Q<cE}(vfS<u!LG&hmVu@Xl$s0OY=r69lroAX}Nvt--FLXcrB
z&oFqKIV>$b!{J{gl%gP{E*8U$XSvW?0h1O2{FU8EVrkhL!w~Z_49P%=z^gNiN?;ib
zDF#vuOk(XO#aJjOR8}c-<b;4><k!Y<<0uC4G!;ZMP&h;jWgBxs3uKg#a@;Y>AykW$
zLs;&CoPkQC7G#%b7fF@c<Cz;yIQ+}_u27EFmK$q^lmjWp4pEK+q#SdwgM}B`3d%iW
zl(rxzLV+%@jvau5XQZXVf;e)iO%pqi)Oy?keK3k~%c)h49cXY53H(kot1R3G*}?X+
zg9Qw^A=V?*RbN>}HK9fi*Kr6Xjkbg;P8GE~V#r%t0ISK|>o_E_S}@}5I8iQwiG~~v
zImQlf9mh(-8ET<$#==WnLEF_3mRU0<MhK6)LIEX=6r2f$>o_pvu}H{6wz`gkOV1Lw
zN}gE|PT{gFHj8ms4jf7v%4in<kb=53+3iuKdU&e3WvLr2_J?uO50W>2FT5%~BGY!o
zB^$^D*5uMh(VL)m)eXK++s9&_#o{Kj(4?%w*M}X0*g@>H&Z3jp?{)n_+9Gx=S4FEs
zHw;M{k%UOv1}mi9DlSc|A%ZGunA$X(q!3Anq;0TCLOUa5j7J2wL{SZXJ=g1lju0H7
zgV5OqBLqawBUVdpHKCNsY~jQvAxR*T5J_8MgpBkUf$<PSq>-8kx9VBD-vQI(f;y0~
z-N;-j)>h_HQCxzAnM+)PgsYybK&{f~nbL%cF21fQtEAGvh%99cM1@N(ZSz7VT`cd;
z>bq5$JK;j6yPWf*ZSD?8xfO0d0xC2iAX@|^PDt(nEw@Y=Zmlpww$916a|~(P;uuoT
z$uWebS_{oxA-v^8Ajd$Cu}K``{mI{|3uKXJ+@MSJ;<eWE?-HEf0me#?C=QC*Drjfj
z{xOt#l3AgJ3;x4jfehgWhOljg)M9RPf7fam7uyzGu5G)7UEe1B)*-xQ8yg@VXTqag
z)23>O$F0UAYug<z%L1#&0=UBl++jm*d&PvrWEFaQ?MIY*{zaRMUxlpDA+9QPXmy~e
zT`S@4RDHl<m~^9>2+~?GhA3Di3VQF`!~mxi3c^)|f~*b$3^9Nhm>&0Q5dt9Tt|Sgv
zlNH9gYHOEDV4emzNTq~S)Y=`%B(zp%EeU24Fq43pggcx`Sp1lsSq{p{6&<_G0nyZ0
z3vRA9ZqXR9bx#Xwh)`NdLl`bD*rC*c&={(`m7}qQUDGy*yLU9!PRAw}*pUzcI1>QP
zdN%Gr0B#uo+uR?+Ik=UV&-m(sIsz+^(PP{6Y=D(jh67)r%y3o*8ikQiM8Rsa0ftsd
z#yT@%LLvqb11rr2aAG}5IhWLMtEhBjNl5secDwRYBaAy7Za2~aqP5~hK`|YG=>SXz
z+}(75Sh&`Z3*%NGXk)FAOalLKu?g#O125FNQgOJ{xM{4@*UX87aLEA5jnmgudz2{y
z3XV#*6FaVg+@b6M$Kd4RQ*<MQ$%@K_7ePE?c7KZtc@coK0FW=7o?}WnuZ0mUgyd8k
zI>ck%nHLQ4xSMzsZB`Okc<<nv38k1MQgO$pM0TQ$B2rjOh_ZS&Ht*rB71r~43+t8P
zR63kW38*6iPen)n7dP+OO)@|iDl6oRSL#dZF$?;|x?{vZx`z2)PZ$Y2K@vsaV&ihV
zMW_Ulfweo-3x&FCXdkQ4zLjNrIBv6ymTWVY3EcL{SWJoRs4Go8!*hRFW6Wf4)MZ+a
zojM(hu!St=!iRk`%RJ}h2&pyUU@O`5ee6o*{Bl^rFYyp8+z@ZtrjBj;?cl6?+AZ%B
zf*S@L&23-pb^wr?!hc+OGsy3b`0cHo`^_017m9v6`t5jJ=$7mQyjUoKEuL2pLq?-K
zt-!8m&aI%7tH$WT0t^bt5t5^dq#NEMIdVH%aHg5CV7)nLWROZFsRw)sREAVSQfc{)
zkfYXW{Q0~!{=_Q4w<OBIPbRVx=#U*DJHo@%+h9j#-iXn#$6X2-3AGqMBI^kl8Oet1
ziekJ7DbZ@<MS@9B8f_(LR4^7Gz>X5R+6P9R0=5*c(A}U!%(g{|n3<CjG2)hY7{Msv
zRvOBY5g{YOcu^~iXpb5VmX<ol4verg2CVO%l%8hH7)mrzuAsMZr9)l!X}xzUy|&mY
zRkrXR7K<Dcb(dDDB9B@Pk7_ftq`C5}Q9?VhW*~$v?WR30h*d%ma0x6Oo^{E#o<_;m
zXq1>Y?PkHX)7-IDVu<BPqmV{z5{){<<+5u&>yn8jPQ@{`i+1pQOQaQZ$gQO<(h9|>
zM>VtHiBQEUs58ealn26Qr4?3z+p;TXXtUNbR&3skm0^raNhPIPfm7ka$H*&?S75>l
zc}0UtnS)n=G}R_|kXInDm_u%5t1Dr2cF0j>r-qdk`A~~nx}o7)?Y4%DRZAKj21+H3
ztHXdm3?K$pd&rSk@t)5_VlgK|qHRe0+y5K}Nn7Qdza;V1WnB(Biu}t#&?`D7d14P>
zA~gUSVWi13s7PFKSkry5MlY;*^IA`AsxMaWjn(>NMSRuyWOV~7@Y|arXLYwPcFy&!
zC+uYpt@tJ?(O;%{#Y>whL|kyk>a|?mR8%xwsZoQrYu0zADQnPqwNI&a(PnYd8r`(A
zqjr-M)#yf*g`gr4MQ3V0fm1gLT;X#iM_)3rr2}Qn@c7K;T@$83zuDMM146yCTcVXW
z63fkz-<r2ww460#Dz`xEx~$VR2Q1K(&7xVRAeb#_%BWQhR%#}Dho)@%OqnEnVN-^f
zZ2_}<Nz7(Qe9)mxAZG6jW(i_;otPDExQvL{yW}v+TXYzWScA7H(OX1saXD}C_l3R1
z>6l=guO*10>S3{63c4wiQJc#^ENudoe!Ge=_Ocu&oI<Oo0AIJ`$4yHlG`oAq3%Aje
zqv**ki`r$9e`ZP4Ht*OvGq@4A+lX6+xLqf1?J6=Ih+52b%;a%q&gNR;Fm;?2<1pwu
z&f`1YQW$0giuD9y>Z}U}Vh~KX4yI+|FgL=9)6COY&*HE;zKdoCk+DQILR+|?Qmqda
z98rO&Sgs)9mQW$|gU+xlQ@CmaP3+1*Ar@v{t#Bf_!JKIDB^(Uuq<jY@oQa`?|ML26
z=dhFb{ZAcsldw-_`*q}ZgNvXT$ZFa!qgYE-Ds8VVEfc2hoWS4w<SgtrRo-M0RXnRF
z`^aSVcQwtp+bp-RA3<rL;o#`%<Cz0nZUE!G`>-x)I~SR0ws?qB)8#^(f-)Dn)Hk2I
zQtmMIpcV{#ZsYPZt#2jpIob@OA~GvqcB9cfMIw&1OY1gwdI@v4Y#jt8vskaW?IW7I
zEi`vVV0M0;SYEb1!sy`|H{&rxGol&S$gLI4vIR#Jv5Z*Ggh}0=BWlVbwuZ7Ot<Vug
zM|2a*qB4juuGb%{(YP`PafKyPTS+9%5{tIAX~iv^C#Ny9Cy8Wo8y!O|!dM#5%Va2*
zEW%()3~Q&<ym&_Ea5wRcTT!hO&xWix1%qhDB+LTadt^eh3BewNJEo#jx{<Y4LyT-f
zw8t=v_8_o!rw$9g#;^=){fpiVR-YhBHnAp4lMMlgK!6}XHn2QP(FYyH>l_&A>b5kI
zp5}^brWBFJz@=3oMq=$2)@~JRw+^z}EwUM<TT0(*=@yl$u6T=itN=%><hpBdRmm5R
ztIB1M<;czfX8PvMsu~Xq#^~ubMo*==mFTHe88@fv*c(}*zlwAmi?r5l$;=X#Y{|k7
zH9Q58PB&z=)(#EX78)`oF{!^!$SzsEp|u>YpyLSHI|Er<SGQKksy6&ZMC@Ji7WM6U
zi-xbkRn+JzqN})xWmFj?7+=c?-O}J%8iW#7PHiP4R7<SPmbMiqZb}NF+0{c%xQ$*M
zR#PmE*(I`mYRPH}BiXPziq3#p^yGF1wYYWFT2V`HKf{5T#T3U(%V>EfC+mp6DBMpK
zoyJY9Y#L!=O#v9iF#v;D+MQx1L)oyHX)Wu)6qaWoFgCG%NlCLlP)Gy?f?^YEmV|xK
zk!EtD6kQ!A6jmi+RZ_7k>GAMHDY_YDNy4oDL@CwEiBgk(qw=7zmyZRGg?D`Q=+f`P
zbdPU3iswy+GWzcSi4)iS&`H#!ak1#hua+(mk5-G8KGw9WwmO@sRWehn0es7v!m2@X
z_A%c$SI$-V)|E*bOG;i&CNCkI+^et`3|_?DPr*3yS!-u=t>1-%@Z&@YUN`^!=~21;
zv)BGvFqXyfcnS{fP{1#s;UO4;G)OAVX{Gu91;OQUeAM-GqZmi)h5<eQTaZkiJQmZE
zEVwUTo$wOD$XdSiFKS**_LxqM@Vtiqrq%YvIl49}QP>Rzvkl{27<`#}+a{dBXydE2
z8EFUo@WM|gU$cH7S<@~s!1%YY|0EowWmK9HEVH&p0cco14dZ^PIq^uTeUuIJX&>}h
zbu7I&tqi6g_-X9&4KITJS$bZ103*Xc%o@Tt*p1Qyv;K1oF#Qr=hBLJ<f(?fFsR=z*
zu>Izr^@DVt_5(oA(RmQ0CBYw~;<vaLo`RWnqkfWm)7Lj;gz7zF_lWFJ^`n$O65IK^
zo4LxfvK!^FI=cKM?w@@)56@$#hKgNY#%VD7mwA<hU^olYJJtuRq7R@o2+PbjcUtK&
zyrEBn8QMA^H$&#-(c?Hx<BOr$ok&8?E{^VZ8pVEE1oY^$AVu&_l-2MF&*CKfGw%Bl
zj63X?M6UXvZrpRayyiiV57Iw-&wh=5e@{v9@5j-H=g;B;@vq1KdVciu{NVi6TO#Sl
z_rFHnvqSiQ@6}uS;r!#XchLb{KYS{WkJ+z(e0cwo96W#j^x)_ZX!G;e=*!_RHxK_s
zaGEYNJx-t8bpO{t@RLHuujclH?0U`J|5B8BnQ~B?Fk^^O;AYt3p`kSobz<=*_zXrS
zsFBrecrR}p-M!5ha~xmZ#_fwbHG2`S`hmlA0~dG`_`O&C=-YI#zlpybMB!O~1_}aA
zOMGiYktenMqZkV4<Kr-`>ZL=!|JffD$iuM51-K&FFr4o&q+AExH2gdZ<H|ZaqwM#)
z!JK{H_QOA}f{Cwj!?%As2%h%SFu7^k-_?`9s`h>#_IvS{W*vKYl_W3|*-Tc+ZPAa)
zIy@aYmj6GT;vBb{*u&BKf+SDta_1H>SFvofBq;d5F7jYX`=N~9ZgJpwXXpC*o2w{D
z8XLeH`sh!){66gEY0Y>lNzIoB)E!0s$7g|qcZr&CUScvMlldk%D=4q=JnT^^9(48?
z$@Yn%MiGx|Eu^tjDap)(4yA^AlrsvRH>X5eT{BG|;`pM%k|eY4kEYyMtwF_=#^yS`
u^P{WE>#Mr8&4~J8rsw4vaZ>oh>%HBoHk$mri~xR*?)?`7r1L9NkO2S$Mingp

diff --git a/paddle/framework/images/multigpu_allreduce.png b/paddle/framework/images/multigpu_allreduce.png
index 52d3eacd55834880d09f4efbee804b5c567a57fd..87a1b3e8f6dd4a713ec9df9f0037d1da04e9178a 100644
GIT binary patch
delta 81600
zcmZs@by!qU+b=u=0}MSwcMRPjUD6<3A|Tx;DM$%!K)So6rA28$MY=@<L<H#&K~V$*
zLF!xhJn#9=xz71R@VeYHd+)XG_{F_UD#VLk$88Ytg@|`AFIh(QjG8J^r?hz3op>mn
z2!CMWGs`mTqLk^Medei9ElsCl=P=23Z8n)1r9xv)&Wf17{*h0afAi-3%U$2iv9Xs+
z(Dci%V|g#1e;UiX=KIX*wLiysnJ-~ZF_V07yZ%TNODZYW>5n#MQ+I?V7bp4>^JlIz
z^ubUvqZ`x=E{=_{Ts7=^cz$gD%FnIpwddliA4_lMUsk5Q^!;jY7J0f6O-%o7=9#uG
zd+EYJkjIzNEIsj)Tc`$l_#}K*zO&NXrI@57m1lymkQeR!2I{)Jcn}AxB@5c-L&HCv
z=9CcW1sx$vHZ<6<wtlVlk<;IJlZ`B?$<>TwVS$kBtYP9qB0~%rCfwI?!i7;R$@lG@
zTNYkPZ4`Gc6lGN*;3{ZM6IeHvOqi<r1ibw^BO~mN=FV|}6~`UeS5Ny{p|(y-B-TsT
zl`U8*Eo-L>wUW2YG)&KU3wG@LSPiec_>`Ox294`$;DvR5`?CxQnK!XVZNUm|Z^0Hc
zplQokk2Jah;_qr55HY>5_DJj0>8GhivKI)rk|GL)`1fmq8m@u2!pm~VVZwcl8T>u0
znw#0w04`Nl7D6|2IoA<5-1Wds={BPO^wxZ79I2gY@a3S(aisD8{}uoJCShuL8daXG
z|EiJTQrGGlAN1^jO1HwUjaP`a!2fkI3Q?wNOZtgWP>c8MH*t^ZzU)pzU*eMm(j1&m
z0|lZMh%Nca(>B}qgQsZeu9Gz8p(<>KCa|9`m%5}?^0v+&`?8rc5v4BkzZn+DLJC*j
zZu9)&Yh0aLMgz(E@60wi5&A%PM!YG05Y&@>Z3!Lt5fxA0<%YLd*V|5PPWABR;ocDQ
zx(KU+LXEGP>3!-9e&gY14j<GOj$3Va)+8LwkH5?ab)Bv8Z^!$0Q|YkSvEsuk(^O%)
zy{ZqRBg0f-#Q0`zrvI~s|6Ks^E4&36%Vl9<R&IboW!&?pA;`tBOqJ!5bTkNL==;;4
zA`~qUsur14R6G>_mOsPWp8{g3L~1Vl=OQE$D<X`A<v+hxaNuR*mq+tns{9v+pKPFu
zldrWmp9TkMefwWmsk5=XP*W!@MhX2NpF{~`y?O9Q(6CH(Z*727W~(IhXwK!m{2#fE
zRF!1epY7`@*Du~lE)%t1bJIgX|9N}?ED&TD(7`)NdP&ca%kw`U{XLDmFan}(Es>`G
z$Nq&kpl4g}emeRw-*x&^wh8I08vvajQqS4m{aE|4R3)b2LCz7*e?lxmP0<hba+PPQ
zY{5GN#QXpLJv*u;+Bt$Z5}_Y$%MWAYtwRrQ?j)vrj_CjQqIYN%qoOA^J9EeX{aJoh
z*m9)d6Lm@+qkc*vR^D2hu8WSKgI`+BsdjcZ{_V^BzmK!=|G#csr))bAPx<V2o8YbI
z3()4f$3sc<G2O@`6s6U@5squlW3zYb-_)&{=Ay55iG>LnOTzNihsw=xd}8ELgkl}q
ztj%*#S8{P0?8jgd{U`8E!H3`6>r>>9%zi&7y?*9hSBl>J?;0d3!$_lej4P+>%nwHS
z>e-j)Ds>+>2;BMH;#Ol+v4RbqpaZOA*`7#|+yDDStT<K~_OS0U;g^#RIYF?md#Wo%
z4}RuJ`))2nmxaR$kGIeNEME@Nd+G3J-u+)Is*iZ#xfpL7+ohGoyLkVT6}5GsU75)i
z2pkQeUMu-H>e~M<{}d*S#P;v+ZIu|3a}Zpz(EV>>?z6)C7DgEdgalrGu%?Ep0fANn
z_XH!6G&1o0+%(}2cHgvlulB|r5>WE0Hm!bqdMm?As?qZKe_s&+Uql6df7WKv;u?OF
z-z82hfl9!#9c4VHYKTTq^O^2!R~64B-7D4jPsXs2gILpq{KrIw<o5a@@-3Kz`|QNC
z+mae4{cKyX23AT#L(bnuZwFp*Q<C$5JqV9LAD?_J7Ge7f-n_6DJWW6x8ArB|jdhdb
zNCa~>-Tx$wA_ejcT==0<SLJ<&$NY=&LRnvD6tM=yI!h?QRce9E*n$aj=F0{M=EMA8
zU*6{ELq9-_*5o=0q1_DqLpGj+^gAOAzkYG-@NK$gr#X#^^>MM7+f*{6Z0t?d{xZqs
zkPMDvo7OKe|C6UBcvy(%4|f`e)Ki#3ztvQg1$j<aPcl8&DaCVBj{474ipf_oq_NT=
z&pCzhlDZ+F#BziTIY0cwrU4~0$olicpD_ongyRy^p<0l{0@C(+U}Y+_b6%+zHO6Sc
z{##)tkl{{mvrD~;u0slWw8s`%0=G@s)*>nX69nM}D3CFF4JQvY*QieVwC@mNiKvFX
z5qDOl<TLG0+)9#M>h5i}>x-^-Oe05c1^@ZEX!bC^8sd&7S23gdSC3o}!qFU_YOQxZ
zKJWbd?b>^}g9M?@ZedeZH5t;p3C@2-h8i&e7rQ%CryNCa{rori*UX@U&9o<qV`^^{
zzqTdMQk|&jd?Up2QL_CyUZ^jBNb%P!f*ieI*BgnC$Kb6<c5wgO^vN!C`S_|SUKNC^
z`V)>sYl=r_8LLM2v04%S+bASf99%@|@yed}N;n?wRFhabKIkt_%Ic#y;m4R?Uvg~L
z{3JXV+VjQl9_=o7S-%NxdY`=OjvD{>rNVgd%rS0hP4AgXUHU{jpG=7u|5B6xhMzo0
zFwGS0;|h!b0kZ?Nt%MTB3m-_JZhp_10`osT&S8Yk_BJO3Gp|BM9=Wm=zi6+sG3`yz
zbwp%MQ4lB+O`I*5SJtz}^9=v$SyQP1S}2cFGcVOl$W#(w#K-!J39Fym7|FojYPkF9
z`D3rnJRO(Q-L9_A4FV_d0b5xta67KxFNPY7PLS-A%XP1}*#a@l4+9ybh#zSr+G!-x
zgo|8d&GRW5(@IyGCp;b&iD5`H0j33EGjc{nn(9{`!MBWWWZtC~dx=7<#dd!!hL)aa
zq&Or?JC7GgKf)um1|<sYICoi@lpfBiIcfqweE-W^0ANmK3@I=I;AL3-ZQVZ5^g-4S
zPrlr&HDL|i2N#^gf2_B$$&=65G1RWgYM`XTN)8jRn-RZz%MW%!gp>)J_v$*|dXmDK
zLn0G)`2ADBw`m@6|2qm+<)Pp~EzJI>yX|Pe0-Cjj-u#|6);|ZYeyh%Gua{7$0CcgV
zW&wt)mBmM<2Ko#pLbRL)bg-oN3a9Q*)RVAWxtVlriL@O*jArpId-oBCypXz8@{QZL
zGSU90Q#>!O<*RFqaCAeyLIlAUHDef7YN24q)0%#A-MJD?jfgNR<QXVx`U)rPdM>#v
zAVi!-GeMctCmfAL{!5b!QaA$cD08{;h!jO|W|LP4U3UFhdQg7BVOvnRgELAw*dSGy
zR?=+x%tAA^?4$QE1+kxdg~%z|8&Ov--B$eZ@=U`I)~Rlu5k@1b_8~ccmZ~C3jQ|Hu
zh@Ne6GX_cJBjc;b2@+&`qV!R0JU@dQ=t#oMaQAUMV!G)PCG4G(cV9W#@VsNfxe9n2
z8lm@^k@%8gV$oSVcY6ELbmGb|INzzhAB=92f`liT1op7cZK>9@l^~YYSnS6F*?d_`
z1C0T9;~EhSyN71Tm#kX`mKoZ|pYPU({#gvr&f={I6U!?jP|OxYz!T7}OC2XKJm!<-
z_BBnP8S6<8DCkLxvy#v&gh&;99<pL#aDzOF_>9F;4YSqlBf|J>L%$XVXl3)e@qJ1r
zS*fWiB4boMVpYfMp4$Meq-!((E;82emEZQxA>~3APd=4n8hHJIFy$MAcHJy$x@+}a
z7h3+;X4XJ;SoM?e_}l^>>K3h`s9M&)#CU=6{-7&KDPTeI9;ckQ|JrklR*1Kk&B-pY
zPmz+^_TC$`P*}B2^;{z_nW3RYfk^bg_6~q**H3qvPvQ$7{Cw&5RLz>(SZtpag+-*S
zh^WIQHq|VTCt+c?m6v$@C||2wjj8eD^L%nXLC(3TH+XAPo&^|;YhN_+)>AIMg1~X<
zuZeWxwR(jz-|*9TYMFeu4IxwmDO`^_iC)r~(KlV<Az%})19l~HbQ1;B%A4iLnBlCk
z+{MN*`=I3D=7IWKBNna8(f1%u$8gu|hOQELh{i(fGVItpTL&h_KY9)f(~53)q0Ruv
zV)CE08#FAU@5(Eplk#5q+<G_eFt;W`_lr9%h)_rSpR@fYG*HJcBYw;F)KCz&U7?_L
z-`|-{<5sit+o;gZE3-C8myux$BYr#c?DolmZDCdPIHRn;ZGUnLwW0D@m(LQM>nU61
z&i(H*hZnK=`5XRmsz&wEue%@kB-w$m_DZz7{ldLwiblfQzxQtH{Tpoq@N@;x+d>jl
z|K`d3Y^tq-?3M8wO0WaBD{{a+Trx<)x+V9i%98`jNMJst`C&9Qj>o)lClpRMFTn~F
zjMOc1Rdt7N(DO`T4D87=gm7h#q&5<Ih{Ka&>$=RBNH+V(_2t6l4bw1cvo39}lMxC0
zAXd_OYGIDs7!v@0DR?`=#V$VRUuWgAXuXpKd4B!M{P4KctkFKwfkr-9Y%N7Gx3x7*
z)xoUO24nwLVFDvCnw;VpFh>|i(!E;}OVTZ5A3Otcf5>Oj_hvwaS~~f=b*VeIh$rJV
z5+*%O%2>8C^cA+JDhMT1T|WUTnDB|I(kv|@UQxNymEwcaUkZeMWH+d`X3s69$K0Vf
zmvZ{~+;TfQm6rAoda}fX;)}x{!kG{ugjD?PTV)A)*U$Yr&127Ezp$5K7^Br)Q3#AG
zj1iWH!7oR#61%-t#|yg`G4FNB0fQ;RZm24m$R9nodOc;Ds!R>{gS9Q(BXl9Jix<YQ
znUXM(+6#VLbC`+JOc<l9hYS;*o<W5HOE-q~Ma3_^M+;mxvjmuu^Td8Cnd{V@|1F*P
z9h0Y~0vn=FZ`Btmga-fk-1@;Zxz?c4)nr(^D<bR=$r7tH`U`;h2A{9WDi8Tg>yAOS
ztv7Fu=2r81bC?}Xp{_<YLlrU?UZIh~lw;*Tmt*tr1Qb8@ElEa7lBs$Yhy@T;7MU=1
z#j(FUGO*^sw`{HP_9B7h>a{0oiuMC1YGV<9&2ZB>qn}AAV-k;wK^FE#5#D{9^Oq+W
zaFQpZNK#a1RQeRRTwsNhQi&oW_FmG|y@ht4Vf*S1&~H9f8$Ri-n!Y8qo_KAT;0!RC
zQeF*x1;^o3Kp~!TK`Kv~M54c}t2K=S5N${F^l`KbUtJ`P5{oCY1l#q~b4KZATg%&w
zu&1w{V6$wc7pba?dZc67%Ym986DX~U6%|?-?~%6W5PyOz5{ksq?WEyfic1Om^W-Ho
zK}!7#OXC+#z+j4jHxWz|K$0tro<6`L91M!({0on#%~18Qoz)&iC@7@Lg_ojgj~-7>
zrL(Ni!$aW0)eP4kcChiT1vMq!tT7r0P+7fh@Fw`a<6ph&vH?F;dp`nnvE~!G-0Uw;
z1D>4mj&dpsiXH%@l8gS4w`j`kI&51VOUCY85eNHgs<gBpC0Nwe<G)1fBl=$qbJ<#$
zx?I<SIvaF&`T!vR@H)Wj=#ItY#bUs6n%Z%mOOe~>mKxJYH%%f3$$+1hqJ^vXW=mD@
z{OH=5?`XPEr8ntM@{?|Q6B?#@Hp}A)ZlFPf(@cOdwX3?Eg6gyIx_`3HpfuvtxR~&s
zK?B3f-`)iIqD}-=31)7Vj1IE^2{HoA3H_z-d-aHe8sL%=A`q{R%{_SUC>UY>VqxoB
zuxvpqt@J$kKZ%7WII=%Z4!-(In7_RD1@r=a#l$DcGFo#5X~`yl<zlj2pZ-jc_iJ{3
z_iI%8;&}b$A1echJ)u<V`^+<u4AzgqM4#p&P=A-ti=m4kfChJ(0ZSNFcmN&k%z>>;
z$-%2g{s)##G2NwW7pJ=#IQszAl&wT*_*Vz#o7uB|-x^^h!$&`r)HB;N@4Jc_@u2+;
zLcw<1wf;DlqM*twa-&=_fF|XxxL*;B_o1a9EFPZH@p?8+nN7oz7GC`DFAu8(nv(Tw
zMz|82?iR*UnI4*p0@!1R89Fnh6;=0Z0U8m&`fGkKLQd{YPqyD+Gf#Lg(iba2hPgtP
z@>nqp_l7;%7S;W;$T+li0Cz=skMdMibM<Mb1U|!ztU1I~BZml7x%k&-o~j#2bseuI
zXvU7qeCNj{bTG3|i>)4h8<{2(p(uy;_Jh@Zje-cyAvRoNKNudPA1YdHZ?1p~w_g#@
zR!cE4u+tQNPhhGH)-zqr!;mw~eKtUC!;)-QtW~F*6nlEzC|19pqxP06t7dY8(s`Hc
z-g$8<71#jG85gVJ-t=uCd>QM@a~S?!lVy$cy^0Q{FmC*0qCz1?$vC`BBefhX%nPW<
zZ|cTw#{mi9gL>;X!p9$@nowJ_ouQ{}DRTFE@wseIK7x`|-}#q-Pk8;TQU#`-55$Xb
z_bWm}FGIIL&KQRK2*rVq5r`){7yiqc^Gp8Ef{HMUe}$fyuFmhnL|8n^aXIu44uXMT
zIAToyUUy7*9NHsY%w?iP>6qvU`u-_p_M;<(b6;pbT7yQZ1I6=?XNV$@N6-*sKqkL=
z6X@4cPxk7Yo}c?q1D;5v=5wR3J=!k`<Z$saRburnO&?Rhf)Anr<zU7ZtrJUuXi)~N
z&Vo2bJ8=tU$-NfBI)#~6hc?IZ06TA>=O#Hq(QoqzV6eEOp}#+2wvK)uZ{D&a`jXwh
zfu(Xz@7Z4Jw>=f<sb+mNbK9%ceNGgem%GB^{*OZ%uwaF(3tO-zKQ_*Ybj491v%_JS
zSXes!vwPihL5lxmk?ey|dZJw#Eh{gXgV~3en_JX~X*L!a*kP{AuAv`nZZGbmRop#~
zdGCy-Hs^OxetY^Ws$IOxgDTcy&4dMHNoq`?AwVovRZhhCB-!jP#k%Z-rjF%mQibSf
zwpqb)@qa+pT{%o&{nK+R`~vs2oCwN`pRad+PS<>hz#-IOrbx6KN@3QY@XTGy4FH8L
zKDzI>2<2a#@M@H>1zy-TH`5RRGD!q)Y)`*H?xHT^4C9=Uc&qNc_-Y*t5o)vjGoU{8
z8@^S)_FM-bIvFKJ{97BxL4YOj%wtb3vK9X<jbYNdRDGcW*Y0Scz2=-7i2I9>mGuoC
z5lxr~y4Kmg^&w~tQl#<K!$OHhHh_4oMJq0Mh5|OEYmZQ&)i7xZI&e-_kDelh=In{?
z{+z}C5)l@Sl#&TcW7EiQzYMq!qr>{1!;T-jO5J!oP3@Kz<)HidK3USc;_J(eA2qhC
z1Tx(Jj!-}PzVIYPK_>8*?OL^)_8peUXV!0Cg`OS_$F;9?H@MF=0d*%qO9JIyaaA3G
zu8VaW0KoWT&&{S0pI8<2rXdvP->Tm5sP_a22iHL!X!a_u+)1=64%xrKbu$Y6n=tLF
zGI9aV7#m`#iibKux$1Q@4Hmv<rTLK9M<P2WHJYW1L)O1D+vouP{63y@DMkkxnSXjH
z`*$^lnP$j-AbwA%hs3jEPaR9{)I;_1FTj@%SEHz-Nf@>I;~`LO{I@1nfAx`mDt1@l
zQN7?!cvR|8ML#!I0feRfmxh%Gj@wI3PH%x$?fo{(tw4-fb94z1iNb3o9O!~=m#aFO
z$#c++P3l-%a#LDXuvcK!PhIbB!T4|~D_ycHo@SpH1AF5e)?a6EJO-426y1^F)8K3+
zCA{uUW;gz*$)9@deyKQQv@*}h%yU2fFf~#n8s}X5g8E3a{j&CQ|E<t?x}Z-&Sw4+j
z20DQnvJPC0%CQdS@yYI0)(CnX+lQJ@q7f-GsWHk81^~lpG`VmzD2wDD5l0fj&ocK`
zl`YVwcm(qUa<yPJs?{|m>Kkq^TQwFcj_i$Hdv6)CH$c;8`qcDY0QvrZ^v^tWxAdRq
zuYZbKm>_(QZxr00F?STBVpQBL&4y%g*;tBPaY?N$rbMh>0)%13qw`rNn5v2V%|^O5
z604Rg=wS_jmXtobNVce6571zM(m23jpsM!JF7q$gP8K8$<1FU>v2f4U$EP`uX9QZ6
zTR3R?F!Z%R6U?^cE-<(tzX>j%B#Hkmh@6cHbCL-g0*Vd*tsg_6+zGfPP&~;{DhGB`
zNpfja+Y*?dNfidT=tMJW<D-@zDO+wa;68ZFAt*-tlcRr-Cr<5sSRqBQgPazpPObq^
zSVHErzW!UHSUxOn_wUm+Dj|38!dUz!fuiX0F290OgQid6_k@aOERrpDGmeNTU@55T
zSof|tq^_LEVv7d?(m*;H{D^xI8XCuyFYln`{T;X>4W|pf`o>XVK!h6*{#N<e)1t{S
z&IkJdNM|Xu*Ki@3oj3!!b)&XV)(>chqEP(*imfTqwrFHZ%Nh2pdNhNEYvHi<lQn1`
zp=7z6a~2mV4nzqbXloxXRER`V5<?mnsak|!6;pBGeD`6If$J1A<u@|;6Ztwh;35>*
zo=yRk@sxE8iE?=u!UyYvNxJV1BkU*$@RWtV#6^z|Ed%}T?(>(p+8Dn9->>6V9f2Y~
zifSh1_mJg6S2;=vm&LkWo;?2YYtii|>cs9rMqWI8Xa_SV18qR?Symtn@I!D(7`C#_
zyjQni6IAOPrP#7yH)0GM9^p0$G&wL_KKe5QI-i>2!w)At!UUge?C)S6U;ovOecw4I
zR3iTh`h@NcP|vEc^7cw-xkN-s{41$r$}1pwIc^j%_B=CA8(Mt|wA5QRq2h+@M!V4#
zn8Ak9UcMwk2rEL73Q_k7{`=>mEa7SC(wh7Jg714DiB*%{xyUdY&gdp$N{rir5SIPz
z4QdhH@8tHdr+TvGl-0NK|N8E}`Ew`!5V|-3C@=h|afnIw&UE!{uQ&IBXes~Ts0BO3
zLxj5bN_O@{+PVb46=+*oy8E<db5esjbl#)w{`M54_GLI;q)w});%d1xmAiu(Vd-?^
zI5#Z$j^jwee$^w2uIN3d9-@e?Fd{adg2gM88Y9qT7t5ua{~JIm>wv}k62b?hq%>dq
zrU5CvzwQc*FTmBM_xi~jxhxuNd&<-jV)t#o-%ywQ6~zSZ5=)5{*~oS+M;9y%6hj)?
z_v-$4>)`E*>DxWoCEgGJZdU>EKWf_(Ktlsra{DLuzKr%9Vg4(R;Q)Ee5zh7XHqf<c
zWr8LMvDV@?aX6lRcVF&g?4#;6%BNqDp<D));c)u{9=TS(NKL@UoSo0Dgib)S*^BEA
z^B+!SjbRF|4i+ufP=c<^EBlCUupnZ<>k+>@Gy85AAh>Jx{jv039q1pLtG!Wuwp0$Y
zlTS1p%I;6y;JUxN#3<x7=^Yu~+;W93a4;xx#_r?2^+`C;z8e%3_iI)V{VVQQ93SAW
zkkm|<<|}JuZck)Iu;<y_+0A<EMVQ5T>hRB?dFX}5mh!nAMaZI8=LzmhN<EiqJ0Obr
z|8CrQp=LrMH!)TA569voS#p3w@>)BSJI%oPD~H=;thJbJcUbqFHqbg93&S7r6JG%e
z4Fm&Ym=jpopa4HuNH2iSDbugRGud(bS4=Q)5iF&fbR&^rnhBz;4(if9&#EB3!oPsP
z<tsgB<`v1hSBCXtwz2QMZ}>;6&XAy|rq2up0farwf-Zzt+D5|ogv-lW|3x>J+6v$l
zN=ENJWhYDzJ$67uajc1xjEASjitw8YjZ3W)HTkaH9aOR(%M}5VF;TgXo9dNh8oqEP
zEPJ+dC~yF$QViO#AjZN77(MdkeslJx>yo4WKIrToN%dpSf6flx54np|9E*x+A*x3*
zxIWdHH=8#*R}}^G8m<jb`tOf$xl{`d{42poEG5jZ@AX||?}pY^W@>EhI<@4yE{wcR
zc!74orIX7t*w9&$F$%6l@m^UXc-E8fjYmat&|@xe`%&_4Zss<j_WQi~=cj*$PN~;F
zPTvmCL1o0My6=33AA#QS_xA^V=|+Z-Up?Ug?EjV;?2DuFN++<b{(y{?`Cg|(S;xOS
z`2;Ubi(1e!mp<UeU<2^|nrQkqqs!$uK%|5N;1*WEO3A2r<^3qJ0y_^XNi|taCSRay
z&XGTwRk^=RaLVS2Q_9zSr13u{2pql5W%J;NczU6@=Ym<mI7M=^-_^H8bYG&(x*p2Y
zMF>Y6yCw^<KELrmr`%}g_L~ge!G@*A^!%&0rM!Auco2A7SiT$?E_hi2?3<Jx_GilN
zqjx^mLN#0bQwk_byTfY^gNfZWw#7-bXm~HsdCWn%1WjL8<<4riY&!10?t(&eu!OZE
z=-;mmBs{2bH!oCAqT7RG!qTWa`)OtfwT!Mz8H@%%%U1x0i~xSZ0^TuXB4KCl=2O!&
z{yLkkhiQf-4G#Ct$+{!LNV?9UquU0Ho=Y9H85IC}<r|c^%6ZFt|J?dr36GJ&<aFvk
zEG{C<22rVBoE*-zcYLPcsvFvC<#6i|@J5$a%|yHAFt)ZUKS~mziU*MK;%528Iw8>a
zPa?DvMiFhCujPU!)-s09cbeY=DMezh<A6C#aw%}V_UZFl$e^@W|N40gK<L1FC>?K|
ze=Gz|l;ivh@?Q0@-#Lq~?kC*z{%)fAHk0R(v;gaYkyo6EJ1?gZ2JAup6%?O@yDF$f
zbzq@UZie9~w_A3AK>}3PdGz^i>l1#R2+&@n?o4whhTZ$_cqJNKzmSq1fBtflE}|+O
zst%lOumkw{-81nZt*Zn9#o7!Zd>{wVBiJ8G_vXRYnWwje>SCw_EH;diYtWrvH5l(_
z()7K`9lkPzwo<~jU^Hx#@U82HQN24ziqVI+e}0RGQHhif6v_s8E{xd$2{s6jAHSaq
zClxu?`{LX-p@)>C-+a~w3nBl{ZpS~ZynAqb9*t958kzrlghM21HllnN0PLH~t)B8!
zJVuUN6Q%9zp~tH+pgKu>V%NIzX90`^){^RCHD&PC`lE&;!J{xZPQVKQ3*3SMeUzaJ
z!3YbA2WFP;nYqE8*@n^rUD0<Ko)iXnqd<HJ@(e<UMeH@b73@<VTy`*>{yA0g)}=V~
z2&nsk8_px?9Jefe)IdMM5A0Vs1$@<gq9mMuh-ot8&cCeZR)MW*?k>Gqrv)|Nv4t)Q
zSc*4mGaZPuad92hF}zwhT3T06b{kYh#mkHH)(Qh!yXPb;fQH<|vNx$U*@i$1+jIw&
zF*7N`FY_AcvYFTX&~nE>GPr&)cCD@p{rj0x&Q4Jy*`4_p0<vzVOF9%li*_2NV0hQw
zO@)uwf=wdDV(i32{m=74#Ztl3CEVv<Bn5^{rkAShJ{Q>;;VMq_cLW}>22)TuNhCh;
zzehj5;*@dTkk37(<rETI;E{iIj~MC;$79MhQvzP72O2ex@rdEfYoNhx(~DlYe%TX!
z0BR6dZ@@PJt<JN3?S<dNen+52K6wWmWX3oeRg+9AzW>ew;33!>qOK8u3RM5*=rd6C
zLO{FU7FJh&^sk$Af*Kq~NG$`y;JYzA*lCPm^ry)v%%+C=$@N);dj%J1;`>-L<udzX
zNSE7v9PAGgsD;2S-##^|QJ%>KM(6t?`PUE6|E~FEBhxJZd2A)DSfPJ^*W9XK0)3%|
z)q_B@7uWyncn{?Ln~T5Sfo=o#glwz8yBFW@$`-&V2Ls2FGL_Y8D+LxLEoH|9N-LB)
z+69!0RL5@<(Dd4kQZRnvL(Qlt_q8k3GJz|f*5MnXZSeI?c9xAOr9}WXK7n>{4z?<T
zOP{j@yQsU`4xp0As47}EUXy?H8ub9t02?sRg&blLgAHuziQUYriPy%vX{gf7r~ZU?
zogb;lMnv)JZ9VwWdW(n^BFiR6vwRxjDh^r8v@Pvc0}zCaZ@87E1plgxQlPLIVf6ws
zVa@LqZfEq6i2~_1?lN28_R+!t^g<Sx!35H4v1~vRUq{~1ns5E*=Db#($pI#P(Y;Zg
z&xTYuN^MKQd?WV4oW0xbz+7{+mv%=TD66U&CPx3}fv(Cruy;ZB`3?ZSSG0cMtPstG
zHOD*LZBQb@zkCj4A{G<T48u|xO@u)@M{vFOVzbM87TMGnh;SAZ5P|lHJgsaMiEiOF
zJoxjIr8`?<OdZI}(Sgaoiy`RbOO6VOw5m$Bszv!~346xxr3bt6KraIvSGN=Wb{PW7
z%(tvt)+c?W*E{w{`5ehv)IJA&Eu_|Fk97N25{cAp!K^{<yCvYJ;U28S_yKVsnNgdU
z!O`(pD?^F)L#e3KsG}FQA^Ie%>tyrs&c}`Che%7mNj2#5yzMRLCt!{EL!(juT~Q{K
ziXSXr$AocMSPhS53c@ftYfI1-tjv(g|4uDh8+0$n3*H5?7(I}#L300+x(nQJ;4GlY
zbc+?_C-lli{yh}}PKoU2ZNQm(=Uw8TWCw(dghV2aX6h1sneB4_AD<7mCi}FVY8*tc
zuIzq5aT$tLFFb&o1(2SDNLO!u91x&DK0=)1^VO~rSC3Lw?Hv2*VcfPNKlaq%gV63*
z!H*?8z;qY==?)Gd6>1sC%O8Ktx49Ib4_}A+5v|D%cHGp6;mm-M>yX@zrQ>K}j`sKc
zM38QXctud6M!0o`ixh8vd@hn`XP#VptDeLORCqnl!G-{!XT1eLVe_#UbZ4jtRZazu
zOjL`fDAm6hIJNfpI`x>v00zPjz$Un{s^=v=9O!!wY+vnrjo!b+=>fw9Y;MS;Cvtd>
zt1TfX`@kkOl4GN4yv`XR{?pmjH179}za#8KR}%r?P60r*y76WIS~OBv{uhnDBf}5A
z)%WDu8y2WvQLuC<c#P&`fr7W3n+@^WLz&4F6l_3msGxP>=QYMvB$uD*8j%iWO@c)9
z5{W#K^6mg3$`xTm-|34T<<cCy@$I#9uI2nP!)5>`<a9m)#xv0MSNN?Qmnffz6p7Rd
zEVOzmGCjBMTgf7rDFQPUy1Bwi<Qpm&kXJ4l71~fvdPRQoh|dsXj*7+@!HzoqHTyTC
z?0DYxhsfgl%0-udzt{IKah#n1FR>Q?wC-EK;k|6XMu90ugQ*Pk`<0oFYit_Hsygao
z-#K;jS+D{@zcN-+Q;hbw1hQj;v=(V`<{&;?PMqw$L-L_{0Ssl019w0*NHwT&I?T!q
zL6xAO1U^|$E_9LIZ{J3G%X~5D>8iYj`F%SYHgy?C&edeqmMD}{SRO_bC_zLc%pz%s
z_PB>+7IQ^WMa1<CFLZ8T6p$=T_okge!T+FPk{Q{hnUD^f+~Vqa<}#KWev+sz7NtW#
zcX-?L<(FKCWTkuWUmQOoRcQbD>if|m6tJwbNF``Vvn=*P*D-s<L}Sz;0tRO!^VPAv
zIStxv?k_>08{)DGgTV(7ee=`{bPi_I^i_rablI*O9Yb4(1-|3>-j)Vs0Fugs5uea7
z4=v+C)-Y*ox>hiUG$S<{W&0W^+F{Fkl#F3E{qE1|yGRt}CdCRL9J3TafP85duA>@}
z&W)tK4H<G+u`$OREW-N=!tN7bfMX1}ii@XgMjx~g-b;CsUPXO$tcDAB`SK_$DOiAv
z68<RB&i9J8Mg9sW?AX(!+t^B0FWjYOjq#CD(JWN!OXG@~olkcR0ESM`3$~QI$uQ23
zKu6yO1-G|Lst;OgSvx~-tO50@pEvw^({|zU`(=Etg6{XO(q#HA{Kwb)XHKb~U~esV
zT`!&E_stRzW>L@f`E!r(5r0SOOubEtRb0+rqaiJX=D|CJBBkmAb|{U`Z&hQ81SRH8
z-BLuCP1$qH(8G_Fr~>-J5Y^(qHT($K70G^u$83@0J(JL_yDHH{tkK-ztg(?lzAM$Z
z(_$6M+z(z)l9WnrD&G7o`ZJPINfC~<V-Z`Kn8(F+^FpdsX2`^v?*++afzjn3QmoMN
z>VlgmISqtMZBmFfPk>h#%P<DL?UaU-K!cEV9;1;sw<9$uRtO!<7D%wD00<iAXaIir
ztZm7n3340(R^su;TQ+FiJurhgM3Qqonhcv;mGmqjH6ej5St^2s4w2C-Q43NloAHF8
zUn_=S_KApD|Fgg!t)FI*tE>0U_J%U>UI6|({vhC258m#=%M8i1TVz!<<9WlxXfBZn
z1W~Q!%Y6U@wW13FHUHL~Jb_|H<YuVy+e6r1?_Q;qV+R5bt}+EL2-zP2OkQceFomT1
zT-~<^CG{5PqSmtCg^aF;hj|U~I;niv5$R(|FX!RFP_y@YS*zqb+}LJkjJmMm2NK#u
zP8$nx9)=q#snIb1`>r-$-xu2ewMwk*8TT4PMCulRMs>I9neLS}kIV2ezmS5$8+`MD
zGpUd?++4{P#`PH}!=s~bBi?FnIMPWJ@|0dzMo5_qv5tz%W`Pkl{)?}qd#I<4cVBel
zympi*>@cPYmw9azNdl%0{Q~3=r>cRGb>)D-GMJM0#c(7QG6qk}CD~;N@16SYZ;YfU
z{N=`a(t>x*@~jAqlFT)`Kx~pO4af+C8Y2!8c6a%>^PvghOaMqG0B<NLb=+bcS$PFr
zm_D~0QM}f<qbS&Qwhv~Sa$U}ixDP{F!_0e(FV9$CJ_RHLOLmFPrdDXJh=odZ@b_AR
z5Lr)H4*J9A*FRr94cuGn<Q}OK)vi!|IgVi2wP<;vX7?SYdHW0knqU(1bp>)P+$^qd
zFsaAf6YvHnHEzXMIzlqAR7YQJ0-Uw`zZx%qy)UA<`a@-FH>3u?w;L<jjP@Dq!3sbJ
z4Y?>hDnrq?+S&EZN)C=xHT|K`&zD~rbYYb}n$Se!UX;^BVGAt8_rzCWG<joq&a%<Q
z-jD-Vg4t(t6#~6J0BG$8c?N?D2L!<kc~8I;QqRvPYzz~*D5P<SoP*}gqq$=|yOxxm
zJEuoKR)1GNNggfsH&tJfCu*A5+g#3Xu@%I!dgVt8%y;7iX=g!d<(HhuSpfn2+i`jk
zk9f;4M;OH1YI%HoZ+$K$kAO3JR<!~8VuMgepx2>}{8!IPLownFJ9kApACB05jTGTO
zQcV_dKiGWdD;#||?{+7&KdAbN`Im*;45T!wdUsdeOG(_BMplbx+hM)F-(lHcr*;sx
z5Ux`<%sAn1|32b@R6%<DTWvz4AvUZI@p6^kAiSX8H*Z-*VLO0Glv8+ILAQ1t%UOiH
z0i`+(X|rgqF;sNMnOhA1pt`18u9jDlplZP`6Q<Lb_QKMIF-+xGCXaE0?bXy$DpeZi
z<8H#_AR%QuGL~|uZ?AVdPlr^G$%~XT6K&}V;k{d6WY8f#_Xy2HL;zp<3wldqmojmT
zym$nYzj^GBvlo{`%uolIC`uNE2`mEK1&v(9Z}C06{`28!g@RT4RN3XP=-k5?DjX4M
z_gu9X0e8?(O3=!ByXR=OJ7ou|Y@^?zUt|zV?ibpQsF%if$G-Dyc&ByOnYCLZZ6F+a
zhq+z-ksQ8B<|3W-lwjGJ#H!$<%K6~r+wO2|L+P2?n|J$T(K|ekdW$PvQ=Q!k1Sg>K
zmjcamPAP9$=LAHyB9!c6lD07q#4{r&jr2GTjH(;Z5kwSk!6=B(xQ6YOwscTP94|rf
zsaz6hnVl0MV1|wcJG+D~#aGt-;RFsQROMUDK7YAn#I}-=)pYL|u?{=ga=&o5%bPVu
zJ3IU*J|E^ynrmgb_ga;`$AZYnA!WoL$HiV-)IgXGQBWiv9x0O!o0!@ovByWo!?d}u
zvxoDasub#RN7Wn!U9}9P*@99cC8)f`3Tv!{z>FvkOeO3ci+M^aRuyB^G@h}xEpaKf
zwHN>P)^=QkhDF=-9iwzEEO$D1-XhlXYqMBv5~IyZe|*Z-fc3|$Vdef1lGbog*jyr5
zkGQ#LHT5pzC5YPy*)m*V{9c;l>d&f0x*Y+L%dS}>DWoYx{cQ#My?1Raz_eqy8LMq^
z3sPPc@FHo&COS4hq?Z_2(5={?TbYc~>YoPV=#Me)3uSYJ2<PvPFCiF_q#`x~Y_IMt
zz7jXJ81`wnW>3ET-3PMCGE6EP+9Vuz#jLRGJgpz9xQD@VeVj(VO4$Vbr~dua=lEUw
z!)z(S4;s;I**1Z9G{GDrt&;c@CK%1Am~E)}rs5e%H*1!rdVl@uJGHt;hn~FWCz%Zw
zep?gaVN!KJ5HH?V@dVAN!M5ou+lgp-Yz*7lzl7VQ(h%_cNP{$(;-sYH@(jybH0350
z`<9tE3wOpCzAGFhYZ9A3Imt0CSRpw}u=p5f5XI3TBxz`hH~lCw+`!ghGsWYwb<X7p
ztA_J?(V^P@&g>5nm(C$_;ZK46AbzXf`o~kYvysBd4VxpQ@vx>Lds-*9an~koXEDad
z`qp~Ph#Mb;(cW%Ls$;PQB|Z!q8B3}yB8hZsx0HuYLdpBsEk8@(!e{L6u2D7GD3MMA
zL1b+bm1dIUFbU)W3EOU%oTtJ3zPT881I~B+W&470Gj-v2iu+NwvpLDB7e$$Lw(_Je
z$weYz9|@4Ro1Jwt!YgPYw-1~|3BR-qMyO|Wt~5M{6nXVJK0RtM7(*BmbK_CLn?hj*
zUh`{ZLjGB$g95!}4M4wQVE*>0{yhP6QN|?jWPSnh^q$cb!`+VE?%#~h<#+E$qCv)J
zDn!{K`iGg}3StFk8cB2mWIgtTg&i+6nR`!v3;P}qHgeiR7ExANsx@zeP<zxWriAIU
z>L`?oZnSL204OA$$|Uv+F}u)sze{lH%|c_!DqG&z7H)E_OWvVZOJM!eiSw|Eij#$7
zi6wmo81k@4r6O<>&u39mSV%<DG<yWhND(WsboJ!c>&puzZvGHD^Vr7oi$9Eg4;x{P
zKO1nRIiZQ;4e~HuVp3Iq)$Eum*X*At2ed>Y+(`^QhAzi8#425((HKm|tLbKbDp5}W
z@y~S#GcX}-^rc5{;$}|vR|t9(O!8a3eI|!q(Y0)GMaSqTPN!0ZUtjMPQ^9^@Mir;<
zNRv8|IwO^U!8GWZ)UJrPN%u^f+lR{BLuZKmz?9vHsSOuy4at_PB@Y+<!I2c95*)zp
zFN!`Zat{Ku@+tHUSMSh$X76!dIg*@-oh^QKbcvDB7W~uXZSQK^fV4OFRsyyQ7lq&L
zi*LC1*{Lh(6m(~}S20pIcww|FfXq#NH*gAb3$J?+x2JInCYn)FMFD-2jJEYZy0)k`
z6@Dl3V;qx+nqyx<O=V`48VuV}A%VSqWA>?aT!cC942^29x^v`fdkX?zKF6!+u(h~B
z`^S5v+T`U-uHg-WtnM2sr1Ft(tA)6Hc@L1J`%+006vWP=xP!2mH*V^8blR1W0nP#i
zYa0x4AK(&79Rji8M$n|Pbr_mpRQ`ol$F1Rp=Rly*7JP|Fk{qIO{I&tD`X+S5*%BAj
z#B#-$7Nnxu7X9htoiGNNI7MMYFfU~PIpZ@`lHjvRPJt?aE)F=Ro3HPL*qqmxaD;90
zHhPJ?h5-9BR^eEUqBwY@r_w1Okh+t$m69TgqZK?)&@1H7U8MyK)*zaYw66?#i}%%L
z7y&Kn$=A#k2VGl|%mY+hCy86Eryto3gqKvOjpO^cR0;ma(L}V$2DoDqH`X3`@+;f_
z-pxN&+F_o(XM0+Q-MVwf;HftY(UGoWO=i1V-e3+n^$LP1iKK^lmu23ttkC(JOvgR0
zo8+f<!^rv^sKmmQ3dQCUVSdYpfXxHzp^Hv#!TiXHl3!z>WV2L>=lhpw?eeiDNe(Pz
zm|ktQVOdApL@@jq<1>YyN@RLJfH%H{TfTuGDz9F76#w7^OqOT(8Z`FLxJ6#|F?1#S
zvxfeRzK(&3b-q*gqx$kck8t|=W*0a`pC8lpkw-#UaVpVC_+&N(>BlO{s+o9OstbMM
zT-;;&9n58B`XaAE`t$t_5`!T;oObn$Tl9tErIS<wb~GE_s&w{gWCp~HO;p?L+Z7D`
z=Z2up@4FckwbH;pRH!Y791x3W9ZFBJnTOG6MhJ&i5*J&Y)f#-ex6gV_gkC7r)feJn
zt-off7vfmBGUFL!Lfv1{+k~G!QDPfINlBakCm}H`D2m{j;<IQibps!}q=b%LpoS8C
z)1`1T%tjNbSZr*;P(2TpJDp2~+pb(N;<o83;S06%M4dcJw!`P6WJi67ZchBl0#|}a
z#slOS(oC(<oUfrNf*{vb*q^2lLWuhV%R}N05;-YozxnP+gnf?qo&Ammwco}to4M~?
zRnt$GCpvE*w4apJDmTP0i{2jmRpxQB#zZfu;e{K7NFDltoENP7GuO82v{<EItT<g-
z-q`<(bVw|6PwKqMt<ov@w_2^qcqHsZLF8DuCEhlft$=TJ++^8Hns|yIO8S7qI9a$%
zRo;1sq!AG4eG-3-wK6TA|LwSY_|=*XCxl#}5@K94`Bw~S`+`-&5U&b2Vrj@$M})IX
z&hB~(`<3-B0*#=C(%w44AlmSQac1;P<TFCJ=WGYgJ|~Edx9+5-SE!BO;B{N#1YBx*
zZ9*~EB*cU~YeA3=`X-e793jcRrd1szJbggvp5mCpCWTH5dgLMTekS+Gbv>0Um0$dQ
zWuRrb+O~*lWr7ZznzCSk!o7YwB6*8sy$=jUGBa8QH~9&0<oW>=+&JFtpJtV2>ezIu
z<Q0w0at}ks<e<{8f7BXYMd{tnDMMLAsHkMEa-kNv@zx45<lFLNiC%S}PS~c3_v>|x
z)Frc;Krfcg;yW%gNc2o~iXz6cs03^}8B%yXbVpf<_F0VP3^@Ha==$#E?k>KT9FL9*
zxc{1y6RQ^4tNmR)SA}%%h4ih4Kx^N3;`^(<L#`KRCj}hCVVl`o{z_X{QV8k|cdDde
zxtjT)q$1>ozD<D$0g)5&0?cT};FQ^+)|FK}IKCQ-)qH*!(4+&yY4f&QHK4uGPARki
zv@`7a*!R~j7qq|d-v=Z4L?qfL;FwkNTV`j``1=8V8VvW>b}_f26CFo?pXzQ!=3R3k
z<H`+{aF$id6Ty4S{S_)0|E0l3LD3g~+pYQ~5yZGgs}M~ltBA$QiZEf>L{f<{T448q
z!Sek$dqVnjS0VDpei~JN#BlYFxICJ)eHdEU#6q<}ZbqJ1vLsFgI|vo$<z;U-`ze^N
z&v`Z9ibsajBcOuhOQg5@nnAs_{JkGes~ie#$n@T+(n}`THk01mMw>E2-q*+6Q)T=g
zpvR1AEItXUc&Rv!&3-W9Ub}BV5?*1WdHKhj(<DdW_F#4(&^YsW`<n}LUm&hcF8qq$
zaYyA<wdQLkT!{DWP4y=svGb5meoTlX6r;ykO^`W?$1W|gxk|c)GUD&ekLyK{Kk-ah
zZ$@b<)3LR(9-9<BVCgEi6OArJ^YPC^LJ=Rg<I-&^r>BSJBiQf;%?soy2t62aw<+!u
z&|wcxoCQ`otCmxfSE#A>KXPX?i9p?mhKt+C7be+Vd*fQ)u}`3KW9FL<<!_!LGooGP
zYBgI5UqUu}gZ;x4iLIJHaXq(ofuq9qkokmpKZu+*nB+~F>y3mEFjfgpmJDP=%Ikix
zi80KL{kzG$C-h;C5ge)aIu&L$)dOWnA`@WrtUEQ$GdQ!(n!0JGq(VtoGh1JeR;R6|
z92mdkf6CgD=)gx5kJG+`F-H>hkENE+J0~2jz54!qohIelE%p&=xd@#IGv(^CRrUjs
zy{a63^R>O6?x!9L?Pz+hN~$AB*MKtcG+TT)lIs)cX8V@(0g-*-bx~CJPg;AH8*P#z
z(9d!l8|1G>F+MK3zk)i3!_nUp-XZy#|3s|(n)Xwt>8l)IHl(vb8Ap;dgk>^+_$+h(
z+e&(N_7f?J^kbqV&z2kC&(;#-6AV%!RfoS`)lI6|gB+?j5*C3D&TlZtHZqChrBA`4
znnh;|Z~qC}h)chyj`C(2>XD1cE60r&X}^n9sv<QxwFAc$6cb;AP1hdt9zQ@*OU-^2
z#$}aXD^KhZ5Uft!Jee3#puZL<0U<UXf`dh!fXRhe5T4EfCvlGs9A?4Kz!bb<6Vbb|
zmiC2`jctR(QBU_32Xx&Ek^3$72)Q4_vLKLnhGowZ=j|iY5tsG^eH*Lbyg-YvCuE$m
zJizf8n9_b_z8;kyIhKxM=AQ$v?rfje$68hfPW-skmbgeiFxDwIf4@%qJHMZLBklok
z#19z?`AyJ$v8o)i_FvC9n@r+83^^kY%s;vT)%Zqt{!?too2CXR6zS$umYjPPKlJ%=
zhjEnnB|Lac6lcN6L-<#C<r9Lx1Pz3p1Lb=@Q3{<FRBn@(MAUoK1J>$FVs~vERGj`*
zY2nvYlhZf4B4(rNXs<lf(Y|z>_JY5|PAJN>L!wqD`zH%@`0G4t)F#yi@LF#lf+wLL
z<{i_a96_u6ydz+5d>bTWSEAWaJLdcrePdxuMAJew-jjbGz-6a38YPK-JdQl^0UC)9
zntSIQsn5pY>DLVPq&f4x7CjJDCYSdOcPf>;y#36g6@zPV@k4bAi%A5@{owH1z~f>2
zwH#bPjFkA9-O{#TWZ~i#cq_unYm6<!48#!G$<!sCIexy^x+8}Y{1m8dKFMsP76!z%
z?p5Geu;ikjF$pJyPY#tGo61k#F@2Nd?Bz}tQjFp8aj5Y7ztv(QY#Q-nU6x<x^7DrZ
zXQ#>)pPN3*kN(+8WDh)oU+X}9Dxg!;*^ZJ-JEXR1S(`Ol+6xyRex!NmRThA$fSd!7
zH0%NxeJpdYzDJmyY*r?!{p|sfXb!9d^k9Rz)BvlbK6jHC(~*FmwyFdeUEwRkbFb+8
zsaJe-C0LrYE8UP|TfIA&Nb5M&v!xKT%>~MuN_KP0FczaCN%~NWX8v<*?_A8PjQaQf
zEuFVdHlw6w?H_0+`@h>JDv*OV_^TCD%kh?)@^PiRQvbLC72x6Pgw~eWoAehqHos7d
zA0D#q1I#h@?%|kF>&~xVHr?q+2ZFlz36H-XthIl^(J*zrlpzD)j0fz@qA6){j^<zb
zx`-|hETmsUb8t%B_<lo$DyXDPwbrPfl@ya_b!VyH1go4~8KIL#We3GXh!Z3WqwTX?
zKB0=6m2}HQ_?bn|lkA#p2eL`@pI;do+KlhK>?$IPacVU8eoF@PcdYBQW6WkG+R_5l
zQ3v84pT%PSwKGNMgYE?u44m7kCOeq0v2Y89jI|3V3F~pnT50LU=eu8`%hR(L<xs;j
zjjP6`GT26R-vk01?S*$+LZ@%0&V2pp-5I_=E9+ig_CM-sW})}(61D=~-H%n`)S6yB
z7?%}d|BU68-EWOzsO6?UC9x+fO>~lWHk9}t%2LO9Z514xWp8S|nEwvWF~krdf%!lb
zQbhBK#rSV5Te9s=;zYkP`*(gs%wJAA$ygtjT%KAe6S}KRBPx9`;!-|h(2JFL96#(@
zCgM)jGB>qq^eia5fdBXFHtUc*wcd8{bHEXKpHQGP>;rIH&AYFR>-f>s!mMS<v<jm&
zDyMwd%NwARt>E?xTj6#0D|(J!TmlDRq=%LJLx+(~au&hnQKQU2gFuho%%>Ns=k0;p
zX}80I&1g;ty83FFG~Ce54D+;NT-LT9t4LnqY)IKj^`<T+GPne4vI-0i8#8Ci8!3x2
zO3WnDiQ6kb#%#O-Lk(%NI*%pZv89-rI*2;_Ony?K)Jx5u$dY@YPm}Shg#sejh$@tX
zAL@ujzZ<7hZvs}H!I4bmf?C+r6AjXPBX~?*mXZiz+RcQ@uFg?5&kplQK@358j^w0=
zak-ixMGdvIW&)UE8kdeWp3}LFk%8e+*$*|)SfCIY{RU>5<|R{_vR1}~pyc@7Eq4Ik
zM}9rAIzxq;rf>bLZ@e~;aD`rewN~o(?;rD$+~H!UJHWakGlxn40j_3Zsi{fGmk&bB
zhN^0aEY(b^OdCIi-+@r}I?ApTIJukW=*M?NlSH;1f4UXf3sL$aIVm}uSC!tx)mUl9
zc_O!j1bW+}vnE`bTl$q^(6b|CqVL6{In1kk1{>lEDy@OW1AVIhCe(@!)siytmcd}H
zJLOnalHLqAj;PH}U79t*S-KZVycQ%6hPQUS?8$;K0$Nz?doCnb-I<00#pYA@!J17(
zqgT{<EVm^fiH|)(AB;UFvB}~^6bFCfDVqfU>4a>Ii+G45a3E*-0-@$LZUIpl7;Ccd
z*e$bR-+idGj)jU0caTN}_{$$C9u$~7hnE9m@ZDocu1@Ue4>goxLQ2DGmN0Uj_snJ+
zMTisdA0Aj;=+=0_n?Na=4iJ!6j|jV%Dg#rm=Gk6oe_P)B?YffDx?i{(5vAB=W#G_{
zbmPSD?+T%RQ(?mjon%8eRnNd_V|Fk@`K~kq+Cm_rbWS-doeav=-0XY^&b9u5Egf<%
zF!}ZdO!Lhk8KojvwX8b=8{-^S3!P^-I^10~_jTOU8Ip$m@_uq=4e43@1|wyRfcAx4
zriTe{aW&Q$eCW~i6stFDECUDBqzYTTmX}-^7>)`R@3yCxy%{4;buI3Fj1Nu(EpV4{
zsD$oH$zf)lA3W>)UJ|^)t?lheE-(2=nZ@*ntDr#e;m2FR8fJf+xM&;m#b>fiZCt0h
zkLE>pn=_c*!BsQiPi3saqC+cuc!W#bi<n(!bH3^Dd>?C{E??~h^_{&v4U5cUPG<<5
z5H$a_$0q>JHBt#Dy%q<v?=6WC6@p4fwdl;X8eW$MaWrjOlhR9!Dd@sFQ`s&50teGd
zqS>-v+ahw5gn8$T*tZE*yGh009b1pQX;LKWqy_Y{M#rA#vsfwSC4)6Z=d@_Yc7sPD
zG;dJvGI#Ojw}k0FHS0RF%}E-Hf4@-_P-_cU4|x-}M4yvy!CLaNWebWR6cn~N%6F@?
zBbccZiGJ7bY%jgp2M(ZV@;9!(3fNsz{qBIb56&XY^%!EP0kW4f#g5~cvFrl59UDKG
z%=q^4Rus=w`y9z>(e#6vr;6=S71bW$GtwYu@DO&K$+6V=KAny48;0!|4ltO5Ksc-Q
zR2i#E^wo)6Ysh=DizKl-nObIKy-zUdW%4UB+6Hl~6)>5e+VChVXZt4(Ug39C36s9F
z%DVa&82{d95$o3OOA;zug#SI55|aoays$zISj)<H+;rU+`{Uro9rm-VpY4bq4_kb`
zd$a&%N^~E*{jIFbZab~h*#Oj&em`rYZLSNp)2Gw-AXzP8h?AO9&~3Nt@?khbTnjg$
zYG=9H72zrOX+s*0#_WK@%)-h0SooZ<pTJ7CuoeIRcsdKWsJid%58W^`bi>fyjpWcB
zN=kQ^grGyi&@J5}Ao|cDh;)OLfPl1wfRv&b_-?+}?|T0L&~w;l@4fDI-=C$*XXmc0
z>!yl+uYV2JH##iyncXw=eC0j2oK4)n%EG!k7W;9RK$lB{NCE#hM_Hz_ybFRdhN5>9
z+d$^e`JmW`{?Pznk{Yr`u{Rh1o7bGVOTD|p4P(f9Q^03!zdjyc(C3%Pj6)W}!tE=g
zd3l;<Cp@RH4I5MFrAP-Eh-jS1QswqQBh<JONkFs$qY`Cp1U7M&umiv%F|5#i6fK~^
z(hHnoB=nMZjD_D)uR#MnR;Ba%l2r_cQuY#l)9ACbzqOpGZ?W3A+Kkbwa6Ohqn@W#j
z9?MQY`-o5xAzsjzF?GV;5VDB6gH8M@_wvUtjyoZ_e9QTX9xj5tu+r}O-<0tU2+6X&
z8!?d!pQhRP-dn%BH`AU`K6aO@a#m_u<}kq5a+fYD_j(EPV9vxf&I#3mIYPF*@~020
zh?-QCP+|XomYOLs<)%Ilf4}ds)fWY_-3H;y-v)q%%}!4=^nB|irLmsvC^#=@^10d`
zzdUfJ(_Wq=(gbw?(+d8i@S;IwK^)C{{&oJQ`>e-Q?VIcy+VzgV4D=1h9?!{C)v`|~
z9i1rs1Gbu<PJ%%AmwH6BGCqJQr~kFmWgKNZ+F*6<??iR$QTyS6W$bqn=HI|76d-Sc
z6}MvpEvWdi#hq5J?7y7Ut(ukQD^U#WZAkxRk2k9QgE!Dgw?A_}z<?`J!h0CU(2bu0
zocba-Pt@jz!FhwcHm;lu+z~kcpBN)@DfquHpC+|?n3Oq6ROD!k#f9MlWrNQIEYk(F
zeq-N_KbFffNcxM62nPcrmT02aS#>Y))v#Or`||R?X2mRSfmN~a`&DO^hkZlw4Gijm
znMip$vxX)6X#rQWCxEiT(o!ZLCQ7{`w;nJlaE(sdqTZB^c1BFKtD0hKa9h$g0Q36@
zqaL3S{#gP~c+`YeODHOVF+W<OWsJ(>R4LNy*eSO4zUdH6Kiz|`5&fq$|M6Rw?{|S5
znKvb6zx0~Vt+V*JdNB^(u8;lCc#0|M=C|Z2M65S~K**ZT??G>564Y_ZLKx@;N=?Dm
z`WJvv?8Mk1k355cuh+@Fugy?uAi5II7Nc5iB^l>KIp?V^s!*I@$%~Q_LzGBPUA6lC
zC28mMJR+qb@v#!W{h8mPExGR?2t`F5j%7@k6`v6jnh%)00|NRAu`RF(g@1KF$VXt8
z+zXc+;mNDJghttuy8NiXPuDz=Awa+Q_xp2+%8EUCBxBhRjp`Z}C<P`%PZI}ey~Mq>
z?PDrD1My>29O;zcK5wv}m+4Ql_v8D|>cL-*g3ou+>?yPOO9d<`z=8r#mPwX+2thDM
z+Yu-oMxvUY=*J0acRtRDKvxx5IgH}$+XZ`Gz$jY1K3)QY*$)MqU~Q@`=3^hzVCm7v
zqXRjG#cyBe6^-z%zIo*NN3AJtgbr#FY;yQMRUd`zlpnP8`eD(zR99R1feEl2NsZh{
zemvmX)~!+Z@@sso!`rdNr3;&_P;%5KlBT3;Rf!D6==^V$Zhs_>4%eRC<rDn3pdlr+
zbll-{i<lGeK;qEBZihuz)0C~7od}fKPq#RTs1pabzJWMR$F1HJW^^-90u6E7qSG=`
zP>w;s{!b6q><Z345w>)4Mr2?2MyEepomT_#NKPW<JYwm3DC=Z-Njo|5TZ)-Ylf!WH
zmG92O{5Rj`b<k_63l^Y48++-Dpzd|s=zSuWrg*;A;%Z<^MrTy;vxeVbj4`ErNhs4V
zVUoD!<SFywJB5gg<t9gszgm=p-YQ$Dw?8%pztbqfit@ET<KPk$IdrsesrMt5*kqM>
ze@o^<OVeqfz&c<Jj(^68eobJoIJtWs=~%EmP~@h6_jRCxF0R8FBk78Bzem5tR0vrr
zVg^m##1vmio<}yo3Vp1m)`Nq@;3T6-@J#oOZJS3#|AZcWWFs88&o++2KT>Phba=C?
zviO%=Um9bRnd@DomCq;+#zE#U*t>Po;H=KRuvv%6I>7n(NbyF_{bcV&DgMx2HQ`UM
zw#*lYGK5iv!-kfqVhHy&?+Q-#ZaQ`waIb~1%&M*7FNl(2;f>A6x&1AoISr}|nZrOT
zVE4l#B|g)-7Tf!u(Bq*5MMtveXz-%#=8_xRK+~5MGdTky@o<+#YeVYn!s#@J%Kzyu
zHYwCY$!wFT0hW&LRQuAMR;@uC^wAB@DwMTB8I*)Unw=LKO8Cq8BTg4Q`0KiRQyI=h
zzFo#x4zE#euR|Jy>wYxF2nO~_tCb~o2&jmC7~<iQGGy6Q(<n^7RS=7%HK_v#j#2w9
z!xk~WrFUSxKmXr)hY%=n_JV5SuDI!wZ1YYUwg>m0)vSYOjlgIp$N8HN!5E${<NTtP
zo7A4|LZ(>MIcambZ|K+N)87ICBJ+I`?(;*)UvFnqYzan8%wIpQ^D->Kz`}*Tj9$^j
zojXZJnVPdbQUPw;C-*pVaiSui=5HUj4Zq1iI@yuCkn4E8QO8m7N(Kg&k3*)MHg?&>
z{C_<ABb({B!%NU6lM}!m9U2W;`OEo5*r^h0)}A7|5~Dzx2kn3m7=S03hA1mjd&|WZ
zs*u|~0l}p`aFD@qhS|2Lw3An&GKVVezeGJ?Vij(sn@6m~TBsiV(3iaCLEcj`Z3Tz-
zW7I@vA){;%jZ8qbzmqHLiVP7r83c70iF~#%CNhOceJgQdoH2;ZS>WbxfhC;k+?a)^
zk)Y+8_MvPF3WR)sJE!y0rbI}hmO<2P|IM<1__qOjB*w+^!;q4y&zz>t#!u5xz;Lp?
z`7B>NBan0k2bf|B@`PQ-PM(s-sPT6!0ojq>=dr@Dn?salsKvc;qC8$FzP&`^_UT$n
zM{h;Y+jWHR<CKhoJ!=o<LG)aWUeb$2W*x}g2QW<4m;LV-)1E^94-LRV+9d4B{dZ$r
z{d2S9C@H)z3^g=Vh^nq1{{918>%=pEwu(FU&fI?Z<>xPqUnE>ag$?|{(Q1%~P?#$b
zB`<|ni@Q}g@W_3X<oGCWJVu?@m*>t*&<>TfU+VtFx3V$lqZc)c!R4baEK-Z@a-GBS
zL~}2#+AOwQpP%5*FW`aLX!oYCB3O=+dflm<rk<4hK9-~sWs<WY_@eV^BbF|>J2pAs
z1bpaYq!)?cg!9UgCvPIDKhz3>=p*GN>vx?1T(9Jvj2L_fOxA0o=VMxAP&M~(p=h;5
z=E}*Ro2wfrxvQ*-9t0~vH{1LK9qat2-Ow`DkCk%{Pis8tWnfK3TTL<KJ5XJ~GRTu)
zc<vmFt%N#xVfNa(;QBQ>?#^m&^y34AC)*I--M-3Jt^6j943nWf7e}}MX_Qy|9Z$iE
z5YPTGkkvz8=$GUOj0t;2MX;X!+55xXe)(~HEth8J94JI;6Md3Aa}4m0wkw%*E3|C(
z+t|+iZq37H#lBMR44K{AU!n<9KV{0b(I8PLiyi>jl860BnZ?Sy{ihx=VQujm=+4i~
z9^Gu+<@nlL>J<tj5o$?(Py0<Dyl*;JqZADprY&Y$#1{6e?qUBPOP$D{L#5RqErjK(
z&A5l8@hQJ%_v+#iKhSl3`}n(G_vDTLC#F3;cKlW!(>P@^uMuS*dfcG-8e8nyE7YCU
z^J&?{D}}6C0&-42!yfRdyJOs7*5Rf}?<9i%fN%$HU?}1Dh7iXihSl_Q*iS;Mn}|gt
zja(iRP9w`$2|{CBjrFbbwAn9!Hbh`$iL|}@A#|Pqjb2KL>S~7F%_CqE!lD@21#F;T
zF4tV%`}w?Ir7yes6JQkrs7k5%HMtKNRIG!;w9`*`Yl*0e0VZO2M@oddJisTxp8~v!
zL-(I7JMS9t#I>S{@r~Ru59ti<%gH~?2xDb1lc~xYF>2zB^d-5#etYg&(Rv7|7Baw%
z|5n)dF=Mls!&f(1>W@dR;yQW!wB(K=)oSSafs?UARLC72fB<!gS0w!MUm|`S6$^Aj
zOF}@J!6f|pL`zd5Z>Py|G(eY8%!Qa%X5j&8RfeMsf5KeAo_9U1wq(`o91Ik~iE!CE
zkOaa0!%CGPUJ+E9B=pdDcvKZm75b2!7%h-DoPw#Z1*~I-z*YYTI59d`*{1sMBKaW<
zu$yiF;0YAZ3x>ys6*JP9{O<Qv;Wsj8<-@)XD`M~ZuGR1iS&%OG?Q)5J^VQ{YLR>5;
zr|GI%7YA^N9<Ex2QB`uXj`7^~;JbnTLyYN|yRfAmWU7+%+d!fBJVK!WU-~0KnEl2#
z*-*1e%2?+G+Hn;{UTT^KAX?_hF0DJrY0m4(t)T1|&4UNqhd*HRfgP}W6RB5DCAtXu
zXZ|3o)P2F~P_+Pbm|L~ll2EA=&*b;D9`W-cK_S7h`LBg-Vl7pdw&Ht~Er;SMlY?+n
z(cYOj^9XahQD$jtU`Rv(${ZEHd8G6iC?C~`W1;=0asER(m=GS|qKPF_r8rDsLlsax
z1t!FdYBk~Gsry7DGs7k?iSTTk*?|D>RU7hkoi<)c2JQ7k9D{~wwMqI9sEF->#{TPA
zi0Yi$HUcKmm%*yq?)7mVXo{}lRUg1N5h0zme}A>CT`T!6_|PzRo@y}Z-58gt0g2NK
z4y{fh$P;>HscJbSW39K?gc_seqQLY_!ioRIqNR;r%>~4AGWz>c53FNt*1=}ZbSg85
zi(KuX$5>LuS(`-?6HyEPS#eB<B}REqmn+A>eR(?D%`JyUa^705&~VbdVY5);ifVz+
zBUGqm{IxQv;@F)v9Y|bY!FO~ht<NKW_fqj=h=x=_;a=sB2Sv5tLyvT&u#v|vp98<_
zXRF<H2L{G91a78O9R6ebNP-qa=Zhj5T({m;Q+|PH$vKRlbS&?3o1mtR?Z;IFkru!X
z!HT{C#m${YJ_C{#peO>RUWt^ijYDayn*@fB>BLF)x}-a-UGgqDG!tkT@+c895vy7X
zr_i87wPS_En&RKL9MVtx)Iy#=s0In_uV!TRmD~S?Jr!~s>#zI{04n@x1o(Wis}gs5
zV$lhR1dCjobz#1WWQzAqp}_}3P@^fi|HST9#_TlKcY1z5;C3c<czfnwl}Dz9%wa)A
z?+)#)RL|1wur9*)v<I0=(+la`VN(b{)2b=dBwiC{8HuV7rt}|stDKea9jR6zB`GCz
z)9$d=i!>}E!4`AR(zseJ-!37?1Q_xJo|2P`r0-b<K{ET8Y+m5+5Xdmw{MMgD7j1y2
z<wrA_c*dn_-JV{F?r&-u6Y;NtDmefslEneNF`!u=c#NzCuWi)k0|)0qoHuG7(JDwd
zMav)#UWz3srek>u)R&_J5YnRMNFTCv>R>y-9CRh$58n2ACw!&F?<A!jy4;AhToSs2
z4o>{Hg=u50NokEd9`>H@<Dua>LNH?wNY}Z%H~<N?FKZoV0M>|X$EdWfikP%fT<NAN
zGs-35sMo1hBWCy~E6IbJ=|~aa-;`!3tc9M`KKD+oaC%0(?0!aPN;A^OisMq2y+@JZ
z=GFsjbma9}upqUkEDtJFa3nQT9BXkM()b6#MtRxAB^x=Gz+#SRD8SX*(=Yk8y(XO#
z@+t<0NUjgD>sS#8BmIII)mvH!F{RjRxf-Ye8^rLY*v0L4QPlWl-C$%T_4ErIq4$W7
zkJvB`nTsf@RT48+^19Mn=uMz_6nwmLI~?SGtko>2h?s?uB*$tR4UM>&ww=(|5j>P$
zr0XBKshiYTQnEd$g#DX>F3Y#WQWfcpw}9m5RLA^NV0iOe-_iT|^5%*l1wWd^Cd%Li
zEV+v0pRbMJ>L`j*(-B&2QX?t`Ug@=6io6Ls1r8RIHjn$(l<VrVAAo6g@~eShF&V?}
zY$Tb)$v};MNm3G@JeJ7Ge<fWW{VN<hb?@7npZ7Ui1VIET?5(#7`vR!dC$+piDnYGg
zH@X8g?(}i`rBxU?R0i|F%m?T_?Z+<$e(`L)vuFKnfa*ROr=WS=vLy?{V6b8MR6-An
zR0;Czw2SSa4-|1)W{Rh+Uh8cpQed(#R%#wfijf$B*)D4CbPe$H^Tn9SCJA6BJ;N6*
z7qDo?mVtG1Uj{lV-1FP&1Ai|A+S`pswiB9!^=!fJ9#kY$>ZE_1R#mp6I5Jfg%f!^F
zU|G>;sKNy{UhKzQsb+={^$(<BrUQO7TuJf|RKUr-+7l+`@DbEG?4{gxdH(ETQV&~X
z{KF1@0Xw#X1&dmt70nh7Jsg+RQ9OBb2(mr#if8yv7~YfA-6$K-H=NeDAqPjoDyskj
zVlrQU{A9Le(aE3lB|lNORYKzOR|#q(cNC`i8bN-MrdEw7k$3Wdbd^hVdH`|^rVfd-
zkib0Q$B%tXv>S%gv?GmU&w5t)oJBw-WR!_6hD{^xC)m}u*Xr!dQs?K>Fh1|XMvC(9
zJzyXbY?-6gz%_D~H_N}@B7^_^QTE7p4M^E&W7Uf;md36lV4VgAU2s!9q;RlD4Y@NH
zzyho~ENiKSU#`56Fb#&&b3LmpWX}glKGhDiF00W*<X+;HVyPQpuczNM2Bip|3Lg>N
zIesrB0+7SfKm!dV$3pKreww;NRpg<`Yh0!aDd7JwKc-rJC1#FhzZUkfkEDKW(gjY|
zNWafZIXi!24j}>boM#69D6fnB{Ppl6cTcch?t*+bL)Fh`8&rya0}LOf^fb7uGbVI{
z?G=7Z*8s2?vVi+DWP_N)_UpX`8?O8aPP?9Etn#y+TEdi&SeVjX>4y;Knvt1&{Sm^G
zK{bb?zLk~}B><cIZ$*pEvrn2q^$E*oTZSD921^QGx&c&?fYRtBP3Z4ho0g%Z<k@6;
zmxs@vY+38XL`Nn=f=O&6F5>9y^}ub@JgP~{Kb(u`AQA(sQNTExfke`tD(@kt*BgAT
z!wf2fWK|BAcJdfgGT*&z>YMrJPuN#fvN{m1>=7~1@7)@6UbItf^sy#5|L!l=_@R{d
zUMriE<INz-v9a$~lIMnVhy`6F;rW2KBq~RG9?`uyq!c;v0)`8yF*lIfwHD$-kf5V9
zd>+FIkyQNm2e6C#d{-Vq=}Eb!`@*+V(B+XIihRF<H=+Qt@YRf_?1d0(kqmnd7uU&w
zPO#lbn>B1xpKbCl@cpy$O2=2qM#<r{_qnX1+O+!JGM)%9`StY8*pRqq(&6xjNvY05
zsS=eMDa{$s;nZG5t$AQwQ5O3<8M%nx9|?@JpepOT$;HKqV#m(VL3C{Z`k`?cjiT#J
zG5QOSmTM>w>;q+o_qkJ;vR7mE@<~*`7B3pHD~?B^Z)tPWqc_UJ_G#5q%g4}9SW(}1
zldDEz`e3r1@w{j!nvf@e4GVoP*z=0>luxsE4;I|s$-o8bv3n{o8;LoL*aUBOfvTss
zARcTUK>3ZSixYw&Wr>%*?Mng4GN7g*%&_7~iQ3%-m#4D;V-_*WWDg+k9^s5<s(OQW
zQmw{*c2%3&qG!pz@$ROFiw>nQS<2IM&jRsIS_H<Gb)W1a<7K3pN~{Qxf?H^1T_)9b
ze2vRl=~y_;GL6<Yls<IwqvP=0LsM}6zU;#n^pk(C3_^%f1;2E!Ga)!d!t`3pgtEhs
zY)1B)Bf?)vJwz)`+$AVBE=~hXVKz0K=?#|m!6~7|<#&0=_TPFbs0fz-k!7%TAe`sO
zU#efBcV%9(HF@8hma=GNYY(2ym}F@b5gUJrDz9i8`**7?{!m|NncPqWQ=31yW6`)s
z6w12FBYxaI3b-Fy#xU~1N$JDTh5w%!{GC3kSv~3Kl{>}*4V7{XIP_KU<aA9lsNzKY
za7L0;r{P*CuuHV5wmGcBVXDyud+1vpDmF+_+RowwCYgjlx{lm#WUn3h6FE|%{2i`5
z`p46)K`Kxh?ROf@<C7y)iG686tYE?*_^<E~GBEhy$z%%W5~8-?L&&RM4dNTV`}Nhm
zn;DOhh2+`kmw->ylK^W4pR3mWP?i4bx;j;CLsb5EH?=%u507G|c>-&YrgIq!X~S9K
zzl`Nhe}+5$-q368;Co-o-Yves6zb0)#IV?!uJ?t0AdK&f`e9|)&{^BlDY9sAd5K>f
z;%*4_gZ4%G0Q9xLQVdqtr+GxYD5(mb?$_~m@dA{a59D^bhxZt*dn07Z0!~U^p~%^6
zQTl3kU>|Hvh^T+@eMg6<(qI*X9tFG7w@UANa1%;2s(h}jd^Qxys%MqWsV9H!T-2H<
zZ56r`w;h;I=(Yz_VjWN8hSaUR;VABF3J5~QKki>HY!J~*v<@S=(U<`_EI^!p)9YRV
zSK94tLF5y5Q{$ZO{-sxU%f)mYM)em4_UJN*;#^pR-Gmat_Yfb`d<E8*JYD<Au#aW?
zDFEr3CVzBlQW1ISG74u+!FJ(P8vOF%C9GN6Y3ti`&AOHh9jm=eh@P%$FekEntP6*Y
z$?3`j0Kla0WK#{@z4R>xym2;B|D_J;{Mz#iL|Cg{_a@k)E7lStlu`IakQ8-)Cy)g$
z9ZmDfrmDc&mPda`*SNi^qYA?|jBvR4(Zn{TR}<A2%($!q|Lal@wF|4yi5g+7h@q1b
zf3w!<U&UB}qdAcX3#Y+)1~v+s`X4*-dime;6GZ!H`Sga#F)+8UQc(aze(J)EYrl|I
z`>#NqP+5{S_SLczs+m&?h=mw$&j*MYxedMGm)T@5E3!lb=I;9G)dG{^%~UyJ;&;$Y
zkObX1=Tuj)xx*FUM{`8Bk^cS#?#3c3E{{A@7kwF2K#6$-k0*tQ83*oO0RBS$`|R&F
zu?sH@xI7X<yo<p$xbOL~**RCHs_Q=uc4|;wWTJvzHJH#K5%C1M(M-UtrQNOp!jAeU
z$AFez-UxbE=s4Dg3$QmqXHTnHGhwDJd$~mK1$AGV`>%)%co}wT$wezg;rqq5{H`X%
zG<l&9a;jlbhWR$Eq^jfA!xd<?`A5rY%W)25MOC+df78oG>+09kRt$x+(&Na@>*5Nc
zIDIRrbq8Lu$wr<xrdB&wkF!5$2a^U?Kzy;1KohxwW=v_zMDM?M;wcff$80Kt<#35Q
zf0U@3t<fz|SwWtyN(0%+Jn=37oJZnq8Qd7))VQ$GTmyY4+#u<IxK%Zc?dk=v`>SHL
z&c$I?c06u@hi_T|u_<hTBb^8nwE?q3ssFkXuxbxQ4K)-t#-E}+lN_indTZZ)segF~
zd;(m8UymosC(L_P1**{2QygpKRIxtf-%~0GZiQ2JvQuRmCv%*sfd(e7PkRwY&c@s;
zji0v57KWTkdhs5xklD)2D&Uc=iB9L}hg<951S6^fsbH?W<A_vG)IU%^p5qR8#AHhx
zGYwG8+aK!&Dzw6kh+M}3Br@E0NZ;K^ccvCw_E8lilsTZq10k!-N=tshEYwJEx7J{<
zts|fxF%+d8KH01iE!P?Ljo(pkWsV#}rM}`#tGTZXaG%h`>fO5_%cuu|E7SjMo&YEm
zMtj|Tlc{-dtw8N`rGrOGA;(#|{R6V7<$C0C&WB|my^B2QKc}g!5(S-@$KJ_~3ce1O
z45m~`q)HsCguf7cM3v_tf32--n;8_<O*P4ap$ZSzAdlpt^LRQWSZH+!<-YF6W0pT{
zT{y>bu73ROP?=j+!8~355}&~$;vAW%rOjar{&uM6YV5z+XFJeE=|ewW^?BN;t+3+;
zMoH-X7dhL%^Gwn|2tb;hxq^)yD!P^<ctqZaJbQYUGq?{ZO~-G07b&nWHMYP#kWt#N
zJaBL3oy054NFMX<&v{m@mAvu}Tdg1ivE9aOKM8=B^omdA>s_vS0bD$o=LWC{KIzfH
zQ2%uYpMKd8J=_6_<`Kx~g!wYS!>jeDjfOOA47nYY9zrUI70)+BsMubLE=fgoG-$;D
zShGW}CIVNl4<zllhl0Mv)BD)f8Pr2>42U0|y*U<4c{h8hA^=A%`SqG;8Zh8c{sSpI
z86JMEJtsUh$c_$zAZujindFR$I}BA?>7S^UzL<Az>T~_Nye2dLv~de)vbiMB<6Cib
z;P#kaqEu{O#B8+@CfLbYKCk*&yjEdF)>2+^rTqAfMcePSRYUUYP3z!Vi7k!33j$t*
zD>firG|O4TU<0se!5~GS&UEkB9U6L1P^6mCUl2HK#Zk9^h6vTIsrgV38o-^kQ<z5i
zRR{v?jR58jn(j;;Qgd=h)(=4Z((F^8=fCtnd63b8kV3~s8p;i8P*w~Ak*i>l$mXH`
zA!0(rQ;AnOr?^@+^AGaX<Zrb^V)i>q*fOBuV$?-cob%|DY!24zmwyE;1`QRD+7ePa
z@kB8l5I=hI9pzBbihqk%aQn<8>+e=#O^N1cx4GhE!x;KH@&K2ycdWFJFu8T@W_%}3
z<ON%W)HDA9{d*wHwLgCr7XTKvkOW|7LisTx&Ww^_jEXBRRo*1QOizDqvu~@kvd4Xr
z)43;k?#CFck+gP@A|CP64`D^Vl$vveeoKBa`j_B0ih*mj-TOF?CMfFR3gFn?%{hs2
z6U;CrPM&@93`mSVz7A1iu@5nDAjSdHe?tPWSWB};@t8s*r~h>N?*?G$9lS7S@IwfK
z2>|d+H_Y{W1QjTJInYM?!TG`H-xWaXoh{apk7R*9UR3oR$q8g=BOZ$yMf<J-1lA<#
z+_~P)O{QYBsWM=R7R4`I>+-khB~ZDU?L>(_hFZ%hw>K|}{Guvk1`k`+b*>))_=p6J
zXK#Da5@%fpOjD~5CoJk4;XM~4^q=75Lin%Zu5&|{kUMW@i9MZCqx6JwpQribpI7uZ
z!YW>gPWW@W^6g^(xt!-IU9LVqH%2Yw$Tnh`)ELrJlt^!8l8|3jmeVa(lz`gV+JBhx
zJg%7w&9uSKUv)E*`J_v)ND?WbLR4Zpnp1}FZQoDfK~zot9u}>Jn7R7SDO%2N&2Nj-
z&OAui`{2<ZT>U%KO0Zs|YV__MlFjcM!Eqok$~aCrke{t}_;STdd8(m!P0-`d%=ZYq
zkn&559!WZYYqi~>fths~HqbM679+bn`lcRaP?<K7AWWgvvCy7~#k;1LJg7cEX9Tg!
zZb}(s?rs9e5b~q^;VHP0sE9LhX~;>bNNzDme0N#P+2_;91zL>h-+<a|jNi_rweIB4
z#KTWe8W9KV$tpCe&SL>Vz-CTqpOJO_dN|HJ-tpH4Xb&!&YJCy`!X`O`{dGE>m;^yz
z3awaTn6iq<Hzn$aL35x4egojCo$m`D`GY_n8)?hF*r-2d3{P70MFcrpS-AN^ONZ+j
zwG0Oa93sd5=d&|eqn=6O06ED!j9ZB~JE5SR8f{IB29&Qc-+vFgsU#q{BmlX>z8xhj
zff!TwXcN|kdi(ikuLY3WTM3&);y7d>j&;{+{Nsp8Tzn#O0z=6`YfdC~Xo*hPF4Eue
zDATToY~{h@1}|la${V}>sN@Emo@!@%Krn64K)3(8^C?Q)jAUqx9I_M^V>3PZHrHxx
zWtfN>?)|S7ld6%Q-z^T}L&RRZ(^F4*UxpluEw)Q*qA4rT)8(G#H-r}<E*X;-)yINb
zwH>8Otdxy~fS`}g&p;?O>8*4$`$CJkA8G%87z=hpq?9Q*M<xK0Ms*-1vXJxOfuJnJ
z*^uurkt5Cg4o}H3n;!j@B1l<@8J=Jto{wr5f<K3;)9tb~7lPRYd`8;8?QY}p(bz>l
z903`4NOE9V8nqha)}KiaRTrjq2x`yQ)}i0r#k@Bcn&AO1_PdkM+F?~ha}K#ewoJIG
z`vXE3(jJU~QzFv+zn;t{{z6XdS?y#c1u-O39z0P$NTGevHJIAe2h?R9Q_g2{C{IB<
z!X>*|y0q9#cI;3*i66)qyXtWb27heV9`>Eo-{Vx|>UN{j6ETR;zrVh_ZTVvo*7w)0
zhJpUtctD>!MuloH-I!I%OUr9n^>sCGC2L|f&1oe23kw@1hL_;4WWhU3E<oPcD?eAP
zo~UB?D=1(t61EV-+u5@h+;0OzIo?Bj*}v}eG03#%x#-J6%TxSmNGkXmZ=_${(^sTM
zpWyl$XAS<x3}BJ&<~yB1o9equ-9A?S3-vW})=Qs9Xd-@j0kOt(zNSV9SSci)94!#(
zzw2p2O3fxzmvbuz=^aX>px1F_1fVNxnN$}-%liCQ9nlul`P7dNXHot9g>MasOd9u}
z*DQ)qU~DM-eXeFF=efj>Nx=1*bp(U-q7lM2V#LJXpHB&Gompp5C|4FHa<W;Wf~mLm
z<=dVd@<TIV43VkQRun@{{WEyZuPTxT@uREfy56Gsb3t}T+#FaWK_`(wS9NmSvG14C
zzJ}^i$oPmTPGAKliaPj<<4RP@l#X79zYuj2j#SVn2FXi~DmDYKL^X63BZ49-M(de~
z+~+_(xnO~2III9wvlzvw>t#1h5sg)S1r&L%R{@bE9`mN4(iQom&wfX<kp6pF$$9%0
zATrU9h~IEZ<6sIkkI*0kqpAl;b^9%&B%Vii;h$sDp;!el?HTOPw?(Nv5)Fv2d<EHV
zO!T45ffZX`N`}MWfQN_2&8peV_N}d2<IrqVh|u<B5M~zxBo|h}96+qT^?PMia@Qw&
zhKddO|BqY)9@$Mh2vTq^-c%fynKN3vVnkM?`A$_#S<J8JVTVtN2GwjtBkyBW!^v~H
zYwbIn@3TUTr7!?TrIBW^#`H3AHqj)OmSf|+e3|Nam*!g<5s22>@q(EVx9|mV{%PGJ
zQ@6VV2{*c6$UU`S`NpFzH5BmUclpeBn5=GrD7fVaqLBegk2nR=^#5FBZ3ZpqVRLIW
zIL-o);&Y+SEzC)TcRwjgx4%#si?!M^1#;vCrkX4;+?%f&!D(jQ=%u*q+Sbm&p8uP5
zDAh4q^(4t@{KP=G$zR1foONV<Rt@+zfT?N#O!81;Aq!w+;T(*PgjD7qgxXt|$vR0f
z%LK4nWr{eCj1|?6cZ!3swXJ?M%3Z~P)JaL6zfY6&z2!B29jf1)^5cqa*yr5%Qmu=t
zffizB@qwODhLXPJ;0I>kjp&NR-J=N;APvBgRH!l1Rv#~e$+N9lm9tz_YurnzX}fW6
zuJ~)dRemtu{G+{87D3jBDtk^yO-K{_#{$?0zY1OS=^J2c0E}b}<0A8nrX}G=F!<fD
zBdrj<KjX%zI4xe<AzgNbTo;<C5(eXd8c7ZgS++h;w<K?vFq|Q(v46;Uk#^p6b(YE@
z6XT+;!@=qjfQTyCL%|JaS|#%DH#ApiqJqr#s=PSZ5c$*^Kc#rj9qlbJR>?G$)ejza
zmx(0c8%QkATr@INyrb(DV>D5NtcOz_#Tu$omo2Nv8}=&m09%6fBbHcOL!gY4syzRp
zZneWIbP}kqtssz`^qZtwu7OlU@!JOqtwWiypO0fF7%rrbem~(|{H5!-bIb^URD5Qy
ziQB6B=+72jkXfE4QS$Adb2is^x2s8x3Ir%|#XEq}L5)<T5yTWyi7;)T5;S3%qSpWO
z`IPyGEZhB4?L;cmrS{#)01Qr^Uv|@y`Q32<2Pcxj=5i62O-Om(41o4>cr#71xG=z`
z(3pM4ow>!&=Mca&=nV;lfXPROC@)4rf90KO-p#hf3xo#V7>lI#+vlzX0Cb@cI*}c)
zA`}<faQXGqR>wu(advDzs)LxXzBP$&My5J^H^m7#1Zl2moTw<DxcZ-K%^d^<5&_~U
z1IL`)_tWF!SDg~}nx0e_NK9qxoe(|M=Sc^J8gnul6(CqmMNShd>i%rW4YsR?oFAPq
zATD>~x!91)Cb`d08MEH%f00Fn!B><ISToCqj=V#Im?@!UD310eC#%6Ps+Xh-v`z--
z9UPPElN6}6cYid_$K(8fXMbi-8ucn9$xFZ!zo}CYHRAhx!(vz5itmChYrZ4NvI^fB
zbd^_FqBIo(c@kU!A;$d=bxNYShKh6iW$S|eu3_e&4G{);&4P?l^-mb)hbHouHg9EL
zyIQ>2KSn8GV#hwYQ#A8DGB;JO(h2AT-u&4_s@r;jJg}R>JYp4G!a40o1h}TDGPN$G
zCF7n_-D~Ekd^;0m7a;4ZIIz_r=RVKg!!5k>pY1yZPIsE1gW1ezjZkvdEaV4u{1l}u
zXIcWYnN?|~^h|FUdE#55{XjDL^T_CN3S(+0ii}9cYe1!vYt`i<jx(le4&w_3)XZ3j
zwhamkKHIow*%$Mr<h(3j+rN;cUitee_;k18%X>YKjP7AZ@RM93124eWMUqoh6R0EJ
z!;*-_^#NSQOjK9Dv5t%U?IXO;oMp^Y04nz^80HUg02b_cNJvTtrZHG}O#aGYfuIm@
za<SmPNDP~6rp0&L@$|fAufhB($tRtf)!0+OKKYbrPtU0zOsGQipqIn6|F^56L+PGO
z8(ub>zOsA-rWy#VAioX~018Z5KupQjEHDOze`z^sfV#3UsoX}_;zz+8sun%o<O)aB
z`qA_Qu9TV7D*T96Hi&yt4i&UtmAZQos+FvivB21eLtk{y-%jOiyjS8gta>8R-@?&s
z>wq)jgZ8HeO@l$aX`Mo0eUlUER7YX*{<|Shl|Di#t{$ud0b}u@=7|}&N0uDm#pvaL
z$w+rYg*@Ek%>I{&{Om@zOwHLkrNo~CHMaX(!S#0!&U(X-Vy**O!pE?x>d%<^7l2?n
zpVQvoM+A%g@cQ9zFwEoPgXmHCdC7a%cVA!1rRVn33Q%kSZv1~_U-X<RL3ot>_C_b!
z4fLK0wj^HmfYO!bA6%l~&}EF#T=7xOzdS1+`-Nk^)5mbRM6*VjM?_jVS~Z*zoPJV+
zN5mx$R?SfiK2J!5ils8is49a=#$(vABRXT51IAdxJa&OqAF~$>K5*1_&G79JBf2Rj
z4<Xq3UYuiOqz2d4mB%+^ro_1UHaU}@oi&PIbMh|ipsj6c3IngfZ=6w8VUD3R^&Ez;
z01g#a1r@q#&td$0J^@Vbl;<*a@IbSdPzV%D3k7wF73kQx$y;RHP8dgzA4NapS%Y9+
zSMX@nlW!Y17_+b<60HhEoiEeZv>&^dU?swrxTMbK3If+3?`3Kr80v;z&&fH|rg@G7
zI3m;5U1}1Uk^LGglWMN!5Rw%s@s-OH|96;89+(7}@bE9CnRUNGgKvP9m5Os2qtFs6
zvzl^3>_o^mB08b(1dyT)FtHm9FCIJvt)14@yq{`WQkhYzcGXsJ0o}fQ)1e8LUgO)<
z&V23Ud=m2whuNs|{<EpX`dd(HL!r&iDM^r$-yEaD1MZ6!vg8HY^m1P7hY>ew4wb$O
z3|rlat+jwqChM7%jZ12LcLsk$vC=(V%+|~;!$8=zw&F7w66}1n1Ta&I#~1@yfhHLk
zA{ja9uSE|wQp2Ht#l#vlGZA;zwjzu?JvjUyAg9X;t8ioQAv0pRW2cU9vKK*o1PE;h
zn`!X^!okv|8E^CbQ}X0V-QDU}o+=pw(g6AWyPd_y5bP{~R*2%vNaf}#5q#|9dOsvO
z)S)M4J?hlCRHeY~BNw;mekCZ`f6$c6hM#p7WeT={JH-kuZR_d#846#71tXMaOZ|mb
zglwy}d5Xh$T$?Y;sx)W)F}j^gap!8e*;!}*82AkPAKXX1X1xRemIS*?&qIS(j|qi(
zf9?QRUz62e6tEUA{wTFd*IW>BbH&|JKX{9iXIVKE_TPVDdOdB_0+vHvbLHsNcxE24
z85Zb&GCeA#EL(`aaX%nqt%N&%OH0M-X8w`4L-ZydGWNd|`!F`mGGn5Qa&jjN7TdL4
zewD+9IdFmi1QT+MbpJJ|pV<zmSLt4Fh8-}Dpu|y+p|P2`g8zcWy`q{_r>%feBY!9&
ziYfjbiC|K({s7c#5E8D8aUw_fR{k;?d#8bHy~B#?!?DTsHOo*jHysdKu#%y;VK-84
z{zUjo<Nm}k?9^<{#5(<3{V@cI@uSd@T=42XMEw8*12X&=-RFRS{qfhn0o)TMcT9%K
z=CF3!^@76zuSQmIq&!xOpV|4qu|)AfS6km4{2|d}X|E+=Rp5M8JNHG<!&+`<Ppho3
zW{by1T(d?Cx@9`bx0F+2#~(-s>}ldWq5XvO^#?nCina3rj;d>VVs5$ZwrrSr=)bo?
z_J)iRxg}>Aw&E05_+cPov+0X^_ctZ5#Dq8}_tV95VR0X$dB4V2%DY3b%$6tf0yv6C
zIj>ojIv|3Cq3^qhqTOUI-8flwp3`Gc6iCo3UvFlvd_Wu?n8s`YhGIVq)0)!gJ(!y!
z+lL#duF!)gV|0aJk;fc4f2`#}yNTVd72At&mzp@p*g(3}tx$2Wp53B|XWPfTvs<J@
ze*nc--R(?w846zKE}WCv0@?@9mKwfcdvgUw#FR|98C)<SjOcjYvx%E-v&X@owf&8w
zwiz2xtFamzl6h6Vp;kIz#@4o;2Io{vgdO5qZLPyU9H6o41e*9+O}LUH?iT-tHi3zU
z`&nG!GK!_|S`%NRGLLprX%nA6m`QO8gARb;WXocvmEheJQ5XFe53AQ@Gi}Qa`hF0W
zHLok2CXxo~N7Klh72qpHjMVLd*x}+*6zxF9l4`ZT|4&hS!Oy)=JNqV8OYG$a`Q6cU
zjj!uKOK^V4!8pfQGXKa&r|iFr=DGj?jh-VH#kx_oM54Tip2;n7I8?#Kqk;;i3_9c&
ziM~ylSNVxLSxqJIXL<YA63<JQRj;RM_2RQN^4$_Kkjn7I|1ySFCelWi>D-yfZ1BGW
z{vJ3C7R2$vue5zl%98a>7bHl;*9#_KHJyZ!sHN}!prfO&Qpy>OV3=%(8J3v9$q&|z
zP<zz1W+dy>P=Nk!q8X!+g$Y58f6(jK18i;4M;tw9pdPT%u_X&shB>=OgkAvdV?fV)
z>FU^uRL8uLhn7Q&#gg81znOYE9~IT4Im+xEsprRkvnxn%i295^cOTMEMlM^j@K?9~
z1{;;uivTcb<a2|xbD74@2Ig<k*8`w=fr=7hy@(Y5lc59;{+)rBIfB85jTa4J*72-X
zBW0nX`*!k?oWvDOc2S{JT9AGS$OE5?Bcw8v#am$mYcgDf2e#K2TB4o`5lB`wWx&ca
zt12W-pXSiX+fqvXMWJY2JFutF(_86H+btkE%Tpe;OFll<_*_M3>?_C0wqsF0N-_(w
z$MEn?@M#i#ceGTd1@C;Nt{4})JM5E4CUht;4_F~N&%g6%xamlWyWYDa_!DT3KFa!l
zX&hP{F^9d?mO3f9OE3ACpBf;k`{i;Klb5d>?SnRu4dLh7s(Yg!qiA@xA5CbMfV7dD
zkMzBZUkTNzj!%<7^eHA$L>NWK6eJ{5wjEPKn4z3RsA=r}oPnRn!*aJR0)WtXK^Pgt
zVW?~mgbgO(56nqgs+*uKb6mHk3dah6;BPfadXmC1sH%7Ie<%KX(Rs}ONtu{kIe$Mb
zz`{Jk>V~avbY9dwGloD8l4A06sqq7iLPTifg@k(&u6y3d(ZDB<J(MC<lSh;jJ_A>A
zVBD1uvs^=}u`<MOD(&ZWlniq}z&5^E8h;h5=^d+5Yy6z>jgONtkdN*0Qtli1yrR7@
zjTDO3cq+5kvOF4`@{_tp8bF4WXMSVHGd+vF8~!Lneeo(^fViyr|K4m2Q)%ni5oqcc
zWGvS+G=+Dgf=`862L8mRvt_D>+AkP)`faN;sAXL5+*e8}%Am+YHjenCKVw#4Vcf$v
z%gM_hQsJJP9YJIcxcyfGPG+`q@?Vt1ZJMu#m*u(ZXZfJ7pQYoxb{!+X`I=CFS2uU~
ziF`)}YNSCmk+)di#s+DWAx>M@m0Ba@)QDGt=JhC$Ak$%aJX)H4FE9H+A1j=ZF6$E*
z6pR%fpf@(6PMnoN^|?;oi4A|Uv&DVDY|E+KEl%0CZU^&4zjwwH$swgID+KigGBW*n
z?mWU(e9izWRIo=jhABMHUdeoc6JRb6$Y5c`XVjJwbk!DmLBjDfR{Smb;VYMsFS0+!
zrlKam7*Elr!CvMLPWbV;-VkYGYcEafK1)d7D2b!_brsEEGp+-jdIsZ?>gc0PGH61Q
zri!hz!}_-XQp(y+2vx%4nS>s6-`!M{ox<J5>whb9SABdVf75MO-R4xD8?!nQP*xh&
zy0PW7cB}Dz?8rebaXabktv0u@a1frFqE<86b;7l^(8gJt$F5(&0HdpWn}PeqVv}NG
zXJ^b3uKNWCHxV}Pi)F;h4qE(Zt82B8<+KNv$e}DQf2h36R_-joTQNoT{NtRk%-Jyx
zTC;Q_ndY{{(i9<zl0)W+G`uy(?i2v_^f&~>VtxL~WBN?OMrY;kO;g|7i<<*f*$Wa$
z7T)sb4;eU0V<Ovno_K(+!Ew=8gS~Ji9cM=n<^{M~sc*db=_1*ol}H+FZ8#2_Id%e3
z$Dbeics|5@=nVCoQ%IaVfc#|@h9Ou24FP0TSQhLX*|rw*A0O++yue_ZDFsE@$uvdg
z_t0+;k_4Wp|L{7*_}lVULQw%bC!@pAg=m$4pD}i(J3fhT<=Z!Zla&7Rd2_^>haZ|a
zsspwE_5Ju+GL3m%m-by;UWbdbY_dgYHM4#fP`>v_kCSTwa(X1H%#4#WZLLX2)Xgvu
z|NKo_TcVsUC?nryCKb`c)cKh*_NV$cI~NB6TXy2L9$DR<b8kOD4Ng#gFZ#(wQRb@6
z6!J3$H(gm%Gn`@n-OX<@qq@Clayc@QKNp2x{`$IkoMJxCxv$OllaKAj!i)xt&TT=@
zB@`g8I1KWfXq<Tq*lOa#hA@YTRK0i$lwGp$EN*L`4~Pj~O%`Fv$kmZdbzNYdmy3c)
zO*R5MAs%van0q3uD9Xh<VzC_KWe%xRCNme?)A)0tUD2)m!H|ruL83@k`kGutO9!Tb
zti%KA>7Sq24PnpB-}R!`8O{X;^2czfOf=}JF!m=~x@B!ofu-jTE*4=B_E!F&IIvUT
zN{KQ#s#%6d-u+5?NC>luw5P%0)qW~<c&n>ooaATXCt|1vg#?CM?R{oUKXq#*Ua%`n
zYKHjmOmTXC$<Oa9$T($svG_C8_2~}+Z~ce$`L6KZBidB@>5<4CRzd><s`;_Fg!^gq
zyWlz{^7s7J9@NIs^ySv@tR?G~E!9e=(Rue2@~X=2p6!K@y6S4+)dBarIq93n*Ltl?
zHw{*A8r1%wdb5`Qvo)Zg`GE}3xr*JJ{t<PI#*=f%F&|~)BM}sv`(yKQu8@nh=l4(6
z6p8)b@ocJ(TMpq3B(=?g31rs*ggQ}Z6|IO;aeSs=T%G6cT#@eT+3Oy$)}+jD)z1HH
z`zxTMePy%e>I!oB+0{^NXh5L{63nT|{Js*U1iFkULSy;H;z=Ry*8>$iS4XschNMGe
z$`lQ17l8?Zd_P{UWD`nSml`!F`n+1oCix-mRfn?5PTQ+UIG72|eiGrk3T!t*wfSU?
zpEJwp*jR$`m{gmT&l5~nsDF}tPA7TFF^)|0B@)VGpJe~V|Bjx;pmj>_QMtvpYooUH
zg#Gw4)aNrD<EXpFzjcRC_i1~zhT|qOVj|JJUw=rvsdJ3oauazx={6>nY4yB8P3s`D
zWL@!I3)QgGmh2pzQi?jhB)IQQXnz?(lZ(ju;cd~6eu^oQ%zM%>(v%DFRQ@C?yG1_n
zDaMTKL5MIl{96(^qJ!~7FD&Z(b~2nWEN2Y$af%m~deuJ1RD|8wbg5#0gFlKQ=Lwl1
z<6c75#B^9b2<d!pr8x0t|I%rg4q}2f6Y|d2_W7x63w->yD1n))57Mti1lcW;vvg(O
z-)DbxAc8pwD`5L1CVoB^S4WeK`<}0d$x%QXksAD0DC+B{?yHOm>Z#RKN0_ZzL(4zZ
z+Lz5T;pr?V3|0O*h8cj^uJ&GE8MmU^53?>#)EbFppvPPj)>mJza|{>#v?ZuYW?fKF
zO4E7W^Wi*v<QEP5#ruHY8p)Lv*mGN)+AiEZzeKcg)AF%{(Ms5_)C4n4J1SWF=!sDL
zhInRdOWB~KG}xV&&U?-t!wfwO__>R!ri=;VTD02QY!&$H!QyE|c-d>7yz?1t!Su=*
z@(N>UkYHW%Wqr@@t$RtEglE!IoX3aeR4eb9H<A3XC2fd7OV49FKz{A}^60521iBT}
zpw@Hkz_^t*bo%KO@2t=My{6sHg%;DOj@y7m*gG`MWxgn1z-H?QY3yW+<jPW12R@ef
z0JE)RN;DE<>nP@#vP$o-*b$3}YSlN-163Ai6StcW*_&Tvz~L<~zklDoewH^FOF-Uf
zwiB1!LKZUEdK&0{4wRRjH|e2Bihwiy!z32bo>Vwc^$9T}OwKQ3xYEaX;K9|0tf#Et
zH^$qsjJ`W^*D%AnBJg`j;E&osT;1rXU=f@jk03~z-r!ESm{f!5HEUi+He&N}P!qcr
z6k$Ew%BiJEXu9)0@8qMWdttknX_W5@Z0S$aNihh^6n}2Ta31Q|QE1_UUVc)Du>Ito
zW54E*KDiUoYhtI#ewY)OF&%8WbsF=-L{8wX@8H9d@MjWVdzX?sZc(YQzPxw>s&}^5
z^nI~f(S7>S4v9C|77Z5a^u=RGK`LfJnjJwiLU>pr*6I({dPK(Mx*r{n9N*h`yka{)
z&$7w&AQD^`Vx|m{x?5+ebJ?w~&WjK(RgbS=vU^6Gc5!}UC4yFY1L_|g6}CF)kZW{c
z<TFJ?kh|5`ef51i85QHoKmGtsGhISZVmHT@kuUB2z>X$5>N3!H@`?0!7EA;SrUf}m
zBJpNygF(n2<kL@MomyEMqv7qc>;-&jt%8qy9_duKy!^X%RO}Uh{Uojv(%=V?yZ15`
zKQVF~8=v7N$Qq@?cUBUIi@!->vyktqqzY8N*R&kvgZqDGpGBP;wL^*Z7{w>zGfLOf
z1eLowCM|Q$yt3%CXhb6ZBpUaBJj>D%WX;NGV+qb`Ny0ZC7`u`yP!%tyO|@$G8b=rF
za|Nj-oK#=^nfY)FB0+B?0N}Qp7*%IwsS(!7aijP{S!LekeQD}T7;AS<H{y0|=IEwv
zn?7wkeMrlR2UT^z_quZ+>B-;~bdm+F2-hEapmylK@t0uwn@AL-?i27KNk4dieqsmL
zI?-_c3g#jD%k6Va)L=Yw5$8<FB$p}CZ~Pp{u|=T`p;Kiiu1^Oq3}J_l5xt~k?wKf@
zM<f6KJYzS@l)obGHX%!mpNS^4s!`@u&!;98mzJhN6gm<=UpLZB#rDgGt<!-R@&B57
z+PBumUP3d`)xDfwqiW5QY4|X(zl?}Jl}^s&xZ*LXkzRWd^%mTe0(;=A6vRJZn6qAz
z*G&b@)~=8;CvoiHEXjFovosydned-ojr)!QzwMN_(i=>B;&up?1^M`}xNni>equCQ
zdJ8ZX)HMCbv!GX545O5ZuGWNW<hY3ikts&&Z5Mp@&Bg~kLo(T`B7gjSOJo8$mz1)g
z+Xh3n(Oppzstrc`8JxrhU!5$kj~2M>-zn5^Dj6|pt8C#R#Sl{CE81VxjsG_FFQ+n;
ztvtzBu1RM(U6IP%(_m*A{9D8`h@5=dppVQqMztIPVdj8-A|*RCOE;iSV|k97d}~<3
z!{T+7{HOU(3K%{EcS*}5u{~yQX3GYS8$5r{Cx9R25N_<7@wnI-b*uMjYP8B4cOsyK
z8O3DU^FE-YOn8-98I$HmlU~c##Bb8c19{vRp_}~&_`YmXHuwGLK7TJTOo1AJyAct}
zTyOjB;Jy%BFI!JW;gNDQAx%p|2d_nP^E)>*FQS1U=MVf`Lqqs8nHZ~jR8N!U$8?HQ
zv94)*HorfBI0a^m66q@LCGGOyk;|!OTxX#yh8&sdC}O&z3Wob)yD+R+n{@L3=>rqt
zyxEm6n9Q6tLrjhPh_EzlzaaJv)=yO7_BOhJ?Z{%aAP4f*?m<C(wMGFQXId7<6w4%M
zUR+{c-seOdh6c^4xcf0I6?g%b**a^n8L`;?-{u%F|Cmo$I?V)W!mIEzhO?W6K=4IE
zyFOod&ujoEw=8{yNmWwQEh(9gj7eU9g#yh%{^il-f790+QrT_XIDt%T&H^ayevO26
z&dC3u1^Y8{XYz8s#`31*>G5x?oW`?!h^TsR$6xC>+=*o$O-bIv1Cp-5&)Y~QcTiD&
z3)sp?jk|{*qHwEH^UUbWw=W+&K~u-a+`N`l={e*UusSVuoNb7jBL2Pme0HXoqL6n=
zcE$YK9!d9tn;S}&N)&UVhHB_`I7-}1Pep5TGud01rfmLqN`Y<M9dtQ4N{5f)c|3Lo
zo%3cd&T8XUR9F?d+m3kfGc?m0Y8$HPIwjTVNKcS1f)2ZRYIB~yYY;V&8t_te<Ec1O
z@LeS;J;gT1<SGhyszH9;(xZ)=+aH5Ro}1^V$!tPyhe5Bw$^b(dOC>AP<%D-{Snj&&
zt+~Kh@~K|*k}3S`E2`DU^G5M^T(aw_wf`K#-jg4cLwn$ZSpp(Nt#$K$^iQs)r7q|L
zRF&5K(ImtZC95QuUrex(O4!Yukpl5bcJsWT)5&qH&!7l%-2*7|<WN62(%5?7;x_P1
z_&Z#4{`R|EE{Y;KI4Kf;2;cBk-1Eb$_lk<`T*o9y_(nQ~W{okVvfRh|h)@Jmd8hM5
zi<|9_@;w3LJkQ^O^?9Z6&ZC+lO_&<z5yf^}DPhjI-On?7fq}Q&tR!Nz4T(=Ct6yk#
z3~L&QGgIFCR33Se)^J~t!{d>v-FKe5x2O(dxgtHIM1k2$l1ScxZcQBxVjPRpDRIhI
z54*1cTHPCnFX}@%pJmysJ@_$U?2&b26K4h8U#sIzXDed<4taE?)1)zp)(4>{h6<s>
zS+)<ztMbPzofd;M^HJo7fV`Ml97>5J<wIpK17_0OyQh_zYBZ?jJS<W~J&owF|1|S$
zricfH$44Wu(6VTA_u`Nc%xXadG@lo{ccHIT#ZboHpCXn?!Xrn=lep`p7S4U^^iwA?
zbSh=JR8s!bDii~Do)586lYH5}221(i<MRMCCOQ_(y;+nneYMM6_tq#WMiH&qZJN|w
z{@9&b+r~(3Nj%?(UT$>w96*gI%bu>F4qZzgQ8hgxAk}E9l9c*&eM%qaWno%Ub@>pu
zjlCSkD`}(Fb4GxD&x-2c3k3}Po6smA!aTE;y_EQw82tiA+vaj_`oH+U-xxNCrVJG0
zUILZsG9~^2jiq|$qB-{eqv^ZDss7*pIh}HdI7s$5_7>Tcy(9E4BwJ=iQ6vtb>^-vg
zmdsFARz?v?NH*D_kWqg3tIzj$UH#eRtmo@~-uL5vjCL*mu&=O^s*#-7PEw@va``bu
zdf|z_(bVQ-qnH^AzvoH(vHA8W?&wo#5;HdY!5D<=sVMSCci+EDTk2BPu4$Y!h5^18
zYr(}3!|;t>t33EJuV3$->$J#Fx<@o@*^qiA@v#vIyl`W(Mm8KT;NHd1D*6AS3A!S8
zun>Q_PuX}mk^g2jr5R+%vUF+4{sKy<@hR}#e=w1>jVD9>Qo{<rdZns>-eGU6b_Xf?
zPubLIKRKQ?2<V*Ff44vV;OpJ|z?soiuff^17=KYe|H_xP`Kgb?McWIyaa&i=Zu^+4
zcf{xX4m7i^Hc#w}$yYEj7;7*}OR0EUxF+7~emajcAon&LIxg&Q=+%(Kz<7>B7Gs_a
zW&H64@|KgDn=CsNJwmTcG{blNN6uj}5`<RzkeZ%Rat_#m1#;SjcDCO##eLd+c<Ez2
z$%mw4bJY?W%fXUBjgQElQ=vtb?n61fkD@<A?u8C7(QbGV-4~{lHga@=a)C4EYBHo+
zJ&i9YFLrYnez$$U5irt<d`mX!7Hjy?7aZ;Vo2KLgG9b|g1&A7%D0w%@CIvJX-jFlZ
z%am3^Lii}J=el^^g#dd>DR5tjKU|MFI@5mye)P^5nm)2-5Ica|M;+S^Ex#H~qR<eh
z!u|u}@QCI;4kHar%%^~YFAy#w&vEFLtM=gGlz<99koADQfVhU>1qMFzA}zxVA>sNG
zO!piP!9>=!lIXQTM2~$Yc3n5R-MBNq-hK!dZY1rOuk-4>n&oDG9!5&rU)U4sE&h00
zz3eL}*wb(EXy;R}c7+A~`8GaVy&36228wNM8pZp@(X^dZl3;^#e@Vwtz=p4>Ts4K#
z&<BeVB*e#2Uz556Ry;YMKRR@2xLh5Q9AtDUxu)#=>XTOS7393TJw@0`;o-8oH_}%I
zg3T)JjCXrXmt{F_-)Tla@u{8~Rl0}OAjAG9M~rf!SgtVN(g9*t&xhz)3y5nQVV$>q
zTWrdB-7@o{_0!&*k0)%aUr%5IgS&$73d5i8zmD=71uwEvoxz=g=EvcW-!gwwghxpG
zp4@y;iuLVE3v6foGH`hp-mw~49dQc&XYKYy*dRYd5;b$H7T~(1&1O2?Jor6B6bjK%
z5He5VXZt8Ad4reW5L7CwYN-haV8r4^<^*Jw;k*f`eiT>~Xq?%;`o1*7H+wMvgN;B}
z&9!-V6hymMAkbLU=)2LWcVtgBG}xBl=4%}CmO6a@h?Br6w#+k&sgvI+Hc;gAe~(be
zlO78$@IQNM9X<`$pVK}Yc4Sa%3z3<Y`aP=hdz;_F>v%r(p&Btk#;Z?4jvhPye=~x!
zLOr;73!%VOyeaY(SL(GrcG54<O8k2;h^O?vYhJQkyWsgGm+VXniJBcrehKEch6MyQ
zfw$TGIb#T$j1Y!IXou+*c^6sQ2qadH<#dgh$NuKG3&Li{aQY`480#bu;=P?WIEJ*p
zXed{m%`~s6&Ff^%If7tSPp~80o)puT!Y3EZa?gT$7^iVb|2(?o2vWLZ%nj%2n@4PZ
zZ_R>W^yfX5NKWL&*dwNhvbo2RTA%@WEWbm|Ywkq)Hs^k%e2KfkF7#%-numv<38^BK
zZQ((u=V5o8LaqAz_M`pH+W?;flYjB9!%Sv-@Y(nq#O%od)dxZbqfx5(7VXSff@&Ju
zAnnNexGi`iv7`}O1CZexb5qJ7)>wtB9yoqhL=@;Jy$4$NQ|Ht51}eJUqB`cxJfI6)
z8Vl9E0413)8T0(dt)RIPG0wB6wAqCeG4}O9la`r}5aYX1PJ8Wu^A#MAC^XPnE<qd}
zP5<`W+ShXh7G3zT^U@F2##d<H5R;_Hc$V6+pRvk@HfmOM6gHU(8foD%)ijGrdf)G;
z%>d`J6!_c_ttCQhDHLkL#y2bj@)9v>#q0;~VJH=~z7RQjy*lfX1Le8bL)UJaz{MWo
zFNlH!LL}@Jyu|L7)!`#uS`L;#T5mukV~7>Pzh1`Wj?53{dWq%S0>i>b&OjoQHC}vG
zuK$ts!#iNeTd7BW+p;C@G@_@}f!o9aER+vQZoKwb2~tMlZ{TW)ulzI?4LMs2-9ASf
z^$h$u!cUqHSIcZJ3mO(i2<Q1o7niG=4M+Vlg5!FYJa#1O2`b!Kpw3b*j5-$Im&gyk
zl-m%c*8pU-s$^BIhiQFrqhS*iM2v3!ANCx%kl$XP^;)(<07U9oLCoM&e(m7jU`jL#
z4deUkx;Nq~FM{ks`T=fj%$}(=^=id3us!F&Suuw!F_0GHT4>}WCYdB`J|Jre9EK0n
zdIr#(NhuL@5U#z&@4aa^QQ@E`mRZl5bJZRepV9gW^d?)rr#DcdoOdjIjKfLO6Wqdr
z#3Tm2RjiFyu@P#yYuPFrGUA-HOPXV~F5hH|WJry#V2KINjhAaO!C4^3FhWs8l!HX4
z7Nt6Y3#E4`&G*-4B>F#LXIR^g8bfi<xg*B>cl$K{;^l({s-4ixN2?{Zv$Fw=nhkk(
zE`{w!9_X{z3G)A(B7cE}LoLVdwv=Kwcg_JXHOT?pp+)<ffA3dLp4Y1FmQEq(i;mZn
z3QMj@lJw$^3qP8?!~oozavSA2X}cTL7wdK2J$xbhKgCU{^$6-zF!ckU=y_1cMJ3=z
zB@#2l?9T0}__^PGbyEzf1|jDN@14Z6eSYj4ex6_L^_mX9kRdJR<?!>m5GAqSWU))9
zfX;VFGMWorTE=%G)@$<dL0CxVOY2qeT6+5|)zQ8YBkeR=Ui_kL@rm~LFn?h*&;2Ba
zA>~Qc8qCD5s;tws$0ytS?=OXGgy3(Yat&7uMsX~y53%%zVAh=llb)WR-rw&(!0ato
zMq#8}Rp$V3Mj1$B@>j7x+$iCDhx17v(L$@RhM}L#-Rhi-S$hQC%+K$G^;h(of+(_>
zYiO|kYMsX5Hrumr*?6I<Dd^h*)Z$l%9OSIV*1Xvcc+8Jma8vwU#^*)r_Pk3dp)JpQ
z{zECT#K2J)rN|L8|4>@dmhv_pc!|WaWG5P2WH(SEqhFK768H&$A}=ew3lD%*mhw~*
zC^|H*-9Xt0_R0EA%*zp$D@nz~SAXwoXw8)OvASVb_witxbTig&dadG3lw-n~{^dY>
z>K^N6CGn@e3NP94@Z&yFU{x5!mFn#Mn7;BUp%^U?lhUKLU}+q__wAwiXu*pSRal<?
z0VcimC@!qC>VpE8UIx21D^M`4p*~){VOSlipn{>WUdt{F_$jEg4+O7dR^nX$jr*8I
zEdpTy7#Y}@UihB#zEwSeO^|y51tsJ1hWm5>^X{W~poK#|?7N3ZD)zhqbDilI3RV(V
zFjj=v(f{NkYrnGTQa{@SuT;6U@|wryJa^>NuPxNx&pGq2X}Czptiww}lSMRPO(;rV
z0U4y|*jYp869H0Xl_cz4^~Gz$%;Eq9-QO0eejILOG>4r_Fx0J@&S@1S&hGfH)Dyop
z`Zp|R2)+eTTukmS0Ct&E^>3Ci5DnF1BHe^@qX<LU4V*6EFFocbjJp1M{bP>ul$H1P
z<oofYx}Nk#%nJ@V;R>UwC$6ebo>J@O{9+w{xxK6{MZFo#sg=OWOs0MqF+%E;D|Pk>
z`@^XSFRr^0jU>uu!^+4t%!>L;iN|QI;&{&UUB_IQh?ooq%6>k?Y*D&A^)7}$S0rll
zkA)!4$QtY-jKX^yFJyZR2{&z_R4ukk_x4kVmf?>)z&KCT65F+bq>hZ`@}%z3<kwpJ
zp*grpJ~XpL`>@(HK>VOw?N=5_8-BtC%A7g$Wr{swbTi<|idp!#1U=Cyr}l9!;oi1j
zy0{L?sC_V^XnNM(Tm-~NIVxe3+MbqF%d6Bxgp=i<qX}L4c$qBOWGCz%q=Yt(G`>EM
zj0C2b%J=^AoXN~ha`o2lqP&QGx84#-5OuX(2fNT?_i~X-YvuD%vQceB>*_|~R+jb>
zvcIb62IcV3xIK<Kc;TC1t{e`h$RZT61KGaI%?<D0SLL)aB}DAW2Gm02mp8ja-ZjRJ
zOrOoF)?t?FL;?X@kcKT0^^484kaIZ2{xD=E#HkYy={nnRg!S8h6Hf@8D{r)ue`f05
zDWSe;5Yvtv8H+T6a3RD7c4O(XB*z;5BN&i@7-IT4qr<e25%pUWl|x5<(ho9NcI<)5
zwg^c~9p*f6zk6j3bROP?ChNKK3?xT$CGy3-$RFlDX#-=ICdkHxhanHt8L`zU8%g%b
zaR59<a+IG<$jR&Q6WWY+xc&;`uc$2XkxK0~p^oxSpi#srm@==VzrlO1sX3mJi*OEO
z?o)AJj@pReST+f)^QJ3KL0WWlumV_HdcEon^@2w&$ZLJ9ioI^Z3pc;e`uN};>Z6t5
zxF2-s7hAT>E_`+P{gkch^Hak(FWk&ULb=0URKPpj-ZJ<pUV*NiHq`A@Cxxq5-{WY-
z0;Y4*xj>qZLE*(6ZkdoN)ljxYW0@`AG2V~3+SM_nGn(u6KCc!7RyI*GB;Ox9Z=@7@
zV$5@;3O7J_lruUA`WWVX-h^`G8)ZeUV4%2NZn{q49qGCb-Ip0?14ryC%|6I7><*l}
z{CmboyT2eLt2O&~mcg65?LoxoxvuLUpTY->#-ovG$nk6(3E3C6I3sIj9Ki+<dS|96
z6W)YH|B5>o+4Vv{PYqaxM#FjD4;?vKh2se}V(WMm;yLv@)co6S(ny9Ryecv3q3kj?
zku|U$<5Cc(LOLSsRzw?uoPmKP7zzbHc6)$(scEHAOK(=cO&O;bbV&ia$#DF%!;Zva
zf%|{iQO|mnv?Sp22~Z3?v+cXUtdyfBXH}_X-4hjcDC8gmu|d@y@kSlGC$mPxZB&sQ
zkCM$El`tiCjHr#~ihL;aAf_c8y@afrWCuScS+QwTKQj>c@jgxR2L!sE8+)($uhXRn
zI3}j>-vd5IyzL6`SMZ$yzaRq_JQ4K%eo*%8e9IXFhq?u9I&*T_ZB^RtqiEfZ_J2e(
zy!p=l%)1Rbkh$D&p>|3;^<KAfbQ5^+1>J<6cQUHa(vjUs|M?xm*XU>hmnyk)aJ@`L
zEn8|)+pUCOaf7aj-H3W{z~dh<jX}o}3@CynFgu<;yvrXBbY`bA$0IMO!GXovg`$I5
zSF%i*6O8(Pb>?r>1057Z$@$M}0o`F$Be1?iykpH$H%mt7>;|ICU(lm9(GBeDGnx{!
z4$4%;v^d!q5*UX&KQR!)x&;$VRVL(iHe>XBWKRWg=`;4h3h089{s`aMu@u`nEFLKy
ztd!Y~gy(9Gb=|FtglI#R1O$Jlusa7tuqj^WV^8tU$hHqObQfB1#bI65P{makm6$fz
zcOh(f4BzHU=-_U0DKdOsJU-u%HU#gqH?`T&RuhbRbYkix<Gx=GD_3n4yH1Xap$Y(i
z2^<>sSpZYXFjI=>Ifu8^9J>y*>gucs{mFR=pm3M#=RLhjmHDpPK?*rVA~6Im;vm56
z&O*?sqGh|LAdVNs#x-2q_S>=@^)N5`^_{@f@>EaUfdj=w1#V2dE)iVX&$FE6K7$ZY
zq&08phg}^*>k!yOr3gDlQOYbE1#t)zTmsgxdMsW}7Qm0eYfh}#o@9cxir@F+mf5Ch
zPNXTKRNRv!xW3SGxjfM-JISdyfjdg2jT%M0&-SH|clcP4XnO?Wft{eGgdbPi!V^>%
zKf@k^@NEh{;+WwMGs%Zc*yI=<3YHw^4C#l3KrZ@?a-&%D-o2S53F~ho?^O-c@HHSu
zKj44a6gQD9@s+5Su8v-#?F+R2{(V__V{lVrL{^LN!0UO&crT<{WXX%^+OujvU?3e?
z<Ex$&WmnD;BD{sq5=T?;-}uWjUg<LDug)Zs(BFi^@)&JW?{fFL#s$frV6GStcRHWn
z5ym*bHg9&T^YRKQ^gDP%jxN9WWCeqe7s@ya3<^-0x|&jY%xx<r%2;Ys!{h>0$iv+(
z=}fv{$oH5538;o-OrMX6M8SIEvh=+sHqYWFIJhgF$3?4e8nE0He-U@8&+Z4@Dlqdd
z1qF8gnF9K-TJtFmjpcQ99qS9GZi>*tXKdVP{)dp_8;D~RQk}6&%Nhg|?9fbsoTbP&
zj;jDmPAV{*%H9+hCvbA~67+ND_DU)Mqq9PT15N3(Zd$VY5?aYZS7}*aiY$dJG;m<e
ziDT+|w)y$^n08EUQo=yi$Z$S`X*j;2=7O0QTTTsVvf_B`V5lTM>Np%4Czz=;h%>|g
zyOX2>fMgNpI~WxMW<&3u??8F@>cH3_(WGnoVB8i)8`AUzN+UEWn?mYiB^pw!n$;LL
zWGYL>*vpmQSLeE%+X~2uGmA!#NuLRIHSw-S1~4RP46&Oc;9t09-$OyxarK$PbQ4MJ
zDCmTVS>bGqGi%+1v;cf~aZt<(rze6wvO1zgtD`49zuZRWVVh_~wWzlc8gB7Z8qwzb
zD2$hazsDV$+U(aiAhZ?Em@m272*m&;kA7~aVEv>L>W1AL@qEdo`p_qn52u$WbaKlY
z*^(ih4{=oish+cGpRwB6+~b!1Cd-a9eRE6a^T1*SA4Ds+J7VZf9_$>=7_KCdW1x;8
zq3aaICUty&dtlVa&gdp|{0YMX^X|U6zfDpxZbmo&Ce&t7oG(<fx__^<>==-mCrvl(
znfvkbI=#Fs3I_aL0H>Njx45n=xhF9C3TY<GRlBv811zd(ya}YQE0>$pt(Mme;9o;m
zg8TcS*Vf?YPr+QhJ`%)V7#f1X0({5dABe8uVa%iNc&ZoQiQFmyUZ|#tsu64}_nW(!
zEFp~#^2^rIZu0`kNeVs0bsj4QlIR;*xS0!tnE_xjk^)|O(Ku-!rZs}S0Qd|Z0F3i%
zd5C>QtJM6S<75P89S_|+=-S>u8i<%o|8ciha|Yz70=RL}{Q5peLS2viE4FoSsMKWi
zmzF_SvPh|1oOzS4pppjB*Z~#H8&q1&ab3^c)uI?f-(GYO$Rr=O*FY*41t9~v;?1TE
zpX!AD^Fm@jw2Cv<JEQQn+c^*__(7b(Na=l3yv=(@F}neerdysos3|Ios+4D6AtUQ2
z*fj!Nb>;_Fjx?G{_SOCG1NoYJX3UYxQURcj2Dc*{7$l?i*R7ukE!Ho+QTS`lvfF2G
zZw#R5`IT<y!xuSE3e#Lx;bO~ITqu-Vyi)HxZU9DY$FJ!`ED+PS;<P05z-;`FhbseX
zlr+N{v9^Kf)Iz&2U64A*N`dC_z7GJ03~u#q{Wmo-@hN7el@I&6Jh$B8NCJ#8;;!ij
z3LuyrWT#%pP!kFhTTnvt9PjqAO~1l*x`j3rv*GZ$G?BraPI&j8qd+O!vpfTv+4kUG
zop;o-s@?m3TaVaZox|@-zT=$br^wb!hKKhoYdTZ4$J3JK)Yg;slAN0+5NKNhVy@F9
z)IF|k9p_}rp~sumUb9$v0|jTtOdNE2S9Y&{jdtDmy*=+edhx1AFFVo^nNiNv_m<3_
zlUdFNkhbbloa!^VH|C&vGEQF}=0(KWaGCrt@N>eW>S4%Tid$H{fo@_`cn}OCR5Uxb
z<sQ6HSK-Wu21UByF0Z+rC%C(g(%aNPUHJy^wV|Z;qXbXmI5emH!WZ47dRfoah3@kV
z<%w|b9=}EfUJ^PuEpp!T$5aMosAAk#4P(Fq-qi+-=X-jA4aR}3xItig5Ux=<SqOjI
zfVThM>bH#e-rhq9sA%|9|L9*s?~K`c0A9ag&I6a`8IKb#cG}H`0p#@onxl<xZ@!Sq
z|GntBdVQ5}R&w^K+}lTaZh`)Xq=bZ1^}SZu#W;f!?rkVy<X3;uMDJ58_6hl5f2dDR
z8P@FV@U)7SqY-Vh3n_iM5B^QLs*CBVoK>o{)8{{WbreZi6;uDzq|5yAiA%c-Wc9mC
zn_$eIW#i#AhM+2o$>t9n$VLToPPWNO9>m~dW#9NH7uzA{y(gqC1j}xqa>;JwKP78d
z(u=@C1-;MEUE;j22VToL&r*LbtvEV+R!;G(zt5NmiX)g(w#4s-kS{)NETy(=lu|X|
z68)qPc$tQNlg{A1mw%wifQH~P_^CEtRH@wDsHSOjCkrYX1T@>XFM;=Z14oUgSMAP!
z@!%F)0bVT*s;ryI$pLUG^6%1RpFz;dS5E4_c?o>*K*Ljxo#H9shyWlifjEKx=0IeG
z#B+R@T$z>Pgfp`UQ1ZiP2IgV@^5$I~tIE^n<DvY(S-CoDjPeSWqAgi1cV7KqjRIr!
zLErsv&wQ}@ubIg=*4-*9&=wY5yq}E&zO7vy#qEzh&G`=11VqgbY>LhXir!gP7-de^
z{nXc<dvUyT`@7CpQ3W#7(u?bags1bo@dU=oFlYSM^;b8N6n(_?Ih?V+Jp7u6^vaOJ
zx{crqxI{Pq^B@^Lyw6&wqx4$Q|33vvm)AozhorcGvqBBopof5jmvgyWN$)vQIFBgF
zPh5%>{xIHSJI%xU4Ebr#z)o<7)8bQo$WP)w2tEfu&DY0ABqM~!`AY*Cc<u8M3Yvdh
zEdi2}_C54SiVfsxvY|n$R~i}-zrmwE=A7<R!z}+uz_cMI>*?g1odTlzWQB!W3;nf@
zAM?>jyQAQP;`&cZ_BPJnM;{6*@2L%g6PJfa!`i~JPun^7v)Tkt^_VQ|?H2x3oxx+q
zehJPoqOvOhZp4@B{~bJ#xa{`OZ|z0B&Kq?@YuYQ~O`o0ZEljEWL;%eS+8&VV<%mJ1
zF)4L_R@8SH?P;pD!RwxwVxJ+{pz{RiSiSr8=+0N1oACK@zS5s41q)DkA|t(jHsufj
zDw096FrS=PoHT}$0c(1j7&94pdo<U%f-Qp`HIB!!a&|%oyP=Y6{Ks1G5FQ#pDf=Uy
zb~Iw6IsCrBrSQ31MVN|U-N2UcCTFy#?o>-GyL$I!N9BJTL)M2#%d|*@1pG8UTy!UI
z(!EyuR{n^doOB_I!)(yH`BFNDgHwU216ii7ux(HWG@*+>D>I#1{kkA;^BbQD);2*5
zKA>PWjT1A{4EJYsiq*dP<n?(l*8*R1Z6(|d!+1vRSK%Cm2q%9<eEjY8BNx~e%tv3j
z{NGvEXhbF~XG8lRy_g|^iruxjUqxAdeZN%d{=w3pFk7`UP$a4H`HU)A=?*n4Y@Ueh
z*|Kh<*N9GzzhBwFXT>>a2j|#o>1^ERP|)FVnUQ{w`$@G8=U*4xC(-eYoSc#3CR2De
z+jMT*ZA!}Dt!Oubw?Az`L~`=2Kv10+a&8xaAJbdhUsmyY6#&7mdHK8=9jA3iTSC_a
z+kkpd$HD!(eMc%1hZJTZY_`Z27(axsg=zn>o4AWd;AD4|(g<A+;_T?MQ4Y7o+d#C|
zqU%l9w?=NWg6?rFtJ7Opa3I;6%IBofYu{Ur`XX6nV;VE-pg_WS{e2?b@_T&qG61(-
zuUwv+mULS31~*Z`AgBA8r`{U6=x#l`BX$3x3ixgg0kvlRIC6zo>9Xj6I3CPl?bWp6
zmWifuTKac9lJq0J$ylYWo}FWgh~qnJNga`N`cekG^%-WQ>9HM=erV17gty@U%-plw
zyXojM*37KUu6gJ6j5LQt(Ee!p8RcV8usI&c8PnFRpU+j8r-mJa9}ZTw>V5>72d*s#
zcm8$}rIVfk4Pc7?J-nLwbD-3Kz=PZ75K^AcmK_@E$tT8|l{5ZcZrFkw0pcTK6k;N5
z28!KhHaA4s>@Kxhx|kMUW1BQu9eu%YhcmE*AqCq4yi;Uh_k8Xo(C$&GLy~~8X!g##
zPM3+=6#%pzACrY^=8_?0+MnR_Cim4;8>-V4v7M*!jNm`K(0*TBvgWGyT$;Y`)94Yq
zkm;01X_OQ%Zq@64>n_Mc_JjZv4ggY=d|3aQY?2|k?9xahJKD{tLoh0iy@`~ZQK}Mr
zJsK)taAmo*0npJ#55+2nEJ}b|*sMwF&HbyF!xiS?>YA;zHrJTEfGMq<o*GYbd6t7`
zd$B_!Eo-}a6|FMELc8N7aIV59mgu3GYXv2xuI1jGW`El3r7+~_&*yTRa47naBF62>
z5M?#*^11Z9OR3b@{IuhpIU6gQn2+hd8>p|l3$&jOjL_ztSn$(`wMXVongB3ZPYIg=
zi=+JFabyA=kp6B|vdCKi=L~;dUo?OqNtWiRZ$yPBF-m_BrkhP@kJk62*)p9z`}q*v
z5`u0E_EnI!o*C%7w87`bPq+hcZ|DR&1hX!KyGiuXh`N|=wFfHpQf$Z~gzMxx(MHUZ
zsGaSW)6~L-oFSAQ#Z6K2K5_$ZyT{M~)JoEj4y1uX-ERY5_a!3nBc(H#!N5u~mx%MS
zb|PP*L#}eikfxkD#;p;0bkE5KcW8tVooNaoOfgH2*^$DIXMC@XvID#A&UI<)qL_2E
zA_PuYF6|;2kC!Y-^}L@}9_&@L8HZ2arge9f`CNlh2~!Ywp$u1IV1buCW4OS&fE2~B
zjxGLDuJjXeY}}21(_-b{e4hTgN4t(@XOWJjH=aAot(|zOOs~s!#Nq3!7hhdu8?&E_
zShmB~(HlId7l%VL?h`XC4C-pvVFReq2<!&s7CMH%yc5eY#OQL1T5!P`LR&H>%tC49
zcKi(h?<43d8P`X*D>RmN@qPR+DW0!*4H7f$@ZE$RPf!LIKEXA%w}yROWu_I~@Odt0
z)!Jb2B+SaEKP*OXTh#guw;}yKJhyDQj%wliXA2Yo_EbC&Xu{0XDa>J!-Od}T+AE^v
zhs*RmWyz-$4AgZN_fD)zA4jqE$h#Q90NSn_?G1_qtx73jlLTvcudK1qEdtM`TU0XZ
z)2$~k%ox2EQG9-E!)lPm@&N*aFSVZ0-9T>hO6MolG}7X`Z`X(z`0#JzWlJf40PL3U
z$?AiBNhm=V(}W~WOEA0{avmwxm-|!Jm&xG#xd=JB3aqXo6Cj6e+@>5Bht=;_mp6h&
zPS20u$Kua>!a#mUSBSIX3M_(7honf$#wLapsPO5j-uUy8huP}iUbbiV>3(r<{Gwuk
zcKl~OwzP6p-kG313S$n<YkI*-@fpp^+E<9WQo(c>EOc=>&U*oo>7p^0xyU|ed|A>d
z`G;lNsQYUb8U@b{E;h%aITMC{v6}Gt;Z4<F=}9jp#Oun0{%5cL=zbHGL0jAC`%AM?
z{AK4(smzSqWs$|yQi}5iR3@p`Bou5iw;aHDakcc#580crU~!vm4;Omf^H7tN8kQKR
zxVt=KHJ;q>7p%Ybt#?Dttip-1m{4T!?szq)dhEsbhT=;A4b*-j|EW#;#Jonw4Ublp
ztaKa4{&yFdfPuTj8Z&b%=qImP0pjevS#134PCGA>Vc+Tj;5hs+UUU}P36s!<GPfFj
z#m&$K9MhZ>r)#SU4t9Q25W=!nKQFm7ni-cj3f?GdLwQ6MdD46H!z+_tWj&S^D%_I~
z$vx_;X-+>n<alg>v6#l^>>LsjF~;O|8sUrYzTLSK^~cLirk4rUJCx<D81hPc3{o`R
zVhH+&f-<|D`DsIcN4izwf4s?k!f?KY1@dNOl_BN|TsN@~OqQnZ0K*x8ppGNn{fd!5
zjE$E6+*4$&T4-0|TZsvhqKyniGJueh3Qg-kEnf5Ox$wH4K&|9LUKc~=|AXy%LJ8EU
z+N&9f|5y+0!VD=Y_nlh*msh$YnM6^z8h1FtHIP9CpS{p+aS1%SJ;(w*(s#eZ#pVS>
zA%|ctVKb;FCB-qXY&XfMZ)M0T@5YD+fJgv#*?eoCJgrzUL0?T&u7YnD{@)h0R_p=@
z>E6IZzft?X`2!%+*TCeMhVR}Fz6ALm6Y4&PL<37Mg^Smx$bdLV7oW%oUShUV#V*#>
zF>z42VTYw{0WX=fMf}$O?Pne?Y0g!op#<!AEb~}`XD_H`5d4F7oMuM;!mWZ>o5QZd
zqf~hp&yJ$Q&N#JzsMppbjGib;p~3LNMJ6P@2H7sd?_8KcPicL>{|}%TpI*Ih4!D6b
z5KE}z5ocjKPe?@!;YISG=2sB-)uw>`&w8v)GgbrIj2t$2kk8-@1w41SUY`p{y>7}Z
z<FO2OC7)K6#wHSY4TmNAtKOEqY{A3t^-0vWZ(%I@&nNF9+;-`oJx4lwRM7Wpc(uhn
zlzeF8Sr(UR+;3`OvvjXZE<YX^=@;AWEcC*UJ^a+1&gB*C1^Nr(-Lm+d^c%k4oCuC5
z2q(8ockc!Qf<xh>{N*$;J6-s4KTr?ymKYs!Mb;ZOm#`pbs#1@@l^$QtapP#~fF|d{
zc(T`H7AZFi6RxQyeXmw@HJ};sUxJ1Q!s*A@wqW}ge`_WPP&`0g-SgUhwhtIPQr$qo
zk_LE>WgUOIeta&w;3>zeko17Exv@T85>aQqm#=a*@w6g?)~pnv0PPJOlNpD<&&o69
ztnJGRJ{pb3^XCy{M*9pwp{w;hdT`1X4-PC|edNvkmom%XhLVVBRiPxrV&n-){xCzp
zTT&{5_KHtI%dA}J2`-ZjAO1pNmou8zP0R9Z0eUiHU-N>!*`o#@MxXU62=#(^<Da1_
z;VIvDUU^s#0=T8FP|qM=riUlKW*U7SFD@q#GWuN{mgIOQ31`|b2hc-Ce&{hDzw{ca
ze>@8qmyA*TIp6Vl$XhnM;U@q>nGX`6*O<N4xzBmGUx2=okMvP1B}GHA0gz@J+Malj
z&`iQrIjeXZws5Syg2uoyp0)(@$evTF6|*NS)8JIU{>r|{SLklTi=X(%EWv#!A|{WI
zz(9C@mOq0Z=~BnB*&pv6hvtx~u?HgcS-APvF!+hUCF}pIc$@FtD_$0N*vG@C6@T@v
zgsgA>n(H~j)h{mx`P1jx(@_lLeqOtaRp~(ja_rj!a<I5Q4(%C)JU#hRIyIVb>S%5S
z%H^Q1KF33Dc<1|n5O^1CQHY_2ff2?{GjGlkGk|+Bp;Zva8~4jzNY<F3P2^cu(hpYE
zGU6{i1@$_E8<UOq3yXe#ZOLx!aPqiGTq#m#@fIc8oXn%&{yAWzQ5d~t2If6`;1if8
z5s%_}yiWgs^q`Tyy@O(Br1U@jf7J<YCVYrL6N1;(Wsq1scrBVw*2v+uN(bIxnHL1S
zSpOrjjyzvE<JHA8g)=&-eC(y^tQdoFur<o(y;+gYmca4!0Buyp{*p$xAY=lTc;Uh=
zUf^1_;4W!?PjmS!VL5k8#M)x^^A61Gg{TjMG~x-8;L2%ENVqn&=~X9MuKFDBgp|<D
zEw7ci>qD@~_{gwQooRJIV;ec}i^>%wXiZ_@*SBEHQ@C=iopTzxhrZPBf#jE;-*^E}
z{)Ifti91o3uGtc*i<DCR>u%cs4i9SN5DbH>IV~os_5*6fpZW{5Z(nO2;%)hq?ve0)
z=Wf+r5fpHW6<^-Oe|t0T1)bfn(BJ}h{+@b#+wxEz^o!nQD(f&TOuTvMni7`l&8guO
zPViSnz!e(xXIqxJ5se=X@42eeZmy1%y0LD+J>#_MB6Ts?de~Nw<eeOs<J*@*P*_AL
z5}*&?ety|Um!K#U{_J?9CGj672psEjH+YC>7V%B**PHI;Z*p(wyZk9X^Kkhi+g)vO
zs1EQ?LpvK5|5ERZ?$WfJ>CE0HP);z`5x<eM51Qa3+E&Fli}#MaX3}#EU#Fw$X{y%W
zlaTe=&Z$;UCz_}w5!^9F-Q@|#8lwzM(R0LGeqn#8WwA;0PA4!X&`z`Zr3F)H$++S9
zqoz}i9vnAkOt}$oq&LEo)ws{t$7~}=GFxkv)>eRd=}<5uZu!97>ie06Vbhoo0S`l=
zd<)kV0$1aT^yf*9wHOhhXcZen+<PgY;s4_tVQds<U<g<|li3xjcqKUEeEy>@1;cy|
zto8*I(M}Cf?9X88k*F^wQFQpC<*Y<IGZT2aXgn(8rpZtO6J2J$REN>G+lY^L_0hSS
z+2ERHb3a<Is6PzjlCS9y!<L{?1-J;OeCqCimlL*Cpa};}r_kDz5<a(F)iTXWRB?{?
z54?V_@kC4srvvrr_DQNUcDg8+l7?Oiul4!a6p!-mNCs;Kokw_{{S2=vfJ9t>&HLI0
zOa|X1hvy$>ArJuTe>(oOuTQ{NEt9XGJJT!IR=A^t`|c{vPBpCg<frUqU^5o*-=lGT
z{)nAsxW)u$ImdMyr$KPUdzlQ*AKmo${B|>wQ|xYVmqP;25csZ0re1E65Wa$3@+4|C
z@co_a;=h=`3r62^7U4)rkfySz;H?^J#h(Cpp5CR$Y#jGs#cVYf92hbmPI-sGi88bz
zg!Gc~e3bpaz7if_8<jTEPg`%gl`dso=AGi%62v9Dsbv)1bv~fEc!}Tb##BmeQL*AB
z;u#`tmRx(?=GuWL2%Yi>=2ydT-X*-{!glkw9YLiCnD8j+E?gxtJEG!tiuJuwSkEZs
z#yTypCFr)2CRDa?KZKz7xN&i|>I&E_q(^KCnzZ`)4z~c$CzRI?=n<OHOV#mX*45fu
z!pv%blXANG`!ATvkWMx>X)JtKW{|qgQ3<`UFnZy649gzy@Tk~7mKfrTs0V7&wU;dB
zx_WM=nqM(qKY=@6D%|l#QTjbCnxgObNFZqdnEXQW%ck#$jdUk9-y~Wq83+J@1Bm!)
z)>|`g(NEl6^|lvryL>CN-u)Uw{}nzHaKhSVBg`tA2{Czj#|?NYdn11O&Yx#`iIb~W
z&W#(XRP^usJh_A~#=pDk@h*xf{Vl)xuuPwv+PqoN&!R{(F{gI<FERl+u>rdQ{lcw`
zMStEuBEIK@(MgxO*h8=J03OgQ#K<r=+5IP>ZN2K?4?SF8pWh*@WhH}m{CV4Xnmmjy
zHN0m$S16vA9?$jNrq};g$kiq$;mvGNKco-Kj0nsx{QCG$$hTx&y6~gvSXTgs<?Wv5
zo;yx4U4D)`TLbO4#P}zT(=+>d1w^80FXDjTapmSH?njF3l{pwvy<+BaUqZ0s&*n1q
zK;p{uF0q0{u^vjQhkeJ#(zi*}K;mYFlH2a2qP`$Po7Rzw4SDyvz9XYl7G4}fk^Mxt
z0590f{}A$l)D^y^UvGlh7ZA6E+_Udqs6wAs0j$X|>^E6(3+CiYYWULnXdr>W;YLN{
zQmW3Nfe*GPia>nH@H$C<g`U`8t|{wjG=kAk&q*`eAGe(%r?A%<Q4LePEz?<c2q~36
z4899_9!(GD3z1PIxNHx#&47^@{s&@OfKKFSA|O&a+a4<*UByZf@c;(pOK`G)f=z3C
zimf&1{SyLgD}<tEw{e)Q-C#ZCRB05>!<}9)S*|KM>2~Tv$jAquOKbAm<}?e9U|ak<
zw*R$c1`l$TW7$L?=ykOd`lKO&GG?#|YjnCQC9;14Ym~*QM7*86)ZbrS(b;zG+@$$M
zFVU<4M?HDeCDn0ZS37-&|G+VGp;!huI8pLL<iOQ6u?^Hu)eE<OjTq`l^GNi*U|JW<
zs{oBwL&QYd$FbgIHL72|x@Sa{Sv*Jdp?EhRETwtxce6)%h`jxy>{qLB;Sw&6Wuv+y
zw+5x?pQvYwUo82o)J%(>s&l&O*P=RB!+)c^r*5{~?4lF4Fc-jAQu<3h<9|P=A!B4v
z1cBn4u}yDOVZ_QU*XG5L5{ybw<|D^S$KsV0ewzyq6h4o`+O_rX?n`KO-zD8XRT|QH
zj*QUp_0LXb4^JafRzs3Z#U}r)xF@_<-n1e!AO70M^#RXSfd%!m9*XS0NDfDAI9eZv
z{|%^%le?AAq`4Xui@mB$|Fdg`rh}zL714R8Dm>=0ts%wgSxM>6K<{Vg@ES<U0Mv5?
z3<2RIZJOeSNNNqSY!!{A3FlR&hd-gc5O2+?Gyeye5}<-2=M6d?S0T97xKIaAN0c)2
zx2>i;km>bAVrGkJSJ}WP%IwW|z=4Wvd;np0uwz>Ra$m82-f(6C!a93^sial7e7VKA
zL73uaa7i$T8o5))pGs!Hyb$okag}){Y#XAUYMs6e=;$*g^R15Jglu!~@UVBOu1lFx
z$t)a&K4|LxsX?F7KY_;Y^oV%x*p!Yyho{Zlqr5%~U!X>I;2OnH!a-z6U5LePyBanN
z!S>rpD(+CQTNbAiV7dBF2&B)YK>XyEOJne@#jV*6@263sB=Wwh-g|#OXS918h5SB(
zkp(e1GPBiWyaAUVzF>G6N5O?MA;o-hR+}&hVyC--*UR;4cjT+CxL%t_NplAkXV3R6
zHhJ2t&(4c3vu{B+O(Ie~i;O{Hp@_?qNAqiM;*Xg`g&Bta(_^J@WhPBG{O_$`xwl|L
z8>bn=V)dl+XMZFmR^`!QSdR)r0hcaeQtUE#Lvdj|Aly|B8|&D|!^Rnb3G(-+Ii-Xb
z3Ek#Xk#WU3yfB{pSvI43e5s-N>Yk@oTkgxRr%Fz|P}eneZntr^RIz`}zsqju6Xqtu
zR~S2n%8O&~FwQW2_|0=5HH$If>TP75@Bahc6E+OQD`2$zh1SytOv%_?_^f1p{aQQ{
zr(+^4X6SF<vV6wWZ9tcE2#KkacpxStmVj0gwtQg&wyam>*(pl8{25%Eu)gxvYhM&%
zU(w8x>8k3qe)N$Ox)MKzm1Dlt*MTJf^I<ijyRnne>kDCdQ*`I9N*yiiOCsFr`k%6^
zAO7-u4-4N(R(c(-fF^#QE?onSjnsi35NvFc!&w8G_|{hVg{d~&@u6*SF+Xf0B!Xn$
zsDvoUKd0yo^d4|`5b%UiAiK+d>-H&l53y6hwU=kpM)K)%9Kt*#H1&oCOffS?BbAoQ
zMa)`TtJlX{Oba{G?rZuMdF)JOkMiIk?9>%piD55UwL3pSN1gI5$xzSSfq3xwMQdO{
zIHZ**nxF*m6W@Rz(J)qYw*)!cjZ`146gEk(uCWb|=7K#;Y`bOQ)llu7%F0HEPh80|
zH|xELrD&h3RY)Bd^L|zTqp>>aJ|trs)W`oy(kEuRbo`mxaYe1`G&55(-Na{oj2z=d
zD;4S{gWFxj1FEb#XlDGeyIudarvy=%EyKFF)8FtV#<8zb=#9``bGYDKdnhu$brr%1
zON#teZC7vSyI4;QY1|4gUZnn2?{f5oDmC{d%*QG0ghl15H!7qu>Gh;$gBVpse|nTT
zLt^r`XN7mP?9EF7SdA5rUtikO)7OTp#*ONlOG4AnvZ9wwI+S~rt1}iC!-57n_G7t)
z@a-Ski;^#eNfDyBPt^~<?D8MEAedNvYX3TggOmR1%7{L;CzXKH6&kcwNQ9SLV=kxi
zHjB_5Z*f})djrp`dK6>IeAAi-(lEpS)O@T5_g%T(mryQ>m^z+3<LowL*anHos79)`
zJw}yZJ~QB0-USoEXGAe${SF`1mKc*%$J3R*8iTRXY~y{0TMX;JhdP#h0m>gxFigMp
zEW_jMZQ$>-U_kLR@k~FS%)5u}S<LPoV8-;Gb2JyhAlYI8K6!450tyTDYkZcTF026n
z_wmnW1pJJv(-1i2W=(Et5nrxqJlLR19q^b~CsQtk`pW_HE4;5l<D7GzXaQSYRj=7E
z_&aI1m<9MOZdq+W#DCeiM8E&8N>&L@Oi8pueceQgI;f3^?~4NL1AsQi`iMI6$uSfG
z{!?SSBG_yvo!m2l1iH^5i=fLU=1oK2`D9t;A3pY9UkUyf)5LO<9UAoCsum#CfX6T0
z9G~YjxrL2!)o)MF1cx}toYovG`KKYKU_}8=W?#)wifPdBKVJ#`Fca(w8l!Ve-5bSz
zwLF8%>3ydB5ll%te)f@Y+0L9LeC*wqB)Hc9k^WKw8g*z8+iwamG7Bi#AWP}KP{#5<
z4xUJof`aS%j-cpXSRCF$k#pwyFz(mQiUw9QG&)7nb?Pys6)RGki6Sr;avoV_T2xis
zDc^i6G2}L=orGfs5{Mn97BqZt3PMSY^K?r_(0P|%>yBBlAr8Blp&siWIGW2dcQ$wc
zK=`gpTEtilB?O<++;dFENNKTVYk%CMbN>2jBN2bo49Qo%NJ+*dQwvl9peimxVA+$j
z%s;v#2FQ(tQen-O1UVJTBT{pOfWkr^5YJ${Oz%4cHi;n>So`7pULZk$x@l$tvuXqs
zF(MT8WkX9YmIifajmf)V)V!2^a*Lf`uE6H_SOo~YbmWs~g`3<!-Ut^K-V)vx|Mg;u
zD(|G%PymK-;DJ!PbZa6^!2kt!Dx-*n0$hUe%%2BBC}tE-ATFL({*i~uuf$YDMnuJ}
zTMoUpN<w0)9~QIBph{x;-1r$OodF;-Kpk}ibPEEByFgSU0><KSpdc(*rYls20|hq*
z0<z)Zj>0~ppx-_9$Titj-@ZbD*9si~oSaQ$^`t~iJj)Mj>Vq)zz;J(lbaCHxaACXo
z`oU|%sf&oBqE+`{`Cn6RijR_E`KBKi^W$|(0d$Dt)CMmNiO1m(|62zO^$Ywr1RQ)e
zs6kr5T?tgA>!sH<TUNJqp1gU1Z&H1mA*(5bv})(5>?a$xVGaRq%)DFcsxEw)S5Smn
zZ~JA>qyz%7aD_!MGMzI(q@ZC6WTU_qdlI`0Y&2*Tjo~+`3CjJX!di;f1vneVL)HI|
zuB|YAOe8p<30~z(5QP_bO2K1PX#s>@;OaPp$Wg)y7Z0Q>U@t`=yE8`<;Ar``=yP!5
z;Xb%z|B*F8`gz=HSq5%Kim>S?L@CaA)vtc^K@m2CNx;INf?PzJfm5`Ado|x7B&gKD
z0Nt)?0ayduHI30+e_bmI-*8~7(O;XOyoDy^cH8OYh6UNQ8yjaP6FgkwIr$R*_-7R7
zQMP>vK>E|YUYnSSqo6?ry&owz7DG>12<CD$8eBN-!pF-8u4zXWp}w#KJFRV@$&XTG
zW1Y^;et+FW`c!Itv4u(3$Ug-Ae>l35o&~_r*PZ6PqpGj6T?67zHK4jkWTSgKF?Wzc
z#&Yqs`5&zo_?0&xyCM3&KX?r!VDdhD<zS@BBTn-KHR>Y`U-W?Uqbn{^?0l4h>i#mf
z|8;Q)MQ9A&({%(<jIla^bQIRu4Hmw(t=v6`{zam8L{9uCiwaqo-oF1qT-q}y9=@n*
z&TN%0<8fkZO|Zi<m>m#Ean@kRJH?V|IR0A=!w-a`FhBuS>@#qf|3az-?E1b#EQJtg
zf9$Cf>v%31uDry-8lYi0H~ctJF)v)fhM0#|u-ee|$y_x}ipbeGHc~W1ZEORY_)L*%
z<sUhEh}GZc1Zu`Z0PM_w&8L#Z1p6SiPq#*MW3CqC%~!!=Y!<{q+fE*l{P0-7o5Z#f
z9RFuNio+i`{feQ0_3cBz0LyN^WTy1ZM_ALCpo}(Hs(PWh$$!DKe{mn#1I>Q=4P7U|
za^In)=_pFQbT6q5dP#u8I=;LecR@OMg-!Irt%QGQJRyLUYDHA4c-V1~To5W>y?N4=
zBZ!o8k>^OP1_*(DfI0aMT&odADcdtfE9e2jZoqR59*PSS;REFCgu|X!vo8O6y;f0X
z+ISCiTks*xIaSE#<+O|1s4D&PZT88JRn8a^epWfV)THSMH}MR`3=wmEFzcOHXU|z{
z@}VB!iQlJ+m%*gcbY9rFIv@Vd1HH4(vLK3fJ3oG^UzXs>JQ)I|#1=aN_Z{ZJW~i2k
z<A`;XG#5dxwhA_cj}ZR%q!WXF#;s4gNv^RgVjKs^MnIQ9{VB5OS~k&u+C)gGqYCRH
zdMURVK-lsu_^kR8Uh|;wUtX~DA=fA39_P}NICwJ~IyK-9n^s!91+Z7hFqS>&b10@k
zU7;j=(z-0ukP0p<$ICK&NX2NREWYt#0$4})HdMdEb+Mo`0}paG$^`r_?``smFagxL
z1~^L*>JeO^r`SDePj@XAflvbpqF|?e<?;<L1Ab6)p+J~70C-j|^eHJFCDyQnod90*
zxq|<QD!_@g#Bmx=i;+um<c{g#{{XHK1r=hXLKuMNztQ6D6Bxrvt1l2Ohnd&AFwq64
zQ$9arc%B99+h%w*aCS5Uhi}pI(p@+Hg!tRRDOjL$Og?jn_<w<&%96qNiU+Vi==MVF
z;|yXV7^Ox3>O6>J5_wZpg^<H<DS*?_KT-~ILMRB;F+rlV8vsizSdtQ@gIzVhIA?k(
z#N(yDj?`E8*<JO71PxivVbH_62_R&q^!{|EQa_b89~5M{Bdm4&;R^SUXQLtbC<-2U
z0MaDxGOqN^>w=7lR*LgFkc54K-U>t*;n7}XS9Kfud<3QTJ6xfAxF3z5ah-$(i-7LP
zKke|p=S-70UmV9F&z~S`JV<K(c>w0S19%`zf!PA}K}p()j<*iV^6|tE7!nVdhW+3;
z5@7LOP!wT-(?&p?o>6Rho0!L}$qTmBHP&-Kj6s(9o~39Lh8I7OD=(a0u*p>SJ|_#O
z4@pgqI66EE!vFo5)xc%@uwq~*qYZ|R*S7sM(_JwHNHJkWT|~hZ%O>q>&BETINS6WE
zu80oBUx>Z|kUjXn*HDM>2<^+#K=ADCt@KA=s8A!NUjz?$h`$56meOJ(qQ@lRLqs}c
zjRJ>b>3fO^B~W5OetHNG{H=1Mi)9Ei88X)MfuxAwdAvpGN+ybXU>6sEH_oZM+h;yf
zQX{+4ZZH!T=J~+ltoj48P+S*5=y?z?HX$Ls(-U}jU2A=lCNFpJ=?RsN79V~31MSkd
zm;|-tQ0VFZ+8k^McE-KNekivpdICd7NCNSHm9xOSd0WMzOx!QAO>%i}%z9yOiH`Kv
z@cMYoJbBf`L+HKk7lZX&{!VDHXUtzZ7S7B%%-#7P@H`j1m&01B;1lqIx01-#4Y&W$
z)?J-j1#JKWXIu~#&RBo|#-5Y)ti4MX&PhKh=tsl+IZC1j8xyd+bX}Z&?nI}@3}}(i
zc^-rw$gG7a{eClDJL~!$d<1q6ISFR*xV~iJ=pstL5hyNodsjfG3QAy{$X$?R%VnE-
z45=oC@77P_{lZH^D%>P)#O2W{Q%MwG^LPJ<*5n98s%(hADgmA<jyO!sh!{Du!1MvI
z3w0(Bc88rlqlhncejtY!!Jq}`iA;cp4?VGid$!*(AczA`{xw-p5$FB8A4PkHbX`fB
z)7TEok{YyFV6>2rwyW{kY$PJWyWJpK`C$Is=;aO-!T#RyIfsAJd~}Ha4U|s00IKB4
zu6hQ>Ns#AZRZ6s}Ox3#7Z;IFjxl!^xR;A91q{cqw63|77v%q%QKj+=VpbYH?ua?<)
zSaUrS3?-vv!$u#1zs#+>zd^<E3QXE17#gT8z1q+6KHv$DWZeCLw^GT7V-G04YCQ7m
z!}iX{6&I?RC!P3!)HN4!>NSKN)t@X}-=t+NTgQfCV~?4C;Rh>TDWa<4JC@S$oxb4a
zyu9-`Bu`RE>7o^w0$MddB?^<@j<$T<X^pw*US<oOk@0HHtAjokw>Cl$v_0RA#Yamo
z(WXDNViEg>@6k7ZIzLpemB^P8wPx%?z0lpY(O-i-btz_Zl`vHxG*vc$(e)uLIb}ir
zM-ez2({xb1MMYJDKl$r1;Clo^zO^R7$+aD|Tl;&UcDNtX<qv&j#=gDF3=9)hx(idf
zPi5sarO)o4JG40;zy|CZX+K`6@U84I3S;L5e6cGH`=k3LkGfn|P9J2qu4AD252U(y
z-WNd0a(Cm588Vv}R@=3Tlj8tu3GdZ>Xa!GByGSQ^M>Z5t3CJ5BgKNT_4BJ0%K#R}p
z(NJhSU%h9c+WwGOkMnZ%(+3-MU%KnRAoA-<%rX2H0_N6+NneyJxpar2l;fW(`}_Mc
zmjQd!HdIA~UQfN@^nXCPff%=ivUUa@E}+&wKZPxcD*rQ<X5Wi@aQs3lX@a=OHT?~;
ztTW=4^M`lBzc-Xj#*Fo)yBYZJdXPlBr6C4|0d1t})FPkd+}WGgtNVXzgpI*coAxNG
zj=*0Zt|+Z4AH1yHY)k-QPaKtSB+`R&i524e(}RThm<vHta17n~i##eWotXmJ?BES+
zlGn{_S2hHQ@3Tz-YC$XIG*bLlbq@;9I~iPR#2kCM3*tWY^_ugQA;*Su%|L%&r;~JH
z+*BeC<^#}Q_MG_Li3I+s!+9tF)7$V?EUw_SZvnYFWVetETdB-io_KnF*~KCa6vDCo
z0&HDR1MX>YU@;^Czgfn!djzC`B51&O@A8h&+foMnNb#a(XoiQ=ZSv=lR?3w<5#$ZK
zv8SbP<n!tdW&#_KZ<ekc=&4mTM);-9-_ccfZ%T%gYEN9*QOTh(D1CoV9x)@pYpN(-
zH~3bENSkozoeDHXUWXRJD)(*T5KBu=|1*zmIQ?x2{>~{sgk`rEVRVLlyOx3ut+gcF
zRCgS`gsN7vx98L+QW0Jqxai+de;x+#H2r$E*T-3Pg>@rmg5yQEAcXh2&C{KvnP9vg
zQ>RGJE>Wp$xf^lP@f9EP$7_<a_KP}f+Y}NZJ$<Kv@L*uhTuF}kc1NhfbIjI`-nq9T
z4ji|KDJ`x(uV%ZhocFwTSDgcO);XyTlg0X|ZAEsJT+g)O(=8TQ6B}f<7~#85LV9>7
zo2@?_f3v8HIeq`$J;RfrGL!$Njx)|h!l*NIHh|(K`gi-M5=H#^&?GWVB}Q)}Z|L*f
zSaI<<_qaeJ^6m?%^o$E>tjv^I>}BTtw|OMaytTK{%O{H#uIhvfy>oD0KB$go=_Dd{
z_uIqiE;8X!H3pnTmj8do{)fIG*IihNe}8Uo@ISy`pr=eHaGyT@O)r!Wj=lshsW=32
z7ZC3bYk54CY<K(^OuX&)so&blN$swWf0B^_IT1et!Q~m4ND53Rq`#+IGiiA#h3_tn
zcf$msXd#Lhpon|t-}yC3@6r3F#S{*DjenW}w=131IzlI4TM!!q{)c@xx^?q+r1l>t
z1zYhGR>|VqwFrbX%F)-pQ5FpWpu_Y%0&m3q)Y(G5*tAiUP338iwK?el&A8p}k0hw|
z-*?med&eHnTL#UO>+v<H)dVy?+hXxs*VI)pCey7^>K$Xfz+S+w;dIqg%rug+n=1>u
z#368XWQFx6_GC`<g9#OL53wNZ5Z5W%&WpXF?*I}&?nj1}L%D>kdwsyV)Tw%kXYXLV
zqAxd59iR8RZ~OSFXKSjs)4ii5EaMef*(W84iyUS_*PrET@b|HIJslqFja`3@)#_?-
z10n1o%_2r~+d8Tm^x96H*|Q2GMS3j5Mq7;HzK6enE^S1vk4`-=XyO5Ek%iWo<XCE2
zj!M?6XZ+O6RzpZ24{i1$Wj%fySlk4JJPqB3xTsYD9FFfLRYS24Z(KLRx*#ERu|m(9
z5Z?{N)m)Xf<|pJx3;$!eUnJmltff$1ACyR7R^d%nH5cw{P8>o^*ZQ`MYQ<@8A8iZ6
zO@tlEiJhE3($LDrf4M{2++^`{YY3L~P<PzLDV*DX{6Y9HS~NnL1R@+mncoRe9~(k`
zKFu+1-fmF%?;JC$z)vbo#zb%aZexZN%ASgD5tl8~pkwN9!1o>@1#5goMoCV!cU#eS
zF@e}*&XYOLXRVYB3z6*0akd7ciZ*=t6aPbSyc7tuNpfzDewYlLouoz8Z$b}xZ9*XD
ze&!nye9&yv2h^GV{^#*V!}HQ%Mq-qw6L|EW>>VfR^AkB?Y1GyuK=sgW(s^hOA<UGI
z?$5K289$sQjLUi1Fh}BiOCwxJPI4|c%Grgq^fFAD?dtyAZkt|gpZL-IgbsE=7!I0K
z!xCaugrBI&uP_dJrJc6jw2~dxG`koRbH3>+K25^-nHzn?-S2nPJ;)0SK>x)ppU|(H
z{Of~i{!C^#5#1}=`o%%pzUqpL5o;8$4geK2o`|(PckyF#v|L_I`P=zA&z@341uN+_
z_B-qXMWYWMoGj(KK44!Jz-lP^pQVI6dLx5c_dN>Tl6=Fr=gNL4p^W@!sP>kv-H5&1
z@ZRxr9zazav!~?2Huq?WpO_pG6Lz#EeC*ziUJefp3${fqm$MkYd$v`t{J+cD{Nj7Y
z+Ydd(zN;EEs<6KKD(jb>F6DRpyUD=v56o~Jl6Q|4WFKxidy~hKEP+FOn^<6zXL|d@
zt2#daD@6R0K{=LBxSNudpD*!VW5%!Tg|J7l>f-idk3~j4M)<zBVwd%;q3xMB{K)@6
z`{oqRUh4SWX%hejjEXvUwk{mm+XXtw7mupTt_J6lHIf(i*jMv;^Pe3nzxaM5F-^2|
zJMQ@NO>_77x>tg>CHv~937r^RjgmAqg&louu0($uef1a7w-6R#C`LaF;_!N&x}L>;
zN_zelC%^mbFZz0af|1T-o!yhq`fG*Dj9<lfnVZCA_mSC^vKVlQ(U|-lsvC@72W8}^
z@&ruo%W|qRJ!M|I2Ef)9cR?F9p?s2bXu)M?9ma34Qd{?spcK6mQ_cRj_*FdRH4c-1
zi(?05iyQm?vIG4A{*lyLA4=n@!<BL+qs<o)Y1@yC9$Gq)Kh^I~VwqO>ip&u+)Pzb_
zV6CY|a;o}*?B}9GEa4C0p6d(p_Yu^S;xY<+NDFh(pUV^d$Fxa6<6o+6pBFuj)v(F9
zEi9vr#;FqgKdRmY9Lo0X17?~ugv?k%izR!AP(m6q_U!w<hZM3z$&Dym_BBfkvad-<
zwh)m*Wfu~Xgd|1z&fD|8@ArMj(eeL3$Mc9WbKU24p1<wvm8^43^h)j&I=EBVv6{Gf
zh1On(`!kEK6T<oh4NxKwNUqnP>!wt(4_?N~1c#r^kyKBjt|I%vK1#c~M;;vt2aYj0
zG7bJ8`a;Au@#m;Yeow#B(aMMmwa2f;n!aOoy4Ib39RMQTj5iPXB~Bj1{$(dmNwh`3
zYx@Gb9pC!I@1RiD^1HcxFlDX59NvRaF8l2+>HIIKgW@_Ol)qUDNcmLUajLBzSMa$l
zp}CI6_N<1*CoZyUvHVbTS8j~UxYihvH6le2T=cqz75E`{eXe`r&)ZRM21N{H>C$9p
z@TBJUku^zNnKs2d=p=^zVPHTJWd}0PBXMlFLSFbO>f6ysynt?;g)8llF~D2JuY1PX
zpHSh{Zo#an#tti>?L!I)DuWA0aW?&VR9F|&>x&l)BZ)=0GRtQv2@BBgK7g>Big<My
zhLX|b)0?=eqhVR^tjwQ4@!T7fZ%~8nMAZ7{ae0h{Qy*{@j8%oKr4I~Eb6{@s=jAzs
z;O40lTA@lOy~3dJG*Wker`BMg#eN1fvh-(kfCcOf=4eO+3Qfdc*`9N;%a|7Yuv0!Z
zW%LHdi;sf^8uoeKVD(y0fzeF$QVzHtQ4Ki|w8@R>F_A<j+%#UTs}}ARfI@FT=p(u{
zFyuq?BF=SX0coZ!SNO(|!zB4$^;@mM5jf}0ge}0Yimi`cn%q|+^<4$}-9tsI@XmS9
z^3=IwX%AiO4c7=v-Y?*5gg%vKrh1&6brMU~_w^Zd#LNI1VaH>5RAeE58n*ZD-{Ic=
z`SqTy40q%<neXY1lmp$=pL-YCB}FK$gu0ir5g>oyic?CV=)DgkF1q^Bff}o|2He7G
zD5G?>NH9nz7CY43KH@@q%;t_L(e0e2+-aHbOMmSD!)c8nxSNo(WCxR-X#I2u$nH>A
zWc=spf>~%<TkrSw;-y%yzA%>E+h?_}iDt~D_zawccwXW;b~2W`G;kVWI<$e>aat|i
z2_xDVfO2z9-l*|s%@s1R%L-#X#<(?lfiP<)$89hK_u+OnhO}X?#KQ~7Kf1FMRo|O$
za`r@Ln;i&8qXv3QEn&MeOao^;rdz?+5neE#zux;)K#&E6o*5`5r~%jv1k$G15%MQd
z0cr3m!q8g1*z}|SMdAfHAvH#8z!P5^@32o?#xz~f&XW0A^@9F~DLvgOj3}e@Z1{<l
z=fA9k73ThZ{n!G@CFKCz6T{Ck^H_nSaI0<KUp;(fIGSXq#7ToK+OLE?f9@)PF+azy
zREk(PXYK@1M;`doRZ-9sItsLnDv4<8j1(^}n#^P27%O7pf#-DyEa_(mZe?tJF1jkC
z3L;UJX#mGAK#o+=#xIZs=@{7ob+h$h#zARm2vkp9e*l)E4LgIPzQAvp>@-yC_pSqe
zx&p%(OeX$yT(qBof{CSW0T~r+K<<#Ix^nRY+2(a>{?9}TUSy76J3HDv{-L7tvIl<p
zz$3Wd0HL5sH5*jkD+-+P6rL|Z2paQIL-8xu#k)zTtmW$@-fFxR7pdh!A8S0<TiD=4
ze@w=PtGxhM%D}P<T6tK7whFl9Hsn+Y<AR&3d(D{QPnx>Q5N*J=DxH{kqcyns9Z4R`
zf{2(Yx#WYo=`v6i(vxKn<uXSPt0FGTzkX3=-sWtZM*0#L6I4=Hcq0GP3_pyp2SCDc
zU`^Bq6q4Oi!xoH{{yEk{v&;i9;lVi$ZH0swK={qTN|4rEhLr<UMCqvcEzE8E)F%RG
z%EuUjC9zd3=mI<!_AegS5G<JBqK;e!^b~Sk+QGYlW;a2D!gD;0wjLU}o^tn#Z}{9h
zftFCsxmFZF(7;~j69dF&`a{fX(_22Am|fo1*fBuE%IuWW?ppzRy|uFvS{Tn-4I1j*
z^9ttwRX-<20SF^#P=`i%q6>_b*P+C{%uhj{2;5aX!p-1>L9+zlY`!SV1&_z~sbi^H
zA^%jrt^&ws@-MSHEWR01WT<e9O}LSE7*?e~gthR@D3;rXigIEfjG_S&4)ILf(-M>^
z7vY_J40>vx<q`4R4%eKI0s&nmGw7)VHf4)tjmMoWy8E=k>AbCEiqDA1y1SEP4NS+2
zI22=nBlALzsgi9aOU~rwcj}?Kkb!o|j6vR0lqGk2ExAvzAis3(Yl+?uywnB)=ifjQ
z>F|nJ1yM8d33U3%F<ga?FQ2nDp+qS1(e$UkT*rNGZZ1a@FGBi?G{}%X?Z<7X2XPO~
z7V4olt1^Qh0$V<0e3qCbSZzxR!~1d+?$nR7b3GhV%mg;lkM+yz*De)H+VDVGV>*ly
z4Y@wPkeTMA44f=-*Qu~**iuj(SDCCu1B|}E+dBYYwL31!on17Xy;0=*M{ybk=v}B8
z-=BSN4jph5QKum9k(+TRDVKD7zYKOF>MK<cjP_-y5L%~DL#AsTGmqJ3%3o!eLsRBe
z@aV_h`*u7SRLOtg7p@*QDsSTS093CZ>N!*ZE80imDA=tXr1JrC0<hMxtIk38cq<a?
zI>7)!V=2S`Q$+5=X8X-jpSDUBGK^%cMZ@Xba)94oL*V)l$-s-Z(nsft!Z*r1Ar`pf
zw;-dhy+SC`FYE`%NIBQ~#y#Edb<|ZdryaWEph%z=`8sXAi*Q*i5{fLJok<?cD@8mc
zFSBu`?30z&+cnoCk@sJF1%f-J@K&JBC!p<z@AP3XU1=EZ=SW8h87O9cReXwwh@y0U
z2`=r@G^}!vG$*S4vc||Yd??bj2LuGQ`0c}i0gYjJ4C?DRg(*)6JY@q|2a(e^Z_b9=
zhiyY#1A-=#ZSHr{hbpd_jAe!ktR5DS&-T(#wAHY^P80ux68G>Ng39j5rPG6ESj`h+
zH#eMzloI3w5!|~{<Jh-_OmyDe_85crU%=AV>AhNN$R`^+b441DI|@um$Z1E=k#B=B
zPRgb=2nIJSX+ABphNV--87MqXRcsZTt^yh&iEg=sreH4WNPbUZ85eJy%?Mp_%8af?
zag=68FE}Q%3F&|pviI_&-PdH)UJnmKaIB;pAjjr9t*h<AboJEIny|R!R`(?G8+1wE
zbiz=NS<(%~Cn{J1_3BT0Yp2gjymx*#K;lsTR89~II=+l%W7Op3d4iUoEh2beIC931
ztb?_}1h$;YUH&$pcM=~iO4)FDzun63C!G~9Wi_CetZ4qTMC?Awi0FMFv(j7*Ctk+@
z-ed91w>)x;uh72LYJ#nzel`+djYklKOcj0tuf(~7{ZMJm5y*QY#}>G~-!A%;PWuWQ
ziNTn6Hi1wR>5?vSg9s<>BGGZ6Fa4lw!p8TH$kMxFG#71XB<>8TSk~-!0Gwkx&3G*;
z_|f9mO@Ulw8G?+~;lsTpV0bW5FXgqoCcD~UG<C)QGj$V(oL^ONjat4i&hLvA%JR*;
zxOevInD@imn*&3}V?C||<ozcNK@_Gnu15!_ms<nG32hH<P8l5sfPU)vtplpdARhs<
zz~QX@gQz}h@dJhB1A-wc>(zcJPosdu&cDm3lq$7_rpOE^ABZ<{X|7FpOHX--U|DJP
z{K21OfPD+&c&X*#U3<kI4Z4+n`fHDYrQ{KJ9ZPXV`$|*HBh#q4$oNe?l+y>|p#eV8
zS`j)I(2xv>zdXvAMLHr$vc4!P8v$2V97iw^;V0FM0zGO!BBwAgEAf029c=@r7^u^w
zNw@z<;bUUI*0~SZ$DUWt>G(f3yJ*v=hc{KHS5Mp8_tAPuX$5E{ndoo(vZI+kATwgn
ziFRbWpmfnY^`!Dx1eE*LWMeDTC|Y>cMfDE64<7dd0qUnqr+(VWN$CM&?Zk0}^fxLf
z7|W-X>6~NmJ0jFDwMP2)vG;-a)ogU*c0FHDL$53A6}zc}IrzyFt$phc*1*B|OcFa|
zaO8}eA5>32%nW{%ouqAj#sw`dRr>>8{1@aGj3bD>dahc*>RUnl-~u$1V_^aHRqm@I
zOCVRTj0ZJ(1GNM)Q$h={h@$?6j}McND8`Avsed;;L`?`Vqt?xD<7?NPA>}h(2@QUM
z8N{K!m2gTQqePbpd4=cX>mI2p!V=0SP3IVYZa5+jR$rJkdKZI22n{)v!YK1Ley|J5
z>WyVx1?Fder@Q3f2W}%bbUi(Y`a^F`Qhh_JWby?6j5UP3SH>zTPs8L$m#z#MByyas
z>@PfD9(y@ND@yVU_Y))ET@53~6h{RDlm{1v15ZPC^ZE-Z@Ld5Z*~xK!cjpIWGqAU)
z?2T*;+N}Z~xkactv@ZpYLlRtTHE;tBzoo=nk^FK}pAwuUfg7Yg{wd_F@}+3~_UUtB
zJ->VSA1dh={xP_ajn}D=R?$I;s+f&DEfkrAAwcn;@abigE~=jw)Bvx~`IqlVJ6g{{
zJD$!xIApA62h23erS0TiseEoU^Kgw=D*~sDND+0RIEvu==cjHIrGHOi;VE~#WXX9q
zNUU_X--rBSlGNdgC4q;>7`8mBPDLc@Li^$hs_M1y^@`(BGt(Yhq54-tH4@7-h3P*T
z^61Ck@q51f>J_swsB%vo_<2P*`P+}ErY>g4M%nSOfd0eibqui6T^xhH0@P5vSA+J^
ztCq4`IZ@xG{mRt0291;I*wBwNZ-<Pn_`v^g<IOYuTWJ5txu0^G9vsYCkyOVyX3G<$
z_@^EL5vzL4FjnQ4jkJZh82iYL&gt_p;em&+UM@J(sDB*7sZ!0&;0C`1%P66wm7;eV
zbR*so^#>+v-rh?NmbCAPk%!<28e;E*-%vt|m~RhVI_o%jF^1o)VU`=Q6VV%OK3d6=
zB?{s<o8M9Qv@A$kvfpYaZs{Ki=4_R?IQC`UlnjYSNfKU0C(hqK*Q>vm927*2tz~7n
zaL{r0U)0AJXPS>GtX+zKviQ;Kp4#X6mR-WU9!H*Lko(h5hhDO+)n4zM*-9BOrMy~J
zbDJ_tI^q}~1}I>GC?}+P%+OZvQsbQBbRXodz<_xD3s;NR&y8dbwr{9~9$*KM<SjLz
zS~HLUWH;XDQu(E(+tbc~XY3%_s6R(5U2vkD0{%8ra4i<iL;>MVq=#ba@u4A_Y14hM
z=AFsBIiTTeg~!B2plH=emlG`=dVJKWOTfL3_@6n*E)Y!7E6=n$zwpRLA=b!f>lI_d
zZ<s7?1wXi3w0ke+!PSzEtv6i%asOBP*_LT1#cpk($@f1_<o$&vs*5mPp=D_CHRYPR
zmax68nizgHp5mik@;9$L2q@n?d3h$gASifNc{gT5tsHA-f9#8T7Ti_gL>kPG1?0{5
z@u%PSAGU}i-`~ruM`<^l()}QnuLZg<gl>~s1u?If7npiN13}58XjT-OrF?tiEX|+*
z69EQ<R!_{WUn{VjVIIa^(gUHvbcj9e+;@Q9JAerSZT@S0xfv|b!5_-3zX49b*uucc
z>}F(;F}5p~LmRnzpi+5AnrhD^56LRxg!>S818fxJ$M~X@CLKYSjK12vKWPeZze>z2
z=!Fy2;bo-~V(EczaUcdYslt_%UPfZ$JQz3x^k5UT3#<{42k;zfv<(kfv#ddHb*IwQ
zS!xTSb-IrZK*RplRiG0=4?Yna%@45)Yx+Etq-l4|oS5FIyyZc7@ePce?q^d%Cbo4C
z%`>(pE`n0ORW<a2E^yjngm}7gxH{)Gjw?|2c<h4w*7xfNx=94;6Fke%aLI=p%|{z+
z7A$Jchi#K6dv2Ops*gTsjtyFrD)*F#$bUnY8y0?L<}=Cp+VWVC3NUl`qtXGOTVbUX
z3<Kvc=%ef3dy@oz02w|%l<50R3spKmxrvjaeJ_TBQ-V~(D(_L=JC_r7=FPxm9j2IE
zCD>-S3+?fp()aSgr2|o|1}#o5k{$^MehJ0Tp`Q-IpnwStb{nY*ewY!m|NQz3!}8BK
zE2oN*rps@Rzp1b?X4YNcW1Kc>fM1n{S;*vb$|aWfw?@5<e`Pl)Gg0Q@bbM-IFo6<F
z6zUCVJ3k~q{e$LCC<Hh?z4irM^1Ujzz}FAzf#JMZc_D@YuGViEYSvF&4ZRIg9_ezn
zBD3mD@1pBtfHv-R|Eu|k9BzF???0t@t2G9Ai5f>Cvxe)uxgiiYm)NrIAOpd+5FWm*
z%8({?g~e%FwMd9}moEr^>~I2rQfhKaoRZinP%IE-FWhtDo-oRXAfIc8a{r-y{NA7I
zTpo}`vmW9P8D^gck%(Ulugp*=kCg%g7a<@$MikiO@P2$fs2>WoAN~Uoip+3<#Cf!l
z1z5-P(W##VUd&d;yGrW|>g*~8!%*%f0-kP$_uqf_KN5{#PDaq4c0cAHo#d(lp8=?k
zQgHz{8jUOO5(FT79VCCy3)v3Mm0BXtHeu!gA&Tn*n}nAhn7K2cb_$wNRF{**LVov^
z9Io-0IoG{RBlP3?zbC1|92-6fF?f&6pO;$Sp`P&jbl)I@kH0|obb`YQm|_d4>Y$lr
z1=QmW4mE((SKd04NwuC$@s(2g>4VL$pUxjZ!AnjAyiQEeVn;uqr?$tU|Byy?AvC9E
zi<&<ZjcdO^r7h(IaG1D#forZkyo4EEkI%`*0O_1-`a`mO4DarJR`FSYN%SvO7g^H`
z?tON&C*X0iW#smx<;=+^QE7&_bAz2@8H7wfu5JZTU7Kq94MHgh%~(L{eN`T`cE_JM
zQI41-eFQmmV97Ex`ydec$G{+9UbhT0zCPew42#^%|9yb%O2q#HA3g5|1w>`eu=ZJ1
z^=hVxwMmZp8+SyLB2Err6XpI9OrE$h>qhSW8E)BOzL>Y5I1M~0G_)S@@S-uV7JoeF
zCLoP)N8e}TC|5d(9~4NxAu6)we-$94X{f<~Y?U~(D%t=(z!iQ9GH2s@)$Gy~r$M>F
z-|GXpV3+W!*Znk5(Gom_v+S$BBO7^wtRm<01fJwfjl#$hb~_h#9)nObI$?=xzz1|l
zs1xkmdKF`Q+n);L*-HE7!n14nB=G)DL!*2<l_vz#1ZNM7AmKT)dZtih1u-f|Jv%_0
zk+pQ`FWgiB4;GTwAbEQTZX}C8Ykwy~nl(5nm0ATbAV%<kR@&hHNRir_Ng7OEnW-qb
z?=N?4X$DZH>!XgTf5Eg-@*$PofGI{4v2JbxJ3ni5cR&CHI*=4Tdn|S*;q}zJZvKN?
zI~E+BThJGEm?<%9c`+G8O18|;LR$ncF3mdc3dnhb85pGhd~ab8AYhos{{H<1h|hW-
zi?a{yO(M0V3N;5i#w2Ragz3R(pBCP#g|6u_bQD`)O+G5+YT&E>1u64j;M76m^*fq_
zDn|j~64uw{e3Wc@LNcgB61>J;eJ^jMpE~_d7P0}mVVkmTQn3Jo+(k|YnB6ffupR`P
zcyjMxa4K_RQJt>c<6_tafM6goL)SMg+?vM8EGjKY<J5D5-_dyio`Ik%fI#GQIsbJC
z52`?#|F|{~A0Xn;kFovG`Mv%b&EW@Y#pxTMU*fI!0&nJNYd^I`LNM4Gu0f1}`vO|l
zl>Y0O-eT+AKSJh01hEu!yE%p0?ErvCYX_yuNR-B+Tzp|6Tn4h0k8et5P-Nr&ZvWyZ
z;g!i~tRCp8oqA6qU;bU9IKX|HQTMVJOxyrC#eZ_`sEmgBCayIk+MY?FH;Nmp@zy_i
z@Hh%-ho25+m>5ZZD)8!Xl{fS!s){)PGx5l3UvK?x2>^r}q@t{lVDCcI#~SB=@3;VP
zV1N#z?Je<nKG<eQo&`(k^<%aEZJ%1W_D+EnhwOkFYNA!p;~co8Y9ba41+k}?|L=!k
zC>8r(+_VQ_Fq5dw7pNC;l)>DjNF5fDJB)H}B{q_WY_~5kW#iB&YN+4`XjxYQt?V_J
z+fCw0w1K2OG%3PYX<KRfBQ{0E+$o*@Wh}=-Yl9H(0q=&r9t2^aMlT;4XEokKZKs1^
z#olEjS*X=yjT{gzslC8za$f&ndb!;6@jO5{5(=z6(`SM~a{eHv8$d}bfbp)y7W{k6
zM-(ChujzQZQH%CGrRyvd=Uot31)!B)!?pRwcPk^oRXLjTbb@<Li7hv=>fAYA!J;h0
zK=8AOdL#+`nM$0k*s5+h&}lmfwPsL7j2O<2^^JO3l)&r+MUD&_ub5wD^v8HShAnRt
zCnevJd+`pn5lDd!kXsGVjd2icF_nbb{5==u3leMeF6JRR=7UlB2JB>j{LxD#Hn^2<
zugsIYolwOC5lf5k5-TPE(O?zI?6{sp4Ad};p(pnbGID;RUEyMN9Xe}x5M68mW;|58
z3J^i&`CGmKP3=CG{k_ZOm#12hwuwqkh^o{&Gu{wxDR*3_pZfyyMGtpl9<#v2QnW?n
zw&A8!uLC@AK&gi(5f_@&3ieC2Kn}gn`AEu1Y`qhlMV}Gdd#_>3X?Zb@I;plh*+r)i
z{hQMmO3Xc||FX92SeoodV$9+-WhBvQ<5cuN{`bd*7ERBz0!3F>_$M+a(~Xb(E3q=s
z@FGt%`~zs%;;59jC941G;G;b*m$r({@a<w~(ha@FUJ!@+&v86lM9YKMjaIj%;rN4X
zg5L%IRKq`N*2Y;Mi_w~{{}l6yD}Atza`@sbJYCr6*4-yJ<|)q_ZttG@!N8Tk^+WUS
zV#?c~gvJD}vz?c?^n2cAE|qkfC^aMqoHhy<5NasuOG*k);IhvRUHtXOtYvfZ*YZN<
zLgtT5+waKjsfE~u(aYbS5C16}HiQ3{|N7&1891EZ*b}z3ET$H42kyR1|L)3iDGWrS
zHLJ__vbF%zgvTWl#$%O4-KEJF@^?o9$=;tV$<RE^?2D@t2rz2g^A03HPWWsmq}Mu8
zx_}B`IXAO{Ri1U6`KiF9(TDwl3N}BHrS}<U-9@dNck<N|QN$;y=&=w?&(rj%Qq%h$
zryzePY>m<lYl1lwlyr{6C1C9<JQ`9f6adO57D2s^OqUO3Y)K>8hhA)ZOhpg9Iu$ku
z2R0cP9Oz8s1}9(7MWiS`S2AP{-eQ#YU+4M3PDKVx!*wx++T9lsVFj0O6eVO(&V_OH
z$=%Ko(VHNPjhm(Z&ZJN$LojU(H2C0yS;!h$cw+by&g_KJG&I&eiPHa49YGfj@pIQe
z(+9(ARc;$==$`am=EbiI#gC=rAW#m9nr=wG>TL<VOBfTlrMV(w=I*eOl<U7(!R29s
z;71yLPBzl|E;&fsi9qoDzrWk|o2;cf`Vt(zA~xR~fG-15h}zl^dy*si+aODU(?jF(
z-FvN|mIrCQdVLMbv(4DRbieCmW{q|ToCJ(??YLb_W^c{$XBiO5E|(zZef1AosheS2
z1J$>UQ|E7W!X<`AQzL+{H$ekClDN6_KI(6YQ(s!!^h4l8#3a;;dm-EVhOZ5DlV{hI
z@-zw!Ci6<&pcf%GS6{qGKQLMP4E=en3^oUpck)d(okSmlDh>z~xL>|NsPvbyC#4Le
zOF?Jpd_{_Ma@*le7k)7==Ji8ig4>@QHu|Bl?+t<tchRn=F(7pap#pPX<)hmT?Iq?O
z81OYgc4w^Ioe;*s^eP-!a2+9HW?pZKxypQHxROI7Aaa+$(j0eq?Kq&j)^=Z6&*ROa
z!Hu}z@#wa8f&WK%CRe3C-BDq#rmEs;fzdodJAJ-WJ(S|)VEXC5A6>aaby=mZB4Y-P
z{eCy1T`yl)6D9qI=W8p22sA`*^2uL78!fy_j(Pybhd6(_uE&)y;DhO+y*;_5lp}U*
zCa9}of4x1RE;5lj;wmzLz~t2roPWA_wded*F0}+@ik}R16ntg3L0k+f_{RXVLsF1{
z3#NE_4(|tQC+FKPQAoltMqcn7@^1%xp4CDZo5tbTH^JjK$YQEVmIYmd9{{>l1VtqH
zBwPBl!rjidXoQG`?Ns!>8mUt+;glZ*g`4;n2TsOT{@0Z-`$={g!NWOPZGV7%jAG6`
zZTwb2dLaR=??mjZIt4FYVJJM?$$(^JDkrLoaJgL3IH;<0V9U;)%>1({NSYtdywDz1
zAadO{Pc=W309_eH9SJKUQ9wS|bw1ee%7+R@p?fr52PgydCb+`$SD7h4Jlc}>VQVgN
zP!xJ$@*KRdPzG#SuF*IQRltNV?i~w@CZH{%kHaLAVI0NsEM0NNxD8%DTi5aliIEMz
zb%}Eo`p2+lai4xd5_@bcWD^u>B^&C)g`dheQtx?!U`NN!$senlIe;Pjl+whQaI2IH
z(LBaMI!--F)SWZp@C}TTF!we5S~>Mh&%NjSOfGef39RfdrGRJb9cLAseCR>(3dF{4
z+B4GZPl3m?DRL~1)Vz!26YD`#+`m&Gbp3RMWLSaj(IN27C#t${to=mw)U<K8Cz?wh
z;QZn#*2+zjXvBoq?MOV2OL5(G-G;BZY!iqSojPE7`8L1{#bEn8vO69ssS!dHf&$pO
zjRn7xOPPd@M~=wsDUfQ^#@_wK)&Xf2Oc%fS0*fMpJWi1+N>nS;V>xBj#~nF2CSY{#
z)>#fkN9yn!8QKX^L8qu!ALU-nx;<LcQ=U{y>iJwbn5bz3a&WgW=QB!=mCkuDGW?LK
zRi*45{;*#LCH;fnVynCW`G{(``}C|_A^1Vsf;V}oUyE#*vS>BrC`Ax*PYsRI5;(Uy
z_a8<qh<ttL!$X3jI4H;9p37)+`|JJ8fE4I@={*U%qmToa{&I;z06D1YRrDOH&>WX#
zC?DmJy00$+m|%O?WzSa=&qmeF=O*hC=fpXK)KND18Q8wR`7;aX1g#~VzXMr<B{|u?
zr)6L~#3d5pMi@mp0>#J3Q0tJ7&>c8E72Fm)rlK?r00`c)GU@~`F^Bsr%R6O1GM-1u
zx;emzOpjI?sJ+TTS<HABAW?xbEwnt7#OX#Q)Z!+aC>cww5h8N+L80C`UC$`JhM)S+
zSCks~A4nY$lC=ylRqFg0{H<CqD}D)_%M2$lfT{C8lHZILM_!#`ek^F&DES^eqM6Ar
zl?<LW$yO^p2+>M3bl$$F1?cz$+{r2rlHDtIfk6q-KWwN3AMH!*$PuP6yv8IHs%qmO
zkWcgmlKGxp7-#fZ=E*5~qQ#f5YM#W$!utpt+?Ou#ee6k;Npk;KP0g#X*MT+3Dooll
z{UEykiFh^Wg>`ctU$eDOg}5O2PA}D~>4A*hfcP;}Yam(0xvD@t54E9#=cMD_-}L15
z0Q9>Cj};vkxNEYiJiH5&I^mr=gyrp{qHWEUt|-{%^qLY6(1}VjRv06voiMzMb}Vtu
z813^vncS&qHZq)75HOboY^`o1x5W;&z={^JKn1$jfxK<*gX!M5>5gI83t$y8+B6aP
z1&yNuj%)&Qzc-bRMG?DD6DK5bzhZl%a&E7e#o#1?8-Wpm3CvrM0kXu0U4}&}FjoS-
zsEtuK#y9hLh*B|4m9S!du*95rzQFSfVYv)eu43h&`=urqoeGb<G0beM0v8TM4szAb
zz<KKqUMQK?3f66f)Zf&Xm-1CjKRk9nt0mu&-SH*4^L$y`MT1Sy_PA|7fi(eETY3v@
zH<sQBwhN&#udrykli4_Q`h}ns#lOv`Eeg~vs1F+7%a|u(pxVUvBeD!h)5-nv#VsET
zV;?jk`8Nv2e$skkmD@ZLh;hY6zr4R@JjlD*XyUEVM7;xKAAF6nZAUZ1j^ZcF%=KSv
zlz(XPL$BjR$GIO$Rg?fu!N5zo?Bib=wLYIe1rJCKs8>GeN3<XGO38nfU={{fi|i%j
zJwyZNe65yWmPcY7rE{^0rIC73ljjy7R_thLDeCD!v(2RG@1wyYLt#PuV}Os3KA_N$
z3Hf4UrcLiWJ9Yb)WpU(SLE%AeUWN;X_-efWv8VDIQ!v(9fQ6IN*DIc;>?>OIDKOBL
zSAdcgCT}(eEKg7?P+-U6$dz0;XKEP)do&BgLaJPtpZ$SNCbsSC(9H+AP?;b03;>+k
zznUdG@i(v%w_%b%#UF3fvMqweKct;?Kc1olcu^zhy3PJeIty<<XRF<joxC^=3$~7&
zIZ1q-gz9jO)C;1B-q=RpZ?3dvU*0%BtWLAnFsj|XZhAkaa0Nk?O2OXL6CkQIGryFY
z;1gU+Vxc*<hj;$M+2n`=2&1^eeG@KG2(M7E@TqE9wYt=FLfoqX;-_+dwoaBo_dbP}
zOHZ&<3+BbEb8F5(3Wr&eQpcn)80X%dwvEaSI;}f2bYBaze4C~CO|e9xsy-wlzspx+
z{EAe-Kjmz^l&*&H<p0Wyo$53Q`=e~HwbjT#$JiglnruIVIs<J|8!lL8JFmE(P}>`%
z6sV!v$@&#$6z@l<DJ5X$&N2R~)P<}1ho-~`+l8O(ptsOtE{IAbnlU_sCS;9$8}L_j
zONXLD^$A%nB=R&_Nd2jWB?w9I(^-L_Kfq5CVxtC3;>&5HX4{PX_XPD^IuPbnPv+ce
zH8i#)z2(^Xiozqk0gyHPdiD75^-oC+Zc(#XqtU%th#NP3^8-LhkHv-Fv;eJ_I0*BT
zQc{+kW2UqxCNB-1PBMS{wqU0#2jb{+<$Vb$noDzv{p_{nFHfEbe61XBBHDm$0*>xS
zh}^jz_7@F%1>Ztf3OI)Nj0wqmc2#bMW0ZDpsxO}O%@dBs%h06ymz6;1Je%(V>xIAo
zZ-2gDqApPhz19WLRD}~auz==1@msT7d6OcN(Gw;nI?}?>8jBn--1xoJUyEZnkR{0L
zsmvsCT13vJ5JKLxY2>rcuopDmK~uHEiP}ltd0yI8I$&?dvUAe_(()0N$=F@bl}NCX
zYAer~MQ;dBBP=YuyFd6a%gXEXD^m7IB!MM)*eAqNe!cqybf)(W6{vev#>k<$Mha*#
zWI*gOiTv7K@AhkRXduCL;gcPs99O~5DlRJg`SFX)dXSn3#&R)Gn77PWun22DF(3su
zZA?mGaYDgt^B{qM``sTx75GVDAjBCsL@Jvrxt*eA{dn_hIXtrbq~pwcM!jY(RLGJO
zqV3?LLqM$^Kz^Ps;VyMXe8q@ip>jr%_vrIQ)^+UF3{aF~iv!~vHbPkr1J9;`oApP(
zPjqEgQo2v=d$z<Q8=^1Iwj8aB7JyK1xUv5lOTF)AfD-r;OlqAPRS$z0e4m55Z-lYj
zq?&vBiG2L~j*E$t{$&JlDMjAD5J?RT!w2<u!F!wnLd=JygDbM!Jfd(34jz%X{ujEh
zhOD<Kl<7<-jiQIwgOn-*43n-@)+WuNDze~Rt}2)p&|Vl!(pwMTdG*^)X*tiJ7HdvR
zYZB*itjFub5`p73lZpIeM5ec_@eL~ZwCe2#&sQ%+Y~5wnCK!_X&F1>EGIXq$BanHy
zg8$j-Dn`v9hkN`QY)SlmU+sO^*2+E4d%d^k{bZO}C-8|j`HnVjd7>flT%ER1o~x2-
zP^TBI%84eyo>a!y><7)=qA+u)>;jaSZKFZnx9Jio{s}7a(*BrbYFGxPXPfH-5Y5>X
zCaoO7bVeP1I$`<TN-*UCQVE-ZVZc1ZD%N^6*=|+Akm>v1n@v+8341`UA_Iu;sK&3X
zvQL01l_9nB6H~9yayTJ2Gcm2OmZTP1NuDIxK*S$vlYHioAQ{13b?etPyDYwv+L5Fm
z0H+cb+Sv|Cxlc6#^2496>lUf@+`^-LeMPx&%~^EzI|3Qkvd7Pz)W)ld1p|##>o$5F
z#D%+CFE0=*ydQK{pJ;=-nNKtAEd;=T_ip_5&|s5B#DCC(3xQSASqCh!@~}G<80CC_
z1q((XNnR{HkdVil<-a_BYZ*YV@GD<DUs-tW)rXt06!HbofBeXvNRVd1iXkSvH%Y?R
z^^RxrzkPJ+$Txw@?j+0^FsZd=lkM@V+>|K~IBqWH!g-qc_L=ZILfausK=G&cTW+IH
zxC&PYYdm{~V81Ymx^$pYtwH+c`RLBKCj9rtwC6e_mD7Mi)uuiHSK^t6#jx4t{J_bV
zgDn091w5B0sB=W@kUKBx59eXf*6tFgqamry_fSvfIj{cAbCZ+K30l}euAtBTfFeJE
zeFtYAsVxkKW{Lvo^H5QNH#D7Q&!|^ReN@&rTAH@0RIHm{+>PN{zd$tbjITs#O#YbH
z$irCH?fY74*hKKSOdS7%d+ZwIuGFAf(%R?J2@>#67HUDqI>h5ZO?8@Hyn;zbF@mT}
zZL_;Om#AG+o}pntVk?hb*$WfLZ$(kqvgr&Nl$qeIKk<}AGTbxqyLgwD>w{EeE@ZTR
zZw=lDy0fdG<b8Snh)TYk@63*d!>PZiSL0L7G-|q%DOzmr9$9964DZHpNKym(03tl8
zt?3e%O^{pL%lCeYayK*qSGG(@wkS4e#C1fp0@VFRnFT(+y~++)Q9Kd$<v{TD*MwDc
z;g=;WBL!?8M4w&HSWIY`1jJy1hew$Qud5uHQYUa?!@IH)mvgA=b+YrzWtQdA$YL}J
zzy%MF#DO>8Ymxq~GvD47>!UtV8Sc!b(a$uz1AvZ!LT`4Qrqu4gg9iT_$UjXmjBo4f
z;tbXP1Fs5lDx!xxBM7_`b-PYVlh9y>d%#9OMzMb9L6+Z5>4xX=(i*MU6Po*3(N+ww
zW*tKHj}uRn)oQPQR;TkWe!Ecu#4du~4Uvpu+Y3VfY(kzX0FLqXU2)>H;wRlq_>6K&
z<esYsb#8(*_om^^9?%moP@@Y#NY@+qD?9HY&|Z!(+UDO+2=9CfuliOEA3<gmn+-%G
zx~#!^_m6SSe%;^y{StTvKVWV~#%p1%9DX8CgEV~#Gl;PKB`NNbrZc_5f1oi=pq!-M
zJJDM~6v3ss1dx4P@%y=*BU7SCpawzFg@}SuXS76v^b7~K-F*X-MbDte8e}Ox65fhX
zgW>wh&eq!Q--;9XRbk5ZFV)+zR(QGbplbXM>@SVHr2LBXKSbQ#eT8&FMsD?=l16TZ
z$X56Qb(0z$<3$S^hLKtq#+k+NrF9@YZ$52&I_V1mjzdi+Au#pj4JR1ys$abh`^^v=
zG}&e;kcMWIqAzOi&#EM)OO)JX071SB8*YZaN5H%acKsHT)yCVLT!U$57Z?zY6ydd`
zX=L}rGPHNk0jP!^;h(l^<ZV-MaLeikz?v%ginAr$o0F@a`wgY=tLKA31A1+|?)9(t
z2K)bm6XaBgW880~(y=k8DANyjrdEFSx&Rm^jBx;OSdlS|$47pyQ!$Hy)d=LBUm-m5
zj@8UpMpVoOK*}IAU`W6k1ZRH#)K~|EEmRr#ikuQYOf!E_d+MZDfb}KA|D^e?URbyu
zEfUqHw+MhQHV0Tf+y+{A^zfnc2NPG(?+q7V#o=>BL4seS>wv2ToT17Hjnb>TGAt^<
z`90G{|1n(jH&v{_^-s<10R>@N+U<4?a`{b6RzJj7F&YF!N6sjCLFhoIfP~{7<M@xO
zwf7C+NYKZJ{J4j>elC8XYCcyypHQ{NPt3W}WISsfqwXOX3?#Ls&wE2sB$I;e2Kc&i
z#jY3U*oOPSaU;B&2bU4eC$onZ?S$Jve7Zb?+Xga_(0JMnl}<!C_W1qkv<pjQvv8st
zH41ahSU&_n9G=NSP-{bklCqmdg!MIS_b1a6<K!*8Q9m4KD;}coeyi?*{9)+j0Ay(z
zsFvby6L_w7c#4Dug)qBZ>@I%#hQY&Ih#A$uQ~^0qY~uC)ccxbinCowCx@ley(uJN9
zf%Y&@Ze5i1!#J#=Z9DPdnPIr<I(!h~?s7Lb1ZZjiQR5}wvJ6}UyLKXP)TWq^|4Fr`
zIfm|s7cR&b&JQ*ETp@3xTaSS(yc8oDpBh;r0Iz^{IR0aaEPBLSA7bVIZs^Zy*=eI$
z%|HR2GOsIxc?7onXbvM4me2KUKeFCm`ZjdZQtLWjI7P`7syr<7BRdOaiYrO*b-+Nq
z>JF1&zbn+C3Q_d=y*zZrnJ`k{yH%W8`69FgbaeGs?gPdIg6ki4hJXz`_FB-YwX}i=
zokASt-Is7}KAUhjq<>5v2%-)3JOSr;V>9*ZkwG`8k0R0@L5CH+zgiV5{{K+FgXwI8
z5M}ssUIaQFAzi^WX<vjCkSCIGHGW$*g9kbxRR&7mi!ZvEOUZ7RFUXaq`NXB=%o=JM
zV8sqx><QsGz(>l~XNJ2Vi}hsFKCsj#U)>^Oax#=S6|hy!-VNWrL7d55I^;-~Vy!s-
z5<_Y<^QR0pQkU4`HK;ApglWS~P_f6@Q;NT4=N*6|@DJt&cEkAOv)?sf8{r5@TI>qy
ze^coHf|0M@v|Y?G@v3fGjyD>0{)RdLoEhj=xDmQ<;E}H``2#WYtKpscW9kFJw{S;6
z)(EIJSPWkx5=X<LZecZy3GMDJw>R&o;kCTb7F)gw%nXiGX)DhgiP!|vH018sJXLr)
z+FoYSkf7=%rqB>u7v2IU2hK54jR484qLvg-82SlEGoH|T=3Z)n>UZ@IKY_2wUxZX!
zDx4041swW~Y&p55)(PZrES2JR5X{!Ae&m5vRFxE%0l6Lp)o*1}m~pY+ggxP#&s}gV
zLtw_(99d5}-9BpKckU8U(a>l?LImYF)wxrvb~GtUFv*_3FnILb^?K_-zG~2oomS`G
z+;Rq%vbr=vsbcm0io2|uGagoiVU<0AOs&&P2n>>P2Cq~!t6cr><e!Vkz@2@Sp-(+B
zHbuVB2~om<wbMcwMOMo$xp8Sn)qpGS%qj1ZJcUjM0?5U5q3)F_HbIjDNdX@yev<OB
zbl_4GTcK9wG*6zk)mqo5G7$Rt8scZg3ciDq_YW@CH40=58e)9H`OSYf&B;Y}hdp+-
zJU==|ZN(dv8i^qO&-uy{;0Qm5xa%Y(rIP-Pv!~8S3N*oTM;DW`U%BxfY?OyxqrzP-
zu^yD=uO+dDP61+;t<(9(N!ivNa}0*7`bdgbhZPSusOK}ZuWH3~UFg`KE&Vy8+R`&h
zZr~%2LkMo#?v-%N-y0Fhy=QoL`qL-<oga5$ZP)F^w~!pX!aOwcxFU29pI8`aCM278
zI%&J5CS8)A9c}6ZwdHOUW5nri=Lv<hY#WlmCZGtzA)3_D$AN3V&&PhkwK#HOcRW6{
z4aa&_x9YK|+GSadc@j+Q*CXU|JL9hEI?KzP1mwvRGgTuZr2mLk;`HSJ?poCjk#<B`
z{u6OFYajV^JWsw;(?NCOzjY-I3J2yHHq3HhnE9y|k7pgr82jB4dusw`H;MSn{5?;2
zjn8uVzU5r8WW_Ho5%;sA_%rLGTT$f|v{EI;l`E6afgBZm4vhlv_am+tfD;O-{S?g|
z<_Uah@mE~nQHR)(4{o|0NOXACyjZNA&k{^Q2dmtD%6DMRw^HjxY?Gt8!v~E2-)0!_
zzs14!u+!kGfoo-ianir<8C;Cto14dq)y*9Wb3YmIx2}f1!L{iKUnB%B)QG76yf!oX
z2jr-(K;I~Bwr%$rV8*#C<r+P*UQiuyiKy?0nY6;o<~*K8##sT1;huKEI7f|?cNrQI
z->-va|42ORSDch;^ch-G`mqM}ehOQL3<ix42dVDfD_XCXSl6K)4r2dO3eWl#>YTey
zwYI*WB6u(I+4Xfp=Q~Cw9elUQ@*avtHboxn5xAc`JGQTu)U5lrm+|$JNzD-6dlW3o
zP0nKn^pQ~(-a-gG5yriJ2uP+&m<zx24dLteZym=5H-CAQAtp@DhrxX<kl`11o(RRh
zU{EN?jXm>ui1Ul4xO-i|!;)9>X3CL&e;5waKD86}#(kI<;>v0VQ7rREhFdb5<<Pm+
z6)r+97(7{|ps=&P_oEH3dA`24za2n$hOY>F=UY&z-McJ)G8_y#$3TG`M7gg>)j}gy
zZQhOR530T3Db2skeN}fbgNUh*p*Ui8<lB?@KC?ewSvdq%Y&+*$2sCFBb-F^m>3rx@
z%$#+*!*_U*O6H*zDGd)Dk0b~593@Pf?fz=nos2KT;#c`8Ye~+kllo6l5DsyumZxRc
z6f=phGniY!zZaxF@mPwFF)riAAwJgr-v2;aHvXUnUcUQmZ}`Q}F-Ro+4Hqnh>L~C5
z>`!8zvmN60UFiNAag0M%5+HtDUQ`mEJ>RBGF8fjoN5souNBmQJN~)dVIasQFW_{%k
zpeZOy3vkGXKBVKd6zu6)ftE>x4!_orzFjS08I0L)Ay9Wl(;KB=w(!&Ce}VCjwEicU
z`~#NJZNdnjeuH+C!$u_}siav5eX>+BW$ZQ0$=w7A6C?P({+^u|Zorf$cXRcQFQS8K
zzY}vzb`w-a^||TV*>e8j+^;^;4yw+6gZ`Hn-yC^arrVH$;uHoABk+}?v^$U^_CWr&
zjFxx>i$ObYEJdkghrlJP-!mzi_1qX%pjJQVwT0fg++O)fMWq#gCZQ@-lgT7tIMI2i
zH@X7W<?7Ln9wEIfRtv(%de90;%jMqm^*@4tKzBr7Ua`AJ$Qw{iOVKKHqam^?N||(c
z^Qv}}9^Q`h*!YQY{%77zn590KXQ=QXzjH2d2QETO^(|RQ`g0QhgTx3urRiP9)<sX)
zF_J7~NWkM&6Dfo`;eI^#TK}4cL(IKlbOHFnd@^r#aoW)%$t)+og`VMgOORoA`}=mB
z4=>2l2_L*BnL@}6J8=;SlA3~+<eAEGZH>XYtLBdwg=H_mG@V@$9tuOB$2m%<;1*g^
zM7b$uNFsSNE=<{JO5NHWelWbA$s_$~GcX_Jn+dfMe8j%PXHwL?i-q2tz~U-D)df{@
zb>s1$;9<V;SW0;9AV@IZ9Hbk!&53w%JkqBMxFvS&#wY|BL8bSzw@aYW+3fzcCaMJ-
zWCpdWIn#>|3g)4w*~8#a*r%1(?BS&7oUxxGVrUqscZm}7^ur=%OxluPOAjBKi<s{;
zO6=1m*vv<q6&(1v7Ubx{^xi#*&41y{l!E``N{fbdP;Z6@Q^nGJ?_Qx)8{m0x`<Wdt
zawYP{0341QX7-9N)w66)@-|>Dlpinwj-PE_q)`*AmlAaE%m+{8$a@H~yYdQDDB~xO
z7&USf?=?Y*FiE@EMflr6wFkb!p(gA%xCU5$a1%0f-J%Gz7M;BJTcMYYIM{0R3|7#Y
z!xA6c!)2BR0Rcp*Vqrt%(d>vahRcO+dMISHo6kr}|8TipLF6drgpumS1usaqh7?u0
zk7tbnQI8=&YWJG0F2nG_$vI9Q@!dB?g#P*1xLopHJ}{O6W&;Y<%x1!cwz0{CstMM(
zuHu|f$CKlGfL{+!YjzaG?=8cmsDD+lTe?#-4LMXt+S)>6vcNZY{^o|NlYP;0Ev_hv
zuPkb1?8#T~V{pxKMj9v{ngZoL%rRWbdS<W5lF*|OF&$3ynu9m~o4Rf=+PqG)ky%L@
z;2&JG>!}+P=9ieMxmDm_dlT@E+8G{SELLokllT{8Y*CPj*mkw~S<+u4B&hXn(z4RZ
zKEtOO;|p+GwU(URIhWj6<np@ScK}8*GS}?ITKDA$z}&)0R0qN*c4XJ@xdr;1a{S(I
zLpLXK$hqZ=PSpplNpm_TGPevbW{q?DyLd2lJmZ5dEPrBj{)F%NRg-AKY6r&}o!$J!
zc>JSAW9+O-kitWgjVsW9U-C4J=b{Y-IJE`Z5Ss`Nug{tmhlCwrUkK_9^L8nU8pC%s
zMO}kzLbMteRly!Bkot4_omq<*Sy_dEx|EV`EvuJnoSeS~9sG&*8okmx5U~_?%}7e8
zhF!q>*<u-vCn~55dpL$VP3b%g&7tp9jVi-yE#%H2ap}Wa5BtuaKU^J_R)fvR8I|L2
zz-rN;YO^U|y43{7o`hkr!WGPKI6|s*|H9a)S7oDMz`UHpM%G&*Ye`oxky}pzcIo*H
zcdz5|3sap%rsI-}+XS{x(brI<<;<NS2~qDMzgc08s3t*>YP2y9cBys4^*@lVfJX{<
zE;?+(@5NBKFHE=U$~$6QgxbD=4q6LZ>o%Rj>y64jhn_Tdg__)_7I)u(xriA?omA_<
z2#Cmi6~`6-mc#Rv9OdZg67!3VZc#b8u@9-mz5^kV{O}5^|0IZ*Uw3)Le=L;c=4=&q
z1YB9EYqa_HEwJf&+9-dCoBxZzuOQ6Th&IWY!v{3bq-T&v-jiQ}hMniGaaC`UC`O#$
zW<rCFQcuLhibCy^mHGrR2%8!CYxsXQ&H-3A19`zS!WwiAX*|`znfJ*|cvWu7`_xp;
zvM)o?pnt*&*Go~&+VzDu4?F%HeL*e|iR99B{H%=0K+d{5O?%}z^dwSq+NfFm7Bo1W
z(WVVTNnnf0uBV<fUV#}}9ulI7f8Ue#1+7ZG7LD{|Fj53H`FINgOMdzFzPiDl&NBOB
z;ly@a_w#o#GV?11n|&8KBrG5<_77~VuscGzo$Sgfd!Oe_oFy$#hXiV8&8inS7SdYl
zkPYbCZ^BB|WiplQBZYf(N4i`oQIY7uuVKx*I+uhrl8dF^A5!Lv^md6N2R$-xQnCSj
zZ5cPu8}aGX$vm~z0NH<zu>9PO_5v)Hf@>Ls=9NHN%PlSYDld`y`)8KPtQ)9t%;!kk
z-H~rojc8akRFcQdO8`0|L>tKULo?Ll4^D7KA91LkOVD_5<|<7}$H5m}^uV&Tw&lZP
zz^|1lhA9fhuyaSyF>yvTLp&+YSrgZ~(k>qS{0jB+$bMM7?Ls+770{_!2GnQAIlBn_
z!H-H^jH*GgCRN-;e8~l`4CLzC3iH8FZu4j9EJi6`C2JXZYUWrbPeG~Cwf=Sz&sYz1
z+Web`5OH>v=&cNc>)A%cN0<Lqkf^%DeUP!PG3@}$i#r<Lkm92OlkS?+)4gAVNL(rC
zH^R+WEj3Bk34%mN(H8Q*D}Zy@f&4r^FR8dh55kdd!Du}t&Ni$hpLFmu4^2+*fp#S9
z;OK|@KaC4p`Sxq*`JSDv0V`|O3@yt(!P9j44-8U-^gaKeTW5~j7#xW<Y5W2Tq$t(H
zdwg#n4hs8$gD2i8Z>fmg`vR8mzM4%J<;A;I$J^Qa*QK|iG>CPde*YRKFBmQY%zO-c
zOR;lSOKuI4lNwp6oq}Oo2p)0?Y+`~VW??DOJEB&9&sEMsIdy!n65(yybpE$~@mzSC
z$~XbxDF-&u`&HSr{$1*;PhosDA<-t)b&n4H*Q2fR5~t0NyFg^DoX0dVKx5r+CA4^D
z>iPECr<dP;c0Z6h$!E671mhAYH=e&6J*xFfZ$WBJ3;6>3>&8!LWlHPP%_)7mOLbix
z_y|>6czU$@D~>UWHgDNwmRG&yIPDN_>9+r2%}~y7kBP6qXLs4Ns+l_+D78oHzv@xS
zodH}&V*PrP5kc7mEc$Pf1c!cKzEEEV=<JbUi6oWT`6pePBL1J@*Z)odKYV|LTCqtr
z!rk2J%E^B<RWeRrtRS3-fm7w7==&p02dH(ptD^skR<rUFgYi~vg&1~mqqf2t>|AN(
ze~2dcG!@408(<870nHVT#_n7l%#S$2lSB*uIr~JACwajkrS-or^~H27wiz)U^u`QA
zW@gr63ifMTQdi)8B!*hPlf@;pxgSU45_MWp`_~@di0g=Ig_wV)%Ch9jJ)UUKege^r
z)Hc^+vT-Li;I3yFFl{T$6k$u8)*oa!*CzD;#Wex<`R`Bbb1Y8v1^{83ivt(o%kMM6
zmq(h<VN3sGeNxZ#P#!u=Y0L;qFKMnh)5w2z%TL2TOsMOoHWEx7=%!>QT;q>Bp+f&O
zYSHvJV1RBzY96WLJ(NteAO_wCK<6xlp=o5QF8rH==<5T3ZuC*&H2POrXc!Ja!_-Mx
zQp1!g)<emRP`&~5u>0{P{x*9B8{keuqn^rxaPL(j{~IWC{{V35;@-~jubTV>^zV@k
z?^=Es+Q-o_-CIvrykvvH9WiJOLB<l#azI57JzI>Lpi$WfY@(2+UmTf81{EQ+DB79Q
z1A*>oFHz72bk|JtozO<1?+5b#zAXo^HCs9KB_TM6I3nnV+#5c#oCf-*54wYUhd59d
z6=;^8L$D@W-?_G@Eg%R3jtYH&1dEDD4T@k;RRgDlx>6uqx^Xu0WV_(N(uhS9AYTR~
z)NTUdfWyGndR!Zp`k%wXiP;O=<-`bT5XZ<#S8p2bv|#_22fqcqbVbwpD`<_K^{$E!
zC=62P@Q}?1`B6<XK$Q*vU0TLtWHvASTZ={Gn_=JU(#RL`(Ek7X?&v?n&0;}qIeU!w
zUi(ffc;OUE#D&ekjRQCNFVIG{+yKisj7a%js3RNH{%|3+z`OuoS)JSoi4CYv9dO;t
zy(#5j`-Ssg=CcHTPZ*cPkDrIK#x8CbntY8OoKvyeN0d2W$t~*{w-I=sT?U6CTm&*Z
znFMqw&XRjYpDlyo=JXfn+1&w80$Kq!5g;KT!Uj#_uY6bpN5k6=Jx2XHD0j{pFI8T^
zWy<)tf?XIz+RL9#{ed6*CQJ5g<jon!2p7E+;aw!&23|e54kZ3Ifxr}H{N8`<_MhvH
z2ZWM54DIpoN^%3hI5&sgTmCvFqfX5huuIN=g)11vasME5@Jzjfp9)F>lhj-?`QAcr
z>Y*pgK#r7o!~ID^V3lx+{`{|x@`BpCdI2xw_vh65!MFq7OnxtZ?b10LnWYRo0h}D1
zui<2i-@D3`m!5_bGjUkIUnkv2nbIfCZ=mNp>l=4*g7P#J{{T=3!MprYj*W}{xLi2$
z>JcPy7TrfTbw#uDKLHo_6ZemSpi_4ad&_H@8{&8Pvie7WPT*&VUk!x#%-0>P9ONq%
z4R@(8j2*2}OqKxTxb!ZFKVkotMB)!1o#E|Yg*zT{ZF2-4)acNy`m~2qH{B$&$^a1%
zyzhVx*P2j7H8q+0I6D@p;AbEh>853Z876w)N#xj8P%Frt4spn9USv!!g7|UC+f2#a
zW$)ol_Iht<?0G6Cg>P&Cm`^b&?lHWKPT~S22dD}2sOJvPtowpK<=BP5iA+T*8Hud7
zLRlYsQN55ceSO#$(dM{CO_+{0(5c_zH>mP5J{~MV3~&17Du#grDR?G3`3}7jrxZ1`
z3Slp49)g_g-a<Ptk6Y<}`2EM<f{t>7Jon|5+ppbAzwEq%-<Qe#1x{lM>>wej*z7Ue
zsfu2Bq%AZkz6VM;|4{GRmOI(eRB&r;_}eFi@h4d?heCQ7Uk|<JpRvC<9l@VX{HAC%
zg3#q=Dd1nc8(!aNGh09>DUeqH@rYh-=r&>$+x(7L#M<uvlLcwjacokrIJ3D?;|KUy
z`UE9q;aqJ(y0WQRhsQqs6Dz0v&$v$=J<N6D53FR?q528Cxop((;o^^B;(HGuF{|%U
zP~~4dr30bmT9EOBFeMLxAU!QJ@gMAONgtJu6rte_J3Iu$2w9FS!{_z)Cmm%YWw~{|
zhpp>7po3-lz{4&+Lv5+Bk;WI!Y^GIlSw?+=TYlnw3xb4LnV(TFdzcrgsGkl(h(<dg
zqIK-9n}Q}`KpXAya}UlggQpbsUd5w8qDl;MG=TSFts>mfWk%)fS!_|~3{>;&kB8Av
zUzRC(VSHPO(~&yo;{v~ZNo?f9<Opc9f5Qs&4|)8KJ@tjNv`z7UcE>W3eU2@Mp>eEk
zaj(fSw@Ucb;Bx>G)bwO!6>~i1-OuB^oa5ZFXEHltcqz+7OvXOnnYwq#>$=sJ2U05T
zC^n?L8&VQpX`(sZYi~db=$txmNXqUI=<^QBo}fJ$U9t#5x7LQ}{RiHiwYbZ>Khfa|
zeUQo#UlzU}AZyKi*a)Tgwei38n~@cf`}oF3ymN#U4~2=LEJ4IksN@-CE8rlm;Nd$<
zu2EyLax<crS!JV?AsjUI=0fyOX$Iq7_S5(_zVFzF`OC+_lq$@dOIz^kL16gGx;wHx
z{-8y^D730<o?T`R5Q0dK^yJ^LCB**NJ?ZJfVAr0a$olB2HE%bj2LhVuRQfB{r1IQ-
zKcc4H!Y91dv3mC@%-%uf_WIG>*`h1e1LiG(=QR$_AE<{hC{SM@95MB4#kxTI%<|BN
zzSn}GecywJ{t%~Q>XxIk^We9d8JnGfA6WdS!y}(c<<YfRdkMXl%kdv`n%GU9p})zt
zw)^=f3TQkvJMhwLBG~_{nBe;ElkP<#mY~VGRK+|%H}+1{WQ2xpjQZLigOgfgWP>Z7
z@6Xa<l$hG(UIzFfDwhL$C0d&ur7QdA>);WCp06uA?6<m1go)m&qP7Ri2M(RjHhD_i
zb&3IgAW<i6aQ}E8dbH+YXZ?$b?)~+vVg>y2cmg}ZK{yvRia&mybyuQC<;>BKn%Dk(
z)xh+bq&*dTV$aueZz`lNoId56qxe$W&%OYib2C@Egp;HZB@^;YXe=Fm?C1}!A_h`U
zfaGM~6j!<?;oM1YQ-^Ts?DwUwJG~8LBtCcqbhfWll9z0XL#aL-e=p!Y4+!=@z99D0
zG;au{;B)0P;w%yt@6=(o&3W}Eg@gWrEfZGGQB2A4{3obU3I}XA%8RZKdYkYiG8}(P
z%UNauOIQ%b?=EtAf=SjzrkwD|uxPVTg`)7r5n^p?S1p5@-s^obeE(XCrd<whc<$Wq
z-#t9!<Q+C~O@4#LjKz_|`t$DgVdlSwgOwPkhYwqhL@rWYEjcXg7Dn|=V9bj03wH>g
zb}g{*vq_vmzIO!_H=h`#m|S{2Fko`!CSTYMXZlo=$CO79yUUl^j95Dv`ro8@GQ89r
zGhKJ2qg*&<_1XV&fd$_oRmal`VWEC=5w~T<A6I{A&6`;&t*kqoYyXjfD6YgL^6dE`
znoE)q!rG>m+?lV)BJI|+`zXoyfgVpMC}F!UE%<#hjA~gGjy|e<DI03rj+^$3YwQwO
zri+xR;+>tENbI3uuV|L{f%c8;0^6M+3f{$YHp<0PFLi#RV5a!P{wdbY6)q3&W*|;7
zDo2>tkItm_zDcQ+bF4b&IH!`q*k_x5SqFE>nVZ+vH*_{FH;C3VLCCUk18PN2V%p8c
zLcc$(;#;3Z&FiA=z1526DPOl93ETY+KGVMG2uEZMUH53$Nt5E$_lRwMS1L;#A+_dV
z?Fe<;jS3lt3pTbVC7o5qvX&{o^j#F2JSw(u8S>ces}oI+$Er4Ivd+97$zJ(p)k3Tz
zS-C6paNNQ)&E-aMX~jug;ak9#OHwJ!3+LLt-LkP&k~?AbNO#4?q~j<B|7q?{4?3C?
zPRM?z;DHH6he%K%W!S$PKUx)drFs1@Bfs3I?qJsRXgffmh=xjy)dSN(gb0fllrq9B
z4}V7{|38hLc{o)6+sDn=#ummlcE*xD`&uDP#=b|g??TAdqC=J}V@uYMHIseImNI1D
zLZv9Oha%EKk>@k|eV^z1T-Wo*<C;HQgPA#><J{+*`}KamZ?)P#rE5#-aI%r^J2XmI
zF0(57Q*@pl-7=>dNHrazP9{#zL=}vDOUch3p*)3>Kd)>7Qf#80oiA_jBeHxMHCcyI
z354%MXkU+BsWS!u)R&u^v#4E4GyfO?tufh;2bFl5{Q##$?PAMlaB3u{a_qi31#3U~
zo8NIM$}D#0{jchJ5`s-sj00;;KgLp94~Rby{8w)Ib#Z03;dn1f(|c~Y*&2V6WdE&c
zr8X`V@CG!s1J19I^m`YW-^8g0S6tI*Ji!d*ADs$iKXNGUF1mT%ZcOpf_Oif{!*k7l
z?F2P#Vhu?PIcZ-&rF*CAY^ChTKzCbyA+OK+FeRsL=EZ6D*if(E58p1hF3w&CHwAw0
zuwSzo_kcb(2YH`<t}`3qWb2f|?3mFLlrO3zUT+W;@YKmx{>S;}3M9}Gp9`eT=Ts$G
zr|zug>eI_^=<mG*w|{R+Dl_WtAm;31`w%mNk55f{q`nZ@hFXGH!{W7Y$?>!~E+rBK
z*QgeQMmLr}Y9r`6H8)H4t5yk)mXT-^U4Fr4{*FcS%#~($Yz=7&a{f_la%}b1=2w^O
zsNYmI;#2P@Zz=Z3D9LaL3B#(}xR=Pg9bHTe_as-NQ4;e}m^Wp$leN;<G}fOB^L@jl
zpOiDONJoi)>#X#w(=9MAh%AG4e>nDXww}$k*i@@>q5@r-TVd-(815OETP(sRz3((V
zZISaKcLFc_XSXWgx(4S@4=JOUPWx4)K_c{PB%P)frWnp%mly|;UKy)xSzjHrz5@Lk
z{t%k@C8PMYE1Wii3FCx1YT%MQa^OA6w0z9W&ro?>)is*WLsWI$R-WU@_}@D-RS`j;
zv+eAah!%SIDO@+=%)mJ7eflr`+2p_^9<^OORv~+(KNc}NPE@>@qAz#*DEXmnO9EHR
zQXeUNc~2HjIE6DDLfq2!L0BQ{E(#hbIK^g@GPq|OnpKSu9lRs60a}E;VxV|8sGcAT
z_yM|!kd9n$rkTWnY%PKI?mL<A9_uI{K>T!Q`O+va#DH)o)>&8!AL;Oz=abJobXJ<G
z+gvb%Tb`l6Qp)btS7>->J>3IT!A5j}YrCS$8Ii#6!?czz5?ZI&-9s`<cV=i=LpA@w
z!Ni5)p@TP_e|n?)^SbAgZLO1v?lbC=bF&&j;;YOLv90y<F`7!NNeR=8<_>f2%ELy+
z-;AY%wLXgip{S0!Qs_GIc9}oIo2BvWx5m{<n_KhQnmidv0oR(XcV55Ell?Tm9<jn0
z8yWp<Rmrmk`-YSw_-+O#WtD3oe3isa#@m4cmWdt@m4;hqvD16Ta{G9g@W^OePhN!H
z?5&b0oYC6^w5d+m^(@iU+JHLxu>DE*T>zWoex7Cjittwi_4VI(=%ftIFJ9JN=IuUu
zqb;$R>-eWrWCiTg8=728+`P}WiQYUudDyL2*<Otrm+jn>-W}lcqZG4t=6Jprj&6oP
zQ1XP{ZYEyy-6HQ8j9!_-*6J_^G6T1qr_2Px&9#TfOPBKJY<YQ#@upsqSK=kT(slTZ
z>|&G@Z~EtD6kkOS-nI8O+DZ20PQU$FJ@fq76WuNp(m}5H!rK>tr^>cN{E%&LR55B}
zGqP?nyTG)t8jx2PZPPB*$fWCMD?O-v#yhK@>t*elV@I5_@x+DoFt36|$XQjT6aLZ8
znh-FQ!y_ToW{tBLh!AD%eC&9x?{xl8^@nR=MeRkuWNPg;Tv6@&<7fS0)EsII_D)z3
zp{@Puu@A;QU*`e1Bntl|o)Ql_?Wn`5w!4U-DbW$5L9)x5X03_Ah)>L+_1#Wp)~On#
zZihzu^}BOER(j*sMas?4mNQQ`HKoHfmDTSvdLMl`<+K!I_+omgf_kPNF;EuON4LtX
zmY>syYe+dBPTNbhnm{f=252Z?RL*0fi1UP(qWG}Bt4=np)I-Ml{-&>`q;-TSFF66p
zw}vn8H@dTv*s(ZFDAi}h1hZwO@%xFErY6sQP<3;!K3M4`eJ)=Gy{<CH?}^U*QFP94
z;EnEc?v$O+=45Y0J%dkvXMuCPJ~KggF7#F--3quncdVuPat?oI+Z4?hDt+SwE}=?%
z_hoZUe>aDaT8nhYc^S|9sgLlApV?)%Wl!m*X}+rnizB_kMD_$X8VAOBdX6P?C?rs_
z8*#2)j*FaaAn!{qfL@2^-i-)amC8KYq#NStq3!HY{W_rWKzE*8A8AFCI!^ofeN58r
z-~>uo`3Pamg>CfNclWU&@p43DKL<U1fNS$s_w@T9b#bB5DE~U;E6Y-sb+ddK<-ajT
z#!)$bio6!Ht6253e!Hc^$NzbKa$EL=n^UCup=sS$$(NwUCv8wL?59}*i+0kDY8MLS
z&)Oox*;lMvz@;Y|O|a+>RILA?+gTj7!wOVJ{ogEA?Q_dpBaFss3D?+CjEz)YW}xAZ
zuq5K@OMf~*OuodP3HA0yN>_3^1*U3GY3J^czGT7-;M_%?(u&n1me%$msuJl~aE;;P
zc<uIfbApFFlzwx>Htcb)Mm!}$isEVPPPjk}75&D^ITIUhtSJ2&>eo?j*SHUiJ^$zm
z=q}fLb?bA}aohFE@|MJ}9DO>d7tsOKh0n@KycpFK6;nNbXc!f;Cz;&EMk6O;Ga^|v
zRH)S-tv<`uDqco9sgi+}GA^<+eKS!bm1Vxz799b#>*dn%OL#eB^@?ZEhM)Ziw`qaU
zpfUI!DpzA|?lR7#weMG*<A9Neh@U|-tXW6U`J=OM2jP~-+mvkawwPLB-V-KbPC`8w
zTo~?X<?IBWyO793{Rufd6Gt~O!b5Xu1Wk5gLRbs^ll_IcAOD78Wv;}<dD>x{OM2BS
z)E%YJPK07hU9TEV5k7?~%+>7nVOf=9IWmvu9I}c}(^m=J;3~ID+7m=mPsZ&i8=F8m
zs^+&Y%yjSU8kuOmf=6a+5(}z)yx&Q}o%v_pqU`DMQKWpi_0THoSqaf1*JXl)=I+^%
zK8~i|>GY5DVgo#99`bbyF8sJGt8JJ<vhY?Hx|?D-9e1MDv{|PsZm?Md#eSje&elO+
z^T5IQh(pjP6bjkg{R&NJj?=a{r>vS2tYBb@*z=TVDB)?!IcVrjvf+2X*;m4=FGQ`E
zrm2=UhiqpYggxh~ajGZ7tv<gn940psneQNP_@Z?PF;c%y4})xAi~IV;ZUrvO9*2BT
zlW*E!s$%d|R+uoXEW389?n_5+slQE~IsfEl_k-bJuguLEZL-z%;mVAUf}xUT5KZ6x
zn)Zao_?X^7K^sx;MESudFdA841PploL?;&%tuz%LcZ718tokbg%he95VQHhMS#tx-
zz0^rg6<X?DyKyUSB|}{|xesrbO)Ws6qpo1sXyyFgQMJeR6x`N-+vd3@OLh8$g0!J?
zh0ACAiBkUrq{?t*bFh9kWk5I5ef`JqU2!T7p;cOnT01dj)x2Vryv+Me%#w*QS%=p@
zWea2am*qEHL%e01YG}h56j=RSPvF9$y{YZAsS94?x8gA=?YnnsxC@-$Hj+VmXjj+q
zzJTm-A5`RMwvV21jHa5r$5q8yjNmKbsU2?Bpt?X?ApcG)-P1>iv7WXL=nBYgWWZdD
zj+YwZZrY?R{jrdhB>MJyoC!beo2uG6eU6P=;>IV!(2XaBY1FkH=Q*xYZz7+c{e#;J
zcM-t-dJ)Y%F99!%bY>@!3A2XiITc4{m@^tZdEzlGH}?ZCrKmJpAq<*v2Sp!DU-mkG
zrP>>|!`F<Y78EMYlTdxL{HFyCny!2|;XS3H_9No?;bPs1!o0uTrYbAWOmOW^vTx5}
zRciq!%zP}R*ety&UR$b+zgQjwye-2967wO<{$1M-r_H7|p{`u(Kb@>-(55p(6e`rE
zyuifbXlI0dbTsVbJt;}ByW&#sfKywBv)G1hJ|u~sey$(yRn-$jkNDWHpnbP2fEb#X
zG!=E1TwSr5M<r0u7FqQ30q{}XKXIRS=ZpuIz*Q2nAXQmN%+ukL^IQrZ_|cUK94Ug*
z)P057S=cTZEyQbA=vTSse_}EY%Q)uS{x_obzq`u+eDmHw1GZl5#|?-XrkwMyRF~L)
z{@3G@9!blkVjHF?`pdN>j8q$UGE%4GW#=vGYTw>*rcS74V4LU0#1#{bvIoOfrKsfk
zT6|j?6qGkHz<abqjK6lo2utvJ?8FMp6M*zHiI@HN%EF9$MvOV<6Lv7P7jdHuv^|$v
zbV)Rqei?ov5GSGn9s~-^OVBP&*tVQX+ELrK*JZ<t`k5Z6Mjr*PhU!1qad;|gE#NfB
zM?lThH0v@w<gDY1T4XJoR@I|5B+#inRceaHs(j*+*h&)=rDLob$JhIpEV7c9RJ7`J
z4YZ1#*%<p$ak1wT>WUM#rs*t=E~P=QWWqkZ9{XQJAXy5=SW=dzpS(!@-No|X|04fd
zn@J%)U(@!upvWxaK>!iJpLivj;p;6ekGTU4pglzYz&d)8{FpL8spo(=S+A(i%JTnQ
zYHdC{Kp}|KEbP7p(c5vbYkcJm|I{QYnMD1V<-QJ9!|oN8k=?67?f;o3b=*wX1%D@W
zw|Php9!YGugH8=r<ZZ9c)^uYzkSI#AN<Y>%XA6d!TiWqNUCh5WL7Oj)_?ZNKGcw5F
z0KeIT%N`_kNI)71EHRPXKyu=hPX1=K)P^fS#s94vCvq%7u&qd}t(#D!2JjdYZA^m9
zpBK@<hZqd(%I(&KS4cy<SJQI<TMO&$Fu>ZfR3Xk`%dIkRzAU9w^V0xfT>^C1e=+!V
z;U=>&kW+!R>^7+^TxM7Uatffl01B_{yv5lI8-Pxbz*&FBXBHzLB~r+Dw0n{;gn#EE
zzU2Lp|JN%jgRs6>dWiQH==)1=2%U#3Wl%>Gi3^`&K>8(7$;5X2)BV>voAN=QP2qv3
zf)(>B-|N5D0EY4MFn}0aiHNc1Q-$U`(E>9h20#&_EetC_%yS0TnJVB5e=8wqh+t#J
zMOp$?6`00N$G`@nNdM@AJyvJJc6jB%cv==jMFfi1YXVPMlD2h3<}HeXGrgXM%WbCG
zN(x`_PFPf=1eG|JJRX2=&}%ONiKR-jK};A7NRF(qX%rqR0fHr^cyKf2*^xYrO#TkL
zt+WskYnf)D6VK1<CAv|(QJnUJ-R{dpcx#pw-UPPk0j@CY#9h(ka$9935GF~_wTOSn
zjFg-ZU`t9KS6+)|xB};3f+KDO7)W=On>hutja&0S>l56&=aYDjVM^42Q9~by`7;6i
z!GZLuhGD5;07yv_zNlNI51F}%8GQ$K_^S`B*euzCuO8j=Szt=N@hCroX7=yuJqc4~
zjU;0j&vv}?xmy!@CuLvfLjEA|ZjmSri`cS3iil3L4&%aP@0>8<dEtlYl$J{V^+V#C
zb^2E3ma)R_bplIxcU&-7Q&0=^s)ycrCq}XeKb86xVMUr_73fPkt(ta1hmR4EYShpi
zH<Q8kBOr1YDt;1Hiu+3k1*#=w{&Tbes@9<e__tK(-yP7gYnxSimcLM@>!r(Ws<S$;
z7#{@nTU!VSca@LzK9FSX7L=kVI}1$DAoXuvWss>~SLk?lGr~2Sq@V(T<&}VRA^qD{
zRzMaNJG~A58N${l<fNX{ttHTU0!?zi6d0lj#C2dcm0I|(CUK}Ip`3v%g&+{&X5TI|
zp&^}v^YCIHwWYhaL*b0?NofT`fY@J%m5i=G`>jy%q}FgUP1oi2d1yeMi;<LSvLY_>
zVItE_>RI-vtITvmcz|O}ua<!NLd&iS{4PYOmP<UAM?)@DgDpm*!2yN1G|3thQDTl}
zu!Wn8lp6u3<`-Z_G_mhj^mwt-H!1LJo7kk3$}~k>dBZB?`Ba`-G&u3NoH0fV7XpNC
z>=R{-wPkL9vtC7fnnkCQN@UMsg!tt02G6O5|5Z%>DJ+vh5izDfVG)N?2DziJa42Q(
zS9lR;7IIQ?UsGvWJvJg`WoFu@g#`&6j@>#fP@hOXy6ikHn0x%zC*w4V0iV6+xhMt_
zt^aw^e}4l`Tr+5;EQvuF2Zc5h5OYmr84SYMDOz(t!S&199Ds39#S4Q4D2}f@eO!$Z
zK%*xIYS#O#8t_93q#-zpwX?<sVC>MmJ1C!RZNY-a+OT6o@n4jEfhrG}4DDI$0)cqC
z^fa-$KchkXhj3p;c-k-IK1}5UE5CpoC$Qk|4Uh?x1c5TC#8lS@EUoX~8{9QK{``C~
z2<sQ1qRWAac&ahpH5GuJhozGUSY}Gtq(ZYC&pK>U_Vn^a&4rUTc66$M-Hgmxo51nf
z@Rt2+)*aW$`$4$&=mw&R4Fn?yI`U*<bB?5IML;VPP-IfKztS>)_DC=};U3Ujx}9s&
z<>EoF#b-BrT7;Tb>Vn!GIZuXqEpA!i`jq5`nQhPgAb_*K33GR31Veu>4*0al!mPgr
zzE~30YSRvDRH$tQIG3m!T$pXYMBECbA)w6U!~jEnXG2M(WaN|KC<EYl0X9UNCQs9(
zjV(WM>`6F!8zN(sx)j$OS3f+jyLQ9h;9^8MW9^sIIr6K0(dDv#x56LkNDb_9G6l*v
zp*)YJd(NZcz}};NbI<?G>SB)#(ZtphP2vLR(-~qrff+sZ3N5ViTbQX1VALfL#RWJ_
zB_{ZTe5Du~eoKge{{w7zC%BZyIVR=ghHlh113taHVDTixH>dM@{k~u$agVm<kkAi<
zgB3h7&(k&*MvLOWj6>)JkW}U?2P^_8NNAv9FYGA}N<-@)eO8uNr_y}D?j010DA@u?
zAA}FT1xbfsXhjW32LSm6E~O3bY=bN>e9PqOIN!Kqg00{T#wG3VW^}(Eo7GBwfv_Sl
zp;Qehj(|GFZ8Z0F$6>)R!>WqvKnD;D0mFT9IduN_Yv9-dS-oqFD{Ue7j)fTWV(`-v
zI{_HW9_nqNb0(6=FB<6A^1v~(4@5R#ioiFJF~I{uR{$LZlFAtnkpbLNH(al>h*JSi
zQfdQP*oqV|{sIL)P_Sb>;{YvZ6EWetP;)7sOY$lFfeOtGMDYWN1)!~WEh$~ZS%84y
z{RDzh>`y&O!E&%*om33A)gWU5_YsA(#GCK~R}OBMXNq;=B2a)&gL((%bevD(cnnxE
zU^HunGavG;9|Nr}Lu43}w0R*2vB3iZoOLOX%I{q(ea{n<8-GT`>m*e}KBG9$K^Ox`
zqI_Wq>Q9k-w7oc){M4Hm{IN|5TcroTmaGC!+8{|Yx*)ET%a&&158j>ja+fTL%O1f0
zL`QdQ=tF6cKjb=d?mDf$sTHvRGLXZX4sdfI59qvBbcRUOH6UR|ElX~8p7c#5=|T*v
z$@AiHd;Xv<Fdu5Q{5{gGA+j*}uXPfjR)Ywp3HF_iCZ5bGg`Kl?t1Jd$*%!Biv}{YS
z>2%bKLtw+CsY=xzBjg_3YsW?AJ8_+I`KEo~W=Y6x1~VYbuHOb4>n(oA7#qjxdw80g
zh=$nJxY{2r4z=Eiw4NAixzDoHy8-m!*yCId^t=lEwDL6#=xc{O{g~_EizQr)XMx{Z
z{K)6->jW?Rv?`2g8`vtCIn7jYy|lBplS2$^nIjqM>Cu&%Kr7##`Z7uCkf)Qxvl}0b
zY^3rObnW}fa~SW9wz(-*8-)p)rkHb-#AV@wn=&E7FqnK}iYv=SeRtN$cu!<IZYxDy
z+X!J~kOuSMm0=Xvri}wS#m56-j6<5oMT;Or_J{X#7{z4C>DU`2;0vo`q+N5)@%DcK
Dcuj(D

delta 89144
zcmZVmcRbbaA3u&C931-`n~dYw86kU*V<bc*E7>EGmCTNL9D8SH7b=dC5y>89lTpY@
zR%J(#-=)|4^Z9+hx9>mn$J05_=k<JC_s4xauCOa43HhXWaUYD{&+~I+?N1)d_f<F3
zQ%hZn`dL32(F<mb+g&SiDWiWiHLhNq%E*&yGIg`zJuPo3%zW=eNoDO<e&2WNH6@=f
zSDlCDeBPdyH;?4WeS2g5Z7|!X>b3E6gVMXc9!q-@uPkML2{yx_<Z3p*<{$ONCWLNv
z^DN(8Rkmz!cb%?c$J}r5m}lKh7E;`tuI5y^o+<j>ce?C)=J(Hy3e%DjZd1aPah?TM
zZv!8`vUpt{!;{`3TP<SQFj&ENPEP2@jg*J9^i9+M*q$Juze*ffWXwxg_3P)aFPJL`
z)mx}5A!Vq=z;oL44#8iE5Z%X;s8DK7WyRMigKja9oS2w&7$GKHm6g!Uhs~=S;TpAP
z495`?ESo)d()v&x7jk~`;iq|kE?kU0%!C*sTu7wj2xSlbGciU`i0w`snY~p^;`!ql
z_RM?hLliQZlQ5AQl8smn@<TwMYfvppMp1{VG@AQDPMcN62zI`m`ualFUp)Tvy)l*d
z`b`{ldvg?Zdk6+S6g9(G(J_m+?*224C#Xxe5_*dfVHdK!4u?ab1Zt}PeuW@BeIZqZ
zH}9%PN$g+z8y8D%OUw|;efe@}N@sxxF1oWNld)b?wH5Q4J!ci}RQBtEH}-}s!~g#y
z{`W~f2)JuXvpti@*X6TIB5j!QH!9HW{lABOjgz$hpNrLy22elGw~NcAD4j2GJ)U<&
z)Uxfe?;||TeeVnku4slX7{A5DJ!`_V*6dp{pIi%;@8<fmKcaHy=a<$Zy~DplEr=vi
zjn(NArsgtQf?nS)ELE^aEQ2xq%7d+`S6kE7{1^ifmc->dOShKuURRgCHN3VWvsIXT
zE32Mse=*atAb5|AgKl{xO)lE;)+6!VEOig|T+!6tYGV7$|E^FuKdq}P75YSADHawP
zB?-;aq9u@|qB~(h5Q2Z8Y|704_uT)_U$rRYQx|jF6m>OKLO49l(~psoHa1izj#hrB
zvjK@r)AwVhFbC^oWNKDw`0oNHOjsle>GglEfWZjELy^dTzc`^d`SY#nceFLfnRg$b
zY&<WxZ2EHjn_va^|GA1eA#{e3MJZUl^?!fz-!IjE*ujta7IX|PDQED`LeS5*A4FTW
zn^qG4{Al@Uq|N^DU`+en$yzSgBfcRr%zs~4t_fj>jGsSvcjh;gBlYv`ucx1e^*8lP
zHTe>rj9&gfU-<9t;8KK`zbD_Gh*-A$`Jqq?^U;5VIXh&UwRv&7dZFR|=klP;BjW$+
zt5Z|e7jm*BI9`0}nfCnu7Tv!k+eR1K_a(Yw;=hBB!Npn!d}Y{<ANLrz@xKKfLcrNo
z9~1F#w*G(rRu~QW5}CrS{o~D6Rh13Q$mnh0q0`9Igxfh$loZ$ht>8>z=sINYl5Gm(
z*XUeD-v8GV*f0<RJh(qr+(X8X$=ex{`4Z2rz!-+u=DPgmcscI5LH?Tw0|oP*S7QH`
zT2%rLZe?Lbc)Yl^D2<dj49$(d_p$QQgVvgW{WYpf(l2klF#Y;IW7e88U^Qj^&YxdD
z3tU_x1OE3QCGd6)+s;TbaM9c4SfTz$4SwG(+~ZiK?wM4Y2x5*^A9vrG9_|+yjx_uC
z7qkiB&dJ&S+7T;XNunSOSCv4le|#}rH=_Ip;Lr2g5vRZ;rK6@L?4_F&UoHO6W*(&r
z19yUCUak&4*=zjr_|M6=f`ZEof!|+GkH&f4zOfv2`M>)KdkHHf%93z=kGJiT_L#%W
zw1xZub&}lrWODU1yz>%!sPMMA*9T$x&&7IrrD*W^v09763O|LdTmA37wd;vZarf_u
zWmvsrAd>0!zeiM4m4Px|v7W1UHGXL)cYUxx1+&l^koEpqY%KD)oroq+F`(8XokHn<
zwcQL?B~|@h$DJs<T!ZQ3&>+PJ9&U=>c_Ukl;!q<ErK-TidKYDyi#BLxXX(U%qNY!1
zzoBv#$ub7&bIoKv;9LInbb%U&fmDEajA!`C{|ji#s*6b~f3qZAX#>~vbEO}ykA_{1
zaFJbRCfzW;SgEjNgeqyM_}EMHx>0B$k?d|ahVH5Yc)NUj3%n?iy<Y7f<b>hv@8HL0
zzrOGr<n=3Gaz)gh@IUpp1Gv(C8D%$?`=D)r_O$xbp2q)OdLjbC+bX9!T_sux@`sbg
zV~R2^qc+j(exnhRMe6@ug}ebplP2?f{^g)^O;O1!eS%E(hqnFwy@iS92y_|D5{K5e
zeR~=N)>$!}v8aewd$6XDqs4PceY*eZt|1xZe=Enzg?AZ4(GUBq4RT#2y8B`!QNe(F
zIg<W=ueSA~P_6OAK|0Hf#9i0#Q4o@$L*Gg{p+R+a#jhvGEVOslJwE$g?vO%<SP%Sj
z^xX{eIJR6Mj84Oh@gGP#L&dPX9-3vk8C7pKE3XdAu6(%iu3O9$ttn5FJ<9h#U>YTm
z^q6aMpRLW6d9aoz5p?A4O^<6X=Ufk;LVPYR#9a$fLkhy&KR0}@c>YwK^SUQ0TtRWK
zpF8MBvoq_yl-tzlPt19)selSG0p@?V&Wz}n#AX<y<GZ--fT9mY!pR9M2*hP`7kBZ~
z)<IF@wUVh61YkF-YrQ89TSe`BynT1cSK4oTrt;PcYG$!QkNG#XBL+X*;luyHM~svp
zeNaG7$8)OI@p`<iSGshxUy;fGUWNi;F;Rcw^@B2$2qu=penSl|MjA#;!H6H>OM>|Q
zInHDcIT%|T6-@`l5UK=QT#uc%Q(v9*B!72AY0Z{*ERLnh6wNMPS{8goc5$!C3<P5Y
z4pExsMLKcmY9j1pguhXt)w9RDD`e|C-wH4bz7zTm@}KRKln!`r3($d;{)~`4UW|yb
z4#3RVM$ZnfpC4Pkdwixw%AO82C|N|<&mg$9Y$$_%Z#;himPL@Y$3-`nksC>98b>Na
zK7}TQ6IXiV{{3h|`}GsCz9=+|2es3~bk}!l>N}1*F-y*8EtyNh^p(ZKE-D#Zxj>Au
znm;4dSADUY?nW>Gci&<Gb&Ey%$K2cPnDJ-g*tgrYA1<4etNvc)z5Zl}0RZ_^i6F`B
z;wJ_|wnCSoYy@s5uiU1p04%Y<eIZJL2eP2r2g)L-7fX_q?|o?taoV0#f7Ac_rQX;}
zJg#0qNcg4zY$gfv!zt6sdjQnH?+fO1=gk2CrE25niD&K<(O}tZ2nl&G)_L*|E6LKa
zsw-DV6ekMRBH*mLT6D}^w6h1RX*Vvm$cejGmFh>MXL>9>b06q$Q8xKH%M+q(q(Yz!
zqgUi1#7KoZ$G44R_;YI2QWBvABJktG>ARrAU&|X7Q|sPv2iV*4v#o_Ooec4B8=Z0z
z!U+3)|8|gaBIz;5?<5q?F@Qr}JG<U4t1wHEAq2`ro?FfNaPwmdyywMIvLNf`jIR9=
z_ae;fnvaLdi}RDDbqHZXEgylnL#z|KRE@}c@;G{OLPiwc@B62!J1xt;Yk7LgV2waK
zj4ki)OksQiCD=!_uLbWcc6`f!#Bp@G=9892)YR3Ep)+8Znfb=k8F`77wd*~-l<hJs
zft4#XG~<k49+Kb^i>22lg}fV+d~oZ91#x$>V9rC2nEQZ0Pzt1*d67`2O7pax?tKub
z8sMD+JK@nnz;ly|o8&raf|U@54Rtu4IgC#~$Nf|N#;693<45_g>BieD>Be-+_@%Ds
zu8I)MTvTVuS}$HgQ-j6Pm4^$~L^uW7M+*Sg2Ji3dkf(v0f@1{!G~)7*VZQ3ON1vpo
ztu$2Q$IGq<P2cvs{mjKQS<@n&fsu%umnE7Yf?g)(-RU01?2uVp9D@*cVz%^Jn&0R9
zGLGm0`dcKgeR*IrvF!kdn$G|JS!oV6iNg5zG6<QX{Y!=8z39uA3H7REmWkmRWKs&I
znyAa~W3Tq&W%HGOua`LIKV-aH<1;MB@Ai6z!ZEg28LA$t2E&DZEbNHInT!#jp)W_A
z_6?0=>ihW{KHc)M>*3{T!Jn&IeM&IagTAtU%NlaHlBDwZXx2H2O|I5yC})!_MG#ZT
z;j>``$I}oPFo6c@!0wX@x?+<Tz{a)Om@(c|-zVD3+(;<mtIwKY+rjvCWF0`@j9z@U
z!||0A!A(w~3PEraQZ+t+>UdFwgx#xT_g5Aj-``>i$DHo={YPx~S7fG*BS@$$n!R~z
zIDBs5EAm8EuJPxi2QloqS+U_v*0F#})QmjddevA378#TgqD&w*@w8shc#=l<rnWN_
zZ##>(B`Ub|a({BP`Qesn{7lLbN>}GtW>{s>;bC|BTd>{#avB+@&573NY<~1)3cOM!
zz6}~mytsRv(~t>k7k1ddm#S8Dqd``6oWCGA{6UJ2%2iC?9n(WSJplqnU&xiV<7H4o
z?&AdTW37eoLBx%Z1V~X6IyiS)yecnQLGjDY3tD_N`}Lnydb#9aZcPiQuAbdqk4@jx
z$YAj%E<7UTaC+stjk4m=jQtGVcR<g!>vE;<zmL)mA#QYty$msbgQEv1u1WqtAmPD6
zD~lXv?9b14)t^sJqB$SS?z9t;^@-h^6zJwG5_5CouIR-Fze9JgDUJT?&5bSFQ2_dU
z4Z#n3aiy&|>h^C<>&=xd<EVj{H6OOuFR<(Ay&e~t%U<vUz?VUDKwLuAZ`t?8cev3Q
zznsz3L~<ERCT<`@$O*~RZBF*S0{-Vo=+w}Cbv)s|-ndu-m&Kgx^L%Vnj>qbTDh**<
zv(Gnc&ko|}1&`&w2%{f&d(#@FP-}2Y!U)BI<_pzJf!@TWU|-((At%8Gfv7^Rq18-E
zh*ae|IR1>NoDW*4k>7!i68Qd@%dR6c)nraV2$3d{iZeaj72i@X)pSaj#=I-<x%Ic<
zvj_Joe>Us+!>%Ws!lD_MZ>0u{oA;%KNI!9;VwEhnek<>t4fcs%wnUj$0;{=D-z@=S
zNzv7^yK3!$P<S`e@%ZTTgV`pZBCdC!*8?s11;CF+F};X&tC<G*3X05)qHVi<{FP<n
z^($`=UR1Ze`Lz&ae~_Dqp&;G34cg-P1$AIWI-Sw8hnlZ^O|#IfVGh2@{V^PctXU36
zIjMyR6J4Y0GQGrw(E2m+Okc>lwW%SvqM-QU=A>E%f*l!5!CPkJyQBLdTr~v-|H@kS
zb*9EXrN=M;kk#^=Wx6RhDyj`%SxD6qo&$I;e?OK+!!BR1kI9wZ92l+Cj<-GWPkC7f
zx;S&!=Ch?yav?4)xDG4fI>e&ULw{J9dD@h;z4FI81oozsKcz-5w?344L~%c!+M2pX
z)%NalJ4=VtEW0)}J<f2|v@0ZWkCwR?&s=_R9&{16wj+W&@^`uEIfx+RMH<oENs6fs
z;;{Gyz$0H^s@p7_jZ(VaEDQeL`*1~D{@gJZD^pU&tA-RKJSMa;l=GOA)i&wgcVKy8
zSfsv)Eyfgje|_J7kpCEfJ2vu#<A*2RE4Vi2?}1{S%%%#Hkw`V9IH1_7XEAA++?U7A
zN!zoU<U_|wwE3NI-V&AfRpayMKO}_0aQBT=+v_nehaaa0+cw|C9&SuLfAn)<`(u6a
z8R%EsaXS+?ROZW#xL?60l9_uwre2ym$aj6k-~ZgOF6pT<nQTCl+WNe_FCt)0-K6xh
z`$z;uzy)2eMTL@zj?F$kKYR2`)JyWi6OpaIli$MG=!LaKmGZ**^s?GFP)4k``<RbF
z2gJ;|Rlhxu!9FCtE_xx*Sip++J4Mjs2+BjYT_Am91j1HL$vjsTeq_n+Q}}%X%yY@T
z7Er&fs+Po1X25c@YUQ}uE6}Q?nwQ5+=b3`mdlN3LWcUD)`Qk@Q`ClDIjSNN+3mZO-
z%0p;XohH}8yFV&t2NU2ebSej(W6pj#7bt-pddF|JXC&gL_CJaygjh860zJ8e@B~1%
z;Z7&Ff{Bv8a82^VgH?D;B<|xS{+FvSmzcZuc3D*(pEiGc`sa8hslcn3#neE@Ru_B*
zIU3+GF=6?!*BF5SiQ>@@T7L2bb}vV0Fu?!F8MH~{j%8OCwF|~7#*^&61C#RB%eu#>
zyV{QZ%wuB!1V6oYi=LPT`od6qg)hNDM6QpLyH^KC54hVQY#Y3v#yM|_*;~3mhX7fv
z(JuHrulOmYiqi7ySf02f_F@#95Ii3rZ&}T%wshJ_aRtdGzdv}DML6CzMiW(xQlZn@
zgb@y)+c9GV_@_)lN9PI5dSe9JZ+|T?Gx9q7JY|Iwv4;v0h^L-dMD@SciBpFfmpt>d
zLY3?QGayD58*;fuS_<{!T!%{;ky&nyh(gMBk{q7@;Q577x6mt4&fh7BME4l~1OzQ`
zv*I~VthG0Pq^Bq#@GLMmZ%0bee%T7&Zm0yNa{Z=i_Y_#=zW_9xY=6$y@sqrJW5`Ir
z4xF@R(592*)|iFg<;Xl}^O%=^*&e^2n+Rccy$RO%LT-dI$+OPUv4@gG!1>8gBC&L5
z0>Ak(J*%zX!brNSCS#t(#I!v{s3O9*0mg-fY7m$wpIoCp77!CJ)=yod<iIF07jytg
zf@sN!?{sgXkz0jx<CswYIJ_Yi(((7f+rugF8bI~aIZOV%HT7^b@s00Bc#(jpEAw=6
z)oxYG5Asjph(mV!nUX~LSO}vXrCWD0SxR`QDGW!g`eMm52EiUmp_zIGZQ<~FUFFYq
z-PxaCkBE9YIPAEU$}pN}NRQzcc1V1DRqLK^e9>HLq?JuMa5SbfMepQ{zUu#Qp=^O*
zG;yE}kH8CFNFac4I|*GbQL-Ke`n1IiCpz&Ao~k}UoMJZZ0Q72BA6?gbr*jpn=@eP6
z6K|`}X<Bp_ZkPmQ{iOYm`I2==>+ZG5J2Bq(KYONOY8!!<(3JXy@#mk@{pVZ-I&pKb
z!sO_DqqE<Kcb+>^kx^&zI2Y*iP?$QDYRcv#u-tQ^rBeS?le>E84K3hS!1COeGs7AF
zCMx`J`O<ba*X*OuNgr?9mvBjY)FW#->roHGZ5+1iK%C(>aaa>v%z`v@Lx7{NkNuIA
z6=pr^^A+c`?5ob+(p-u#XF3Cc<*#tjPQC`d22H-i9HyZr2|!5Ppai^DDJ$P@wFL)M
zlP<StE6<-DeNOrt;~q`;#NQS8vQoR0aC&w1KpHk#xIM~bsQcd8a@-Y94tgUaL*}wJ
zpyGL`!hqK>Rc;i~4Z}(#mahXTy@YXb5I9Q~>BX-l<{OH_TyUW!ME>hV(Ry>|E%Fem
zFe=?tXg4g@%zHDgzQ41E@4I|V@7iGY%FJfT!Nikm(r2r29S?tmk#aCgIKaFn>v^T<
zJ567jjXl%9w2HR!v@xr9F-qgpNAq4HZy3jcyO>FUgo$V+vRw<7@tIwf#dH_J_RCFL
zEVyUDMxFB+`B-%)mxmHs)ysS|z+d2d^A@w(JNU(04@4p5LudLleIYe8VQbKBo4FIV
zaPcA8#ZaR7602Dgq}t%rE6cuk7LB9=%wM@M{)~t0^(CB-kHfqa0kJvBv>}z5^CTj8
z8X;YbIRid4;8Oct7%UaHon{Lld=2-5M%73+ZZs6pqcG_`9iuf93GQp-SgK3Ma{?0q
zwZ7GN-#rGJr0MH>NrPM(9!Hg*t^2;ip-;3ts|tYGA!7V&Q0doP+KX`0uiH(q#>fd4
zxmQ5QL}P&g!j7bR=at-jgl*ry5aF$hE;B7JH#s80G1Q?$D&yoO^ZA||ND3YaUn^e4
ztdHMKEmSFo`9LL1GjB9kf*TzfP8IcvTh3uNI*cn3wB_@Xd?-I^ZUT*B>td4+xG~!F
z6qtfQ?U8vsDOT4GXu);CKPNDY7XvYp6?i9bqwIZR{`-Io>VF#>M^12>zPqGufw+Ys
zCI8%`SZUSlU9P1gU5)?Tey1j1bzt{k;>KOyF-_(!Q(%&L((V1{iZ&zK<$eLCBV%C#
z^*O$l7qFJ?Aj!zC`lINH0tq)TUrA`JO~=LWGy_b&Bd9keAIMls_ZE11WQ3Lc0csV2
z=V!lSg0z_h={>GG4VpE1a$!WnP?yQcOD35v{EznacE;?kXfmylr%Lnn9@z}bA|1?(
zvw*j}#;>X`MeY@phodgXJ~AwQq}Hws6`mka1S+o}<S2WsE0TH<@ZGdf<{X+dkJXrZ
zokoNpj$;{?9QHJ-u+Vflh`Hkws3B#b`;T=5>c74F`<p^v7!ydcT{?Pc#R%9VNhUAN
zYDWFO{^uN~#zM$H=SaEHe)+@-f!(PGf(0-Hw;n_<9&|=dT2v>jH~>Y|b-!_*@9~61
zLJFK4HWLOw#v~Cyj7jnRyLXpkoJO)}=iXR0dTc8{nj16WTKxOxXOitCP+-Tux2iA6
zh+)29&d=7zN_L-jJ%yEti#}BNUZ3V=1?%`Hm0hABYuUoD97slED+yz~I`3S7mG#>|
z9x)O8>rT!-X2$1NP0pIG+fxd$CU$aVFkY-75rNP{NTmM?z28z4Q=nMY@e^pbA!Sp>
zeD-Biytf#+L#*&NkXn*T+GJQGi~t+#tqGF#GUtT`uWnvr6^<EY7y-}S5l-Aqz+jZ}
zo-4WS;$uyr+c5VOIr(pSJ_djw<G&8Hmc2yB+h+VTT*5=+0CF%(x+GiLbwmLQZW*5z
zO}(w7O3AL(LCxCT(4XjaO0!^jL+Oy7;6(o$X_tT%CZ-|=<Ng?^+VO1mqgfX_V3&7X
zd3Q=1B!JeO(b~|%|IhI<hbe;Q%@b3{-Sm)ePd4~gx#3dfiz-wsdTn0kdN;H5P5&}4
z0LeS5?Sbs7KNsFA%_JQ#I;Xi#JoA!XGU=R+3}xoS)+W%STonZf?JVi(#r3!7>>w`;
z6fmU><(6c>eSj-)FZK}gcYprTm$0K0Z4No_)qVhw0pCBpo^N^7NECcuoT(G61JV^{
z<NT@F$9XR3zSPHy@_W$y^+_VTLM5BIIQCXd!@-5w?FuuiCQ`lG)r?GVBL*=M{BK6P
zPggWeUw2HqPpL87D1~tK)lVQnFw&H}bT<@t`#3RHLEm3Xi|y0UFn{Zse4iooaK&Y|
z))9na7izACKS$xwkJFR=m_3`VYg#fpQ5@ibb?6BrYq+mw>UX@l8ch?@1&EV%@NUnb
z<<gTv&!>t3-T<$}=arTs=?oR<>{fuGL*2;as{fDU(EvwukpN?`&<LDS8Q;A^PG~pX
z5RI<IH278utVm>;qzqW@Hc}3Mn*!w!u_82ETEOj)0RXNs5|PNlW3-K`g75k$=tDP@
zSjJ2Kg&$;Ui-5{-fXxl`q>IeE7T|2*5ZoS3OC<H&J#kD);pXwS+6k;tF;n1?zr1$C
z+-_~v>_QL^k3YLGnqySeeIYBo2^yq^z>h-_2$qJ^!6WJhZr=^zB@ic9yZQ7`OVvHD
zbys{*wByz6>8&cOS%^hMa{rG2Ad1%2BNXW$&jXj5v6;cZx!gABU_DIv$$SGYwv*Jt
z;(|P?A&CiA_a4o8X~)p=-h97}Y2^`N9g8EOW@DBQ%I?LdzP#$2T&8wWOfjIK@xovU
zekM|f-kACM?)+(hOnT~MD07}J4^m5@T$j5U4#Ff+hd}}Uro-+SI$-TtU+t$-=Y4k5
zz@!j}J(2*?`Zw2~2)iz{%DtZfo38bAgp=e{n#D5;n=Z)%3I*hp<uFCS$8|Pm??->C
zY)eoA%?jPx+aX6!D~{#e(tLgIWB+hWH4AAB&@h=ow?)%?%Mkzg$DYQ&c<0IQ^mWM1
zv1B(X7_MH2YB=R!@5}k$%*CmYKS!^n+9Y!4<sWX^G_aBd%dzqjP6Ai-h3TtLO<pU;
zH<U6k0$!^w3LB+q-X@uL|2`EOCWbivnt2zuX7A;>rH2>0%llLYlif&|(bS7iQryEa
zcaek}T)s#~N&&+wg!^lf8o~ni+{3`vm#acng6G;6r?OpuDGd2}H~#k32T|7QdMMhO
z8aL9J&*0sFBxR+$7Q`u$3n`KqfeS(hEL@7P77!=p0fPx_b>Ps{wgS{dU(T0#VEuTm
zjHT<s`>~_U<xJg62Za*(FJ8i$_(B3S_-U|FAmr;*zxhzXz|(Ee+Hb~|^kaz7c2Xfj
zWEJicr|s@?4`|Rmp}2#z>bBpP2ZmXt-31Js$@|RTVE%=pix^li5KJzxV%TztTw7!l
z-*@p8-I@o7H!7LCe)s9#14gu}6?j@C&8jA8GO5d+4n25U0dkXD^)b5?OP7UiRhAal
znf*~UiTpS+Sz)a4JUm3&2@Qok0K^%frKdc%3{^#x#!nsAe^m&|D+01i>#m5QQ9ia&
z$q;z~tXCpsRjAlrS0?yMwQE40gsW~}P-sASZm4q=!l7sj6Ifu4(jU-6(u7|EHd13$
zh`0AXqiqilr3UTH&20Yh{*X+TF#F@<rAR`}m9PnOKu?cA1d0EY=?9AcB6JQ%i8;Qf
zr&NsTV>(cC!)((1U~Q?ae)soN1h+W$5TfmIVX1rk_r-69t`+ga2~i^C&`ib*G7o*f
zfW1Oq{-i(V&iS`Ww5XN?n3N$8T6cRhZd;o-xX%vd%3QH(@;us}E%*6;v?jzR>8f@C
zHeuxm<e4KN^t23f@^s@M{jxE*-rqWCX@{ZA{g>B=hpzqil{q(zxT<w7o|Omv>6SmR
z@ERX{S*W}v^j#n>71+}x32bt-+=Zv|vq1j?b)}nF=T_B>U98}_{CKB>G)dt{)2YS(
z4LdtEAy-F$mb_hUtu*J9;P%RS<f$3WP=U&qGB#uG;1L>(!nT+=_MvzR1z4O>PLLyr
zX%!QVVv5P57GU)IZ6c2zYCz(c$Crn5q=a18#DGEHdN^eXfNuxTYZBJ<hg(xcFF6$a
zwl93Ep9??==Sr><cJOBWC!vx6+tvBq=KWRo@XprM3Q$Y(?lWR@yHBp%{X1p(zzd^a
zT)#Gyd&Q!Dea1fNV}eD|KLs9kQEcsMB<$?~CX=Wax9djybVmu2pSw(!8+8{Dm;FB6
z`Y_8gEu6YHgMfc7o{2^>69_N^Z|SDHL%5QY=6BGM9)p+|xL~a}JgkA90zuL#XaIbR
zh(Fs**8WSa`@?BpXE6MMH~;GgjxZ$K255B>5th#3@unqaHFh9|@{YG1Xb&CKL2G8K
zUc1NwUH<iXF`Li^`pevd9|=x@n`EFGzxVaSIpB<W`aDQ%tx<B~-UUqWVMVc|UqAp3
z{JWe(8El)>dg}aDQ+0smQVOo73s&??7+-1IPAJ<@Q#&IWX+7Z2S8qEC{^NqGa2yjs
zsx+|9WBh}6Q;Rfy0zVaawvm1g8{=glwkxNOB%}IAlc4HglT>(md8>U%qP78s*P?rP
zUzc^z_Y4cQDq@2BV1-ckqxtgjP?AQ2=tY(D4(vX-S^myPYZwvTw%?mVa|ivtM?ifP
z4Vk^Z#}LDeHc4j?68vwgl0td#QCG=bd+`7~HuYuZeLsjau8V1@NBnDfVHYZ#f=M2N
za%BVgGbx)V<Ou!pyG|V1WObRF={GlW9Cg;Wer|jWUNDzfr|3L$YzbGoVmhD$ck3`t
z`uFxJ7=@eu7XeLWfeHh8WC>t-%f3&@Q>^}_afMujA-@kQEE|8YIOi2{-6z-SB=G|_
zIq)P~0c-+Rs7XRXyUyk_DDE5b`F<mp`bF|gpUib<JqZ#ZCu{oypl<T}ci>$#I&6+&
zCR48!twZv>m@w?B65rm1SjF+@u-I=x8+%)B@gP^D)aE}`7(WMm_Z+bOmA)kYLZYYb
z<+cU>%VffeL(E1+7ybh@teZQOSYWvGG*^{J49hNV2mK5CSb3|9nK8xu)2|RfD76*B
zql69C0jV`6<f46Ij6hEJfnwN9w<&4|>3?1_=HD9v709ky70SF0@xMPtkoR!PJcsg0
zGJ=~>=X4s#u9d{|x_OXH8})NVv+V&{*sXJOs;Z{}1YMsPYTG35O|GXK8;m8^%EI1&
z!HR2)5#k5~XE4Epru$$Uk;McggS}+wf0GS+=+Jh?JRxfN^+Lr0nSG47BBI|^jRcEI
zus!*aDPTlcl$sq+{?Qf{nkxJ;C~6B#azxLD)$_t4!vG)9aq`?k5!nGl5eUd=f`)oq
zX<*oib^DZ_pI!{e`~=PvYc0<>ii(sszx2_UkYCx(Pd7d{KKPx0{`>fA#_g)6<yfFe
zmq3H%%Y_JzWA05?<?HZ7y8gpvN-dl5(rZ~m7yFfDA0FuoTm&F!FS9u~9P6g3u}l;J
z==5(4E61I_2Qvn*9!WL3DG@SRd3FF9@x6B7AsED&FYxAqxr75D(n@3p+DBSUB@xTB
zbqLy`w99?BxP#0i1l-n!pvM79X)*`{`o>m&3OBGR<NC}&L?Z1z6PoPuK3%ZL50&sQ
zT}8smRgp$vBK;m%qStPeEda|L90an@AR!mv3>;(7s2^Xx&K`%D0gf(+8#a%5_8tJy
z?>Fnn6nDZfrFY>4788kiE%a*E0betPu}4bd^ys=CsT~U@S$g3o(bPxF7Qg$TKg>9f
z<SS$zPYId6beVWY^=JFf`3W~)`9n4ejT5Bmv(~utlNH-<sM|kUgDo<hFX)^JH1O9C
zeJxw{qsUK{ehIdeAG}SPy&VODfC;~U75f%M^wvQB%6thnRyc}8zJ?(k<{nPp*+|3e
zw%z{Rj5*+g(O;-GG^t@Xeah!^t#*3$cIvFE(_HU)jM@(H<#brpR|H0g%w?5!yXo!%
zdhMVkY=_Z<y3GxZzNFnqJ4kb7Uho-F0(QVK*k!q<qAGMa$OKW+u1W7$8|`1k*XSzZ
zFjW|`qAxyWM8C!^Rze;BnQwJz^W%#x%yjjGT^Z_?9H6-Ns$U`EK>okg89}4mPv6PW
z075C{8;IsLe%2tp_>%>>B{82{Y2I51fje&`pWOrSynHt#%ke2)b_O#W@D-jbo4Bf9
zQ#XA$`s35<t!MhV9098c{p2JZr7t8;pHb9G=gGZcJ}`k~!oe;sgulc)h{GISPE%~Z
zHpv}4MHcWE^ayxR=FYe0PXjCYw(BbCv<J+`dm41o)CLT+Kx5b+Ry-KhzzyJ>LV&hJ
z5w3rEhj~^G>&JfwD)v{6+<WSRbUrcyMX8%3_KAn)sSvm|F@ZfCFN@a{wVxsEd*-#$
z>tC{xc*jOzuW1A3DgPB?*zxY!Rn&nc7K}du5tR4+jbm{0b}4apd!6lKd)P@#tV^24
z@nT0r4P1{(Ywd`aX+RCKn_MET0C@5tf}w2b%6PG)s^sr$o?8I2?37Co`iQh}t+zhk
zhPBzeps7I(Jr}z2^|W=%XOEVzn+^WE%g4`Z?xX7Ts^o8Cp2DH9_2e9T!WYFr;qyj5
zvNeW_!NOwb_&sK8djOMV?kZl8{a!zTM{|Eu(yDeY2=T+MgwvJ*N&z-q2=2bFaW7Y>
z91-hlkN~o1H^bd!z<2{pz770{&3bisI8A`#&w1uU!x;$z1@;ObS`=f4@R1s`U#d1_
z7XvR~gKGUH#ZvtS$PO4j&3TfJz7V24z!|V8YPU$VSM@6g)2y4Jq5j2D)DqYsevp{O
z!j(OXWH_!zu+5K^@1LPo#SENb6-tt9#5ZwoMXvevC9+#yEb$u5=H}TfxMeWNa2j`2
z*g$BQQ%w}uiC13I9&_g|)SP2Z*+I`ZjnCxXMeny4)YL7!V`gs;B_M(gIc2|aPazOj
z?42JGp{(6e2O)40UTD|&?xtA`#I-FOQXY;ADx`_1c<K^sC%)d*Dww|b>-uEIi>f#8
z>BIL+(vsG}#D$h8=_z)`*(cHyB};Pe{`zQ9H=I`tnt~@%637tIz!U<N0vIGopV85T
zm1S=}o*O2h8hVFfEM0tmwe&sbbJK5G@7;Q)mI{nF&;P0wj<96ZT|TGv>o{kOUYr85
z{V6c}4Lh3{jK<`?<y{RtaFISUqA>)D71*(|lb3G+|El&1b$J}qmB$yR@|C+6K6)qe
zFKBEtOM93;Lpw;+dKd=b?d1j93mhF`b8|)uI1C-N+;>yVHVg@ZBuV0$7%qgvB`!s)
zMCIQud+kqzZVw0i1W`r-h=X?d_H>u&yX5Gaxt4iIazfQWy1U1lX{C<M2V+ckN&O<R
ztH|~|9y`P(f48XwN%r>Dfi&)Mf;`{xYZ##K(Vy=CvUjHC(*c+LK>WATipQCjwL_3|
z8O+Pd=Rl5Vz69@6Y4mP&1twNg7T9RBv@id3Av~U|Gq)44|7PGGXU7>X48_0D_#^K;
zoEOU=<kWm$6Y&imiN2wH;xPZ_VG3P7aO$G)J!Fyj6X}A+<gVzaQtSkZV@IDc4^&L8
z8Lg~Q{1f;MJ+d=#RC|#;nNihj#bBPCPlpNy7j2esr;b>3hPQ5GHNO)!Ih%p<HLG(<
zXj$Q!;qT;j-~|G88K1+dEPQ`^d$feng)$d=ZL|Q82Fzj@DV*6+DQju@7vNSyS>i;^
zauJ&FNI({SkhWXbwpFkgY|$JbCBWgl;Vi%ss4da7S?~^Idg7z;R}_*j1}BT-M3-j^
z>2=RkVK$~alPVaUANMB9!`c&n2_$dEHSyz!Q^eoybW)(j79G3mkHBkH0Tv&T(M=u4
zC%Q4-`4p-Pul?`G6;BZ)H~8<b4FPxILsY2R*NTgY3UZ7u#N8Z0o-8_%L=EAYol8{@
z^Jfo!DdRDRQzL)CsgC(}T^~-u=-K_=g?NSt@PqK4NLf5C)!bUKZI~n>#k`dESIDP&
z*F-+ZgiWc+`pj{s#JYh6q^<c>j?+=wd)2*a=1NUDEm3Zg?Eb}g3dn1ht`^0ItT7zO
zN@%BG7zHDgRJCOb#kt_mp#HtdmD_c@G@UM-5g(=Re`?>bjb=BM{r)n(HHg`vXhoz|
z%C*A^D(|-~3sU-Lj6Q=_qLJSEaJ;2OcLQxcl=K!_b6XQu+jWhBLDNuN=%&RW3Lj&0
z{{9^-9){`-kn*wA_=CP@Gdt}N9Ff{FOlFTU+m2zgp8+!}tG1Mv>L&(&<Ms`;>)tmv
z%5sR|@}}Agyxd0%2@LI@bVB{Mj+<?|V%O-JiL%X76biq0KPCR<Da*@{S-@DDcmwM1
zrV|~xPrbJ;RNYET+N+y!`6PI%gmqbk&`F#Q4@Xa4N9ruIIkv@ch^#(}%jO#|wi?Cs
zphRc^_gm_>sH&rG=kf>BLS}f71or8~>7?l|68E(-6%moaYb6eWGk0c>Iu*RZrK=^D
zFbuKvjizQ7(*$=+kBx#nI6K&wn7vzs?Rv_pM$;nP?Wgt`^rSCe0=Go8Swf%uh!v``
z!Dt?26mtG5JiA#SYUMhH@z=l%{?>>K&tj7vM159yjWZgBaXtTajzdU|1r#G3A8LFa
zz+O5rXmThq)U@e#l9_WVld6L+mh>;TzjI0vncR$l_{7xZy}uRp;A*t!@Mg;6m37C_
zLbY->FvO*7*9|*Lw_UnnEU8l`X~5(%KKEpHySDFZ^7a#E-;P3jSFbN7<t?jjr@tUS
zoq(Xp3`~tA_N&6rR=RK8Vi;gZoD5?%FU7izUs}r0y6d8RN!lGpaNR6{znNWHWt}jk
zjq6Y&fN|_n#{sfJ^q%b+2_J3613UV~I_up;JRUvPv~j=E{3Cn%u?3gxnqmwYS9nmZ
zhpiaS4k>_D?2ORBQWjwlImK!@L%#)ADfJ+K3%_p}7u+WLijrWG<SoxjY3)<LL?N3q
zJXrN@C^wogl&jq-`K7i_SY4v}T0}N49y>}<E%U_p-~iph{mRlwN>C6^Ae|68=s_^e
z;vR;$#4uC%M|QjzI6HAAqSEj3O2j#R9%y=XRQ?d%SR1g~e!__9U^gPlf*Mf<Tir*E
z@GU5GKB#}6)u*jZvTQlbKlb{H$c52V49DNFh~y{^`t^eiX`X<PN6X}Z1@%Qg@_nDk
zR^qnOO%<HtJe1Ru8TDPt`9U}}I@+x83F6LcmoY*ikyY6lmYNvK1OLMJ<yvF*`^3(Y
z2v7Kt)ac4hISd0q{ZN0*4!Jo9s_q4#rz7uPgs$0dXK;uQSLCHib>bAY3=`^c*1#<l
zC(mxI|41NPN;`l%mTGOrJ;W~zmA*@+ohkrk1T8%Jb$+{FlqB!S8yGUAi$0T!`>W6<
zW252$1+4~GnG&E*2vTkStGXGCG5oA9&_7+RX_)+M`}w{FL<N_d<oS3k-CP*P=!HC@
z5)Y+b<$J#w8kj5wCwCtVuVHVM{^KsW+P`hSW|lEC?QH>D&<o}FJbY>KdJ{lC;loIk
z=BqiUAFSWA7#^7km$-Mc==%-CKeOK-_`nj1tN_;Ea4wD|g-$YY#pP2jP8w^ZajLtH
zX)1hRmmYcjb1$-8-d%qVbtJUT6?n9LO_3FqTInVH8$?f})NHsus(Y5oaU{eIHcCsD
z@e?c`V9!`|+B`puJ1iGl_C}B?UHyhZVc~@q1hzol*n*ivFX`2?c1n(t2rzsixIUdD
zdTg@9bTNBP8MmWrIKVWe2d2gYF~>{MF%Nq-iZ6VWpJJ}bb&Ygb4PG~`gPtmh@g_|(
z`lGGsq!kTtt=UqEkHqiA2$~<jW;LjszVnEkT|9rr3Xv9-(1;*ff)6|wiw$Lt;%<+|
zrkw*#V~#R&FG`zc+|ckX;Bcb;B0B}udx3wED)2CX2KCpj|DGD{XAF$cu6)2Vy78F(
z(VSYpTKu1#vyt3mE1KF$pUP!=J>rXz*NwYh&88I)m4>xk#_jtP2$Gs5zN##4<3$4P
z(a;*sON82xu)>raG+dQjCsbTg@%-MeFDkI8s2kiA;tgNU+w5nqFe0ber;vS6)`DO{
ze8o~&e`5dM4*lUehFYA_wu{+Ni>`xBO&-n*Cr|Wp*!E8*Kdrg9U%AtlIz-X66t;7i
zUSj2_%A#+zS=-BUxNhOWfP*|(262E~uw5y(9(KyY^o)g-xw}^9fCc@=rZp?lWDyT&
zQ$SK!31LH90;^Q#%jrM*!pS-TbF5JRv+v1_a;TsMra4TvSra`CJcxBKsP-%SRj2sB
z6_;c`)|vM`piV>S`6tH<U_Z1dFt&rvRn)j>*zu#Y#e4nHM*h6m(4N$)dzY=2nqJ;$
zv*8z{_f3Q8nOOo>mi0+G)g*Ua#Fae>)C5<`-KGbq>t#A|-c)YB<EY!D?Y$%&HR}*>
zsxV`K_!tQJ>}zl4CC_1M##vE&R!MB3%%A2Ts0~%A&P{Yu7OK{E_?GeeI#eiuM1f}N
zZSU<{)#wgkA~#BuE%(*VJK{o7Ck(Y0BWSy5`<kJX=rXdKH1(=xnQ-YnaIm4vQ8GDA
zE+Zh4G$2pQC6>g8g~&d%YeAA<if;}XsZhs=G4S$E?Ad_iA{Q9DRkK~VuC&95zj9de
zE(kf8f!OilK*E^6PrTYQgTGA5js3Z>srKwxe{Yu_GBPjV0~MO_5%Q+OpmZ5RxA~qa
zRoow2A7{Nk?(OvGCM}AvRU+?8a+$>IWJTJy3`<HWm12b=;{$o*CM6O{_?&&&gi#DV
z)mk8(MyXYf1S^4Z4P&<m%7(3GM@%@@tE||Kx;&XcJkWqlc?!CwDo2yDzLkt!*Wyn>
z?_gK{XddM;yLQ#O0!j^aY<ExNQC{!0<Iy^=is9`38zw^9Fd}dGp^D@+IccZo7c}5P
zmg#Q6a$W=m>tpNi(_QLXcHU2}n5wE0p+`5RYB;g+>w#o1932V7_xQ3NGCUU>=UaT#
z8#&OD-^z(*ihFc5j%f3!_0qLZ5wK6PBq%Oat4VyT21&2Wh2zDa4<E3(_2_xAPFyX~
zhdu3Ba^>=-%!@gr08)nQ_wz5--j&f-_Y?VbMY%ysN~Jb%j^bQ5Pz7VA+8RD||ME&~
zwYZ|1K(f5UOcrvtM=(Z(2aOI0HRx8d@(J%e?LOy<J*2PuY>+QsRG>|yI%<0KLsIk4
zL2=duk8fWhv_)}MAs8L-sGl<7cNx6*znAU$q<5n@ndGZ^9761l#Q~VFp)D~f;8sc{
z#)a{~CY(R*%=Tca$;(9fIVI@ocE_WoGimqa?Z0+wPmy_if716*&FZpnE|QFj+|SG_
zr6}NCKP>BZ`DpCqy>bUA1JhJf|DRznM1WsFr!WS@+t81rv`Ir@3+G24BRWq#VyKZS
za}oue$4rM;B^*uT2U{**pP`|H<nm;B7Yow*4RVWd+c2PeoL=?6C%<e{+%aS8OFK0K
z#%J9o;qc-+-GjN`?a{TDa(ZTxMA!RMv}@ODZ#d5XD%SfD)O1o<sg95RB4OFPUVNW8
zZN8F4+HkU{w5EH9rcW|*SN5#d^_gQJCA!jNI0|y2EPkwBA8V7yohP(3d`rbej%r*u
zVS<=_?9OPK@QX{BkiYP_XqR;1MK6}F-{?^2n>=0|GwsTNSenVHo2Qs>ap!#C)~jp=
zTI+B_tkHyVI(91Z6$OFEOl#C0Uu=E9eM7VmU~!?OlRrebeIeAZWhLQ!44fZLcpE?9
zep5UX>Y$2gqIKn<JPeNd%su-RW`K5B@XQs#OFdKiB)f_s<GC(^7vx}aHS(iO*+3b$
zYZ5z)6MuGlQ^@_M>v#too6^$G%oo0Mos4fD43@mSiofG<yObe&`|)V*5(Jlri|m_$
z@1}*^Uw0*DtBNs*`bKD&Je`a|a1ndoo+OuN$Jrjw^w_<LLgkCG7^R^vh=YKi`mBt$
zKEA%15<}oqrU_EIjuKtVG;dghBF;oQy@I{D7tgi^7)L^#?dF~Ns+gG2%i%lYZxCyv
z8m;$-8HArsTlpjeAJ2*BTy>^)CX*;^=6!Pq{`)tWHsc925_kaif?}@>+7)3*4-GCR
z+EM)sW;?qG$HZ`6MMVv%`+_+p_unaS(5YfNM6AqYO?qiPdT1U2xz;5OqJ_2{<Jpo#
zow)H1Tfv`WC8VO5tO0?k`7w5{$(VO>RxA?Zf)hVwis+t(9GNde>qdIr`Gv2pd9r*x
zO`2f&>k=7mXXYi8AemXDq1Xj9E~yTZCo+`0n{6rlnTDH<|KZ2m`!RMDmr`7=Fky0i
zwV(S^5ooo>G&85{!4YC67HGWUloFLpVTA@!06gZwgL}4<(}ApAfu2KE^!d>ZEjOdR
z)n&Q8$}Jm3n~xOzuMLZYm*Rcb-;t`mi8hu#c?H*IJpNZ2fpsM?UXPS@h@z_Wby5g7
z+cOYS_iBsz{j9Sp7qzIziR@y)Scdm49Va>Q`(#eV%}Wfkv&dRiH>niHGdJu?S<EE_
ziG;Y;=OQ;)C}qAzIb1YsXT${|9+4yKWfxYrbJrfeKvS+>NuClZsdZ0dr-Ja8s@k<K
z2TBEQQ?rJ<&=(VQWrcbb6-kZ6+SKuNvOxljpX4dW1vlv>Ns^0BcJ(krIGrnex^Xxh
z>8w^clq;2z=w4f({H3n7z%lE`Cyv<EmmixiuNx=UFfT+I#<dG9F=}-eu5NKY7&Shp
zi>O{G%e2-#Ro+vf-{X@wU<=e7OJz7QEY$>uTw+C{dN&}2*_x@V(`>iMWgr8@*9E=?
zhu~A<cYWE&nYg6YXN)n<uPADUa5<<~9M-INF=r1KzIRo$l=92Jm#P*TyYEj2Q_h#k
z;ZlaN;Ig@{H4ZUu^#C{PgE)tCm09YQV`w(CCel!|>)M+A@{J2SSDBOY?I++T;&+|y
z*?xiEDr$US9--&&eP)vw_q47ZITgiCM6msoizL-Ih>qf$0mcFr3E{k6=UrU%vy(o`
zy(w%azD6SG3M|n;-#Jp;-SiYuwo5FoxelSwIT3&L%Dj$G)YV*0SDh<nWOThZw_a3W
zxV#B=sG}39{tpY=t()p~d?DsbQbyp7y9*B_tZa+c?v>r*8cs@<e|+lysysG|Sut*m
zm%A~c>=#B+o);})-noB7U^k~}ad6iT>+@8*sl7uvNgTJ1;CH0-@Uo4Pyb@_}G@I2?
zcfyVQ1%jeF;3cYC#QmO~t#jUI1tB3v8Y=INy|}@(m6Wdx+}x|w4hH%yyefB~S(~^c
z*j_YpP9*+}&<+{n=_TJ9Q&W;+MwoY&r%Ta+Q6FX~edO@!sf1&!`ASqy<X|eCQ*>5~
zxogcXyNj#17#vMntt@nw3o_O#H6KZ1Imqc|T8#+;NyH%=J543d7F|s1UbFw0&ipb6
zC73#+UT)S4%7;qY3<0qa4-z|~0by6siR1Ajenz6za3o_c2+xq)7w5tDArnRz;*h8r
z;xBgN$4%-5sv>dq>lkd@ah%<i5@gQ9!kQup9yR9Go~%1@E}Nxy)wb2A;s?_BUnmL_
zM}4SBvLZ1;n2#84aY?_gP7!#DdaMS<H;Txao(p`Ed8^MC0J5@d*AJOjrW2TS!o~XI
zcJ(Kz(n`ndlk0Kx^JWd;D8@LZTRhwyM{`}mEmqvUDrNc4b);oJ5zn{cJa0XT5xH+{
zk$iz`g!_f9JC-2e<tH&*`rbWG$RQ3B2s+vx$nCeIO(y{z5NQg=4?qsr&iabTlbLIm
z+QK1`q#}3&lkgHc)aX^4iVwZh=)wL?cCRq%W8r&0FjyY4k|48Yi&0PUbwx-^tX6fN
zqQANjf2Y_((nj4FC8c<oq)T_zyR`rG!^5yw6#;WGBKqX3z*lfxxBh?xc{8m_TaA5R
zB|&$(wLhIXW=vBkQRFKp=RXKP=ljin^PX08sM<Jck0<$1T3(>ogJs#aDA}6L-VB9j
zwx!(a5TtG<R>v%j`A;^EiXdk3V2~2WfzL>MqY&~5vt83?xP;~1jgP7Z&T{HC(vm-l
z&%l7FB3Af;YYMS^s3K32d5Yv6(nT@N&WG_+F=N>o+Cvmx_@1L=w(mI;Eb;;U=7wR&
zYcy{irKr3gbN!qEvoPU)nJ%t~7t*I`gF*uLHn%iZF&Ae=3(m|8@69ePy$j_<^wOI#
z6n&=rq;7S1F)NeDkyK>y`riHdH&UV!vD=Lr(bNuY*PD2qFYD^Lr%t6|g^yG<D@+`%
zH8{``Et9P=*-)Kp@vtnENGCP18$Fw?U3X=B6_@gZh#V7pGi8V0m%0LNdY(|<r~(G|
zd`Ti^tr35-V<g<9)c^M-CC0yAd{ziyx7^SDD%`jea85S8FIQk3MUy<MC?<o`f#`C4
zlz70^K*tUt5EmXL5Nmo#n`bI(>quQM$Sd3CDJ&KB3BT^pfBd8<k^F_);dh%4Sl1aa
z4#ZsUHYN0Hk<R@q`J_HN1F>ft$Lu@w5T}OW)16!bA-y)*a(yUuh9aDs;j7&jH9b^%
zjc}jvd#ATFkslrS<~+No&<#_Vp!pZKuSg&-f{3CKx)oLjT~6<QuK0XJjv3A&KWb6u
z3wnh_oD6zZ^DxO<AEv?Q%ykkk$)al(nxcwe(U-&?s^*HYi+;>;9@s7X^7gQk0&@l4
zdkAbslt#Iw%Pw&qP@E%#B$$)0OeWavm*H(O!Wp!U0@Kj`8x}nB5E&$PLmt9M;_Y*E
zeh~(6P^&esW&HMjX(HBmV^R=%pUkf|tbup4ujfV*bh4{c!0?#e^pHn4F6-;eaG~5T
z(F^5=qaw(@m`((C+i6v{O9MmhDz5ePH-0;fHgyAE;{p6Ox=-|6x51cBvVvZY<N$w?
zsj0n${g1?wGPO=7;do@${tfyE#Tg~;q{do0aqvvK`7hM%wot#=)u?<Kf<ry<kWN(G
zvtSByq$N@C5uy;BOiZxaE9$#4O+}!GJt^L(uHK8e^nkG_4oA3w`KIstncg*pLUT=%
z<ATH%`cN<HN!L9Au5?(`jJPnC5!%5cPS1Fgv0Sp&V0y6g;$#Ht#tayFV04K{0TgF^
z4tRbNaJ}p_zA9P=dEn_&T}KWn2Q7EUpcz$%=igy4l&_bUUnD<1!J#gr%6W?G@u$TU
zMSbUWY|$7P1u$w*=SedrX|D0t_<e5d(+6c|X^2$_h*QvP=?`wJ>-uV_Vq6iuQZFO5
z{uGu`%g47TYWFKnGe2?2?`7$F=A9Fu<TmxPu=3rY)`8%XVaEmz7N2PySfOkZWk6PG
zp&n>|ogwgpxHRtDVS5<S+b`XcduWSv5}nNv#ZnCD)M5fitxYrMw_0|Ik&XxJW9C0u
zZ|gt{!1OPxwjT#n1)M8Jmwavu3E~*JciayQ7yRv=b0=OS3kP3_I-6})HQsP>Rl{6O
zn1?c69D8xdLBFPNV3k$XKykMlxU29l1YH}0<T;thRg`FGOga;_u=kDdQ=l*_cQDT;
z8zHlB8j8)9Cmi-_;6aVR+@8pXMwo(FhTPn<OI=Rh=@#Lm1V&@v^iaBm%kPIE-q3+K
zaNez037?n(scWW%U58XOc`+TFUScaZD?=@X(J`t=kn-|BTAszX0497df^+wT(J*f2
z;#JVmvSb<mx?=T*KL8J&XQYR3U<{|1x#eLIuw1Ii369PYkg~P^x<Unucubp1F2x?d
zs=W_R9a`4AfT<!7^2Yae*nT0a#|jC4`fv(~`6bwz?RnOdfU=gD{FT_n9e%%Idb5T4
zg?3^s^qcgU&sAY4?Z?Em@#5Rye;NR1MH&ZZsQk=0zKs<89K=W}#M{ONyK!Jny&_PT
z83z)L2@+rh8yyj(_JrH}h`LJg+W`atR70rGG_Q}3xkYqC=-w94x~7mStCfLww*#4I
zvb;$9Ke(vbbeOCTBw<<GSx;_XcKdM%l2n%hKFD1iTG0>?|A^*tFb$+K7S(o<ws>~2
z(Wri$3n#`flhGH&zNh2UE4O%k`HAohgXN+)#U!vnRJg+_neHVg6*CoQ@15CPV%J;=
zY#4Co`NLP~q%^R9gF(1mr%+}LvvSi|2r8M$m=8i&tq;~d2~v6;Ia0m&<{3SazF_|!
zbIldJepbJJ%;&m<Y^#C;4eq4c6;_HM6)~eKeS$IbIoo_ujk?Tw)rs;!G3PV1#cJ41
zr0WG~PL|wdd|G)##?B0zB8A%WCE_@9yxM9;@cp)I`zX}@F#(tPcriFtHC18UM<e=s
zIt($htHB=-JX0n<ZFCz&)Sv}v1;-U>{qTDd^Jj~NKJ0{(D+BH>bLEPmyw(c6FBp$8
zUpffm+YLC7{oGDmg?07oz&6*u=c?~MKN&+{+-f(&^~shp_r1lG+tC~tI2jE5aG<Wv
zV>7AOh6g%j#@AY6ga0r0pBIDP3c6w#J|+fH*NaKW6gHq`iHcRnQQ!b8(sfQmJ$;WW
z2z31Tl2)B?vd|!KE{4&3Xp|gdW&x-BWVrlx;QvR{d&g7#|Ns9S2gi1dWAD93WM&>5
zA$zZ5Co(cp<k*hAWo8pe_AYymkc^V-Eu)3J>ih6{|9<~>QJ$Uic-|lP`}KCc{{@dh
zT<o0sNYd1*887MlbFnfkMq6>>+Tbtf{=?T?*ZZTlZb@QCY|0%|`+w)J2@taQ8`Q~P
zSdcT_PRvP&`nf+-g2y;vBdsUg12C=UBGv@tvfIZX`@-9pdiR#2n68NdPX)s2Nm4Ej
zaW<$O9CMs-F_A|lzOmPU2m)FI2V8}IHBsmdL&$w{BQ+7qes!;X8#Z_Yq5w|<0Iz~E
zn|1H59;k7wgnXY`rz+O=h+5jz83{F^e#*&%S3$1&FpihfUCdW{`X>J;tJIxCmnCW|
zFDj{a^1N0B(14E53|eKQrXwClhP=`K9fLuV_DQuKuVz^aDqn6<Z!HEr2wv1(v;Fr<
z9t7%1h~m4!9L$1%CfN2oDm62viD<mF(PhGF7;P4aI#9a=>72$@=}t-^Xy$JK!E$-z
z%qBYXyqNk3XgPowJpWm@AAN&A<~fh1a;zrS*Kq9zX8TW6Xlx0u034J45yl%y+W8qk
za#1ey2UW9^YKd+Luk59o)Buvt-c;TdA0__d<KuUJyH{XnlI}X>A6Se5RCI{s%}gM_
z6aOfvzBf6e>!qdQK~BJgQu1WKJz2}Ay9xF0)}49)ERyW@px9uAg!7gO3-b_MivOFH
z?_<>5+3T1OwvWG<`Yt}ebgK6aTvATlzn*>iW3Qyr82@dm6J7Z10EKdG@f(ZIrS2!U
zmUZGp-^aC$${3bwusUto-@J(_=##pJ$l1arCH6<zU7629VXR6avbhXCqd$$8i{+Vb
z`$`Op2UZmZN@6|#mEXj)xvlNnyb6DFRnL{OynV&^j821)_qH=;TRbm8YGbzDv4z~j
zORnnxuvZ*SY|5dEbV~{#JG!i72*2L2)~eTyS`Ng<ufa5gA_JzR7u(C+xAE8i6L%3a
zl5<C#)Z-b~yCRm=UbDbb9~LI2w#LWPb6E_G7mBsBY;O1AKX{QK+Bs@w2v2M}-m%xr
z@El~7rGHE^AcHAK52qa$p>1GeMSphky?M9vex&FkrDt*1mhPXGcLKl**j32#;kYlY
zo659Ocy;jVqC4RT#GMObc9DKN;aUP|_uV+Y7h%%Y8X94f{mHYG`%Tw)UQNv~E}>FZ
zN}VNT=BvfQ&!-@Bz$Y(ReonjF6#>aU{Rq%|#@hp!8-OTT;geh?Ox0~NC)_vQrA85n
z&WT)p#^29H6?{)^WB*_1Wff)M5)?hM`wtEx*HO_pgx;QU0}u4x2VQjY^ifYeEx~I_
z=!j=wW}^N3(XSD|;sC%f1LI6TO=k$FYTozW0m(0-C)7@CjI+3dO3-zxuo>Kcj(-Y>
zZ84cXKiMy-v~INxCX-C|9W88d@!wylT?SwIV@&r^rf{$F!k3pvKFL2*q6MVog7x#H
z;>?>h5)j~yM3N-UUU^hGQ^ZyX8x`NLH7p}o_dfSfeP0#0`ot~q`@_cf`UAPTNlR<!
zpV?+!efl|^#p-Fp1qdHOVUVskgZS9DV5BFLNTf;oc_%x0e%TfavMXi`jk~#xiuPPT
zlH$BhaA{_Vyoqa2Bj?EN6x&sp396pk2Il_sE+vGGgFc@M1L!fSNhZ(g0g8(D2dxub
zju_>duqH(9<SDTEls7g=(i;n!p%N@U8|!qxd7XuT?=!sV#-mW2$;VKi^ke34V1X5{
zf@v_=McIi^cfgR}f#1jHI{YiZ&j_E+V8;h#T%M?XLql6s<DwL7T*X}|u8RNuTu<l$
zjro238iV`Pkdt?RMC(Q_;+6(Yc*Vp~OEg`DcO-~;A>esSF1J-jorVmC@tfCy7aqyc
z8j0qEg<5w)F;<&3N=zr1c$J+u{MGw`qdFxri~F+A(o$?mXwA><?WF@?!bUG#GmLhD
z@uS7gWYAiL27Mmp+y3iP(?Upu6HAfy;#n!$Mmjw))o$=gq*6f?`9*mr$LaJx-*b}o
z?LZq!AJ%HNvx#FgUlK_Hp-5}N!E#fTUP;%7;d`M60QkR$sR?^;#`GlqMHwZrDQU&e
z5uz~1Hxahta(__<H_n|{h;4V0xJZV$PuNj+<U&u}**(6ebSx@I>Xf3m7RF|;%%(P$
zWXNrt6M3Y?$)W3|bWR>0s`Ry>2rk?Qox`f>-QIsU$36*&Y+Mf$F|*Hy)xD#$;9{lr
zO_+boqLV_8*=wBTR<jd+E3`+0$IL5oc1rW3qLJhg4=|`qIVzKENc6F%37D1&KI37=
zy;lPj(1pP2-P@XB-=+a_!cEHK`G}QQBoE)L88p19D~)b+HB<!T(h7KP|K-#FF^DH;
z|HT1I!*SkVd-lT+9G&W(#-A%Kn0MY6y%Bk$N`wgp+5xJ@9?rRje>uj1cN5YnZ&MWw
z5K;3&pONX-_V{jUZaJ2RcLB&july^;)I}CrF?zeEXq@09gRBpGgCKUwnQ|%|6)kPV
zM@lcwK4WZTNy!fl2k&Z&S`uvyg4+(aTLvpVOxOoo`R(6z>*;Q9;NKqW?e9U5@^``C
zVHP`NTWYTX-opzE`{G?nuFflu!=lL{%$ORNk<QOC$0_WK3ZO&ypS7yL6p+GQAcO>}
z)LAre8XAN^yh@s}EcDvUAfWn<8lwJ+LnJ~kM>;d~pR-eu+B_aoqwBae|6f(&s~cV$
zP`jhDjdx&y(`q0c@$02ipNOKJ9Oj&7U9TBkFO0?|K|)>u4Db`gF5VgUw!?x_nNI8(
z0GH2;Bgq^8bnTE!><duJ%l_+(j<J(X<Y)i;$>U8x^GZ~W?!?J#g$~{IAEOL{N7H>B
z#rj1`>r^|WxnH~;+p*h=dB8?Bp!dFtUQd{kNbD8glXatL&uHef^;QORj7RP9jRr`q
zzq|N5c-b`RzZA(t0AoeomJfjCGWB7kuL%l&g#Y^g4Fi#~gZ;H7KmPDBWlez0p*9XU
z6RADYYY>ZHmH%E0AUB%k5-;Wdm>&)d^|5>Qdzw)K-587x*IyM`AOg~n?NhD~Ut|9L
z^~|61WEI%Lkm#T%YEz=DVfo^R_W#j(ofx)Ns_cbP%W&l|nm}#^-%V92f21`hq-Ic+
z9ToZvj7+@&6+IhuIKyPV!mh;yPOE-`S%JIW&y8)AmvAfP!-Q=+X*wS4B`47<Q4?hq
z#tnIwCOojAE5*Z-Tw<?y<nUTGv`1B;VKA0=mpd3H1;|`lDUHWC)>Y#{(@n%&UAJJL
zUK*E_SyIojp^7T|@ROQ5ygxb$$;Qn?SU24fvyk@)GXKW5s}+<)3qF}ya4MXJBD+>k
z%Y`mpgc3l-kYCLGYLnn>^ig^_NofT8@^UbrB{PIL{vZ5ChCm=>5Q&Zye+x?Du4kA(
zzs3>HIV~#hQUgLC{Y|mDJlZ<GOXhnqF7$@o1)ZPyfIRL<>;Oe}l$ciU6VZs4THX+v
zZp*)&Jo?2aKs9>!+i0N`DQDmVp!f>V>F;Du6%n?+=8eHOSv>}k4ez}lA*bLTfc<1?
z%BI%~5d85=d}GTyBb=0`qbsy-KA2=f>fYY8^YnS@Q1(UxEAEDkYtn9kfK+OS_V7;`
z_1l#*1u}(TKs|IUdNL9aoGO#XKGSJPi+`s+7uxZVW=1<-6|VBW2Y8b3p9%I{q11qg
z3{D>3_9c)>gx#L~V=us#AAyp1XA#Y+U-IIow{W4>9|<Lhy`fMMg~Ac09#FmIpkUHe
z9bdyTF|Wd{o&kE*P$yx}17sUHW*hO#;u`#rOjN=>2lLf<2J|Tu43dL|1+;=D=$iJ;
z(Om*I&8SxFD$4uBMBOTil+I1+cXhp}g%3yoxFviYtPdOSVi`6BVVSb3^+@c$Qq~zM
zB1!ha3Zh~UMbl%<slVz@w2#ep;$0@Kp*);-(u34L2Hr%^mEMTV5^qhP4s`Daj80=a
z%Jyx7KD@u5Eo>>tG3<y+Ffxr^9o@q`2G}tX(K%th_^W?dSTus2@{6{#kSw!t*W;6W
zDCz7C38!HnU6wmfr5NOz%#uq}o#ceIW*;3rtYNr&yQ~hAiHqTMCI;gaGEVN_*6Ktt
z%77`0XQYgw3NHh{`W`oNoJjUiBIKSn1&H$A5cq)X>H5Sq**C5DUH36F{Bqqla9p73
z2>g4eWg)`urd(8)awo+@-j4IPdPaYNfD$*-)fAnTa3xfo#f#%=Njt?@Ry*I1Phz`Z
zjgLW;@^Xv{{OyL_w`%o_GZ}Rkx6{L!EG7RjkoRzoQz;iFyG(THu=+zj$h?#`%H;#Y
zSutA{2`PwD50oWKw=I#U`+p!0q?gYA^75gtbRRJ|xW364QooEQmFA2wDL850BoDzZ
z?cFqCFz^Ret4UX%Hg47AmH3l>DP-ZO#0VwQl`N!6w{v)>A{rcqIBok3fKMa58HQLo
zizQzKb&ze1Ni}exa87ptz8#Lof%FM>U_ML7QMt?P4qqRj#FyVhRYrqry9B2@d{W-;
zV?=}B+5AC&xH*G1_|h{T08CMX!{8p5HJctzOl+VqIXF{wXy!`6ydVLCe;rL{r$QzU
z(tJM;7a#A<^f62Y-F*AmlAE7o{j7R8R|@b)Vj;<p&8A08UJgfa6^9HKRspB}ZJk#O
zsR&<kbLs%$z+s~3-aN<f5+zhR6=Z)5*e#f)YrJaP>EE_HP`f_+<flo7ZT6Od)W#CG
zmQ>S&Yw-1YEE_K+2Jdqoxa(J~8u>hIXvQW91h>4D(3SSrY7DyvapNi1h!TCYkq+br
zsPWWc#!1ej8^P3v@uWXvOmKWUY%HUUm+mGxyJxUbk{^Q>FX@|gpuo6XkZq&a8fnF<
zZs~DE*wLW-g-tB4c{OpZif_~EBoyoU<Af9N-**v)(qd|lukJ(83*_#GNXK{D8N4F)
z??Q$X8Em@(-(DTO0+FG@JU$LPy)dS%x#0BAydwPzJ3#07eMhdgOCE{i($Aq!l)y#y
zIb1<a5Gg3)$T#?HnP_sIGpqysDLJ)`#qSwprz@bez4$BOF}dTjY1_sOk-m2Q^Xh2>
zwRgq5FgaN>8u7apl3jj|SpJ>D#az}|RbLC!m6j{NOi6-<)!!Y4LQ&kDcT<)7d>tQo
z8r>&5qkKU0sKSBby_Q~5Gt3n(TVuFG8)1?5UXw!xm-%nK$lc>RKbfqmjd+R&zL1_#
zexx^h0NTNMtRez)oVxW~;{xm5k%*qS>YR~BA{hBF3v^1IG4w2|Af@hE)%LuecXu96
zHnOO?rF6Lb^w$#x5^wI;_=w|-d-Bzw{eDmGYae*;bp<%iLwjqT-W*V69QNOKhg;F%
z?~sz`foCFOSPJdWv?=pYm4h=g7;v^}jgTnRgHUS|kzbcjt9{O8Iiit?tJ`Bf{4g1F
z=a+@TAjSt*OthR>bmQkntLqv~Vy3?#3}RSd@^=3c^oxKG>ZWy&{xFPxk^gJ=VQlas
z_zE?aAOqr>vEiHcBfLYG*$vU#OtI?JwEZkCwXY>IkJ8uW4{alf%jjjxdyo9)3Lr*g
zNx;zMre-P`Q8!94IG!ysK@jJQNt{=iOmnfg<{ei04ctb5D5G?6YPRQ}%d8B4S@<yv
zEe@zhI}lnphY0(N_B+_J&ny~Khs8&P1coj(>sggw1oN);gz{Q)1kfA&D2;Aj+qLBG
z@<QSqIAuhbz~EG$C4c({IuUP#Icm6Qbvr{uku+N=-aZQ!zKJxOdoO1&9cIk$v&1SQ
zrot7NwmBPMNMEC$IMj4Hd=!i?7B9*;@V|YiO40gn##`WTW$e8kt;q41xyxP$_F=On
zJ&L0!(Ty%mZ5qQMNd`B5Ii#h}%QepwV+-EM=*%!(po;W0^<h#q|2L;2)}h?tqvUg>
zZ**tAo+Mw^ms7BAk94pe(-XK;1I(-ORZ5HF*^uWhv$*XBNZ)V|$b>D8NL5R994}5u
zBuj!ia=kNczsenAgc*KGgwBp{=07fsKd1`&rx|^ueB8)qjBcNfI7C8yf_(iSg9ku7
z`+7YB{>5u@DAhKY#x>RJ&$#GC5YiWNhNUI+NaT9=vSH9K+7e@7M)1Zz(*jB(26t8o
z25q`D>V^bPWtp81;PUYl#E7!iPwanJB$n69J3g;;okcrr_3BrO(6Cxi$^G~)=j06M
z5V)w5yHl!G{CF-X8Xs+QoT|k`9P3*;>clhO<o4eC-6cVTCuois>(hu&$foD!*DV*c
zcl+HaG8%<3shm?Hq-POiM0l|hdElji_1Rv`biC=moip+N(|6!|>B{#8JOo#OO~Jzi
z9)A3(yvHhSSuiFz(f@6MW1}T_O2Z}(XCV->Q)!eAYNgMmU?SSFR>ll}lnBmBvm~6)
zc;P5-2gNt+X3SOW!dtdYH02t5-<H&!e70U<DCxxP61}S;Y^Cq&C=g_(Xo2Sb<62pS
zjceJhWYW^{pgQqau6~sJhrCILDL6_ghFbc5s%XCtZ-ak>*7E!t`>z6n{n!sml3ZJ@
zUM(mwv{{3`NBL3t?R%mc(%Rzsvg=U){81XC3Q|<E?7z*p%C#<nMP>uIAA55nyP+!v
z89H(!3Iy&H;4FzXmtL4U63LhL$d-4gF?)TN<4i2|OZ)cpDF@1tViOS)v@V}@Q(nQ$
zL-hn!l|hBo<8c0uW&}|bH}+p!Zc8O^Uf4}oT3^?UcuH`v;{>!5DGGOh=O}x?7C1Vw
z0c2*NU|>=X4S68ea33OzvHy2EK%uk=xj9;kb$m_mhZTll#XbN1VuI%x?-sZ$dWO-U
z680cmK3e0w+1Nevqw+LqG;?x^QtG)Fxi2xfNW2W$>=dGS(}ncyGm|&04I7R>HexKz
z-Pz<9=Ze|aIKylQfVZbw)8YeY;hUj7jaF92`sIAqKuB?(8>1DemlN&s94B9tlS<w>
z8au_mT|LLyII?k9|3c%Pza<+{Tk6*9Q3j%~#cQ`1MYnTbYsmPFIxL5Z@6O&y-Tn?*
zFQ|an5p?)}53SDiu&PutAgre=uP)$Qt`XQ1SAt`T;vy{t`bJ!XHaCVDfQktg-zC(;
z@eDm1qN}gPgnfP4#eSRlUS7%Xeg<k)+*dMRS0c&f?+;?(?g}q0q;T6E*$3`UP4D!z
zgL*{Pymp=L#6D%;^g^Un3tBADUC3#TzIFVm94$5a4nZMnhjx1t(~L@jWOwJo-C+qx
z_7!5`<iWQjZF*MUlWDc^-I!<5xAFuxJNkta^!KhYcO!3X*`qvZHlU)v{aP`;2&A)x
zYpSFQ*G<iBhxTdm%lMz_dQgLVByoqIl~%(wHy-bmn~4(fQw&I!Fi1ymR{e^ePZ3LI
zq=0DctufY=r#r0tgljRYW_Iqi<~6@zFG<&5THbh6Q2$5o2a$&VM{vxtv^F#+UR{2F
zk5P6TGo1-O%EN)K+-C`i=`qAXaoD2yRlF+r36fJz01!Q(9X@FyT*D&oUrfiXMcx8F
zHmQiT?)O!<r*FHb1^q*uzqo%P99+5g%ene`V-Q(@GP^UVUI#%C!H0z}ir3>RHLaqU
z^77sW3!WI-IB`d-d+vz@kHnWLj3oTQROWj*a+oo^r*);J?~6{ZlDf|c+2S=W8dE|P
zqp4k;_N`Q_=+TXlgHpDt_6?75^GPeddmjzbpm1EvUluPj(JUpyQ>ab4fEo_`cQ;T(
z?csn(h7VCM^aTnt%F_SvwueW92Siq00Hv9k^?Qb7dH9x(VN``5N~1EnG898ni2sjl
z!!I`Hf4>iWpjOj(H&N?owdY<P@xP3L_B_&;d9q5$Ul~-Hmmr-)T9iy6HbOeUdEbCb
zL~Rcp8%I1>%1=-?YT^6RGDzU!sl>;$JO%0(4GHougeE_gSV?%rq1gL$&rSZ(B*4S;
zw0s{3r4$UvC4C;Z6|}F%#JyvQ#n(MEo*;YcY9-GBBdz$#x4h-zM1n+X9eza?g(1yB
zxistW>Ya-wwFi8D@}SG?D_6|*_QZ4ZuinTzoKxDPT2H`~<L?0M6Zj673(A+EKh4i`
zK=>!GnmKj2Rx^uGwPlp@&W-w%gbR-cF{2(Uvg9{bKSy&&7l!rcd;1;ZlILhtbeQlD
z*N-yT2l(nMGk-)R%?dV$l15K}JM-D)_Hz~p&7j|qf%4P@m;Wk|H1w+C;B|sg7B?;#
z`BtD#0{u8QjHUhu(k?}W_N$q~Mi<W(i%p|5QPHEL!q*sTC;NLE|MzxB;Yj)xm{L22
z%7VLYf*9tVEb_<3!x{i8V+}XJ&pjCo;?VmHYT~^nvCriigUkKYV#GxQx++d7IK<k4
zNZirE1(9{=x23pto>|+|7G4ihDo%4d^}lJr*y{um|8?G0{%;jTHmBp!f(Ldq%_(8L
z@iT%N!|{>QyFpKY4odfh;_|wQs-ihKcknU2cj+Q<5P{HB$6F6%Xv9(N_KQXHcU>)O
z4JK=r8wO$Q;kiUJ^^QZtQNrX5vd@|X9vdYOqbyc!sXwF^62bY;z4`#UbUhCS<3!$$
zM+0e;mt3;eTu+YXDoCD?Dg+!LcVa=dE(9dXpp|`p645@_4exmRkkKi0V`UFNVg3Wk
zP5LjIM<i<VP_$G$mkxl?o&`GRH^W$9yV|v(A?Nm}mL;yRM#1$HC#{@(8@%vi`MPNC
zt0_~C!ADx-sl4=$LTh~0BcB7e@KTy(Sa~9SgO-HPYQskF%-P1brG`zAkKBzxrAWOH
z^6amDub9a(ZQ(AM3!{V;d0%6IdDcnhk2owp5c%JGX(L!85r^kn-VEsu5cenD19ZfR
zO5gb8SUkn6l61*OvvxpCR4y_h-H6D%C3O>V0x$NAQ&7Cwfve6lSOis=aF*WAOo_U8
z*YWLI!c@un;`L7h8)3}3E#!Z$ZZ1jpn_NHg{bY)r0ke}%XG|GhZ{qw)PzVroY<+ir
zOScX(QrJ)H_DpX1&DhJQE50KFewu#vDHfvtI}I$W#IZ<HYC-%vj!FH$UIJ@i^|Kp7
zlIFJes$z?HsfWDU^KV~Mq&Ow}A7pnwnq7{jWe_6Qm$In_YZRyv3arQI_pl~b+O2GM
zN>K<iX8lO18nOrkBKg3aE$YSCnjKZ*T(>bZ$hqHnSbC_|5IZ#sB6(8GG|Jo(a=E3U
z4;ZZo@fGHD2}C)(9lHm(<->x`-j_W#!-1+RVtL^3TUf{N4NBE95OmeDQZFClHfs%z
z3wxYcGXpIz-5n$!{xpjG?$(&81guAcp%_hr7vq^Oi_m~VaR-P4^?x3l!kGvMj*VfV
z2G6h$OMT;MefNZ#kN83zEFN+ruVY`r{GL5C+C|7d_o+^h;T{3@hkf}n+R=pHn<R9i
z!lpHuk6Ai|U{P3{Ker~g7v9*jIp$kQ<9jT<Xs_hAnnj382pa=2!8Paz_CGW$!YtGW
z#nNk1cjV|~Wr)DO@E$OzQ(rr1_3|<Kp`Qy(h+aH(6iq9Tk@&T)6S1~LKE`EK0@Oah
z)fZfCxI`(j_G!Rc{o2@B6Ki=X7kYiMv};-pWf$RO2zfv!`@5K^xBn#=7F2J`+i!kK
zNupmZz6kVbbp4{9qbBcJ=q?(Cxy|`NBH(kg`(_VZp+e*Ah2ei?_GmpwVW%MuvFF{-
zBk;oFR9);v3f&S}{p(}w1s^G&(uT_A8@8uYATfJ$RU}S!py%#-@<VM`4G2CY;q2v%
zH984Dvq1_vLUn0!>q}!g70+FhI-A_AamjZ!fjJd7NWLFiM~?Mq9H!qFT#tEYDkDrk
z9n05}kjEaiPPHVR590zYz5{5K-kZ+fBf~~?mEsxTJX*Cy|4A9`0{Rc;E>zDgkOOP}
zS=WrcUp4rD?t)NVFFI!zRhpZWW7}XO$x4d_frDa}m=rE-!r(hs;OvT8tV#oTbSF2$
z3O0VOkFPP-GB&VF^cRrf^>IGyqvqfh$bz`vd<N2vGE~0udMO!wZR33Ngw`R0eo|vW
z>McZ#Fv0%Q{OptcisuDO+SSoJcHb)o_jon0B<8^>I!+?iO({)QzYIvvm%(Yj@WV}@
zUZBqizRewJ9Y)@{p=8GO3qYX+8Y$Y!1&=VWGt1@L2LLJnH+kk7*I$Al_!t=X4;1Gs
zdW`TlaWsdkl6xEAE#05lJ8rQEibwUn<dMgvtAjziFZOUYhLk-{uzC`=+?N}-L91=9
z*78;;IO{UZU*s&*eooHw0a=l4ziZ&zU;%U8>Lj$VrkP8vw-L$1PT+as#=Jr@YTbwU
zOqEf7m<1ajUQ{T(Ry$HoyZQpQdeCDeLJSfHi*UCUqJIn;8l0T~jG_RM%^lb{z=?Cl
zlLpmxqSIxHZJ)1IUO5KIenNxq>`ZZDTF4Y)#7chwt^y4Sy+As?Gk@(+_y3;Ws;+aA
zZwI=D>%SrTIT49(n25Kof?5arY+(G4&`kS95?RmSPB#~113N8Q2GU_`HW%pW_-4={
z{Wk6LJ}`_-AgV?m44Z}tSlmiJ>Lzln<}0RXsQd1v`k8Rqhf90+LkvPb3-ySo;@dL`
z&Yd4wR(+X=U|mAmhcbcrtMUwXVWco8YtcWMv{~nM$IElfvPNmmrTxd7I5UKxjI@=4
zEr&CYd)4}?X_&64*cL`pf5q)1m_X|YS2J@ndgnYw<@)#L=RKJa;4sAdqrTGk$W*+I
zS&k$;S2m@p!3U&r9j&*7PojqU&QTf0q)zc+DJk%8uD~v$)j9H9CMoDQ9~DpOW8)E#
zh8@Tk45_F>=qi$6)^<*$StR0)V~XX8@WED%j~%M(ACEka4wWQV*YcJz(nq%Zi)g-Z
zd-vSZcn8n0EoI=rtQK2R+7XzKC#s{$pX51E-geqD?7SxoISMjJXozbd!d==2A@vm2
zJD!f?`E=P&J=fMhb>;a!o)w{BQ({)7XYx0~Ojph653>E{J8JUi0S%uoRFR^xp~V>&
zK(s?v+JQ2X5LP?1tJKXce+%TH)B*UyjKJ1j-0|;TD7z0o*BNYO`*(X|71lYzQeHbC
z&6ku_F~^)q3>XDJH)37p!l<bw5)}&+zjf@ga`S?v8&e?*Kd2Cd1z0fK`X{A{!gy_d
z)9$6@kUTGImOsFxzfJBAKz(zaE+EndcVD&_gmorD8LS~6k`>QUF=F4dXL+JnwFuX0
zZ}rhj+_M_4D@A?g^u#Y@w;uN0%|PGl0WVUe!+g8fK<f!zk|`wdz6$<%5N#Q3#X=>@
zF2E$=h_vbuP$qgE0P?v{M^9)m;%dS;8)C?Lcx_2!^c42fb62ne=1ui<t43I)R4xtc
zSt*Uz&3JpgwHo6g`!$Ws53HBC$~Y+MS9*4Z{zPQYA3XPBV~TT*F#B}R`|Ny8(Y5Xl
zzM8cK%4ue$C@8tOQmue)PtREe(6Niff7(P!@k`$_1IFm=Tizpybj)M9qtvyazt@*f
z+KP`<Hx^<K$HRQde3JpoEaoE2zlVkT44#6V#F3EnB~^lAc#{8VWS^a6hRF$)V$g-6
zgVK)eRBczgE-T3SD}6<p1x!3lkS%Wf(k$!AUCS{Uqc<!`ggF{s4f(ED?u59NfaH>P
zAuTx1K9AG5^SMF?L-)&p_xxLwgX57Sx6<6?{RD&v=VXcAY^5W!z`i^Ml#-`OF)yd)
zwosFJgejar!1Bb(1{Fl|sIl2&aXRh-ja%ihSQpJD=~OyAsS6PgI4~(tkj*I}>|Bc=
zIsNJL$S{BOCSDdUE4=#N2HWmi(2I^QLrZ=gq)eN9BT$RkU;eBN9B-F*>UNFmBm2n3
zx_GyqKtFJFGmh+2?sK2hh2mW^Ip#OWd<2k&+PLGh4DB({OEbhn&xCDqtcK(;u0ygd
zMdGQ*_3pDYwkEC!dUR-UxDwZg@ktX(q&y(_;&@-&Z?mjmB>#RDA7iAJ=j(7hLPn|h
zaMaMBu!k6zVy(nFU2=%9EOx?0ijH+~vv&Gvrt2$xSKLy$Q<@1jf})rgoU%kW9!=J6
zt!D`qUi^pkh~!c(;a_+awAcs&_giBy^pbwR!1SxYFdVL+l*(<3hR7Dqj$UG_{0I*9
z7#hnDSYi}2%0Fy2*eqG}?T<NNJ!+!UN_BpUz>rLHTu7rCJ^M-`8h{&|xP}|tqI&E>
zHRlft(oge8e$sfbDeMc#nS;-5CxE~)o|t-T_bx;(ZlM0o<T!-m0Cp952Dj7$TOhLd
zU7ppSs!byMr~XK|jA!07uI)Gb-9MDzYk8^2_b6f;%&d?I^_3P`o*mryos?5KjF!O^
zhmD{708h#@J|;<$KqpHF)kf`_PtDe-gn}Oh^zx19oAS<`Ht{0vB~00YLhs(ZP4UK>
z$DN0tnr9Ru1ZuB0vZAA>Go|h`L!eZlJGpL4MGRU`82fH!@`+1%Snj0dTIE^g?;{V2
ze%7f*FP&2VExI0rhEFq|3~XbDAnx$)=KHgyU}w;vyziyt8RU&MaPs>&u2KEv9aaG8
z5iAYn*GKwp@Fc7@0RmD9<?yuXC{d1+(7OL`601S3(9WuWYXKh<P>Q6^WGow{G~CwX
zz9%=C-hOM))ybV0m+hY5R{exAlUw!(%RHwTzob{%0L`o$gX<<eMmQav6Vbx?%F+zC
z%06U2&I`?8Uq7%MHsb-tnNExzz)uLh*YvGX%~4jerRt-^24A|KBXk%g$U3nuU#igE
zOH(^E`rMyI8q)nNZ$#t~WscG_dAIXPY3>IhYMyKx4!D|Xo*%H4quyb{ObO3)@8^1>
z+Tz~0dr`89Yb7RQ7KvIlqTnKuMeGXlg48@C*uHTM?gw~AmJJ#bGE8D;_tRE5UQM0o
zIUn3zDYf~)Ww63o1rHQ1_Z*oPGS9EfEE^6><G+nwg@NxsAy_T+v##u1^a-g28!*+D
zHGbf~>X*d#E8(OI?A&>od^D^-rQ$F6LdSmum+Tc!8J9k$1P6V!UPcYOV-_Iem1fR*
zrLtL7+lX<W1t)mU2lpq|S+fj=TJ?T|aIIb|<niIk(M@4}7r#(oGNqNAMZVZ?G~~vz
z#*+(q<8i>oYMLu8Eyf{>-_WToQX*FvJ{rjap0A3&whPFPf4rq;1+9_oUmI{ibXU)-
zg}7pJ{>>9-+fx5tpEuSYTFh~Q5~Vq5-u(giXXu#(9q}{))RbFGtic>+yX7_y6LG79
zB>Kd2uiva?7z<G0eCVvno&9T|#VcURoBfS*WZ~ERQ6sSQ$;K%!wi#Oq^o}I@%G>nr
zR24(|deQ+F^bPIngUL@V^im9esLhZvW4z5@TWvS_O0m_(D+&Fn)>KIciR}sfDau1d
zAZ8$C&c&*E8oL;yNM#kc1c_^soJ{qjnz4PoyQ#1QO&zFl6-{C<xo!o5cb75XQ0dRn
zfa@pT>}3zf_qnD`@!wfa1gl`eiLx4}t@_~FYp^Q!&&GF%HioJw%X2Bby0#5dKEntl
zO-HQ1Nc<?37O4kiYvf=>=yEGj<c*>F=kNYe_XIsoWh-^6q*4;UEJds2|KN(1L?40y
zpnoq%ypK<!-F*#byQ$IaWBoB7yJ&=Rx}j{;O*y+23e*f4X8s6mQi0Eg>UAKkbRv^B
z1~y6A$nNrqlmf(RD`Flq_^5oBp8!;=u6wzed!pQrxqd}pPA!=RXul0+5f@l2CKkRo
zar}E`XiSaqacpyXPT8zr_7g<I!(wK>?(nkfGg(Ov)l=vQ>CWq`(cRRTQ4&ot9tEgn
zJ2#hIumd8>vXvyN9pq8FnieD~DDA~{jmFqriThUl`(a00n6&f{U6viMm>?m-dJQ5F
z`H(yqlV^cE?n4d)1Ld2~ds=rX?$z$$*o4EAhi{B0pSd^+W}xqY$HglSPRGohN?oEq
z{nKz_d<X#P!`)L)bv~<=5}(_O0)=>G8%U|HsPz+@hn9<wh&o{7eFz5aic@buI?)%o
zbWd~=FyIqHRB3z;YALu98{w)A?GND_`t87BV{K3Sg^Z~Mi(}}XE#I*y?vZR{6)oa2
zgNs!&S;NfOGF&DtNOws{*t%8Mr6>CL$0~g`qz2f4Frm)$T-Y5x{R+vJsKaV`f$9<|
z_C_W|vlLsWPm-knbC7&Uc(Rp+G)d`Q&Bw5mS``3tb~WHfPeM6Gg-{>-2j9nhHH$F7
zvexlF7E$R1KP&9}+X1}$pzx>+OCh1+`v?tPMDZKfT1aSFH4l(06w~`;1x+udO+%b1
znQ|pFkM6>+(4`TKxA~`{YR^x;P&S{<9A-r2w2=nfGaI7aW!9~b<zj$KtW`Y6*f~;B
zZ4~#)q-g>`W>I&-+6oz2isxgf0nr(!IhXgna@k!0xkGBMq#NS1UHa$A*RR_XSQ5@~
zc2T0v$`Yd?zp1Ldf-=g!pWqS`Z@_yv$0%md9f!KIxQxT45##wA?ABNkukN2(bUWGf
z%a`mW=C)D%Og++u<s_v_HH@NgFLuJjQS3so{wg(o&L6sDyb&VdK3C4kungEoVUPBw
z2<<0U_~+_P1NVfazVU%S?^Q#wiI$2$D`gWbBji%4l%s>(G+~JaQ33=+$o{<M3%kOg
zjkv)(E)H(V>yg=}W)X)HR2AUfCsrF;=0R8_OfLtdMzSVb4T~7H-l%<l|E=hclHmtS
zalomX%RI;W@I0V5kkp}Q$CqIC+NpndMd(`3`3G1w<8haUQ>1h%>JTT#oL*$%m3hiX
zDvl%2pXmPj&ZewbU(T1*$AeR>ONsl5aet}S%X84H3^Jd1`g`)$`vQf>UJKUuI2?Xz
z@B4w~w^RLl{!I+&MCC+TXkuFsH3wB0nhUDQKGv8RCFK2(IfYNhYH$_*dk6S<@C|F5
zp%(!r|HQMO_{jYI2+^K;h=IB-#ugRld8}N9=<*>C|JD>^NBK8{URU!L<g7l7DWXK3
z-I-7=ETzJCj*}kPdMQenrES?=Y0Y_qY4|?gXu*6jJ@Wn|>&*MgMdBOO@LY0=VjB~=
zN;Zp<otk2Gxj-Y7=Cu4_Edd$hR+~<r>LwVZeGGqYBgKxxDDTgID-c+!LVvN&-z4vd
zq^cj1a*1p0CSHkT%B10Im1}hH>W9y9=IP~;<Gf)&<^pXTU#KD?;pkEuTibCYy}0L=
zjcE$2w{kRk5QV)4&Z4zrP4~hvY!Ow2`tQA2t|ndup+N?0Y-$g?pc>j4X=J{N9O2J#
z#^mbm8%qM0WcFi%I?Dg=NnQAM;<fuYH3{V-QpwI9sRZD<v1$550q@lmxMw<oJF&&w
zl|J>t+u+9}wU6j4f#-|ZunmJ5Zeg?>jR~@ti5KMHm4{wf<nfC@Qh|x^^y5rmof`d?
zs<Xe78*(DqGPW&hNCIc}_FKG;*wiN+hz@XH`VVW^m`pO<ALJIanfg=JB^uTb3_WLH
z0}R1I8tl!*G+m=T#eCuW@wLPf`Be-QLv*O;TulSaxz{Am5wnjlYW_IFj{$k0#TToH
zkNMX<#rY^Hns>W95-Fqpt@o;$NZWF3Z>8f_{;G=m;)hSF-4U&yMY!jG11p#|Dy56z
zp4(n?eB#i$(yOuP@4V-Ia|>`PTsF~laNc@{e`>$U-Z+C!Z(pY9tG8{u#^OQ@A=7bP
z&@;;u-quv^Oiai<yJtU5so`k1No@fGsIjnI!>&}?vTh!e^aEtP%XysZYJ6wZ(47-n
zoCvT8&mWtThjO=PgHC24!a0WL3QXcUUS&}W4XRK@F*r=T9y0;gY9^IlB`Kb~R4Bt>
zxr!Q+vJEgkG^}k|Kj!2KDS2x3L)Xe~?Y3*wN$C+_$lr3s9JFTxwHzJCyEw3GhyE*w
zk2V)US$fqe?nJ$IS<Mz^Ei3*k0@ValEIF;Gwch8})|}f|7O{mi=)92lk(hmm4Q#eE
z?NuxkW>K>7`Wp039<+!M#dCXe;>Gr&+PJN{WpxD|T`29VeIDAVRcK5S2zyKHsmxIN
zyJ7}SF(7aPgN<3h>2T$L{{gZljsnfXid$*9%?$&?Nun?9xeYfge2!85LyzI^@B^-&
zdCo|UOMJR06G&1S#q4R>3%QYUZeeI*HUbwSsJdeN%A`6%O-xvq*qu<4ogkff))b)@
zw&4~z4E@QTGKDkY3NT3E{vb?wFz$sZ`M%5U+fT&DPMj<5ptk<sMt~YJz5nxUJ^pYV
z*mwxqvOsm|)V<}4Qa2v@g_qv{j9SBLOQO+1J7Euec35TegGe@6w#YrjNNC2RCKOcD
zB<OhBU%bcpxcYX}OLMCl$WR^K%0oY&=vl8menNeghAa<bD&$p4zn@i4KZHpq@%U=c
z?0Z9nU#1oG01upm{VR}CHZ#T4aZo+Ch4J>p{GG3<It3xIn{~EKdjJNjKDpDwnpWvF
zGfg^}Bx*ncI7g;5={P(1p1>WU&DVD@e?z8|4)|`nF>y}p;F4T$xePNvH@*kiu)?8^
zs$m)Hi43>ue-H1mFkuL9Yc`~+-bAp0?jaTG;A>b~YoCk-%nsQrg`>yfT0`9pV67xK
z0r9J4#ZP@-jsRN2y&`<ZYe5SDt%t)>DS>mWs+5|H?agWmp^||wpIlz8sivgCacXu+
z&L}r(K#fZI51)lQKXt1+?qD{#PD}T46^N03*wN?vm19!sxc-T#6Wn}$RqfI!Bx&Jf
zbe19<v@xU%YARBSG2(SV>RUo?qhBabyO{~H2LS$zc<sH;-Oze#50-c;%6-p%o$O!+
zV$4G4<oqCK@Z2zvlXO^vUQ5ScWN-b(138%ru+D>BJyOkwtB+dRS&?liUSz!Ad2x!g
zSut+kcD_VJ5K2(-T65gaCBBmEBw0Dprljv0pQebJW`zs&T}K@vRrWr$yedjgd>Q$s
z@jJ^n1(}PYnjL7cqO52V4htl2SCrMlPAkNvybKu?JUwP)eUV$9MS(hf`tlVXuo;ms
zo$_u!?D*%CkCEce3!czr%!s3TLUEBX876{)G#KnTF+&VrQ*Ap-VxMLYE(MA;O%mEc
zqqKT+;#8147rga)oJul*E&N4XPeIO?$IGw|PCDB<dK}RtBl#OlR}?ja-8y4%M#Jcz
z*K>-^&<r;Q(rQ~I()L1e$>QcvPgu)Ji1`FE+iQ>ctG<t>hwB#9rsaxr>aV}Tq*;1t
z5oNFq!u<Wlc%Glmn!l$hc~>v&Flw=d2U<?)<2BQtwoNANcZ{56;HLaGR!|2d$p}`S
zM&~vEQk3{;<S0L5+ncG?x52lnnnaDR#C75cL*AhNr2noeMDNP}cQ9bMp2@DXIqwUM
zz#fOaqz<S<j=6T(km!`;W*i7cSU3=suoj)=2EJx{g&>Jp<=Ils)mdiTIUT&8QJoWm
zb4b--NR$`$H|#PuP^PqKFSkK+-FdV%ha$I0q|Uf%8o0N(4h)|)g{0r?p?LXcn?B|r
zBz8c+xDt6~L|KYuh$zwlT@`UV?2rE){dKw2(aqR*HY_T_`<UGB036*T&<wjn$8Pi@
z4kzLfQUjKVm5n47ZBi@eNu+a48SLuSO_WNeuR+JsH*0ktV8`<=X_CtPq&c@OdCx-c
z22leAH&RLyVU7GgK6i1tVGllaK}Ei(I5nn`nin8nJfQ48$NlS0knf*dV!=b%<n7d|
zMq~9q{gP}LOS}o{^$*y>%Bpt|f%m{pr(N|3xLOgVvfGxz|D=hm;0sX>%Hd)hXJJy;
zsVY=obw)Gv^eqM++9=%CPGCNjkYexSY;e8lUb}ZPFjkSsmv(VS37aanU7o_DGgJD(
z6g~QZygcMIY&e|pr$A^s^vl53>hUY@ukQ*dS1?7D_UJb0C+E)iHa;Po3=FL%d-%(5
zqZC3sX_qvq!V1cCfW{k@R<foCQko;paVRTfiY_<wtiZ+p11=@45J%s>;_#gmYt*{$
zxdwIE;%(8-iQcS9(5F$`jv&hzRCT^vcWgnlZcF>JFv>h|6<HhHOP0pwM(9|Mj3wg2
zMAr~ZkyY{8{lhl<a*AC2<HAaaGqU=NMHj5A7m%f@#uKBA+$)$l{~OQSXej_~db<|j
z^8X432n7R*0y-8B(v#`KYKgrsE{vn!{PA4=UQcd%Tl6vjDvTO|{}Qok=I=q21`#{5
z6ZK0y&P<4S?EXQcSIIqXfU$}$(f9()e?HPHg#=;Gq}_og23TAAO?5j`t)9R)5s+T(
z>VY<j$xB$wHr=o|JOf7N7fnn%6Tp?x33&KrTqn~^`H0$#LGODa5?AtrbbWl}%;(F8
zo0)po3I9e`a|u^XUOxlI{S4CCuu>|GI2eP`m<6YR6V{+KvjfHGob7yoxt$A|xBTA~
zQ{p6}rAFu9W*)r$?f3cjx8QFKzvbL0JdXb>P+sbJwqSSo<Xv!4%i`0RlK~1hfTRh$
zp>F?K&JDQ9a2dL5Gy3!PXLbKk4O!9bCM&p!ND_eP3wvd$?0fb8H%kBgFfP@o|D}df
z%jJq?&K9@GXdm<spT$+9!0g5A_T@z9eCPV2h$=-VMhmp=5)O-{F`0wg&9xBpfuYUB
z@Hz5=w)k+{MfHH;7jpGa^3%;hs~=za;!$Kd|8TzeF>*mwNayTA5Q^vawa2oKHH->@
z!pI+V_8}7QB|K`u(U8xwsXow1k_$Q8+R^r=z{1tJYc*PBDN0uU&q3VAC?_SKZ1&B)
zv0{UnDkk7|KYP0Re>!cQ>}%#8Y2nJZ$uH6c{`PjJvg=jSUEmw#D0W1VbQJuivy+xD
zx27QD7ge^U_fW)mG=1ZH9oH}4P3<dau<$%=f#!&Os~t`{@iY@{v7@3kz`pFDBxOW5
zR^@68zx!;+wvqvpOCsv@Eq{H<WNiatL3OaqcCG_`529<o4;LxkQ{q1LeL$6Vf4HR<
z>+>X?uun4bWfU-Mrt_r!=|MIihMw>iwjvnRyN>X}?6fRzr^x<RD|>@x82FT+5$0Y3
zCVw=zU?*A1{PW+^hB{0CJ+Y)*y#ROe83HCs+8MRNG7C(>*ea~6;1+jnbYZ6Q0Ht_n
z+=?`*W{`2toc16n0Zkq##+aDl2dWvY!`u`sfn9i&5n9o?mPFEmD2z?NemzSVK1PWv
z$;2~K5jh+FtiA8ko+M)}fkgpu)u;PwRnt%shUgMqE9Oks&ofSeLJ(QcMiKdhmYi%u
zRa{U4!{jI?^S%qydvyYYqC*<JoG<uQqn?1X`U8y<*n1q6309gYOoh&O7qoZ-zjnzF
z41mdV9PYrWqKMYqCI*w7<bNRv7))58sJG#CF$_CELF2C_$SWh6=)<bw=ensN`tDTM
zOgG1*ZCf2lL@uFdKJUGuE~w;jH(Mb@1rzB)Nj3K(?DEauD|W#w8<?XTd?5Cr;N-|z
z?Dyq1&qW}3evLJP=1k^ECv?Tl#wy@{=MmDO*X+nz;qoH;4NqWYBKU6hh_E9=*!j{x
zCt6zc_wz!$g2H)%olgcIEncda3@-F5jzwMpkn_WmA{QY7&i(~Y@Fn2a@{UzsVCKSQ
zFNq{Kq`_RDZoSs^d|hVrEf;ZE3t3whg@^=2pM~lc-BggbTAkL*=jJ^~Uc#3+;=Jc!
zxd)@+;W)o1S}Y5Tet}dIVBrQ&GLN@GtZ=A#G9q65y4;IU=wY^cH+(bz-4t(K$RhUZ
zKI<X?si}sBKj?@%fgN8N2WA*Xz66{Xu2eKRh>NjXv_9%d!Ng~lMz#_aJ~C(3PL8gU
z;m54Xgt<emeBHmbAKuoXHJ-_!M}T#wc;prtr#cjOxJC@&IQ$4h<kJmQBnibUX6j84
z$rb<a^wpM7>F`UlzTufu%2TmAI$q+{iIw}IF2h&x7zxI?7h^@<PqZ-9TOQX`wicO;
zI!aD=={#n8yng@_@EFADUX*nnZMs4a;R`-Q;1>V-2$MJ(hvSwZn}Qq>q&5osZByfZ
zX`g;{7Q)kKhhY=P<ZJ>SsrTAH#7nrmpRxUvPvb$yYuPIc>+d?7i^*i1&x9yy8l(o@
zj~2D?-xK{rR{tHQilW3=21<*r2s^=;T?`c0rnkw67mK~eLplkH#RxYvD4`yPdvCdb
zHDKJP8a?2W<QS5JNz5YiV{D!8snuR=+H?5A@Su_6STFewGK@@c2hZ~frHK=C@LR7W
z4HJ~Dy#HOVLM8=aH;)4+Uu$#{k8`eb5l9(W&OZ9er<)V@;WNg7OpIIz{r9KA2bvhy
z{>R_{qcT;i^^EcYRkp}M+m6^7Fsr-yB^P2pNBO1Y-@m3xVk_h9K~~;vKxn*8DRBPl
zKV{`S-s|W1oB}6?##22e4a*RwVgjKTn4p&$35$EpM+i22^=HU~nmmzw66bm&=T1&@
z!;zEp?K`+J>X<L$@|4u$M|}%;C*(V7>2^eF&Y${LX{sTA?0*RiCiE>QKF1D;@2HWQ
zripCpszK`;8cWz^GqX#2kmH5lAwyU1kJ07{k7J*#1_^whc5#WO2(#8U5qsF4rF*gR
z4uVg7JwdnIDU<A$NT^+_Vdy1uD9F!bsfm`0keicJ$HP#IRyqnZ)|xM|I*G({CKX)S
zL&XD{P}_vA1-$jxBF=^7e-D42^oUQXHHL8@PPD)vq_ov{gS$(3@_UplSkPyklRCvw
zPe$!LsQd7R_Aw55NcPXa<iQT+1Aj>W(W9t}hL}X9Yd36$e#Kd+Em^y(&l;rPhf&`U
z6hzd$7=DIbwM#(*7<9&p=H(xu@yak91%#69%OsC=m)>mERWD-OCjmr%%RvjIe)0S+
z&aexQIx}Uk+=eez{2AD4VCWr=p>F+A6K;{P!G03g!f(R*;rxC(uJ9GkQne$;hPp2C
ztaeNqjOy#mi|wUn(9@d6-yBaf+i8EvuIOFwVJr}`?_=#bvta|&y$`Zq6E0v#$%5WE
zUE&yoPc$GJj)&xX69<1JFTVsReu2NeDqA8-NvS%-Q*FRM?wZGObTiw5+F*t$7G2r0
zXYcFs^x1_3j*Lt%N5ei_8GGNyC~ry^bg-Q$%bZAIM&7gv>250LyCcM1)Q9}OAI}Z6
zFspu(S>3y^Uk<l*my`+#z>&+v$aXf$@BEGYFy2I~G^zjIc#$vV+H<u`wxC9&t2RrO
z%LDs{Fs{1^(+m;ep8qPUc(ZIOXdvD9u_BUoTi;;AT?}+@)Al!~v6Ar`P;7C-(%jdn
zpsw|D4)aBW3_X&}gs0Vt4?ju0$BzO0NdY9m(ZC>Ug)<@Pj{C65*x=BQ?XkUhikjJ4
zFAK8b=^;TRcQAmQIKG<HJHOs|Y(sYzRIMtx)aas98mVnP92@_5A?SlctmBJC4S2=Y
z)^9a-tcJXDAls3CR{B?(7|SmObMJR4E~7ZZM98Rx-$|-5g%M>y;|jTlQoj>LU`j+@
zv-xp6|3G9r%kgDS?hh`3qwL{-D4kIS2P`^RznS&d^QbUN(FEExz{G_y#3vGMJmY_w
zep*w`Na>84vW<9jW=b`PM9^|{r3q|s<28NS)GILARrQ)oh#!Bv^yTvMbbvz=(P%2E
z`C%mLUj3fxuIecrD}4;Q*h2NeK8Awt&dm~_1sIx*L!)ri>~%}MPSC2CxR=`beSfSn
z!46lKd==ZM4bfs-!@y{r3A=Up)XdXb%6DX3EbK;fa$xZq7RN-|c_y{+E-{tF`IZX0
zIy20J32s&*1Yv5(QOBrPLgYkO{2Sl1zjjJM+Q~r~VMG<=b&|B@AMX)hHXafUM!r~y
z-}*FpTuG+=9IwhLq2C7`+!J)>c<#8QE9un#LvxPUSskh{cqVmDOLkw=X+_IsblYFe
z{GL{O)Y!>{ayX6XM`u&6W~vOb>c30=7Qd@3*QhhtE(a!Cb5b4ta*=<9sf5{tot^_k
z4jCW2{la8q^!p=TCCD`v16*C2>~@9n0TvP&(Id~wYxu`LBh<fxzs1!@bnB7)`Y|^A
z{l@m$P=@!jL>#f;IFa+S|5{2JacfTFB*(1%X|48+f8U-g&@D-Q%oyKWkABMj!JNzk
zzD~OTyX+B#fnmLVLZWIi&#f04W+@HtOxG2SD+Uomo$zaU|2TRcjD&JF;*s5B=;%av
zUgSez4|JooI81ZZi<2oQPu!tc*jd>3SBF4xpJv?tooJGuol*13hYzT7nF0o0MvPA$
z?Bsu`kQXx}^?cCE0gOBlw!wjSD&&04X`VJ?e|fmW*Y0fCT&$m!;3|tNZC#z`bJIXX
zg*Gfv_FFPz@AdWYF#(zJ8iM)%oO$?a5-3VkGUL$_ca501zl&=c#*$dG$_Od#@?`L`
znkbd~3_QkLXXZu@QiG;YVyLde=@{h?WT#k0^piPRwD1eI3}&e#>&sgR2rwJm?qNTW
zu6J~O4%wz$NxBKn^W+pUEwB}QUQzX8bG7{b9CitZItycF+g-K-{Qd&Kh;x8u!%waD
zGWyX-N7a#^W0KG=*yF=K3ZsmtSii&LzJ}n|l^!*S+P|kO)x6NXvlQgx8AaECB=^Yv
zRAk6tVSIRZ`BhN0A(et8fn0WP_kN<ebuGaoJIZ=2PN?2!ij(O)vF()AkwIQ{3XYd&
zsV$b(u?M5>;9^tN@Jm5BW%-m!=QHTea&^OqA9m41g@`mm%Wc01etZwMIV?S;*TfTw
z*q3b6(UTH-eCsJwX%xz%|8#h2N1e-VHw0_v(BJZo9J|_8@E-+4*KY~t`Ss%mzGms1
z&|=%(%CO56j^ejl_H!%kDND~Tru^I_#2vhP%}^LP*n@5nF^1mw3VKq1wTAy)77<3u
z<LBJ!H+xIpfj#Bs+{ai=u=&D!_q;OQv7h3t2a*Q#Cs+o*ds|@dyIUml3x(0f1v|{M
z!wzxpdG!Bhg3^Xj>;5$%#0~6MeHB+pe!+UT-#KEIt*Rfq%JqpH&Q0&oXfBsHoN>EH
zzr_$57ER90{1V_1OjGcTr_>dwri?j$`Yz<sBB;Kgp8qotrJ>?D|F$fx{^zQj)g8S)
z;&Qz8cP~C=RC4aWy!#E};na~V#9l`hWLw?;DA|9KR?2ZYzdkCrFv&=XDWd%_ce*~X
zd~`uUxhus<iR<WSMvjPOZ;TqGyZMn*u9(S#9N9zx+Xa<su9Q2l38A9-u2dmvj^4Y4
zoXPhrEU=6NbmLXPx&>+rk`kSZ=i4W5C^Xpye#Ks<%!nTy`VtdI9rQ)w2LC^rzB``k
z{{5fh$T8v^dpnN3ciDUIy|O7gTPVE|G9sH)viBZE5eZ2u*;G~<LLn`!-*xWa=lgiv
ze|F=X^FFWFbv-BUqJubB#rqB!Ew*ub`)}M8?IRjv6fc2TbPSI9l8b#HWPeHV{o}FA
zp~lU6Vfef@bYeO0&~Jq199g2=Bwgr=XBaIzYW{1AA+7x^(JLMOk+3a05<q#JBAPIr
z<*l%{r`plMIB@%{sJ3kpzgNosUeQl~Kjyh-h3CI9SYDhaXHg8AU{)9&$RP6oS|aE|
zX#S~w<I+SD+MpYzXo91GPH~@Wv)=sYhVDoX(u_3R?48hcZDdD5@=9>kF5NjXF@KB{
z=C8cK<u6x0R72QXv53u>st83G)`R_~jQ26_CyK$=e~As0)z?1A+SE1Ix)A%@m~rKY
z=m=Cc>4kXFM`4b(kG9S%dKscqDT*seffi|+dFZrr%VQIkQ5>!R0R?YkS4`pcryKV)
zD~$=C8|snv@yGR;p%`s`_A{^QS2G%fMa9(>je3|}UieVdwNY2}%;pLHb8gl%N__k|
zyItHZbH0rb^XMj5YgBIK?SfK2E6?2g6}~+v$<oH@?e{uv^{iyFo<ePWZ)e82jRdN6
zGmRK+-cKpZeWQr+s1~|@HKQzSx%xNLm`X`CzQY*lqv*4QKBM1>B#e4FrMV6%I)#?@
zGE7eWEyRF~UsQRW>!IB-=~Yh+KSO2<C)J1DS$QTdrEp@m0<pi%S~15^J5tlS*znB|
zA@u0bf836&>7obIBU|)foy12LVao&;y%VyF>Oou5_?!H8TlA)Brv)<wn48kA%hP)h
z!-ki<!@oHaTV3yMycTFA#_E+OY4Z!IbBaH$ICyzlf>7Y9ke6nuu&Zny{mv}fCC{ur
z`*PjM$@=0TjcGUd)8+L@8qeqDUc1(S(-D}N^0PEk2)7BU&oo5WQYBu$q|Zftp6YvU
z3>B3&V7PO<JUX)#`N~FS1nsLW?U#{~*8$yC(x;2a=y2t)D7Lz}sqi}IlSr)en44~&
z!qJ7C+`{E$29t|b9Y?S9-ptF`ofc@vkdHe1{4vTYlH?uY*E7Tw1f4LyAesJm@FCXi
z5cW^ay5O?fMwY&)t(Sr&S24;8_rAU*wR>NG;SuvQrCr0#+J3GJn~MTBIrQQXW+-Mn
zjbjDG!bUz*mHt+Wm1N3G^dp84<sa<Uz4NUmjuMs!d4(rKj&C1wMrWNJ;BosyY}<i)
zku!hjosHR@twYHxWb~faJL}~*V+)wL%+U3{W_IphEDY$%-zCTuY`uPA^?O$7(=M^&
zrEsg^#Vg5qf`(d?`iqohPlQ+?OJ}-l`+E4HkrVTmF>I@aKh3bpQ@L;~b)ruo&7+S(
z&tf}MK7)Wy_)q!0>>1Ph#M;0?!c@Ig6_G;B)fMH^N?l1hq>w(ej0yP8sTz~h&Q&R@
z%E-4q1-R!hg?*>b<75dP4!`ay4AZIXUA+XX9N`mGdeZ0}A)pQGi2oI>ur0jQ)N{z3
zL4qq_te{hS^C1lCmC7%91VlwLB9Z3WY*OWl3bB=qB9rf)*ZyA^L5Cjk@lQ$+K#%u6
zyQy-grgq|F(tHKf)7Vbhxmow1ljWiS*L{5CVKbHJBfoR5>&#Q@%QXt*5}Wi;I%RmA
z6zO27Nj>$!4pQ3${$-|wd@xeV*2VX{3gDH<(3Ms6&F0Z~=A5>pH?eYVYdU2*MiCEk
z$_$mPJi7-EwB*T2%Sb1>dXJH-pPxQ2!oIegc0{W?&wW}lq<1~F+IDK30pFo#^xDaP
z=NQT~?JUik3D(|(Q1rgrxzqChXtXb|MIx*PIT6kR;Yan>4%LRZJf54}CP)9j4;aBn
zBi3q{*?s=nJk`%KzoDr0Lao<3f7hmpORuxh=$9h~t9%HdQJo3mw)NN*anu}$@dqyc
zi8Fj_QYb30zts^s^GEg7Nyin3R+ok}LIDR6gZRF)RF$SWk?}xMYH5fNiM;ST%MzH|
z`MeDCjsd+EIrNpPxY6P;OfPHp%nGCKlRN?Ho*ZctT<G=Tx8cMNy!xvo7i5TTLU;3c
zfsubSxm3q+@phrk;o927Ixp>3PAQR$Gl~*rhixd7t}v(sC6T|slXHml^Y3OLqfU5_
zU4eiXWBy#KljFPyn>1~BTUf4jeetB!&S%==1wnVF6motWrs{UG9BcdoHDVOxq6*A;
zb8rSeCJ645ldcj-j7U?wt94Vpa?tdB1;4W^b+5qKFHkR!Ap~Z%?eB)pR48`dgon&-
zj#=n=b6c8E`f9Sf{O8Qp3S9BYLQl><?RmMmt{KWo$B8KBTwqR9xz?FV%SFZO0IZKH
zqfu%!E$IbT$kcnl9065w!%wfSdq9%uCTpjYWk%+khPpCmxS&BGb>ouu_cAE?Q*%R{
zy3Ux{<75}gKzZQPc2NB%cB|Mo-03V@J@$a#g>2dAldNR`twH(*kb181U+17$0AnsK
zfjF;%o(kYpzV|Lqe^nOP{&3v0deabX&<Gt)>pc&fbgZf=crHxc)H=;5f%V894jkDY
z>j`iYXXBmrYQ=Awt%OCZ@!RegH=gNpeRHguCofjyrrZS}0Y)=zT+Ng4gwMZ-L4GHK
zen5uJ&#0wKZtfR<zGZT7;my1;eoqSNisKhdR3kg}6R0iq-+MYA*}J-jBqrV$ojPm9
zhJ6=JMe+QaoWW^cwSwc+pTVZ_q(J0N1&uk)(`H((3Yi)c!BPf{$t@Wi^v=);Yw7J_
zM!CFB9K2>>{61=sTsMBMbg2zH*W4g;LmL%3QRy>TvJcAQ=fzUD_^WEJRJcF~YW<aj
zk_^rPmHxH84ELxxJ7Ui2m*p27^1C;N2pK1S%-L3NneFTDK#o_7>%IB>iqn>KxFIp6
z^WTEb7+%>@i#rsyEGIcp#ibGrRE&X>&5L%=hquqZrt$-Y;<^XiQ{1TfovBv_%&dRq
zm{Le3=PRW(H!Ub|<s<k8<QJ$Cq#Rl~bI;ReSGz6r)59YW{&801tFDbnB@o`TX@_Ap
z&EO6lmPq-6<|)7aX2S8m;>8cyB*(t1LMNym1bM`c@~B^MJeRdoc#DS5uZtV9ZqDB;
zT#3|8*yO7@xDq*HX>jea(y-?R=qJ>&N?xh~(X8S#CI>WfZJ!j+zm=2x+k4a6#lvF6
zL}_%9!QY14wpmX^M~Fo*i{$ZQbyUvE5_F`-{6h7}3CUd(7s}vDV8^##FlE2*;gaK=
zwJzV`dFSErM$3w%<mKA*Y8OdNOxn<|&x4=WKgh~|HTCoyQXp6+6NvI$H%W*gJPqd@
ztstt`1bKXlvi6hKd>pX^k45U;Cys9{66N)719zP&*}YF1@j^EE3+?cQWYzd1d0ub$
z45ET{+m6u(<~od1gt-(Q03Ji=;<|);M^B@ls140%Zmf`X5hdxdS#h9O9(kvJXaCKk
z2A|b>E56L#-QZ)dd-$uCs2iYqt~Z@TW*s?8E99+5*aO<#YL6p(WxQ?URs&$JC0#c+
zo>reVg9Oli>|c@oSE&jgq1FOicXeI`KU+E}%6c^m^d5ds+BY>;ul@Wv+?Jd-Ha)H^
zAG)j4z7{gSuQ=<ou?t~K%+q%TH<KwD%1b7%$<+n&#VO=CX7+2Vhram>p?Bg4VJW9d
z$^3*%u+|qc8>SveK!CYtxvA}b#(;RJMFM$BMNj345eTXBs%ZIR#>)Pj@;>SpWL3gy
zG!%<2SB<!cW#i>W*`^_&e1$qbt~RUg&wKoy-bc`0`c&##mQ?Rrl4ZyVmls4&uK=H)
z&RhI3S=ZHfc8y#^e`t`tD9FLIWiZ_g!;l6<7C`<_#j~1Ir*T7}2^x-0`_n8mWfzMc
zL8+Eer)#<p-f4s}NNS!y9+IJ<M_|1kRpu0t;Uk~c+TilZBG}t6#gluu798J#W8bwj
zc7su&5EM+*-c#A3|C~ts``?3GoiPL%rpNzq7HZ~yM9Rb8PIO<M8|}@IIH_<weUDBe
zts`Jg^oK4ik<rWTF$sL>p|jQcVS_SK3iZLRVvQf}Cna*p*grL1bb~UIm7!sJ$5E&(
zG*qm<Iwv%pU#=7_(^2-@@$RoV29hr)KOK)g-of~NCbon5=hj)#7HUsE+id>EG{bQ9
z81idjibzBER5UdOli$nR!R~Qe0{2wD=^%_$56Z?l7IaQ?gQo;|)EMCe8VQ)SDRJc<
z+QTQNhs4mGXP#2M{t64#tpXCpV}!=pZsuHD!&g{&9`w7Czf8?~P;*VOO%YccCzy<b
zW-y}w#$t}Hcf6^k0e)Lo?m==-`OmrO#3#sB;QCGeo*YA;GSj+8d4CnO6J&kwOT~?Q
zTlD*OiWAhoq0x2IHIA9`5R=Co_3!T=c&#39zrRD&`Sh=pD{D?Q>%AJ5biHeOOPxzo
z{{GEpX`g~Uaw{O(4Dulr%@6wI8co{ZefOXwrI~=*Qz5>`)`9+;I8>-BgeqP{ppt9;
zkuQVx#k>9z1k$W4U8h0@nC{cT68xP$Iv*<BSmTGuFoELa@-1}h7fmJv8AYu5uUr*z
zlG&@O=g}@~wqEX81)<zUb?eCtW(ngDoIWUR`<B;crtWL<$dfJVwOBf7|24`X#}HR~
z_c#<B{w#`LoxA-!xh{_vT|y}6=0_Ru^OfT-g?x7mNSd&uo0i{HxexpoTDeroW2kp|
zEY`W5ensyr<^~i{J^7I6$2k7beLXhIGo#AtV@ZKL9X^Mbh!aSnh~1vBu3|!lpPpRh
zyi;*uO9o|H4ed3?Mf8Qq<U0xaYbBl%${?%VH{qDo16W!)7CET)m)I~`D@jzfSNODd
zyH^tmyKP}$glM1QY&=Pm^0)g?24;DE6d%k61OWfM?L)s;y)txld{xGERD1~)mY!R$
zuvy|JdHmz3+v+7!YNSDqP~l0&)e^-?IJL36mr<u+UGl@%r!({F)eC`SDE%=ffSb~m
zF}tm914vwQ0Y%m}J9>Qhx(4!>xTa;K7QTTBMEWh`*5;)cxIS(q>O50@2=5C7O1#mm
zYLwYTUPJC|nM%g~<;%ky8I7F8o>~VC)=i6w{{Ej_K7(h_1q}zMn-SrSJ|SU{D*{>k
z2{TV#qmF6V;!j1yi<Qt71ed-oEt2x>19Uvp%HZ_%gK*gbMOgNIihNCjdRyRU-qee`
zi=>g9JPX4|PvF8C2UVYOY)o3r6&T{={HE~sBHCItYZ6@!)sUR8V1F4>gv^@Tk|N6X
z2p?Io?*m;=|BFQhHAJSra4PnS^^)lP+imiM5x1syer<^KB9rHWjVh8deW~4@;RVaL
z!oh$1ktrACQCs<Ta^k;=pHPlI3K|6IjpeeE{ah-UeQ~3Vu-B8x>*8g(yD|=*<C&^b
z#1_&a2DIUF4L>TDC%Gn%h90YLWKnJ&Y{{o3KFOr(z~UN-<Lx#|VS#&sN(H}lrsOi@
zhkgXcYMO4oGX59xjZ6JP$#?eVX!hMX=w@}z7^1hk$=)7b_ymJB8fk>+U2T@wMGwbE
zt7G(;i;_ml5kK8xWI3(8(PUsYj@Vi3D8rP`mct=>1cOCYhVF#4pe^&R)nHcKv~L+n
zkpR?F<Kcklyss)UPjFwc?<p9rpG=}RFsp_pzo!WCaRYl3VesJlH6pXBuuuG;tiLoi
zwSv(p=kd-^Px6PZTA7JQtggAQ9x;YEW=1x!>y*qvYL5(;U`i@O)>0*;8$C@gt0A`R
z8d8zgDl`Uag*a5|QcmRL*NQgq7?FN3_f9(HUDy#aT}^4F4Mx#!Vaj8*aP%VkNpjOu
zSDMb=-4fJRZRfe>hRqy^eR)OUIXM6G1qXZXh@yidwKHg(<++z3_224%02{Iv8E?Xo
zSNcUrg-K)Q(?5W;$*Lq?P>*b28)dhKV*EDHzic>LKw>ek_DB5-+Cv0}SvIL=%c<Y-
z;`0P#)%}IJ9@;V2gAchREGaVBR6)rcH{-0^D^Bg_I4{l1)^XImPV?m4$_36qD}`f9
z`|r!iwX8Xx5c|1N*+W1gu2G1%^hyiVv@?AD@InVAA9;bl{nG%%jJ{I0`H_5&`VF~(
z=H=YfGcIXg$keR}S7@Ztmab1w)F-qgs^v{+X9QJXu$rK3M>IABeMIP~GTpBlNhBK=
zPTE&KUgL>@J8Q3N*icV6u)LYX>a&zaBC+QA&xGRR$!LfN%fFAUGyC9h@c`{%%y+h6
zvB{D=9)TCOy6R<RsRUT=LyMKWQGxYrY0}2V>ZqPHZh}`)6JBCt6{cMJrjWga1EF~W
z6?StHxt8xO(KzOzatZBvq5YNAPYjCGAR&!rajBg_M&_U*`3zY5$+FYuKk8IIS}w(V
z<;5B9E5hh`Au8qpZu4q-PI2i{p7(g8^!@fWISbK?<%VZ_HiY}6xWA`F75{oOk(Pni
z)YYK&CNOsx@TkRgL-ZYMI);q@msZ%s1w9f(eiVQE17tc~^=X#lNap%6#xV2*b0iwT
zmn!n%v&N|N=I38x%MEEH=XsO9bo2<BY$wQ?_Wlk`X7S`5raNY%nCV}ic_pu?V*SY3
zHtQ$k7-q0Yna31l+%fhc?&){1i?@S0%2t4q1-pPJm_jD=&t{3gQA{C8Yz-0}NBYMu
zlUnqxkEO}6{tfsR)Za%;{_MJjvrgtUDlVrBxbqZBiFH+N$dxQ9NO?u=-@Ek+QcAJ>
zs4Lgzi%hG(FsZ{|ew;`BMxS1>)2KqTIBoLiarOtE4^y3jq{>J#H5-;poAm-#SaL<N
zn4BLO5tmH;x0erYLo35Ae+gMmq`iO7Fqp4|xt2~bL7r;CAna5C%_#~?*x-^9N->uH
zHLwTG3XEV{Ym*6fofs^oO6c0qt&2yL>WX657IIkgT9}#tklupyR>-<BkB)3xws9-9
zB>xx_Bam=~gPOFm1it0$+ozO1JD?4N-+dqF)UMDcc(ogQkAG|GY(=(+y8Vn}oxj>;
zD>YO}bt@nIB5k<FRQP9Bzgh3%-RqT^XE19s<$~RX3X%0*jvs)NtA6Kx=@scoajRTY
z&&{)%Oj?(wzKQa@SpmN)v8&CNAx~4H=ohi0Iv^=sH7UUzEYZRVnbLzRX7p3_isqGv
z5Pzykvi0Fbl_yeh(A4P}xAXy3VYxso3YZ`K%|?+3@mo=!SJceOeu^ewJcf4Xnpiw{
zNM~Ow6;>QIf|H|2QB3E@Lxw##^77rdoXBRWP|kO7fw}W{ERq+vNr%0hXrEn^dvsTj
zzsl2-_bbi21#tVbv(WqKda?Wa-FQ)-L$N;uHf(F8U6@<{4rJ8%*|sblec_WJCsn}Z
z+`3-+U4Xm+Lo4cjUv<f`zfm~o(RTXbEuw@zrM3H>4@@ngv8*YA(=2~s=YnO%Z19%m
zSG;^$;@#J&ovc+^<-bTieyW&QT+`3y^tWt?6HX1(dHI1d!S|+%;R3%!1iphA68Kys
z5lnRo$&2=aO8)*d!}FP3IAyeh@D6vg!PC5HpIa@C9Vmu8eSg1^&mI!iw>yWxZO_&u
z+dZlx@RV49O;5TV;kwxqt}lY_Ej;F4Cc<P#jVT$ec`#z`Ds5et)fV$K_cn=l+R}6z
zMUYUt{bJ#J?KNrL{)?lcTmxQad^v6i?jXnfd3Z}UX7ZNCs_^g1s=cF$8yd((%yPYC
zT^rTBfNwtex;Saqyj1!O{WGV$<~Q<|AH_PDMWR%6OYnqHxU+Bq@^Fb{PO!ja1}>ke
z1CDCFF+0cXwNyL<L{j4*1=8C?qo=<2`cqnNRmWz>4##b)lLEONXi^P?m!yN8SLouj
z@c3QJdeYU-bP`Vm^c{MOZUDtHE|$JmO6Vq9<2I{O$_`g$vN_$}!hZbyR)DBqvVY!>
zTyC|v9BW<6+f;;fnhd;UmHzQzRwkUFrPtc7;Mv3LA@}zO=1&}BPGo8=08!C4suZrT
zJ{v;aE!@!OiBgVAgd)?JeNwk#zS5;H+%uZDP=_zcu`rM(|D;xpUFIE~Qeq9d9=aX3
z8vQbunZrbG6qk_KZ+cA@^LTy$>}C*F=p|&~rI2_nHxFX<o?x$eTcrP<IOz+VjsDi1
z`j5Z#4Zbt+cF`TwGD|pbKD~YiR>+ayEWm+aK>B2Y5cuLUp_PEbt1zf|`b_j|_Q|7y
z)s7X8!3ofn-dHZ4mwM45@HYnYqx3{@W7Xc)6Y=@VLm8S#I>ii;7Xso}{MFI375y6-
zVt@nrcJLHcCl6Q-$#$#FAMYfI^mJE&OavG<@DfyZS`FR=^W6zSOJgaVK6VeA_-*&)
zWrW$_yCprFay*(Q5SRZG23OO^s1F%{m#{XS`px!3N!}|Ut;X=i7two~8)+@?9H4iV
zUB(Oj>72cRZVe1I0f2=ix2(H%iLvRLjp{WzV!DHtrtdl^*b9ZL)1m<79N{i*9=HeJ
zkr7+3Cz&%{pWMxEDDH!04o=oP$ft~I!9mAN>xX8oM=U4^jTG!pIqIdF@3#I)dlTRV
zco%5$bNW|7#YT}h|G<)SW)U)_+Hkz*7XRdF5C?B^^a$!Y$}hmJGh|ggBiH*EA8GL3
z+5=Dr8~5g=PN$KRnL8^e7~3>T{X4QWm3w+I>4MT(=|x{qtHuT)nI)IU1NdulH>!UW
zNy`IVj~`!4`!pM#t5;QZ^+Bax<`wHFMp=1$2DH;8OT+w8g>G47pD3|`xo%gRw5K~I
zWli<2sve|idu`#V$~pk9J83!%bB$Jk&V{JO*ulU4A*gnDEBXO6sTmHFy?$zjqu~DH
z8Fu!2)QjJnBV&+<$4`2gvdR196pk(Tc&hBf4QA{=ID7ULoyKnx6-6hB-zL8?DMk1H
z{XPG~GRdh;S%D@#l!^49TKtD5&djb{UQ75<N|Z!d>^ru!u|IL2G5$WvjnFI+rFmiA
zx>-;CUWH^s*SXJbf{eqSAX43#&QWJiHM%TdRjoax>|uydq0aF=dkB{d+m1=D;y4?O
za8vB0h8v#sFPzD7>b$7h9$}GRn2|;Ps@(pfc4UCva_AR)0>#U_BC86rZ0SN9;I2&V
zIsNnC?bEzAk?$fbACpKT&qzC4x;_zJugqLj&u0+Zg}KAJiy6wDTi{n!=+XT~(|!u>
z#9u;G`O<sFEuMFo+}&9nyfu6(x!{d{gufwj{0l^1>Ye`73knw}-rV}nQc+cYRSq7w
zXGO$(XDc0Z_-1O#qlK^1+tZ6Viz(He{O?ayAjb1=<ibMmwYt@M5I_jT<wx7qBS6l1
zV||GsQumWfH_~m*@k0{<zkeY6p-!4+3Xr6c(T%+^-C!dmG?u5%Kp>G~eOy)`YQ88N
zKo+hC0CFD)lImAKo~K{^253q<^H!7gQC?<`a6HYWxrYnxh*`%s7PlDQwZR*apW>Pq
z2}jk<ny@YZ+rNrb&kSxXlqp3VD>|DMK=V*e0g_@yt(f-Fq)}L;TznQ3!Zq2dgmei^
zd(Xo;&&U*((zV~S6E8b1;@>atOfWC0ejedv!Q8Nfjq$)m<sLdAC++K1*8wAqQo@iN
zNpLJ2Lc)1+kUK`f9g`3`n?lV3*zAc$Av+s|9tIVKX@bpD0Dr!aeFi)zc8rHYcVNGk
zTG1cQ!r#cg^u20w0Z~y12M6wXLg21^HP7XAZwXOQ?lNft7`S0D@UMl~94@V>|8&&G
zCNPE}`G(2UQ+(FMqcIQ@37fHT?pP>%!3FXJd+2YB;Bq?j-t`3Lsi<Me>wt@J9lRL-
zl47{|Hl!H6y!W;7At2T2yi9i;Iv@8U=6nfV*G7Y3wajD`>AdhW{2NX@73zHFAjrcn
z?<TkFkww}Nl#ke$=*XQHtSvE}1`$Px?}gAc?4ZIR`41%OfkCt0-B%)?->QahVy^!N
z8{};mA687NTF|{d7B>`eLbF|v2n`&rK$+etCI-x>smbdev$lh&`cy^8aYYg9UCmYo
z^o2}KpJ0R4M$wEvyCDVrj|;_$6VapThK3ewg=b&iwbIFY?LDC;9L0NDj0nq!J9REf
zGR-yH$UR<?Oqbx&b`O0(&ayaN7+Hx!l_ZbL^J`wx?1SscgP+htp>cyjgDuO$%GlP}
z&g;Ou0`FYU%qHuTa1F^m#TjS4?OnMt-op3$X|%)-N<Uj7_<37{<nIo4fvvW(?(fkL
z8=X7d5D@S9?1s`C@e(R#@kM|`|GiLVB>~EVW2FgHBzIHJ-F<^SrQ-XHb6$BXnty>B
zrLLV|M=LfA(PTdjg{%ZfQT^8^nf;wJG19S62osi7!sU!;c<la#h%Kj(&ygk>_Xoc`
zQ=#2ImEU_)A<(|guDt_s4FC0Y+t>T@bRiu!!POHER@U^i!Y0VtKS9nT{8o;s-cSGS
zwZq+Wx*1w7hSsO3eykOz@C&C<-@J$Z9u1en%^0t*2k5k(@jl;44Yh|-2@Umd1r`_{
zUbW};*G8@_CZ{h=DTDYR^<<NQtg@o`CM~GO)4ataLU(NM2_6qV*nHgJCb2+?y05(k
zU$oP&5#S`)|9ZM_2~IfbB#^5`Nz8}`7~POf{lTcc4>cY#9zzMlEbh|w>^sAu`fBB~
z!Z+xt>K3w5QJuceSAddFjx)NX$vbqBgBi)h7Z+Ini+*0F_6vqd3wQIP?~Gc;Sp#W*
za5LwRWEbbiLfBMEAzUHu=`J7$+Z6}F8bOp8Mc>O7VY;iLTa{kGWEdJ29n%sgefd{R
z>(9%O!FaHu)Bg0hdG9jgP6Cp9?R6h-wXc?DJUXqS{NnLRzGS++2k)7Ms_f&V5Zz%%
z48Qz{rim2@zD@Z%;@ZQxyHKN7$Io_6D>ff5?3!JQRW3c%vXGtha^#FfN%}DN0{*wr
zVU9cB#--3-n2$d<^DTlpm`Pdu`_OvUg0fB@TKf--Gd2Xy7e?%+-<i0j@efjyH%GAl
zy@2oywmpkmFsV?#eYnIs6ol;Zw5ZPcKxcFBTZRnHqDQ@9Z)jniX|K^gzUL~v>ulvN
zQUj3<*8LwaZ3VW3EBG}*6g9KB2ah5&oagOofDawwU|u|Zaaw!<J7inXo}pLj?^i21
zb*5_wLM1Q9>fgHr&1xMY0n^p#I(cG+>8XS7VPAjK`a1lH1TC^9_BtSg#EQ4#T4QMd
z&atCndJsA<>NPy5QDApofodu37#z3YDdodP1_L!xR>zsc-d@T~$IMr{rv^JfPdy7;
z$Kuu@gR5P-=dOrty`}b6=<-8Bb5-BrB6fw)`muIujG=5PHU0@UfS`Fi<YP5R_DR1P
zCm9oiT#(uughDuYlg-9I_!swJvT1%37+r7cg<B{;D28?ciT9$)4QoiPBt-zd)KF)#
zeC)?-#sRm5=&ny(+8|F3wK+`*LVC<%Vd@!LtrBa9gQ_#Tex~fjCz3p;#bQp;G}Awv
zy*cuPw=gk8)3sxoPokK;GcEIG2d~NbLI67A5R&qbPc6)Pd!en4Y78;|aEJgpXE^X6
zt2uv)(1g$7Mw!|N;bMR6vspf=g@b78tB^xr!Vn<x)^HyvCeeFX!|fDFRhbgK0>?f~
z$6o?4Y7H+sW+bt6-kIeE?9dYsDYyCOX^Bh~RLR|V;C3ak4FV_FU-RMS5J@zIanwce
zm%A$dXybX=qNk}FW7Xz1AWYUG;sQ~g>hg@VDgWx%pf2lv8jJ$H)gAa8N#cuOO-0v3
zj2}C4kxHj?<O1{<Soco#nbgZnmgV~?yiW~n^>t%~nj^YzD$EnlvIuI_s0@69Lkm#N
zkj=sSb`wa30Oa-E9({9ZJ?U9nnmo*g$wiQXYgrBt?q7HM&HPVpU9>cn1!%1Fjf7~n
zXO)C^2enBpufe5r3rxWrXM0Hp+F8!+E*CD5JhB4@=ga0}20KF~HY|FNZLzSK5j)at
zna`ux@eqjmu7PR$7U#k|b<2Romo0$<N*^ghW4#ay9(h>Bu2vl4F^+m9#%Y%*=$s$x
z%Ef`h0!*wBqW<2V;)jJwz^4t;+LG6ZdCYH(r6>F(Dz;8TT(2>m_-BEe<GHiR;l8?T
zde}WQY1x<I6l!l;d;!T9zaxPtwHMmQ=EzAWv8v&$B9XYog8W^y9tZvM!sNHd-0dC-
zh|2U$9=sUCV)3{_)uF^nE)XzQ$F6^Oh4PS&4+3I+-QNXJvoefS5&vF0UHqDpt5m_#
z6Ibs&bY>*|8P{$ESOQFUfE7CD$2ee^Nut%PN2^9T-GaxGC-NC&OG}+{Kko|O(wJ|d
z1NvKHFGS8DJ4flxeTQnZCE~ILqFKnb=6626IWnYuQg-xt1NO_*z#MD(Y(T{GWg50T
zF$c-45oW7=?S#3Hus`MB9jO>2jyW{(%PGOys>F~LczjgQ)UWSeOVww!f@}QTr&mTp
zj2!*2a4XjpA2y4lHk0X(wKm5PO589W!4;iP;aK=e&*9H?T)YxN(j}%p<zf^6Ws;EM
ze-zJB8vbnaxu}#RaZquY+t~Ozu}=daXz{hYyT(0Z+4jt#22Py)F-{Qe5)h5~bVaxd
zgDS&oeq(#yC#3VF{1ZOmF~`G`6=I)gibD(F;@b~5oyT5He5li=fNtGMX}$ZgF?qS>
z)sR`>3z#oo*CJ*dqK<XqEiUyUsA?YIF$!F&0rU0zqx$axCEwJcd86=&9X)8Fki8wn
zl#<EI(8?Gd0{OViZ_aa=XZDOA!`4t+U_IvaX7?e0zR6rOLr76kS@?lM%uYT#s*=Ym
zQ&{iXtevsK84eO9FMtgS^F+x#4fR#v5PH=Y@fC^tQd8dGLI8H>zHY)&iJmn1NsS+C
zWKl5wX`MF64T`~CVTJ0DT!%Z4A@6)wg-SO4S((~4lb39(H73XAA0(0u(G8XI)QbKU
zIL3LV(ts!4RzyjUwFO6<oI4t?vdR!dB=$ICSbH^}8Z^J>PvRGdu+9p;^)jDbEoBVx
z(l#fRN8C^*<euWQ%KLS3%7Nr^!)0pc?LjorK;OL8P<u`7e6-oylcTM0>2q4^x_L;;
zT1+J3%SBLA4uPV5X*bzXFp>*NC7^ZQ$r$uqv%y~dDChDO^4*;Wi|Dan?!L3KGFJ3E
zI{NcUJm3XOi@rV%w?7MGgrD3%lp;1Y!vz&F*miwKP62~O8ZiE}Id*s@w}!mx#hnVy
zCFmtxy5#qf9mUU1$10tWz_rase_zu&<?FX^w<DF0ANNnWGrn?4Z~&&!p;hV44)OVu
z{o^Uu7H-y|heamcs=49tS4nsx{ERpBVABY7{KmF>oIMrk?`}|U;XRQ?^diIRN7}IP
zPE1Z$Lh{GWN29s7h6viy9xfR(f`8hAp6f;w<g3KT*HFC2dJNYQV&Hfd4)_KxqF10?
zIM$<DV@t&-y!NVhx(lC>djxweEE)At!qOvOnw>*JyiIyv;(xT+x0GhZ>qnoyCxj2j
zoPTZ@lsgb(9FmlP@L}YGmkZvie|^EZ_4=_nuNi+x6YP)Xv>R@#{8hV<Uk1HrlV3k&
ziSkjRqWKbzv(t6G9xz5k2}6<{G5Z_fae93__!Jbs+A;1+UBj$Q=JGuPokC&Z6m|zf
z#)x(y7vX(kT%Vvx<T$}|vE4I+On9U33dvP5+eidZLnxhmu>SC}=P^GxjH}X5NQwzR
z6v@_vp22oz&rVwx^O1)GX4lN?mhN?TF)MAeTK((5Qj(KOV5R-JASAjmWvDjiKBc&<
z{P4P_@fWz5hKH%o*xQ&XVbO=$32~Wml;YE1D^Lv#q6XXdW#V59Q5qtL*f`2WL_^xr
zV}?@A|FaBJ@%mEyX0E{v+LuNKL9)4RARmOYv~DK7Fwlr&)4Qf-DfL`FoiNHQYIBGg
zeHQp+xS!neu)(A@%0{G#+9_1I3a<u{f6&Dtk@Cd#9$K~yxcB^$t;4jSx01bUs5*m=
z4mfZ+ai&|w*ibBKt+O=}tH%G;RO=+4&dLk+ha!fB&jcnvw?Cx+DDc>CYe;y32*Xxk
zEY}&mtXTEeg}ipJ-)~5xr8LD?3|V!6?KJXr)IzSvn-9YN(XE}-=zE#~m(SQ*N`7ZU
zjWA_2jQf6<kB>^rg$p(A=#S$LQJMuY*uznt9Ros7$7kOmHL=hprl=P8I8UHLGb)ae
z;}^)0vOEIjnNHZMmr`<hgaH$x>c{0F<u!Mf>sZn2g1cQSD=S4XQ=JB1i0FA@Ee4#m
z=C{;?rCg@3<2_VRfcSO}SZrB_p`rN9M9|IpE>J}?EvHVNx=_-eVCdUb)AZxRe(5Eq
zudk54i2duaUw6TC1VKQBK}jVN)7D!T;BhKwYjo_Wl;AQJy01mzKnO$bTNyBhk&>GJ
zE8vds#Cu~{HP++!;%0krHmu_*#>FCrHNyg464vXIy%J>NdDvMwvwF^*-4nvfSBPz<
zK}(&vl&4)CW1`FDx`{8~;?3AT>hB252X2`2>v=0^(w8_q<Fjw-H9Dg>k5xlX>}h$|
z@uE(y4A}mPfjyJz-c@@!&C4k5u`Ot(I2TdQ%2%!cc%$^pmv2_goTsow=GHnMlO-|@
zuEGwKs84v!*J{lC2HBipzP&yMKcU3yWXOsOsOo&AXtaNQ$;+^@B*h`<RN+ke_GP11
zR2<>bWsFCem`$xItXVU;i8;(M!VsJBCh6))E~P$ZUfw&ky$MH^wdEq7HY34(Ti1-V
z4Ey-a&}nrEu4M~ZfvHiwbnsBVgaPLPE#K!}`4D$$srMa;tr+W5@pG#k-W2bCkWtX~
z`WJMdSQdvJ<-wGUm^#}p*Uh9}$Ea}gphhr)m7GFRNqf)hX6hmqWbNo*=nqKn>Sm&s
z#+<m^+h%^>%RP9)rZpsqmS+6}YPJC96VOs04|<&jdXJUBGf@|q#tI}m^aRDj@fgK4
z5}D0j3eR%W2pU7-#bawLrd>QNl9yi*jgj}|nG)T0mysT)5y<;_Mb3X31JDl^ElwDs
zvIw-n1>_y*8#a&Hsp!`L?dF6QdU^)nPMAuXB5DcERvagS63jS_sZq_qPS3Wws?f|6
zH5#B|C{$poNpEyHizCh8@M`E@bOZ>MiZBJAA)DNpH98x!6FcpS39j{#4W%2*ZxI0P
z^~$TS;d7W#tY(H!It8>7`&WU6whDVD()l4>F&=`RrnSx1GsEH-{yYe@@MP2LCgIhz
z`%L05)_BB4TAvBdB%c%xXlFAOwY3c;rGFpZl+GXbd;d`z>_PQvG3SR*7~+}$nJezb
z?AS=w3+mKg9XfvU)gCCe8WE{FLt)H__8KI+#Djt*Z6kdSf2i6(Ihj%6dqXAl%k?Qk
zPNrwaE25rzNlYJoOW)JdO{f$SFfIU@Pxr(Gh{G<+a@yXpavGOE{qXC}i!~MUr*oc=
zCo%|`YhHMA)}|c8j#58?%iFU9A2;B<g8g-j=Lw6x0GCmos$D`7_*1TG{OyIY&;J|4
z*8|9X*dI#{04bp>5a5?$jp?&1QB0fpfSf%A3?>JMeI1`b)UW*c-*ieD#ao&uat@#-
z_;IU4N*0<&oefi$95ZcusR@_xIh-hX9OGeQhN<EDtJ9LV?fb^sU8MnuE}pyYI)E5y
zBx!bHq{%EbGvXB}?h5e;G~MeK(n_CX<awhx`u5IB9Cco1vU1+-+i=+=Kb;Pe_8dlU
z7|r-_UlAa?q1*e4C>NSKw$p%jI}Tn*b^U}x&X3Pytj|7wRG$&h$V);;`_s?#vpxDA
zB(ktbIQ!I`Rt}DNBE571@@sCIBfsNY0EUW3^i&XnW{?&5yNOYa3UgJP4<L*HW`VZa
zm}sMv?ehS1M1K|hgF(&X@D3?i^5>u2R5{n`K3L{GNT2NVh~G~ly*8|F<t?DvknXvo
z-p{J(bf3X5Kh*DBG~?Di&AZXNrIy80$41y1b{MFb?6Ph26INM}SEvz@)v!lg$yV6d
z+;E5R3Vb;^irCfCe~S@sBHW@9q0-L=XRPMYNGIi-3<p&DGQj$r`Lphb0#3Vs+%;mJ
z@1NalsO1gJjtsNii5t#5`u>^DeZVXHj|t}@v*X)e>OQ68bDT=y!fAtPo8IWP2Y_pb
z!(1a=g#M05RM&a#W7OweZK?H0^sfU}pIR2S!J5<$jAfj{3pG%Sn0DDx`goct28(a2
zj_Uu+C@_VS^3|I$j*uI0q5<yzSl47mwK&Md`cXEQkwurQ$ktQ@x@2s>9_OibTm7YN
zMlDcX8Sgpoiw5bJWeTBwu#%QoR7`)j*NR3cLVps~6cwzDkN9t=t(puMiO44`E%%DX
zTWEGx1s2<h&R6<`be!4kSOd98DF(DrF9K4l;x>lQ89GVEXtt#3>)nQ6Qz!px7ddL9
z=D%yzi|0VnQ$SPQ)e_=TrAtB>gZl}XpORaPzkhxqhHYfB7|Z?R4TW6~=_|Jdkg=Ze
zjh#~@YcqtFfnSVH=ie(}CBu6v8%=*=0w$%LI{81|*}74kO`5WfD~GIhsDZL6jrDvy
z#$3bK=h7sI3W|3s(itCSPgnZ_gj~8%(k<#*(MGmL5wKEnpwl>r<%hv{Ucc+pSN-P>
z3lV@o!U0CypO7E>Z>lm1q}<ihI?^k$kAjtj!bq^WS))a(_*4AT3e#H>OM%2bj2(fM
z+k;at*H1?p;D%)M$n8(C!X+6IiQjs~fKGEev%Lr(STeWCJKc%>i%<a#AZ)a=Ewj!z
z-!I3yu%f?YDI{BGf)qeKZJewJEGeqyo-l_g8d=N3*Aufhblrac{_;El`^eFoUwqL`
zCc_o|myR!$#f#eOsJ!rjQNEBxS^Y_J4oO}^;2R8QL*DvG<M7F(ZW1SI>Q}1_=mhuR
zjCgR8#Z+E5kPe4GTMlJC+*BU4S$8f~o16)va?f+Jpd5E3M_=iC{Vm|MW2W?ib#p4h
zRwnG4B+4%Dx0X#ehSsVggN5wF837?;gQtn@3wR!1nU0#8>-yj1MiE=O)Sm+e_uBa6
zC!c==N-MgnYTrtja;vjI8O^Wj0ZPyv$sg@SNsp2*B^Mi#Y4C>r{t-4QaB+IPQnEd)
z=j`j!ddG*hBX*3o*-<SzQi=}|XCE44Yi<f+JK}}!lYscJnHVbZn4*dNSwfi=AsV*7
z#XqCAjFUoI-6yd4vU)hrAWT>&zC_%+*2&^sbh4-YBYQm(f4F-dGlde;9`l?8pUdCg
ztLx5tC!Fs_idg9_b@|f$GyfaSMcNqw6T{X>d-Y$CVi#>FB9toU{A9RJNZM$utif05
z^3CG^kxIl;`i$d7R8yTE$xNc#jMB4z@_@oaf4QH3X$+MrF876YBjqnJAx|}Q=<t+d
z&CI?}aLbEFVf8(jJ96CY<EORtsJ0tHY3K8tm@Pn-O){y=zii3Ox2BkOl$4N2^2=tY
z4p~fZpJnw0K>4!c8kR=Sh@y7G@wl}LkMf-vGSY<R9%2cxbIAQca1E*Vk1h&gL<j-P
zVrPtbyh>Az(H(N6N|90?5gt-nOuNY_?dOy&WW2?26U-zC;DU^+*<gB(Ug*?I|I#>P
z&V_S?lz^BA=7*p*{aGgG#x-e&<P)2L0gJLX6J^;<>_`f?V)c_fGwDb3s10AFhIz)a
z5_WsR`9zk@;nnxCFkFT7@9Teg_te$WJ#~tj^hWf%`IDyMUr1?Hq<~6#pj&QLrQhCn
zd3tLDm#)`Jr6YXJ>$Y6PW#`P8E=?7)w>cgwv|`S%H&Yoa_asv086y6y0UH!1+tqpR
zI<?vL5~bf4;7jt8UjkY$gU0>RjRE`0J$jYrkc6)P=fu-|-Hw||5vk^{+Qf5eH*dRS
zTU@{AVd5~Ij>l$@eCZpxoQk3!dV^o+e*y|LyP}Z3x6BXS^erv4KN8O0!bbeF@WAoZ
z8>LHB{^%H}G7M4IL=yIbyGHWpW^qht()ZY*i^is8)p{UiSI}R-@+3lKp74g^$%uK;
z2jG!7Ct2Mx&p9>gF(*z}u@&M(>P<0}w;k2^nMG#w6V98ZXTg-e{$=DR%<}fdv;tKR
z22Zs(D~y0q<XCh*N#}_iH0dcD@x{0O@NEIpt7ZUZ4LR%7;YBVb#rZ<_Xk^I>dkS3~
z8c?O{_7$#pvum>Cs~qk(AysE&uk)s2%pU&p5-~yIcCj3YbQBfSd8vNL;Xj@-NH8Ln
zotHD`vw3bVzdDQYA7SPlJ6-Imf5t_cEGEC#eg3+!$tdmv<W~)<8B6Sa5R+_7^JPK#
z6d^KiaUIX^7VR;6W!9H6HbtCgnRzF?$<6S0Xi%5ciZ*qGls@li$fZzCVfSX%WqiGq
za_v>30BZWPZX6JWl%{D>zhr&JHM2DV>wklO9xgHYU^DZ91A7yX*!!gp8Xon2kS@(C
z)^T26k9lY=4@_G>%08~W>A2Dk%t%Og)5)zz%;+olPxY=w9@@5!o7H$M1*xC>F@V(x
ztP$sL@HjxIWW{CQVq9I5&!B9xUO^(jm`vH<e5B}{iGJ9}T<^b+G2g5w5bnfR=GA+W
zk_Nn$erokE+y8K9smtX-5t-;zeep2U9Vn!h>sAlA&S)<Cp4~x4UbcN1NzG;n38ixN
zkE6s`DmvM$>jgc>B`~f0*pHJ;$E-1-;8gKu{~mBS9u{jeq{$My(j-@s*uG#H&pGy8
zQMQ^>!{j2H0H0y3nCsm`P|jf%YAWUQ2`;PE!?+V1zLuBo6M9AL-THL+jPLB@JhFmE
zpkzvLl?v6!tGULZ`vPe+Iz6r<a`kb(P?wC>xFfwATu!FXN}{|J4A)f^Vi&vG1IHfT
z9Dcjyl5~m9L04I=7uiKJ&3ZL%U-4r>!Pn$8ub%A-@H2Qi+?u}1j9Gi;WyUt)YNjag
zt+d|Fvg9_1?XINUhvoQ0fUk4r6?I%WNU#`L+QTrT_Xs$|B8*Z`wGihwQ=6qD{zEiS
zqLU3{`QQ#n=fE_3X0(}B5aiV0oKl#}eUyVP-^fhvfqrJEAdQWanZJt`1#XK0&X~Mi
z@Cx0f1rzI2?_xU4*-EeUCzmEd!>|cw^c*)g*7)2T(n%=$P2U0eYiFZcUFm`!6F1ck
z<CjOm`AXK`bQ^taf~|cQbYS24`Ey$CMn}qcNgREj(Ytc&STRgvn;<5}H}D6M-I(S3
zVCOxZd!-`w$VgudNJ<7>_=w*8Lk;BFkM)p~%3eGfysSHgwc6Gvkf|>^&J*<I1zp}Z
z>y49<`GMp#_=UML4xN;sSz}z7R9=gGU`-!_qrQQ`9qdQ18o?j%XiFIgS{lb(luFpx
zw>cK62eX?#UvidH-r02?$yQ7qCe$$6I$5Q#*jN#t)(s3upKpr0dPUe4$yHqH1Bf1!
z(d|~`<1N8E5)G}OY^<$bZkqZrD9OlFy>EQ1q$0g<u_urvC1dIQ8^nv?sIy9UP(^;X
z92bFJjwNrtvW-H)*ezCZFsoAKM)OcBA*n}rzE_0e<D-u`UuylCY;cdM$&S&a%{cyN
zMT4~s|9$V8Z<!~#>C!!r4Onv7;O?a2e}eyE_2xiF^Zm^`<tP*|Vth5C@LfE8*VQh^
z*awTAHt-lna(Md#r#}nzaEBfnDoCm-O<+GEQgUGPt3cE)!**)6wP#y&(||fsi|^3*
zIM%E!Z?SUFvnAP=dO5fLv5NAF!$sBtllMA7IHs8n!8K8SI(GwC-7{hsMp0YLf;^7O
zHYjn(4@UQn*5Qj+1#+`8oKo@jCfn~jm_40SYO|M+3&Y8j^HK#xNt&Tpym!HIN{zX-
zdhyaRZIV{Bc1B4SRE2!Gp%gI|@NVuWAgtF-<yrKL=ZcARP8e1Zj6T3@%PeTy+rdl)
zkEz*BWd-x|KaH_d)QNLV4qc>&8OgN1Bo{9H?!>qYLo;D5^S8P<QdUt36)x6Ki}nb|
zx-&rjX8{%SBje$rc<nq=WjarDx;iE;KKA<0_)4yqF-6r6Vw;)z3c<`*Y5bb#DoIk^
z<c9x+6vcH#%g1Jn9mi({XTB$N+OC}U?lJnNI!`#L<jvrVeUKJne4D~w3iE?I5gv>V
zkAs2IN{<U@VXZ0Xz^I+*K)fGyre?1$=7%h_xm(1t2TUod-EF>qL*Mb}Y6?lpmv-#_
z(jcPY896y=``}55jCb=laFkCmlA<!pKhnMRiZH{49qT>TLrq4m%#fui+g#?e4~`)p
zRZYJ;9uJbtZz_{Zgg#4IwfM~|d~5wooG983*7%f%A3$rD##ge7Joe+izb-jsUkJ3i
z;+tdZVxkGw@02j58QG5sBwB4~W&!Hb=gT(UnI}=RC-~Ew{6$frs7yG3-^w3!9)+`e
zZfVeyCZS-=u>%6L$O!P3Ex2Y1KO2Vd9%!5D<%T|HdyVr5yASIlj8DC7Y3npY=Xb9+
z-;}1wvZO&|^QYu@#N7X~`32&-=bIM|XAj=C9JgzPe_oKFw>p&{z9WyaHR^v~cK!yp
zKsVy|(j(yNE!s2pdN7a3O%Pt2_xg<Bu^l4ylhp~y03|v;2S|$tyzQ8r<V7Gefy0U0
z7)fVb$<V9XR#FEm#I+N_&M?8;EYh`iH%Iuy|6Ir9kYDlcg}2MLm(Pypyl^ccTw<MP
zznYLIQo@NrWfS&b*HVR){!hQcM%mUr1b?mmqIoJa?N}Y5R;c%$hH#tF4(8KUjzydi
z$w=_&Dv7+G3serAxVqjp%5UjkW$H7%;<um0@<fcT8(Kw+u`cpd3IFF%%3wmoFuwTl
zbFjE&EDT=iCZ}{Qe^oSBe3{N?Ds=@;VL(%5L2WC{_3Qg*cVfk|k`|fP;c<3U5Zg-p
z_ty~OIuSn#f0Oz<^Q)}z79L8D<S8BRApEKF=Tzp<OFqnJMKj|@y%#Ez9SuZY*8P1$
z`t|12<0j<#jr<&UwFd~_v#(NW+i63akMp6R=u(RYgr#OY`E1V4jvxOhww~V(YzQpc
zf<c1E=?Mult>CdtUK9!vIgTb>#cO71{4o0R{`u$o`WrN%JP~yUG~GRzI2NkQ|NRYb
z7FDBnN(jSfEgFuQskIEg?9$G7{zaSj*_Y=f`3f9PqJAP-rSX0v=WAeW{P3Yr`b4(L
z2JdKH*w*4=U1SY8YA-tMdk0gIq^mP^qMqRb7SE-bS?2|7M$X@Zj4)C*r+5MMwWr+8
zJe1ds$jC_TUD~@b?BF(mU_$_azz0XvMOjQ?eKjvEItg2UCS2s0yMMg9C(YUX?&rHL
z8WxCO#{K!1CQsyX?N6NfUBdiVFejx352sBU1km^Tu3wsPTp<i`U%&G1E~J}-hm%p4
znDskQXrS+KoYb|~3L~CGH^=3F`fK5y$mljcD1p*SmZN=x@OtvM4J|qm&o?Amkh}q!
zfY%Jz5jHGOwCL%3U&^wi^sQHqEjt%5BG%K6*!FA^fe;w%3NBLe&u3cSey+SD+CAbU
z4XLqA{5ZB7`fCtk07=Tpm}XEZN_%ymea)^dSw~U$mz-1x`(yXcsya|c!d5pwxEz_1
zl6Ei)kh7CN3nvnj0`TMa#85h^H2b3M7{gIE?EAhWunmFUVO#xc3@4J?d_iEIASP_Y
zp@mmM?@)=;xAty^I+RBN%>naFG(d~wJ~-SB?F4!U@zJw=Z;P5*?^03C!S|7NyWQrJ
z(O@Y)^urYJ7Fh&A&Y=Q>UWg{Ix*$~7ubFTZxx6dPIZCO?e|CXbd57TJVMhfB6xzz3
z^L+eaO8p9f@W*9=6QBX}eO^&tg+{?4o3x=EZ<Q=28BgQ-DKW`!Z($PP*tcP)70+SU
z%vS3xZV`#!&umJ5un86(Qk=_ixKMBl`MZ?=Q7-xc$|M`jyhL6r@y9U8iuDQrNE%@q
zx{F3XZ2;UH#<Hr$gu<7t=SdXR%y1hLfGBsEEcUze7`}KShhEf|bm~O^eRn8;ZS6gj
zi5$Faz#R*gO&RDAas_0nI|AQJS1id_2KWZNxkL^UvU5Fe4ZJH2jLdb}j3iA2-5Jvg
z56iJ|KTPEb$-5s2mz9gp+GuKeWe#gL0dTR%_P;-5&FpZ5!J+^Wu@T(q|J{H=l{#ng
z-!{yBZ!oe)hl&|Wd;c7|B9BBV9bArB@2=`VX6{1EDfsW?#F={<cdvXvqjagY-LsfV
z=nf^lM8i=&Ex#r~!~#=)Z5x$jNrqwc3=7Y9frUt$Tw2Q<UkS(d1l83@<=jAWcq(&v
zb&jPm3bY#{OTL;j>98+PfJ>pI6|<=4ocwo~sqT$Vj#9)sIOaVqOH^Y!vkx*Sb1@zg
zrOHvF`z)dNbT5h#y)Srml+VY<7g%{J+e0ofoEquU`b3oc!0GynSUKy_OLNKjFC)nC
zwT8IRe?bbT>A3#P@kBOZp1|>-pcS2nAiN$`J%%K6jHG!l03uR6_sM1cL1`o2>oP$T
z<qQam&hzmw58Myfv&ovyPh@!O+)@nlnkc5&(>8ORn*ThzHB<NKDi0ltlzt@)O-*oG
z!K3gK;2q3N*GhW#f4^Ll4=sOlan!l`g`8kzKmy6%Ay7%!*q$PD(nE9~VH1|m0Ne_Y
z?vTgkB7V99wzJ+Wl9cDw(l+8#*AB}MGrqRmxRXtCHWgG#WypQH;t!h(54RtQ$%3;R
zm`?<{bne5c0HAC)D9;>R!cwBrs%!pR@|Jn=^Mu{A*D&Ik#j7MZfdbQsJb}zVt|nKH
z@I%`F{vy9V*Y)4!cmEJE^H)GbYnaHD-MumdIGJZkXso4c>O)1z&+4bn7S*a`DBkrx
z1Kgq?Os-TclKqieL_q_wEgtxu0GP?MIh8hCgx!g<Jp395yyTZPf~6s4npV1$=yC<z
z83ut*$qU*!SH3|&wDp|%%(+VoKag1Izkn+*Zm$49&V5BfCm=}^zJkdP0`mhOfuL?~
z+q__n`LWfKC<%|FU_zQSAwvV;12ybEcz@(0YlZ3mFN}bg<e)XggTaL4YJ95l<0Y=R
z3CA(EEzTh2ADERX>}TSbimajjAqxrN>l2o;X{{JEvSfSs{uTo17txSbdUr)Ne)Ai{
z=vQJ<U(B8d{nQH7aDlijqrvdjBTIfJ_!ZZ2Q?m?f53BvIhcrhD{kKOSU|eF>Q;2-7
zUQfEC(cgf+VKQ|2^jd4IbUTd*CMIzRP!ymE3PODpBf$cq^Kt5Ju{$_qw!a|HRHW7g
zrtxNI`PWvub4Ph5lDxOai#biLz7|5T!Y>RB$n+o-?9){-KEcd{+K79OnRx>*{tMnf
zlbT=-w--Kx#uTpRr3RZiaaVj6ya}4e(NKBDp@0svq}r$kGxQ5bGAzx(_rZSyFpw2A
zdx=^1$AqCW!BmJmxWT+luN1<9QXXTLaOxb9{93~U6McL35YPfKU~sawA3YM7o)$xo
zIBzl?$hQ9XScwvH<b0ojJ5qB-I^vxv(_89lQ=geMYRQ~)>XbD`*C>Z4?g8Tldphky
z@<xTpR|RsHvr3HZU^|fhJpx8&=w<YNkVO6vVSjRqmNy{<xP5)vs85!6RLyC>ioXWh
zQ<D=QWJtnhu1ma7LyIasU^bc^P`;XDNYpB*>n7~ZgF}l<sg*8j#<B*eog60y^r?i@
zlaJ2l(s3$kPB3jNO`M38$~r~wq3sHB0Xu{0fRpo;rA?F(d$@CqN?#4DA0djPe~)zk
zna_0*F(^CvToX8^(WC-cSV+whQRD?!0XANK1_`$A>X!drxe<XVgN0$06Iv=BZ&&y6
zVX*Qqy8Ueo?iQIx>61vVfcK_Q(v`h7;$nNEXz<VFz&5pCJ6-Ml7S~T?8CC+(turlE
zOSsA7`|GquJaTksW1)5$8Jw>A@5d)<mBSV5;2KpJ{n0omH~Nk#n)qDb#Fh)Iq<GX?
z+{!l566%W=-6r-P+C0^5nZ@D%G{UWB8<<r3{_b|G$pD{P+X&SZAhK$wSsGlm*8qb6
z?QoR8G@Jf^uT4vn4Gky#wNt!Fe%Xj7Y_G&jiavV6eLhUi6~V$6oLct(kEizzr~3c@
z$B)xdha|_=<k(yG%slp9*?Vt=NOg{pz4u-rJ0c?^WRnp|QXvVcgo=`W_w#yxe&6eQ
z{q?%KoO7Pf$GG2b>n=@)dZD~br0*e%n8&mZ{a)m?>s4_&drUm339L1TJM0inbs~(n
zp?$TX^DDI|an-?J504_4#fJwC<80Na4xpR%*Lm;v>`c<WK%qwxq6tS^F9`+N`Ykus
zW3cb3C10w?C$?j&)X=b2KDNH`u>@RY;8V_@J67jG4EglibZLZ|(0Jo5W~ij1K&`fS
zV_MlN)$o5mqnD1jB*`a)W8Pz!@0Efl;_fTMeJu>hrDVtyAZ9Rg+X!Zxs}ijJq3;H_
zLoTP@rEvz0XT2<lY>|#@bE#nuKjG8$u8x8i>M6@eCNBf&<Wc>7w`?SGIuOd8gz%e(
zd!Kyg>t4ecQ2gF=i3iik|0!jN=i4>nhrg2Hy+u6kc+fzvtbyCZ0C%o<jD*Y+${RSv
z^1#;vm$W9A`0ZQ8-uj+!S$oBXs0V#r+T<L&QO5A=+k5<Fx{}j^M-mzz<^LT{0n_>S
z{lCAqb1v6~nbA~d5>qpF@n$b!it4~qz){$)rPgU{Xi<!aFnf3;BINCBXbPkQH9h%N
zf~~1Q;3sfXbB%c6SYwz|O2NecMDt<S60v~=!Sh|9;Hb@1C|;E)M#Pls+jwB6&!dFs
z@Ny6w0l3Ni<LoKqG*{?9!2JfJ0|D`Zz1ssb#8xFB_${^<c1KkkM9#1l-M)(imDQWB
zkUfCNDP%eYP3S97B(~Vb&|f|^p4PWED|+U1GYrp=RL52>Uc`hD4`pAp-AHfoH;H=M
z=~Hp5&_@5M$hLvB1f(gR01^vc=8TvUQ6zBJ=daNASUEjCT+b&fC2AIkX7d{;=&Mb@
zgyY%JF+{I{R>+YH{R!=3JG5w^#CuhH&39wE^r}eaA~GWYi{1z)I3;Xde{jU|V1WMc
z!__!?cwNkB6gmu^#Elo_8adH0-tG_%O^72h!E574p8R#MvVdgWJ36?OkIgH3%Ap;`
zXHg?oQ3fLd=G9Nb4y$3BUG&+@MPN|=kSs4yiS2C6y!+pZZ;a^$<P09QcT<z$^8t&D
z{V&WP!^XJne>6&w$=G>3W-Ge;+kNa28-KLb|1ei(OpeS=fuvuspu2&Q#DkWVFe{xR
z6!St+aTz9hUJ$ZVfae=)$mYw%<-z%44~k{iE2;>})(;2XAeK3XNs@!U_vhLGm2xUo
zy0=Ze32&F9u?iU6Wji7U_un}Zu3(6^f=t^p=S0=FsB!1ODX@87k&I59b4qqEL4yim
zz7CL<It#u?=4(6a=GWt&60LG@1>YoR7-Rq610xRubT2hv9_GG1nwoF0)+$ghnPYHH
zS7I*3uvEXZ^3ir9yP#|h1cza{IyiACFGwOfC}iiMTMII$c8>l7{C30B<6G;IMEwMB
zxo0@$CcS4J?9;*Dnv+D!Hgm~2Jo4@jloeo(Cbgy}XS{iLL8wix`ENFA5eq3xl=0$O
zKDYZIuJaM`(FFbe=k9u#42nORJg1HhQRCAjniB<*{!0A0D(4h?AD|Ux7GYLSRo&;g
z>JLVAIQ~$o5<A&~o}c?49Rx?TuRFt@9HG0lRsBHNby601<lQtpIx*}z7qlk=V`cmL
z8G+~FRJW0?vLS}bR5%CPNM1^YyG2qyDdE4@NBDq_;q3K~Nxz@~kzmN5*Y+%ZI!Hf>
zKXi5Ihq9h{=fR$L0{g&*C+I=zwyv*Fj7mR*Qt9M4u?UcivYNf%3yN~dA$#|W&<o(9
zcAD6u@>_yBgNoO<*3=uO3=JI;yPNnIAcB0&ntH=WJnO_1rgydzvW2DwsYn7839D`y
zUti!XUaqnEW~Y5B|69Y6GjjhNl24-9X|ycQxM}V1=Nn&h90()+#PK#TLR?t2VTGCc
zPU!^Ga8m>M?ja0W-|qa@6N=v`jYe+|>xaMLpgr@#m`IM!+byI)HAnNC?p+V0*zB1=
zU01RT!)@EY^Y2d$a5c?iA16*h1z%;Lqq+e1P{?cNY=6QXo|elMNh}RR2@M(QpwlKk
zFK%Y$A_A?#jo;MfHy*_7z-(58za2amGTKkJA}tA-!`=8_@3ctcuG!wn#F1(I=o$>!
zH{UV%mNMsmL)L>0tajPz?5kCvuyrl~^R8$p3-v6BDhH7j0A{Ji`~Xwd6}lO^z9FUS
zP7YQ5B)p;3!uR_Vr*ToxI7N=ASeKjV=r++noKHq(bTEx|5xqz5`+!-U_QK#lm*JtT
z^yBYEPOME8+nRfsP;C>9UB=baLlD3XX|HDz>o(tZO1L=ZkeEee!hHbME*v4vEFyie
zKsz#0N^wn%xUeIhPi#LsjOL|`ySzGJ9?>U7kWYSW4uW<v#ITII?>nBPA3Rxz$hP7C
zx9Bm*RWkkWE~)AjucGlzcb_qUBWo!+K>EtI<N%f8b)TORaxt|26v{(P_n(6h@}L>6
z1;$$mdk1}PL%*Zs!B8K=AA2Gu1!Q_Z02pHQKFCgFk0iZ7Wf-{<%f+@=OH27;PU0Ts
zW11)iy~uw$ugCrBrK>#!Uy8{WzbjBbRd^owD)Mkx{+Ohq36C+^S}YMxX+t;K`Tb^c
zMDuaHt6`G9m1p!9IvX;|{%V(fvVUz&hO2t>(WqNASSQL;ijvO~8Lxzkl#UA>vpK}*
z``>3bzuW#h!O>@!s0weiQ~aQJOibSIlUXkwy;NQ;_^fH4U%y1*(T6AVmq>Hlb9$`K
z8{7|bGJUmycd6#%Y4uVfH_n<);eWo_0Oul<XPnll;m;=-hLfwW!!@u`5E2R)T01J{
z>^+IaoHrnGAG#)aqAF_~6lw*LJ|C3UgBZWYuzlp|m)W2ZL@IZI3aX2{J+9a0M$G>`
zjJF-A?g?&dhj<0nLE@N&4r(I|Wd7k9R8BN;ux`h^Jo)M0@vo<Qbi+q)fG|ruFR?1M
z#`uXP*%!En=clC>Amiio#Ul6HRzs(Fsef`{GY{x5TYwu;>CHH?W81jb3&kxAsXV)2
zcFtVz2ft>#Jbt=Fk?7e}R5*=VP2p;?V;)<O;=lTPgLVG--7;wo`(kJOOCZkN2&%&8
zc-dPGk2>t1oh;!vH!>h18gjXY_1q5KPY&6I+e#;batab_+b^PSMyK6xLV&BW;PnH}
zQ+yaXrfVmOT?rkP9nXz@ozn26NF8ESx;|MX%jujNj>yz=;fk2?F{(e4P@cr%Vx3mN
z6U6w(^BkD?m>3E8cKR`SWS}yqw)VfA^mg(apn!N?5@80D>nquBlLaFe{A)`($Am$8
zKiH|<J`H||@UUjk$2MC1+Z`8lQIm}a+^DZJa7ykMV=K!C-bXLd7aiUX6~k5K!b078
z>rI>QW!slk1`(X^4yafYNinJ?CifJGW+l~YyAPT1b{LYUru2<fJs!{P8txJJ5*(Dp
zSmd&!6PIZ8WLG*Eu>^W_WRFuOT@G3-2Hj1a%*Rvww}iPOd@r>dWg~HQuFAP$K!ZxW
zC>rGaiz??iMXoFl6-%z|s2dd&iP%xMo&?nTehA&Ms{qB{S=pd{LgQG~U)HmyN2+W`
zH5Bm*R6BCW#?<Ylj%K()kLN{miV#1(Rf8k6L#<9LLE6phsHFZ7V%|#1M5X(UBMtUo
zR2_SGk|gr_@B~RrHY-)qbxsxfi9(rILWOznuOkvM{~V(Gx0rsjEXM4qEZq~&a#ZXk
zM;e;<l56_3vLrDF(5$XcKS?0HQs2=zi^sX7&eEV3j-Ih>j81Eu8Ph8g$|@IVs=e}=
zX5mq6PgAY()E^=O@2p%B)c?%u_J~SvpNm<144INrc#n>hxj#nN_lT%jY5v{+Px}ay
zYovcJ_j#9GuX#*%es%J1PT7}+AD_Q@!*c&LaDIs5b50(N!zMHnHX*mS$r>5(T-l*F
zgbrRi>nvSJqE(~k1r;3`iHZ<6^@O<JvVy7O+%my8{I<G75=BUYAB&NA+KioE`I&Ks
zbjCMFSw|sd(T(&VWWO_{kx&Osa}!gPYiq^{;ty7JCfWK06!N{7oD-C$4Qb;4RTXYw
zh>-VPx9KpMZVde=t6lNuxRaw_zV>r#Qu#i6DdG#N9JgI4&~xBG<3TN|PqY`*-M5o)
zT_H;RKD;44P0#q6e1MRCN1>yII4AOe$LpTU&-KMVP*X5>u0=#pT{;odmY5ST+x0Pz
zO*jQMCRaQ{<R$oMWjPmbp1V9tqwreu1|Ltd$fpnV!g3_|^@+>tWXw!VdWe_&Wal8U
zinEW&g){Dgom|OXQf`&F5l3qZknASCd};|BC*zC6>*V#n9Q`r9mrZ&2=%sp>lz*QU
z3>}GWMBT_u)f^(RO2}v+%;#;nM4r2sSO~hF2@SKns#%h2n19s5i4SAZYmHgE518#p
zNxEKFzhuM@G3b;^tjd^Wz>K0>-&Y5R{DdY9e4;A&IS;eUq~_ln|KbJKXjVeY?`Xjk
zl`bBRKt~GqkTHiVkfV2A7<0-+uUbKEs;hYVlR#fyRB>-hzNAEor`&laSGL{!^CunM
z=8eq;{OVI9R1pRMT$eW4UBp#%fK%sOJQwwIywIDb&G3KW60Ee08R`)ntdtcQG`C9o
zfs!;=r53;Y@9!^GM0I!_b<^h+6&2%Fn#d}dD=qunaJfrkr=Q~SKci4w8(dw1xjK3X
z3Z3=nUkuKD%5o$w(e_0M@mF5T2}cf7awL^0rKh|KOCFNR-5+lP1f$1jK%Ry~G++a-
zzo*&yJd&!p&*eMGC;A3l(Mjx7$a}pUp%qAgaUdnI#Qv;1p|0?OB}rRZ<d$QHf(a!(
zRn#FFwVL6(iLf<d;7iz?2L;W^DdiFAKajCEnAxX1u2dkz9yPR+?2V4VB(wjcW>@Q?
zKIN6J{$xN0K-z@z>8v3Q76L=e6Fi-UH8zoyBokxTbLK}(4KokBogG<4Ru*~x0``Rc
zTQJv^w3wUu)^++VW0De2ocJZY&SrJ1?*52a1Q)`>XmL-#ak9@M@*YO!0~l1A=vV$l
zoE58+;5=Tt82A1jDVuVe$4l}z^e@Uze^ynROMM=z_+L!T^hFZwQ6bSk_@0T>v$7;)
z#iGI~eka+yDjiPvB->m@PAxn~I{$Wf^FpJXQGG;p1K^UOYku^3!>Nh)i~3r?y-`|a
zTXEwB%bi(_#(U3LW=G%UmZQz3M;V?JG}4jYMRDU%L34SE6#F0b<~I&p%x%uw_Dv05
z%gDKMJm3}iaZXbA%lm+R9lW>7*@svP#p=9y0p+mJ;xFmr5galX&az!Wmp8G6+K+u9
z)4ZokD&tjh_q6Hq>SMQx-22|a9r-l(@1J=6Qp$4{I^;Jm^ArEJ?*todZ^Kl))O3nw
zfg+C^U)f11PNXN+;J%RGxni(l#t}9doU7_x`Avh?Y1QHPi)*Bx*s{&>Z>78<rbT`v
z#pKPX&(2leTvh1s1Nn1mp2J^gqGi&>{Y&5i7AcF4X}JXx$zvx7ZAn@?-sf#dR;70Q
z(xaltiuP@K`bG7^t~bB=+|ZdMOSY$XU|&`0ew^E{=<&N!C}ljmBwMQDv@gbO;b?{A
zOyI}V@n@XXSAypQA790DR6LkWT)A3;y+3wN#gx8REbL;s!rv;fs1O#WEun5<?i&*v
zhtMk7rHLga`d0r^<G25ilP21(QOY28ye2#D^TF&`2ZYWoj6IzJn`w;|g6Efi{;IGk
z%b-E0%s=-`H7TSi<lIg_Od0Uw$$gZ<o<!~PF*3Vqnh?PGr~r+pHBn^z<;?tAzC}2X
z8zGbE`LKKbC|x>pz3Kj`$l~!QlSJY|u@{eC{>5%~9l)AnBYy4spslk%Lnh;Xq+!t%
za!Fo}u-xJzeXopCIr+N@G&wPx6}_}%a#2!b*Di~5^vFTE5F&NTK7q-Gt^`e|@z=45
zz8asbqC|(m_uBt~3$2>#HQq^>_cS9XMSe>k9mCThp)cHDqElnS3so&X;-4J7`ZvVq
zRnRBov9POrexwP9vXgFJFa&9J7><y!0H|VWKa{De*@`7K-0Lg}D*CK6bV^Z@1G#ZB
zWJ)akA2dB0%O&qQtgbZW#L1#;qv(mT!Jkm6croG7$ahy?%94~mI{aY<kuTlNW!VvE
zyWAnudf{v4(wU?}bc!8f5WK~6#wY@jkJ+T4EMppJ(5PkECYoT_pz%Ub4Qd7pDb8J;
zH!!|INg^A!AhZ1KzF(v#su-}kex-rv6C=l^+C#aqZ<nOR3O5cxOf@jaK4+(Ehv&UR
zv{*1NBaxAhg`>}NTL@5aDx;B2fTtMnIjwauQ_%p|y?`4yvIrA|dEce?9P8#7Rf8mr
z#<KXe{*o%672`e4Mew)4`U$GLM0Pb2mDGXdTZ1`9-21aIIyh~@B-3Eckbz|N*7AuV
z<1?L{7kYJ1lXpIcDe6sBc+3^elLGeeyc{PI`K-5E*n)oO)m=3eAN|#GT5ppW_=F<K
z#`L~cl{>)V;@V{!C^*rPEX-mv{fJa~(y#$Emk&R7lfu)$LoCXm=*&sc0?9oxiZFpS
z80YhmEhurIT|M53irK{;GTBD0O4TgI-|_h`V+L4Mk(eLYANUl@a^sU>;8UULAnTxT
zrneTNGjJ?o*5r_uq;#3cA?bJn++Vd*&BJQ6xX)H?9zEXKrw&cTVBx7js2+CO{3_t<
zZy{8V4J{c<7@(JmZP@>!hC=6+#C(IOf((<%sw)?ABkblG*Lh{9fJcg;p|tTZlvPg@
zpS_2_wZEb0ZqGW^;J4E67H1vNO(SsA*<{{8Xd+|@>1N)dbv-shJm0(BdRgXy=ntS8
zFFmNZ&5l;VA}TCk;OzkeW$-I%yDM5>mO5xAkJUp&<&>$fJnaL;uLlplcW~$;sevb4
zt0$5osszy}rZ9kO7<i0h`h^uwDh6Cb;luVEuB7l{ZOa+ax51y~ky<GzwVk|qlbXBq
zk!yJ6LuHBZRY^APl-6baf2D7w?~Dr}ax+PlhCkj^yz>i8_-m7e`FpBwH9ttU<bD~}
z{p{J<>mFqG^zL6^4WN`3ety@uz2FHos7ODIma4S)jtPBNFf@&LbtX6BE8%Dgk3UYA
zy!t}kbFK<g^@?UJiAm(7(FXwGo=2>mTmHIanyt37(nnKro!V`>@^|^Z_jC!#!;{B>
zhfNx;lC(!qf;SmNaXfu6TUy+3X1ng|gZzz00$o&~nOawjIJw!Lr|I$J29p_QRCCog
z!0;^EeB^+R<aF{Y)7!@m4&>lpSX+?Dlt#gg9vn_sb$8RSr>=d3xS;~r_e6v6zaLlr
zWt`}wVs?vX7JOTe13Ht_?A+cev;qAb(b2;Kx66wO%!I%?fN#tyV&nNNTB^QS!1U{1
zk$m?|S03L%2TB76zW}j>3jiJls!(gBDG_w;c@})tAmf007f)p7%+yk2zD$}4$H|}p
z%?YPVl5u=}HExIgHDz_c#tqbw6aVgEy`fDv6eSO>JF~WMV_N5|dj)$A$^arv*`yxn
z{jPjgj;*q6Lw)r10-6G(-gcoe6&i%Ve&c}=6`nH}rA1rOJ-<&<a>%}*E~r@68NB3y
z<54nTQ^F6E>%aIPE^bXVeaj(x0c(9pB-Mxvg)S*E(*x8bBSkfG`T~hVD@d1Kw)xsV
zn$>y7#UBv<mHY+!h5@MhHSN41z^;YLGSc`=*oWa{qm4V@vxHJr+g4otBWPI}U!U!W
zyp5eA8fU~%Kfk^W{rlrp`ZOSCo0|Acp0|+fL#6c*bVP=TSpK*HmC671RtX={@0JzB
zMSd#2EQnI`^fZti%;@ID<}iaQ5Je9UTnEY3)&Cy<AuVCZv;&M?^lvBQm>U5s1@^A_
zUujxKP8@9qw{#qd)FNyLuPbz^p2S>^i;7ga5d?%Y7J`VK@qCKFmOA!xx-(>5PEa#W
z$q{uo<U(5b@+i@zI{pvSCm;gJVm{sTh%+pg)MxRJUn`uu<-p)+Ky}G+W=0&vcILS_
zSH=CORFGQx8;CTnfBOKgdZILdv9<v?SWBQog*6vvOp0OdniQypz(2N+j@wA1-~LDv
zb^KDU6dC@0&j6qe5Sen#;d9kl#3=2Y#d-|2rIQlNssFkL2)(Aqa|BgYLYmcUa2u|w
zn!ibjqP8eFH?#qeTlGUNV=|b0M@KKkjaE@4@)ou&ZWu@>?@2cbSI|W7gabK7`LgqY
zc5a#>3>insKan^wUe%|%j0b&A94XO`L-m<gN)XKz-V6Hu8*3vN9*W5M6c2!)UddKJ
zKD{`nMEM`$3kM%j{|;}si3&p6OwjT2A>Z>Qe4`w~dD>2*E%de>_{m4o`8*a0h}ROR
z^k*y{%C;|X6eu_hb|Go<?*5D#D4^gx_zZ5GDm?2B{@*YBmApSIqN8qekrwTKEZ?YK
z?jHisY)qrEHbYC|5m><|bzuuYdS2U}+t4N$epL~mMailWg0(AuPp3Qq&ROoqoKb1G
zdv`V3L?sDAi4>^ddjgm@4+ZfZ$OXyyy-LOinMzE<LrnFYK~40zZ6+WOL80+>{AF84
z;i|iVx6t}Yg+i0CeB=Pd0J2rUrc2UaV(rAUOKQ14_kWx2dj#_`Be1b2&*vIql`7`C
zLT*&3P}_^*K)6IYI-fMC+6CubOjF|w%?vTf37(_tKnx}i%j|F(c@o(*2gn_%BLm(+
z90t26aY^uOx>K>@HFqjy`Ty4%MiS#*mU*tok!XJ)C8u?J-_u<kZWgc21y%lz2(tZg
zHR7Y|8M^5+H5N+J8owjv!|G^Qi)cE8Ui=Z8MOiZ9He$_<_?Z$dr%RPjkzOLlD2Aeh
z{w_gxYC30!6Z2kd0mfo|H2sv30odwXCvyBZrtzgf$T?m+6S!o)$uH^MD2%Z7+ka_c
zw8R3)dC64;NO9`dtn=u;KmEHhW+$A`hSvAU@`ydab<ruS>M@Zx#G4JBKgnvw!7M`l
z1BLcA80#Io_utQnWF@A<Nq8+#?dfkYY*U+?MHhXay)x<s`H?JXt8>*xTN^k<y#APM
zEo-uRnIIB4#e~ZO{;>3r9V~Y~RkJeufRWq_J$E_{1a=?}eCC?}E4JaDR)C0JMWu(m
zP+5kDTp$}FZ1O_sPcC?o^sS<i*kPiOk+2)(!D;-8JZbEBHlsopiXb662<hSR+AMp_
zxvcc>EzgWPR2Bhh%Xqj^U3{f{mL#(#H^Qrnu)?71B)ne%ihka@Q~eNTDfho2wa9|I
zqt$H&`zWjcgF1>sbUg)Li)Tg|{;0>>Ct|58_E`>3d3+w))ZjQ4YYYF~`oQv_M={4{
z`ltWpJBLZ9XN`$6j4&aT*~TuJ3^(2*8u6K_Ozkv1o{y13YQ7BQ!S@R1KI*OTzxb#d
zTM(|$NL50~9z+8<e+|y7KERlb9*wJPKaD(HW@10+@%j<_JC|R+d+=&M9hV4~G21aK
zY*OVgS2OGCvvB7+@T8q$ZA1rb&A9{Rrisfr`O_w5^}n}LPCzu+mc?<oQ3ch{H{l%c
z{N$ba(2ql~emFv#UC0JPkSo?oQmIMLZ$N=YM4lHCMo3I?`uS^B)}7+w_u+(pGgti5
z2Sfz+kE>f~P^(xWQ1!LHS3695(oDx|N~=`HH9b;IU?JRef;KJZ8%XCa3%$fy|96@m
zB_MndcN~X-J^Z9*6o99Jw_pc2A;hRNint_{qE%kOCrEgy8f!Y(>qO;!mCoDU!?G|%
z4_w$wkU2#h7a?<El&eQ=8N{CPbKfDvaK}WgC|d-69#w2FU@<yw{N|&nwxh%-En@X}
z5!3tZFJx_l^~XYO1ym>>tZ8yd!w@VW0sVxZeuQ-MSxCqdv^@q#-#cLZfsxofb73yx
zhCCBL1W)yl=+l0uCmGAdk?#N(WkkZW+Xah&_%f?%PJ}WH2E^3L-9GSJX#4Lm67Q}g
zwEmtCY<)<CGVcm*3fhf-tlnt}xh-Ike@fNsqC(2F#nlZXR5N+;*MRS6KI>P?F$ZkN
zJ@t`OZNdZ|{TJ*Ljh+NQ;4Fy(3=))zF+C~`S3(qJj__Hw^7&9tsE*nQeEILFqe9WD
z5RA!IRaabZ!|D2%^#P4%YPeoTO!)=mQP<z$y1ENv++@bER>9q9Ah(!)p&>&h=_Y1B
zm`HHr=Mag(+2#j6a3(lJJ|d)xYG6pNKZk2Ne|7MBF070rK_rl76OiHY;lar3e=qfp
ziOHUd%nIlwFz(HDI!(O*ej0p#pDQ}d;`g7}V(U+#_3J)CUb^Aj{E3bFGa8c0U_pl5
z-C7{yEjyJqO`&*kfHv0~Pn(+2O$73{U(@vXo$t(S2v0y5tJrVR(5q1x{GV*hP#N*=
z7w{|}fl=4{a=k+n-s_wiA+5yNQXcs9-(_);K`hg?*B-KKWqB>N&ooe4Z*We}V{L|1
zCOZj_$dyZ~X$v*Js1i+=g+U&##!MPm>=w4I&rf#X!>SHcGr25(17Q&gjdtv<!Q?)F
zsAs%3O{K~Bh%&}gUAt`WKvE$@fs%DGn4Rv=8!Z;cHUb&dZI$5_H<WZy3o`_{O?sxb
zWW}f&r`2k#gp|FOJCOjA#u7shellcR2jwR9$4^wV^!j{!{L1DtlyV-~q1R3DCH1(F
z-L=?&lv~%&TQTl~QZe&zgtP$Mz@UkFa`W3=m?b1C^ZF?s55Dz<UPu1lQ1f1qZ<5n$
z*6WZ3m9691F9IvG<+ov`cs(=(Tj0&Io6%SEpBFOwI+x0s#ReFC*<EK4Epa)K^|=E3
z^t}(1rSw+oArlY3;GrF1d<kPgVf!p|qKM(ET#|pwF;CRFRQu0sD&|kg|GN#a8JQ)|
z3+CzhG5``!4!cKGZq$xAayr0bSKvdTwREG|;s+oJ<Mtz9`i$eBxu|UP;TWuGJ^7%d
z7*IsieuM?G?98WM(hGh+$OLPGZU;04mniXkc&9r<GuH7B8#>m{{WyJY7flQ`1f>d8
z2&}E><_}QW4s5g43i3`Z03v<`k}i?I>qN8u2d4%n5h5Mfx=2AJ6byn>NbZP;<fn6|
z0wwOsw1~@u{+g~N{}9oUDC~EOcGH$hpDGLx8(uCGzSPDoLwZj9e4GCyBrW8nxCn55
zNC?IT{Q&b(;;aow`>_~id5B~PlvS6Fq;*+ra#RC~GsNKQ>*iC>2EZa6rpLF-0|W2|
zln4Pl2S0$-_cmqbOHFin^gE99+Cs?4y1FH^xGcyEiEtq?PF9^OD)SF=6AtCt`WE9-
zosxPy4)&z;d9uQ2@|*nzOi(0+S8jaX;Cj))T*84Aa<1aeu{Qh*uN<x==J5z?Cc4$C
z+3LFkVb)Ky<Xf4EClb*D9rqSgZ-b7>a#7D?Y|kXm0y`Oev&C&DT)PaS#MdVa-(;<n
z(-gB%MMt8)?x|0znBOtSBVwL`|J*bC$Ilo4la!prP`MK5b|4$~DO|GFtEwlw`THH8
zwZ>@*9RGY%VhtMm8QMGbvVG7o>@Xo|IJ9?iRZF*y{(i#`>$U`aS%lNy55h~2r+q(~
zQN<)9Mx6t<=EBrYKAVOJtjOm0@=_wx*;@w9fnU^13bLOn_?UrjYCUK8kZ^@?8?FC)
zBeO@St!(}GR1{+{BU{6k`C_@$iHC(Ov67<nf#|e?3%esL`%7g*a(AI}T4(YZ*%In%
z{j`&D!hh}*&3iumbaS#}>ds`xFR9Qk-A_Ly|9R*7rTf*BdDr^ppEj?aG<g2yt=Rhf
zg_qL`3Y$f!;P7j&f!4%G@(JUHvCK%U9X|>GNw!&0TpVp*cz!)kTF1hk=WeJ#SR0b6
zHQkH?ltv(?lIoM`rsJTgaFn_y?FHfJbD}hdyi*C(LQShm)Wtb3Hy7C5uI-!*7HkB<
zG@T$!Jnd?l#jf@8q$8+$clSZ%mMiIxy(*}S4-Dq}x%Kk&-wyy4K*Z=)`DX=`1DTKy
zQ3Kty8EdrvlP-Iz=gyxp?DMy9Ir%>{E<?u<&}Qnr<`j0y6m^;W8PjyB3EU^eUybHC
zP=V?9PQ=MGaeR3uPcF&?sZ^SUUI_BWoKaR;04}@D>os+~CV*&qvfouQ)Ow1F@gkpH
z+AvQz`Z)Zsy*27JSWkRdEU8G-DqQ~A2aoIuoY!8q-bvMQFOZ>5#(CxK0k5w1uEG(c
z9;>h|-XJT9PJ)Yr%c|W^vtx4=LDM_Nr<rQc6E6*lvt4&%D_`5DY3TE)^Q&kvR3w-k
z3_M;HG5S5ppWI2B`!lcb`k`e~UsT#SxcDZa5<TlG$nq{ar}Xew#L-;%6Mfb#c9YcE
z>FP@`aY|-m(Xe}b@G~9nBrd0Cocw+=6x8hW`Sv{DNu_y_K(u}GPhmTip%&RbWQ-aX
zt&U5Jq>mbhmgx}oPp6Gh_nnR;W2QKGc)El}f|ry>M@k5uHhB=S?ISculuZNa-3>1o
zwen@x%rwYH?XHKSbrt3(yY|^?$243sy$ys{RK<@@3x3-<b=dHoDi{vN0ly!GYX;pa
zwqH0ES}0|;Udt`=8o-gO89WZ4QYgEE__*^C?r8|fr(B!$Gue@3=pAOx7b3~|M~BT$
z(fB`rb=W(#!G2p+dGaHHl@b#o4)!NSokAlRRKzm&xcO#gdc%oO^fF6~q9{wnBW~)k
z&(K3UjhIX3(;1lOGWY6_F0$Ehj!HT^DIxOrCI+f}ZrowFbU+S>5B<r1&}A!18e5xw
zz*mTeuzGM8*ejdRH=v7B4sPX7?cxuftwfmIwwa1~p7E<GDW?WzJIa1-|FB&r_V)(J
zMl17P(7%4Inx=_5)*|?~g3Y9zW;gGW7gif*7QP%P^6i7KaLKeJ)K+;H3%!#}RqKpe
zOQg3OpH&$^Bm!}2m(I&HcPoa6WKpdnIOegj8%PlFdv3m{Fvp?&WYPF37A>e<)Gadz
zIW0J{<;We{&sVw@hJO>HsqWQyiG-<pr^?QF`EZZY2O+C;8w2NB?=<|1fFVZq11jx_
zpoD5K>ab*bbS?aYLgXZ@-?rY9!o?&fPTmHLA*?Qshfaflvnb<z(!c{&<V~3N<qtDr
zRI^|O!*Zk|!84HUOiJDP9iY8wJz-=Xt58YiH2UP~>`VP)ER``bjxM`2iyy{YovdIQ
zTgvo;MfPoth-!W^JB6jvDgl9ZgC~`PmO<Pn)WWsFdb3Dq;u&<kx2x21cpQu$yU(se
z4Aal9aB2mViUKNJ(+&h!d<ajq?2=Yr{m<;)cmsV)(;wuNB;FTj3R;bIRG39j5Htx%
zc0_eM6itRta15ck%Wzo^wZ<$MhNP~6d@K}JlL0m6abDGJu2{by4i_t%-MMb!`NfRv
zpk|IaD|2$~R;~aQNiAC4(#hn8IwK}Vk1*&|86@lCr$cU5CwZyhb*u?(aUI^d8d!Vf
znKB$djpk~wHlc8wfzOPnX+_QxH8~;KQ3rYOHPTwSzOgkIocLV`>$Q;Tm3u}fr2UHO
zaK(ac@G(>~_t`*v*$Il+3{}%7>@ukJsbamg80e;*A4q0mk=PW2|K@W*lGpWlb64jU
zT4w=d15FCPbtJHSOdlnzgvcZ>4LWdQyXovf686TY^9%4_X{{sSm5oX$<7n+|$nXV?
z_jaD5b}qo}9ayH$u2OD4c3$T+9i)4EslA3M*fla#7HGtLmKfPFT9i=nj8!i1@M<8p
zttDap01RlmAs6Ii3&XEKeO$EG{X`q3g73ZlU)n^dySYI508N+~@&)~%ut6O0j15*G
z$Stm&T_w3W9_JJ{vUI6*huQ%{%{~)@VO=VGr^V`Sp3liX36PJ_l~7n!X-|F^$8+cv
zB_PW{P7UT-+1(M=e;V`n^JI27vl0aoRwkMAT(k{|iEOzl%=}2NL<vXyPy#4H9|X4f
z>`IWB{w8DSrKG!BAJu}p>Ma?|kHk`mA`fx~>R-T@zZ|;Vbo}?vhfJzpNcP`_i0@Gl
z{>bV;w&Jl5!kyxP7aDULAZe1>pH}{BBR2Ytp5(VGa3T8G@(h;sI78`jm5FEZ22{2;
zo{L=~FDb(*poCQrc$~aV1!R^5MD1Q{6{l&3TL<2@2ip<IPo;KV{W<Crf0pTmU8W;x
zy?+<Np}KgiKe`kno=GJ3gzT+vGZcEp9)R5`EM2ML+7|LX#?FX`V-LmlWF_i6h)mQF
zA0UU}9Vm{B2k1;@K*4%vSOq!EP8c)?P4|Z)i#EJZVF*d0T)zBMdQ!(~E*x=}JkqoS
zQ{oF!*UWaEdtzP-jq0gb^%nQpN>B=CirM9tk6Y8b$6t3X=T%ZCkev@xQZoBKZOSAj
zubB%MrAr=Sh6^oJ;K2<gnm;%gjJma6z0pI^DS$&Ui_cu?Vf(-1KOq%<{)1$lzwiL`
zES^=QUe~mP)2_YrF+ez9GePAali#57zawIb4CHwXMC5mo@F`6l;Y()RV3kW&`AvJG
zoPW3VO;vyvS)Ex^ICnuszee2zRXgNof;Y_i(pK*1G5#1*f{5WxwWeQyb<wUyyTk&G
zXRFz^;Q*~*6ut>A9j3)R^3e2xrSpIL)vljIf^?*k`}wL9UphD_JFA=$p4)Q&Z*RNP
z^ftyciapxbK40RqpG9cFz1zLEDNYU3sK@(GUAC&!W;%Z_U7LwjzunKiD+~s&t=}Tk
zUWDYExX0)IST@B5*3caM{_*)NlW^q4M^+5lOn8?tgnvx1(S(v>6OQT)<G0p@fQo}8
zgo{JY!#}!S&;h|Gvv4X{DZL}1l^0lh+xzNecI(Ki=bS4dN{IJo{+(=?LZ?nUEka&M
zP`=A&at6<dp_eFr7uEz&Z9&4gL}EiAgew%^n3QrU=3^i=GJVam1A5{=lUtH9PR4l6
zG!$_p5Vpiqt(ndLa9?=-gpg0*ugwoewA7ZUa1}zff)dSqjR~)MrlGYGPGPZ3U#yKM
zpO(ZuKCPd2Ab!ET3yuTSrJ#!I2O>p~N(oz!IF%i6!ZDu>ch~W|<W)^CdaNaGsh7Yv
zVzBZHi$+$tJ*5eT3qnr{|1krkbcfy8G<i;>8MAJl=w^;%lD#+c1J4WrBz9R1Daor{
zAUTtqrsaYWLY4Cy>QRNcA;~Rhb=b7ulPMp7y!~U-5OJm*94fk|Di3TrgS<g+y_dr>
z<INywGm(HzKHT_8fe8ksVXN=^qSA>Q&$H@7Zgi_MEa=(cp&0YO$A-KGH|T1Sp@R@Q
zgUcZNz!@rO3e`a4)35}i6#$WNGFzLT4!aB&<=U`m`a~`}^P94X-N!OF_k~dzI+m)&
zC%g7P+%koK2^TmOdNi{=%BwQzYwQBM`{~N_BAN)nQmFr2TVmRCecyF42@ZmDk>{&9
zJN|b#V(cSCr#M8+gEG1u`5wF_ktK1Vnz4#lApkJI1tF;}W@rujQiXW}_4C{N%ZPst
z?(zLg(I&qnQ884C8J<Y$VW|h@R8#;E{e;vw&CBYK8}<uAzjZUPW42gYHM`+vK<5{F
zaF8_&Gkt-V4_RDl;hq6wyhwa7Sw|VYw#{9o2Vju?u)$!~#&*ecS7J+33;3u}PB5#n
zr7+#XwmN_?81Y12%PF0)h$<M8qQrfcHZ*2<jEN_PO0PAttEKt=l|Eibu>a~aX+Fu1
z&l)%K$J$ASfz-1OxQoTag<Y4is$i2EWBXkQSM12p++6Cwi@c&Kt^xpp`)$tJATjY4
z5aUFn^FKA=3{c+{u+F-P^m5yUOPtyzC&?ey%qO{LU7rZjX=hCJ8TqZ+8=)i$nbMbX
zibm)}X>nsdA07eh30NzWm<R157HR3EhRAdf2_w3q8lO=%n=6fS5cd=|g)MBH@-`mC
zyu&h<X>AK3xi<RSUJM!^nMhi2!hE|4Jr_wu{^~k9xsrqlRWKRSkyHZEtfI(cRGAPy
z=ky?7=*ImL;D1UbwqI17KIha$+iVmNZg>nSHo$z9+V;*9^_<x?T%Zcoo$(;f<i1)%
z$78JUeAqW+Ypxowf^%1I<Ab3QD(89hpyCn7DEu3GOM1%Fan~=#<NiBxmTi$Fj0?_|
zYwy2uj>3koJ;UVV6j_HC*3q-fkdjMP{z^3Lgp-j!?+*P%&uuuM@hU9-Kl9B*%t225
z_Z{snIBc6F0-9o4nk~9-1;SC3B)wGyRW`f_?k~6T0gm(NW$8tjMFtS-qqlTfDqa<J
zFJtQZ+FDEC+?-`vFqY9&dQFcbnYl(hk9TCk5M52<ts~zG^>^b5=yvTAU_gStUJZiK
z9~!D`szrz$iyIH5yc4ux3rSG65PrTovV1ej&;_m>{k|U)C%Y-K&|KtX1_STReDPi1
zqt0OO^^CKTB2Hj$(f;uQ3*s6yIq=}PJ^H#gVEr=_NHN0;DE`(g60`86`V7PW+Y3mc
z#_?9JiHl}|3?!Hs<r0N3WMo~^jaB?Hoau5^bd%Db>)KsK3e_)s1l|dIV$uydvWWnH
z`w=`a1|S%`P<a`zcENuDh87ajoC~_(%5*h(3n^WZKgn??{;&@&vYN5iM8;yLcX~N_
z3~Zli5F>slmc9HAL9{S<(FrM3jdmTcnI|qN!gz@Osi4>N)uHfMVtW6nmFSl$dBy+y
z$NqQE84`}F=#CH9`|t%b05tvrCgvDKbC>Y6m)o&0&Sn#^1NtI;kYw+;KPaUCW&h2j
zyGdnZ_Q4e3I0HeUg5CRR#yA@>mDTH@K8-rzxat`ozJ_A7PuQXod{uB1^D&lwSGQ~n
zKKdY|7k=&n34g;5iP^~%Z-b>Q6X;nLg2`xkP5Bz~5h_3$0?%WAcBUW%rWOXP+yCn?
z(u^Gcp`T_)o7g(C9tS?ZlMhCK5TF&bS#pt57I*_R)goX|Da!gkXsGP;3cz^d0N^qh
zg-mp_n|@(yY0WNu3<!GxRSmoyJ~l$LRP8M(Z$;&$GLW5Ha0TsRM_ai<OrIFSF-D7;
z9wFw4`PXjY>ub=d3s+;gwl|uu-{KpjlbMJGAm<A}Mt3xqVK2Gfm#dEB^X-9_!i|@A
z{LKjO_hZ=LEXq;K8bm0LQ=g8-1~FgU9bR{x_~E_0NWh&WKNZ{{`p$@A?MxPgV`mG_
z-_kUe_?qmvNgO^xi&rqh;ML}YuZ|IqeGKOh{J2Fn1M){^j~Fb;sZZ@2Vka9v-BI9u
z;NcvHIPQm9=E@B+<R?d}y!IT#1Q+!+eOKbgG25ZwRrY`o+)-{!JX^|y|Gy8D5^Eqn
z3MEa#wcDm-b@j5Att%1pkDRss<xM!H^9v|#vu8i2HZk47qxo!m3YLHM)n<ESNFn!g
z;W)c^W8w;=mf8S+d7WQfG5$KP{Pcf^EHzS+lE@DbwXM&Lg0AW?XA*n1jSOP7e2|8?
zG$O^6v1oL;-+G+sxqA|Ef|%TT;X`RElAC0rP=SLD>+5>;5>k5^%syv($WS~E_!Sln
zIY)#GEs2l5i4O7IT|jPoDo&99<aSBU0K=}-rvKR(%Z%B{ll~9egbZ47^x|<UyHN@-
zaE(%WgZaYkE^eC&3+3emkx5C2Vz0+=_AtXSe(9>|c0bj>fDyxF&>OQoa0&|Fgy{=~
z;rG;d-hBdmH+<MUxTY=-?-Gvq!w3xhci?P4S(gDCH_E``ZLnJxP=Q|1AaIUA`X#_7
zO1;7|>>0)(oI>C^1uJ7CK{RC7>8Jjr*fT`JmlyOO7ssB!x?+3}PS$aoMJWC$aQ93`
z3&^N6q(7y%7OY?f=_%qW92QSJ^*8QrZPD4fegS&XYW(|h87t;S$YlYvfT#)v=h27f
z-%^w*{aK8hF%C#cH1uPB(4T=6REsP^cE+&3A|BVvgIu`vA!sa%ma{Ox5V8!-@813K
z@>+eGe*an_LbI1Rh~h6f(0%i8u(F9{v~p?aL%ZX}Nl1%f&Hzl)abTF~WspBRzW6)}
zt_(eiN8kC*T2(eh?bIE3(M?$Vg_k!A`3F!kIoD2jWzRwyjTZyc*Ilze7sWh`!5j$8
zm@v476BN~RA)YPzPqyQH($MT-2_?tSA)vkeZ}pFB<w*FZ<tt}WkwW)(Y&|s`Fn$rl
zp&TKX9bejmDCIr=A6~8|5}fyTp}SB>b)<IhHijcubifBE27n;k5`}OyjgHe2n%<t$
z+8Aw%ZlDd@M-y*Yz@=p!!GQjSi4yEgxi(jyjH{s*rX<-VNKV(P?yM6H9$*Y=)o<;J
z9}rBnC6;G4V4?!Q%EuAb<St-e_KLiN@Ykm&<YfvZe3sNRt~}998@vOF;{c%bPqjc<
zxsVRn+-KNd&lHmT-bXbOIZBm4=d19?y`Dv#A*xJ3?!;~`W9@Px?3|cKIp;|U8|be#
z-_~}*QtrYMeS0uId51`RLy)ErU%y}v6sTMYa44ykT{Eo4TiqnvsqAYYg;L-3JGF_O
zX6`9wAA{ozp1OD(a|$ZPJ-PWx4b~yE5NtaOO#T-}@@fpJdesP03t)`^?&19D$bGoC
z=NXC=Q44RL1l9^3tvg74%ILp+DUhFIsqS4t;JkU=8Zie7uz4wKUvFs<Je$Po_lDk1
z*6s$p(`XtnD;1B0|KNls<eS8liyih=7CtoUNx4W$A_waG!OCqTyVLhA;fKw;?O2_m
zDbBBhmpFtwB_Ue=B8k4MpjCUy`|#fh&)9lTJwAIIX{8AJ7^y)}(}Un-^Vzt%ivsTL
z;!mLr*nS%3>9NLr@lO(h(rw#92$UiFc!E1RS&I5|q5G2C6#slT|LX|Vw8&`07K;~&
zw%H%bi_&lsM=$r?Ehve4F#4i)li1XO<(e2|4D6MB(y;#zJ~PVM-#ydxBrLgz(9Nc?
zb?bK2uHj%^&e=FRubANP4h%fDr&4Do<dv=6VD&vH3_53ldpG0HexFtrKhntP@vckC
z|DO4)L>>mbfZ<|``yM0%6kH$UuF8Vcyq~<jRks~Ho$7n30WVNM^{WzI#&GE54Ol!V
z`<*;@9DPPBWy|xO!E6x&)u(}PqR<nDp4L3>q6W$8saD*!orAE_jf7BS?R=f}Co#{r
zzGp}*LV(nP`@i7(HeS6Xs+V){Ty=yXdW%I#u(Z@sGnom~;0meTXPq94?cJ<3XR3XD
z#t*RfoO}7Q#P~M{W03%w*IVEhjpsq%F9RioD*NKUK-dd_$TOp&RM}+C$Cu3U$u3iR
zJqQNqPK8+pWfrv`Y8J6g*!AT$!#=3z+4ZewWzTNKbEQQkJ%cuy$3H&C_brT|G5QAp
z-Dyu&_xSLc^NlLO(DVx_GpFzy+$9!nc5}RpkAuM!yF+#ktnf8=myc_DT#kJnGibH}
zkg@0SwJJ_~vV1xo?N4W!l5C7qQ5Ta()o2v22!2m9imPO@yi?&6(<?G18<A)kC43Fv
z3U}gFETJ<uxjeH9o5BNqogsVWXHaVo!Zfi&?0qc(+nyu1FOtdJ>01_~s;{bkz6p5{
z_G}64$o$Lqs(&6=c@o;BW+&untW$Cva^4g`rQMW!O5rKd6b^~<IW?I4Sc7;m3QDLu
z&RN=41ZxrW&HLv4L*uVi;NuClXez;P#DUA1Kfw&!r?XO-?_+2cZzievodLeh(My}O
zd1oZlNZfq^&!gOQ=Rr>3fMeh{oZEeTBaCvJQ^B|c43Zd`sP3tSivGJ{5j-sNSAx9Z
z1l<d_=quQ&$i?eA?WQF?mv6fU89jozW}&`tWRum=b6>(QVuM~Df;0W1#E8XOru3vR
zNp;*K3ai0bLLo|5as}0*<W)-L{A6;RGwKU)K#fLT8g$vc(8Iqz;QLxF#Chq$`dIB(
zEJ1Sv%DLpl-285en7euj=Vj+2M{c5klvN`*DP_uoqy?OoDK#T+=#K!Ah~@rJPf?WN
z<AOidK0e&(QTHPrcUSg?*Phdagd4jBPaHSJv#p=wdBrSNyan(bGct^Y@M9X_yMELB
zRIycD#vQr^c}_;lR25VHh2cncvLI3f7UTW@hY`%y7PjCv{b#L#Ce8?@M=j{*J{P>d
zbpxz&E8i4$-c|+Nd0dRhXUAJ#A7x)asBx2g$(Bu_roNw@*`{F1|5DR^y_lzI;d9d=
z=aq)n@*HK6obla*uef3TpWGc!7<{981PY-pI)w_TIL)+5fP4z3g522)*((sPOpIpU
zp42!B<-icLz)Zc!=(7>xDF0dIA}+m_l$!+oYx1Ba0<RA~HhwwoL~;9r3^yr~0Gg|D
zo}cNpi~PXtw(r#JU1<ny4e{V!s?pNlgA?Q%e|+J&uN5e@Z;;Jk>~mIbKjr;R{@3uY
z;I8r6!Uf;0i6u{PP33(0nX9YxZyf6v$h4&Mlx?}FK*tbfMFxd!C19em5iubi0>s>R
zJgkxrztjvH=+3?)U1_;D_4r~1{}^2=V7HhZnlQhU2RZWitEr3MpN{v0kS><NpZilC
z?}Rgbek2l>Yx@g?g_7V?5f^m5iF^<P6g^F`nk=$_R{m0*dGk>0PT;iSDNN2&Xh@9b
zlLN^TCLqU@n3ZwD=%R@V91NhpV5~iCJWHI=Vf5I;1Z&|`M*EQ)Swtln!*3zSE%`+W
zlhfws5PPV*+frM19*W9G3IzUI1_7%PT?uop@wZq;3P>r)f{}&VB_T#Dy>E|nd@}ke
z#A&1~^ErP&HvakKeB)H=`el%21a3LI97}c!@5uv5S-8ItQH;O)JN6Y3vjDOZD?25L
z9q0z{8t`S<b5lmTQtb3vPUM1!f4^TnEs1o<Ks$r4kxkXIDkVny5CE@8Bl8QRYQtZt
z7oa4?s9jj`De4<fWJ%g+Y}ckn3kN?QcTDV;s`-S;^MMv^Ckt`^H&n{cQySWF?^^F^
z!vDjorz^n0Iol-?{RSO8U};UsS+7fUeSFU(f(-h>Qbl||8WXygp-tYJE<NF&_jXJ~
z=+Nu!lOJXA0&(?}92*<A3eKk1zn<`fAkSD6T+<;+hSZfeD6wO!rWSPSIA%Z5BH|d0
z29qHgZy-KY*gwB%`_?pk$?}nt2L8GN%`b3QX_{p4%2}uIgiTb(p*{OQx9^n_`{^|p
zE;QsLV8RmEO{}qF6n0V2y=5N_J`rZsn(pTzZp>T`1<A8F<0XIe4!@OcPCu_0uPgIK
zw2~<}G8ds5BAD%+lXG{kuP-_{5o%_*>q6c&pE0D=1Ib){+dKRs6n4xeJf{Qu%pAe>
zYDNvwEXn+*-?X^Hgy)hBq?DuHKvCw#GL%yy;=Bsa14dY)ofNOru>g9{EZlI#twR3(
zogs%Le)E}QUbD>4Rhgy4Z_0=~sPF7)J!jYYf}XV9-$V?97AAq%9Op7Qq3Ile9#-f^
zFPtv9i~sw3r6%Gp?woVe6)8%^iM2yJhKQ&th)9$)p#vf!aS3_*^bIMc+>#%?Z|i`n
zC26RRH+hzZ)WT+{<n$Lpal;k6U?(Zqcd>C$5)dTo41M!^1qhY}!yi)5WtyNUD*#YX
z_v*kHZZp~60nY9Mm5anj3DAN1@+RmXgRZm{JX)=|1<Gv3%xg8q)zCIeS-aq$$j)s0
zvTAS<pRp8mFbr^jX*8}PkCz2m?4W^2`|$F+?($RdyGOw;-|Af6gBHt+EKFktitdH=
z1QVa?IBagy@l;W}*#OnG80+g5JaLgUzj72TsAN7Qi0qsTg@HW{-T0ptN{Y`5q|&Ky
zAqI4V{93Wt6mgPwo7sZ4ML)mzyp0)(s{`zhUhg9-@q09(Ky`mUf5U;i@6>`udWrnO
zt7frF0z@HL8`C>$K6X!YXmdTQOj2O2NTXud%X@RldRK1ELqNwRoYk|vhfuGUFF_^+
znUBBjTd2SPaz*)lviHv)5d=d$8>X<ZSvm&01E-k-tr1Wb`zpcmxWbpO+EJicP)lIZ
zez)Vg|6|c&X_5O1BMO8Cp->B~-u>(y2qBT_kV=r##D4v$yf4mh6TU9|$BkkBdErgI
zgTz?B8I;T+!$UQBzr2G!F-rJYX_}kkuQ#>evP2^X;JgC}1ysD-<bvJ?7er5`5!K`P
z<J*)MPc!$o+x=5Q$X_y+dh7h+yBKS@n1I)StkQ|J1Bb$U2><C^Lu%K^-WFsUTiCar
zKe~Mn(w!>CWm$wkmW7?ZFCiiQi*MWy3a%2CvS}hz3*VEOgF%{g(*)nVM9!m2tm~Gr
z)15bcGmcR~z-9}DIPhwk8QZ)f3krd16Gk(<tPg*bmE3|eSFR6-e{IM?Fj0b>0;3*&
zQ3>rLSti`sqN_>ANN-GMY_t3#m|AlfPqYyUpxB2quho$`G|(8zfXIMK;Y-X7b6xwI
z3LZnwrYt-{kzMGS75r9!XL*gQmSYQod8w*K00)U#uwfLM1oI}`%ta&j0sDsYS_pOb
z=)0Qn&g(i#O@IvOb&9XPt%@rsd)<eiqu!O|zACN1guN`5r9u}rjp-fc2?*4!=y$tk
z0UrP<1o(e@eS&$t%G)8y_Y8QXHW!!hWFtsk0aiX%^w{}d8PrBth4K^f4bqjyYM}xg
z+tO1EKT&4wex=uHux`txqaMwNjZ(}_>5R`HyTtn}mj;p8%OQ@x@RQWMojB=I+?#EC
zchra9T;L{%Tr$+%#;dM}hzv8;LxE(l*>k|-ohq28p#&U!jDlzw@f$UmH?M<5KSp*2
zj&M)}Q(GK>tTRFTRTJ7N2GI}#$|>_G;j<>KVEg6*$#_-Gwuy5{DW|F2BDOBL%#_eg
zJ(>}^U>}Xy(%QP_YpGZfub_!)g{14vs!{II`|&IaJIoW}=IYOd&zRovlWX2;O1lXp
z-^aqz>u7Ws9`NcssUQWO4{aWe$mRMGFOsO0sy5+r3+b7<rpCcmuB{<r;M%7W0BY4L
zK6|F2cJx?y?9&*c*B-&OzivDWT=d5P%}rS7W%8C^Dg)vLbR&AhNl!Cgy0WI&^z~q;
zGF6Ym$9_lAq&(M^V}VBxGeSk8^ukUmt#l$2?|lwcsavSYeA*@reqqkgmynKouY7}4
zMztvE-|>2<-b=)$cCiRS24U;jUzOWb*-rp0g7SF}me@`fiI(z!^bgZ0V=dR{Pzay8
z<WaM9z5CKDo_{_8d7fWhv=a_Gi4hndRNtmi6ESF}*|ZFwnZasmVhs<i4qgPHYEY=N
zy$lucmy@CHI-YVpI`|GsxMxydp%5NZF&{I6bw`78vdkc>asM>B&JU%+o~3X1pwClx
z*R-FBL_j@|U;HY@qMTUrLPer3<9gONqjg02fM-zvSx3=$)h=5*EGw#;jhN)(1{*Nj
z)%Ne@d-M?|__Av%>|36aWJE_F9PcZ2p-UDtjqZSU`0H`c;NVSlphbL4`NU{G5~Cpo
zvXVj-f6{;uYusjh9cHSu{r&T+nTI&G%-NPY6AtT?OC+!9D~6x{=xlu61(l9D1@q)S
z8E3)rt>aNAx1UP(Ps6L?A?$N%U4-wuD!$^+YVA*zJ)DTfh!<5#E7$o9)-gM!$#@xo
z9?$^As?XcsM&`7RC2(Cm*JcDRm6f2Z;}O9JZ<PV7<vd%>C%OWhBa(ib`Sf&}AacAA
za6%&F?WLW5WS!gj2u3H@iMtBmKLU}KauNB6e~j|eK*jRd9iSBV8v2>>-&WFTC0uD$
zA%epliV9*8Gcs{$U?fWLUivC{4CxQ6n%F#Chzg0DJw!6bp1DBp_~;Pwd}90Xbvhi*
z#39w82Gw!YcEcy!KvM}y+A6D#U82%zS+kHjtc9#GLh|GbU>{{EueiIGo!`?{XvKxN
z<`_{f{KG>APjPo5zX9Dxwg*JW@V1)5B6cdDBLCu?=<vmx7kY*+)x<6LQn<n7NxsPY
zjJH6Oy3n->y0dRk#egDDv0QdtrKnj`83B)-M5hgvj(BPxb$*)>`VGd3b1t=}L@35U
z?IO;-cOdMZ9BWfUvD6J0<3rH-#cASat|=pGUCP|Pvn*}Kau%sy@Vq27d?~c<AckOx
zg4ZWF{~q86`|fyk0a)}DohlufPj@vUfckV-(Av!cS}{72FRjZgJNAF&QLqWGO;_;l
ztIwLR$Y>#i0dlpnA}@Q-l#%wCL2SShFp2LTSMj}#cp0$7ZS@cz93%|rm`s65+*x$V
zSplO;3N6%v|AB&4JWlp=T{E*g%Bc=&{ZA16x~87lKIo@m))o96ifespdsP+2crC&1
z0AQ;Z{dlglbrpPh*DW*kt!r($Gn8loBG%&<F}gosK%AO}D117YqJB4E!)O_QNOxX1
z`N*RN{~vaW$|RUyKI8G_$WTzr7<ZE4d;}4dr7L;t`D!)5s>uPjo^zLliT6PZu1PzC
z+6Fnqq_c><8QKdQI`&%wU+b#jTS(X!pjOo|29e8Yfp$BHN32Vs;^z4JLo*<iK&HCk
z5H;_DQ@CJt&Vaf!zI>i5R$s@Xq`4m79aH0&ub!qxyaI{;F`q$tU0(#;O4O(23()Ld
zo|w@O>?HJW+PvSaSD^ZM@*8EmM1MjAoizw4dQ59w%gh%zh^#BDNcR`zWSZEcTywDL
zD~^e9A|;mmlyx5<)0581(|at?l!mKl46Is$s|aZ%L+n2k@c-{~ZveSEcER%Fwmfnm
z#2Vs=fEoSE`9z1B#^cj+#J}^3#gh752|LUGUtw<nRn^+{kFsH-djld}0@5ulN{DoK
z2_oInp@O2D66sC}krol@R=Pt#kQ8Y_5h)Q+;6L{{@Ar=T-8;trp5ZvdAp`bet!F*Y
zoWGh3VlHW#xVM=J%{hV4huppNI>b?_B{Q?IQb6kjS2#9ssQFmLC`_0NEAaWx9~2X`
z^0-sQC3-X1N(S)0`fB&d=uk{63=wm<FTvV43x%7x?;`%%>%Z8n4iqUVH-}r7&<<0v
zeqrFktQR2S2OA%i37RzWYDLa~nwsTK%P0GvpAR2Yp4p0-Jo`n80G{~U<XdfH;n7~!
zKtSo{pDc9L&SA@|rkC$PU;!^iuT8B<(+NmvFkS93#rz7{!!|Fkf%qJb^xZVonL5_9
zeAUE~TW^Q??|}kYdFJTwSczvV?|tgIyL^YJ$I0E%jsHj`{`1gekYw1s3qhYnz<|K*
zDZe)f1tE3gbBD@y2=eF_`T@PykS(*@+K|zb%dHRHwwn>pd;F(m_r@{$+Cw%RPG5$p
zHZEOr%a;IwQx7Zfa(^E!14KEMZVm3J(>HQu33)u$U!v~_e8h*E?h{$@?V4|(E-0nc
z4#K{70jLnH7~!SsLqjQPL^^fX*Mw)|ptD_0FoH^1Y9;=i>2w=*kSy#?Ogw^%fr;GL
z10U{$kZWPK1)G1vxHN6aAJu@q&^~48=^JMkoFow>{8I~5^lvbJH0<{C;_1Jgiol)`
zz#&v@sKHI%CPYxY-^^a7c49C3w>R<yu@gmht&wIT3zMs^UK*MCJ`i~}eeDh~6bv~8
zt(Q8WXaY@}l_A%ry~sXO{aLEP03)w@2yHpw%tf6Ei?$99DmieZKv>-mpf+Wnd<(kC
z)Kt!1rfxa>1jfq3$J@O*g)7|vcn{}WSD}4~;zb#(xyvK96*9oXmFADI+1ru~l07R(
z%CEBC0910;%M`(rDrJYx0{@f38_EwqgLZbSfRz4)`98V)cky@Qp}`ZUNdFlRGD&?T
z$|w4_TA|lP<@Z=ouits;%}~fbjt!(p<!W*13%M{1Muix53W7X<zh4Tf%{dMlp1a7H
zc}S-{oZD^vt`XkFq+yF7S8Yg|B3G?6;WFefqHw(i?HstkatjCTH`00GB-6Ej(Kd*o
ze3^NJQzl1f1NqVMDQr)b!dB9qj?Po%i%^DX^wx@2?ygYHWI05;L`ns|l~ZK;iG(=d
zh<1seM_kXiaAB4UzQBs(*!10_`sJAA=q*6H4|FNC<`qq9qDzu4x2XLqSfFJa0R<0q
zW47f1G7ear(cGUEqI@u$ttC$dW!@M6h8M*gDg)S0DV#v0KAW#fo}|o2`xx$@_>C8^
z80{S<;xdO_#@V&ZHYp~qyuvc)&7Ppz0cNS;a4GV0|EEy@k(1RAZ*OS0Iz|xn2i#z~
zIYC54GEzz#vLk-=vgmSg7igQ$pxr($Xe^E(zOk=^84^1gGf&V*=Az2G-aKj1)!02v
zrgmS_vexcoTLKZ3!vuD685X{V*75o8!Ky@g%P6H^2kBYFd3yfx0n=B5<J~)l%5sVW
zgH9AW>F_UQ0{hv156XQgjAXw;(&P@H001cQHI?rf%s{5E$1B!Is1Kh<h4qABW}JbI
z2{Z_;6fg!Ey2m8Uh`$5x++OE!UIuoU8PunAuBpnauZ@4ADfsWu{A&E|b$AE2>W(K^
zJ^mCh65L)%O-aavx6*_a8knE3s7PmmI9Z<x=T|uq3DiNKPoc}bbA9+Zn-evgtM1IY
zPaeE30%*(GGnhM#q2$tOO&#Op6rFu$CEqC=pYBdGLG$PN1zKUY1|pY+2$|2-<Dhq=
z|G_xJa;Lf&GR>qNBa`_10>zC%4Lh=9rMDvKPAvB3(N3!abkHCy4yjh;zeiw4u7|sM
z*1Lt|Q@BK~YoUW0jB2N@X3vMuK0)e`!O84+yuYD@F<e1uXh;~SfchgZ(RvS2xCLcb
zg-JV`U81XQLXJ;5DkMt=NU`7&9&0CR^-K^#o`Um7aw&`z^{CC<9(a?U-(SIdpg_AQ
z=}LMV<eH{r7E+ay)<f!i7d5`(k|P|zO|Tvgl8<nQ>ycXV$u)n|wr7=8LsUt%yfN4T
zxo*LK-=|%>&EL>Tj+jO9nKw^uHiV~Kb|u9?w5@hE;q1zwDkCV)Sn_;GNdi5{jNd)H
zO38WXWbkVsnfEK}S!^qo0#y7Z|CloHj6L8j-)&d=c(AKbS`&m3X7ryY!rj@kzkImB
z@VKA5Xk{QZ_QXffJ9;kT+|yAE96x{rN$V+*c#!gDKq8(%$-AkPEm))7W&=g<Z|hDh
z@=}TeoqQP2=+(<y;oz!nMb>27N0!INlI95_jxe*&5VPLO_0t89-%n01z*0NLKJ)yw
z7Gaik&xO=&s$B&_Rb|7gP}(K1wf+$=bI=V{RGoMJU<bn9k5rfh1>_x$R*j6T=JH`n
zsvhkVK*CulS_bU6B;TT&K@^P~?7%s{Ou@13E4YXcYxTQ|A*rx%K?^ASMwAa#Ohc-%
zn4Dpxb<_pI+myrSluSEgv0eb(iRvG`(gAST<;0S{q8qsSkV4lNRCMkcnNV%`s|BQe
z3aJRgP%yfZntuKJ=gZl_Qywuvk%(T7kqYd7p061&aaHux>5E%mbRg}lovE#AYyyoD
zNo&_WaJ@`qf-hr+7>x(Fyk>|72*O2}kVY4^99M{J-PKd3d;p)Dhz~uIBDDqi_eOqN
zFdVGb1nT?FT1V{x_!Xn~cEd|h78`>um_~|>tuGXR3>ASV<}&)&L+?$b#!}jt|KPyg
zVc`8rGq25$v=c8p9hO;>7{(jNaAbECO`U%0t?fM(O9%lK4<dS2HAu~$He*nbl~tP5
z`9=J#Tt%!FHtm(wEyi@WMWNivF9kn?oVwY$U3Ks9sS$Ap0>(y;_te7I_5?FZulVG`
zwh$%f9G-ke>{2j?8}Xe%*e5A|(~Zl-Zy+1O?mqa+=NXx4!(f}(3Pt>&v>TWzJgx6J
zRQcTMm(#9>{*!=`r#>M=7>$RJGpYC^$=p39n3hF5%_vP&bH6SIi<#3tw^Z6%+q8D$
zyH)Kx8FOH#R1JBRgLxR8b?o@D{`V7IWXZ|Um2;rBbA#|UVdMj(AIR#+eM$>=i3<;E
zPb@MaT=|u8h>ou@Z+UXVv}PQlu!Tp5ZHOY%w2yxOx|&c@tr#!rLw_HaNP8-6p=b{h
z3u&nS4$E7cD&**t{tmB_gpIm&KNtcRgQF>N^|N#gSN1R;ZuXc(US|5!pUP)4R-`=_
z$=}Fj@TRc8P&45Cg$yld9jaiFTNk!HXTv<6`SCF{Za%vBwmIhHp0~hSVhEk>KoDp_
zg375k?9#HDNHS;(_kscmrPf|U(l=MA^L?3D6r|-TIQ@_1Oy7ftAZmhUZO^-#$)F^+
zxsIA#2xD)%DJUO+8ei861z$5N-WIwH4bHBa;MV$=vWJ$=-0Fg)+aWkqo6U|&$MpE~
zwP8b{elw)wp$wk@1BYI~Q@ht8SLaoPu<TVrzuixYbL^kKbG?R%Y|WQ#VOU8-s|-%l
z`u^608Os6*6W<Nf<;NgYx-Oi<!;%$tD<0$CH=GQJTR`?|Ge)}Jzp{p%O2Wl~_FlTH
zg)=n4z*wuiDtG?@t}fe%7<h?YIIxo|G$M5LL>w;-xyPBmyulWG;0oPYXQPskLfR;3
zjLL(ifV#(G&cHL4GCkxnO;EgiZ(J9!F}DP2(<kWwxm}Cv+zY^Yh#}8=CuS$9j$sT-
zSY%0$CXe|F+({2jwn;xE8>TPfe*W+T8$$K1{>59b{Un^kcR+fFo`#pDG`tvEO!0Wr
z0vV}N1qJQFGB}310PszRjJ_gr9)FJ-+eB^lA(n5dQGM;mX+s86I$3PmM)1%-GVh44
zUHcr24&|gVEXKf+qeD)$<N`Duuau+9vmm-VekBw~NPaptZ5aRo*&0x!0I!L)cm3Jl
zSSD0Z2vD@Q2z&lI?(DaD!uNtkpk(C(vL2+D(nW^m%up?Q_hiyvxG^e(O}YNq%|1Nd
za{5c)1%W`tGFcjAs>7fKh6|*HRD&l|>BGYqc1irQnAtEjjbxl*LSIB)Q;%%l-1px@
z!TRhZ?cdR)Dxz+CF@av&z_t=HBMnWEf7NlKTfxSNwa|Qx?%MTmI+gnj(n;v?ma%aw
zw<(lBycuOMIBh~56$C&<lfwr_mXUOAm^e6PnBH=qAT!EfjEY@ebuuI#2UCHkLxZ(~
zUhhJdo?gEV(8^GnOB9U|ZmOllBtH+1dCK?pW-TGd&vwJHF#)XI<UM`qT;)Cn7hNBw
zkAfUQBZ9d|(I*a#3vzds_rD<Z#1;4JPHuX*UQ*EwBx)%10);oTf;HuH@uC}*s$_6}
z*`l@w?5J<Eh^PIga>d+@C#p^=V>S{Et>gZ@wAj_$d9Xejyn1<4>1g2m38`egRp{DF
z13$jo*^Yk5;@MygoIGXvn!nBu$Y;fTC#(d!3fn<K4bGUXlM<HuT`<TF>-C~ArM{%6
z!;O`giK&ng(fSrzd%1jCH+JH5B#aSXcP#%U@BZD2!Q>{U)ZkSD2y)lsNG$PpVR-Y^
zM_Woi=;_IDtzNG|mVUQSke$PUEz0z>p#>@G96}CLnQiO;tOA2pDlmxLM{Sq+0zJ1Y
zHc)br7ePZCe~~nGY+xv%$rg?ac@-jcG)ddAUY{N{;bax)eBnxC9<S04=zU;Pe?Wq=
zpKr$aA&FNyKK=%!SI-oix?PbmQJMedlp)Yj6hA+%?x0CpNG9bt`h2M?DL&SF+HA&c
z;O)E}w*CfI)E8Z21_s4wh#-&M{tICPDd$oI_wlL~Asa#7O2L0iaQxOiia;4Mh?F4v
zInn$c>Sl>JN`ry~P8ky8PYDb?NtJe(flA_}zdxwVx&MXwgcSpR5bbcVH%c>>yC6fq
zUDIbbNfh{*b9hWf7<};$fenw{4_@r(QMHfTR{B==3t8WPzEU@5?H6`kqk8bYncG$0
zv?i{*vQ+C+1Xup3ds}kuw?Z%YOsK(B^Oe5>lPfMrcRHp*!n~Zm=3F|F!(a!-8S7(B
z9Px<Z6PD%`3T*#syvnjY2SpVh6AOe1JbT9Yu-BtC?=R#H?MspaCq5sPaMJ`AtHCJ@
zaH&%7Wz5AR2^&7Obn5$b<?Ht;?ockpDd&aW;7fFfVDI;zf<*?$zJ7QSdx2DqWj=Y7
zdPG7qL+n@AeX&qXYGekL-J8f<o-UWR>Byeo&{(dqja*3ee+$4cMDJ7;UYTj|<|M80
z<ltl(18rjHePKC%hKK+Mq;Jd0gGG}ap6gA>@lVsMZ^jguSdr!7pWMY+ZBJ6TcgUwc
zC*UZdhBe6X^japEZkszr=7z67VnUaHUBbI4B?x$o1m@m1WMz=JVT32aB^7Bd>k(Qg
zDBYoWhYk0`#&K!Evhx8fVu9P$D0kMuuqjfJ73PqtX~j1)i&y`c`9Pxr%1AP`T#za)
z{=5&s=1n-TI#v2_f4wF9F9GuJ{{z`tBiVJWO|<OW6&KW7CUs|6TIzCEK47wNI4&WJ
zzck)+?!8vqnM;5|23c`>eK7nCXz=>PYidsfUW2-5Tf;O(`iYy@&#d!hI@DKlgl806
z4wnc;75t?k)Y;nU<B1L>1C)oVwekGqppXigdzVD=E5&`Q@LXkc;Mkwf8~mgy83EnM
zXtdJ<M(<^apO%10aJbnRB<uUY1BC$_%*c;ga?6n(geyoBm~T<0-^Xo-z9)wHqH(jU
z__*QV6y7TyTo-7dh1MEj-Uc)|DZCE%MpYEb9*n|0103;ko}B*>j3ZW9mt7*S_zaF1
zb1p+&Ha`Z3Ae;!xpto;Ks?J?Db(z*-SuP<tyf~3xX^@VIAU~o+rCVlP*LHEDeTJsW
zJ`}zhck4n@y_t^@2&i&|+S-KzPrAY=Tk*PHo5$a6@bzL6d7Fhue{{U{=4o0bG9zc&
zpR)mZ5Yh^!S1`t{sE&T3A|NTZy)9+GmS%0=Na4Ka;2Z_U>XUV_3>vl(KS!?^P)9#q
zHloHq#R&D#ca&nM_bp75n(8c3Fp94KsTA)hnk|LVwkqZiyV>37EHdq!;;&&NnvdjW
zh|mqzT?7&$T*671AeV6EZUS+Fyv#CPJlXE#bGhNP^)&Cj<QUd`qQ6PseH@K;i$<4S
zli1<yhsF~W3ys;{p~AK*Z2QWg!XM;nqcGow<M?M;9M{QhaR-}bfGuC4D>Scz`kB;F
z;@2O&reKu`kfW*n$i^`n;#&pI+cppzuLIZ86-ZkkKX?2pvfEm9L~&YD;NUn@PvXJW
zUr@T`#<EHo(d5p2?Udb^=4{2i$~Y2*V?~%k{dA^2mF$|}A1Wo-Lyn(a#ysh*MZ35$
zuv1!0$b^PJyFf9^1?b8}w4;XL_78gUVNkkhv_D>%dKzw2VcjLQ`GKG6ifeqNU~+Oa
z5<;;DS{?Fi=lhT(<@I{f5f0LxRqccU_X6sR+7o*Z%s~d}4`NK?)iR?uGk~Ciu(q!F
z+v`Xr#G>;Ye}pHA5k?L$PJ8XTUQ_FM2?uODMGpqpN+p@Cg7{*y2LjQ9QaeH&T>#x+
zqnT?-!<zLjP^flvoQt54?_pVBMnt3IvDZ13@&uM9aM>{wSBq@uut-rUl@9g69s75J
z+m~U;-z#CtqGR!RRaUc^T+y5js+**cEFvk12|7|;8XV|%B)j@Wt4U)CX+!<|;0JR<
zeDoqV9_W79i>uB?fK>aAE6&iut-c>T#UCN-4r&0MI#QQ)Kzuyiu=0;rMcLo{CjwDZ
zGM3=7#0pU330+9K&(UHvJO}TQ>IYA)Y@UG`YRPg8x^x&*Gl9xE7jYMZx_F68jqEWc
zLL23%5!UXunASR96JoCD0Cu({E~mWr)*OXlf&4RjYNiwuZTnXpkzp#a65@_E{ics`
za8sIBcUm3@7?;bOQ&3Hi+Q?6||BxL$gSsZ;qXfm6NF(NL<2n!&V2lyZe36)RGb2^;
zo-PX({lP0@wB_)?Fs<E=5d@`gd&cLZ5<JrKz<@b*8&l@MM^2kvTZQ0sqJ26FEwvWV
zyMTSgH_&mVl*20FjnY$K8bFG2bmG=|Xjpa&ffTg4*#r|!IhAjz+P|fA2TP2LWwu!}
zv3->J&;7qiE+N)mEXHZwP^fO2#LHsFgo{~Z4`2yS-9%RH;Nla6q-x#1>ONJ@Or=yD
zy&Y>w><wP2qU318YZc6&A>wp!uJ7rla0-4SbLq44P;V(%otq*=l^TKLxsIc`w7hp{
zBZ2i^F{e|%QU<?(9jzMklTJvQNw=}gJ>0DNP!eb@{%BUFVi$~Nm8vo}a(oz9$(|6X
zJu8A>Is5^&P-lPXf_^C4weH?`qdJyzaMx!)Cb8gJakkQJg8!_`cre8r&V~mAB;1?E
ztlzX3OaU2y={FYS@tHs5m+%<@ubt+<n!JnB|NJjIr92-JLq`6lmx3WOzz}<Jv-A~P
z1r?LF6067GF>f6nb)v~RiUX~6XGdfz^tj>5<gr)g;M@}eeH}0j=Lk=*(;_0?Ea=CJ
zI!e?32H+&|H;ocN4nY$YsE||~1OwPD-E&bep6EBgb}kbol%br(+vrwObN=*F@l!|S
zQlBmQbodf$DF}KQ{I}bb84njKpG;X{o-IBM-Bp?U`?*>+b)NUPUPgmbL+V5JZ;<Tq
zOrRW|b_+W^?KF7WgxIGoSL>%?S+JKED{bAP&ZmUdu?)Gj{i;~WhxgFnwAlZbAutFA
zNKhTeZqc62!#KWJEBvkz=$bW+bqd#CF6H#!R>67u4p8aJ4Msv%T?_^=-aCZ&DJpPI
z9+1c3)ttDpq5q+}g%H;Jmb?-85{L~fE}v^kR(f9ldFO+tKls{;rdTfM2Q(YpnaOP)
z`S^}y!3OvGVGf(MH$TTb49;Rn3SMeCtQQaFdI+d*u<?}%)&JocuoQ<jQczr3JB0Mu
zsR)y$q~L)6BBEeJZJXOE(z=aD5h*8|11K6c;Ae43PS?~BECc^?BmZGYWRNJ8NvH*L
zOL-0rTwXVl`K!(38keh)2I`Pl2KVWV>nB4I^<U}%{(@o&1ZZ_B^gejN$eg*ueyO);
zl_6i~IVN1$za;YjONak|{`e{`fC_x4WW6<&*Hw2R!<do+k-A-gpuhxD0MMO;%V+kF
z{3npj1E^eS2*85f&;HQ{*n=hG?im#PuWJVIhcl;Mv9D?@8C#Els}w{}k3i}S0XO2X
z(EuE|`wQei;8>{&GSZ&-FCFxsUzrkk*aA8AN$izkWCBqB@U3v@gZ5WYjiuxlz!LPZ
zsCG|?bqHVc{6Fg+mN*WCUv2?fs}4W76$HE5^}b&_3>v`E327m_SiQ(7LFH!~j=^hb
ze?L6G%A1$K>wOVDUSb&e{unGSiv8GF7hoEmCjsq!0fOz8i!m`nW-gSv;3R_gMqwUY
zu|4y!TI0E)aEY9tV(?puzy3<myzoB`@BeuaftKXpMJb=XRihpu0|g#7lY##PN*vG%
zNPx){m_!FSSg~GMz2BzZqkw}+0Avn5^DtdFTbMHz_Qo1l_A!!JYwv23enQ{m#?+m^
z&<W)6fT8Lmn8fBm<ni?YU}FquLpphm;V4#7Ff}OFVg6c5{Qr6)?W$mA^r&sRdvZn!
ziR;}BU$einO~CsOuPE^7S}XZ8;;-?r;DZ`VBl}}7=;Z;Uaumi$WEICe?qg3);xqAr
z&y8oeRFwzuaq#j}|5^ullv`U=o9@=r|GlbQkt^Z$rjaU)KhY%b@R_5mMuF9Y+_sqM
zE(m|q0qNhE_!cm_V1uv#H!Fw~72TL>VA>$y@(_-c?brsV;6s?Sm5J$xl*4t{O}q1o
z;GsZh!-KJ*Hozb~Fdgf`Ou{fV;{@)2ri9XEI0wx^=kLiX0Vfmcrpp*MpyeaT)4&NQ
z`K>q)1PE1rd-h{M^+l3d!9%%o1ls&G-e1>fkRZZ)V6BU_o4_=iGq|@{nZMGz?Mh6i
zC^zQx%!~~%e*ymHZMkD{wPJvc!nPE$l}tsgJHr4m82ygL&^D4AEZcA&rlP%Uk@XN3
z2^MN0?`HTi(AO`dO@J$fnJQ~g8r}<0P*lWL`CNCW0VBA+Awd#ILWad07=AktAFAed
zE2jT}^36UeR0@;xu{ERl%-x`BtTL($fNW#9d33A&XV)$`0xKu1z|`MlVZ|%lX)-+n
zkC;VrR#DGSX#qGRKj1PzE7>hY%M6KL;r(kcy+FZr9GH&%>(^i107ug1b!jNetL4@c
zGIeB|BSi{+g12v0UC8x40E&a1Fz8L>14b$leRGDsQZ9K9{6~^k>7kwkggZSmmO-)>
zT+%Rx?K>8wk0|6b!1HH-fgW$OxkAfNv3hhKHgk3%F<FC7`QX-HNV#?%g2Y4sSv(h3
z-Uj24U<b>9N88Dfg&b=)5PeLS%u+hx@Z#}(AjiA~QWQ95MtkEitU_0hKVVMfi{x%}
zz5?^WQ)87RxwjpwZ=eZ+tgmZb%y?PpTI)%}KUX1hf06h6T)Pjog^Euh@ET137Y!!c
z4*$fT;JMKZ>IgU^yFEVi$wHih7A`0S_+lxgu`SgH{3j3oYWxkuy|@m)`fzN*S03;D
zGxsv;3=CnW{p!XS*zG_Lk(P?go>84I8~BG`3wY2%TkXfb{5YG^dVcKW@H_E#j;rg}
zof^wsg}r-ws=g0U*x6~tX1#<>%+(@X)@jEQ+0w?}{h@Kg3cO8~hth<Wu~TH1qp$g%
zEk7jU9r=mv?`g00Gi9@22lu$vVLo?ri%P=wft|i%a=RIJ2*ImM{E3|R>FZInqX4G{
z3m0p@kVdMr6aWmEFZc=zJXS0pLfbw<h;{vB@&FseQv|A4L2=H<tpN!8+F0vIIn~9k
zefS7@KZSc5YsZ&paDDJ@w-~(A8!v`e$|nu1NmlKN^+A|*Ax#qA>dZSIe3I%F{p-I2
z<;JmIW0Ti;EMCIJVrz4)$GH=#fGD!W0sf{<;eGwOIWP64pzG>?^T2A;48w&uRK#d=
zDQ^RK36#yopz+{Kxh3JjJRRmm4dr`YL;BQVG6CqEFJC!rc)}uHB!|OHm%qS*4d-#T
z-usPFpP=#8|8*AaVq~yAMQGvKMLETkp~%o;aK{VwoKCW^-K|?w5ylE?Xf}}I;8&U$
zz8Xhay0(q+U*4DYPfOfue`Xrp_cUGXnA$+mzsnlef-#S0iT6O-x0v*iFSOKCv{uGe
z28xi>Ksjv3cRlC*qA(wK8k?jyj?_$f6-F>!0LDa~j6;r=^ZVV@Nqrwfdco=LLOes#
zhg>0kFy2OGr5-%6w)emi<;^bCK$@&sxoz$3H6q)J8i_aV%_U5KXhooK{k~DT_ZtpI
zlJ#pxIrR$XLxhv&osXFt2QVk!Ee1}Ha>&{%kHc7FWc2&|DQ^mM?qSaUd>Qt{1mQiQ
zD^AHg8An~x?ooOf-k#(nxO%ETE-WF~f*R)N)htqAazT#d?&7V8%XR5aTIXRWpQlg_
zUi5$v%fE#>Bvso1=rPddGf0bz^Lfj{8i#GW0mdltiRo}X8T!kXMI^){wHQKPF4{T!
zd^GgSTX9OUi;mf~e!tJ`F!jOarDs1uEYmAY=1xs~kogzmcT&Gmn2|Z)A=Sf>R!5x3
za%S-?iWpm=j}e>3&*y0V@FFmMDJwakdAW8e{|bR=<m>mIC=y-N9{t`TL)Oui4*|Vb
zI{fa5)qx!B74VhuCTVZ;6T{f?1@@5+RwoEUVlg!~Yu?UhxBg;!o5IX(-#?F&k9Q!X
z;IedE!M)>~Hb$;<4jtWcF+8(t6BWxCb(X%udMU=8<J!puHiSLjplRBSP~wWbyR50g
z`GJ~BnqLbK?WKz|%P(0|snN6Wje4~AWL|KMIU+F#W6_a*ZlY=X%vM-+rL@y7o|(f}
z>#Z2|M_pK{9Xpg5!X`|_-LL{O@BH3krErl+6H=5IV!XaHAB4VNb?_>1W$nVbLrYM2
zhEHn$dK{wjm=?H$nR&j`!kfx3fy#4NOeUYSO4lIsl@SRjw-$1p5p6Eoi7M22^3pyq
zB<;792@MHWMdE@pihJp!92+|0d^PWBOYw#i;lAhJXv5vJtF@W=dmCR`Fj}WAX<~W2
zqNzKn^Kt{Zn-GTHLQKUFCX<5R{;n?IqWCaatY(1Q<>|Q}qR-6{8+t(;+SN%O8%sJQ
zGLrHXFMDf@Qj9f+Mn|uw-cLcWV95H0g-Qh@m@{86EHlNVS68hJ(BTeCHuvuht2Jma
z#`W6<5ukT^$^YE9wzv5hsN6y8kV$G15`!m?(?z0N&GO^1j$^K>tzZ8JS)T!wE9t2<
z4T>Z>Oe;oY+3=Pbr621#AC3eGvw@8`4Ze-U_*6y4jIhZx<{X-c#wc!nd(1=s`fQu(
z;`nQ8C4mvm#yu<~Qn<#c+i5xGQp7rloVP-!y@iDW&D*#ZkA80r-P$DZi`gEy`f}Z`
z-_pCkH&%?3^O;??ELtijJSk+FG{Pnnceh9N8zd^E3tHj&k?CJNq0qj}dah*-O2HR4
zKB`uG0rm@;9cRB^K4TOP=ES6Q0Y_v48L$K5_8+6_Pg@jtY`62@te!bHW}Zd_)FAJm
zvCDa=#9PaXSJbg+#)<`REd7~fnYiWf2P*zM-PW8{QjGDw>4NPU6*nelPo4z2{DW40
zd5qh<@-mWHOULDt#xVM}6H(BhiCjm{PFs7JuTrx{-dtOGm^dYlmaRg>A{~X&Ai*GR
zUIV`N3sAA7GNmCXG0^lW8!9wbj4lL_h2QBwNbpg;IEuJ|dV89`PA-Lu9Bh9^cukdF
zFY}_n1r5OjA%O^F(ED@RR8=+SZ=`QrmJ~|AXfEg&<565{TF3u5R%><c5VlmC`|u{2
zkj!Z8B{RA`A%qe(Z&vxf&ERE~Hp&<x@qx^}7rAP?%hu)?g<Sm*c%ws*f{Wq{IFDKX
zFk3DM>xJ&j9lfc*X*CIcYC%=jqhzR{29+n8+5L$dc59)%S)D1B(~#<VBdgAK^SsFq
znWj!Nuz+;me2z!|+D~go=yV+W3>dg>oLOuM!~K3Le-zW0>8iKb$60nJ@ber_j}e#G
zjewV!`MHEcfgfCFUpm78nY+Wbpf(k;f9?nJ-{;WTEJYqi9Q@eGE-jwjTT=Gs)L808
zv0A<wYuf%MpOwO?KHH@)&ftx6pH=mmM8Q3X;jTDZIEf4=silmM$IVReeP+`?q{jEr
z@{aoyRrJZI$A#9Ly;GM-RcM?EH)E$T#Oc3pDv@W-)CW*M7?;tgnof<Fnu>WvDsmT&
zyl>jRmi*m^FYwKmNb+o0<kzOS2nDAXu8?<-bhcuinIP{MmGamOtC5!S=pYHWhpjKQ
z^Ge%^yE(_KNNjT>vhS`tPJNTsdqU*xakK4SSGuWEwx-8mv{HCqkXT(+jwX!@hNuMH
z(v?XPjPF0P-neggOv!PP>+uvaWZrStHZ=S$8D0TaN<19C92Nex%jB5zBikObKqT>I
z!;X}Fp~5rejlRxMqNB(T{6amgY;^tvHz`UGp`jf8M}#pCs}1YN*LY%lU(9f=&R(=6
z>P*Vn`bG|Xk-F!evGf8?7N!?ZqTijuFlkh2$9R!i%`=L!Swk|uw|xzfReL6I`)0S2
zbbahguXZk?r^ExJ=)ahuz1ycPh|`MKciWg2jIa;Ol}0IsjE|P?C-TA!_b-0)49Ylu
zcI!KQuDT_iK96zGB-7Nc6-V1oqD)2rG(ii{nHsMBm?C+3sl=iMBdru(iY9k!Rj5^3
z9aLx5cb;!R9|u^_WF(RP4Vxk<WwWeNNxR!lBy^r>P%tC^diu@1L6Nkk<<Ni~MQnL`
zI}g3I3He)r_7PVKj1TS118LdnY>vY!&P_{$yopAJ<zgAZMX7C^Mm#s!b17M)_Z1(C
zUdP{PeSj1^difc%aWh7*`xS21g<nfBBVsL|&AavWxkzTHY&CPVbLZM-P~><%*Y>#*
z@iu&(>?&J?^XvhgWY=z^W#!RxrL=ow2$dg1OjtzeP4JOz&Um{$=hB{1*Jju&0e6v1
zyCTkGPU=MF0}80J3SMM1?cV|`KZ#6DgvsS&lf4PPn;4Hlrr7@OdMU}$!@Dk8>G!sM
zT6a+$GIADsmqaI5z9FI|`ud{rjD&R%-T{>~H3Gpn>ZI$EG>FWBsPm#7#I+)WdNeHb
zPFc>l-T}`U3l5;07=9ji1g2d40qeh*#J8-jJ!S!`-ZBgiXyTQ1MwFjSEsTj4RD~*>
z!ozgJP_xRwX~VV3aQ3A)=z6!LwBO>r0$@+AasL(zeY6k!)G$a}ccj}xJaeb}(W9AQ
zu{!1BKi*MB&+8N}%n+a168p9aO;vvTko0X_Ut4oR<qHLAV}MWd$3NdY<SDN5yjc~f
zV1IK9DLiiIYtwO_vsTYXi^XJ2eTaLWG|NV>9COgK%XPr7Cm7l}8MRIs@6eE{=PP+F
zeC|7a9)p8xdS%*?@_fRIXJ|Q_Nh5zG+k&g9V4uPb7;hwAXffL6OB#Md(V}3v>qb|X
zkJl6YiU(GRo>x=*nR?A#*7<VhG%60?tg_{Q-)KG)QtTm|2>sS%-n4C-R(~`i?SUY8
zu7Y7|Zcpe{!wD~To{kz6kCd(hC(H0HMsjSV$xY`nwxCLHf+~ICw%c(OTh<Ki;$3E?
z9%&R*Y$Gz_E&5M3o3CXglZhH}#bh<5Ju7c-yJa5Z_wN3f-)#3_oF=o1l_xGkg*Q)|
ztTi9n&F93xS{0`kzx1YZ3L*da7aE4dUW~<^y+IS7^-%dadpp1IQdDAFmnh3YnpU<C
z4wbK<t9?FxlhZ@*PkeYeua|g_ISwj4ww8p2JFh&Y7#sZ~Su|Lh+|ul8nHE~}+?#;%
zrlRuu=<7u+7s+hGn}xrnUBc1^xnsX5ap@)=k6s{q@TBU(Du8TsCe-s!buKREDPUN(
zEaSxmc`C~WwPza`37wNR=i0ZkyK`LS3Ug<|2IB%#{bW_nxfU&<iCHq3iUy{lilG@?
z+k_NezqrL}tM?o<Qw*k-o9iZ@i@4D$4qMhoH!EeQgVPpMZz(g<u9Ksi7gXBsBFLTx
zW;K=TdT8x1EhN!S`EBM-KbRjh!T7CGFmD?#FEYmr4x3%i$k_Ut-+4pjocd<;+$zyV
z<&~gNX_h!)gv$-Vw_)!te1FXMmqFmtoJXb&{~G0sU#(M3DvJ$4J1x_n=NbYYad)9}
zyWb59;nCk`NwOQfmTBsPwF+65EML!1m{)1Z4c?(5OLvUfc<s|%UsNBUHH>-7bimK}
zA~QnsRs<>}5y$9B_lY0qI*qqqQY=hd4$d;Bqp4W>(y7iPP{VZ>J0mc+ZK_)MmX)9Q
zr+;fb8@vJXkUB%Dyz~{D`S#aW&Z$WS_*O5fM*O_Qt>LS4b0@U_n@X`%j@ka%)!>)Y
zxWYHiz5t(O`7tC_f=&X{7-1u=L_TnK?Vu-@g0z<&QP*yiFXZ@@qBiLFwF@(Sb=S{X
z@D1kJ3e}bK1QRompWg9aDAPHgA8E*8k}cPiTb5^bB~^dPL{7$x-Ax{)eupoct>ns*
zbsd6GxVM(e?K_u{P+;e~Nt%})(R)^HaaqGUb7?MlE|Dso!(G$9n81_a^2<WEya+8c
zrRj4F+SS6|b~6z_6JA~oDWkeU@rgaAhV)Wy`VFf3G({ABu4bshmPmHdoZPsp(N&Y`
zR@#)R`Xq?~N?+AwgsYoX<O3=>-OJ?Lz4<Hqsg+(5$*cRw1WSDMHx*VUWicCR8I3xm
z`iL%vO|kDt^Ky!tFh+JGSSxPK#M?f6);Q3U;4yu-wUXxtZvGCu&wR4p5d|)8iyaAM
zYD&a|WTHPiTuNn4>8~H%tS6{cj^z#$RINXGPNe&hRYq7WVP)AstN_m~$yHaY6I><a
zw6Dj=A3U)!wSD5aCeV0~hzM<Jb(hr6RDp$unLcZ<k*<&n>5mbmns&N4qw!>Vsk@X!
zM0$BNwtx)(sqKhXVI7;Bf>*2HGiPHyMBdT%gwVW(m*rP#&5t5w!nocp$MWMAMghL3
z3F~QJX^u>8s<@ci3g>Iud{>uBEY=&m<uDcdt}EZeM5~A_Uzq?yGTC)yc18iYwlVNJ
z`6pW)l~XzXvgR#}8GCK`Q}&ebTikae>sMS`gNPg>edo5_NN=f3lBX|Qg$C&J?8#ra
zQ?@*GJCyd@yM3iXE-gnQeTHLnwlbf<B}3x1@9Wg`HN=x1DQ`W8ZijmYZ;-gZLv6OU
zIb1kzSgoZr*(w_TDyHI*Vq!bD+6v#>nc*v{jJ7ECgaT#rHOw0}$D}k7n&x=Dmm`Z5
zL!{0Ps_LVa#=eT?Kj}YquBTZ?OQdH#C(7a_znmZa^inz=E91-OWYnbXRoE?M-RZ$;
zXHADb(W4YDi;!lbnd`*l)GtV<qWTK=1&L<D#WSiJ#8_U6Rm8;31$aRjs*tcGz+7y+
zZP|t2W91bri0Nn5@X_7MadlvdrhhV>7GxHSqHOmVIwok!f4LQ#6vrw2$hvY{9(k@z
zsAz;=Xmj~P*q3W!?ZsyrQ;1;}9AFX<B*D*qft(;qQ&&84E#A=wf-w+fJ@k^iNDfn_
z)6Sp84c0<O3MZ@l2Dj{0K8#)RNJ1X}gK-0*jKQMHaZ(8Du^#h5Ub7X05Nb{7_*C(?
zCGp{FX{p1-zqQ8vvcTb5;$2bkRLH`Qy-RjNydZcVb=}@tw=OSW0PQ}P;^p??Rru=J
zcc!sGl&>7l8`+gZ7u()}9M^Uug$c#r`xJ5Y7M|1QwJzz?Bts22b<RY7+tNy!%WP}X
z4`r_3EMrdKiynvuNcuiw65nA(#V?S?P9>n;odIk+<uhd)_ETIPca7!>(F1XOKn8p1
zMI4v!^&Ygvq{`;QjLH}0Ic`Pkw!}-F`6_E-W~`~Cg$N1~CR{Oc9ht^zQ{|c5>d_4)
zEXq%bQAaPNsEayVZ(IGKWM`2*DJK(_)K$vr-{3uXnHB`CIKJ;Q(p;=9hvux3k~`F9
zllQJ08-s)BZV#pzkT|4>uoXf)@)|wjeA&B$+kbnf1TIdB74c=xObsnSyVY*KfyHp$
z<8gwze`Tx(g2Z$f7<-jag+k}q+AEmF*O`|@9xSxP@$2Z%<qqBpTUZ_ZnIcWz{BX5G
zsXakCfPcQMHjMD&Ilf+C7CcRr!^6>2Bag@-rp4q@`4nq7=Y@Fvmrgc|N?lj_!%LQ^
zRu#HS_Nxg^wFz)=2v1ZMWDO)+3~M~hXfRegMl%8U`FMp(R1K%l(_kNUDtRG-l#rQ)
zWFf!@B&@nOQ?C>vXi>MY*1ZT7>eNxVM^OEQVKx7`seFY9y8qkPF#q#?hZ!=EObiK5
z?wH9o*9`<NRh_rt6DKkq>S+8Te!2k#?*GkRXu{{&CD!)u2FL)v42tJ(vTan<M=&)a
zg-f^H8qvZ2ywTC+(d)?W|AdDAm&mbA1V_Yl{+miECXc_Eou0RM8peWyuJDsFSymXg
zdW0A9)6LF*%LPNvKk*9x{n%+XFfWKXz2>`!0uhpRcN8rng^OoAC}Oaa%&$-h-pK_X
zy0)&-k&nCpTv9;SUK-%}t#>tU6^OcQAnNk4MBP%&-H{kdZEh1*_;H^wI<J}BQ9eIr
z?!l1>%lDwr_m7%cg*CoRtMv;<=42H!(UQdp#k_#^*FRCc??h|+O1%UO@l57rDwcqe
z6jsABlIE|^{V?81O1O%>Cl_i6c0*eMmX`+Z1n|vR-4B=0E<BVd2rfB~GImMe9j`2o
zR1#}xbY#20fcca3&xH>E|6M3t=Lggsm|YJ;a2+7Lv%+Ls14T&C-C6>G9^`@GhT$)%
zgq*SF@ZU8qjy$z^F%tg*jlJMs|JMt~990=55Riec!3PWjFh&9N_?^AQQ0#+BlBGW{
zm{_7$KE6_4)V#+9-E)l7mf<1O0dd&5n{nDT6!&!RjT-!kDyvJu*(w0qOZksUkrspV
z)4@W^?sq<5wVLJ=l{LAW!*#v(s!ozG>{;m?IYP`4<(GSq<o-NFiA|W|Znf=+<OTi8
zD{3;vMYxJMSNb5s1HcY(|HJJ>Es0kU9!JC#c&4Mtqn`Zg`b26;@0#=*{u2JX-e*ZJ
zedKYEoY;5?KdVp3q+HX<_m!Xu<RckLBSBBM%S!4sXU0(a^P~`Int)t%YClO2PE|5~
zvrTtgDIz!N?VSf#F>H@L(8ZDG>Rb)0JV<4Znw%&3qm031ZVuyxJq)FDYiW$F#++ZC
zucqs*PN=y+F&gf}g*(jz@*AClpJ|ps+8~|dzx26_)2<#%ce1>wuI>CU$+>+I(U5w3
z%e(|DHu}3~nwpPbhVPuT2|`ONj>iZ1C}XabEUIn43f^|`mQUMU4QvGsgftL9{+xaX
zG_TnH;phUT$XoUA0Tm?Ww)^$NSbx9v&6AMnZ6N=AH%3MHKU{EBJXeji-FzKPuL6M#
zmKK66Nu#i2Rb~`ppfb)tvyV+~prvqn-`MW+Iy$bO7dmyqFMFm7+wEeq4b3w?zW;Nz
zWOkCHK_y?Pt>~sLT})UY(#fz>=3^IH`sr3wn|Sm&JA*Gx58gKD&g9++CCd8OL5AW4
z?^yfvR!I$PIBH;1Qwva^$whkjs3GUrw?R4h`^W49SgTxbKMzemXsFl^6Zi`(!?Q>`
zx_0Hek(Jfv&yk@?jf$xi1qdoqGIE`5?v9?%3;G}e-MS~i`|lxOc@=^G!t2IiQnjMd
zZ(lSg_FpNPcoBS2piWTiY=v<2XCl@MP|be~GfW~k{@0H1&rw0~g+%6-*Ir<(J6o;s
zNbr>u&D`*S)DieFLfZ`NCd5&;t005@bP9r2dw6-L=Ohv||G64;r6!P&F66iTQZE)F
zs0b`Y;8z7XP#?!XS4GH#JIw+6M{W3PDeOm45tgUJ4-~75u_gbPj3x+lAqVH#Y^P!N
zKc9%1(g4aU!&JHQUtidNx5LBaR@w6mE4N13>)<Y~gP)b+m><R|!B!83GLn(z_UDCh
zOz`(UvXRzga_;!P@F=Vw)rDgr$_#tS8VUcz!znN+ATu4jv70#Dd6J1chLqq1<3uQx
zz|=3+CX9&xQ59o)@-Y}yJ{4!1boQG;BhQXIT<o^H_qsNWGjTrO>>F9X(m$L&)AIm7
z<A=g$R_30ALOpzDjiR?X9z!Y!31R>RV^*-Sp($v)`0G*%BjiG^qs4BzuT>z&J}?t4
zSTb3x+u?^eD5wp)wgIu%>zWc@jXu$0{<D-x7p(dQZp3CtttKoD4L9dP%oea;>N{>`
zkSdG4tJs-2U~0-G@hqg8PHElRWzsLVY?HruqZ;BY=_Y=-@v3s-vTR%6LNuJBY3G!k
zhwZN$Q^`~dPMzN+wK<|2%C-$S{=It@$TVE6vXztg#T2HBR?Tyb5<4sfJe>P!10)W1
zypUr{faop;Yw|vf8h9EeQscq?wbd3kMV?rE1qj#8nVQ#1zgk6sy2q7eZ!Mm0!JI*A
z*$C~}o>2dvidtZPmlnr!Wbx8ljw_DXs<3K>DUE<<t6*Nac7M~DQD`uzvQY3C%X?#O
zub^)F1>9U~C7L#;GuH=Rz)F`@Tkr2GVsLld(F?`kUr7M+9-_Uz!U<9a#0Lc2(Z0%@
zep<IQZ6tx$<mqZ};J}WvBU&>1NeS#Ca$2oX-+*@C1cn-X1$RJa4RWEhx-Y>r?f%N=
z)4`l72ASpi`xwU-a2Ac0Tgl(Z^`DF(GsZsWQkK>ZvdazzilsO($E6%}Oozm_`kH%_
z78uw>uxd}N>>Bh%?|~!c)-Uvj-vOv`C9&7Wsx<JpN5~eEn}F&Di8VIZ8g)SzmLKFb
zfC1ooMqYu%f{26cpAW6?;Zb3Kph#LQtq3-46daXop9N~-u<ooj4jC7iz*8cWZKGlU
z##JD|qdez(W>{`JLs;k~F&8YGjh$5hrG?u}wW|AA+g4{u7WP)#J`YI8P3AW{e(qc*
z4y2IBZ1-cPk584>J~ZRb<^petzCbF$T1bVh4=|7ro^EjH>BL!iR|2*#kW8;#9ONyJ
z5&yg0p`OGQm>t$aFG)DsR$LfaV@Vqz<PYZ|B*5c!Nb08ny&3I9xWS*2TZ+{xfEC#B
zi1r;JB~M5widXmC9t_|V?3J9sSj<o%f=E*yI7-vd?kvIvB6|GH4J?_UbisI_wjEcj
z{J@-c*|U>$J<8Wy;-dQ3=Ngx<?-6iNYFzPce0gQ8ndReDatrJO2}rN3+idCD#r!sw
z4m3RjG&e{wCX^n9S8k(GhT1xIimi;eT50&7)WREEpD{T*#vIZe9els#8jl!4Ift4l
zSxFy^q_X|{8oXz14Nk<IXZxVOx3I7FgwcUdVb4ci+t!|X7*A`D;4PivD$tV(npcb?
zyW7KFDX#ETbFC@_jQ(lyPR6^-8p8nn=p-*}7ublS*5>2g)LRzP^xNTllV4_7iZ~ty
z-q%_Gj~vDoSNB72^<Lj{wAA>aVAethf}H4dO-TDQzcBXg_Ty}23OCAX5!bx3YaZhD
guM(&{r1cE%QxGXPKF69r4*aL8sHIRQXBqT=0K#T@vj6}9


From 1e5302c9a20ce62486c6490accd5387f57c7330e Mon Sep 17 00:00:00 2001
From: dongzhihong <dzhwinter@gmail.com>
Date: Wed, 13 Sep 2017 13:40:14 -0700
Subject: [PATCH 05/94] "redraw the graph"

---
 paddle/framework/multigpu.md | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/paddle/framework/multigpu.md b/paddle/framework/multigpu.md
index c8501725f5..1c843326ee 100644
--- a/paddle/framework/multigpu.md
+++ b/paddle/framework/multigpu.md
@@ -30,13 +30,13 @@ As mentioned above, we summarise that several kinds of operators are needed. Cur
 
 ### Graph Converter
 
-To be compatible with parameter server design doc, the graph converter converts the user defined operation graph into sub-graphs to be executed on different devices.
+To be compatible with [parameter server design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md), the graph converter converts the user defined operation graph into sub-graphs to be executed on different devices.
 
 1. The user-defined operator graph will be partitioned into sub-graph. 
 
 2. Control operators between GPUs will be inserted into the graph.
 
-   *Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, Send, Recv in multiple machines*
+   *Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, [Send, Recv](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md#graph-converter) in multiple machines*
 
    <img src="images/multigpu_before_convert.png" width="300"/>
 
@@ -53,12 +53,12 @@ These two operators need the Multi-GPU context support.
 
 Need to notice that Allreduce operator force GPUs synchronized at that point. Every device only need runs sub-graph in a loop style forever, the whole training process in asynchronous or synchronous mode depends on the Allreduce point in the graph.
 
-For the simplest implement, when each GPU compute the gradient of `W`, followed with a `AllReduce` operator, accumulate the `dW` to full batch of data, then run the optimize process individually and apply the gradient to its `W`.
+As it shown in the picture, when each GPU compute the gradient of `W`, followed with a `AllReduce` operator, accumulate the `dW` to full batch of data, then run the optimize process individually and apply the gradient to its `W`.
 
 In fact, in the way of every GPU optimized full batch of data, wasted (n-1) GPU compute resources. We will enhance it in the next stage.
 
 ### Benefits
 
 - can easily move the optimize sub-graph to parameter server,  multi-GPU feature can be  compatible with distributed support design.
-- easily plug-in with NCCL2 library.
+- easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library.
 - GPU Model parallelism becomes easier to implement. we only need to replace different GPU's sub-graph with different part of the whole graph.

From e0a8b5915dd45d8c732a28d09d1026e78e41e341 Mon Sep 17 00:00:00 2001
From: dongzhihong <dzhwinter@gmail.com>
Date: Wed, 13 Sep 2017 13:40:39 -0700
Subject: [PATCH 06/94] "redo the graph"

---
 .../images/multigpu_before_convert.graffle      | Bin 0 -> 3056 bytes
 .../images/multigpu_before_convert.png          | Bin 0 -> 33557 bytes
 2 files changed, 0 insertions(+), 0 deletions(-)
 create mode 100644 paddle/framework/images/multigpu_before_convert.graffle
 create mode 100644 paddle/framework/images/multigpu_before_convert.png

diff --git a/paddle/framework/images/multigpu_before_convert.graffle b/paddle/framework/images/multigpu_before_convert.graffle
new file mode 100644
index 0000000000000000000000000000000000000000..6c35ab1b21fb76ceae82d3693ed0d085b5bc0855
GIT binary patch
literal 3056
zcmV<M3lH=kiwFP!000030PS5_Q`^WEexCdaU0!b09k6@RnjuV?C4pQ542CdCx$0IU
zwQUVTawIuGQvCOOTJoZ$<pmrdDe-`0_v!AV?(>~Z*3sslx7VKXm4}fV_;1#51Fb3C
zw*$xZdvDgx&)#mBYkzLN+Wgn<@y^+&lRahNxlyc~oNpf<>?ms+&1P#b@OZPid$y~b
z93HgK6d0q~+&fxR*7|Wgc-?H?+}t#nplaB`wIGa|Ct)z)VSIN8gKa>Eh7&t$Fikpo
zjve^caqW2P)o+_u{BEmd$L?2t$nJP};5+<wvnjrXM%RydkB3`mvpGH;cY(QGo5e8X
z$8_XLCl-cGocy*K#UWhs7I4@Iu6?%`2E#$)7|z~>tlRatzS$g+WT>%jH3$XviJ|K{
z$&hR|$3v$q4CA2X{4<Q=$<%Q;<jryWHnXq9l)eMg4f?J<Asb!$h)hsK<F*MK-@l`E
z1>LVJ-!VaCT`^30f1l%5blCA&RPNgjJa;;2q07PUlzzOZB)h&TSv$V-_>^+Ya@q((
z23OP8Omo`48?u`zTT{l4j<f8*f=*Uqd^SBvmoVa=gW!4rf!JaGR~8+IZqJ>H!gwu{
zrOfW+v`5?h#!K|mB`<Xfyy3Oq^4y+ZVhWd;N~qfdX2TE@`j9hki-PbaCu36hAf$Ph
z#U*|S!o~(NH;Aba@*30Elq$a?_$zTI1_{Dq|9$SczH-3?5BUzeZZz=NUE5|}X;uKL
zGKiHoN^){r?E2RA_}Sf{)QtfwinNI=?gsYon)`7Puao5?rS%)ha3oZU1mAg~EC4Bk
zB#NvC#*}K<q?V2~Vy!Ex2HzPLMLITB(=rW9V2y&LP?QIn3~5uS_1S>Sur^+#_>AAi
z<$j;`+7fCnxG9sK@+gtfB6SAH(?!(OJeE_^L)V{y*6cvpk#K&bl=viw+&Hs+;`>C7
zQA4$`VW<WL+^Ddyf|HSik2{|RaU$edxyjCn3cNS-JDk@^rR#|L%*ifbmO!qTsuWXp
zX1T^&&Dg~YEXYcMhkWZ)zB^&u#U18HZt~9_^f1v##}4d(i6#xHufBJ>fgi`0UZ;yX
zUEhVH-DhFMW2At3_{w9jrbm3p|F{QQJMe-Ki2vw>+##CQ>4lv8;^fjBayTYb|LO4C
z0Rw~WOzG{Y2k7kBSg0oiWSs#EolbnK=?MN~(5&L1o8mx-njB0d4wNP&)JjNHH6a1s
zQQBJ^bexdr29Stx-Oy+8Fl1imi!Hv0-CjTDewW#x1<|Wl$;C#LC>^_tl#9;4Eqq^L
zfzl<u03eqUnBT#S)yj9SP!R%r)d~W3<-qR-)w?bzB)dM@w1RLYILX+w2|jTaj>YS)
z&?Q)+OHKm#Dtnm|eVh?Vs+tW;1x;XCszx-^)Ylaa94l(VhFTizF*IrS1>PsUk#NC4
z96$zC9aB^2Yiu+y(M?crSTi*Xmv(Z>$IDk1`<ywgFw8M986PaG$H&dtiAKtt_n3W`
z>)2&plsKMKHO=CfhklM~;_VgG>{z`p*Q?|SEw000n5Auv@P_1?|5rqt*!>kQ(M$-V
zheBAz-~y}2XduhNga8_;Mm4JvvkyRgjoIc)0U)+&0R9jFQtP1rB)TS|7J>{_)hISh
zoZ=CnWmw>Ysk%v2QseR0#G|@29;@SDY{9-mXr}5q(U7K7;ZF(qFu||ZRLeq0w+s>g
z68y8rztmh+{7a1`<6qsRSQ9waG(a6VuXX%e$G<Nq{%zH<@5-EkYW+M@IT;Ui20*bi
zhz0t^0#$8-bb$)j2yLjQ27ZEubOepG${BcwD^WWG&mK@ez|>PDar;Qg?E_P9qIJbw
zp{cKscwHj#FItjzg3#yTDRbOmeu_ka<ZXIw$)3rQIqdp=-mSzXKw}Vad-YPP<H=Bh
zWkF&xQh9<MH?%#8JNzu5s>S{(#ooDCq0kdj_6Y<A3O>QpVL+`hP-9^JO}rICz*G|_
z`srjGX|*?0Lt@pC_?vxodAKw#Dg^tT_?4Ul27%|XLBySWfISUR5g<rI5&(^V#8VOB
zmLx9E*8E&TWi}x{xHO|MH>V)aDoFDRd3?!o3o?ds`Gt(GPz_QFDoexXJ7@Z0X+T=c
zG~{=pc<+xGM#IplNeoIY6<h`VzO*)ln}zzE4%wW7T>HtEsH6c4L={@FS`QYMvO*V1
zXOZR^7RfYJ<{FCGhFQsj%Je}|I7*U|zX`FNAR4m-F?U{^wP<=JL03s+s*LeG^-yh)
z=YfSt!BTKNXF8uRN0uc!pSP|Oo{AK;1%7N`wgg0KbGL|=nS)^ZICF<FTEu!SZ69lM
z_Y;~sQ+=?xtFgQSEUS;ivS~E5bp;c%L25MDXvQ^~pA*g63Ifp@%Qco0VbXtGAZivx
zOnrtxRIdZkIuLzaAo}USf#_U_5NbSsuEvl<<mb}eRX0_nw^eFfy#!o+7<zSz=aG)&
zY+rm1@XeKkh0EO2SU8fboynguFeG(g_*gtY<_@LxI6M<w72sSZX}horsPX(W;+fWX
zenvc-=!tK^)M&16!Ym!fVl@$(myql+ah>d`gVIL@r9URuGs4JA$o80|&i2&6dU{|z
zCfPF!T5uTxS});8ZziqHPmPjQK8dj+j2P&J07R_;QUm01&t?1&farK#(WxweQIkh7
zbhBYl9UDZ~k$C=FeKN#o5BAnp;zwyO*pG^C+$fSk;1?8sqNd{i?&Mr)DIxQ(&@3QK
z7yHCxHWw9wrSerwFy$UME0n5`K7FU^vg#WjSlDy@s)v&jfwFQ>xh1DWD}N^;!=~0C
zIWXqxkQBIfqxSH!CxDmXW?WJ{TOvDGhWuUVI@v5_zPh+*E!3YL^EPE=v=ILUp&;K8
z(p6RYo0T8^LJeGg=<8BKSElMML=`SX6~R?X9!!hE-epAvo7`2G7HJ|jxvJE;^5pM^
zuEV3lz`o*6u?%+3B4?;Sx{-TXF1DSohn1%NPFRmer83#HR3-c18VGm5WW+pNpeXr&
zIUk$_`I6t0Y+#x&6qWd+xdpeyh0iM(lu{nCYsuCm(<l<*If0Lh)(mTVuFa#x*!bx3
zo4LbQ;S9J9hH*90TE2Uosoh>xKuNUP;420nTz@xH)_kHd3b0IQ+Z?nl%at7)7ac<d
zdM{yf&+lbRo{f2Ng1E$tFG|LXI~|AI@5TMn1O%mj#4-_~U_%NMCjFOSFun-}?n3l=
zumOu#Rp=>$?E~xiJYI%A0`#=|oW}*hpP=GH;JDq<J11sU>8EH^rC+t=Rr=F`c&SYP
zZs7Mm_uYO_uG4=M#C-9KISyU0qPgXjzv^L34Z^bU;6f{H!wy|V3(yK6S3~Bgy&c4H
zaGl!VsU&3V;%L0C7qB=F=+U`~^5C5+t8|9_AauV4KJ$QY*Dr`%c_KS-gf5p%wB3ro
zIs5OuPZt>Rzqh^52m3)w{cHQLgZ5s()jvK*8os=E@7cW_`0N~?<Inxe{f}M?+IRM}
zvorGk%jb(D)H=A>Yqh@s&D&e=X6M~(_78bCe$4U+eKKqO6L9v!+`UH0@a6i8V^3Je
zq>(5G-9W@NktjE1Soum~l1YVFoO1C>7%q|3pJ884h3oqxH|8uD{D|T6IyKpdW1kC5
zSNH;_oH@t7cQ+sGr@_sl??nO9g<gu1=;JdtF6*WA<t;hqA?<MuVM)S<2qaRMD}BAt
z0v=b=Sr}#J+kDBs=f3-8$ftqFjH^G7_?{oT;jCr<D0lwK+WW%wo#3Wg$L<Wn5LO~t
z$+B~c$}j%e8wHvFOP4s~TTSiZ=$8#))~(B=xsfV<HkVYH!v+SGV5wO0@}hP>&v-W2
z8MmJfJswsPzz$uqdCeEDlXYvRMf)Td52$N<>~fzAco$p~N(-LM$Yh@KUQT&U)?q`V
zWL?>yYFb(sV!#HXhN+p@pcFpABB_Sapqi-}h+>GgHBGWi*%u&9yd;fS;?EE-M=%Uf
yaRuH)r;p4VmbS05#H0jKzb!OhhKQovAC*VDWo<M)CNB%ydi8&Ll8F|>YybfN&izdQ

literal 0
HcmV?d00001

diff --git a/paddle/framework/images/multigpu_before_convert.png b/paddle/framework/images/multigpu_before_convert.png
new file mode 100644
index 0000000000000000000000000000000000000000..9c8f7711165d80a2fa3911280fdee91855a401b1
GIT binary patch
literal 33557
zcmeFZbyQW+`#wlq;8K#3m+tQFkZyzSl8#Gv!<7c<MiiBhkVfKy(j`bYNQfW^C@9M8
z>*q7y-^{EvYt5gt){JYh<eq!ZKKtym_xnEY^E?}bfu1@c9t|EE8XBRdhN=-78u|<x
z8b%lt103NWBd`Vkp~H>TmC@=aVTa%YuD6CM91RVh4E2MKR#;30PWa^Vz{Jl)S4Z03
z%Y)z6!OPB(Kgh!yoQ;Mi8zc>W^>FmFWeoCg_k>Fa$ua*sLmK>!IxN7<`0o@yH#ue#
zT?0lHFJDJSF@8~gL1uY8Mn*<iUk4{?BUQEkUJgFVF+cY6^OhD62n-D54;10|@^uyv
zl9G}V5EK>=7UlzI@WF#U{cMByJmD<=Y2<%zN7WH-@9X02=i=qbh-%l?&dc9Vj+q&C
zqyPE$pMLteIQ_3XdBXquSl|H#P~Qj$@e2z4Put*9S=3Q!6)z8OUq?6`v@b6v`|r&E
z)3N_`pZ}b1@Yu`G3ygxVi@m0&pQA6h*v}T#J9**%Uj2W6#{ab~eP0(x@YMgV7W(hi
z|HothJzrJ;_2B=9f%wlT|2qnXSsqVT;D273JYK_&@ERJLBATYE(t{xMuO-2Kx(|O2
zZ9JK-l2T>mDkO+k%f#b)g551j%7m*Ko5iVINPw$JmQ$5gk)*~3S0yKXUKOm@JGFBC
zZvWlMyQ%QWcR!2aKjnY^zUlAy{nfVa+q!d!|GZUbUCH0SzrRHu>O-OEOpFj^C7XQ%
zwHZcbMbzh01_&#><xX4_F)_v#6Jr4_{XC84zY_~;#26vqIFtx|Wc9RRUgO^dBNlb2
zM&cBMpc;D#i7jfR7~r_Zo-5(M#*r$D!V<EzVp*U$Wkq3Qw|_SwrpL(Y*s`t5{CAtI
z4l_(tThXV)8RLXRZmF!O+hmM?qCs5)jZ2Jj@C-6>V@2I2%AtN9)iq^CS<ugc0W%g{
zbZ{JW8T<e5^8eRHIjL_${MKW=yTfrTcZfdJJ(KyK@}Q2&zZqe~6s4-n&s36O=)*k!
z$PkMbpOW8wv0>|S!8&ryf8kTDDLl_IKr=r2<lpH?<;^#}e}8Sp5;=|M8+Q5~V&PNF
zy)?A1E(!3(-oV|uOF`X}ia9|m;VHYDq2e@I9K2WH{A>P=^uOtT^`x)Xw0Sy!lF#fx
ztGlWCS2Hen<nQCQ$lKo=>I~r*9&KxTJn~<3Q0r(@1OpSBKygq0@72!xCwHH@{r)j+
zupD{)IWLFD*v{mXn%18KHUCd(qp=BTYE2q3Ud4Z{^!k0Vx)FK%r{+!1Yl8|T=qopE
z{l=KeAJBWyV$$;V`{Xy~Z6{HfcNYB@ERSd5w>RhdReZNvo4>k6=SskmL1&Dr@_kOR
zw!#5#^V+zBiO%?ppHFc!UE!}ui|gGo7m8_a(s)Gz3=KX9N~|9v7R3~>kM@f*ju#}G
zJJ`K_e|~QC-b<!<u9YKwHb~ZZDC6E-G0L$*ANsZN_emE+*lGXOors@nfvcUTNPMq%
zmyak5>&)Ald~UAy%evSmvDh9nTXy;(c&5JYuXe!o<u2#iTWgHpoGiDv?R-f6*0E1t
zw=U`ZeV60>L=v&Y0Y7(07<7VR$c6r}XmwxWC<{aOT1WozU+eVuA@dVwd%5&XJ)I%)
z=B&Cb;@9Z_QMGQ7xaVqn`yN+>{Pk}35co=DN_OW*Zly1F8{^mR^T}uDfB*dMJ^Mf_
zu^wcPUGh@X0mPWp@ds)JBoAf`Z*Pv<HX=?<-#(s^Ica}pKa^%2aqS3O3H|m?;v@sU
zH;&-ZKO^IypVPkyA~l!0nzBBculv@0t{2iLaW>47Ui+wyW?exqF_lKFP}mWV`S-L6
z(R0tQoC@GQU{P8`DGmwIziu~w7jk*B)|E~zqW0R8*WfyUub{|F#FMReEKfk;pgflQ
zQSE_Iy3p2Mkr)5{7b)i37QCiS4%>r>BxL*dn21WT$nyya6U;&hufKEm_1`~Kng4k>
zixd0dJ*9O@@@Jkm-NM1&9ey#$he=`@7VHJC`BS=zJrSSB+^wC(3~-i~J?+CZZ*gsL
znxGkcdp}~)Hc<!`60)lJt3UyS?7b-q^Kq^wqK7lKn@i4R_co;?!TT(`pR_sfXpA9|
zSE@Ye^--qz(^YN$j<|$;L}qmm^9M26tvE8O&)^LRm^QI$xqW*5@a(GJb$mpxnw|Sp
zXb3j@>OM#9ICvdG@%bczLeElI3a=5_(^jF!pC7znvbV+biTJW1tJ*gogdZ&V@OoT$
zpM6Q_H%L$<R*Px~S_~TKqn|7#;mzir9r4PoGpF%>4a#Cp`xEa8gGa5XwQ9_j;e(t=
zpC4bBEKp1CaTAyb{yg4FoJ9N12vRn|$BcQIe;e02JH?BF(p27eee3xRMoSSJ8Mt2(
zxDHn4Vh>!p0%HKIKK=lx<^ElakMMKXP7g<l<1D$2w)GHv>}l`DFO~ig7c=Vo3Q$$N
zYd98v+v^9fwoTBAT{S#v>nE?Zt04@SRi|;dWGB715nt=_tRsRegZt?)Xir)_R;C?%
z9<ew)HBWos|E2{0sa*=w@FR6X0qY*w*Y)t^<k1`+9s(<vUGSwyM91#kbGvy8$TVEw
z-G$Md&JgLZ!|F0C_=H$mYf(%*SS-DNrWpJi-r7Y<9#&=B9od_-A5<pm{rzU$6>w}7
zcJ@9-{)Xf%NUQv@i1f#ev)_p0iFY4Tl09WW@aqZx?YaE+a1}lkkGJ~1eCwp^TiUaJ
zkLSs<N+3(?voD8;`|xdM!-}|84B{m0vhl8S3VBv>69}l05j2H+J-<)J^8Tt2$?b~w
z%SX9T6&&Txy}FN+DeeE|K^yCjO8EX{zQ8HI=j;pJl~xa*8zcpZInzm})w_`4$FREw
z^*Dqyaj6)iOYj}(`WF#<cQJmvH^16SWPt7-gt_m1dVnw`-aIP(^Ohw|GDaXUJ<8_g
zdzz?Ecj0kE>{t&6XtZmIejL1Pa2%t!Sc-9F&u#W}p*;%8G*M?Wycd;vAGvzk(Yjw6
z>Pp1-x*4Y1^W%D7fN#B_&2CY*MpG%Lkwyk156=bH;>&r%nX$nX_Y?!f!F;%^gQ6N>
zj9q$wX;!DAIZ?<W62ygx{|8@O^X*358Lvh&eiUYi&`cU2Q^%IH9;1*|x0J%74LunP
zQ&Y8lK*H#o(JX&WNpD&_N%~%3MKqTUu3MD3IvcBU^zpei;$js2QCjEsp_qV2g16rj
zcRx>=Zr2`a<#78f=ST{u?r$5oZ<{WzHnZYY#-n*IpYKhu)W@S5lta^(r{MTVVO8=T
zp&-VMkvx07Gd|{0U6Rmd>)AhDkhHq(HGKe`*UV;rck_MBtLK`1=$)imJe5BM&we`|
zv)`XzmxJ66Gb}F<3fk^@_w}Sc7sDsF!t$cI)s!q=u7*J*-TzF=e3SiiF$>a-^w1lr
z2Hn+=v>#I1b0jSY3&t?Ij_!bCGo+jM<0AWLQC3A#bv&up8H=n*S&JehSshA@?)S@7
zSC|=jwW;4zm}<(?#yNG~x#b@yo>X?ohNCZ~E^FtjGEb4_e5fnOXUgEmfU&SD3u%)(
z<~Gni##H^vDrwmH$Dr(d1Eg+MDJq2sS#H6j>^?P^+jx1++rq)o)yT6S4M*uw5yh3U
zwq3*G8D<%*N$81oBrW=~XAqj`bk&cv;U7i*h~U{@FLrc}l0gjIaqC8mDREPwB>Msj
zej06%8wV)%_qOx)^$cSTzJ0C+eP7MJ54z6S3b@|)3aNYkTqDi8gM6^%g1nS(Q({O`
zNvIb>RX&Y7L-wZ1(l5)E<MHCyki4`o@~=c95KJLhEVgKdgQsGj*)-p8`zg7fF!avh
zvv@;Z(I*Jnv80r-gK6UBcA8;n$gt1Q&xpg1&k>F~5#PNJoEa1B4H6R2HgIyrFX@bA
zq;I}9PkmW<6FFz7-dK=9_Jlfbrg2C~Qx7^_p1^Vn;w@QF>8`JokeU#uiVB*;75-C)
z`re@B8CqLg&0R6SgXw9%#Hw~dX;DSo3)Z@F<c%sCj<?arOqznl(~f&+7>u#xl^MH~
zY>mRk`_qan+mS)+84oM8^Y~en4(993ooIU^YzGgf(b(e9I6?yLCe?efjT9)D6zNlU
z(qz)~LwR2_1(8&#VS|R;D>5LquXvojjfEvJr`=tVTUf<7Nx2hN6dkSvOmteO`{B2L
zepV3T5msn9>_VdV%OZPBo1LnvcBVo(e<^DE`z|RF#VM8c+H3CpO0G72x5amKmAWH%
zw8t$Gr`ja0cvt?|V7QT4<s~UK6HC<Rr=piIIAaNXScOs=#;QcMk@QS0mHTbePrR&-
z^~R=Jc6#xb+1)SI(_k*Vhtt^clAW~L6;S*;(KFA~v**%OM5~tZJfFIUBnXb<Q}s)Z
z=jtjnda%LmXObqxu6;D+l0zSr$%hl{<KZ=a@r{`sIdqe3VV)9z&<AZs|2STO|C%|4
z`4rz}Z&mp`n!~h6xy}X;kDjU3lbq9NGN-rzX~l)Y5z9a%{Fh%ERKjzWY83xv<+cWs
zq63krg)C(8wQ!kG(#>wRJ|k6AbM+%(iBN)xFX+y-`26ire8!Qf21+6$N6`ddg7B3h
zr|n<et}+XT{&0717jYUN?nsUagiQx9Q8h8Ght^byvB9&Ec0NqV<i}6aWxpr+%cjuD
z<g&N|h}n?lOY)Tt{fGa`TNe=k3#=N~%?k|3V$N)@e)irH;xA3J6ciDLd}#SxPWk2d
zS-QL{%yAuBC}@j3-5S_Jn)`ohB0x{0dh7b>wPjbpU~zNGMdUEs?)oQVYuy`Ss*(JN
zk=2w6k$R?31NTC?AB*;JTEc0PUJw$M?`J~{`;BuEg%W+dHSwAk0AjXZ&YSZ+lS!~h
zO2nc<aA|)tI)l#+jO{K%Es!ji-!?WJnBPqY0B}P_mS~V+#a)xe2PGcQFrXdF<r|Zf
zZe}XBGiNMTPzrk2)&Bm&R8Fbeu&k3dwpCnM)`s7#?oxn*VrY9*9%Gs%O`pZzEN=ku
zg_>LOE%t~gY#;%`McN~)<K7}Qcg^*SbE$<LpwSl-mAovjn|7z0{a;B2jMeB~X8J%I
ze!q2@W#5~V92I7Us@ZTf+xKU6P%C+wuym6r&BZ6z*!nr@-O;@NV<R$>_ML$6zEmfl
z)VZ<^OF&>UIYQ&3IaM@W{wSkHwnP0T51nI1Gs3~Pp){M0dNYhhGra1{Nz#*?UX6Bj
z8$wbc6U}gR)qy)Y{buMt@p*H{`G{%EJvYB3MdJ!km_vVPaGa>V>?{cH35F?uLMqv8
z_C*8ii051@s2<H_q8+F@vg$`7QXlbvQdBXUS&+6Gw)28p0GAn4w1Q+x`l$B6hv+^P
z^$XpiTne|W4wrsM3rFSYEQ7eg=N}CBaFkB50#6UroxiY1KvSrMDxIfFJE&!@3=RD?
zZ42fl|8v1-0Ij99%$*TeNIF^e+?qi=NYC{iZH(`|ha5I|NsZc%MHesR=}>D<xSsvD
z_ArZ{2eo=ZB&45C)|F#IK~?4W5a39*K}%5!s%#7#x4{9HUlAX>H1YpUT6HEG2v=2_
z>~{^B;fW-z_E>eviI*)jpQ+GE^T}q|hg2c5vW8wZv%eYtbB8@pVg(J+sON4#`dNy9
zt=0U(heek~{=I(pbKPyaTxpZ#vmY9j!x*}goDoyO5eX+*O~s#*fMpkUqdgv7R-OA*
zLe?aNubKnbgA=PhzfYzla_NQ<V#bQO&Yea)uoRY1C(F5P07c*tG9F)F`mm}tRQO#i
znZB&gOy#{(H|dH>FLg-J{qH@ijl6}u2I1XBy{vE#OUU{D)c&#7M-g-CQymFeB-c%d
zXT=jW#=Zpk=)n|5HXCLeGAFq^%%~piysk5A8LgaSH$~!9pB6cCxY7$_=$O8h2(py5
zo&Ejm>A1g7^VgW&sol|*lmnfS`1=gmmd;x7%g>|H4pn>q{4nP+(s|MU5dEPoXw%6c
zXPgi1-cGVgc83G=5-9tO>n*5pG@G3!et>Zo=)DoHmrW`InYf(haarvY02L%PGSh~3
zGAO8yD1j;<%Rz}?{FSf1EbP1H=!i0@@yVEeNP=m)YqUeXK>mI?S(ZZLv;B2Y1W85=
z-IHg-`GPOd-8XuNgNV2o>2=3oKBIt{c==ZfOGecKz+U_XCKB$8o9n+o0Fa8b=#2;u
zb6Hc46zCQ(9I=SZBzhx}YbZGs3OW++&PIkfy3&<@S#Q}j4}QXb|JV}j237nGNNrmw
z&ke(r7~ftNXdK59$tP`(kotQu#f&s0Qhd$5ir5NAXzl&F6QM!Z^rB$T8=GZ;>yG(a
zS0KkW0m~g)arcQrF&6bS)>omDkUp-P1d&{|GpuOh`Tg*pYl*jBpeQgYzQX8oVDKGg
z0=F#@b{qwjlk(P=Pc_$%Sg3T%Lnqhs9&1dc;3x9mC^N>19cMk6SA1~ursvm_;e#Rx
zuakv08!Zb~=5=ObZOYS>>y&=ePhUo+Lm}_vO`kZhvUp!DHrgM}y@F3l?E~l~!0*A?
zQ9H$N*||BdD^}$_{R-oqz)VursCVOJ8NU4Kfb5n^<?PX$-iU+g3Yfl4^LCod36#uL
zGJZo)ea^&``E<pr7i(Wu4<^ubdwDPutHV4!A}&yh*+{}tvMP1Zt6L}<eB83I7y<od
z8*OUx1EcW5LV&Zt1joMq{5b=S&tnQekbL$g+3@TBq(Yk)y(tRY@z2+}7g{U#bCvAp
ztEeE>9`rO*z=ss9L~hS3gGnkK@HMHA4fYsHGHXkbYZ+o>p|SS@ZRo&!z1Q)^E>=RS
z2nooG9qZm-7mNDS46o@&48*=cA)WFd-9gztkL3j(F1fvV_hT*4`e!wKD~Ri@=bAKX
zHm>Krbnqm;N}O>ncaz4#TnxCsN_dd_6N`=*{foFI1jDk-p(CNxW2Lp@0g0_tTPYvv
zf_fRZT+58WEM2~_q{*iiU|@92Yb@^&EAnoa-kt6nq6>t{i-3YDYej)#W!3EF9csP>
z&hvI{t|tFvS<gMoRXhPa4EZE_MqK8@3?__m%H~|#9b}>H!#=H*u0SssQn0FNXu3ir
zu{bQc<7tR>WmX6O<6KKdxKGx<l}LD84u1i8Z7GWZQohPm>gC3hH<@eS)aircMBRdT
zz0v5@g<!sNp0<t>?%(GdaGM*7Z+@gb&+3rbww)5VNNRCgNYD|uBi%1|k1Htp{$U5#
zsn98f*}iQe9mOibBfxJcogrOitkeJKijpA#)Tk$gZuN+GjNY~0+rLH6`WoYvESlxt
z;p72^)D91M0M>Zny2@;#l0k;O_h=Gvr&3yZ&mVQhI5xAlj}5;PRFow0LZ1b0gy)g@
zV6CO{<c+6d4i=W(<#`l0zj*<LM=2r{U3{;^61feHI5Aut^*P^4Wb;(zkf;bR?7rfS
zYn9a`&*gC0z$<Xa_tM^DFA*JCt!h{_4Epk_I%EpzfTKO~GEQAmH$dG{jU@`Vbf@8H
z;f+M$Xe}aEX?3n)Kjf>X@B)hg`y&~^f1KqY0({lzmjUjh*V8RVn|Vr#&Z`VFWi<&J
zHJjgvQ<|?1)E67ENUkm(Zgs!k4b}|BxdGLoH%_{@naRC136khXb!PXh_DGFs^2EL!
zXzzs#jSd#-w9;y=s{U$vf#As97eDfxQ}QHcoM+~@0A$M4#A$yzo*Qgb^7GUC-9&5!
zM|3=7Zr_i^Lw|hTykokT>wFc{b*l1ZH}{A$b&^to$%#JvVun-b?ihY;e|g{lhDRXM
zS+W)X{0+iE=|RQohc#yE!HSEd^03=eKq1JK|AjB|CCcE^eec60+ReZwEMdd24m-no
zHSFW6lkz*qaZ1WdkIH&791wVweup1lA(S}pry9J3<z|-_(eU7(WC`zs(s$Cs;5nmB
z3`^F1;ymf3$xqIYp*G20GnH6Uqsp{GZ=-Eq2cPeOVqG$X6~2Sk_>E)Vn?~F{X|W~!
z;wd?c_ZE-1n(OG7P8WU*2XDq(`$3qy4bBU@?E68*XiF5b7><5oO<zm=K58;VT-iz%
zOsVwQ0X+5a^@FWpFVsAy(o=z?_C3$DdLmE7hQF99v995%XeS>O%|GADSa5_Tg#TsL
zdw@{&6ei(WjP*8z@eD*W4z#0krC8ayhj;ObufXS}SSEfYQM<b&2We>1`H$sXr9es1
ziiJO4LrMDk;!tIA@A`Z{^p$&1Nr8wn5+6pbS3P{5hSqclNG_+|g;26=Hzf3H5P7>#
zipJgtbh{&cqY46g*>(Ur>9f;m#yBnlPgV<3DP<{A*o@*FqtIr5q1dUu5?Gm!MKvHF
zo}xoVGsh<gB*@Ht0-~Z4eV9~rSw-1&c^!=$$j+aBJt1S)(J(cZ@KCl{una5_bs?p(
z^S)+Lj*8voZ4mU?R?TX3l7v~+2mB!^qY}YeUA`9}Iaf(*C!mr8b(Y*vkM=!C#1F9o
z%wXdX{YsT)Hei^p<ujzm?_vNB(#Z@FR>p(TL`~VmL{IVY7!n*mryMY25aOBeja#JX
z7%b(mE0fn+<$wK-moE_<rBV2sb&u+nW3cd{t%^t`u5PFLHiDO(-$!|Z>x5^h-%bRR
zmGZ(|Y?yeoKHgoFdLtm$HqDhU8!Y_8K<_WP*=#W$ggw!TQ=e56&x@J;QQpnX`IP&w
z9T$5_%Jyx$%a|*IVKp`FsfjQ59E0qc$JQc#A3K)@v;U1v$o`UnQ&L<Jw^t=ZYZ;$6
z)d|w=I9{5n10F={DW{Of^4nMUtM9-h1@DPH^AvNNCx}7#;+!p!=0=ZCdXOD<(4X}l
zx2~8lxHp8<mesfd>?Or<mPknZh!Z`TQsBd+VwQ0sxtZx*0(CS7>=cb!qk+RX+5I)e
zn}!t4dpe*?KP(+1Lg&;V)P{Sf8{)ZUk6Bh3m{_{0u87;efjEcDQ5>~*bas>8n361r
z(c`_G@#kfcXXl86s3+i)lN?yu97eMKP8T%pPCisQ>~LXHa2ltQZOh{|&3aVh&8>1*
z+=5dlP*Q!%W3|1JeBG}?q)tAO15_VDFbIkcpM4x#1=uJhdQ~S)TBKP>y@D6-d1%M>
z6CV2D-Aur?11OB-bgjU74V}P@KO{{+CTpu4FSm4LAdC1z)IOmIc)&qlf}zPUxa$@x
z^;U_1^(-J!Ht8a}f%kLZ)x`6Cr>JYd=K*3{@6>DPQY@A}%pHT!07K3o$Fc44Xp$wl
z)%nD{sV~j(-s7i>Iz)hxrL+p_hv+Tw$5Ls0*Z!^MpgnbY{?0MN8j{N1t<mB8S#tG#
z(<D=W%ANl-!^RdQEQu#Tu@2>6875^rFZ$Lr#^0XCbO>_ChwDq&o}4WRB7x|3jJU`f
z^{KcvEi+M#DP8>f_R+yKCbey_b|w}qR_n?mg_(DRrA4pg=K&8z(gld?EaZ_)t8IYB
zchSKUFl*F(Gmcl<5~V~3ggBib&?th)L^bcOkos_r$87Ls3qnA?qf(~nJe^FyE~|3x
zyI_||&Xo5klEbRd!632pTl>zP(t<|&A*X2z-Ek$fvj+eX)~dm@2t51SZ|oFp7}u8w
zw1{2XR)_FMEExLVw(zX_LDdpr-B;VmhNPHwv89J)hE&NUm^5sy=4yjRIx)Ch-Hi>4
zL4Q7K3p6JC4>KiZU~rU&Lpz_t6Ku*Ut;2*QpR&SZ2j-hfN2J6ZhKav#!2IV7|NQ>>
z^Qc|$&-8!jAg%oNa_Y;Y1${iP{CpgVc>r(KCmY}WSU$?q=C>8;%hTnPuHccbGMDD~
z;^%!KBR4c^@fm@Z8g@I7b&moTvm<*xE~25CW>^fW3zYsK|NH9>T_FH1TMtFN*1IJ(
z!r*{i122}Tkxi<s=X|mQ@$9YplJNKb&7u?)nUkOE)P}m4B#WK?JjvM&@UY*_k|hev
zc`xp9N6YL@mW+{pCW<cyn61`G9{KSLYx8JhFf%J$WI(X6F6b6MJ!YEc)%!;r#|w)7
zeuG3E)-Pr|bIoEZ8p6Zicr(V(-zTa4fnnv&uEUP)rby~DyL(T_N2JsmR50p+QbSZ-
zM$jDv*XxC9s79UJ)UOVPJ!rJM=Ztvt(!jQDEdZX&`DO)>yieJ*WDEU|=B19Zz1{#2
zp8x)kcLqRZ)&P17goCH|qFS}{Q|q~h&hMIV6KKYm_r8IV*v{`}2_qTI;eO~#A&2K%
z|6qBp9Pr$*_`I*K#e|tikovk#z84Nre=w~q)=<Ut&OP|L)I0@PMrQ=GbIIk1^5yZV
z=PyCcy%*^PW{W9kcml6aKz%pXyM(RKLY*jEL5H4r_Y<d+y`!h6yN6molU+AU9}Z0o
z&SzvO8JVlyL7z{Im(d;A4$HX2E8GW&gd0^rZ7b}{snq+_G7*yBIxWgqPb`XWozhas
z$qM1tX5Tjx4f@M$Q%q`g@W7@;N&l5u%WIc1qIe_E(0EV`ChJ<IJPuk61Jpz`ESUg{
z<wp-7x06MvDXjFE5`|x2;}W4SQ6Pa3jrC6iTr{zVt_O%XPZIdmD$1yDy+0z0&zmJn
zT2mY{+E-@015VyXo0-BO>8t8f$8Nh`FglMJqt%kGXgZsPN%`)YI+r*;AXm!Muv8@l
zoV12s&8<#mK16Qe_$MG<o)7+UImDvio@D$~V@wpE_lzvpm#OI(DC-OL*Q3-eFm!IS
zw=ZR67mCf{k;|dcNEEmwx2mkpZn=;5=G}c4{Lfh(t>X{U4!c2}D@C|P+J6aZfStc@
z|6s&Z<|T+sN@B)SZn??GRBkvxd?OWS9sE>_k`>eG<&y49P94rxleHw{e^>1YWbuwh
zhVG<F_(mLp__Om^VbUS4=n0y&UB4(~m&a*2C0{Q+I!*Gga)LID`)0F07Keapx+gp|
zQpVVeiP2`0G@dmE1AbBL&DO*h*MKb=rwf(D`$p^4y|sJ7dJv~Om;7jlBx77G;)$va
zKQUt>@fqLJ!y41(y_bg1e3fE8q_6`E!A!BSPmUZ0XE&fmm|=n=5sOyY$RfF_FX$*o
z070B0H$+2EhtrmMr~K46N||Hwe)RKDdO+`AqGi3_l>ir3t*LFXm4wD(6IPw>{q!Bq
z4INo^CwcH&RtYWsXI<MmQeO|xt8_d_(Oqi~8&cbu$_Oq3nxyI$orEV0PaxQudAQ01
zj|#l1DD+1^GgZ*QFk+Ki>BSR6EGgE9C%hX7H21KR--M*IVmL&oY`vcojJOH1ps6}l
z^UcMy(-M|owE5(5LBdEU%ukp~E1|YUCm+^MzJ8$J!2knQZgl0ZVPl9v5JEK?DPxO_
z`*nGwy*05ApMkvrSemHkuaz0OK$nu%*mVmxKQhKkaH(?pfr0a{n2skNhM|+M1HG}i
zZI81X#3`4(@K=si;M>%Tizgl4Brh#Gnx37OdLwa(<SsIJySI>DLyxmM0>-mnzQsaU
z-*WUi27ra}XbQ-$cKDwwMPMk?MZf8@T}p6ISP3IW2&ff@7r5djdS588wPkg1ZO!O1
zWuSV*DvCsEb9c<@D_G)ry--xED{=6gb9!(?72gMduin);s9-d=#9Xa>HBuZr@ANNB
z2S=mfd9``2fj9BqA=Ji5cmY%6H8Z41t?yxwj|<>laU8ZLXbjC|GR9l{_f*^bQ8VV|
z`kX_r*aTb#N-VRX80aKQ%K0kvfJm8<5W@ZNwI~S4fLJXLI9~p_5!p+=F4~Z+`Vg!v
zooD$Ccwl^v>bp+X0yh9I{0P(&_V|U@OJlhv|E57GQwBA7ly*^o{gm;%oY2gcwS*Tw
zBFL5u6>Weu;LCDLP0eY&X)`;X1tRe&&`@8*sY}Ewo<iI4yh^0~fxsqj8Wfu-o5up=
zRnqWOJ_rl~LoGDtvhW{3!OcO^ZpAYd&~A;~b<1V)27>I?NAH<?l(t)R;_i!pYv!|8
z+zn9#%lqF@AeF(aTy}edqSHC_OH;IC$=4gAh`eC`HRSZMV1Np*SSIiUB?u!yUi+!F
zb(vIj_Y*6;E}xyGAP`!98gTd#D?V`aDHmrqP-T7Am52=!*x)SEjdnbw;Oia`7&yB*
zN`orN*EMtw0mt=BD1Do`2{qWD!9;+G1^`pSGUgf~$X-yyaL&umr`~56WZvo?{85vG
zkWTsmg;nVOec^@m4zkbJ@!*~APnI;E>Xk@eKE1D(iF7->(B3nX_F<OMCy~6ER@omf
znC<j;BVT7yNzmA8@ubt(Gm~&+mfRpw+&Z3nMTYlHTi~7F_Yn^5trj;WjlE?t>&i#a
z8wK!Unw_VB+DOPxXcWX}Mv>J)p_{957ej_I3K=>fwg4;%IBCc(z+cmO9wlaiw?iYv
z$E*bnQbWl$X~^UfTfqJ{orb9KOWb9H!(+SD?qbL@Z*q3yQ}E`Cd(ecs0q;g;dPxmW
zBv><-l=_$K5Z?qeyl;Tn2snJ6=j*p!8hUaKm>XZUC)oU`AQ00yL&s!83;G5?(cp92
zUl)h@V6D?gMrU<k>pjsR#V}=xLjL|z^TsM*$q`lpkmcQ{=%TsfglG@^xgfr&V7L<R
z(0RKAea<lk3n4^2fXZ4p5J&Pt1*(c0{r&UU&PUKe5U;Z07ve^8#>H$YRAY&u#9+0S
z2-#!v6Y?0>6L_=wAEQSz(|>vS1Q<U8SfWrb4FpbamxHBdXAsQ6y(zm)K}`bYJ~N{F
z{%WnM2>)=?OjqEyxF=HlizA(blL~-|`;bao^Qg|1?z2>>YgxjO0q1TwQW$3qWu$q+
zQ-1|idca#)15XIdfOs(@5#~D#CN{;2+_xOydk|@5HC~Vv<9SreY2Nx3j1<?3a=nw%
z4qh~jG0Fz{eCu^hz48{Fq_@pe-fjs{XnJ-X@7axt6^N~t?E${2P=vBl6&e?(`1{8(
zCBKEF+dS{~PI+X!sB7N<P$js3+>J@l9vq7r8m|ITTgTu&Et8E5Mm<tHW(-xLKQc2N
zT?W;c`5P;_yr6n6pU6TuQ-Ry0MG`3!u{d}`$rOBcp><$hk{ASyJd5rpC%(=*YM8kY
zFLkJwD1M*~RV9!1{XA;xpm<ez+H~pFf?R2^RSC?c5p&(OqV)%D&G{u!F=fy~fGnSg
z)ShR!srd>*)q{3GU?kt=`5;7WU$Vsthc^$3y(ShU1hEEGz`jvP*Z92Jdw|U?YD~al
zw-jD@n<Eesr^rvzmpcA12sP{$Sy?lSz>x4EvN%%2l}?RiWL3~$o=Ulo8DbF+LR}}2
zcLaA$qVJ)-V1FnqAtoQN`Ug<1nxjU&5q|)hr{pq<QxWL@6RqK0GAnX1fYsec61F6S
zZdgScA9%L?9w_zipW9x4|7_YswMR<m7w*)nih}go9shhKXKe3#a1sv?Yd?XF=G4RS
z8^gaQW<cD&6WEa`Y0b4+%MIj`n6x);w}1bLJPK-24Kz`jmn(+d4gB%2?wKq!1{i{r
z-Xhtwb0IUqlfMTWhV9~qitlIbLx_hqfwEa|`0PWI<JiY|Wo`07lMldiv=CW8drxrx
zk`2Db6$Dm$!5ApBJJ@ErqGYVNy=mm;Yj+-x=L`M*)*A^PvO7}S1tHzcTi~|ls;?|~
z_K8@nxKQ6$!0aASFag6sexsJm$J3#-^?4N?PaxZf6Vt4cUlkA&KP;vjzI*KsXV(4&
zQOsO3Tw0X@J}y+u6|><-^U9MEYO}?*L?qwf?z<kmUy>FdIG9B8l{Yhk*;9jWMpCgD
z(CM1CjpZ57H@bu2_|VJPK?*)Iog(px^wPORA}Db*2=2f%0#=O~(CUxY-^rAavWJ0R
z8{fGhK2&a5Y43F!m;@+g<$$<ks^<4%7x^k{)B(?m@%WDXm5nO*(7Qr$k6++1MLq!8
zJ)|Ft@>J|Nt(xtNGD#@N6v+v?3n0{T0e6rtxw8^a?h&A+xs&hAi4sF1ejRc^>d7b)
zQ^iuc<Kq3)Ka%=aB)tRa{@%a~HwtMr(IU>tm`9Q=((4Z?YKvC<fSGCm7;&Z)Zog&?
zU(yAxMu)cCuVnkeO!l8!)yD%L+7q;`u7-W^luvD1EMz7c1dBsYI;0izF@OX`zIl)f
zA`BLDRDYsu*FkHXEC=sEfpaP1lntMn{-V0rS1Wqr``-lLy8S<-#AwQl_>iP@AEZJg
zaN)_V)mnD_n2^|51>v;mKzLghWFh%}JAG&zqR;mBCSI!2-+rXkX6-4$23S0K)j$^9
z9w9pfb`29?39a|P&mIH%*8sw<MKXte6+SwX4IKc5^ST943Jm2wJAhUAN@X+oft$`B
z1b<XgF#!&|WDv_+g1)PEf{xDvhpXHF3vXZ=1pz{eHwt0^pubH8>hS$XfVv~|dY=OW
zc@4z>Cp>fdOq-M#V2Qzi-|Jl}wMc@PX`|iy!BXq+%ZHxF0G1=#wzWWn@7zgOkgpe<
z26HzW>E^hPGT#9%gHHPwigf|2@k8!`8I&rf4^_+lSn(PnIltOTAIix;2U)zhLog<b
zOiyxaOrW<|!pn+R3<yi#0rwL;E;KYq@x=_9AJsR)QHIj9KNtZ%uve#4w&44sGQ6G*
zK2d<r!np)!=te<VlxuMA^c=8N8^Cq7ip<uNcKI^*>O{ZiJDBlb{?5e{pOQp_3iVCM
z<rg-sr;^};4X_5=OU>In<p53RhtYNeaLwTB>2@FcwkrKHc|Z~RN7h^S(y!S8I@41#
zpE9e%w$fwBgx8K2caaR{+Y6e@vkE~VO};2xZN`EHwamBF9efUA<XTxLF-8QF_lifG
zt?xY-kgnYUVgrf|Y;JLJvPjJB$N7}PRI$`7-(bUZ<-KIP_fI<5fKE07-pE=RN+tvh
z9%{w4v5w<|_&n2UUkjg-qpY)Ce(Q3l16V&+fo>N5f!cXd>D|i|VvG$?LuN3kgj^h)
z0h?)4Bnq9a1MM2nDC)e)z($F)VRDkU1iTm3`epRh%!Bh&*DVg7lhu8a7*|OHS)2TJ
zZxev^Zg7mE!KXNhaf~Dgz`z+bfv&H@c|<qPt8PCraDOB51tdd`fb|g$@Ny~OCr&J9
z`B{PGh=k?9mc(8km9tTv#18m=tLvvxXp^dHDN}>bj+!LDbx-%$PY@2?=f?h??T|dG
zZGJ*2An!Oz9#P3$PV%k0xg)Di19tg~3!;dojbWAH{@3oU+}DNy5Xf?9axDrz7~Ox~
z)ym@ChmO;S1hzLpM#*mXjh;V=3=v5?HIHX1DJ}g-tex`1&u9mTNsO6JoIUeuI=o^f
z^DjhnAPuY*z`e!`p0P^n1OVqApv31@0qVGG;Xf~G)n}-p@%zZ<07?4~DUzYUIVaEz
zTqK0F;<Z3f1inM~kql*$d(Yx3a+Tj0=s(PUjhGm`+aSVujAXHCI-UUfRwm-+3QWg9
zu$BNRG?Z$a_~925<VF##2IJe`^>?B65?5dW2*LdirlT>P5aB!BCcu3PqH%Be=T7&f
zCZC-T6~fH};=o@sNUzM3IZjCEI);rbjd~bup9wOy9&5?k?ncCK!pIT8l{C!E0IThw
zG&c#x63Z!YiEidbAp!vO?>9`zL7COq1CC%i`!^0o1!Mdcz{FDDzK!SJ59T3_T!^0?
zQAN&I^1xA%x6vZ>P`5j|m;aF=KJapQ3^R}MVWdo|p0c*LwpRK5dYBK7_+!T!`RmMQ
zw1LBBj2lS7;zjn4N3VY5@>@!mb7}Mov^}adiDsX~R|$QLzQypz(^O3sb7ucOj3x@L
z%b{@{zrZlANGh_R<+3ry9bZ1?RsBCleLN#$fjLyMlkcA?nFIt|hW-b}|NQ#Oppxc;
zFs^}Us|dYH0x>z=<~0w9)p3eUAmCD9WO#z0%m@hs9L$2c;HMlAah1!7L0ra1*JuN=
zn;5`tpfj-Kf6+o&qYa=K(ut4$b1y|B$S{Z6oT`z;*%0D};r5zEcM|Y>GjFVW0fo!v
z@m%@8I}e1Fq6X7cP1wsl-0ioSi1(X;b+<9TGwq&YYrKi5AFA8Nkl{PjMf;gffCys}
zm;@ktWN))2@1};MpePc~pB5j*0L|#c&6z(&2#NwQYu4_)#gqu!ruf1Po_AGM=K(W#
zq-ZFy#Py6iLmeO?fMlYj2II)u2hmk)^$qYvV^MN7D#o~X0aUn>g0N!BV1a0O!BK{m
zKKREfO<+3M2Z3W`Eb1*=_AIIZKKKP3)e7oWOQN+7^}5Y$X7P%8#6jQ!Lzi^<UysAc
z)<cf!TxMTn21KKcYM8gqfQTc|Bw9dRD9*82CG+xz;^RG?%Xq~e##2y$pghl<%~i$+
zAD@o_@tF)BGgoIW(RQs>TSPEoar8n5f*;*Sas^_20uh5Ae-l8gL^PrY<04b4HH2Yv
zH&kG>Zt|UdxFFq3ik@qG6Ghp(FgD6$^x#S0CgH`k@Y)Q1%_g@%rt_K=QkjZ2AB%EW
zN8`{luIvCQ{P&`L`ukte_lkaIbYzbcQhO6}K*p7A00|0+1>)A_Y<>A)AZ`=Vhzch6
z3>&hnDDS>_K!A(x1syLC*#{}%jo~xw{lK$}kga%1R_}LAc@OyxJ52klKPdM}L!c@m
zwn)f0^orI-Eod9T5wPMRan#<BXES0Ey07;XKl+0B>S9)3h$V$q;^ghD-T-E8#GhY>
zEyf-S#2Fs_MC9-poRfp=9ZFnay1>t85#9rcV;)$2Kw67#w~kW{`g^^vLfmdvscD>k
z0d!6i0}V8kp_~q+T8imHF$56(<u{`+n@-rjNnZg&?x~{UbGB_tr#O%?J5>>+6&2-d
zh2*Y>|FB^L2uDB<X^k>a(1^Q#b@{A1?9d*_2mo!dh$Df6U)pzf7_hs5u;AB;#-B#r
zw#dQL$b^@3x9>N|A<pO(5gDj=1sv^*>7~r;IBNNzau9d<SOuyT;9Rk8a==ZF+v$+9
zHP;rE`l^#<8L4_~(P)P?zId<o+gui5e0#9$gPx!6W{HaoTjUYCEaRLZeK^K(C~GZ?
z1BjnnDJqjAv!wpjX#eCYLuNUN>zd&kqsmkQYT+u|frLfriE43sysbMpMv}TI^;te;
zByLxr#t1xKzAJ(RDO4?;K9z0{RDp^k6H-Os8LAaS9|LIo3^;UiL9Wv+5O$n~03pN^
zXXenL!|%|@iknqBZq{%4^F2$^7t|t74GOI`{TS{o4ueA8Di5_qN6h?Mtj+ly%#(_^
zDv!?&!@-^ivNQ*<JD`~BNsY~dK9J#pqFJtDWV#NBqRr@0SRq|-tNXD}D;t$AG}cK(
zby)i%MPza~VOe>_sX;ZfyWXQ;z6KZ$whQ~b13*gN^g(1at)<`nwS{4|sUg1F-eJ|^
z`VjCFh&vHy?<vXDWl!oy6wbpCA4XS8@~c^w-}K+pp${<t*#%TJ#^r#K`3dra1Ytcu
z-FUYXKv;VL7JrBGhg_cm86y!JHO8$)ltGRiPN3CF%phN)S}N+%qCy(txzyzNa}}<?
z1Rt;_-AYZ;zG2I=h)LW-835K_uB-X_I=8sY?tx7R^s+&M`_H;t#@?8PeIhx-PmBY;
zl{^(ffu1Yo>!I)XF~dGJ@$?;Be*{$A=4+Xv?_yAfS_2yHuetl^ZTIzAL1L@nGc})^
zNWfFSFmS@A1B^iTUZLBt`CkA?ND>m-t-H%uy(jXEYhMDRGoHLX?KzhjXaHnV#yM|V
z)b<WS)|{PNgM0Zr5WMes%|_Uy)WVuyr*ulBb1Wn;zkC|_@f`%hM)ddSi5Vl)-u<Y3
zVzAbZ!ve|M@Lw1G8c%Gaaj2M|+~r9q3-f%bDaBnmD3yw;X;9l-Zdeh4qZEcc*Wg*V
z6H@Lz?u$YP))Q*#kp^TQ_e1Vc*?l&ztPaDX0AFh&m}G{%2ERSjTI65)HV5X0JUw!&
z)*Zv~wJh0=D_|Cq@jLK7s7UlJj&8eSI&9YFNtdisx|zWN7+W=uN|gHN+wy_EnfrRG
zEIN!Cp4rTAd(cg3;f+qxu$uMQ`t&YH)a1pQjjWrAQPj}}+iCCEmQ(hn0zntd(ZY&-
zvEKrBT)t0<z)$-$sSi5rXGK6l3olQV#p-7Lkj|j5m>ZVS;ZC(3o?@p>b6!*5`6R_5
zKlYCF0i#H@=7Z~GkQRG`Xl7{ds7Yda3zC0hhIRtSehMQGHAxWfW?z($!(xLitD@*c
zoiO!?X^xL_8!#nlnDaMaY;g=7niP;UA?I2@e2^ld#4DdvCC4L)zK?>8$PxoMh$wft
zA8@|atmk%L>%1$K{D%LwDKVi<VMV=znZ4cT7NiVeS>Kno;GG3U`P3^hEyk*tJ2S@&
zJd)#3LQCgal{!1|J1OAxuJ8tY1pHuEidsJGN@~AYh28AylqOYkb5b6jfI?G7j(~OP
zR9d&N$%>SH^nJb+<xS#5Su|H5=>4%UrTBiZL|_aoDS*;l16gay&rZ)=nbtb^b58!k
zjE9thS`Z{f_S&XS+!=4Yiq-l@(uC?p;OoV!4+~F*BYJ+q^b*@&pd?s#o~@6>lSw`4
z-+@&mYb$fM6J<=R)+-s#wBPwX6=*V7ff`Qy_)ALnk~QE=D7I^o2tYT)GWbaH(5w3_
z!T<aWZ16%+C?*=y9zrZOtc`LXiu~`}VD&fgnF#5YCJk;Jfc+0GYyD(wbk&wRA4`M8
zpOJ3$B>gadWyI!Zu4g$_B9jM?B5illN%MK2F(u}yCCrla6c08&^hUmUZ||+LCWDZj
zD%j(IQmtS?6HL6Bq~lOYa(gw#daO<<3`#k<#zL()MP9Iv#D1pXs<5tE=3BQ2&6KR+
zw6{$=5YSUfh+=MI#FctA>?8hME<}PI9zef%f_?SDj-DC`_*TK3bh#f?wswmlPcTGo
zPrbn|oPnp=n@gN8H00-7J<Pvvm#<`f3Y#pJYL6UAv6%ywiWG<XEwTra!5iT}>9i=!
zC4P98unCWF;eIqf?Jo`D;b2pB3kB94C<0MIQHDou-Z}==FDA1SefbGh8Ow#=GL_be
z=#<5^HvY4#>d=^`0$RP@-e<8FDvY=oN<pF5)u5eJDv!{MgRM9Ow1_;F5*P9|(dnmX
zlYheByU}&cgD#35LuOB`5FUZx0W@lf{cv$<B<r-X?Ub3je=NIIZQ_ThUX~F-HKtmE
zauJ4o^QM44de#{_&=v+N+S5fscveG-0@xTgIKY}?0TgNIuXu)dh-dzN;^W4ueKXnK
znH6d6VuDIKdZ6kWTll;O>St!q!ba9ak;F4ovMj>nW4^Z4m<T#HvW{MpS8XXI8Kp%+
zeo~$X?j;p5By{XURS+pzLq|m4G$?==ajq#}QJb~6#tV;`HUUb4E)k-ar?H9sBzQf5
zzrgKi?pcyg=hJq<R^iP8PY6E^XEl**+o>6e2*w7UyFwN|PPNTA6oSUG8N2Y~H3#IQ
zIzUkuLL%i@LN$xi)-A3%=x-r<80&9>kbiz}+<MNvB%x*tqxJ%mCyd!ft<NqLIM=4(
zZxtxR&~;0$VxlfAtZfKd$XX{)pLGIR;`@+2!jH7NG=#e=ehKr>yt!BAAt|j;Z~)?s
zB#5(soYmfRBB~8JO#)Y@D^fp5?Y7;evoRlGER5lB1l)G9lDG1RgCGqPzt3wHaSUov
zu(2_wxo)c;CcJ5=TZ_^ap1AyC$b9S^v`5y>z44$K6vik&b<&uzRq*FX@1XYOb|01(
zxN*EB3rSO<z*f3X>qoC>lmoWvdN9e>hEpTzMDASK$6q{`=#oBMZs9px)8O69bdNc;
z3fhX-Qwwa8>j7*iNG(tME|W@MW5q>XtVTr<R;>USqX1;|Oj-XuECJG$^2V$xQp+5j
z?^t5}DT-H?<4}uo%5+)@evY9<JZAmR!D-+b-6$04!Ni`_1?xqFX#B73HlG&I-bDxm
z0OkNbPP{}ejX}+)Fx=HDlUIVqU~2E-UB4Xy&MtM5E_deVYr<xP33upF+dN6#*mL=e
zM1VSU)4r27@Htl!cQ;HH)VvOAN%0yB35<*@F-&BJQ0mEf>NMRf7955!@2OI`H(e*v
z!o7b0%`vFPNJ_{>tD*D(JrT%9?}$15&7r&G(Io~YzFLaQv9};KeN~0)%u0{WI*GlA
zCezHlZzSY+9f}avzQPhGm8uRM<*H`%n3CK{xrtXP{+5Imf^b2iXF>bYKj9PIJM56r
z%ggN5@Ie8=y^&YV%FZ@mlN)C%U_;n?wt?=Yk`#WQQ}Z<=q??5?v_7!b_kOm7c=hD)
z!{(psq5BCKB5G{Zf=a!T3JH(KMj8Qpz*q|arHGRO2REuFs@e~^$8CJ+?Y}($!Fgpt
zl5y^%88`5PTCy0h8Wx+Kn;Cw9p?FmD=opwS*bpj{2y%2Obu=%Eo=})@p_tnTETxRl
zpC=`7<Tg}r@{~l~dV8w7Y2S0DwGIH_;_la=PGcs|6bQcjWGuwuUAzTcwu6-=oaUN6
zUEW*Bk!%iidORT$P7Ge6Vl4X_5<0}hr^PG_%b{hsqxtYB-XNNvz#+yirQ3tM3V10*
zGf}s@hP~cv3HqvNOdO!N5f4U36T8KAUB<oOzDE;H>)FmE{uQ<mhFdT!w5Qzlx99e6
zC<j)XrGv8~AI&PLRgzv8DVb{*0UtRwhv4?}E%#*o7)pMqoZ%~@=b&bv(02TU5X5k=
zq2PHmezytmW;|>Or}z(18)7dsru_GZYw`D9s3hJw2)u6a$8HAOF&HWDIFC!tC4YHR
zrUsJnJ008#ubCt%&aLN7UsS2kn=m?CBNlg;$7Q3JRYd$q(&%I;+{kHF6F%;fnW{e8
z4Wnx$xgs58igqC8!r&qo-2+HZy&fl~a7XD@j|C0wKYt70*2I7=XCD)zikwMa*j&E7
zAG>6V%U%|EjR-8G2pc(w+lY(UU;IQ%9C2Ad?Xy&Fwdn8#j5Do>6M_6N%xzL}6FFFs
zNUz09o6;-gs1dtQkG}n(L-iE>e`Jgs1tI8aGJhvjos$ePmc~8~ZMQv@Ee32?C@>a-
zg88iRGp}jDyT8fP;&G;%?W(zdUUqzfp1KSd(bn+y<XF&-;!j0cMrLlP{056_QJSSM
z#`4^VvxSwLUfiAPw`K9t`%=a7_sT_852H=&s?u)&CHc|y`6!7>pH%%wui9Ds#R7d{
z#~a1g$s`q%CI?LwO`S4cdbAWb;3ejGE`^8>dSz3cwq~c-JI2Zf_(HP^Wi0nD&68K0
zUc^Ag7XFwUnfloz^s$RRVWI&pYm!u!LgjG-e%wT9<YAIQ9o4t%yout0+v^&WM!VW6
zn~GT5C0<~a1sEB>&Ltuq5UloUn+BK_^8nbpuAlEN;68Q#u`L-g2RN6v*vy$-B1Jc2
z`|%<bS5UIb@d8~7^?V&szrTF<U+9$czJz~+O6oIvT>yodK|rM^YNu2+#dw@XssK8g
zd6IollH%v~1y;y0*z@ZP59zyx6_w&gr8+m~beW?ky1<WK%#oc|Jd}eL3p4ho-vPnW
zfUJ#!1AdH-1Iw`M23z;A+c;!z5ST@kh1YF@ESrJ-QkKR3i<~S|{F7N*8M-IT>q+|Z
z*A7p%f~YrtkH-lYRsDT`I)oWKt%M_#K2AsYs*ho2Y;j<h=p7=lK)sOQ_^Exklk$j4
zN~XhpNefFBhRdzj8S7+$2#&Gx4OvFV=2UmX_|>3N$JSAXro)51<a-)eSyE>vOko9*
z+zrWnOf0LzMhb|Y*$Hmteedljz^@aN=dxqLMRW>SQ9bEH)?IcY;E!+VR&W^?4V95y
z8Frg&*!~ztp{Cqi^#ZdNZ<w8Y;DKmCc={XxXOY7q8v)@*iaHU8kl^u#FQgChMxM~p
z%pgo-m%;!#0;3v<jIJE{t|;6gJCaBjkpFDFQ@~JO(gdwr{V==PyIa-xjg>AFMj@|~
zs1hY8wb;nanw@mPhU=}1S#?N$5-u70=E#6L?rgt~*yCxmnUUIxkrx4hw)<#&L7*~+
zy4lSLA)wYharGU0U~(<9hLeU@tsZhe+WR}ksohLzhg?z_@P7sDP~G%EEeSKqxNZ_r
z;qbec=r<NSO5Aw!efOEGX{W#1NuZGm=>a|kkTS*TQ%vTSd(?25c|yNFJ$qW|ckcaN
zJuO5>VtQm}RIDrbrE)io94X>uIA9;7coVadNf7o+vhV!v-KK0el05_-S!>pAd7*EN
zCcy#2XfZdJ>#~jP6Nhp22N|A+#Z06%*XA+1{y(RIHMp@v%#FguQgFrREf*uMMq4q&
z8H#Nu$+=mej^z}|RZ6oH`HW*7PzztvD?VmjXMXb`(@?`js$*y|egm5gHdRR9(Z0jJ
zdw14qqbFSGuY*2^=iFnfe46Q;#kX7dd8O1sze_)1LNlrhyhzB&t-Ku2`YkpEc)<or
zze!?(LGQmm7*+s<!Ao=7wf8KRgDjc5b(q)zM@5i$V5@fdjkSV1<t@wm1Re{wl{T;R
z7r?PqM5UHF-h-z^3zbfJ`+k->Jhx0zyP=2@Qetw9`s)GZeYO3)P&@KlM(ojS`GOw%
zbh+FqXiu>q6fz-(YL^bEb_8^T-dP=IFTZ~ncohK(7!}9!zXjAA+G@Sjo&f$(<u^q9
z`pZ6*RH;lu0wbc8c=6WgxmJa81gJveMTg&elroP)O{h-uY|Og+j}nx`A1mj}zsU0Q
zaL+mACR8$kOc-AMiDeq1HjtGfCZEP{VthfBrbJDH5f(|x1`mv)bGyTk!09ISR#{jS
zqNoOa@tiJSe;>!NusHw8%iXV8?`d>q57`$x4eseEz_dF9j-PxT;7rF7Ut|jWS~5lH
zXGL$RLkLqBCwoq(DgK#~WJro|FcE`3o2f?r4uoK|nxjlLBU+#%y<KDPuR?wpn%$*6
zeo-AC!`VbT&#)p_EbDBKP9DEy$ft+KXM(4}lk<@u2(Z_P7cjcDCS}icCfAy~;w4O`
zSt#Zd{<^CEb#$q4Q+cTL>A?wOUkt|sQBNy(566ysxpGy+QO+cTidokBl@2G7d-0Y|
zrFis)^yy#sN8Zpcu0*pE(6PJ6*sFQlaW}20t0nN6w_d$>BL0(+M1L8F>GiU0P<RSa
zNj^l$@5b3J`>vP}(e=EHAA{H5D&y)n6?kg-#;>sk6^*jlb;E+e$lW%VgaaC{vMav9
zIQQgBG%GZ=8+H-}oV5!MdCEbJTna$#IakQY^n_Z-l6hmOT)jvR_P$j3b-?o7N%<~A
zU4N>b*A%bs9g4<}xqRkMG?ZZfX-3|k?EG)Sl;Np=93L>v?n&m-565b3VD=B*$-G!b
zyh=JJJ1f|hP0F5^sF+^r@SSAeo#wPACf8x6?>BKmRi80mcAcd_6lZA#DYPS0ofgSR
z#M~T4dTzGY#DaV4-X1z6XDr5)HHX{dRl)9vqdSjyHJnUGNGT`bB8G{to=Wrnr_4r_
z1J>Snel)o>E6En+H}a3RRNA(~ag7Z=Q^<#m``D{<DJt};O`T1_r02$a|66fy{TJ2u
zy?+yf4n1@uUD8N*NJt~n-O?d~h%^Y&%>YuOfYK--C@n~Xq9901h)OAf)OXG6{rMN}
z$NifDXV09o&)#dVbzRTpPa7#u9J*)p#B<DAVv$v<<Ov-~!6yBTd&7)B%jInW39En^
zp)7KVR`XFPP&qA*oMyRUGAyMyJ6~0fABE!OTzeZg@>uB8RLOFbX3{>Ya4$2jLNvDy
zw>q=%3sFei>CY>j*5&p&#S&Ut{tkTmqlw3~<1yudI)ZBmZV9WDhWZk(d$)XLqAPME
zNV#wsu6gmhkRTjnhRtH|bnP83+}cqitUDt`eP-waA;WYmGDYsUO2oRIBZZW^8cJr>
zN2+l|k+C?>%T1GlSJ2#D<YCf`_c#s*HYcA*M`qdxwbdrGA7q=_ykEZk10R7aCdu2C
zqEwpKL?T_a2j=o0!Q8fC4M~%%W-)3H2)N%S2@N`QM%f*<>>Z6AYL&mS4oCO5vS#7$
z_dg;<_iaUZ#y^bhv%H<{>q5jJQmk3Xw#ttdcx}&?iq)em<Pfs#sODSOM-Z2rCYXt&
zk}xWgug6q*E+-xZ%mz_~k){#v)u;WM7(6?D<kw5*K1Pqa!eU)MQgs0ddH-78)!wB(
z)uloP7X%xJS#q=(JFV)B^;%z~;oIp2Q^LiHA90Rb`wJZO;v)9;7Rs|NxlEzwGyk+d
zQ|H=>=8j{lM7Y**tcPZI;UtxRGthB%cmUbEnnwoYv+Z>2j0j|xZQhKX*{4?F$f6Oq
zE+3lY_ab!8pP~b6%#!=waIB>TW(|-L5ejE~%FMGBrMb_Oezs!J^qV`WD}`pI6BN<l
z0%H=x(W6&u{<DpIl3R*f6a9zUM<V6$?)Q(p6Ok&Lg)H-C_z4bVJ!V`)-QL3Yi^%FL
z;lp&L1Q{P_uCE|tZ*?238m=UP-J;8QMW7{ShOv;uE-0#+!i-~n-_7dxudfU(^&ojM
zch65Ph;15IV3-TBvgZgdL)hJ_|M!98t__%J?f9C7`vaNZTUiCX4%#WgA>8)vCeh){
zNDkt|&eLjnNTcUF`H5nHxG9F{YF}ti{Vs+#A?@w!XrIbOg_53)V(3XLk{%u_X6-vt
zQMX|HmN?=(kf3^>t5|o)?$_g;{F%{f<m}{fX{k(%6~5-lNp!4ajLj!nmEM!-Icn78
z+TO3XxEk+s=4ezQ^nGmv8WJUN=HV0EBWlVO5~UH+j4?#GNT%SWh%JXHv1$njwXKnN
zkO_M?a#NTJI4I)Fq;$u0kd`0BlY4mfFQxCp>E?Pum6Q##b!1qyRv1SSYQO7-_r?j)
ziB)*^PM$xKJgFIV8vd&6lOO)VZ}vFy`{OB<>@|$t16oXqnQP6$^q(Pnu@Vz>Em&*K
z0n*io{M|}YE(yuDm}u$xbxu63>XBs6+Q1d!1G*H=sJYFvbcL8+Q%3ILOv=R$Q$FL$
zu39`Y)_5Vt2K@fhShA!#>QHR<cl19{PwBa@dw!a9Bu?>dTo-eOYI#*#;^-^p*``i}
zv2z`-IO=eLbBpjC{~nt!nro^cb1^9ZF-g(t`!H3$I|(Yy_B&BJ%F;XJ4!aeVc7FuG
zRU#}vUB(-r92q$BImp7y%$V-eZsmBf*a?J>9E*7n3pKf=OtEE8xba)Cxj&m8Ivs)t
zcX@(I&~H66pDxzZs+UD@ccDPz=6VKedhgMKVSNKvgM^f$3;VaNsU$^FADZT|qZ_&{
zH3EP3q}0;~wm(ou$xprC8UW*HJ=B1iw=GSHKS|}pU0;y>MwZpaqlWP9t{apkkocsw
z-orKWMKjdm+LNp43zs4jJ@hPye?CttATAC3Stc$1ieC?>9d#}cC_cC2&>N$>qwj!l
z7VO4{+L~*#m1ptr$KSmy*hKacfv=GBoN>(UI(?f7SVFvJOTU)@kL1UsZ8%VCxpvo~
zq6Nw*chesWz7j?j2yZvo!(~!=uw~@;eF9&BjbiMv>}PlC4iuzM>oQ5u5uG-CbC#Ds
ztE*`7o7H#OP_OV@0rBr;_C!iva#QAp;{UFaUWXjA>O=JmvR{~(u1@z5zH_H0;AFm?
zNFI5@S5Ovf()<ReltZ|DW-L})QJ>9H=hryWw(D8YPwpsDnjMd^H)_FA)}N$sqfMJn
z#4>0`xcm<G*Lv~LiAGn@_tvb)FTY~%)gw>`Gx1I1_%~GI=tNEi3T{6e2XxYR@r?=O
z$;#JYA9fPV0g$LD<a?C_d9mQA)J#B}>{k<)Pp+Rdvl7MP&}+xqRGu>Ck#FuWh@*$4
z<;s}+S;h$u%s<gd`La}yubkfJak{(%!snsJ0HIO)OfliYn%_pVQ!;~ACSpD`+FSaG
zo}#vC!U-cuw_k}iHD;#7OlT`-YWxdfRwe0M?Hk2AN&6KKmgtI@j&N^$8pz8H7*knh
zoQZqxr?B36o?BbS6)<@sRkId*@M9Zg(x9G9kEx4YFZjTK6Q#NQST=~ntM_4798tc!
z(7QvdpmIM#<8|^!LEzl_Rv*ld<hU+mQcc@3ud-1t&YJW!gGG>*Mt_YZgrrUXF1^Yj
zlmA^m2U3F3B6)^IW}bvIYPBnDE)6?wUDX<y-XEO=&ktBTEFR-Kc4n+;cGx{0$iDZj
zgtCh6*XwP;Mc0=lB+?bBntYqyh#}BO1aQ3!-dpM3>)Ggm!q=7iQ4|N1S9B^18)D^N
z=m@q6ML&rtO4z1N&gHS{(k?OU4-5mrsiSdHb{pWLy8F}INVRkL{rH|_?sXto1K#Lc
z!T2eK_TIh}S><qEX>DwpR(<DhbB9#%RJg)_)kM`-5xln1+u&B{Wnwl~&WwDdmanHm
zDvMish1HU*PwpjbUGpHQ{(43?fk*%5@7KHtx80|fJ>+=5S?umF)8bq~^c{N6H4Cs!
zbLQQ<<1AP-rZf93*EWUaZV(r5_ZL*3+&)LhH3F^m*9Y0lc&s7EU9MypgWu`!r+_#n
z=a2mRqPa`HtY#Z~iFV++m&uH%<n-nU|JC$J7u~MBsp0n_h0(scRfQqk!8e*_k#}P?
z`h5rl(^?wp-d*shgSMvnnPM&78bd>jqC|4TrR%%oUhn8R35Jljvv^Z<3fg^oCGzZ-
zSd^$_JX#N%&7_lzU42){L7FYu<59lyeAKMlb{hi4YUYcY=s!$2d%P8Ji-Xy)BRs91
zr*qK3oxGn|Knj7@?ig~@;qggfDN?Kc4oVL8D+Km4@oqA}+qlLY#OJ8{T8Ax+#q4`}
zTizYeU+9SgkR=n@#B7^mnAsLnxWpT{!BR-*k`)Oem36)w%SppLg*0}zVkNAOa9Igd
z)7pp(!P>2{-l+Yq7OS{et|9c{epZk8g~JN6TWRZTk~my&ik@sU3hC9Gs{y6KuW3d6
zsWZVO51B)FKCF#AIcUHcMVi0;x98eL(7kqvBfy^eA*T@P-t)Dm?5f{OxPn9Hj;;&e
zPXG6JNB=G(qrr?Sq4~-airCwohlJO5&g$<|>mhLzw7<H(0wokjd*3s|_PbsiP1mH(
zct{9(q}eo92)s_;d+>jn`c6am!z)L~R-w5?Y6t1ce2)mNyP+1>bdXT*qCwf989TEc
zX5Gu*GobAQU!8SM(%Kep-oHV(-W1MZwnD8Zx47Yvn>j3Zs2%h7TVoS{o=(kUzr$8^
ztD0bj8fPkX#YDx4i!%1H*Gs8ijTZ?Ngjcc}&ic;oG^-x}cKWt7wDs|h&g%{-Ui_Yw
zZXV(7y=Z=S(cBs%g%-yOjtP$r%7x%n;&Opi2Z}W@S2&R^uoF4!Gmw-|ehR>rU_5!|
zsHH@uGWeR?<emw>Om7nb@$#S*vvulsqr*h*31*a+vJPW0%Mmht`wM5$*|op<gdfeA
z2{Qccnfw(G(XvH15q`e#VAXgZJuNTUu6%EGs-hCW>$>7c$u1QFavH|a$-Bo~*Y_LS
zEl?}xIlZ1QNo5?12v;U`3-MBf9GF%(AAM9km!X>Z^G<Kz<vupUU*dw9XdzCXKiF+%
z$0|jC_3B^OTDFYce~>C(_xd~Ef|h{AOFQnLAh>|!;>r6Ss<}g`i>T3rfm%D*_Vu?x
z+d_ii(o5U{`0v!RT!Z>3Y2dd1@Q<O6ia2n|TX6VbwGypj7y5gCf|0i8ubh5EUQ}zp
z42gicFdz9ipZ}-q5UHmqq1&J!YasPSdA@$oD340<4bmXm#1VN&yZ<KO?V_Qn$mf;)
zKoD8Ufp|V)Jm0Y)`LU=?<;M+F=843g$HIBB&iY>+(qC?;{-rXUU9Mqi6hTu&m6Ws_
z@;%heOD5My=tyFeChvau3mrWCmz?{~JCNg1&xtXV^EX4Dq>uViknT&$4%?`0#gU?8
zHKg~7@Tz|qzs_BwB)kIQf-)R!B!)<+_wlz?F#Cdv+d5v{@bVSUIBF|4Q_mijoNnKg
z-i(<?8d33d4I>zV8zmS|6C%OSfI^#D&CEufsCNg`f8o8D`IGiL-^U|QH{@!AaFl4S
zP$EdR1k$HBvwhXX_mxSGr!JsGyuZfzsO#H(Y0=w@!WmMv+mrWyLs|}WGOKbkD|~(D
zNzX<`Gtpcf(E+<$ch%YNCS$2Y4`$}_2vT&@G4|Thni)5wK2-ERVt3dJdSFe!BoEtd
zu|S8Gib8<_&%|3(rIYB<(82ibCk5Q17>iZlVR)#mfb<sX3NB?=AraGlNxfDzRcf@l
z!McBFB~&;=gYL(4O}PulN4K>1dh^oQ405pwmNrY(EIn<U4%}r;KJ)GFNGvX-;OSC-
z8=wAeEMq`(K|*wMf+)Iv>&z+`f(@l5F@}v-t~*K8YSOY>$R|<1)T<baZ!A37j4*WI
zOQ-2DmmZ6@d~rW94R7X_N@Dm<%JwSWdhiau-vdU;ky<;H#G6!FUG#kr%lON1N(^=e
zsa03vdq5mwVo<=DUdYAv3wt@C>BqqCy4>qrD}Q`{i5$J)fxslJ(S}%R>%YIg-akFs
z<M}~7Sx8+~f4KQ1wddJilgX^MwQLnxmZgFqmLJdYJ<R;=KL|F8J<B-=i{bZMDRY@m
zSSHrmp!&PqgT6*Y*kSt~$GK`F&avoj)MdJqaF;b_lq7e)D9w?~oz1TF`Vrr2ozvhH
zhm;!q!Pf6YmFhx9&Je(6o_?Qr8K%h`9en+9DrZuMTFN5r$y0#wF|!Uh?0oV63J&X;
zru^o#pXqueSBe3m!F*MAn(ZoNul`z<>1z0MC+mh1h4X)V>GI`p$TJE`o?j2Wd7i$z
za>gP-Qhk#Sy{&e888?SEu5e8AYXUlmtQF7@QxL3=atJXzw|gWsc4v_%J-cnD|A^%q
z@93qQVsqsFv2sd&af)JEf6|{TJos0iZ~;QfqnQd;Yri;#?a-Bngvwd-lwM}E%%E#$
zE2L9t3KEr4kKb|6OE7=)A;(!aFY49HXw`j!SKD<H!Gwact(zH@DT5W>S_nbA3dqOT
zQ)AmqIY*}L4loAjysjCk(^nVbZ%(+gugU#RC}`{lw~0g>C}<VB4`PJvxG2zqUiG*)
z>(`&M_Q((IQO^g_G#JLa%z=T<)ZEv{JOrD)UX_K@tSKwjb5MP9^-gtWD{DbmlKb$J
zlRgT4aGHy!^F>h!ujbJp(~~aq>@bw=&d^%tW=Vr|;cgjQ@nGY1PRsx~f}`oL1f-Of
zbgzf@gKXq!Vj}tveq+ChQBi9=rye$X(RT<$Xvt68wGtQ&qcVoWgcb!8f=JIdQqxKc
z`+8qSeV{cACVydW7$qyc5y8INg~blnMMvyh7m;|L5;Pa3DslBsNi<2%=pc`vD9feT
zc%>VjzOEAzH+16*do+TsNv$l=DbLRS&>bBE!ENe|?AIyBL9<q@HyHEVkISDuLty|;
zMy82x65Z~45#p(rCVSa^416t0v+~|B?wQ>mCsjHAu6V<MbP9)hw}2JkC5%ebeH~EV
z>t;!lc;8F-GxO!EdL6+r6%Trw$c|0F<pg4GDc*baSnN8s7#0yYN&CDAQ`W<W1zK2|
zsu!eVgUuq+b>2FcO<&?R8N0hmiO*sZ@TtjFUxxq$h<>G3^<rb_C1hxxW1_2M&-eOp
z6e*!psmm{^EMvX`tGvh}LfKZCefyLb$?vPF)EF?IkJ6ZRamY{v$?Fg6*4*&X@2w9F
ze8G!D1A8Jk^lS(HezMJG9b(=eDODx<dR1N<?^%or%|+o(B+-ctIyJ$8{LbEx*BQtH
zk@v6lZR8MrP~S+xYClF4iYE&T{kEfB=6v*S?P=pp*Yy!}%%?v#i);B#X=`gT45cuW
zG#L;S!<SYle5fokDnx0lkP>nJ-K!Qn>T$+hmht1`6B&MVJ+CxGpXs1`G$bRDa-Ss2
z>A7W*J)pe~dUl=El}&hSN}|TZ#bjL!53SwoS*_|~l&48m-k7XKnfr32)myPaR8r_i
zTkV@_NRUgZ6oE$1sV;yp`Avi-Tbhcm)ce+?{!*46NUY1)bzv3RjQLR+54`k0?53CW
z9wF!x&Y<=}*hoRye=k7uMK(ISA;w{b)3EkBA;an9bff6%d64L%)4y!C=R1>6F6*t+
z(iyOVsmP6gPU<3{I*la^Qe)S01`8?TY-PBbp`5?0qQTWr_#rbN>kcPL(~^y@UDikO
z%mrt8l=~BqwrTBrxfTCH+*}Wt^zamqn!7p%&0jLJ@6Dnw>yZ@xfMser%&TqClT;u@
zmRkON9jqLWu1c<P#{f5gUL|$eNeUe@S)lVQvtBJ!*VeZi5B;F9hu*RZg!+=`fjhTP
zFw+>QR#pZ{fzi>zQCiAQr3xRb?r9iptTJ0L@*G`nOzpPMXx%Toh~YpUtTdwoi~{tq
z<_zQ-Jhi};PWtq{*_@z0KV7~b;jqhUd^6)IoRB=m?Ej`?^%y%a_2|9!nv=NF{^#jc
z&6O;cHD#w9rW=qTj-bt+G3x!2Hkz6C&8bh~AyKFVD}Il|a3Iq|tmkYRO(`<pc@(A-
za8~8W*UVWmlOv0WVw9@@TYL97N!%*;foJ!)!b$as7(Gqb6nI;7RQoO@9<s+ncJ{#e
ziN5IHaIwTsO6Atzt=u`UyGCP*ir^-=B1(4C>9n7G??y%u$f%Zh-pRPQvVZcRddRn8
zh;q$IH3=UWZIyg$YS6V;Hs1PhChm9i1I1<z(a%zEL@Tyl2O#>wKmx+bg{wFpwV6JW
ztg#Mt+OBDI%^H7=#NJuR$!G}?F}CQ^vr9y{IAiuHFP+H_Ogn=2q;^6?5eKD1)Zt^K
zmsh{WhE^L_O#!#-=E1kKU!c~P3jF@&t+n``B9wLsHcIWqOtZwUA?&(hX|LS2s~IX{
zNae2R7*vlH`?2;uv+3ee($1TRMJQwt4Nb_nm|$sM*d}?n&)145LV>VghQN4_6+;3(
zmNYF8gKdaWg6**V{p}o0`*)o6@7kq%v|~x{8M)XsW3&z!&p}e;Pzftw6I4|rgbriO
z=s7?NG_#y9PiM>#b!iWqVRQAy2Fwv8u85VzM*3YA$?-6hY+eBN#iY;0vFdx><$D;0
z-m1y1x1sk!`Fi^j9>|rE*?Sbp26>}aR~)|85l|!2%@5zNQw;#Nz(Z)ibm?i4vrnwV
ze`TLb#BNtp@vh3ZbR{Td+Yu7BFek+!1a$X|X=bhuF<^=|DQb2saoXnAl0iyX53QmO
zWJPoPvSjDA52Fhu3$NM}Chp)bX6LqIZ;WFAYa6P}%Zno-JOgC6C-sZ6AO?!$Mz1J<
z%8G@j!cqpt+&|Mw<untOq)x`E%N4K$4}+wh{u1Rq5D>_q%?Ya??_Y|_UcaMaiW$79
zG1i+p9*YD^j)SLznaZs1itWEpebm$i)TqBSI%`K-yWYWV7hrJSY~K86wx*>Ubz-3|
zUu8&+$g8_r+lq6UYRUQGC`~)T;|Z#5Y>iw{)PDKPnLZfj<jux~ms7&H=Xr1|kqL`}
zRgdHcQ@~WQq|>8vr};9R=}Mg(Px`4sy-hwhuNO*sY<F-q{dP5TS?33Z>mBMrtIE9P
z+j#2MZ=mfdTy|q$cStMru$3uQRqm?r1VxaD^}cZ@3nDmPc>cg}@V3ug18KoNIr5m8
zrhH~aiYj&Io|=g?%$KwebH?aLmp&XDHGQ-FBy-5$lMtsbt<T$R%J}zFwRv=#FTByg
zT%`dbv8UK9CvvByM@u``WpG!B8@)!CL*7Ou5W5iem`uE9SBsh?()=bj1EAeWoqKD(
zU2By+;IDemkp5mJ<k@E0U!L>OQ~OHJU)bwypvlPKf7sdTH~<1)Wki6U<aM9d-<YtO
z55a+vSMPpJIRrc&g+v=2yRA!ps&Nhq4wujhSqy)*I3{FWO+@O%oOznF9GwPS2Iule
z{8@`7*5kzp<GgO!90P%`Lg6hJ{Z-!Z;LE2P6X!?lEf?$|e`j#b?jb3}n%}j2_VGcM
zNbi>~KiaN&{n_GoLhNPiA+$x2j!hYnGLq!H-?v+ioFd|34*I&z!NmT%Zys5>JPY&S
zaH7AU&ce@%jcg(_KN6M-eyaY5Kq*2fZdkCL$kAi!M2DC{qvNK_Zh_BRzcqewrsKra
z%Z+($OR?Wi{%PINSB}kS!()Bozlz%YSoHzYe1`79nRvR4Z{Glr>S6=$UvW6<@fbr^
z>9J0C(WbRDWX3A$f2y-IsZQk4d{Mk_h14B$rMCF62JBF-@f5d{D>$CH%S2sUjUR2#
zE<Ef;sN#ea>@Jfq4u<86sLg61eH7jfOxh<XIb8W8LOC!rfBLAH0XIreSEB^PFnMEI
zY4(r1@XU)u#!_WwqjfHSvHQpP5H*MA&DfRE<`=P9L555_CzQm!W<_#BuWay7tzHbs
zLWU|+<%=(OsXf`239W+Vhdt^U61=+oxmMnPX10)H`}a5iVL&Q99AV3JpGBIRivcIt
z)e4juo@VzY5w2oZeP5y&1UaQD)$C*~nq%Vy#wr^QAb#!m_x%MhFkQY?H;t{d)3!`h
zjvDZs`brgHA<0m!er{Uh$+&8Y0jXHb!G6qjZ56yX34GPq39p>%%+QJcWWO>;mWCLZ
z12GZlMF(<MNL~^rlh&D<jolZ`ful`cjLH@mqM*j|3EgG{_tGR>ao3GhfCP0$iKD<C
zfs30SF8yvjpfmkfO3TMvrq)zi)L`(Fu5lP*LAZacRKA2Vl44NHoGq{S!RWyE1M_uV
z;y!AvyC4hmo#U`d!|T1slGsjt>Cj|9w6FfQ8q-)>-qH3+-*Nge(?HuzC}ZIzdjd3z
z@e=3Xihv(w4AV&OKKmp()x?s!$=50mG7A1gh5e_ze!^*OhT{Cfi5YFjjaj-&naE30
zdV#={GV_3tsuG_gRvkBT1iwTFl{2CK@=ghuipai35nt2Z63X&?Kd-`Sgtv(5+Rl*@
z#SmTgNE6tLOc%fOF>v_(h2u@As#qD*_5B~_KdgWpTvobmM}|Qa8=}C7nmoffB<vtL
zvq!XzQSWhfv76%)JsD>4N6P8O5F0hpt7R^hYp(xGuKafu3)^8bLuwTWG|LKGOGxEp
zq+p5nwZ!*Vl#8M;Ql$JJ$K6!xVMdv=w9|vKe=$r<qtR+q9^kBIJ&xYq4e9RalAzyn
zZ~6{u;h2oBG+o4<IV5Sq0<!(vYvc`Zoo=$3Aq?y_FDJxlpW~ft_9e8_#&)N;o9)Ed
zf;XcV?4S0u*6L{oV7}Tn;j+vi`-o7FSrI^7s*y=iUV^iUJq3O$9^1Y4ehl{NO0gNj
zTT8(T@{P#|MKBU7+>6+cwB3<gxuVYZ7{plJ_bDvh-@UWsaUh&5`{0h4aMjX`<<0bz
z6lBNvaAf!l?T|N?sMXJ-P8ZfXTh9*X4s-|f<(s4$>ML%zG<}?A^EGXbMe3Ao@k>b{
zxDD^}*D(I-8%*<&z0DF~9ZTc`y7r{&y~E;LF_U?ks0|aYATmzBy<w4jaP3ciWK})X
z;(n}FDADq^frf%x55uuOV3luns%=OVmWfG6pF$?}+cj54ZKro!f5q>698+J!mgj^j
zugXwg)+|*7yubG2nU6VYce#6hfZxGk+~{D-0KDYx<ZIZ+-q#D82jRwMUPW)KlqQy-
zy0~c)Cg9}ShrP?(TW$#<N}uJVAZxM6>;lF*OI;9?)la~9H2_09c&5uRl>`jVzAbk^
zRhGvFISG2RGM@HsiPsqR)Flmd0xQv`YnhbkJRaJlIM$NIc6HW!FZvG)w5B{w#Cu+Q
zH30ud@(pcKJVXOSXB2rgacIQ{u?E6tdXN9XmwGCDWAM+Dd~!D4B#c?vuoQ)4j0F4z
znKQAMWKH^!n^I<=&P&>W+u9|Aa4LSXD|5yYkShC2io1zyoiK?sjzMdW)KOL>q>YO(
z-vpQjX@+emBsF&9pQSfxwa*zmrmm0q7p6YYbQ0vUerdsTH@VHqw8nypJ$hN0Obb8q
zvi*7R7(h=>tg3=o({M0ZR>00DOepex@B7MTtqV>FrUM*Nn%mlVh^sDsr*y(U92f$V
z+Wd%<&(Nf<;||?GE4;4yS%t+I;VZqQWAXaA;5wd@;287ieXvt1`9VxK_&QFDMjP+m
z*D4ergE<j~cj_m{@tl`60|%(al+;uNk!dKFUws&_>Pn(e$Q9B99Df2w90os*b3b|+
zE^8vftSo->srx={HCn;`W%1v#S06cfd*>Faj|U-kMOd^rpL<tFd$}7JI3~nseZnuK
zm~ekQG()i)t&XyEGK&wr5nq15^M6C_qNKpvF=-~vu;K^#K)DPv>AH^nM~)+b7Ode3
zIk1$op%YmP4(Nl;gPr8)fyRMhYV>#`3<5EPykBgoo>J6vX6SZMA+mNL?IXFP2lV`H
z1j5IP`c(ulrtjJh>Jird&67a=D7>pJ9<K~5&!zJ5JYjnk9rNeI;6u4&iKr5Ayz2Ly
z*!J@bk4i(3Zu{87ER>c=?uafyB7r=s5+#pu4RClirr$PkyG_-pO}v`(Zcl#Vye0Xn
z5l<-ZU`%$r?59GfLC(k@!uK89u3#ES=?a|sxyE!6CZ;3k6ctH!b7m)n_@~g~y#LqR
z|1oB+(<6OPDI*izb7%$llWmLxH4wj`DF#t{YsybE#%|j}i*v}~@$a)A`jK!1JV1!q
zzGmCPEm)O0unpOIdfXGh6KYat^Vqiqd}0PXPrmu5b=>mDwr;!U-?s0`Sg;%6R1J(f
z06=(^FZ$-sX08qPf^{vS!9%KwE>ojWEXt@Bh`PH=vUFm3Gbo`SfIq398+mzST(?hy
z6OXOlzV{M`%Fe<YPG7=4o^}StWiLQifbX`zT2j4btQhW3g|et5FF&776jp+bZwLJ7
z{ecvw%z2Y^edO<-bGc$J<<EatDXzYLuis)71cAIY&z*Dv>>gWI-}cjk+~-BVoRhKA
zdyzfDzDr#|3~s<ZC7dO70{~q@U#v=|lQ)mBrNSU24kEkWz;>)v>nwfsZkm4e9wsKE
zl)al)K$3kg8{Spn|K8Q*3zTn$M>^Is@SI#+oxYWZtnZJ}AB@1pjbh}u{9wZprX7k#
z%i_5T>k^Annn^}`+CV)7<UJfZ2&juc1A|wYMk+J7p$P-#l?K?P88)&>Wq;HwX^3<L
zY_-Md8>GBvufGjN6HedW;!8N*4W<bccaBsqOFAgPa#s1@KB0oli^Y0Z2p)_g_HdW<
zY+mv5m|zIyq3uY~?aO4hxzP;i36u<zs&?>?Z=rTJcoU87Q7B8R5E^`mINC$hdtT}-
zdY_L|%}5x7dM;C8wL{<x&cnu<o*{4Teisb564ky^S}pcUe-h*pEPr9b1A`0dz^iDF
zNvDL59jXy_t)TJl3FOBegT59wyoZSBFHOc|p)ekzSMJkMoAb2_oB==c>+Gto)0ww;
zKb9Ucte>}-IuZYDrAk>#s&Ldr9+H`f-hkyq8P1j0+3V_#DVanZ`oP*aYM8PgqHsPe
zcgf)Q{GEe_i<Vbp7nsf}PChqh^j$?c0+P+LxKM@=ml^Z1xfUoTxy4^!0{Jor!Dk3V
z=`7tQ0~emfXv%2(-})+y#I|`DEHv}t`;dU-U6KFd01s9>9vna_c8{fyMwkAk(xFyp
zb9&2yaO=Lczn$Xb9X8t1`U>RYK?~>+c}8S&N*HgO-%Tzw!{zue0&G<>sFmZL6bg0h
zMwR~NWKVoTS03B>qj5On>&OSu_w#a(gJ-yEg3g@|Q@XBYF2)jNn+5u99gLAwbovLg
zc!A8DBR`8<=SL{!v!n}shH_6Q_>G_G7u%G_;|$MU+uA4$Ai;lC1W>3QHH!07u+8n-
zEH38z59{{bl;)ujF-$-S(?m)KtiSa5O!B$G`tEGM72icE5!<ZH2^cm<xy9W$vZF=a
zVx~J$NCND}`CK8}b8v+byJKDay9LwKY^XU6vk+qkkg<d#e!Y&HTgfIbE5YMHTcSky
zXmH}#44|wP)t^K!K4kFkoOK(#{Tx1_;J?-366JoWwZ4wHOU#F8R4K9T3<FN}XAoW%
z-V#!l$g%WR-n_9+9g{Q^8_|X=*(mZK;)N$(g64n@Kzks@?=PLv+4tAZh6NpVvO7i{
zg`g4O+E%kKLA8k(1VST{-@T-Yw1a3PmGV4F48}rj*#vAx3qU;t@{t2z@DgzpSC%`G
zX;iJyKm$Zh$3Du>FED}t3~u8V@x?D*PBnLa&VZY%j(@F1@lf0O6Hle4=0+1L_2aM>
z2tl%{IC(qs>wGB@sd<KE9cbYnPT7TFTriLx@PVZ@dTZgI4QD`2aZeQOCSXG7e?hhn
z{<|*fd*!@40#qTs;79}m+yo&_p}H!4NEO<d>9M>vafy?+d#L9dZSop$tUQ2>9}pm3
zW?!ECiH&WfnZzvrMO)^nuT!m&3Z=~L4CtBTWQ*5RQnEnq5eUZfhSU(lLJ}P3p{zDY
zRU|yP>GlLdfB<i)nj@*$L$uj4q#Xu^baIaREcyJv-ktpA#^zj$(DJz^)PM!tI}(bW
zkRTH(ke`C^+Bb7^2}2cYw75`_xdMPLaJ}k<NDiyKtwhdgazngAsfCH38>?_CREk5s
z^M0h;x0q9FaT~?KUZ>y5!X(;@n_<s;8qY?Zb}v!~1=(o{r)%yX-5~k`Yi$Viyjo|K
zllPtVBM3^vel!dK9q~StzSJIG^~)f~_QX4i%oeb`3k{NZIyT!1B)s@4@z!I3aYS=R
za&;eAo!MTv=hx>Tw)B~uWdOv(yEoV0miAPrJ3YEg$~>Z?vBZ35LYd57%QIMM2^`J*
zrnU2s@kkg9cuv813=jhC*FiUQM2#d+58}#mnExGG;Fa!>j7}Cno-VTz<33dU^YP}J
zyXpFW$~V;Qn*oL@;lHM-L-15PHw9);w-}DpLKt8gQUsuHW0@L$xy=;G%*`6BwtX;g
zWG;-hPGq+UVgm*=0^zM3j7j5(ZW2~L3Jqm#9xPHyvQIn*JxVjQSy2#X$}hblxP+Io
z^rEEA^d17W6wEqcGvTgfv{m0r&!f3+#ELvnsUC@CP8tRkMG96kb^{Cg<ci&F(8^gS
z&?}hbL+YcBk9u@m{*UKFRvhe0Gx6Q9g^Q&tp)Qt?W+ijU$>JaE*G7Z$EX|Wv=)|sm
z!=#t`TwM<Q#El#0+mN8?d6T|o2i(}|hJ!3>dp8-WOM9OI#}OJG%(Z8%aiaNx;#M;<
z60DzqJJ0e=h3D8?-wqp8RQyI}@m~=W(<`e!ECGfFfR-5BbLjX)KyUGl1n0PRqYQVS
zsgExysVmuX7IEgtUXc6f<yR>uwnGZnYz(*!gd#v3w`}uwR-R{duQ~v42(>Y5w7-*J
zL2&C;)U2wxE+L+9owH!GUKIJRHTBOL?$%YaFJK20n)5ClkDjK2Ga@NT=xQOxpiG%`
z<ZaWH)m=eFj%jkGFH-9*ja+**;#}7W8RBCU)>fWlO1CXj*sEVplCkvcAZFJ-i`;1+
z;S6P}3*ZQQ+3+W^y(6bBltA2PVP1%bc;{isnOt%l1@FEQON7<IibXVTub~E0?27qB
zvE!7$rnl6$;~y58pG9e!Xx7IlU&!5PS-KXB7P?WWo<Y<0NzKLL;ENA-6HUQMI(qnH
zlHD}1z)dCh8+RnF5o;Eif(~4t$34))VOMR9xb#y~Uv^h*Yiv2TeCKmRFn*>l)GvGe
zVY~IylB+sL2Pk4MCA*HDjC;(fj7`#DtFVaU_}6G9rzQt+26z>UPZkOx2VZf8h(Tca
zMw&`vQdy~1d=fo;;m?LXvE}&Pi4rEoZu^uc$vtbtfi4h;au-5Gslv~6+ZVhj0!~_N
z$2%&VSaz1r`8^_yHE~Y0USeD!7!nU1z#$4;^<QF0JPbhMX&%knqyiF;3lJZ?isUCT
zBpxp?WqpxwEHee*Lji`wbMH{C9Y7Gx7!uEa<OCLkCkFMiK1%^yOXN%(7(V}h8{qc%
zOjjgs5KvwdD7fD>m-|9VBy78;#eq=X#uoM@STZ^WhMsf57sD7SpUFRfHUz*jPzOjC
z#(q=*%V~Hhv?T+CH4qx~_1fFR&t5=@!nM#({SxuC8r+3#)XWwfpPLgmF_dnQ58irV
zGFs&Sae71sdcWc4-v%BWbe_QusWh!yfW;d2S)iJ=7ZAC;JIx<4etCPig*TC}bN3az
z<ABuq)wWSI6(Dxc1x|FWu<a0EQ8)#BE3JIUZ|J*+8Nzf)5Fnv2%`rO>xKDuX69Y@8
z*gK$acK-o-@z*^F3otxkNc|Rr-wD9BHI@oMAArysBEAJ!r4ONF!b;ZP%}LyYen`5(
z%KHUslP!&czznG;1ZW1)#x`U;<X7U^fobvx(gBg-<Df~ecmhoUivmBH!>a!iqS|*U
zOv5KXqX-3A9Ph$e7$n|Lz!=uPs+%Xg3{^uuqw>KxVus+$GT%VX;SQ|7XL0H58qC0$
zf#~!cbQohbKx1a{-TDrH&e{9iZ$P8crGPlP*1S!6^*MWKBX_~<&@&1R`m%8_rj7o*
zT!m@p)c4}%%13kPdZ-9yy5!b7Ll{H#b0GEvoc;h0BRQJT48*o)m|kubBo{XR?>T7&
zTb15i4}C(`-XhvMn6XHtZk)+aHu%3gANz3y+f`n3950#Z0>VoTZNx7OU`TL62mJ0d
z#!KL6joue=eCM0!aqcz!d>Sk=!s-F?pc1nV`uYlxGTy3!)Ph#=`JvDb&9Tw<Jgr~T
z>!r&tAh-O3fFQZ1&I2ea)tFYQy}YV{KNma0Ay`RK__qnC`ZI`0yftPn1BEB{UJING
zQ(^+W?ZcIlAAmaGNE&*x8Jikgw@}e|e6Ibr2&-r*iEKC!OxOBb?;W@dy)do<P*i*0
zr_ZlYJd}~3MN1PZC!pf1Pct$Vgf=0TlvUfC-iI)`U7}crX{0Tm8}lB%Qct3dtTO|z
z$m0}LE3|cN`#q(Np$dUBz>e7WG7mVbfUZ{rwF59#!4Z6ThIrm`3~$|9Fm2uT*U`gu
z?PKt5QkdE*#wwK}<zQv5zH&&!C}vvv=N79QEEi(Rv_g2NsI+FDrlU2mqGm^5CVgie
z`F8hOB_PWG3CY@>-Z3@}A{rr-k~X;9a1#FWv6dU9Htj${!bxk1Ul*PDcV$D$BuUb!
zVDjjSrHle*M1SV4Ac5dzW#yH)sLP~e#&SHtDC088Zq@nSk`-mL+<1mZBCXd?-<a-9
zZ`17mc(3F9?D;>%ho!G8p8q=un?5N1UibD_!?5?^%_-Kr_hoIz;WnZAgJu)IKd~9v
z8z9U&<O9D`qE|ulx3#%RYixOonGI)ZUOD=~@jhpQgO`^V0$az?rCmqG(1yH|hb7-?
zB3(mgB8P#e(lH8`5}x(YkG$(*qNu#XLs^Nf`|+Y7UuS2vh;oCx4Oi!SUfFl%f1!*t
zLR!JO2uh7mtgr$25n*%nJeMT~8Qa!|PEsFhfDWY`iuA&Z=>$Lloqy#)L)g!u&=@ij
z$Y3Zg>f2h!ojWW5MF5YtPKtb_N)f|r^)}+;mUh`mP024VI3&dqW<W6q`8|j$9hGjj
zhMk2pm^o!SU%Od+?`py4TI;4!uLYd@M<yj-xyEEcaOcAFjJ4Z#%xA8=17|>>uGC`s
zky5PcyAeCc^3dv+C)#?C8EqrKYi*o%6-x)i72jPa(7^#8q&nU{83O}@7Wu}F6#aBl
zl{X{YNAdFoa71i~QzrTeTyYGg<ar9r)F1YwHD<l_4|$##ly*8ZHpuAi4zCNu{N=1}
zTYpAE?uYBGeg9xRxoUQ9&SSXBOBZj<2gjGwMDUG8HTRt@MAQrO_h}DKkJ6?26Bgci
z)HdiWuojdLpr}_j7I9`f!xaZa^G{AfHSXkS%YR4<B(w+&3`FZb>(bHiEOw^k+iOYc
zy;XxBW<Gk$6PrH8R)+^V#CH3<-O4>!$kCP~$F|~X<1{f5C!<ZnlDy`h{_)EfcLmOl
zu6?W~Yzf2n<Hd@AELxDzS(ra62-WzWnW_4v?6w&XB{DwU(NSWySZAr>@ryJLmQmd)
z19QX*tI|6h8l6El^ioGKAH@fmyDOzy_MYzUjSDGP6pMJDe?9FC)!^fxU_{&OVqwqS
zh&4i9%QDgmR8vCqu<ai0IU0GGk(#yomcA-<jV@`MZ}!e`Z+-kQR_K5iTg2pn?(CW@
z07lZ1)ZP6*NNamFQ&922*fpq<{KSY<#Dqzd1(Y#fJ%5xfHy=%9QM9jP!j^jfE$Coj
zVN3o4R4I`l?Z(>&OrsiS`%I&HZ*p~&8e$tKt4fi<H@f;zgMnHEg3TGFnrX{WSXHCK
zwDy7R13LJnWowQ}k#fD_P3RmC`u%akcO+vyPRyjy7;%GL$#@x7{k6b9FUD|HwD?@T
zlCx_eWy><eky%<VARsWWd7Hfnj<6M6`1Xi<ZB_dqpjeajCnTL*4r6j}BEoT2c-fxB
z-sZm2D>iI>Z1gQGZ%`Qb`Mz>s>kTB)pENRkmc2v{Iu&9qS!w+u`TOb?d0K7D4Z3U6
zL=tCQv-Sz!N)KDq#0pJ&h5Y+R#D=NH15-?y9w`YUT1nZ^Z|r`mvkwNnEQpD_H>*p<
z$GTg5rR9c<ktG=ln+z3c3~#xy+H)2hHWr*Qq-Mb5QezGHBU{=HrTl_NnpE&))Vk=f
zZ`X=teb+?5-;D4z18$O4MIqxOkW%a{qk(ZoObHRlcqKmU+unR^w=h@5c14u8jTt;=
z$Bd!L3A5vRIRWOXOGy#eKGWR@<iHG91wT|3Z>)^DY9U$Lx{PX73T8N#?1*cUROx&$
zKzxK4&Xkoa&vZIwxN`Ee^=E8Gc$lmDsw?Gh6Frq7|KFEDf13aI@86G8Q{ti(pYQmk
zn9)>kuzbib7~eP+t7QGQt-ta0>sN3C><uuRv*FBgq}?8UEa|fV5s#SYZt$Z{IzJQd
zM-eRu(VDi?o!>xid&9)g5(j4yZw#F8#MlR}7|jX4^%c-dVC&ua+AeA4-aXwJut5T2
z`sNJVk;Nw|JrkGA9e&l)QfJl>W#;O%nt$i=4z-X?QpHihd|e~lra$=Cj4#q`m>Ozo
zeuBLxR%IEmJfowdHnZz}v@Uz6m5lNsl=_eFRO&s~>MH;AaeDgJLCGr_6Llkx#@e<*
z8^C9xuiZA%)+YP{FY9Z!Nj46t&8EVD+)yIw(RdRIBfGe(-8I=M#=%yZ7B~5+um0eU
zU#HWmNnby|C&8u#dJq|)$bknPUZh<J$lktv%cA!IJK4#<FDRQIj~YLfQMxRxCn76q
z=C+ZCXj$9;e)nuzHShdro2t*$P~9aaLaSz(ji2z`h2s_12M;%Msnl<d6NoGeir>4$
zZXMTxl=#2!;@F@J+>>c~mp&r<sS8((UD=lMBaA$(C~Ew|r{?KNBS?f`3S?PiKH@1M
z2<(<O+x}3Y4+~DQVtANu&gDB6gT|!zzkmNuvj~7jlY>Sc@#Y%E7OKCt;I93zU$WHg
z+K0A+P|jmRQAL(R({_R9rP*TeoQf#Y)<YDOMD<}Z@xcVIjJr7o^Cwsj=FBnHdpVf*
zl`z+z(CbOF!=#jk>Aiu=aR8IAjE!Y9*<I>e63n*^n4d+YX(JTaWbuz+N<LrC`v2(S
zN|pI@6B84kKhsPb4AQ};qs1>V-pu~PZ#?=Kwh+*!*D(C=btpyvKR4FKe6Ht%)%iCE
zVDI!hG&D7RcwPG63wfj!(OIurXCUyBkg)fIm5yuKTL{n&5AVLOE#^j7!^<L!n5cv~
z|HQPC`b+;<NONyl@8Vh5$zd`IUy%vAk98eEn@Ro0GSf#CIYNyla074PAdO8;O*J$I
zKo}7O)l0kXs8EIR;`2|RKR^82*(ue#SknnhEa3G?uPDgDjt3lOo+ibi?#!F;EPgfo
z>Bhps>ef<IHtr@<vMk8Y=h83gfUf}z6M(>EN;_2M6Q02~6Lz#)+TPCa4FE!3K0Z5t
zcNWnX`vl&ePWABGI3eQa_UQrd1AI8xc?n~2XjWj6g{A`>Mdt<GsI)nD^&}`L;9Odz
zQe={sm*@J@#r)n|X_e>ZbruDjN&!8Q@$1k0+Zb<JJJGyx_#nk#<Jmov++5Z);(HH*
zoNs`FRt!H+@;P0DwMHo(F=BK9%;~)EjY~00^aAg4ZQEE~?w$Aj;DRrMrFC6TtyaY@
G>i+?8$zyo{

literal 0
HcmV?d00001


From fac96456c241a1a34975453d1db0b8418446fe2c Mon Sep 17 00:00:00 2001
From: xzl <zlx_hg@163.com>
Date: Fri, 1 Dec 2017 20:41:34 +0800
Subject: [PATCH 07/94] add prelu neon impl

---
 paddle/math/Matrix.cpp        | 23 +++++++++++++++++++-
 paddle/math/NEONFunctions.cpp | 40 +++++++++++++++++++++++++++++++++++
 paddle/math/NEONFunctions.h   |  1 +
 3 files changed, 63 insertions(+), 1 deletion(-)

diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp
index 88e9180690..be87a4c296 100644
--- a/paddle/math/Matrix.cpp
+++ b/paddle/math/Matrix.cpp
@@ -28,6 +28,7 @@ limitations under the License. */
 #include "hl_top_k.h"
 #include "paddle/utils/Logging.h"
 
+#include "NEONFunctions.h"
 #include "paddle/function/GemmFunctor.h"
 #include "paddle/utils/ThreadLocal.h"
 
@@ -4157,16 +4158,36 @@ void CpuMatrix::print(std::ostream& os) const {
 void CpuMatrix::paramReluForward(Matrix& data, Matrix& W) {
   real* input = data.getData();
   real* w = W.getData();
+  real* output = data_;
   size_t numElements = data.getWidth();
   size_t numSamples = data.getHeight();
   size_t paraSize = W.getHeight() * W.getWidth();
   CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
+
   size_t partial_sum = numElements / paraSize;
+  if (paraSize == numElements) {
+    for (size_t n = 0; n < numSamples * numElements; ++n) {
+      output[n] = input[n] > 0 ? input[n] : input[n] * w[n % numElements];
+    }
+    return;
+  }
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+  for (size_t n = 0; n < numSamples; ++n) {
+    for (size_t i = 0; i < paraSize; i++) {
+      neon::prelu(
+          input + i * partial_sum, w[i], output + i * partial_sum, partial_sum);
+    }
+    input = input + numElements;
+    output = output + numElements;
+  }
+#else
   for (size_t n = 0, k = 0; n < numSamples; ++n) {
     for (size_t i = 0; i < numElements; ++i, ++k) {
-      data_[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum];
+      output[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum];
     }
   }
+#endif
 }
 
 void CpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) {
diff --git a/paddle/math/NEONFunctions.cpp b/paddle/math/NEONFunctions.cpp
index 3bf47901f1..0f83149422 100644
--- a/paddle/math/NEONFunctions.cpp
+++ b/paddle/math/NEONFunctions.cpp
@@ -49,6 +49,46 @@ void relu(const float* a, float* b, int len) {
   }
 }
 
+// b[i] = a[i] > 0.0f ? a[i] : a[i] * w
+void prelu(const float* a, float w, float* b, int len) {
+  int offset = len % 16;
+  float32x4_t ma0, ma1, ma2, ma3;
+
+  float32x4_t zero = vdupq_n_f32(0.f);
+  float32x4_t vw = vdupq_n_f32(w);
+
+  for (int k = 0; k < len / 16; k++, a += 16, b += 16) {
+    ma0 = vld1q_f32(a);
+    ma1 = vld1q_f32(a + 4);
+    ma2 = vld1q_f32(a + 8);
+    ma3 = vld1q_f32(a + 12);
+
+    uint32x4_t flag0 = vcgtq_f32(ma0, zero);
+    uint32x4_t flag1 = vcgtq_f32(ma1, zero);
+    uint32x4_t flag2 = vcgtq_f32(ma2, zero);
+    uint32x4_t flag3 = vcgtq_f32(ma3, zero);
+
+    float32x4_t mul0 = vmulq_f32(ma0, vw);
+    float32x4_t mul1 = vmulq_f32(ma1, vw);
+    float32x4_t mul2 = vmulq_f32(ma2, vw);
+    float32x4_t mul3 = vmulq_f32(ma3, vw);
+
+    ma0 = vbslq_f32(flag0, ma0, mul0);
+    ma1 = vbslq_f32(flag1, ma1, mul1);
+    ma2 = vbslq_f32(flag2, ma2, mul2);
+    ma3 = vbslq_f32(flag3, ma3, mul3);
+
+    vst1q_f32(b, ma0);
+    vst1q_f32(b + 4, ma1);
+    vst1q_f32(b + 8, ma2);
+    vst1q_f32(b + 12, ma3);
+  }
+
+  for (int i = 0; i < offset; i++) {
+    b[i] = a[i] > 0.0f ? a[i] : a[i] * w;
+  }
+}
+
 }  // namespace neon
 }  // namespace paddle
 
diff --git a/paddle/math/NEONFunctions.h b/paddle/math/NEONFunctions.h
index 69085e3335..d67b2f47a8 100644
--- a/paddle/math/NEONFunctions.h
+++ b/paddle/math/NEONFunctions.h
@@ -18,6 +18,7 @@ namespace paddle {
 namespace neon {
 
 void relu(const float* a, float* b, int len);
+void prelu(const float* a, float w, float* b, int len);
 
 }  // namespace neon
 }  // namespace paddle

From 5c057f95529656e379ef404a2e388e1be3e88de1 Mon Sep 17 00:00:00 2001
From: sweetsky0901 <work@yq01-idl-gpu-online20.yq01.baidu.com>
Date: Sun, 3 Dec 2017 14:40:33 +0800
Subject: [PATCH 08/94] add spp op only can test ok

---
 paddle/operators/spp_op.cc                  |  98 +++++++++++++
 paddle/operators/spp_op.h                   | 148 ++++++++++++++++++++
 python/paddle/v2/fluid/tests/test_spp_op.py |  48 +++++++
 3 files changed, 294 insertions(+)
 create mode 100644 paddle/operators/spp_op.cc
 create mode 100644 paddle/operators/spp_op.h
 create mode 100644 python/paddle/v2/fluid/tests/test_spp_op.py

diff --git a/paddle/operators/spp_op.cc b/paddle/operators/spp_op.cc
new file mode 100644
index 0000000000..62fc2112a8
--- /dev/null
+++ b/paddle/operators/spp_op.cc
@@ -0,0 +1,98 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+Indicesou may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/spp_op.h"
+namespace paddle {
+namespace operators {
+
+class SppOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SppOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput(
+        "X",
+        "(Tensor) The input tensor of spp operator. "
+        "The format of input tensor is NCHW. Where N is batch size, C is the "
+        "number of channels, H and W is the height and width of feature.");
+    AddOutput("Out",
+              "(Tensor) The output tensor of spp operator."
+              "N * M."
+              "M = C * H * W");
+    AddAttr<int>("pyramid_height", ">= 1");
+    AddComment(R"DOC(
+        "Input shape: $(N, C_{in}, H_{in}, W_{in})$
+        Output shape: $(H_{out}, W_{out})$
+        Where
+          $$
+            H_{out} = (H_{in}−1) * strides[0] − 2 * paddings[0] + ksize[0] \\
+            W_{out} = (W_{in}−1) * strides[1] − 2 * paddings[1] + ksize[1]
+          $$
+        )DOC");
+  }
+};
+
+int OutputSize(int pyramid_level, int input_size) {
+  int bins = std::pow(2, pyramid_level);
+  int ksize = std::ceil(input_size / static_cast<double>(bins));
+  int padding = (ksize * bins - input_size + 1) / 2;
+  int output_size = (input_size - ksize + 2 * padding) / ksize + 1;
+  // output_size = bins
+  return output_size;
+}
+
+class SppOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of SppOp"
+                   "should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of SppOp should not be null.");
+    auto in_x_dims = ctx->GetInputDim("X");
+    int pyramid_height = ctx->Attrs().Get<int>("pyramid_height");
+    PADDLE_ENFORCE(in_x_dims.size() == 4,
+                   "Spping intput must be of 4-dimensional.");
+    int outlen = 0;
+    for (int p = 0; p < pyramid_height; ++p) {
+      int outh = OutputSize(p, in_x_dims[2]);
+      int outw = OutputSize(p, in_x_dims[3]);
+      int p_level_outlen = outh * outw * in_x_dims[1];
+      outlen += p_level_outlen;
+    }
+    std::vector<int64_t> output_shape({in_x_dims[0], outlen});
+    ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
+  }
+};
+
+class SppOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
+                   "Input(X@GRAD) should not be null.");
+    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(spp, ops::SppOp, ops::SppOpMaker, spp_grad, ops::SppOpGrad);
+REGISTER_OP_CPU_KERNEL(spp, ops::SppKernel<paddle::platform::CPUPlace, float>,
+                       ops::SppKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(spp_grad,
+                       ops::SppGradKernel<paddle::platform::CPUPlace, float>,
+                       ops::SppGradKernel<paddle::platform::CPUPlace, double>);
diff --git a/paddle/operators/spp_op.h b/paddle/operators/spp_op.h
new file mode 100644
index 0000000000..2a2824bb31
--- /dev/null
+++ b/paddle/operators/spp_op.h
@@ -0,0 +1,148 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+Indicesou may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/math_function.h"
+#include "paddle/operators/math/pooling.h"
+#include "paddle/operators/strided_memcpy.h"
+
+namespace paddle {
+namespace operators {
+template <typename Place, typename T>
+class SppKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const framework::Tensor* in_x = context.Input<framework::Tensor>("X");
+    auto* out = context.Output<framework::Tensor>("Out");
+    int pyramid_height = context.template Attr<int>("pyramid_height");
+    out->mutable_data<T>(context.GetPlace());
+    auto out_stride = framework::stride(out->dims());
+    int input_h = in_x->dims()[2];
+    int input_w = in_x->dims()[3];
+    size_t output_offset = 0;
+    for (int p = 0; p < pyramid_height; ++p) {
+      int bins = std::pow(2, p);
+      int ksize_h = std::ceil(input_h / static_cast<double>(bins));
+      int ksize_w = std::ceil(input_w / static_cast<double>(bins));
+      int padding_h = (ksize_h * bins - input_h + 1) / 2;
+      int padding_w = (ksize_w * bins - input_w + 1) / 2;
+      std::vector<int> ksize({ksize_h, ksize_w});
+      std::vector<int> strides({ksize_h, ksize_w});
+      std::vector<int> paddings({padding_h, padding_w});
+      // pooling output shape
+      std::vector<int64_t> output_shape_vec({in_x->dims()[0], in_x->dims()[1]});
+      output_shape_vec.push_back((input_h - ksize_h + 2 * padding_h) / ksize_h +
+                                 1);
+      output_shape_vec.push_back((input_w - ksize_w + 2 * padding_w) / ksize_w +
+                                 1);
+      framework::DDim output_shape(framework::make_ddim(output_shape_vec));
+      // flatten pooling output shape
+      int output_flatten_w = in_x->dims()[1] * bins * bins;
+      std::vector<int64_t> output_flatten_shape_vec(
+          {in_x->dims()[0], output_flatten_w});
+      framework::DDim output_flatten_shape(
+          framework::make_ddim(output_flatten_shape_vec));
+      framework::Tensor out_level;
+      framework::Tensor out_flatten_level;
+      out_level.mutable_data<T>(output_shape, context.GetPlace());
+      // pooling
+      math::Pool2dFunctor<Place, math::MaxPool<T>, T> pool_forward;
+      math::MaxPool<T> max_process;
+      pool_forward(context.device_context(), *in_x, ksize, strides, paddings,
+                   max_process, &out_level);
+      out_flatten_level.ShareDataWith(out_level);
+      out_flatten_level.Resize(output_flatten_shape);
+      auto in_stride = framework::stride(out_flatten_level.dims());
+      const T* src_data = out_flatten_level.data<T>();
+      StridedMemcpy<T>(context.device_context(), src_data, in_stride,
+                       out_flatten_level.dims(), out_stride,
+                       out->data<T>() + output_offset);
+      output_offset += out_flatten_level.dims()[1] * in_stride[1];
+    }
+  }
+};
+template <typename Place, typename T>
+class SppGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const framework::Tensor* in_x = context.Input<framework::Tensor>("X");
+    const framework::Tensor* out = context.Input<framework::Tensor>("Out");
+    const framework::Tensor* out_grad =
+        context.Input<framework::Tensor>(framework::GradVarName("Out"));
+    framework::Tensor* in_x_grad =
+        context.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto& device_ctx = context.device_context();
+    math::SetConstant<Place, T> zero;
+    in_x_grad->mutable_data<T>(context.GetPlace());
+    zero(device_ctx, in_x_grad, static_cast<T>(0));
+    int pyramid_height = context.template Attr<int>("pyramid_height");
+    auto outgrad_stride = framework::stride(out_grad->dims());
+    auto out_stride = framework::stride(out->dims());
+    int input_h = in_x->dims()[2];
+    int input_w = in_x->dims()[3];
+    size_t out_offset = 0;
+    for (int p = 0; p < pyramid_height; ++p) {
+      int bins = std::pow(2, p);
+      int ksize_h = std::ceil(input_h / static_cast<double>(bins));
+      int ksize_w = std::ceil(input_w / static_cast<double>(bins));
+      int padding_h = (ksize_h * bins - input_h + 1) / 2;
+      int padding_w = (ksize_w * bins - input_w + 1) / 2;
+      std::vector<int> ksize({ksize_h, ksize_w});
+      std::vector<int> strides({ksize_h, ksize_w});
+      std::vector<int> paddings({padding_h, padding_w});
+      // split outgrad and get flatten
+      std::vector<int64_t> out_shape_vec({in_x->dims()[0], in_x->dims()[1]});
+      out_shape_vec.push_back((input_h - ksize_h + 2 * padding_h) / ksize_h +
+                              1);
+      out_shape_vec.push_back((input_w - ksize_w + 2 * padding_w) / ksize_w +
+                              1);
+      framework::DDim out_shape(framework::make_ddim(out_shape_vec));
+      int out_flatten_w = in_x->dims()[1] * bins * bins;
+      std::vector<int64_t> out_flatten_shape_vec(
+          {in_x->dims()[0], out_flatten_w});
+      framework::DDim out_flatten_shape(
+          framework::make_ddim(out_flatten_shape_vec));
+      framework::Tensor out_level;
+      framework::Tensor outgrad_level;
+      framework::Tensor out_flatten_level;
+      framework::Tensor outgrad_flatten_level;
+      out_flatten_level.mutable_data<T>(out_flatten_shape, context.GetPlace());
+      outgrad_flatten_level.mutable_data<T>(out_flatten_shape,
+                                            context.GetPlace());
+
+      auto flatten_stride = framework::stride(out_flatten_level.dims());
+      // memcpy
+      StridedMemcpy<T>(context.device_context(), out->data<T>() + out_offset,
+                       out_stride, out_flatten_level.dims(), flatten_stride,
+                       out_flatten_level.data<T>());
+
+      StridedMemcpy<T>(context.device_context(),
+                       out_grad->data<T>() + out_offset, outgrad_stride,
+                       outgrad_flatten_level.dims(), flatten_stride,
+                       outgrad_flatten_level.data<T>());
+      out_offset += out_flatten_level.dims()[1] * out_stride[1];
+      // flatten backward
+      out_level.ShareDataWith(out_flatten_level);
+      out_level.Resize(out_shape);
+      outgrad_level.ShareDataWith(outgrad_flatten_level);
+      outgrad_level.Resize(out_shape);
+      math::MaxPool2dGradFunctor<Place, T> pool2d_backward;
+      pool2d_backward(context.device_context(), *in_x, *&out_level,
+                      *&outgrad_level, ksize, strides, paddings, in_x_grad);
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle
diff --git a/python/paddle/v2/fluid/tests/test_spp_op.py b/python/paddle/v2/fluid/tests/test_spp_op.py
new file mode 100644
index 0000000000..806d5e7736
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/test_spp_op.py
@@ -0,0 +1,48 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+from test_pool2d_op import max_pool2D_forward_naive
+
+
+class TestSppOp(OpTest):
+    def setUp(self):
+        self.op_type = "spp"
+        self.init_test_case()
+        input = np.random.random(self.shape).astype("float32")
+        nsize, csize, hsize, wsize = input.shape
+        out_level_flatten = []
+        for i in xrange(self.pyramid_height):
+            bins = np.power(2, i)
+            ksize = [0, 0]
+            padding = [0, 0]
+            ksize[0] = np.ceil(hsize / bins.astype("double")).astype("int32")
+            padding[0] = ((ksize[0] * bins - hsize + 1) / 2).astype("int32")
+
+            ksize[1] = np.ceil(wsize / bins.astype("double")).astype("int32")
+            padding[1] = ((ksize[1] * bins - wsize + 1) / 2).astype("int32")
+            out_level = max_pool2D_forward_naive(input, ksize, ksize, padding)
+            out_level_flatten.append(
+                out_level.reshape(nsize, bins * bins * csize))
+            if i == 0:
+                output = out_level_flatten[i]
+            else:
+                output = np.concatenate((output, out_level_flatten[i]), 1)
+        # output = np.concatenate(out_level_flatten.tolist(), 0);
+        self.inputs = {'X': input.astype('float32'), }
+        self.attrs = {'pyramid_height': self.pyramid_height}
+
+        self.outputs = {'Out': output.astype('float32')}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+    def init_test_case(self):
+        self.shape = [1, 1, 2, 2]
+        self.pyramid_height = 2
+
+
+if __name__ == '__main__':
+    unittest.main()

From 531e7b6fa694e7c15ca3831704097946162248fe Mon Sep 17 00:00:00 2001
From: sweetsky0901 <work@yq01-idl-gpu-online20.yq01.baidu.com>
Date: Sun, 3 Dec 2017 17:26:43 +0800
Subject: [PATCH 09/94] gpu test ok

---
 paddle/operators/spp_op.cc                  | 28 +++-------
 paddle/operators/spp_op.cu.cc               | 22 ++++++++
 paddle/operators/spp_op.h                   | 61 +++++++++++----------
 python/paddle/v2/fluid/tests/test_spp_op.py |  6 +-
 4 files changed, 64 insertions(+), 53 deletions(-)
 create mode 100644 paddle/operators/spp_op.cu.cc

diff --git a/paddle/operators/spp_op.cc b/paddle/operators/spp_op.cc
index 62fc2112a8..ff607c5769 100644
--- a/paddle/operators/spp_op.cc
+++ b/paddle/operators/spp_op.cc
@@ -29,28 +29,22 @@ class SppOpMaker : public framework::OpProtoAndCheckerMaker {
               "(Tensor) The output tensor of spp operator."
               "N * M."
               "M = C * H * W");
-    AddAttr<int>("pyramid_height", ">= 1");
+    AddAttr<int>("pyramid_height", "int");
     AddComment(R"DOC(
-        "Input shape: $(N, C_{in}, H_{in}, W_{in})$
+        "Does spatial pyramid pooling on the input image by taking the max,
+        etc. within regions so that the result vector of different sized
+        images are of the same size
+        Input shape: $(N, C_{in}, H_{in}, W_{in})$
         Output shape: $(H_{out}, W_{out})$
         Where
           $$
-            H_{out} = (H_{in}−1) * strides[0] − 2 * paddings[0] + ksize[0] \\
-            W_{out} = (W_{in}−1) * strides[1] − 2 * paddings[1] + ksize[1]
+            H_{out} = N \\
+            W_{out} = ((std::pow(4, pyramid_height) - 1) / (4 - 1)) * C_{in}
           $$
         )DOC");
   }
 };
 
-int OutputSize(int pyramid_level, int input_size) {
-  int bins = std::pow(2, pyramid_level);
-  int ksize = std::ceil(input_size / static_cast<double>(bins));
-  int padding = (ksize * bins - input_size + 1) / 2;
-  int output_size = (input_size - ksize + 2 * padding) / ksize + 1;
-  // output_size = bins
-  return output_size;
-}
-
 class SppOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -64,13 +58,7 @@ class SppOp : public framework::OperatorWithKernel {
     int pyramid_height = ctx->Attrs().Get<int>("pyramid_height");
     PADDLE_ENFORCE(in_x_dims.size() == 4,
                    "Spping intput must be of 4-dimensional.");
-    int outlen = 0;
-    for (int p = 0; p < pyramid_height; ++p) {
-      int outh = OutputSize(p, in_x_dims[2]);
-      int outw = OutputSize(p, in_x_dims[3]);
-      int p_level_outlen = outh * outw * in_x_dims[1];
-      outlen += p_level_outlen;
-    }
+    int outlen = ((std::pow(4, pyramid_height) - 1) / (4 - 1)) * in_x_dims[1];
     std::vector<int64_t> output_shape({in_x_dims[0], outlen});
     ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
   }
diff --git a/paddle/operators/spp_op.cu.cc b/paddle/operators/spp_op.cu.cc
new file mode 100644
index 0000000000..a7057907ce
--- /dev/null
+++ b/paddle/operators/spp_op.cu.cc
@@ -0,0 +1,22 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+Indicesou may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/spp_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(spp, ops::SppKernel<paddle::platform::GPUPlace, float>,
+                       ops::SppKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_GPU_KERNEL(spp_grad,
+                       ops::SppGradKernel<paddle::platform::GPUPlace, float>,
+                       ops::SppGradKernel<paddle::platform::GPUPlace, double>);
diff --git a/paddle/operators/spp_op.h b/paddle/operators/spp_op.h
index 2a2824bb31..7a385352a0 100644
--- a/paddle/operators/spp_op.h
+++ b/paddle/operators/spp_op.h
@@ -42,34 +42,36 @@ class SppKernel : public framework::OpKernel<T> {
       std::vector<int> strides({ksize_h, ksize_w});
       std::vector<int> paddings({padding_h, padding_w});
       // pooling output shape
+      framework::Tensor out_level;
       std::vector<int64_t> output_shape_vec({in_x->dims()[0], in_x->dims()[1]});
       output_shape_vec.push_back((input_h - ksize_h + 2 * padding_h) / ksize_h +
                                  1);
       output_shape_vec.push_back((input_w - ksize_w + 2 * padding_w) / ksize_w +
                                  1);
       framework::DDim output_shape(framework::make_ddim(output_shape_vec));
-      // flatten pooling output shape
-      int output_flatten_w = in_x->dims()[1] * bins * bins;
-      std::vector<int64_t> output_flatten_shape_vec(
-          {in_x->dims()[0], output_flatten_w});
-      framework::DDim output_flatten_shape(
-          framework::make_ddim(output_flatten_shape_vec));
-      framework::Tensor out_level;
-      framework::Tensor out_flatten_level;
       out_level.mutable_data<T>(output_shape, context.GetPlace());
       // pooling
       math::Pool2dFunctor<Place, math::MaxPool<T>, T> pool_forward;
       math::MaxPool<T> max_process;
       pool_forward(context.device_context(), *in_x, ksize, strides, paddings,
                    max_process, &out_level);
+      // flatten pooling output shape
+      framework::Tensor out_flatten_level;
+      int output_flatten_w = in_x->dims()[1] * bins * bins;
+      std::vector<int64_t> output_flatten_shape_vec(
+          {in_x->dims()[0], output_flatten_w});
+      framework::DDim output_flatten_shape(
+          framework::make_ddim(output_flatten_shape_vec));
       out_flatten_level.ShareDataWith(out_level);
       out_flatten_level.Resize(output_flatten_shape);
-      auto in_stride = framework::stride(out_flatten_level.dims());
-      const T* src_data = out_flatten_level.data<T>();
-      StridedMemcpy<T>(context.device_context(), src_data, in_stride,
-                       out_flatten_level.dims(), out_stride,
-                       out->data<T>() + output_offset);
-      output_offset += out_flatten_level.dims()[1] * in_stride[1];
+      // concat
+      auto out_flatten_level_stride =
+          framework::stride(out_flatten_level.dims());
+      StridedMemcpy<T>(context.device_context(), out_flatten_level.data<T>(),
+                       out_flatten_level_stride, out_flatten_level.dims(),
+                       out_stride, out->data<T>() + output_offset);
+      output_offset +=
+          out_flatten_level.dims()[1] * out_flatten_level_stride[1];
     }
   }
 };
@@ -83,12 +85,11 @@ class SppGradKernel : public framework::OpKernel<T> {
         context.Input<framework::Tensor>(framework::GradVarName("Out"));
     framework::Tensor* in_x_grad =
         context.Output<framework::Tensor>(framework::GradVarName("X"));
+    int pyramid_height = context.template Attr<int>("pyramid_height");
     auto& device_ctx = context.device_context();
     math::SetConstant<Place, T> zero;
     in_x_grad->mutable_data<T>(context.GetPlace());
     zero(device_ctx, in_x_grad, static_cast<T>(0));
-    int pyramid_height = context.template Attr<int>("pyramid_height");
-    auto outgrad_stride = framework::stride(out_grad->dims());
     auto out_stride = framework::stride(out->dims());
     int input_h = in_x->dims()[2];
     int input_w = in_x->dims()[3];
@@ -102,26 +103,17 @@ class SppGradKernel : public framework::OpKernel<T> {
       std::vector<int> ksize({ksize_h, ksize_w});
       std::vector<int> strides({ksize_h, ksize_w});
       std::vector<int> paddings({padding_h, padding_w});
-      // split outgrad and get flatten
-      std::vector<int64_t> out_shape_vec({in_x->dims()[0], in_x->dims()[1]});
-      out_shape_vec.push_back((input_h - ksize_h + 2 * padding_h) / ksize_h +
-                              1);
-      out_shape_vec.push_back((input_w - ksize_w + 2 * padding_w) / ksize_w +
-                              1);
-      framework::DDim out_shape(framework::make_ddim(out_shape_vec));
+      // split out and outgrad  ...  to flatten
+      framework::Tensor out_flatten_level;
+      framework::Tensor outgrad_flatten_level;
       int out_flatten_w = in_x->dims()[1] * bins * bins;
       std::vector<int64_t> out_flatten_shape_vec(
           {in_x->dims()[0], out_flatten_w});
       framework::DDim out_flatten_shape(
           framework::make_ddim(out_flatten_shape_vec));
-      framework::Tensor out_level;
-      framework::Tensor outgrad_level;
-      framework::Tensor out_flatten_level;
-      framework::Tensor outgrad_flatten_level;
       out_flatten_level.mutable_data<T>(out_flatten_shape, context.GetPlace());
       outgrad_flatten_level.mutable_data<T>(out_flatten_shape,
                                             context.GetPlace());
-
       auto flatten_stride = framework::stride(out_flatten_level.dims());
       // memcpy
       StridedMemcpy<T>(context.device_context(), out->data<T>() + out_offset,
@@ -129,15 +121,24 @@ class SppGradKernel : public framework::OpKernel<T> {
                        out_flatten_level.data<T>());
 
       StridedMemcpy<T>(context.device_context(),
-                       out_grad->data<T>() + out_offset, outgrad_stride,
+                       out_grad->data<T>() + out_offset, out_stride,
                        outgrad_flatten_level.dims(), flatten_stride,
                        outgrad_flatten_level.data<T>());
       out_offset += out_flatten_level.dims()[1] * out_stride[1];
-      // flatten backward
+      // flatten backward to nchw
+      framework::Tensor out_level;
+      framework::Tensor outgrad_level;
+      std::vector<int64_t> out_shape_vec({in_x->dims()[0], in_x->dims()[1]});
+      out_shape_vec.push_back((input_h - ksize_h + 2 * padding_h) / ksize_h +
+                              1);
+      out_shape_vec.push_back((input_w - ksize_w + 2 * padding_w) / ksize_w +
+                              1);
+      framework::DDim out_shape(framework::make_ddim(out_shape_vec));
       out_level.ShareDataWith(out_flatten_level);
       out_level.Resize(out_shape);
       outgrad_level.ShareDataWith(outgrad_flatten_level);
       outgrad_level.Resize(out_shape);
+      // pooling backward
       math::MaxPool2dGradFunctor<Place, T> pool2d_backward;
       pool2d_backward(context.device_context(), *in_x, *&out_level,
                       *&outgrad_level, ksize, strides, paddings, in_x_grad);
diff --git a/python/paddle/v2/fluid/tests/test_spp_op.py b/python/paddle/v2/fluid/tests/test_spp_op.py
index 806d5e7736..89b12e885c 100644
--- a/python/paddle/v2/fluid/tests/test_spp_op.py
+++ b/python/paddle/v2/fluid/tests/test_spp_op.py
@@ -37,11 +37,11 @@ class TestSppOp(OpTest):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out')
+        self.check_grad(['X'], 'Out', max_relative_error=0.05)
 
     def init_test_case(self):
-        self.shape = [1, 1, 2, 2]
-        self.pyramid_height = 2
+        self.shape = [3, 2, 4, 4]
+        self.pyramid_height = 3
 
 
 if __name__ == '__main__':

From 8368e55be93ea6d2912e3ebebb35e284c5428a28 Mon Sep 17 00:00:00 2001
From: sweetsky0901 <work@yq01-idl-gpu-online20.yq01.baidu.com>
Date: Mon, 4 Dec 2017 11:19:16 +0800
Subject: [PATCH 10/94] modify some doc

---
 paddle/operators/spp_op.cc                  |  4 +-
 paddle/operators/spp_op.h                   | 47 +++++++++++----------
 python/paddle/v2/fluid/tests/test_spp_op.py | 19 ++++++---
 3 files changed, 38 insertions(+), 32 deletions(-)

diff --git a/paddle/operators/spp_op.cc b/paddle/operators/spp_op.cc
index ff607c5769..026b35de1e 100644
--- a/paddle/operators/spp_op.cc
+++ b/paddle/operators/spp_op.cc
@@ -29,7 +29,7 @@ class SppOpMaker : public framework::OpProtoAndCheckerMaker {
               "(Tensor) The output tensor of spp operator."
               "N * M."
               "M = C * H * W");
-    AddAttr<int>("pyramid_height", "int");
+    AddAttr<int>("pyramid_height", "int", "multi level pooling");
     AddComment(R"DOC(
         "Does spatial pyramid pooling on the input image by taking the max,
         etc. within regions so that the result vector of different sized
@@ -39,7 +39,7 @@ class SppOpMaker : public framework::OpProtoAndCheckerMaker {
         Where
           $$
             H_{out} = N \\
-            W_{out} = ((std::pow(4, pyramid_height) - 1) / (4 - 1)) * C_{in}
+            W_{out} = (((4^pyramid_height) - 1) / (4 - 1))$ * C_{in}
           $$
         )DOC");
   }
diff --git a/paddle/operators/spp_op.h b/paddle/operators/spp_op.h
index 7a385352a0..0f2c43ee65 100644
--- a/paddle/operators/spp_op.h
+++ b/paddle/operators/spp_op.h
@@ -34,27 +34,27 @@ class SppKernel : public framework::OpKernel<T> {
     size_t output_offset = 0;
     for (int p = 0; p < pyramid_height; ++p) {
       int bins = std::pow(2, p);
-      int ksize_h = std::ceil(input_h / static_cast<double>(bins));
-      int ksize_w = std::ceil(input_w / static_cast<double>(bins));
-      int padding_h = (ksize_h * bins - input_h + 1) / 2;
-      int padding_w = (ksize_w * bins - input_w + 1) / 2;
-      std::vector<int> ksize({ksize_h, ksize_w});
-      std::vector<int> strides({ksize_h, ksize_w});
+      int kernel_size_h = std::ceil(input_h / static_cast<double>(bins));
+      int kernel_size_w = std::ceil(input_w / static_cast<double>(bins));
+      int padding_h = (kernel_size_h * bins - input_h + 1) / 2;
+      int padding_w = (kernel_size_w * bins - input_w + 1) / 2;
+      std::vector<int> kernel_size({kernel_size_h, kernel_size_w});
+      std::vector<int> strides({kernel_size_h, kernel_size_w});
       std::vector<int> paddings({padding_h, padding_w});
       // pooling output shape
       framework::Tensor out_level;
       std::vector<int64_t> output_shape_vec({in_x->dims()[0], in_x->dims()[1]});
-      output_shape_vec.push_back((input_h - ksize_h + 2 * padding_h) / ksize_h +
-                                 1);
-      output_shape_vec.push_back((input_w - ksize_w + 2 * padding_w) / ksize_w +
-                                 1);
+      output_shape_vec.push_back(
+          (input_h - kernel_size_h + 2 * padding_h) / kernel_size_h + 1);
+      output_shape_vec.push_back(
+          (input_w - kernel_size_w + 2 * padding_w) / kernel_size_w + 1);
       framework::DDim output_shape(framework::make_ddim(output_shape_vec));
       out_level.mutable_data<T>(output_shape, context.GetPlace());
       // pooling
       math::Pool2dFunctor<Place, math::MaxPool<T>, T> pool_forward;
       math::MaxPool<T> max_process;
-      pool_forward(context.device_context(), *in_x, ksize, strides, paddings,
-                   max_process, &out_level);
+      pool_forward(context.device_context(), *in_x, kernel_size, strides,
+                   paddings, max_process, &out_level);
       // flatten pooling output shape
       framework::Tensor out_flatten_level;
       int output_flatten_w = in_x->dims()[1] * bins * bins;
@@ -96,12 +96,12 @@ class SppGradKernel : public framework::OpKernel<T> {
     size_t out_offset = 0;
     for (int p = 0; p < pyramid_height; ++p) {
       int bins = std::pow(2, p);
-      int ksize_h = std::ceil(input_h / static_cast<double>(bins));
-      int ksize_w = std::ceil(input_w / static_cast<double>(bins));
-      int padding_h = (ksize_h * bins - input_h + 1) / 2;
-      int padding_w = (ksize_w * bins - input_w + 1) / 2;
-      std::vector<int> ksize({ksize_h, ksize_w});
-      std::vector<int> strides({ksize_h, ksize_w});
+      int kernel_size_h = std::ceil(input_h / static_cast<double>(bins));
+      int kernel_size_w = std::ceil(input_w / static_cast<double>(bins));
+      int padding_h = (kernel_size_h * bins - input_h + 1) / 2;
+      int padding_w = (kernel_size_w * bins - input_w + 1) / 2;
+      std::vector<int> kernel_size({kernel_size_h, kernel_size_w});
+      std::vector<int> strides({kernel_size_h, kernel_size_w});
       std::vector<int> paddings({padding_h, padding_w});
       // split out and outgrad  ...  to flatten
       framework::Tensor out_flatten_level;
@@ -129,10 +129,10 @@ class SppGradKernel : public framework::OpKernel<T> {
       framework::Tensor out_level;
       framework::Tensor outgrad_level;
       std::vector<int64_t> out_shape_vec({in_x->dims()[0], in_x->dims()[1]});
-      out_shape_vec.push_back((input_h - ksize_h + 2 * padding_h) / ksize_h +
-                              1);
-      out_shape_vec.push_back((input_w - ksize_w + 2 * padding_w) / ksize_w +
-                              1);
+      out_shape_vec.push_back(
+          (input_h - kernel_size_h + 2 * padding_h) / kernel_size_h + 1);
+      out_shape_vec.push_back(
+          (input_w - kernel_size_w + 2 * padding_w) / kernel_size_w + 1);
       framework::DDim out_shape(framework::make_ddim(out_shape_vec));
       out_level.ShareDataWith(out_flatten_level);
       out_level.Resize(out_shape);
@@ -141,7 +141,8 @@ class SppGradKernel : public framework::OpKernel<T> {
       // pooling backward
       math::MaxPool2dGradFunctor<Place, T> pool2d_backward;
       pool2d_backward(context.device_context(), *in_x, *&out_level,
-                      *&outgrad_level, ksize, strides, paddings, in_x_grad);
+                      *&outgrad_level, kernel_size, strides, paddings,
+                      in_x_grad);
     }
   }
 };
diff --git a/python/paddle/v2/fluid/tests/test_spp_op.py b/python/paddle/v2/fluid/tests/test_spp_op.py
index 89b12e885c..b57f4a795d 100644
--- a/python/paddle/v2/fluid/tests/test_spp_op.py
+++ b/python/paddle/v2/fluid/tests/test_spp_op.py
@@ -13,14 +13,19 @@ class TestSppOp(OpTest):
         out_level_flatten = []
         for i in xrange(self.pyramid_height):
             bins = np.power(2, i)
-            ksize = [0, 0]
+            kernel_size = [0, 0]
             padding = [0, 0]
-            ksize[0] = np.ceil(hsize / bins.astype("double")).astype("int32")
-            padding[0] = ((ksize[0] * bins - hsize + 1) / 2).astype("int32")
-
-            ksize[1] = np.ceil(wsize / bins.astype("double")).astype("int32")
-            padding[1] = ((ksize[1] * bins - wsize + 1) / 2).astype("int32")
-            out_level = max_pool2D_forward_naive(input, ksize, ksize, padding)
+            kernel_size[0] = np.ceil(hsize /
+                                     bins.astype("double")).astype("int32")
+            padding[0] = (
+                (kernel_size[0] * bins - hsize + 1) / 2).astype("int32")
+
+            kernel_size[1] = np.ceil(wsize /
+                                     bins.astype("double")).astype("int32")
+            padding[1] = (
+                (kernel_size[1] * bins - wsize + 1) / 2).astype("int32")
+            out_level = max_pool2D_forward_naive(input, kernel_size,
+                                                 kernel_size, padding)
             out_level_flatten.append(
                 out_level.reshape(nsize, bins * bins * csize))
             if i == 0:

From 141a323c3410f9d65b75cd546af69715735db23a Mon Sep 17 00:00:00 2001
From: sweetsky0901 <work@yq01-idl-gpu-online20.yq01.baidu.com>
Date: Mon, 4 Dec 2017 11:43:14 +0800
Subject: [PATCH 11/94] fix a bug

---
 paddle/operators/spp_op.cc | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/paddle/operators/spp_op.cc b/paddle/operators/spp_op.cc
index 026b35de1e..5e51e73ecc 100644
--- a/paddle/operators/spp_op.cc
+++ b/paddle/operators/spp_op.cc
@@ -29,7 +29,7 @@ class SppOpMaker : public framework::OpProtoAndCheckerMaker {
               "(Tensor) The output tensor of spp operator."
               "N * M."
               "M = C * H * W");
-    AddAttr<int>("pyramid_height", "int", "multi level pooling");
+    AddAttr<int>("pyramid_height", "(int), multi level pooling");
     AddComment(R"DOC(
         "Does spatial pyramid pooling on the input image by taking the max,
         etc. within regions so that the result vector of different sized

From 4f1381eac3708ce92b07a01f6cfc9d4131c996af Mon Sep 17 00:00:00 2001
From: typhoonzero <typhoonzero1986@gmail.com>
Date: Fri, 8 Dec 2017 16:20:09 +0800
Subject: [PATCH 12/94] recv_op use serialized program

---
 paddle/operators/recv_op.cc           | 11 +++++++----
 paddle/operators/send_recv_op_test.cc |  2 +-
 2 files changed, 8 insertions(+), 5 deletions(-)

diff --git a/paddle/operators/recv_op.cc b/paddle/operators/recv_op.cc
index c69e416e10..45222f6b76 100644
--- a/paddle/operators/recv_op.cc
+++ b/paddle/operators/recv_op.cc
@@ -72,8 +72,10 @@ class RecvOp : public framework::OperatorBase {
     // FIXME(typhoonzero): do not copy
     framework::CopyFrom(t, dev_ctx.GetPlace(), dev_ctx, tensor);
 
-    auto *block = Attr<framework::BlockDescBind *>("OptimizeBlock");
-    auto *program = block->Program();
+    std::string program_str = Attr<std::string>("OptimizeProgram");
+    framework::Program program_desc;
+    program_desc.ParseFromString(program_str);
+    framework::ProgramDescBind program(program_desc);
     framework::Executor executor(dev_ctx);
     // Run sub graph to get optimized tensor
     executor.Run(*program, &recv_scope, block->ID(),
@@ -108,8 +110,9 @@ This operator will recv tensor from send_op
                          "IP address to listen on.")
         .SetDefault("127.0.0.1:6164")
         .AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
-    AddAttr<framework::BlockDescBind *>("OptimizeBlock", "type BlockDescBind*",
-                                        "optimize network run in server");
+    AddAttr<framework::BlockDescBind *>(
+        "OptimizeProgram", "type string",
+        "Serialized ProgramDesc string for recv to run.");
   }
 };
 
diff --git a/paddle/operators/send_recv_op_test.cc b/paddle/operators/send_recv_op_test.cc
index ac03eb3752..c35dc8fa50 100644
--- a/paddle/operators/send_recv_op_test.cc
+++ b/paddle/operators/send_recv_op_test.cc
@@ -85,7 +85,7 @@ void StartServerNet() {
 
   paddle::framework::AttributeMap attrs;
   attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
-  attrs.insert({"OptimizeBlock", block});
+  attrs.insert({"OptimizeProgram", program.Proto()->SerializeToString()});
   recv_op = paddle::framework::OpRegistry::CreateOp("recv", {{"RX", {"RX"}}},
                                                     {{"Out", {"Out"}}}, attrs);
   paddle::platform::CPUDeviceContext ctx(place);

From 986ca03ce24f6d84eb9cfaef64b59fda2298823b Mon Sep 17 00:00:00 2001
From: typhoonzero <typhoonzero1986@gmail.com>
Date: Fri, 8 Dec 2017 19:45:15 +0800
Subject: [PATCH 13/94] update

---
 paddle/operators/recv_op.cc           | 9 ++++-----
 paddle/operators/send_recv_op_test.cc | 5 ++++-
 2 files changed, 8 insertions(+), 6 deletions(-)

diff --git a/paddle/operators/recv_op.cc b/paddle/operators/recv_op.cc
index 45222f6b76..eed482c1b4 100644
--- a/paddle/operators/recv_op.cc
+++ b/paddle/operators/recv_op.cc
@@ -73,12 +73,12 @@ class RecvOp : public framework::OperatorBase {
     framework::CopyFrom(t, dev_ctx.GetPlace(), dev_ctx, tensor);
 
     std::string program_str = Attr<std::string>("OptimizeProgram");
-    framework::Program program_desc;
+    framework::ProgramDesc program_desc;
     program_desc.ParseFromString(program_str);
     framework::ProgramDescBind program(program_desc);
     framework::Executor executor(dev_ctx);
     // Run sub graph to get optimized tensor
-    executor.Run(*program, &recv_scope, block->ID(),
+    executor.Run(program, &recv_scope, 0, /*global_block*/
                  false /*create_local_scope*/);
 
     auto *out_var = recv_scope.FindVar("Out");
@@ -110,9 +110,8 @@ This operator will recv tensor from send_op
                          "IP address to listen on.")
         .SetDefault("127.0.0.1:6164")
         .AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
-    AddAttr<framework::BlockDescBind *>(
-        "OptimizeProgram", "type string",
-        "Serialized ProgramDesc string for recv to run.");
+    AddAttr<std::string>("OptimizeProgram", "type string",
+                         "Serialized ProgramDesc string for recv to run.");
   }
 };
 
diff --git a/paddle/operators/send_recv_op_test.cc b/paddle/operators/send_recv_op_test.cc
index c35dc8fa50..3e2e2051af 100644
--- a/paddle/operators/send_recv_op_test.cc
+++ b/paddle/operators/send_recv_op_test.cc
@@ -85,7 +85,10 @@ void StartServerNet() {
 
   paddle::framework::AttributeMap attrs;
   attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
-  attrs.insert({"OptimizeProgram", program.Proto()->SerializeToString()});
+  std::string program_proto;
+  PADDLE_ENFORCE(program.Proto()->SerializeToString(&program_proto));
+
+  attrs.insert({"OptimizeProgram", program_proto});
   recv_op = paddle::framework::OpRegistry::CreateOp("recv", {{"RX", {"RX"}}},
                                                     {{"Out", {"Out"}}}, attrs);
   paddle::platform::CPUDeviceContext ctx(place);

From 578ad6d23251c0fc08cbf49aa1d6bf9daae55a88 Mon Sep 17 00:00:00 2001
From: hedaoyuan <hedaoyuan@github.com>
Date: Mon, 11 Dec 2017 11:21:47 +0800
Subject: [PATCH 14/94] Use PADDLE_WITH_NATIVE_FP16 for float16_t.

---
 paddle/math/float16.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/paddle/math/float16.h b/paddle/math/float16.h
index f805cad08b..76ad3a0123 100644
--- a/paddle/math/float16.h
+++ b/paddle/math/float16.h
@@ -101,7 +101,7 @@ public:
     half tmp = __float2half(val);
     x = *reinterpret_cast<uint16_t*>(&tmp);
 
-#elif defined(PADDLE_NEON)
+#elif defined(PADDLE_WITH_NATIVE_FP16)
     float32x4_t tmp = vld1q_dup_f32(&val);
     float16_t res = vget_lane_f16(vcvt_f16_f32(tmp), 0);
     x = *reinterpret_cast<uint16_t*>(&res);
@@ -252,7 +252,7 @@ public:
     half tmp = *reinterpret_cast<const half*>(this);
     return __half2float(tmp);
 
-#elif defined(PADDLE_NEON)
+#elif defined(PADDLE_WITH_NATIVE_FP16)
     float16x4_t res = vld1_dup_f16(reinterpret_cast<const float16_t*>(this));
     return vgetq_lane_f32(vcvt_f32_f16(res), 0);
 

From a6ef875885486eea5573e4da52ec2c1c02f1d1e7 Mon Sep 17 00:00:00 2001
From: chengduoZH <zhaochengduo@163.com>
Date: Mon, 11 Dec 2017 14:05:15 +0800
Subject: [PATCH 15/94] refine conv

---
 paddle/operators/conv_op.h           | 7 +++++--
 paddle/operators/conv_transpose_op.h | 2 +-
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/paddle/operators/conv_op.h b/paddle/operators/conv_op.h
index 09bff0a68d..66728d75a6 100644
--- a/paddle/operators/conv_op.h
+++ b/paddle/operators/conv_op.h
@@ -260,8 +260,11 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
 
     if (input_grad) {
       input_grad->mutable_data<T>(context.GetPlace());
-      set_zero(context.device_context(), input_grad, static_cast<T>(0));
-
+      // if is_expand is false, the operation of set_zero is unnecessary,
+      // because math::matmul will reset input_grad.
+      if (is_expand) {
+        set_zero(context.device_context(), input_grad, static_cast<T>(0));
+      }
       math::Col2VolFunctor<Place, T> col2vol;
       math::Col2ImFunctor<math::ColFormat::kCFO, Place, T> col2im;
 
diff --git a/paddle/operators/conv_transpose_op.h b/paddle/operators/conv_transpose_op.h
index 1cacb770e6..a43dd5b8c0 100644
--- a/paddle/operators/conv_transpose_op.h
+++ b/paddle/operators/conv_transpose_op.h
@@ -225,7 +225,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
 
       if (input_grad) {
         input_grad->mutable_data<T>(context.GetPlace());
-        set_zero(context.device_context(), input_grad, static_cast<T>(0));
+        // set_zero is unnecessary, math::matmul will reset input_grad.
       }
       if (filter_grad) {  // filter size (m, c, k_h, k_w)
         filter_grad->mutable_data<T>(context.GetPlace());

From 8d428bd9b89ec59dfcf16eb9e319a9106415b766 Mon Sep 17 00:00:00 2001
From: ranqiu <ranqiu@baidu.com>
Date: Mon, 4 Dec 2017 19:28:12 +0800
Subject: [PATCH 16/94] Update annotations of layers.py

---
 .../paddle/trainer_config_helpers/layers.py   | 121 ++++++++++--------
 1 file changed, 66 insertions(+), 55 deletions(-)

diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
index 4bd94861af..48858f4c34 100644
--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -1516,34 +1516,33 @@ def lstmemory(input,
     NOTE: This is a low level user interface. You can use network.simple_lstm
     to config a simple plain lstm layer.
 
-    Please refer to **Generating Sequences With Recurrent Neural Networks** for
-    more details about LSTM.
-
-    Link_ goes as below.
-
-    .. _Link: http://arxiv.org/abs/1308.0850
+    Reference:
+        `Generating Sequences With Recurrent Neural Networks
+        <https://arxiv.org/pdf/1308.0850.pdf>`_
 
-    :param name: The lstmemory layer name.
+    :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param size: DEPRECATED. size of the lstm cell
+    :param size: DEPRECATED. The dimension of the lstm cell.
     :type size: int
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param reverse: is sequence process reversed or not.
+    :param reverse: Whether the input sequence is processed in a reverse order.
     :type reverse: bool
     :param act: Activation type. TanhActivation is the default activation.
     :type act: BaseActivation
-    :param gate_act: gate activation type, SigmoidActivation by default.
+    :param gate_act: Activation type of this layer's gates. SigmoidActivation is the
+                     default activation.
     :type gate_act: BaseActivation
-    :param state_act: state activation type, TanhActivation by default.
+    :param state_act: Activation type of the state. TanhActivation is the default activation.
     :type state_act: BaseActivation
     :param bias_attr: The bias attribute. If the parameter is set to False or an object
                       whose type is not ParameterAttribute, no bias is defined. If the
                       parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
-    :param param_attr: Parameter Attribute.
-    :type param_attr: ParameterAttribute | None | False
-    :param layer_attr: Extra Layer attribute
+    :param param_attr: The parameter attribute. See ParameterAttribute for details.
+    :type param_attr: ParameterAttribute
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -1632,14 +1631,14 @@ def grumemory(input,
         h_t = (1 - z_t) h_{t-1} + z_t {\\tilde{h_t}}
 
     NOTE: In PaddlePaddle's implementation, the multiplication operations
-    :math:`W_{r}x_{t}`, :math:`W_{z}x_{t}` and :math:`W x_t` are not computed in
-    gate_recurrent layer. Consequently, an additional mixed_layer with
+    :math:`W_{r}x_{t}`, :math:`W_{z}x_{t}` and :math:`W x_t` are not performed
+    in gate_recurrent layer. Consequently, an additional mixed_layer with
     full_matrix_projection or a fc_layer must be included before grumemory
     is called.
 
-    More details can be found by referring to `Empirical Evaluation of Gated
-    Recurrent Neural Networks on Sequence Modeling.
-    <https://arxiv.org/abs/1412.3555>`_
+    Reference:
+        `Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling
+        <https://arxiv.org/abs/1412.3555>`_
 
     The simple usage is:
 
@@ -1647,28 +1646,29 @@ def grumemory(input,
 
        gru = grumemory(input)
 
-    :param name: The gru layer name.
-    :type name: None | basestring
+    :param name: The name of this layer. It is optional.
+    :type name: basestring
     :param input: The input of this layer.
     :type input: LayerOutput.
-    :param size: DEPRECATED. size of the gru cell
+    :param size: DEPRECATED. The dimension of the gru cell.
     :type size: int
-    :param reverse: Whether sequence process is reversed or not.
+    :param reverse: Whether the input sequence is processed in a reverse order.
     :type reverse: bool
     :param act: Activation type, TanhActivation is the default. This activation
                 affects the :math:`{\\tilde{h_t}}`.
     :type act: BaseActivation
-    :param gate_act: gate activation type, SigmoidActivation by default.
-                     This activation affects the :math:`z_t` and :math:`r_t`. It is the
-                     :math:`\\sigma` in the above formula.
+    :param gate_act: Activation type of this layer's two gates. SigmoidActivation is
+                     the default activation. This activation affects the :math:`z_t`
+                     and :math:`r_t`. It is the :math:`\\sigma` in the above formula.
     :type gate_act: BaseActivation
     :param bias_attr: The bias attribute. If the parameter is set to False or an object
                       whose type is not ParameterAttribute, no bias is defined. If the
                       parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
-    :param param_attr: Parameter Attribute.
-    :type param_attr: ParameterAttribute | None | False
-    :param layer_attr: Extra Layer attribute
+    :param param_attr: The parameter attribute. See ParameterAttribute for details.
+    :type param_attr: ParameterAttribute
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -1712,10 +1712,10 @@ def last_seq(input,
     """
     Get Last Timestamp Activation of a sequence.
 
-    If stride > 0, this layer slides a window whose size is determined by stride,
-    and return the last value of the window as the output. Thus, a long sequence
-    will be shorten. Note that for sequence with sub-sequence, the default value
-    of stride is -1.
+    If stride > 0, this layer will slide a window whose size is determined by stride,
+    and return the last value of the sequence in the window as the output. Thus, a
+    long sequence will be shortened. Note that for sequence with sub-sequence, the
+    default value of stride is -1.
 
     The simple usage is:
 
@@ -1724,14 +1724,16 @@ def last_seq(input,
        seq = last_seq(input=layer)
 
     :param agg_level: Aggregated level
+    :type agg_level: AggregateLevel
     :param name: The name of this layer. It is optional.
     :type name: basestring
     :param input: The input of this layer.
     :type input: LayerOutput
     :param stride: The step size between successive pooling regions.
-    :type stride: Int
-    :param layer_attr: extra layer attributes.
-    :type layer_attr: ExtraLayerAttribute.
+    :type stride: int
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
+    :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -1768,10 +1770,10 @@ def first_seq(input,
     """
     Get First Timestamp Activation of a sequence.
 
-    If stride > 0, this layer slides a window whose size is determined by stride,
-    and return the first value of the window as the output. Thus, a long sequence
-    will be shorten. Note that for sequence with sub-sequence, the default value
-    of stride is -1.
+    If stride > 0, this layer will slide a window whose size is determined by stride,
+    and return the first value of the sequence in the window as the output. Thus, a
+    long sequence will be shortened. Note that for sequence with sub-sequence, the
+    default value of stride is -1.
 
     The simple usage is:
 
@@ -1780,13 +1782,15 @@ def first_seq(input,
        seq = first_seq(input=layer)
 
     :param agg_level: aggregation level
+    :type agg_level: AggregateLevel
     :param name: The name of this layer. It is optional.
     :type name: basestring
     :param input: The input of this layer.
     :type input: LayerOutput
     :param stride: The step size between successive pooling regions.
-    :type stride: Int
-    :param layer_attr: extra layer attributes.
+    :type stride: int
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute.
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -1844,8 +1848,8 @@ def expand_layer(input,
                  expand_level=ExpandLevel.FROM_NO_SEQUENCE,
                  layer_attr=None):
     """
-    A layer for "Expand Dense data or (sequence data where the length of each
-    sequence is one) to sequence data."
+    A layer for expanding dense data or (sequence data where the length of each
+    sequence is one) to sequence data.
 
     The example usage is:
 
@@ -1857,7 +1861,9 @@ def expand_layer(input,
 
     :param input: The input of this layer.
     :type input: LayerOutput
-    :param expand_as: Expand as this layer's sequence info.
+    :param expand_as: Expand the input according to this layer's sequence infomation. And
+                      after the operation, the input expanded will have the same number of
+                      elememts as this layer.
     :type expand_as: LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
@@ -1865,9 +1871,10 @@ def expand_layer(input,
                       whose type is not ParameterAttribute, no bias is defined. If the
                       parameter is set to True, the bias is initialized to zero.
     :type bias_attr: ParameterAttribute | None | bool | Any
-    :param expand_level: whether input layer is timestep(default) or sequence.
+    :param expand_level: Whether the input layer is a sequence or the element of a sequence.
     :type expand_level: ExpandLevel
-    :param layer_attr: extra layer attributes.
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
     :type layer_attr: ExtraLayerAttribute.
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -3294,7 +3301,7 @@ def row_l2_norm_layer(input, name=None, layer_attr=None):
     A layer for L2-normalization in each row.
 
     .. math::
-       out[i] = \frac{in[i]}{\sqrt{\sum_{k=1}^N in[k]^{2}}}
+       out[i] = \\frac{in[i]} {\\sqrt{\\sum_{k=1}^N in[k]^{2}}}
 
     where the size of :math:`in` is (batchSize x dataDim) ,
     and the size of :math:`out` is a (batchSize x dataDim) .
@@ -6161,9 +6168,11 @@ def huber_regression_cost(input,
     Given a prediction f(x), a label y and :math:`\delta`, the loss function
     is defined as:
 
-    .. math:
-       loss = 0.5*\left ( y-f(x) \right )^2, \left | y-f(x) \right |\leq \delta
-       loss = \delta \left | y-f(x) \right |-0.5\delta ^2, otherwise
+    .. math::
+
+       loss = 0.5*(y-f(x))^{2}, | y-f(x) | < \delta
+
+       loss = \delta | y-f(x) | - 0.5 \delta ^2, otherwise
 
     The example usage is:
 
@@ -6210,12 +6219,14 @@ def huber_classification_cost(input,
     """
     For classification purposes, a variant of the Huber loss called modified Huber
     is sometimes used. Given a prediction f(x) (a real-valued classifier score) and
-    a true binary class label :math:`y\in \left \{-1, 1 \right \}`, the modified Huber
+    a true binary class label :math:`y\in \{-1, 1 \}`, the modified Huber
     loss is defined as:
 
     .. math:
-       loss = \max \left ( 0, 1-yf(x) \right )^2, yf(x)\geq 1
-       loss = -4yf(x), \text{otherwise}
+
+       loss = \max ( 0, 1-yf(x) )^2, yf(x) \geq -1
+
+       loss = -4yf(x), otherwise
 
     The example usage is:
 
@@ -6959,7 +6970,7 @@ def clip_layer(input, min, max, name=None):
 
     .. math::
 
-        out[i] = \min\left(\max\left(in[i],p_{1}\right),p_{2}\right)
+        out[i] = \min (\max (in[i],p_{1} ),p_{2} )
 
     .. code-block:: python
 

From ddf20e589fad724f077b0613ebf3872d2311647a Mon Sep 17 00:00:00 2001
From: Luo Tao <luotao02@baidu.com>
Date: Mon, 11 Dec 2017 18:31:14 +0800
Subject: [PATCH 17/94] typo WITH_TEST to WITH_TESTING

---
 doc/howto/dev/contribute_to_paddle_cn.md | 2 +-
 paddle/scripts/docker/README.md          | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/doc/howto/dev/contribute_to_paddle_cn.md b/doc/howto/dev/contribute_to_paddle_cn.md
index 6993901452..3eb477eb65 100644
--- a/doc/howto/dev/contribute_to_paddle_cn.md
+++ b/doc/howto/dev/contribute_to_paddle_cn.md
@@ -87,7 +87,7 @@ no changes added to commit (use "git add" and/or "git commit -a")
 随后可以用这个开发镜像开始build PaddlePaddle的源码。比如如果要build一个不依赖GPU，但是支持AVX指令集，并且包括unit tests的PaddlePaddle，可以：
 
 ```bash
-➜  docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TEST=ON" paddle:dev
+➜  docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TESTING=ON" paddle:dev
 ```
 
 这个过程除了编译PaddlePaddle为 `./build/libpaddle.so`，并且输出一个 `./build/paddle.deb`文件之外，还会输出一个 `build/Dockerfile`。我们只需要运行下面命令把编译好的PaddlePaddle打包成一个*生产镜像*（`paddle:prod`）：
diff --git a/paddle/scripts/docker/README.md b/paddle/scripts/docker/README.md
index f3a6f1dba7..1e1fcc50dc 100644
--- a/paddle/scripts/docker/README.md
+++ b/paddle/scripts/docker/README.md
@@ -192,7 +192,7 @@ For developers who are interested in the C++ source code, please use -e "WOBOQ=O
 - The following command builds PaddlePaddle, generates HTML pages from C++ source code, and writes HTML pages into `$HOME/woboq_out` on the host:
 
 ```bash
-docker run -v $PWD:/paddle -v $HOME/woboq_out:/woboq_out -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TEST=ON" -e "WOBOQ=ON" paddlepaddle/paddle:latest-dev
+docker run -v $PWD:/paddle -v $HOME/woboq_out:/woboq_out -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TESTING=ON" -e "WOBOQ=ON" paddlepaddle/paddle:latest-dev
 ```
 
 - You can open the generated HTML files in your Web browser. Or, if you want to run a Nginx container to serve them for a wider audience, you can run:

From 9f44af9d7c9dd2f1dc775f59d059a87eb9e64fd6 Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Mon, 11 Dec 2017 18:37:18 +0800
Subject: [PATCH 18/94] Fix #6460 (#6461)

---
 python/paddle/v2/fluid/layers.py | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)

diff --git a/python/paddle/v2/fluid/layers.py b/python/paddle/v2/fluid/layers.py
index b4426bad14..fd8a2ed18c 100644
--- a/python/paddle/v2/fluid/layers.py
+++ b/python/paddle/v2/fluid/layers.py
@@ -762,7 +762,7 @@ def sequence_conv(input,
     helper = LayerHelper('sequence_conv', **locals())
     dtype = helper.input_dtype()
     filter_shape = [filter_size * input.shape[1], num_filters]
-    filter = helper.create_parameter(
+    filter_param = helper.create_parameter(
         attr=helper.param_attr, shape=filter_shape, dtype=dtype)
     pre_bias = helper.create_tmp_variable(dtype)
 
@@ -770,7 +770,7 @@ def sequence_conv(input,
         type='sequence_conv',
         inputs={
             'X': [input],
-            'Filter': [filter],
+            'Filter': [filter_param],
         },
         outputs={"Out": pre_bias},
         attrs={
@@ -785,7 +785,7 @@ def sequence_conv(input,
 def conv2d(input,
            num_filters,
            filter_size,
-           stride=[1, 1],
+           stride=None,
            padding=None,
            groups=None,
            param_attr=None,
@@ -802,6 +802,8 @@ def conv2d(input,
     conv-2d output, if mentioned in the input parameters.
     """
 
+    if stride is None:
+        stride = [1, 1]
     helper = LayerHelper('conv2d', **locals())
     dtype = helper.input_dtype()
 
@@ -827,7 +829,7 @@ def conv2d(input,
         std = (2.0 / (filter_size[0]**2 * num_channels))**0.5
         return Normal(0.0, std, 0)
 
-    filter = helper.create_parameter(
+    filter_param = helper.create_parameter(
         attr=helper.param_attr,
         shape=filter_shape,
         dtype=dtype,
@@ -839,7 +841,7 @@ def conv2d(input,
         type='conv2d_cudnn',
         inputs={
             'Input': input,
-            'Filter': filter,
+            'Filter': filter_param,
         },
         outputs={"Output": pre_bias},
         attrs={'strides': stride,
@@ -875,8 +877,8 @@ def sequence_pool(input, pool_type, **kwargs):
 def pool2d(input,
            pool_size,
            pool_type,
-           pool_stride=[1, 1],
-           pool_padding=[0, 0],
+           pool_stride=None,
+           pool_padding=None,
            global_pooling=False,
            main_program=None,
            startup_program=None):
@@ -884,6 +886,10 @@ def pool2d(input,
     This function adds the operator for pooling in 2 dimensions, using the
     pooling configurations mentioned in input parameters.
     """
+    if pool_padding is None:
+        pool_padding = [0, 0]
+    if pool_stride is None:
+        pool_stride = [1, 1]
     if pool_type not in ["max", "avg"]:
         raise ValueError(
             "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",

From 7389ea98eaae7ca79442811991cba37f908e5a74 Mon Sep 17 00:00:00 2001
From: dzhwinter <dzhwinter@gmail.com>
Date: Mon, 11 Dec 2017 03:24:25 -0800
Subject: [PATCH 19/94] "add NCCL multi-GPU design doc"

---
 .../design}/images/multigpu_allreduce.graffle     | Bin
 .../design}/images/multigpu_allreduce.png         | Bin
 .../images/multigpu_before_convert.graffle        | Bin
 .../design}/images/multigpu_before_convert.png    | Bin
 .../multigpu.md => doc/design/paddle_nccl.md      |  14 ++++++--------
 5 files changed, 6 insertions(+), 8 deletions(-)
 rename {paddle/framework => doc/design}/images/multigpu_allreduce.graffle (100%)
 rename {paddle/framework => doc/design}/images/multigpu_allreduce.png (100%)
 rename {paddle/framework => doc/design}/images/multigpu_before_convert.graffle (100%)
 rename {paddle/framework => doc/design}/images/multigpu_before_convert.png (100%)
 rename paddle/framework/multigpu.md => doc/design/paddle_nccl.md (83%)

diff --git a/paddle/framework/images/multigpu_allreduce.graffle b/doc/design/images/multigpu_allreduce.graffle
similarity index 100%
rename from paddle/framework/images/multigpu_allreduce.graffle
rename to doc/design/images/multigpu_allreduce.graffle
diff --git a/paddle/framework/images/multigpu_allreduce.png b/doc/design/images/multigpu_allreduce.png
similarity index 100%
rename from paddle/framework/images/multigpu_allreduce.png
rename to doc/design/images/multigpu_allreduce.png
diff --git a/paddle/framework/images/multigpu_before_convert.graffle b/doc/design/images/multigpu_before_convert.graffle
similarity index 100%
rename from paddle/framework/images/multigpu_before_convert.graffle
rename to doc/design/images/multigpu_before_convert.graffle
diff --git a/paddle/framework/images/multigpu_before_convert.png b/doc/design/images/multigpu_before_convert.png
similarity index 100%
rename from paddle/framework/images/multigpu_before_convert.png
rename to doc/design/images/multigpu_before_convert.png
diff --git a/paddle/framework/multigpu.md b/doc/design/paddle_nccl.md
similarity index 83%
rename from paddle/framework/multigpu.md
rename to doc/design/paddle_nccl.md
index 1c843326ee..7c889fdd7f 100644
--- a/paddle/framework/multigpu.md
+++ b/doc/design/paddle_nccl.md
@@ -1,15 +1,17 @@
-# Design Doc: Multi-GPU support in Operation Graph
+# Design Doc: NCCL support in Paddle Fluid
 
 ## Abstract
 
-This Design Doc refers to the multi-GPU feature in  paddle.  We propose an approach to support multi-GPU both on a single machine and multiple machines. Every device only run sub-graphs which our framework issued. We use `Broadcast`, `Allreduce` operators to join different device sub-graph to the whole graph.
-
+This Design Doc refers to the NCCL feature in  paddle.  We propose an approach to support NCCL library both on a single machine and multiple machines. We wrapper the NCCL primitives `Broadcast`, `Allreduce`, `Reduce` as operators to utilize Multi-GPU powers in one script.
 
 
 ## Motivation
 
-Paddle supports training with multiple CPUs and GPUs, refer to different physical devices. We need to support multi-GPU training in parallel for acceleration, in detail, there are two aspects. 
+NCCL is a Nvidia library support Multi-GPU communicating. [NCCL](https://developer.nvidia.com/nccl). With NCCL library, we can easily accelerate the training in parallel.
 
+- can easily move the optimize sub-graph to parameter server,  multi-GPU feature can be compatible with distributed support design.
+- easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library.
+- GPU Model parallelism becomes easier to implement. we only need to replace different GPU's sub-graph with different part of the whole graph.
 - GPU Data Parallelism 
 
   Suppose to we have `n`GPUs, every GPU has `1/n`part of training data, and store a complete model in GPU memory.  
@@ -58,7 +60,3 @@ As it shown in the picture, when each GPU compute the gradient of `W`, followed
 In fact, in the way of every GPU optimized full batch of data, wasted (n-1) GPU compute resources. We will enhance it in the next stage.
 
 ### Benefits
-
-- can easily move the optimize sub-graph to parameter server,  multi-GPU feature can be  compatible with distributed support design.
-- easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library.
-- GPU Model parallelism becomes easier to implement. we only need to replace different GPU's sub-graph with different part of the whole graph.

From 69b44f2f198509e29b8ab50edab9ab34f56fd1af Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Tue, 12 Dec 2017 15:31:44 +0800
Subject: [PATCH 20/94] unify MKL macro definition

---
 cmake/cblas.cmake                                 | 2 +-
 cmake/external/mkldnn.cmake                       | 2 +-
 paddle/gserver/activations/ActivationFunction.cpp | 4 ++--
 paddle/gserver/gradientmachines/NeuralNetwork.cpp | 4 ++--
 paddle/math/Allocator.h                           | 2 +-
 paddle/math/MathFunctions.cpp                     | 2 +-
 paddle/math/MathFunctions.h                       | 2 +-
 paddle/memory/detail/system_allocator.cc          | 2 +-
 paddle/operators/math/math_function.cc            | 2 +-
 paddle/operators/math/math_function.h             | 2 +-
 paddle/parameter/FirstOrderOptimizer.h            | 2 +-
 paddle/parameter/ParameterUpdateFunctions.cpp     | 2 +-
 paddle/utils/Flags.cpp                            | 2 +-
 13 files changed, 15 insertions(+), 15 deletions(-)

diff --git a/cmake/cblas.cmake b/cmake/cblas.cmake
index b21fc43904..13294c0548 100644
--- a/cmake/cblas.cmake
+++ b/cmake/cblas.cmake
@@ -17,7 +17,7 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB)
   set(CBLAS_INC_DIR ${MKLML_INC_DIR})
   set(CBLAS_LIBRARIES ${MKLML_LIB})
 
-  add_definitions(-DPADDLE_USE_MKLML)
+  add_definitions(-DPADDLE_WITH_MKLML)
   add_definitions(-DLAPACK_FOUND)
 
   message(STATUS "Found cblas and lapack in MKLML "
diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake
index fc52d339d7..5d24caebdc 100644
--- a/cmake/external/mkldnn.cmake
+++ b/cmake/external/mkldnn.cmake
@@ -67,5 +67,5 @@ ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
 ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
 MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
-add_definitions(-DPADDLE_USE_MKLDNN)
+add_definitions(-DPADDLE_WITH_MKLDNN)
 LIST(APPEND external_project_dependencies mkldnn)
diff --git a/paddle/gserver/activations/ActivationFunction.cpp b/paddle/gserver/activations/ActivationFunction.cpp
index f5a41b66bf..57c890e488 100644
--- a/paddle/gserver/activations/ActivationFunction.cpp
+++ b/paddle/gserver/activations/ActivationFunction.cpp
@@ -24,7 +24,7 @@ limitations under the License. */
 #include "paddle/utils/ClassRegistrar.h"
 #include "paddle/utils/Logging.h"
 
-#ifdef PADDLE_USE_MKLDNN
+#ifdef PADDLE_WITH_MKLDNN
 #include "MKLDNNActivation.h"
 #endif
 
@@ -490,7 +490,7 @@ Error __must_check backward(Argument& act) {
 END_DEFINE_ACTIVATION(log)
 
 ActivationFunction* ActivationFunction::create(const std::string& type) {
-#ifdef PADDLE_USE_MKLDNN
+#ifdef PADDLE_WITH_MKLDNN
   if (!type.empty() && type.compare(0, 7, "mkldnn_") == 0) {
     return MKLDNNActivation::create(type);
   }
diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/gserver/gradientmachines/NeuralNetwork.cpp
index be112b4123..68bf37d59d 100644
--- a/paddle/gserver/gradientmachines/NeuralNetwork.cpp
+++ b/paddle/gserver/gradientmachines/NeuralNetwork.cpp
@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/utils/Logging.h"
 #include "paddle/utils/Stat.h"
 
-#ifdef PADDLE_USE_MKLDNN
+#ifdef PADDLE_WITH_MKLDNN
 #include "paddle/gserver/layers/MKLDNNLayer.h"
 #endif
 
@@ -307,7 +307,7 @@ void NeuralNetwork::backward(const UpdateCallback& callback) {
 }
 
 void NeuralNetwork::finish() {
-#ifdef PADDLE_USE_MKLDNN
+#ifdef PADDLE_WITH_MKLDNN
   FOR_EACH_R(layer, layers_) {
     MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast<MKLDNNLayer>(*layer);
     if (dnnLayer) {
diff --git a/paddle/math/Allocator.h b/paddle/math/Allocator.h
index 94ef561f06..17563bf5e1 100644
--- a/paddle/math/Allocator.h
+++ b/paddle/math/Allocator.h
@@ -48,7 +48,7 @@ public:
    */
   virtual void* alloc(size_t size) {
     void* ptr;
-#ifdef PADDLE_USE_MKLDNN
+#ifdef PADDLE_WITH_MKLDNN
     // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
     // memory alignment
     CHECK_EQ(posix_memalign(&ptr, 4096ul, size), 0);
diff --git a/paddle/math/MathFunctions.cpp b/paddle/math/MathFunctions.cpp
index ba86eacbb5..28ab54b450 100644
--- a/paddle/math/MathFunctions.cpp
+++ b/paddle/math/MathFunctions.cpp
@@ -206,7 +206,7 @@ double dotProduct<double>(const int n, const double* x, const double* y) {
 }
 #endif
 
-#if defined(PADDLE_USE_MKLML)
+#if defined(PADDLE_WITH_MKLML)
 
 template <>
 void vExp<float>(const int n, const float* a, float* r) {
diff --git a/paddle/math/MathFunctions.h b/paddle/math/MathFunctions.h
index f6e77029bd..29fe36e3a4 100644
--- a/paddle/math/MathFunctions.h
+++ b/paddle/math/MathFunctions.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #ifndef MATHFUNCTIONS_H_
 #define MATHFUNCTIONS_H_
 
-#ifdef PADDLE_USE_MKLML
+#ifdef PADDLE_WITH_MKLML
 #include <mkl_cblas.h>
 #include <mkl_lapacke.h>
 #include <mkl_vml_functions.h>
diff --git a/paddle/memory/detail/system_allocator.cc b/paddle/memory/detail/system_allocator.cc
index b543b767e8..6a815a1b57 100644
--- a/paddle/memory/detail/system_allocator.cc
+++ b/paddle/memory/detail/system_allocator.cc
@@ -43,7 +43,7 @@ void* CPUAllocator::Alloc(size_t& index, size_t size) {
 
   void* p;
 
-#ifdef PADDLE_USE_MKLDNN
+#ifdef PADDLE_WITH_MKLDNN
   // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
   // memory alignment
   PADDLE_ENFORCE_EQ(posix_memalign(&p, 4096ul, size), 0);
diff --git a/paddle/operators/math/math_function.cc b/paddle/operators/math/math_function.cc
index 2e333a8cde..e099a6a439 100644
--- a/paddle/operators/math/math_function.cc
+++ b/paddle/operators/math/math_function.cc
@@ -132,7 +132,7 @@ void matmul<platform::CPUPlace, double>(
       matrix_b.data<double>(), beta, matrix_out->data<double>());
 }
 
-#ifdef PADDLE_USE_MKLML
+#ifdef PADDLE_WITH_MKLML
 // Use cblas_{s,d}gemm_batched if available: Run with 1 group of size batchSize.
 template <>
 void batched_gemm<platform::CPUPlace, float>(
diff --git a/paddle/operators/math/math_function.h b/paddle/operators/math/math_function.h
index 5a42854f22..f2b025b78b 100644
--- a/paddle/operators/math/math_function.h
+++ b/paddle/operators/math/math_function.h
@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#ifdef PADDLE_USE_MKLML
+#ifdef PADDLE_WITH_MKLML
 #include <mkl_cblas.h>
 #include <mkl_lapacke.h>
 #include <mkl_vml_functions.h>
diff --git a/paddle/parameter/FirstOrderOptimizer.h b/paddle/parameter/FirstOrderOptimizer.h
index f157188a4f..5b0c52a30d 100644
--- a/paddle/parameter/FirstOrderOptimizer.h
+++ b/paddle/parameter/FirstOrderOptimizer.h
@@ -38,7 +38,7 @@ public:
     real torch_learningRate = optConfig_.learning_method() == "torch_momentum"
                                   ? 1.0 - paraConfig.momentum()
                                   : 1.0;
-#ifdef PADDLE_USE_MKLDNN
+#ifdef PADDLE_WITH_MKLDNN
     sgdUpdate(learningRate_ * paraConfig.learning_rate() *
                   (firstTime_ ? 1.0 : torch_learningRate),
               paraConfig.momentum(),
diff --git a/paddle/parameter/ParameterUpdateFunctions.cpp b/paddle/parameter/ParameterUpdateFunctions.cpp
index 1898598e49..d60cb36383 100644
--- a/paddle/parameter/ParameterUpdateFunctions.cpp
+++ b/paddle/parameter/ParameterUpdateFunctions.cpp
@@ -30,7 +30,7 @@ void sgdUpdateCpu(real learningRate,
                   const real* grad,
                   real* momentumVec) {
   decayRate *= learningRate;
-#ifdef PADDLE_USE_MKLML
+#ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for
 #endif
   for (size_t i = 0; i < size; ++i) {
diff --git a/paddle/utils/Flags.cpp b/paddle/utils/Flags.cpp
index 8f100f02e9..9a7dc0e356 100644
--- a/paddle/utils/Flags.cpp
+++ b/paddle/utils/Flags.cpp
@@ -20,7 +20,7 @@ DEFINE_bool(use_gpu, false, "Only support CPU training");
 DEFINE_bool(use_gpu, true, "Whether to use GPU for training");
 #endif
 
-#ifdef PADDLE_USE_MKLDNN
+#ifdef PADDLE_WITH_MKLDNN
 // TODO(TJ): change to true when MKLDNN layers support multi-inputs
 DEFINE_bool(use_mkldnn, false, "Default still keep use CPU training");
 #else

From ebd0cf1408f3408551c342c8eb249f4313981695 Mon Sep 17 00:00:00 2001
From: dzhwinter <dzhwinter@gmail.com>
Date: Mon, 11 Dec 2017 06:55:04 -0800
Subject: [PATCH 21/94] "add manual allreduce"

---
 doc/design/paddle_nccl.md | 44 +++++++++++++++++++--------------------
 1 file changed, 21 insertions(+), 23 deletions(-)

diff --git a/doc/design/paddle_nccl.md b/doc/design/paddle_nccl.md
index 7c889fdd7f..dba451b9b3 100644
--- a/doc/design/paddle_nccl.md
+++ b/doc/design/paddle_nccl.md
@@ -7,36 +7,34 @@ This Design Doc refers to the NCCL feature in  paddle.  We propose an approach t
 
 ## Motivation
 
-NCCL is a Nvidia library support Multi-GPU communicating. [NCCL](https://developer.nvidia.com/nccl). With NCCL library, we can easily accelerate the training in parallel.
+NCCL is a NVIDIA library support Multi-GPU communicating and optimized for NVIDIA GPUs, it provides routines such as all-gather, all-reduce, broadcast, reduce, reduce-scatter, that can achieve high bandwidth over PCIe and NVLink high-speed interconnect. [NCCL](https://developer.nvidia.com/nccl). With NCCL library, we can easily accelerate the training in parallel. 
 
-- can easily move the optimize sub-graph to parameter server,  multi-GPU feature can be compatible with distributed support design.
-- easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library.
-- GPU Model parallelism becomes easier to implement. we only need to replace different GPU's sub-graph with different part of the whole graph.
-- GPU Data Parallelism 
+- Pros
+1. easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library.
+1. high performance in NVIDIA GPUs.
+1. MPI like primitives, which have low learning cost for users.
 
-  Suppose to we have `n`GPUs, every GPU has `1/n`part of training data, and store a complete model in GPU memory.  
+- Cons
+1. Only design for NVIDIA GPUs, not a general multi-device solution.
+1. Although NCCL1 is opensourced under BSD license, but NCCL2 is not opensourced anymore.
 
-- GPU Model Parallelism
+At the beginning of training, the framework needs to distribute the same parameters to every GPU, and merge the gradients at any time user interests.
 
-  every GPU have part of a complete model in GPU memory.
+As a result, during training, we need the operations of peer to peer copy between different GPUs, aggregating gradients/parameters from GPUs, and broadcasting parameters to GPUs. Every GPU only need to run the operator with correct place information.
 
-At the beginning of training, the framework needs to issue the same sub-graph to every GPU in Data Parallelism, or different sub-graph in Model Parallelism.
-
-During training, we need the operations of peer to peer copy between different GPUs, aggregating gradients/parameters from GPUs, and broadcasting parameters to GPUs. Every GPU only need to run the sub-graph with correct place information.
-
-Besides, it needs interfaces to synchronize model update with each other, and issue/merge model from different GPU Cards. 
+Besides, it needs interfaces to synchronize model update with each different GPU Cards. 
 
 ## Implementation
 
-As mentioned above, we summarise that several kinds of operators are needed. Currently, we need to issue parameters to different GPUs,  named it with Broadcast operator.  And also synchronize parameters between GPUs, called it with AllReduce. 
+As mentioned above, we wrap the NCCL routines as several kinds of operators. Need to note that NCCL need to create Communicator between gpu at the beginning, so there is a NCCLInit operator created.
 
 ### Graph Converter
 
 To be compatible with [parameter server design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md), the graph converter converts the user defined operation graph into sub-graphs to be executed on different devices.
 
-1. The user-defined operator graph will be partitioned into sub-graph. 
+1. The user-defined model will be a single device program
 
-2. Control operators between GPUs will be inserted into the graph.
+2. Broadcast/Reduce operators between GPUs will be inserted into the program, even for the multi-node, may insert the `Send`, `Recv` operator.
 
    *Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, [Send, Recv](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md#graph-converter) in multiple machines*
 
@@ -49,14 +47,14 @@ After convert, the graph as shows
 Operators are added to the sub-graphs. Every GPU assigned a role of `rank0`, `rank1` etc. 
 
 - **Broadcast**. Broadcast operator distribute initialized parameter to all the GPUs from the GPU who owns it. e.g. from`rank0` GPU.
-- **Allreduce**. Allreduce operator synchronizes parameters/gradients between GPUs. AllReduce implemented in the Ring-Based  communicating method, avoid of the bottle neck in a single GPU.
+- **AllReduce**. AllReduce operator synchronizes parameters/gradients between GPUs. AllReduce implemented in the Ring-Based  communicating method, avoid of the bottle neck in a single GPU.
 
-These two operators need the Multi-GPU context support.
-
-Need to notice that Allreduce operator force GPUs synchronized at that point. Every device only need runs sub-graph in a loop style forever, the whole training process in asynchronous or synchronous mode depends on the Allreduce point in the graph.
+Need to notice that AllReduce operator force GPUs synchronized at that point. The whole training process in asynchronous or synchronous mode depends on the AllReduce point in the graph.
 
 As it shown in the picture, when each GPU compute the gradient of `W`, followed with a `AllReduce` operator, accumulate the `dW` to full batch of data, then run the optimize process individually and apply the gradient to its `W`.
 
-In fact, in the way of every GPU optimized full batch of data, wasted (n-1) GPU compute resources. We will enhance it in the next stage.
-
-### Benefits
+- **AllReduce2**
+If we use the NCCL2 AllReduce primitive, every GPU optimized full batch of data, wasted (n-1) GPU compute resources. In addition, AllReduce will only utilize the communicate resource during synchronization, then update the gradient will be a seperated phase. In fact, we can amortize the update gradient time cost into the communicating phase.
+- Every parameter has its root card. That card will call **Reduce** operator and collect the gradients from GPUs.
+- The whole model's parameter will be hashed to different root card, ensure the load balance between GPUs.
+Then we have another version AllReduce operator. Other part keep the same with before.

From 0ca6274451d3693f363f2b8b5d6b29ce722febaf Mon Sep 17 00:00:00 2001
From: dzhwinter <dongzhihong01@baidu.com>
Date: Mon, 11 Dec 2017 23:15:35 +0800
Subject: [PATCH 22/94] "add global regularization" (#6443)

* "add global regularization"

* Polish `append_regularization_ops`
---
 python/paddle/v2/fluid/optimizer.py   | 38 +++++++++++----------------
 python/paddle/v2/fluid/regularizer.py | 15 ++++++++---
 2 files changed, 27 insertions(+), 26 deletions(-)

diff --git a/python/paddle/v2/fluid/optimizer.py b/python/paddle/v2/fluid/optimizer.py
index 719e3b2563..bbdfab2df9 100644
--- a/python/paddle/v2/fluid/optimizer.py
+++ b/python/paddle/v2/fluid/optimizer.py
@@ -18,8 +18,9 @@ class Optimizer(object):
     but need to use one of it's implementation.
     """
 
-    def __init__(self, global_step=None):
+    def __init__(self, global_step=None, regularization=None):
         self._global_step = global_step
+        self.regularization = regularization
         # Dictionary of accumulators. Some optimizer subclasses need to
         # allocate and manage extra variables associated with the parameters
         # to train. These variables are called accumulators.
@@ -199,7 +200,8 @@ class Optimizer(object):
         """
         params_grads = append_backward_ops(loss, parameter_list, no_grad_set)
         # Add regularization if any
-        params_grads = append_regularization_ops(params_grads)
+        params_grads = append_regularization_ops(params_grads,
+                                                 self.regularization)
         optimize_ops = self.create_optimization_pass(params_grads, loss,
                                                      startup_program)
         return optimize_ops
@@ -209,9 +211,9 @@ class SGDOptimizer(Optimizer):
     """ Simple SGD optimizer without any state.
     """
 
-    def __init__(self, learning_rate, global_step=None):
+    def __init__(self, learning_rate, **kwargs):
         assert learning_rate is not None
-        super(SGDOptimizer, self).__init__(global_step)
+        super(SGDOptimizer, self).__init__(**kwargs)
         self.type = "sgd"
         self._learning_rate = learning_rate
 
@@ -236,14 +238,10 @@ class MomentumOptimizer(Optimizer):
     """
     _velocity_acc_str = "velocity"
 
-    def __init__(self,
-                 learning_rate,
-                 momentum,
-                 use_nesterov=False,
-                 global_step=None):
+    def __init__(self, learning_rate, momentum, use_nesterov=False, **kwargs):
         assert learning_rate is not None
         assert momentum is not None
-        super(MomentumOptimizer, self).__init__(global_step)
+        super(MomentumOptimizer, self).__init__(**kwargs)
         self.type = "momentum"
         self._learning_rate = learning_rate
         self._momentum = momentum
@@ -284,10 +282,10 @@ class AdagradOptimizer(Optimizer):
     """
     _moment_acc_str = "moment"
 
-    def __init__(self, learning_rate, epsilon=1.0e-6, global_step=None):
+    def __init__(self, learning_rate, epsilon=1.0e-6, **kwargs):
         assert learning_rate is not None
         assert epsilon is not None
-        super(AdagradOptimizer, self).__init__(global_step)
+        super(AdagradOptimizer, self).__init__(**kwargs)
         self.type = "adagrad"
         self._learning_rate = learning_rate
         self._epsilon = epsilon
@@ -331,12 +329,12 @@ class AdamOptimizer(Optimizer):
                  beta1=0.9,
                  beta2=0.999,
                  epsilon=1e-8,
-                 global_step=None):
+                 **kwargs):
         assert learning_rate is not None
         assert beta1 is not None
         assert beta2 is not None
         assert epsilon is not None
-        super(AdamOptimizer, self).__init__(global_step)
+        super(AdamOptimizer, self).__init__(**kwargs)
         self.type = "adam"
         self._learning_rate = learning_rate
         self._beta1 = beta1
@@ -436,12 +434,12 @@ class AdamaxOptimizer(Optimizer):
                  beta1=0.9,
                  beta2=0.999,
                  epsilon=1e-8,
-                 global_step=None):
+                 **kwargs):
         assert learning_rate is not None
         assert beta1 is not None
         assert beta2 is not None
         assert epsilon is not None
-        super(AdamaxOptimizer, self).__init__()
+        super(AdamaxOptimizer, self).__init__(**kwargs)
         self.type = "adamax"
         self._learning_rate = learning_rate
         self._beta1 = beta1
@@ -514,16 +512,12 @@ class DecayedAdagradOptimizer(Optimizer):
     """
     _moment_acc_str = "moment"
 
-    def __init__(self,
-                 learning_rate,
-                 decay=0.95,
-                 epsilon=1.0e-6,
-                 global_step=None):
+    def __init__(self, learning_rate, decay=0.95, epsilon=1.0e-6, **kwargs):
         assert learning_rate is not None
         assert decay is not None
         assert epsilon is not None
 
-        super(DecayedAdagradOptimizer, self).__init__(global_step)
+        super(DecayedAdagradOptimizer, self).__init__(**kwargs)
         self.type = "decayed_adagrad"
         self._learning_rate = learning_rate
         self._decay = decay
diff --git a/python/paddle/v2/fluid/regularizer.py b/python/paddle/v2/fluid/regularizer.py
index bb1ac8911e..d1955b0047 100644
--- a/python/paddle/v2/fluid/regularizer.py
+++ b/python/paddle/v2/fluid/regularizer.py
@@ -3,7 +3,7 @@ import framework
 __all__ = ['append_regularization_ops', 'L1Decay', 'L2Decay']
 
 
-def append_regularization_ops(parameters_and_grads):
+def append_regularization_ops(parameters_and_grads, regularization=None):
     """Create and add backward regularization Operators
 
     Creates and adds backward regularization operators in the BlockDesc.
@@ -14,6 +14,8 @@ def append_regularization_ops(parameters_and_grads):
     Args:
         parameters_and_grads: A list of (parameters, gradients) pairs
                               that need to be regularized.
+        regularization: A global regularizer. If the parameter is not
+                        set. It will be applied with regularizer.
 
     Returns:
         list of (parameters, gradients) pair with the regularized gradient
@@ -23,14 +25,19 @@ def append_regularization_ops(parameters_and_grads):
     """
     params_and_grads = []
     for param, grad in parameters_and_grads:
+        regularization_term = None
+        if param.regularizer is not None:
+            # Add variable for regularization term in grad block
+            regularization_term = param.regularizer(param, grad.block)
+        elif regularization is not None:
+            regularization_term = regularization(param, grad.block)
+
         # If no gradient or no regularization specified,
         # then we don't need to do anything
-        if grad is None or param.regularizer is None:
+        if grad is None or regularization_term is None:
             params_and_grads.append((param, grad))
             continue
 
-        # Add variable for regularization term in grad block
-        regularization_term = param.regularizer(param, grad.block)
         assert grad.shape == regularization_term.shape
 
         grad.block.append_op(

From 4ff6bc175a18d03f1159e78aef0c305703adbe37 Mon Sep 17 00:00:00 2001
From: Siddharth Goyal <vi.siddharth78@gmail.com>
Date: Mon, 11 Dec 2017 11:33:54 -0800
Subject: [PATCH 23/94] Add row conv operator (#6013)

* Fix documentation

* Address review comments
---
 paddle/operators/row_conv_op.cc               | 257 +++++++++++
 paddle/operators/row_conv_op.cu               | 408 ++++++++++++++++++
 paddle/operators/row_conv_op.h                |  33 ++
 .../paddle/v2/fluid/tests/test_row_conv_op.py |  95 ++++
 4 files changed, 793 insertions(+)
 create mode 100644 paddle/operators/row_conv_op.cc
 create mode 100644 paddle/operators/row_conv_op.cu
 create mode 100644 paddle/operators/row_conv_op.h
 create mode 100644 python/paddle/v2/fluid/tests/test_row_conv_op.py

diff --git a/paddle/operators/row_conv_op.cc b/paddle/operators/row_conv_op.cc
new file mode 100644
index 0000000000..ea0bb99f8d
--- /dev/null
+++ b/paddle/operators/row_conv_op.cc
@@ -0,0 +1,257 @@
+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/row_conv_op.h"
+#include "paddle/framework/eigen.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using framework::Tensor;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+class RowConvOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of RowConvOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Filter"),
+                   "Input(Filter) of RowConvOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of RowConvOp should not be null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto filter_dims = ctx->GetInputDim("Filter");
+    PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(filter_dims.size(), 2, "Input(Y)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(
+        x_dims[1], filter_dims[1],
+        "The 2nd dimension of Input(X) and Input(Filter) should be same.");
+    ctx->SetOutputDim("Out", x_dims);
+    ctx->ShareLoD("X", "Out");
+  }
+};
+
+class RowConvGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Filter"),
+                   "Input(Filter) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Gradient of output(Out) should not be null.");
+
+    auto x_grad_name = framework::GradVarName("X");
+    if (ctx->HasOutput(x_grad_name)) {
+      auto x_dims = ctx->GetInputDim("X");
+      ctx->SetOutputDim(x_grad_name, x_dims);
+    }
+
+    auto filter_grad_name = framework::GradVarName("Filter");
+    if (ctx->HasOutput(filter_grad_name)) {
+      auto filter_dims = ctx->GetInputDim("Filter");
+      ctx->SetOutputDim(filter_grad_name, filter_dims);
+    }
+  }
+};
+
+class RowConvOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  RowConvOpMaker(framework::OpProto *proto,
+                 framework::OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(LoDTensor), the input(X) is a LodTensor, which supports "
+             "variable time-length input sequences. The underlying tensor "
+             "in this LoDTensor is a matrix with shape (T x N), where T "
+             "is the total time steps in this mini-batch and N is the input "
+             "data dimension.");
+    AddInput("Filter",
+             "(Tensor), the input(Filter) is a learnable parameter. It "
+             "is a 2-D tensor with shape (future_context x N), where, "
+             "future_context is the future context length and N is the data "
+             "dimension.");
+    AddOutput("Out",
+              "(LoDTensor), the output(Out) is a LodTensor, which supports "
+              "variable time-length input sequences. The underlying tensor "
+              "in this LodTensor is a matrix with shape T x N, i.e., the "
+              "same shape as X.");
+    AddComment(R"DOC(
+Row-convolution Operator.
+
+The row convolution is called lookahead convolution.  This operator was 
+introduced in the following paper for DeepSpeech2:
+http://www.cs.cmu.edu/~dyogatam/papers/wang+etal.iclrworkshop2016.pdf 
+
+The main motivation is that a bidirectional RNN, useful in DeepSpeech 
+like speech models, learns representation for a sequence by performing a 
+forward and a backward pass through the entire sequence. However, unlike 
+unidirectional RNNs, bidirectional RNNs are challenging to deploy in an online
+and low-latency setting. The lookahead convolution incorporates information 
+from future subsequences in a computationally efficient manner to improve 
+unidirectional recurrent neural networks. The row convolution operator is 
+different from the 1D sequence convolution, and is computed as follows:
+
+Given an input sequence $in$ of length $t$ and input dimension $d$, 
+and a filter ($W$) of size $context \times d$, 
+the output sequence is convolved as:
+
+$$
+out_{i, :} = \sum_{j=i}^{i + context} in_{j,:} \dot W_{i-j, :}
+$$
+
+)DOC");
+  }
+};
+
+template <typename T>
+class RowConvKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto *x = context.Input<LoDTensor>("X");
+    auto *filter = context.Input<Tensor>("Filter");
+    auto *out = context.Output<LoDTensor>("Out");
+
+    out->mutable_data<T>(context.GetPlace());
+
+    auto batch_indices = x->lod()[0];
+    auto input_dim = x->dims()[1];  // 'in' is of size T x N
+    size_t num_sequence = batch_indices.size() - 1;
+
+    auto future_context = filter->dims()[0];
+    auto weights = EigenMatrix<T>::From(*filter);
+
+    for (size_t i = 0; i < num_sequence; i++) {
+      int start = static_cast<int>(batch_indices[i]);
+      int end = static_cast<int>(batch_indices[i + 1]);
+      int current_timesteps = end - start;
+      Tensor cur_input_sequence =
+          x->Slice(start, end);  // Current input sequence
+      Tensor cur_output_sequence =
+          out->Slice(start, end);  // Current output sequence
+      auto cip_seq = EigenMatrix<T>::From(cur_input_sequence);
+      auto cot_seq = EigenMatrix<T>::From(cur_output_sequence);
+
+      for (int k = 0; k < current_timesteps;
+           k++) {  // For different time steps in the same sequence
+        for (int w = 0; (w < future_context) && ((k + w) < current_timesteps);
+             w++) {
+          for (int d = 0; d < input_dim; d++) {
+            if (w == 0) {
+              cot_seq(k, d) = weights(w, d) * cip_seq(k + w, d);
+            } else {
+              cot_seq(k, d) += weights(w, d) * cip_seq(k + w, d);
+            }
+          }
+        }
+      }
+    }
+  }
+};
+
+template <typename T>
+class RowConvGradKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto *x = context.Input<LoDTensor>("X");
+    auto *filter = context.Input<Tensor>("Filter");
+    auto *d_out = context.Input<LoDTensor>(framework::GradVarName("Out"));
+    auto *dx = context.Output<LoDTensor>(framework::GradVarName("X"));
+    auto *d_filter = context.Output<Tensor>(framework::GradVarName("Filter"));
+
+    auto input_dim = x->dims()[1];  // 'x' is of size T x N
+    auto batch_indices = x->lod()[0];
+    size_t num_sequence = batch_indices.size() - 1;
+    auto future_context = filter->dims()[0];
+
+    if (d_filter) {
+      d_filter->mutable_data<T>(context.GetPlace());
+      auto dweights =
+          EigenMatrix<T>::From(*d_filter);  // Gradient of weight matrix
+      dweights.setZero();
+
+      for (size_t i = 0; i < num_sequence; i++) {  // For different sequences
+        int start = static_cast<int>(batch_indices[i]);
+        int end = static_cast<int>(batch_indices[i + 1]);
+
+        Tensor cur_input = x->Slice(start, end);  // Current input sequence
+        Tensor cur_doutput =
+            d_out->Slice(start, end);  // Current output grad sequence
+
+        auto cur_ip = EigenMatrix<T>::From(cur_input);
+        auto cur_dout = EigenMatrix<T>::From(cur_doutput);
+        int current_timesteps = end - start;
+
+        for (int k = 0; k < current_timesteps;
+             k++) {  // For different time steps in the same sequence
+          for (int w = 0; (w < future_context) && ((k + w) < current_timesteps);
+               w++) {
+            // For dweights (Updating the gradient of weight matrix)
+            for (int d = 0; d < input_dim; d++) {
+              dweights(w, d) += cur_ip(k + w, d) * cur_dout(k, d);
+            }
+          }
+        }
+      }
+    }
+
+    if (dx) {
+      dx->mutable_data<T>(context.GetPlace());
+      auto weights = EigenMatrix<T>::From(*filter);
+      for (size_t i = 0; i < num_sequence; i++) {  // For different sequences
+        int start = static_cast<int>(batch_indices[i]);
+        int end = static_cast<int>(batch_indices[i + 1]);
+
+        Tensor cur_doutput =
+            d_out->Slice(start, end);  // Current output grad sequence
+        Tensor cur_dinput =
+            dx->Slice(start, end);  // Current input grad sequence
+
+        auto cur_dout = EigenMatrix<T>::From(cur_doutput);
+        auto cur_dip = EigenMatrix<T>::From(cur_dinput);
+        cur_dip.setZero();
+        int current_timesteps = end - start;
+
+        for (int k = 0; k < current_timesteps;
+             k++) {  // For different time steps in the same sequence
+          for (int w = 0; (w < future_context) && ((k + w) < current_timesteps);
+               w++) {
+            // For dinput (Updating the gradient wrt input)
+            for (int d = 0; d < input_dim; d++) {
+              cur_dip(k + w, d) += weights(w, d) * cur_dout(k, d);
+            }
+          }
+        }
+      }
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(row_conv, ops::RowConvOp, ops::RowConvOpMaker, row_conv_grad,
+            ops::RowConvGradOp);
+REGISTER_OP_CPU_KERNEL(row_conv,
+                       ops::RowConvKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    row_conv_grad, ops::RowConvGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/row_conv_op.cu b/paddle/operators/row_conv_op.cu
new file mode 100644
index 0000000000..e0d7ebda7e
--- /dev/null
+++ b/paddle/operators/row_conv_op.cu
@@ -0,0 +1,408 @@
+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/math/math_function.h"
+#include "paddle/operators/row_conv_op.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using framework::Tensor;
+
+namespace {
+
+inline int DivUp(int x, int y) { return (x + y - 1) / y; }
+
+// Forward prop (shared memory version, for small future_context)
+template <typename T>
+__global__ void RowConvForwardSharedMemory(const T *in, const T *wt,
+                                           int num_sequence, int input_dim,
+                                           int future_context,
+                                           const size_t *batch_indices,
+                                           T *out) {
+  int blx = blockDim.x;
+  int bly = blockDim.y;
+  int thx = threadIdx.x;
+  int thy = threadIdx.y;
+  int d = blockIdx.x * blx + thx;  // index along input dim
+
+  extern __shared__ T mem[];
+  T *sw = mem;
+
+  if (thy < future_context) {
+    sw[thy * blx + thx] =
+        (d < input_dim) ? wt[thy * input_dim + d] : static_cast<T>(0);
+  }
+  __syncthreads();
+
+  for (size_t i = 0; i < num_sequence; i++) {
+    int start = static_cast<int>(batch_indices[i]);
+    int end = static_cast<int>(batch_indices[i + 1]);
+    int current_timesteps = end - start;
+    for (int k = thy; k < current_timesteps; k += bly) {
+      T sum = 0;
+      for (int w = 0; (w < future_context) && ((k + w) < current_timesteps);
+           w++) {
+        sum += (d < input_dim)
+                   ? sw[w * blx + thx] * in[(start + k + w) * input_dim + d]
+                   : static_cast<T>(0);
+      }
+      if (d < input_dim) {
+        out[(start + k) * input_dim + d] = sum;
+      }
+    }
+  }
+}
+
+// Forward prop (naive version)
+template <typename T>
+__global__ void RowConvForward(const T *in, const T *wt, int num_sequence,
+                               int input_dim, int future_context,
+                               const size_t *batch_indices, T *out) {
+  int d = blockIdx.x * blockDim.x + threadIdx.x;  // index along input_dim
+  int bly = blockDim.y;
+  int thy = threadIdx.y;
+
+  if (d >= input_dim) return;
+
+  for (size_t i = 0; i < num_sequence; i++) {
+    int start = static_cast<int>(batch_indices[i]);
+    int end = static_cast<int>(batch_indices[i + 1]);
+    int current_timesteps = end - start;
+    for (int k = thy; k < current_timesteps; k += bly) {
+      T sum = 0;
+      for (int w = 0; (w < future_context) && ((k + w) < current_timesteps);
+           w++) {
+        sum += (wt[w * input_dim + d] * in[(start + k + w) * input_dim + d]);
+      }
+      out[(start + k) * input_dim + d] = sum;
+    }
+  }
+}
+
+// Compute input gradient (shared memory version, for small future_context)
+template <typename T>
+__global__ void RowConvGradInputSharedMemory(const T *dout, const T *wt,
+                                             int num_sequence, int input_dim,
+                                             int future_context,
+                                             const size_t *batch_indices,
+                                             T *din) {
+  int blx = blockDim.x;
+  int bly = blockDim.y;
+  int thx = threadIdx.x;
+  int thy = threadIdx.y;
+  int d = blockIdx.x * blx + thx;  // index along input dim
+
+  extern __shared__ T mem[];
+  T *sw = mem;
+  if (thy < future_context) {
+    sw[thy * blx + thx] =
+        (d < input_dim) ? wt[thy * input_dim + d] : static_cast<T>(0);
+  }
+  __syncthreads();
+
+  for (int i = 0; i < num_sequence; i++) {
+    int start = static_cast<int>(batch_indices[i]);
+    int end = static_cast<int>(batch_indices[i + 1]);
+    int current_timesteps = end - start;
+    for (int k = thy; k < current_timesteps; k += bly) {
+      T sum = 0;
+      for (int w = 0; (w < future_context) && ((k - w) >= 0); w++) {
+        sum += (d < input_dim)
+                   ? (sw[w * blx + thx] * dout[(k + start - w) * input_dim + d])
+                   : static_cast<T>(0);
+      }
+      if (d < input_dim) {
+        din[(k + start) * input_dim + d] = sum;
+      }
+    }
+  }
+}
+
+// Compute input gradient (Naive version)
+template <typename T>
+__global__ void RowConvGradInput(const T *dout, const T *wt, int num_sequence,
+                                 int input_dim, int future_context,
+                                 const size_t *batch_indices, T *din) {
+  int d = blockIdx.x * blockDim.x + threadIdx.x;  // index along input_dim
+  int bly = blockDim.y;
+  int thy = threadIdx.y;
+
+  if (d >= input_dim) return;
+  for (int i = 0; i < num_sequence; i++) {
+    int start = static_cast<int>(batch_indices[i]);
+    int end = static_cast<int>(batch_indices[i + 1]);
+    int current_timesteps = end - start;
+    for (int k = thy; k < current_timesteps; k += bly) {
+      T sum = 0;
+      for (int w = 0; (w < future_context) && ((k - w) >= 0); w++) {
+        sum += (wt[w * input_dim + d] * dout[(k + start - w) * input_dim + d]);
+      }
+      din[(k + start) * input_dim + d] = sum;
+    }
+  }
+}
+
+// Compute W gradient (small future_context version)
+template <typename T>
+__global__ void RowConvGradFilterImproved(const T *in, const T *dout,
+                                          int num_sequence, int input_dim,
+                                          int future_context, int block_x,
+                                          int block_y,
+                                          const size_t *batch_indices,
+                                          T *dfilter) {
+  int blx = blockDim.x;
+  int bly = blockDim.y;
+  int thx = threadIdx.x;
+  int thy = threadIdx.y;
+  int gx = blockIdx.x * blx;
+  int d = gx + thx;  // index along input dim
+
+  extern __shared__ T mem[];
+
+  int xdim_sh_in = block_y;
+  int xdim_sh_dout = block_y;
+  // int xdim_sh_dfilter = future_context;
+  int ydim_sh_in = block_x;
+  int ydim_sh_dout = block_x + future_context - 1;
+  int ydim_sh_dfilter = block_y;
+
+  T *sh_in = mem;
+  T *sh_dout = &mem[xdim_sh_in * ydim_sh_in];
+  T *sh_dfilter = &mem[xdim_sh_in * ydim_sh_in + xdim_sh_dout * ydim_sh_dout];
+
+  if (thy < future_context) {
+    sh_dfilter[thy * ydim_sh_dfilter + thx] = static_cast<T>(0);
+  }
+  __syncthreads();
+
+  for (int i = 0; i < num_sequence; i++) {
+    int start = static_cast<int>(batch_indices[i]);
+    int end = static_cast<int>(batch_indices[i + 1]);
+    int current_timesteps = end - start;
+    int scaled_cur_steps =
+        ((current_timesteps + block_x - 1) / block_x) * block_x;
+
+    for (int k = thy; k < scaled_cur_steps; k += block_x) {
+      int pos = start + k;
+      sh_in[thx * ydim_sh_in + thy] =
+          (d < input_dim && pos < end) ? in[pos * input_dim + d] : T(0);
+      sh_dout[thx * ydim_sh_dout + thy + future_context - 1] =
+          (d < input_dim && pos < end) ? dout[pos * input_dim + d] : T(0);
+      __syncthreads();
+
+      if (thy < future_context - 1) {
+        int pos_offset = pos - future_context + 1;
+        sh_dout[thx * ydim_sh_dout + thy] =
+            (d < input_dim && pos_offset >= start)
+                ? dout[pos_offset * input_dim + d]
+                : T(0);
+      }
+      __syncthreads();
+
+      for (int w = 0; w < future_context; w++) {
+        T val = sh_in[thy * ydim_sh_in + thx] *
+                sh_dout[thy * ydim_sh_dout + thx + future_context - 1 - w];
+        __syncthreads();
+
+        for (int offset = 16; offset > 0;
+             offset = offset / 2) {  // blockDim.x is 32.
+          val += __shfl_down(val, offset);
+        }
+        __syncthreads();
+
+        if (thx == 0) {
+          sh_dfilter[w * ydim_sh_dfilter + thy] += val;
+        }
+        __syncthreads();
+      }
+    }
+  }
+  for (int w = thy; (w < future_context) && (d < input_dim); w += bly) {
+    dfilter[w * input_dim + d] += sh_dfilter[w * ydim_sh_dfilter + thx];
+  }
+}
+
+// Compute weight(filter) gradient
+template <typename T>
+__global__ void RowConvGradFilter(const T *in, const T *dout, int num_sequence,
+                                  int input_dim, int future_context,
+                                  int block_x, int block_y,
+                                  const size_t *batch_indices, T *dfilter) {
+  int blx = blockDim.x;
+  int bly = blockDim.y;
+  int thx = threadIdx.x;
+  int thy = threadIdx.y;
+  int gx = blockIdx.x * blx;
+  int d = gx + thx;  // index along input dim
+  extern __shared__ T mem[];
+  T *sh_in = mem;
+  T *sh_dout = &mem[block_x * block_y];
+
+  for (int i = 0; i < num_sequence; i++) {
+    int start = static_cast<int>(batch_indices[i]);
+    int end = static_cast<int>(batch_indices[i + 1]);
+    int current_timesteps = end - start;
+    int scaled_cur_steps =
+        ((current_timesteps + block_x - 1) / block_x) * block_x;
+
+    for (int k = thy; k < scaled_cur_steps; k += block_x) {
+      int pos = start + k;
+      sh_in[thx * block_y + thy] =
+          (d < input_dim && pos < end) ? in[pos * input_dim + d] : 0.0;
+      __syncthreads();
+
+      for (int w = 0; w < future_context; w++) {
+        sh_dout[thx * block_y + thy] =
+            (d < input_dim && (k - w) >= 0 && (k - w) < current_timesteps)
+                ? dout[(pos - w) * input_dim + d]
+                : 0.0;
+        __syncthreads();
+
+        T val = sh_in[thy * block_y + thx] * sh_dout[thy * block_y + thx];
+        __syncthreads();
+
+        for (int offset = 16; offset > 0;
+             offset = offset / 2) {  // blockDim.x is 32.
+          val += __shfl_down(val, offset);
+        }
+        __syncthreads();
+
+        if (thx == 0 && (gx + thy) < input_dim) {
+          dfilter[w * input_dim + gx + thy] += val;
+        }
+      }
+    }
+  }
+}
+
+}  // namespace
+
+template <typename T>
+class RowConvKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto *X = context.Input<LoDTensor>("X");
+    auto *Filter = context.Input<Tensor>("Filter");
+    auto *Out = context.Output<LoDTensor>("Out");
+
+    const T *in = X->data<T>();
+    const T *weight = Filter->data<T>();
+    T *out = Out->mutable_data<T>(context.GetPlace());
+
+    auto batch_indices = X->lod()[0];
+    int input_dim = X->dims()[1];
+    int num_sequence = batch_indices.size() - 1;
+    int future_context = Filter->dims()[0];
+    size_t *idx = batch_indices.data();
+    auto stream = context.cuda_device_context().stream();
+
+    if (future_context <= 32) {
+      dim3 block_dim = dim3(32, 32);
+      dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1);
+      int mem_per_block = (future_context * block_dim.x) * sizeof(T);
+      RowConvForwardSharedMemory<
+          T><<<grid_dim, block_dim, mem_per_block, stream>>>(
+          in, weight, num_sequence, input_dim, future_context, idx, out);
+    } else {
+      dim3 block_dim = dim3(32, 32);
+      dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1);
+      RowConvForward<T><<<grid_dim, block_dim, 0, stream>>>(
+          in, weight, num_sequence, input_dim, future_context, idx, out);
+    }
+  }
+};
+
+template <typename T>
+class RowConvGradKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto *X = context.Input<LoDTensor>("X");
+    auto *Filter = context.Input<Tensor>("Filter");
+    auto *dOut = context.Input<LoDTensor>(framework::GradVarName("Out"));
+    const T *in = X->data<T>();
+    const T *weights = Filter->data<T>();
+    const T *dout = dOut->data<T>();
+
+    Tensor *dX = context.Output<LoDTensor>(framework::GradVarName("X"));
+    Tensor *dFilter = context.Output<Tensor>(framework::GradVarName("Filter"));
+
+    auto batch_indices = X->lod()[0];
+    int input_dim = X->dims()[1];
+    int num_sequence = batch_indices.size() - 1;
+    int future_context = Filter->dims()[0];
+    size_t *idx = batch_indices.data();
+
+    auto &device_ctx = context.cuda_device_context();
+    math::SetConstant<platform::GPUPlace, T> zero;
+
+    if (dFilter) {
+      T *dfilter = dFilter->mutable_data<T>(context.GetPlace());
+      zero(device_ctx, dFilter, static_cast<T>(0.0));
+
+      if (future_context <= 32) {
+        dim3 block_dim = dim3(32, 32);
+        dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1);
+        int block_x = block_dim.x;
+        int block_y = block_dim.y;
+        int mem_per_block =
+            (block_y * block_x + block_y * (block_x + future_context - 1) +
+             future_context * block_y) *
+            sizeof(T);
+        RowConvGradFilterImproved<
+            T><<<grid_dim, block_dim, mem_per_block, device_ctx.stream()>>>(
+            in, dout, num_sequence, input_dim, future_context, block_x, block_y,
+            idx, dfilter);
+      } else {
+        dim3 block_dim = dim3(32, 32);
+        dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1);
+        int block_x = block_dim.x;
+        int block_y = block_dim.y;
+        int mem_per_block =
+            (block_x * block_y * 2) * sizeof(T);  // For 2 arrays of size 32x32
+        RowConvGradFilter<
+            T><<<grid_dim, block_dim, mem_per_block, device_ctx.stream()>>>(
+            in, dout, num_sequence, input_dim, future_context, block_x, block_y,
+            idx, dfilter);
+      }
+    }
+
+    if (dX) {
+      T *din = dX->mutable_data<T>(context.GetPlace());
+      if (future_context <= 32) {
+        dim3 block_dim = dim3(32, 32);
+        dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1);
+        int mem_per_block = (future_context * block_dim.x) * sizeof(T);
+        RowConvGradInputSharedMemory<
+            T><<<grid_dim, block_dim, mem_per_block, device_ctx.stream()>>>(
+            dout, weights, num_sequence, input_dim, future_context, idx, din);
+      } else {
+        dim3 block_dim = dim3(32, 32);
+        dim3 grid_dim = dim3(DivUp(input_dim, block_dim.x), 1);
+        RowConvGradInput<T><<<grid_dim, block_dim, 0, device_ctx.stream()>>>(
+            dout, weights, num_sequence, input_dim, future_context, idx, din);
+      }
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(row_conv,
+                       ops::RowConvKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    row_conv_grad, ops::RowConvGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/row_conv_op.h b/paddle/operators/row_conv_op.h
new file mode 100644
index 0000000000..525e83908d
--- /dev/null
+++ b/paddle/operators/row_conv_op.h
@@ -0,0 +1,33 @@
+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class RowConvKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override;
+};
+
+template <typename Place, typename T>
+class RowConvGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override;
+};
+}  // namespace operators
+}  // namespace paddle
diff --git a/python/paddle/v2/fluid/tests/test_row_conv_op.py b/python/paddle/v2/fluid/tests/test_row_conv_op.py
new file mode 100644
index 0000000000..1ed86e23ac
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/test_row_conv_op.py
@@ -0,0 +1,95 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+def row_conv_forward(x, lod, wt):
+    out = np.zeros_like(x)
+    seq_info = lod[0]
+    num_sequences = len(seq_info) - 1
+    context_length = wt.shape[0]
+
+    for i in range(num_sequences):  # loop over number of sequences
+        start = seq_info[i]
+        end = seq_info[i + 1]
+        curinput = x[start:end, :]
+        curoutput = out[start:end, :]
+
+        cur_timesteps = end - start
+        for j in range(cur_timesteps):  # loop over different timesteps
+            for k in range(context_length):
+
+                if j + k >= cur_timesteps:
+                    continue
+                curoutput[j, :] += curinput[j + k, :] * wt[k, :]
+
+    return out
+
+
+class TestRowConvOp1(OpTest):
+    def setUp(self):
+
+        self.op_type = "row_conv"
+        lod = [[0, 2, 5, 7]]
+        T = lod[0][-1]
+        D = 16
+        context_length = 2
+
+        x = np.random.random((T, D)).astype("float32")
+        wt = np.random.random((context_length, D)).astype("float32")
+        self.inputs = {'X': (x, lod), 'Filter': wt}
+
+        out = row_conv_forward(x, lod, wt)
+        self.outputs = {'Out': (out, lod)}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Filter'], 'Out', max_relative_error=0.05)
+
+    def test_check_grad_ignore_x(self):
+        self.check_grad(
+            ['Filter'], 'Out', max_relative_error=0.05, no_grad_set=set('X'))
+
+    def test_check_grad_ignore_wt(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Filter'))
+
+
+class TestRowConvOp2(OpTest):
+    def setUp(self):
+
+        self.op_type = "row_conv"
+        lod = [[0, 20, 50, 100]]
+        T = lod[0][-1]
+        D = 35
+        context_length = 35
+
+        x = np.random.random((T, D)).astype("float32")
+        wt = np.random.random((context_length, D)).astype("float32")
+        self.inputs = {'X': (x, lod), 'Filter': wt}
+
+        out = row_conv_forward(x, lod, wt)
+        self.outputs = {'Out': (out, lod)}
+
+    def test_check_output(self):
+        self.check_output()
+
+    #max_relative_error is increased from 0.05 to 0.06 as for higher
+    #dimensional input, the dX on CPU for some values has max_rel_error 
+    #slightly more than 0.05
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Filter'], 'Out', max_relative_error=0.06)
+
+    def test_check_grad_ignore_x(self):
+        self.check_grad(
+            ['Filter'], 'Out', max_relative_error=0.06, no_grad_set=set('X'))
+
+    def test_check_grad_ignore_wt(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.06, no_grad_set=set('Filter'))
+
+
+if __name__ == '__main__':
+    unittest.main()

From 35420cdf63dd1369972c26f70cac2d4d75b1492a Mon Sep 17 00:00:00 2001
From: kavyasrinet <kavyasrinet@baidu.com>
Date: Mon, 11 Dec 2017 15:33:28 -0800
Subject: [PATCH 24/94] Updating the Latex equation for Adagrad (#6009)

* Updating the Latex equation for Adagrad

* Fixing Latex euqations for adadelta, adam and adamax
---
 paddle/operators/adadelta_op.cc | 12 ++++++------
 paddle/operators/adagrad_op.cc  |  4 ++--
 paddle/operators/adam_op.cc     | 12 +++++++-----
 paddle/operators/adamax_op.cc   |  8 ++++----
 4 files changed, 19 insertions(+), 17 deletions(-)

diff --git a/paddle/operators/adadelta_op.cc b/paddle/operators/adadelta_op.cc
index 16a7794d5b..29434a0ee2 100644
--- a/paddle/operators/adadelta_op.cc
+++ b/paddle/operators/adadelta_op.cc
@@ -92,12 +92,12 @@ for gradient descent.
 
 Adadelta updates are as follows:
 
-$$avgSquaredGradOut = \rho * avgSquaredGrad + (1 - \rho) * grad * grad \break
-paramUpdate =  - $\sqrt{((avgSquaredUpdate + \epsilon) /
-                       (avgSquaredGrad_out + \epsilon))}$ * grad \break
-avgSquaredUpdateOut = \rho * avgSquaredUpdate + (1 - \rho) *
-                                  {(paramUpdate)}^2 \break
-paramOut = param + paramUpdate$$
+$$
+avg\_squared\_grad\_out = \rho * avg\_squared\_grad + (1 - \rho) * grad * grad \\
+param\_update =  - \sqrt{\frac{avg\_squared\_update + \epsilon}{avg\_squared\_grad\_out + \epsilon}} * grad \\
+avg\_squared\_update\_out = \rho * avg\_squared\_update + (1 - \rho) * {param\_update}^2 \\
+param\_out = param + param\_update
+$$
 
 )DOC");
   }
diff --git a/paddle/operators/adagrad_op.cc b/paddle/operators/adagrad_op.cc
index d6686e3ef3..d19602244b 100644
--- a/paddle/operators/adagrad_op.cc
+++ b/paddle/operators/adagrad_op.cc
@@ -80,8 +80,8 @@ Adaptive Gradient Algorithm (Adagrad).
 
 The update is done as follows:
 
-$$momentOut = moment + grad * grad \break
-paramOut = param - learningRate * grad / ($\sqrt{momentOut}$ + \epsilon) \break
+$$moment\_out = moment + grad * grad \\
+param\_out = param - \frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon}
 $$
 
 The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
diff --git a/paddle/operators/adam_op.cc b/paddle/operators/adam_op.cc
index 03faa2a7c5..a268d05484 100644
--- a/paddle/operators/adam_op.cc
+++ b/paddle/operators/adam_op.cc
@@ -112,11 +112,13 @@ adaptive estimates of lower-order moments.
 
 Adam updates:
 
-$$moment_1_{out} = \beta_1 * moment_1 + (1 - \beta_1) * grad \break
-moment_2_{out} = \beta_2 * moment_2 + (1 - \beta_2) * grad * grad \break
-learningRate = learningRate *
-                  $\sqrt{(1 - \beta_2_{pow})}$ / (1 - \beta_1_{pow}) \break
-paramOut = param - learningRate * moment_1/ ($\sqrt{(moment_2)} + \epsilon)$$
+$$
+moment\_1\_out = \beta_1 * moment\_1 + (1 - \beta_1) * grad \\
+moment\_2_\out = \beta_2 * moment\_2 + (1 - \beta_2) * grad * grad \\
+learning\_rate = learning\_rate *
+                  \frac{\sqrt{1 - \beta_{2\_pow}}}{1 - \beta_{1\_pow}} \\
+param\_out = param - learning\_rate * \frac{moment\_1}{\sqrt{moment\_2} + \epsilon}
+$$
 
 )DOC");
   }
diff --git a/paddle/operators/adamax_op.cc b/paddle/operators/adamax_op.cc
index 867ddd9790..9e7576c961 100644
--- a/paddle/operators/adamax_op.cc
+++ b/paddle/operators/adamax_op.cc
@@ -108,10 +108,10 @@ Adam algorithm based on the infinity norm.
 Adamax updates:
 
 $$
-  momentOut = \beta_{1} * moment + (1 - \beta_{1}) * grad \\
-  infNormOut = max(\beta_{2} * infNorm + \epsilon, |grad|) \\
-  learningRate = \frac{learningRate}{1 - \beta_{1}^{Beta1Pow}} \\
-  paramOut = param - learningRate * \frac{momentOut}{infNormOut}
+moment\_out = \beta_1 * moment + (1 - \beta_1) * grad \\
+inf\_norm\_out = max(\beta_2 * inf\_norm + \epsilon, |grad|) \\
+learning\_rate = \frac{learning\_rate}{1 - \beta_{1\_pow}} \\
+param\_out = param - learning\_rate * \frac{moment\_out}{inf\_norm\_out}
 $$
 
 The original paper does not have an epsilon attribute.

From f4f17e539b70a6cdb4245267ae4027cd5202b0fa Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Tue, 12 Dec 2017 16:35:13 +0800
Subject: [PATCH 25/94] skip mkl setting in v1 with Mac

---
 paddle/scripts/submit_local.sh.in | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/paddle/scripts/submit_local.sh.in b/paddle/scripts/submit_local.sh.in
index d71cb84df3..43d2d1b410 100755
--- a/paddle/scripts/submit_local.sh.in
+++ b/paddle/scripts/submit_local.sh.in
@@ -140,7 +140,11 @@ else:
   sys.exit(0)
 EOF
 
-cpu_config
+if [ "`uname -s`" == "Linux" ]; then
+  # only support on linux yet, with mac can use v2
+  cpu_config
+fi
+
 # echo $KMP_AFFINITY $OMP_DYNAMIC
 
 case "$1" in

From c175eeb387ee90d8ba8a549a2e29a1680c9f1f35 Mon Sep 17 00:00:00 2001
From: wangmeng28 <wangmeng28@baidu.com>
Date: Tue, 12 Dec 2017 11:08:23 +0800
Subject: [PATCH 26/94] Fix wrong index in dataset downloading exception

---
 python/paddle/v2/dataset/common.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/python/paddle/v2/dataset/common.py b/python/paddle/v2/dataset/common.py
index e31e501ce9..191d9ecfb1 100644
--- a/python/paddle/v2/dataset/common.py
+++ b/python/paddle/v2/dataset/common.py
@@ -71,7 +71,7 @@ def download(url, module_name, md5sum):
         if retry < retry_limit:
             retry += 1
         else:
-            raise RuntimeError("Cannot download {0} within retry limit {2}".
+            raise RuntimeError("Cannot download {0} within retry limit {1}".
                                format(url, retry_limit))
         print "Cache file %s not found, downloading %s" % (filename, url)
         r = requests.get(url, stream=True)

From f3acdd3af94cb3018134b13893cffebdf09c827c Mon Sep 17 00:00:00 2001
From: Luo Tao <luotao02@baidu.com>
Date: Tue, 12 Dec 2017 11:17:33 +0800
Subject: [PATCH 27/94] fix warning in row_conv_op.cu

---
 paddle/operators/row_conv_op.cu | 1 -
 1 file changed, 1 deletion(-)

diff --git a/paddle/operators/row_conv_op.cu b/paddle/operators/row_conv_op.cu
index e0d7ebda7e..79b7086b24 100644
--- a/paddle/operators/row_conv_op.cu
+++ b/paddle/operators/row_conv_op.cu
@@ -243,7 +243,6 @@ __global__ void RowConvGradFilter(const T *in, const T *dout, int num_sequence,
                                   int block_x, int block_y,
                                   const size_t *batch_indices, T *dfilter) {
   int blx = blockDim.x;
-  int bly = blockDim.y;
   int thx = threadIdx.x;
   int thy = threadIdx.y;
   int gx = blockIdx.x * blx;

From 59c74fb14adcd23fc244275ef9e74b8430cc43fc Mon Sep 17 00:00:00 2001
From: Luo Tao <luotao02@baidu.com>
Date: Tue, 12 Dec 2017 11:35:38 +0800
Subject: [PATCH 28/94] change paddle:dev to paddle:latest-dev

---
 doc/howto/dev/contribute_to_paddle_cn.md | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/doc/howto/dev/contribute_to_paddle_cn.md b/doc/howto/dev/contribute_to_paddle_cn.md
index 3eb477eb65..3e0bf7b397 100644
--- a/doc/howto/dev/contribute_to_paddle_cn.md
+++ b/doc/howto/dev/contribute_to_paddle_cn.md
@@ -76,18 +76,18 @@ no changes added to commit (use "git add" and/or "git commit -a")
 
 ## 构建和测试
 
-编译 PaddlePaddle 的源码以及生成文档需要多种开发工具。为了方便大家，我们的标准开发流程是把这些工具都装进一个Docker image，称为*开发镜像*，通常名字是 `paddle:dev`。然后所有用 `cmake && make` 的地方（比如IDE配置里）都用 `docker run paddle:dev`来代替。
+编译 PaddlePaddle 的源码以及生成文档需要多种开发工具。为了方便大家，我们的标准开发流程是把这些工具都装进一个Docker image，称为*开发镜像*，通常名字是 `paddle:latest-dev` 或者 `paddle:[version tag]-dev` 如 `paddle:0.11.0-dev`。然后所有用 `cmake && make` 的地方（比如IDE配置里）都用 `docker run paddle:latest-dev`来代替。
 
 如要build这个开发镜像，在源码目录树的根目录中运行：
 
 ```bash
-➜  docker build -t paddle:dev .
+➜  docker build -t paddle:latest-dev .
 ```
 
 随后可以用这个开发镜像开始build PaddlePaddle的源码。比如如果要build一个不依赖GPU，但是支持AVX指令集，并且包括unit tests的PaddlePaddle，可以：
 
 ```bash
-➜  docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TESTING=ON" paddle:dev
+➜  docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TESTING=ON" paddle:latest-dev
 ```
 
 这个过程除了编译PaddlePaddle为 `./build/libpaddle.so`，并且输出一个 `./build/paddle.deb`文件之外，还会输出一个 `build/Dockerfile`。我们只需要运行下面命令把编译好的PaddlePaddle打包成一个*生产镜像*（`paddle:prod`）：
@@ -99,7 +99,7 @@ no changes added to commit (use "git add" and/or "git commit -a")
 如果要运行所有的单元测试，可以用如下命令：
 
 ```bash
-➜  docker run -it -v $(pwd):/paddle paddle:dev bash -c "cd /paddle/build && ctest"
+➜  docker run -it -v $(pwd):/paddle paddle:latest-dev bash -c "cd /paddle/build && ctest"
 ```
 
 关于构建和测试的更多信息，请参见[这篇文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)。

From 936f0546e3aaa772514bba721167e897769ec2a9 Mon Sep 17 00:00:00 2001
From: chengduoZH <zhaochengduo@163.com>
Date: Tue, 12 Dec 2017 10:24:50 +0800
Subject: [PATCH 29/94] fix img_pool maxout doc

---
 .../paddle/trainer_config_helpers/layers.py   | 48 ++++++++++++-------
 1 file changed, 31 insertions(+), 17 deletions(-)

diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
index d0b14cf63c..3a82e858f7 100644
--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -2729,15 +2729,17 @@ def img_pool_layer(input,
 
     ..  math::
 
-        w = 1 + \frac{ceil(input\_width + 2 * padding - pool\_size)}{stride} \\\\
-        h = 1 + \frac{ceil(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y}
+        w & = 1 + \\frac{ceil(input\_width + 2 * padding - pool\_size)}{stride}
+
+        h & = 1 + \\frac{ceil(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y}
 
     - ceil_mode=False:
 
     ..  math::
 
-        w = 1 + \frac{floor(input\_width + 2 * padding - pool\_size)}{stride} \\\\
-        h = 1 + \frac{floor(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y}
+        w & = 1 + \\frac{floor(input\_width + 2 * padding - pool\_size)}{stride}
+
+        h & = 1 + \\frac{floor(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y}
 
     The example usage is:
 
@@ -2870,17 +2872,21 @@ def img_pool3d_layer(input,
 
     ..  math::
 
-        w = 1 + \frac{ceil(input\_width + 2 * padding - pool\_size)}{stride} \\\\
-        h = 1 + \frac{ceil(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y} \\\\
-        d = 1 + \frac{ceil(input\_depth + 2 * padding\_z - pool\_size\_z)}{stride\_z}
+        w & = 1 + \\frac{ceil(input\_width + 2 * padding - pool\_size)}{stride}
+
+        h & = 1 + \\frac{ceil(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y}
+
+        d & = 1 + \\frac{ceil(input\_depth + 2 * padding\_z - pool\_size\_z)}{stride\_z}
 
     - ceil_mode=False:
 
     ..  math::
 
-        w = 1 + \frac{floor(input\_width + 2 * padding - pool\_size)}{stride} \\\\
-        h = 1 + \frac{floor(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y} \\\\
-        d = 1 + \frac{floor(input\_depth + 2 * padding\_z - pool\_size\_z)}{stride\_z} \\\\
+        w & = 1 + \\frac{floor(input\_width + 2 * padding - pool\_size)}{stride}
+
+        h & = 1 + \\frac{floor(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y}
+
+        d & = 1 + \\frac{floor(input\_depth + 2 * padding\_z - pool\_size\_z)}{stride\_z}
 
     The example usage is:
 
@@ -5437,13 +5443,21 @@ def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None):
 
 
     .. math::
-       out = \max_k (in[n, k, o_c , s])   \\\\
-       out_{i * s + j} = \max_k in_{  k * o_{c} * s + i * s + j}  \\\\
-       s = \frac{input.size}{ num\_channels}       \\\\
-       o_{c} =\frac{num\_channels}{groups}         \\\\
-       0 \le i < o_{c}                             \\\\
-       0 \le j < s                                 \\\\
-       0 \le k < groups                            \\\\
+
+       out & = \max_k (in[n, k, o_c , s])
+
+       out_{i * s + j} & = \max_k in_{  k * o_{c} * s + i * s + j}
+
+       s & = \\frac{input.size}{ num\_channels}
+
+       o_{c} & = \\frac{num\_channels}{groups}
+
+       0 \le i & < o_{c}
+
+       0 \le j & < s
+
+       0 \le k & < groups
+
 
     The simple usage is:
 

From 7d8802938149df6b0dcf8934c0eb7fe0e4e5145c Mon Sep 17 00:00:00 2001
From: chengduoZH <zhaochengduo@163.com>
Date: Tue, 12 Dec 2017 12:55:58 +0800
Subject: [PATCH 30/94] equation align

---
 python/paddle/trainer_config_helpers/layers.py | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
index 3a82e858f7..7e118b24a4 100644
--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -5444,19 +5444,19 @@ def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None):
 
     .. math::
 
-       out & = \max_k (in[n, k, o_c , s])
+       & out = \max_k (in[n, k, o_c , s])
 
-       out_{i * s + j} & = \max_k in_{  k * o_{c} * s + i * s + j}
+       & out_{i * s + j} = \max_k in_{  k * o_{c} * s + i * s + j}
 
-       s & = \\frac{input.size}{ num\_channels}
+       & s = \\frac{input.size}{ num\_channels}
 
-       o_{c} & = \\frac{num\_channels}{groups}
+       & o_{c} = \\frac{num\_channels}{groups}
 
-       0 \le i & < o_{c}
+       & 0 \le i < o_{c}
 
-       0 \le j & < s
+       & 0 \le j < s
 
-       0 \le k & < groups
+       & 0 \le k < groups
 
 
     The simple usage is:

From 61ec0b951656fb402e61c4dc519e80ba3fbc61d0 Mon Sep 17 00:00:00 2001
From: QI JUN <qijun1994@hotmail.com>
Date: Tue, 12 Dec 2017 14:00:28 +0800
Subject: [PATCH 31/94] Refine device context (#6433)

There are mainly following fixes:

- take `DeviceContext` as the template parameter of math functors and OpKernel instead of `Place`
- remove `eigen_device` interface in base class  `DeviceContext`
- remove `GetEigenDevice` interface in `ExecutionContext` and base class `DeviceContext`
- remove unused `platform::EigenDeviceConverter`
- rename `REGISTER_OP_GPU_KERNEL` to `REGISTER_OP_CUDA_KERNEL`
- rename `USE_GPU_ONLY_OP` to `USE_CUDA_ONLY_OP`
---
 paddle/framework/op_registry.h                |  12 +-
 paddle/framework/operator.cc                  |  16 +-
 paddle/framework/operator.h                   |  26 +--
 paddle/framework/operator_test.cc             |   2 +-
 paddle/operators/CMakeLists.txt               |   2 +-
 paddle/operators/accuracy_op.cc               |   2 +-
 paddle/operators/accuracy_op.cu               |   5 +-
 paddle/operators/accuracy_op.h                |   2 +-
 paddle/operators/activation_op.cc             |  21 +-
 paddle/operators/activation_op.cu             |  23 +-
 paddle/operators/activation_op.h              |  14 +-
 paddle/operators/adadelta_op.cc               |   4 +-
 paddle/operators/adadelta_op.cu               |   6 +-
 paddle/operators/adadelta_op.h                |   4 +-
 paddle/operators/adagrad_op.cc                |  18 +-
 paddle/operators/adagrad_op.cu                |  20 +-
 paddle/operators/adagrad_op.h                 |  17 +-
 paddle/operators/adam_op.cc                   |   6 +-
 paddle/operators/adam_op.cu                   |   6 +-
 paddle/operators/adam_op.h                    |  10 +-
 paddle/operators/adamax_op.cc                 |   6 +-
 paddle/operators/adamax_op.cu                 |   6 +-
 paddle/operators/adamax_op.h                  |  10 +-
 paddle/operators/auc_op.h                     |   2 +-
 paddle/operators/batch_norm_op.cc             |  14 +-
 paddle/operators/batch_norm_op.cu.cc          |  32 +--
 paddle/operators/batch_norm_op.h              |   4 +-
 .../operators/bilinear_tensor_product_op.cc   |  11 +-
 .../operators/bilinear_tensor_product_op.cu   |  16 +-
 paddle/operators/bilinear_tensor_product_op.h |  41 ++--
 paddle/operators/cast_op.cc                   |   2 +-
 paddle/operators/cast_op.cu                   |   6 +-
 paddle/operators/cast_op.h                    |  13 +-
 paddle/operators/chunk_eval_op.h              |   2 +-
 paddle/operators/clip_by_norm_op.cc           |   3 +-
 paddle/operators/clip_by_norm_op.cu           |   5 +-
 paddle/operators/clip_by_norm_op.h            |   5 +-
 paddle/operators/clip_op.cc                   |   8 +-
 paddle/operators/clip_op.cu                   |   8 +-
 paddle/operators/clip_op.h                    |  16 +-
 paddle/operators/compare_op.cu                |  10 +-
 paddle/operators/compare_op.h                 |  31 ++-
 paddle/operators/concat_op.cu.cc              |   9 +-
 paddle/operators/concat_op.h                  |   4 +-
 paddle/operators/conv_cudnn_op.cc             |  22 +-
 paddle/operators/conv_cudnn_op.cu.cc          |  32 +--
 paddle/operators/conv_op.cc                   |  22 +-
 paddle/operators/conv_op.cu.cc                |  26 ++-
 paddle/operators/conv_op.h                    |  55 +++--
 paddle/operators/conv_shift_op.cu             |  22 +-
 paddle/operators/conv_shift_op.h              |   4 +-
 paddle/operators/conv_transpose_cudnn_op.cc   |  18 +-
 .../operators/conv_transpose_cudnn_op.cu.cc   |  32 +--
 paddle/operators/conv_transpose_op.cc         |  18 +-
 paddle/operators/conv_transpose_op.cu.cc      |  28 ++-
 paddle/operators/conv_transpose_op.h          |  58 ++---
 paddle/operators/cos_sim_op.cc                |   7 +-
 paddle/operators/cos_sim_op.cu                |   9 +-
 paddle/operators/cos_sim_op.h                 |  10 +-
 paddle/operators/crf_decoding_op.cc           |   5 +-
 paddle/operators/crf_decoding_op.h            |   6 +-
 paddle/operators/crop_op.cc                   |   4 +-
 paddle/operators/crop_op.cu                   |   6 +-
 paddle/operators/crop_op.h                    |  19 +-
 paddle/operators/cross_entropy_op.cu          |  23 +-
 paddle/operators/cross_entropy_op.h           |  12 +-
 paddle/operators/decayed_adagrad_op.cc        |   2 +-
 paddle/operators/decayed_adagrad_op.cu        |   4 +-
 paddle/operators/decayed_adagrad_op.h         |   4 +-
 paddle/operators/dropout_op.cc                |   6 +-
 paddle/operators/dropout_op.cu                |  12 +-
 paddle/operators/dropout_op.h                 |  10 +-
 paddle/operators/elementwise_add_op.cc        |  16 +-
 paddle/operators/elementwise_add_op.cu        |  21 +-
 paddle/operators/elementwise_add_op.h         |  10 +-
 paddle/operators/elementwise_div_op.cc        |  16 +-
 paddle/operators/elementwise_div_op.cu        |  21 +-
 paddle/operators/elementwise_div_op.h         |   8 +-
 paddle/operators/elementwise_mul_op.cc        |  16 +-
 paddle/operators/elementwise_mul_op.cu        |  21 +-
 paddle/operators/elementwise_mul_op.h         |   8 +-
 paddle/operators/elementwise_op_function.h    |  94 ++++----
 paddle/operators/elementwise_sub_op.cc        |  16 +-
 paddle/operators/elementwise_sub_op.cu        |  21 +-
 paddle/operators/elementwise_sub_op.h         |   8 +-
 paddle/operators/expand_op.cc                 |   7 +-
 paddle/operators/expand_op.cu                 |   9 +-
 paddle/operators/expand_op.h                  |  10 +-
 .../fill_constant_batch_size_like_op.cc       |  11 +-
 .../fill_constant_batch_size_like_op.cu.cc    |  13 +-
 .../fill_constant_batch_size_like_op.h        |   7 +-
 paddle/operators/fill_zeros_like_op.cc        |  11 +-
 paddle/operators/fill_zeros_like_op.cu.cc     |  13 +-
 paddle/operators/fill_zeros_like_op.h         |   7 +-
 paddle/operators/ftrl_op.cc                   |   4 +-
 paddle/operators/ftrl_op.cu                   |   4 +-
 paddle/operators/ftrl_op.h                    |   4 +-
 paddle/operators/gather.cu.h                  |   2 +-
 paddle/operators/gather_op.cu                 |   7 +-
 paddle/operators/gather_op.h                  |   3 +-
 paddle/operators/gaussian_random_op.cu        |   4 +-
 paddle/operators/gru_op.cc                    |  11 +-
 paddle/operators/gru_op.cu.cc                 |  11 +-
 paddle/operators/gru_op.h                     |  48 ++--
 paddle/operators/gru_unit_op.cc               |  11 +-
 paddle/operators/gru_unit_op.cu               |  13 +-
 paddle/operators/gru_unit_op.h                |  64 +++---
 paddle/operators/hinge_loss_op.cc             |   7 +-
 paddle/operators/hinge_loss_op.cu             |   9 +-
 paddle/operators/hinge_loss_op.h              |  10 +-
 paddle/operators/huber_loss_op.cc             |   7 +-
 paddle/operators/huber_loss_op.cu             |   9 +-
 paddle/operators/huber_loss_op.h              |  10 +-
 paddle/operators/l1_norm_op.cc                |   7 +-
 paddle/operators/l1_norm_op.cu                |   9 +-
 paddle/operators/l1_norm_op.h                 |  10 +-
 paddle/operators/linear_chain_crf_op.cc       |   9 +-
 paddle/operators/linear_chain_crf_op.cu       |  13 +-
 paddle/operators/linear_chain_crf_op.h        |  24 +-
 paddle/operators/lod_reset_op.cu              |  13 +-
 paddle/operators/lod_reset_op.h               |   4 +-
 paddle/operators/log_loss_op.cc               |   7 +-
 paddle/operators/log_loss_op.cu               |   9 +-
 paddle/operators/log_loss_op.h                |   8 +-
 paddle/operators/logical_op.cu                |   8 +-
 paddle/operators/logical_op.h                 |  21 +-
 paddle/operators/lookup_table_op.cu           |  20 +-
 paddle/operators/lrn_op.cc                    |  21 +-
 paddle/operators/lrn_op.cu                    |  30 +--
 paddle/operators/lrn_op.h                     |  10 +-
 paddle/operators/lstm_op.cc                   |  11 +-
 paddle/operators/lstm_op.cu.cc                |  11 +-
 paddle/operators/lstm_op.h                    |  94 ++++----
 paddle/operators/lstm_unit_op.cu              |   8 +-
 paddle/operators/lstm_unit_op.h               |   4 +-
 paddle/operators/margin_rank_loss_op.cc       |   4 +-
 paddle/operators/margin_rank_loss_op.cu       |   8 +-
 paddle/operators/margin_rank_loss_op.h        |   8 +-
 paddle/operators/math/context_project.cc      |   4 +-
 paddle/operators/math/context_project.cu      |   4 +-
 paddle/operators/math/context_project.h       |  20 +-
 paddle/operators/math/cross_entropy.cc        |  10 +-
 paddle/operators/math/cross_entropy.cu        |  14 +-
 paddle/operators/math/cross_entropy.h         |   6 +-
 paddle/operators/math/gru_compute.cc          |  28 +--
 paddle/operators/math/gru_compute.cu          |  34 ++-
 paddle/operators/math/gru_compute.h           |  13 +-
 paddle/operators/math/im2col.cc               |  32 +--
 paddle/operators/math/im2col.cu               |  48 ++--
 paddle/operators/math/im2col.h                |  11 +-
 paddle/operators/math/im2col_test.cc          |  27 +--
 paddle/operators/math/lstm_compute.cc         |  16 +-
 paddle/operators/math/lstm_compute.cu         |  16 +-
 paddle/operators/math/lstm_compute.h          |  13 +-
 paddle/operators/math/math_function.cc        | 144 ++++++------
 paddle/operators/math/math_function.cu        | 207 ++++++++----------
 paddle/operators/math/math_function.h         |  81 ++++---
 paddle/operators/math/math_function_impl.h    |  39 ++--
 paddle/operators/math/math_function_test.cc   |   9 +-
 paddle/operators/math/math_function_test.cu   |  10 +-
 paddle/operators/math/matmul.h                |  19 +-
 paddle/operators/math/maxouting.cc            |  16 +-
 paddle/operators/math/maxouting.cu            |  33 ++-
 paddle/operators/math/maxouting.h             |  14 +-
 paddle/operators/math/pooling.cc              | 128 ++++++-----
 paddle/operators/math/pooling.cu              | 182 +++++++--------
 paddle/operators/math/pooling.h               |  68 +++---
 .../operators/math/selected_rows_functor.cc   |  45 ++--
 .../operators/math/selected_rows_functor.cu   |  77 +++----
 paddle/operators/math/selected_rows_functor.h |  16 +-
 .../math/selected_rows_functor_test.cc        |  12 +-
 .../math/selected_rows_functor_test.cu        |  12 +-
 paddle/operators/math/sequence2batch.cc       |  16 +-
 paddle/operators/math/sequence2batch.cu       |  19 +-
 paddle/operators/math/sequence2batch.h        |  22 +-
 paddle/operators/math/sequence_pooling.cc     |  18 +-
 paddle/operators/math/sequence_pooling.cu     |  24 +-
 paddle/operators/math/sequence_pooling.h      |   8 +-
 paddle/operators/math/softmax.cc              |   8 +-
 paddle/operators/math/softmax.cu              |   8 +-
 paddle/operators/math/softmax.h               |  13 +-
 paddle/operators/math/softmax_impl.h          |  32 ++-
 paddle/operators/math/unpooling.cc            |  16 +-
 paddle/operators/math/unpooling.cu            |  36 ++-
 paddle/operators/math/unpooling.h             |  10 +-
 paddle/operators/math/vol2col.cc              |  16 +-
 paddle/operators/math/vol2col.cu              |  24 +-
 paddle/operators/math/vol2col.h               |  10 +-
 paddle/operators/math/vol2col_test.cc         |  24 +-
 paddle/operators/matmul_op.cc                 |   7 +-
 paddle/operators/matmul_op.cu.cc              |   9 +-
 paddle/operators/matmul_op.h                  |  45 ++--
 paddle/operators/maxout_op.cc                 |   7 +-
 paddle/operators/maxout_op.cu.cc              |  13 +-
 paddle/operators/maxout_op.h                  |  18 +-
 paddle/operators/mean_op.cc                   |  11 +-
 paddle/operators/mean_op.cu                   |  11 +-
 paddle/operators/mean_op.h                    |  10 +-
 paddle/operators/minus_op.cc                  |   4 +-
 paddle/operators/minus_op.cu                  |   5 +-
 paddle/operators/minus_op.h                   |   5 +-
 paddle/operators/modified_huber_loss_op.cc    |   2 +-
 paddle/operators/modified_huber_loss_op.cu    |   8 +-
 paddle/operators/modified_huber_loss_op.h     |   5 +-
 paddle/operators/momentum_op.cu               |   4 +-
 paddle/operators/mul_op.cc                    |   7 +-
 paddle/operators/mul_op.cu.cc                 |   7 +-
 paddle/operators/mul_op.h                     |  18 +-
 paddle/operators/multiplex_op.cc              |   5 +-
 paddle/operators/multiplex_op.cu              |  16 +-
 paddle/operators/multiplex_op.h               |  11 +-
 paddle/operators/nccl_op.cu.cc                |   6 +-
 paddle/operators/nccl_op_test.cu.cc           |   6 +-
 paddle/operators/nce_op.cc                    |   4 +-
 paddle/operators/nce_op.h                     |   8 +-
 paddle/operators/pad_op.cc                    |   7 +-
 paddle/operators/pad_op.cu                    |   7 +-
 paddle/operators/pad_op.h                     |  38 ++--
 paddle/operators/pool_cudnn_op.cc             |  26 ++-
 paddle/operators/pool_cudnn_op.cu.cc          |  18 +-
 paddle/operators/pool_op.cc                   |  24 +-
 paddle/operators/pool_op.cu.cc                |  26 ++-
 paddle/operators/pool_op.h                    |  59 ++---
 paddle/operators/pool_with_index_op.cc        |  22 +-
 paddle/operators/pool_with_index_op.cu.cc     |  32 ++-
 paddle/operators/pool_with_index_op.h         |  26 ++-
 paddle/operators/positive_negative_pair_op.h  |   2 +-
 paddle/operators/precision_recall_op.h        |   2 +-
 paddle/operators/prelu_op.cc                  |   9 +-
 paddle/operators/prelu_op.cu                  |  11 +-
 paddle/operators/prelu_op.h                   |  16 +-
 paddle/operators/proximal_adagrad_op.cc       |   2 +-
 paddle/operators/proximal_adagrad_op.cu       |   4 +-
 paddle/operators/proximal_adagrad_op.h        |  10 +-
 paddle/operators/proximal_gd_op.cc            |   3 +-
 paddle/operators/proximal_gd_op.cu            |   5 +-
 paddle/operators/proximal_gd_op.h             |   4 +-
 paddle/operators/rank_loss_op.cc              |   7 +-
 paddle/operators/rank_loss_op.cu              |  12 +-
 paddle/operators/rank_loss_op.h               |   8 +-
 paddle/operators/reduce_op.cc                 |  15 +-
 paddle/operators/reduce_op.cu                 |  15 +-
 paddle/operators/reduce_op.h                  |  44 ++--
 paddle/operators/reshape_op.cu                |   4 +-
 paddle/operators/reshape_op.h                 |   4 +-
 paddle/operators/rmsprop_op.cc                |   4 +-
 paddle/operators/rmsprop_op.cu                |   4 +-
 paddle/operators/rmsprop_op.h                 |   4 +-
 paddle/operators/roi_pool_op.cc               |   9 +-
 paddle/operators/roi_pool_op.cu               |  15 +-
 paddle/operators/roi_pool_op.h                |   9 +-
 paddle/operators/row_conv_op.cc               |  13 +-
 paddle/operators/row_conv_op.cu               |  17 +-
 paddle/operators/row_conv_op.h                |   4 +-
 paddle/operators/scale_op.cc                  |  10 +-
 paddle/operators/scale_op.cu                  |  12 +-
 paddle/operators/scale_op.h                   |   5 +-
 paddle/operators/scatter_op.cu                |   4 +-
 paddle/operators/seq_expand_op.cc             |   7 +-
 paddle/operators/seq_expand_op.cu             |   9 +-
 paddle/operators/seq_expand_op.h              |  14 +-
 paddle/operators/sequence_concat_op.cc        |   4 +-
 paddle/operators/sequence_concat_op.cu.cc     |  10 +-
 paddle/operators/sequence_concat_op.h         |   4 +-
 paddle/operators/sequence_conv_op.cc          |   9 +-
 paddle/operators/sequence_conv_op.cu.cc       |  13 +-
 paddle/operators/sequence_conv_op.h           |  65 +++---
 paddle/operators/sequence_pool_op.cc          |   5 +-
 paddle/operators/sequence_pool_op.cu          |   9 +-
 paddle/operators/sequence_pool_op.h           |  26 ++-
 paddle/operators/sequence_slice_op.cc         |   4 +-
 paddle/operators/sequence_slice_op.cu         |   8 +-
 paddle/operators/sequence_slice_op.h          |   9 +-
 paddle/operators/sequence_softmax_op.cc       |   4 +-
 paddle/operators/sequence_softmax_op.cu.cc    |   8 +-
 paddle/operators/sequence_softmax_op.h        |  12 +-
 paddle/operators/sgd_op.cc                    |  13 +-
 paddle/operators/sgd_op.cu                    |  22 +-
 paddle/operators/sgd_op.h                     |  14 +-
 .../sigmoid_cross_entropy_with_logits_op.cc   |   4 +-
 .../sigmoid_cross_entropy_with_logits_op.cu   |  12 +-
 .../sigmoid_cross_entropy_with_logits_op.h    |   9 +-
 paddle/operators/sign_op.cc                   |   4 +-
 paddle/operators/sign_op.cu                   |   5 +-
 paddle/operators/sign_op.h                    |   5 +-
 paddle/operators/smooth_l1_loss_op.cc         |   5 +-
 paddle/operators/smooth_l1_loss_op.cu         |   9 +-
 paddle/operators/smooth_l1_loss_op.h          |  30 +--
 paddle/operators/softmax_op.cc                |   7 +-
 paddle/operators/softmax_op.cu.cc             |   9 +-
 paddle/operators/softmax_op.h                 |  10 +-
 .../softmax_with_cross_entropy_op.cu          |  40 ++--
 .../operators/softmax_with_cross_entropy_op.h |  18 +-
 paddle/operators/split_op.cu.cc               |   4 +-
 paddle/operators/split_op.h                   |   2 +-
 paddle/operators/squared_l2_distance_op.cc    |   8 +-
 paddle/operators/squared_l2_distance_op.cu    |  10 +-
 paddle/operators/squared_l2_distance_op.h     |  10 +-
 paddle/operators/squared_l2_norm_op.cc        |   4 +-
 paddle/operators/squared_l2_norm_op.cu        |   8 +-
 paddle/operators/squared_l2_norm_op.h         |  14 +-
 paddle/operators/sum_op.cc                    |   9 +-
 paddle/operators/sum_op.cu                    |   9 +-
 paddle/operators/sum_op.h                     |  23 +-
 paddle/operators/top_k_op.cu                  |   2 +-
 paddle/operators/top_k_op.h                   |   2 +-
 paddle/operators/transpose_op.cc              |   6 +-
 paddle/operators/transpose_op.cu.cc           |   9 +-
 paddle/operators/transpose_op.h               |  29 +--
 paddle/operators/uniform_random_op.cc         |   2 +-
 paddle/operators/uniform_random_op.cu         |   6 +-
 paddle/operators/unpool_op.cc                 |  11 +-
 paddle/operators/unpool_op.cu.cc              |  13 +-
 paddle/operators/unpool_op.h                  |  22 +-
 paddle/platform/device_context.cc             |  12 -
 paddle/platform/device_context.h              |  17 --
 paddle/platform/device_context_test.cc        |   5 +-
 paddle/platform/transform.h                   |  24 +-
 paddle/platform/transform_test.cu             |   8 +-
 319 files changed, 2624 insertions(+), 2546 deletions(-)

diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h
index daade439e5..b29238432b 100644
--- a/paddle/framework/op_registry.h
+++ b/paddle/framework/op_registry.h
@@ -181,8 +181,8 @@ class OpKernelRegistrar : public Registrar {
     return 0;                                                             \
   }
 
-#define REGISTER_OP_GPU_KERNEL(op_type, ...) \
-  REGISTER_OP_KERNEL(op_type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
+#define REGISTER_OP_CUDA_KERNEL(op_type, ...) \
+  REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::GPUPlace, __VA_ARGS__)
 
 #define REGISTER_OP_CPU_KERNEL(op_type, ...) \
   REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
@@ -217,7 +217,7 @@ class OpKernelRegistrar : public Registrar {
 #else
 #define USE_OP_KERNEL(op_type)        \
   USE_OP_DEVICE_KERNEL(op_type, CPU); \
-  USE_OP_DEVICE_KERNEL(op_type, GPU)
+  USE_OP_DEVICE_KERNEL(op_type, CUDA)
 #endif
 
 #define USE_NO_KERNEL_OP(op_type) USE_OP_ITSELF(op_type);
@@ -226,9 +226,9 @@ class OpKernelRegistrar : public Registrar {
   USE_OP_ITSELF(op_type);        \
   USE_OP_DEVICE_KERNEL(op_type, CPU);
 
-#define USE_GPU_ONLY_OP(op_type) \
-  USE_OP_ITSELF(op_type);        \
-  USE_OP_DEVICE_KERNEL(op_type, GPU)
+#define USE_CUDA_ONLY_OP(op_type) \
+  USE_OP_ITSELF(op_type);         \
+  USE_OP_DEVICE_KERNEL(op_type, CUDA)
 
 #define USE_OP(op_type)   \
   USE_OP_ITSELF(op_type); \
diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc
index f1444eeee9..e83d754783 100644
--- a/paddle/framework/operator.cc
+++ b/paddle/framework/operator.cc
@@ -22,20 +22,6 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-template <>
-Eigen::DefaultDevice& ExecutionContext::GetEigenDevice<
-    platform::CPUPlace, Eigen::DefaultDevice>() const {
-  return *device_context_.GetEigenDevice<platform::CPUPlace>();
-}
-
-#ifdef PADDLE_WITH_CUDA
-template <>
-Eigen::GpuDevice&
-ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
-  return *device_context_.GetEigenDevice<platform::GPUPlace>();
-}
-#endif
-
 std::string OperatorBase::Input(const std::string& name) const {
   auto& ins = Inputs(name);
   PADDLE_ENFORCE_LE(ins.size(), 1UL,
@@ -429,7 +415,7 @@ void OperatorWithKernel::Run(const Scope& scope,
 }
 OpKernelType OperatorWithKernel::GetKernelType(
     const ExecutionContext& ctx) const {
-  return OpKernelType(IndicateDataType(ctx), ctx.device_context());
+  return OpKernelType(IndicateDataType(ctx), ctx.GetPlace());
 }
 DataType OperatorWithKernel::IndicateDataType(
     const ExecutionContext& ctx) const {
diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h
index 60861d9293..e60dbfc313 100644
--- a/paddle/framework/operator.h
+++ b/paddle/framework/operator.h
@@ -276,17 +276,25 @@ class ExecutionContext {
     out_tensor->set_lod(in_tensor.lod());
   }
 
-  template <typename PlaceType,
-            typename DeviceType = typename platform::EigenDeviceConverter<
-                PlaceType>::EigenDeviceType>
-  DeviceType& GetEigenDevice() const;
-
   platform::Place GetPlace() const { return device_context_.GetPlace(); }
 
+  template <typename DeviceContextType>
+  const DeviceContextType& device_context() const {
+    return *reinterpret_cast<const DeviceContextType*>(&device_context_);
+  }
+
   const platform::DeviceContext& device_context() const {
     return device_context_;
   }
 
+#ifdef PADDLE_WITH_CUDA
+  const inline platform::CUDADeviceContext& cuda_device_context() const {
+    PADDLE_ENFORCE(platform::is_gpu_place(device_context_.GetPlace()));
+    return *reinterpret_cast<const platform::CUDADeviceContext*>(
+        &device_context_);
+  }
+#endif
+
   //! Get actual name vector for this input.
   const std::vector<std::string>& Inputs(const std::string& name) const {
     return op_.Inputs(name);
@@ -297,14 +305,6 @@ class ExecutionContext {
     return op_.Outputs(name);
   }
 
-#ifdef PADDLE_WITH_CUDA
-  const inline platform::CUDADeviceContext& cuda_device_context() const {
-    PADDLE_ENFORCE(platform::is_gpu_place(device_context_.GetPlace()));
-    return *reinterpret_cast<const platform::CUDADeviceContext*>(
-        &device_context_);
-  }
-#endif
-
  private:
   const OperatorBase& op_;
   const Scope& scope_;
diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc
index 59ddbc7791..b678178454 100644
--- a/paddle/framework/operator_test.cc
+++ b/paddle/framework/operator_test.cc
@@ -115,7 +115,7 @@ class OpWithKernelTest : public OperatorWithKernel {
  protected:
   void InferShape(framework::InferShapeContext* ctx) const override {}
   OpKernelType GetKernelType(const ExecutionContext& ctx) const override {
-    return OpKernelType(DataType::FP32, ctx.device_context());
+    return OpKernelType(DataType::FP32, ctx.GetPlace());
   }
 };
 
diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index 38b89b9eb1..5aaaf99332 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -138,7 +138,7 @@ function(op_library TARGET)
     if ("${TARGET}" STREQUAL "nccl_op")
         set(pybind_flag 1)
         # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_GPU_ONLY_OP(ncclAllReduce);\n")
+        file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(ncclAllReduce);\n")
     endif()
 
     # reduce_op contains several operators
diff --git a/paddle/operators/accuracy_op.cc b/paddle/operators/accuracy_op.cc
index 2785a8c6fb..76da21c472 100644
--- a/paddle/operators/accuracy_op.cc
+++ b/paddle/operators/accuracy_op.cc
@@ -57,7 +57,7 @@ class AccuracyOp : public framework::OperatorWithKernel {
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Out")->type()),
-        ctx.device_context());
+        ctx.GetPlace());
   }
 };
 
diff --git a/paddle/operators/accuracy_op.cu b/paddle/operators/accuracy_op.cu
index d2dcab4e54..539a935302 100644
--- a/paddle/operators/accuracy_op.cu
+++ b/paddle/operators/accuracy_op.cu
@@ -104,5 +104,6 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
 
 // FIXME(typhoonzero): types of T is for inference data.
 // label data is always int64
-REGISTER_OP_GPU_KERNEL(accuracy, paddle::operators::AccuracyOpCUDAKernel<float>,
-                       paddle::operators::AccuracyOpCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(accuracy,
+                        paddle::operators::AccuracyOpCUDAKernel<float>,
+                        paddle::operators::AccuracyOpCUDAKernel<double>);
diff --git a/paddle/operators/accuracy_op.h b/paddle/operators/accuracy_op.h
index d060e6eddd..04104a695f 100644
--- a/paddle/operators/accuracy_op.h
+++ b/paddle/operators/accuracy_op.h
@@ -21,7 +21,7 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class AccuracyKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
diff --git a/paddle/operators/activation_op.cc b/paddle/operators/activation_op.cc
index 7f3118f176..63490f0ec9 100644
--- a/paddle/operators/activation_op.cc
+++ b/paddle/operators/activation_op.cc
@@ -611,16 +611,17 @@ REGISTER_OP(hard_sigmoid, ops::ActivationOp, ops::HardSigmoidOpMaker,
 REGISTER_OP(swish, ops::ActivationOp, ops::SwishOpMaker, swish_grad,
             ops::ActivationOpGrad);
 
-#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor)       \
-  REGISTER_OP_CPU_KERNEL(                                                     \
-      act_type,                                                               \
-      ops::ActivationKernel<paddle::platform::CPUPlace, ops::functor<float>>, \
-      ops::ActivationKernel<paddle::platform::CPUPlace,                       \
-                            ops::functor<double>>);                           \
-  REGISTER_OP_CPU_KERNEL(                                                     \
-      act_type##_grad, ops::ActivationGradKernel<paddle::platform::CPUPlace,  \
-                                                 ops::grad_functor<float>>,   \
-      ops::ActivationGradKernel<paddle::platform::CPUPlace,                   \
+#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor)   \
+  REGISTER_OP_CPU_KERNEL(                                                 \
+      act_type, ops::ActivationKernel<paddle::platform::CPUDeviceContext, \
+                                      ops::functor<float>>,               \
+      ops::ActivationKernel<paddle::platform::CPUDeviceContext,           \
+                            ops::functor<double>>);                       \
+  REGISTER_OP_CPU_KERNEL(                                                 \
+      act_type##_grad,                                                    \
+      ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,       \
+                                ops::grad_functor<float>>,                \
+      ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,       \
                                 ops::grad_functor<double>>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL);
diff --git a/paddle/operators/activation_op.cu b/paddle/operators/activation_op.cu
index 97737857ab..856d3fc35d 100644
--- a/paddle/operators/activation_op.cu
+++ b/paddle/operators/activation_op.cu
@@ -17,16 +17,17 @@
 
 namespace ops = paddle::operators;
 
-#define REGISTER_ACTIVATION_GPU_KERNEL(act_type, functor, grad_functor)       \
-  REGISTER_OP_GPU_KERNEL(                                                     \
-      act_type,                                                               \
-      ops::ActivationKernel<paddle::platform::GPUPlace, ops::functor<float>>, \
-      ops::ActivationKernel<paddle::platform::GPUPlace,                       \
-                            ops::functor<double>>);                           \
-  REGISTER_OP_GPU_KERNEL(                                                     \
-      act_type##_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace,  \
-                                                 ops::grad_functor<float>>,   \
-      ops::ActivationGradKernel<paddle::platform::GPUPlace,                   \
+#define REGISTER_ACTIVATION_CUDA_KERNEL(act_type, functor, grad_functor)   \
+  REGISTER_OP_CUDA_KERNEL(                                                 \
+      act_type, ops::ActivationKernel<paddle::platform::CUDADeviceContext, \
+                                      ops::functor<float>>,                \
+      ops::ActivationKernel<paddle::platform::CUDADeviceContext,           \
+                            ops::functor<double>>);                        \
+  REGISTER_OP_CUDA_KERNEL(                                                 \
+      act_type##_grad,                                                     \
+      ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,       \
+                                ops::grad_functor<float>>,                 \
+      ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,       \
                                 ops::grad_functor<double>>);
 
-FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_GPU_KERNEL);
+FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL);
diff --git a/paddle/operators/activation_op.h b/paddle/operators/activation_op.h
index ac0e0a3b01..75eefca8b8 100644
--- a/paddle/operators/activation_op.h
+++ b/paddle/operators/activation_op.h
@@ -19,7 +19,7 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename Functor>
+template <typename DeviceContext, typename Functor>
 class ActivationKernel
     : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
  public:
@@ -32,18 +32,19 @@ class ActivationKernel
 
     auto x = framework::EigenVector<T>::Flatten(*X);
     auto y = framework::EigenVector<T>::Flatten(*Y);
-    auto place = context.GetEigenDevice<Place>();
+    auto* place =
+        context.template device_context<DeviceContext>().eigen_device();
     Functor functor;
 
     auto attrs = functor.GetAttrs();
     for (auto& attr : attrs) {
       *attr.second = context.Attr<float>(attr.first);
     }
-    functor(place, x, y);
+    functor(*place, x, y);
   }
 };
 
-template <typename Place, typename Functor>
+template <typename DeviceContext, typename Functor>
 class ActivationGradKernel
     : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
  public:
@@ -59,13 +60,14 @@ class ActivationGradKernel
     auto x = framework::EigenVector<T>::Flatten(*X);
     auto y = framework::EigenVector<T>::Flatten(*Y);
     auto dx = framework::EigenVector<T>::Flatten(*dX);
-    auto place = context.GetEigenDevice<Place>();
+    auto* place =
+        context.template device_context<DeviceContext>().eigen_device();
     Functor functor;
     auto attrs = functor.GetAttrs();
     for (auto& attr : attrs) {
       *attr.second = context.Attr<float>(attr.first);
     }
-    functor(place, x, y, dy, dx);
+    functor(*place, x, y, dy, dx);
   }
 };
 
diff --git a/paddle/operators/adadelta_op.cc b/paddle/operators/adadelta_op.cc
index 29434a0ee2..507811e7b5 100644
--- a/paddle/operators/adadelta_op.cc
+++ b/paddle/operators/adadelta_op.cc
@@ -109,5 +109,5 @@ $$
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(adadelta, ops::AdadeltaOp, ops::AdadeltaOpMaker);
 REGISTER_OP_CPU_KERNEL(
-    adadelta, ops::AdadeltaOpKernel<paddle::platform::CPUPlace, float>,
-    ops::AdadeltaOpKernel<paddle::platform::CPUPlace, double>);
+    adadelta, ops::AdadeltaOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::AdadeltaOpKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/adadelta_op.cu b/paddle/operators/adadelta_op.cu
index 9fb6185207..eee2d0a2f5 100644
--- a/paddle/operators/adadelta_op.cu
+++ b/paddle/operators/adadelta_op.cu
@@ -16,6 +16,6 @@
 #include "paddle/operators/adadelta_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    adadelta, ops::AdadeltaOpKernel<paddle::platform::GPUPlace, float>,
-    ops::AdadeltaOpKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    adadelta, ops::AdadeltaOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::AdadeltaOpKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/adadelta_op.h b/paddle/operators/adadelta_op.h
index a8c5f0c8aa..819d0845db 100644
--- a/paddle/operators/adadelta_op.h
+++ b/paddle/operators/adadelta_op.h
@@ -19,7 +19,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class AdadeltaOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -51,7 +51,7 @@ class AdadeltaOpKernel : public framework::OpKernel<T> {
         framework::EigenVector<T>::Flatten(*avg_squared_grad_out_tensor);
     auto avg_squared_update_out =
         framework::EigenVector<T>::Flatten(*avg_squared_update_out_tensor);
-    auto place = ctx.GetEigenDevice<Place>();
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
 
     avg_squared_grad_out.device(place) =
         rho * avg_squared_grad + (1 - rho) * grad.square();
diff --git a/paddle/operators/adagrad_op.cc b/paddle/operators/adagrad_op.cc
index d19602244b..5d00716316 100644
--- a/paddle/operators/adagrad_op.cc
+++ b/paddle/operators/adagrad_op.cc
@@ -100,8 +100,8 @@ size_t FindPos(const std::vector<int64_t>& rows, int64_t value) {
 }  // namespace
 
 template <typename T>
-struct SparseAdagradFunctor<platform::CPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SparseAdagradFunctor<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::SelectedRows& grad,
                   const framework::Tensor& learning_rate, T epsilon,
                   framework::Tensor* moment, framework::Tensor* param) {
@@ -120,7 +120,7 @@ struct SparseAdagradFunctor<platform::CPUPlace, T> {
             {static_cast<int64_t>(merge_rows.size()), grad_width}),
         context.GetPlace());
 
-    math::SetConstant<platform::CPUPlace, T> constant_functor;
+    math::SetConstant<platform::CPUDeviceContext, T> constant_functor;
     constant_functor(context, grad_merge->mutable_value(), 0.0);
 
     auto* grad_merge_data = grad_merge->mutable_value()->data<T>();
@@ -144,9 +144,9 @@ struct SparseAdagradFunctor<platform::CPUPlace, T> {
     auto gs =
         framework::EigenVector<T>::Flatten(*(grad_square->mutable_value()));
     auto gm = framework::EigenVector<T>::Flatten(grad_merge->value());
-    gs.device(*context.GetEigenDevice<platform::CPUPlace>()) = gm * gm;
+    gs.device(*context.eigen_device()) = gm * gm;
 
-    math::SelectedRowsAddToTensor<platform::CPUPlace, T> functor;
+    math::SelectedRowsAddToTensor<platform::CPUDeviceContext, T> functor;
     functor(context, *grad_square, moment);
 
     // 3. update parameter
@@ -164,13 +164,13 @@ struct SparseAdagradFunctor<platform::CPUPlace, T> {
   }
 };
 
-template struct SparseAdagradFunctor<platform::CPUPlace, float>;
-template struct SparseAdagradFunctor<platform::CPUPlace, double>;
+template struct SparseAdagradFunctor<platform::CPUDeviceContext, float>;
+template struct SparseAdagradFunctor<platform::CPUDeviceContext, double>;
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(adagrad, ops::AdagradOp, ops::AdagradOpMaker);
 REGISTER_OP_CPU_KERNEL(
-    adagrad, ops::AdagradOpKernel<paddle::platform::CPUPlace, float>,
-    ops::AdagradOpKernel<paddle::platform::CPUPlace, double>);
+    adagrad, ops::AdagradOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::AdagradOpKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/adagrad_op.cu b/paddle/operators/adagrad_op.cu
index 1c870214b2..585b2d9289 100644
--- a/paddle/operators/adagrad_op.cu
+++ b/paddle/operators/adagrad_op.cu
@@ -72,8 +72,8 @@ __global__ void SparseAdagradFunctorKernel(const T* grad, const int64_t* rows,
 }  // namespace
 
 template <typename T>
-struct SparseAdagradFunctor<platform::GPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SparseAdagradFunctor<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::SelectedRows& grad,
                   const framework::Tensor& learning_rate, T epsilon,
                   framework::Tensor* moment, framework::Tensor* param) {
@@ -92,7 +92,7 @@ struct SparseAdagradFunctor<platform::GPUPlace, T> {
             {static_cast<int64_t>(merge_rows.size()), grad_width}),
         context.GetPlace());
 
-    math::SetConstant<platform::GPUPlace, T> constant_functor;
+    math::SetConstant<platform::CUDADeviceContext, T> constant_functor;
     constant_functor(context, grad_merge->mutable_value(), 0.0);
 
     auto* grad_merge_data = grad_merge->mutable_value()->data<T>();
@@ -119,9 +119,9 @@ struct SparseAdagradFunctor<platform::GPUPlace, T> {
     auto gs =
         framework::EigenVector<T>::Flatten(*(grad_square->mutable_value()));
     auto gm = framework::EigenVector<T>::Flatten(grad_merge->value());
-    gs.device(*context.GetEigenDevice<platform::GPUPlace>()) = gm * gm;
+    gs.device(*context.eigen_device()) = gm * gm;
 
-    math::SelectedRowsAddToTensor<platform::GPUPlace, T> functor;
+    math::SelectedRowsAddToTensor<platform::CUDADeviceContext, T> functor;
     functor(context, *grad_square, moment);
 
     // 3. update parameter
@@ -139,13 +139,13 @@ struct SparseAdagradFunctor<platform::GPUPlace, T> {
   }
 };
 
-template struct SparseAdagradFunctor<platform::GPUPlace, float>;
-template struct SparseAdagradFunctor<platform::GPUPlace, double>;
+template struct SparseAdagradFunctor<platform::CUDADeviceContext, float>;
+template struct SparseAdagradFunctor<platform::CUDADeviceContext, double>;
 
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    adagrad, ops::AdagradOpKernel<paddle::platform::GPUPlace, float>,
-    ops::AdagradOpKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    adagrad, ops::AdagradOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::AdagradOpKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/adagrad_op.h b/paddle/operators/adagrad_op.h
index 4d4a6434c7..0d77dbcbac 100644
--- a/paddle/operators/adagrad_op.h
+++ b/paddle/operators/adagrad_op.h
@@ -19,15 +19,15 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct SparseAdagradFunctor {
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::SelectedRows& grad,
                   const framework::Tensor& learning_rate, T epsilon,
                   framework::Tensor* moment, framework::Tensor* param);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class AdagradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -52,11 +52,11 @@ class AdagradOpKernel : public framework::OpKernel<T> {
 
       auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
       auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
-      auto place = ctx.GetEigenDevice<Place>();
+      auto* place = ctx.template device_context<DeviceContext>().eigen_device();
 
-      moment_out.device(place) = moment + grad * grad;
+      moment_out.device(*place) = moment + grad * grad;
       Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
-      param_out.device(place) =
+      param_out.device(*place) =
           param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
     } else if (grad_var->IsType<framework::SelectedRows>()) {
       auto* param_tensor = ctx.Input<framework::Tensor>("Param");
@@ -65,8 +65,9 @@ class AdagradOpKernel : public framework::OpKernel<T> {
       auto* moment_tensor = ctx.Input<framework::Tensor>("Moment");
       PADDLE_ENFORCE_EQ(moment_tensor, moment_out_tensor);
 
-      SparseAdagradFunctor<Place, T> functor;
-      functor(ctx.device_context(), *ctx.Input<framework::SelectedRows>("Grad"),
+      SparseAdagradFunctor<DeviceContext, T> functor;
+      functor(ctx.template device_context<DeviceContext>(),
+              *ctx.Input<framework::SelectedRows>("Grad"),
               *ctx.Input<framework::Tensor>("LearningRate"), epsilon,
               moment_out_tensor, param_out_tensor);
     } else {
diff --git a/paddle/operators/adam_op.cc b/paddle/operators/adam_op.cc
index a268d05484..cf6ef6dd53 100644
--- a/paddle/operators/adam_op.cc
+++ b/paddle/operators/adam_op.cc
@@ -128,6 +128,6 @@ $$
 
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(adam, ops::AdamOp, ops::AdamOpMaker);
-REGISTER_OP_CPU_KERNEL(adam,
-                       ops::AdamOpKernel<paddle::platform::CPUPlace, float>,
-                       ops::AdamOpKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(
+    adam, ops::AdamOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::AdamOpKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/adam_op.cu b/paddle/operators/adam_op.cu
index 6e34f7818c..c135b37378 100644
--- a/paddle/operators/adam_op.cu
+++ b/paddle/operators/adam_op.cu
@@ -16,6 +16,6 @@
 #include "paddle/operators/adam_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(adam,
-                       ops::AdamOpKernel<paddle::platform::GPUPlace, float>,
-                       ops::AdamOpKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    adam, ops::AdamOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::AdamOpKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/adam_op.h b/paddle/operators/adam_op.h
index 7f7fa1da1c..45157842a6 100644
--- a/paddle/operators/adam_op.h
+++ b/paddle/operators/adam_op.h
@@ -19,7 +19,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class AdamOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -52,17 +52,17 @@ class AdamOpKernel : public framework::OpKernel<T> {
     auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
     auto moment1_out = framework::EigenVector<T>::Flatten(*moment1_out_tensor);
     auto moment2_out = framework::EigenVector<T>::Flatten(*moment2_out_tensor);
-    auto place = ctx.GetEigenDevice<Place>();
+    auto* place = ctx.template device_context<DeviceContext>().eigen_device();
 
-    moment1_out.device(place) = beta1 * moment1 + (1 - beta1) * grad;
-    moment2_out.device(place) = beta2 * moment2 + (1 - beta2) * grad.square();
+    moment1_out.device(*place) = beta1 * moment1 + (1 - beta1) * grad;
+    moment2_out.device(*place) = beta2 * moment2 + (1 - beta2) * grad.square();
 
     // All of these are tensors of 1 element
     auto lr_t = lr * (1 - beta2_pow).sqrt() / (1 - beta1_pow);
     // Eigen does not support automatic broadcast
     // Get dimensions of moment vector to broadcast lr_t
     Eigen::DSizes<int, 1> m_dsize(moment1_out_tensor->numel());
-    param_out.device(place) =
+    param_out.device(*place) =
         param -
         lr_t.broadcast(m_dsize) *
             (moment1_out / (moment2_out.sqrt() + epsilon));
diff --git a/paddle/operators/adamax_op.cc b/paddle/operators/adamax_op.cc
index 9e7576c961..49ce497bb7 100644
--- a/paddle/operators/adamax_op.cc
+++ b/paddle/operators/adamax_op.cc
@@ -127,6 +127,6 @@ division by 0 error.
 
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(adamax, ops::AdamaxOp, ops::AdamaxOpMaker);
-REGISTER_OP_CPU_KERNEL(adamax,
-                       ops::AdamaxOpKernel<paddle::platform::CPUPlace, float>,
-                       ops::AdamaxOpKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(
+    adamax, ops::AdamaxOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::AdamaxOpKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/adamax_op.cu b/paddle/operators/adamax_op.cu
index 057ef39025..2d143905c4 100644
--- a/paddle/operators/adamax_op.cu
+++ b/paddle/operators/adamax_op.cu
@@ -16,6 +16,6 @@
 #include "paddle/operators/adamax_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(adamax,
-                       ops::AdamaxOpKernel<paddle::platform::GPUPlace, float>,
-                       ops::AdamaxOpKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    adamax, ops::AdamaxOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::AdamaxOpKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/adamax_op.h b/paddle/operators/adamax_op.h
index bf36ed7860..172c179c5f 100644
--- a/paddle/operators/adamax_op.h
+++ b/paddle/operators/adamax_op.h
@@ -19,7 +19,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class AdamaxOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -51,14 +51,14 @@ class AdamaxOpKernel : public framework::OpKernel<T> {
     auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
     auto inf_norm_out =
         framework::EigenVector<T>::Flatten(*inf_norm_out_tensor);
-    auto place = ctx.GetEigenDevice<Place>();
+    auto* place = ctx.template device_context<DeviceContext>().eigen_device();
 
-    moment_out.device(place) = beta1 * moment + (1 - beta1) * grad;
-    inf_norm_out.device(place) =
+    moment_out.device(*place) = beta1 * moment + (1 - beta1) * grad;
+    inf_norm_out.device(*place) =
         grad.abs().cwiseMax((beta2 * inf_norm) + epsilon);
     auto lr_t = lr / (1 - beta1_pow);
     Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
-    param_out.device(place) =
+    param_out.device(*place) =
         param - lr_t.broadcast(m_dsize) * (moment_out / inf_norm_out);
   }
 };
diff --git a/paddle/operators/auc_op.h b/paddle/operators/auc_op.h
index e5ac57b038..b80509e2a9 100644
--- a/paddle/operators/auc_op.h
+++ b/paddle/operators/auc_op.h
@@ -25,7 +25,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class AucKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
diff --git a/paddle/operators/batch_norm_op.cc b/paddle/operators/batch_norm_op.cc
index ac97bd83ab..94a972b7ab 100644
--- a/paddle/operators/batch_norm_op.cc
+++ b/paddle/operators/batch_norm_op.cc
@@ -135,7 +135,8 @@ The required data format for this layer is one of the following:
 };
 
 template <typename T>
-class BatchNormKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
+class BatchNormKernel<platform::CPUDeviceContext, T>
+    : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     const float epsilon = ctx.Attr<float>("epsilon");
@@ -318,12 +319,12 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
       PADDLE_THROW("can't find Y@GRAD");
     }
     return framework::OpKernelType(framework::ToDataType(t->type()),
-                                   ctx.device_context());
+                                   ctx.GetPlace());
   }
 };
 
 template <typename T>
-class BatchNormGradKernel<platform::CPUPlace, T>
+class BatchNormGradKernel<platform::CPUDeviceContext, T>
     : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
@@ -436,8 +437,9 @@ class BatchNormGradKernel<platform::CPUPlace, T>
 namespace ops = paddle::operators;
 REGISTER_OP(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker,
             batch_norm_grad, ops::BatchNormGradOp);
-REGISTER_OP_CPU_KERNEL(batch_norm,
-                       ops::BatchNormKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    batch_norm,
+    ops::BatchNormKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     batch_norm_grad,
-    ops::BatchNormGradKernel<paddle::platform::CPUPlace, float>);
+    ops::BatchNormGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/batch_norm_op.cu.cc b/paddle/operators/batch_norm_op.cu.cc
index 7b2f318700..c7adc3d80e 100644
--- a/paddle/operators/batch_norm_op.cu.cc
+++ b/paddle/operators/batch_norm_op.cu.cc
@@ -47,7 +47,8 @@ void ExtractNCWHD(const framework::DDim &dims,
 }
 
 template <typename T>
-class BatchNormKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
+class BatchNormKernel<platform::CUDADeviceContext, T>
+    : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
@@ -121,11 +122,12 @@ class BatchNormKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
     saved_mean->mutable_data<T>(ctx.GetPlace());
     saved_variance->mutable_data<T>(ctx.GetPlace());
 
-    math::SetConstant<platform::GPUPlace, T> functor;
-    functor(ctx.device_context(), saved_mean, 0);
-    functor(ctx.device_context(), saved_variance, 0);
+    auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    math::SetConstant<platform::CUDADeviceContext, T> functor;
+    functor(dev_ctx, saved_mean, 0);
+    functor(dev_ctx, saved_variance, 0);
 
-    auto handle = ctx.cuda_device_context().cudnn_handle();
+    auto handle = dev_ctx.cudnn_handle();
 
     // Now, depending on whether we are running test or not, we have two paths.
     if (is_test) {
@@ -171,7 +173,7 @@ class BatchNormKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
 };
 
 template <typename T>
-class BatchNormGradKernel<platform::GPUPlace, T>
+class BatchNormGradKernel<platform::CUDADeviceContext, T>
     : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &ctx) const override {
@@ -244,11 +246,12 @@ class BatchNormGradKernel<platform::GPUPlace, T>
     const void *saved_mean_data = saved_mean->template data<T>();
     const void *saved_var_data = saved_var->template data<T>();
 
+    auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
     CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationBackward(
-        ctx.cuda_device_context().cudnn_handle(), mode_,
-        CudnnDataType<T>::kOne(), CudnnDataType<T>::kZero(),
-        CudnnDataType<T>::kOne(), CudnnDataType<T>::kZero(), data_desc_,
-        x->template data<T>(), data_desc_, d_y->template data<T>(), data_desc_,
+        dev_ctx.cudnn_handle(), mode_, CudnnDataType<T>::kOne(),
+        CudnnDataType<T>::kZero(), CudnnDataType<T>::kOne(),
+        CudnnDataType<T>::kZero(), data_desc_, x->template data<T>(),
+        data_desc_, d_y->template data<T>(), data_desc_,
         d_x->template mutable_data<T>(ctx.GetPlace()), bn_param_desc_,
         scale->template data<T>(),
         d_scale->template mutable_data<T>(ctx.GetPlace()),
@@ -266,8 +269,9 @@ class BatchNormGradKernel<platform::GPUPlace, T>
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(batch_norm,
-                       ops::BatchNormKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    batch_norm,
+    ops::BatchNormKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     batch_norm_grad,
-    ops::BatchNormGradKernel<paddle::platform::GPUPlace, float>);
+    ops::BatchNormGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/batch_norm_op.h b/paddle/operators/batch_norm_op.h
index 4e80134a1a..8d99b68647 100644
--- a/paddle/operators/batch_norm_op.h
+++ b/paddle/operators/batch_norm_op.h
@@ -34,13 +34,13 @@ inline TensorFormat StringToTensorFormat(const std::string& str) {
   }
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class BatchNormKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override;
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class BatchNormGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override;
diff --git a/paddle/operators/bilinear_tensor_product_op.cc b/paddle/operators/bilinear_tensor_product_op.cc
index c88b2c9beb..217fd52366 100644
--- a/paddle/operators/bilinear_tensor_product_op.cc
+++ b/paddle/operators/bilinear_tensor_product_op.cc
@@ -159,9 +159,12 @@ REGISTER_OP(bilinear_tensor_product, ops::BilinearTensorProductOp,
             ops::BilinearTensorProductOpGrad);
 REGISTER_OP_CPU_KERNEL(
     bilinear_tensor_product,
-    ops::BilinearTensorProductKernel<paddle::platform::CPUPlace, float>,
-    ops::BilinearTensorProductKernel<paddle::platform::CPUPlace, double>);
+    ops::BilinearTensorProductKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::BilinearTensorProductKernel<paddle::platform::CPUDeviceContext,
+                                     double>);
 REGISTER_OP_CPU_KERNEL(
     bilinear_tensor_product_grad,
-    ops::BilinearTensorProductGradKernel<paddle::platform::CPUPlace, float>,
-    ops::BilinearTensorProductGradKernel<paddle::platform::CPUPlace, double>);
+    ops::BilinearTensorProductGradKernel<paddle::platform::CPUDeviceContext,
+                                         float>,
+    ops::BilinearTensorProductGradKernel<paddle::platform::CPUDeviceContext,
+                                         double>);
diff --git a/paddle/operators/bilinear_tensor_product_op.cu b/paddle/operators/bilinear_tensor_product_op.cu
index 858d2668d0..0f48010716 100644
--- a/paddle/operators/bilinear_tensor_product_op.cu
+++ b/paddle/operators/bilinear_tensor_product_op.cu
@@ -16,11 +16,15 @@ limitations under the License. */
 #include "paddle/operators/bilinear_tensor_product_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     bilinear_tensor_product,
-    ops::BilinearTensorProductKernel<paddle::platform::GPUPlace, float>,
-    ops::BilinearTensorProductKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
+    ops::BilinearTensorProductKernel<paddle::platform::CUDADeviceContext,
+                                     float>,
+    ops::BilinearTensorProductKernel<paddle::platform::CUDADeviceContext,
+                                     double>);
+REGISTER_OP_CUDA_KERNEL(
     bilinear_tensor_product_grad,
-    ops::BilinearTensorProductGradKernel<paddle::platform::GPUPlace, float>,
-    ops::BilinearTensorProductGradKernel<paddle::platform::GPUPlace, double>);
+    ops::BilinearTensorProductGradKernel<paddle::platform::CUDADeviceContext,
+                                         float>,
+    ops::BilinearTensorProductGradKernel<paddle::platform::CUDADeviceContext,
+                                         double>);
diff --git a/paddle/operators/bilinear_tensor_product_op.h b/paddle/operators/bilinear_tensor_product_op.h
index 1113a4c6f3..ba9a2c5ce3 100644
--- a/paddle/operators/bilinear_tensor_product_op.h
+++ b/paddle/operators/bilinear_tensor_product_op.h
@@ -27,7 +27,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class BilinearTensorProductKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -46,7 +46,8 @@ class BilinearTensorProductKernel : public framework::OpKernel<T> {
     int out_dim = weight_dims[0];
     auto x_dim = weight_dims[1];
     auto y_dim = weight_dims[2];
-    auto place = ctx.GetEigenDevice<Place>();
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
 
     // Create the intermediate variable to caculate the result of
     // Input(X) multiplied by Input(Weight_i), the formula is:
@@ -60,9 +61,9 @@ class BilinearTensorProductKernel : public framework::OpKernel<T> {
       auto output_col_vec = output_mat.chip(i, 1);
       Tensor weight_mat =
           weight->Slice(i, i + 1).Resize(framework::make_ddim({x_dim, y_dim}));
-      math::gemm<Place, T>(ctx.device_context(), CblasNoTrans, CblasNoTrans,
-                           batch_size, y_dim, x_dim, 1, x->data<T>(),
-                           weight_mat.data<T>(), 0, left_mul.data<T>());
+      math::gemm<DeviceContext, T>(dev_ctx, CblasNoTrans, CblasNoTrans,
+                                   batch_size, y_dim, x_dim, 1, x->data<T>(),
+                                   weight_mat.data<T>(), 0, left_mul.data<T>());
       output_col_vec.device(place) =
           (left_mul_mat * y_mat).sum(Eigen::DSizes<int, 1>(1));
     }
@@ -74,7 +75,7 @@ class BilinearTensorProductKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -96,8 +97,8 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
     auto x_mat = EigenMatrix<T>::From(*x);
     auto y_mat = EigenMatrix<T>::From(*y);
     auto d_out_mat = EigenMatrix<T>::From(*d_out);
-    auto place = ctx.GetEigenDevice<Place>();
-
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
     // Create the intermediate variable to caculate the Output(Y@Grad).
     Tensor x_scale;
     x_scale.mutable_data<T>(framework::make_ddim({batch_size, x_dim}),
@@ -110,18 +111,18 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
                             ctx.GetPlace());
     auto y_scale_mat = EigenMatrix<T>::From(y_scale);
 
-    math::SetConstant<Place, T> set_zero;
+    math::SetConstant<DeviceContext, T> set_zero;
 
     // Set Output(X@Grad) be zero.
     if (d_x) {
       d_x->mutable_data<T>(ctx.GetPlace());
-      set_zero(ctx.device_context(), d_x, static_cast<T>(0));
+      set_zero(dev_ctx, d_x, static_cast<T>(0));
     }
 
     // Set Output(Y@Grad) be zero.
     if (d_y) {
       d_y->mutable_data<T>(ctx.GetPlace());
-      set_zero(ctx.device_context(), d_y, static_cast<T>(0));
+      set_zero(dev_ctx, d_y, static_cast<T>(0));
     }
 
     // Caculate the Output(X@Grad) and Output(Y@Grad).
@@ -137,18 +138,18 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
               output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
                   .broadcast(bcast_for_x) *
               y_mat;
-          math::gemm<Place, T>(ctx.device_context(), CblasNoTrans, CblasTrans,
-                               batch_size, x_dim, y_dim, 1, y_scale.data<T>(),
-                               weight_i.data<T>(), 1, d_x->data<T>());
+          math::gemm<DeviceContext, T>(
+              dev_ctx, CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1,
+              y_scale.data<T>(), weight_i.data<T>(), 1, d_x->data<T>());
         }
         if (d_y) {
           x_scale_mat.device(place) =
               output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
                   .broadcast(bcast_for_y) *
               x_mat;
-          math::gemm<Place, T>(ctx.device_context(), CblasNoTrans, CblasNoTrans,
-                               batch_size, y_dim, x_dim, 1, x_scale.data<T>(),
-                               weight_i.data<T>(), 1, d_y->data<T>());
+          math::gemm<DeviceContext, T>(
+              dev_ctx, CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1,
+              x_scale.data<T>(), weight_i.data<T>(), 1, d_y->data<T>());
         }
       }
     }
@@ -165,9 +166,9 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
             output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
                 .broadcast(bcast_for_weight) *
             x_mat;
-        math::gemm<Place, T>(ctx.device_context(), CblasTrans, CblasNoTrans,
-                             x_dim, y_dim, batch_size, 1, x_scale.data<T>(),
-                             y->data<T>(), 0, d_weight_i.data<T>());
+        math::gemm<DeviceContext, T>(dev_ctx, CblasTrans, CblasNoTrans, x_dim,
+                                     y_dim, batch_size, 1, x_scale.data<T>(),
+                                     y->data<T>(), 0, d_weight_i.data<T>());
       }
     }
 
diff --git a/paddle/operators/cast_op.cc b/paddle/operators/cast_op.cc
index 3082a53ccf..42bff69a1e 100644
--- a/paddle/operators/cast_op.cc
+++ b/paddle/operators/cast_op.cc
@@ -68,7 +68,7 @@ class CastOpGradMaker : public framework::SingleGradOpDescMaker {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-using CPU = paddle::platform::CPUPlace;
+using CPU = paddle::platform::CPUDeviceContext;
 REGISTER_OP_WITH_KERNEL(cast, ops::CastOpGradMaker, ops::CastOpInferShape,
                         ops::CastOpProtoMaker);
 REGISTER_OP_CPU_KERNEL(cast, ops::CastOpKernel<CPU, float>,
diff --git a/paddle/operators/cast_op.cu b/paddle/operators/cast_op.cu
index fb75ddbabf..4681deaa62 100644
--- a/paddle/operators/cast_op.cu
+++ b/paddle/operators/cast_op.cu
@@ -16,7 +16,7 @@
 
 template <typename T>
 using CastOpKernel =
-    paddle::operators::CastOpKernel<paddle::platform::GPUPlace, T>;
+    paddle::operators::CastOpKernel<paddle::platform::CUDADeviceContext, T>;
 
-REGISTER_OP_GPU_KERNEL(cast, CastOpKernel<float>, CastOpKernel<double>,
-                       CastOpKernel<int>, CastOpKernel<int64_t>);
+REGISTER_OP_CUDA_KERNEL(cast, CastOpKernel<float>, CastOpKernel<double>,
+                        CastOpKernel<int>, CastOpKernel<int64_t>);
diff --git a/paddle/operators/cast_op.h b/paddle/operators/cast_op.h
index 850dc8e349..a6773f13a8 100644
--- a/paddle/operators/cast_op.h
+++ b/paddle/operators/cast_op.h
@@ -27,13 +27,13 @@ struct CastOpTransformFunctor {
   HOSTDEVICE OutT operator()(InT in) const { return static_cast<OutT>(in); }
 };
 
-template <typename Place, typename InT>
+template <typename DeviceContext, typename InT>
 struct CastOpFunctor {
   const framework::Tensor* in_;
   framework::Tensor* out_;
-  const platform::DeviceContext& ctx_;
+  const DeviceContext& ctx_;
   CastOpFunctor(const framework::Tensor* in, framework::Tensor* out,
-                const platform::DeviceContext& ctx)
+                const DeviceContext& ctx)
       : in_(in), out_(out), ctx_(ctx) {}
 
   template <typename OutT>
@@ -42,13 +42,13 @@ struct CastOpFunctor {
     auto numel = in_->numel();
     auto* in_end = in_begin + numel;
     auto* out_begin = out_->mutable_data<OutT>(ctx_.GetPlace());
-    platform::Transform<Place> trans;
+    platform::Transform<DeviceContext> trans;
     trans(ctx_, in_begin, in_end, out_begin,
           CastOpTransformFunctor<InT, OutT>());
   }
 };
 
-template <typename Place, typename InT>
+template <typename DeviceContext, typename InT>
 class CastOpKernel : public framework::OpKernel<InT> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -56,7 +56,8 @@ class CastOpKernel : public framework::OpKernel<InT> {
     auto* out = context.Output<framework::Tensor>("Out");
     framework::VisitDataType(
         static_cast<framework::DataType>(context.Attr<int>("out_dtype")),
-        CastOpFunctor<Place, InT>(in, out, context.device_context()));
+        CastOpFunctor<DeviceContext, InT>(
+            in, out, context.template device_context<DeviceContext>()));
   }
 };
 
diff --git a/paddle/operators/chunk_eval_op.h b/paddle/operators/chunk_eval_op.h
index dd88f2553b..9cd758a825 100644
--- a/paddle/operators/chunk_eval_op.h
+++ b/paddle/operators/chunk_eval_op.h
@@ -23,7 +23,7 @@ namespace operators {
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ChunkEvalKernel : public framework::OpKernel<T> {
  public:
   struct Segment {
diff --git a/paddle/operators/clip_by_norm_op.cc b/paddle/operators/clip_by_norm_op.cc
index f73d55bbe3..0b7975a63f 100644
--- a/paddle/operators/clip_by_norm_op.cc
+++ b/paddle/operators/clip_by_norm_op.cc
@@ -71,4 +71,5 @@ namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(clip_by_norm, ops::ClipByNormOp,
                              ops::ClipByNormOpMaker);
 REGISTER_OP_CPU_KERNEL(
-    clip_by_norm, ops::ClipByNormKernel<paddle::platform::CPUPlace, float>);
+    clip_by_norm,
+    ops::ClipByNormKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/clip_by_norm_op.cu b/paddle/operators/clip_by_norm_op.cu
index 2593a24ebb..acd7543823 100644
--- a/paddle/operators/clip_by_norm_op.cu
+++ b/paddle/operators/clip_by_norm_op.cu
@@ -15,5 +15,6 @@
 #include "paddle/operators/clip_by_norm_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    clip_by_norm, ops::ClipByNormKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    clip_by_norm,
+    ops::ClipByNormKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/clip_by_norm_op.h b/paddle/operators/clip_by_norm_op.h
index b26476cae9..d8db1566b0 100644
--- a/paddle/operators/clip_by_norm_op.h
+++ b/paddle/operators/clip_by_norm_op.h
@@ -26,7 +26,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ClipByNormKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -38,7 +38,8 @@ class ClipByNormKernel : public framework::OpKernel<T> {
     auto x = EigenVector<T>::Flatten(*input);
     auto out = EigenVector<T>::Flatten(*output);
     auto x_norm = x.square().sum().sqrt();
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     auto temp = (x_norm <= max_norm).template cast<T>().eval();
     auto scaling = temp + (static_cast<T>(1) - temp) * max_norm / x_norm;
diff --git a/paddle/operators/clip_op.cc b/paddle/operators/clip_op.cc
index 4ddf24dea3..6092212de4 100644
--- a/paddle/operators/clip_op.cc
+++ b/paddle/operators/clip_op.cc
@@ -83,7 +83,7 @@ class ClipOpGrad : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(clip, ops::ClipOp, ops::ClipOpMaker<float>, clip_grad,
             ops::ClipOpGrad);
-REGISTER_OP_CPU_KERNEL(clip,
-                       ops::ClipKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(clip_grad,
-                       ops::ClipGradKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    clip, ops::ClipKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    clip_grad, ops::ClipGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/clip_op.cu b/paddle/operators/clip_op.cu
index ca9701298f..bb7dcc671a 100644
--- a/paddle/operators/clip_op.cu
+++ b/paddle/operators/clip_op.cu
@@ -15,7 +15,7 @@
 #include "paddle/operators/clip_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(clip,
-                       ops::ClipKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(clip_grad,
-                       ops::ClipGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    clip, ops::ClipKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    clip_grad, ops::ClipGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/clip_op.h b/paddle/operators/clip_op.h
index ac702e9935..0c40797410 100644
--- a/paddle/operators/clip_op.h
+++ b/paddle/operators/clip_op.h
@@ -55,7 +55,7 @@ class ClipGradFunctor {
   T max_;
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ClipKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -66,13 +66,13 @@ class ClipKernel : public framework::OpKernel<T> {
     T* out_data = out->mutable_data<T>(context.GetPlace());
     const T* x_data = x->data<T>();
     int64_t numel = x->numel();
-    Transform<Place> trans;
-    trans(context.device_context(), x_data, x_data + numel, out_data,
-          ClipFunctor<T>(min, max));
+    Transform<DeviceContext> trans;
+    trans(context.template device_context<DeviceContext>(), x_data,
+          x_data + numel, out_data, ClipFunctor<T>(min, max));
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ClipGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -86,9 +86,9 @@ class ClipGradKernel : public framework::OpKernel<T> {
       auto* d_x_data = d_x->mutable_data<T>(context.GetPlace());
       const T* d_out_data = d_out->data<T>();
       const T* x_data = x->data<T>();
-      Transform<Place> trans;
-      trans(context.device_context(), d_out_data, d_out_data + numel, x_data,
-            d_x_data, ClipGradFunctor<T>(min, max));
+      Transform<DeviceContext> trans;
+      trans(context.template device_context<DeviceContext>(), d_out_data,
+            d_out_data + numel, x_data, d_x_data, ClipGradFunctor<T>(min, max));
     }
   }
 };
diff --git a/paddle/operators/compare_op.cu b/paddle/operators/compare_op.cu
index 6ac8c124b9..596a878bcf 100644
--- a/paddle/operators/compare_op.cu
+++ b/paddle/operators/compare_op.cu
@@ -14,10 +14,10 @@
 
 #include "paddle/operators/compare_op.h"
 
-REGISTER_LOGICAL_KERNEL(less_than, GPU, paddle::operators::LessThanFunctor);
-REGISTER_LOGICAL_KERNEL(less_equal, GPU, paddle::operators::LessEqualFunctor);
-REGISTER_LOGICAL_KERNEL(greater_than, GPU,
+REGISTER_LOGICAL_KERNEL(less_than, CUDA, paddle::operators::LessThanFunctor);
+REGISTER_LOGICAL_KERNEL(less_equal, CUDA, paddle::operators::LessEqualFunctor);
+REGISTER_LOGICAL_KERNEL(greater_than, CUDA,
                         paddle::operators::GreaterThanFunctor);
-REGISTER_LOGICAL_KERNEL(greater_equal, GPU,
+REGISTER_LOGICAL_KERNEL(greater_equal, CUDA,
                         paddle::operators::GreaterEqualFunctor);
-REGISTER_LOGICAL_KERNEL(equal, GPU, paddle::operators::EqualFunctor);
+REGISTER_LOGICAL_KERNEL(equal, CUDA, paddle::operators::EqualFunctor);
diff --git a/paddle/operators/compare_op.h b/paddle/operators/compare_op.h
index afdf3ab3e0..a56536e155 100644
--- a/paddle/operators/compare_op.h
+++ b/paddle/operators/compare_op.h
@@ -59,7 +59,7 @@ struct EqualFunctor {
   }
 };
 
-template <typename Place, typename Functor>
+template <typename DeviceContext, typename Functor>
 class CompareOpKernel
     : public framework::OpKernel<typename Functor::ELEM_TYPE> {
  public:
@@ -69,24 +69,23 @@ class CompareOpKernel
     auto* y = context.Input<framework::Tensor>("Y");
     auto* out = context.Output<framework::Tensor>("Out");
     Functor binary_func;
-    platform::Transform<Place> trans;
-    trans(context.device_context(), x->data<T>(), x->data<T>() + x->numel(),
-          y->data<T>(), out->mutable_data<bool>(context.GetPlace()),
-          binary_func);
+    platform::Transform<DeviceContext> trans;
+    trans(context.template device_context<DeviceContext>(), x->data<T>(),
+          x->data<T>() + x->numel(), y->data<T>(),
+          out->mutable_data<bool>(context.GetPlace()), binary_func);
   }
 };
 
 }  // namespace operators
 }  // namespace paddle
 
-#define REGISTER_LOGICAL_KERNEL(op_type, dev, functor)                     \
-  REGISTER_OP_##dev##_KERNEL(                                              \
-      op_type,                                                             \
-      ::paddle::operators::CompareOpKernel<::paddle::platform::dev##Place, \
-                                           functor<int>>,                  \
-      ::paddle::operators::CompareOpKernel<::paddle::platform::dev##Place, \
-                                           functor<int64_t>>,              \
-      ::paddle::operators::CompareOpKernel<::paddle::platform::dev##Place, \
-                                           functor<float>>,                \
-      ::paddle::operators::CompareOpKernel<::paddle::platform::dev##Place, \
-                                           functor<double>>);
+#define REGISTER_LOGICAL_KERNEL(op_type, dev, functor)                    \
+  REGISTER_OP_##dev##_KERNEL(                                             \
+      op_type, ::paddle::operators::CompareOpKernel<                      \
+                   ::paddle::platform::dev##DeviceContext, functor<int>>, \
+      ::paddle::operators::CompareOpKernel<                               \
+          ::paddle::platform::dev##DeviceContext, functor<int64_t>>,      \
+      ::paddle::operators::CompareOpKernel<                               \
+          ::paddle::platform::dev##DeviceContext, functor<float>>,        \
+      ::paddle::operators::CompareOpKernel<                               \
+          ::paddle::platform::dev##DeviceContext, functor<double>>);
diff --git a/paddle/operators/concat_op.cu.cc b/paddle/operators/concat_op.cu.cc
index ede832ddcd..7b46452d3d 100644
--- a/paddle/operators/concat_op.cu.cc
+++ b/paddle/operators/concat_op.cu.cc
@@ -14,7 +14,8 @@ limitations under the License. */
 
 #include "paddle/operators/concat_op.h"
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(concat,
-                       ops::ConcatKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    concat_grad, ops::ConcatGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    concat, ops::ConcatKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    concat_grad,
+    ops::ConcatGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/concat_op.h b/paddle/operators/concat_op.h
index c113f19fb5..de4011585a 100644
--- a/paddle/operators/concat_op.h
+++ b/paddle/operators/concat_op.h
@@ -21,7 +21,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ConcatKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -43,7 +43,7 @@ class ConcatKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ConcatGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
diff --git a/paddle/operators/conv_cudnn_op.cc b/paddle/operators/conv_cudnn_op.cc
index 0dd8c13b2a..008bf01885 100644
--- a/paddle/operators/conv_cudnn_op.cc
+++ b/paddle/operators/conv_cudnn_op.cc
@@ -57,18 +57,20 @@ REGISTER_OP(conv2d_cudnn, ops::ConvOp, ops::CudnnConv2DOpMaker,
 REGISTER_OP(conv3d_cudnn, ops::ConvOp, ops::CudnnConv3DOpMaker,
             conv3d_cudnn_grad, ops::ConvOpGrad);
 
-REGISTER_OP_CPU_KERNEL(conv2d_cudnn,
-                       ops::GemmConvKernel<paddle::platform::CPUPlace, float>,
-                       ops::GemmConvKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(
+    conv2d_cudnn,
+    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     conv2d_cudnn_grad,
-    ops::GemmConvGradKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvGradKernel<paddle::platform::CPUPlace, double>);
+    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
 
-REGISTER_OP_CPU_KERNEL(conv3d_cudnn,
-                       ops::GemmConvKernel<paddle::platform::CPUPlace, float>,
-                       ops::GemmConvKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(
+    conv3d_cudnn,
+    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     conv3d_cudnn_grad,
-    ops::GemmConvGradKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvGradKernel<paddle::platform::CPUPlace, double>);
+    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/conv_cudnn_op.cu.cc b/paddle/operators/conv_cudnn_op.cu.cc
index bc265dcc4f..3da0a9001a 100644
--- a/paddle/operators/conv_cudnn_op.cu.cc
+++ b/paddle/operators/conv_cudnn_op.cu.cc
@@ -118,7 +118,8 @@ class CudnnConvOpKernel : public framework::OpKernel<T> {
     }
     // ------------------- cudnn conv algorithm ---------------------
     cudnnConvolutionFwdAlgo_t algo;
-    auto handle = ctx.cuda_device_context().cudnn_handle();
+    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto handle = dev_ctx.cudnn_handle();
 
     PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm(
         handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
@@ -238,7 +239,8 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
       workspace_size_limit = user_workspace_size * 1024 * 1024;
     }
 
-    auto handle = ctx.cuda_device_context().cudnn_handle();
+    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto handle = dev_ctx.cudnn_handle();
     if (input_grad) {
       PADDLE_ENFORCE(
           platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
@@ -313,16 +315,16 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_GPU_KERNEL(conv2d_cudnn,
-                       paddle::operators::CudnnConvOpKernel<float>,
-                       paddle::operators::CudnnConvOpKernel<double>);
-REGISTER_OP_GPU_KERNEL(conv2d_cudnn_grad,
-                       paddle::operators::CudnnConvGradOpKernel<float>,
-                       paddle::operators::CudnnConvGradOpKernel<double>);
-
-REGISTER_OP_GPU_KERNEL(conv3d_cudnn,
-                       paddle::operators::CudnnConvOpKernel<float>,
-                       paddle::operators::CudnnConvOpKernel<double>);
-REGISTER_OP_GPU_KERNEL(conv3d_cudnn_grad,
-                       paddle::operators::CudnnConvGradOpKernel<float>,
-                       paddle::operators::CudnnConvGradOpKernel<double>);
+REGISTER_OP_CUDA_KERNEL(conv2d_cudnn,
+                        paddle::operators::CudnnConvOpKernel<float>,
+                        paddle::operators::CudnnConvOpKernel<double>);
+REGISTER_OP_CUDA_KERNEL(conv2d_cudnn_grad,
+                        paddle::operators::CudnnConvGradOpKernel<float>,
+                        paddle::operators::CudnnConvGradOpKernel<double>);
+
+REGISTER_OP_CUDA_KERNEL(conv3d_cudnn,
+                        paddle::operators::CudnnConvOpKernel<float>,
+                        paddle::operators::CudnnConvOpKernel<double>);
+REGISTER_OP_CUDA_KERNEL(conv3d_cudnn_grad,
+                        paddle::operators::CudnnConvGradOpKernel<float>,
+                        paddle::operators::CudnnConvGradOpKernel<double>);
diff --git a/paddle/operators/conv_op.cc b/paddle/operators/conv_op.cc
index 462e6d9cbc..7ef805fd44 100644
--- a/paddle/operators/conv_op.cc
+++ b/paddle/operators/conv_op.cc
@@ -235,16 +235,18 @@ namespace ops = paddle::operators;
 REGISTER_OP(conv3d, ops::ConvOp, ops::Conv3DOpMaker, conv3d_grad,
             ops::ConvOpGrad);
 
-REGISTER_OP_CPU_KERNEL(conv2d,
-                       ops::GemmConvKernel<paddle::platform::CPUPlace, float>,
-                       ops::GemmConvKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
-    conv2d_grad, ops::GemmConvGradKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvGradKernel<paddle::platform::CPUPlace, double>);
+    conv2d, ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    conv2d_grad,
+    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
 
-REGISTER_OP_CPU_KERNEL(conv3d,
-                       ops::GemmConvKernel<paddle::platform::CPUPlace, float>,
-                       ops::GemmConvKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
-    conv3d_grad, ops::GemmConvGradKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvGradKernel<paddle::platform::CPUPlace, double>);
+    conv3d, ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    conv3d_grad,
+    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/conv_op.cu.cc b/paddle/operators/conv_op.cu.cc
index 546451234a..38615a8bef 100644
--- a/paddle/operators/conv_op.cu.cc
+++ b/paddle/operators/conv_op.cu.cc
@@ -16,16 +16,18 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(conv2d,
-                       ops::GemmConvKernel<paddle::platform::GPUPlace, float>,
-                       ops::GemmConvKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
-    conv2d_grad, ops::GemmConvGradKernel<paddle::platform::GPUPlace, float>,
-    ops::GemmConvGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    conv2d, ops::GemmConvKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GemmConvKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    conv2d_grad,
+    ops::GemmConvGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GemmConvGradKernel<paddle::platform::CUDADeviceContext, double>);
 
-REGISTER_OP_GPU_KERNEL(conv3d,
-                       ops::GemmConvKernel<paddle::platform::GPUPlace, float>,
-                       ops::GemmConvKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
-    conv3d_grad, ops::GemmConvGradKernel<paddle::platform::GPUPlace, float>,
-    ops::GemmConvGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    conv3d, ops::GemmConvKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GemmConvKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    conv3d_grad,
+    ops::GemmConvGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GemmConvGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/conv_op.h b/paddle/operators/conv_op.h
index 09bff0a68d..749258183b 100644
--- a/paddle/operators/conv_op.h
+++ b/paddle/operators/conv_op.h
@@ -72,7 +72,7 @@ class ConvOpGrad : public framework::OperatorWithKernel {
   void InferShape(framework::InferShapeContext* ctx) const override;
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class GemmConvKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -141,9 +141,10 @@ class GemmConvKernel : public framework::OpKernel<T> {
     int in_step = static_cast<int>(input->dims()[1]) / groups;
     int out_step = static_cast<int>(output->dims()[1]) / groups;
 
-    math::Vol2ColFunctor<Place, T> vol2col;
-    math::Im2ColFunctor<math::ColFormat::kCFO, Place, T> im2col;
+    math::Vol2ColFunctor<DeviceContext, T> vol2col;
+    math::Im2ColFunctor<math::ColFormat::kCFO, DeviceContext, T> im2col;
 
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     for (int i = 0; i < batch_size; i++) {
       Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
       Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
@@ -157,27 +158,26 @@ class GemmConvKernel : public framework::OpKernel<T> {
           col_matrix.Resize(col_matrix_shape);
         } else if (data_dim == 2U) {
           // im2col
-          im2col(context.device_context(), in_slice, dilations, strides,
+          im2col(dev_ctx, in_slice, dilations, strides,
                  std::vector<int>{paddings[0], paddings[1], paddings[0],
                                   paddings[1]},
                  &col);
         } else if (data_dim == 3U) {
           // vol2col
-          vol2col(context.device_context(), in_slice, dilations, strides,
-                  paddings, &col);
+          vol2col(dev_ctx, in_slice, dilations, strides, paddings, &col);
         }
 
         // gemm
         Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
         Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
-        math::matmul<Place, T>(context.device_context(), filter_slice, false,
-                               col_matrix, false, T(1.0), &out_slice, T(0.0));
+        math::matmul<DeviceContext, T>(dev_ctx, filter_slice, false, col_matrix,
+                                       false, T(1.0), &out_slice, T(0.0));
       }
     }
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class GemmConvGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -256,14 +256,15 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
       col_matrix.Resize(col_matrix_shape);
     }
 
-    math::SetConstant<Place, T> set_zero;
+    math::SetConstant<DeviceContext, T> set_zero;
+    auto& dev_ctx = context.template device_context<DeviceContext>();
 
     if (input_grad) {
       input_grad->mutable_data<T>(context.GetPlace());
-      set_zero(context.device_context(), input_grad, static_cast<T>(0));
+      set_zero(dev_ctx, input_grad, static_cast<T>(0));
 
-      math::Col2VolFunctor<Place, T> col2vol;
-      math::Col2ImFunctor<math::ColFormat::kCFO, Place, T> col2im;
+      math::Col2VolFunctor<DeviceContext, T> col2vol;
+      math::Col2ImFunctor<math::ColFormat::kCFO, DeviceContext, T> col2im;
 
       for (int i = 0; i < batch_size; i++) {
         Tensor out_grad_batch =
@@ -282,18 +283,17 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
             col_matrix.ShareDataWith(in_grad_slice);
             col_matrix.Resize(col_matrix_shape);
           }
-          math::matmul<Place, T>(context.device_context(), filter_slice, true,
-                                 out_grad_slice, false, T(1.0), &col_matrix,
-                                 T(0.0));
+          math::matmul<DeviceContext, T>(dev_ctx, filter_slice, true,
+                                         out_grad_slice, false, T(1.0),
+                                         &col_matrix, T(0.0));
 
           if (is_expand && data_dim == 2U) {
-            col2im(context.device_context(), col, dilations, strides,
+            col2im(dev_ctx, col, dilations, strides,
                    std::vector<int>{paddings[0], paddings[1], paddings[0],
                                     paddings[1]},
                    &in_grad_slice);
           } else if (is_expand && data_dim == 3U) {
-            col2vol(context.device_context(), col, dilations, strides, paddings,
-                    &in_grad_slice);
+            col2vol(dev_ctx, col, dilations, strides, paddings, &in_grad_slice);
           }
         }
       }
@@ -303,9 +303,9 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
       filter_grad->mutable_data<T>(context.GetPlace());
       Tensor filter_grad_ = *filter_grad;
       filter_grad_.Resize(filter_matrix_shape);
-      set_zero(context.device_context(), filter_grad, static_cast<T>(0));
-      math::Im2ColFunctor<math::ColFormat::kCFO, Place, T> im2col;
-      math::Vol2ColFunctor<Place, T> vol2col;
+      set_zero(dev_ctx, filter_grad, static_cast<T>(0));
+      math::Im2ColFunctor<math::ColFormat::kCFO, DeviceContext, T> im2col;
+      math::Vol2ColFunctor<DeviceContext, T> vol2col;
       for (int i = 0; i < batch_size; i++) {
         Tensor out_grad_batch =
             output_grad->Slice(i, i + 1).Resize(output_matrix_shape);
@@ -321,21 +321,20 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
             col_matrix.ShareDataWith(col);
             col_matrix.Resize(col_matrix_shape);
           } else if (data_dim == 2U) {
-            im2col(context.device_context(), in_slice, dilations, strides,
+            im2col(dev_ctx, in_slice, dilations, strides,
                    std::vector<int>{paddings[0], paddings[1], paddings[0],
                                     paddings[1]},
                    &col);
           } else if (data_dim == 3U) {
-            vol2col(context.device_context(), in_slice, dilations, strides,
-                    paddings, &col);
+            vol2col(dev_ctx, in_slice, dilations, strides, paddings, &col);
           }
 
           // gemm
           Tensor filter_grad_slice =
               filter_grad_.Slice(g * out_step, (g + 1) * out_step);
-          math::matmul<Place, T>(context.device_context(), out_grad_slice,
-                                 false, col_matrix, true, T(1.0),
-                                 &filter_grad_slice, T(1.0));
+          math::matmul<DeviceContext, T>(dev_ctx, out_grad_slice, false,
+                                         col_matrix, true, T(1.0),
+                                         &filter_grad_slice, T(1.0));
         }
       }
     }
diff --git a/paddle/operators/conv_shift_op.cu b/paddle/operators/conv_shift_op.cu
index 95e13c38a8..f7ca82ce26 100644
--- a/paddle/operators/conv_shift_op.cu
+++ b/paddle/operators/conv_shift_op.cu
@@ -111,7 +111,8 @@ __global__ void ConvShiftDy(const T *x, const T *dout, int x_width, int y_width,
 }  // namespace
 
 template <typename T>
-class ConvShiftKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
+class ConvShiftKernel<platform::CUDADeviceContext, T>
+    : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
     const Tensor *X = context.Input<Tensor>("X");
@@ -132,7 +133,8 @@ class ConvShiftKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
 
     dim3 grid_dim(num_x_blocks, batch_size);
 
-    auto stream = context.cuda_device_context().stream();
+    auto stream =
+        context.template device_context<platform::CUDADeviceContext>().stream();
 
     ConvShiftForward<T><<<grid_dim, x_per_block, mem_per_block, stream>>>(
         x_data, y_data, x_width, y_width, y_half_width, batch_size, out_data);
@@ -140,7 +142,7 @@ class ConvShiftKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
 };
 
 template <typename T>
-class ConvShiftGradKernel<platform::GPUPlace, T>
+class ConvShiftGradKernel<platform::CUDADeviceContext, T>
     : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
@@ -159,8 +161,9 @@ class ConvShiftGradKernel<platform::GPUPlace, T>
     int y_width = Y->dims()[1];
     int y_half_width = (y_width - 1) / 2;
 
-    auto &device_ctx = context.cuda_device_context();
-    math::SetConstant<platform::GPUPlace, T> zero;
+    auto &device_ctx =
+        context.template device_context<platform::CUDADeviceContext>();
+    math::SetConstant<platform::CUDADeviceContext, T> zero;
 
     const int x_per_block = 256;
     int num_x_blocks = DivUp(x_width, x_per_block);
@@ -186,8 +189,9 @@ class ConvShiftGradKernel<platform::GPUPlace, T>
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(conv_shift,
-                       ops::ConvShiftKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    conv_shift,
+    ops::ConvShiftKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     conv_shift_grad,
-    ops::ConvShiftGradKernel<paddle::platform::GPUPlace, float>);
+    ops::ConvShiftGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/conv_shift_op.h b/paddle/operators/conv_shift_op.h
index 5a160b0f16..1a70b38a0d 100644
--- a/paddle/operators/conv_shift_op.h
+++ b/paddle/operators/conv_shift_op.h
@@ -18,13 +18,13 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ConvShiftKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override;
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ConvShiftGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override;
diff --git a/paddle/operators/conv_transpose_cudnn_op.cc b/paddle/operators/conv_transpose_cudnn_op.cc
index 0192178ce3..4cb6a2ccff 100644
--- a/paddle/operators/conv_transpose_cudnn_op.cc
+++ b/paddle/operators/conv_transpose_cudnn_op.cc
@@ -61,12 +61,13 @@ REGISTER_OP(conv2d_transpose_cudnn, ops::ConvTransposeOp,
 
 REGISTER_OP_CPU_KERNEL(
     conv2d_transpose_cudnn,
-    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, double>);
+    ops::GemmConvTransposeKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvTransposeKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     conv2d_transpose_cudnn_grad,
-    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, double>);
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUDeviceContext,
+                                     double>);
 
 REGISTER_OP(conv3d_transpose_cudnn, ops::ConvTransposeOp,
             ops::CudnnConv3DTransposeOpMaker, conv3d_transpose_cudnn_grad,
@@ -74,9 +75,10 @@ REGISTER_OP(conv3d_transpose_cudnn, ops::ConvTransposeOp,
 
 REGISTER_OP_CPU_KERNEL(
     conv3d_transpose_cudnn,
-    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, double>);
+    ops::GemmConvTransposeKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvTransposeKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     conv3d_transpose_cudnn_grad,
-    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, double>);
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUDeviceContext,
+                                     double>);
diff --git a/paddle/operators/conv_transpose_cudnn_op.cu.cc b/paddle/operators/conv_transpose_cudnn_op.cu.cc
index 494904fe52..f0297f6c40 100644
--- a/paddle/operators/conv_transpose_cudnn_op.cu.cc
+++ b/paddle/operators/conv_transpose_cudnn_op.cu.cc
@@ -83,7 +83,8 @@ class CudnnConvTransposeOpKernel : public framework::OpKernel<T> {
     }
     // ------------------- cudnn conv algorithm ---------------------
     cudnnConvolutionBwdDataAlgo_t algo;
-    auto handle = ctx.cuda_device_context().cudnn_handle();
+    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto handle = dev_ctx.cudnn_handle();
     // Get the algorithm
     PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
         handle, cudnn_filter_desc, cudnn_input_desc, cudnn_conv_desc,
@@ -165,7 +166,8 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel<T> {
       workspace_size_limit = user_workspace_size * 1024 * 1024;
     }
 
-    auto handle = ctx.cuda_device_context().cudnn_handle();
+    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto handle = dev_ctx.cudnn_handle();
     if (input_grad) {
       // choose backward algorithm for data
       PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm(
@@ -234,16 +236,16 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel<T> {
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(conv2d_transpose_cudnn,
-                       ops::CudnnConvTransposeOpKernel<float>,
-                       ops::CudnnConvTransposeOpKernel<double>);
-REGISTER_OP_GPU_KERNEL(conv2d_transpose_cudnn_grad,
-                       ops::CudnnConvTransposeGradOpKernel<float>,
-                       ops::CudnnConvTransposeGradOpKernel<double>);
-
-REGISTER_OP_GPU_KERNEL(conv3d_transpose_cudnn,
-                       ops::CudnnConvTransposeOpKernel<float>,
-                       ops::CudnnConvTransposeOpKernel<double>);
-REGISTER_OP_GPU_KERNEL(conv3d_transpose_cudnn_grad,
-                       ops::CudnnConvTransposeGradOpKernel<float>,
-                       ops::CudnnConvTransposeGradOpKernel<double>);
+REGISTER_OP_CUDA_KERNEL(conv2d_transpose_cudnn,
+                        ops::CudnnConvTransposeOpKernel<float>,
+                        ops::CudnnConvTransposeOpKernel<double>);
+REGISTER_OP_CUDA_KERNEL(conv2d_transpose_cudnn_grad,
+                        ops::CudnnConvTransposeGradOpKernel<float>,
+                        ops::CudnnConvTransposeGradOpKernel<double>);
+
+REGISTER_OP_CUDA_KERNEL(conv3d_transpose_cudnn,
+                        ops::CudnnConvTransposeOpKernel<float>,
+                        ops::CudnnConvTransposeOpKernel<double>);
+REGISTER_OP_CUDA_KERNEL(conv3d_transpose_cudnn_grad,
+                        ops::CudnnConvTransposeGradOpKernel<float>,
+                        ops::CudnnConvTransposeGradOpKernel<double>);
diff --git a/paddle/operators/conv_transpose_op.cc b/paddle/operators/conv_transpose_op.cc
index 678b192dea..ca063e94bb 100644
--- a/paddle/operators/conv_transpose_op.cc
+++ b/paddle/operators/conv_transpose_op.cc
@@ -197,21 +197,23 @@ REGISTER_OP(conv2d_transpose, ops::ConvTransposeOp, ops::Conv2DTransposeOpMaker,
 
 REGISTER_OP_CPU_KERNEL(
     conv2d_transpose,
-    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, double>);
+    ops::GemmConvTransposeKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvTransposeKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     conv2d_transpose_grad,
-    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, double>);
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUDeviceContext,
+                                     double>);
 
 REGISTER_OP(conv3d_transpose, ops::ConvTransposeOp, ops::Conv3DTransposeOpMaker,
             conv3d_transpose_grad, ops::ConvTransposeOpGrad);
 
 REGISTER_OP_CPU_KERNEL(
     conv3d_transpose,
-    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, double>);
+    ops::GemmConvTransposeKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvTransposeKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     conv3d_transpose_grad,
-    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, float>,
-    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, double>);
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUDeviceContext,
+                                     double>);
diff --git a/paddle/operators/conv_transpose_op.cu.cc b/paddle/operators/conv_transpose_op.cu.cc
index 4165eb0c7b..b91ebd7922 100644
--- a/paddle/operators/conv_transpose_op.cu.cc
+++ b/paddle/operators/conv_transpose_op.cu.cc
@@ -16,20 +16,24 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     conv2d_transpose,
-    ops::GemmConvTransposeKernel<paddle::platform::GPUPlace, float>,
-    ops::GemmConvTransposeKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
+    ops::GemmConvTransposeKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GemmConvTransposeKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
     conv2d_transpose_grad,
-    ops::GemmConvTransposeGradKernel<paddle::platform::GPUPlace, float>,
-    ops::GemmConvTransposeGradKernel<paddle::platform::GPUPlace, double>);
+    ops::GemmConvTransposeGradKernel<paddle::platform::CUDADeviceContext,
+                                     float>,
+    ops::GemmConvTransposeGradKernel<paddle::platform::CUDADeviceContext,
+                                     double>);
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     conv3d_transpose,
-    ops::GemmConvTransposeKernel<paddle::platform::GPUPlace, float>,
-    ops::GemmConvTransposeKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
+    ops::GemmConvTransposeKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GemmConvTransposeKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
     conv3d_transpose_grad,
-    ops::GemmConvTransposeGradKernel<paddle::platform::GPUPlace, float>,
-    ops::GemmConvTransposeGradKernel<paddle::platform::GPUPlace, double>);
+    ops::GemmConvTransposeGradKernel<paddle::platform::CUDADeviceContext,
+                                     float>,
+    ops::GemmConvTransposeGradKernel<paddle::platform::CUDADeviceContext,
+                                     double>);
diff --git a/paddle/operators/conv_transpose_op.h b/paddle/operators/conv_transpose_op.h
index 1cacb770e6..80600b5361 100644
--- a/paddle/operators/conv_transpose_op.h
+++ b/paddle/operators/conv_transpose_op.h
@@ -52,7 +52,7 @@ class ConvTransposeOpGrad : public framework::OperatorWithKernel {
   void InferShape(framework::InferShapeContext* ctx) const override;
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class GemmConvTransposeKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -109,11 +109,12 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
     filter.Resize(filter_matrix_shape);
 
     output->mutable_data<T>(context.GetPlace());
-    math::SetConstant<Place, T> set_zero;
-    set_zero(context.device_context(), output, static_cast<T>(0));
+    math::SetConstant<DeviceContext, T> set_zero;
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    set_zero(dev_ctx, output, static_cast<T>(0));
 
-    math::Col2ImFunctor<math::ColFormat::kCFO, Place, T> col2im;
-    math::Col2VolFunctor<Place, T> col2vol;
+    math::Col2ImFunctor<math::ColFormat::kCFO, DeviceContext, T> col2im;
+    math::Col2VolFunctor<DeviceContext, T> col2vol;
     std::vector<int> dilations({1, 1, 1});
 
     // convolution transpose: gemm + col2im or col2vol (similar to conv-backward
@@ -127,29 +128,27 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
 
       // col_matrix = filter * input_batch
       // of shape (c * k_h * k_w, h * w) or (c * k_d * k_h * k_w, d * h * w)
-      math::matmul<Place, T>(context.device_context(), filter, true,
-                             input_batch, false, static_cast<T>(1.0),
-                             &col_matrix, static_cast<T>(0.0));
+      math::matmul<DeviceContext, T>(dev_ctx, filter, true, input_batch, false,
+                                     static_cast<T>(1.0), &col_matrix,
+                                     static_cast<T>(0.0));
 
       if (data_dim == 2U) {
         // col2im: col_matrix -> dy
         // from (c * k_h * k_w, h * w) to (c, o_h, o_w)
-        col2im(context.device_context(), col,
-               std::vector<int>{dilations[0], dilations[1]}, strides,
-               std::vector<int>{paddings[0], paddings[1], paddings[0],
-                                paddings[1]},
+        col2im(dev_ctx, col, std::vector<int>{dilations[0], dilations[1]},
+               strides, std::vector<int>{paddings[0], paddings[1], paddings[0],
+                                         paddings[1]},
                &output_batch);
       } else if (data_dim == 3U) {
         // col2vol: col_matrix -> dy
         // from (c * k_d * k_h * k_w, d * h * w) to (c, o_d, o_h, o_w)
-        col2vol(context.device_context(), col, dilations, strides, paddings,
-                &output_batch);
+        col2vol(dev_ctx, col, dilations, strides, paddings, &output_batch);
       }
     }
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -206,6 +205,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
     // convolution transpose grad on input:
     // im2col + gemm (similar to conv-forward)
     // input need to compute gradient
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     if (input_grad || filter_grad) {
       Tensor col;
       col.mutable_data<T>(col_shape, context.GetPlace());
@@ -217,19 +217,19 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
       col_matrix.Resize(col_matrix_shape);
 
       Tensor filter_grad_;
-      math::SetConstant<Place, T> set_zero;
+      math::SetConstant<DeviceContext, T> set_zero;
 
-      math::Im2ColFunctor<math::ColFormat::kCFO, Place, T> im2col;
-      math::Vol2ColFunctor<Place, T> vol2col;
+      math::Im2ColFunctor<math::ColFormat::kCFO, DeviceContext, T> im2col;
+      math::Vol2ColFunctor<DeviceContext, T> vol2col;
       std::vector<int> dilations({1, 1, 1});
 
       if (input_grad) {
         input_grad->mutable_data<T>(context.GetPlace());
-        set_zero(context.device_context(), input_grad, static_cast<T>(0));
+        set_zero(dev_ctx, input_grad, static_cast<T>(0));
       }
       if (filter_grad) {  // filter size (m, c, k_h, k_w)
         filter_grad->mutable_data<T>(context.GetPlace());
-        set_zero(context.device_context(), filter_grad, static_cast<T>(0));
+        set_zero(dev_ctx, filter_grad, static_cast<T>(0));
         filter_grad_ = *filter_grad;
         filter_grad_.Resize(filter_matrix_shape);
       }
@@ -242,7 +242,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
         if (data_dim == 2U) {
           // im2col: dy -> col matrix
           // from (c, o_h, o_w) to (c * k_h * k_w, h * w)
-          im2col(context.device_context(), output_grad_batch,
+          im2col(dev_ctx, output_grad_batch,
                  std::vector<int>{dilations[0], dilations[1]}, strides,
                  std::vector<int>{paddings[0], paddings[1], paddings[0],
                                   paddings[1]},
@@ -250,8 +250,8 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
         } else if (data_dim == 3U) {
           // vol2col: dy -> col_matrix
           // from (c, o_d, o_h, o_w) to (c * k_d * k_h * k_w, d * h * w)
-          vol2col(context.device_context(), output_grad_batch, dilations,
-                  strides, paddings, &col);
+          vol2col(dev_ctx, output_grad_batch, dilations, strides, paddings,
+                  &col);
         }
 
         if (input_grad) {
@@ -263,9 +263,9 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
           // or
           // (m, c * k_d * k_h * k_w) * (c * k_d * k_h * k_w, d * h * w) -> (m,
           // d, h, w)
-          math::matmul<Place, T>(context.device_context(), filter, false,
-                                 col_matrix, false, static_cast<T>(1.0),
-                                 &input_grad_batch, static_cast<T>(0.0));
+          math::matmul<DeviceContext, T>(
+              dev_ctx, filter, false, col_matrix, false, static_cast<T>(1.0),
+              &input_grad_batch, static_cast<T>(0.0));
         }
         if (filter_grad) {
           // input batch
@@ -275,9 +275,9 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
           // or
           // (m, d * h * w) * (d * h * w, c * k_d * k_h * k_w) -> (m, c * k_d *
           // k_h * k_w)
-          math::matmul<Place, T>(context.device_context(), in_batch, false,
-                                 col_matrix, true, static_cast<T>(1.0),
-                                 &filter_grad_, static_cast<T>(1.0));
+          math::matmul<DeviceContext, T>(dev_ctx, in_batch, false, col_matrix,
+                                         true, static_cast<T>(1.0),
+                                         &filter_grad_, static_cast<T>(1.0));
         }
       }
     }
diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc
index 312264ccd4..440c427cba 100644
--- a/paddle/operators/cos_sim_op.cc
+++ b/paddle/operators/cos_sim_op.cc
@@ -155,7 +155,8 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(cos_sim, ops::CosSimOp, ops::CosSimOpMaker, cos_sim_grad,
             ops::CosSimOpGrad);
-REGISTER_OP_CPU_KERNEL(cos_sim,
-                       ops::CosSimKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
-    cos_sim_grad, ops::CosSimGradKernel<paddle::platform::CPUPlace, float>);
+    cos_sim, ops::CosSimKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    cos_sim_grad,
+    ops::CosSimGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/cos_sim_op.cu b/paddle/operators/cos_sim_op.cu
index 0cb8fd26de..1cb01f5945 100644
--- a/paddle/operators/cos_sim_op.cu
+++ b/paddle/operators/cos_sim_op.cu
@@ -16,7 +16,8 @@
 #include "paddle/operators/cos_sim_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(cos_sim,
-                       ops::CosSimKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    cos_sim_grad, ops::CosSimGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    cos_sim, ops::CosSimKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    cos_sim_grad,
+    ops::CosSimGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h
index 62a4e484ec..fecb5a79b2 100644
--- a/paddle/operators/cos_sim_op.h
+++ b/paddle/operators/cos_sim_op.h
@@ -27,7 +27,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class CosSimKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -51,7 +51,8 @@ class CosSimKernel : public framework::OpKernel<T> {
     auto y_norm = EigenVector<T>::Flatten(*out_y_norm);
 
     // compute
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
     auto row_along = Eigen::array<int, 1>({{1}});
     x_norm.device(place) = x.square().sum(row_along).sqrt();
     y_norm.device(place) = y.square().sum(row_along).sqrt();
@@ -66,7 +67,7 @@ class CosSimKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class CosSimGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -96,7 +97,8 @@ class CosSimGradKernel : public framework::OpKernel<T> {
     auto z_bcast = z.broadcast(bcast_cols);
     auto dz_bcast = dz.broadcast(bcast_cols);
     auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast_cols);
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
     if (rows_x == rows_y) {
       auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast_cols);
       auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast_cols);
diff --git a/paddle/operators/crf_decoding_op.cc b/paddle/operators/crf_decoding_op.cc
index 291b23ed1b..1ce189fa6e 100644
--- a/paddle/operators/crf_decoding_op.cc
+++ b/paddle/operators/crf_decoding_op.cc
@@ -135,5 +135,6 @@ namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(crf_decoding, ops::CRFDecodingOp,
                              ops::CRFDecodingOpMaker);
 REGISTER_OP_CPU_KERNEL(
-    crf_decoding, ops::CRFDecodingOpKernel<paddle::platform::CPUPlace, float>,
-    ops::CRFDecodingOpKernel<paddle::platform::CPUPlace, double>);
+    crf_decoding,
+    ops::CRFDecodingOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::CRFDecodingOpKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/crf_decoding_op.h b/paddle/operators/crf_decoding_op.h
index 57b5e21b3a..f6827b7b11 100644
--- a/paddle/operators/crf_decoding_op.h
+++ b/paddle/operators/crf_decoding_op.h
@@ -24,7 +24,7 @@ using framework::LoDTensor;
 using framework::LoD;
 using framework::Tensor;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class CRFDecodingOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -44,8 +44,8 @@ class CRFDecodingOpKernel : public framework::OpKernel<T> {
     const size_t seq_num = lod[level].size() - 1;
 
     int64_t* path = decoded_path->mutable_data<int64_t>(platform::CPUPlace());
-    math::SetConstant<platform::CPUPlace, int64_t>()(ctx.device_context(),
-                                                     decoded_path, 0);
+    math::SetConstant<DeviceContext, int64_t>()(
+        ctx.template device_context<DeviceContext>(), decoded_path, 0);
     for (size_t i = 0; i < seq_num; ++i) {
       int start_pos = static_cast<int>(lod[level][i]);
       int end_pos = static_cast<int>(lod[level][i + 1]);
diff --git a/paddle/operators/crop_op.cc b/paddle/operators/crop_op.cc
index 6752eb8c1c..7c2a0ac7a7 100644
--- a/paddle/operators/crop_op.cc
+++ b/paddle/operators/crop_op.cc
@@ -133,5 +133,5 @@ class CropOpGrad : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(crop, ops::CropOp, ops::CropOpMaker, crop_grad, ops::CropOpGrad);
 REGISTER_OP_CPU_KERNEL(crop, ops::CropKernel<float>);
-REGISTER_OP_CPU_KERNEL(crop_grad,
-                       ops::CropGradKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    crop_grad, ops::CropGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/crop_op.cu b/paddle/operators/crop_op.cu
index f8ee18a1d6..90fd83ca10 100644
--- a/paddle/operators/crop_op.cu
+++ b/paddle/operators/crop_op.cu
@@ -16,6 +16,6 @@
 #include "paddle/operators/crop_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(crop, ops::CropKernel<float>);
-REGISTER_OP_GPU_KERNEL(crop_grad,
-                       ops::CropGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(crop, ops::CropKernel<float>);
+REGISTER_OP_CUDA_KERNEL(
+    crop_grad, ops::CropGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/crop_op.h b/paddle/operators/crop_op.h
index 2e72583d68..d531a19c78 100644
--- a/paddle/operators/crop_op.h
+++ b/paddle/operators/crop_op.h
@@ -49,7 +49,7 @@ class CropKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T, size_t D>
+template <typename DeviceContext, typename T, size_t D>
 void CropGradFunction(const framework::ExecutionContext& context) {
   auto* d_x = context.Output<Tensor>(framework::GradVarName("X"));
   if (d_x != nullptr) {
@@ -63,12 +63,13 @@ void CropGradFunction(const framework::ExecutionContext& context) {
     }
     auto d_x_tensor = EigenTensor<T, D>::From(*d_x);
     auto d_out_tensor = EigenTensor<T, D>::From(*d_out);
-    d_x_tensor.device(context.GetEigenDevice<Place>()) =
+    d_x_tensor.device(
+        *context.template device_context<DeviceContext>().eigen_device()) =
         d_out_tensor.pad(paddings, 0);
   }
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class CropGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -76,22 +77,22 @@ class CropGradKernel : public framework::OpKernel<T> {
         context.Input<Tensor>(framework::GradVarName("Out"))->dims().size();
     switch (rank) {
       case 1:
-        CropGradFunction<Place, T, 1>(context);
+        CropGradFunction<DeviceContext, T, 1>(context);
         break;
       case 2:
-        CropGradFunction<Place, T, 2>(context);
+        CropGradFunction<DeviceContext, T, 2>(context);
         break;
       case 3:
-        CropGradFunction<Place, T, 3>(context);
+        CropGradFunction<DeviceContext, T, 3>(context);
         break;
       case 4:
-        CropGradFunction<Place, T, 4>(context);
+        CropGradFunction<DeviceContext, T, 4>(context);
         break;
       case 5:
-        CropGradFunction<Place, T, 5>(context);
+        CropGradFunction<DeviceContext, T, 5>(context);
         break;
       case 6:
-        CropGradFunction<Place, T, 6>(context);
+        CropGradFunction<DeviceContext, T, 6>(context);
         break;
       default:
         PADDLE_THROW(
diff --git a/paddle/operators/cross_entropy_op.cu b/paddle/operators/cross_entropy_op.cu
index 6212e39dfd..0546964588 100644
--- a/paddle/operators/cross_entropy_op.cu
+++ b/paddle/operators/cross_entropy_op.cu
@@ -53,8 +53,9 @@ class CrossEntropyOpCUDAKernel : public framework::OpKernel<T> {
     Tensor* y = ctx.Output<Tensor>("Y");
     y->mutable_data<T>(ctx.GetPlace());
 
-    math::CrossEntropyFunctor<platform::GPUPlace, T>()(
-        ctx.device_context(), y, x, label, ctx.Attr<bool>("soft_label"));
+    math::CrossEntropyFunctor<platform::CUDADeviceContext, T>()(
+        ctx.template device_context<platform::CUDADeviceContext>(), y, x, label,
+        ctx.Attr<bool>("soft_label"));
   }
 };
 
@@ -80,15 +81,17 @@ class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel<T> {
 
     int block = 512;
     int grid = (batch_size * class_num + block - 1) / block;
-    auto stream = ctx.cuda_device_context().stream();
+
+    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto stream = dev_ctx.stream();
 
     if (ctx.Attr<bool>("soft_label")) {
       auto* label_data = label->data<T>();
       SoftCrossEntropyGradientKernel<T><<<grid, block, 0, stream>>>(
           dx_data, dy_data, x_data, label_data, batch_size, class_num);
     } else {
-      math::SetConstant<platform::GPUPlace, T> functor;
-      functor(ctx.device_context(), dx, 0);
+      math::SetConstant<platform::CUDADeviceContext, T> functor;
+      functor(dev_ctx, dx, 0);
       auto* label_data = label->data<int64_t>();
       grid = (batch_size + block - 1) / block;
       CrossEntropyGradientKernel<T><<<grid, block, 0, stream>>>(
@@ -101,8 +104,8 @@ class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(cross_entropy, ops::CrossEntropyOpCUDAKernel<float>,
-                       ops::CrossEntropyOpCUDAKernel<double>);
-REGISTER_OP_GPU_KERNEL(cross_entropy_grad,
-                       ops::CrossEntropyGradientOpCUDAKernel<float>,
-                       ops::CrossEntropyGradientOpCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(cross_entropy, ops::CrossEntropyOpCUDAKernel<float>,
+                        ops::CrossEntropyOpCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(cross_entropy_grad,
+                        ops::CrossEntropyGradientOpCUDAKernel<float>,
+                        ops::CrossEntropyGradientOpCUDAKernel<double>);
diff --git a/paddle/operators/cross_entropy_op.h b/paddle/operators/cross_entropy_op.h
index 37db0a930a..5623d2ded1 100644
--- a/paddle/operators/cross_entropy_op.h
+++ b/paddle/operators/cross_entropy_op.h
@@ -37,8 +37,9 @@ class CrossEntropyOpKernel : public framework::OpKernel<T> {
     Tensor* y = ctx.Output<Tensor>("Y");
     y->mutable_data<T>(ctx.GetPlace());
 
-    math::CrossEntropyFunctor<platform::CPUPlace, T>()(
-        ctx.device_context(), y, x, labels, ctx.Attr<bool>("soft_label"));
+    math::CrossEntropyFunctor<platform::CPUDeviceContext, T>()(
+        ctx.template device_context<platform::CPUDeviceContext>(), y, x, labels,
+        ctx.Attr<bool>("soft_label"));
   }
 };
 
@@ -61,7 +62,8 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel<T> {
       auto lbl_mat = EigenMatrix<T>::From(*label);
       auto dx_mat = EigenMatrix<T>::From(*dx);
 
-      dx_mat.device(ctx.GetEigenDevice<platform::CPUPlace>()) =
+      dx_mat.device(*ctx.template device_context<platform::CPUDeviceContext>()
+                         .eigen_device()) =
           -(lbl_mat *
             dy_mat.broadcast(Eigen::DSizes<int64_t, 2>(1, class_num)) / x_mat);
     } else {
@@ -70,8 +72,8 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel<T> {
       const T* x_data = x->data<T>();
       const int64_t* label_data = label->data<int64_t>();
 
-      math::SetConstant<platform::CPUPlace, T> functor;
-      functor(ctx.device_context(), dx, 0);
+      math::SetConstant<platform::CPUDeviceContext, T> functor;
+      functor(ctx.template device_context<platform::CPUDeviceContext>(), dx, 0);
 
       for (int64_t i = 0; i < batch_size; ++i) {
         PADDLE_ASSERT(label_data[i] >= 0 || label_data[i] < class_num);
diff --git a/paddle/operators/decayed_adagrad_op.cc b/paddle/operators/decayed_adagrad_op.cc
index 640b4e7744..fd29c7270b 100644
--- a/paddle/operators/decayed_adagrad_op.cc
+++ b/paddle/operators/decayed_adagrad_op.cc
@@ -99,4 +99,4 @@ REGISTER_OP_WITHOUT_GRADIENT(decayed_adagrad, ops::DecayedAdagradOp,
                              ops::DecayedAdagradOpMaker);
 REGISTER_OP_CPU_KERNEL(
     decayed_adagrad,
-    ops::DecayedAdagradOpKernel<paddle::platform::CPUPlace, float>);
+    ops::DecayedAdagradOpKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/decayed_adagrad_op.cu b/paddle/operators/decayed_adagrad_op.cu
index 6fce77fe4e..282b90f275 100644
--- a/paddle/operators/decayed_adagrad_op.cu
+++ b/paddle/operators/decayed_adagrad_op.cu
@@ -16,6 +16,6 @@
 #include "paddle/operators/decayed_adagrad_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     decayed_adagrad,
-    ops::DecayedAdagradOpKernel<paddle::platform::GPUPlace, float>);
+    ops::DecayedAdagradOpKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/decayed_adagrad_op.h b/paddle/operators/decayed_adagrad_op.h
index 0fe0fc5acd..fec9705cfc 100644
--- a/paddle/operators/decayed_adagrad_op.h
+++ b/paddle/operators/decayed_adagrad_op.h
@@ -19,7 +19,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class DecayedAdagradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -43,7 +43,7 @@ class DecayedAdagradOpKernel : public framework::OpKernel<T> {
 
     auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
     auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
-    auto place = ctx.GetEigenDevice<Place>();
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
 
     moment_out.device(place) = decay * moment + (1 - decay) * grad * grad;
     Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
diff --git a/paddle/operators/dropout_op.cc b/paddle/operators/dropout_op.cc
index 932c0bf8fb..acd526ae80 100644
--- a/paddle/operators/dropout_op.cc
+++ b/paddle/operators/dropout_op.cc
@@ -100,6 +100,8 @@ namespace ops = paddle::operators;
 REGISTER_OP(dropout, ops::DropoutOp, ops::DropoutOpMaker<float>, dropout_grad,
             ops::DropoutOpGrad<float>);
 REGISTER_OP_CPU_KERNEL(
-    dropout, ops::CPUDropoutKernel<paddle::platform::CPUPlace, float, float>);
+    dropout,
+    ops::CPUDropoutKernel<paddle::platform::CPUDeviceContext, float, float>);
 REGISTER_OP_CPU_KERNEL(
-    dropout_grad, ops::DropoutGradKernel<paddle::platform::CPUPlace, float>);
+    dropout_grad,
+    ops::DropoutGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/dropout_op.cu b/paddle/operators/dropout_op.cu
index db3578b9bf..10c670751d 100644
--- a/paddle/operators/dropout_op.cu
+++ b/paddle/operators/dropout_op.cu
@@ -58,7 +58,7 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
     auto X = EigenMatrix<T>::Reshape(*x, 1);
     auto Y = EigenMatrix<T>::Reshape(*y, 1);
 
-    auto place = context.GetEigenDevice<Place>();
+    auto& place = *context.template device_context<Place>().eigen_device();
     if (!context.Attr<bool>("is_test")) {
       auto* mask = context.Output<Tensor>("Mask");
       auto* mask_data = mask->mutable_data<T>(context.GetPlace());
@@ -80,7 +80,9 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    dropout, ops::GPUDropoutKernel<paddle::platform::GPUPlace, float, float>);
-REGISTER_OP_GPU_KERNEL(
-    dropout_grad, ops::DropoutGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    dropout,
+    ops::GPUDropoutKernel<paddle::platform::CUDADeviceContext, float, float>);
+REGISTER_OP_CUDA_KERNEL(
+    dropout_grad,
+    ops::DropoutGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/dropout_op.h b/paddle/operators/dropout_op.h
index d9a130fdc0..84ad39f0bb 100644
--- a/paddle/operators/dropout_op.h
+++ b/paddle/operators/dropout_op.h
@@ -25,7 +25,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename Place, typename T, typename AttrType>
+template <typename DeviceContext, typename T, typename AttrType>
 class CPUDropoutKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -55,13 +55,14 @@ class CPUDropoutKernel : public framework::OpKernel<T> {
     } else {
       auto X = EigenMatrix<T>::Reshape(*x, 1);
       auto Y = EigenMatrix<T>::Reshape(*y, 1);
-      auto place = context.GetEigenDevice<Place>();
+      auto& place =
+          *context.template device_context<DeviceContext>().eigen_device();
       Y.device(place) = X * dropout_prob;
     }
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class DropoutGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -77,7 +78,8 @@ class DropoutGradKernel : public framework::OpKernel<T> {
     auto dX = EigenMatrix<T>::Reshape(*grad_x, 1);
     auto dY = EigenMatrix<T>::Reshape(*grad_y, 1);
 
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
     dX.device(place) = dY * M;
   }
 };
diff --git a/paddle/operators/elementwise_add_op.cc b/paddle/operators/elementwise_add_op.cc
index 432b9ba6f7..a62eeeeb95 100644
--- a/paddle/operators/elementwise_add_op.cc
+++ b/paddle/operators/elementwise_add_op.cc
@@ -34,13 +34,13 @@ REGISTER_OP(elementwise_add, ops::ElementwiseOp, ops::ElementwiseAddOpMaker,
             elementwise_add_grad, ops::ElementwiseOpGrad);
 REGISTER_OP_CPU_KERNEL(
     elementwise_add,
-    ops::ElementwiseAddKernel<paddle::platform::CPUPlace, float>,
-    ops::ElementwiseAddKernel<paddle::platform::CPUPlace, double>,
-    ops::ElementwiseAddKernel<paddle::platform::CPUPlace, int>,
-    ops::ElementwiseAddKernel<paddle::platform::CPUPlace, int64_t>);
+    ops::ElementwiseAddKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ElementwiseAddKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ElementwiseAddKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ElementwiseAddKernel<paddle::platform::CPUDeviceContext, int64_t>);
 REGISTER_OP_CPU_KERNEL(
     elementwise_add_grad,
-    ops::ElementwiseAddGradKernel<paddle::platform::CPUPlace, float>,
-    ops::ElementwiseAddGradKernel<paddle::platform::CPUPlace, double>,
-    ops::ElementwiseAddGradKernel<paddle::platform::CPUPlace, int>,
-    ops::ElementwiseAddGradKernel<paddle::platform::CPUPlace, int64_t>);
+    ops::ElementwiseAddGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ElementwiseAddGradKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ElementwiseAddGradKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ElementwiseAddGradKernel<paddle::platform::CPUDeviceContext, int64_t>);
diff --git a/paddle/operators/elementwise_add_op.cu b/paddle/operators/elementwise_add_op.cu
index 7591428ac7..78642bb424 100644
--- a/paddle/operators/elementwise_add_op.cu
+++ b/paddle/operators/elementwise_add_op.cu
@@ -17,15 +17,16 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     elementwise_add,
-    ops::ElementwiseAddKernel<paddle::platform::GPUPlace, float>,
-    ops::ElementwiseAddKernel<paddle::platform::GPUPlace, double>,
-    ops::ElementwiseAddKernel<paddle::platform::GPUPlace, int>,
-    ops::ElementwiseAddKernel<paddle::platform::GPUPlace, int64_t>);
-REGISTER_OP_GPU_KERNEL(
+    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
     elementwise_add_grad,
-    ops::ElementwiseAddGradKernel<paddle::platform::GPUPlace, float>,
-    ops::ElementwiseAddGradKernel<paddle::platform::GPUPlace, double>,
-    ops::ElementwiseAddGradKernel<paddle::platform::GPUPlace, int>,
-    ops::ElementwiseAddGradKernel<paddle::platform::GPUPlace, int64_t>);
+    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext,
+                                  int64_t>);
diff --git a/paddle/operators/elementwise_add_op.h b/paddle/operators/elementwise_add_op.h
index 921dc5f6a6..069bdaf0ab 100644
--- a/paddle/operators/elementwise_add_op.h
+++ b/paddle/operators/elementwise_add_op.h
@@ -24,7 +24,7 @@ struct AddFunctor {
   inline HOSTDEVICE T operator()(T a, T b) const { return a + b; }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ElementwiseAddKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -34,8 +34,8 @@ class ElementwiseAddKernel : public framework::OpKernel<T> {
     auto* y = ctx.Input<Tensor>("Y");
     auto* z = ctx.Output<Tensor>("Out");
     z->mutable_data<T>(ctx.GetPlace());
-    TransformFunctor<AddFunctor<T>, T, Place> functor(
-        x, y, z, ctx.device_context(), AddFunctor<T>());
+    TransformFunctor<AddFunctor<T>, T, DeviceContext> functor(
+        x, y, z, ctx.template device_context<DeviceContext>(), AddFunctor<T>());
 
     auto x_dims = x->dims();
     auto y_dims = y->dims();
@@ -137,11 +137,11 @@ struct ElementwiseAddBroadCast2GradFunctor {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ElementwiseAddGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseGradCompute<Place, T, ElementwiseAddGradFunctor<T>,
+    ElementwiseGradCompute<DeviceContext, T, ElementwiseAddGradFunctor<T>,
                            ElementwiseAddOneGradFunctor<T>,
                            ElementwiseAddBroadCastGradFunctor<T>,
                            ElementwiseAddBroadCast2GradFunctor<T>>(ctx);
diff --git a/paddle/operators/elementwise_div_op.cc b/paddle/operators/elementwise_div_op.cc
index 7a325199bd..1c3e9e70ee 100644
--- a/paddle/operators/elementwise_div_op.cc
+++ b/paddle/operators/elementwise_div_op.cc
@@ -35,13 +35,13 @@ REGISTER_OP(elementwise_div, ops::ElementwiseOp, ops::ElementwiseDivOpMaker,
             elementwise_div_grad, ops::ElementwiseOpGrad);
 REGISTER_OP_CPU_KERNEL(
     elementwise_div,
-    ops::ElementwiseDivKernel<paddle::platform::CPUPlace, float>,
-    ops::ElementwiseDivKernel<paddle::platform::CPUPlace, double>,
-    ops::ElementwiseDivKernel<paddle::platform::CPUPlace, int>,
-    ops::ElementwiseDivKernel<paddle::platform::CPUPlace, int64_t>);
+    ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext, int64_t>);
 REGISTER_OP_CPU_KERNEL(
     elementwise_div_grad,
-    ops::ElementwiseDivGradKernel<paddle::platform::CPUPlace, float>,
-    ops::ElementwiseDivGradKernel<paddle::platform::CPUPlace, double>,
-    ops::ElementwiseDivGradKernel<paddle::platform::CPUPlace, int>,
-    ops::ElementwiseDivGradKernel<paddle::platform::CPUPlace, int64_t>);
+    ops::ElementwiseDivGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ElementwiseDivGradKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ElementwiseDivGradKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ElementwiseDivGradKernel<paddle::platform::CPUDeviceContext, int64_t>);
diff --git a/paddle/operators/elementwise_div_op.cu b/paddle/operators/elementwise_div_op.cu
index de4d0c3344..502c528936 100644
--- a/paddle/operators/elementwise_div_op.cu
+++ b/paddle/operators/elementwise_div_op.cu
@@ -17,15 +17,16 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     elementwise_div,
-    ops::ElementwiseDivKernel<paddle::platform::GPUPlace, float>,
-    ops::ElementwiseDivKernel<paddle::platform::GPUPlace, double>,
-    ops::ElementwiseDivKernel<paddle::platform::GPUPlace, int>,
-    ops::ElementwiseDivKernel<paddle::platform::GPUPlace, int64_t>);
-REGISTER_OP_GPU_KERNEL(
+    ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
     elementwise_div_grad,
-    ops::ElementwiseDivGradKernel<paddle::platform::GPUPlace, float>,
-    ops::ElementwiseDivGradKernel<paddle::platform::GPUPlace, double>,
-    ops::ElementwiseDivGradKernel<paddle::platform::GPUPlace, int>,
-    ops::ElementwiseDivGradKernel<paddle::platform::GPUPlace, int64_t>);
+    ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext,
+                                  int64_t>);
diff --git a/paddle/operators/elementwise_div_op.h b/paddle/operators/elementwise_div_op.h
index 8946ff3d25..d91313db42 100644
--- a/paddle/operators/elementwise_div_op.h
+++ b/paddle/operators/elementwise_div_op.h
@@ -19,11 +19,11 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ElementwiseDivKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseCompute<EigenDivFunctor, Place, T>(ctx);
+    ElementwiseCompute<EigenDivFunctor, DeviceContext, T>(ctx);
   }
 };
 
@@ -102,11 +102,11 @@ struct ElementwiseDivBroadCast2GradFunctor {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ElementwiseDivGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseGradCompute<Place, T, ElementwiseDivGradFunctor<T>,
+    ElementwiseGradCompute<DeviceContext, T, ElementwiseDivGradFunctor<T>,
                            ElementwiseDivGradFunctor<T>,
                            ElementwiseDivBroadCastGradFunctor<T>,
                            ElementwiseDivBroadCast2GradFunctor<T>>(ctx);
diff --git a/paddle/operators/elementwise_mul_op.cc b/paddle/operators/elementwise_mul_op.cc
index 8851267a52..aadb95cbe3 100644
--- a/paddle/operators/elementwise_mul_op.cc
+++ b/paddle/operators/elementwise_mul_op.cc
@@ -36,13 +36,13 @@ REGISTER_OP(elementwise_mul, ops::ElementwiseOp, ops::ElementwiseMulOpMaker,
             elementwise_mul_grad, ops::ElementwiseOpGrad);
 REGISTER_OP_CPU_KERNEL(
     elementwise_mul,
-    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, float>,
-    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, double>,
-    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, int>,
-    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, int64_t>);
+    ops::ElementwiseMulKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ElementwiseMulKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ElementwiseMulKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ElementwiseMulKernel<paddle::platform::CPUDeviceContext, int64_t>);
 REGISTER_OP_CPU_KERNEL(
     elementwise_mul_grad,
-    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, float>,
-    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, double>,
-    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, int>,
-    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, int64_t>);
+    ops::ElementwiseMulGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CPUDeviceContext, int64_t>);
diff --git a/paddle/operators/elementwise_mul_op.cu b/paddle/operators/elementwise_mul_op.cu
index b0dfdee1cc..089451b3e1 100644
--- a/paddle/operators/elementwise_mul_op.cu
+++ b/paddle/operators/elementwise_mul_op.cu
@@ -17,15 +17,16 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     elementwise_mul,
-    ops::ElementwiseMulKernel<paddle::platform::GPUPlace, float>,
-    ops::ElementwiseMulKernel<paddle::platform::GPUPlace, double>,
-    ops::ElementwiseMulKernel<paddle::platform::GPUPlace, int>,
-    ops::ElementwiseMulKernel<paddle::platform::GPUPlace, int64_t>);
-REGISTER_OP_GPU_KERNEL(
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
     elementwise_mul_grad,
-    ops::ElementwiseMulGradKernel<paddle::platform::GPUPlace, float>,
-    ops::ElementwiseMulGradKernel<paddle::platform::GPUPlace, double>,
-    ops::ElementwiseMulGradKernel<paddle::platform::GPUPlace, int>,
-    ops::ElementwiseMulGradKernel<paddle::platform::GPUPlace, int64_t>);
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext,
+                                  int64_t>);
diff --git a/paddle/operators/elementwise_mul_op.h b/paddle/operators/elementwise_mul_op.h
index 4469b07eaa..16fa5ec4b3 100644
--- a/paddle/operators/elementwise_mul_op.h
+++ b/paddle/operators/elementwise_mul_op.h
@@ -18,11 +18,11 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ElementwiseMulKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseCompute<EigenMulFunctor, Place, T>(ctx);
+    ElementwiseCompute<EigenMulFunctor, DeviceContext, T>(ctx);
   }
 };
 
@@ -101,11 +101,11 @@ struct ElementwiseMulBroadCast2GradFunctor {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ElementwiseMulGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseGradCompute<Place, T, ElementwiseMulGradFunctor<T>,
+    ElementwiseGradCompute<DeviceContext, T, ElementwiseMulGradFunctor<T>,
                            ElementwiseMulGradFunctor<T>,
                            ElementwiseMulBroadCastGradFunctor<T>,
                            ElementwiseMulBroadCast2GradFunctor<T>>(ctx);
diff --git a/paddle/operators/elementwise_op_function.h b/paddle/operators/elementwise_op_function.h
index ca3542e783..7ebfc7df8c 100644
--- a/paddle/operators/elementwise_op_function.h
+++ b/paddle/operators/elementwise_op_function.h
@@ -59,17 +59,17 @@ inline void get_mid_dims(const framework::DDim& x_dims,
   }
 }
 
-template <typename T, typename Place>
+template <typename T, typename DeviceContext>
 class RowwiseTransformIterator;
-template <typename T, typename Place>
+template <typename T, typename DeviceContext>
 class MidWiseTransformIterator;
 
 template <typename T>
-class RowwiseTransformIterator<T, platform::CPUPlace> {
+class RowwiseTransformIterator<T, platform::CPUDeviceContext> {
  public:
   RowwiseTransformIterator(const T* ptr, int n) : ptr_(ptr), i_(0), n_(n) {}
 
-  RowwiseTransformIterator<T, platform::CPUPlace>& operator++() {
+  RowwiseTransformIterator<T, platform::CPUDeviceContext>& operator++() {
     ++i_;
     if (UNLIKELY(i_ == n_)) {
       i_ = 0;
@@ -77,13 +77,13 @@ class RowwiseTransformIterator<T, platform::CPUPlace> {
     return *this;
   }
 
-  bool operator==(
-      const RowwiseTransformIterator<T, platform::CPUPlace>& rhs) const {
+  bool operator==(const RowwiseTransformIterator<T, platform::CPUDeviceContext>&
+                      rhs) const {
     return (ptr_ + i_) == &(*rhs);
   }
 
-  bool operator!=(
-      const RowwiseTransformIterator<T, platform::CPUPlace>& rhs) const {
+  bool operator!=(const RowwiseTransformIterator<T, platform::CPUDeviceContext>&
+                      rhs) const {
     return (ptr_ + i_) != &(*rhs);
   }
 
@@ -96,12 +96,12 @@ class RowwiseTransformIterator<T, platform::CPUPlace> {
 };
 
 template <typename T>
-class MidWiseTransformIterator<T, platform::CPUPlace> {
+class MidWiseTransformIterator<T, platform::CPUDeviceContext> {
  public:
   MidWiseTransformIterator(const T* ptr, int n, int post)
       : ptr_(ptr), i_(0), j_(0), n_(n), post_(post) {}
 
-  MidWiseTransformIterator<T, platform::CPUPlace>& operator++() {
+  MidWiseTransformIterator<T, platform::CPUDeviceContext>& operator++() {
     ++j_;
     i_ = j_ / post_;
     if (UNLIKELY(i_ == n_)) {
@@ -111,13 +111,13 @@ class MidWiseTransformIterator<T, platform::CPUPlace> {
     return *this;
   }
 
-  bool operator==(
-      const MidWiseTransformIterator<T, platform::CPUPlace>& rhs) const {
+  bool operator==(const MidWiseTransformIterator<T, platform::CPUDeviceContext>&
+                      rhs) const {
     return (ptr_ + i_) == &(*rhs);
   }
 
-  bool operator!=(
-      const MidWiseTransformIterator<T, platform::CPUPlace>& rhs) const {
+  bool operator!=(const MidWiseTransformIterator<T, platform::CPUDeviceContext>&
+                      rhs) const {
     return (ptr_ + i_) != &(*rhs);
   }
 
@@ -133,12 +133,12 @@ class MidWiseTransformIterator<T, platform::CPUPlace> {
 
 #ifdef __NVCC__
 template <typename T>
-class RowwiseTransformIterator<T, platform::GPUPlace>
+class RowwiseTransformIterator<T, platform::CUDADeviceContext>
     : public thrust::iterator_adaptor<
-          RowwiseTransformIterator<T, platform::GPUPlace>, const T*> {
+          RowwiseTransformIterator<T, platform::CUDADeviceContext>, const T*> {
  public:
   typedef thrust::iterator_adaptor<
-      RowwiseTransformIterator<T, platform::GPUPlace>, const T*>
+      RowwiseTransformIterator<T, platform::CUDADeviceContext>, const T*>
       super_t;
   HOSTDEVICE RowwiseTransformIterator(const T* x, int n)
       : super_t(x), begin_(x), n_(n){};
@@ -153,12 +153,12 @@ class RowwiseTransformIterator<T, platform::GPUPlace>
 };
 
 template <typename T>
-class MidWiseTransformIterator<T, platform::GPUPlace>
+class MidWiseTransformIterator<T, platform::CUDADeviceContext>
     : public thrust::iterator_adaptor<
-          MidWiseTransformIterator<T, platform::GPUPlace>, const T*> {
+          MidWiseTransformIterator<T, platform::CUDADeviceContext>, const T*> {
  public:
   typedef thrust::iterator_adaptor<
-      MidWiseTransformIterator<T, platform::GPUPlace>, const T*>
+      MidWiseTransformIterator<T, platform::CUDADeviceContext>, const T*>
       super_t;
   HOSTDEVICE MidWiseTransformIterator(const T* x, int n, int post)
       : super_t(x), begin_(x), n_(n), post_(post){};
@@ -174,12 +174,11 @@ class MidWiseTransformIterator<T, platform::GPUPlace>
 };
 #endif
 
-template <typename Functor, typename T, typename Place>
+template <typename Functor, typename T, typename DeviceContext>
 class TransformFunctor {
  public:
   TransformFunctor(const framework::Tensor* x, const framework::Tensor* y,
-                   framework::Tensor* z, const platform::DeviceContext& ctx,
-                   Functor func)
+                   framework::Tensor* z, const DeviceContext& ctx, Functor func)
       : x_(x->data<T>()),
         y_(y->data<T>()),
         z_(z->mutable_data<T>(ctx.GetPlace())),
@@ -188,20 +187,20 @@ class TransformFunctor {
         func_(func) {}
 
   inline void Run() const {
-    platform::Transform<Place> trans;
+    platform::Transform<DeviceContext> trans;
     trans(ctx_, x_, x_ + nx_, y_, z_, func_);
   }
 
   inline void RunRowWise(int n, int pre) const {
-    platform::Transform<Place> trans;
-    trans(ctx_, x_, x_ + nx_, RowwiseTransformIterator<T, Place>(y_, n), z_,
-          func_);
+    platform::Transform<DeviceContext> trans;
+    trans(ctx_, x_, x_ + nx_, RowwiseTransformIterator<T, DeviceContext>(y_, n),
+          z_, func_);
   }
 
   inline void RunMidWise(int n, int pre, int post) const {
-    platform::Transform<Place> trans;
-    trans(ctx_, x_, x_ + nx_, MidWiseTransformIterator<T, Place>(y_, n, post),
-          z_, func_);
+    platform::Transform<DeviceContext> trans;
+    trans(ctx_, x_, x_ + nx_,
+          MidWiseTransformIterator<T, DeviceContext>(y_, n, post), z_, func_);
   }
 
  private:
@@ -209,22 +208,24 @@ class TransformFunctor {
   const T* y_;
   T* z_;
   int64_t nx_;
-  const platform::DeviceContext& ctx_;
+  const DeviceContext& ctx_;
   Functor func_;
 };
 
 #define EIGEN_FUNCTOR(name, eigen_op)                                          \
   struct Eigen##name##Functor {                                                \
-    template <typename Place, typename T>                                      \
+    template <typename DeviceContext, typename T>                              \
     inline void Run(const framework::Tensor* x, const framework::Tensor* y,    \
                     framework::Tensor* z,                                      \
                     const framework::ExecutionContext& ctx) {                  \
       auto x_e = framework::EigenVector<T>::Flatten(*x);                       \
       auto y_e = framework::EigenVector<T>::Flatten(*y);                       \
       auto z_e = framework::EigenVector<T>::Flatten(*z);                       \
-      z_e.device(ctx.GetEigenDevice<Place>()) = eigen_op(x_e, y_e);            \
+      z_e.device(                                                              \
+          *ctx.template device_context<DeviceContext>().eigen_device()) =      \
+          eigen_op(x_e, y_e);                                                  \
     }                                                                          \
-    template <typename Place, typename T>                                      \
+    template <typename DeviceContext, typename T>                              \
     inline void RunBroadCast(const framework::Tensor* x,                       \
                              const framework::Tensor* y, framework::Tensor* z, \
                              const framework::ExecutionContext& ctx, int pre,  \
@@ -235,9 +236,11 @@ class TransformFunctor {
       auto y_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))                  \
                          .broadcast(Eigen::DSizes<int, 2>(pre, 1))             \
                          .reshape(Eigen::DSizes<int, 1>(x_e.size()));          \
-      z_e.device(ctx.GetEigenDevice<Place>()) = eigen_op(x_e, y_bcast);        \
+      z_e.device(                                                              \
+          *ctx.template device_context<DeviceContext>().eigen_device()) =      \
+          eigen_op(x_e, y_bcast);                                              \
     }                                                                          \
-    template <typename Place, typename T>                                      \
+    template <typename DeviceContext, typename T>                              \
     inline void RunBroadCast2(const framework::Tensor* x,                      \
                               const framework::Tensor* y,                      \
                               framework::Tensor* z,                            \
@@ -249,11 +252,13 @@ class TransformFunctor {
       auto y_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))               \
                          .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))       \
                          .reshape(Eigen::DSizes<int, 1>(x_e.size()));          \
-      z_e.device(ctx.GetEigenDevice<Place>()) = eigen_op(x_e, y_bcast);        \
+      z_e.device(                                                              \
+          *ctx.template device_context<DeviceContext>().eigen_device()) =      \
+          eigen_op(x_e, y_bcast);                                              \
     }                                                                          \
   }
 
-template <class functor, typename Place, typename T>
+template <class functor, typename DeviceContext, typename T>
 void ElementwiseCompute(const framework::ExecutionContext& ctx) {
   using Tensor = framework::Tensor;
 
@@ -269,7 +274,7 @@ void ElementwiseCompute(const framework::ExecutionContext& ctx) {
 
   if (x_dims == y_dims) {
     functor f;
-    f.template Run<Place, T>(x, y, z, ctx);
+    f.template Run<DeviceContext, T>(x, y, z, ctx);
     return;
   }
 
@@ -282,11 +287,11 @@ void ElementwiseCompute(const framework::ExecutionContext& ctx) {
   get_mid_dims(x_dims, y_dims, axis, pre, n, post);
   if (post == 1) {
     functor f;
-    f.template RunBroadCast<Place, T>(x, y, z, ctx, pre, n);
+    f.template RunBroadCast<DeviceContext, T>(x, y, z, ctx, pre, n);
     return;
   } else {
     functor f;
-    f.template RunBroadCast2<Place, T>(x, y, z, ctx, pre, n, post);
+    f.template RunBroadCast2<DeviceContext, T>(x, y, z, ctx, pre, n, post);
     return;
   }
 }
@@ -303,8 +308,9 @@ EIGEN_FUNCTOR(Mul, EIGEN_MUL);
 #define EIGEN_DIV(x, y) ((x) / (y))
 EIGEN_FUNCTOR(Div, EIGEN_DIV);
 
-template <typename Place, typename T, typename functor, typename functor1,
-          typename broadcastfunctor, typename broadcast2functor>
+template <typename DeviceContext, typename T, typename functor,
+          typename functor1, typename broadcastfunctor,
+          typename broadcast2functor>
 void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
   using Tensor = framework::Tensor;
 
@@ -313,7 +319,7 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
   auto* out = ctx.Input<Tensor>("Out");
   auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
 
-  auto place = ctx.GetEigenDevice<Place>();
+  auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
 
   auto x_dims = x->dims();
   auto y_dims = y->dims();
diff --git a/paddle/operators/elementwise_sub_op.cc b/paddle/operators/elementwise_sub_op.cc
index 95d7979e39..3e4d19361e 100644
--- a/paddle/operators/elementwise_sub_op.cc
+++ b/paddle/operators/elementwise_sub_op.cc
@@ -34,13 +34,13 @@ REGISTER_OP(elementwise_sub, ops::ElementwiseOp, ops::ElementwiseSubOpMaker,
             elementwise_sub_grad, ops::ElementwiseOpGrad);
 REGISTER_OP_CPU_KERNEL(
     elementwise_sub,
-    ops::ElementwiseSubKernel<paddle::platform::CPUPlace, float>,
-    ops::ElementwiseSubKernel<paddle::platform::CPUPlace, double>,
-    ops::ElementwiseSubKernel<paddle::platform::CPUPlace, int>,
-    ops::ElementwiseSubKernel<paddle::platform::CPUPlace, int64_t>);
+    ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext, int64_t>);
 REGISTER_OP_CPU_KERNEL(
     elementwise_sub_grad,
-    ops::ElementwiseSubGradKernel<paddle::platform::CPUPlace, float>,
-    ops::ElementwiseSubGradKernel<paddle::platform::CPUPlace, double>,
-    ops::ElementwiseSubGradKernel<paddle::platform::CPUPlace, int>,
-    ops::ElementwiseSubGradKernel<paddle::platform::CPUPlace, int64_t>);
+    ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext, int64_t>);
diff --git a/paddle/operators/elementwise_sub_op.cu b/paddle/operators/elementwise_sub_op.cu
index ec23bec35f..0b2f0f7d4d 100644
--- a/paddle/operators/elementwise_sub_op.cu
+++ b/paddle/operators/elementwise_sub_op.cu
@@ -17,15 +17,16 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     elementwise_sub,
-    ops::ElementwiseSubKernel<paddle::platform::GPUPlace, float>,
-    ops::ElementwiseSubKernel<paddle::platform::GPUPlace, double>,
-    ops::ElementwiseSubKernel<paddle::platform::GPUPlace, int>,
-    ops::ElementwiseSubKernel<paddle::platform::GPUPlace, int64_t>);
-REGISTER_OP_GPU_KERNEL(
+    ops::ElementwiseSubKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseSubKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseSubKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseSubKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
     elementwise_sub_grad,
-    ops::ElementwiseSubGradKernel<paddle::platform::GPUPlace, float>,
-    ops::ElementwiseSubGradKernel<paddle::platform::GPUPlace, double>,
-    ops::ElementwiseSubGradKernel<paddle::platform::GPUPlace, int>,
-    ops::ElementwiseSubGradKernel<paddle::platform::GPUPlace, int64_t>);
+    ops::ElementwiseSubGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseSubGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseSubGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseSubGradKernel<paddle::platform::CUDADeviceContext,
+                                  int64_t>);
diff --git a/paddle/operators/elementwise_sub_op.h b/paddle/operators/elementwise_sub_op.h
index 3f40c1c5bc..731a30c5e3 100644
--- a/paddle/operators/elementwise_sub_op.h
+++ b/paddle/operators/elementwise_sub_op.h
@@ -18,11 +18,11 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ElementwiseSubKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseCompute<EigenSubFunctor, Place, T>(ctx);
+    ElementwiseCompute<EigenSubFunctor, DeviceContext, T>(ctx);
   }
 };
 
@@ -101,11 +101,11 @@ struct ElementwiseSubBroadCast2GradFunctor {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ElementwiseSubGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseGradCompute<Place, T, ElementwiseSubGradFunctor<T>,
+    ElementwiseGradCompute<DeviceContext, T, ElementwiseSubGradFunctor<T>,
                            ElementwiseSubOneGradFunctor<T>,
                            ElementwiseSubBroadCastGradFunctor<T>,
                            ElementwiseSubBroadCast2GradFunctor<T>>(ctx);
diff --git a/paddle/operators/expand_op.cc b/paddle/operators/expand_op.cc
index 282775fcda..8b3cddbb94 100644
--- a/paddle/operators/expand_op.cc
+++ b/paddle/operators/expand_op.cc
@@ -130,7 +130,8 @@ class ExpandGradOp : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(expand, ops::ExpandOp, ops::ExpandOpMaker, expand_grad,
             ops::ExpandGradOp);
-REGISTER_OP_CPU_KERNEL(expand,
-                       ops::ExpandKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
-    expand_grad, ops::ExpandGradKernel<paddle::platform::CPUPlace, float>);
+    expand, ops::ExpandKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    expand_grad,
+    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/expand_op.cu b/paddle/operators/expand_op.cu
index 6744562b6c..99ee584d08 100644
--- a/paddle/operators/expand_op.cu
+++ b/paddle/operators/expand_op.cu
@@ -17,7 +17,8 @@
 #include "paddle/operators/expand_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(expand,
-                       ops::ExpandKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    expand_grad, ops::ExpandGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    expand, ops::ExpandKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    expand_grad,
+    ops::ExpandGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/expand_op.h b/paddle/operators/expand_op.h
index 4d7996ad1e..14ef8b0912 100644
--- a/paddle/operators/expand_op.h
+++ b/paddle/operators/expand_op.h
@@ -56,7 +56,7 @@ template <typename T, size_t D, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ExpandKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -83,12 +83,13 @@ class ExpandKernel : public framework::OpKernel<T> {
     auto x = EigenTensor<T, Rank>::From(*in0);
     out0->mutable_data<T>(context.GetPlace());
     auto y = EigenTensor<T, Rank>::From(*out0);
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
     y.device(place) = x.broadcast(bcast_dims);
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ExpandGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -164,7 +165,8 @@ class ExpandGradKernel : public framework::OpKernel<T> {
       reduce_dims[i] = reduce_dims_vec[i];
     }
     auto out_grad = EigenVector<T>::Flatten(*in0);
-    x_grad.device(context.GetEigenDevice<Place>()) =
+    x_grad.device(
+        *context.template device_context<DeviceContext>().eigen_device()) =
         out_grad.reshape(reshape_dims).sum(reduce_dims).reshape(x.dimensions());
   }
 };
diff --git a/paddle/operators/fill_constant_batch_size_like_op.cc b/paddle/operators/fill_constant_batch_size_like_op.cc
index 892922cd3a..7fb74e2b95 100644
--- a/paddle/operators/fill_constant_batch_size_like_op.cc
+++ b/paddle/operators/fill_constant_batch_size_like_op.cc
@@ -100,8 +100,11 @@ REGISTER_OPERATOR(fill_constant_batch_size_like,
                   ops::FillConstantBatchSizeLikeOpMaker);
 REGISTER_OP_CPU_KERNEL(
     fill_constant_batch_size_like,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, float>,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, double>,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, int>,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUDeviceContext,
+                                           float>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUDeviceContext,
+                                           double>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUDeviceContext,
+                                           int>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUDeviceContext,
                                            int64_t>);
diff --git a/paddle/operators/fill_constant_batch_size_like_op.cu.cc b/paddle/operators/fill_constant_batch_size_like_op.cu.cc
index 9e7a1eeab8..2e0e15f36b 100644
--- a/paddle/operators/fill_constant_batch_size_like_op.cu.cc
+++ b/paddle/operators/fill_constant_batch_size_like_op.cu.cc
@@ -16,10 +16,13 @@
 #include "paddle/framework/op_registry.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     fill_constant_batch_size_like,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, float>,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, double>,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, int>,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CUDADeviceContext,
+                                           float>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CUDADeviceContext,
+                                           double>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CUDADeviceContext,
+                                           int>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CUDADeviceContext,
                                            int64_t>);
diff --git a/paddle/operators/fill_constant_batch_size_like_op.h b/paddle/operators/fill_constant_batch_size_like_op.h
index 339d97a30a..66da9d0307 100644
--- a/paddle/operators/fill_constant_batch_size_like_op.h
+++ b/paddle/operators/fill_constant_batch_size_like_op.h
@@ -19,7 +19,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class FillConstantBatchSizeLikeOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -27,8 +27,9 @@ class FillConstantBatchSizeLikeOpKernel : public framework::OpKernel<T> {
     out->mutable_data<T>(ctx.GetPlace());
     auto value = ctx.Attr<float>("value");
 
-    math::SetConstant<Place, T> setter;
-    setter(ctx.device_context(), out, static_cast<T>(value));
+    math::SetConstant<DeviceContext, T> setter;
+    setter(ctx.template device_context<DeviceContext>(), out,
+           static_cast<T>(value));
   }
 };
 
diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc
index 95fb5932b8..720c11f5f1 100644
--- a/paddle/operators/fill_zeros_like_op.cc
+++ b/paddle/operators/fill_zeros_like_op.cc
@@ -54,8 +54,9 @@ namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, ops::FillZerosLikeOp,
                              ops::FillZerosLikeOpMaker);
 REGISTER_OP_CPU_KERNEL(
-    fill_zeros_like, ops::FillZerosLikeKernel<paddle::platform::CPUPlace, int>,
-    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, int64_t>,
-    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, float>,
-    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, double>,
-    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, bool>);
+    fill_zeros_like,
+    ops::FillZerosLikeKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUDeviceContext, int64_t>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUDeviceContext, bool>);
diff --git a/paddle/operators/fill_zeros_like_op.cu.cc b/paddle/operators/fill_zeros_like_op.cu.cc
index 1501a17441..9f412306bb 100644
--- a/paddle/operators/fill_zeros_like_op.cu.cc
+++ b/paddle/operators/fill_zeros_like_op.cu.cc
@@ -16,9 +16,10 @@
 #include "paddle/framework/op_registry.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    fill_zeros_like, ops::FillZerosLikeKernel<paddle::platform::GPUPlace, int>,
-    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, int64_t>,
-    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, float>,
-    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, double>,
-    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, bool>);
+REGISTER_OP_CUDA_KERNEL(
+    fill_zeros_like,
+    ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, int64_t>,
+    ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, bool>);
diff --git a/paddle/operators/fill_zeros_like_op.h b/paddle/operators/fill_zeros_like_op.h
index 7e7d78eea2..a6e2941f52 100644
--- a/paddle/operators/fill_zeros_like_op.h
+++ b/paddle/operators/fill_zeros_like_op.h
@@ -19,15 +19,16 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class FillZerosLikeKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto* out = context.Output<framework::Tensor>("Y");
     out->mutable_data<T>(context.GetPlace());
 
-    math::SetConstant<Place, T> setter;
-    setter(context.device_context(), out, static_cast<T>(0));
+    math::SetConstant<DeviceContext, T> setter;
+    setter(context.template device_context<DeviceContext>(), out,
+           static_cast<T>(0));
   }
 };
 
diff --git a/paddle/operators/ftrl_op.cc b/paddle/operators/ftrl_op.cc
index cb7ae69196..b14913ff21 100644
--- a/paddle/operators/ftrl_op.cc
+++ b/paddle/operators/ftrl_op.cc
@@ -135,5 +135,5 @@ The paper that proposed Follow The Regularized Leader (FTRL):
 
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(ftrl, ops::FTRLOp, ops::FTRLOpMaker);
-REGISTER_OP_CPU_KERNEL(ftrl,
-                       ops::FTRLOpKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    ftrl, ops::FTRLOpKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/ftrl_op.cu b/paddle/operators/ftrl_op.cu
index 97b36dade6..abbbe7adbe 100644
--- a/paddle/operators/ftrl_op.cu
+++ b/paddle/operators/ftrl_op.cu
@@ -15,5 +15,5 @@ specific language governing permissions and limitations under the License. */
 #include "paddle/operators/ftrl_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(ftrl,
-                       ops::FTRLOpKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    ftrl, ops::FTRLOpKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/ftrl_op.h b/paddle/operators/ftrl_op.h
index b040162f8d..4eea04cd8d 100644
--- a/paddle/operators/ftrl_op.h
+++ b/paddle/operators/ftrl_op.h
@@ -24,7 +24,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class FTRLOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -53,7 +53,7 @@ class FTRLOpKernel : public framework::OpKernel<T> {
     auto p_out = EigenVector<T>::Flatten(*param_out);
     auto s_acc_out = EigenVector<T>::Flatten(*sq_accum_out);
     auto l_acc_out = EigenVector<T>::Flatten(*lin_accum_out);
-    auto place = ctx.GetEigenDevice<Place>();
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
 
     Eigen::DSizes<int, 1> grad_dsize(grad->numel());
 
diff --git a/paddle/operators/gather.cu.h b/paddle/operators/gather.cu.h
index 8d04ecd284..c806aa5f05 100644
--- a/paddle/operators/gather.cu.h
+++ b/paddle/operators/gather.cu.h
@@ -20,7 +20,7 @@ namespace paddle {
 namespace operators {
 
 using framework::Tensor;
-using platform::Place;
+using platform::DeviceContext;
 
 #define CUDA_1D_KERNEL_LOOP(i, n)                              \
   for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \
diff --git a/paddle/operators/gather_op.cu b/paddle/operators/gather_op.cu
index 92219d6a43..b37f0576e2 100644
--- a/paddle/operators/gather_op.cu
+++ b/paddle/operators/gather_op.cu
@@ -49,7 +49,8 @@ class GatherGradOpCUDAKernel : public framework::OpKernel<T> {
 
     dX->mutable_data<T>(ctx.GetPlace());
     auto dxt = framework::EigenVector<T>::Flatten(*dX);
-    auto place = ctx.GetEigenDevice<platform::GPUPlace>();
+    auto &place = *ctx.template device_context<platform::CUDADeviceContext>()
+                       .eigen_device();
     dxt.device(place) = dxt.constant(static_cast<T>(0));
 
     GPUScatterAssign<T>(ctx.device_context(), *dO, *Index, dX);
@@ -60,5 +61,5 @@ class GatherGradOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(gather, ops::GatherOpCUDAKernel<float>);
-REGISTER_OP_GPU_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel<float>);
+REGISTER_OP_CUDA_KERNEL(gather, ops::GatherOpCUDAKernel<float>);
+REGISTER_OP_CUDA_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel<float>);
diff --git a/paddle/operators/gather_op.h b/paddle/operators/gather_op.h
index 8276ed0d3d..1a1ba0c41a 100644
--- a/paddle/operators/gather_op.h
+++ b/paddle/operators/gather_op.h
@@ -53,7 +53,8 @@ class GatherGradientOpKernel : public framework::OpKernel<T> {
 
     dX->mutable_data<T>(ctx.GetPlace());
     auto dxt = framework::EigenVector<T>::Flatten(*dX);
-    auto place = ctx.GetEigenDevice<platform::CPUPlace>();
+    auto &place = *ctx.template device_context<platform::CPUDeviceContext>()
+                       .eigen_device();
     dxt.device(place) = dxt.constant(static_cast<T>(0));
 
     ScatterAssign<T>(ctx.device_context(), *dO, *Index, dX);
diff --git a/paddle/operators/gaussian_random_op.cu b/paddle/operators/gaussian_random_op.cu
index 315560bf1b..ffce6f7138 100644
--- a/paddle/operators/gaussian_random_op.cu
+++ b/paddle/operators/gaussian_random_op.cu
@@ -60,5 +60,5 @@ class GPUGaussianRandomKernel : public framework::OpKernel<T> {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_GPU_KERNEL(gaussian_random,
-                       paddle::operators::GPUGaussianRandomKernel<float>);
+REGISTER_OP_CUDA_KERNEL(gaussian_random,
+                        paddle::operators::GPUGaussianRandomKernel<float>);
diff --git a/paddle/operators/gru_op.cc b/paddle/operators/gru_op.cc
index 5aa03f8916..311e7edcf1 100644
--- a/paddle/operators/gru_op.cc
+++ b/paddle/operators/gru_op.cc
@@ -213,8 +213,9 @@ class GRUGradOp : public framework::OperatorWithKernel {
 
 namespace ops = paddle::operators;
 REGISTER_OP(gru, ops::GRUOp, ops::GRUOpMaker, gru_grad, ops::GRUGradOp);
-REGISTER_OP_CPU_KERNEL(gru, ops::GRUKernel<paddle::platform::CPUPlace, float>,
-                       ops::GRUKernel<paddle::platform::CPUPlace, double>);
-REGISTER_OP_CPU_KERNEL(gru_grad,
-                       ops::GRUGradKernel<paddle::platform::CPUPlace, float>,
-                       ops::GRUGradKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(
+    gru, ops::GRUKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GRUKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    gru_grad, ops::GRUGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GRUGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/gru_op.cu.cc b/paddle/operators/gru_op.cu.cc
index 0ceff94ec3..458630ca61 100644
--- a/paddle/operators/gru_op.cu.cc
+++ b/paddle/operators/gru_op.cu.cc
@@ -15,8 +15,9 @@
 #include "paddle/operators/gru_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(gru, ops::GRUKernel<paddle::platform::GPUPlace, float>,
-                       ops::GRUKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(gru_grad,
-                       ops::GRUGradKernel<paddle::platform::GPUPlace, float>,
-                       ops::GRUGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    gru, ops::GRUKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GRUKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    gru_grad, ops::GRUGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GRUGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/gru_op.h b/paddle/operators/gru_op.h
index 564489d3a9..6d02dff578 100644
--- a/paddle/operators/gru_op.h
+++ b/paddle/operators/gru_op.h
@@ -27,16 +27,16 @@ namespace operators {
 using LoDTensor = framework::LoDTensor;
 using Tensor = framework::Tensor;
 
-template <typename Place, typename T>
-inline void ReorderInitState(const platform::DeviceContext& ctx,
+template <typename DeviceContext, typename T>
+inline void ReorderInitState(const DeviceContext& ctx,
                              const framework::Tensor& src, const size_t* index,
                              framework::Tensor* dst, bool indexed_src) {
-  math::CopyMatrixRowsFunctor<Place, T> row_shuffle;
+  math::CopyMatrixRowsFunctor<DeviceContext, T> row_shuffle;
   dst->mutable_data<T>(src.dims(), ctx.GetPlace());
   row_shuffle(ctx, src, index, *dst, indexed_src);
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class GRUKernel : public framework::OpKernel<T> {
  public:
   void BatchCompute(const framework::ExecutionContext& context) const {
@@ -60,12 +60,12 @@ class GRUKernel : public framework::OpKernel<T> {
     auto hidden_dims = hidden->dims();
 
     bool is_reverse = context.Attr<bool>("is_reverse");
-    math::LoDTensor2BatchFunctor<Place, T> to_batch;
-    auto& dev_ctx = context.device_context();
+    math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     to_batch(dev_ctx, *input, *batch_gate, true, is_reverse);
 
     if (bias) {
-      math::RowwiseAdd<Place, T> add_bias;
+      math::RowwiseAdd<DeviceContext, T> add_bias;
       add_bias(dev_ctx, *batch_gate, *bias, batch_gate);
     }
 
@@ -80,8 +80,9 @@ class GRUKernel : public framework::OpKernel<T> {
       // Since the batch computing for GRU reorders the input sequences
       // according to their length. The initialized cell state also needs
       // to reorder.
-      ReorderInitState<Place, T>(context.device_context(), *h0, order,
-                                 &ordered_h0, true);
+      ReorderInitState<DeviceContext, T>(
+          context.template device_context<DeviceContext>(), *h0, order,
+          &ordered_h0, true);
       gru_value.prev_out_value = ordered_h0.data<T>();
     } else {
       gru_value.prev_out_value = nullptr;
@@ -99,14 +100,14 @@ class GRUKernel : public framework::OpKernel<T> {
       gru_value.output_value = hidden_t.data<T>();
       gru_value.gate_value = gate_t.data<T>();
       gru_value.reset_output_value = reset_hidden_prev_t.data<T>();
-      math::GRUUnitFunctor<Place, T>::compute(
+      math::GRUUnitFunctor<DeviceContext, T>::compute(
           dev_ctx, gru_value, frame_size, cur_batch_size,
           math::ActiveType(context.Attr<std::string>("activation")),
           math::ActiveType(context.Attr<std::string>("gate_activation")));
       gru_value.prev_out_value = gru_value.output_value;
     }
 
-    math::Batch2LoDTensorFunctor<Place, T> to_seq;
+    math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
     batch_hidden->set_lod(batch_gate->lod());
     to_seq(dev_ctx, *batch_hidden, *hidden);
   }
@@ -116,7 +117,7 @@ class GRUKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class GRUGradKernel : public framework::OpKernel<T> {
  public:
   void BatchCompute(const framework::ExecutionContext& context) const {
@@ -141,14 +142,14 @@ class GRUGradKernel : public framework::OpKernel<T> {
     auto hidden_dims = hidden->dims();
     int frame_size = hidden_dims[1];
 
-    math::LoDTensor2BatchFunctor<Place, T> to_batch;
+    math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
     LoDTensor batch_hidden_grad, batch_gate_grad, batch_reset_hidden_prev_grad;
     batch_hidden_grad.mutable_data<T>(hidden_dims, context.GetPlace());
     batch_gate_grad.mutable_data<T>(gate_dims, context.GetPlace());
     batch_reset_hidden_prev_grad.mutable_data<T>(hidden_dims,
                                                  context.GetPlace());
-    math::SetConstant<Place, T> zero;
-    auto& dev_ctx = context.device_context();
+    math::SetConstant<DeviceContext, T> zero;
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     zero(dev_ctx, &batch_hidden_grad, static_cast<T>(0.0));
     zero(dev_ctx, &batch_gate_grad, static_cast<T>(0.0));
     zero(dev_ctx, &batch_reset_hidden_prev_grad, static_cast<T>(0.0));
@@ -156,12 +157,13 @@ class GRUGradKernel : public framework::OpKernel<T> {
     Tensor ordered_h0, ordered_h0_grad;
     const size_t* order = batch_gate->lod()[2].data();
     if (h0) {
-      ReorderInitState<Place, T>(context.device_context(), *h0, order,
-                                 &ordered_h0, true);
+      ReorderInitState<DeviceContext, T>(dev_ctx, *h0, order, &ordered_h0,
+                                         true);
     }
     if (h0_grad) {
       ordered_h0_grad.mutable_data<T>(h0_grad->dims(), context.GetPlace());
-      zero(context.device_context(), &ordered_h0_grad, static_cast<T>(0.0));
+      zero(context.template device_context<DeviceContext>(), &ordered_h0_grad,
+           static_cast<T>(0.0));
     }
 
     bool is_reverse = context.Attr<bool>("is_reverse");
@@ -216,25 +218,25 @@ class GRUGradKernel : public framework::OpKernel<T> {
         gru_grad.prev_out_grad = hidden_prev_grad_t.data<T>();
       }
 
-      math::GRUUnitGradFunctor<Place, T>::compute(
+      math::GRUUnitGradFunctor<DeviceContext, T>::compute(
           dev_ctx, gru_value, gru_grad, frame_size, cur_batch_size,
           math::ActiveType(context.Attr<std::string>("activation")),
           math::ActiveType(context.Attr<std::string>("gate_activation")));
     }
     if (input_grad) {
       input_grad->mutable_data<T>(context.GetPlace());
-      math::Batch2LoDTensorFunctor<Place, T> to_seq;
+      math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
       batch_gate_grad.set_lod(batch_gate->lod());
       to_seq(dev_ctx, batch_gate_grad, *input_grad);
     }
     if (bias_grad) {
       bias_grad->mutable_data<T>(context.GetPlace());
-      math::ColwiseSum<Place, T> col_sum;
+      math::ColwiseSum<DeviceContext, T> col_sum;
       col_sum(dev_ctx, batch_gate_grad, bias_grad);
     }
     if (h0 && h0_grad) {
-      ReorderInitState<Place, T>(context.device_context(), ordered_h0_grad,
-                                 order, h0_grad, false);
+      ReorderInitState<DeviceContext, T>(dev_ctx, ordered_h0_grad, order,
+                                         h0_grad, false);
     }
   }
 
diff --git a/paddle/operators/gru_unit_op.cc b/paddle/operators/gru_unit_op.cc
index 877c969103..705de87be5 100644
--- a/paddle/operators/gru_unit_op.cc
+++ b/paddle/operators/gru_unit_op.cc
@@ -201,9 +201,10 @@ class GRUUnitGradOp : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(gru_unit, ops::GRUUnitOp, ops::GRUUnitOpMaker, gru_unit_grad,
             ops::GRUUnitGradOp);
-REGISTER_OP_CPU_KERNEL(gru_unit,
-                       ops::GRUUnitKernel<paddle::platform::CPUPlace, float>,
-                       ops::GRUUnitKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
-    gru_unit_grad, ops::GRUUnitGradKernel<paddle::platform::CPUPlace, float>,
-    ops::GRUUnitGradKernel<paddle::platform::CPUPlace, double>);
+    gru_unit, ops::GRUUnitKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GRUUnitKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    gru_unit_grad,
+    ops::GRUUnitGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GRUUnitGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/gru_unit_op.cu b/paddle/operators/gru_unit_op.cu
index 821c8c6421..7c752db494 100644
--- a/paddle/operators/gru_unit_op.cu
+++ b/paddle/operators/gru_unit_op.cu
@@ -16,9 +16,10 @@
 #include "paddle/operators/gru_unit_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(gru_unit,
-                       ops::GRUUnitKernel<paddle::platform::GPUPlace, float>,
-                       ops::GRUUnitKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
-    gru_unit_grad, ops::GRUUnitGradKernel<paddle::platform::GPUPlace, float>,
-    ops::GRUUnitGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    gru_unit, ops::GRUUnitKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GRUUnitKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    gru_unit_grad,
+    ops::GRUUnitGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GRUUnitGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/gru_unit_op.h b/paddle/operators/gru_unit_op.h
index 3398c0934e..8fe60c750d 100644
--- a/paddle/operators/gru_unit_op.h
+++ b/paddle/operators/gru_unit_op.h
@@ -34,7 +34,7 @@ using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
 enum GRUActivationType { identity = 0, sigmoid = 1, tanh = 2, relu = 3 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class GRUUnitKernel : public framework::OpKernel<T> {
  public:
   template <typename Device, typename X, typename Y>
@@ -71,7 +71,8 @@ class GRUUnitKernel : public framework::OpKernel<T> {
     auto g = EigenMatrix<T>::From(*gate);
     auto r_h_p = EigenMatrix<T>::From(*reset_hidden_prev);
     auto h = EigenMatrix<T>::From(*hidden);
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     // calculate unactivated gate outputs
     if (bias) {
@@ -86,10 +87,10 @@ class GRUUnitKernel : public framework::OpKernel<T> {
     const T* weight_data = weight->data<T>();
     T* gate_data = gate->data<T>();
     T* reset_hidden_prev_data = reset_hidden_prev->data<T>();
-    math::gemm<Place, T>(context.device_context(), false, false, batch_size,
-                         2 * frame_size, frame_size, 1, hidden_prev_data,
-                         frame_size, weight_data, frame_size * 2, 1, gate_data,
-                         frame_size * 3);
+    math::gemm<DeviceContext, T>(
+        context.template device_context<DeviceContext>(), false, false,
+        batch_size, 2 * frame_size, frame_size, 1, hidden_prev_data, frame_size,
+        weight_data, frame_size * 2, 1, gate_data, frame_size * 3);
 
     // calculate activited gate
     Eigen::array<int, 2> extents({{batch_size, frame_size}});
@@ -102,11 +103,11 @@ class GRUUnitKernel : public framework::OpKernel<T> {
                g.slice(r_offsets, extents), g.slice(r_offsets, extents));
     auto r = g.slice(r_offsets, extents);  // reset gate
     r_h_p.device(place) = r * h_p;         // reset previous hidden state
-    math::gemm<Place, T>(context.device_context(), false, false, batch_size,
-                         frame_size, frame_size, 1, reset_hidden_prev_data,
-                         frame_size, weight_data + frame_size * frame_size * 2,
-                         frame_size, 1, gate_data + frame_size * 2,
-                         frame_size * 3);
+    math::gemm<DeviceContext, T>(
+        context.template device_context<DeviceContext>(), false, false,
+        batch_size, frame_size, frame_size, 1, reset_hidden_prev_data,
+        frame_size, weight_data + frame_size * frame_size * 2, frame_size, 1,
+        gate_data + frame_size * 2, frame_size * 3);
 
     Eigen::array<int, 2> c_offsets({{0, frame_size * 2}});
     ActCompute(context.Attr<int>("activation"), place,
@@ -118,7 +119,7 @@ class GRUUnitKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class GRUUnitGradKernel : public framework::OpKernel<T> {
  public:
   template <typename Device, typename X, typename Y, typename DX, typename DY>
@@ -166,7 +167,8 @@ class GRUUnitGradKernel : public framework::OpKernel<T> {
     auto d_h = EigenMatrix<T>::From(*hidden_grad);
     auto d_g = EigenMatrix<T>::From(gate_grad);
     auto d_r_h_p = EigenMatrix<T>::From(reset_hidden_prev_grad);
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     int batch_size = input->dims()[0];
     int frame_size = hidden_prev->dims()[1];
@@ -186,11 +188,11 @@ class GRUUnitGradKernel : public framework::OpKernel<T> {
     ActGradCompute(context.Attr<int>("activation"), place, c, c,
                    d_g.slice(c_offsets, extents), d_h * u);
     // backward for reset_hidden_prev
-    math::gemm<Place, T>(context.device_context(), false, true, batch_size,
-                         frame_size, frame_size, 1,
-                         gate_grad_data + frame_size * 2, frame_size * 3,
-                         weight_data + frame_size * frame_size * 2, frame_size,
-                         0, reset_hidden_prev_grad_data, frame_size);
+    math::gemm<DeviceContext, T>(
+        context.template device_context<DeviceContext>(), false, true,
+        batch_size, frame_size, frame_size, 1, gate_grad_data + frame_size * 2,
+        frame_size * 3, weight_data + frame_size * frame_size * 2, frame_size,
+        0, reset_hidden_prev_grad_data, frame_size);
     // backward for unactivated reset gate
     ActGradCompute(context.Attr<int>("gate_activation"), place, r, r,
                    d_g.slice(r_offsets, extents), d_r_h_p * h_p);
@@ -198,17 +200,18 @@ class GRUUnitGradKernel : public framework::OpKernel<T> {
     if (weight_grad) {
       T* weight_grad_data = weight_grad->mutable_data<T>(context.GetPlace());
       // backward for state_weight
-      math::gemm<Place, T>(
-          context.device_context(), true, false, frame_size, frame_size,
-          batch_size, 1, reset_hidden_prev_data, frame_size,
-          gate_grad_data + frame_size * 2, frame_size * 3, 0,
+      math::gemm<DeviceContext, T>(
+          context.template device_context<DeviceContext>(), true, false,
+          frame_size, frame_size, batch_size, 1, reset_hidden_prev_data,
+          frame_size, gate_grad_data + frame_size * 2, frame_size * 3, 0,
           weight_grad_data + frame_size * frame_size * 2, frame_size);
 
       // backward for update_gate_weight and reset_gate_weight
-      math::gemm<Place, T>(context.device_context(), true, false, frame_size,
-                           frame_size * 2, batch_size, 1, hidden_prev_data,
-                           frame_size, gate_grad_data, frame_size * 3, 0,
-                           weight_grad_data, frame_size * 2);
+      math::gemm<DeviceContext, T>(
+          context.template device_context<DeviceContext>(), true, false,
+          frame_size, frame_size * 2, batch_size, 1, hidden_prev_data,
+          frame_size, gate_grad_data, frame_size * 3, 0, weight_grad_data,
+          frame_size * 2);
     }
     // backward for hidden_prev
     if (hidden_prev_grad) {
@@ -216,10 +219,11 @@ class GRUUnitGradKernel : public framework::OpKernel<T> {
           hidden_prev_grad->mutable_data<T>(context.GetPlace());
       auto d_h_p = EigenMatrix<T>::From(*hidden_prev_grad);
       d_h_p.device(place) = d_r_h_p * r + d_h * (u.constant(T(1)) - u);
-      math::gemm<Place, T>(context.device_context(), false, true, batch_size,
-                           frame_size, frame_size * 2, 1, gate_grad_data,
-                           frame_size * 3, weight_data, frame_size * 2, 1,
-                           hidden_prev_grad_data, frame_size);
+      math::gemm<DeviceContext, T>(
+          context.template device_context<DeviceContext>(), false, true,
+          batch_size, frame_size, frame_size * 2, 1, gate_grad_data,
+          frame_size * 3, weight_data, frame_size * 2, 1, hidden_prev_grad_data,
+          frame_size);
     }
     // backward for input
     if (input_grad) {
diff --git a/paddle/operators/hinge_loss_op.cc b/paddle/operators/hinge_loss_op.cc
index 1e13897bb6..373b4d99b4 100644
--- a/paddle/operators/hinge_loss_op.cc
+++ b/paddle/operators/hinge_loss_op.cc
@@ -106,8 +106,9 @@ class HingeLossGradOp : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(hinge_loss, ops::HingeLossOp, ops::HingeLossOpMaker<float>,
             hinge_loss_grad, ops::HingeLossGradOp);
-REGISTER_OP_CPU_KERNEL(hinge_loss,
-                       ops::HingeLossKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    hinge_loss,
+    ops::HingeLossKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     hinge_loss_grad,
-    ops::HingeLossGradKernel<paddle::platform::CPUPlace, float>);
+    ops::HingeLossGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/hinge_loss_op.cu b/paddle/operators/hinge_loss_op.cu
index ec20b08e30..31a5bde292 100644
--- a/paddle/operators/hinge_loss_op.cu
+++ b/paddle/operators/hinge_loss_op.cu
@@ -16,8 +16,9 @@
 #include "paddle/operators/hinge_loss_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(hinge_loss,
-                       ops::HingeLossKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    hinge_loss,
+    ops::HingeLossKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     hinge_loss_grad,
-    ops::HingeLossGradKernel<paddle::platform::GPUPlace, float>);
+    ops::HingeLossGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/hinge_loss_op.h b/paddle/operators/hinge_loss_op.h
index c0be496f9c..91369cfb8a 100644
--- a/paddle/operators/hinge_loss_op.h
+++ b/paddle/operators/hinge_loss_op.h
@@ -19,14 +19,15 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T, typename AttrType = T>
+template <typename DeviceContext, typename T, typename AttrType = T>
 class HingeLossKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto* pred = context.Input<framework::Tensor>("Logits");
     auto* label = context.Input<framework::Tensor>("Labels");
     auto* loss = context.Output<framework::Tensor>("Loss");
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     auto x = framework::EigenVector<T>::Flatten(*pred);
     auto y = framework::EigenVector<T>::Flatten(*label);
@@ -38,7 +39,7 @@ class HingeLossKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T, typename AttrType = T>
+template <typename DeviceContext, typename T, typename AttrType = T>
 class HingeLossGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -48,7 +49,8 @@ class HingeLossGradKernel : public framework::OpKernel<T> {
         context.Input<framework::Tensor>(framework::GradVarName("Loss"));
     auto* dpred =
         context.Output<framework::Tensor>(framework::GradVarName("Logits"));
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     auto x = framework::EigenVector<T>::Flatten(*pred);
     auto y = framework::EigenVector<T>::Flatten(*label);
diff --git a/paddle/operators/huber_loss_op.cc b/paddle/operators/huber_loss_op.cc
index 938803d5b3..11828d083a 100644
--- a/paddle/operators/huber_loss_op.cc
+++ b/paddle/operators/huber_loss_op.cc
@@ -124,8 +124,9 @@ class HuberLossGradOp : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(huber_loss, ops::HuberLossOp, ops::HuberLossOpMaker<float>,
             huber_loss_grad, ops::HuberLossGradOp);
-REGISTER_OP_CPU_KERNEL(huber_loss,
-                       ops::HuberLossKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    huber_loss,
+    ops::HuberLossKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     huber_loss_grad,
-    ops::HuberLossGradKernel<paddle::platform::CPUPlace, float>);
+    ops::HuberLossGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/huber_loss_op.cu b/paddle/operators/huber_loss_op.cu
index 317321dc6c..d49a4d9d42 100644
--- a/paddle/operators/huber_loss_op.cu
+++ b/paddle/operators/huber_loss_op.cu
@@ -16,8 +16,9 @@
 #include "paddle/operators/huber_loss_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(huber_loss,
-                       ops::HuberLossKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    huber_loss,
+    ops::HuberLossKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     huber_loss_grad,
-    ops::HuberLossGradKernel<paddle::platform::GPUPlace, float>);
+    ops::HuberLossGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/huber_loss_op.h b/paddle/operators/huber_loss_op.h
index 4e7bc55432..4dd20e8b08 100644
--- a/paddle/operators/huber_loss_op.h
+++ b/paddle/operators/huber_loss_op.h
@@ -41,7 +41,7 @@ struct HuberLossForward {
   T delta;
 };
 
-template <typename Place, typename T, typename AttrType = T>
+template <typename DeviceContext, typename T, typename AttrType = T>
 class HuberLossKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -50,7 +50,8 @@ class HuberLossKernel : public framework::OpKernel<T> {
     auto* out0 = context.Output<Tensor>("Residual");
     auto* out1 = context.Output<Tensor>("Out");
     auto delta = static_cast<T>(context.Attr<AttrType>("delta"));
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     auto x = EigenVector<T>::Flatten(*in0);
     auto y = EigenVector<T>::Flatten(*in1);
@@ -85,7 +86,7 @@ struct HuberLossBackward {
   T delta;
 };
 
-template <typename Place, typename T, typename AttrType = T>
+template <typename DeviceContext, typename T, typename AttrType = T>
 class HuberLossGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -94,7 +95,8 @@ class HuberLossGradKernel : public framework::OpKernel<T> {
     auto* out0 = context.Output<Tensor>(framework::GradVarName("X"));
     auto* out1 = context.Output<Tensor>(framework::GradVarName("Y"));
     auto delta = static_cast<T>(context.op().Attr<AttrType>("delta"));
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     auto residual = EigenVector<T>::Flatten(*in0);
     auto out_grad = EigenVector<T>::Flatten(*in1);
diff --git a/paddle/operators/l1_norm_op.cc b/paddle/operators/l1_norm_op.cc
index 02ebf02296..c0b51202c6 100644
--- a/paddle/operators/l1_norm_op.cc
+++ b/paddle/operators/l1_norm_op.cc
@@ -69,7 +69,8 @@ $$Out = \sum{|X|}$$
 namespace ops = paddle::operators;
 REGISTER_OP(l1_norm, ops::L1NormOp, ops::L1NormOpMaker, l1_norm_grad,
             ops::L1NormGradOp);
-REGISTER_OP_CPU_KERNEL(l1_norm,
-                       ops::L1NormKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
-    l1_norm_grad, ops::L1NormGradKernel<paddle::platform::CPUPlace, float>);
+    l1_norm, ops::L1NormKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    l1_norm_grad,
+    ops::L1NormGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/l1_norm_op.cu b/paddle/operators/l1_norm_op.cu
index 1c206e04cc..fd725f86f6 100644
--- a/paddle/operators/l1_norm_op.cu
+++ b/paddle/operators/l1_norm_op.cu
@@ -16,7 +16,8 @@
 #include "paddle/operators/l1_norm_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(l1_norm,
-                       ops::L1NormKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    l1_norm_grad, ops::L1NormGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    l1_norm, ops::L1NormKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    l1_norm_grad,
+    ops::L1NormGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/l1_norm_op.h b/paddle/operators/l1_norm_op.h
index 3c60dc3dc7..ae3878f2b7 100644
--- a/paddle/operators/l1_norm_op.h
+++ b/paddle/operators/l1_norm_op.h
@@ -20,7 +20,7 @@ namespace paddle {
 namespace operators {
 
 // Out = sum(abs(X))
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class L1NormKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
@@ -30,14 +30,15 @@ class L1NormKernel : public framework::OpKernel<T> {
 
     auto x = framework::EigenVector<T>::Flatten(*X);
     auto out = framework::EigenScalar<T>::From(*Out);
-    auto place = context.GetEigenDevice<Place>();
+    auto &place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     out.device(place) = x.abs().sum();
   }
 };
 
 // dX = dout * sign(X)
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class L1NormGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
@@ -52,7 +53,8 @@ class L1NormGradKernel : public framework::OpKernel<T> {
     auto x_eigen = framework::EigenVector<T>::Flatten(*x);
     auto d_out_eigen = framework::EigenVector<T>::Flatten(*d_out);
     auto dx_eigen = framework::EigenVector<T>::Flatten(*dx);
-    auto place = context.GetEigenDevice<Place>();
+    auto &place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     Eigen::DSizes<int, 1> x_dsize(x->numel());
     dx_eigen.device(place) = d_out_eigen.broadcast(x_dsize) * x_eigen.sign();
diff --git a/paddle/operators/linear_chain_crf_op.cc b/paddle/operators/linear_chain_crf_op.cc
index 8e079a14e0..896e3657d4 100644
--- a/paddle/operators/linear_chain_crf_op.cc
+++ b/paddle/operators/linear_chain_crf_op.cc
@@ -261,9 +261,10 @@ REGISTER_OP(linear_chain_crf, ops::LinearChainCRFOp, ops::LinearChainCRFOpMaker,
             linear_chain_crf_grad, ops::LinearChainCRFGradOp);
 REGISTER_OP_CPU_KERNEL(
     linear_chain_crf,
-    ops::LinearChainCRFOpKernel<paddle::platform::CPUPlace, float>,
-    ops::LinearChainCRFOpKernel<paddle::platform::CPUPlace, double>);
+    ops::LinearChainCRFOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::LinearChainCRFOpKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     linear_chain_crf_grad,
-    ops::LinearChainCRFGradOpKernel<paddle::platform::CPUPlace, float>,
-    ops::LinearChainCRFGradOpKernel<paddle::platform::CPUPlace, double>);
+    ops::LinearChainCRFGradOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::LinearChainCRFGradOpKernel<paddle::platform::CPUDeviceContext,
+                                    double>);
diff --git a/paddle/operators/linear_chain_crf_op.cu b/paddle/operators/linear_chain_crf_op.cu
index 6fc8995f4c..3b105ec341 100644
--- a/paddle/operators/linear_chain_crf_op.cu
+++ b/paddle/operators/linear_chain_crf_op.cu
@@ -16,11 +16,12 @@ limitations under the License. */
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     linear_chain_crf,
-    ops::LinearChainCRFOpKernel<paddle::platform::GPUPlace, float>,
-    ops::LinearChainCRFOpKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
+    ops::LinearChainCRFOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::LinearChainCRFOpKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
     linear_chain_crf_grad,
-    ops::LinearChainCRFGradOpKernel<paddle::platform::GPUPlace, float>,
-    ops::LinearChainCRFGradOpKernel<paddle::platform::GPUPlace, double>);
+    ops::LinearChainCRFGradOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::LinearChainCRFGradOpKernel<paddle::platform::CUDADeviceContext,
+                                    double>);
diff --git a/paddle/operators/linear_chain_crf_op.h b/paddle/operators/linear_chain_crf_op.h
index 014bbfa758..694584e79c 100644
--- a/paddle/operators/linear_chain_crf_op.h
+++ b/paddle/operators/linear_chain_crf_op.h
@@ -50,7 +50,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LinearChainCRFOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -137,7 +137,8 @@ class LinearChainCRFOpKernel : public framework::OpKernel<T> {
         framework::make_ddim({static_cast<int64_t>(batch_size), 1}),
         platform::CPUPlace());
 
-    auto place = ctx.GetEigenDevice<platform::CPUPlace>();
+    auto& place = *ctx.template device_context<platform::CPUDeviceContext>()
+                       .eigen_device();
     auto x = EigenMatrix<T>::From(*emission_weights);
     auto x_row_max = EigenMatrix<T>::From(emission_row_max);
     x_row_max.device(place) =
@@ -287,7 +288,7 @@ class LinearChainCRFOpKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LinearChainCRFGradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -359,8 +360,7 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> {
     emission_grad->mutable_data<T>(platform::CPUPlace());
     if (transition_grad) {
       transition_grad->mutable_data<T>(platform::CPUPlace());
-      math::SetConstant<platform::CPUPlace, T>()(ctx.device_context(),
-                                                 transition_grad, 0.);
+      math::set_constant(ctx.device_context(), transition_grad, 0.);
     }
     // Now, all the inputs and outputs should be on the CPU memory.
 
@@ -384,10 +384,10 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> {
       Tensor one_seq_beta = beta.Slice(start_pos, end_pos);
       Tensor one_seq_emission_grad = emission_grad->Slice(start_pos, end_pos);
 
-      BackwardOneSequence(ctx.device_context(), ll_grad[i],
-                          one_seq_emission_exps, *transition_exps,
-                          one_seq_alpha, one_seq_label, &one_seq_beta,
-                          transition_grad, &one_seq_emission_grad);
+      BackwardOneSequence(
+          ctx.template device_context<platform::CPUDeviceContext>(), ll_grad[i],
+          one_seq_emission_exps, *transition_exps, one_seq_alpha, one_seq_label,
+          &one_seq_beta, transition_grad, &one_seq_emission_grad);
     }
 
     if (platform::is_gpu_place(ctx.GetPlace())) {
@@ -441,8 +441,8 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> {
     copyTensor(ctx, transition_grad_src, transition_grad_dst);
   }
 
-  void BackwardOneSequence(const platform::DeviceContext& ctx, const T ll_grad,
-                           const Tensor& emission_exps,
+  void BackwardOneSequence(const platform::CPUDeviceContext& ctx,
+                           const T ll_grad, const Tensor& emission_exps,
                            const Tensor& transition_exps, const Tensor& alpha,
                            const Tensor& label, Tensor* beta,
                            Tensor* transition_grad,
@@ -481,7 +481,7 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> {
     auto alpha_mat = EigenMatrix<T>::From(alpha);
     auto beta_mat = EigenMatrix<T>::From(*beta);
 
-    auto* place = ctx.GetEigenDevice<platform::CPUPlace>();
+    auto* place = ctx.eigen_device();
     auto prob = alpha_mat * beta_mat;
     auto row_sum = prob.sum(Eigen::DSizes<int, 1>(1))
                        .reshape(Eigen::DSizes<int, 2>(seq_length, 1))
diff --git a/paddle/operators/lod_reset_op.cu b/paddle/operators/lod_reset_op.cu
index 5244a17c3a..f7c2358980 100644
--- a/paddle/operators/lod_reset_op.cu
+++ b/paddle/operators/lod_reset_op.cu
@@ -16,9 +16,10 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(lod_reset,
-                       ops::LoDResetKernel<paddle::platform::GPUPlace, float>,
-                       ops::LoDResetKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
-    lod_reset_grad, ops::LoDResetGradKernel<paddle::platform::GPUPlace, float>,
-    ops::LoDResetGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    lod_reset, ops::LoDResetKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::LoDResetKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    lod_reset_grad,
+    ops::LoDResetGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::LoDResetGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/lod_reset_op.h b/paddle/operators/lod_reset_op.h
index cbcbf80adc..b86f8b1313 100644
--- a/paddle/operators/lod_reset_op.h
+++ b/paddle/operators/lod_reset_op.h
@@ -20,7 +20,7 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LoDResetKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
@@ -65,7 +65,7 @@ class LoDResetKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LoDResetGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
diff --git a/paddle/operators/log_loss_op.cc b/paddle/operators/log_loss_op.cc
index 257e5c8a49..4524229a33 100644
--- a/paddle/operators/log_loss_op.cc
+++ b/paddle/operators/log_loss_op.cc
@@ -109,7 +109,8 @@ class LogLossGradOp : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(log_loss, ops::LogLossOp, ops::LogLossOpMaker<float>, log_loss_grad,
             ops::LogLossGradOp);
-REGISTER_OP_CPU_KERNEL(log_loss,
-                       ops::LogLossKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
-    log_loss_grad, ops::LogLossGradKernel<paddle::platform::CPUPlace, float>);
+    log_loss, ops::LogLossKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    log_loss_grad,
+    ops::LogLossGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/log_loss_op.cu b/paddle/operators/log_loss_op.cu
index 6c189ef341..e87ac7d12a 100644
--- a/paddle/operators/log_loss_op.cu
+++ b/paddle/operators/log_loss_op.cu
@@ -16,7 +16,8 @@
 #include "paddle/operators/log_loss_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(log_loss,
-                       ops::LogLossKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    log_loss_grad, ops::LogLossGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    log_loss, ops::LogLossKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    log_loss_grad,
+    ops::LogLossGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/log_loss_op.h b/paddle/operators/log_loss_op.h
index 73404fce91..743eddb740 100644
--- a/paddle/operators/log_loss_op.h
+++ b/paddle/operators/log_loss_op.h
@@ -24,7 +24,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-template <typename Place, typename T, typename AttrType = T>
+template <typename DeviceContext, typename T, typename AttrType = T>
 class LogLossKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -38,7 +38,7 @@ class LogLossKernel : public framework::OpKernel<T> {
     auto label = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Labels"));
 
     auto loss = EigenVector<T>::Flatten(*loss_out);
-    auto place = ctx.GetEigenDevice<Place>();
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
 
     loss.device(place) = (-(label * (prediction + epsilon).log()) -
                           ((static_cast<T>(1) - label) *
@@ -46,7 +46,7 @@ class LogLossKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T, typename AttrType = T>
+template <typename DeviceContext, typename T, typename AttrType = T>
 class LogLossGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -59,7 +59,7 @@ class LogLossGradKernel : public framework::OpKernel<T> {
     auto* dpred = ctx.Output<Tensor>(framework::GradVarName("Predicted"));
 
     auto dl = EigenVector<T>::Flatten(*dloss);
-    auto place = ctx.GetEigenDevice<Place>();
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
 
     if (dpred) {
       dpred->mutable_data<T>(ctx.GetPlace());
diff --git a/paddle/operators/logical_op.cu b/paddle/operators/logical_op.cu
index d41239b2ca..7fef60e0c9 100644
--- a/paddle/operators/logical_op.cu
+++ b/paddle/operators/logical_op.cu
@@ -14,11 +14,11 @@
 
 #include "paddle/operators/logical_op.h"
 
-REGISTER_BINARY_LOGICAL_KERNEL(logical_and, GPU,
+REGISTER_BINARY_LOGICAL_KERNEL(logical_and, CUDA,
                                paddle::operators::LogicalAndFunctor);
-REGISTER_BINARY_LOGICAL_KERNEL(logical_or, GPU,
+REGISTER_BINARY_LOGICAL_KERNEL(logical_or, CUDA,
                                paddle::operators::LogicalOrFunctor);
-REGISTER_UNARY_LOGICAL_KERNEL(logical_not, GPU,
+REGISTER_UNARY_LOGICAL_KERNEL(logical_not, CUDA,
                               paddle::operators::LogicalNotFunctor);
-REGISTER_BINARY_LOGICAL_KERNEL(logical_xor, GPU,
+REGISTER_BINARY_LOGICAL_KERNEL(logical_xor, CUDA,
                                paddle::operators::LogicalXorFunctor);
diff --git a/paddle/operators/logical_op.h b/paddle/operators/logical_op.h
index 6e78a7d6ed..629388cac8 100644
--- a/paddle/operators/logical_op.h
+++ b/paddle/operators/logical_op.h
@@ -47,7 +47,7 @@ struct LogicalXorFunctor {
   }
 };
 
-template <typename Place, typename Functor>
+template <typename DeviceContext, typename Functor>
 class BinaryLogicalOpKernel
     : public framework::OpKernel<typename Functor::ELEM_TYPE> {
  public:
@@ -57,14 +57,14 @@ class BinaryLogicalOpKernel
     auto* y = context.Input<framework::Tensor>("Y");
     auto* out = context.Output<framework::Tensor>("Out");
     Functor binary_func;
-    platform::Transform<Place> trans;
-    trans(context.device_context(), x->data<T>(), x->data<T>() + x->numel(),
-          y->data<T>(), out->mutable_data<bool>(context.GetPlace()),
-          binary_func);
+    platform::Transform<DeviceContext> trans;
+    trans(context.template device_context<DeviceContext>(), x->data<T>(),
+          x->data<T>() + x->numel(), y->data<T>(),
+          out->mutable_data<bool>(context.GetPlace()), binary_func);
   }
 };
 
-template <typename Place, typename Functor>
+template <typename DeviceContext, typename Functor>
 class UnaryLogicalOpKernel
     : public framework::OpKernel<typename Functor::ELEM_TYPE> {
  public:
@@ -73,8 +73,9 @@ class UnaryLogicalOpKernel
     auto* x = context.Input<framework::Tensor>("X");
     auto* out = context.Output<framework::Tensor>("Out");
     Functor unary_func;
-    platform::Transform<Place> trans;
-    trans(context.device_context(), x->data<T>(), x->data<T>() + x->numel(),
+    platform::Transform<DeviceContext> trans;
+    trans(context.template device_context<DeviceContext>(), x->data<T>(),
+          x->data<T>() + x->numel(),
           out->mutable_data<bool>(context.GetPlace()), unary_func);
   }
 };
@@ -85,9 +86,9 @@ class UnaryLogicalOpKernel
 #define REGISTER_BINARY_LOGICAL_KERNEL(op_type, dev, functor) \
   REGISTER_OP_##dev##_KERNEL(                                 \
       op_type, ::paddle::operators::BinaryLogicalOpKernel<    \
-                   ::paddle::platform::dev##Place, functor<bool>>);
+                   ::paddle::platform::dev##DeviceContext, functor<bool>>);
 
 #define REGISTER_UNARY_LOGICAL_KERNEL(op_type, dev, functor) \
   REGISTER_OP_##dev##_KERNEL(                                \
       op_type, ::paddle::operators::UnaryLogicalOpKernel<    \
-                   ::paddle::platform::dev##Place, functor<bool>>);
+                   ::paddle::platform::dev##DeviceContext, functor<bool>>);
diff --git a/paddle/operators/lookup_table_op.cu b/paddle/operators/lookup_table_op.cu
index 84b044184a..9431030a53 100644
--- a/paddle/operators/lookup_table_op.cu
+++ b/paddle/operators/lookup_table_op.cu
@@ -85,6 +85,8 @@ template <typename T>
 class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
+    auto& dev_ctx =
+        context.template device_context<platform::CUDADeviceContext>();
     bool is_sparse = context.Attr<bool>("is_sparse");
     if (is_sparse) {
       auto* ids = context.Input<LoDTensor>("Ids");
@@ -95,7 +97,7 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
       auto* ids_data = ids->data<int64_t>();
       auto ids_dim = ids->dims();
 
-      auto stream = context.cuda_device_context().stream();
+      auto stream = dev_ctx.stream();
       // copy GPU memory to CPU pinned memory
       framework::Vector<int64_t> new_rows;
       new_rows.resize(ids_dim[0]);
@@ -129,14 +131,11 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
       T* d_table = d_table_t->mutable_data<T>(context.GetPlace());
 
       auto t = framework::EigenVector<T>::Flatten(*d_table_t);
-      t.device(context.GetEigenDevice<platform::GPUPlace>()) =
-          t.constant(static_cast<T>(0));
+      t.device(*dev_ctx.eigen_device()) = t.constant(static_cast<T>(0));
 
       dim3 threads(128, 8);
       dim3 grids(8, 1);
-      LookupTableGrad<
-          T, 128, 8,
-          8><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
+      LookupTableGrad<T, 128, 8, 8><<<grids, threads, 0, dev_ctx.stream()>>>(
           d_table, d_output, ids, N, K, D);
     }
   }
@@ -146,7 +145,8 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(lookup_table, ops::LookupTableCUDAKernel<float>,
-                       ops::LookupTableCUDAKernel<double>);
-REGISTER_OP_GPU_KERNEL(lookup_table_grad, ops::LookupTableGradCUDAKernel<float>,
-                       ops::LookupTableGradCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(lookup_table, ops::LookupTableCUDAKernel<float>,
+                        ops::LookupTableCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(lookup_table_grad,
+                        ops::LookupTableGradCUDAKernel<float>,
+                        ops::LookupTableGradCUDAKernel<double>);
diff --git a/paddle/operators/lrn_op.cc b/paddle/operators/lrn_op.cc
index e20340e77b..b5b7bc940a 100644
--- a/paddle/operators/lrn_op.cc
+++ b/paddle/operators/lrn_op.cc
@@ -20,7 +20,7 @@ namespace operators {
 using framework::Tensor;
 
 template <typename T>
-struct LRNFunctor<platform::CPUPlace, T> {
+struct LRNFunctor<platform::CPUDeviceContext, T> {
   void operator()(const framework::ExecutionContext& ctx,
                   const framework::Tensor& input, framework::Tensor* out,
                   framework::Tensor* mid, int N, int C, int H, int W, int n,
@@ -55,11 +55,11 @@ struct LRNFunctor<platform::CPUPlace, T> {
     out_e = x_v * e_mid.reshape(Eigen::DSizes<int, 1>(e_mid.size())).pow(-beta);
   }
 };
-template struct LRNFunctor<platform::CPUPlace, float>;
-template struct LRNFunctor<platform::CPUPlace, double>;
+template struct LRNFunctor<platform::CPUDeviceContext, float>;
+template struct LRNFunctor<platform::CPUDeviceContext, double>;
 
 template <typename T>
-struct LRNGradFunctor<platform::CPUPlace, T> {
+struct LRNGradFunctor<platform::CPUDeviceContext, T> {
   void operator()(const framework::ExecutionContext& ctx,
                   const framework::Tensor& x, const framework::Tensor& out,
                   const framework::Tensor& mid, framework::Tensor* x_g,
@@ -113,8 +113,8 @@ struct LRNGradFunctor<platform::CPUPlace, T> {
     }
   }
 };
-template struct LRNGradFunctor<platform::CPUPlace, float>;
-template struct LRNGradFunctor<platform::CPUPlace, double>;
+template struct LRNGradFunctor<platform::CPUDeviceContext, float>;
+template struct LRNGradFunctor<platform::CPUDeviceContext, double>;
 
 class LRNOp : public framework::OperatorWithKernel {
  public:
@@ -204,7 +204,7 @@ Input(i, x, y), Output(i, x, y) represents an element in an image.
 C is the number of feature maps of one image. n is a hyper-parameter
 configured when operator is initialized. The sum in the denominator
 is the sum of the same positions in the neighboring maps.
-    
+
 )DOC");
   }
 };
@@ -230,6 +230,7 @@ class LRNOpGrad : public framework::OperatorWithKernel {
 
 namespace ops = paddle::operators;
 REGISTER_OP(lrn, ops::LRNOp, ops::LRNOpMaker<float>, lrn_grad, ops::LRNOpGrad);
-REGISTER_OP_CPU_KERNEL(lrn, ops::LRNKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(lrn_grad,
-                       ops::LRNGradKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    lrn, ops::LRNKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    lrn_grad, ops::LRNGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/lrn_op.cu b/paddle/operators/lrn_op.cu
index e9a8671233..c6857c2b6d 100644
--- a/paddle/operators/lrn_op.cu
+++ b/paddle/operators/lrn_op.cu
@@ -69,19 +69,18 @@ void CrossMapNormal(const framework::ExecutionContext& ctx, const T* inputs,
   const int block_size = 1024;
   int grid_size = (img_size + block_size - 1) / block_size;
 
-  KeCMRNormFillScale<
-      T><<<grid_size, block_size, 0, ctx.cuda_device_context().stream()>>>(
+  auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+  KeCMRNormFillScale<T><<<grid_size, block_size, 0, dev_ctx.stream()>>>(
       img_size, inputs, mid, C, H, W, n, k, alpha);
 
   int input_size = N * H * W * C;
   grid_size = (input_size + block_size - 1) / block_size;
-  KeCMRNormOutput<
-      T><<<grid_size, block_size, 0, ctx.cuda_device_context().stream()>>>(
+  KeCMRNormOutput<T><<<grid_size, block_size, 0, dev_ctx.stream()>>>(
       input_size, inputs, mid, -beta, outputs);
 }
 
 template <typename T>
-struct LRNFunctor<platform::GPUPlace, T> {
+struct LRNFunctor<platform::CUDADeviceContext, T> {
   void operator()(const framework::ExecutionContext& ctx,
                   const framework::Tensor& input, framework::Tensor* out,
                   framework::Tensor* mid, int N, int C, int H, int W, int n,
@@ -92,8 +91,8 @@ struct LRNFunctor<platform::GPUPlace, T> {
   }
 };
 
-template struct LRNFunctor<platform::GPUPlace, float>;
-template struct LRNFunctor<platform::GPUPlace, double>;
+template struct LRNFunctor<platform::CUDADeviceContext, float>;
+template struct LRNFunctor<platform::CUDADeviceContext, double>;
 
 template <typename T>
 __global__ void KeCMRNormDiff(int img_size, const T* x, const T* out,
@@ -148,14 +147,14 @@ void CrossMapNormalGrad(const framework::ExecutionContext& ctx, const T* x,
   const int block_size = 1024;
   int grid_size = (img_size + block_size - 1) / block_size;
 
-  KeCMRNormDiff<
-      T><<<grid_size, block_size, 0, ctx.cuda_device_context().stream()>>>(
+  auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+  KeCMRNormDiff<T><<<grid_size, block_size, 0, dev_ctx.stream()>>>(
       img_size, x, out, mid, x_g, out_g, C, H, W, n, -beta,
       2.0f * alpha * beta);
 }
 
 template <typename T>
-struct LRNGradFunctor<platform::GPUPlace, T> {
+struct LRNGradFunctor<platform::CUDADeviceContext, T> {
   void operator()(const framework::ExecutionContext& ctx,
                   const framework::Tensor& x, const framework::Tensor& out,
                   const framework::Tensor& mid, framework::Tensor* x_g,
@@ -167,12 +166,13 @@ struct LRNGradFunctor<platform::GPUPlace, T> {
   }
 };
 
-template struct LRNGradFunctor<platform::GPUPlace, float>;
-template struct LRNGradFunctor<platform::GPUPlace, double>;
+template struct LRNGradFunctor<platform::CUDADeviceContext, float>;
+template struct LRNGradFunctor<platform::CUDADeviceContext, double>;
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(lrn, ops::LRNKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(lrn_grad,
-                       ops::LRNGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    lrn, ops::LRNKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    lrn_grad, ops::LRNGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/lrn_op.h b/paddle/operators/lrn_op.h
index aa7539db4a..44063d3e03 100644
--- a/paddle/operators/lrn_op.h
+++ b/paddle/operators/lrn_op.h
@@ -29,7 +29,7 @@ struct LRNFunctor {
                   T k, T alpha, T beta);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LRNKernel : public framework::OpKernel<T> {
  public:
   using Tensor = framework::Tensor;
@@ -65,12 +65,12 @@ class LRNKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE(beta >= 0.0, "beta should >= 0.0");
     PADDLE_ENFORCE(k >= 0.0, "k should >= 0.0");
 
-    LRNFunctor<Place, T> f;
+    LRNFunctor<DeviceContext, T> f;
     f(ctx, x, out, mid, N, C, H, W, n, k, alpha, beta);
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct LRNGradFunctor {
   void operator()(const framework::ExecutionContext& ctx,
                   const framework::Tensor& x, const framework::Tensor& out,
@@ -98,7 +98,7 @@ struct LRNGradFunctor {
  * The upper and lower is the same as forward. The logic of the sum
  * is also the same as forward.
  */
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LRNGradKernel : public framework::OpKernel<T> {
  public:
   using Tensor = framework::Tensor;
@@ -121,7 +121,7 @@ class LRNGradKernel : public framework::OpKernel<T> {
     T alpha = ctx.Attr<T>("alpha");
     T beta = ctx.Attr<T>("beta");
 
-    LRNGradFunctor<Place, T> f;
+    LRNGradFunctor<DeviceContext, T> f;
     f(ctx, x, out, mid, x_g, out_g, N, C, H, W, n, alpha, beta);
   }
 };
diff --git a/paddle/operators/lstm_op.cc b/paddle/operators/lstm_op.cc
index fa8e5f2da8..2db7da30db 100644
--- a/paddle/operators/lstm_op.cc
+++ b/paddle/operators/lstm_op.cc
@@ -273,8 +273,9 @@ class LSTMGradOp : public framework::OperatorWithKernel {
 
 namespace ops = paddle::operators;
 REGISTER_OP(lstm, ops::LSTMOp, ops::LSTMOpMaker, lstm_grad, ops::LSTMGradOp);
-REGISTER_OP_CPU_KERNEL(lstm, ops::LSTMKernel<paddle::platform::CPUPlace, float>,
-                       ops::LSTMKernel<paddle::platform::CPUPlace, double>);
-REGISTER_OP_CPU_KERNEL(lstm_grad,
-                       ops::LSTMGradKernel<paddle::platform::CPUPlace, float>,
-                       ops::LSTMGradKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(
+    lstm, ops::LSTMKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::LSTMKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    lstm_grad, ops::LSTMGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::LSTMGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/lstm_op.cu.cc b/paddle/operators/lstm_op.cu.cc
index 610cbb03e8..48519bed6f 100644
--- a/paddle/operators/lstm_op.cu.cc
+++ b/paddle/operators/lstm_op.cu.cc
@@ -15,8 +15,9 @@
 #include "paddle/operators/lstm_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(lstm, ops::LSTMKernel<paddle::platform::GPUPlace, float>,
-                       ops::LSTMKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(lstm_grad,
-                       ops::LSTMGradKernel<paddle::platform::GPUPlace, float>,
-                       ops::LSTMGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    lstm, ops::LSTMKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::LSTMKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    lstm_grad, ops::LSTMGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::LSTMGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/lstm_op.h b/paddle/operators/lstm_op.h
index a78f548aaf..14abd4bf0a 100644
--- a/paddle/operators/lstm_op.h
+++ b/paddle/operators/lstm_op.h
@@ -24,16 +24,16 @@ namespace operators {
 using LoDTensor = framework::LoDTensor;
 using Tensor = framework::Tensor;
 
-template <typename Place, typename T>
-inline void ReorderInitState(const platform::DeviceContext& ctx,
+template <typename DeviceContext, typename T>
+inline void ReorderInitState(const DeviceContext& ctx,
                              const framework::Tensor& src, const size_t* index,
                              framework::Tensor* dst, bool indexed_src) {
-  math::CopyMatrixRowsFunctor<Place, T> row_shuffle;
+  math::CopyMatrixRowsFunctor<DeviceContext, T> row_shuffle;
   dst->mutable_data<T>(src.dims(), ctx.GetPlace());
   row_shuffle(ctx, src, index, *dst, indexed_src);
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LSTMKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -52,8 +52,8 @@ class LSTMKernel : public framework::OpKernel<T> {
     cell_out->mutable_data<T>(ctx.GetPlace());
 
     bool is_reverse = ctx.Attr<bool>("is_reverse");
-    math::LoDTensor2BatchFunctor<Place, T> to_batch;
-    auto& device_ctx = ctx.device_context();
+    math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
+    auto& device_ctx = ctx.template device_context<DeviceContext>();
     to_batch(device_ctx, *input, *batch_gate, true, is_reverse);
 
     auto in_dims = input->dims();
@@ -64,7 +64,7 @@ class LSTMKernel : public framework::OpKernel<T> {
       Tensor b = *bias;
       b.Resize({bias->numel(), 1});
       Tensor gate_bias = b.Slice(0, 4 * frame_size);
-      math::RowwiseAdd<Place, T> add_bias;
+      math::RowwiseAdd<DeviceContext, T> add_bias;
       add_bias(device_ctx, *batch_gate, gate_bias, batch_gate);
     }
 
@@ -88,8 +88,8 @@ class LSTMKernel : public framework::OpKernel<T> {
       // Since the batch computing for LSTM reorders the input sequence
       // according to their length. The initialized cell state also needs
       // to reorder.
-      ReorderInitState<Place, T>(device_ctx, *cell_t0, order, &ordered_c0,
-                                 true);
+      ReorderInitState<DeviceContext, T>(device_ctx, *cell_t0, order,
+                                         &ordered_c0, true);
       lstm_value.prev_state_value = ordered_c0.data<T>();
     }
 
@@ -121,9 +121,9 @@ class LSTMKernel : public framework::OpKernel<T> {
         int pre_h_start = static_cast<int>(batch_starts[n - 1]);
         int pre_h_end = pre_h_start + cur_batch_size;
         auto pre_hidden_t = batch_hidden.Slice(pre_h_start, pre_h_end);
-        math::matmul<Place, T>(device_ctx, pre_hidden_t, false, *weight, false,
-                               static_cast<T>(1.0), &gate_t,
-                               static_cast<T>(1.0));
+        math::matmul<DeviceContext, T>(device_ctx, pre_hidden_t, false, *weight,
+                                       false, static_cast<T>(1.0), &gate_t,
+                                       static_cast<T>(1.0));
       } else if (hidden_t0) {
         // If n == 0 and there is no initialized hidden state, that is to say
         // the H0 is zeros, the calculation W_h * H0 will be skiped.
@@ -133,24 +133,24 @@ class LSTMKernel : public framework::OpKernel<T> {
         // according to their length. The initialized hidden state also needs
         // to reorder.
         Tensor ordered_h0;
-        ReorderInitState<Place, T>(device_ctx, *hidden_t0, order, &ordered_h0,
-                                   true);
-        math::matmul<Place, T>(device_ctx, ordered_h0, false, *weight, false,
-                               static_cast<T>(1.0), &gate_t,
-                               static_cast<T>(1.0));
+        ReorderInitState<DeviceContext, T>(device_ctx, *hidden_t0, order,
+                                           &ordered_h0, true);
+        math::matmul<DeviceContext, T>(device_ctx, ordered_h0, false, *weight,
+                                       false, static_cast<T>(1.0), &gate_t,
+                                       static_cast<T>(1.0));
       }
 
       lstm_value.gate_value = gate_t.data<T>();
       lstm_value.output_value = out_t.data<T>();
       lstm_value.state_value = cell_t.data<T>();
       lstm_value.state_active_value = cell_pre_act_t.data<T>();
-      math::LstmUnitFunctor<Place, T>::compute(device_ctx, lstm_value,
-                                               frame_size, cur_batch_size,
-                                               gate_act, cell_act, cand_act);
+      math::LstmUnitFunctor<DeviceContext, T>::compute(
+          device_ctx, lstm_value, frame_size, cur_batch_size, gate_act,
+          cell_act, cand_act);
       lstm_value.prev_state_value = lstm_value.state_value;
     }
 
-    math::Batch2LoDTensorFunctor<Place, T> to_seq;
+    math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
     batch_hidden.set_lod(batch_gate->lod());
     // restore the output hidden in LoDTensor from the batch hidden
     to_seq(device_ctx, batch_hidden, *hidden_out);
@@ -161,7 +161,7 @@ class LSTMKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LSTMGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -187,8 +187,8 @@ class LSTMGradKernel : public framework::OpKernel<T> {
     auto* h0_g = ctx.Output<Tensor>(framework::GradVarName("H0"));
     auto* c0_g = ctx.Output<Tensor>(framework::GradVarName("C0"));
 
-    auto& device_ctx = ctx.device_context();
-    math::SetConstant<Place, T> zero;
+    auto& device_ctx = ctx.template device_context<DeviceContext>();
+    math::SetConstant<DeviceContext, T> zero;
     if (weight_g) {
       weight_g->mutable_data<T>(ctx.GetPlace());
       zero(device_ctx, weight_g, static_cast<T>(0.0));
@@ -200,7 +200,8 @@ class LSTMGradKernel : public framework::OpKernel<T> {
     Tensor ordered_h0, ordered_c0, ordered_h0_g, ordered_c0_g;
     const size_t* order = batch_gate->lod()[2].data();
     if (c0) {
-      ReorderInitState<Place, T>(device_ctx, *c0, order, &ordered_c0, true);
+      ReorderInitState<DeviceContext, T>(device_ctx, *c0, order, &ordered_c0,
+                                         true);
     }
     if (c0 && c0_g) {
       ordered_c0_g.mutable_data<T>(c0_g->dims(), ctx.GetPlace());
@@ -240,10 +241,10 @@ class LSTMGradKernel : public framework::OpKernel<T> {
       lstm_grad.check_og_grad = nullptr;
     }
 
-    math::LoDTensor2BatchFunctor<Place, T> to_batch;
+    math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
 
     auto ToBatch = [&batch_gate, &to_batch](
-        const platform::DeviceContext& ctx, const framework::LoDTensor& src,
+        const DeviceContext& ctx, const framework::LoDTensor& src,
         const framework::DDim& dims, framework::LoDTensor& dst) {
       dst.mutable_data<T>(dims, ctx.GetPlace());
       dst.set_lod(batch_gate->lod());
@@ -299,7 +300,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
       }
 
       int cur_batch_size = bend - bstart;
-      math::LstmUnitGradFunctor<Place, T>::compute(
+      math::LstmUnitGradFunctor<DeviceContext, T>::compute(
           device_ctx, lstm_value, lstm_grad, frame_size, cur_batch_size,
           gate_act, cell_act, cand_act);
 
@@ -307,33 +308,34 @@ class LSTMGradKernel : public framework::OpKernel<T> {
         int pre_h_start = static_cast<int>(batch_starts[n - 1]);
         int pre_h_end = pre_h_start + cur_batch_size;
         auto pre_hidden_g = batch_hidden_g.Slice(pre_h_start, pre_h_end);
-        math::matmul<Place, T>(device_ctx, gate_g, false, *weight, true,
-                               static_cast<T>(1.0), &pre_hidden_g,
-                               static_cast<T>(1.0));
+        math::matmul<DeviceContext, T>(device_ctx, gate_g, false, *weight, true,
+                                       static_cast<T>(1.0), &pre_hidden_g,
+                                       static_cast<T>(1.0));
         if (weight_g) {
           /* backward weight */
           auto pre_hidden = batch_hidden.Slice(pre_h_start, pre_h_end);
-          math::matmul<Place, T>(device_ctx, pre_hidden, true, gate_g, false,
-                                 static_cast<T>(1.0), weight_g,
-                                 static_cast<T>(1.0));
+          math::matmul<DeviceContext, T>(device_ctx, pre_hidden, true, gate_g,
+                                         false, static_cast<T>(1.0), weight_g,
+                                         static_cast<T>(1.0));
         }
       } else {
         if (h0 && weight_g) {
-          ReorderInitState<Place, T>(device_ctx, *h0, order, &ordered_h0, true);
-          math::matmul<Place, T>(device_ctx, ordered_h0, true, gate_g, false,
-                                 static_cast<T>(1.0), weight_g,
-                                 static_cast<T>(1.0));
+          ReorderInitState<DeviceContext, T>(device_ctx, *h0, order,
+                                             &ordered_h0, true);
+          math::matmul<DeviceContext, T>(device_ctx, ordered_h0, true, gate_g,
+                                         false, static_cast<T>(1.0), weight_g,
+                                         static_cast<T>(1.0));
         }
         if (h0 && h0_g) {
           ordered_h0_g.mutable_data<T>(h0_g->dims(), ctx.GetPlace());
-          math::matmul<Place, T>(device_ctx, gate_g, false, *weight, true,
-                                 static_cast<T>(1.0), &ordered_h0_g,
-                                 static_cast<T>(0.0));
+          math::matmul<DeviceContext, T>(device_ctx, gate_g, false, *weight,
+                                         true, static_cast<T>(1.0),
+                                         &ordered_h0_g, static_cast<T>(0.0));
         }
       }
     }
 
-    math::Batch2LoDTensorFunctor<Place, T> to_seq;
+    math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
     if (in_g) {
       /* backward data */
       in_g->mutable_data<T>(ctx.GetPlace());
@@ -344,15 +346,17 @@ class LSTMGradKernel : public framework::OpKernel<T> {
       Tensor b_g = *bias_g;
       b_g.Resize({bias_g->numel(), 1});
       Tensor gate_bias_g = b_g.Slice(0, 4 * frame_size);
-      math::ColwiseSum<Place, T> col_sum;
+      math::ColwiseSum<DeviceContext, T> col_sum;
       col_sum(device_ctx, batch_gate_g, &gate_bias_g);
     }
 
     if (h0 && h0_g) {
-      ReorderInitState<Place, T>(device_ctx, ordered_h0_g, order, h0_g, false);
+      ReorderInitState<DeviceContext, T>(device_ctx, ordered_h0_g, order, h0_g,
+                                         false);
     }
     if (c0 && c0_g) {
-      ReorderInitState<Place, T>(device_ctx, ordered_c0_g, order, c0_g, false);
+      ReorderInitState<DeviceContext, T>(device_ctx, ordered_c0_g, order, c0_g,
+                                         false);
     }
   }
 };
diff --git a/paddle/operators/lstm_unit_op.cu b/paddle/operators/lstm_unit_op.cu
index e192283aa0..291f2c295e 100644
--- a/paddle/operators/lstm_unit_op.cu
+++ b/paddle/operators/lstm_unit_op.cu
@@ -173,7 +173,7 @@ class LstmUnitGradOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(lstm_unit, ops::LstmUnitOpCUDAKernel<float>,
-                       ops::LstmUnitOpCUDAKernel<double>);
-REGISTER_OP_GPU_KERNEL(lstm_unit_grad, ops::LstmUnitGradOpCUDAKernel<float>,
-                       ops::LstmUnitGradOpCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(lstm_unit, ops::LstmUnitOpCUDAKernel<float>,
+                        ops::LstmUnitOpCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(lstm_unit_grad, ops::LstmUnitGradOpCUDAKernel<float>,
+                        ops::LstmUnitGradOpCUDAKernel<double>);
diff --git a/paddle/operators/lstm_unit_op.h b/paddle/operators/lstm_unit_op.h
index 38cb298f92..61705675d9 100644
--- a/paddle/operators/lstm_unit_op.h
+++ b/paddle/operators/lstm_unit_op.h
@@ -35,7 +35,7 @@ inline T tanh(T x) {
   return 2. * sigmoid(2. * x) - 1.;
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LstmUnitKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -78,7 +78,7 @@ class LstmUnitKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LstmUnitGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
diff --git a/paddle/operators/margin_rank_loss_op.cc b/paddle/operators/margin_rank_loss_op.cc
index d7e8a0ea76..42e8961c0e 100644
--- a/paddle/operators/margin_rank_loss_op.cc
+++ b/paddle/operators/margin_rank_loss_op.cc
@@ -117,7 +117,7 @@ REGISTER_OP(margin_rank_loss, ops::MarginRankLossOp,
             ops::MarginRankLossGradOp);
 REGISTER_OP_CPU_KERNEL(
     margin_rank_loss,
-    ops::MarginRankLossKernel<paddle::platform::CPUPlace, float>);
+    ops::MarginRankLossKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     margin_rank_loss_grad,
-    ops::MarginRankLossGradKernel<paddle::platform::CPUPlace, float>);
+    ops::MarginRankLossGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/margin_rank_loss_op.cu b/paddle/operators/margin_rank_loss_op.cu
index 3a639f25d4..1c2afccc5b 100644
--- a/paddle/operators/margin_rank_loss_op.cu
+++ b/paddle/operators/margin_rank_loss_op.cu
@@ -16,9 +16,9 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     margin_rank_loss,
-    ops::MarginRankLossKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+    ops::MarginRankLossKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     margin_rank_loss_grad,
-    ops::MarginRankLossGradKernel<paddle::platform::GPUPlace, float>);
+    ops::MarginRankLossGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/margin_rank_loss_op.h b/paddle/operators/margin_rank_loss_op.h
index 8d0830147e..9c1f96cac1 100644
--- a/paddle/operators/margin_rank_loss_op.h
+++ b/paddle/operators/margin_rank_loss_op.h
@@ -34,7 +34,7 @@ struct Heaviside {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MarginRankLossKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
@@ -56,13 +56,13 @@ class MarginRankLossKernel : public framework::OpKernel<T> {
     auto x1 = framework::EigenVector<T>::Flatten(*x1_t);
     auto x2 = framework::EigenVector<T>::Flatten(*x2_t);
 
-    auto& dev = ctx.GetEigenDevice<Place>();
+    auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
     out.device(dev) = (-label * (x1 - x2) + margin).unaryExpr(ReLU<T>());
     act.device(dev) = out.unaryExpr(Heaviside<T>());
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MarginRankLossGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
@@ -78,7 +78,7 @@ class MarginRankLossGradKernel : public framework::OpKernel<T> {
     auto d_out = framework::EigenVector<T>::Flatten(*d_out_t);
     auto act = framework::EigenVector<T>::Flatten(*act_t);
     auto label = framework::EigenVector<T>::Flatten(*label_t);
-    auto& dev = ctx.GetEigenDevice<Place>();
+    auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
 
     // compute d_x1
     if (d_x1_t) {
diff --git a/paddle/operators/math/context_project.cc b/paddle/operators/math/context_project.cc
index f82ea5d7be..980dd90df8 100644
--- a/paddle/operators/math/context_project.cc
+++ b/paddle/operators/math/context_project.cc
@@ -18,8 +18,8 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template class ContextProjectFunctor<platform::CPUPlace, float>;
-template class ContextProjectFunctor<platform::CPUPlace, double>;
+template class ContextProjectFunctor<platform::CPUDeviceContext, float>;
+template class ContextProjectFunctor<platform::CPUDeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/context_project.cu b/paddle/operators/math/context_project.cu
index 04eeed543c..934e3df645 100644
--- a/paddle/operators/math/context_project.cu
+++ b/paddle/operators/math/context_project.cu
@@ -20,8 +20,8 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template class ContextProjectFunctor<platform::GPUPlace, float>;
-template class ContextProjectFunctor<platform::GPUPlace, double>;
+template class ContextProjectFunctor<platform::CUDADeviceContext, float>;
+template class ContextProjectFunctor<platform::CUDADeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/context_project.h b/paddle/operators/math/context_project.h
index d853507188..4036614086 100644
--- a/paddle/operators/math/context_project.h
+++ b/paddle/operators/math/context_project.h
@@ -81,17 +81,17 @@ using LoDTensor = framework::LoDTensor;
  *
  */
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ContextProjectFunctor {
  public:
-  void operator()(const platform::DeviceContext& context, const LoDTensor& in,
+  void operator()(const DeviceContext& context, const LoDTensor& in,
                   const Tensor& padding_data, bool padding_trainable,
                   const int context_start, const int context_length,
                   const int context_stride, const int up_pad,
                   const int down_pad, Tensor* col) {
     auto lod_level_0 = in.lod()[0];
 
-    math::Im2ColFunctor<math::ColFormat::kOCF, Place, float> im2col_ocf;
+    math::Im2ColFunctor<math::ColFormat::kOCF, DeviceContext, float> im2col_ocf;
 
     std::vector<int> dilation({1, 1});
     std::vector<int> padding({up_pad, 0, down_pad, 0});
@@ -188,17 +188,17 @@ class ContextProjectFunctor {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ContextProjectGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context, const LoDTensor& in,
+  void operator()(const DeviceContext& context, const LoDTensor& in,
                   bool padding_trainable, const int context_start,
                   const int context_length, const int context_stride,
                   const int up_pad, const int down_pad, bool pad_grad,
                   bool input_grad, Tensor* padding_data, Tensor* col) {
     auto lod_level_0 = in.lod()[0];
 
-    math::Col2ImFunctor<math::ColFormat::kOCF, Place, float> col2im_ocf;
+    math::Col2ImFunctor<math::ColFormat::kOCF, DeviceContext, float> col2im_ocf;
 
     std::vector<int> dilation({1, 1});
     std::vector<int> padding({up_pad, 0, down_pad, 0});
@@ -258,8 +258,8 @@ class ContextProjectGradFunctor {
               Tensor out_t_sub = out_t.Slice(k * context_length,
                                              k * context_length + padding_size);
               Tensor w_sub = padding_data->Slice(k, k + padding_size);
-              axpy<Place, T>(context, w_sub.numel(), static_cast<T>(1),
-                             out_t_sub.data<T>(), w_sub.data<T>());
+              axpy<DeviceContext, T>(context, w_sub.numel(), static_cast<T>(1),
+                                     out_t_sub.data<T>(), w_sub.data<T>());
             }
           }
           if (down_pad > 0) {
@@ -290,8 +290,8 @@ class ContextProjectGradFunctor {
                   (down_pad_begin_row + t) * context_length);
               Tensor w_sub = padding_data->Slice(
                   up_pad + padding_idx, up_pad + padding_idx + padding_size);
-              axpy<Place, T>(context, w_sub.numel(), static_cast<T>(1),
-                             out_t_sub.data<T>(), w_sub.data<T>());
+              axpy<DeviceContext, T>(context, w_sub.numel(), static_cast<T>(1),
+                                     out_t_sub.data<T>(), w_sub.data<T>());
             }
           }
           out_t.Resize({sequence_height, context_length * sequence_width});
diff --git a/paddle/operators/math/cross_entropy.cc b/paddle/operators/math/cross_entropy.cc
index cf238a58e0..6011a196d4 100644
--- a/paddle/operators/math/cross_entropy.cc
+++ b/paddle/operators/math/cross_entropy.cc
@@ -24,9 +24,9 @@ template <typename T, int MajorType = Eigen::RowMajor,
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
 template <typename T>
-class CrossEntropyFunctor<platform::CPUPlace, T> {
+class CrossEntropyFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& ctx, framework::Tensor* out,
+  void operator()(const platform::CPUDeviceContext& ctx, framework::Tensor* out,
                   const framework::Tensor* prob,
                   const framework::Tensor* labels, const bool softLabel) {
     const int batch_size = prob->dims()[0];
@@ -35,7 +35,7 @@ class CrossEntropyFunctor<platform::CPUPlace, T> {
       auto lbl = EigenMatrix<T>::From(*labels);
       auto loss = EigenMatrix<T>::From(*out);
 
-      loss.device(*ctx.GetEigenDevice<platform::CPUPlace>()) =
+      loss.device(*ctx.eigen_device()) =
           -((lbl * in.log().unaryExpr(math::TolerableValue<T>()))
                 .sum(Eigen::DSizes<int, 1>(1))
                 .reshape(Eigen::DSizes<int, 2>(batch_size, 1)));
@@ -53,8 +53,8 @@ class CrossEntropyFunctor<platform::CPUPlace, T> {
   }
 };
 
-template class CrossEntropyFunctor<platform::CPUPlace, float>;
-template class CrossEntropyFunctor<platform::CPUPlace, double>;
+template class CrossEntropyFunctor<platform::CPUDeviceContext, float>;
+template class CrossEntropyFunctor<platform::CPUDeviceContext, double>;
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/math/cross_entropy.cu b/paddle/operators/math/cross_entropy.cu
index 651c08f740..2132d49c93 100644
--- a/paddle/operators/math/cross_entropy.cu
+++ b/paddle/operators/math/cross_entropy.cu
@@ -95,10 +95,10 @@ __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label,
 using Tensor = framework::Tensor;
 
 template <typename T>
-class CrossEntropyFunctor<platform::GPUPlace, T> {
+class CrossEntropyFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& ctx, framework::Tensor* out,
-                  const framework::Tensor* prob,
+  void operator()(const platform::CUDADeviceContext& ctx,
+                  framework::Tensor* out, const framework::Tensor* prob,
                   const framework::Tensor* labels, bool softLabel) {
     const T* prob_data = prob->data<T>();
     T* loss_data = out->mutable_data<T>(ctx.GetPlace());
@@ -118,16 +118,14 @@ class CrossEntropyFunctor<platform::GPUPlace, T> {
       const int64_t* label_data = labels->data<int64_t>();
       int block = 512;
       int grid = (batch_size + block - 1) / block;
-      CrossEntropyKernel<T><<<
-          grid, block, 0,
-          reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream()>>>(
+      CrossEntropyKernel<T><<<grid, block, 0, ctx.stream()>>>(
           loss_data, prob_data, label_data, batch_size, class_num);
     }
   }
 };
 
-template class CrossEntropyFunctor<platform::GPUPlace, float>;
-template class CrossEntropyFunctor<platform::GPUPlace, double>;
+template class CrossEntropyFunctor<platform::CUDADeviceContext, float>;
+template class CrossEntropyFunctor<platform::CUDADeviceContext, double>;
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/math/cross_entropy.h b/paddle/operators/math/cross_entropy.h
index 70ed9ddd55..677adb5ada 100644
--- a/paddle/operators/math/cross_entropy.h
+++ b/paddle/operators/math/cross_entropy.h
@@ -33,11 +33,11 @@ struct TolerableValue {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class CrossEntropyFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  framework::Tensor* out, const framework::Tensor* prob,
+  void operator()(const DeviceContext& context, framework::Tensor* out,
+                  const framework::Tensor* prob,
                   const framework::Tensor* labels, const bool softLabel);
 };
 }  // namespace math
diff --git a/paddle/operators/math/gru_compute.cc b/paddle/operators/math/gru_compute.cc
index ae4e47b014..d570c68cd4 100644
--- a/paddle/operators/math/gru_compute.cc
+++ b/paddle/operators/math/gru_compute.cc
@@ -19,14 +19,14 @@ namespace operators {
 namespace math {
 
 template <typename T>
-struct GRUUnitFunctor<platform::CPUPlace, T> {
-  static void compute(const platform::DeviceContext &context,
+struct GRUUnitFunctor<platform::CPUDeviceContext, T> {
+  static void compute(const platform::CPUDeviceContext &context,
                       hl_gru_value<T> value, int frame_size, int batch_size,
                       activation_mode_t active_node,
                       activation_mode_t active_gate) {
 #ifndef __NVCC__
     if (value.prev_out_value) {
-      math::gemm<platform::CPUPlace, T>(
+      math::gemm<platform::CPUDeviceContext, T>(
           context, false, false, batch_size, frame_size * 2, frame_size, 1,
           value.prev_out_value, frame_size, value.gate_weight, frame_size * 2,
           1, value.gate_value, frame_size * 3);
@@ -36,7 +36,7 @@ struct GRUUnitFunctor<platform::CPUPlace, T> {
                                  frame_size, batch_size, active_gate);
 
     if (value.prev_out_value) {
-      math::gemm<platform::CPUPlace, T>(
+      math::gemm<platform::CPUDeviceContext, T>(
           context, false, false, batch_size, frame_size, frame_size, 1,
           value.reset_output_value, frame_size, value.state_weight, frame_size,
           1, value.gate_value + frame_size * 2, frame_size * 3);
@@ -49,8 +49,8 @@ struct GRUUnitFunctor<platform::CPUPlace, T> {
 };
 
 template <typename T>
-struct GRUUnitGradFunctor<platform::CPUPlace, T> {
-  static void compute(const platform::DeviceContext &context,
+struct GRUUnitGradFunctor<platform::CPUDeviceContext, T> {
+  static void compute(const platform::CPUDeviceContext &context,
                       hl_gru_value<T> value, hl_gru_grad<T> grad,
                       int frame_size, int batch_size,
                       activation_mode_t active_node,
@@ -60,13 +60,13 @@ struct GRUUnitGradFunctor<platform::CPUPlace, T> {
                                 grad, frame_size, batch_size, active_node);
 
     if (value.prev_out_value && grad.prev_out_grad) {
-      math::gemm<platform::CPUPlace, T>(
+      math::gemm<platform::CPUDeviceContext, T>(
           context, false, true, batch_size, frame_size, frame_size, 1,
           grad.gate_grad + frame_size * 2, frame_size * 3, value.state_weight,
           frame_size, 0, grad.reset_output_grad, frame_size);
 
       if (grad.state_weight_grad) {
-        math::gemm<platform::CPUPlace, T>(
+        math::gemm<platform::CPUDeviceContext, T>(
             context, true, false, frame_size, frame_size, batch_size, 1,
             value.reset_output_value, frame_size,
             grad.gate_grad + frame_size * 2, frame_size * 3, 1,
@@ -78,13 +78,13 @@ struct GRUUnitGradFunctor<platform::CPUPlace, T> {
                                 grad, frame_size, batch_size, active_gate);
 
     if (grad.prev_out_grad && value.prev_out_value) {
-      math::gemm<platform::CPUPlace, T>(
+      math::gemm<platform::CPUDeviceContext, T>(
           context, false, true, batch_size, frame_size, frame_size * 2, 1,
           grad.gate_grad, frame_size * 3, value.gate_weight, frame_size * 2, 1,
           grad.prev_out_grad, frame_size);
 
       if (grad.gate_weight_grad) {
-        math::gemm<platform::CPUPlace, T>(
+        math::gemm<platform::CPUDeviceContext, T>(
             context, true, false, frame_size, frame_size * 2, batch_size, 1,
             value.prev_out_value, frame_size, grad.gate_grad, frame_size * 3, 1,
             grad.gate_weight_grad, frame_size * 2);
@@ -94,10 +94,10 @@ struct GRUUnitGradFunctor<platform::CPUPlace, T> {
   }
 };
 
-template struct GRUUnitFunctor<platform::CPUPlace, float>;
-template struct GRUUnitFunctor<platform::CPUPlace, double>;
-template struct GRUUnitGradFunctor<platform::CPUPlace, float>;
-template struct GRUUnitGradFunctor<platform::CPUPlace, double>;
+template struct GRUUnitFunctor<platform::CPUDeviceContext, float>;
+template struct GRUUnitFunctor<platform::CPUDeviceContext, double>;
+template struct GRUUnitGradFunctor<platform::CPUDeviceContext, float>;
+template struct GRUUnitGradFunctor<platform::CPUDeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/gru_compute.cu b/paddle/operators/math/gru_compute.cu
index 0252bdbdb6..dd518cd1e4 100644
--- a/paddle/operators/math/gru_compute.cu
+++ b/paddle/operators/math/gru_compute.cu
@@ -19,13 +19,12 @@ namespace operators {
 namespace math {
 
 template <typename T>
-struct GRUUnitFunctor<platform::GPUPlace, T> {
-  static void compute(const platform::DeviceContext &context,
+struct GRUUnitFunctor<platform::CUDADeviceContext, T> {
+  static void compute(const platform::CUDADeviceContext &context,
                       hl_gru_value<T> value, int frame_size, int batch_size,
                       activation_mode_t active_node,
                       activation_mode_t active_gate) {
-    auto stream =
-        reinterpret_cast<const platform::CUDADeviceContext &>(context).stream();
+    auto stream = context.stream();
     dim3 threads;
     dim3 grid;
     if (batch_size == 1) {
@@ -39,7 +38,7 @@ struct GRUUnitFunctor<platform::GPUPlace, T> {
     }
 
     if (value.prev_out_value) {
-      math::gemm<platform::GPUPlace, T>(
+      math::gemm<platform::CUDADeviceContext, T>(
           context, false, false, batch_size, frame_size * 2, frame_size, 1,
           value.prev_out_value, frame_size, value.gate_weight, frame_size * 2,
           1, value.gate_value, frame_size * 3);
@@ -62,7 +61,7 @@ struct GRUUnitFunctor<platform::GPUPlace, T> {
     }
 
     if (value.prev_out_value) {
-      math::gemm<platform::GPUPlace, T>(
+      math::gemm<platform::CUDADeviceContext, T>(
           context, false, false, batch_size, frame_size, frame_size, 1,
           value.reset_output_value, frame_size, value.state_weight, frame_size,
           1, value.gate_value + frame_size * 2, frame_size * 3);
@@ -87,14 +86,13 @@ struct GRUUnitFunctor<platform::GPUPlace, T> {
 };
 
 template <typename T>
-struct GRUUnitGradFunctor<platform::GPUPlace, T> {
-  static void compute(const platform::DeviceContext &context,
+struct GRUUnitGradFunctor<platform::CUDADeviceContext, T> {
+  static void compute(const platform::CUDADeviceContext &context,
                       hl_gru_value<T> value, hl_gru_grad<T> grad,
                       int frame_size, int batch_size,
                       activation_mode_t active_node,
                       activation_mode_t active_gate) {
-    auto stream =
-        reinterpret_cast<const platform::CUDADeviceContext &>(context).stream();
+    auto stream = context.stream();
     dim3 threads;
     dim3 grid;
     if (batch_size == 1) {
@@ -124,13 +122,13 @@ struct GRUUnitGradFunctor<platform::GPUPlace, T> {
     }
 
     if (value.prev_out_value && grad.prev_out_grad) {
-      math::gemm<platform::GPUPlace, T>(
+      math::gemm<platform::CUDADeviceContext, T>(
           context, false, true, batch_size, frame_size, frame_size, 1,
           grad.gate_grad + frame_size * 2, frame_size * 3, value.state_weight,
           frame_size, 0, grad.reset_output_grad, frame_size);
 
       if (grad.state_weight_grad) {
-        math::gemm<platform::GPUPlace, T>(
+        math::gemm<platform::CUDADeviceContext, T>(
             context, true, false, frame_size, frame_size, batch_size, 1,
             value.reset_output_value, frame_size,
             grad.gate_grad + frame_size * 2, frame_size * 3, 1,
@@ -155,13 +153,13 @@ struct GRUUnitGradFunctor<platform::GPUPlace, T> {
     }
 
     if (grad.prev_out_grad && value.prev_out_value) {
-      math::gemm<platform::GPUPlace, T>(
+      math::gemm<platform::CUDADeviceContext, T>(
           context, false, true, batch_size, frame_size, frame_size * 2, 1,
           grad.gate_grad, frame_size * 3, value.gate_weight, frame_size * 2, 1,
           grad.prev_out_grad, frame_size);
 
       if (grad.gate_weight_grad) {
-        math::gemm<platform::GPUPlace, T>(
+        math::gemm<platform::CUDADeviceContext, T>(
             context, true, false, frame_size, frame_size * 2, batch_size, 1,
             value.prev_out_value, frame_size, grad.gate_grad, frame_size * 3, 1,
             grad.gate_weight_grad, frame_size * 2);
@@ -170,10 +168,10 @@ struct GRUUnitGradFunctor<platform::GPUPlace, T> {
   }
 };
 
-template struct GRUUnitFunctor<platform::GPUPlace, float>;
-template struct GRUUnitFunctor<platform::GPUPlace, double>;
-template struct GRUUnitGradFunctor<platform::GPUPlace, float>;
-template struct GRUUnitGradFunctor<platform::GPUPlace, double>;
+template struct GRUUnitFunctor<platform::CUDADeviceContext, float>;
+template struct GRUUnitFunctor<platform::CUDADeviceContext, double>;
+template struct GRUUnitGradFunctor<platform::CUDADeviceContext, float>;
+template struct GRUUnitGradFunctor<platform::CUDADeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/gru_compute.h b/paddle/operators/math/gru_compute.h
index 58ea59f68e..ca1343cb2c 100644
--- a/paddle/operators/math/gru_compute.h
+++ b/paddle/operators/math/gru_compute.h
@@ -40,19 +40,18 @@ struct hl_gru_grad {
   T *prev_out_grad;
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct GRUUnitFunctor {
-  static void compute(const platform::DeviceContext &context,
-                      hl_gru_value<T> value, int frame_size, int batch_size,
+  static void compute(const DeviceContext &context, hl_gru_value<T> value,
+                      int frame_size, int batch_size,
                       activation_mode_t active_node,
                       activation_mode_t active_gate);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct GRUUnitGradFunctor {
-  static void compute(const platform::DeviceContext &context,
-                      hl_gru_value<T> value, hl_gru_grad<T> grad,
-                      int frame_size, int batch_size,
+  static void compute(const DeviceContext &context, hl_gru_value<T> value,
+                      hl_gru_grad<T> grad, int frame_size, int batch_size,
                       activation_mode_t active_node,
                       activation_mode_t active_gate);
 };
diff --git a/paddle/operators/math/im2col.cc b/paddle/operators/math/im2col.cc
index c10c44c520..707ebf0596 100644
--- a/paddle/operators/math/im2col.cc
+++ b/paddle/operators/math/im2col.cc
@@ -25,9 +25,9 @@ namespace math {
  */
 template <class T>
 class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
-                    platform::CPUPlace, T> {
+                    platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& im, const std::vector<int>& dilation,
                   const std::vector<int>& stride,
                   const std::vector<int>& padding, framework::Tensor* col) {
@@ -90,9 +90,9 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
  */
 template <class T>
 class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
-                    platform::CPUPlace, T> {
+                    platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& col,
                   const std::vector<int>& dilation,
                   const std::vector<int>& stride,
@@ -149,13 +149,13 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
 };
 
 template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
-                             platform::CPUPlace, float>;
+                             platform::CPUDeviceContext, float>;
 template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
-                             platform::CPUPlace, double>;
+                             platform::CPUDeviceContext, double>;
 template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
-                             platform::CPUPlace, float>;
+                             platform::CPUDeviceContext, float>;
 template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
-                             platform::CPUPlace, double>;
+                             platform::CPUDeviceContext, double>;
 
 /*
  * im = [input_channels, input_height, input_width]
@@ -164,9 +164,9 @@ template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
  */
 template <class T>
 class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
-                    platform::CPUPlace, T> {
+                    platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& im, const std::vector<int>& dilation,
                   const std::vector<int>& stride,
                   const std::vector<int>& padding, framework::Tensor* col) {
@@ -235,9 +235,9 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
  */
 template <class T>
 class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
-                    platform::CPUPlace, T> {
+                    platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& col,
                   const std::vector<int>& dilation,
                   const std::vector<int>& stride,
@@ -300,13 +300,13 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
 };
 
 template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
-                             platform::CPUPlace, float>;
+                             platform::CPUDeviceContext, float>;
 template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
-                             platform::CPUPlace, double>;
+                             platform::CPUDeviceContext, double>;
 template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
-                             platform::CPUPlace, float>;
+                             platform::CPUDeviceContext, float>;
 template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
-                             platform::CPUPlace, double>;
+                             platform::CPUDeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/im2col.cu b/paddle/operators/math/im2col.cu
index bf78942439..a88e837b03 100644
--- a/paddle/operators/math/im2col.cu
+++ b/paddle/operators/math/im2col.cu
@@ -58,9 +58,9 @@ __global__ void im2col(const T* data_im, int num_outs, int im_height,
  */
 template <class T>
 class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
-                    platform::GPUPlace, T> {
+                    platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& im, const std::vector<int>& dilation,
                   const std::vector<int>& stride,
                   const std::vector<int>& padding, framework::Tensor* col) {
@@ -96,9 +96,7 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
     int block_y = (blocks + 512 - 1) / 512;
     dim3 threads(1024, 1);
     dim3 grid(block_x, block_y);
-    im2col<T><<<grid, threads, 0,
-                reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                    .stream()>>>(
+    im2col<T><<<grid, threads, 0, context.stream()>>>(
         im.data<T>(), num_outputs, im_height, im_width, dilation[0],
         dilation[1], filter_height, filter_width, stride[0], stride[1],
         padding[0], padding[1], col_height, col_width, col->data<T>());
@@ -160,9 +158,9 @@ __global__ void col2im(int n, const T* data_col, int im_height, int im_width,
  */
 template <class T>
 class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
-                    platform::GPUPlace, T> {
+                    platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& col,
                   const std::vector<int>& dilation,
                   const std::vector<int>& stride,
@@ -203,9 +201,7 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
 
     // To avoid involving atomic operations, we will launch one kernel per
     // bottom dimension, and then in the kernel add up the top dimensions.
-    col2im<T><<<grid, threads, 0,
-                reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                    .stream()>>>(
+    col2im<T><<<grid, threads, 0, context.stream()>>>(
         num_kernels, col.data<T>(), im_height, im_width, dilation[0],
         dilation[1], filter_height, filter_width, stride[0], stride[1],
         padding[0], padding[2], col_height, col_width, im->data<T>());
@@ -213,13 +209,13 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
 };
 
 template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
-                             platform::GPUPlace, float>;
+                             platform::CUDADeviceContext, float>;
 template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
-                             platform::GPUPlace, double>;
+                             platform::CUDADeviceContext, double>;
 template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
-                             platform::GPUPlace, float>;
+                             platform::CUDADeviceContext, float>;
 template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
-                             platform::GPUPlace, double>;
+                             platform::CUDADeviceContext, double>;
 
 template <class T>
 __global__ void im2colOCF(const T* im_data, int im_channels, int im_height,
@@ -260,9 +256,9 @@ __global__ void im2colOCF(const T* im_data, int im_channels, int im_height,
  */
 template <class T>
 class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
-                    platform::GPUPlace, T> {
+                    platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& im, const std::vector<int>& dilation,
                   const std::vector<int>& stride,
                   const std::vector<int>& padding, framework::Tensor* col) {
@@ -310,9 +306,7 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
     int block_dim_z = 1024 / block_dim_x / block_dim_y;
     dim3 threads(block_dim_x, block_dim_y, std::min(block_dim_z, im_channels));
     dim3 grid(col_width, col_height);
-    im2colOCF<T><<<grid, threads, 0,
-                   reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                       .stream()>>>(
+    im2colOCF<T><<<grid, threads, 0, context.stream()>>>(
         im.data<T>(), im_channels, im_height, im_width, filter_height,
         filter_width, stride[0], stride[1], padding[0], padding[1], col_height,
         col_width, col->data<T>());
@@ -358,9 +352,9 @@ __global__ void col2imOCF(const T* col_data, int im_channels, int im_height,
  */
 template <class T>
 class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
-                    platform::GPUPlace, T> {
+                    platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& col,
                   const std::vector<int>& dilation,
                   const std::vector<int>& stride,
@@ -409,9 +403,7 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
     int block_dim_z = 1024 / block_dim_x / block_dim_y;
     dim3 threads(block_dim_x, block_dim_y, std::min(block_dim_z, im_channels));
     dim3 grid(col_width, col_height);
-    col2imOCF<T><<<grid, threads, 0,
-                   reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                       .stream()>>>(
+    col2imOCF<T><<<grid, threads, 0, context.stream()>>>(
         col.data<T>(), im_channels, im_height, im_width, filter_height,
         filter_width, stride[0], stride[1], padding[0], padding[1], col_height,
         col_width, im->data<T>());
@@ -419,13 +411,13 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
 };
 
 template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
-                             platform::GPUPlace, float>;
+                             platform::CUDADeviceContext, float>;
 template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
-                             platform::GPUPlace, double>;
+                             platform::CUDADeviceContext, double>;
 template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
-                             platform::GPUPlace, float>;
+                             platform::CUDADeviceContext, float>;
 template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
-                             platform::GPUPlace, double>;
+                             platform::CUDADeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/im2col.h b/paddle/operators/math/im2col.h
index 24fd9a06e9..38f2c9fe0a 100644
--- a/paddle/operators/math/im2col.h
+++ b/paddle/operators/math/im2col.h
@@ -79,20 +79,19 @@ enum class ColFormat { kCFO = 0, kOCF = 1 };
  * \note The caller needs to ensure that imShape.inputChannels is equal to
  *       colShape.inputChannels.
  */
-template <ColFormat Format, typename Place, typename T>
+template <ColFormat Format, typename DeviceContext, typename T>
 class Im2ColFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& im, const std::vector<int>& dilation,
+  void operator()(const DeviceContext& context, const framework::Tensor& im,
+                  const std::vector<int>& dilation,
                   const std::vector<int>& stride,
                   const std::vector<int>& padding, framework::Tensor* col);
 };
 
-template <ColFormat Format, typename Place, typename T>
+template <ColFormat Format, typename DeviceContext, typename T>
 class Col2ImFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& col,
+  void operator()(const DeviceContext& context, const framework::Tensor& col,
                   const std::vector<int>& dilation,
                   const std::vector<int>& stride,
                   const std::vector<int>& padding, framework::Tensor* im);
diff --git a/paddle/operators/math/im2col_test.cc b/paddle/operators/math/im2col_test.cc
index ae197a97ed..256f3bc9bd 100644
--- a/paddle/operators/math/im2col_test.cc
+++ b/paddle/operators/math/im2col_test.cc
@@ -16,7 +16,7 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <iostream>
 
-template <typename Place>
+template <typename DeviceContext, typename Place>
 void testIm2col() {
   paddle::framework::Tensor input_tmp;
   paddle::framework::Tensor input;
@@ -59,18 +59,7 @@ void testIm2col() {
   memcpy(input_ptr, arr, 6 * sizeof(float));
 
   auto* place = new Place();
-  paddle::platform::DeviceContext* context;
-  if (paddle::platform::is_cpu_place(*place)) {
-    context =
-        new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace());
-  } else {
-#ifdef PADDLE_WITH_CUDA
-    context =
-        new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace());
-#else
-    PADDLE_THROW("no GPU support");
-#endif  // PADDLE_WITH_CUDA
-  }
+  DeviceContext* context = new DeviceContext(*place);
   if (paddle::platform::is_cpu_place(*place)) {
     input = input_tmp;
   } else {
@@ -83,10 +72,10 @@ void testIm2col() {
 
   // Im2Col
   paddle::operators::math::Im2ColFunctor<
-      paddle::operators::math::ColFormat::kCFO, Place, float>
+      paddle::operators::math::ColFormat::kCFO, DeviceContext, float>
       im2col;
   paddle::operators::math::Im2ColFunctor<
-      paddle::operators::math::ColFormat::kOCF, Place, float>
+      paddle::operators::math::ColFormat::kOCF, DeviceContext, float>
       im2col_ocf;
 
   im2col(*context, input, dilation, stride, padding, &output_cfo);
@@ -119,10 +108,10 @@ void testIm2col() {
 
   // Col2Im: kCFO
   paddle::operators::math::Col2ImFunctor<
-      paddle::operators::math::ColFormat::kCFO, Place, float>
+      paddle::operators::math::ColFormat::kCFO, DeviceContext, float>
       col2im;
   paddle::operators::math::Col2ImFunctor<
-      paddle::operators::math::ColFormat::kOCF, Place, float>
+      paddle::operators::math::ColFormat::kOCF, DeviceContext, float>
       col2im_ocf;
   float col2im_data[] = {0, 2, 2, 3, 8, 5};
 
@@ -168,8 +157,8 @@ void testIm2col() {
 }
 
 TEST(math, im2col) {
-  testIm2col<paddle::platform::CPUPlace>();
+  testIm2col<paddle::platform::CPUDeviceContext, paddle::platform::CPUPlace>();
 #ifdef PADDLE_WITH_CUDA
-  testIm2col<paddle::platform::GPUPlace>();
+  testIm2col<paddle::platform::CUDADeviceContext, paddle::platform::GPUPlace>();
 #endif
 }
diff --git a/paddle/operators/math/lstm_compute.cc b/paddle/operators/math/lstm_compute.cc
index ad3a59bcdb..2c2e8bb82e 100644
--- a/paddle/operators/math/lstm_compute.cc
+++ b/paddle/operators/math/lstm_compute.cc
@@ -21,8 +21,8 @@ namespace operators {
 namespace math {
 
 template <class T>
-struct LstmUnitFunctor<platform::CPUPlace, T> {
-  static void compute(const platform::DeviceContext& context,
+struct LstmUnitFunctor<platform::CPUDeviceContext, T> {
+  static void compute(const platform::CPUDeviceContext& context,
                       LstmMetaValue<T> value, int frame_size, int batch_size,
                       const std::string& gate_act, const std::string& cell_act,
                       const std::string& cand_act) {
@@ -42,8 +42,8 @@ struct LstmUnitFunctor<platform::CPUPlace, T> {
 };
 
 template <class T>
-struct LstmUnitGradFunctor<platform::CPUPlace, T> {
-  static void compute(const platform::DeviceContext& context,
+struct LstmUnitGradFunctor<platform::CPUDeviceContext, T> {
+  static void compute(const platform::CPUDeviceContext& context,
                       LstmMetaValue<T> value, LstmMetaGrad<T> grad,
                       int frame_size, int batch_size,
                       const std::string& gate_act, const std::string& cell_act,
@@ -72,10 +72,10 @@ struct LstmUnitGradFunctor<platform::CPUPlace, T> {
   }
 };
 
-template class LstmUnitFunctor<platform::CPUPlace, float>;
-template class LstmUnitFunctor<platform::CPUPlace, double>;
-template class LstmUnitGradFunctor<platform::CPUPlace, float>;
-template class LstmUnitGradFunctor<platform::CPUPlace, double>;
+template class LstmUnitFunctor<platform::CPUDeviceContext, float>;
+template class LstmUnitFunctor<platform::CPUDeviceContext, double>;
+template class LstmUnitGradFunctor<platform::CPUDeviceContext, float>;
+template class LstmUnitGradFunctor<platform::CPUDeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/lstm_compute.cu b/paddle/operators/math/lstm_compute.cu
index b2122f2a5c..92b1f4228b 100644
--- a/paddle/operators/math/lstm_compute.cu
+++ b/paddle/operators/math/lstm_compute.cu
@@ -21,8 +21,8 @@ namespace operators {
 namespace math {
 
 template <class T>
-struct LstmUnitFunctor<platform::GPUPlace, T> {
-  static void compute(const platform::DeviceContext& context,
+struct LstmUnitFunctor<platform::CUDADeviceContext, T> {
+  static void compute(const platform::CUDADeviceContext& context,
                       LstmMetaValue<T> value, int frame_size, int batch_size,
                       const std::string& gate_act, const std::string& cell_act,
                       const std::string& cand_act) {
@@ -33,8 +33,8 @@ struct LstmUnitFunctor<platform::GPUPlace, T> {
 };
 
 template <class T>
-struct LstmUnitGradFunctor<platform::GPUPlace, T> {
-  static void compute(const platform::DeviceContext& context,
+struct LstmUnitGradFunctor<platform::CUDADeviceContext, T> {
+  static void compute(const platform::CUDADeviceContext& context,
                       LstmMetaValue<T> value, LstmMetaGrad<T> grad,
                       int frame_size, int batch_size,
                       const std::string& gate_act, const std::string& cell_act,
@@ -45,10 +45,10 @@ struct LstmUnitGradFunctor<platform::GPUPlace, T> {
   }
 };
 
-template class LstmUnitFunctor<platform::GPUPlace, float>;
-template class LstmUnitFunctor<platform::GPUPlace, double>;
-template class LstmUnitGradFunctor<platform::GPUPlace, float>;
-template class LstmUnitGradFunctor<platform::GPUPlace, double>;
+template class LstmUnitFunctor<platform::CUDADeviceContext, float>;
+template class LstmUnitFunctor<platform::CUDADeviceContext, double>;
+template class LstmUnitGradFunctor<platform::CUDADeviceContext, float>;
+template class LstmUnitGradFunctor<platform::CUDADeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/lstm_compute.h b/paddle/operators/math/lstm_compute.h
index 9652399d4c..5f74e27358 100644
--- a/paddle/operators/math/lstm_compute.h
+++ b/paddle/operators/math/lstm_compute.h
@@ -67,21 +67,20 @@ inline activation_mode_t ActiveType(const std::string &type) {
   }
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LstmUnitFunctor {
  public:
-  static void compute(const platform::DeviceContext &context,
-                      LstmMetaValue<T> value, int frame_size, int batch_size,
+  static void compute(const DeviceContext &context, LstmMetaValue<T> value,
+                      int frame_size, int batch_size,
                       const std::string &gate_act, const std::string &cell_act,
                       const std::string &cand_act);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LstmUnitGradFunctor {
  public:
-  static void compute(const platform::DeviceContext &context,
-                      LstmMetaValue<T> value, LstmMetaGrad<T> grad,
-                      int frame_size, int batch_size,
+  static void compute(const DeviceContext &context, LstmMetaValue<T> value,
+                      LstmMetaGrad<T> grad, int frame_size, int batch_size,
                       const std::string &gate_act, const std::string &cell_act,
                       const std::string &cand_act);
 };
diff --git a/paddle/operators/math/math_function.cc b/paddle/operators/math/math_function.cc
index e099a6a439..2b35e4532a 100644
--- a/paddle/operators/math/math_function.cc
+++ b/paddle/operators/math/math_function.cc
@@ -21,13 +21,11 @@ namespace operators {
 namespace math {
 
 template <>
-void gemm<platform::CPUPlace, float>(const platform::DeviceContext& context,
-                                     const CBLAS_TRANSPOSE transA,
-                                     const CBLAS_TRANSPOSE transB, const int M,
-                                     const int N, const int K,
-                                     const float alpha, const float* A,
-                                     const float* B, const float beta,
-                                     float* C) {
+void gemm<platform::CPUDeviceContext, float>(
+    const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA,
+    const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
+    const float alpha, const float* A, const float* B, const float beta,
+    float* C) {
   int lda = (transA == CblasNoTrans) ? K : M;
   int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
@@ -36,13 +34,11 @@ void gemm<platform::CPUPlace, float>(const platform::DeviceContext& context,
 }
 
 template <>
-void gemm<platform::CPUPlace, double>(const platform::DeviceContext& context,
-                                      const CBLAS_TRANSPOSE transA,
-                                      const CBLAS_TRANSPOSE transB, const int M,
-                                      const int N, const int K,
-                                      const double alpha, const double* A,
-                                      const double* B, const double beta,
-                                      double* C) {
+void gemm<platform::CPUDeviceContext, double>(
+    const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA,
+    const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
+    const double alpha, const double* A, const double* B, const double beta,
+    double* C) {
   int lda = (transA == CblasNoTrans) ? K : M;
   int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
@@ -51,35 +47,32 @@ void gemm<platform::CPUPlace, double>(const platform::DeviceContext& context,
 }
 
 template <>
-void gemm<platform::CPUPlace, float>(const platform::DeviceContext& context,
-                                     const bool transA, const bool transB,
-                                     const int M, const int N, const int K,
-                                     const float alpha, const float* A,
-                                     const int lda, const float* B,
-                                     const int ldb, const float beta, float* C,
-                                     const int ldc) {
+void gemm<platform::CPUDeviceContext, float>(
+    const platform::CPUDeviceContext& context, const bool transA,
+    const bool transB, const int M, const int N, const int K, const float alpha,
+    const float* A, const int lda, const float* B, const int ldb,
+    const float beta, float* C, const int ldc) {
   cblas_sgemm(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans,
               transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A,
               lda, B, ldb, beta, C, ldc);
 }
 
 template <>
-void gemm<platform::CPUPlace, double>(const platform::DeviceContext& context,
-                                      const bool transA, const bool transB,
-                                      const int M, const int N, const int K,
-                                      const double alpha, const double* A,
-                                      const int lda, const double* B,
-                                      const int ldb, const double beta,
-                                      double* C, const int ldc) {
+void gemm<platform::CPUDeviceContext, double>(
+    const platform::CPUDeviceContext& context, const bool transA,
+    const bool transB, const int M, const int N, const int K,
+    const double alpha, const double* A, const int lda, const double* B,
+    const int ldb, const double beta, double* C, const int ldc) {
   cblas_dgemm(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans,
               transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A,
               lda, B, ldb, beta, C, ldc);
 }
 
 template <>
-void matmul<platform::CPUPlace, float>(
-    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
-    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, float alpha,
+void matmul<platform::CPUDeviceContext, float>(
+    const platform::CPUDeviceContext& context,
+    const framework::Tensor& matrix_a, bool trans_a,
+    const framework::Tensor& matrix_b, bool trans_b, float alpha,
     framework::Tensor* matrix_out, float beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
@@ -99,15 +92,16 @@ void matmul<platform::CPUPlace, float>(
   CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
-  gemm<platform::CPUPlace, float>(
+  gemm<platform::CPUDeviceContext, float>(
       context, transA, transB, M, N, K, alpha, matrix_a.data<float>(),
       matrix_b.data<float>(), beta, matrix_out->data<float>());
 }
 
 template <>
-void matmul<platform::CPUPlace, double>(
-    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
-    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, double alpha,
+void matmul<platform::CPUDeviceContext, double>(
+    const platform::CPUDeviceContext& context,
+    const framework::Tensor& matrix_a, bool trans_a,
+    const framework::Tensor& matrix_b, bool trans_b, double alpha,
     framework::Tensor* matrix_out, double beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
@@ -127,7 +121,7 @@ void matmul<platform::CPUPlace, double>(
   CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
-  gemm<platform::CPUPlace, double>(
+  gemm<platform::CPUDeviceContext, double>(
       context, transA, transB, M, N, K, alpha, matrix_a.data<double>(),
       matrix_b.data<double>(), beta, matrix_out->data<double>());
 }
@@ -135,8 +129,8 @@ void matmul<platform::CPUPlace, double>(
 #ifdef PADDLE_WITH_MKLML
 // Use cblas_{s,d}gemm_batched if available: Run with 1 group of size batchSize.
 template <>
-void batched_gemm<platform::CPUPlace, float>(
-    const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
+void batched_gemm<platform::CPUDeviceContext, float>(
+    const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA,
     const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
     const float alpha, const float* A, const float* B, const float beta,
     float* C, const int batchCount, const int strideA, const int strideB) {
@@ -157,8 +151,8 @@ void batched_gemm<platform::CPUPlace, float>(
 }
 
 template <>
-void batched_gemm<platform::CPUPlace, double>(
-    const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
+void batched_gemm<platform::CPUDeviceContext, double>(
+    const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA,
     const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
     const double alpha, const double* A, const double* B, const double beta,
     double* C, const int batchCount, const int strideA, const int strideB) {
@@ -183,8 +177,8 @@ void batched_gemm<platform::CPUPlace, double>(
 // functions of Intel MKL are not available. In the future, this computation
 // should be parallelized.
 template <>
-void batched_gemm<platform::CPUPlace, float>(
-    const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
+void batched_gemm<platform::CPUDeviceContext, float>(
+    const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA,
     const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
     const float alpha, const float* A, const float* B, const float beta,
     float* C, const int batchCount, const int strideA, const int strideB) {
@@ -192,14 +186,14 @@ void batched_gemm<platform::CPUPlace, float>(
     const float* Ak = &A[k * strideA];
     const float* Bk = &B[k * strideB];
     float* Ck = &C[k * M * N];
-    gemm<platform::CPUPlace, float>(context, transA, transB, M, N, K, alpha, Ak,
-                                    Bk, beta, Ck);
+    gemm<platform::CPUDeviceContext, float>(context, transA, transB, M, N, K,
+                                            alpha, Ak, Bk, beta, Ck);
   }
 }
 
 template <>
-void batched_gemm<platform::CPUPlace, double>(
-    const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
+void batched_gemm<platform::CPUDeviceContext, double>(
+    const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA,
     const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
     const double alpha, const double* A, const double* B, const double beta,
     double* C, const int batchCount, const int strideA, const int strideB) {
@@ -207,55 +201,53 @@ void batched_gemm<platform::CPUPlace, double>(
     const double* Ak = &A[k * strideA];
     const double* Bk = &B[k * strideB];
     double* Ck = &C[k * M * N];
-    gemm<platform::CPUPlace, double>(context, transA, transB, M, N, K, alpha,
-                                     Ak, Bk, beta, Ck);
+    gemm<platform::CPUDeviceContext, double>(context, transA, transB, M, N, K,
+                                             alpha, Ak, Bk, beta, Ck);
   }
 }
 #endif
 
 template <>
-void gemv<platform::CPUPlace, float>(const platform::DeviceContext& context,
-                                     const bool trans_a, const int M,
-                                     const int N, const float alpha,
-                                     const float* A, const float* B,
-                                     const float beta, float* C) {
+void gemv<platform::CPUDeviceContext, float>(
+    const platform::CPUDeviceContext& context, const bool trans_a, const int M,
+    const int N, const float alpha, const float* A, const float* B,
+    const float beta, float* C) {
   CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
   cblas_sgemv(CblasRowMajor, transA, M, N, alpha, A, N, B, 1, beta, C, 1);
 }
 
 template <>
-void gemv<platform::CPUPlace, double>(const platform::DeviceContext& context,
-                                      const bool trans_a, const int M,
-                                      const int N, const double alpha,
-                                      const double* A, const double* B,
-                                      const double beta, double* C) {
+void gemv<platform::CPUDeviceContext, double>(
+    const platform::CPUDeviceContext& context, const bool trans_a, const int M,
+    const int N, const double alpha, const double* A, const double* B,
+    const double beta, double* C) {
   CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
   cblas_dgemv(CblasRowMajor, transA, M, N, alpha, A, N, B, 1, beta, C, 1);
 }
 
 template <>
-void axpy<platform::CPUPlace, float>(const platform::DeviceContext& context,
-                                     const int n, const float alpha,
-                                     const float* x, float* y) {
+void axpy<platform::CPUDeviceContext, float>(
+    const platform::CPUDeviceContext& context, const int n, const float alpha,
+    const float* x, float* y) {
   cblas_saxpy(n, alpha, x, 1, y, 1);
 }
 
 template <>
-void axpy<platform::CPUPlace, double>(const platform::DeviceContext& context,
-                                      const int n, const double alpha,
-                                      const double* x, double* y) {
+void axpy<platform::CPUDeviceContext, double>(
+    const platform::CPUDeviceContext& context, const int n, const double alpha,
+    const double* x, double* y) {
   cblas_daxpy(n, alpha, x, 1, y, 1);
 }
 
-template struct SetConstant<platform::CPUPlace, float>;
-template struct SetConstant<platform::CPUPlace, double>;
-template struct SetConstant<platform::CPUPlace, int>;
-template struct SetConstant<platform::CPUPlace, int64_t>;
-template struct SetConstant<platform::CPUPlace, bool>;
+template struct SetConstant<platform::CPUDeviceContext, float>;
+template struct SetConstant<platform::CPUDeviceContext, double>;
+template struct SetConstant<platform::CPUDeviceContext, int>;
+template struct SetConstant<platform::CPUDeviceContext, int64_t>;
+template struct SetConstant<platform::CPUDeviceContext, bool>;
 
-#define DEFINE_CPU_TRANS(RANK)                                \
-  template struct Transpose<platform::CPUPlace, float, RANK>; \
-  template struct Transpose<platform::CPUPlace, double, RANK>;
+#define DEFINE_CPU_TRANS(RANK)                                        \
+  template struct Transpose<platform::CPUDeviceContext, float, RANK>; \
+  template struct Transpose<platform::CPUDeviceContext, double, RANK>;
 
 DEFINE_CPU_TRANS(1);
 DEFINE_CPU_TRANS(2);
@@ -310,10 +302,10 @@ void set_constant(const platform::DeviceContext& context,
 #endif
 }
 
-template struct RowwiseAdd<platform::CPUPlace, float>;
-template struct RowwiseAdd<platform::CPUPlace, double>;
-template struct ColwiseSum<platform::CPUPlace, float>;
-template struct ColwiseSum<platform::CPUPlace, double>;
+template struct RowwiseAdd<platform::CPUDeviceContext, float>;
+template struct RowwiseAdd<platform::CPUDeviceContext, double>;
+template struct ColwiseSum<platform::CPUDeviceContext, float>;
+template struct ColwiseSum<platform::CPUDeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/math_function.cu b/paddle/operators/math/math_function.cu
index 3018e50a4f..1b560a7e2d 100644
--- a/paddle/operators/math/math_function.cu
+++ b/paddle/operators/math/math_function.cu
@@ -22,13 +22,11 @@ namespace operators {
 namespace math {
 
 template <>
-void gemm<platform::GPUPlace, float>(const platform::DeviceContext& context,
-                                     const CBLAS_TRANSPOSE transA,
-                                     const CBLAS_TRANSPOSE transB, const int M,
-                                     const int N, const int K,
-                                     const float alpha, const float* A,
-                                     const float* B, const float beta,
-                                     float* C) {
+void gemm<platform::CUDADeviceContext, float>(
+    const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA,
+    const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
+    const float alpha, const float* A, const float* B, const float beta,
+    float* C) {
   // Note that cublas follows fortran order, so the order is different from
   // the cblas convention.
   int lda = (transA == CblasNoTrans) ? K : M;
@@ -39,19 +37,16 @@ void gemm<platform::GPUPlace, float>(const platform::DeviceContext& context,
       (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
 
   PADDLE_ENFORCE(platform::dynload::cublasSgemm(
-      reinterpret_cast<const platform::CUDADeviceContext&>(context)
-          .cublas_handle(),
-      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N));
+      context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A,
+      lda, &beta, C, N));
 }
 
 template <>
-void gemm<platform::GPUPlace, double>(const platform::DeviceContext& context,
-                                      const CBLAS_TRANSPOSE transA,
-                                      const CBLAS_TRANSPOSE transB, const int M,
-                                      const int N, const int K,
-                                      const double alpha, const double* A,
-                                      const double* B, const double beta,
-                                      double* C) {
+void gemm<platform::CUDADeviceContext, double>(
+    const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA,
+    const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
+    const double alpha, const double* A, const double* B, const double beta,
+    double* C) {
   // Note that cublas follows fortran order, so the order is different from
   // the cblas convention.
   int lda = (transA == CblasNoTrans) ? K : M;
@@ -61,51 +56,45 @@ void gemm<platform::GPUPlace, double>(const platform::DeviceContext& context,
   cublasOperation_t cuTransB =
       (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
   PADDLE_ENFORCE(platform::dynload::cublasDgemm(
-      reinterpret_cast<const platform::CUDADeviceContext&>(context)
-          .cublas_handle(),
-      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N));
+      context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A,
+      lda, &beta, C, N));
 }
 
 template <>
-void gemm<platform::GPUPlace, float>(const platform::DeviceContext& context,
-                                     const bool transA, const bool transB,
-                                     const int M, const int N, const int K,
-                                     const float alpha, const float* A,
-                                     const int lda, const float* B,
-                                     const int ldb, const float beta, float* C,
-                                     const int ldc) {
+void gemm<platform::CUDADeviceContext, float>(
+    const platform::CUDADeviceContext& context, const bool transA,
+    const bool transB, const int M, const int N, const int K, const float alpha,
+    const float* A, const int lda, const float* B, const int ldb,
+    const float beta, float* C, const int ldc) {
   // Note that cublas follows fortran order, so the order is different from
   // the cblas convention.
   cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T;
   cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T;
   PADDLE_ENFORCE(platform::dynload::cublasSgemm(
-      reinterpret_cast<const platform::CUDADeviceContext&>(context)
-          .cublas_handle(),
-      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, ldc));
+      context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A,
+      lda, &beta, C, ldc));
 }
 
 template <>
-void gemm<platform::GPUPlace, double>(const platform::DeviceContext& context,
-                                      const bool transA, const bool transB,
-                                      const int M, const int N, const int K,
-                                      const double alpha, const double* A,
-                                      const int lda, const double* B,
-                                      const int ldb, const double beta,
-                                      double* C, const int ldc) {
+void gemm<platform::CUDADeviceContext, double>(
+    const platform::CUDADeviceContext& context, const bool transA,
+    const bool transB, const int M, const int N, const int K,
+    const double alpha, const double* A, const int lda, const double* B,
+    const int ldb, const double beta, double* C, const int ldc) {
   // Note that cublas follows fortran order, so the order is different from
   // the cblas convention.
   cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T;
   cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T;
   PADDLE_ENFORCE(platform::dynload::cublasDgemm(
-      reinterpret_cast<const platform::CUDADeviceContext&>(context)
-          .cublas_handle(),
-      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, ldc));
+      context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A,
+      lda, &beta, C, ldc));
 }
 
 template <>
-void matmul<platform::GPUPlace, float>(
-    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
-    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, float alpha,
+void matmul<platform::CUDADeviceContext, float>(
+    const platform::CUDADeviceContext& context,
+    const framework::Tensor& matrix_a, bool trans_a,
+    const framework::Tensor& matrix_b, bool trans_b, float alpha,
     framework::Tensor* matrix_out, float beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
@@ -125,15 +114,16 @@ void matmul<platform::GPUPlace, float>(
   CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
-  gemm<platform::GPUPlace, float>(
+  gemm<platform::CUDADeviceContext, float>(
       context, transA, transB, M, N, K, alpha, matrix_a.data<float>(),
       matrix_b.data<float>(), beta, matrix_out->data<float>());
 }
 
 template <>
-void matmul<platform::GPUPlace, double>(
-    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
-    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, double alpha,
+void matmul<platform::CUDADeviceContext, double>(
+    const platform::CUDADeviceContext& context,
+    const framework::Tensor& matrix_a, bool trans_a,
+    const framework::Tensor& matrix_b, bool trans_b, double alpha,
     framework::Tensor* matrix_out, double beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
@@ -153,14 +143,14 @@ void matmul<platform::GPUPlace, double>(
   CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
-  gemm<platform::GPUPlace, double>(
+  gemm<platform::CUDADeviceContext, double>(
       context, transA, transB, M, N, K, alpha, matrix_a.data<double>(),
       matrix_b.data<double>(), beta, matrix_out->data<double>());
 }
 
 template <>
-void batched_gemm<platform::GPUPlace, float>(
-    const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
+void batched_gemm<platform::CUDADeviceContext, float>(
+    const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA,
     const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
     const float alpha, const float* A, const float* B, const float beta,
     float* C, const int batchCount, const int strideA, const int strideB) {
@@ -176,15 +166,13 @@ void batched_gemm<platform::GPUPlace, float>(
   const int strideC = M * N;
 
   PADDLE_ENFORCE(platform::dynload::cublasSgemmStridedBatched(
-      reinterpret_cast<const platform::CUDADeviceContext&>(context)
-          .cublas_handle(),
-      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, strideB, A, lda, strideA,
-      &beta, C, ldc, strideC, batchCount));
+      context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb,
+      strideB, A, lda, strideA, &beta, C, ldc, strideC, batchCount));
 }
 
 template <>
-void batched_gemm<platform::GPUPlace, double>(
-    const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
+void batched_gemm<platform::CUDADeviceContext, double>(
+    const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA,
     const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
     const double alpha, const double* A, const double* B, const double beta,
     double* C, const int batchCount, const int strideA, const int strideB) {
@@ -200,68 +188,58 @@ void batched_gemm<platform::GPUPlace, double>(
   const int strideC = M * N;
 
   PADDLE_ENFORCE(platform::dynload::cublasDgemmStridedBatched(
-      reinterpret_cast<const platform::CUDADeviceContext&>(context)
-          .cublas_handle(),
-      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, strideB, A, lda, strideA,
-      &beta, C, ldc, strideC, batchCount));
+      context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb,
+      strideB, A, lda, strideA, &beta, C, ldc, strideC, batchCount));
 }
 
 template <>
-void gemv<platform::GPUPlace, float>(const platform::DeviceContext& context,
-                                     const bool trans_a, const int M,
-                                     const int N, const float alpha,
-                                     const float* A, const float* B,
-                                     const float beta, float* C) {
+void gemv<platform::CUDADeviceContext, float>(
+    const platform::CUDADeviceContext& context, const bool trans_a, const int M,
+    const int N, const float alpha, const float* A, const float* B,
+    const float beta, float* C) {
   cublasOperation_t cuTransA = (trans_a == false) ? CUBLAS_OP_T : CUBLAS_OP_N;
 
-  PADDLE_ENFORCE(platform::dynload::cublasSgemv(
-      reinterpret_cast<const platform::CUDADeviceContext&>(context)
-          .cublas_handle(),
-      cuTransA, N, M, &alpha, A, N, B, 1, &beta, C, 1));
+  PADDLE_ENFORCE(platform::dynload::cublasSgemv(context.cublas_handle(),
+                                                cuTransA, N, M, &alpha, A, N, B,
+                                                1, &beta, C, 1));
 }
 
 template <>
-void gemv<platform::GPUPlace, double>(const platform::DeviceContext& context,
-                                      const bool trans_a, const int M,
-                                      const int N, const double alpha,
-                                      const double* A, const double* B,
-                                      const double beta, double* C) {
+void gemv<platform::CUDADeviceContext, double>(
+    const platform::CUDADeviceContext& context, const bool trans_a, const int M,
+    const int N, const double alpha, const double* A, const double* B,
+    const double beta, double* C) {
   cublasOperation_t cuTransA = (trans_a == false) ? CUBLAS_OP_T : CUBLAS_OP_N;
-  PADDLE_ENFORCE(platform::dynload::cublasDgemv(
-      reinterpret_cast<const platform::CUDADeviceContext&>(context)
-          .cublas_handle(),
-      cuTransA, N, M, &alpha, A, N, B, 1, &beta, C, 1));
+  PADDLE_ENFORCE(platform::dynload::cublasDgemv(context.cublas_handle(),
+                                                cuTransA, N, M, &alpha, A, N, B,
+                                                1, &beta, C, 1));
 }
 
 template <>
-void axpy<platform::GPUPlace, float>(const platform::DeviceContext& context,
-                                     const int n, const float alpha,
-                                     const float* x, float* y) {
-  PADDLE_ENFORCE(platform::dynload::cublasSaxpy(
-      reinterpret_cast<const platform::CUDADeviceContext&>(context)
-          .cublas_handle(),
-      n, &alpha, x, 1, y, 1));
+void axpy<platform::CUDADeviceContext, float>(
+    const platform::CUDADeviceContext& context, const int n, const float alpha,
+    const float* x, float* y) {
+  PADDLE_ENFORCE(platform::dynload::cublasSaxpy(context.cublas_handle(), n,
+                                                &alpha, x, 1, y, 1));
 }
 
 template <>
-void axpy<platform::GPUPlace, double>(const platform::DeviceContext& context,
-                                      const int n, const double alpha,
-                                      const double* x, double* y) {
-  PADDLE_ENFORCE(platform::dynload::cublasDaxpy(
-      reinterpret_cast<const platform::CUDADeviceContext&>(context)
-          .cublas_handle(),
-      n, &alpha, x, 1, y, 1));
+void axpy<platform::CUDADeviceContext, double>(
+    const platform::CUDADeviceContext& context, const int n, const double alpha,
+    const double* x, double* y) {
+  PADDLE_ENFORCE(platform::dynload::cublasDaxpy(context.cublas_handle(), n,
+                                                &alpha, x, 1, y, 1));
 }
 
-template struct SetConstant<platform::GPUPlace, float>;
-template struct SetConstant<platform::GPUPlace, double>;
-template struct SetConstant<platform::GPUPlace, int>;
-template struct SetConstant<platform::GPUPlace, int64_t>;
-template struct SetConstant<platform::GPUPlace, bool>;
+template struct SetConstant<platform::CUDADeviceContext, float>;
+template struct SetConstant<platform::CUDADeviceContext, double>;
+template struct SetConstant<platform::CUDADeviceContext, int>;
+template struct SetConstant<platform::CUDADeviceContext, int64_t>;
+template struct SetConstant<platform::CUDADeviceContext, bool>;
 
-#define DEFINE_GPU_TRANS(RANK)                                \
-  template struct Transpose<platform::GPUPlace, float, RANK>; \
-  template struct Transpose<platform::GPUPlace, double, RANK>;
+#define DEFINE_GPU_TRANS(RANK)                                         \
+  template struct Transpose<platform::CUDADeviceContext, float, RANK>; \
+  template struct Transpose<platform::CUDADeviceContext, double, RANK>;
 
 DEFINE_GPU_TRANS(1);
 DEFINE_GPU_TRANS(2);
@@ -277,8 +255,9 @@ struct TensorSetConstantGPU {
 
   template <typename T>
   void operator()() const {
-    SetConstant<platform::GPUPlace, T> functor;
-    functor(context_, tensor_, static_cast<T>(value_));
+    SetConstant<platform::CUDADeviceContext, T> functor;
+    functor(reinterpret_cast<const platform::CUDADeviceContext&>(context_),
+            tensor_, static_cast<T>(value_));
   }
 
   const platform::DeviceContext& context_;
@@ -294,27 +273,27 @@ void set_constant_with_place<platform::GPUPlace>(
                            TensorSetConstantGPU(context, tensor, value));
 }
 
-template struct RowwiseAdd<platform::GPUPlace, float>;
-template struct RowwiseAdd<platform::GPUPlace, double>;
-template struct ColwiseSum<platform::GPUPlace, float>;
-// template struct ColwiseSum<platform::GPUPlace, double>;
-// The ColwiseSum<platform::GPUPlace, double> failed in debug mode,
+template struct RowwiseAdd<platform::CUDADeviceContext, float>;
+template struct RowwiseAdd<platform::CUDADeviceContext, double>;
+template struct ColwiseSum<platform::CUDADeviceContext, float>;
+// template struct ColwiseSum<platform::CUDADeviceContext, double>;
+// The ColwiseSum<platform::CUDADeviceContext, double> failed in debug mode,
 // and only failed for this case. So reimplemented it.
 template <>
-void ColwiseSum<platform::GPUPlace, double>::operator()(
-    const platform::DeviceContext& context, const framework::Tensor& input,
+void ColwiseSum<platform::CUDADeviceContext, double>::operator()(
+    const platform::CUDADeviceContext& context, const framework::Tensor& input,
     framework::Tensor* vector) {
   auto in_dims = input.dims();
   auto size = input.numel() / in_dims[0];
   PADDLE_ENFORCE_EQ(vector->numel(), size);
   framework::Tensor one;
   one.mutable_data<double>({in_dims[0]}, context.GetPlace());
-  SetConstant<platform::GPUPlace, double> set;
+  SetConstant<platform::CUDADeviceContext, double> set;
   set(context, &one, static_cast<double>(1.0));
-  gemv<platform::GPUPlace, double>(context, true, static_cast<int>(in_dims[0]),
-                                   static_cast<int>(in_dims[1]), 1.0,
-                                   input.data<double>(), one.data<double>(),
-                                   0.0, vector->data<double>());
+  gemv<platform::CUDADeviceContext, double>(
+      context, true, static_cast<int>(in_dims[0]), static_cast<int>(in_dims[1]),
+      1.0, input.data<double>(), one.data<double>(), 0.0,
+      vector->data<double>());
 }
 
 }  // namespace math
diff --git a/paddle/operators/math/math_function.h b/paddle/operators/math/math_function.h
index f2b025b78b..8cc03c2ba0 100644
--- a/paddle/operators/math/math_function.h
+++ b/paddle/operators/math/math_function.h
@@ -62,53 +62,51 @@ namespace math {
 // Then matrixA: M * K, matrixB: K * N, matrixC : M * N
 // For more detailed info, please refer to
 // http://www.netlib.org/lapack/explore-html/d4/de2/sgemm_8f.html
-template <typename Place, typename T>
-void gemm(const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
+template <typename DeviceContext, typename T>
+void gemm(const DeviceContext& context, const CBLAS_TRANSPOSE transA,
           const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
           const T alpha, const T* A, const T* B, const T beta, T* C);
 
 // gemm wrapper with stride args for matrix uncontinuous in memory
-template <typename Place, typename T>
-void gemm(const platform::DeviceContext& context, const bool transA,
-          const bool transB, const int M, const int N, const int K,
-          const T alpha, const T* A, const int lda, const T* B, const int ldb,
-          const T beta, T* C, const int ldc);
+template <typename DeviceContext, typename T>
+void gemm(const DeviceContext& context, const bool transA, const bool transB,
+          const int M, const int N, const int K, const T alpha, const T* A,
+          const int lda, const T* B, const int ldb, const T beta, T* C,
+          const int ldc);
 
 // matrix multiply with continuous memory
-template <typename Place, typename T>
-void matmul(const platform::DeviceContext& context,
-            const framework::Tensor& matrix_a, bool trans_a,
-            const framework::Tensor& matrix_b, bool trans_b, T alpha,
-            framework::Tensor* matrix_out, T beta);
+template <typename DeviceContext, typename T>
+void matmul(const DeviceContext& context, const framework::Tensor& matrix_a,
+            bool trans_a, const framework::Tensor& matrix_b, bool trans_b,
+            T alpha, framework::Tensor* matrix_out, T beta);
 
 // Batched gemm
-template <typename Place, typename T>
-void batched_gemm(const platform::DeviceContext& context,
-                  const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB,
-                  const int M, const int N, const int K, const T alpha,
-                  const T* A, const T* B, const T beta, T* C,
-                  const int batchCount, const int strideA, const int strideB);
-
-template <typename Place, typename T>
-void gemv(const platform::DeviceContext& context, const bool trans_a,
-          const int M, const int N, const T alpha, const T* A, const T* B,
-          const T beta, T* C);
-
-template <typename Place, typename T>
-void axpy(const platform::DeviceContext& context, const int n, const T alpha,
-          const T* x, T* y);
-
-template <typename Place, typename T, int Rank>
+template <typename DeviceContext, typename T>
+void batched_gemm(const DeviceContext& context, const CBLAS_TRANSPOSE transA,
+                  const CBLAS_TRANSPOSE transB, const int M, const int N,
+                  const int K, const T alpha, const T* A, const T* B,
+                  const T beta, T* C, const int batchCount, const int strideA,
+                  const int strideB);
+
+template <typename DeviceContext, typename T>
+void gemv(const DeviceContext& context, const bool trans_a, const int M,
+          const int N, const T alpha, const T* A, const T* B, const T beta,
+          T* C);
+
+template <typename DeviceContext, typename T>
+void axpy(const DeviceContext& context, const int n, const T alpha, const T* x,
+          T* y);
+
+template <typename DeviceContext, typename T, int Rank>
 struct Transpose {
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& in, framework::Tensor* out,
-                  const std::vector<int>& axis);
+  void operator()(const DeviceContext& context, const framework::Tensor& in,
+                  framework::Tensor* out, const std::vector<int>& axis);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct SetConstant {
-  void operator()(const platform::DeviceContext& context,
-                  framework::Tensor* tensor, T num);
+  void operator()(const DeviceContext& context, framework::Tensor* tensor,
+                  T num);
 };
 
 template <typename Place>
@@ -118,17 +116,16 @@ void set_constant_with_place(const platform::DeviceContext& context,
 void set_constant(const platform::DeviceContext& context,
                   framework::Tensor* tensor, float value);
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct RowwiseAdd {
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input, const framework::Tensor& vec,
-                  framework::Tensor* output);
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  const framework::Tensor& vec, framework::Tensor* output);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct ColwiseSum {
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input, framework::Tensor* vec);
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  framework::Tensor* vec);
 };
 
 }  // namespace math
diff --git a/paddle/operators/math/math_function_impl.h b/paddle/operators/math/math_function_impl.h
index 4dc17a4e52..3e6d833865 100644
--- a/paddle/operators/math/math_function_impl.h
+++ b/paddle/operators/math/math_function_impl.h
@@ -20,16 +20,17 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template <typename Place, typename T>
-void SetConstant<Place, T>::operator()(const platform::DeviceContext& context,
-                                       framework::Tensor* tensor, T num) {
+template <typename DeviceContext, typename T>
+void SetConstant<DeviceContext, T>::operator()(const DeviceContext& context,
+                                               framework::Tensor* tensor,
+                                               T num) {
   auto t = framework::EigenVector<T>::Flatten(*tensor);
-  t.device(*context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(num));
+  t.device(*context.eigen_device()) = t.constant(static_cast<T>(num));
 }
 
-template <typename Place, typename T, int Rank>
-void Transpose<Place, T, Rank>::operator()(
-    const platform::DeviceContext& context, const framework::Tensor& in,
+template <typename DeviceContext, typename T, int Rank>
+void Transpose<DeviceContext, T, Rank>::operator()(
+    const DeviceContext& context, const framework::Tensor& in,
     framework::Tensor* out, const std::vector<int>& axis) {
   Eigen::array<int, Rank> permute;
   for (int i = 0; i < Rank; i++) {
@@ -40,15 +41,15 @@ void Transpose<Place, T, Rank>::operator()(
 
   auto eigen_in = framework::EigenTensor<T, Rank>::From(in);
   auto eigen_out = framework::EigenTensor<T, Rank>::From(*out);
-  auto* dev = context.GetEigenDevice<Place>();
+  auto* dev = context.eigen_device();
   eigen_out.device(*dev) = eigen_in.shuffle(permute);
 }
 
-template <typename Place, typename T>
-void RowwiseAdd<Place, T>::operator()(const platform::DeviceContext& context,
-                                      const framework::Tensor& input,
-                                      const framework::Tensor& vector,
-                                      framework::Tensor* output) {
+template <typename DeviceContext, typename T>
+void RowwiseAdd<DeviceContext, T>::operator()(const DeviceContext& context,
+                                              const framework::Tensor& input,
+                                              const framework::Tensor& vector,
+                                              framework::Tensor* output) {
   auto in_dims = input.dims();
   auto size = input.numel() / in_dims[0];
   PADDLE_ENFORCE_EQ(vector.numel(), size);
@@ -59,14 +60,14 @@ void RowwiseAdd<Place, T>::operator()(const platform::DeviceContext& context,
   auto out = framework::EigenMatrix<T>::From(*output);
   Eigen::array<int, 2> shape({{1, static_cast<int>(size)}});
   Eigen::array<int, 2> bcast({{static_cast<int>(in_dims[0]), 1}});
-  out.device(*context.GetEigenDevice<Place>()) =
+  out.device(*context.eigen_device()) =
       in + vec.reshape(shape).broadcast(bcast);
 }
 
-template <typename Place, typename T>
-void ColwiseSum<Place, T>::operator()(const platform::DeviceContext& context,
-                                      const framework::Tensor& input,
-                                      framework::Tensor* vector) {
+template <typename DeviceContext, typename T>
+void ColwiseSum<DeviceContext, T>::operator()(const DeviceContext& context,
+                                              const framework::Tensor& input,
+                                              framework::Tensor* vector) {
   auto in_dims = input.dims();
   auto size = input.numel() / in_dims[0];
   PADDLE_ENFORCE_EQ(vector->numel(), size);
@@ -74,7 +75,7 @@ void ColwiseSum<Place, T>::operator()(const platform::DeviceContext& context,
   auto vec = framework::EigenMatrix<T>::From(*vector);
   auto in = framework::EigenMatrix<T>::From(input);
   Eigen::array<int, 2> shape({{1, static_cast<int>(size)}});
-  vec.reshape(shape).device(*context.GetEigenDevice<Place>()) =
+  vec.reshape(shape).device(*context.eigen_device()) =
       in.sum(Eigen::array<int, 1>({{0}})).reshape(shape);
 }
 
diff --git a/paddle/operators/math/math_function_test.cc b/paddle/operators/math/math_function_test.cc
index 983c9fdcff..7c6f098ca9 100644
--- a/paddle/operators/math/math_function_test.cc
+++ b/paddle/operators/math/math_function_test.cc
@@ -21,7 +21,7 @@ TEST(math_function, gemm_notrans_cblas) {
   memcpy(input3_ptr, arr3, 8 * sizeof(float));
 
   paddle::platform::CPUDeviceContext context(*cpu_place);
-  paddle::operators::math::gemm<paddle::platform::CPUPlace, float>(
+  paddle::operators::math::gemm<paddle::platform::CPUDeviceContext, float>(
       context, false, false, m, n, k, 1, input1_ptr, 3, input2_ptr + 1, 4, 1,
       input3_ptr + 1, 4);
 
@@ -55,7 +55,7 @@ TEST(math_function, gemm_trans_clbas) {
   memcpy(input3_ptr, arr3, 8 * sizeof(float));
 
   paddle::platform::CPUDeviceContext context(*cpu_place);
-  paddle::operators::math::gemm<paddle::platform::CPUPlace, float>(
+  paddle::operators::math::gemm<paddle::platform::CPUDeviceContext, float>(
       context, false, true, m, n, k, 1, input1_ptr, 3, input2_ptr + 3, 3, 1,
       input3_ptr + 1, 4);
 
@@ -74,7 +74,8 @@ TEST(math_function, zero) {
   auto* cpu_place = new paddle::platform::CPUPlace();
   float* t = tensor.mutable_data<float>({2, 2}, *cpu_place);
   paddle::platform::CPUDeviceContext context(*cpu_place);
-  paddle::operators::math::SetConstant<paddle::platform::CPUPlace, float>
+  paddle::operators::math::SetConstant<paddle::platform::CPUDeviceContext,
+                                       float>
       functor;
   functor(context, &tensor, 0);
   EXPECT_EQ(t[0], 0);
@@ -110,7 +111,7 @@ void GemvTest(int m, int n, bool trans) {
   }
 
   paddle::platform::CPUDeviceContext context(*cpu_place);
-  paddle::operators::math::gemv<paddle::platform::CPUPlace, T>(
+  paddle::operators::math::gemv<paddle::platform::CPUDeviceContext, T>(
       context, trans, static_cast<int>(m), static_cast<int>(n), 1., data_a,
       data_b, 0., data_c);
 
diff --git a/paddle/operators/math/math_function_test.cu b/paddle/operators/math/math_function_test.cu
index d5d6f0c73b..32e96d9487 100644
--- a/paddle/operators/math/math_function_test.cu
+++ b/paddle/operators/math/math_function_test.cu
@@ -21,7 +21,7 @@ TEST(math_function, notrans_mul_trans) {
 
   out_gpu.mutable_data<float>({2, 2}, *gpu_place);
 
-  paddle::operators::math::matmul<paddle::platform::GPUPlace, float>(
+  paddle::operators::math::matmul<paddle::platform::CUDADeviceContext, float>(
       context, input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0);
 
   paddle::framework::CopyFrom(out_gpu, *cpu_place, context, &out);
@@ -55,7 +55,7 @@ TEST(math_function, trans_mul_notrans) {
 
   out_gpu.mutable_data<float>({3, 3}, *gpu_place);
 
-  paddle::operators::math::matmul<paddle::platform::GPUPlace, float>(
+  paddle::operators::math::matmul<paddle::platform::CUDADeviceContext, float>(
       context, input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0);
 
   paddle::framework::CopyFrom(out_gpu, *cpu_place, context, &out);
@@ -106,7 +106,7 @@ TEST(math_function, gemm_notrans_cublas) {
   float* b = input2_gpu.data<float>();
   float* c = input3_gpu.mutable_data<float>(*gpu_place);
 
-  paddle::operators::math::gemm<paddle::platform::GPUPlace, float>(
+  paddle::operators::math::gemm<paddle::platform::CUDADeviceContext, float>(
       context, false, false, m, n, k, 1, a, 3, b + 1, 4, 1, c + 1, 4);
 
   paddle::framework::CopyFrom(input3_gpu, *cpu_place, context, &input3);
@@ -161,7 +161,7 @@ TEST(math_function, gemm_trans_cublas) {
   float* b = input2_gpu.data<float>();
   float* c = input3_gpu.mutable_data<float>(*gpu_place);
 
-  paddle::operators::math::gemm<paddle::platform::GPUPlace, float>(
+  paddle::operators::math::gemm<paddle::platform::CUDADeviceContext, float>(
       context, false, true, m, n, k, 1, a, 3, b + 3, 3, 1, c + 1, 4);
 
   paddle::framework::CopyFrom(input3_gpu, *cpu_place, context, &input3);
@@ -208,7 +208,7 @@ void GemvTest(int m, int n, bool trans) {
   paddle::framework::CopyFrom(mat_a, *gpu_place, context, &g_mat_a);
   paddle::framework::CopyFrom(vec_b, *gpu_place, context, &g_vec_b);
 
-  paddle::operators::math::gemv<paddle::platform::GPUPlace, T>(
+  paddle::operators::math::gemv<paddle::platform::CUDADeviceContext, T>(
       context, trans, static_cast<int>(m), static_cast<int>(n), 1., g_data_a,
       g_data_b, 0., g_data_c);
 
diff --git a/paddle/operators/math/matmul.h b/paddle/operators/math/matmul.h
index 6ba9a0ba9a..7048e11e6f 100644
--- a/paddle/operators/math/matmul.h
+++ b/paddle/operators/math/matmul.h
@@ -26,13 +26,12 @@ namespace math {
 //
 // Both a & b can be 1- to 3-dimensional. Higher rank tensors are not supported
 // yet.
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MatMulFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& a, bool trans_a,
-                  const framework::Tensor& b, bool trans_b, T alpha,
-                  framework::Tensor* out, T beta) {
+  void operator()(const DeviceContext& context, const framework::Tensor& a,
+                  bool trans_a, const framework::Tensor& b, bool trans_b,
+                  T alpha, framework::Tensor* out, T beta) {
     auto dim_a = a.dims();
     auto dim_b = b.dims();
 
@@ -108,13 +107,13 @@ class MatMulFunctor {
 
     if (!batchCount) {
       // regular matrix multiplication
-      gemm<Place, T>(context, transA, transB, M, N, kA, alpha, a.data<T>(),
-                     b.data<T>(), beta, out->data<T>());
+      gemm<DeviceContext, T>(context, transA, transB, M, N, kA, alpha,
+                             a.data<T>(), b.data<T>(), beta, out->data<T>());
     } else {
       // batched matrix multiplication
-      batched_gemm<Place, T>(context, transA, transB, M, N, kA, alpha,
-                             a.data<T>(), b.data<T>(), beta, out->data<T>(),
-                             batchCount, strideA, strideB);
+      batched_gemm<DeviceContext, T>(
+          context, transA, transB, M, N, kA, alpha, a.data<T>(), b.data<T>(),
+          beta, out->data<T>(), batchCount, strideA, strideB);
     }
   }
 };
diff --git a/paddle/operators/math/maxouting.cc b/paddle/operators/math/maxouting.cc
index c9003962d3..fea86675f7 100644
--- a/paddle/operators/math/maxouting.cc
+++ b/paddle/operators/math/maxouting.cc
@@ -20,9 +20,9 @@ namespace math {
 
 // All tensors are in NCHW format, and the groups must be greater than 1
 template <typename T>
-class MaxOutFunctor<platform::CPUPlace, T> {
+class MaxOutFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input, framework::Tensor* output,
                   int groups) {
     const int batch_size = input.dims()[0];
@@ -54,9 +54,9 @@ class MaxOutFunctor<platform::CPUPlace, T> {
 };
 
 template <class T>
-class MaxOutGradFunctor<platform::CPUPlace, T> {
+class MaxOutGradFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input, framework::Tensor* input_grad,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, int groups) {
@@ -91,10 +91,10 @@ class MaxOutGradFunctor<platform::CPUPlace, T> {
   }
 };
 
-template class MaxOutGradFunctor<platform::CPUPlace, float>;
-template class MaxOutGradFunctor<platform::CPUPlace, double>;
-template class MaxOutFunctor<platform::CPUPlace, float>;
-template class MaxOutFunctor<platform::CPUPlace, double>;
+template class MaxOutGradFunctor<platform::CPUDeviceContext, float>;
+template class MaxOutGradFunctor<platform::CPUDeviceContext, double>;
+template class MaxOutFunctor<platform::CPUDeviceContext, float>;
+template class MaxOutFunctor<platform::CPUDeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/maxouting.cu b/paddle/operators/math/maxouting.cu
index c3fabcae08..6056ad251c 100644
--- a/paddle/operators/math/maxouting.cu
+++ b/paddle/operators/math/maxouting.cu
@@ -78,9 +78,9 @@ __global__ void KernelMaxoutGrad(const int nthreads, const T* input_data,
  * All tensors are in NCHW format.
  */
 template <typename T>
-class MaxOutFunctor<platform::GPUPlace, T> {
+class MaxOutFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input, framework::Tensor* output,
                   int groups) {
     const int batch_size = input.dims()[0];
@@ -98,20 +98,18 @@ class MaxOutFunctor<platform::GPUPlace, T> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelMaxOut<
-        T><<<grid, threads, 0,
-             reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                 .stream()>>>(nthreads, input_data, input_channels,
-                              input_height, input_width, groups, output_data);
+    KernelMaxOut<T><<<grid, threads, 0, context.stream()>>>(
+        nthreads, input_data, input_channels, input_height, input_width, groups,
+        output_data);
   }
 };
 /*
  * All tensors are in NCHW format.
  */
 template <typename T>
-class MaxOutGradFunctor<platform::GPUPlace, T> {
+class MaxOutGradFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input, framework::Tensor* input_grad,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, int groups) {
@@ -132,20 +130,17 @@ class MaxOutGradFunctor<platform::GPUPlace, T> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelMaxoutGrad<
-        T><<<grid, threads, 0,
-             reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                 .stream()>>>(nthreads, input_data, output_data,
-                              output_grad_data, input_grad_data, input_channels,
-                              input_height, input_width, groups);
+    KernelMaxoutGrad<T><<<grid, threads, 0, context.stream()>>>(
+        nthreads, input_data, output_data, output_grad_data, input_grad_data,
+        input_channels, input_height, input_width, groups);
   }
 };
 
-template class MaxOutGradFunctor<platform::GPUPlace, float>;
-template class MaxOutGradFunctor<platform::GPUPlace, double>;
+template class MaxOutGradFunctor<platform::CUDADeviceContext, float>;
+template class MaxOutGradFunctor<platform::CUDADeviceContext, double>;
 
-template class MaxOutFunctor<platform::GPUPlace, float>;
-template class MaxOutFunctor<platform::GPUPlace, double>;
+template class MaxOutFunctor<platform::CUDADeviceContext, float>;
+template class MaxOutFunctor<platform::CUDADeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/maxouting.h b/paddle/operators/math/maxouting.h
index 2d9069b0b3..68f4743db0 100644
--- a/paddle/operators/math/maxouting.h
+++ b/paddle/operators/math/maxouting.h
@@ -23,20 +23,18 @@ namespace math {
 
 #define FLT_MAX __FLT_MAX__
 
-template <typename Place, typename T>
-
+template <typename DeviceContext, typename T>
 class MaxOutFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input, framework::Tensor* output,
-                  int groups);
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  framework::Tensor* output, int groups);
 };
 
-template <typename Place, class T>
+template <typename DeviceContext, class T>
 class MaxOutGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input, framework::Tensor* input_grad,
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  framework::Tensor* input_grad,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, int groups);
 };
diff --git a/paddle/operators/math/pooling.cc b/paddle/operators/math/pooling.cc
index 135984586a..150de6fd59 100644
--- a/paddle/operators/math/pooling.cc
+++ b/paddle/operators/math/pooling.cc
@@ -24,9 +24,9 @@ namespace math {
  * height and width, respectively.
  */
 template <typename PoolProcess, typename T>
-class Pool2dFunctor<platform::CPUPlace, PoolProcess, T> {
+class Pool2dFunctor<platform::CPUDeviceContext, PoolProcess, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   PoolProcess pool_process, framework::Tensor* output) {
@@ -84,9 +84,9 @@ class Pool2dFunctor<platform::CPUPlace, PoolProcess, T> {
 * and width, respectively.
 */
 template <typename PoolProcess, class T>
-class Pool2dGradFunctor<platform::CPUPlace, PoolProcess, T> {
+class Pool2dGradFunctor<platform::CPUDeviceContext, PoolProcess, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
@@ -152,9 +152,9 @@ class Pool2dGradFunctor<platform::CPUPlace, PoolProcess, T> {
  * height and width, respectively.
  */
 template <class T>
-class MaxPool2dGradFunctor<platform::CPUPlace, T> {
+class MaxPool2dGradFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
@@ -213,25 +213,29 @@ class MaxPool2dGradFunctor<platform::CPUPlace, T> {
   }
 };
 
-template class MaxPool2dGradFunctor<platform::CPUPlace, float>;
-template class MaxPool2dGradFunctor<platform::CPUPlace, double>;
+template class MaxPool2dGradFunctor<platform::CPUDeviceContext, float>;
+template class MaxPool2dGradFunctor<platform::CPUDeviceContext, double>;
 
-template class Pool2dFunctor<platform::CPUPlace,
+template class Pool2dFunctor<platform::CPUDeviceContext,
                              paddle::operators::math::MaxPool<float>, float>;
-template class Pool2dFunctor<platform::CPUPlace,
+template class Pool2dFunctor<platform::CPUDeviceContext,
                              paddle::operators::math::AvgPool<float>, float>;
-template class Pool2dGradFunctor<
-    platform::CPUPlace, paddle::operators::math::MaxPoolGrad<float>, float>;
-template class Pool2dGradFunctor<
-    platform::CPUPlace, paddle::operators::math::AvgPoolGrad<float>, float>;
-template class Pool2dFunctor<platform::CPUPlace,
+template class Pool2dGradFunctor<platform::CPUDeviceContext,
+                                 paddle::operators::math::MaxPoolGrad<float>,
+                                 float>;
+template class Pool2dGradFunctor<platform::CPUDeviceContext,
+                                 paddle::operators::math::AvgPoolGrad<float>,
+                                 float>;
+template class Pool2dFunctor<platform::CPUDeviceContext,
                              paddle::operators::math::MaxPool<double>, double>;
-template class Pool2dFunctor<platform::CPUPlace,
+template class Pool2dFunctor<platform::CPUDeviceContext,
                              paddle::operators::math::AvgPool<double>, double>;
-template class Pool2dGradFunctor<
-    platform::CPUPlace, paddle::operators::math::MaxPoolGrad<double>, double>;
-template class Pool2dGradFunctor<
-    platform::CPUPlace, paddle::operators::math::AvgPoolGrad<double>, double>;
+template class Pool2dGradFunctor<platform::CPUDeviceContext,
+                                 paddle::operators::math::MaxPoolGrad<double>,
+                                 double>;
+template class Pool2dGradFunctor<platform::CPUDeviceContext,
+                                 paddle::operators::math::AvgPoolGrad<double>,
+                                 double>;
 
 /*
  * All tensors are in NCDHW format.
@@ -239,9 +243,9 @@ template class Pool2dGradFunctor<
  * depth, height and width, respectively.
  */
 template <typename PoolProcess, class T>
-class Pool3dFunctor<platform::CPUPlace, PoolProcess, T> {
+class Pool3dFunctor<platform::CPUDeviceContext, PoolProcess, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   PoolProcess pool_process, framework::Tensor* output) {
@@ -314,9 +318,9 @@ class Pool3dFunctor<platform::CPUPlace, PoolProcess, T> {
  * depth, height and width, respectively.
  */
 template <typename PoolProcess, class T>
-class Pool3dGradFunctor<platform::CPUPlace, PoolProcess, T> {
+class Pool3dGradFunctor<platform::CPUDeviceContext, PoolProcess, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
@@ -398,9 +402,9 @@ class Pool3dGradFunctor<platform::CPUPlace, PoolProcess, T> {
  * depth, height and width, respectively.
  */
 template <class T>
-class MaxPool3dGradFunctor<platform::CPUPlace, T> {
+class MaxPool3dGradFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
@@ -473,25 +477,29 @@ class MaxPool3dGradFunctor<platform::CPUPlace, T> {
   }
 };
 
-template class MaxPool3dGradFunctor<platform::CPUPlace, float>;
-template class MaxPool3dGradFunctor<platform::CPUPlace, double>;
+template class MaxPool3dGradFunctor<platform::CPUDeviceContext, float>;
+template class MaxPool3dGradFunctor<platform::CPUDeviceContext, double>;
 
-template class Pool3dFunctor<platform::CPUPlace,
+template class Pool3dFunctor<platform::CPUDeviceContext,
                              paddle::operators::math::MaxPool<float>, float>;
-template class Pool3dFunctor<platform::CPUPlace,
+template class Pool3dFunctor<platform::CPUDeviceContext,
                              paddle::operators::math::AvgPool<float>, float>;
-template class Pool3dGradFunctor<
-    platform::CPUPlace, paddle::operators::math::MaxPoolGrad<float>, float>;
-template class Pool3dGradFunctor<
-    platform::CPUPlace, paddle::operators::math::AvgPoolGrad<float>, float>;
-template class Pool3dFunctor<platform::CPUPlace,
+template class Pool3dGradFunctor<platform::CPUDeviceContext,
+                                 paddle::operators::math::MaxPoolGrad<float>,
+                                 float>;
+template class Pool3dGradFunctor<platform::CPUDeviceContext,
+                                 paddle::operators::math::AvgPoolGrad<float>,
+                                 float>;
+template class Pool3dFunctor<platform::CPUDeviceContext,
                              paddle::operators::math::MaxPool<double>, double>;
-template class Pool3dFunctor<platform::CPUPlace,
+template class Pool3dFunctor<platform::CPUDeviceContext,
                              paddle::operators::math::AvgPool<double>, double>;
-template class Pool3dGradFunctor<
-    platform::CPUPlace, paddle::operators::math::MaxPoolGrad<double>, double>;
-template class Pool3dGradFunctor<
-    platform::CPUPlace, paddle::operators::math::AvgPoolGrad<double>, double>;
+template class Pool3dGradFunctor<platform::CPUDeviceContext,
+                                 paddle::operators::math::MaxPoolGrad<double>,
+                                 double>;
+template class Pool3dGradFunctor<platform::CPUDeviceContext,
+                                 paddle::operators::math::AvgPoolGrad<double>,
+                                 double>;
 
 /*
  * All tensors are in NCHW format.
@@ -499,9 +507,9 @@ template class Pool3dGradFunctor<
  * height and width, respectively.
  */
 template <typename T1, typename T2>
-class MaxPool2dWithIndexFunctor<platform::CPUPlace, T1, T2> {
+class MaxPool2dWithIndexFunctor<platform::CPUDeviceContext, T1, T2> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   framework::Tensor* output, framework::Tensor* mask) {
@@ -564,9 +572,9 @@ class MaxPool2dWithIndexFunctor<platform::CPUPlace, T1, T2> {
  * height and width, respectively.
  */
 template <typename T1, typename T2>
-class MaxPool2dWithIndexGradFunctor<platform::CPUPlace, T1, T2> {
+class MaxPool2dWithIndexGradFunctor<platform::CPUDeviceContext, T1, T2> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& output_grad,
                   const framework::Tensor& mask, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
@@ -602,10 +610,14 @@ class MaxPool2dWithIndexGradFunctor<platform::CPUPlace, T1, T2> {
   }
 };
 
-template class MaxPool2dWithIndexFunctor<platform::CPUPlace, float, int>;
-template class MaxPool2dWithIndexGradFunctor<platform::CPUPlace, float, int>;
-template class MaxPool2dWithIndexFunctor<platform::CPUPlace, double, int>;
-template class MaxPool2dWithIndexGradFunctor<platform::CPUPlace, double, int>;
+template class MaxPool2dWithIndexFunctor<platform::CPUDeviceContext, float,
+                                         int>;
+template class MaxPool2dWithIndexGradFunctor<platform::CPUDeviceContext, float,
+                                             int>;
+template class MaxPool2dWithIndexFunctor<platform::CPUDeviceContext, double,
+                                         int>;
+template class MaxPool2dWithIndexGradFunctor<platform::CPUDeviceContext, double,
+                                             int>;
 
 /*
  * All tensors are in NCDHW format.
@@ -613,9 +625,9 @@ template class MaxPool2dWithIndexGradFunctor<platform::CPUPlace, double, int>;
  * depth, height and width, respectively.
  */
 template <typename T1, typename T2>
-class MaxPool3dWithIndexFunctor<platform::CPUPlace, T1, T2> {
+class MaxPool3dWithIndexFunctor<platform::CPUDeviceContext, T1, T2> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   framework::Tensor* output, framework::Tensor* mask) {
@@ -692,9 +704,9 @@ class MaxPool3dWithIndexFunctor<platform::CPUPlace, T1, T2> {
  * depth, height and width, respectively.
  */
 template <typename T1, typename T2>
-class MaxPool3dWithIndexGradFunctor<platform::CPUPlace, T1, T2> {
+class MaxPool3dWithIndexGradFunctor<platform::CPUDeviceContext, T1, T2> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& output_grad,
                   const framework::Tensor& mask, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
@@ -735,10 +747,14 @@ class MaxPool3dWithIndexGradFunctor<platform::CPUPlace, T1, T2> {
   }
 };
 
-template class MaxPool3dWithIndexFunctor<platform::CPUPlace, float, int>;
-template class MaxPool3dWithIndexGradFunctor<platform::CPUPlace, float, int>;
-template class MaxPool3dWithIndexFunctor<platform::CPUPlace, double, int>;
-template class MaxPool3dWithIndexGradFunctor<platform::CPUPlace, double, int>;
+template class MaxPool3dWithIndexFunctor<platform::CPUDeviceContext, float,
+                                         int>;
+template class MaxPool3dWithIndexGradFunctor<platform::CPUDeviceContext, float,
+                                             int>;
+template class MaxPool3dWithIndexFunctor<platform::CPUDeviceContext, double,
+                                         int>;
+template class MaxPool3dWithIndexGradFunctor<platform::CPUDeviceContext, double,
+                                             int>;
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/math/pooling.cu b/paddle/operators/math/pooling.cu
index ca3560f264..0243cf8316 100644
--- a/paddle/operators/math/pooling.cu
+++ b/paddle/operators/math/pooling.cu
@@ -155,9 +155,9 @@ __global__ void KernelMaxPool2DGrad(
  * height and width, respectively.
  */
 template <typename PoolProcess, typename T>
-class Pool2dFunctor<platform::GPUPlace, PoolProcess, T> {
+class Pool2dFunctor<platform::CUDADeviceContext, PoolProcess, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   PoolProcess pool_process, framework::Tensor* output) {
@@ -183,11 +183,7 @@ class Pool2dFunctor<platform::GPUPlace, PoolProcess, T> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelPool2D<
-        PoolProcess,
-        T><<<grid, threads, 0,
-             reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                 .stream()>>>(
+    KernelPool2D<PoolProcess, T><<<grid, threads, 0, context.stream()>>>(
         nthreads, input_data, input_channels, input_height, input_width,
         output_height, output_width, ksize_height, ksize_width, stride_height,
         stride_width, padding_height, padding_width, pool_process, output_data);
@@ -200,9 +196,9 @@ class Pool2dFunctor<platform::GPUPlace, PoolProcess, T> {
  * height and width, respectively.
  */
 template <typename PoolProcess, typename T>
-class Pool2dGradFunctor<platform::GPUPlace, PoolProcess, T> {
+class Pool2dGradFunctor<platform::CUDADeviceContext, PoolProcess, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
@@ -231,11 +227,7 @@ class Pool2dGradFunctor<platform::GPUPlace, PoolProcess, T> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelPool2DGrad<
-        PoolProcess,
-        T><<<grid, threads, 0,
-             reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                 .stream()>>>(
+    KernelPool2DGrad<PoolProcess, T><<<grid, threads, 0, context.stream()>>>(
         nthreads, input_data, output_data, output_grad_data, input_channels,
         input_height, input_width, output_height, output_width, ksize_height,
         ksize_width, stride_height, stride_width, padding_height, padding_width,
@@ -249,9 +241,9 @@ class Pool2dGradFunctor<platform::GPUPlace, PoolProcess, T> {
  * height and width, respectively.
  */
 template <typename T>
-class MaxPool2dGradFunctor<platform::GPUPlace, T> {
+class MaxPool2dGradFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
@@ -281,10 +273,7 @@ class MaxPool2dGradFunctor<platform::GPUPlace, T> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelMaxPool2DGrad<
-        T><<<grid, threads, 0,
-             reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                 .stream()>>>(
+    KernelMaxPool2DGrad<T><<<grid, threads, 0, context.stream()>>>(
         nthreads, input_data, output_data, output_grad_data, input_channels,
         input_height, input_width, output_height, output_width, ksize_height,
         ksize_width, stride_height, stride_width, padding_height, padding_width,
@@ -292,25 +281,29 @@ class MaxPool2dGradFunctor<platform::GPUPlace, T> {
   }
 };
 
-template class MaxPool2dGradFunctor<platform::GPUPlace, float>;
-template class MaxPool2dGradFunctor<platform::GPUPlace, double>;
+template class MaxPool2dGradFunctor<platform::CUDADeviceContext, float>;
+template class MaxPool2dGradFunctor<platform::CUDADeviceContext, double>;
 
-template class Pool2dFunctor<platform::GPUPlace,
+template class Pool2dFunctor<platform::CUDADeviceContext,
                              paddle::operators::math::MaxPool<float>, float>;
-template class Pool2dFunctor<platform::GPUPlace,
+template class Pool2dFunctor<platform::CUDADeviceContext,
                              paddle::operators::math::AvgPool<float>, float>;
-template class Pool2dGradFunctor<
-    platform::GPUPlace, paddle::operators::math::MaxPoolGrad<float>, float>;
-template class Pool2dGradFunctor<
-    platform::GPUPlace, paddle::operators::math::AvgPoolGrad<float>, float>;
-template class Pool2dFunctor<platform::GPUPlace,
+template class Pool2dGradFunctor<platform::CUDADeviceContext,
+                                 paddle::operators::math::MaxPoolGrad<float>,
+                                 float>;
+template class Pool2dGradFunctor<platform::CUDADeviceContext,
+                                 paddle::operators::math::AvgPoolGrad<float>,
+                                 float>;
+template class Pool2dFunctor<platform::CUDADeviceContext,
                              paddle::operators::math::MaxPool<double>, double>;
-template class Pool2dFunctor<platform::GPUPlace,
+template class Pool2dFunctor<platform::CUDADeviceContext,
                              paddle::operators::math::AvgPool<double>, double>;
-template class Pool2dGradFunctor<
-    platform::GPUPlace, paddle::operators::math::MaxPoolGrad<double>, double>;
-template class Pool2dGradFunctor<
-    platform::GPUPlace, paddle::operators::math::AvgPoolGrad<double>, double>;
+template class Pool2dGradFunctor<platform::CUDADeviceContext,
+                                 paddle::operators::math::MaxPoolGrad<double>,
+                                 double>;
+template class Pool2dGradFunctor<platform::CUDADeviceContext,
+                                 paddle::operators::math::AvgPoolGrad<double>,
+                                 double>;
 
 template <typename PoolProcess, typename T>
 __global__ void KernelPool3D(const int nthreads, const T* input_data,
@@ -478,9 +471,9 @@ __global__ void KernelMaxPool3DGrad(
  * depth, height and width, respectively.
  */
 template <typename PoolProcess, class T>
-class Pool3dFunctor<platform::GPUPlace, PoolProcess, T> {
+class Pool3dFunctor<platform::CUDADeviceContext, PoolProcess, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   PoolProcess pool_process, framework::Tensor* output) {
@@ -512,11 +505,7 @@ class Pool3dFunctor<platform::GPUPlace, PoolProcess, T> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelPool3D<
-        PoolProcess,
-        T><<<grid, threads, 0,
-             reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                 .stream()>>>(
+    KernelPool3D<PoolProcess, T><<<grid, threads, 0, context.stream()>>>(
         nthreads, input_data, input_channels, input_depth, input_height,
         input_width, output_depth, output_height, output_width, ksize_depth,
         ksize_height, ksize_width, stride_depth, stride_height, stride_width,
@@ -531,9 +520,9 @@ class Pool3dFunctor<platform::GPUPlace, PoolProcess, T> {
  * depth, height and width, respectively.
  */
 template <typename PoolProcess, class T>
-class Pool3dGradFunctor<platform::GPUPlace, PoolProcess, T> {
+class Pool3dGradFunctor<platform::CUDADeviceContext, PoolProcess, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
@@ -569,11 +558,7 @@ class Pool3dGradFunctor<platform::GPUPlace, PoolProcess, T> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelPool3DGrad<
-        PoolProcess,
-        T><<<grid, threads, 0,
-             reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                 .stream()>>>(
+    KernelPool3DGrad<PoolProcess, T><<<grid, threads, 0, context.stream()>>>(
         nthreads, input_data, output_data, output_grad_data, input_channels,
         input_depth, input_height, input_width, output_depth, output_height,
         output_width, ksize_depth, ksize_height, ksize_width, stride_depth,
@@ -588,9 +573,9 @@ class Pool3dGradFunctor<platform::GPUPlace, PoolProcess, T> {
  * depth, height and width, respectively.
  */
 template <class T>
-class MaxPool3dGradFunctor<platform::GPUPlace, T> {
+class MaxPool3dGradFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
@@ -626,10 +611,7 @@ class MaxPool3dGradFunctor<platform::GPUPlace, T> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelMaxPool3DGrad<
-        T><<<grid, threads, 0,
-             reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                 .stream()>>>(
+    KernelMaxPool3DGrad<T><<<grid, threads, 0, context.stream()>>>(
         nthreads, input_data, output_data, output_grad_data, input_channels,
         input_depth, input_height, input_width, output_depth, output_height,
         output_width, ksize_depth, ksize_height, ksize_width, stride_depth,
@@ -638,25 +620,29 @@ class MaxPool3dGradFunctor<platform::GPUPlace, T> {
   }
 };
 
-template class MaxPool3dGradFunctor<platform::GPUPlace, float>;
-template class MaxPool3dGradFunctor<platform::GPUPlace, double>;
+template class MaxPool3dGradFunctor<platform::CUDADeviceContext, float>;
+template class MaxPool3dGradFunctor<platform::CUDADeviceContext, double>;
 
-template class Pool3dFunctor<platform::GPUPlace,
+template class Pool3dFunctor<platform::CUDADeviceContext,
                              paddle::operators::math::MaxPool<float>, float>;
-template class Pool3dFunctor<platform::GPUPlace,
+template class Pool3dFunctor<platform::CUDADeviceContext,
                              paddle::operators::math::AvgPool<float>, float>;
-template class Pool3dGradFunctor<
-    platform::GPUPlace, paddle::operators::math::MaxPoolGrad<float>, float>;
-template class Pool3dGradFunctor<
-    platform::GPUPlace, paddle::operators::math::AvgPoolGrad<float>, float>;
-template class Pool3dFunctor<platform::GPUPlace,
+template class Pool3dGradFunctor<platform::CUDADeviceContext,
+                                 paddle::operators::math::MaxPoolGrad<float>,
+                                 float>;
+template class Pool3dGradFunctor<platform::CUDADeviceContext,
+                                 paddle::operators::math::AvgPoolGrad<float>,
+                                 float>;
+template class Pool3dFunctor<platform::CUDADeviceContext,
                              paddle::operators::math::MaxPool<double>, double>;
-template class Pool3dFunctor<platform::GPUPlace,
+template class Pool3dFunctor<platform::CUDADeviceContext,
                              paddle::operators::math::AvgPool<double>, double>;
-template class Pool3dGradFunctor<
-    platform::GPUPlace, paddle::operators::math::MaxPoolGrad<double>, double>;
-template class Pool3dGradFunctor<
-    platform::GPUPlace, paddle::operators::math::AvgPoolGrad<double>, double>;
+template class Pool3dGradFunctor<platform::CUDADeviceContext,
+                                 paddle::operators::math::MaxPoolGrad<double>,
+                                 double>;
+template class Pool3dGradFunctor<platform::CUDADeviceContext,
+                                 paddle::operators::math::AvgPoolGrad<double>,
+                                 double>;
 
 template <typename T1, typename T2>
 __global__ void KernelMaxPool2dWithIdx(
@@ -747,9 +733,9 @@ __global__ void KernelMaxPool2DWithIdxGrad(
  * height and width, respectively.
  */
 template <typename T1, typename T2>
-class MaxPool2dWithIndexFunctor<platform::GPUPlace, T1, T2> {
+class MaxPool2dWithIndexFunctor<platform::CUDADeviceContext, T1, T2> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   framework::Tensor* output, framework::Tensor* mask) {
@@ -776,10 +762,7 @@ class MaxPool2dWithIndexFunctor<platform::GPUPlace, T1, T2> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelMaxPool2dWithIdx<
-        T1, T2><<<grid, threads, 0,
-                  reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                      .stream()>>>(
+    KernelMaxPool2dWithIdx<T1, T2><<<grid, threads, 0, context.stream()>>>(
         nthreads, input_data, input_channels, input_height, input_width,
         output_height, output_width, ksize_height, ksize_width, stride_height,
         stride_width, padding_height, padding_width, output_data, mask_data);
@@ -792,9 +775,9 @@ class MaxPool2dWithIndexFunctor<platform::GPUPlace, T1, T2> {
  * height and width, respectively.
  */
 template <typename T1, typename T2>
-class MaxPool2dWithIndexGradFunctor<platform::GPUPlace, T1, T2> {
+class MaxPool2dWithIndexGradFunctor<platform::CUDADeviceContext, T1, T2> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& output_grad,
                   const framework::Tensor& mask, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
@@ -821,10 +804,7 @@ class MaxPool2dWithIndexGradFunctor<platform::GPUPlace, T1, T2> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelMaxPool2DWithIdxGrad<
-        T1, T2><<<grid, threads, 0,
-                  reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                      .stream()>>>(
+    KernelMaxPool2DWithIdxGrad<T1, T2><<<grid, threads, 0, context.stream()>>>(
         nthreads, output_grad_data, mask_data, input_channels, input_height,
         input_width, output_height, output_width, ksize_height, ksize_width,
         stride_height, stride_width, padding_height, padding_width,
@@ -832,10 +812,14 @@ class MaxPool2dWithIndexGradFunctor<platform::GPUPlace, T1, T2> {
   }
 };
 
-template class MaxPool2dWithIndexFunctor<platform::GPUPlace, float, int>;
-template class MaxPool2dWithIndexGradFunctor<platform::GPUPlace, float, int>;
-template class MaxPool2dWithIndexFunctor<platform::GPUPlace, double, int>;
-template class MaxPool2dWithIndexGradFunctor<platform::GPUPlace, double, int>;
+template class MaxPool2dWithIndexFunctor<platform::CUDADeviceContext, float,
+                                         int>;
+template class MaxPool2dWithIndexGradFunctor<platform::CUDADeviceContext, float,
+                                             int>;
+template class MaxPool2dWithIndexFunctor<platform::CUDADeviceContext, double,
+                                         int>;
+template class MaxPool2dWithIndexGradFunctor<platform::CUDADeviceContext,
+                                             double, int>;
 
 template <typename T1, typename T2>
 __global__ void KernelMaxPool3DWithIdx(
@@ -950,9 +934,9 @@ __global__ void KernelMaxPool3DWithIdxGrad(
  * depth, height and width, respectively.
  */
 template <typename T1, typename T2>
-class MaxPool3dWithIndexFunctor<platform::GPUPlace, T1, T2> {
+class MaxPool3dWithIndexFunctor<platform::CUDADeviceContext, T1, T2> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   framework::Tensor* output, framework::Tensor* mask) {
@@ -985,10 +969,7 @@ class MaxPool3dWithIndexFunctor<platform::GPUPlace, T1, T2> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelMaxPool3DWithIdx<
-        T1, T2><<<grid, threads, 0,
-                  reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                      .stream()>>>(
+    KernelMaxPool3DWithIdx<T1, T2><<<grid, threads, 0, context.stream()>>>(
         nthreads, input_data, input_channels, input_depth, input_height,
         input_width, output_depth, output_height, output_width, ksize_depth,
         ksize_height, ksize_width, stride_depth, stride_height, stride_width,
@@ -1002,9 +983,9 @@ class MaxPool3dWithIndexFunctor<platform::GPUPlace, T1, T2> {
  * depth, height and width, respectively.
  */
 template <typename T1, typename T2>
-class MaxPool3dWithIndexGradFunctor<platform::GPUPlace, T1, T2> {
+class MaxPool3dWithIndexGradFunctor<platform::CUDADeviceContext, T1, T2> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& output_grad,
                   const framework::Tensor& mask, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
@@ -1037,10 +1018,7 @@ class MaxPool3dWithIndexGradFunctor<platform::GPUPlace, T1, T2> {
     dim3 threads(1024, 1);
     dim3 grid(blocks, 1);
 
-    KernelMaxPool3DWithIdxGrad<
-        T1, T2><<<grid, threads, 0,
-                  reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                      .stream()>>>(
+    KernelMaxPool3DWithIdxGrad<T1, T2><<<grid, threads, 0, context.stream()>>>(
         nthreads, output_grad_data, mask_data, input_channels, input_depth,
         input_height, input_width, output_depth, output_height, output_width,
         ksize_depth, ksize_height, ksize_width, stride_depth, stride_height,
@@ -1049,10 +1027,14 @@ class MaxPool3dWithIndexGradFunctor<platform::GPUPlace, T1, T2> {
   }
 };
 
-template class MaxPool3dWithIndexFunctor<platform::GPUPlace, float, int>;
-template class MaxPool3dWithIndexGradFunctor<platform::GPUPlace, float, int>;
-template class MaxPool3dWithIndexFunctor<platform::GPUPlace, double, int>;
-template class MaxPool3dWithIndexGradFunctor<platform::GPUPlace, double, int>;
+template class MaxPool3dWithIndexFunctor<platform::CUDADeviceContext, float,
+                                         int>;
+template class MaxPool3dWithIndexGradFunctor<platform::CUDADeviceContext, float,
+                                             int>;
+template class MaxPool3dWithIndexFunctor<platform::CUDADeviceContext, double,
+                                         int>;
+template class MaxPool3dWithIndexGradFunctor<platform::CUDADeviceContext,
+                                             double, int>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/pooling.h b/paddle/operators/math/pooling.h
index 19fbd8b4bb..2759f06cb6 100644
--- a/paddle/operators/math/pooling.h
+++ b/paddle/operators/math/pooling.h
@@ -84,62 +84,58 @@ class AvgPoolGrad {
  * This is different from average pooling. So we rewrite the max_pool_grad:
  * MaxPool2dGradFunctor, MaxPool3dGradFunctor.
  */
-template <typename Place, typename PoolProcess, typename T>
+template <typename DeviceContext, typename PoolProcess, typename T>
 class Pool2dFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input, std::vector<int>& ksize,
-                  std::vector<int>& strides, std::vector<int>& paddings,
-                  PoolProcess pool_compute, framework::Tensor* output);
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  std::vector<int>& ksize, std::vector<int>& strides,
+                  std::vector<int>& paddings, PoolProcess pool_compute,
+                  framework::Tensor* output);
 };
 
-template <typename Place, typename PoolProcess, typename T>
+template <typename DeviceContext, typename PoolProcess, typename T>
 class Pool2dGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input,
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   PoolProcess pool_compute, framework::Tensor* input_grad);
 };
 
-template <typename Place, class T>
+template <typename DeviceContext, class T>
 class MaxPool2dGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input,
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   framework::Tensor* input_grad);
 };
 
-template <typename Place, typename PoolProcess, typename T>
+template <typename DeviceContext, typename PoolProcess, typename T>
 class Pool3dFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input, std::vector<int>& ksize,
-                  std::vector<int>& strides, std::vector<int>& paddings,
-                  PoolProcess pool_compute, framework::Tensor* output);
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  std::vector<int>& ksize, std::vector<int>& strides,
+                  std::vector<int>& paddings, PoolProcess pool_compute,
+                  framework::Tensor* output);
 };
 
-template <typename Place, typename PoolProcess, typename T>
+template <typename DeviceContext, typename PoolProcess, typename T>
 class Pool3dGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input,
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   PoolProcess pool_compute, framework::Tensor* input_grad);
 };
 
-template <typename Place, class T>
+template <typename DeviceContext, class T>
 class MaxPool3dGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input,
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
@@ -153,38 +149,38 @@ class MaxPool3dGradFunctor {
  * In pool2d, all tensors are in NCHW format. In pool3d, all tensors are in
  * NCDHW format.
  */
-template <typename Place, typename T1, typename T2>
+template <typename DeviceContext, typename T1, typename T2>
 class MaxPool2dWithIndexFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input, std::vector<int>& ksize,
-                  std::vector<int>& strides, std::vector<int>& paddings,
-                  framework::Tensor* output, framework::Tensor* mask);
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  std::vector<int>& ksize, std::vector<int>& strides,
+                  std::vector<int>& paddings, framework::Tensor* output,
+                  framework::Tensor* mask);
 };
 
-template <typename Place, typename T1, typename T2>
+template <typename DeviceContext, typename T1, typename T2>
 class MaxPool2dWithIndexGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::Tensor& output_grad,
                   const framework::Tensor& mask, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
                   framework::Tensor* input_grad);
 };
 
-template <typename Place, typename T1, typename T2>
+template <typename DeviceContext, typename T1, typename T2>
 class MaxPool3dWithIndexFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input, std::vector<int>& ksize,
-                  std::vector<int>& strides, std::vector<int>& paddings,
-                  framework::Tensor* output, framework::Tensor* mask);
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  std::vector<int>& ksize, std::vector<int>& strides,
+                  std::vector<int>& paddings, framework::Tensor* output,
+                  framework::Tensor* mask);
 };
 
-template <typename Place, typename T1, typename T2>
+template <typename DeviceContext, typename T1, typename T2>
 class MaxPool3dWithIndexGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::Tensor& output_grad,
                   const framework::Tensor& mask, std::vector<int>& ksize,
                   std::vector<int>& strides, std::vector<int>& paddings,
diff --git a/paddle/operators/math/selected_rows_functor.cc b/paddle/operators/math/selected_rows_functor.cc
index 514f2adef2..ab758d1e7f 100644
--- a/paddle/operators/math/selected_rows_functor.cc
+++ b/paddle/operators/math/selected_rows_functor.cc
@@ -19,8 +19,8 @@ namespace paddle {
 namespace operators {
 namespace math {
 template <typename T>
-struct SelectedRowsAdd<platform::CPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SelectedRowsAdd<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::SelectedRows& input1,
                   const framework::SelectedRows& input2,
                   framework::SelectedRows* output) {
@@ -67,12 +67,12 @@ struct SelectedRowsAdd<platform::CPUPlace, T> {
   }
 };
 
-template struct SelectedRowsAdd<platform::CPUPlace, float>;
-template struct SelectedRowsAdd<platform::CPUPlace, double>;
+template struct SelectedRowsAdd<platform::CPUDeviceContext, float>;
+template struct SelectedRowsAdd<platform::CPUDeviceContext, double>;
 
 template <typename T>
-struct SelectedRowsAddTensor<platform::CPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SelectedRowsAddTensor<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::SelectedRows& input1,
                   const framework::Tensor& input2, framework::Tensor* output) {
     auto in1_height = input1.height();
@@ -88,7 +88,7 @@ struct SelectedRowsAddTensor<platform::CPUPlace, T> {
     PADDLE_ENFORCE_EQ(in1_row_numel, input2.numel() / in1_height);
     PADDLE_ENFORCE_EQ(in1_row_numel, output->numel() / in1_height);
 
-    SetConstant<platform::CPUPlace, T> functor;
+    SetConstant<platform::CPUDeviceContext, T> functor;
     functor(context, output, 0.0);
 
     auto* in1_data = in1_value.data<T>();
@@ -103,17 +103,16 @@ struct SelectedRowsAddTensor<platform::CPUPlace, T> {
 
     auto out_eigen = framework::EigenVector<T>::Flatten(*output);
     auto in2_eigen = framework::EigenVector<T>::Flatten(input2);
-    out_eigen.device(*context.GetEigenDevice<platform::CPUPlace>()) =
-        out_eigen + in2_eigen;
+    out_eigen.device(*context.eigen_device()) = out_eigen + in2_eigen;
   }
 };
 
-template struct SelectedRowsAddTensor<platform::CPUPlace, float>;
-template struct SelectedRowsAddTensor<platform::CPUPlace, double>;
+template struct SelectedRowsAddTensor<platform::CPUDeviceContext, float>;
+template struct SelectedRowsAddTensor<platform::CPUDeviceContext, double>;
 
 template <typename T>
-struct SelectedRowsAddTo<platform::CPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SelectedRowsAddTo<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::SelectedRows& input1,
                   const int64_t input2_offset,
                   framework::SelectedRows* input2) {
@@ -143,14 +142,14 @@ struct SelectedRowsAddTo<platform::CPUPlace, T> {
   }
 };
 
-template struct SelectedRowsAddTo<platform::CPUPlace, float>;
-template struct SelectedRowsAddTo<platform::CPUPlace, double>;
-template struct SelectedRowsAddTo<platform::CPUPlace, int>;
-template struct SelectedRowsAddTo<platform::CPUPlace, int64_t>;
+template struct SelectedRowsAddTo<platform::CPUDeviceContext, float>;
+template struct SelectedRowsAddTo<platform::CPUDeviceContext, double>;
+template struct SelectedRowsAddTo<platform::CPUDeviceContext, int>;
+template struct SelectedRowsAddTo<platform::CPUDeviceContext, int64_t>;
 
 template <typename T>
-struct SelectedRowsAddToTensor<platform::CPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SelectedRowsAddToTensor<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::SelectedRows& input1,
                   framework::Tensor* input2) {
     auto in1_height = input1.height();
@@ -175,10 +174,10 @@ struct SelectedRowsAddToTensor<platform::CPUPlace, T> {
   }
 };
 
-template struct SelectedRowsAddToTensor<platform::CPUPlace, float>;
-template struct SelectedRowsAddToTensor<platform::CPUPlace, double>;
-template struct SelectedRowsAddToTensor<platform::CPUPlace, int>;
-template struct SelectedRowsAddToTensor<platform::CPUPlace, int64_t>;
+template struct SelectedRowsAddToTensor<platform::CPUDeviceContext, float>;
+template struct SelectedRowsAddToTensor<platform::CPUDeviceContext, double>;
+template struct SelectedRowsAddToTensor<platform::CPUDeviceContext, int>;
+template struct SelectedRowsAddToTensor<platform::CPUDeviceContext, int64_t>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/selected_rows_functor.cu b/paddle/operators/math/selected_rows_functor.cu
index c1dd323ba2..c44577e00a 100644
--- a/paddle/operators/math/selected_rows_functor.cu
+++ b/paddle/operators/math/selected_rows_functor.cu
@@ -20,8 +20,8 @@ namespace paddle {
 namespace operators {
 namespace math {
 template <typename T>
-struct SelectedRowsAdd<platform::GPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SelectedRowsAdd<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::SelectedRows& input1,
                   const framework::SelectedRows& input2,
                   framework::SelectedRows* output) {
@@ -64,16 +64,15 @@ struct SelectedRowsAdd<platform::GPUPlace, T> {
         reinterpret_cast<const platform::CUDADeviceContext&>(context).stream());
 
     auto* in2_data = in2_value.data<T>();
-    memory::Copy(
-        boost::get<platform::GPUPlace>(out_place), out_data + in1_value.numel(),
-        boost::get<platform::GPUPlace>(in2_place), in2_data,
-        in2_value.numel() * sizeof(T),
-        reinterpret_cast<const platform::CUDADeviceContext&>(context).stream());
+    memory::Copy(boost::get<platform::GPUPlace>(out_place),
+                 out_data + in1_value.numel(),
+                 boost::get<platform::GPUPlace>(in2_place), in2_data,
+                 in2_value.numel() * sizeof(T), context.stream());
   }
 };
 
-template struct SelectedRowsAdd<platform::GPUPlace, float>;
-template struct SelectedRowsAdd<platform::GPUPlace, double>;
+template struct SelectedRowsAdd<platform::CUDADeviceContext, float>;
+template struct SelectedRowsAdd<platform::CUDADeviceContext, double>;
 
 namespace {
 template <typename T, int block_size>
@@ -96,8 +95,8 @@ __global__ void SelectedRowsAddTensorKernel(const T* selected_rows,
 }  // namespace
 
 template <typename T>
-struct SelectedRowsAddTensor<platform::GPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SelectedRowsAddTensor<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::SelectedRows& input1,
                   const framework::Tensor& input2, framework::Tensor* output) {
     auto in1_height = input1.height();
@@ -117,30 +116,28 @@ struct SelectedRowsAddTensor<platform::GPUPlace, T> {
     auto* in2_data = input2.data<T>();
     auto* out_data = output->data<T>();
 
-    SetConstant<platform::GPUPlace, T> functor;
+    SetConstant<platform::CUDADeviceContext, T> functor;
     functor(context, output, 0.0);
 
     const int block_size = 256;
     dim3 threads(block_size, 1);
     dim3 grid(1, in1_rows.size());
-    SelectedRowsAddTensorKernel<T, block_size><<<
-        grid, threads, 0,
-        reinterpret_cast<const platform::CUDADeviceContext&>(context)
-            .stream()>>>(in1_data, in1_rows.data(), out_data, in1_row_numel);
+    SelectedRowsAddTensorKernel<
+        T, block_size><<<grid, threads, 0, context.stream()>>>(
+        in1_data, in1_rows.data(), out_data, in1_row_numel);
 
     auto out_eigen = framework::EigenVector<T>::Flatten(*output);
     auto in2_eigen = framework::EigenVector<T>::Flatten(input2);
-    out_eigen.device(*context.GetEigenDevice<platform::GPUPlace>()) =
-        out_eigen + in2_eigen;
+    out_eigen.device(*context.eigen_device()) = out_eigen + in2_eigen;
   }
 };
 
-template struct SelectedRowsAddTensor<platform::GPUPlace, float>;
-template struct SelectedRowsAddTensor<platform::GPUPlace, double>;
+template struct SelectedRowsAddTensor<platform::CUDADeviceContext, float>;
+template struct SelectedRowsAddTensor<platform::CUDADeviceContext, double>;
 
 template <typename T>
-struct SelectedRowsAddTo<platform::GPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SelectedRowsAddTo<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::SelectedRows& input1,
                   const int64_t input2_offset,
                   framework::SelectedRows* input2) {
@@ -163,18 +160,17 @@ struct SelectedRowsAddTo<platform::GPUPlace, T> {
 
     auto* in1_data = in1_value.data<T>();
     auto* in2_data = in2_value->data<T>();
-    memory::Copy(
-        boost::get<platform::GPUPlace>(in2_place), in2_data + input2_offset,
-        boost::get<platform::GPUPlace>(in1_place), in1_data,
-        in1_value.numel() * sizeof(T),
-        reinterpret_cast<const platform::CUDADeviceContext&>(context).stream());
+    memory::Copy(boost::get<platform::GPUPlace>(in2_place),
+                 in2_data + input2_offset,
+                 boost::get<platform::GPUPlace>(in1_place), in1_data,
+                 in1_value.numel() * sizeof(T), context.stream());
   }
 };
 
-template struct SelectedRowsAddTo<platform::GPUPlace, float>;
-template struct SelectedRowsAddTo<platform::GPUPlace, double>;
-template struct SelectedRowsAddTo<platform::GPUPlace, int>;
-template struct SelectedRowsAddTo<platform::GPUPlace, int64_t>;
+template struct SelectedRowsAddTo<platform::CUDADeviceContext, float>;
+template struct SelectedRowsAddTo<platform::CUDADeviceContext, double>;
+template struct SelectedRowsAddTo<platform::CUDADeviceContext, int>;
+template struct SelectedRowsAddTo<platform::CUDADeviceContext, int64_t>;
 
 namespace {
 template <typename T, int block_size>
@@ -197,8 +193,8 @@ __global__ void SelectedRowsAddToTensorKernel(const T* selected_rows,
 }  // namespace
 
 template <typename T>
-struct SelectedRowsAddToTensor<platform::GPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SelectedRowsAddToTensor<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::SelectedRows& input1,
                   framework::Tensor* input2) {
     auto in1_height = input1.height();
@@ -216,17 +212,16 @@ struct SelectedRowsAddToTensor<platform::GPUPlace, T> {
     const int block_size = 256;
     dim3 threads(block_size, 1);
     dim3 grid(1, in1_rows.size());
-    SelectedRowsAddToTensorKernel<T, block_size><<<
-        grid, threads, 0,
-        reinterpret_cast<const platform::CUDADeviceContext&>(context)
-            .stream()>>>(in1_data, in1_rows.data(), in2_data, in1_row_numel);
+    SelectedRowsAddToTensorKernel<
+        T, block_size><<<grid, threads, 0, context.stream()>>>(
+        in1_data, in1_rows.data(), in2_data, in1_row_numel);
   }
 };
 
-template struct SelectedRowsAddToTensor<platform::GPUPlace, float>;
-template struct SelectedRowsAddToTensor<platform::GPUPlace, double>;
-template struct SelectedRowsAddToTensor<platform::GPUPlace, int>;
-template struct SelectedRowsAddToTensor<platform::GPUPlace, int64_t>;
+template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, float>;
+template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, double>;
+template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, int>;
+template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, int64_t>;
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/math/selected_rows_functor.h b/paddle/operators/math/selected_rows_functor.h
index d6dc6c03c9..1149075abf 100644
--- a/paddle/operators/math/selected_rows_functor.h
+++ b/paddle/operators/math/selected_rows_functor.h
@@ -21,33 +21,33 @@ namespace math {
 
 // SelectedRows + SelectedRows will simplely concat value and rows.
 // The real computation happens in dealing with LoDTensor.
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct SelectedRowsAdd {
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::SelectedRows& input1,
                   const framework::SelectedRows& input2,
                   framework::SelectedRows* output);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct SelectedRowsAddTensor {
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::SelectedRows& input1,
                   const framework::Tensor& input2, framework::Tensor* output);
 };
 
 // input2 = input1 + input2
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct SelectedRowsAddTo {
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::SelectedRows& input1,
                   const int64_t input2_offset, framework::SelectedRows* input2);
 };
 
 // input2 = input1 + input2
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct SelectedRowsAddToTensor {
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::SelectedRows& input1,
                   framework::Tensor* input2);
 };
diff --git a/paddle/operators/math/selected_rows_functor_test.cc b/paddle/operators/math/selected_rows_functor_test.cc
index a3649b6875..8c74cab0a1 100644
--- a/paddle/operators/math/selected_rows_functor_test.cc
+++ b/paddle/operators/math/selected_rows_functor_test.cc
@@ -23,7 +23,7 @@ TEST(selected_rows_functor, cpu_add) {
 
   CPUPlace cpu_place;
   CPUDeviceContext ctx(cpu_place);
-  SetConstant<CPUPlace, float> functor;
+  SetConstant<CPUDeviceContext, float> functor;
   int64_t height = 10;
   int64_t row_numel = 10;
 
@@ -47,7 +47,7 @@ TEST(selected_rows_functor, cpu_add) {
   // simplely concat two SelectedRows
   out_value->mutable_data<float>(make_ddim({7, 10}), cpu_place);
 
-  SelectedRowsAdd<CPUPlace, float> add_functor;
+  SelectedRowsAdd<CPUDeviceContext, float> add_functor;
   add_functor(ctx, *selected_rows1, *selected_rows2, output.get());
 
   auto out_height = output->height();
@@ -85,7 +85,7 @@ TEST(selected_rows_functor, cpu_add) {
   std::unique_ptr<Tensor> tensor2{new Tensor()};
   tensor2->mutable_data<float>(make_ddim({height, row_numel}), cpu_place);
 
-  SelectedRowsAddTensor<CPUPlace, float> add_tensor_functor;
+  SelectedRowsAddTensor<CPUDeviceContext, float> add_tensor_functor;
   add_tensor_functor(ctx, *output, *tensor1, tensor2.get());
 
   auto* tensor2_data = tensor2->data<float>();
@@ -112,7 +112,7 @@ TEST(selected_rows_functor, cpu_add_to) {
 
   CPUPlace cpu_place;
   CPUDeviceContext ctx(cpu_place);
-  SetConstant<CPUPlace, float> functor;
+  SetConstant<CPUDeviceContext, float> functor;
   int64_t height = 10;
   int64_t row_numel = 10;
 
@@ -137,7 +137,7 @@ TEST(selected_rows_functor, cpu_add_to) {
   // simplely concat two SelectedRows
   out_value->mutable_data<float>(make_ddim({7, 10}), cpu_place);
 
-  SelectedRowsAddTo<CPUPlace, float> add_to_functor;
+  SelectedRowsAddTo<CPUDeviceContext, float> add_to_functor;
   add_to_functor(ctx, *selected_rows1, 0, output.get());
   add_to_functor(ctx, *selected_rows2, in1_value->numel(), output.get());
 
@@ -173,7 +173,7 @@ TEST(selected_rows_functor, cpu_add_to) {
   tensor1->mutable_data<float>(make_ddim({height, row_numel}), cpu_place);
   functor(ctx, tensor1.get(), 3.0);
 
-  SelectedRowsAddToTensor<CPUPlace, float> add_to_tensor_functor;
+  SelectedRowsAddToTensor<CPUDeviceContext, float> add_to_tensor_functor;
   add_to_tensor_functor(ctx, *output, tensor1.get());
 
   auto* tensor1_data = tensor1->data<float>();
diff --git a/paddle/operators/math/selected_rows_functor_test.cu b/paddle/operators/math/selected_rows_functor_test.cu
index 7de9291c17..777caf5635 100644
--- a/paddle/operators/math/selected_rows_functor_test.cu
+++ b/paddle/operators/math/selected_rows_functor_test.cu
@@ -24,7 +24,7 @@ TEST(selected_rows_functor, gpu_add) {
   GPUPlace gpu_place(0);
   CPUPlace cpu_place;
   CUDADeviceContext ctx(gpu_place);
-  SetConstant<GPUPlace, float> functor;
+  SetConstant<CUDADeviceContext, float> functor;
   int64_t height = 10;
   int64_t row_numel = 10;
 
@@ -48,7 +48,7 @@ TEST(selected_rows_functor, gpu_add) {
   // simplely concat two SelectedRows
   out_value->mutable_data<float>(make_ddim({7, 10}), gpu_place);
 
-  SelectedRowsAdd<GPUPlace, float> add_functor;
+  SelectedRowsAdd<CUDADeviceContext, float> add_functor;
   add_functor(ctx, *selected_rows1, *selected_rows2, output.get());
 
   auto out_height = output->height();
@@ -90,7 +90,7 @@ TEST(selected_rows_functor, gpu_add) {
   std::unique_ptr<Tensor> tensor2{new Tensor()};
   tensor2->mutable_data<float>(make_ddim({height, row_numel}), gpu_place);
 
-  SelectedRowsAddTensor<GPUPlace, float> add_tensor_functor;
+  SelectedRowsAddTensor<CUDADeviceContext, float> add_tensor_functor;
   add_tensor_functor(ctx, *output, *tensor1, tensor2.get());
 
   Tensor tensor2_cpu;
@@ -122,7 +122,7 @@ TEST(selected_rows_functor, gpu_add_to) {
   GPUPlace gpu_place(0);
   CPUPlace cpu_place;
   CUDADeviceContext ctx(gpu_place);
-  SetConstant<GPUPlace, float> functor;
+  SetConstant<CUDADeviceContext, float> functor;
   int64_t height = 10;
   int64_t row_numel = 10;
 
@@ -147,7 +147,7 @@ TEST(selected_rows_functor, gpu_add_to) {
   // simplely concat two SelectedRows
   out_value->mutable_data<float>(make_ddim({7, 10}), gpu_place);
 
-  SelectedRowsAddTo<GPUPlace, float> add_to_functor;
+  SelectedRowsAddTo<CUDADeviceContext, float> add_to_functor;
   add_to_functor(ctx, *selected_rows1, 0, output.get());
   add_to_functor(ctx, *selected_rows2, in1_value->numel(), output.get());
 
@@ -187,7 +187,7 @@ TEST(selected_rows_functor, gpu_add_to) {
   tensor1->mutable_data<float>(make_ddim({height, row_numel}), gpu_place);
   functor(ctx, tensor1.get(), 3.0);
 
-  SelectedRowsAddToTensor<GPUPlace, float> add_to_tensor_functor;
+  SelectedRowsAddToTensor<CUDADeviceContext, float> add_to_tensor_functor;
   add_to_tensor_functor(ctx, *output, tensor1.get());
 
   Tensor tensor1_cpu;
diff --git a/paddle/operators/math/sequence2batch.cc b/paddle/operators/math/sequence2batch.cc
index 5b3bde02fb..88977be1f8 100644
--- a/paddle/operators/math/sequence2batch.cc
+++ b/paddle/operators/math/sequence2batch.cc
@@ -19,9 +19,9 @@ namespace operators {
 namespace math {
 
 template <typename T>
-class CopyMatrixRowsFunctor<platform::CPUPlace, T> {
+class CopyMatrixRowsFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& src, const size_t* index,
                   framework::Tensor& dst, bool is_src_index) {
     auto src_dims = src.dims();
@@ -48,13 +48,13 @@ class CopyMatrixRowsFunctor<platform::CPUPlace, T> {
   }
 };
 
-template class CopyMatrixRowsFunctor<platform::CPUPlace, float>;
-template class CopyMatrixRowsFunctor<platform::CPUPlace, double>;
+template class CopyMatrixRowsFunctor<platform::CPUDeviceContext, float>;
+template class CopyMatrixRowsFunctor<platform::CPUDeviceContext, double>;
 
-template class LoDTensor2BatchFunctor<platform::CPUPlace, float>;
-template class LoDTensor2BatchFunctor<platform::CPUPlace, double>;
-template class Batch2LoDTensorFunctor<platform::CPUPlace, float>;
-template class Batch2LoDTensorFunctor<platform::CPUPlace, double>;
+template class LoDTensor2BatchFunctor<platform::CPUDeviceContext, float>;
+template class LoDTensor2BatchFunctor<platform::CPUDeviceContext, double>;
+template class Batch2LoDTensorFunctor<platform::CPUDeviceContext, float>;
+template class Batch2LoDTensorFunctor<platform::CPUDeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/sequence2batch.cu b/paddle/operators/math/sequence2batch.cu
index c5d968aeb2..452ae89510 100644
--- a/paddle/operators/math/sequence2batch.cu
+++ b/paddle/operators/math/sequence2batch.cu
@@ -39,9 +39,9 @@ __global__ void CopyMatrixRowsKernel(const T* src, T* dst, const size_t* index,
 }
 
 template <typename T>
-class CopyMatrixRowsFunctor<platform::GPUPlace, T> {
+class CopyMatrixRowsFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& src, const size_t* index,
                   framework::Tensor& dst, bool is_src_index) {
     auto src_dims = src.dims();
@@ -59,20 +59,19 @@ class CopyMatrixRowsFunctor<platform::GPUPlace, T> {
 
     dim3 threads(128, 8);
     dim3 grid(8, 1);
-    auto stream =
-        reinterpret_cast<const platform::CUDADeviceContext&>(context).stream();
+    auto stream = context.stream();
     CopyMatrixRowsKernel<T, 128, 8, 8><<<grid, threads, 0, stream>>>(
         src_data, dst_data, index, height, width, is_src_index);
   }
 };
 
-template class CopyMatrixRowsFunctor<platform::GPUPlace, float>;
-template class CopyMatrixRowsFunctor<platform::GPUPlace, double>;
+template class CopyMatrixRowsFunctor<platform::CUDADeviceContext, float>;
+template class CopyMatrixRowsFunctor<platform::CUDADeviceContext, double>;
 
-template class LoDTensor2BatchFunctor<platform::GPUPlace, float>;
-template class LoDTensor2BatchFunctor<platform::GPUPlace, double>;
-template class Batch2LoDTensorFunctor<platform::GPUPlace, float>;
-template class Batch2LoDTensorFunctor<platform::GPUPlace, double>;
+template class LoDTensor2BatchFunctor<platform::CUDADeviceContext, float>;
+template class LoDTensor2BatchFunctor<platform::CUDADeviceContext, double>;
+template class Batch2LoDTensorFunctor<platform::CUDADeviceContext, float>;
+template class Batch2LoDTensorFunctor<platform::CUDADeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/sequence2batch.h b/paddle/operators/math/sequence2batch.h
index 73295ddbcb..a5c43a2c7d 100644
--- a/paddle/operators/math/sequence2batch.h
+++ b/paddle/operators/math/sequence2batch.h
@@ -26,7 +26,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class CopyMatrixRowsFunctor {
  public:
   // If is_src_index is true,
@@ -34,12 +34,12 @@ class CopyMatrixRowsFunctor {
   // If is_src_index is false,
   // copy the input src to the indexed rows of output dst.
   // The indexed rows are based on the input index.
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& src, const size_t* index,
-                  framework::Tensor& dst, bool is_src_index);
+  void operator()(const DeviceContext& context, const framework::Tensor& src,
+                  const size_t* index, framework::Tensor& dst,
+                  bool is_src_index);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class LoDTensor2BatchFunctor {
   // Calculate the length of each sequence and
   // sort sequence index by the length.
@@ -56,7 +56,7 @@ class LoDTensor2BatchFunctor {
   };
 
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::LoDTensor& lod_tensor,
                   framework::LoDTensor& batch, bool is_cal_batch_lod,
                   bool is_reverse = false) const {
@@ -65,7 +65,7 @@ class LoDTensor2BatchFunctor {
       PADDLE_ENFORCE_GT(lods.size(), 2UL);
       PADDLE_ENFORCE_EQ(lods[1].size(),
                         static_cast<size_t>(lod_tensor.dims()[0]));
-      CopyMatrixRowsFunctor<Place, T> to_batch;
+      CopyMatrixRowsFunctor<DeviceContext, T> to_batch;
       to_batch(context, lod_tensor, lods[1].data(), batch, true);
       return;
     }
@@ -143,22 +143,22 @@ class LoDTensor2BatchFunctor {
     }
     batch.set_lod(batch_lods);
 
-    CopyMatrixRowsFunctor<Place, T> to_batch;
+    CopyMatrixRowsFunctor<DeviceContext, T> to_batch;
     to_batch(context, lod_tensor, seq2batch_idx, batch, true);
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class Batch2LoDTensorFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::LoDTensor& batch,
                   framework::LoDTensor& lod_tensor) const {
     auto in_lod = batch.lod();
     PADDLE_ENFORCE_GT(in_lod.size(), 2UL);
     PADDLE_ENFORCE_EQ(in_lod[1].size(),
                       static_cast<size_t>(lod_tensor.dims()[0]));
-    CopyMatrixRowsFunctor<Place, T> to_seq;
+    CopyMatrixRowsFunctor<DeviceContext, T> to_seq;
     size_t* index = in_lod[1].data();
     to_seq(context, batch, index, lod_tensor, false);
   }
diff --git a/paddle/operators/math/sequence_pooling.cc b/paddle/operators/math/sequence_pooling.cc
index 5913c99fdb..8fb92b1a13 100644
--- a/paddle/operators/math/sequence_pooling.cc
+++ b/paddle/operators/math/sequence_pooling.cc
@@ -20,9 +20,9 @@ namespace operators {
 namespace math {
 
 template <typename T>
-class MaxSeqPoolFunctor<platform::CPUPlace, T> {
+class MaxSeqPoolFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::LoDTensor& input, framework::Tensor* output,
                   framework::Tensor* index) {
     auto in_dims = input.dims();
@@ -60,9 +60,9 @@ class MaxSeqPoolFunctor<platform::CPUPlace, T> {
 };
 
 template <typename T>
-class MaxSeqPoolGradFunctor<platform::CPUPlace, T> {
+class MaxSeqPoolGradFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& out_grad,
                   const framework::Tensor& index,
                   framework::LoDTensor* in_grad) {
@@ -80,7 +80,7 @@ class MaxSeqPoolGradFunctor<platform::CPUPlace, T> {
     const int* max_index = index.data<int>();
     T* ig_data = in_grad->data<T>();
 
-    SetConstant<platform::CPUPlace, T> set_zero;
+    SetConstant<platform::CPUDeviceContext, T> set_zero;
     set_zero(context, in_grad, static_cast<T>(0.0));
     int64_t num_seq = og_dims[0];
     int64_t dim = out_grad.numel() / num_seq;
@@ -93,10 +93,10 @@ class MaxSeqPoolGradFunctor<platform::CPUPlace, T> {
   }
 };
 
-template class MaxSeqPoolFunctor<platform::CPUPlace, float>;
-template class MaxSeqPoolFunctor<platform::CPUPlace, double>;
-template class MaxSeqPoolGradFunctor<platform::CPUPlace, float>;
-template class MaxSeqPoolGradFunctor<platform::CPUPlace, double>;
+template class MaxSeqPoolFunctor<platform::CPUDeviceContext, float>;
+template class MaxSeqPoolFunctor<platform::CPUDeviceContext, double>;
+template class MaxSeqPoolGradFunctor<platform::CPUDeviceContext, float>;
+template class MaxSeqPoolGradFunctor<platform::CPUDeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/sequence_pooling.cu b/paddle/operators/math/sequence_pooling.cu
index 5ed951402f..4c9e6b375c 100644
--- a/paddle/operators/math/sequence_pooling.cu
+++ b/paddle/operators/math/sequence_pooling.cu
@@ -46,9 +46,9 @@ __global__ void KeMaxSequencePool(const T* input, const size_t* starts,
 }
 
 template <typename T>
-class MaxSeqPoolFunctor<platform::GPUPlace, T> {
+class MaxSeqPoolFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::LoDTensor& input, framework::Tensor* output,
                   framework::Tensor* index) {
     auto in_dims = input.dims();
@@ -71,8 +71,7 @@ class MaxSeqPoolFunctor<platform::GPUPlace, T> {
 
     dim3 threads(256, 1);
     dim3 grid(num_seq, 1);
-    auto stream =
-        reinterpret_cast<const platform::CUDADeviceContext&>(context).stream();
+    auto stream = context.stream();
     KeMaxSequencePool<T><<<grid, threads, 0, stream>>>(
         in_data, starts.data(), out_data, max_index, num_seq, dim);
   }
@@ -91,9 +90,9 @@ __global__ void KeMaxSequencePoolGrad(const T* out_grad, const int* max_index,
 }
 
 template <typename T>
-class MaxSeqPoolGradFunctor<platform::GPUPlace, T> {
+class MaxSeqPoolGradFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& out_grad,
                   const framework::Tensor& index,
                   framework::LoDTensor* in_grad) {
@@ -111,7 +110,7 @@ class MaxSeqPoolGradFunctor<platform::GPUPlace, T> {
     const int* max_index = index.data<int>();
     T* ig_data = in_grad->data<T>();
 
-    SetConstant<platform::GPUPlace, T> set_zero;
+    SetConstant<platform::CUDADeviceContext, T> set_zero;
     set_zero(context, in_grad, static_cast<T>(0.0));
     int64_t num_seq = og_dims[0];
     int64_t dim = out_grad.numel() / num_seq;
@@ -119,17 +118,16 @@ class MaxSeqPoolGradFunctor<platform::GPUPlace, T> {
     unsigned int blocks = (num_seq * dim + 128 - 1) / 128;
     dim3 threads(128, 1);
     dim3 grid(blocks, 1);
-    auto stream =
-        reinterpret_cast<const platform::CUDADeviceContext&>(context).stream();
+    auto stream = context.stream();
     KeMaxSequencePoolGrad<T><<<grid, threads, 0, stream>>>(
         og_data, max_index, ig_data, num_seq, dim);
   }
 };
 
-template class MaxSeqPoolFunctor<platform::GPUPlace, float>;
-template class MaxSeqPoolFunctor<platform::GPUPlace, double>;
-template class MaxSeqPoolGradFunctor<platform::GPUPlace, float>;
-template class MaxSeqPoolGradFunctor<platform::GPUPlace, double>;
+template class MaxSeqPoolFunctor<platform::CUDADeviceContext, float>;
+template class MaxSeqPoolFunctor<platform::CUDADeviceContext, double>;
+template class MaxSeqPoolGradFunctor<platform::CUDADeviceContext, float>;
+template class MaxSeqPoolGradFunctor<platform::CUDADeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/sequence_pooling.h b/paddle/operators/math/sequence_pooling.h
index 35dfe26de1..13ffb2ebef 100644
--- a/paddle/operators/math/sequence_pooling.h
+++ b/paddle/operators/math/sequence_pooling.h
@@ -23,18 +23,18 @@ namespace math {
 
 #define FLT_MAX __FLT_MAX__
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MaxSeqPoolFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::LoDTensor& input, framework::Tensor* output,
                   framework::Tensor* index);
 };
 
-template <typename Place, class T>
+template <typename DeviceContext, class T>
 class MaxSeqPoolGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::Tensor& out_grad,
                   const framework::Tensor& index,
                   framework::LoDTensor* in_grad);
diff --git a/paddle/operators/math/softmax.cc b/paddle/operators/math/softmax.cc
index 3e2f15d6c2..72f10f35f4 100644
--- a/paddle/operators/math/softmax.cc
+++ b/paddle/operators/math/softmax.cc
@@ -19,10 +19,10 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template class SoftmaxFunctor<platform::CPUPlace, float>;
-template class SoftmaxFunctor<platform::CPUPlace, double>;
-template class SoftmaxGradFunctor<platform::CPUPlace, float>;
-template class SoftmaxGradFunctor<platform::CPUPlace, double>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, float>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, double>;
+template class SoftmaxGradFunctor<platform::CPUDeviceContext, float>;
+template class SoftmaxGradFunctor<platform::CPUDeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/softmax.cu b/paddle/operators/math/softmax.cu
index 4dbab51d46..9e73f6a371 100644
--- a/paddle/operators/math/softmax.cu
+++ b/paddle/operators/math/softmax.cu
@@ -21,10 +21,10 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template class SoftmaxFunctor<platform::GPUPlace, float>;
-template class SoftmaxFunctor<platform::GPUPlace, double>;
-template class SoftmaxGradFunctor<platform::GPUPlace, float>;
-template class SoftmaxGradFunctor<platform::GPUPlace, double>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, float>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, double>;
+template class SoftmaxGradFunctor<platform::CUDADeviceContext, float>;
+template class SoftmaxGradFunctor<platform::CUDADeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/softmax.h b/paddle/operators/math/softmax.h
index fe10746502..471f44d340 100644
--- a/paddle/operators/math/softmax.h
+++ b/paddle/operators/math/softmax.h
@@ -19,19 +19,18 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SoftmaxFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor* X, framework::Tensor* Y);
+  void operator()(const DeviceContext& context, const framework::Tensor* X,
+                  framework::Tensor* Y);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SoftmaxGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor* y, const framework::Tensor* y_grad,
-                  framework::Tensor* x_grad);
+  void operator()(const DeviceContext& context, const framework::Tensor* y,
+                  const framework::Tensor* y_grad, framework::Tensor* x_grad);
 };
 
 }  // namespace math
diff --git a/paddle/operators/math/softmax_impl.h b/paddle/operators/math/softmax_impl.h
index 05793eeb3e..82f597ff79 100644
--- a/paddle/operators/math/softmax_impl.h
+++ b/paddle/operators/math/softmax_impl.h
@@ -32,10 +32,10 @@ struct ValueClip {
   }
 };
 
-template <typename Place, typename T>
-void SoftmaxFunctor<Place, T>::operator()(
-    const platform::DeviceContext& context, const framework::Tensor* X,
-    framework::Tensor* Y) {
+template <typename DeviceContext, typename T>
+void SoftmaxFunctor<DeviceContext, T>::operator()(const DeviceContext& context,
+                                                  const framework::Tensor* X,
+                                                  framework::Tensor* Y) {
   auto logits = EigenMatrix<T>::From(*X);
   auto softmax = EigenMatrix<T>::From(*Y);
 
@@ -56,19 +56,18 @@ void SoftmaxFunctor<Place, T>::operator()(
                              .broadcast(one_by_class))
                             .unaryExpr(ValueClip<T>());
 
-  softmax.device(*context.GetEigenDevice<Place>()) = shifted_logits.exp();
-  softmax.device(*context.GetEigenDevice<Place>()) =
-      (softmax *
-       softmax.sum(along_class)
-           .inverse()
-           .eval()
-           .reshape(batch_by_one)
-           .broadcast(one_by_class));
+  softmax.device(*context.eigen_device()) = shifted_logits.exp();
+  softmax.device(*context.eigen_device()) = (softmax *
+                                             softmax.sum(along_class)
+                                                 .inverse()
+                                                 .eval()
+                                                 .reshape(batch_by_one)
+                                                 .broadcast(one_by_class));
 }
 
-template <typename Place, typename T>
-void SoftmaxGradFunctor<Place, T>::operator()(
-    const platform::DeviceContext& context, const framework::Tensor* y,
+template <typename DeviceContext, typename T>
+void SoftmaxGradFunctor<DeviceContext, T>::operator()(
+    const DeviceContext& context, const framework::Tensor* y,
     const framework::Tensor* y_grad, framework::Tensor* x_grad) {
   auto softmax = EigenMatrix<T>::From(*y);
   auto softmax_grad = EigenMatrix<T>::From(*y_grad);
@@ -89,8 +88,7 @@ void SoftmaxGradFunctor<Place, T>::operator()(
                  .eval()
                  .reshape(batch_by_one)
                  .broadcast(one_by_class);
-  logits_grad.device(*context.GetEigenDevice<Place>()) =
-      (softmax_grad - dot) * softmax;
+  logits_grad.device(*context.eigen_device()) = (softmax_grad - dot) * softmax;
 }
 
 }  // namespace math
diff --git a/paddle/operators/math/unpooling.cc b/paddle/operators/math/unpooling.cc
index b57d3dc141..ecd3a647e0 100644
--- a/paddle/operators/math/unpooling.cc
+++ b/paddle/operators/math/unpooling.cc
@@ -17,9 +17,9 @@ namespace paddle {
 namespace operators {
 namespace math {
 template <typename T>
-class Unpool2dMaxFunctor<platform::CPUPlace, T> {
+class Unpool2dMaxFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& indices, framework::Tensor* output) {
     const int batch_size = input.dims()[0];
@@ -48,9 +48,9 @@ class Unpool2dMaxFunctor<platform::CPUPlace, T> {
   }
 };
 template <class T>
-class Unpool2dMaxGradFunctor<platform::CPUPlace, T> {
+class Unpool2dMaxGradFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& indices,
                   const framework::Tensor& output,
@@ -82,10 +82,10 @@ class Unpool2dMaxGradFunctor<platform::CPUPlace, T> {
     }
   }
 };
-template class Unpool2dMaxGradFunctor<platform::CPUPlace, float>;
-template class Unpool2dMaxGradFunctor<platform::CPUPlace, double>;
-template class Unpool2dMaxFunctor<platform::CPUPlace, float>;
-template class Unpool2dMaxFunctor<platform::CPUPlace, double>;
+template class Unpool2dMaxGradFunctor<platform::CPUDeviceContext, float>;
+template class Unpool2dMaxGradFunctor<platform::CPUDeviceContext, double>;
+template class Unpool2dMaxFunctor<platform::CPUDeviceContext, float>;
+template class Unpool2dMaxFunctor<platform::CPUDeviceContext, double>;
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/math/unpooling.cu b/paddle/operators/math/unpooling.cu
index 37c3c8b689..ecbde0f6a7 100644
--- a/paddle/operators/math/unpooling.cu
+++ b/paddle/operators/math/unpooling.cu
@@ -67,9 +67,9 @@ __global__ void KernelUnpool2dMaxGrad(
  * All tensors are in NCHW format.
  */
 template <typename T>
-class Unpool2dMaxFunctor<platform::GPUPlace, T> {
+class Unpool2dMaxFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& indices, framework::Tensor* output) {
     const int batch_size = input.dims()[0];
@@ -83,21 +83,18 @@ class Unpool2dMaxFunctor<platform::GPUPlace, T> {
     T* output_data = output->mutable_data<T>(context.GetPlace());
     int threads = 1024;
     int grid = (input.numel() + threads - 1) / threads;
-    KernelUnpool2dMax<
-        T><<<grid, threads, 0,
-             reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                 .stream()>>>(input.numel(), input_data, indices_data,
-                              input_height, input_width, output_channels,
-                              output_data, output_height, output_width);
+    KernelUnpool2dMax<T><<<grid, threads, 0, context.stream()>>>(
+        input.numel(), input_data, indices_data, input_height, input_width,
+        output_channels, output_data, output_height, output_width);
   }
 };
 /*
  * All tensors are in NCHW format.
  */
 template <typename T>
-class Unpool2dMaxGradFunctor<platform::GPUPlace, T> {
+class Unpool2dMaxGradFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input,
                   const framework::Tensor& indices,
                   const framework::Tensor& output,
@@ -116,19 +113,16 @@ class Unpool2dMaxGradFunctor<platform::GPUPlace, T> {
     T* input_grad_data = input_grad->mutable_data<T>(context.GetPlace());
     int threads = 1024;
     int grid = (input.numel() + threads - 1) / threads;
-    KernelUnpool2dMaxGrad<
-        T><<<grid, threads, 0,
-             reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                 .stream()>>>(input.numel(), input_data, indices_data,
-                              input_height, input_width, output_channels,
-                              output_data, output_grad_data, output_height,
-                              output_width, input_grad_data);
+    KernelUnpool2dMaxGrad<T><<<grid, threads, 0, context.stream()>>>(
+        input.numel(), input_data, indices_data, input_height, input_width,
+        output_channels, output_data, output_grad_data, output_height,
+        output_width, input_grad_data);
   }
 };
-template class Unpool2dMaxGradFunctor<platform::GPUPlace, float>;
-template class Unpool2dMaxGradFunctor<platform::GPUPlace, double>;
-template class Unpool2dMaxFunctor<platform::GPUPlace, float>;
-template class Unpool2dMaxFunctor<platform::GPUPlace, double>;
+template class Unpool2dMaxGradFunctor<platform::CUDADeviceContext, float>;
+template class Unpool2dMaxGradFunctor<platform::CUDADeviceContext, double>;
+template class Unpool2dMaxFunctor<platform::CUDADeviceContext, float>;
+template class Unpool2dMaxFunctor<platform::CUDADeviceContext, double>;
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/math/unpooling.h b/paddle/operators/math/unpooling.h
index 7077d7c227..0f0ff1371e 100644
--- a/paddle/operators/math/unpooling.h
+++ b/paddle/operators/math/unpooling.h
@@ -18,18 +18,16 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 namespace math {
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class Unpool2dMaxFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input,
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
                   const framework::Tensor& indices, framework::Tensor* output);
 };
-template <typename Place, class T>
+template <typename DeviceContext, class T>
 class Unpool2dMaxGradFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& input,
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
                   const framework::Tensor& indices,
                   const framework::Tensor& output,
                   const framework::Tensor& output_grad,
diff --git a/paddle/operators/math/vol2col.cc b/paddle/operators/math/vol2col.cc
index 99eb7fd46d..d574ed9234 100644
--- a/paddle/operators/math/vol2col.cc
+++ b/paddle/operators/math/vol2col.cc
@@ -25,9 +25,9 @@ namespace math {
  *                    output_depth, output_height, output_width]
  */
 template <class T>
-class Vol2ColFunctor<platform::CPUPlace, T> {
+class Vol2ColFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& vol,
                   const std::vector<int>& dilations,
                   const std::vector<int>& strides,
@@ -111,9 +111,9 @@ class Vol2ColFunctor<platform::CPUPlace, T> {
  *                    output_depth, output_height, output_width]
  */
 template <class T>
-class Col2VolFunctor<platform::CPUPlace, T> {
+class Col2VolFunctor<platform::CPUDeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& col,
                   const std::vector<int>& dilations,
                   const std::vector<int>& strides,
@@ -190,10 +190,10 @@ class Col2VolFunctor<platform::CPUPlace, T> {
   }
 };
 
-template class Vol2ColFunctor<platform::CPUPlace, float>;
-template class Vol2ColFunctor<platform::CPUPlace, double>;
-template class Col2VolFunctor<platform::CPUPlace, float>;
-template class Col2VolFunctor<platform::CPUPlace, double>;
+template class Vol2ColFunctor<platform::CPUDeviceContext, float>;
+template class Vol2ColFunctor<platform::CPUDeviceContext, double>;
+template class Col2VolFunctor<platform::CPUDeviceContext, float>;
+template class Col2VolFunctor<platform::CPUDeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/vol2col.cu b/paddle/operators/math/vol2col.cu
index dae3be858e..b029442fe4 100644
--- a/paddle/operators/math/vol2col.cu
+++ b/paddle/operators/math/vol2col.cu
@@ -68,9 +68,9 @@ __global__ void vol2col(int num_kernels, const T* data_vol, int depth,
  *                    output_depth, output_height, output_width]
  */
 template <class T>
-class Vol2ColFunctor<platform::GPUPlace, T> {
+class Vol2ColFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& vol,
                   const std::vector<int>& dilations,
                   const std::vector<int>& strides,
@@ -117,9 +117,7 @@ class Vol2ColFunctor<platform::GPUPlace, T> {
 
     const int threads = 1024;
     const int blocks = (num_outputs + 1024 - 1) / 1024;
-    vol2col<T><<<blocks, threads, 0,
-                 reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                     .stream()>>>(
+    vol2col<T><<<blocks, threads, 0, context.stream()>>>(
         num_outputs, vol.data<T>(), input_depth, input_height, input_width,
         dilations[0], dilations[1], dilations[2], filter_depth, filter_height,
         filter_width, strides[0], strides[1], strides[2], paddings[0],
@@ -196,9 +194,9 @@ __global__ void col2vol(int num_kernels, const T* data_col, int depth,
  *                    output_depth, output_height, output_width]
  */
 template <class T>
-class Col2VolFunctor<platform::GPUPlace, T> {
+class Col2VolFunctor<platform::CUDADeviceContext, T> {
  public:
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& col,
                   const std::vector<int>& dilations,
                   const std::vector<int>& strides,
@@ -245,9 +243,7 @@ class Col2VolFunctor<platform::GPUPlace, T> {
     const int threads = 1024;
     const int blocks = (num_kernels + 1024 - 1) / 1024;
 
-    col2vol<T><<<blocks, threads, 0,
-                 reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                     .stream()>>>(
+    col2vol<T><<<blocks, threads, 0, context.stream()>>>(
         num_kernels, col.data<T>(), input_depth, input_height, input_width,
         dilations[0], dilations[1], dilations[2], filter_depth, filter_height,
         filter_width, strides[0], strides[1], strides[2], paddings[0],
@@ -256,10 +252,10 @@ class Col2VolFunctor<platform::GPUPlace, T> {
   }
 };
 
-template class Vol2ColFunctor<platform::GPUPlace, float>;
-template class Vol2ColFunctor<platform::GPUPlace, double>;
-template class Col2VolFunctor<platform::GPUPlace, float>;
-template class Col2VolFunctor<platform::GPUPlace, double>;
+template class Vol2ColFunctor<platform::CUDADeviceContext, float>;
+template class Vol2ColFunctor<platform::CUDADeviceContext, double>;
+template class Col2VolFunctor<platform::CUDADeviceContext, float>;
+template class Col2VolFunctor<platform::CUDADeviceContext, double>;
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/vol2col.h b/paddle/operators/math/vol2col.h
index dc64d1d977..dcd80370e8 100644
--- a/paddle/operators/math/vol2col.h
+++ b/paddle/operators/math/vol2col.h
@@ -63,22 +63,20 @@ namespace math {
  * \note The caller needs to ensure that volShape.inputChannels is equal to
  *       colShape.inputChannels.
  */
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class Vol2ColFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& vol,
+  void operator()(const DeviceContext& context, const framework::Tensor& vol,
                   const std::vector<int>& dilations,
                   const std::vector<int>& strides,
                   const std::vector<int>& paddings,
                   framework::Tensor* col) const;
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class Col2VolFunctor {
  public:
-  void operator()(const platform::DeviceContext& context,
-                  const framework::Tensor& col,
+  void operator()(const DeviceContext& context, const framework::Tensor& col,
                   const std::vector<int>& dilations,
                   const std::vector<int>& strides,
                   const std::vector<int>& paddings,
diff --git a/paddle/operators/math/vol2col_test.cc b/paddle/operators/math/vol2col_test.cc
index 62c3152304..f46db3c567 100644
--- a/paddle/operators/math/vol2col_test.cc
+++ b/paddle/operators/math/vol2col_test.cc
@@ -16,7 +16,7 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <iostream>
 
-template <typename Place>
+template <typename DeviceContext, typename Place>
 void testVol2col() {
   paddle::framework::Tensor input;
   paddle::framework::Tensor input_tmp;
@@ -24,18 +24,7 @@ void testVol2col() {
   paddle::framework::Tensor output_tmp;
 
   auto* place = new Place();
-  paddle::platform::DeviceContext* context;
-  if (paddle::platform::is_cpu_place(*place)) {
-    context =
-        new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace());
-  } else {
-#ifdef PADDLE_WITH_CUDA
-    context =
-        new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace());
-#else
-    PADDLE_THROW("no GPU support");
-#endif  // PADDLE_WITH_CUDA
-  }
+  DeviceContext* context = new DeviceContext(*place);
 
   /**
    * input = [[0, 1, 2,
@@ -88,7 +77,7 @@ void testVol2col() {
                               output_depth, output_height, output_width},
                              *place);
 
-  paddle::operators::math::Vol2ColFunctor<Place, float> vol2col;
+  paddle::operators::math::Vol2ColFunctor<DeviceContext, float> vol2col;
   vol2col(*context, input, dilations, strides, paddings, &output);
 
   float vol_2_col[] = {0, 1, 1, 2, 3, 4, 4, 5, 6, 7, 7, 8, 9, 10, 10, 11};
@@ -113,7 +102,7 @@ void testVol2col() {
     CopyFrom(input_tmp, *place, *context, &input);
   }
 
-  paddle::operators::math::Col2VolFunctor<Place, float> col2vol;
+  paddle::operators::math::Col2VolFunctor<DeviceContext, float> col2vol;
   col2vol(*context, output, dilations, strides, paddings, &input);
 
   float* in_ptr;
@@ -130,8 +119,9 @@ void testVol2col() {
 }
 
 TEST(math, vol2col) {
-  testVol2col<paddle::platform::CPUPlace>();
+  testVol2col<paddle::platform::CPUDeviceContext, paddle::platform::CPUPlace>();
 #ifdef PADDLE_WITH_CUDA
-  testVol2col<paddle::platform::GPUPlace>();
+  testVol2col<paddle::platform::CUDADeviceContext,
+              paddle::platform::GPUPlace>();
 #endif  // PADDLE_WITH_CUDA
 }
diff --git a/paddle/operators/matmul_op.cc b/paddle/operators/matmul_op.cc
index 5a1a615420..ee0bc0c370 100644
--- a/paddle/operators/matmul_op.cc
+++ b/paddle/operators/matmul_op.cc
@@ -206,7 +206,8 @@ class MatMulOpGrad : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(matmul, ops::MatMulOp, ops::MatMulOpMaker, matmul_grad,
             ops::MatMulOpGrad);
-REGISTER_OP_CPU_KERNEL(matmul,
-                       ops::MatMulKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
-    matmul_grad, ops::MatMulGradKernel<paddle::platform::CPUPlace, float>);
+    matmul, ops::MatMulKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    matmul_grad,
+    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/matmul_op.cu.cc b/paddle/operators/matmul_op.cu.cc
index b7e66382f0..6a3772c004 100644
--- a/paddle/operators/matmul_op.cu.cc
+++ b/paddle/operators/matmul_op.cu.cc
@@ -15,7 +15,8 @@
 #include "paddle/operators/matmul_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(matmul,
-                       ops::MatMulKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    matmul_grad, ops::MatMulGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    matmul, ops::MatMulKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    matmul_grad,
+    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/matmul_op.h b/paddle/operators/matmul_op.h
index 1e4aa48b70..de9da487b3 100644
--- a/paddle/operators/matmul_op.h
+++ b/paddle/operators/matmul_op.h
@@ -27,7 +27,7 @@ using DDim = framework::DDim;
 using framework::make_ddim;
 using framework::vectorize;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MatMulKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -38,8 +38,9 @@ class MatMulKernel : public framework::OpKernel<T> {
     bool transpose_x = context.Attr<bool>("transpose_X");
     bool transpose_y = context.Attr<bool>("transpose_Y");
 
-    math::MatMulFunctor<Place, T>()(context.device_context(), x, transpose_x, y,
-                                    transpose_y, T(1), out, T(0));
+    math::MatMulFunctor<DeviceContext, T>()(
+        context.template device_context<DeviceContext>(), x, transpose_x, y,
+        transpose_y, T(1), out, T(0));
   }
 };
 
@@ -68,17 +69,16 @@ Tensor CombineBatchAndM(const Tensor& input) {
 // Reshape a rank-3 tensor from P x M x N to M x (P * N).
 // (Warning: This requires transposing data and writes into new memory.)
 // Identity op if the tensor is not of rank 3.
-template <typename Place, typename T>
-Tensor CombineBatchAndN(const framework::ExecutionContext& context,
-                        const Tensor& input) {
+template <typename DeviceContext, typename T>
+Tensor CombineBatchAndN(const DeviceContext& context, const Tensor& input) {
   Tensor output;
   auto in_dims = input.dims();
   if (in_dims.size() == 3) {
     output.Resize({in_dims[1], in_dims[0], in_dims[2]});
     output.mutable_data<T>(context.GetPlace());
     std::vector<int> axis = {1, 0, 2};
-    math::Transpose<Place, T, 3> trans;
-    trans(context.device_context(), input, &output, axis);
+    math::Transpose<DeviceContext, T, 3> trans;
+    trans(context, input, &output, axis);
     std::vector<int64_t> out_dims = {in_dims[1], in_dims[0] * in_dims[2]};
     output.Resize({in_dims[1], in_dims[0] * in_dims[2]});
   } else {
@@ -112,7 +112,7 @@ Tensor CombineBatchAndN(const framework::ExecutionContext& context,
 //
 // To handle this sort of scenario, we reshape X : P x M x K, dOut: P x M x N
 // to X: (P * M) x K, dOut: (P * M) x N.
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MatMulGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -178,24 +178,23 @@ class MatMulGradKernel : public framework::OpKernel<T> {
     Tensor Y = Reshape<T>(y, make_ddim(y_dims));
     Tensor dOut = Reshape<T>(dout, make_ddim(dout_dims));
 
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     if (dx) {
       dx->mutable_data<T>(context.GetPlace());
       const Tensor& dOut_for_dX =
           (x_dims.size() == 2 && y_dims.size() == 3)
-              ? CombineBatchAndN<Place, T>(context, dOut)
+              ? CombineBatchAndN<DeviceContext, T>(dev_ctx, dOut)
               : dOut;
       if (x_dims.size() == 2 && y_dims.size() == 3) {
         Y = transpose_y ? CombineBatchAndM<T>(Y)
-                        : CombineBatchAndN<Place, T>(context, Y);
+                        : CombineBatchAndN<DeviceContext, T>(dev_ctx, Y);
       }
       if (transpose_x) {
-        math::MatMulFunctor<Place, T>()(context.device_context(), Y,
-                                        transpose_y, dOut_for_dX, transpose_x,
-                                        T(1), dx, T(0));
+        math::MatMulFunctor<DeviceContext, T>()(
+            dev_ctx, Y, transpose_y, dOut_for_dX, transpose_x, T(1), dx, T(0));
       } else {
-        math::MatMulFunctor<Place, T>()(context.device_context(), dOut_for_dX,
-                                        transpose_x, Y, !transpose_y, T(1), dx,
-                                        T(0));
+        math::MatMulFunctor<DeviceContext, T>()(
+            dev_ctx, dOut_for_dX, transpose_x, Y, !transpose_y, T(1), dx, T(0));
       }
     }
 
@@ -205,18 +204,16 @@ class MatMulGradKernel : public framework::OpKernel<T> {
                                       ? CombineBatchAndM<T>(dOut)
                                       : dOut;
       if (y_dims.size() == 2 && x_dims.size() == 3) {
-        X = transpose_x ? CombineBatchAndN<Place, T>(context, X)
+        X = transpose_x ? CombineBatchAndN<DeviceContext, T>(dev_ctx, X)
                         : CombineBatchAndM<T>(X);
         dOut = CombineBatchAndM<T>(dOut);
       }
       if (transpose_y) {
-        math::MatMulFunctor<Place, T>()(context.device_context(), dOut_for_dY,
-                                        transpose_y, X, transpose_x, T(1), dy,
-                                        T(0));
+        math::MatMulFunctor<DeviceContext, T>()(
+            dev_ctx, dOut_for_dY, transpose_y, X, transpose_x, T(1), dy, T(0));
       } else {
-        math::MatMulFunctor<Place, T>()(context.device_context(), X,
-                                        !transpose_x, dOut_for_dY, transpose_y,
-                                        T(1), dy, T(0));
+        math::MatMulFunctor<DeviceContext, T>()(
+            dev_ctx, X, !transpose_x, dOut_for_dY, transpose_y, T(1), dy, T(0));
       }
     }
   }
diff --git a/paddle/operators/maxout_op.cc b/paddle/operators/maxout_op.cc
index 44bf402e95..011616e615 100644
--- a/paddle/operators/maxout_op.cc
+++ b/paddle/operators/maxout_op.cc
@@ -101,7 +101,8 @@ class MaxOutOpGrad : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(maxout, ops::MaxOutOp, ops::MaxOutOpMaker, maxout_grad,
             ops::MaxOutOpGrad);
-REGISTER_OP_CPU_KERNEL(maxout,
-                       ops::MaxOutKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
-    maxout_grad, ops::MaxOutGradKernel<paddle::platform::CPUPlace, float>);
+    maxout, ops::MaxOutKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    maxout_grad,
+    ops::MaxOutGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/maxout_op.cu.cc b/paddle/operators/maxout_op.cu.cc
index decd43913d..2904f0ff96 100644
--- a/paddle/operators/maxout_op.cu.cc
+++ b/paddle/operators/maxout_op.cu.cc
@@ -15,9 +15,10 @@
 #include "paddle/operators/maxout_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(maxout,
-                       ops::MaxOutKernel<paddle::platform::GPUPlace, float>,
-                       ops::MaxOutKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
-    maxout_grad, ops::MaxOutGradKernel<paddle::platform::GPUPlace, float>,
-    ops::MaxOutGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    maxout, ops::MaxOutKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::MaxOutKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    maxout_grad,
+    ops::MaxOutGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::MaxOutGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/maxout_op.h b/paddle/operators/maxout_op.h
index 44a0d073dd..e8b12552b9 100644
--- a/paddle/operators/maxout_op.h
+++ b/paddle/operators/maxout_op.h
@@ -23,7 +23,7 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MaxOutKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -31,12 +31,13 @@ class MaxOutKernel : public framework::OpKernel<T> {
     Tensor* out = context.Output<Tensor>("Out");
     int groups = context.template Attr<int>("groups");
 
-    math::MaxOutFunctor<Place, T> maxout_forward;
-    maxout_forward(context.device_context(), *in_x, out, groups);
+    math::MaxOutFunctor<DeviceContext, T> maxout_forward;
+    maxout_forward(context.template device_context<DeviceContext>(), *in_x, out,
+                   groups);
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MaxOutGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -46,14 +47,13 @@ class MaxOutGradKernel : public framework::OpKernel<T> {
         context.Input<Tensor>(framework::GradVarName("Out"));
     Tensor* in_x_grad = context.Output<Tensor>(framework::GradVarName("X"));
     int groups = context.template Attr<int>("groups");
-    auto& device_ctx = context.device_context();
-    math::SetConstant<Place, T> zero;
+    auto& device_ctx = context.template device_context<DeviceContext>();
+    math::SetConstant<DeviceContext, T> zero;
     if (in_x_grad) {
       in_x_grad->mutable_data<T>(context.GetPlace());
       zero(device_ctx, in_x_grad, static_cast<T>(0.0));
-      math::MaxOutGradFunctor<Place, T> maxout_backward;
-      maxout_backward(context.device_context(), *in_x, in_x_grad, *out,
-                      *out_grad, groups);
+      math::MaxOutGradFunctor<DeviceContext, T> maxout_backward;
+      maxout_backward(device_ctx, *in_x, in_x_grad, *out, *out_grad, groups);
     }
   }
 };
diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc
index dcc5b4286f..8932d700c2 100644
--- a/paddle/operators/mean_op.cc
+++ b/paddle/operators/mean_op.cc
@@ -76,8 +76,9 @@ class MeanGradMaker : public framework::SingleGradOpDescMaker {
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(mean, ops::MeanOp, ops::MeanOpMaker, ops::MeanGradMaker);
 REGISTER_OPERATOR(mean_grad, ops::MeanGradOp);
-REGISTER_OP_CPU_KERNEL(mean, ops::MeanKernel<paddle::platform::CPUPlace, float>,
-                       ops::MeanKernel<paddle::platform::CPUPlace, double>);
-REGISTER_OP_CPU_KERNEL(mean_grad,
-                       ops::MeanGradKernel<paddle::platform::CPUPlace, float>,
-                       ops::MeanGradKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(
+    mean, ops::MeanKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::MeanKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    mean_grad, ops::MeanGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::MeanGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/mean_op.cu b/paddle/operators/mean_op.cu
index ca089938c0..93062bf540 100644
--- a/paddle/operators/mean_op.cu
+++ b/paddle/operators/mean_op.cu
@@ -17,8 +17,9 @@
 #include "paddle/operators/mean_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(mean, ops::MeanKernel<paddle::platform::GPUPlace, float>,
-                       ops::MeanKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(mean_grad,
-                       ops::MeanGradKernel<paddle::platform::GPUPlace, float>,
-                       ops::MeanGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    mean, ops::MeanKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::MeanKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    mean_grad, ops::MeanGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::MeanGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h
index c99286a5b9..351b345959 100644
--- a/paddle/operators/mean_op.h
+++ b/paddle/operators/mean_op.h
@@ -27,7 +27,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MeanKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -38,13 +38,14 @@ class MeanKernel : public framework::OpKernel<T> {
 
     auto X = EigenVector<T>::Flatten(*input);
     auto y = EigenScalar<T>::From(*output);
-    auto& place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     y.device(place) = X.mean();
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MeanGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -56,7 +57,8 @@ class MeanGradKernel : public framework::OpKernel<T> {
     T ig_size = static_cast<T>(IG->numel());
     Eigen::DSizes<int, 1> bcast(ig_size);
 
-    EigenVector<T>::Flatten(*IG).device(context.GetEigenDevice<Place>()) =
+    EigenVector<T>::Flatten(*IG).device(
+        *context.template device_context<DeviceContext>().eigen_device()) =
         (EigenVector<T>::From(*OG) / ig_size).broadcast(bcast);
   }
 };
diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc
index 4684c20208..27f0c8de20 100644
--- a/paddle/operators/minus_op.cc
+++ b/paddle/operators/minus_op.cc
@@ -102,5 +102,5 @@ class MinusGradMaker : public framework::GradOpDescMakerBase {
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(minus, ops::MinusOp, ops::MinusOpMaker, ops::MinusGradMaker);
-REGISTER_OP_CPU_KERNEL(minus,
-                       ops::MinusKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    minus, ops::MinusKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/minus_op.cu b/paddle/operators/minus_op.cu
index a8375cc630..3b202ea92e 100644
--- a/paddle/operators/minus_op.cu
+++ b/paddle/operators/minus_op.cu
@@ -14,5 +14,6 @@
 
 #include "paddle/operators/minus_op.h"
 
-REGISTER_OP_GPU_KERNEL(
-    minus, paddle::operators::MinusKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    minus,
+    paddle::operators::MinusKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/minus_op.h b/paddle/operators/minus_op.h
index bd9a2790aa..78e1e1be6d 100644
--- a/paddle/operators/minus_op.h
+++ b/paddle/operators/minus_op.h
@@ -19,7 +19,7 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MinusKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -28,7 +28,8 @@ class MinusKernel : public framework::OpKernel<T> {
     auto* out_tensor = context.Output<framework::Tensor>("Out");
 
     out_tensor->mutable_data<T>(context.GetPlace());
-    auto& dev = context.GetEigenDevice<Place>();
+    auto& dev =
+        *context.template device_context<DeviceContext>().eigen_device();
     framework::EigenVector<T>::Flatten(*out_tensor).device(dev) =
         framework::EigenVector<T>::Flatten(*left_tensor) -
         framework::EigenVector<T>::Flatten(*right_tensor);
diff --git a/paddle/operators/modified_huber_loss_op.cc b/paddle/operators/modified_huber_loss_op.cc
index 28528848af..f0a42491bf 100644
--- a/paddle/operators/modified_huber_loss_op.cc
+++ b/paddle/operators/modified_huber_loss_op.cc
@@ -115,6 +115,6 @@ REGISTER_OP(modified_huber_loss, ops::ModifiedHuberLossOp,
 
 REGISTER_OP_CPU_KERNEL(
     modified_huber_loss,
-    ops::ModifiedHuberLossKernel<paddle::platform::CPUPlace, float>);
+    ops::ModifiedHuberLossKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(modified_huber_loss_grad,
                        ops::ModifiedHuberLossGradCPUKernel<float>);
diff --git a/paddle/operators/modified_huber_loss_op.cu b/paddle/operators/modified_huber_loss_op.cu
index 8854e166cd..40a8447da4 100644
--- a/paddle/operators/modified_huber_loss_op.cu
+++ b/paddle/operators/modified_huber_loss_op.cu
@@ -71,8 +71,8 @@ class ModifiedHuberLossGradGPUKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     modified_huber_loss,
-    ops::ModifiedHuberLossKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(modified_huber_loss_grad,
-                       ops::ModifiedHuberLossGradGPUKernel<float>);
+    ops::ModifiedHuberLossKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(modified_huber_loss_grad,
+                        ops::ModifiedHuberLossGradGPUKernel<float>);
diff --git a/paddle/operators/modified_huber_loss_op.h b/paddle/operators/modified_huber_loss_op.h
index aba75efad9..157ae0682e 100644
--- a/paddle/operators/modified_huber_loss_op.h
+++ b/paddle/operators/modified_huber_loss_op.h
@@ -46,7 +46,7 @@ struct ModifiedHuberLossForward {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ModifiedHuberLossKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -57,7 +57,8 @@ class ModifiedHuberLossKernel : public framework::OpKernel<T> {
 
     out0->mutable_data<T>(context.GetPlace());
     out1->mutable_data<T>(context.GetPlace());
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     auto x = EigenVector<T>::Flatten(*in0);
     auto y = EigenVector<T>::Flatten(*in1);
diff --git a/paddle/operators/momentum_op.cu b/paddle/operators/momentum_op.cu
index be0c8ea071..00f1253465 100644
--- a/paddle/operators/momentum_op.cu
+++ b/paddle/operators/momentum_op.cu
@@ -74,5 +74,5 @@ class MomentumOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(momentum, ops::MomentumOpCUDAKernel<float>,
-                       ops::MomentumOpCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(momentum, ops::MomentumOpCUDAKernel<float>,
+                        ops::MomentumOpCUDAKernel<double>);
diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc
index 3c39ae10dc..bc4a5fdf0b 100644
--- a/paddle/operators/mul_op.cc
+++ b/paddle/operators/mul_op.cc
@@ -149,6 +149,7 @@ REGISTER_OPERATOR(mul, paddle::framework::OperatorWithKernel, ops::MulOpMaker,
                   ops::MulOpShapeInference,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(mul_grad, ops::MulOpGrad);
-REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(mul_grad,
-                       ops::MulGradKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    mul, ops::MulKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    mul_grad, ops::MulGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/mul_op.cu.cc b/paddle/operators/mul_op.cu.cc
index 66dc3d6d10..6095de58d0 100644
--- a/paddle/operators/mul_op.cu.cc
+++ b/paddle/operators/mul_op.cu.cc
@@ -15,6 +15,7 @@
 #include "paddle/operators/mul_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(mul_grad,
-                       ops::MulGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    mul, ops::MulKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    mul_grad, ops::MulGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h
index 0eb9df41e9..1b467dca83 100644
--- a/paddle/operators/mul_op.h
+++ b/paddle/operators/mul_op.h
@@ -23,7 +23,7 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MulKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -46,15 +46,16 @@ class MulKernel : public framework::OpKernel<T> {
     if (z_dim.size() != 2) {
       z->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
     }
-    math::matmul<Place, T>(context.device_context(), x_matrix, false, y_matrix,
-                           false, 1, z, 0);
+    math::matmul<DeviceContext, T>(
+        context.template device_context<DeviceContext>(), x_matrix, false,
+        y_matrix, false, 1, z, 0);
     if (z_dim.size() != 2) {
       z->Resize(z_dim);
     }
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MulGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -77,6 +78,7 @@ class MulGradKernel : public framework::OpKernel<T> {
 
     Tensor* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
     Tensor* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
     if (dx) {
       dx->mutable_data<T>(ctx.GetPlace());
       Tensor dx_matrix = dx->dims().size() > 2
@@ -84,8 +86,8 @@ class MulGradKernel : public framework::OpKernel<T> {
                              : *dx;
 
       // dx = dout * y'. dx: M x K, dout : M x N, y : K x N
-      math::matmul<Place, T>(ctx.device_context(), dout_mat, false, y_matrix,
-                             true, 1, &dx_matrix, 0);
+      math::matmul<DeviceContext, T>(dev_ctx, dout_mat, false, y_matrix, true,
+                                     1, &dx_matrix, 0);
     }
     if (dy) {
       dy->mutable_data<T>(ctx.GetPlace());
@@ -93,8 +95,8 @@ class MulGradKernel : public framework::OpKernel<T> {
                              ? framework::ReshapeToMatrix(*dy, y_num_col_dims)
                              : *dy;
       // dy = x' * dout. dy K x N, dout : M x N, x : M x K
-      math::matmul<Place, T>(ctx.device_context(), x_matrix, true, dout_mat,
-                             false, 1, &dy_matrix, 0);
+      math::matmul<DeviceContext, T>(dev_ctx, x_matrix, true, dout_mat, false,
+                                     1, &dy_matrix, 0);
     }
   }
 };
diff --git a/paddle/operators/multiplex_op.cc b/paddle/operators/multiplex_op.cc
index 8e7f544e0d..b1ee8051c4 100644
--- a/paddle/operators/multiplex_op.cc
+++ b/paddle/operators/multiplex_op.cc
@@ -119,7 +119,8 @@ REGISTER_OPERATOR(multiplex, ops::MultiplexOp, ops::MultiplexOpMaker,
                   paddle::framework::DefaultGradOpDescMaker<false>);
 REGISTER_OPERATOR(multiplex_grad, ops::MultiplexGradOp);
 REGISTER_OP_CPU_KERNEL(
-    multiplex, ops::MultiplexCPUKernel<paddle::platform::CPUPlace, float>);
+    multiplex,
+    ops::MultiplexCPUKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     multiplex_grad,
-    ops::MultiplexGradCPUKernel<paddle::platform::CPUPlace, float>);
+    ops::MultiplexGradCPUKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/multiplex_op.cu b/paddle/operators/multiplex_op.cu
index 10dff8d021..47986e9ff8 100644
--- a/paddle/operators/multiplex_op.cu
+++ b/paddle/operators/multiplex_op.cu
@@ -36,7 +36,7 @@ class MultiplexGPUKernel : public framework::OpKernel<T> {
     CopyFrom(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
     auto* index = index_t_cpu.data<int32_t>();
     auto stream = ctx.cuda_device_context().stream();
-    Place place = boost::get<Place>(ctx.GetPlace());
+    platform::GPUPlace place = boost::get<platform::GPUPlace>(ctx.GetPlace());
     for (auto i = 0; i < rows; i++) {
       int32_t k = index[i];
       PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative.");
@@ -60,7 +60,8 @@ class MultiplexGradGPUKernel : public framework::OpKernel<T> {
       if (d_ins[i]) {
         d_ins[i]->mutable_data<T>(ctx.GetPlace());
         auto t = framework::EigenVector<T>::Flatten(*d_ins[i]);
-        t.device(ctx.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));
+        t.device(*ctx.template device_context<Place>().eigen_device()) =
+            t.constant(static_cast<T>(0));
       }
     }
 
@@ -72,7 +73,7 @@ class MultiplexGradGPUKernel : public framework::OpKernel<T> {
     auto* index = index_t_cpu.data<int32_t>();
 
     auto stream = ctx.cuda_device_context().stream();
-    Place place = boost::get<Place>(ctx.GetPlace());
+    platform::GPUPlace place = boost::get<platform::GPUPlace>(ctx.GetPlace());
     for (auto i = 0; i < rows; i++) {
       size_t k = static_cast<size_t>(index[i]);
       if (d_ins[k]) {
@@ -87,8 +88,9 @@ class MultiplexGradGPUKernel : public framework::OpKernel<T> {
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
-    multiplex, ops::MultiplexGPUKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    multiplex,
+    ops::MultiplexGPUKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     multiplex_grad,
-    ops::MultiplexGradGPUKernel<paddle::platform::GPUPlace, float>);
+    ops::MultiplexGradGPUKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/multiplex_op.h b/paddle/operators/multiplex_op.h
index ab3cafaa32..3443151161 100644
--- a/paddle/operators/multiplex_op.h
+++ b/paddle/operators/multiplex_op.h
@@ -22,7 +22,7 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MultiplexCPUKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
@@ -35,7 +35,7 @@ class MultiplexCPUKernel : public framework::OpKernel<T> {
     auto rows = ins[0]->dims()[0];
     auto cols = ins[0]->numel() / rows;
     auto index = ids->data<int32_t>();
-    Place place = boost::get<Place>(ctx.GetPlace());
+    platform::CPUPlace place = boost::get<platform::CPUPlace>(ctx.GetPlace());
     for (auto i = 0; i < rows; i++) {
       int32_t k = index[i];
       PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative.");
@@ -47,7 +47,7 @@ class MultiplexCPUKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class MultiplexGradCPUKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
@@ -60,14 +60,15 @@ class MultiplexGradCPUKernel : public framework::OpKernel<T> {
       if (d_ins[i]) {
         d_ins[i]->mutable_data<T>(ctx.GetPlace());
         auto t = framework::EigenVector<T>::Flatten(*d_ins[i]);
-        t.device(ctx.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));
+        t.device(*ctx.template device_context<DeviceContext>().eigen_device()) =
+            t.constant(static_cast<T>(0));
       }
     }
 
     auto rows = ins[0]->dims()[0];
     auto cols = ins[0]->numel() / rows;
     auto* index = ids->data<int32_t>();
-    Place place = boost::get<Place>(ctx.GetPlace());
+    platform::CPUPlace place = boost::get<platform::CPUPlace>(ctx.GetPlace());
     for (auto i = 0; i < rows; i++) {
       size_t k = static_cast<size_t>(index[i]);
       if (d_ins[k]) {
diff --git a/paddle/operators/nccl_op.cu.cc b/paddle/operators/nccl_op.cu.cc
index 4f0a2a79ed..6ca6db7253 100644
--- a/paddle/operators/nccl_op.cu.cc
+++ b/paddle/operators/nccl_op.cu.cc
@@ -204,6 +204,6 @@ class NCCLBcastKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(ncclAllReduce, ops::NCCLAllReduceKernel<float>);
-REGISTER_OP_GPU_KERNEL(ncclBcast, ops::NCCLBcastKernel<float>);
-REGISTER_OP_GPU_KERNEL(ncclReduce, ops::NCCLReduceKernel<float>);
+REGISTER_OP_CUDA_KERNEL(ncclAllReduce, ops::NCCLAllReduceKernel<float>);
+REGISTER_OP_CUDA_KERNEL(ncclBcast, ops::NCCLBcastKernel<float>);
+REGISTER_OP_CUDA_KERNEL(ncclReduce, ops::NCCLReduceKernel<float>);
diff --git a/paddle/operators/nccl_op_test.cu.cc b/paddle/operators/nccl_op_test.cu.cc
index bb7ae20286..d747cc0cf5 100644
--- a/paddle/operators/nccl_op_test.cu.cc
+++ b/paddle/operators/nccl_op_test.cu.cc
@@ -33,9 +33,9 @@
 #include "paddle/platform/place.h"
 
 USE_NO_KERNEL_OP(ncclInit);
-USE_GPU_ONLY_OP(ncclAllReduce);
-USE_GPU_ONLY_OP(ncclReduce);
-USE_GPU_ONLY_OP(ncclBcast);
+USE_CUDA_ONLY_OP(ncclAllReduce);
+USE_CUDA_ONLY_OP(ncclReduce);
+USE_CUDA_ONLY_OP(ncclBcast);
 
 namespace f = paddle::framework;
 namespace p = paddle::platform;
diff --git a/paddle/operators/nce_op.cc b/paddle/operators/nce_op.cc
index 952da10434..5ad1610fde 100644
--- a/paddle/operators/nce_op.cc
+++ b/paddle/operators/nce_op.cc
@@ -67,7 +67,7 @@ class NCEOp : public framework::OperatorWithKernel {
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
-        ctx.device_context());
+        ctx.GetPlace());
   }
 };
 
@@ -170,7 +170,7 @@ class NCEOpGrad : public framework::OperatorWithKernel {
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
-        ctx.device_context());
+        ctx.GetPlace());
   }
 };
 
diff --git a/paddle/operators/nce_op.h b/paddle/operators/nce_op.h
index 0a8a95de5f..6636dad060 100644
--- a/paddle/operators/nce_op.h
+++ b/paddle/operators/nce_op.h
@@ -28,7 +28,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 void PrepareSamples(const framework::ExecutionContext& context) {
   auto label = context.Input<Tensor>("Label");
   const int64_t* label_data = label->data<int64_t>();
@@ -67,11 +67,11 @@ void PrepareSamples(const framework::ExecutionContext& context) {
   }
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class NCEKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    PrepareSamples<Place, T>(context);
+    PrepareSamples<DeviceContext, T>(context);
     auto sample_labels = context.Output<Tensor>("SampleLabels");
     const int64_t* sample_labels_data = sample_labels->data<int64_t>();
     auto sample_out = context.Output<Tensor>("SampleLogits");
@@ -135,7 +135,7 @@ class NCEKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class NCEGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
diff --git a/paddle/operators/pad_op.cc b/paddle/operators/pad_op.cc
index adb75df6ef..936dde22c3 100644
--- a/paddle/operators/pad_op.cc
+++ b/paddle/operators/pad_op.cc
@@ -134,6 +134,7 @@ namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(pad, ops::PadOp, ops::PadOpMaker, ops::PadOpGradMaker);
 REGISTER_OPERATOR(pad_grad, ops::PadOpGrad);
-REGISTER_OP_CPU_KERNEL(pad, ops::PadKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(pad_grad,
-                       ops::PadGradKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    pad, ops::PadKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    pad_grad, ops::PadGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/pad_op.cu b/paddle/operators/pad_op.cu
index 555a7dba23..c309fb625c 100644
--- a/paddle/operators/pad_op.cu
+++ b/paddle/operators/pad_op.cu
@@ -16,6 +16,7 @@
 #include "paddle/operators/pad_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(pad, ops::PadKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(pad_grad,
-                       ops::PadGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    pad, ops::PadKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    pad_grad, ops::PadGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/pad_op.h b/paddle/operators/pad_op.h
index 9534dbf545..1b95942af3 100644
--- a/paddle/operators/pad_op.h
+++ b/paddle/operators/pad_op.h
@@ -26,7 +26,7 @@ template <typename T, size_t D, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
 
-template <typename Place, typename T, size_t D>
+template <typename DeviceContext, typename T, size_t D>
 void PadFunction(const framework::ExecutionContext& context) {
   auto pads = context.Attr<std::vector<int>>("paddings");
   Eigen::array<std::pair<int, int>, D> paddings;
@@ -42,33 +42,34 @@ void PadFunction(const framework::ExecutionContext& context) {
 
   auto x_tensor = EigenTensor<T, D>::From(*x);
   auto out_tensor = EigenTensor<T, D>::From(*out);
-  auto place = context.GetEigenDevice<Place>();
+  auto& place =
+      *context.template device_context<DeviceContext>().eigen_device();
   out_tensor.device(place) = x_tensor.pad(paddings, pad_value);
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class PadKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     int rank = context.Input<Tensor>("X")->dims().size();
     switch (rank) {
       case 1:
-        PadFunction<Place, T, 1>(context);
+        PadFunction<DeviceContext, T, 1>(context);
         break;
       case 2:
-        PadFunction<Place, T, 2>(context);
+        PadFunction<DeviceContext, T, 2>(context);
         break;
       case 3:
-        PadFunction<Place, T, 3>(context);
+        PadFunction<DeviceContext, T, 3>(context);
         break;
       case 4:
-        PadFunction<Place, T, 4>(context);
+        PadFunction<DeviceContext, T, 4>(context);
         break;
       case 5:
-        PadFunction<Place, T, 5>(context);
+        PadFunction<DeviceContext, T, 5>(context);
         break;
       case 6:
-        PadFunction<Place, T, 6>(context);
+        PadFunction<DeviceContext, T, 6>(context);
         break;
       default:
         PADDLE_THROW(
@@ -77,7 +78,7 @@ class PadKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T, size_t D>
+template <typename DeviceContext, typename T, size_t D>
 void PadGradFunction(const framework::ExecutionContext& context) {
   auto pads = context.Attr<std::vector<int>>("paddings");
   Eigen::array<std::pair<int, int>, D> paddings;
@@ -91,12 +92,13 @@ void PadGradFunction(const framework::ExecutionContext& context) {
     d_x->mutable_data<T>(context.GetPlace());
     auto d_x_tensor = EigenTensor<T, D>::From(*d_x);
     auto d_out_tensor = EigenTensor<T, D>::From(*d_out);
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
     d_x_tensor.device(place) = d_out_tensor.pad(paddings, 0);
   }
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class PadGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -104,22 +106,22 @@ class PadGradKernel : public framework::OpKernel<T> {
         context.Input<Tensor>(framework::GradVarName("Out"))->dims().size();
     switch (rank) {
       case 1:
-        PadGradFunction<Place, T, 1>(context);
+        PadGradFunction<DeviceContext, T, 1>(context);
         break;
       case 2:
-        PadGradFunction<Place, T, 2>(context);
+        PadGradFunction<DeviceContext, T, 2>(context);
         break;
       case 3:
-        PadGradFunction<Place, T, 3>(context);
+        PadGradFunction<DeviceContext, T, 3>(context);
         break;
       case 4:
-        PadGradFunction<Place, T, 4>(context);
+        PadGradFunction<DeviceContext, T, 4>(context);
         break;
       case 5:
-        PadGradFunction<Place, T, 5>(context);
+        PadGradFunction<DeviceContext, T, 5>(context);
         break;
       case 6:
-        PadGradFunction<Place, T, 6>(context);
+        PadGradFunction<DeviceContext, T, 6>(context);
         break;
       default:
         PADDLE_THROW(
diff --git a/paddle/operators/pool_cudnn_op.cc b/paddle/operators/pool_cudnn_op.cc
index be9fcc5661..77407f5cdf 100644
--- a/paddle/operators/pool_cudnn_op.cc
+++ b/paddle/operators/pool_cudnn_op.cc
@@ -19,19 +19,21 @@ namespace ops = paddle::operators;
 REGISTER_OP(pool2d_cudnn, ops::PoolOp, ops::Pool2dOpMaker, pool2d_cudnn_grad,
             ops::PoolOpGrad);
 
-REGISTER_OP_CPU_KERNEL(pool2d_cudnn,
-                       ops::PoolKernel<paddle::platform::CPUPlace, float>,
-                       ops::PoolKernel<paddle::platform::CPUPlace, double>);
-REGISTER_OP_CPU_KERNEL(pool2d_cudnn_grad,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, double>)
+REGISTER_OP_CPU_KERNEL(
+    pool2d_cudnn, ops::PoolKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::PoolKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    pool2d_cudnn_grad,
+    ops::PoolGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::PoolGradKernel<paddle::platform::CPUDeviceContext, double>)
 
 REGISTER_OP(pool3d_cudnn, ops::PoolOp, ops::Pool3dOpMaker, pool3d_cudnn_grad,
             ops::PoolOpGrad);
 
-REGISTER_OP_CPU_KERNEL(pool3d_cudnn,
-                       ops::PoolKernel<paddle::platform::CPUPlace, float>,
-                       ops::PoolKernel<paddle::platform::CPUPlace, double>);
-REGISTER_OP_CPU_KERNEL(pool3d_cudnn_grad,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, double>)
+REGISTER_OP_CPU_KERNEL(
+    pool3d_cudnn, ops::PoolKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::PoolKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    pool3d_cudnn_grad,
+    ops::PoolGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::PoolGradKernel<paddle::platform::CPUDeviceContext, double>)
diff --git a/paddle/operators/pool_cudnn_op.cu.cc b/paddle/operators/pool_cudnn_op.cu.cc
index 66dd194ccd..fc2b37bd0f 100644
--- a/paddle/operators/pool_cudnn_op.cu.cc
+++ b/paddle/operators/pool_cudnn_op.cu.cc
@@ -162,12 +162,12 @@ class PoolCudnnGradOpKernel : public framework::OpKernel<T> {
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(pool2d_cudnn, ops::PoolCudnnOpKernel<float>,
-                       ops::PoolCudnnOpKernel<double>);
-REGISTER_OP_GPU_KERNEL(pool2d_cudnn_grad, ops::PoolCudnnGradOpKernel<float>,
-                       ops::PoolCudnnGradOpKernel<double>);
-
-REGISTER_OP_GPU_KERNEL(pool3d_cudnn, ops::PoolCudnnOpKernel<float>,
-                       ops::PoolCudnnOpKernel<double>);
-REGISTER_OP_GPU_KERNEL(pool3d_cudnn_grad, ops::PoolCudnnGradOpKernel<float>,
-                       ops::PoolCudnnGradOpKernel<double>);
+REGISTER_OP_CUDA_KERNEL(pool2d_cudnn, ops::PoolCudnnOpKernel<float>,
+                        ops::PoolCudnnOpKernel<double>);
+REGISTER_OP_CUDA_KERNEL(pool2d_cudnn_grad, ops::PoolCudnnGradOpKernel<float>,
+                        ops::PoolCudnnGradOpKernel<double>);
+
+REGISTER_OP_CUDA_KERNEL(pool3d_cudnn, ops::PoolCudnnOpKernel<float>,
+                        ops::PoolCudnnOpKernel<double>);
+REGISTER_OP_CUDA_KERNEL(pool3d_cudnn_grad, ops::PoolCudnnGradOpKernel<float>,
+                        ops::PoolCudnnGradOpKernel<double>);
diff --git a/paddle/operators/pool_op.cc b/paddle/operators/pool_op.cc
index e26ffd86e5..45fa20280c 100644
--- a/paddle/operators/pool_op.cc
+++ b/paddle/operators/pool_op.cc
@@ -216,19 +216,19 @@ namespace ops = paddle::operators;
 REGISTER_OP(pool2d, ops::PoolOp, ops::Pool2dOpMaker, pool2d_grad,
             ops::PoolOpGrad);
 
-REGISTER_OP_CPU_KERNEL(pool2d,
-                       ops::PoolKernel<paddle::platform::CPUPlace, float>,
-                       ops::PoolKernel<paddle::platform::CPUPlace, double>);
-REGISTER_OP_CPU_KERNEL(pool2d_grad,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, double>)
+REGISTER_OP_CPU_KERNEL(
+    pool2d, ops::PoolKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::PoolKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    pool2d_grad, ops::PoolGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::PoolGradKernel<paddle::platform::CPUDeviceContext, double>)
 
 REGISTER_OP(pool3d, ops::PoolOp, ops::Pool3dOpMaker, pool3d_grad,
             ops::PoolOpGrad);
 
-REGISTER_OP_CPU_KERNEL(pool3d,
-                       ops::PoolKernel<paddle::platform::CPUPlace, float>,
-                       ops::PoolKernel<paddle::platform::CPUPlace, double>);
-REGISTER_OP_CPU_KERNEL(pool3d_grad,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(
+    pool3d, ops::PoolKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::PoolKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    pool3d_grad, ops::PoolGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::PoolGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/pool_op.cu.cc b/paddle/operators/pool_op.cu.cc
index 1010cb7622..39a9dfbf79 100644
--- a/paddle/operators/pool_op.cu.cc
+++ b/paddle/operators/pool_op.cu.cc
@@ -16,16 +16,18 @@ limitations under the License. */
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(pool2d,
-                       ops::PoolKernel<paddle::platform::GPUPlace, float>,
-                       ops::PoolKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(pool2d_grad,
-                       ops::PoolGradKernel<paddle::platform::GPUPlace, float>,
-                       ops::PoolGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    pool2d, ops::PoolKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::PoolKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    pool2d_grad,
+    ops::PoolGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::PoolGradKernel<paddle::platform::CUDADeviceContext, double>);
 
-REGISTER_OP_GPU_KERNEL(pool3d,
-                       ops::PoolKernel<paddle::platform::GPUPlace, float>,
-                       ops::PoolKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(pool3d_grad,
-                       ops::PoolGradKernel<paddle::platform::GPUPlace, float>,
-                       ops::PoolGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    pool3d, ops::PoolKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::PoolKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    pool3d_grad,
+    ops::PoolGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::PoolGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/pool_op.h b/paddle/operators/pool_op.h
index 63492a89e8..ab85d587a3 100644
--- a/paddle/operators/pool_op.h
+++ b/paddle/operators/pool_op.h
@@ -50,7 +50,7 @@ class Pool3dOpMaker : public framework::OpProtoAndCheckerMaker {
                 framework::OpAttrChecker* op_checker);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class PoolKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -67,41 +67,41 @@ class PoolKernel : public framework::OpKernel<T> {
         ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
       }
     }
-
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     switch (ksize.size()) {
       case 2: {
         if (pooling_type == "max") {
           paddle::operators::math::Pool2dFunctor<
-              Place, paddle::operators::math::MaxPool<T>, T>
+              DeviceContext, paddle::operators::math::MaxPool<T>, T>
               pool2d_forward;
           paddle::operators::math::MaxPool<T> pool_process;
-          pool2d_forward(context.device_context(), *in_x, ksize, strides,
-                         paddings, pool_process, out);
+          pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process,
+                         out);
 
         } else if (pooling_type == "avg") {
           paddle::operators::math::Pool2dFunctor<
-              Place, paddle::operators::math::AvgPool<T>, T>
+              DeviceContext, paddle::operators::math::AvgPool<T>, T>
               pool2d_forward;
           paddle::operators::math::AvgPool<T> pool_process;
-          pool2d_forward(context.device_context(), *in_x, ksize, strides,
-                         paddings, pool_process, out);
+          pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process,
+                         out);
         }
       } break;
       case 3: {
         if (pooling_type == "max") {
           paddle::operators::math::Pool3dFunctor<
-              Place, paddle::operators::math::MaxPool<T>, T>
+              DeviceContext, paddle::operators::math::MaxPool<T>, T>
               pool3d_forward;
           paddle::operators::math::MaxPool<T> pool_process;
-          pool3d_forward(context.device_context(), *in_x, ksize, strides,
-                         paddings, pool_process, out);
+          pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process,
+                         out);
         } else if (pooling_type == "avg") {
           paddle::operators::math::Pool3dFunctor<
-              Place, paddle::operators::math::AvgPool<T>, T>
+              DeviceContext, paddle::operators::math::AvgPool<T>, T>
               pool3d_forward;
           paddle::operators::math::AvgPool<T> pool_process;
-          pool3d_forward(context.device_context(), *in_x, ksize, strides,
-                         paddings, pool_process, out);
+          pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process,
+                         out);
         }
       } break;
       default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }
@@ -109,7 +109,7 @@ class PoolKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class PoolGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -130,42 +130,43 @@ class PoolGradKernel : public framework::OpKernel<T> {
         ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
       }
     }
-
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     if (in_x_grad) {
       in_x_grad->mutable_data<T>(context.GetPlace());
       auto temp = framework::EigenVector<T>::Flatten(*in_x_grad);
-      temp.device(context.GetEigenDevice<Place>()) =
+      temp.device(
+          *context.template device_context<DeviceContext>().eigen_device()) =
           temp.constant(static_cast<T>(0));
 
       switch (ksize.size()) {
         case 2: {
           if (pooling_type == "max") {
-            paddle::operators::math::MaxPool2dGradFunctor<Place, T>
+            paddle::operators::math::MaxPool2dGradFunctor<DeviceContext, T>
                 pool2d_backward;
-            pool2d_backward(context.device_context(), *in_x, *out, *out_grad,
-                            ksize, strides, paddings, in_x_grad);
+            pool2d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides,
+                            paddings, in_x_grad);
           } else if (pooling_type == "avg") {
             paddle::operators::math::Pool2dGradFunctor<
-                Place, paddle::operators::math::AvgPoolGrad<T>, T>
+                DeviceContext, paddle::operators::math::AvgPoolGrad<T>, T>
                 pool2d_backward;
             paddle::operators::math::AvgPoolGrad<T> pool_process;
-            pool2d_backward(context.device_context(), *in_x, *out, *out_grad,
-                            ksize, strides, paddings, pool_process, in_x_grad);
+            pool2d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides,
+                            paddings, pool_process, in_x_grad);
           }
         } break;
         case 3: {
           if (pooling_type == "max") {
-            paddle::operators::math::MaxPool3dGradFunctor<Place, T>
+            paddle::operators::math::MaxPool3dGradFunctor<DeviceContext, T>
                 pool3d_backward;
-            pool3d_backward(context.device_context(), *in_x, *out, *out_grad,
-                            ksize, strides, paddings, in_x_grad);
+            pool3d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides,
+                            paddings, in_x_grad);
           } else if (pooling_type == "avg") {
             paddle::operators::math::Pool3dGradFunctor<
-                Place, paddle::operators::math::AvgPoolGrad<T>, T>
+                DeviceContext, paddle::operators::math::AvgPoolGrad<T>, T>
                 pool3d_backward;
             paddle::operators::math::AvgPoolGrad<T> pool_process;
-            pool3d_backward(context.device_context(), *in_x, *out, *out_grad,
-                            ksize, strides, paddings, pool_process, in_x_grad);
+            pool3d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides,
+                            paddings, pool_process, in_x_grad);
           }
         } break;
         default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }
diff --git a/paddle/operators/pool_with_index_op.cc b/paddle/operators/pool_with_index_op.cc
index b9c42a6912..1a2383f8b8 100644
--- a/paddle/operators/pool_with_index_op.cc
+++ b/paddle/operators/pool_with_index_op.cc
@@ -266,12 +266,15 @@ REGISTER_OP(max_pool2d_with_index, ops::MaxPoolWithIndexOp,
 
 REGISTER_OP_CPU_KERNEL(
     max_pool2d_with_index,
-    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, float, int>,
-    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, double, int>);
+    ops::MaxPoolWithIndexKernel<paddle::platform::CPUDeviceContext, float, int>,
+    ops::MaxPoolWithIndexKernel<paddle::platform::CPUDeviceContext, double,
+                                int>);
 REGISTER_OP_CPU_KERNEL(
     max_pool2d_with_index_grad,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, float, int>,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, double, int>)
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUDeviceContext, float,
+                                    int>,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUDeviceContext, double,
+                                    int>)
 
 REGISTER_OP(max_pool3d_with_index, ops::MaxPoolWithIndexOp,
             ops::MaxPool3dWithIndexOpMaker, max_pool3d_with_index_grad,
@@ -279,9 +282,12 @@ REGISTER_OP(max_pool3d_with_index, ops::MaxPoolWithIndexOp,
 
 REGISTER_OP_CPU_KERNEL(
     max_pool3d_with_index,
-    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, float, int>,
-    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, double, int>);
+    ops::MaxPoolWithIndexKernel<paddle::platform::CPUDeviceContext, float, int>,
+    ops::MaxPoolWithIndexKernel<paddle::platform::CPUDeviceContext, double,
+                                int>);
 REGISTER_OP_CPU_KERNEL(
     max_pool3d_with_index_grad,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, float, int>,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, double, int>)
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUDeviceContext, float,
+                                    int>,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUDeviceContext, double,
+                                    int>)
diff --git a/paddle/operators/pool_with_index_op.cu.cc b/paddle/operators/pool_with_index_op.cu.cc
index 335064a7ee..4c9804da63 100644
--- a/paddle/operators/pool_with_index_op.cu.cc
+++ b/paddle/operators/pool_with_index_op.cu.cc
@@ -16,20 +16,28 @@ limitations under the License. */
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     max_pool2d_with_index,
-    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, float, int>,
-    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, double, int>);
-REGISTER_OP_GPU_KERNEL(
+    ops::MaxPoolWithIndexKernel<paddle::platform::CUDADeviceContext, float,
+                                int>,
+    ops::MaxPoolWithIndexKernel<paddle::platform::CUDADeviceContext, double,
+                                int>);
+REGISTER_OP_CUDA_KERNEL(
     max_pool2d_with_index_grad,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, float, int>,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, double, int>)
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CUDADeviceContext, float,
+                                    int>,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CUDADeviceContext, double,
+                                    int>)
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     max_pool3d_with_index,
-    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, float, int>,
-    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, double, int>);
-REGISTER_OP_GPU_KERNEL(
+    ops::MaxPoolWithIndexKernel<paddle::platform::CUDADeviceContext, float,
+                                int>,
+    ops::MaxPoolWithIndexKernel<paddle::platform::CUDADeviceContext, double,
+                                int>);
+REGISTER_OP_CUDA_KERNEL(
     max_pool3d_with_index_grad,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, float, int>,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, double, int>)
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CUDADeviceContext, float,
+                                    int>,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CUDADeviceContext, double,
+                                    int>)
diff --git a/paddle/operators/pool_with_index_op.h b/paddle/operators/pool_with_index_op.h
index 40766c7e82..4f4087d1dd 100644
--- a/paddle/operators/pool_with_index_op.h
+++ b/paddle/operators/pool_with_index_op.h
@@ -24,7 +24,7 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-template <typename Place, typename T1, typename T2>
+template <typename DeviceContext, typename T1, typename T2>
 class MaxPoolWithIndexKernel : public framework::OpKernel<T1> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -35,6 +35,8 @@ class MaxPoolWithIndexKernel : public framework::OpKernel<T1> {
     std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
+
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     if (context.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
@@ -44,23 +46,23 @@ class MaxPoolWithIndexKernel : public framework::OpKernel<T1> {
 
     switch (ksize.size()) {
       case 2: {
-        paddle::operators::math::MaxPool2dWithIndexFunctor<Place, T1, T2>
+        paddle::operators::math::MaxPool2dWithIndexFunctor<DeviceContext, T1,
+                                                           T2>
             pool2d_forward;
-        pool2d_forward(context.device_context(), *in_x, ksize, strides,
-                       paddings, out, mask);
+        pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, out, mask);
       } break;
       case 3: {
-        paddle::operators::math::MaxPool3dWithIndexFunctor<Place, T1, T2>
+        paddle::operators::math::MaxPool3dWithIndexFunctor<DeviceContext, T1,
+                                                           T2>
             pool3d_forward;
-        pool3d_forward(context.device_context(), *in_x, ksize, strides,
-                       paddings, out, mask);
+        pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, out, mask);
       } break;
       default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }
     }
   }
 };
 
-template <typename Place, typename T1, typename T2>
+template <typename DeviceContext, typename T1, typename T2>
 class MaxPoolWithIndexGradKernel : public framework::OpKernel<T1> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -81,18 +83,20 @@ class MaxPoolWithIndexGradKernel : public framework::OpKernel<T1> {
 
     if (in_x_grad) {
       in_x_grad->mutable_data<T1>(context.GetPlace());
-      auto& device_ctx = context.device_context();
+      auto& device_ctx = context.template device_context<DeviceContext>();
       math::set_constant(device_ctx, in_x_grad, 0);
 
       switch (ksize.size()) {
         case 2: {
-          paddle::operators::math::MaxPool2dWithIndexGradFunctor<Place, T1, T2>
+          paddle::operators::math::MaxPool2dWithIndexGradFunctor<DeviceContext,
+                                                                 T1, T2>
               pool2d_backward;
           pool2d_backward(device_ctx, *out_grad, *mask, ksize, strides,
                           paddings, in_x_grad);
         } break;
         case 3: {
-          paddle::operators::math::MaxPool3dWithIndexGradFunctor<Place, T1, T2>
+          paddle::operators::math::MaxPool3dWithIndexGradFunctor<DeviceContext,
+                                                                 T1, T2>
               pool3d_backward;
           pool3d_backward(device_ctx, *out_grad, *mask, ksize, strides,
                           paddings, in_x_grad);
diff --git a/paddle/operators/positive_negative_pair_op.h b/paddle/operators/positive_negative_pair_op.h
index 2efd3777e0..977e59b7d2 100644
--- a/paddle/operators/positive_negative_pair_op.h
+++ b/paddle/operators/positive_negative_pair_op.h
@@ -22,7 +22,7 @@ namespace operators {
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class PositiveNegativePairKernel : public framework::OpKernel<T> {
  public:
   struct PredictionResult {
diff --git a/paddle/operators/precision_recall_op.h b/paddle/operators/precision_recall_op.h
index 4a871ce674..c0d55405a3 100644
--- a/paddle/operators/precision_recall_op.h
+++ b/paddle/operators/precision_recall_op.h
@@ -26,7 +26,7 @@ using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
 enum StateVariable { TP = 0, FP, TN, FN };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class PrecisionRecallKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
diff --git a/paddle/operators/prelu_op.cc b/paddle/operators/prelu_op.cc
index 055c471b45..317a2a4015 100644
--- a/paddle/operators/prelu_op.cc
+++ b/paddle/operators/prelu_op.cc
@@ -85,7 +85,8 @@ namespace ops = paddle::operators;
 
 REGISTER_OP(prelu, ops::PReluOp, ops::PReluOpMaker, prelu_grad,
             ops::PReluGradOp);
-REGISTER_OP_CPU_KERNEL(prelu,
-                       ops::PReluKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(prelu_grad,
-                       ops::PReluGradKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    prelu, ops::PReluKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    prelu_grad,
+    ops::PReluGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/prelu_op.cu b/paddle/operators/prelu_op.cu
index 9e391dabae..12033dee0e 100644
--- a/paddle/operators/prelu_op.cu
+++ b/paddle/operators/prelu_op.cu
@@ -14,8 +14,9 @@
 
 #include "paddle/operators/prelu_op.h"
 
-REGISTER_OP_GPU_KERNEL(
-    prelu, paddle::operators::PReluKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    prelu_grad,
-    paddle::operators::PReluGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    prelu,
+    paddle::operators::PReluKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(prelu_grad,
+                        paddle::operators::PReluGradKernel<
+                            paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/prelu_op.h b/paddle/operators/prelu_op.h
index 5ad31c2203..56f9a553ec 100644
--- a/paddle/operators/prelu_op.h
+++ b/paddle/operators/prelu_op.h
@@ -39,7 +39,7 @@ class PReluFunctor {
   const T* alpha_;
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class PReluKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -54,9 +54,9 @@ class PReluKernel : public framework::OpKernel<T> {
 
     int numel = x->numel();
 
-    Transform<Place> trans;
-    trans(context.device_context(), x_ptr, x_ptr + numel, o_ptr,
-          PReluFunctor<T>(alpha_ptr));
+    Transform<DeviceContext> trans;
+    trans(context.template device_context<DeviceContext>(), x_ptr,
+          x_ptr + numel, o_ptr, PReluFunctor<T>(alpha_ptr));
   }
 };
 
@@ -76,7 +76,7 @@ class PReluGradFunctor {
   const T* alpha_;
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class PReluGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -92,9 +92,9 @@ class PReluGradKernel : public framework::OpKernel<T> {
     const T* out_ptr = out->data<T>();
     int numel = dx->numel();
 
-    Transform<Place> trans;
-    trans(context.device_context(), out_ptr, out_ptr + numel, dout_ptr, dx_ptr,
-          PReluGradFunctor<T>(alpha_ptr));
+    Transform<DeviceContext> trans;
+    trans(context.template device_context<DeviceContext>(), out_ptr,
+          out_ptr + numel, dout_ptr, dx_ptr, PReluGradFunctor<T>(alpha_ptr));
 
     // TODO(Zhuoyuan): add dalpha upgrade when GPU kernels ready
   }
diff --git a/paddle/operators/proximal_adagrad_op.cc b/paddle/operators/proximal_adagrad_op.cc
index 36e460103a..cc350f6d26 100644
--- a/paddle/operators/proximal_adagrad_op.cc
+++ b/paddle/operators/proximal_adagrad_op.cc
@@ -114,4 +114,4 @@ REGISTER_OP_WITHOUT_GRADIENT(proximal_adagrad, ops::ProximalAdagradOp,
                              ops::ProximalAdagradOpMaker);
 REGISTER_OP_CPU_KERNEL(
     proximal_adagrad,
-    ops::ProximalAdagradOpKernel<paddle::platform::CPUPlace, float>);
+    ops::ProximalAdagradOpKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/proximal_adagrad_op.cu b/paddle/operators/proximal_adagrad_op.cu
index d0ae039518..42a178f94b 100644
--- a/paddle/operators/proximal_adagrad_op.cu
+++ b/paddle/operators/proximal_adagrad_op.cu
@@ -15,6 +15,6 @@ specific language governing permissions and limitations under the License. */
 #include "paddle/operators/proximal_adagrad_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     proximal_adagrad,
-    ops::ProximalAdagradOpKernel<paddle::platform::GPUPlace, float>);
+    ops::ProximalAdagradOpKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/proximal_adagrad_op.h b/paddle/operators/proximal_adagrad_op.h
index 7a1560e8cb..523924d80e 100644
--- a/paddle/operators/proximal_adagrad_op.h
+++ b/paddle/operators/proximal_adagrad_op.h
@@ -24,7 +24,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ProximalAdagradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -45,20 +45,20 @@ class ProximalAdagradOpKernel : public framework::OpKernel<T> {
 
     auto p_out = EigenVector<T>::Flatten(*param_out);
     auto m_out = EigenVector<T>::Flatten(*moment_out);
-    auto place = ctx.GetEigenDevice<Place>();
+    auto* place = ctx.template device_context<DeviceContext>().eigen_device();
 
     Eigen::DSizes<int, 1> grad_dsize(grad->numel());
 
-    m_out.device(place) = m + g * g;
+    m_out.device(*place) = m + g * g;
     auto prox_param = p - lr.broadcast(grad_dsize) * g / m_out.sqrt();
     if (l1 > static_cast<T>(0)) {
-      p_out.device(place) =
+      p_out.device(*place) =
           prox_param.sign() *
           (((prox_param.abs() - (lr * l1).broadcast(grad_dsize))
                 .cwiseMax(static_cast<T>(0.0))) /
            (static_cast<T>(1.0) + (lr * l2).broadcast(grad_dsize)));
     } else {
-      p_out.device(place) =
+      p_out.device(*place) =
           prox_param / (static_cast<T>(1.0) + (lr * l2).broadcast(grad_dsize));
     }
   }
diff --git a/paddle/operators/proximal_gd_op.cc b/paddle/operators/proximal_gd_op.cc
index 5693d0ec9e..0b26beb3ac 100644
--- a/paddle/operators/proximal_gd_op.cc
+++ b/paddle/operators/proximal_gd_op.cc
@@ -94,4 +94,5 @@ namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(proximal_gd, ops::ProximalGDOp,
                              ops::ProximalGDOpMaker);
 REGISTER_OP_CPU_KERNEL(
-    proximal_gd, ops::ProximalGDOpKernel<paddle::platform::CPUPlace, float>);
+    proximal_gd,
+    ops::ProximalGDOpKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/proximal_gd_op.cu b/paddle/operators/proximal_gd_op.cu
index 26f4ebaa0f..b7dd840d19 100644
--- a/paddle/operators/proximal_gd_op.cu
+++ b/paddle/operators/proximal_gd_op.cu
@@ -15,5 +15,6 @@ specific language governing permissions and limitations under the License. */
 #include "paddle/operators/proximal_gd_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    proximal_gd, ops::ProximalGDOpKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    proximal_gd,
+    ops::ProximalGDOpKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/proximal_gd_op.h b/paddle/operators/proximal_gd_op.h
index bebda02041..64648b3cca 100644
--- a/paddle/operators/proximal_gd_op.h
+++ b/paddle/operators/proximal_gd_op.h
@@ -24,7 +24,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ProximalGDOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -42,7 +42,7 @@ class ProximalGDOpKernel : public framework::OpKernel<T> {
     auto lr = EigenVector<T>::Flatten(*ctx.Input<Tensor>("LearningRate"));
 
     auto p_out = EigenVector<T>::Flatten(*param_out);
-    auto place = ctx.GetEigenDevice<Place>();
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
 
     Eigen::DSizes<int, 1> grad_dsize(grad->numel());
 
diff --git a/paddle/operators/rank_loss_op.cc b/paddle/operators/rank_loss_op.cc
index 912f88f455..b80b175792 100644
--- a/paddle/operators/rank_loss_op.cc
+++ b/paddle/operators/rank_loss_op.cc
@@ -123,7 +123,8 @@ namespace ops = paddle::operators;
 
 REGISTER_OP(rank_loss, ops::RankLossOp, ops::RankLossOpMaker, rank_loss_grad,
             ops::RankLossGradOp);
-REGISTER_OP_CPU_KERNEL(rank_loss,
-                       ops::RankLossKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
-    rank_loss_grad, ops::RankLossGradKernel<paddle::platform::CPUPlace, float>);
+    rank_loss, ops::RankLossKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    rank_loss_grad,
+    ops::RankLossGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/rank_loss_op.cu b/paddle/operators/rank_loss_op.cu
index 5382e3a629..5aee66443d 100644
--- a/paddle/operators/rank_loss_op.cu
+++ b/paddle/operators/rank_loss_op.cu
@@ -14,9 +14,9 @@
 
 #include "paddle/operators/rank_loss_op.h"
 
-REGISTER_OP_GPU_KERNEL(
-    rank_loss,
-    paddle::operators::RankLossKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    rank_loss_grad,
-    paddle::operators::RankLossGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(rank_loss,
+                        paddle::operators::RankLossKernel<
+                            paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(rank_loss_grad,
+                        paddle::operators::RankLossGradKernel<
+                            paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/rank_loss_op.h b/paddle/operators/rank_loss_op.h
index 703c77a0b2..ea24b61fd9 100644
--- a/paddle/operators/rank_loss_op.h
+++ b/paddle/operators/rank_loss_op.h
@@ -20,7 +20,7 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class RankLossKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
@@ -35,13 +35,13 @@ class RankLossKernel : public framework::OpKernel<T> {
     auto left = framework::EigenVector<T>::Flatten(*left_t);
     auto right = framework::EigenVector<T>::Flatten(*right_t);
 
-    auto& dev = ctx.GetEigenDevice<Place>();
+    auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
     out.device(dev) =
         (1. + (left - right).exp()).log() - label * (left - right);
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class RankLossGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
@@ -55,7 +55,7 @@ class RankLossGradKernel : public framework::OpKernel<T> {
     auto* left_t = ctx.Input<framework::Tensor>("Left");
     auto* right_t = ctx.Input<framework::Tensor>("Right");
 
-    auto& dev = ctx.GetEigenDevice<Place>();
+    auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
     auto d_out = framework::EigenVector<T>::Flatten(*d_out_t);
     auto label = framework::EigenVector<T>::Flatten(*label_t);
     auto left = framework::EigenVector<T>::Flatten(*left_t);
diff --git a/paddle/operators/reduce_op.cc b/paddle/operators/reduce_op.cc
index 2589a54cfc..b754637bf2 100644
--- a/paddle/operators/reduce_op.cc
+++ b/paddle/operators/reduce_op.cc
@@ -180,12 +180,13 @@ REGISTER_OP(reduce_max, ops::ReduceOp, ops::ReduceMaxOpMaker, reduce_max_grad,
 REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker, reduce_min_grad,
             ops::ReduceGradOp);
 
-#define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor)     \
-  REGISTER_OP_CPU_KERNEL(                                                  \
-      reduce_type,                                                         \
-      ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::functor>); \
-  REGISTER_OP_CPU_KERNEL(reduce_type##_grad,                               \
-                         ops::ReduceGradKernel<paddle::platform::CPUPlace, \
-                                               float, ops::grad_functor>);
+#define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor)         \
+  REGISTER_OP_CPU_KERNEL(reduce_type,                                          \
+                         ops::ReduceKernel<paddle::platform::CPUDeviceContext, \
+                                           float, ops::functor>);              \
+  REGISTER_OP_CPU_KERNEL(                                                      \
+      reduce_type##_grad,                                                      \
+      ops::ReduceGradKernel<paddle::platform::CPUDeviceContext, float,         \
+                            ops::grad_functor>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_CPU_KERNEL);
diff --git a/paddle/operators/reduce_op.cu b/paddle/operators/reduce_op.cu
index d306e1a240..a10ace5253 100644
--- a/paddle/operators/reduce_op.cu
+++ b/paddle/operators/reduce_op.cu
@@ -17,12 +17,13 @@
 
 namespace ops = paddle::operators;
 
-#define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor)     \
-  REGISTER_OP_GPU_KERNEL(                                                  \
-      reduce_type,                                                         \
-      ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::functor>); \
-  REGISTER_OP_GPU_KERNEL(reduce_type##_grad,                               \
-                         ops::ReduceGradKernel<paddle::platform::GPUPlace, \
-                                               float, ops::grad_functor>);
+#define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor)    \
+  REGISTER_OP_CUDA_KERNEL(                                                \
+      reduce_type, ops::ReduceKernel<paddle::platform::CUDADeviceContext, \
+                                     float, ops::functor>);               \
+  REGISTER_OP_CUDA_KERNEL(                                                \
+      reduce_type##_grad,                                                 \
+      ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, float,   \
+                            ops::grad_functor>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_GPU_KERNEL);
diff --git a/paddle/operators/reduce_op.h b/paddle/operators/reduce_op.h
index dd6547542d..47ce910f28 100644
--- a/paddle/operators/reduce_op.h
+++ b/paddle/operators/reduce_op.h
@@ -32,55 +32,55 @@ template <typename T, int MajorType = Eigen::RowMajor,
 using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
 
 struct SumFunctor {
-  template <typename Place, typename X, typename Y, typename Dim>
-  void operator()(const Place& place, X& x, Y& y, const Dim& dim) {
+  template <typename DeviceContext, typename X, typename Y, typename Dim>
+  void operator()(const DeviceContext& place, X& x, Y& y, const Dim& dim) {
     y.device(place) = x.sum(dim);
   }
 };
 
 struct SumGradFunctor {
-  template <typename Place, typename X, typename Y, typename DX, typename DY,
-            typename Dim>
-  void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy,
+  template <typename DeviceContext, typename X, typename Y, typename DX,
+            typename DY, typename Dim>
+  void operator()(const DeviceContext& place, X& x, Y& y, DX& dx, DY& dy,
                   const Dim& dim, int size) {
     dx.device(place) = dy.broadcast(dim);
   }
 };
 
 struct MeanFunctor {
-  template <typename Place, typename X, typename Y, typename Dim>
-  void operator()(const Place& place, X& x, Y& y, const Dim& dim) {
+  template <typename DeviceContext, typename X, typename Y, typename Dim>
+  void operator()(const DeviceContext& place, X& x, Y& y, const Dim& dim) {
     y.device(place) = x.mean(dim);
   }
 };
 
 struct MeanGradFunctor {
-  template <typename Place, typename X, typename Y, typename DX, typename DY,
-            typename Dim>
-  void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy,
+  template <typename DeviceContext, typename X, typename Y, typename DX,
+            typename DY, typename Dim>
+  void operator()(const DeviceContext& place, X& x, Y& y, DX& dx, DY& dy,
                   const Dim& dim, int size) {
     dx.device(place) = dy.broadcast(dim) / dx.constant(size);
   }
 };
 
 struct MaxFunctor {
-  template <typename Place, typename X, typename Y, typename Dim>
-  void operator()(const Place& place, X& x, Y& y, const Dim& dim) {
+  template <typename DeviceContext, typename X, typename Y, typename Dim>
+  void operator()(const DeviceContext& place, X& x, Y& y, const Dim& dim) {
     y.device(place) = x.maximum(dim);
   }
 };
 
 struct MinFunctor {
-  template <typename Place, typename X, typename Y, typename Dim>
-  void operator()(const Place& place, X& x, Y& y, const Dim& dim) {
+  template <typename DeviceContext, typename X, typename Y, typename Dim>
+  void operator()(const DeviceContext& place, X& x, Y& y, const Dim& dim) {
     y.device(place) = x.minimum(dim);
   }
 };
 
 struct MaxOrMinGradFunctor {
-  template <typename Place, typename X, typename Y, typename DX, typename DY,
-            typename Dim>
-  void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy,
+  template <typename DeviceContext, typename X, typename Y, typename DX,
+            typename DY, typename Dim>
+  void operator()(const DeviceContext& place, X& x, Y& y, DX& dx, DY& dy,
                   const Dim& dim, int size) {
     auto equals = x == y.broadcast(dim);
     auto ones = dx.constant(1);
@@ -91,7 +91,7 @@ struct MaxOrMinGradFunctor {
   }
 };
 
-template <typename Place, typename T, typename Functor>
+template <typename DeviceContext, typename T, typename Functor>
 class ReduceKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -139,7 +139,8 @@ class ReduceKernel : public framework::OpKernel<T> {
       dims = framework::make_ddim(dims_vector);
     }
 
-    auto& place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
     Functor functor;
 
     if (D == 1) {
@@ -152,7 +153,7 @@ class ReduceKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T, typename Functor>
+template <typename DeviceContext, typename T, typename Functor>
 class ReduceGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -201,7 +202,8 @@ class ReduceGradKernel : public framework::OpKernel<T> {
     Eigen::array<int, D> broadcast_dim;
     for (size_t i = 0; i < D; ++i) broadcast_dim[i] = 1;
     broadcast_dim[dim] = input0->dims()[dim];
-    auto& place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
     Functor functor;
     functor(place, x, x_reduce, x_grad, x_reduce_grad, broadcast_dim,
             broadcast_dim[dim]);
diff --git a/paddle/operators/reshape_op.cu b/paddle/operators/reshape_op.cu
index dca6c15007..b7329238c0 100644
--- a/paddle/operators/reshape_op.cu
+++ b/paddle/operators/reshape_op.cu
@@ -14,9 +14,9 @@
 
 #include "paddle/operators/reshape_op.h"
 
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     reshape,
     paddle::operators::ReshapeKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     reshape_grad,
     paddle::operators::ReshapeGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/reshape_op.h b/paddle/operators/reshape_op.h
index 73fd1da642..92d8cbbb56 100644
--- a/paddle/operators/reshape_op.h
+++ b/paddle/operators/reshape_op.h
@@ -20,7 +20,7 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ReshapeKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
@@ -33,7 +33,7 @@ class ReshapeKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ReshapeGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
diff --git a/paddle/operators/rmsprop_op.cc b/paddle/operators/rmsprop_op.cc
index a9c45f639c..fc3f9b8988 100644
--- a/paddle/operators/rmsprop_op.cc
+++ b/paddle/operators/rmsprop_op.cc
@@ -116,5 +116,5 @@ http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf)
 
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(rmsprop, ops::RmspropOp, ops::RmspropOpMaker);
-REGISTER_OP_CPU_KERNEL(rmsprop,
-                       ops::RmspropOpKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    rmsprop, ops::RmspropOpKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/rmsprop_op.cu b/paddle/operators/rmsprop_op.cu
index 52634a5481..2a9fd6e104 100644
--- a/paddle/operators/rmsprop_op.cu
+++ b/paddle/operators/rmsprop_op.cu
@@ -16,5 +16,5 @@
 #include "paddle/operators/rmsprop_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(rmsprop,
-                       ops::RmspropOpKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    rmsprop, ops::RmspropOpKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/rmsprop_op.h b/paddle/operators/rmsprop_op.h
index 7bf2129010..16a561835d 100644
--- a/paddle/operators/rmsprop_op.h
+++ b/paddle/operators/rmsprop_op.h
@@ -24,7 +24,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class RmspropOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -51,7 +51,7 @@ class RmspropOpKernel : public framework::OpKernel<T> {
     auto p_out = EigenVector<T>::Flatten(*param_out);
     auto mom_out = EigenVector<T>::Flatten(*moment_out);
     auto ms_out = EigenVector<T>::Flatten(*mean_square_out);
-    auto place = ctx.GetEigenDevice<Place>();
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
 
     Eigen::DSizes<int, 1> grad_dsize(grad->numel());
 
diff --git a/paddle/operators/roi_pool_op.cc b/paddle/operators/roi_pool_op.cc
index 2b5e66c96b..75fcea8401 100644
--- a/paddle/operators/roi_pool_op.cc
+++ b/paddle/operators/roi_pool_op.cc
@@ -157,9 +157,10 @@ namespace ops = paddle::operators;
 REGISTER_OP(roi_pool, ops::ROIPoolOp, ops::ROIPoolOpMaker, roi_pool_grad,
             ops::ROIPoolGradOp);
 REGISTER_OP_CPU_KERNEL(
-    roi_pool, ops::CPUROIPoolOpKernel<paddle::platform::CPUPlace, float>,
-    ops::CPUROIPoolOpKernel<paddle::platform::CPUPlace, double>);
+    roi_pool,
+    ops::CPUROIPoolOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::CPUROIPoolOpKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     roi_pool_grad,
-    ops::CPUROIPoolGradOpKernel<paddle::platform::CPUPlace, float>,
-    ops::CPUROIPoolOpKernel<paddle::platform::CPUPlace, double>);
+    ops::CPUROIPoolGradOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::CPUROIPoolOpKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/roi_pool_op.cu b/paddle/operators/roi_pool_op.cu
index 9a4c8ca752..a874befe4d 100644
--- a/paddle/operators/roi_pool_op.cu
+++ b/paddle/operators/roi_pool_op.cu
@@ -177,7 +177,7 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel<T> {
     if (x_grad) {
       x_grad->mutable_data<T>(ctx.GetPlace());
       math::SetConstant<Place, T> set_zero;
-      set_zero(ctx.device_context(), x_grad, static_cast<T>(0));
+      set_zero(ctx.cuda_device_context(), x_grad, static_cast<T>(0));
 
       int output_grad_size = out_grad->numel();
       int blocks = NumBlocks(output_grad_size);
@@ -199,10 +199,11 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    roi_pool, ops::GPUROIPoolOpKernel<paddle::platform::GPUPlace, float>,
-    ops::GPUROIPoolOpKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    roi_pool,
+    ops::GPUROIPoolOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GPUROIPoolOpKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
     roi_pool_grad,
-    ops::GPUROIPoolGradOpKernel<paddle::platform::GPUPlace, float>,
-    ops::GPUROIPoolOpKernel<paddle::platform::GPUPlace, double>);
+    ops::GPUROIPoolGradOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::GPUROIPoolOpKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/roi_pool_op.h b/paddle/operators/roi_pool_op.h
index 3812c66c65..09a9d3d870 100644
--- a/paddle/operators/roi_pool_op.h
+++ b/paddle/operators/roi_pool_op.h
@@ -19,7 +19,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class CPUROIPoolOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -126,7 +126,7 @@ class CPUROIPoolOpKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class CPUROIPoolGradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -145,8 +145,9 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel<T> {
       const T* out_grad_data = out_grad->data<T>();
       const int64_t* argmax_data = argmax->data<int64_t>();
       T* in_grad_data = in_grad->mutable_data<T>(ctx.GetPlace());
-      math::SetConstant<Place, T> set_zero;
-      set_zero(ctx.device_context(), in_grad, static_cast<T>(0));
+      math::SetConstant<DeviceContext, T> set_zero;
+      set_zero(ctx.template device_context<DeviceContext>(), in_grad,
+               static_cast<T>(0));
 
       auto in_stride = framework::stride(in->dims());
       auto argmax_stride = framework::stride(argmax->dims());
diff --git a/paddle/operators/row_conv_op.cc b/paddle/operators/row_conv_op.cc
index ea0bb99f8d..5203a5079c 100644
--- a/paddle/operators/row_conv_op.cc
+++ b/paddle/operators/row_conv_op.cc
@@ -124,7 +124,8 @@ $$
 };
 
 template <typename T>
-class RowConvKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
+class RowConvKernel<platform::CPUDeviceContext, T>
+    : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
     auto *x = context.Input<LoDTensor>("X");
@@ -169,7 +170,8 @@ class RowConvKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
 };
 
 template <typename T>
-class RowConvGradKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
+class RowConvGradKernel<platform::CPUDeviceContext, T>
+    : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
     auto *x = context.Input<LoDTensor>("X");
@@ -251,7 +253,8 @@ class RowConvGradKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
 namespace ops = paddle::operators;
 REGISTER_OP(row_conv, ops::RowConvOp, ops::RowConvOpMaker, row_conv_grad,
             ops::RowConvGradOp);
-REGISTER_OP_CPU_KERNEL(row_conv,
-                       ops::RowConvKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
-    row_conv_grad, ops::RowConvGradKernel<paddle::platform::CPUPlace, float>);
+    row_conv, ops::RowConvKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    row_conv_grad,
+    ops::RowConvGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/row_conv_op.cu b/paddle/operators/row_conv_op.cu
index e0d7ebda7e..3fc5eabcf5 100644
--- a/paddle/operators/row_conv_op.cu
+++ b/paddle/operators/row_conv_op.cu
@@ -292,7 +292,8 @@ __global__ void RowConvGradFilter(const T *in, const T *dout, int num_sequence,
 }  // namespace
 
 template <typename T>
-class RowConvKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
+class RowConvKernel<platform::CUDADeviceContext, T>
+    : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
     auto *X = context.Input<LoDTensor>("X");
@@ -327,7 +328,8 @@ class RowConvKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
 };
 
 template <typename T>
-class RowConvGradKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
+class RowConvGradKernel<platform::CUDADeviceContext, T>
+    : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
     auto *X = context.Input<LoDTensor>("X");
@@ -347,7 +349,7 @@ class RowConvGradKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
     size_t *idx = batch_indices.data();
 
     auto &device_ctx = context.cuda_device_context();
-    math::SetConstant<platform::GPUPlace, T> zero;
+    math::SetConstant<platform::CUDADeviceContext, T> zero;
 
     if (dFilter) {
       T *dfilter = dFilter->mutable_data<T>(context.GetPlace());
@@ -402,7 +404,8 @@ class RowConvGradKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(row_conv,
-                       ops::RowConvKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    row_conv_grad, ops::RowConvGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    row_conv, ops::RowConvKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    row_conv_grad,
+    ops::RowConvGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/row_conv_op.h b/paddle/operators/row_conv_op.h
index 525e83908d..80912ad8f7 100644
--- a/paddle/operators/row_conv_op.h
+++ b/paddle/operators/row_conv_op.h
@@ -18,13 +18,13 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class RowConvKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override;
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class RowConvGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override;
diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc
index e5c10fec4d..d848be823e 100644
--- a/paddle/operators/scale_op.cc
+++ b/paddle/operators/scale_op.cc
@@ -75,8 +75,8 @@ namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker<float>,
                   ops::ScaleGradMaker);
-REGISTER_OP_CPU_KERNEL(scale,
-                       ops::ScaleKernel<paddle::platform::CPUPlace, float>,
-                       ops::ScaleKernel<paddle::platform::CPUPlace, double>,
-                       ops::ScaleKernel<paddle::platform::CPUPlace, int>,
-                       ops::ScaleKernel<paddle::platform::CPUPlace, int64_t>);
+REGISTER_OP_CPU_KERNEL(
+    scale, ops::ScaleKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::ScaleKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::ScaleKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::ScaleKernel<paddle::platform::CPUDeviceContext, int64_t>);
diff --git a/paddle/operators/scale_op.cu b/paddle/operators/scale_op.cu
index 0d70775159..0c7980430f 100644
--- a/paddle/operators/scale_op.cu
+++ b/paddle/operators/scale_op.cu
@@ -14,8 +14,10 @@
 
 #include "paddle/operators/scale_op.h"
 
-REGISTER_OP_GPU_KERNEL(
-    scale, paddle::operators::ScaleKernel<paddle::platform::GPUPlace, float>,
-    paddle::operators::ScaleKernel<paddle::platform::GPUPlace, double>,
-    paddle::operators::ScaleKernel<paddle::platform::GPUPlace, int>,
-    paddle::operators::ScaleKernel<paddle::platform::GPUPlace, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    scale,
+    paddle::operators::ScaleKernel<paddle::platform::CUDADeviceContext, float>,
+    paddle::operators::ScaleKernel<paddle::platform::CUDADeviceContext, double>,
+    paddle::operators::ScaleKernel<paddle::platform::CUDADeviceContext, int>,
+    paddle::operators::ScaleKernel<paddle::platform::CUDADeviceContext,
+                                   int64_t>);
diff --git a/paddle/operators/scale_op.h b/paddle/operators/scale_op.h
index 4931294c9d..02a8c97a83 100644
--- a/paddle/operators/scale_op.h
+++ b/paddle/operators/scale_op.h
@@ -19,7 +19,7 @@
 
 namespace paddle {
 namespace operators {
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class ScaleKernel : public framework::OpKernel<T> {
  public:
   virtual void Compute(const framework::ExecutionContext& context) const {
@@ -31,7 +31,8 @@ class ScaleKernel : public framework::OpKernel<T> {
 
     auto eigen_out = framework::EigenVector<T>::Flatten(*tensor);
     auto eigen_in = framework::EigenVector<T>::Flatten(*in);
-    auto& dev = context.GetEigenDevice<Place>();
+    auto& dev =
+        *context.template device_context<DeviceContext>().eigen_device();
     eigen_out.device(dev) = scale * eigen_in;
   }
 };
diff --git a/paddle/operators/scatter_op.cu b/paddle/operators/scatter_op.cu
index 3b32ae2fb7..6b43a1389f 100644
--- a/paddle/operators/scatter_op.cu
+++ b/paddle/operators/scatter_op.cu
@@ -59,5 +59,5 @@ class ScatterGradOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(scatter, ops::ScatterOpCUDAKernel<float>);
-REGISTER_OP_GPU_KERNEL(scatter_grad, ops::ScatterGradOpCUDAKernel<float>);
+REGISTER_OP_CUDA_KERNEL(scatter, ops::ScatterOpCUDAKernel<float>);
+REGISTER_OP_CUDA_KERNEL(scatter_grad, ops::ScatterGradOpCUDAKernel<float>);
diff --git a/paddle/operators/seq_expand_op.cc b/paddle/operators/seq_expand_op.cc
index b862056ad4..ede9754697 100644
--- a/paddle/operators/seq_expand_op.cc
+++ b/paddle/operators/seq_expand_op.cc
@@ -148,8 +148,9 @@ class SeqExpandOpGrad : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(seq_expand, ops::SeqExpandOp, ops::SeqExpandOpMaker,
             seq_expand_grad, ops::SeqExpandOpGrad);
-REGISTER_OP_CPU_KERNEL(seq_expand,
-                       ops::SeqExpandKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    seq_expand,
+    ops::SeqExpandKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     seq_expand_grad,
-    ops::SeqExpandGradKernel<paddle::platform::CPUPlace, float>);
+    ops::SeqExpandGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/seq_expand_op.cu b/paddle/operators/seq_expand_op.cu
index f1e4b82a76..8e67ce9ccb 100644
--- a/paddle/operators/seq_expand_op.cu
+++ b/paddle/operators/seq_expand_op.cu
@@ -16,8 +16,9 @@
 #include "paddle/operators/seq_expand_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(seq_expand,
-                       ops::SeqExpandKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    seq_expand,
+    ops::SeqExpandKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     seq_expand_grad,
-    ops::SeqExpandGradKernel<paddle::platform::GPUPlace, float>);
+    ops::SeqExpandGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/seq_expand_op.h b/paddle/operators/seq_expand_op.h
index 4ef0d02cf8..fbee0db454 100644
--- a/paddle/operators/seq_expand_op.h
+++ b/paddle/operators/seq_expand_op.h
@@ -23,7 +23,7 @@ namespace operators {
 
 using LoDTensor = framework::LoDTensor;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SeqExpandKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -37,7 +37,8 @@ class SeqExpandKernel : public framework::OpKernel<T> {
                       "The size of last lod level in Input(Y)"
                       "must be equal to dims[0] of Input(X).");
     out->set_lod(y->lod());
-    auto place = context.GetEigenDevice<Place>();
+    auto* place =
+        context.template device_context<DeviceContext>().eigen_device();
     size_t element_len = framework::product(x_dims) / x_dims[0];
     T* out_data = out->mutable_data<T>(context.GetPlace());
     auto out_starts = out->lod().back();
@@ -50,7 +51,7 @@ class SeqExpandKernel : public framework::OpKernel<T> {
       Eigen::TensorMap<Eigen::Tensor<T, 2, Eigen::RowMajor, Eigen::DenseIndex>>
           out_t(out_data, scale, element_len);
       Eigen::array<int, 2> cast({{scale, 1}});
-      out_t.device(place) = x_t.broadcast(cast);
+      out_t.device(*place) = x_t.broadcast(cast);
       x_data += element_len;
       out_data += element_len * scale;
     }
@@ -69,7 +70,7 @@ class SeqExpandKernel : public framework::OpKernel<T> {
  *    Grad(X).lod = Input(X).lod
  *
  * */
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SeqExpandGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -89,8 +90,9 @@ class SeqExpandGradKernel : public framework::OpKernel<T> {
       d_out_t(d_out_data, static_cast<int>(repeat), element_len);
       Eigen::TensorMap<Eigen::Tensor<T, 1, Eigen::RowMajor, Eigen::DenseIndex>>
       d_x_t(d_x_data, static_cast<int>(element_len));
-      auto place = context.GetEigenDevice<Place>();
-      d_x_t.device(place) = d_out_t.sum(Eigen::array<int, 1>({{0}}));
+      auto place =
+          context.template device_context<DeviceContext>().eigen_device();
+      d_x_t.device(*place) = d_out_t.sum(Eigen::array<int, 1>({{0}}));
       d_out_data += (repeat * element_len);
       d_x_data += element_len;
     }
diff --git a/paddle/operators/sequence_concat_op.cc b/paddle/operators/sequence_concat_op.cc
index d1de0b4447..9c7e5456e8 100644
--- a/paddle/operators/sequence_concat_op.cc
+++ b/paddle/operators/sequence_concat_op.cc
@@ -129,7 +129,7 @@ REGISTER_OP(sequence_concat, ops::SequenceConcatOp, ops::SequenceConcatOpMaker,
             sequence_concat_grad, ops::SequenceConcatGradOp);
 REGISTER_OP_CPU_KERNEL(
     sequence_concat,
-    ops::SequenceConcatOpKernel<paddle::platform::CPUPlace, float>);
+    ops::SequenceConcatOpKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     sequence_concat_grad,
-    ops::SequenceConcatGradOpKernel<paddle::platform::CPUPlace, float>);
+    ops::SequenceConcatGradOpKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/sequence_concat_op.cu.cc b/paddle/operators/sequence_concat_op.cu.cc
index 9ca99c2258..144bdb5af6 100644
--- a/paddle/operators/sequence_concat_op.cu.cc
+++ b/paddle/operators/sequence_concat_op.cu.cc
@@ -15,9 +15,9 @@ limitations under the License. */
 #include "paddle/operators/sequence_concat_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     sequence_concat,
-    ops::SequenceConcatOpKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    sequence_concat_grad,
-    ops::SequenceConcatGradOpKernel<paddle::platform::GPUPlace, float>);
+    ops::SequenceConcatOpKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(sequence_concat_grad,
+                        ops::SequenceConcatGradOpKernel<
+                            paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/sequence_concat_op.h b/paddle/operators/sequence_concat_op.h
index 09212070aa..8445224f46 100644
--- a/paddle/operators/sequence_concat_op.h
+++ b/paddle/operators/sequence_concat_op.h
@@ -59,7 +59,7 @@ LoD ConcatLoD(const std::vector<const T*> ins, const size_t level) {
   return out_lod;
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SequenceConcatOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -119,7 +119,7 @@ class SequenceConcatOpKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SequenceConcatGradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
diff --git a/paddle/operators/sequence_conv_op.cc b/paddle/operators/sequence_conv_op.cc
index c5533732d4..f5c4f1c133 100644
--- a/paddle/operators/sequence_conv_op.cc
+++ b/paddle/operators/sequence_conv_op.cc
@@ -179,9 +179,10 @@ REGISTER_OP(sequence_conv, ops::SequenceConvOp, ops::SequenceConvOpMaker,
             sequence_conv_grad, ops::SequenceConvGradOp);
 
 REGISTER_OP_CPU_KERNEL(
-    sequence_conv, ops::SequenceConvKernel<paddle::platform::CPUPlace, float>,
-    ops::SequenceConvKernel<paddle::platform::CPUPlace, double>);
+    sequence_conv,
+    ops::SequenceConvKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SequenceConvKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     sequence_conv_grad,
-    ops::SequenceConvGradKernel<paddle::platform::CPUPlace, float>,
-    ops::SequenceConvGradKernel<paddle::platform::CPUPlace, double>);
+    ops::SequenceConvGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SequenceConvGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/sequence_conv_op.cu.cc b/paddle/operators/sequence_conv_op.cu.cc
index c8136dbcb3..eacba79ace 100644
--- a/paddle/operators/sequence_conv_op.cu.cc
+++ b/paddle/operators/sequence_conv_op.cu.cc
@@ -15,10 +15,11 @@
 #include "paddle/operators/sequence_conv_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    sequence_conv, ops::SequenceConvKernel<paddle::platform::GPUPlace, float>,
-    ops::SequenceConvKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    sequence_conv,
+    ops::SequenceConvKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SequenceConvKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
     sequence_conv_grad,
-    ops::SequenceConvGradKernel<paddle::platform::GPUPlace, float>,
-    ops::SequenceConvGradKernel<paddle::platform::GPUPlace, double>);
+    ops::SequenceConvGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SequenceConvGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/sequence_conv_op.h b/paddle/operators/sequence_conv_op.h
index b8fbe2647c..bb584b7bfa 100644
--- a/paddle/operators/sequence_conv_op.h
+++ b/paddle/operators/sequence_conv_op.h
@@ -23,7 +23,7 @@ namespace operators {
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SequenceConvKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -56,21 +56,23 @@ class SequenceConvKernel : public framework::OpKernel<T> {
     Tensor col;
     col.mutable_data<T>(col_shape, context.GetPlace());
     // Because if padding_trainable is false, padding data should be zeros.
-    math::SetConstant<Place, T> set_zero;
-    set_zero(context.device_context(), &col, static_cast<T>(0));
+    math::SetConstant<DeviceContext, T> set_zero;
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    set_zero(dev_ctx, &col, static_cast<T>(0));
 
-    math::ContextProjectFunctor<Place, T> seq_project_functor;
+    math::ContextProjectFunctor<DeviceContext, T> seq_project_functor;
 
-    seq_project_functor(context.device_context(), *in, *padding_data,
-                        padding_trainable, context_start, context_length,
-                        context_stride, up_pad, down_pad, &col);
+    seq_project_functor(dev_ctx, *in, *padding_data, padding_trainable,
+                        context_start, context_length, context_stride, up_pad,
+                        down_pad, &col);
 
-    math::matmul<Place, T>(context.device_context(), col, false, filter, false,
-                           static_cast<T>(1.0), out, static_cast<T>(0.0));
+    math::matmul<DeviceContext, T>(dev_ctx, col, false, filter, false,
+                                   static_cast<T>(1.0), out,
+                                   static_cast<T>(0.0));
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SequenceConvGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -95,7 +97,8 @@ class SequenceConvGradKernel : public framework::OpKernel<T> {
     int down_pad = std::max(0, context_start + context_length - 1);
     int sequence_width = static_cast<int>(in->dims()[1]);
 
-    math::SetConstant<Place, T> set_zero;
+    math::SetConstant<DeviceContext, T> set_zero;
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     // use col_shape in the im2col calculation
     framework::DDim col_shape = {in->dims()[0],
                                  sequence_width * context_length};
@@ -104,38 +107,36 @@ class SequenceConvGradKernel : public framework::OpKernel<T> {
     if (in_g || filter_g || (padding_trainable && padding_data_g)) {
       col.mutable_data<T>(col_shape, context.GetPlace());
       // Because if padding_trainable is false, padding data should be zeros.
-      set_zero(context.device_context(), &col, static_cast<T>(0));
-      math::matmul<Place, T>(context.device_context(), *out_g, false, *filter,
-                             true, T(1.0), &col, T(1.0));
+      set_zero(dev_ctx, &col, static_cast<T>(0));
+      math::matmul<DeviceContext, T>(dev_ctx, *out_g, false, *filter, true,
+                                     T(1.0), &col, T(1.0));
     }
-    math::ContextProjectFunctor<Place, T> seq_project_functor;
-    math::ContextProjectGradFunctor<Place, T> seq_project_grad_functor;
+    math::ContextProjectFunctor<DeviceContext, T> seq_project_functor;
+    math::ContextProjectGradFunctor<DeviceContext, T> seq_project_grad_functor;
 
     if (in_g) {
       in_g->mutable_data<T>(context.GetPlace());
       in_g->set_lod(in->lod());
-      set_zero(context.device_context(), in_g, static_cast<T>(0));
+      set_zero(dev_ctx, in_g, static_cast<T>(0));
 
-      seq_project_grad_functor(context.device_context(), *in_g,
-                               padding_trainable, context_start, context_length,
-                               context_stride, up_pad, down_pad, false, true,
-                               padding_data_g, &col);
+      seq_project_grad_functor(dev_ctx, *in_g, padding_trainable, context_start,
+                               context_length, context_stride, up_pad, down_pad,
+                               false, true, padding_data_g, &col);
     }
 
     if (padding_trainable && padding_data_g) {
       padding_data_g->mutable_data<T>(context.GetPlace());
-      set_zero(context.device_context(), padding_data_g, static_cast<T>(0));
+      set_zero(dev_ctx, padding_data_g, static_cast<T>(0));
 
       LoDTensor* input = const_cast<LoDTensor*>(in);
-      seq_project_grad_functor(context.device_context(), *input,
-                               padding_trainable, context_start, context_length,
-                               context_stride, up_pad, down_pad, true, false,
-                               padding_data_g, &col);
+      seq_project_grad_functor(
+          dev_ctx, *input, padding_trainable, context_start, context_length,
+          context_stride, up_pad, down_pad, true, false, padding_data_g, &col);
     }
 
     if (filter_g) {
       filter_g->mutable_data<T>(context.GetPlace());
-      set_zero(context.device_context(), filter_g, static_cast<T>(0));
+      set_zero(dev_ctx, filter_g, static_cast<T>(0));
 
       Tensor filter_grad = *filter_g;
       LoDTensor out_grad = *out_g;
@@ -145,12 +146,12 @@ class SequenceConvGradKernel : public framework::OpKernel<T> {
         padding_data = context.Input<Tensor>("PaddingData");
       }
 
-      seq_project_functor(context.device_context(), *in, *padding_data,
-                          padding_trainable, context_start, context_length,
-                          context_stride, up_pad, down_pad, &col);
+      seq_project_functor(dev_ctx, *in, *padding_data, padding_trainable,
+                          context_start, context_length, context_stride, up_pad,
+                          down_pad, &col);
 
-      math::matmul<Place, T>(context.device_context(), col, true, out_grad,
-                             false, T(1.0), &filter_grad, T(1.0));
+      math::matmul<DeviceContext, T>(dev_ctx, col, true, out_grad, false,
+                                     T(1.0), &filter_grad, T(1.0));
     }
   }
 };
diff --git a/paddle/operators/sequence_pool_op.cc b/paddle/operators/sequence_pool_op.cc
index bfda8649cd..3526e45a1b 100644
--- a/paddle/operators/sequence_pool_op.cc
+++ b/paddle/operators/sequence_pool_op.cc
@@ -123,7 +123,8 @@ namespace ops = paddle::operators;
 REGISTER_OP(sequence_pool, ops::SequencePoolOp, ops::SequencePoolOpMaker,
             sequence_pool_grad, ops::SequencePoolGradOp);
 REGISTER_OP_CPU_KERNEL(
-    sequence_pool, ops::SequencePoolKernel<paddle::platform::CPUPlace, float>);
+    sequence_pool,
+    ops::SequencePoolKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     sequence_pool_grad,
-    ops::SequencePoolGradKernel<paddle::platform::CPUPlace, float>);
+    ops::SequencePoolGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/sequence_pool_op.cu b/paddle/operators/sequence_pool_op.cu
index 66850772d5..fcd6508435 100644
--- a/paddle/operators/sequence_pool_op.cu
+++ b/paddle/operators/sequence_pool_op.cu
@@ -17,8 +17,9 @@
 #include "paddle/operators/sequence_pool_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    sequence_pool, ops::SequencePoolKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    sequence_pool,
+    ops::SequencePoolKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     sequence_pool_grad,
-    ops::SequencePoolGradKernel<paddle::platform::GPUPlace, float>);
+    ops::SequencePoolGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/sequence_pool_op.h b/paddle/operators/sequence_pool_op.h
index 7f136d8cf0..7519aa1d72 100644
--- a/paddle/operators/sequence_pool_op.h
+++ b/paddle/operators/sequence_pool_op.h
@@ -30,7 +30,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SequencePoolKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -54,17 +54,18 @@ class SequencePoolKernel : public framework::OpKernel<T> {
     auto lod_level_0 = lod[0];
 
     out->mutable_data<T>(context.GetPlace());
-
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     if (pooltype == "MAX") {
-      math::MaxSeqPoolFunctor<Place, T> max_pool;
+      math::MaxSeqPoolFunctor<DeviceContext, T> max_pool;
       auto* index = context.Output<Tensor>("MaxIndex");
       index->Resize({dims});
       index->mutable_data<int>(context.GetPlace());
-      max_pool(context.device_context(), *in, out, index);
+      max_pool(dev_ctx, *in, out, index);
       return;
     }
 
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
     for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
       Tensor in_t = in->Slice(static_cast<int>(lod_level_0[i]),
                               static_cast<int>(lod_level_0[i + 1]));
@@ -91,7 +92,7 @@ class SequencePoolKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SequencePoolGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -105,20 +106,23 @@ class SequencePoolGradKernel : public framework::OpKernel<T> {
     int64_t w = in->numel() / dims[0];
 
     in_g->mutable_data<T>(context.GetPlace());
+    auto& dev_ctx = context.template device_context<DeviceContext>();
 
     if (pooltype == "MAX") {
-      math::MaxSeqPoolGradFunctor<Place, T> max_pool_grad;
+      math::MaxSeqPoolGradFunctor<DeviceContext, T> max_pool_grad;
       auto* index = context.Input<Tensor>("MaxIndex");
-      max_pool_grad(context.device_context(), *out_g, *index, in_g);
+      max_pool_grad(dev_ctx, *out_g, *index, in_g);
       return;
     }
 
     if (pooltype == "LAST" || pooltype == "FIRST") {
       // set X@Grad be zero at first when pooltype is LAST/FIRST
-      math::SetConstant<Place, T> functor;
-      functor(context.device_context(), in_g, 0);
+      math::SetConstant<DeviceContext, T> functor;
+      functor(dev_ctx, in_g, 0);
     }
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
+
     for (int i = 0; i < static_cast<int>(lod.size()) - 1; ++i) {
       auto in_g_t =
           in_g->Slice(static_cast<int>(lod[i]), static_cast<int>(lod[i + 1]));
diff --git a/paddle/operators/sequence_slice_op.cc b/paddle/operators/sequence_slice_op.cc
index 255683a572..481db8f9e5 100644
--- a/paddle/operators/sequence_slice_op.cc
+++ b/paddle/operators/sequence_slice_op.cc
@@ -125,7 +125,7 @@ REGISTER_OP(sequence_slice, ops::SequenceSliceOp, ops::SequenceSliceOpMaker,
             sequence_slice_grad, ops::SequenceSliceGradOp);
 REGISTER_OP_CPU_KERNEL(
     sequence_slice,
-    ops::SequenceSliceOpKernel<paddle::platform::CPUPlace, float>);
+    ops::SequenceSliceOpKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     sequence_slice_grad,
-    ops::SequenceSliceGradOpKernel<paddle::platform::CPUPlace, float>);
+    ops::SequenceSliceGradOpKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/sequence_slice_op.cu b/paddle/operators/sequence_slice_op.cu
index a9f59dadba..43a21d619f 100755
--- a/paddle/operators/sequence_slice_op.cu
+++ b/paddle/operators/sequence_slice_op.cu
@@ -15,9 +15,9 @@ limitations under the License. */
 #include "paddle/operators/sequence_slice_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     sequence_slice,
-    ops::SequenceSliceOpKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+    ops::SequenceSliceOpKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     sequence_slice_grad,
-    ops::SequenceSliceGradOpKernel<paddle::platform::GPUPlace, float>);
+    ops::SequenceSliceGradOpKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/sequence_slice_op.h b/paddle/operators/sequence_slice_op.h
index 428ef556da..14bcaebbb4 100644
--- a/paddle/operators/sequence_slice_op.h
+++ b/paddle/operators/sequence_slice_op.h
@@ -39,7 +39,7 @@ inline LoD SequenceSliceLoD(const T& in, const int64_t* offset_data,
   return out_lod;
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SequenceSliceOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -108,7 +108,7 @@ class SequenceSliceOpKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SequenceSliceGradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -143,8 +143,9 @@ class SequenceSliceGradOpKernel : public framework::OpKernel<T> {
     if (x_grad) {
       x_grad->mutable_data<T>(ctx.GetPlace());
       x_grad->set_lod(in->lod());
-      math::SetConstant<Place, T> set_zero;
-      set_zero(ctx.device_context(), x_grad, static_cast<T>(0));
+      math::SetConstant<DeviceContext, T> set_zero;
+      set_zero(ctx.template device_context<DeviceContext>(), x_grad,
+               static_cast<T>(0));
 
       auto out_grad_stride = framework::stride(out_grad->dims());
 
diff --git a/paddle/operators/sequence_softmax_op.cc b/paddle/operators/sequence_softmax_op.cc
index 32c1502566..37d5452e6b 100644
--- a/paddle/operators/sequence_softmax_op.cc
+++ b/paddle/operators/sequence_softmax_op.cc
@@ -99,7 +99,7 @@ REGISTER_OP(sequence_softmax, ops::SequenceSoftmaxOp,
             ops::SequenceSoftmaxGradOp);
 REGISTER_OP_CPU_KERNEL(
     sequence_softmax,
-    ops::SequenceSoftmaxKernel<paddle::platform::CPUPlace, float>);
+    ops::SequenceSoftmaxKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     sequence_softmax_grad,
-    ops::SequenceSoftmaxGradKernel<paddle::platform::CPUPlace, float>);
+    ops::SequenceSoftmaxGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/sequence_softmax_op.cu.cc b/paddle/operators/sequence_softmax_op.cu.cc
index 7023795a3b..5f65b4daf9 100644
--- a/paddle/operators/sequence_softmax_op.cu.cc
+++ b/paddle/operators/sequence_softmax_op.cu.cc
@@ -15,9 +15,9 @@ limitations under the License. */
 #include "paddle/operators/sequence_softmax_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     sequence_softmax,
-    ops::SequenceSoftmaxKernel<paddle::platform::GPUPlace, float>)
-REGISTER_OP_GPU_KERNEL(
+    ops::SequenceSoftmaxKernel<paddle::platform::CUDADeviceContext, float>)
+REGISTER_OP_CUDA_KERNEL(
     sequence_softmax_grad,
-    ops::SequenceSoftmaxGradKernel<paddle::platform::GPUPlace, float>);
+    ops::SequenceSoftmaxGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/sequence_softmax_op.h b/paddle/operators/sequence_softmax_op.h
index 1b68dd0662..e889e88cb3 100644
--- a/paddle/operators/sequence_softmax_op.h
+++ b/paddle/operators/sequence_softmax_op.h
@@ -23,7 +23,7 @@ namespace operators {
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SequenceSoftmaxKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -52,12 +52,13 @@ class SequenceSoftmaxKernel : public framework::OpKernel<T> {
       framework::DDim dims_i = framework::make_ddim({1UL, end_pos - start_pos});
       x_i.Resize(dims_i);
       out_i.Resize(dims_i);
-      math::SoftmaxFunctor<Place, T>()(ctx.device_context(), &x_i, &out_i);
+      math::SoftmaxFunctor<DeviceContext, T>()(
+          ctx.template device_context<DeviceContext>(), &x_i, &out_i);
     }
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SequenceSoftmaxGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -83,8 +84,9 @@ class SequenceSoftmaxGradKernel : public framework::OpKernel<T> {
       out_i.Resize(dims_i);
       out_grad_i.Resize(dims_i);
       x_grad_i.Resize(dims_i);
-      math::SoftmaxGradFunctor<Place, T>()(ctx.device_context(), &out_i,
-                                           &out_grad_i, &x_grad_i);
+      math::SoftmaxGradFunctor<DeviceContext, T>()(
+          ctx.template device_context<DeviceContext>(), &out_i, &out_grad_i,
+          &x_grad_i);
     }
   }
 };
diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc
index 5576d7b8be..121bf60b27 100644
--- a/paddle/operators/sgd_op.cc
+++ b/paddle/operators/sgd_op.cc
@@ -62,8 +62,8 @@ $$param\_out = param - learning\_rate * grad$$
 };
 
 template <typename T>
-struct SparseSGDFunctor<platform::CPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SparseSGDFunctor<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& context,
                   const framework::SelectedRows& input,
                   const framework::Tensor& learning_rate,
                   framework::Tensor* output) {
@@ -90,13 +90,14 @@ struct SparseSGDFunctor<platform::CPUPlace, T> {
   }
 };
 
-template struct SparseSGDFunctor<platform::CPUPlace, float>;
-template struct SparseSGDFunctor<platform::CPUPlace, double>;
+template struct SparseSGDFunctor<platform::CPUDeviceContext, float>;
+template struct SparseSGDFunctor<platform::CPUDeviceContext, double>;
 
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(sgd, ops::SGDOp, ops::SGDOpMaker);
-REGISTER_OP_CPU_KERNEL(sgd, ops::SGDOpKernel<paddle::platform::CPUPlace, float>,
-                       ops::SGDOpKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(
+    sgd, ops::SGDOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SGDOpKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/sgd_op.cu b/paddle/operators/sgd_op.cu
index 7b6c5ec306..a3c0db7e50 100644
--- a/paddle/operators/sgd_op.cu
+++ b/paddle/operators/sgd_op.cu
@@ -41,8 +41,8 @@ __global__ void SparseSGDFunctorKernel(const T* selected_rows,
 }  // namespace
 
 template <typename T>
-struct SparseSGDFunctor<platform::GPUPlace, T> {
-  void operator()(const platform::DeviceContext& context,
+struct SparseSGDFunctor<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& context,
                   const framework::SelectedRows& input,
                   const framework::Tensor& learning_rate,
                   framework::Tensor* output) {
@@ -62,21 +62,19 @@ struct SparseSGDFunctor<platform::GPUPlace, T> {
     const int block_size = 256;
     dim3 threads(block_size, 1);
     dim3 grid(1, in_rows.size());
-    SparseSGDFunctorKernel<
-        T, 256><<<grid, threads, 0,
-                  reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                      .stream()>>>(in_data, in_rows.data(),
-                                   learning_rate.data<T>(), out_data,
-                                   in_row_numel);
+    SparseSGDFunctorKernel<T, 256><<<grid, threads, 0, context.stream()>>>(
+        in_data, in_rows.data(), learning_rate.data<T>(), out_data,
+        in_row_numel);
   }
 };
 
-template struct SparseSGDFunctor<platform::GPUPlace, float>;
-template struct SparseSGDFunctor<platform::GPUPlace, double>;
+template struct SparseSGDFunctor<platform::CUDADeviceContext, float>;
+template struct SparseSGDFunctor<platform::CUDADeviceContext, double>;
 
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel<paddle::platform::GPUPlace, float>,
-                       ops::SGDOpKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    sgd, ops::SGDOpKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SGDOpKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h
index 78b595fc6c..c920025a91 100644
--- a/paddle/operators/sgd_op.h
+++ b/paddle/operators/sgd_op.h
@@ -20,15 +20,15 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 struct SparseSGDFunctor {
-  void operator()(const platform::DeviceContext& context,
+  void operator()(const DeviceContext& context,
                   const framework::SelectedRows& input,
                   const framework::Tensor& learning_rate,
                   framework::Tensor* output);
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SGDOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
@@ -46,7 +46,8 @@ class SGDOpKernel : public framework::OpKernel<T> {
       auto g = framework::EigenVector<T>::Flatten(*grad);
       auto o = framework::EigenVector<T>::Flatten(*param_out);
       auto lr = framework::EigenVector<T>::Flatten(*learning_rate);
-      auto place = ctx.GetEigenDevice<Place>();
+      auto& place =
+          *ctx.template device_context<DeviceContext>().eigen_device();
 
       Eigen::DSizes<int, 1> grad_dsize(grad->numel());
       o.device(place) = p - lr.broadcast(grad_dsize) * g;
@@ -56,8 +57,9 @@ class SGDOpKernel : public framework::OpKernel<T> {
       // It's better to find a more elegant solution.
       PADDLE_ENFORCE_EQ(param, param_out);
       auto* grad = ctx.Input<framework::SelectedRows>("Grad");
-      SparseSGDFunctor<Place, T> functor;
-      functor(ctx.device_context(), *grad, *learning_rate, param_out);
+      SparseSGDFunctor<DeviceContext, T> functor;
+      functor(ctx.template device_context<DeviceContext>(), *grad,
+              *learning_rate, param_out);
     } else {
       PADDLE_THROW("Unsupported Variable Type of Grad");
     }
diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc
index 782f4c7936..b8a1bf122a 100644
--- a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc
+++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc
@@ -142,7 +142,7 @@ REGISTER_OP(sigmoid_cross_entropy_with_logits,
             ops::SigmoidCrossEntropyWithLogitsGradOp);
 REGISTER_OP_CPU_KERNEL(sigmoid_cross_entropy_with_logits,
                        ops::SigmoidCrossEntropyWithLogitsKernel<
-                           paddle::platform::CPUPlace, float>);
+                           paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(sigmoid_cross_entropy_with_logits_grad,
                        ops::SigmoidCrossEntropyWithLogitsGradKernel<
-                           paddle::platform::CPUPlace, float>);
+                           paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu
index 32a39956a1..1b569c93ed 100644
--- a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu
+++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu
@@ -16,9 +16,9 @@
 #include "paddle/operators/sigmoid_cross_entropy_with_logits_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(sigmoid_cross_entropy_with_logits,
-                       ops::SigmoidCrossEntropyWithLogitsKernel<
-                           paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(sigmoid_cross_entropy_with_logits_grad,
-                       ops::SigmoidCrossEntropyWithLogitsGradKernel<
-                           paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(sigmoid_cross_entropy_with_logits,
+                        ops::SigmoidCrossEntropyWithLogitsKernel<
+                            paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(sigmoid_cross_entropy_with_logits_grad,
+                        ops::SigmoidCrossEntropyWithLogitsGradKernel<
+                            paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.h b/paddle/operators/sigmoid_cross_entropy_with_logits_op.h
index 2a9d9bbc77..8fe7c5ba82 100644
--- a/paddle/operators/sigmoid_cross_entropy_with_logits_op.h
+++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.h
@@ -20,7 +20,7 @@ namespace paddle {
 namespace operators {
 
 // Out = max(X, 0) - X * Labels + log(1 + exp(-abs(X)))
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SigmoidCrossEntropyWithLogitsKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
@@ -32,7 +32,7 @@ class SigmoidCrossEntropyWithLogitsKernel : public framework::OpKernel<T> {
     auto x = framework::EigenVector<T>::Flatten(*X);
     auto labels = framework::EigenVector<T>::Flatten(*Labels);
     auto out = framework::EigenVector<T>::Flatten(*Out);
-    auto place = context.GetEigenDevice<Place>();
+    auto &place = *context.device_context<DeviceContext>().eigen_device();
 
     // term1 = max(x, 0)
     auto term1 = x.cwiseMax(static_cast<T>(0));
@@ -46,7 +46,7 @@ class SigmoidCrossEntropyWithLogitsKernel : public framework::OpKernel<T> {
 };
 
 // dX = sigmoid(X) - labels
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SigmoidCrossEntropyWithLogitsGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
@@ -62,7 +62,8 @@ class SigmoidCrossEntropyWithLogitsGradKernel : public framework::OpKernel<T> {
     auto labels = framework::EigenVector<T>::Flatten(*Labels);
     auto dout = framework::EigenVector<T>::Flatten(*dOut);
     auto dx = framework::EigenVector<T>::Flatten(*dX);
-    auto place = context.GetEigenDevice<Place>();
+    auto &place =
+        *context.template device_context<DeviceContext>().eigen_device();
 
     auto sigmoid_x = static_cast<T>(1) / (static_cast<T>(1) + (-x).exp());
     dx.device(place) = dout * (sigmoid_x - labels);
diff --git a/paddle/operators/sign_op.cc b/paddle/operators/sign_op.cc
index 08bf2e4e7c..d5a7ccb77e 100644
--- a/paddle/operators/sign_op.cc
+++ b/paddle/operators/sign_op.cc
@@ -67,5 +67,5 @@ namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(sign, ops::SignOp, ops::SignOpMaker<float>,
                   ops::SignGradMaker);
-REGISTER_OP_CPU_KERNEL(sign,
-                       ops::SignKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    sign, ops::SignKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/sign_op.cu b/paddle/operators/sign_op.cu
index 4d0638cb97..9bc1c65d21 100644
--- a/paddle/operators/sign_op.cu
+++ b/paddle/operators/sign_op.cu
@@ -14,5 +14,6 @@
 
 #include "paddle/operators/sign_op.h"
 
-REGISTER_OP_GPU_KERNEL(
-    sign, paddle::operators::SignKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    sign,
+    paddle::operators::SignKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/sign_op.h b/paddle/operators/sign_op.h
index ab5cd4bac0..2e476ed665 100644
--- a/paddle/operators/sign_op.h
+++ b/paddle/operators/sign_op.h
@@ -19,7 +19,7 @@
 
 namespace paddle {
 namespace operators {
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SignKernel : public framework::OpKernel<T> {
  public:
   virtual void Compute(const framework::ExecutionContext& context) const {
@@ -29,7 +29,8 @@ class SignKernel : public framework::OpKernel<T> {
 
     auto eigen_out = framework::EigenVector<T>::Flatten(*out);
     auto eigen_in = framework::EigenVector<T>::Flatten(*in);
-    auto& place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
     eigen_out.device(place) = eigen_in.sign();
   }
 };
diff --git a/paddle/operators/smooth_l1_loss_op.cc b/paddle/operators/smooth_l1_loss_op.cc
index 50543fcc14..56e8d9058f 100644
--- a/paddle/operators/smooth_l1_loss_op.cc
+++ b/paddle/operators/smooth_l1_loss_op.cc
@@ -138,7 +138,8 @@ REGISTER_OP(smooth_l1_loss, ops::SmoothL1LossOp,
             ops::SmoothL1LossOpMaker<float>, smooth_l1_loss_grad,
             ops::SmoothL1LossGradOp);
 REGISTER_OP_CPU_KERNEL(
-    smooth_l1_loss, ops::SmoothL1LossKernel<paddle::platform::CPUPlace, float>);
+    smooth_l1_loss,
+    ops::SmoothL1LossKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     smooth_l1_loss_grad,
-    ops::SmoothL1LossGradKernel<paddle::platform::CPUPlace, float>);
+    ops::SmoothL1LossGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/smooth_l1_loss_op.cu b/paddle/operators/smooth_l1_loss_op.cu
index 1c3172f438..8e94ebac64 100644
--- a/paddle/operators/smooth_l1_loss_op.cu
+++ b/paddle/operators/smooth_l1_loss_op.cu
@@ -17,8 +17,9 @@
 #include "paddle/operators/smooth_l1_loss_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    smooth_l1_loss, ops::SmoothL1LossKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    smooth_l1_loss,
+    ops::SmoothL1LossKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     smooth_l1_loss_grad,
-    ops::SmoothL1LossGradKernel<paddle::platform::GPUPlace, float>);
+    ops::SmoothL1LossGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/smooth_l1_loss_op.h b/paddle/operators/smooth_l1_loss_op.h
index 39d0070b6c..1a70c9c63c 100644
--- a/paddle/operators/smooth_l1_loss_op.h
+++ b/paddle/operators/smooth_l1_loss_op.h
@@ -44,7 +44,7 @@ struct SmoothL1LossForward {
   T sigma2;
 };
 
-template <typename Place, typename T, typename AttrType = T>
+template <typename DeviceContext, typename T, typename AttrType = T>
 class SmoothL1LossKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -57,7 +57,8 @@ class SmoothL1LossKernel : public framework::OpKernel<T> {
 
     out0->mutable_data<T>(context.GetPlace());
     out1->mutable_data<T>(context.GetPlace());
-    auto place = context.GetEigenDevice<Place>();
+    auto* place =
+        context.template device_context<DeviceContext>().eigen_device();
 
     auto sigma = static_cast<T>(context.Attr<AttrType>("sigma"));
     T sigma2 = sigma * sigma;
@@ -67,12 +68,12 @@ class SmoothL1LossKernel : public framework::OpKernel<T> {
     auto y = EigenVector<T>::Flatten(*in1);
     auto diff = EigenVector<T>::Flatten(*out0);
 
-    diff.device(place) = x - y;
+    diff.device(*place) = x - y;
     // multiply inside weight
     if (has_weight) {
       auto inside_weight = EigenVector<T>::Flatten(*in2);
       // cache diff, reused in bp
-      diff.device(place) = diff * inside_weight;
+      diff.device(*place) = diff * inside_weight;
     }
 
     auto in_counts = in0->numel();
@@ -81,12 +82,12 @@ class SmoothL1LossKernel : public framework::OpKernel<T> {
                                    context.GetPlace());
     auto errors = EigenVector<T>::Flatten(ptensor_errors);
     // apply smooth l1 forward
-    errors.device(place) = diff.unaryExpr(SmoothL1LossForward<T>(sigma2));
+    errors.device(*place) = diff.unaryExpr(SmoothL1LossForward<T>(sigma2));
 
     // multiply outside weight
     if (has_weight) {
       auto outside_weight = EigenVector<T>::Flatten(*in3);
-      errors.device(place) = errors * outside_weight;
+      errors.device(*place) = errors * outside_weight;
     }
     auto loss = EigenVector<T>::Flatten(*out1);
     // first dimension of 'X' is the number of samples
@@ -94,7 +95,7 @@ class SmoothL1LossKernel : public framework::OpKernel<T> {
         framework::make_ddim({static_cast<int>(in0->dims()[0]),
                               static_cast<int>(in_counts / in0->dims()[0])});
     auto errors_mat_view = EigenMatrix<T>::From(ptensor_errors, mat_dims);
-    loss.device(place) = errors_mat_view.sum(Eigen::array<int, 1>({{1}}));
+    loss.device(*place) = errors_mat_view.sum(Eigen::array<int, 1>({{1}}));
   }
 };
 
@@ -114,7 +115,7 @@ struct SmoothL1LossBackward {
   T sigma2;
 };
 
-template <typename Place, typename T, typename AttrType = T>
+template <typename DeviceContext, typename T, typename AttrType = T>
 class SmoothL1LossGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -126,7 +127,8 @@ class SmoothL1LossGradKernel : public framework::OpKernel<T> {
     T sigma2 = sigma * sigma;
     bool has_weight = (in0 != nullptr) && (in1 != nullptr);
 
-    auto place = context.GetEigenDevice<Place>();
+    auto* place =
+        context.template device_context<DeviceContext>().eigen_device();
 
     auto in_dims = in2->dims();
     auto counts = in2->numel();
@@ -139,7 +141,7 @@ class SmoothL1LossGradKernel : public framework::OpKernel<T> {
                                  context.GetPlace());
     auto diff = EigenVector<T>::Flatten(ptensor_diff);
     // apply smooth l1 backwoard
-    diff.device(place) = EigenVector<T>::Flatten(*in2).unaryExpr(
+    diff.device(*place) = EigenVector<T>::Flatten(*in2).unaryExpr(
         SmoothL1LossBackward<T>(sigma2));
 
     // compute weights
@@ -147,11 +149,11 @@ class SmoothL1LossGradKernel : public framework::OpKernel<T> {
     ptensor_weights.mutable_data<T>(mat_dims, context.GetPlace());
     auto weights = EigenMatrix<T>::From(ptensor_weights);
     // initialize to 1.0
-    weights.device(place) = weights.constant(static_cast<T>(1.0));
+    weights.device(*place) = weights.constant(static_cast<T>(1.0));
     if (has_weight) {
       auto inside_weight = EigenMatrix<T>::From(*in0, mat_dims);
       auto outside_weight = EigenMatrix<T>::From(*in1, mat_dims);
-      weights.device(place) = inside_weight * outside_weight;
+      weights.device(*place) = inside_weight * outside_weight;
     }
 
     // compute gradients
@@ -167,13 +169,13 @@ class SmoothL1LossGradKernel : public framework::OpKernel<T> {
     if (out0) {
       out0->mutable_data<T>(context.GetPlace());
       auto x_grad = EigenMatrix<T>::From(*out0, mat_dims);
-      x_grad.device(place) = gradients;
+      x_grad.device(*place) = gradients;
     }
 
     if (out1) {
       out1->mutable_data<T>(context.GetPlace());
       auto y_grad = EigenMatrix<T>::From(*out1, mat_dims);
-      y_grad.device(place) = -1 * gradients;
+      y_grad.device(*place) = -1 * gradients;
     }
   }
 };
diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc
index 93e0525bad..0988c83d43 100644
--- a/paddle/operators/softmax_op.cc
+++ b/paddle/operators/softmax_op.cc
@@ -89,7 +89,8 @@ namespace ops = paddle::operators;
 
 REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, softmax_grad,
             ops::SoftmaxOpGrad);
-REGISTER_OP_CPU_KERNEL(softmax,
-                       ops::SoftmaxKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
-    softmax_grad, ops::SoftmaxGradKernel<paddle::platform::CPUPlace, float>);
+    softmax, ops::SoftmaxKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    softmax_grad,
+    ops::SoftmaxGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/softmax_op.cu.cc b/paddle/operators/softmax_op.cu.cc
index 013ace19ae..7b9882cbcf 100644
--- a/paddle/operators/softmax_op.cu.cc
+++ b/paddle/operators/softmax_op.cu.cc
@@ -16,7 +16,8 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(softmax,
-                       ops::SoftmaxKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    softmax_grad, ops::SoftmaxGradKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    softmax, ops::SoftmaxKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    softmax_grad,
+    ops::SoftmaxGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h
index 44d1e63f1b..0f8998b99e 100644
--- a/paddle/operators/softmax_op.h
+++ b/paddle/operators/softmax_op.h
@@ -21,7 +21,7 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SoftmaxKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -31,11 +31,12 @@ class SoftmaxKernel : public framework::OpKernel<T> {
     // allocate memory on device.
     Y->mutable_data<T>(context.GetPlace());
 
-    math::SoftmaxFunctor<Place, T>()(context.device_context(), X, Y);
+    math::SoftmaxFunctor<DeviceContext, T>()(
+        context.template device_context<DeviceContext>(), X, Y);
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SoftmaxGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -46,7 +47,8 @@ class SoftmaxGradKernel : public framework::OpKernel<T> {
     // allocate memory on device.
     dX->mutable_data<T>(context.GetPlace());
 
-    math::SoftmaxGradFunctor<Place, T>()(context.device_context(), Y, dY, dX);
+    math::SoftmaxGradFunctor<DeviceContext, T>()(
+        context.template device_context<DeviceContext>(), Y, dY, dX);
   }
 };
 
diff --git a/paddle/operators/softmax_with_cross_entropy_op.cu b/paddle/operators/softmax_with_cross_entropy_op.cu
index b1faddac3f..6100c63f9a 100644
--- a/paddle/operators/softmax_with_cross_entropy_op.cu
+++ b/paddle/operators/softmax_with_cross_entropy_op.cu
@@ -69,10 +69,10 @@ class SoftmaxWithCrossEntropyCUDAKernel : public framework::OpKernel<T> {
     softmax->mutable_data<T>(context.GetPlace());
     loss->mutable_data<T>(context.GetPlace());
 
-    math::SoftmaxFunctor<platform::GPUPlace, T>()(context.device_context(),
-                                                  logits, softmax);
-    math::CrossEntropyFunctor<platform::GPUPlace, T>()(
-        context.device_context(), loss, softmax, labels,
+    math::SoftmaxFunctor<platform::CUDADeviceContext, T>()(
+        context.cuda_device_context(), logits, softmax);
+    math::CrossEntropyFunctor<platform::CUDADeviceContext, T>()(
+        context.cuda_device_context(), loss, softmax, labels,
         context.Attr<bool>("soft_label"));
   }
 };
@@ -98,18 +98,18 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel<T> {
 
     if (context.Attr<bool>("soft_label")) {
       const T* label_data = labels->data<T>();
-      SoftCrossEntropyGradientKernel<T><<<
-          grid, block, 0, reinterpret_cast<const platform::CUDADeviceContext&>(
-                              context.device_context())
-                              .stream()>>>(logit_grad_data, loss_grad_data,
-                                           label_data, batch_size, class_num);
+      SoftCrossEntropyGradientKernel<
+          T><<<grid, block, 0,
+               context.template device_context<platform::CUDADeviceContext>()
+                   .stream()>>>(logit_grad_data, loss_grad_data, label_data,
+                                batch_size, class_num);
     } else {
       const int64_t* label_data = labels->data<int64_t>();
-      CrossEntropyGrad<T><<<
-          grid, block, 0, reinterpret_cast<const platform::CUDADeviceContext&>(
-                              context.device_context())
-                              .stream()>>>(logit_grad_data, loss_grad_data,
-                                           label_data, batch_size, class_num);
+      CrossEntropyGrad<
+          T><<<grid, block, 0,
+               context.template device_context<platform::CUDADeviceContext>()
+                   .stream()>>>(logit_grad_data, loss_grad_data, label_data,
+                                batch_size, class_num);
     }
   }
 };
@@ -118,9 +118,9 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(softmax_with_cross_entropy,
-                       ops::SoftmaxWithCrossEntropyCUDAKernel<float>,
-                       ops::SoftmaxWithCrossEntropyCUDAKernel<double>);
-REGISTER_OP_GPU_KERNEL(softmax_with_cross_entropy_grad,
-                       ops::SoftmaxWithCrossEntropyGradCUDAKernel<float>,
-                       ops::SoftmaxWithCrossEntropyGradCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(softmax_with_cross_entropy,
+                        ops::SoftmaxWithCrossEntropyCUDAKernel<float>,
+                        ops::SoftmaxWithCrossEntropyCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(softmax_with_cross_entropy_grad,
+                        ops::SoftmaxWithCrossEntropyGradCUDAKernel<float>,
+                        ops::SoftmaxWithCrossEntropyGradCUDAKernel<double>);
diff --git a/paddle/operators/softmax_with_cross_entropy_op.h b/paddle/operators/softmax_with_cross_entropy_op.h
index c4ab3f74b4..9c3431605b 100644
--- a/paddle/operators/softmax_with_cross_entropy_op.h
+++ b/paddle/operators/softmax_with_cross_entropy_op.h
@@ -40,11 +40,12 @@ class SoftmaxWithCrossEntropyKernel : public framework::OpKernel<T> {
     softmax->mutable_data<T>(context.GetPlace());
     loss->mutable_data<T>(context.GetPlace());
 
-    math::SoftmaxFunctor<platform::CPUPlace, T>()(context.device_context(),
-                                                  logits, softmax);
-    math::CrossEntropyFunctor<platform::CPUPlace, T>()(
-        context.device_context(), loss, softmax, labels,
-        context.Attr<bool>("soft_label"));
+    auto& dev_ctx =
+        context.template device_context<platform::CPUDeviceContext>();
+    math::SoftmaxFunctor<platform::CPUDeviceContext, T>()(dev_ctx, logits,
+                                                          softmax);
+    math::CrossEntropyFunctor<platform::CPUDeviceContext, T>()(
+        dev_ctx, loss, softmax, labels, context.Attr<bool>("soft_label"));
   }
 };
 
@@ -62,14 +63,15 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
     const int class_num = logit_grad->dims()[1];
     auto out_grad_mat = EigenMatrix<T>::From(*out_grad);
     auto logit_grad_mat = EigenMatrix<T>::From(*logit_grad);
-
+    auto& place = *context.template device_context<platform::CPUDeviceContext>()
+                       .eigen_device();
     if (context.Attr<bool>("soft_label")) {
       auto lbl_mat = EigenMatrix<T>::From(*labels);
-      logit_grad_mat.device(context.GetEigenDevice<platform::CPUPlace>()) =
+      logit_grad_mat.device(place) =
           out_grad_mat.broadcast(Eigen::DSizes<int, 2>(1, class_num)) *
           (logit_grad_mat - lbl_mat);
     } else {
-      logit_grad_mat.device(context.GetEigenDevice<platform::CPUPlace>()) =
+      logit_grad_mat.device(place) =
           logit_grad_mat *
           out_grad_mat.broadcast(Eigen::DSizes<int, 2>(1, class_num));
 
diff --git a/paddle/operators/split_op.cu.cc b/paddle/operators/split_op.cu.cc
index 93d1fc3c44..dbad0bbf68 100644
--- a/paddle/operators/split_op.cu.cc
+++ b/paddle/operators/split_op.cu.cc
@@ -14,5 +14,5 @@ limitations under the License. */
 
 #include "paddle/operators/split_op.h"
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(split,
-                       ops::SplitOpKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_CUDA_KERNEL(
+    split, ops::SplitOpKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/split_op.h b/paddle/operators/split_op.h
index fa26e5f677..a38c435d53 100644
--- a/paddle/operators/split_op.h
+++ b/paddle/operators/split_op.h
@@ -21,7 +21,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SplitOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc
index bec2a2c18a..50bc6da196 100644
--- a/paddle/operators/squared_l2_distance_op.cc
+++ b/paddle/operators/squared_l2_distance_op.cc
@@ -115,7 +115,7 @@ REGISTER_OP(squared_l2_distance, ops::SquaredL2DistanceOp,
             ops::SquaredL2DistanceGradOp);
 REGISTER_OP_CPU_KERNEL(
     squared_l2_distance,
-    ops::SquaredL2DistanceKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(
-    squared_l2_distance_grad,
-    ops::SquaredL2DistanceGradKernel<paddle::platform::CPUPlace, float>);
+    ops::SquaredL2DistanceKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(squared_l2_distance_grad,
+                       ops::SquaredL2DistanceGradKernel<
+                           paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/squared_l2_distance_op.cu b/paddle/operators/squared_l2_distance_op.cu
index 3fe62f1a9c..ecc82ed1e4 100644
--- a/paddle/operators/squared_l2_distance_op.cu
+++ b/paddle/operators/squared_l2_distance_op.cu
@@ -17,9 +17,9 @@
 #include "paddle/operators/squared_l2_distance_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     squared_l2_distance,
-    ops::SquaredL2DistanceKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
-    squared_l2_distance_grad,
-    ops::SquaredL2DistanceGradKernel<paddle::platform::GPUPlace, float>);
+    ops::SquaredL2DistanceKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(squared_l2_distance_grad,
+                        ops::SquaredL2DistanceGradKernel<
+                            paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/squared_l2_distance_op.h b/paddle/operators/squared_l2_distance_op.h
index 259ef40296..5bd5f4819a 100644
--- a/paddle/operators/squared_l2_distance_op.h
+++ b/paddle/operators/squared_l2_distance_op.h
@@ -27,7 +27,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SquaredL2DistanceKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -51,7 +51,8 @@ class SquaredL2DistanceKernel : public framework::OpKernel<T> {
     auto sub_result = EigenMatrix<T>::From(*out0);
     auto z = EigenVector<T>::Flatten(*out1);
 
-    auto place = context.GetEigenDevice<Place>();
+    auto& place =
+        *context.template device_context<DeviceContext>().eigen_device();
     auto x_dims = x.dimensions();
     auto y_dims = y.dimensions();
     // buffer the substraction result
@@ -67,7 +68,7 @@ class SquaredL2DistanceKernel : public framework::OpKernel<T> {
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SquaredL2DistanceGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -89,7 +90,8 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel<T> {
                     sub_result;
 
     // propagate back to input
-    auto eigen_place = context.GetEigenDevice<Place>();
+    auto& eigen_place =
+        *context.template device_context<DeviceContext>().eigen_device();
     if (x_g) {
       x_g->mutable_data<T>(context.GetPlace());
       // eigen matrix
diff --git a/paddle/operators/squared_l2_norm_op.cc b/paddle/operators/squared_l2_norm_op.cc
index 3c10e6159f..3cff61a02f 100644
--- a/paddle/operators/squared_l2_norm_op.cc
+++ b/paddle/operators/squared_l2_norm_op.cc
@@ -72,7 +72,7 @@ REGISTER_OP(squared_l2_norm, ops::SquaredL2NormOp, ops::SquaredL2NormOpMaker,
             squared_l2_norm_grad, ops::SquaredL2NormGradOp);
 REGISTER_OP_CPU_KERNEL(
     squared_l2_norm,
-    ops::SquaredL2NormKernel<paddle::platform::CPUPlace, float>);
+    ops::SquaredL2NormKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     squared_l2_norm_grad,
-    ops::SquaredL2NormGradKernel<paddle::platform::CPUPlace, float>);
+    ops::SquaredL2NormGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/squared_l2_norm_op.cu b/paddle/operators/squared_l2_norm_op.cu
index d384e9c28c..2d6567d090 100644
--- a/paddle/operators/squared_l2_norm_op.cu
+++ b/paddle/operators/squared_l2_norm_op.cu
@@ -16,9 +16,9 @@
 #include "paddle/operators/squared_l2_norm_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
     squared_l2_norm,
-    ops::SquaredL2NormKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+    ops::SquaredL2NormKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     squared_l2_norm_grad,
-    ops::SquaredL2NormGradKernel<paddle::platform::GPUPlace, float>);
+    ops::SquaredL2NormGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/squared_l2_norm_op.h b/paddle/operators/squared_l2_norm_op.h
index 48d7b1c2d5..0ced7e7d70 100644
--- a/paddle/operators/squared_l2_norm_op.h
+++ b/paddle/operators/squared_l2_norm_op.h
@@ -20,7 +20,7 @@ namespace paddle {
 namespace operators {
 
 // Out = sum(square(X))
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SquaredL2NormKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
@@ -30,14 +30,15 @@ class SquaredL2NormKernel : public framework::OpKernel<T> {
 
     auto x = framework::EigenVector<T>::Flatten(*X);
     auto out = framework::EigenScalar<T>::From(*Out);
-    auto place = context.GetEigenDevice<Place>();
+    auto *place =
+        context.template device_context<DeviceContext>().eigen_device();
 
-    out.device(place) = x.square().sum();
+    out.device(*place) = x.square().sum();
   }
 };
 
 // dX = X
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SquaredL2NormGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
@@ -53,10 +54,11 @@ class SquaredL2NormGradKernel : public framework::OpKernel<T> {
     auto x = framework::EigenVector<T>::Flatten(*X);
     auto dout = framework::EigenVector<T>::Flatten(*dOut);
     auto dx = framework::EigenVector<T>::Flatten(*dX);
-    auto place = context.GetEigenDevice<Place>();
+    auto *place =
+        context.template device_context<DeviceContext>().eigen_device();
 
     Eigen::DSizes<int, 1> x_dsize(X->numel());
-    dx.device(place) = (dout.broadcast(x_dsize) * x) * static_cast<T>(2.0);
+    dx.device(*place) = (dout.broadcast(x_dsize) * x) * static_cast<T>(2.0);
   }
 };
 
diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc
index 744b2fe3f2..cd52672f78 100644
--- a/paddle/operators/sum_op.cc
+++ b/paddle/operators/sum_op.cc
@@ -195,7 +195,8 @@ namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker,
                   ops::SumOpVarTypeInference);
-REGISTER_OP_CPU_KERNEL(sum, ops::SumKernel<paddle::platform::CPUPlace, float>,
-                       ops::SumKernel<paddle::platform::CPUPlace, double>,
-                       ops::SumKernel<paddle::platform::CPUPlace, int>,
-                       ops::SumKernel<paddle::platform::CPUPlace, int64_t>);
+REGISTER_OP_CPU_KERNEL(
+    sum, ops::SumKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SumKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::SumKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::SumKernel<paddle::platform::CPUDeviceContext, int64_t>);
diff --git a/paddle/operators/sum_op.cu b/paddle/operators/sum_op.cu
index 5c30dd4d47..873155076c 100644
--- a/paddle/operators/sum_op.cu
+++ b/paddle/operators/sum_op.cu
@@ -13,7 +13,8 @@ limitations under the License. */
 #include "paddle/operators/sum_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(sum, ops::SumKernel<paddle::platform::GPUPlace, float>,
-                       ops::SumKernel<paddle::platform::GPUPlace, double>,
-                       ops::SumKernel<paddle::platform::GPUPlace, int>,
-                       ops::SumKernel<paddle::platform::GPUPlace, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    sum, ops::SumKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SumKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::SumKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::SumKernel<paddle::platform::CUDADeviceContext, int64_t>);
diff --git a/paddle/operators/sum_op.h b/paddle/operators/sum_op.h
index ed6c80ce60..eaa36aa1ae 100644
--- a/paddle/operators/sum_op.h
+++ b/paddle/operators/sum_op.h
@@ -26,7 +26,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SumKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
@@ -43,12 +43,14 @@ class SumKernel : public framework::OpKernel<T> {
       auto result = EigenVector<T>::Flatten(*out);
 
       if (!in_place) {
-        math::SetConstant<Place, T> constant_functor;
-        constant_functor(context.device_context(), out, 0.0);
+        math::SetConstant<DeviceContext, T> constant_functor;
+        constant_functor(context.template device_context<DeviceContext>(), out,
+                         0.0);
       }
 
-      math::SelectedRowsAddToTensor<Place, T> functor;
-      auto place = context.GetEigenDevice<Place>();
+      math::SelectedRowsAddToTensor<DeviceContext, T> functor;
+      auto &place =
+          *context.template device_context<DeviceContext>().eigen_device();
       // If in_place, just skip the first tensor
       for (int i = in_place ? 1 : 0; i < N; i++) {
         if (in_vars[i]->IsType<framework::LoDTensor>()) {
@@ -60,7 +62,7 @@ class SumKernel : public framework::OpKernel<T> {
           result.device(place) = result + in;
         } else if (in_vars[i]->IsType<framework::SelectedRows>()) {
           auto &in_t = in_vars[i]->Get<framework::SelectedRows>();
-          functor(context.device_context(), in_t, out);
+          functor(context.template device_context<DeviceContext>(), in_t, out);
         } else {
           PADDLE_THROW("Variable type must be LoDTensor/SelectedRows.");
         }
@@ -82,14 +84,14 @@ class SumKernel : public framework::OpKernel<T> {
       out_value->Resize(framework::make_ddim(in_dim_vec));
       out_value->mutable_data<T>(context.GetPlace());
 
-      math::SelectedRowsAddTo<Place, T> functor;
+      math::SelectedRowsAddTo<DeviceContext, T> functor;
 
       int64_t offset = 0;
       for (int i = 0; i < N; i++) {
         PADDLE_ENFORCE_EQ(out->height(),
                           in_vars[i]->Get<SelectedRows>().height());
-        functor(context.device_context(), in_vars[i]->Get<SelectedRows>(),
-                offset, out);
+        functor(context.template device_context<DeviceContext>(),
+                in_vars[i]->Get<SelectedRows>(), offset, out);
         offset += in_vars[i]->Get<SelectedRows>().value().numel();
       }
     } else if (out_var->IsType<framework::LoDTensorArray>()) {
@@ -112,7 +114,8 @@ class SumKernel : public framework::OpKernel<T> {
               PADDLE_ENFORCE(out_array[i].lod() == in_array[i].lod());
               auto in = EigenVector<T>::Flatten(in_array[i]);
               auto result = EigenVector<T>::Flatten(out_array[i]);
-              result.device(context.GetEigenDevice<Place>()) = result + in;
+              result.device(*context.template device_context<DeviceContext>()
+                                 .eigen_device()) = result + in;
             }
           }
         }
diff --git a/paddle/operators/top_k_op.cu b/paddle/operators/top_k_op.cu
index 7851c71bbe..453bd07267 100644
--- a/paddle/operators/top_k_op.cu
+++ b/paddle/operators/top_k_op.cu
@@ -317,4 +317,4 @@ class TopkOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_GPU_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel<float>);
+REGISTER_OP_CUDA_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel<float>);
diff --git a/paddle/operators/top_k_op.h b/paddle/operators/top_k_op.h
index bc8563717a..e9cd9bbd4d 100644
--- a/paddle/operators/top_k_op.h
+++ b/paddle/operators/top_k_op.h
@@ -27,7 +27,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class TopkKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
diff --git a/paddle/operators/transpose_op.cc b/paddle/operators/transpose_op.cc
index 94de3d5069..de5ff561ad 100644
--- a/paddle/operators/transpose_op.cc
+++ b/paddle/operators/transpose_op.cc
@@ -112,8 +112,8 @@ class TransposeOpGrad : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(transpose, ops::TransposeOp, ops::TransposeOpMaker, transpose_grad,
             ops::TransposeOpGrad);
-REGISTER_OP_CPU_KERNEL(transpose,
-                       ops::TransposeKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    transpose, ops::TransposeKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     transpose_grad,
-    ops::TransposeGradKernel<paddle::platform::CPUPlace, float>);
+    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/transpose_op.cu.cc b/paddle/operators/transpose_op.cu.cc
index af3f581462..7d23f1493e 100644
--- a/paddle/operators/transpose_op.cu.cc
+++ b/paddle/operators/transpose_op.cu.cc
@@ -15,8 +15,9 @@
 #include "paddle/operators/transpose_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(transpose,
-                       ops::TransposeKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(
+REGISTER_OP_CUDA_KERNEL(
+    transpose,
+    ops::TransposeKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
     transpose_grad,
-    ops::TransposeGradKernel<paddle::platform::GPUPlace, float>);
+    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/transpose_op.h b/paddle/operators/transpose_op.h
index e296032f41..d995271a6b 100644
--- a/paddle/operators/transpose_op.h
+++ b/paddle/operators/transpose_op.h
@@ -20,33 +20,33 @@
 namespace paddle {
 namespace operators {
 
-template <typename Place, typename T>
-inline void TransCompute(const int dim, const platform::DeviceContext& dev_ctx,
+template <typename DeviceContext, typename T>
+inline void TransCompute(const int dim, const DeviceContext& dev_ctx,
                          const framework::Tensor& in, framework::Tensor* out,
                          const std::vector<int>& axis) {
   switch (dim) {
     case 1:
-      math::Transpose<Place, T, 1> trans1;
+      math::Transpose<DeviceContext, T, 1> trans1;
       trans1(dev_ctx, in, out, axis);
       break;
     case 2:
-      math::Transpose<Place, T, 2> trans2;
+      math::Transpose<DeviceContext, T, 2> trans2;
       trans2(dev_ctx, in, out, axis);
       break;
     case 3:
-      math::Transpose<Place, T, 3> trans3;
+      math::Transpose<DeviceContext, T, 3> trans3;
       trans3(dev_ctx, in, out, axis);
       break;
     case 4:
-      math::Transpose<Place, T, 4> trans4;
+      math::Transpose<DeviceContext, T, 4> trans4;
       trans4(dev_ctx, in, out, axis);
       break;
     case 5:
-      math::Transpose<Place, T, 5> trans5;
+      math::Transpose<DeviceContext, T, 5> trans5;
       trans5(dev_ctx, in, out, axis);
       break;
     case 6:
-      math::Transpose<Place, T, 6> trans6;
+      math::Transpose<DeviceContext, T, 6> trans6;
       trans6(dev_ctx, in, out, axis);
       break;
     default:
@@ -54,7 +54,7 @@ inline void TransCompute(const int dim, const platform::DeviceContext& dev_ctx,
   }
 }
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class TransposeKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -64,12 +64,12 @@ class TransposeKernel : public framework::OpKernel<T> {
 
     std::vector<int> axis = context.Attr<std::vector<int>>("axis");
     int ndims = axis.size();
-    auto& dev_ctx = context.device_context();
-    TransCompute<Place, T>(ndims, dev_ctx, *x, out, axis);
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    TransCompute<DeviceContext, T>(ndims, dev_ctx, *x, out, axis);
   }
 };
 
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class TransposeGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -88,8 +88,9 @@ class TransposeGradKernel : public framework::OpKernel<T> {
     }
 
     int ndims = axis.size();
-    auto& dev_ctx = context.device_context();
-    TransCompute<Place, T>(ndims, dev_ctx, *out_grad, x_grad, reversed_axis);
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    TransCompute<DeviceContext, T>(ndims, dev_ctx, *out_grad, x_grad,
+                                   reversed_axis);
   }
 };
 
diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc
index fff1dc7ccd..2a49ee471f 100644
--- a/paddle/operators/uniform_random_op.cc
+++ b/paddle/operators/uniform_random_op.cc
@@ -67,7 +67,7 @@ class UniformRandomOp : public framework::OperatorWithKernel {
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         static_cast<framework::DataType>(ctx.Attr<int>("dtype")),
-        ctx.device_context());
+        ctx.GetPlace());
   }
 };
 
diff --git a/paddle/operators/uniform_random_op.cu b/paddle/operators/uniform_random_op.cu
index 8b20bb8287..cfe9d293cf 100644
--- a/paddle/operators/uniform_random_op.cu
+++ b/paddle/operators/uniform_random_op.cu
@@ -63,6 +63,6 @@ class GPUUniformRandomKernel : public framework::OpKernel<T> {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_GPU_KERNEL(uniform_random,
-                       paddle::operators::GPUUniformRandomKernel<float>,
-                       paddle::operators::GPUUniformRandomKernel<double>);
+REGISTER_OP_CUDA_KERNEL(uniform_random,
+                        paddle::operators::GPUUniformRandomKernel<float>,
+                        paddle::operators::GPUUniformRandomKernel<double>);
diff --git a/paddle/operators/unpool_op.cc b/paddle/operators/unpool_op.cc
index 89c48e071c..49df2a530c 100644
--- a/paddle/operators/unpool_op.cc
+++ b/paddle/operators/unpool_op.cc
@@ -135,9 +135,10 @@ class UnpoolOpGrad : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(unpool, ops::UnpoolOp, ops::Unpool2dOpMaker, unpool_grad,
             ops::UnpoolOpGrad);
-REGISTER_OP_CPU_KERNEL(unpool,
-                       ops::UnpoolKernel<paddle::platform::CPUPlace, float>,
-                       ops::UnpoolKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
-    unpool_grad, ops::UnpoolGradKernel<paddle::platform::CPUPlace, float>,
-    ops::UnpoolGradKernel<paddle::platform::CPUPlace, double>);
+    unpool, ops::UnpoolKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::UnpoolKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    unpool_grad,
+    ops::UnpoolGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::UnpoolGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/unpool_op.cu.cc b/paddle/operators/unpool_op.cu.cc
index 18aafb7dc7..9b002e35c4 100644
--- a/paddle/operators/unpool_op.cu.cc
+++ b/paddle/operators/unpool_op.cu.cc
@@ -15,9 +15,10 @@ limitations under the License. */
 #include "paddle/operators/unpool_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(unpool,
-                       ops::UnpoolKernel<paddle::platform::GPUPlace, float>,
-                       ops::UnpoolKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(
-    unpool_grad, ops::UnpoolGradKernel<paddle::platform::GPUPlace, float>,
-    ops::UnpoolGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    unpool, ops::UnpoolKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::UnpoolKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    unpool_grad,
+    ops::UnpoolGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::UnpoolGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/unpool_op.h b/paddle/operators/unpool_op.h
index 243eb7e532..ee18b118c9 100644
--- a/paddle/operators/unpool_op.h
+++ b/paddle/operators/unpool_op.h
@@ -20,7 +20,7 @@ limitations under the License. */
 
 namespace paddle {
 namespace operators {
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class UnpoolKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -32,15 +32,16 @@ class UnpoolKernel : public framework::OpKernel<T> {
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
     T* output_data = out->mutable_data<T>(context.GetPlace());
+    auto& dev_ctx = context.template device_context<DeviceContext>();
     if (output_data) {
-      math::SetConstant<Place, T> set_zero;
-      set_zero(context.device_context(), out, static_cast<T>(0));
+      math::SetConstant<DeviceContext, T> set_zero;
+      set_zero(dev_ctx, out, static_cast<T>(0));
     }
-    math::Unpool2dMaxFunctor<Place, T> unpool2d_max_forward;
-    unpool2d_max_forward(context.device_context(), *in_x, *in_y, out);
+    math::Unpool2dMaxFunctor<DeviceContext, T> unpool2d_max_forward;
+    unpool2d_max_forward(dev_ctx, *in_x, *in_y, out);
   }
 };
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class UnpoolGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -56,15 +57,14 @@ class UnpoolGradKernel : public framework::OpKernel<T> {
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
 
-    auto& device_ctx = context.device_context();
-    math::SetConstant<Place, T> zero;
+    auto& device_ctx = context.template device_context<DeviceContext>();
+    math::SetConstant<DeviceContext, T> zero;
     if (in_x_grad) {
       in_x_grad->mutable_data<T>(context.GetPlace());
       zero(device_ctx, in_x_grad, static_cast<T>(0));
     }
-    math::Unpool2dMaxGradFunctor<Place, T> unpool2d_max_backward;
-    unpool2d_max_backward(context.device_context(), *in_x, *in_y, *out,
-                          *out_grad, in_x_grad);
+    math::Unpool2dMaxGradFunctor<DeviceContext, T> unpool2d_max_backward;
+    unpool2d_max_backward(device_ctx, *in_x, *in_y, *out, *out_grad, in_x_grad);
   }
 };
 }  // namespace operators
diff --git a/paddle/platform/device_context.cc b/paddle/platform/device_context.cc
index ae4f0bf896..2c7f964216 100644
--- a/paddle/platform/device_context.cc
+++ b/paddle/platform/device_context.cc
@@ -15,12 +15,6 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
-template <>
-Eigen::DefaultDevice* DeviceContext::GetEigenDevice<
-    platform::CPUPlace, Eigen::DefaultDevice>() const {
-  return reinterpret_cast<const CPUDeviceContext*>(this)->eigen_device();
-}
-
 CPUDeviceContext::CPUDeviceContext() {
   eigen_device_.reset(new Eigen::DefaultDevice());
 }
@@ -37,12 +31,6 @@ Place CPUDeviceContext::GetPlace() const { return CPUPlace(); }
 
 #ifdef PADDLE_WITH_CUDA
 
-template <>
-Eigen::GpuDevice*
-DeviceContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
-  return reinterpret_cast<const CUDADeviceContext*>(this)->eigen_device();
-}
-
 class EigenCudaStreamDevice : public Eigen::StreamInterface {
  public:
   EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) {
diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h
index ef5f19214d..596d9d0bba 100644
--- a/paddle/platform/device_context.h
+++ b/paddle/platform/device_context.h
@@ -27,24 +27,11 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
-template <typename T>
-struct EigenDeviceConverter;
-
-template <>
-struct EigenDeviceConverter<platform::CPUPlace> {
-  using EigenDeviceType = Eigen::DefaultDevice;
-};
-
 class DeviceContext {
  public:
   virtual ~DeviceContext() {}
   virtual Place GetPlace() const = 0;
 
-  template <typename PlaceType,
-            typename DeviceType =
-                typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
-  DeviceType* GetEigenDevice() const;
-
   virtual void Wait() const {}
 };
 
@@ -62,10 +49,6 @@ class CPUDeviceContext : public DeviceContext {
 };
 
 #ifdef PADDLE_WITH_CUDA
-template <>
-struct EigenDeviceConverter<platform::GPUPlace> {
-  using EigenDeviceType = Eigen::GpuDevice;
-};
 
 class EigenCudaStreamDevice;
 
diff --git a/paddle/platform/device_context_test.cc b/paddle/platform/device_context_test.cc
index 8bf5174c4a..4893cd92f6 100644
--- a/paddle/platform/device_context_test.cc
+++ b/paddle/platform/device_context_test.cc
@@ -22,9 +22,8 @@ TEST(Device, Init) {
 
   int count = paddle::platform::GetCUDADeviceCount();
   for (int i = 0; i < count; i++) {
-    DeviceContext* device_context = new CUDADeviceContext(GPUPlace(i));
-    Eigen::GpuDevice* gpu_device =
-        device_context->template GetEigenDevice<GPUPlace>();
+    CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i));
+    Eigen::GpuDevice* gpu_device = device_context->eigen_device();
     ASSERT_NE(nullptr, gpu_device);
     delete device_context;
   }
diff --git a/paddle/platform/transform.h b/paddle/platform/transform.h
index bb9d59ec0a..148ebaed3d 100644
--- a/paddle/platform/transform.h
+++ b/paddle/platform/transform.h
@@ -31,7 +31,7 @@ namespace paddle {
 namespace platform {
 
 // Transform on host or device. It provides the same API in std library.
-template <typename Place>
+template <typename DeviceContext>
 struct Transform {
   template <typename InputIter, typename OutputIter, typename UnaryOperation>
   void operator()(const DeviceContext& context, InputIter first, InputIter last,
@@ -45,16 +45,16 @@ struct Transform {
 };
 
 template <>
-struct Transform<platform::CPUPlace> {
+struct Transform<platform::CPUDeviceContext> {
   template <typename InputIter, typename OutputIter, typename UnaryOperation>
-  void operator()(const DeviceContext& context, InputIter first, InputIter last,
-                  OutputIter result, UnaryOperation op) {
+  void operator()(const platform::CPUDeviceContext& context, InputIter first,
+                  InputIter last, OutputIter result, UnaryOperation op) {
     std::transform(first, last, result, op);
   }
 
   template <typename InputIter1, typename InputIter2, typename OutputIter,
             typename BinaryOperation>
-  void operator()(const DeviceContext& context, InputIter1 first1,
+  void operator()(const platform::CPUDeviceContext& context, InputIter1 first1,
                   InputIter1 last1, InputIter2 first2, OutputIter result,
                   BinaryOperation op) {
     std::transform(first1, last1, first2, result, op);
@@ -63,27 +63,25 @@ struct Transform<platform::CPUPlace> {
 
 #ifdef __NVCC__
 template <>
-struct Transform<platform::GPUPlace> {
+struct Transform<platform::CUDADeviceContext> {
   template <typename InputIter, typename OutputIter, typename UnaryOperation>
-  void operator()(const DeviceContext& context, InputIter first, InputIter last,
-                  OutputIter result, UnaryOperation op) {
+  void operator()(const platform::CUDADeviceContext& context, InputIter first,
+                  InputIter last, OutputIter result, UnaryOperation op) {
     auto place = context.GetPlace();
     PADDLE_ENFORCE(is_gpu_place(place), "It must use GPU place.");
-    auto& ctx = reinterpret_cast<const CUDADeviceContext&>(context);
-    thrust::transform(thrust::cuda::par.on(ctx.stream()),
+    thrust::transform(thrust::cuda::par.on(context.stream()),
                       details::DevPtrCast(first), details::DevPtrCast(last),
                       details::DevPtrCast(result), op);
   }
 
   template <typename InputIter1, typename InputIter2, typename OutputIter,
             typename BinaryOperation>
-  void operator()(const DeviceContext& context, InputIter1 first1,
+  void operator()(const platform::CUDADeviceContext& context, InputIter1 first1,
                   InputIter1 last1, InputIter2 first2, OutputIter result,
                   BinaryOperation op) {
     auto place = context.GetPlace();
     PADDLE_ENFORCE(is_gpu_place(place), "It must use GPU place.");
-    auto& ctx = reinterpret_cast<const CUDADeviceContext&>(context);
-    thrust::transform(thrust::cuda::par.on(ctx.stream()),
+    thrust::transform(thrust::cuda::par.on(context.stream()),
                       details::DevPtrCast(first1), details::DevPtrCast(last1),
                       details::DevPtrCast(first2), details::DevPtrCast(result),
                       op);
diff --git a/paddle/platform/transform_test.cu b/paddle/platform/transform_test.cu
index c76cab80e4..d36eac8379 100644
--- a/paddle/platform/transform_test.cu
+++ b/paddle/platform/transform_test.cu
@@ -39,7 +39,7 @@ TEST(Transform, CPUUnary) {
   using namespace paddle::platform;
   CPUDeviceContext ctx;
   float buf[4] = {0.1, 0.2, 0.3, 0.4};
-  Transform<paddle::platform::CPUPlace> trans;
+  Transform<paddle::platform::CPUDeviceContext> trans;
   trans(ctx, buf, buf + 4, buf, Scale<float>(10));
   for (int i = 0; i < 4; ++i) {
     ASSERT_NEAR(buf[i], static_cast<float>(i + 1), 1e-5);
@@ -54,7 +54,7 @@ TEST(Transform, GPUUnary) {
   float cpu_buf[4] = {0.1, 0.2, 0.3, 0.4};
   float* gpu_buf = static_cast<float*>(Alloc(gpu0, sizeof(float) * 4));
   Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf));
-  Transform<paddle::platform::GPUPlace> trans;
+  Transform<paddle::platform::CUDADeviceContext> trans;
   trans(ctx, gpu_buf, gpu_buf + 4, gpu_buf, Scale<float>(10));
   ctx.Wait();
   Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf));
@@ -68,7 +68,7 @@ TEST(Transform, CPUBinary) {
   using namespace paddle::platform;
   using namespace paddle::memory;
   int buf[4] = {1, 2, 3, 4};
-  Transform<paddle::platform::CPUPlace> trans;
+  Transform<paddle::platform::CPUDeviceContext> trans;
   CPUDeviceContext ctx;
   trans(ctx, buf, buf + 4, buf, buf, Multiply<int>());
   for (int i = 0; i < 4; ++i) {
@@ -84,7 +84,7 @@ TEST(Transform, GPUBinary) {
   CUDADeviceContext ctx(gpu0);
   int* gpu_buf = static_cast<int*>(Alloc(gpu0, sizeof(buf)));
   Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf));
-  Transform<paddle::platform::GPUPlace> trans;
+  Transform<paddle::platform::CUDADeviceContext> trans;
   trans(ctx, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply<int>());
   ctx.Wait();
   Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf));

From de8c4627776b2b9a60dbcc64c48e858c80a2715f Mon Sep 17 00:00:00 2001
From: QI JUN <qijun1994@hotmail.com>
Date: Tue, 12 Dec 2017 14:42:33 +0800
Subject: [PATCH 32/94] Update new op docs (#6505)

* update docs about how to add a new operator
---
 doc/howto/dev/new_op_cn.md | 98 ++++++++++++++++----------------------
 doc/howto/dev/new_op_en.md | 93 ++++++++++++++++--------------------
 2 files changed, 80 insertions(+), 111 deletions(-)

diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md
index 6cfc9536f2..44dbeecbbd 100644
--- a/doc/howto/dev/new_op_cn.md
+++ b/doc/howto/dev/new_op_cn.md
@@ -30,8 +30,8 @@
 --------------  | :----------------------
 OpProtoMake定义  | `.cc`文件，Backward Op不需要定义OpProtoMake
 Op定义           | `.cc`文件
-Kernel实现       | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，GPU 实现在`.cu`文件中。
-注册Op           | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中
+Kernel实现       | CPU、CUDA共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，CUDA 实现在`.cu`文件中。
+注册Op           | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，CUDA实现在`.cu`文件中
 
 
 实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。**系统会根据文件名自动构建op和其对应的Python扩展。**
@@ -153,7 +153,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
 `MulKernel`继承自`framework::OpKernel`，带有下面两个模板参数:
 
-- `typename  Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
+- `typename DeviceContext`: 表示设备类型，不同设备(CPU、CUDA)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
 
 - `typename T` : 表示数据类型，如`float`, `double`等。
 
@@ -165,7 +165,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 下面是 `MulKernel` `Compute`的实现：
 
   ```cpp
-  template <typename Place, typename T>
+  template <typename DeviceContext, typename T>
   class MulKernel : public framework::OpKernel {
   public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -173,18 +173,16 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
     auto* Y = context.Input<Tensor>("Y");
     auto* Z = context.Output<Tensor>("Out");
     Z->mutable_data<T>(context.GetPlace());
-    auto* device_context =
-        const_cast<platform::DeviceContext*>(context.device_context_);
-    math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
+    auto& device_context = context.template device_context<DeviceContext>();
+    math::matmul<DeviceContext, T>(*X, false, *Y, false, 1, Z, 0, device_context);
   }
   };
-  ```
 
-需要注意：**不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。**
+需要注意：**不同设备(CPU、CUDA)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。**
 
-`MulOp`的CPU、GPU实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
+`MulOp`的CPU、CUDA实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
 
-为了使`OpKernel`的计算过程书写更加简单，并且CPU、GPU的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)。
+为了使`OpKernel`的计算过程书写更加简单，并且CPU、CUDA的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)。
 
 
 到此，前向Op实现完成。接下来，需要在`.cc`文件中注册该op和kernel。
@@ -197,9 +195,9 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
     ```cpp
     namespace ops = paddle::operators;
     REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
-    REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
+    REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUDeviceContext, float>);
     REGISTER_OP_CPU_KERNEL(mul_grad,
-                  ops::MulGradKernel<paddle::platform::CPUPlace, float>);
+                  ops::MulGradKernel<paddle::platform::CPUDeviceContext, float>);
     ```
 
    在上面的代码中：
@@ -209,17 +207,17 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
     - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulGradKernel`类。
 
 
-- 在 `.cu`文件中注册GPU Kernel。
-    - 请注意，如果GPU Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下：
+- 在 `.cu`文件中注册CUDA Kernel。
+    - 请注意，如果CUDA Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下：
 
     ```cpp
     // if use Eigen unsupported module before include head files
-    // #define EIGEN_USE_GPU
+    #define EIGEN_USE_GPU
 
     namespace ops = paddle::operators;
-    REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
-    REGISTER_OP_GPU_KERNEL(mul_grad,
-                           ops::MulGradKernel<paddle::platform::GPUPlace, float>);
+    REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel<paddle::platform::CUDADeviceContext, float>);
+    REGISTER_OP_CUDA_KERNEL(mul_grad,
+                           ops::MulGradKernel<paddle::platform::CUDADeviceContext, float>);
     ```
 
 ### 5. 编译
@@ -236,71 +234,55 @@ make mul_op
 
 ## 实现单元测试
 
-单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
+单测包括对比前向Op不同设备(CPU、CUDA)的实现、对比反向OP不同设备(CPU、CUDA)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
 
-### 前向Operator单元测试
 
-前向Op单元测试继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`。各项更加具体的单元测试在`OpTestMeta`里完成。测试前向Operator，需要：
+Op单元测试继承自`OpTest`。各项更加具体的单元测试在`TestMulOp`里完成。测试Operator，需要：
 
 1. 在`setUp`函数定义输入、输出，以及相关的属性参数。
 2. 生成随机的输入数据。
 3. 在Python脚本中实现与前向operator相同的计算逻辑，得到输出值，与operator前向计算的输出进行对比。
+4. 反向计算已经自动集成进测试框架，直接调用相应接口即可。
 
 
   ```python
   import unittest
   import numpy as np
-  from gradient_checker import GradientChecker, create_op
-  from op_test_util import OpTestMeta
+  from op_test import OpTest
 
-  class TestMulOp(unittest.TestCase):
-      __metaclass__ = OpTestMeta
 
+  class TestMulOp(OpTest):
       def setUp(self):
-          self.type = "mul"
+          self.op_type = "mul"
           self.inputs = {
               'X': np.random.random((32, 84)).astype("float32"),
               'Y': np.random.random((84, 100)).astype("float32")
           }
           self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
-  ```
 
-上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释：
-
-- `self.type = "mul" ` : 定义类型，与operator注册时注册的类型一致。
-- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。
-- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。
-
-
-### 反向Operator单元测试
+      def test_check_output(self):
+          self.check_output()
 
-反向Op单元测试继承自`GradientChecker`，而`GradientChecker`继承自`unittest.TestCase`，因此，**反向单元测试函数需要以`test_`开头**。
+      def test_check_grad_normal(self):
+          self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
 
-```python
-class TestMulGradOp(GradientChecker):
-    def setUp(self):
-        self.op = create_op("mul")
-        self.inputs = {
-            'X': np.random.random((32, 84)).astype("float32"),
-            'Y': np.random.random((84, 100)).astype("float32")
-        }
+      def test_check_grad_ingore_x(self):
+          self.check_grad(
+              ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
 
-    def test_check_grad_normal(self):
-        # mul op will enlarge the relative error
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
+      def test_check_grad_ingore_y(self):
+          self.check_grad(
+              ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
 
-    def test_check_grad_ingore_x(self):
-        self.check_grad(
-            ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
+    ```
 
-    def test_check_grad_ingore_y(self):
-        self.check_grad(
-            ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
-```
+上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释：
 
-下面解释代码中一些关键的地方:
+- `self.op_type = "mul" ` : 定义类型，与operator注册时注册的类型一致。
+- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。
+- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。
 
-- 调用`create_op("mul")`创建反向Op对应的前向Op。
+而反向测试中：
 - `test_check_grad_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。
   - 第一个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。
   - 第二个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`。
@@ -328,5 +310,5 @@ ctest -R test_mul_op
 
 - 为每个Op创建单独的`*_op.h`（如有）、`*_op.cc`和`*_op.cu`（如有）。不允许一个文件中包含多个Op，这将会导致编译出错。
 - 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OP(B, ...)`等，这将会导致单元测试出错。
-- 如果Op没有实现GPU Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
+- 如果Op没有实现CUDA Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
 - 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。
diff --git a/doc/howto/dev/new_op_en.md b/doc/howto/dev/new_op_en.md
index 1e88e1f5b4..510233306c 100644
--- a/doc/howto/dev/new_op_en.md
+++ b/doc/howto/dev/new_op_en.md
@@ -28,8 +28,8 @@ An operator can be differentiated by whether in has kernel methods. An operator
 --------------  | :----------------------
 OpProtoMake definition  | `.cc`files, Backward Op does not need an OpProtoMake interface.
 Op definition           | `.cc` files
-Kernel implementation       | The kernel methods shared between CPU and GPU are defined in `.h` files. CPU-specific kernels live in `.cc` files, while GPU-specific kernels are implemented in `.cu`files.
-Registering the Op           | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the GPU implementation.
+Kernel implementation       | The kernel methods shared between CPU and CUDA are defined in `.h` files. CPU-specific kernels live in `.cc` files, while CUDA-specific kernels are implemented in `.cu`files.
+Registering the Op           | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the CUDA implementation.
 
 
 New Operator implementations are added to the list [paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators), with file names in the format `*_op.h` (if applicable), `*_op.cc`, `*_op.cu` (if applicable).** The system will use the naming scheme to automatically build operators and their corresponding Python extensions. **
@@ -151,7 +151,7 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w
 
 `MulKernel` inherits `framework::OpKernel`, which includes the following templates:
 
-- `typename  Place` denotes device type. When different devices, namely the CPU and the GPU, share the same kernel, this template needs to be added. If they don't share kernels, this must not be added. An example of a non-sharing kernel is [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
+- `typename  DeviceContext` denotes device context type. When different devices, namely the CPUDeviceContext and the CUDADeviceContext, share the same kernel, this template needs to be added. If they don't share kernels, this must not be added. An example of a non-sharing kernel is [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
 
 - `typename T` denotes data type, such as `float` or `double`.
 
@@ -163,7 +163,7 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w
 `MulKernel`'s implementation of `Compute` is as follows:
 
   ```cpp
-  template <typename Place, typename T>
+  template <typename DeviceContext, typename T>
   class MulKernel : public framework::OpKernel {
   public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -171,16 +171,15 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w
     auto* Y = context.Input<Tensor>("Y");
     auto* Z = context.Output<Tensor>("Out");
     Z->mutable_data<T>(context.GetPlace());
-    auto* device_context =
-        const_cast<platform::DeviceContext*>(context.device_context_);
-    math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
+    auto& device_context = context.template device_context<DeviceContext>();
+    math::matmul<DeviceContext, T>(*X, false, *Y, false, 1, Z, 0, device_context);
   }
   };
   ```
 
-Note that **different devices (CPU, GPU)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.**
+Note that **different devices (CPU, CUDA)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.**
 
-`MulOp`'s CPU and GPU share the same `Kernel`. A non-sharing  `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
+`MulOp`'s CPU and CUDA share the same `Kernel`. A non-sharing  `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
 
 To ease the writing of `OpKernel` compute, and for reusing code cross-device, [`Eigen-unsupported Tensor`](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md?fileviewer=file-view-default) module is used to implement `Compute` interface. To learn about how the Eigen library is used in PaddlePaddle, please see [usage document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md).
 
@@ -196,9 +195,9 @@ The definition of its corresponding backward operator, if applicable, is similar
     ```cpp
     namespace ops = paddle::operators;
     REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
-    REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
+    REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUDeviceContext, float>);
     REGISTER_OP_CPU_KERNEL(mul_grad,
-                  ops::MulGradKernel<paddle::platform::CPUPlace, float>);
+                  ops::MulGradKernel<paddle::platform::CPUDeviceContext, float>);
     ```
 
    In that code block,
@@ -208,17 +207,17 @@ The definition of its corresponding backward operator, if applicable, is similar
     - `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`.
 
 
-- Registering GPU Kernel in `.cu` files
-    - Note that if GPU Kernel is implemented using the `Eigen unsupported` module, then on top of `.cu`, a macro definition `#define EIGEN_USE_GPU` is needed, such as
+- Registering CUDA Kernel in `.cu` files
+    - Note that if CUDA Kernel is implemented using the `Eigen unsupported` module, then on top of `.cu`, a macro definition `#define EIGEN_USE_GPU` is needed, such as
 
     ```cpp
     // if use Eigen unsupported module before include head files
     #define EIGEN_USE_GPU
 
     namespace ops = paddle::operators;
-    REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
-    REGISTER_OP_GPU_KERNEL(mul_grad,
-                           ops::MulGradKernel<paddle::platform::GPUPlace, float>);
+    REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel<paddle::platform::CUDADeviceContext, float>);
+    REGISTER_OP_CUDA_KERNEL(mul_grad,
+                           ops::MulGradKernel<paddle::platform::CUDADeviceContext, float>);
     ```
 
 ### 5. Compilation
@@ -253,62 +252,50 @@ A forward operator unit test inherits `unittest.TestCase` and defines metaclass
 
 2. Generating random input data.
 
-3. Implementing the same computation logic in a Python script:
+3. Implementing the same computation logic in a Python script.
+
+4. Call check gradient function to check the backward operator.
 
   ```python
   import unittest
   import numpy as np
-  from gradient_checker import GradientChecker, create_op
-  from op_test_util import OpTestMeta
+  from op_test import OpTest
 
-  class TestMulOp(unittest.TestCase):
-      __metaclass__ = OpTestMeta
 
+  class TestMulOp(OpTest):
       def setUp(self):
-          self.type = "mul"
+          self.op_type = "mul"
           self.inputs = {
               'X': np.random.random((32, 84)).astype("float32"),
               'Y': np.random.random((84, 100)).astype("float32")
           }
           self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
-  ```
-Get its output, and compare it with the forward operator's own output.
-
-The code above first loads required packages. In addition, we have
-
-- `self.type = "mul" ` defines the type that is identical to what the operator's registered type.
-- `self.inputs` defines input, with type `numpy.array` and initializes it.
-- `self.outputs` defines output and completes the same operator computation in the Python script, and returns its result from the Python script.
 
-### Testing Backward Operators
+      def test_check_output(self):
+          self.check_output()
+          
+      def test_check_grad_normal(self):
+          self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
 
-A backward operator unit test inherits `GradientChecker`, which inherits `unittest.TestCase`. As a result, **a backward operator unit test needs to be have the prefix `test_`**.
+      def test_check_grad_ingore_x(self):
+          self.check_grad(
+              ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
 
-```python
-class TestMulGradOp(GradientChecker):
-    def setUp(self):
-        self.op = create_op("mul")
-        self.inputs = {
-            'X': np.random.random((32, 84)).astype("float32"),
-            'Y': np.random.random((84, 100)).astype("float32")
-        }
+      def test_check_grad_ingore_y(self):
+          self.check_grad(
+              ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
 
-    def test_check_grad_normal(self):
-        # mul op will enlarge the relative error
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
+    ```
+Get its output, and compare it with the forward operator's own output.
 
-    def test_check_grad_ingore_x(self):
-        self.check_grad(
-            ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
+The code above first loads required packages. In addition, we have
 
-    def test_check_grad_ingore_y(self):
-        self.check_grad(
-            ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
-```
+- `self.op_type = "mul" ` defines the type that is identical to what the operator's registered type.
+- `self.inputs` defines input, with type `numpy.array` and initializes it.
+- `self.outputs` defines output and completes the same operator computation in the Python script, and returns its result from the Python script.
 
-Some key points in the code above include:
+Some key points in checking gradient above include:
 
-- `create_op("mul")` creates the backward operator's corresponding forward operator.
 - `test_normal` calls `check_grad` to validate scaling tests' correctness and stability through numeric methods.
   - The first variable `["X", "Y"]` appoints `X` and `Y` to be scale tested.
   - The second variable `"Out"` points to the network's final output target `Out`.
@@ -338,5 +325,5 @@ ctest -R test_mul_op
 
 - Every `*_op.h` (if applicable), `*_op.cc`, and `*_op.cu` (if applicable) must be created for a unique Op. Compiling will fail if multiple operators are included per file.
 - The type with which an operator is registered needs to be identical to the Op's name. Registering `REGISTER_OP(B, ...)` in `A_op.cc` will cause unit testing failures.
-- If the operator does not implement a GPU kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail.
+- If the operator does not implement a CUDA kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail.
 - If multiple operators rely on some shared methods, a file NOT named `*_op.*` can be created to store them, such as `gather.h`.

From d918ccded35d972eba42b07767972bc4c2dd0921 Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Tue, 12 Dec 2017 15:22:52 +0800
Subject: [PATCH 33/94] Add fill_op (#6477)

* Add fill_op

* Fix bug
---
 paddle/operators/fill_op.cc                  | 111 +++++++++++++++++++
 python/paddle/v2/fluid/tests/test_fill_op.py |  24 ++++
 2 files changed, 135 insertions(+)
 create mode 100644 paddle/operators/fill_op.cc
 create mode 100644 python/paddle/v2/fluid/tests/test_fill_op.py

diff --git a/paddle/operators/fill_op.cc b/paddle/operators/fill_op.cc
new file mode 100644
index 0000000000..382e161c5d
--- /dev/null
+++ b/paddle/operators/fill_op.cc
@@ -0,0 +1,111 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/framework/data_type.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/detail/safe_ref.h"
+
+namespace paddle {
+namespace operators {
+
+struct FillOpVisitor {
+  FillOpVisitor(framework::LoDTensor *tensor, const std::vector<float> &value)
+      : tensor_(tensor), value_(value) {}
+
+  template <typename T>
+  void operator()() const {
+    platform::CPUPlace cpu;
+    auto *data = tensor_->mutable_data<T>(cpu);
+    std::transform(value_.data(), value_.data() + tensor_->numel(), data,
+                   [](float dat) { return static_cast<T>(dat); });
+  }
+
+  framework::LoDTensor *tensor_;
+  const std::vector<float> &value_;
+};
+
+class FillOp : public framework::OperatorBase {
+ public:
+  FillOp(const std::string &type, const framework::VariableNameMap &inputs,
+         const framework::VariableNameMap &outputs,
+         const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto &out =
+        detail::Ref(detail::Ref(scope.FindVar(Output("Out")),
+                                "Cannot find variable %s", Output("Out"))
+                        .GetMutable<framework::LoDTensor>());
+    out.Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
+    auto dtype = static_cast<framework::DataType>(Attr<int>("dtype"));
+    platform::CPUPlace cpu;
+    auto force_cpu = Attr<bool>("force_cpu");
+    out.mutable_data(force_cpu ? cpu : dev_ctx.GetPlace(),
+                     framework::ToTypeIndex(dtype));
+
+    framework::LoDTensor tensor;
+
+    if (force_cpu || platform::is_cpu_place(dev_ctx.GetPlace())) {
+      tensor.ShareDataWith(out);
+    } else {
+      // Always make tensor in CPU memory.
+      tensor.Resize(out.dims());
+      tensor.mutable_data(cpu, framework::ToTypeIndex(dtype));
+    }
+
+    framework::VisitDataType(
+        dtype, FillOpVisitor(&tensor, Attr<std::vector<float>>("value")));
+
+    if (!force_cpu && platform::is_gpu_place(dev_ctx.GetPlace())) {
+      // Copy tensor to out
+      framework::CopyFrom(tensor, dev_ctx.GetPlace(), dev_ctx, &out);
+    }
+  }
+};
+
+class FillOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  FillOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddComment(R"DOC(Fill operator
+
+Fill an tensor with `value` and `shape`. The type of the tensor is specify by
+`dtype`.
+)DOC");
+    AddOutput("Out", "(LoDTensor) The output tensor.");
+    AddAttr<std::vector<float>>(
+        "value", "The float values of tensor, which are flatten in row major");
+    AddAttr<std::vector<int>>("shape", "The shape of output tensor");
+    AddAttr<int>("dtype", "The data type of output tensor, Default is float")
+        .SetDefault(framework::DataType::FP32);
+    AddAttr<bool>("force_cpu",
+                  "Whether the output tensor must be at CPU memory or not. "
+                  "Default is false.")
+        .SetDefault(false);
+  }
+};
+
+class FillOpInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {
+    context->SetOutputDim(
+        "Out",
+        framework::make_ddim(context->Attrs().Get<std::vector<int>>("shape")));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(fill, ops::FillOp, ops::FillOpInferShape, ops::FillOpMaker);
diff --git a/python/paddle/v2/fluid/tests/test_fill_op.py b/python/paddle/v2/fluid/tests/test_fill_op.py
new file mode 100644
index 0000000000..88337598c8
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/test_fill_op.py
@@ -0,0 +1,24 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+import paddle.v2.fluid.core as core
+
+
+class TestFillOp(OpTest):
+    def setUp(self):
+        self.op_type = "fill"
+        val = np.random.random(size=[100, 200])
+        self.inputs = {}
+        self.attrs = {
+            'value': val.flatten().tolist(),
+            'shape': [100, 200],
+            'dtype': int(core.DataType.FP64)
+        }
+        self.outputs = {'Out': val.astype('float64')}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()

From 9a89b041ba12ab8bcf9e36e218486b224eae1d33 Mon Sep 17 00:00:00 2001
From: guosheng <guosheng@baidu.com>
Date: Mon, 11 Dec 2017 23:23:46 +0800
Subject: [PATCH 34/94] Add ChunkEvaluator for multi-batches

---
 paddle/operators/chunk_eval_op.cc             | 18 +++++
 paddle/operators/chunk_eval_op.h              | 42 +++++++----
 python/paddle/v2/fluid/evaluator.py           | 73 ++++++++++++++++++-
 python/paddle/v2/fluid/layers.py              | 14 +++-
 .../tests/book/test_label_semantic_roles.py   | 12 +--
 .../v2/fluid/tests/test_chunk_eval_op.py      |  8 +-
 6 files changed, 143 insertions(+), 24 deletions(-)

diff --git a/paddle/operators/chunk_eval_op.cc b/paddle/operators/chunk_eval_op.cc
index 94127ab33e..ff2a08ccac 100644
--- a/paddle/operators/chunk_eval_op.cc
+++ b/paddle/operators/chunk_eval_op.cc
@@ -32,6 +32,13 @@ class ChunkEvalOp : public framework::OperatorWithKernel {
                    "Output(Recall) of ChunkEvalOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("F1-Score"),
                    "Output(F1-Score) of ChunkEvalOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("NumInferChunks"),
+                   "Output(NumInferChunks) of ChunkEvalOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("NumLabelChunks"),
+                   "Output(NumLabelChunks) of ChunkEvalOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("NumCorrectChunks"),
+        "Output(NumCorrectChunks) of ChunkEvalOp should not be null.");
 
     auto inference_dim = ctx->GetInputDim("Inference");
     auto label_dim = ctx->GetInputDim("Label");
@@ -42,6 +49,9 @@ class ChunkEvalOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("Precision", {1});
     ctx->SetOutputDim("Recall", {1});
     ctx->SetOutputDim("F1-Score", {1});
+    ctx->SetOutputDim("NumInferChunks", {1});
+    ctx->SetOutputDim("NumLabelChunks", {1});
+    ctx->SetOutputDim("NumCorrectChunks", {1});
   }
 
  protected:
@@ -70,6 +80,14 @@ class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker {
               "sensitivity) of chunks on the given mini-batch.");
     AddOutput("F1-Score",
               "(float). The evaluated F1-Score on the given mini-batch.");
+    AddOutput(
+        "NumInferChunks",
+        "(int). The number of chunks in Inference on the given mini-batch.");
+    AddOutput("NumLabelChunks",
+              "(int). The number of chunks in Label on the given mini-batch.");
+    AddOutput("NumCorrectChunks",
+              "(int). The number of chunks both in Inference and Label on the "
+              "given mini-batch.");
     AddAttr<int>("num_chunk_types",
                  "(int). The number of chunk type. See below for details.");
     AddAttr<std::string>(
diff --git a/paddle/operators/chunk_eval_op.h b/paddle/operators/chunk_eval_op.h
index dd88f2553b..5d703333c7 100644
--- a/paddle/operators/chunk_eval_op.h
+++ b/paddle/operators/chunk_eval_op.h
@@ -111,9 +111,7 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
     std::vector<Segment> label_segments;
     std::vector<Segment> output_segments;
     std::set<int> excluded_chunk_types;
-    int64_t num_output_segments = 0;
-    int64_t num_label_segments = 0;
-    int64_t num_correct = 0;
+
     if (context.Attr<std::string>("chunk_scheme") == "IOB") {
       num_tag_types = 2;
       tag_begin = 0;
@@ -151,12 +149,24 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
     auto* precision = context.Output<Tensor>("Precision");
     auto* recall = context.Output<Tensor>("Recall");
     auto* f1 = context.Output<Tensor>("F1-Score");
+    auto* num_infer_chunks = context.Output<Tensor>("NumInferChunks");
+    auto* num_label_chunks = context.Output<Tensor>("NumLabelChunks");
+    auto* num_correct_chunks = context.Output<Tensor>("NumCorrectChunks");
 
     const int64_t* inference_data = inference->data<int64_t>();
     const int64_t* label_data = label->data<int64_t>();
     T* precision_data = precision->mutable_data<T>(context.GetPlace());
     T* racall_data = recall->mutable_data<T>(context.GetPlace());
     T* f1_data = f1->mutable_data<T>(context.GetPlace());
+    int64_t* num_infer_chunks_data =
+        num_infer_chunks->mutable_data<int64_t>(context.GetPlace());
+    int64_t* num_label_chunks_data =
+        num_label_chunks->mutable_data<int64_t>(context.GetPlace());
+    int64_t* num_correct_chunks_data =
+        num_correct_chunks->mutable_data<int64_t>(context.GetPlace());
+    *num_infer_chunks_data = 0;
+    *num_label_chunks_data = 0;
+    *num_correct_chunks_data = 0;
 
     auto lod = label->lod();
     PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
@@ -166,17 +176,23 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
     for (int i = 0; i < num_sequences; ++i) {
       int seq_length = lod[0][i + 1] - lod[0][i];
       EvalOneSeq(inference_data + lod[0][i], label_data + lod[0][i], seq_length,
-                 output_segments, label_segments, num_output_segments,
-                 num_label_segments, num_correct, num_chunk_types,
-                 num_tag_types, other_chunk_type, tag_begin, tag_inside,
-                 tag_end, tag_single, excluded_chunk_types);
+                 output_segments, label_segments, *num_infer_chunks_data,
+                 *num_label_chunks_data, *num_correct_chunks_data,
+                 num_chunk_types, num_tag_types, other_chunk_type, tag_begin,
+                 tag_inside, tag_end, tag_single, excluded_chunk_types);
     }
-    *precision_data = !num_output_segments ? 0 : static_cast<T>(num_correct) /
-                                                     num_output_segments;
-    *racall_data = !num_label_segments ? 0 : static_cast<T>(num_correct) /
-                                                 num_label_segments;
-    *f1_data = !num_correct ? 0 : 2 * (*precision_data) * (*racall_data) /
-                                      ((*precision_data) + (*racall_data));
+    *precision_data = !(*num_infer_chunks_data)
+                          ? 0
+                          : static_cast<T>(*num_correct_chunks_data) /
+                                (*num_infer_chunks_data);
+    *racall_data = !(*num_label_chunks_data)
+                       ? 0
+                       : static_cast<T>(*num_correct_chunks_data) /
+                             (*num_label_chunks_data);
+    *f1_data = !(*num_correct_chunks_data)
+                   ? 0
+                   : 2 * (*precision_data) * (*racall_data) /
+                         ((*precision_data) + (*racall_data));
   }
 
   void EvalOneSeq(const int64_t* output, const int64_t* label, int length,
diff --git a/python/paddle/v2/fluid/evaluator.py b/python/paddle/v2/fluid/evaluator.py
index 137c573622..2d23ff0a16 100644
--- a/python/paddle/v2/fluid/evaluator.py
+++ b/python/paddle/v2/fluid/evaluator.py
@@ -4,7 +4,7 @@ import layers
 from framework import Program, unique_name, Variable
 from layer_helper import LayerHelper
 
-__all__ = ['Accuracy']
+__all__ = ['Accuracy', 'ChunkEvaluator']
 
 
 def _clone_var_(block, var):
@@ -132,3 +132,74 @@ class Accuracy(Evaluator):
         correct = layers.cast(correct, dtype='float32', **kwargs)
         out = layers.elementwise_div(x=correct, y=total, **kwargs)
         return np.array(executor.run(eval_program, fetch_list=[out])[0])
+
+
+class ChunkEvaluator(Evaluator):
+    """
+    Accumulate counter numbers output by chunk_eval from mini-batches and 
+    compute the precision recall and F1-score using the accumulated counter 
+    numbers.
+    """
+
+    def __init__(self,
+                 input,
+                 label,
+                 chunk_scheme,
+                 num_chunk_types,
+                 excluded_chunk_types=None,
+                 **kwargs):
+        super(ChunkEvaluator, self).__init__("chunk_eval", **kwargs)
+        main_program = self.helper.main_program
+        if main_program.current_block().idx != 0:
+            raise ValueError("You can only invoke Evaluator in root block")
+
+        self.num_infer_chunks = self.create_state(
+            dtype='int64', shape=[1], suffix='num_infer_chunks')
+        self.num_label_chunks = self.create_state(
+            dtype='int64', shape=[1], suffix='num_label_chunks')
+        self.num_correct_chunks = self.create_state(
+            dtype='int64', shape=[1], suffix='num_correct_chunks')
+        kwargs = {'main_program': main_program}
+        precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks = layers.chunk_eval(
+            input=input,
+            label=label,
+            chunk_scheme=chunk_scheme,
+            num_chunk_types=num_chunk_types,
+            excluded_chunk_types=excluded_chunk_types,
+            **kwargs)
+        layers.sums(
+            input=[self.num_infer_chunks, num_infer_chunks],
+            out=self.num_infer_chunks,
+            **kwargs)
+        layers.sums(
+            input=[self.num_label_chunks, num_label_chunks],
+            out=self.num_label_chunks,
+            **kwargs)
+        layers.sums(
+            input=[self.num_correct_chunks, num_correct_chunks],
+            out=self.num_correct_chunks,
+            **kwargs)
+
+        self.metrics.extend([precision, recall, f1_score])
+
+    def eval(self, executor, eval_program=None):
+        if eval_program is None:
+            eval_program = Program()
+        block = eval_program.current_block()
+        kwargs = {'main_program': eval_program}
+        num_infer_chunks, num_label_chunks, num_correct_chunks = executor.run(
+            eval_program,
+            fetch_list=[_clone_var_(block, state) for state in self.states])
+        num_infer_chunks = num_infer_chunks[0]
+        num_label_chunks = num_label_chunks[0]
+        num_correct_chunks = num_correct_chunks[0]
+        precision = float(
+            num_correct_chunks) / num_infer_chunks if num_infer_chunks else 0
+        recall = float(
+            num_correct_chunks) / num_label_chunks if num_label_chunks else 0
+        f1_score = float(2 * precision * recall) / (
+            precision + recall) if num_correct_chunks else 0
+        return np.array(
+            [precision], dtype='float32'), np.array(
+                [recall], dtype='float32'), np.array(
+                    [f1_score], dtype='float32')
diff --git a/python/paddle/v2/fluid/layers.py b/python/paddle/v2/fluid/layers.py
index 98a04ea9c2..809534884b 100644
--- a/python/paddle/v2/fluid/layers.py
+++ b/python/paddle/v2/fluid/layers.py
@@ -640,8 +640,8 @@ def chunk_eval(input,
                excluded_chunk_types=None,
                **kwargs):
     """
-    This function computes the accuracy using the input and label.
-    The output is the top_k inputs and their indices.
+    This function computes and outputs the precision, recall and 
+    F1-score of chunk detection.
     """
     helper = LayerHelper("chunk_eval", **kwargs)
 
@@ -649,6 +649,9 @@ def chunk_eval(input,
     precision = helper.create_tmp_variable(dtype="float32")
     recall = helper.create_tmp_variable(dtype="float32")
     f1_score = helper.create_tmp_variable(dtype="float32")
+    num_infer_chunks = helper.create_tmp_variable(dtype="int64")
+    num_label_chunks = helper.create_tmp_variable(dtype="int64")
+    num_correct_chunks = helper.create_tmp_variable(dtype="int64")
 
     helper.append_op(
         type="chunk_eval",
@@ -657,14 +660,17 @@ def chunk_eval(input,
         outputs={
             "Precision": [precision],
             "Recall": [recall],
-            "F1-Score": [f1_score]
+            "F1-Score": [f1_score],
+            "NumInferChunks": [num_infer_chunks],
+            "NumLabelChunks": [num_label_chunks],
+            "NumCorrectChunks": [num_correct_chunks]
         },
         attrs={
             "num_chunk_types": num_chunk_types,
             'chunk_scheme': chunk_scheme,
             'excluded_chunk_types': excluded_chunk_types or []
         })
-    return precision, recall, f1_score
+    return precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks
 
 
 def sequence_conv(input,
diff --git a/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py b/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py
index d2693b602e..caa51b5df4 100644
--- a/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py
+++ b/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py
@@ -150,7 +150,7 @@ def main():
     crf_decode = fluid.layers.crf_decoding(
         input=feature_out, param_attr=fluid.ParamAttr(name='crfw'))
 
-    precision, recall, f1_score = fluid.layers.chunk_eval(
+    chunk_evaluator = fluid.evaluator.ChunkEvaluator(
         input=crf_decode,
         label=target,
         chunk_scheme="IOB",
@@ -176,14 +176,16 @@ def main():
 
     batch_id = 0
     for pass_id in xrange(PASS_NUM):
+        chunk_evaluator.reset(exe)
         for data in train_data():
             outs = exe.run(fluid.default_main_program(),
                            feed=feeder.feed(data),
-                           fetch_list=[avg_cost, precision, recall, f1_score])
+                           fetch_list=[avg_cost] + chunk_evaluator.metrics)
+            precision, recall, f1_score = chunk_evaluator.eval(exe)
             avg_cost_val = np.array(outs[0])
-            precision_val = np.array(outs[1])
-            recall_val = np.array(outs[2])
-            f1_score_val = np.array(outs[3])
+            precision_val = np.array(precision)
+            recall_val = np.array(recall)
+            f1_score_val = np.array(f1_score)
 
             if batch_id % 10 == 0:
                 print("avg_cost=" + str(avg_cost_val))
diff --git a/python/paddle/v2/fluid/tests/test_chunk_eval_op.py b/python/paddle/v2/fluid/tests/test_chunk_eval_op.py
index 819e65a653..53bf6f815b 100644
--- a/python/paddle/v2/fluid/tests/test_chunk_eval_op.py
+++ b/python/paddle/v2/fluid/tests/test_chunk_eval_op.py
@@ -147,7 +147,13 @@ class TestChunkEvalOp(OpTest):
             'Recall': np.asarray(
                 [recall], dtype='float32'),
             'F1-Score': np.asarray(
-                [f1], dtype='float32')
+                [f1], dtype='float32'),
+            'NumInferChunks': np.asarray(
+                [self.num_infer_chunks], dtype='int64'),
+            'NumLabelChunks': np.asarray(
+                [self.num_label_chunks], dtype='int64'),
+            'NumCorrectChunks': np.asarray(
+                [self.num_correct_chunks], dtype='int64')
         }
 
     def setUp(self):

From 8f7d0b18145c8310fb36ad017aa0f1d4bce98c65 Mon Sep 17 00:00:00 2001
From: fengjiayi <fengjiayi@baidu.com>
Date: Tue, 12 Dec 2017 15:54:38 +0800
Subject: [PATCH 35/94] add param_attr for nets (#6509)

---
 python/paddle/v2/fluid/layers.py | 6 ++++--
 python/paddle/v2/fluid/nets.py   | 7 +++++++
 2 files changed, 11 insertions(+), 2 deletions(-)

diff --git a/python/paddle/v2/fluid/layers.py b/python/paddle/v2/fluid/layers.py
index fd8a2ed18c..1f45487902 100644
--- a/python/paddle/v2/fluid/layers.py
+++ b/python/paddle/v2/fluid/layers.py
@@ -1732,8 +1732,10 @@ def conv2d_transpose(input,
 
         h_in = input.shape[2]
         w_in = input.shape[3]
-        filter_size_h = output_size[0] - (h_in - 1) * stride[0] + 2 * padding[0]
-        filter_size_w = output_size[1] - (w_in - 1) * stride[1] + 2 * padding[1]
+        filter_size_h = output_size[0] - \
+            (h_in - 1) * stride[0] + 2 * padding[0]
+        filter_size_w = output_size[1] - \
+            (w_in - 1) * stride[1] + 2 * padding[1]
         filter_size = [filter_size_h, filter_size_w]
     elif isinstance(filter_size, int):
         filter_size = [filter_size, filter_size]
diff --git a/python/paddle/v2/fluid/nets.py b/python/paddle/v2/fluid/nets.py
index 05728ad75a..7ef524318e 100644
--- a/python/paddle/v2/fluid/nets.py
+++ b/python/paddle/v2/fluid/nets.py
@@ -9,6 +9,7 @@ def simple_img_conv_pool(input,
                          pool_size,
                          pool_stride,
                          act,
+                         param_attr=None,
                          pool_type='max',
                          main_program=None,
                          startup_program=None):
@@ -16,6 +17,7 @@ def simple_img_conv_pool(input,
         input=input,
         num_filters=num_filters,
         filter_size=filter_size,
+        param_attr=param_attr,
         act=act,
         main_program=main_program,
         startup_program=startup_program)
@@ -36,6 +38,7 @@ def img_conv_group(input,
                    conv_padding=1,
                    conv_filter_size=3,
                    conv_act=None,
+                   param_attr=None,
                    conv_with_batchnorm=False,
                    conv_batchnorm_drop_rate=None,
                    pool_stride=1,
@@ -57,6 +60,7 @@ def img_conv_group(input,
 
     conv_padding = __extend_list__(conv_padding)
     conv_filter_size = __extend_list__(conv_filter_size)
+    param_attr = __extend_list__(param_attr)
     conv_with_batchnorm = __extend_list__(conv_with_batchnorm)
     conv_batchnorm_drop_rate = __extend_list__(conv_batchnorm_drop_rate)
 
@@ -70,6 +74,7 @@ def img_conv_group(input,
             num_filters=conv_num_filter[i],
             filter_size=conv_filter_size[i],
             padding=conv_padding[i],
+            param_attr=param_attr[i],
             act=local_conv_act,
             main_program=main_program,
             startup_program=startup_program)
@@ -101,6 +106,7 @@ def img_conv_group(input,
 def sequence_conv_pool(input,
                        num_filters,
                        filter_size,
+                       param_attr=None,
                        act="sigmoid",
                        pool_type="max",
                        main_program=None,
@@ -109,6 +115,7 @@ def sequence_conv_pool(input,
         input=input,
         num_filters=num_filters,
         filter_size=filter_size,
+        param_attr=param_attr,
         act=act,
         main_program=main_program,
         startup_program=startup_program)

From e5dcefc4d3bf666627dac6047f71b99415c23a5a Mon Sep 17 00:00:00 2001
From: Luo Tao <luotao02@baidu.com>
Date: Tue, 12 Dec 2017 18:28:56 +0800
Subject: [PATCH 36/94] remove ATLAS library

---
 README.md                                |  2 +-
 cmake/cblas.cmake                        | 38 +-----------------------
 paddle/math/tests/test_matrixCompare.cpp |  2 +-
 3 files changed, 3 insertions(+), 39 deletions(-)

diff --git a/README.md b/README.md
index db0fbd88b2..bbb2d49858 100644
--- a/README.md
+++ b/README.md
@@ -36,7 +36,7 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl
     examples:
 
       - Optimized math operations through SSE/AVX intrinsics, BLAS libraries
-      (e.g. MKL, ATLAS, cuBLAS) or customized CPU/GPU kernels.
+      (e.g. MKL, OpenBLAS, cuBLAS) or customized CPU/GPU kernels.
       - Highly optimized recurrent networks which can handle **variable-length**
       sequence without padding.
       - Optimized local and distributed training for models with high dimensional
diff --git a/cmake/cblas.cmake b/cmake/cblas.cmake
index 13294c0548..6320b17520 100644
--- a/cmake/cblas.cmake
+++ b/cmake/cblas.cmake
@@ -3,7 +3,7 @@
 # It will search MKLML, atlas, OpenBlas, reference-cblas in order.
 #
 # If any cblas implementation found, the following variable will be set.
-#    CBLAS_PROVIDER  # one of MKLML, ATLAS, OPENBLAS, REFERENCE
+#    CBLAS_PROVIDER  # one of MKLML, OPENBLAS, REFERENCE
 #    CBLAS_INC_DIR   # the include directory for cblas.
 #    CBLAS_LIBS      # a list of libraries should be linked by paddle.
 #                    # Each library should be full path to object file.
@@ -25,42 +25,6 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB)
   return()
 endif()
 
-## Then find atlas.
-set(ATLAS_ROOT $ENV{ATLAS_ROOT} CACHE PATH "Folder contains Atlas")
-set(ATLAS_INCLUDE_SEARCH_PATHS
-        ${ATLAS_ROOT}/include
-        /usr/include
-        /usr/include/atlas)
-set(ATLAS_LIB_SEARCH_PATHS
-        ${ATLAS_ROOT}/lib
-        /usr/lib
-        /usr/lib/blas/atlas
-        /usr/lib/atlas
-        /usr/lib/atlas-base   # special for ubuntu 14.04.
-    )
-find_path(ATLAS_INC_DIR NAMES cblas.h
-  PATHS ${ATLAS_INCLUDE_SEARCH_PATHS})
-find_path(ATLAS_CLAPACK_INC_DIR NAMES clapack.h
-  PATHS ${ATLAS_INCLUDE_SEARCH_PATHS})
-find_library(ATLAS_CBLAS_LIB NAMES cblas libcblas.so.3
-  PATHS ${ATLAS_LIB_SEARCH_PATHS})
-find_library(ATLAS_CLAPACK_LIB NAMES lapack_atlas liblapack_atlas.so.3
-  PATHS ${ATLAS_LIB_SEARCH_PATHS})
-
-if(ATLAS_CLAPACK_INC_DIR AND ATLAS_INC_DIR AND ATLAS_CBLAS_LIB AND ATLAS_CLAPACK_LIB)
-  set(CBLAS_FOUND ON)
-  set(CBLAS_PROVIDER ATLAS)
-  set(CBLAS_INC_DIR ${ATLAS_INC_DIR} ${ATLAS_CLAPACK_INC_DIR})
-  set(CBLAS_LIBRARIES ${ATLAS_CLAPACK_LIB} ${ATLAS_CBLAS_LIB})
-
-  add_definitions(-DPADDLE_USE_ATLAS)
-  add_definitions(-DLAPACK_FOUND)
-
-  message(STATUS "Found ATLAS (include: ${ATLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
-  message(STATUS "Found lapack in ATLAS (include: ${ATLAS_CLAPACK_INC_DIR})")
-  return()
-endif()
-
 ## Then find openblas.
 set(OPENBLAS_ROOT $ENV{OPENBLAS_ROOT} CACHE PATH "Folder contains Openblas")
 set(OPENBLAS_INCLUDE_SEARCH_PATHS
diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp
index 7e5a1db44a..afb8d9d599 100644
--- a/paddle/math/tests/test_matrixCompare.cpp
+++ b/paddle/math/tests/test_matrixCompare.cpp
@@ -244,7 +244,7 @@ TEST(Matrix, unary) {
     LOG(WARNING) << "This version of PaddlePaddle was not built with LAPACK"
                  << "support so we cannot test matrix inverse. To test "
                  << "matrix inverse, please install LAPACKE "
-                 << "and MKL/Openblas/ATLAS, and re-build PaddlePaddle.";
+                 << "and MKL/Openblas, and re-build PaddlePaddle.";
 #endif
   }
 }

From 6ecf08b17395944febedd9867c48000a9d7a9f9f Mon Sep 17 00:00:00 2001
From: guosheng <guosheng@baidu.com>
Date: Tue, 12 Dec 2017 19:39:07 +0800
Subject: [PATCH 37/94] Enhance init_from_tar to support indicating parameters
 excluded from the initialized model

---
 python/paddle/v2/parameters.py | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/python/paddle/v2/parameters.py b/python/paddle/v2/parameters.py
index bd97dc1199..0dd7d38d8e 100644
--- a/python/paddle/v2/parameters.py
+++ b/python/paddle/v2/parameters.py
@@ -383,19 +383,21 @@ class Parameters(object):
             params.deserialize(param_name, f)
         return params
 
-    def init_from_tar(self, f):
+    def init_from_tar(self, f, exclude_params=[]):
         """
         Different from `from_tar`, this interface can be used to
         init partial network parameters from another saved model.
 
         :param f: the initialized model file.
         :type f: tar file
+        :param exclude_params: the names of parameters that shouldn't be initialized from the model file.
+        :type exclude_params: list of strings
         :return: Nothing.
         """
 
         tar_param = Parameters.from_tar(f)
         for pname in tar_param.names():
-            if pname in self.names():
+            if pname in self.names() and pname not in exclude_params:
                 self.set(pname, tar_param.get(pname))
 
 

From 63ce906b088c641c3bf33a2b8aa6324a39310ffe Mon Sep 17 00:00:00 2001
From: guosheng <guosheng@baidu.com>
Date: Tue, 12 Dec 2017 20:33:44 +0800
Subject: [PATCH 38/94] Refine ChunkEvalutor by following comments

---
 paddle/operators/chunk_eval_op.cc             | 16 +++++++------
 .../tests/book/test_label_semantic_roles.py   | 23 +++++++++----------
 2 files changed, 20 insertions(+), 19 deletions(-)

diff --git a/paddle/operators/chunk_eval_op.cc b/paddle/operators/chunk_eval_op.cc
index ff2a08ccac..894f355deb 100644
--- a/paddle/operators/chunk_eval_op.cc
+++ b/paddle/operators/chunk_eval_op.cc
@@ -80,14 +80,16 @@ class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker {
               "sensitivity) of chunks on the given mini-batch.");
     AddOutput("F1-Score",
               "(float). The evaluated F1-Score on the given mini-batch.");
+    AddOutput("NumInferChunks",
+              "(int64_t). The number of chunks in Inference on the given "
+              "mini-batch.");
     AddOutput(
-        "NumInferChunks",
-        "(int). The number of chunks in Inference on the given mini-batch.");
-    AddOutput("NumLabelChunks",
-              "(int). The number of chunks in Label on the given mini-batch.");
-    AddOutput("NumCorrectChunks",
-              "(int). The number of chunks both in Inference and Label on the "
-              "given mini-batch.");
+        "NumLabelChunks",
+        "(int64_t). The number of chunks in Label on the given mini-batch.");
+    AddOutput(
+        "NumCorrectChunks",
+        "(int64_t). The number of chunks both in Inference and Label on the "
+        "given mini-batch.");
     AddAttr<int>("num_chunk_types",
                  "(int). The number of chunk type. See below for details.");
     AddAttr<std::string>(
diff --git a/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py b/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py
index caa51b5df4..c3591a613a 100644
--- a/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py
+++ b/python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py
@@ -178,20 +178,19 @@ def main():
     for pass_id in xrange(PASS_NUM):
         chunk_evaluator.reset(exe)
         for data in train_data():
-            outs = exe.run(fluid.default_main_program(),
-                           feed=feeder.feed(data),
-                           fetch_list=[avg_cost] + chunk_evaluator.metrics)
-            precision, recall, f1_score = chunk_evaluator.eval(exe)
-            avg_cost_val = np.array(outs[0])
-            precision_val = np.array(precision)
-            recall_val = np.array(recall)
-            f1_score_val = np.array(f1_score)
+            cost, precision, recall, f1_score = exe.run(
+                fluid.default_main_program(),
+                feed=feeder.feed(data),
+                fetch_list=[avg_cost] + chunk_evaluator.metrics)
+            pass_precision, pass_recall, pass_f1_score = chunk_evaluator.eval(
+                exe)
 
             if batch_id % 10 == 0:
-                print("avg_cost=" + str(avg_cost_val))
-                print("precision_val=" + str(precision_val))
-                print("recall_val:" + str(recall_val))
-                print("f1_score_val:" + str(f1_score_val))
+                print("avg_cost:" + str(cost) + " precision:" + str(
+                    precision) + " recall:" + str(recall) + " f1_score:" + str(
+                        f1_score) + " pass_precision:" + str(
+                            pass_precision) + " pass_recall:" + str(pass_recall)
+                      + " pass_f1_score:" + str(pass_f1_score))
 
             # exit early for CI
             exit(0)

From 3ef8ec37bb80427ab8abe23384c42a9e9056c87d Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Wed, 13 Dec 2017 10:35:57 +0800
Subject: [PATCH 39/94] fix the new_op doc (#6540)

* fix the ending symbol

* fix invalid content link
---
 doc/howto/dev/new_op_cn.md | 36 ++++++++++++++++++++----------------
 doc/howto/dev/new_op_en.md | 19 ++++++++++---------
 2 files changed, 30 insertions(+), 25 deletions(-)

diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md
index 44dbeecbbd..757a5840bc 100644
--- a/doc/howto/dev/new_op_cn.md
+++ b/doc/howto/dev/new_op_cn.md
@@ -1,17 +1,18 @@
 # 如何写新的Operator
 
  - [概念简介](#概念简介)
- - [实现C++类](#实现C++类)
-   - [定义ProtoMaker类](#定义ProtoMaker类)
-   - [定义Operator类](#定义Operator类)
-   - [定义OpKernel类](#定义OpKernel类)
-   - [注册Operator](#注册Operator)
+ - [实现C++类](#实现c类)
+   - [定义ProtoMaker类](#定义protomaker类)
+   - [定义Operator类](#定义operator类)
+   - [定义OpKernel类](#定义opkernel类)
+   - [注册Operator](#注册operator)
    - [编译](#编译)
- - [绑定Python](#绑定Python)
+ - [绑定Python](#绑定python)
  - [实现单元测试](#实现单元测试)
-   - [前向Operator单测](#前向Operator单测)
-   - [反向Operator单测](#反向Operator单测)
+   - [前向Operator单测](#前向operator单测)
+   - [反向Operator单测](#反向operator单测)
    - [编译和执行](#编译和执行)
+ - [注意事项](#注意事项)
 
 
 ## 概念简介
@@ -43,7 +44,7 @@ Kernel实现       | CPU、CUDA共享Kernel实现在`.h`文件中，否则，CPU
 ## 实现C++类
 
 
-### 1. 定义ProtoMaker类
+### 定义ProtoMaker类
 
 矩阵乘法的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。
 
@@ -100,7 +101,7 @@ The equation is: Out = scale*X
 - `AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。
 
 
-### 2. 定义Operator类
+### 定义Operator类
 
 下面的点实现了MulOp的定义：
 
@@ -149,7 +150,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
 通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和下面将要介绍的注册函数一起放在`.cc`中
 
-### 3. 定义OpKernel类
+### 定义OpKernel类
 
 `MulKernel`继承自`framework::OpKernel`，带有下面两个模板参数:
 
@@ -177,6 +178,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
     math::matmul<DeviceContext, T>(*X, false, *Y, false, 1, Z, 0, device_context);
   }
   };
+  ```
 
 需要注意：**不同设备(CPU、CUDA)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。**
 
@@ -188,7 +190,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 到此，前向Op实现完成。接下来，需要在`.cc`文件中注册该op和kernel。
 反向Op类的定义，反向OpKernel的定义与前向Op类似，这里不再赘述。**但需注意反向Op没有`ProtoMaker`**。
 
-### 4. 注册Operator
+### 注册Operator
 
 - 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。
 
@@ -220,7 +222,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
                            ops::MulGradKernel<paddle::platform::CUDADeviceContext, float>);
     ```
 
-### 5. 编译
+### 编译
 
 运行下面命令可以进行编译：
 
@@ -236,6 +238,7 @@ make mul_op
 
 单测包括对比前向Op不同设备(CPU、CUDA)的实现、对比反向OP不同设备(CPU、CUDA)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
 
+### 前向Operator单测
 
 Op单元测试继承自`OpTest`。各项更加具体的单元测试在`TestMulOp`里完成。测试Operator，需要：
 
@@ -273,8 +276,7 @@ Op单元测试继承自`OpTest`。各项更加具体的单元测试在`TestMulOp
       def test_check_grad_ingore_y(self):
           self.check_grad(
               ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
-
-    ```
+  ```
 
 上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释：
 
@@ -282,6 +284,8 @@ Op单元测试继承自`OpTest`。各项更加具体的单元测试在`TestMulOp
 - `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。
 - `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。
 
+### 反向operator单测
+
 而反向测试中：
 - `test_check_grad_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。
   - 第一个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。
@@ -290,7 +294,7 @@ Op单元测试继承自`OpTest`。各项更加具体的单元测试在`TestMulOp
 - `test_check_grad_ingore_x`和`test_check_grad_ingore_y`分支用来测试只需要计算一个输入梯度的情况。
 
 
-### 编译和执行单元测试
+### 编译和执行
 
 `python/paddle/v2/framework/tests` 目录下新增的 `test_*.py` 单元测试会被自动加入工程进行编译。
 
diff --git a/doc/howto/dev/new_op_en.md b/doc/howto/dev/new_op_en.md
index 510233306c..fe86936bc1 100644
--- a/doc/howto/dev/new_op_en.md
+++ b/doc/howto/dev/new_op_en.md
@@ -1,8 +1,8 @@
 # How to write a new operator
 
  - [Background](#background)
- - [Implementing C++ Types](#implementing-c++-types)
-   - [Defining ProtoMaker](#defining-protoMaker)
+ - [Implementing C++ Types](#implementing-c-types)
+   - [Defining ProtoMaker](#defining-protomaker)
    - [Defining Operator](#defining-operator)
    - [Registering Operator](#registering-operator)
    - [Compilation](#compilation)
@@ -41,7 +41,7 @@ Let's take matrix multiplication operator, [MulOp](https://github.com/PaddlePadd
 ## Implementing C++ Types
 
 
-### 1. Defining Class ProtoMaker
+### Defining ProtoMaker
 
 Matrix Multiplication can be written as $Out = X * Y$, meaning that the operation consists of two inputs and pne output.
 
@@ -98,7 +98,7 @@ There are two changes in this example:
 - `AddAttr<AttrType>("scale", "...").SetDefault(1.0);`  adds `scale`constant as an attribute, and sets the default value to 1.0.
 
 
-### 2. Defining Operator
+### Defining Operator
 
 The following code defines the interface for MulOp:
 
@@ -147,7 +147,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
 Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, which also include the registration methods introduced later.
 
-### 3. Defining OpKernel
+### Defining OpKernel
 
 `MulKernel` inherits `framework::OpKernel`, which includes the following templates:
 
@@ -188,7 +188,7 @@ This concludes the forward implementation of an operator. Next its operation and
 
 The definition of its corresponding backward operator, if applicable, is similar to that of an forward operator. **Note that a backward operator does not include a `ProtoMaker`**.
 
-### 4. Registering Operator
+### Registering Operator
 
 - In `.cc` files, register forward and backward operator classes and the CPU kernel.
 
@@ -220,7 +220,7 @@ The definition of its corresponding backward operator, if applicable, is similar
                            ops::MulGradKernel<paddle::platform::CUDADeviceContext, float>);
     ```
 
-### 5. Compilation
+### Compilation
 
 Run the following commands to compile.
 
@@ -284,8 +284,7 @@ A forward operator unit test inherits `unittest.TestCase` and defines metaclass
       def test_check_grad_ingore_y(self):
           self.check_grad(
               ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
-
-    ```
+  ```
 Get its output, and compare it with the forward operator's own output.
 
 The code above first loads required packages. In addition, we have
@@ -294,6 +293,8 @@ The code above first loads required packages. In addition, we have
 - `self.inputs` defines input, with type `numpy.array` and initializes it.
 - `self.outputs` defines output and completes the same operator computation in the Python script, and returns its result from the Python script.
 
+### Testing Backward Operators
+
 Some key points in checking gradient above include:
 
 - `test_normal` calls `check_grad` to validate scaling tests' correctness and stability through numeric methods.

From 8ad36cdb5d2c3a79c82c36cd7c2e79bc2d4cc4bf Mon Sep 17 00:00:00 2001
From: "Yang Yang(Tony)" <yangyang62@baidu.com>
Date: Wed, 13 Dec 2017 10:41:31 +0800
Subject: [PATCH 40/94] PaddlePaddle Fluid Source Overview (#6485)

* first commit

* Update read_source.md
---
 doc/howto/read_source.md | 67 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 67 insertions(+)
 create mode 100644 doc/howto/read_source.md

diff --git a/doc/howto/read_source.md b/doc/howto/read_source.md
new file mode 100644
index 0000000000..383acb0c82
--- /dev/null
+++ b/doc/howto/read_source.md
@@ -0,0 +1,67 @@
+# PaddlePaddle Fluid Source Code Overview
+
+Examples: https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/v2/fluid/tests/book
+
+Core: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework
+
+Operator: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators
+
+Optimizer: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/optimizer
+
+Memory: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory
+
+# Compile Time
+
+The following **defines** the NN. The definition goes into this [protocol buffer](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto).
+
+```python
+x = fluid.layers.data(name='x', shape=[13], dtype='float32')
+y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+
+y_predict = fluid.layers.fc(input=x, size=1, act=None)
+cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+avg_cost = fluid.layers.mean(x=cost)
+
+sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
+sgd_optimizer.minimize(avg_cost)
+```
+
+- Variables: `x`,  `y`, `y_predict`, `cost` and `avg_cost`. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/framework.py#L93)
+- Layers: `fluid.layers.data`, `fluid.layers.fc` and `fluid.layers.mean` are layers. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/layers.py)
+  - Every Layer has one or more operators and variables/parameters
+    - All the operators are defined at [`paddle/operators/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators). Other worth-looking files:
+      - Base class: [`paddle/framework/operator.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h)
+      - Operator Registration: [`paddle/framework/op_registry.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h) 
+      - Operator Lookup: [`paddle/framework/op_info.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_info.h)
+- Optimizer: `fluid.optimizer.SGD`. It does the following
+  - Add backward operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/backward.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/backward.cc)]
+  - Add optimizer operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/optimizer.py), [C++](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/optimizer)]
+
+# Run Time
+
+The following **evaluates** the NN. Instantiates all the variables, operators.
+
+```python
+place = fluid.CPUPlace()
+feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
+exe = fluid.Executor(place)
+
+# Allocate memory. Initialize Parameter.
+exe.run(fluid.default_startup_program())
+
+# Allocate memory. Do computation.
+exe.run(fluid.default_main_program(),
+        feed=feeder.feed(data),
+        fetch_list=[avg_cost])
+```
+
+- Place: `place`. one of CPU, GPU or FPGA. [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h)
+  - The device handle are at [paddle/platform/device_context.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h)
+- Executor: `fluid.Executor(place)`. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/executor.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.cc)]
+  - Feeds the data: `feed=feeder.feed(data)`
+  - Evaluates all the operators
+  - Fetches the result: `fetch_list=[avg_cost]`
+- Other worth looking files:
+  - Scope: [paddle/framework/scope.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/scope.h). Where all the variables live
+    - Variable: [paddle/framework/variable.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h). Where all the data (most likely tensors) live
+      - Tensor: [paddle/framework/tensor.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h). Where we allocate memory through [`paddle/memory/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory)

From 697facc92f58e6b65c76db2f1b6efc282b3c57a0 Mon Sep 17 00:00:00 2001
From: dzhwinter <dongzhihong01@baidu.com>
Date: Wed, 13 Dec 2017 10:43:20 +0800
Subject: [PATCH 41/94] "add registry interface" (#6449)

* "add registry interface"

* "move function to registry"

* "rename with meaningful name"

* "add exposed layers"

* "fixed based on comments"

* "remove unsed comments"
---
 python/paddle/v2/fluid/layers.py              | 185 ++---------------
 python/paddle/v2/fluid/registry.py            | 186 ++++++++++++++++++
 python/paddle/v2/fluid/tests/test_registry.py |  22 +++
 3 files changed, 221 insertions(+), 172 deletions(-)
 create mode 100644 python/paddle/v2/fluid/registry.py
 create mode 100644 python/paddle/v2/fluid/tests/test_registry.py

diff --git a/python/paddle/v2/fluid/layers.py b/python/paddle/v2/fluid/layers.py
index 1f45487902..9f5a219b20 100644
--- a/python/paddle/v2/fluid/layers.py
+++ b/python/paddle/v2/fluid/layers.py
@@ -1,12 +1,12 @@
-import core
+import contextlib
+
 import proto.framework_pb2 as framework_pb2
+import core
 from framework import OpProtoHolder, Variable, Program, Operator
 from initializer import Constant, Normal, Xavier, Initializer
 from paddle.v2.fluid.layer_helper import LayerHelper, unique_name
-import re
-import cStringIO
+from registry import register_layer
 from param_attr import ParamAttr
-import contextlib
 
 __all__ = [
     'fc', 'data', 'cross_entropy', 'conv2d', 'pool2d', 'embedding', 'concat',
@@ -14,6 +14,15 @@ __all__ = [
     'batch_norm', 'accuracy', 'split_lod_tensor', 'While'
 ]
 
+_REGISTER_LAYER_FROM_OPS = [
+    'mean', 'mul', 'elementwise_add', 'elementwise_div', 'dropout', 'reshape',
+    'sigmoid', 'scale', 'transpose', 'sigmoid_cross_entropy_with_logits'
+]
+
+for _OP in set(_REGISTER_LAYER_FROM_OPS):
+    globals()[_OP] = register_layer(_OP)
+    __all__.append(_OP)
+
 
 def fc(input,
        size,
@@ -309,174 +318,6 @@ def create_tensor(dtype, name=None, main_program=None, startup_program=None):
     return helper.create_variable(name=helper.name, dtype=dtype)
 
 
-def _convert_(name):
-    """
-    Formatting.
-
-    Args:
-       name: The name/alias
-
-    This function takes in a name and converts it to a standard format of
-    group1_group2. Where as per the regular expression, group1 can have
-    alphabets and numbers and group2 has capital alphabets.
-
-    """
-    s1 = re.sub('(.)([A-Z][a-z]+)', r'\1_\2', name)
-    return re.sub('([a-z0-9])([A-Z])', r'\1_\2', s1).lower()
-
-
-def _generate_doc_string_(op_proto):
-    """
-    Generate docstring by OpProto
-
-    Args:
-        op_proto (framework_pb2.OpProto): a protobuf message typed OpProto
-
-    Returns:
-        str: the document string
-    """
-
-    def _type_to_str_(tp):
-        return framework_pb2.AttrType.Name(tp)
-
-    if not isinstance(op_proto, framework_pb2.OpProto):
-        raise TypeError("OpProto should be `framework_pb2.OpProto`")
-
-    buf = cStringIO.StringIO()
-    buf.write(op_proto.comment)
-    buf.write('\nArgs:\n')
-    for each_input in op_proto.inputs:
-        line_begin = '    {0}: '.format(_convert_(each_input.name))
-        buf.write(line_begin)
-        buf.write(each_input.comment)
-        buf.write('\n')
-        buf.write(' ' * len(line_begin))
-        buf.write('Duplicable: ')
-        buf.write(str(each_input.duplicable))
-        buf.write('  Optional: ')
-        buf.write(str(each_input.dispensable))
-        buf.write('\n')
-
-    for each_attr in op_proto.attrs:
-        buf.write('    ')
-        buf.write(each_attr.name)
-        buf.write(' (')
-        buf.write(_type_to_str_(each_attr.type))
-        buf.write('): ')
-        buf.write(each_attr.comment)
-        buf.write('\n')
-
-    if len(op_proto.outputs) != 0:
-        buf.write('\nReturns:\n')
-        buf.write('    ')
-        for each_opt in op_proto.outputs:
-            if not each_opt.intermediate:
-                break
-        buf.write(each_opt.comment)
-
-    return buf.getvalue()
-
-
-def _create_op_func_(op_type):
-    """
-    Create an Operator for a Function.
-
-    Args:
-       op_type: The name of the operator to be created
-
-    This function takes in the operator type (sigmoid, mean , average etc) and
-    creates the operator functionality.
-
-    """
-    op_proto = OpProtoHolder.instance().get_op_proto(op_type)
-    not_intermediate_outputs = \
-        filter(lambda output: not output.intermediate, op_proto.outputs)
-    intermediate_outputs = \
-        filter(lambda output: output.intermediate, op_proto.outputs)
-
-    if len(not_intermediate_outputs) != 1:
-        raise ValueError("Only one non intermediate output operator can be",
-                         "automatically generated")
-
-    if not_intermediate_outputs[0].duplicable:
-        raise ValueError(
-            "Only non duplicable op can be automatically generated")
-
-    for output in intermediate_outputs:
-        if output.duplicable:
-            raise ValueError("The op can be automatically generated only when ",
-                             "all intermediate ops are not duplicable")
-
-    o_name = not_intermediate_outputs[0].name
-    intermediate_output_names = [output.name for output in intermediate_outputs]
-
-    def infer_and_check_dtype(op_proto, **kwargs):
-        """
-        This function performs the sanity check for dtype and
-        instance type.
-        """
-        dtype = None
-        for ipt in op_proto.inputs:
-            name = _convert_(ipt.name)
-            val = kwargs.pop(name, [])
-            if not isinstance(val, list) and not isinstance(val, tuple):
-                val = [val]
-            for each in val:
-                if not isinstance(each, Variable):
-                    raise ValueError("input of {0} must be variable".format(
-                        op_type))
-
-                if dtype is None:
-                    dtype = each.dtype
-                elif dtype != each.dtype:
-                    raise ValueError(
-                        "operator {0} must input same dtype. {1} vs {2}".format(
-                            op_type, dtype, each.dtype))
-
-        return dtype
-
-    def func(**kwargs):
-        helper = LayerHelper(op_type, **kwargs)
-
-        dtype = infer_and_check_dtype(op_proto, **kwargs)
-
-        inputs = dict()
-        for ipt in op_proto.inputs:
-            name = _convert_(ipt.name)
-            val = kwargs.pop(name, [])
-            if not isinstance(val, list) and not isinstance(val, tuple):
-                val = [val]
-            inputs[ipt.name] = val
-
-        outputs = dict()
-        out = helper.create_tmp_variable(dtype=dtype)
-        outputs[o_name] = [out]
-        for name in intermediate_output_names:
-            outputs[name] = [helper.create_tmp_variable(dtype=dtype)]
-        helper.append_op(
-            type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs)
-        return helper.append_activation(out)
-
-    func.__name__ = op_type
-    globals()[op_type] = func
-    func.__doc__ = _generate_doc_string_(op_proto)
-    global __all__
-    __all__.append(op_type)
-
-
-_create_op_func_('mean')
-_create_op_func_('mul')
-_create_op_func_('elementwise_add')
-_create_op_func_('elementwise_div')
-_create_op_func_('dropout')
-_create_op_func_('reshape')
-_create_op_func_('sigmoid')
-_create_op_func_('scale')
-_create_op_func_('reshape')
-_create_op_func_('transpose')
-_create_op_func_('sigmoid_cross_entropy_with_logits')
-
-
 def cast(x, dtype, main_program=None):
     """
     This function takes in the input with input_dtype
diff --git a/python/paddle/v2/fluid/registry.py b/python/paddle/v2/fluid/registry.py
new file mode 100644
index 0000000000..6f5dd365de
--- /dev/null
+++ b/python/paddle/v2/fluid/registry.py
@@ -0,0 +1,186 @@
+import re
+import cStringIO
+import warnings
+import functools
+import inspect
+
+import proto.framework_pb2 as framework_pb2
+from framework import OpProtoHolder, Variable, Program, Operator
+from paddle.v2.fluid.layer_helper import LayerHelper, unique_name
+
+__all__ = ['deprecated', 'register_layer']
+
+
+def _convert_(name):
+    """
+    Formatting.
+
+    Args:
+       name: The name/alias
+
+    This function takes in a name and converts it to a standard format of
+    group1_group2. Where as per the regular expression, group1 can have
+    alphabets and numbers and group2 has capital alphabets.
+
+    """
+    s1 = re.sub('(.)([A-Z][a-z]+)', r'\1_\2', name)
+    return re.sub('([a-z0-9])([A-Z])', r'\1_\2', s1).lower()
+
+
+def _generate_doc_string_(op_proto):
+    """
+    Generate docstring by OpProto
+
+    Args:
+        op_proto (framework_pb2.OpProto): a protobuf message typed OpProto
+
+    Returns:
+        str: the document string
+    """
+
+    def _type_to_str_(tp):
+        return framework_pb2.AttrType.Name(tp)
+
+    if not isinstance(op_proto, framework_pb2.OpProto):
+        raise TypeError("OpProto should be `framework_pb2.OpProto`")
+
+    buf = cStringIO.StringIO()
+    buf.write(op_proto.comment)
+    buf.write('\nArgs:\n')
+    for each_input in op_proto.inputs:
+        line_begin = '    {0}: '.format(_convert_(each_input.name))
+        buf.write(line_begin)
+        buf.write(each_input.comment)
+        buf.write('\n')
+        buf.write(' ' * len(line_begin))
+        buf.write('Duplicable: ')
+        buf.write(str(each_input.duplicable))
+        buf.write('  Optional: ')
+        buf.write(str(each_input.dispensable))
+        buf.write('\n')
+
+    for each_attr in op_proto.attrs:
+        buf.write('    ')
+        buf.write(each_attr.name)
+        buf.write(' (')
+        buf.write(_type_to_str_(each_attr.type))
+        buf.write('): ')
+        buf.write(each_attr.comment)
+        buf.write('\n')
+
+    if len(op_proto.outputs) != 0:
+        buf.write('\nReturns:\n')
+        buf.write('    ')
+        for each_opt in op_proto.outputs:
+            if not each_opt.intermediate:
+                break
+        buf.write(each_opt.comment)
+
+    return buf.getvalue()
+
+
+def register_layer(op_type):
+    """
+    Register an Python layer for an Operator
+
+    Args:
+       op_type: The name of the operator to be created
+
+    This function takes in the operator type (sigmoid, mean , average etc) and
+    creates the operator functionality.
+
+    """
+    op_proto = OpProtoHolder.instance().get_op_proto(op_type)
+    not_intermediate_outputs = \
+        filter(lambda output: not output.intermediate, op_proto.outputs)
+    intermediate_outputs = \
+        filter(lambda output: output.intermediate, op_proto.outputs)
+
+    if len(not_intermediate_outputs) != 1:
+        raise ValueError("Only one non intermediate output operator can be",
+                         "automatically generated")
+
+    if not_intermediate_outputs[0].duplicable:
+        raise ValueError(
+            "Only non duplicable op can be automatically generated")
+
+    for output in intermediate_outputs:
+        if output.duplicable:
+            raise ValueError("The op can be automatically generated only when ",
+                             "all intermediate ops are not duplicable")
+
+    o_name = not_intermediate_outputs[0].name
+    intermediate_output_names = [output.name for output in intermediate_outputs]
+
+    def infer_and_check_dtype(op_proto, **kwargs):
+        """
+        This function performs the sanity check for dtype and
+        instance type.
+        """
+        dtype = None
+        for ipt in op_proto.inputs:
+            name = _convert_(ipt.name)
+            val = kwargs.pop(name, [])
+            if not isinstance(val, list) and not isinstance(val, tuple):
+                val = [val]
+            for each in val:
+                if not isinstance(each, Variable):
+                    raise ValueError("input of {0} must be variable".format(
+                        op_type))
+
+                if dtype is None:
+                    dtype = each.dtype
+                elif dtype != each.dtype:
+                    raise ValueError(
+                        "operator {0} must input same dtype. {1} vs {2}".format(
+                            op_type, dtype, each.dtype))
+
+        return dtype
+
+    def func(**kwargs):
+        helper = LayerHelper(op_type, **kwargs)
+
+        dtype = infer_and_check_dtype(op_proto, **kwargs)
+
+        inputs = dict()
+        for ipt in op_proto.inputs:
+            name = _convert_(ipt.name)
+            val = kwargs.pop(name, [])
+            if not isinstance(val, list) and not isinstance(val, tuple):
+                val = [val]
+            inputs[ipt.name] = val
+
+        outputs = dict()
+        out = helper.create_tmp_variable(dtype=dtype)
+        outputs[o_name] = [out]
+        for name in intermediate_output_names:
+            outputs[name] = [helper.create_tmp_variable(dtype=dtype)]
+        helper.append_op(
+            type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs)
+        return helper.append_activation(out)
+
+    func.__name__ = op_type
+    func.__doc__ = _generate_doc_string_(op_proto)
+    return func
+
+
+def deprecated(func_or_class):
+    """
+    Deprecated warning decorator. It will result a warning message.
+    Should be used before class or function, member function
+    """
+
+    @functools.wraps(func)
+    def func_wrapper(*args, **kwargs):
+        """
+        Wrap func with deprecated warning
+        """
+        warnings.simplefilter('always', DeprecationWarning)  #turn off filter
+        warnings.warn(
+            "Call to deprecated function {}.".format(func.__name__),
+            category=DeprecationWarning,
+            stacklevel=2)
+        warnings.simplefilter('default', DeprecationWarning)  #reset filter
+        return func(*args, **kwargs)
+
+    return func_wrapper
diff --git a/python/paddle/v2/fluid/tests/test_registry.py b/python/paddle/v2/fluid/tests/test_registry.py
new file mode 100644
index 0000000000..f8328f31cf
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/test_registry.py
@@ -0,0 +1,22 @@
+import unittest
+import warnings
+
+import paddle.v2.fluid as fluid
+import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
+import paddle.v2.fluid.registry as registry
+
+
+class TestRegistry(unittest.TestCase):
+    def test_registry_layer(self):
+        self.layer_type = "mean"
+        program = framework.Program()
+
+        x = fluid.layers.data(name='X', shape=[10, 10], dtype='float32')
+        output = layers.mean(x)
+        place = fluid.CPUPlace()
+        exe = fluid.Executor(place)
+
+        X = np.random.random((10, 10)).astype("float32")
+        mean_out = exe.run(program, feed={"X": X}, fetch_list=[output])
+        self.assertAlmostEqual(np.mean(X), mean_out)

From 2cc34532437cb73c00cfe77facb63cb3578ae651 Mon Sep 17 00:00:00 2001
From: guosheng <guosheng@baidu.com>
Date: Wed, 13 Dec 2017 10:58:50 +0800
Subject: [PATCH 42/94] Refine the doc of parameters.init_from_tar

---
 python/paddle/v2/parameters.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/python/paddle/v2/parameters.py b/python/paddle/v2/parameters.py
index 0dd7d38d8e..7b7d1a1d16 100644
--- a/python/paddle/v2/parameters.py
+++ b/python/paddle/v2/parameters.py
@@ -390,7 +390,8 @@ class Parameters(object):
 
         :param f: the initialized model file.
         :type f: tar file
-        :param exclude_params: the names of parameters that shouldn't be initialized from the model file.
+        :param exclude_params: the names of parameters that should  
+            not be initialized from the model file.
         :type exclude_params: list of strings
         :return: Nothing.
         """

From ea093283e6cdc89bed0ffa0f2e534f9583765465 Mon Sep 17 00:00:00 2001
From: sweetsky0901 <work@yq01-idl-gpu-online20.yq01.baidu.com>
Date: Wed, 13 Dec 2017 12:39:29 +0800
Subject: [PATCH 43/94] for code review by zhaolong

---
 paddle/operators/spp_op.cc    | 24 +++++++----
 paddle/operators/spp_op.cu.cc | 11 ++---
 paddle/operators/spp_op.h     | 75 +++++++++++++++--------------------
 3 files changed, 55 insertions(+), 55 deletions(-)

diff --git a/paddle/operators/spp_op.cc b/paddle/operators/spp_op.cc
index 5e51e73ecc..c4bd4f5ab3 100644
--- a/paddle/operators/spp_op.cc
+++ b/paddle/operators/spp_op.cc
@@ -31,9 +31,15 @@ class SppOpMaker : public framework::OpProtoAndCheckerMaker {
               "M = C * H * W");
     AddAttr<int>("pyramid_height", "(int), multi level pooling");
     AddComment(R"DOC(
-        "Does spatial pyramid pooling on the input image by taking the max,
-        etc. within regions so that the result vector of different sized
-        images are of the same size
+        "With spatial pyramid pooling, the input image can
+        be of any sizes. This not only allows arbitrary aspect
+        ratios, but also allows arbitrary scales. We can resize
+        the input image to any scale (e.g., min(w, h)=180, 224,
+        ...) and apply the same deep network. When the
+        input image is at different scales, the network (with
+        the same filter sizes) will extract features at different
+        scales. The scales play important roles in traditional
+        methods.
         Input shape: $(N, C_{in}, H_{in}, W_{in})$
         Output shape: $(H_{out}, W_{out})$
         Where
@@ -41,6 +47,7 @@ class SppOpMaker : public framework::OpProtoAndCheckerMaker {
             H_{out} = N \\
             W_{out} = (((4^pyramid_height) - 1) / (4 - 1))$ * C_{in}
           $$
+        paper https://arxiv.org/pdf/1406.4729v4.pdf
         )DOC");
   }
 };
@@ -79,8 +86,9 @@ class SppOpGrad : public framework::OperatorWithKernel {
 
 namespace ops = paddle::operators;
 REGISTER_OP(spp, ops::SppOp, ops::SppOpMaker, spp_grad, ops::SppOpGrad);
-REGISTER_OP_CPU_KERNEL(spp, ops::SppKernel<paddle::platform::CPUPlace, float>,
-                       ops::SppKernel<paddle::platform::CPUPlace, double>);
-REGISTER_OP_CPU_KERNEL(spp_grad,
-                       ops::SppGradKernel<paddle::platform::CPUPlace, float>,
-                       ops::SppGradKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(
+    spp, ops::SppKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SppKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    spp_grad, ops::SppGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SppGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/spp_op.cu.cc b/paddle/operators/spp_op.cu.cc
index a7057907ce..761e4d6c4a 100644
--- a/paddle/operators/spp_op.cu.cc
+++ b/paddle/operators/spp_op.cu.cc
@@ -15,8 +15,9 @@ limitations under the License. */
 #include "paddle/operators/spp_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(spp, ops::SppKernel<paddle::platform::GPUPlace, float>,
-                       ops::SppKernel<paddle::platform::GPUPlace, double>);
-REGISTER_OP_GPU_KERNEL(spp_grad,
-                       ops::SppGradKernel<paddle::platform::GPUPlace, float>,
-                       ops::SppGradKernel<paddle::platform::GPUPlace, double>);
+REGISTER_OP_CUDA_KERNEL(
+    spp, ops::SppKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SppKernel<paddle::platform::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    spp_grad, ops::SppGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SppGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/spp_op.h b/paddle/operators/spp_op.h
index 0f2c43ee65..16510cb826 100644
--- a/paddle/operators/spp_op.h
+++ b/paddle/operators/spp_op.h
@@ -20,7 +20,7 @@ limitations under the License. */
 
 namespace paddle {
 namespace operators {
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SppKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -43,39 +43,32 @@ class SppKernel : public framework::OpKernel<T> {
       std::vector<int> paddings({padding_h, padding_w});
       // pooling output shape
       framework::Tensor out_level;
-      std::vector<int64_t> output_shape_vec({in_x->dims()[0], in_x->dims()[1]});
-      output_shape_vec.push_back(
-          (input_h - kernel_size_h + 2 * padding_h) / kernel_size_h + 1);
-      output_shape_vec.push_back(
-          (input_w - kernel_size_w + 2 * padding_w) / kernel_size_w + 1);
+      std::vector<int64_t> output_shape_vec(
+          {in_x->dims()[0], in_x->dims()[1], bins, bins});
       framework::DDim output_shape(framework::make_ddim(output_shape_vec));
       out_level.mutable_data<T>(output_shape, context.GetPlace());
       // pooling
-      math::Pool2dFunctor<Place, math::MaxPool<T>, T> pool_forward;
+      math::Pool2dFunctor<DeviceContext, math::MaxPool<T>, T> pool_forward;
       math::MaxPool<T> max_process;
-      pool_forward(context.device_context(), *in_x, kernel_size, strides,
-                   paddings, max_process, &out_level);
+      pool_forward(context.template device_context<DeviceContext>(), *in_x,
+                   kernel_size, strides, paddings, max_process, &out_level);
       // flatten pooling output shape
-      framework::Tensor out_flatten_level;
       int output_flatten_w = in_x->dims()[1] * bins * bins;
       std::vector<int64_t> output_flatten_shape_vec(
           {in_x->dims()[0], output_flatten_w});
       framework::DDim output_flatten_shape(
           framework::make_ddim(output_flatten_shape_vec));
-      out_flatten_level.ShareDataWith(out_level);
-      out_flatten_level.Resize(output_flatten_shape);
+      out_level.Resize(output_flatten_shape);
       // concat
-      auto out_flatten_level_stride =
-          framework::stride(out_flatten_level.dims());
-      StridedMemcpy<T>(context.device_context(), out_flatten_level.data<T>(),
-                       out_flatten_level_stride, out_flatten_level.dims(),
+      auto out_level_stride = framework::stride(out_level.dims());
+      StridedMemcpy<T>(context.template device_context<DeviceContext>(),
+                       out_level.data<T>(), out_level_stride, out_level.dims(),
                        out_stride, out->data<T>() + output_offset);
-      output_offset +=
-          out_flatten_level.dims()[1] * out_flatten_level_stride[1];
+      output_offset += out_level.dims()[1] * out_level_stride[1];
     }
   }
 };
-template <typename Place, typename T>
+template <typename DeviceContext, typename T>
 class SppGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -86,8 +79,8 @@ class SppGradKernel : public framework::OpKernel<T> {
     framework::Tensor* in_x_grad =
         context.Output<framework::Tensor>(framework::GradVarName("X"));
     int pyramid_height = context.template Attr<int>("pyramid_height");
-    auto& device_ctx = context.device_context();
-    math::SetConstant<Place, T> zero;
+    auto& device_ctx = context.template device_context<DeviceContext>();
+    math::SetConstant<DeviceContext, T> zero;
     in_x_grad->mutable_data<T>(context.GetPlace());
     zero(device_ctx, in_x_grad, static_cast<T>(0));
     auto out_stride = framework::stride(out->dims());
@@ -104,45 +97,43 @@ class SppGradKernel : public framework::OpKernel<T> {
       std::vector<int> strides({kernel_size_h, kernel_size_w});
       std::vector<int> paddings({padding_h, padding_w});
       // split out and outgrad  ...  to flatten
-      framework::Tensor out_flatten_level;
-      framework::Tensor outgrad_flatten_level;
+      framework::Tensor out_level;
+      framework::Tensor outgrad_level;
       int out_flatten_w = in_x->dims()[1] * bins * bins;
       std::vector<int64_t> out_flatten_shape_vec(
           {in_x->dims()[0], out_flatten_w});
       framework::DDim out_flatten_shape(
           framework::make_ddim(out_flatten_shape_vec));
-      out_flatten_level.mutable_data<T>(out_flatten_shape, context.GetPlace());
-      outgrad_flatten_level.mutable_data<T>(out_flatten_shape,
-                                            context.GetPlace());
-      auto flatten_stride = framework::stride(out_flatten_level.dims());
+      out_level.mutable_data<T>(out_flatten_shape, context.GetPlace());
+      outgrad_level.mutable_data<T>(out_flatten_shape, context.GetPlace());
+      auto flatten_stride = framework::stride(out_level.dims());
       // memcpy
-      StridedMemcpy<T>(context.device_context(), out->data<T>() + out_offset,
-                       out_stride, out_flatten_level.dims(), flatten_stride,
-                       out_flatten_level.data<T>());
+      StridedMemcpy<T>(context.template device_context<DeviceContext>(),
+                       out->data<T>() + out_offset, out_stride,
+                       out_level.dims(), flatten_stride, out_level.data<T>());
 
-      StridedMemcpy<T>(context.device_context(),
+      StridedMemcpy<T>(context.template device_context<DeviceContext>(),
                        out_grad->data<T>() + out_offset, out_stride,
-                       outgrad_flatten_level.dims(), flatten_stride,
-                       outgrad_flatten_level.data<T>());
-      out_offset += out_flatten_level.dims()[1] * out_stride[1];
+                       outgrad_level.dims(), flatten_stride,
+                       outgrad_level.data<T>());
+      out_offset += out_level.dims()[1] * out_stride[1];
       // flatten backward to nchw
-      framework::Tensor out_level;
-      framework::Tensor outgrad_level;
+
       std::vector<int64_t> out_shape_vec({in_x->dims()[0], in_x->dims()[1]});
       out_shape_vec.push_back(
           (input_h - kernel_size_h + 2 * padding_h) / kernel_size_h + 1);
       out_shape_vec.push_back(
           (input_w - kernel_size_w + 2 * padding_w) / kernel_size_w + 1);
       framework::DDim out_shape(framework::make_ddim(out_shape_vec));
-      out_level.ShareDataWith(out_flatten_level);
+      out_level.ShareDataWith(out_level);
       out_level.Resize(out_shape);
-      outgrad_level.ShareDataWith(outgrad_flatten_level);
+      outgrad_level.ShareDataWith(outgrad_level);
       outgrad_level.Resize(out_shape);
       // pooling backward
-      math::MaxPool2dGradFunctor<Place, T> pool2d_backward;
-      pool2d_backward(context.device_context(), *in_x, *&out_level,
-                      *&outgrad_level, kernel_size, strides, paddings,
-                      in_x_grad);
+      math::MaxPool2dGradFunctor<DeviceContext, T> pool2d_backward;
+      pool2d_backward(context.template device_context<DeviceContext>(), *in_x,
+                      *&out_level, *&outgrad_level, kernel_size, strides,
+                      paddings, in_x_grad);
     }
   }
 };

From 316ab4e07d0dd4fce3efa8f3b32380a829ffbd48 Mon Sep 17 00:00:00 2001
From: liaogang <liaogang@baidu.com>
Date: Wed, 13 Dec 2017 12:40:11 +0800
Subject: [PATCH 44/94] christmas gift: fix a typo

---
 paddle/scripts/docker/README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/paddle/scripts/docker/README.md b/paddle/scripts/docker/README.md
index 1e1fcc50dc..f0620498cf 100644
--- a/paddle/scripts/docker/README.md
+++ b/paddle/scripts/docker/README.md
@@ -20,7 +20,7 @@ binaries.
 
 ## Run The Build
 
-### Build Evironments
+### Build Environments
 
 The pre-built build environment images are:
 

From 1ba8f7fe71fc8df40669a24fc5af88d0904240f1 Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Wed, 13 Dec 2017 14:18:19 +0800
Subject: [PATCH 45/94] The comments in reshape_op is wrong (#6565)

---
 paddle/operators/reshape_op.cc | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc
index 39bf2118d6..7fd33bf662 100644
--- a/paddle/operators/reshape_op.cc
+++ b/paddle/operators/reshape_op.cc
@@ -84,9 +84,9 @@ Given a 2-D tensor X with 2 rows and 2 columns
     [[1, 2], [3, 4]]
 
 and target shape = [1, 4], the reshape operator will transform
-the tensor X into a 1-D tensor:
+the tensor X into a 2-D tensor:
 
-    [1, 2, 3, 4]
+    [[1, 2, 3, 4]]
 
 )DOC");
   }

From d069f2ca0a92dd07eb06c9a02528c41b0702f131 Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Wed, 13 Dec 2017 14:37:05 +0800
Subject: [PATCH 46/94] Make fluid.layers.fc support multiple param_attr
 (#6532)

Fix #6531
---
 python/paddle/v2/fluid/param_attr.py        | 2 ++
 python/paddle/v2/fluid/tests/test_layers.py | 5 ++++-
 2 files changed, 6 insertions(+), 1 deletion(-)

diff --git a/python/paddle/v2/fluid/param_attr.py b/python/paddle/v2/fluid/param_attr.py
index 86088fdd7c..7952a5ea51 100644
--- a/python/paddle/v2/fluid/param_attr.py
+++ b/python/paddle/v2/fluid/param_attr.py
@@ -36,6 +36,8 @@ class ParamAttr(object):
     def to_attr(arg):
         if arg is None:
             return ParamAttr()
+        elif isinstance(arg, list) or isinstance(arg, tuple):
+            return [ParamAttr.to_attr(a) for a in arg]
         elif isinstance(arg, ParamAttr):
             return arg
         elif isinstance(arg, str) or isinstance(arg, unicode):
diff --git a/python/paddle/v2/fluid/tests/test_layers.py b/python/paddle/v2/fluid/tests/test_layers.py
index 57f6a362de..9b88080158 100644
--- a/python/paddle/v2/fluid/tests/test_layers.py
+++ b/python/paddle/v2/fluid/tests/test_layers.py
@@ -29,7 +29,10 @@ class TestBook(unittest.TestCase):
             label = layers.data(name='label', shape=[1], dtype='int32')
             hidden1 = layers.fc(input=images, size=128, act='relu')
             hidden2 = layers.fc(input=hidden1, size=64, act='relu')
-            predict = layers.fc(input=hidden2, size=10, act='softmax')
+            predict = layers.fc(input=[hidden2, hidden1],
+                                size=10,
+                                act='softmax',
+                                param_attr=["sftmax.w1", "sftmax.w2"])
             cost = layers.cross_entropy(input=predict, label=label)
             avg_cost = layers.mean(x=cost)
             self.assertIsNotNone(avg_cost)

From 0e18bc88366e3987e1cf0dfef382725bf7391bf3 Mon Sep 17 00:00:00 2001
From: Luo Tao <luotao02@baidu.com>
Date: Wed, 13 Dec 2017 14:45:15 +0800
Subject: [PATCH 47/94] update paddledev to paddlepaddle

---
 doc/faq/build_and_install/index_cn.rst                        | 2 +-
 doc/getstarted/build_and_install/docker_install_cn.rst        | 4 ++--
 doc/getstarted/build_and_install/docker_install_en.rst        | 4 ++--
 .../cluster_train_v2/openmpi/docker_cluster/Dockerfile        | 2 +-
 paddle/scripts/tools/build_docs/build_docs.sh                 | 2 +-
 5 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/doc/faq/build_and_install/index_cn.rst b/doc/faq/build_and_install/index_cn.rst
index f1677e216f..a2bdeead78 100644
--- a/doc/faq/build_and_install/index_cn.rst
+++ b/doc/faq/build_and_install/index_cn.rst
@@ -14,7 +14,7 @@
 
     $ export CUDA_SO="$(\ls usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
     $ export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-    $ docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddlepaddle:latest-gpu
+    $ docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu
 
 更多关于Docker的安装与使用, 请参考 `PaddlePaddle Docker 文档 <http://www.paddlepaddle.org/doc_cn/build_and_install/install/docker_install.html>`_ 。
 
diff --git a/doc/getstarted/build_and_install/docker_install_cn.rst b/doc/getstarted/build_and_install/docker_install_cn.rst
index f78b1fb0e1..1eb06e4182 100644
--- a/doc/getstarted/build_and_install/docker_install_cn.rst
+++ b/doc/getstarted/build_and_install/docker_install_cn.rst
@@ -114,7 +114,7 @@ PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Note
 
   .. code-block:: bash
 
-     nvidia-docker run -it -v $PWD:/work paddledev/paddle:latest-gpu /bin/bash
+     nvidia-docker run -it -v $PWD:/work paddlepaddle/paddle:latest-gpu /bin/bash
 
 **注: 如果没有安装nvidia-docker，可以尝试以下的方法，将CUDA库和Linux设备挂载到Docker容器内：**
 
@@ -122,7 +122,7 @@ PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Note
 
      export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
      export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-     docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:latest-gpu
+     docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu
 
 **关于AVX：**
 
diff --git a/doc/getstarted/build_and_install/docker_install_en.rst b/doc/getstarted/build_and_install/docker_install_en.rst
index d7acc7aeb7..5a46c598f2 100644
--- a/doc/getstarted/build_and_install/docker_install_en.rst
+++ b/doc/getstarted/build_and_install/docker_install_en.rst
@@ -122,7 +122,7 @@ GPU driver installed before move on.
 
   .. code-block:: bash
 
-     nvidia-docker run -it -v $PWD:/work paddledev/paddle:latest-gpu /bin/bash
+     nvidia-docker run -it -v $PWD:/work paddlepaddle/paddle:latest-gpu /bin/bash
 
 **NOTE: If you don't have nvidia-docker installed, try the following method to mount CUDA libs and devices into the container.**
 
@@ -130,7 +130,7 @@ GPU driver installed before move on.
 
      export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
      export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-     docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:latest-gpu
+     docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu
 
 **About AVX:**
 
diff --git a/paddle/scripts/cluster_train_v2/openmpi/docker_cluster/Dockerfile b/paddle/scripts/cluster_train_v2/openmpi/docker_cluster/Dockerfile
index 1a2d19e823..c2f631bdf4 100644
--- a/paddle/scripts/cluster_train_v2/openmpi/docker_cluster/Dockerfile
+++ b/paddle/scripts/cluster_train_v2/openmpi/docker_cluster/Dockerfile
@@ -1,7 +1,7 @@
 # Build this image:  docker build -t mpi .
 #
 
-FROM paddledev/paddle:0.10.0rc3
+FROM paddlepaddle/paddle:0.10.0rc3
 
 ENV DEBIAN_FRONTEND noninteractive
 
diff --git a/paddle/scripts/tools/build_docs/build_docs.sh b/paddle/scripts/tools/build_docs/build_docs.sh
index c6cbbc4eef..f9bc8bf63a 100755
--- a/paddle/scripts/tools/build_docs/build_docs.sh
+++ b/paddle/scripts/tools/build_docs/build_docs.sh
@@ -5,4 +5,4 @@ docker run --rm \
        -e "WITH_AVX=ON" \
        -e "WITH_DOC=ON" \
        -e "WOBOQ=ON" \
-       ${1:-"paddledev/paddle:dev"}
+       ${1:-"paddlepaddle/paddle:latest-dev"}

From 0a8addf802e2198084b9cc66e49ca4ae2fdd8125 Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Wed, 13 Dec 2017 15:24:53 +0800
Subject: [PATCH 48/94] Make cast op support bool (#6562)

Also add `elemwise_sub/mul/abs/clip` layers
---
 paddle/operators/cast_op.cc      | 3 ++-
 paddle/operators/cast_op.cu      | 3 ++-
 python/paddle/v2/fluid/layers.py | 5 +++--
 3 files changed, 7 insertions(+), 4 deletions(-)

diff --git a/paddle/operators/cast_op.cc b/paddle/operators/cast_op.cc
index 42bff69a1e..d641b8fc9f 100644
--- a/paddle/operators/cast_op.cc
+++ b/paddle/operators/cast_op.cc
@@ -74,4 +74,5 @@ REGISTER_OP_WITH_KERNEL(cast, ops::CastOpGradMaker, ops::CastOpInferShape,
 REGISTER_OP_CPU_KERNEL(cast, ops::CastOpKernel<CPU, float>,
                        ops::CastOpKernel<CPU, double>,
                        ops::CastOpKernel<CPU, int>,
-                       ops::CastOpKernel<CPU, int64_t>);
+                       ops::CastOpKernel<CPU, int64_t>,
+                       ops::CastOpKernel<CPU, bool>);
diff --git a/paddle/operators/cast_op.cu b/paddle/operators/cast_op.cu
index 4681deaa62..91e6fb391c 100644
--- a/paddle/operators/cast_op.cu
+++ b/paddle/operators/cast_op.cu
@@ -19,4 +19,5 @@ using CastOpKernel =
     paddle::operators::CastOpKernel<paddle::platform::CUDADeviceContext, T>;
 
 REGISTER_OP_CUDA_KERNEL(cast, CastOpKernel<float>, CastOpKernel<double>,
-                        CastOpKernel<int>, CastOpKernel<int64_t>);
+                        CastOpKernel<int>, CastOpKernel<int64_t>,
+                        CastOpKernel<bool>);
diff --git a/python/paddle/v2/fluid/layers.py b/python/paddle/v2/fluid/layers.py
index 9f5a219b20..f67d6d08c7 100644
--- a/python/paddle/v2/fluid/layers.py
+++ b/python/paddle/v2/fluid/layers.py
@@ -15,8 +15,9 @@ __all__ = [
 ]
 
 _REGISTER_LAYER_FROM_OPS = [
-    'mean', 'mul', 'elementwise_add', 'elementwise_div', 'dropout', 'reshape',
-    'sigmoid', 'scale', 'transpose', 'sigmoid_cross_entropy_with_logits'
+    'mean', 'mul', 'dropout', 'reshape', 'sigmoid', 'scale', 'transpose',
+    'sigmoid_cross_entropy_with_logits', 'elementwise_add', 'elementwise_div',
+    'elementwise_sub', 'elementwise_mul', 'clip', 'abs'
 ]
 
 for _OP in set(_REGISTER_LAYER_FROM_OPS):

From 842b485f6a15e0ddba7ff345cb17d440cc950374 Mon Sep 17 00:00:00 2001
From: guosheng <guosheng@baidu.com>
Date: Wed, 13 Dec 2017 17:09:15 +0800
Subject: [PATCH 49/94] Enhance ReduceOp to support reducing over all elements

---
 paddle/operators/reduce_op.cc                 |  32 +++--
 paddle/operators/reduce_op.h                  | 119 ++++++++++++------
 .../paddle/v2/fluid/tests/test_reduce_op.py   |  14 +++
 3 files changed, 113 insertions(+), 52 deletions(-)

diff --git a/paddle/operators/reduce_op.cc b/paddle/operators/reduce_op.cc
index b754637bf2..fedc2a5c37 100644
--- a/paddle/operators/reduce_op.cc
+++ b/paddle/operators/reduce_op.cc
@@ -37,18 +37,23 @@ class ReduceOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_LT(
         dim, x_rank,
         "The dim should be in the range [-rank(input), rank(input)).");
-    bool keep_dim = ctx->Attrs().Get<bool>("keep_dim");
-    auto dims_vector = vectorize(x_dims);
-    if (keep_dim || x_rank == 1) {
-      dims_vector[dim] = 1;
+    bool reduce_all = ctx->Attrs().Get<bool>("reduce_all");
+    if (reduce_all) {
+      ctx->SetOutputDim("Out", {1});
     } else {
-      dims_vector.erase(dims_vector.begin() + dim);
-    }
-    auto out_dims = framework::make_ddim(dims_vector);
-    ctx->SetOutputDim("Out", out_dims);
-    if (dim != 0) {
-      // Only pass LoD when not reducing on the first dim.
-      ctx->ShareLoD("X", /*->*/ "Out");
+      bool keep_dim = ctx->Attrs().Get<bool>("keep_dim");
+      auto dims_vector = vectorize(x_dims);
+      if (keep_dim || x_rank == 1) {
+        dims_vector[dim] = 1;
+      } else {
+        dims_vector.erase(dims_vector.begin() + dim);
+      }
+      auto out_dims = framework::make_ddim(dims_vector);
+      ctx->SetOutputDim("Out", out_dims);
+      if (dim != 0) {
+        // Only pass LoD when not reducing on the first dim.
+        ctx->ShareLoD("X", /*->*/ "Out");
+      }
     }
   }
 };
@@ -95,11 +100,16 @@ class ReduceOpMaker : public framework::OpProtoAndCheckerMaker {
                   "(bool, default false) "
                   "If true, retain the reduced dimension with length 1.")
         .SetDefault(false);
+    AddAttr<bool>("reduce_all",
+                  "(bool, default false) "
+                  "If true, output a scalar reduced along all dimensions.")
+        .SetDefault(false);
     comment_ = R"DOC(
 {ReduceOp} Operator.
 
 This operator computes the {reduce} of input tensor along the given dimension. 
 The result tensor has 1 fewer dimension than the input unless keep_dim is true.
+If reduce_all is true, just reduce along all dimensions and output a scalar.
 
 )DOC";
     AddComment(comment_);
diff --git a/paddle/operators/reduce_op.h b/paddle/operators/reduce_op.h
index 47ce910f28..7bd99cb1e6 100644
--- a/paddle/operators/reduce_op.h
+++ b/paddle/operators/reduce_op.h
@@ -26,10 +26,12 @@ using DDim = framework::DDim;
 template <typename T, size_t D, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
-
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
 struct SumFunctor {
   template <typename DeviceContext, typename X, typename Y, typename Dim>
@@ -95,26 +97,41 @@ template <typename DeviceContext, typename T, typename Functor>
 class ReduceKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    int rank = context.Input<Tensor>("X")->dims().size();
-    switch (rank) {
-      case 1:
-        ReduceCompute<1>(context);
-        break;
-      case 2:
-        ReduceCompute<2>(context);
-        break;
-      case 3:
-        ReduceCompute<3>(context);
-        break;
-      case 4:
-        ReduceCompute<4>(context);
-        break;
-      case 5:
-        ReduceCompute<5>(context);
-        break;
-      case 6:
-        ReduceCompute<6>(context);
-        break;
+    bool reduce_all = context.Attr<bool>("reduce_all");
+    if (reduce_all) {
+      // Flatten and reduce 1-D tensor
+      auto* input = context.Input<Tensor>("X");
+      auto* output = context.Output<Tensor>("Out");
+      output->mutable_data<T>(context.GetPlace());
+      auto x = EigenVector<T>::Flatten(*input);
+      auto out = EigenScalar<T>::From(*output);
+      auto& place =
+          *context.template device_context<DeviceContext>().eigen_device();
+      auto reduce_dim = Eigen::array<int, 1>({{0}});
+      Functor functor;
+      functor(place, x, out, reduce_dim);
+    } else {
+      int rank = context.Input<Tensor>("X")->dims().size();
+      switch (rank) {
+        case 1:
+          ReduceCompute<1>(context);
+          break;
+        case 2:
+          ReduceCompute<2>(context);
+          break;
+        case 3:
+          ReduceCompute<3>(context);
+          break;
+        case 4:
+          ReduceCompute<4>(context);
+          break;
+        case 5:
+          ReduceCompute<5>(context);
+          break;
+        case 6:
+          ReduceCompute<6>(context);
+          break;
+      }
     }
   }
 
@@ -157,26 +174,46 @@ template <typename DeviceContext, typename T, typename Functor>
 class ReduceGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    int rank = context.Input<Tensor>("X")->dims().size();
-    switch (rank) {
-      case 1:
-        ReduceGradCompute<1>(context);
-        break;
-      case 2:
-        ReduceGradCompute<2>(context);
-        break;
-      case 3:
-        ReduceGradCompute<3>(context);
-        break;
-      case 4:
-        ReduceGradCompute<4>(context);
-        break;
-      case 5:
-        ReduceGradCompute<5>(context);
-        break;
-      case 6:
-        ReduceGradCompute<6>(context);
-        break;
+    bool reduce_all = context.Attr<bool>("reduce_all");
+    if (reduce_all) {
+      auto* input0 = context.Input<Tensor>("X");
+      auto* input1 = context.Input<Tensor>("Out");
+      auto* input2 = context.Input<Tensor>(framework::GradVarName("Out"));
+      auto* output = context.Output<Tensor>(framework::GradVarName("X"));
+      output->mutable_data<T>(context.GetPlace());
+      auto x = EigenVector<T>::Flatten(*input0);
+      auto x_reduce = EigenVector<T>::From(*input1);
+      auto x_reduce_grad = EigenVector<T>::From(*input2);
+      auto x_grad = EigenVector<T>::Flatten(*output);
+      auto& place =
+          *context.template device_context<DeviceContext>().eigen_device();
+      auto broadcast_dim =
+          Eigen::array<int, 1>({{static_cast<int>(input0->numel())}});
+      Functor functor;
+      functor(place, x, x_reduce, x_grad, x_reduce_grad, broadcast_dim,
+              broadcast_dim[0]);
+    } else {
+      int rank = context.Input<Tensor>("X")->dims().size();
+      switch (rank) {
+        case 1:
+          ReduceGradCompute<1>(context);
+          break;
+        case 2:
+          ReduceGradCompute<2>(context);
+          break;
+        case 3:
+          ReduceGradCompute<3>(context);
+          break;
+        case 4:
+          ReduceGradCompute<4>(context);
+          break;
+        case 5:
+          ReduceGradCompute<5>(context);
+          break;
+        case 6:
+          ReduceGradCompute<6>(context);
+          break;
+      }
     }
   }
 
diff --git a/python/paddle/v2/fluid/tests/test_reduce_op.py b/python/paddle/v2/fluid/tests/test_reduce_op.py
index 70359d60cb..a021d4dd91 100644
--- a/python/paddle/v2/fluid/tests/test_reduce_op.py
+++ b/python/paddle/v2/fluid/tests/test_reduce_op.py
@@ -85,5 +85,19 @@ class Test1DReduce(OpTest):
         self.check_grad(['X'], 'Out')
 
 
+class TestReduceAll(OpTest):
+    def setUp(self):
+        self.op_type = "reduce_sum"
+        self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float32")}
+        self.attrs = {'reduce_all': True}
+        self.outputs = {'Out': self.inputs['X'].sum()}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
 if __name__ == '__main__':
     unittest.main()

From fd12776c570207697b790bdbd44cf466fb1141d7 Mon Sep 17 00:00:00 2001
From: QI JUN <qijun1994@hotmail.com>
Date: Wed, 13 Dec 2017 21:16:50 +0800
Subject: [PATCH 50/94] Design doc: how to support a new device/library doc
 (#6577)

* add support new device doc

* add more info
---
 doc/design/support_new_device.md | 248 +++++++++++++++++++++++++++++++
 1 file changed, 248 insertions(+)
 create mode 100644 doc/design/support_new_device.md

diff --git a/doc/design/support_new_device.md b/doc/design/support_new_device.md
new file mode 100644
index 0000000000..92443e4392
--- /dev/null
+++ b/doc/design/support_new_device.md
@@ -0,0 +1,248 @@
+# Design Doc: Support new Device/Library
+
+## Background
+
+Deep learning has a high demand for computing resources. New high-performance device and computing library are coming constantly. The deep learning framework has to integrate these high-performance device and computing library flexibly.
+
+On the one hand, hardware and computing library are not usually one-to-one coresponding relations. For example, in Intel CPU, there are Eigen and MKL computing library. And in Nvidia GPU, there are Eigen and cuDNN computing library. We have to implement specific kernels for an operator for each computing library.
+
+On the other hand, users usually do not want to care about the low-level hardware and computing library when writing a neural network configuration. In Fluid, `Layer` is exposed in `Python`, and `Operator` is exposed in `C++`. Both `Layer` and `Operator` are independent on hardwares.
+
+So, how to support a new Device/Library in Fluid becomes a challenge.
+
+
+## Basic: Integrate A New Device/Library
+
+For a general overview of fluid, please refer to [overview doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/read_source.md).
+
+There are mainly there parts we have to consider in integrating a new device/library:
+
+- Place and DeviceContext: indicates the device id and manages hardware resources
+
+- Memory and Tensor: malloc/free data on certain device
+
+- Math Functor and OpKernel: implement computing unit on certain device/library
+
+### Place and DeviceContext
+
+
+#### Place
+Fluid use class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent specific device and computing library. There are inheritance relationships between different kinds of `Place`.
+
+```
+        |   CPUPlace   --> MKLDNNPlace
+Place --|   CUDAPlace  --> CUDNNPlace
+        |   FPGAPlace
+```
+
+And `Place` is defined as follows:
+
+```
+typedef boost::variant<CUDAPlace, CPUPlace, FPGAPlace> Place;
+```
+
+#### DeviceContext
+
+Fluid use class [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L30) to manage the resources in certain hardware, such as CUDA stream in `CDUADeviceContext`. There are also inheritance relationships between different kinds of `DeviceContext`.
+
+
+```
+                /->  CPUDeviceContext   --> MKLDeviceContext
+DeviceContext ---->  CUDADeviceContext  --> CUDNNDeviceContext
+                \->  FPGADeviceContext
+```
+
+A example of Nvidia GPU is as follows:
+
+- DeviceContext
+
+
+```
+class DeviceContext {
+  virtual Place GetPlace() const = 0;
+};  
+```
+
+
+- CUDADeviceContext
+
+
+```
+class CUDADeviceContext : public DeviceContext {
+  Place GetPlace() const override { return place_; }
+private:
+  CUDAPlace place_;
+  cudaStream_t stream_; 
+  cublasHandle_t cublas_handle_;
+  std::unique_ptr<Eigen::GpuDevice> eigen_device_;  // binds with stream_
+};
+```
+
+- CUDNNDeviceContext
+
+```
+class CUDNNDeviceContext : public CUDADeviceContext {
+  private:
+    cudnnHandle_t cudnn_handle_;
+};
+```
+
+
+### Memory and Tensor
+
+
+#### memory module
+
+Fluid provide following [memory interfaces](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/memory/memory.h#L36):
+
+```
+template <typename Place>
+void* Alloc(Place place, size_t size);
+
+template <typename Place>
+void Free(Place place, void* ptr);
+
+template <typename Place>
+size_t Used(Place place);
+```
+
+To implementing these interfaces, we have to implement MemoryAllocator for specific Device
+
+
+#### Tensor
+
+[Tensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h#L36) holds data with some shape in certain Place.
+
+```cpp
+class Tensor {
+ public:
+  /*! Return a pointer to mutable memory block. */
+  template <typename T>
+  inline T* data();
+
+  /**
+   * @brief   Return a pointer to mutable memory block.
+   * @note    If not exist, then allocation.
+   */
+  template <typename T>
+  inline T* mutable_data(platform::Place place);
+
+  /**
+   * @brief     Return a pointer to mutable memory block.
+   *
+   * @param[in] dims    The dimensions of the memory block.
+   * @param[in] place   The place of the memory block.
+   *
+   * @note      If not exist, then allocation.
+   */
+  template <typename T>
+  inline T* mutable_data(DDim dims, platform::Place place);
+
+  /*! Resize the dimensions of the memory block. */
+  inline Tensor& Resize(const DDim& dims);
+
+  /*! Return the dimensions of the memory block. */
+  inline const DDim& dims() const;
+
+ private:
+  /*! holds the memory block if allocated. */
+  std::shared_ptr<Placeholder> holder_;
+
+  /*! points to dimensions of memory block. */
+  DDim dim_;
+};
+```
+
+`Placeholder` is used to delay memory allocation; that is, we can first define a tensor, using `Resize` to configure its shape, and then call `mutuable_data` to allocate the actual memory.
+
+```cpp
+paddle::framework::Tensor t;
+paddle::platform::CPUPlace place;
+// set size first
+t.Resize({2, 3});
+// allocate memory on CPU later
+t.mutable_data(place);
+```
+
+
+
+### Math Functor and OpKernel
+
+Fluid implements computing unit based on different DeviceContext. Some computing unit is shared between operators. These common part will be put in operators/math directory as basic Functors.
+
+Let's take [MaxOutFunctor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/math/maxouting.h#L27) as an example:
+
+The interface is defined in header file.
+
+```
+template <typename DeviceContext, typename T>
+class MaxOutFunctor {
+ public:
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  framework::Tensor* output, int groups);
+};
+```
+
+CPU implement in .cc file
+
+```
+template <typename T>
+class MaxOutFunctor<platform::CPUDeviceContext, T> {
+  public:
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor* output,
+                  int groups) {
+                  ...
+                  }
+};
+```
+
+CUDA implement in .cu file
+
+```
+template <typename T>
+class MaxOutFunctor<platform::CUDADeviceContext, T> {
+ public:
+  void operator()(const platform::CUDADeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor* output,
+                  int groups) {
+                  ...
+                  }
+};                  
+```
+
+
+We get computing handle from concrete DeviceContext, and make compution on tensors.
+
+The implement of `OpKernel` is similar to math functors, the extra thing we need to do is registering the OpKernel to global map.
+
+Fluid provides different register interface in op_registry.h
+
+
+Let's take [Crop](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/crop_op.cc#L134) operator as an example:
+
+In .cc file:
+
+```
+REGISTER_OP_CPU_KERNEL(crop, ops::CropKernel<float>);
+REGISTER_OP_CPU_KERNEL(
+    crop_grad, ops::CropGradKernel<paddle::platform::CPUDeviceContext, float>);
+```
+
+In .cu file:
+
+```
+REGISTER_OP_CUDA_KERNEL(crop, ops::CropKernel<float>);
+REGISTER_OP_CUDA_KERNEL(
+    crop_grad, ops::CropGradKernel<paddle::platform::CUDADeviceContext, float>);
+```
+
+
+## Advanced topics: How to switch between different Device/Library
+
+Generally, we will impelement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not sutibale in a specific Device. For example, crf operator can be only run at CPU, whereas most other operators can be run at GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library.
+
+
+We will discuss how to implement an efficient OpKernel switch policy. 
+
+- TBD

From 16a61004da780fb1025ea7d537e425403d6ce465 Mon Sep 17 00:00:00 2001
From: Yi Wang <yi.wang.2005@gmail.com>
Date: Wed, 13 Dec 2017 16:47:24 -0800
Subject: [PATCH 51/94] Add fluid.md (#6547)

* Add fluid.md

* in response to comments from Tony and Abhinav
---
 doc/design/fluid-compiler.graffle | Bin 0 -> 3405 bytes
 doc/design/fluid-compiler.png     | Bin 0 -> 124118 bytes
 doc/design/fluid.md               | 122 ++++++++++++++++++++++++++++++
 3 files changed, 122 insertions(+)
 create mode 100644 doc/design/fluid-compiler.graffle
 create mode 100644 doc/design/fluid-compiler.png
 create mode 100644 doc/design/fluid.md

diff --git a/doc/design/fluid-compiler.graffle b/doc/design/fluid-compiler.graffle
new file mode 100644
index 0000000000000000000000000000000000000000..c933df2cb855462c52b2d25f7f9a99b95652961d
GIT binary patch
literal 3405
zcmV-T4YKkdiwFP!000030PS5{bK*D_e(wAVZ(eqHCILyl_hd3fZk_a`lT65EyNjwV
zVFY+%Y}z)tbXWiPm27;s38BLfKzWF5bR-?i=lgUdAIUp^JPs`3p17Xr*uQVW8rT$w
zZ8#m%?)|=beRZ_0ZT_*lvGb3E^Zl#O7l*>oGCg0oxZXQG*%vmq>-ENPXpwsT;OaoQ
zI6Y}z3ABuQ{qSs4*zEiM@O{1h@bFN>OsZx$119LzFI;CxT>t5m7Pd_@)H;4=lU5U!
zo>7PX+A$4(cjMj89eLVq7`}N=PVp0QPwWnP+^Mr)X`*TSq(|Icuv3ptqb#&;)4)D0
z@+K@f&V*eTvy*o_p6}9I-lZDWoPlljTxT?_ozt@qF79?MqU_YiA}ML8=rtr#^^vM5
z3QCFW)T5$96-K_(==^Ww`Ek|0J0kTceGeOVtV+8>s~h%BBNiLq`&f)gcu`uc#`o`F
zO8`H%gzr$6^eq8GQT*{EqcF>{Z(*;Px#?JDQfn^D$w49?wJJ(Ca46RFpDdC{MmmRy
zv`FVJNlhh(=?5-;NYqLcJ3CLSLt8X%is&pk3ER*kUmRyJg+=UR`yPAeuGupaTZkGn
zXv_4YOmZ~MZ=@ifYV$%a#~KanhGq8bf>yAoRUq9QVuKbD%ew?yyCOBe;3N{I7tEdy
zuwPIJ0jO?6v<*c8!1oe-FKfcDfc_P<V}&?wKL2M^xW!C?tn`8D4K4iCG_X~u5(=g*
z{4TIVKHHILk*lX+Aq%F6$jb&+c;FbL0kQo&RO5C-g@zkjugBP=1ln11%wZ@6B1IN!
z@|GaVQUE2$&OsJOWk5=h)W<_o2BoMa{44V47xR5SY?eq{&O=f3l6Zj$=A~0OO|_Ji
zL24OP3t;S;mX*e0l3WhfDzv8H8qq)>cha}?s%7UW=5@+WpR$a(r<c<qo0^za%k{gk
zg>BCa{u#p_L^5g_j^SXIBr7ny@J4NR4Zz0tR@bq8ziqX;pw+cadNlgj^@tCIV`AMC
zpLS<wWJG@ZL1hica$G9>TgxRKB+IRyONh-*+SZ8BV<7cQi#!f7?J+yCJpH}ZAzdH0
zRRlI#L+o~ff6-5^uLisJ&0gOpb{89zEWpM_pp+*fp=ES|@afM}!FTDM4Jpvw9~G1x
zHqkAGM$x9TPas9})8iH~+)5CJxK)=Ow<5<a$L*YP+qnF+aNJ5-4Je307wKvMU8F2I
zbajqgj@&sTcmMjJv1sVZx?Gd#Ktq=#Ntbnn0T%&K(?z6+w6!Iu=?a6izC@kz5jBAm
zyX#b2+&+7dlZ?JyX7HdUWi39G?Y%(Ool15Bl1kZ1EE^!4n)U>?N|%!=2xbckWiK4h
z^ixAP_Iogh1vOdKL?BXE4Jn!mbXAUrw^8Ox#}7uh>9J$FN{~}McfUVLy7>p4ktli8
z@r`M+q*GWZx+EfciyBPdSlNSrQBHS;oEhroUBsK|B3J?hC`@1{EP);?`;M03^@vSe
z({Ow3mLahzo$P^iFF~K;KgtTA{~<)dLDAR&At^X$fE`3J5E1o2M3Mp#83ZCUH#;aW
z5K&mq#L6{Y8~Y;{TT}E?NO&Olo^%9?QfH2RR7E$!Sgf)W*{2WEb*u}A>XyzMPZ^~w
z1xh0A)l}-FDnD{YSv6t!E;ai*b=SfqN-Q<DtX$mz$E9W?V>)2dV=?;kEqzRI=f&Kv
zPG;KDo<#yBJ<C+>ttrxYP#~eo7T0yXN~qBE&-4U~Wxzf1%_P~5CVDUMF8$CNuA5OZ
z*5QbDWF$&1`VFU~hmEI9#xAzJYy}J}xo);%nmXZCY+98%f$7Pc7V}bjL}HaaE{-Ea
zY0ME|O*ulDfg?oy=Q%=D^lFX(RYII2R-Gfz+v5lnOj09BTu+LC;w)|ioRuP^@Le9P
z_90XVx%;qw)BxWiHOOo3KFG7U585lZ4^XaZaXRGJVkyK)V!iwa_+pZ%o-5ZCP?N*K
zYWRj)UG^8$g((64nv1{YdhOR-AI`3Yo?$HH^_DKH5%-6|V9K(;m@bUaIb=CxUmLO~
z)W9skTo`Ec;UoYoJ8ET)T8`Q|qxS6P<lv+s*!OeA?c4{^x)e<K=vw87POgc%B(v}o
z>WbG?z|K~9!cCfGP>p7ua6<}fe_fwA-^Qon*KWQIzP<T2MP;T+J^Kn%6>CU=D%9AN
z+%||WR+lUES*T+6xp`4$?XRK<RKcg-)|D^7D^9&BK%wYWkfMw96o~4U5UkjwE}H@+
zpeTw4gYmtJc$n(j4O0cL6sC&w7lx@)`hhZ3L={p4Qt2rVQ{`c*JWO?^!c;BOz9SuB
zse@ESG4N=&t4Ckk)aMdK1j5lT1*52IlDNt~?JOZ_Xy%YKQ5ByV{>sYOr-iCslULoV
zow;WU7=r%Ex&kDH;(h9#X_}nw3^~&+#9j6>w@Q6`LwpskZNNU7CDINlvoc?K9Nkm_
z{d_RL>Vx?=7^^QyYa3@*9Tq)_8RVa0h6xe+6T<j8Gpsr@NN<oCqLqGY8lqnv^*D(h
zq@SRNs#*T2Vs$x3tT;#LulOcF1M!N{Xq%{I4-P_nn<zd-JMZA2nqdf@Gja_gP}?C3
z3G7R-8*k0H^w>olyBxdgg<Swv1iO%9mt%MC*d6^%*q!&V8o_9Fs|NYf_vy#gVD_c&
z`t!?M`Rezz98$|~Zq^L}C7N|x)Aw~V1?oe6{`A0S<vBepd{SQ1)_ZS^Swi!nHg9s^
z;rW~)-b|<u7&C;vo(wTdpd(c0%Mj}o?8rIdt#gEoqCrMYUQd>oWs5C1b5I~s<IOh_
zsBZEJWWrfu#aTjn!&4D!I@XwJB4Xz7Kz#~FVj_=lidb=q5MMC>NK_T}k<KvoM_&3*
zJM#De{tj~9?;w8(J`#6E99y^}u1zd&At3~!3fa;+ZAC(VAP*n-8=;%|_VR><`6~<b
zDNxQn;ksCWv)y>cUx0+m4sXEW&Efr);2nG~a-sNUOYX!l`VZ+#j&F^_n>*ffcD$V-
zaWPF042B7^7}O%5#;AqMK3?V<)^pU(8?^ygg~LZ;jA-&gFss7oQxmi?gU>IsC{J~X
zfy=k8=fIseaQ!|JhPczQh|t;?x=zo<gN49X0&=v2y@tXa?6oaHQI;6Ld;@!q-+ANr
zB82P#@r(e!zSACcTN?x7dALUw<UB_mar7+;QD5?&SmH?LNS-&6pZvaKvronocjyxO
zzqRo%%WQY8k=dzv&c7GaElaDmgq80c!LdAVEFUq<3X627p+@^$s4>Wj3OiNMD`Ixt
zc40?zm_y_Hmvb6ghdssvE^Pf%N1pJ(^!uZ>;F94p8;-N*bzxzH3`ja$G{IvAj&bfJ
zC^9JJgKKuuG!<J#2;aW*;4zW7{0h$_l(zBEyK<(GPqYbipf^vvQ^&X?amG^fPD$UG
zp4m>QrJ|Z3X7Yj#@?z+eyEq}M4XEHg?JfKy46{7;F%bWgkl~fnG;m6bIG{2Nro>Og
zO=^yVQF9#>yYc6Gw@~s74@gOTGpj~`Mez%IIjz&!v`m9|)2VS|l80xDtui>Otuykg
zO|D^^1MHK60p)somTH);;e0af1Jg@xa~%4d+M+v=Z4!z^ocWeVMBX4LZv|=|+P(B=
zL?T@}j$cs5CfIja`0OdMdw#!A0UGZij&uzm5wUF%GZhB;U!b7>;0(>F<g;L-D&AG0
zrwF!BxMvgpCGs9cPqR;mpELXz6`!1r*`@r?v_42Z>8MZypt=pwHWUQ_-^<#2fP`OJ
z%N6Ph3QA!$GDK6j#da?*e&+aO`WrcSO-ieorY&dINSYeT)P*xf+M`eB&YMD(fw&qj
zXU#pw_npBf>{1O9D@aAYkIU;?4)(KP9-m7v3*3aL!Wxbp*Zk(#*rIxycFx?26&X&4
z@pwr^dkz0v=lG-b`4$55pFQi#$+6Rr{@VNNq<PqH^v|z>4BNLKEu*(juRG`0@Jqja
zd}B3e`u?GOb%j2D{c?K-8Yj1hjpkP>bM$CE?0=Z7{wW=n{7myOE}2yR1sw;tnFSan
z#TN(Pk$<H?v?Zu|cA<1hEzAxs$vy2+NQrSS!-KeFKF~a$#FaCgNAXz#4Iad~O&EwP
zI{r(7JLk6b^gOyRorg?_+5~@tiZ+MXFztBU5>h?;a(Be^q0dIi_saB(M6{3Xd+cRA
zg)ql}`lo>!%#{rNtA^p5_tUsnL1xwiW7{Av%zJH{Uq>VvO-{J`YmXe-zUfYC_M37~
zuSnOorcH-i)ug&Va$VXIK}#m4R317QWH=m;K>iyxamtZOI=}I>fSdN=QZED)jV(XX
z*<=I)ibyYzs%57mVwXHKuaK+~iqkJgR%#_tUVuKLHhwPo*6gI6m}J+flDns{Yg)K{
zOc=g%j)*Ac{FZSic}aR1_KF)KSV0%8!I9XM3>3wZM)=}5gS^L=tA~l)U^-6pO39#P
j!C47B-j$SI_EVaQ3bc_NhZe=j?#BNCVJ-kgRI300VN{3*

literal 0
HcmV?d00001

diff --git a/doc/design/fluid-compiler.png b/doc/design/fluid-compiler.png
new file mode 100644
index 0000000000000000000000000000000000000000..1b0ffed2039c91a3a00bbb719da08c91c3acf7bb
GIT binary patch
literal 124118
zcmeFZcU)6h8#RiI3>rijC8Bh@fPw-M=}iR-f{;Tfp^1nD2qkm~2ntvMbr7W^(h?wa
z2vuMdq$7kN9hKf|Xo0)qJ38vjd*}XfzrVhBevZP)$tnBn{XFYgYdz<l_7!z@7!Qnz
ziHZG!##tRECRQ{P(=U44eg#_$lQ)FGKU;7*>Z(i`jeMivm+cN31~?|BUCxaExAfi+
zbZ25>X0t}>JL_v+mbI|A6EnYVf6Y?N-Od5*&BP?{E(?CNvvf8Wa<{X^;$+<w4t&`|
z7W~ZEEPg=f%P!6~3J3HxwT0B|oh*eAVp3u!4=BQfgoNasuHTT=IeY%kuY=zd4%~Eh
zc90bpcXM+SbCVRacd`;cB_ksveo{hQLgED2;{?tF>um0R0*gENpM!io&RI*Gg_E^|
zv$Z`|h;dx=YxXYA3I`4_PW0`+|6HfD^^L#IgvI@NTi^!88J~!s5<4mW?bzU}@{Fyr
zI+i$lTNlRRA*{8tqJ;dHoqyi^*V+E_Ej4>P2PaD$4tz}!@#j~*-}dwVT29uM;Qkn=
zQ9SkM{@-u=d4KJj_RjWTtevbaE?}K4oxsWd{Px!i{O^1Gc`SKx#&CT%eE)eiU$%mW
zqX?51|Ms92VS(>64lyw)GhH~Vf^^?9(X+jz9ulmcjg7wf7?J(^U2VhPpCCj;j3V{I
zmKvU&hQqBBOOc(w2|pRRr}{G2Nax8#Vax6N(eLLsV`Oo|CX%$q`qFN)o8s_dgq-B&
z;(A}sf_&M$T=m01CgyEIxBmCn|Hj~d^WgusWe`eW4$xVwBxAR7Y6$xo-6l>(=%DJ~
z9QZm8?@r8IY%Z01TKr&P=a#LU%1nR#<=3%w8`=VQO>`?~O&_x<T-;<mjt7DHJWY(F
zeT^r|TR8QLoKq7M6Hn|AOZ3@N$S1btuOs~DC(-*`z`P5EF~f|KuS7?{!fO%VPJiJ<
zbCM>nQHi^w4()N{{;yZ{=ixa+ln2~T-!S^|x_&-B<5NH9EzDGBHi4gy@Y4aoua4U}
z3*WwZ^Y-f<{MWFkT*TyOQ^XST!<hWzxeLKKefoIK;$JUQ$oST4M32Avzn>e;yfwqo
zG3(U7F4#~p<^Bk4#le3+w-^|q!orTd|8xsKY}4Qta?mX3|LtGT%^40x2#vOV_+Q8H
z>j$scI17`EX14usp6}<{*AMPq0wW|QcKZBx^YH6uKkomYWgEKng!`SZr}(eI{g1=S
z{XE><ya)e27k+vUqhQcS_sFaM>$yMf5^{L-;K8Hce!Qf=&y@d7%fFAwkB?L6f79~c
zX7K-sY3cWJl*8l8+6nF3`u_d<iRtpq`~&Ci>(73CGjBx8nV24Gh@6b|-FW8|Suc9&
z_haS(Z1YAh>&4{Nnfz*)SA$#J9vQxSayVT(UiGo`b>V-E6$aZl3vc*UiGbK7^nt~5
zam;e}(JQ$b#>JfqwC<-9J&%aKnh{c22%Gjl@|tW&r8m_19M7WHGC|aPmGSz?;Y$;R
zpMr5rxS{&2+Az^%v0qWa|1egenM{_-+gq~^{^uGz#G8v;Y;4+69)n{GPs^!sz;c<U
z^O)&#PL4mV_b<Q7aK=?4SluV?GY*|O&m!wQVph=OYION%C%B|*of$%QxvrCjsO@P3
ziPqaDE^q0zk<$&g_@`*{*$RZ%_4Tn9H^v>aY?UHi3xL6O`Y)2t{<rY#w|?w(XywW*
zl+XlH<Syp%_VgUzjg_O+45)f-zDenmY|PS!Tx(tQnr%-J1wv!{$`mr_Pdq+7;*~zq
zR@x9LT|F7zeD7noHGOqa#?#(<6KU%)`fS8YoAi_ryJ+&TZNbC1)HAQ^$othOvARn+
z-5cwN=lCN=Xyym{N<Gt^8@lrC9*kRF$_;UhzVi<O&5v13IlyL%uOC?+zyMR`gXbS+
zy7hWg4>iS3tA>7j^T6m44`Lz;*O-;lC|ie|+-IX(GxwI)zT=X3>GcPDPQX~iWQVe<
z1#EhX-P~v|LZrQxY<U&E3-QYXK|^XW?-MT><=kjUGb+Hk)}N}r-us4YXUR+@+iBMc
z{k75LSoCvLp@TV<rmLkuX;j+u?!q-GxOLcD9bcx?`X&j)45uwR(XuWR8*9^j15&UU
z@$z-2u1xdW{4Q-*+exP7{P{2BJ+dDOD0G7PJZ$p7Fb_kv(~=#1br-wARJ2ho;8j0K
zn=I%nnn^&Do!Vbr*j+qVE1Iry`&S=bx4vX8K>$s%BxNUaTP4!wwFKPzw?v+@2+Xr?
zeVb9bWau_wgeAX8Z$C-GbPj#6iggJAE5@VCYr3?@VskjXpf}s`)7uXyQ@myQ#wu~L
zY;8)f+y_r3iDndD3!=3bPnX$nC@h3_-I(dGm~sgl?6NPnD-tMG(usdFo<v<Q8qdH<
zd2V_yeXv?uXw~ZSXJNCY5Fov*Yv-hgu>eb5o=!|NESavjBu4&YVMFkYLma>NOvP!Z
z&l+x1U_tjz6;CV5eR?R_J-XoIHANooc546&NzFLLq_il{zOUDFX~H`3Vwh9i;zpC!
zbe_ACZ~TLh8KIWx=CkD-yH}>5f<^w5*K0V}K0np!cEjJ>b$H_L!Qjk<5XH_$xoLTt
zL6f(~y`@w0j_&`oFlR_iexlP`8EawtR<3p7S65$ujAZ73!bSM@oiw`z=1r}M*JR$^
zP4_9<EX{Y5Uckl$k61}+MncK5LtqfCz&(_)^Pa8`aQES5k&^Cle(F1>sskwryF<-O
z<(u@TfO`WfEfENtG#hHhQiF6~A)RaFnXv9#h`cdV#UVC>RBH1T{CMX|=57Hb0r~EV
z{OsA{`9?Xrscy1shk4Kk@XVqIF2nS#8X_z$>z?H$$9RfV$wGH~eAXt#(&j|37hHe0
z%gt*r423RiA>*vdWt~am6BJ@NTU?%Yd`=|?#3@qPI@#+rohR8OFPR%s>A*4pt9_Z5
zcqzQd=IX7qB!Y^uk-Kj5tH=lw$-a%%adBzZDI%$CjQd4~diC;1LP)nJDP8sBFFVf&
z<vub&MXQmLdvAL_g6~bbc&SZuZH0P1c9<7iSP`%ZLWxuS(tWrnx;a%)@0bu}I9=oM
zfqnYB@f>34!scfujD}2%@K(o=4c$ZaFD!;*ys#-RA^Q97YS`q$hR58yqsAI()6Aty
zE2E{)YKFygLVROJ+-|<?3%ch`Bal2577sq`^I2<fG49Lq+#Fdmmd59s`m9)XS%fI8
zCz7X!uCF%kba}}7_aLgQxW`Omz<d2;i_2EdV`rGDroQ-Nwvd@Le$|=VOhmc4=;gTw
zMYV_1A0!HEP|Duihv#h2*SdZe)<s(;sJ4LJdkDzXk0Rn#zw8SrD|?b|Q60QATP;|8
zv6fW6(YZJYo~2tDkx+vi47Z2fbVfs&1PQ$^-`O|ZNi*uhde2`C8x{~F2yP*AQ)k>{
zSxAa^+n58|cOSXzFfQOdqe4WJv%MxZk8wiF@wZf=1j6OEy!t4aR%AM{*^qRbIcFxR
zA2|#SKKH*0iQ1S3*uz{x<mWUB$jqk&YXYVPn=Bx4f~SGBsZ4jV+jMjG%vgIDwz$E&
zgV1k4jUT@%h$ItkohQW-Qs8aZu|+1OLog#=X{;?+_<3v&lJ=<PnBg`}0ngZD_1;4-
zth%i(Ou@|Q9D9%}t6&05DTj}EH4AFQ;_>;=n=~pqRt-uTUo2hz_@HOFz`8YQ>Ei?O
zA|L$9VE7c>;e%y3SuI89eYyI4o!5y1CEt7AK``;h(^Y57z2Mm<4izk7wiH^li5q@%
z44>qa5@AS6V0LAxDLedLy<GfT-S&MReGHKF!QxCV0L(fpnRz}!VIdf|9+z<5MqD|4
z3D*QUrVKbDj=$<`OVynWQQB|_$yrT+=^@x7y;{z3lbN)`Y%#Da&~J%LE5otHUi7Oq
zA2+X>#K(!8!^@4t-A3O`xvvS;te!kSPH)a8h)_ZcU_lhuE<0&AEZw3!BCziAE0C(+
z8Rp)6JlVjJOvH^CBG%J0RNz6dGA=$5cPKO|$=ul}-*!}BO^IixUezqdm;G_tejm7D
zs&yg6-j!;mgB!;b_RJ(X=+!MoSl_f{?@GI<!xC{c%E6zlUFuNoLqx9qm59~0l&c5b
z=6iWOJH>UIGm`qx;jgxy%!28=@MB=i@VAflNxn5ERYulR$~v!1|Hdv#>^H0)>nb4S
z*-)<#VYYSb(gP{}h9nn|GmUSN1O-QOUSX~U&@eD#`1`eIPib*Vfr@NFd`@?X*fyJt
z>05YO4;9M5oUwy4M7FjM*P;|f3d>i9BGqEO=o+dg%1fwDYxChXZYd?ZRunHU6uULY
zsZ6b<Lqt978Am*Ar1(;w%<99j>cgIC)8k+!q<Urie|?N)b2%6a0UOh<adfIBA;@L)
z{Tr$2Par++vVOXwj_VLc5F(t%z|d`6=S2=7iF)Y~#njl@Pt`#JYpjEH;RW}0A1xB8
z1xdC~eCK5He$T4m+Y7wMMsyF~c;%r(KW)%MTph0#^sQeS8kShB)(IP~;!v7kzN@7~
zZ-czwovm(N5A=!AEH9mcBZWC3_q6!C2lIICOIn~wf+wW5g(zG@aj3)a9BD^I7qM*8
zuPzhKO%PmJl_SP2TtlOaJyx-J^lQG+D1zAn&I_|vuelcMR4a@P{!l*-8=LQTc>2ia
zSbUr$PWM<+zhOmT-~==>Y4mbhrM$cx#q?sO8@9f<uW+&5D4~oMu{A$J@#t)B_sL5O
zrTLPIj3rZB=Ns~N9zQ~B*~s#tjj`@3(^2x?>&{;#O)pfPw<4xu`lNl!+*e0vg91Mm
zEMJO4q3qY?mKu+#S&T)aY{tDU<)x-36iN_!)P@w6b{=WgARZ8F%_O~u99mol*rX@Q
z_i^AU$OSpsnb+eT?be4cASO|olX>nV)jkvK7sshh-ugCuWu<J_5N@wT(z%WioUT+g
ziMS(Y_9^}`xs$c#hRXa<L(j>S54JZ5GxgmOQq~`tkFfK5H!08xDTIl(#kJbz8T`6@
zF3%19s<iVqu;X0JHY3A<?7Xs)<>~S-^8h>(#iR(>#YP>cb`vqM5U%`lHlEWAXoRGv
zkgrz}BlAK6<24hX8DU|B1VlhT;WSmgXhTl9>&U9W4y6zid;Bv=>1eHjdd;eH<0TQX
zXS|mu>$@x$<qPU}d5~6cL|^ja%B9b8ZLy|Gm+uX)oqRlP!XBTN)RgCQX7}V;tuL+d
z`Kc=Uva0!J<3^(^|0j;WR}G=pKn*Uu<c}i*IPA@=*smR6gFVeoC)Vn)3#RTj4;M4$
z;gQf+i#(4JFI{*-Twm|M&2bruFG{u6zsf@5mvwg7)E5-kA^JP?++%IBpctdjz?K?m
z*i+<U+>tCoJWJ9VNM`On8yTq8u3b;ifOOxaS7vB5uu4a4N_hKpX>?&ej8TX^A|D80
z!n#MK&JvA6t9^Zz2Uw}MXYPO2Fov0uv=wpv1m$C4hFG#?#>ocT6v6Gs92SRLSI_p{
z9;j}dT6}M%QuW*ypf)UaWM<3(FTEW9rYht`kA*M$z;vR<+8t6;I<2#D^(zuD=_^%y
zMJ*}eSrIDEx9d7hyA7jnJJO<f!)<5#Khzni#{uP2OH!l)(B&Z(=W!14D#Z}HQ<4~e
z0E!<R_JV5SJq{tHZA(2DySrZk+}$^@3O`%6K39irEC*AMdZRIFG_U~@L|o6nvhzqe
zROtqCM0VxbpiuEd4`vtLySdliz)*<fTA2oJ8HfJY#)dtk;#cB>IkIvutL4Y|td5HD
zEq0BVRJ|a;^flES34}AnL&I*1IESGjo98{`93v7<&(JWUtDa*wuU+iS7>ec0fF{lx
zBi4zuCJ!DV-BhxjdA1S623u-#niE*4Zo`L(K$Wq90;0~X9GvS>EuQ5BRii#On04f7
zyRo4&?NKkm@|F0ye0_pyPk@Z)UhYMA3kytaFL|An3+kc3Ksl{fAz4wk?fy<W_#xqd
z)S{qb#?LePz2emFe<^YXB8!=p$=4|gZ_^b@h<8{FnP`a!yi+=nU55_!T4+)4Dy8*~
zGygF&2|seyhQ}%UYQp)cc2E+udO)*;`&vCj*(>u|Hw2#PZ%G__LDAHtE{qXL5h4Zc
zX$BS%R@zlBFzz3XDO_hn%sR6zAMkrkWQob^l}iO#4(*R{>B+-TvIE$-i##|G+%{<q
z99C&Ews6cjo(%nJoxwN`l}lkF@bQ6Y(|X>`g7{vevGB?prue;+dtr%qHj;8%w)^U6
z$m7^)T%G2xjN*b#GQs1fz!k*i#=Cc3C)fE}OeFE#0E995R`a_RHW6Z*WV|=s1&5y*
z<oZ}q1*gM}V#}j^f(YgJyanIR_ly&tFRreT(K)!<62-@gHxz$u`X_R01AbTDKuGor
z1T4O+OtHimKw-j6+Loz-BlzKHx9o{7kQDd`5%|xK)m^t4<>8h&Uv8Lzs))Q6<kb@-
z4JqY=+Ufkmz2@rtF25Tk_CuOD)-k~~7u`jqyKYztbJx)r_sSD``tn{&6Em8ws}}4o
zX-v?3K1P7L<(FRzab*BQ|E2~LyteEimEfWsPwB|A=!q)qd|Mt^be_GHCw+9?lf7zq
zX@QdbLiAEq$ZU^WUx3o1l`Iae98ddEGtZ_J;hHAwUUY%^md&R|SORBAJCZUMG_m9t
zpRc)~uy;~^pbn9nA0bi#4eS-ReYDWS>GSYr2(IpU=WLE)%+n^*SD%ioeyyN4@Gqy0
z8@PGXtuw6gC~Fggyt8Y2Czp?_{{x4(jJowY%#9fYz0hiAm?TAtI<~$te{g9C)VZ!z
zJ(#=iZ!|=l>wYV~KzYbBSD+dI^KlGZJ7w|qPK@XwWTd23uraOy;abHmeW%^hr<nkW
z7Sy0RL^4~JI_=*+-1kD-tqH)^nr4T0kC|ic$_-Yk1{m|y>T6c*t5c;E@E96Q4n^Ac
z?oXMakR#UGo6^Tm?_5iWe?}^qyU2aSOOAbXbU1+jw5@(e`zQyZZ+xUBp~t^H-6%-!
zX0TyiL7$IY?(ucMN@M8a4XU?kWx<!iQWrX`PW0D>;(ARnpm@UMSM?XMvn|JVJ3Y0@
z6*8V?QHpY3=8L*l#+u`GVPQjiIVtJll!j(Ste}tF`?Yt|yRD6}FII4W;9?mAJAdfg
zz`_bSWIl6J@$}3iI<4AGgc&M_={oJwOZ-D;?`QwS@_$D$*X;JoX;PF!?*EW(B)taE
zv&4d0)n+iC%=;6%LP;Jwb1-Fd7v^~gPb5o(7sf#~rx~lLklPWY$m63w32N{>Q2tMH
za?#i4n}%+89?^^_v_m*&kUUH2OI>c;9>VL7-9kcP5q(Vx_hv&x3b}mVI?dzEF;ktr
zrC!C7x%PdEYUf7#D{dRVA!Zb-hX`PwoA$w3YR3|W)R8Q4VDWkyKVWt=J?!<zZ|Zma
zMWNKC$u27YdA~G}luo`a8%1|&)gezYN`+MjyHiu)aj|z5EmFerK3gFEHGik@gkP+>
zD=1;*N^ZdBLu7{b7A!7OO$9ya>r<xq#jx=Q4-v7+{`(@x_ouoYTD|mCuDhFMa8Be=
zO>k<fwT6^NRsnajy9zBdvk|7N0?PyU^9+YK0of#Jo4WD>!9}*;@TyRv;<L-BPID`4
z<%*kU{$d177ZZNVLz#HIS*1<K4$PA-4h9K$)-+V7+b!Uep4}AeX<S$Nhq-@OMElqn
zj~;`hXHoAbcIMd#ehzBA{I*Uf(+w0hkpy-M@<gid!scMOscKdG*h-{***?M^iYv<4
zC5%@@XY+Z+0FJ_yD{_6I>hiBtEP%)^AWlB>SA!H3mO3o3idWN7XHxM~MdM=Jp<`z(
zYk$uIgd_Qu8>lvu9IlvO1^?Os{2w2$1s8Fd>i(dYZkX89?ZkeGNWKgLD4?x~?WJCB
z#gqA6-JkR~ZVDJQ(0R0U&k;rR#mknj>FOI~pvlenWyjC%ixgDxIR&pJ=l!{+q?OAE
zU!ve!6jgNX+=D#<8Rx*G8zUO#k<f5`k$N57!v<A>>^&Bf#pYGN()8zV2C`-t_U2l*
zHlJ)FT+T2pFOw_pM$pnkmSa4}gr&JA0OXzy{<JxQm-LvszS-7eqtSLTOoUwRiO(qf
z2$QB|?77){o?mM^gh4Ubaed_2CvgZ#CkF5YBB=?VmiGtsycM21dbI#R{r9r26C=Mb
zkoQVsqrB4g=P$}2Z%+5Iu?l);5-XQTVaXugrO+OmPyrFu)hIiD+_$iqJ~L3IreZ((
zKCwr>Vb#&EcYO7+f3MrB;$9`660gmC+Yb^*dfM$n^Ij)fnqo8P42JeRK?-`~aivUc
zH(l}a=VDZJyY^_&=;~tCj)uqSUUVk4y0(R?U7NpfDQM+}bcP2iW4Dj6s;Hc!SN6Lv
z;>MrzvHk?H{YtowJqbTL+zk@VYq4F0BHMaJq5v&gJmNeXn*@dfdj`dlMhMhhj`!t4
z^{0%b1Gcd(a3?6^E>w2Cy6_Zj`D(~*QEiamuT!;CH1{e*2u_SY3(dZ|SadYj`#pO)
z@&3i0Tobq+PmIoh$;#ZIF=w)e1%*2x7q;Qty^~+I(5$hEno@R5s7)z~aNx>7fIli)
zi;jAeI&>U@(gXQ}lT%|`0xcf7@-0pSIt&(YRzf!=jQmc8nlYdYKQdOd$de)*iiBRO
zavPVcI>?L}O1;Xd)ZwSvPlu$#DtBKwQxT8@ix?!(!^KUf{OwB5hbXS5biJKU@M6gU
zDexs=cAM;oFmTr4_p`Ynfu4{O{4T%m`s_e)@wid;^?P1;yh|Fh1<Me|6E9_z?2jld
zUtb(qMP{YYYC-fHuPm7P#b(N=fjjNfh<j@LE*F4?`_OOVOORP<w5L94af2|m*7xtb
zL-s5PwAv8ouVX=wUX&#`L9yzBOF@EbnN>S+)6$FTv8d9fg_ve9?Xab>kRdgwBg1D*
z?qJCn6q|S*PddE1(mK5GhTGKn@&@kd({IdSGW3{q^;fCs+%m7l(sWuxL|HOq;5!h&
zT{OI*^%B3Zh6_=~ZX3aIdv7d==2YW8x%2(<w^^_UQ1c|yEsY^e%*CogqoQcZgU7aA
zJ(F<%Ob-_zG!{apWuGy_Aw^-0ZT73`o+pR~IX9NCq4p*E=A!gd;;zoe>JjxCz<dca
zFRs^;a?V^$zon-VEkdg#DofhO3dW2Za!I5b>OuEkuqPPbK92lg2-s%sYnIYnxxopX
z$l)i^BJA)3gRZ)FNpS{Y-kzFnZ)Kc{o8CE;==W5LqT#k<c}UAbW*^9lXef(2XX8o*
zT+Yc(M|eH4$Y4t}b9;U4Rj7Zglqh)V`o?4RT#NID0z`vXcwEu(#~|O~0FWS99G=6c
z5Zca`dATSCd|qvN5V(!%1|6TyIlG#IYvHU(WrAnOz_0uMC?WV=-(5rZeW#f(`br6`
zl3NC4Y-o=bK(+e9&ce|1!(q)fjLLKvsNGZYM!`Qxw4u_>0rNw-QShIJ{O`ZkY=VMp
zWfSok_IFVI<K99cKn9=R7SsAo-2XdM_=R>--rznKu_^8+`b|B3qSk-wmz|l6wBnM4
zOx1bUBJAn+Jw^4`k>sFTuU(SooE-if4voWp5|*_cE$V{N7sKoG{_)(QbId!0qvq_l
zF#YR2Tm=fw*})EuA8+mN2~EcY=+uLK%uN4PcK-7N<(FXm?$u@q{=L@yQICFA0>t#%
zau(nR`+q;V-&w%d?GCe0`f)=1{Yx4i;AVdJWn=pBME`Kb_h6ibM?|W9{u<_g{s{~u
zsLnU{F#XFVZrQqvQEnlF>Hj+5gWzU#H%>79^V9zBG7gD=syBsi?VqsU_qU{6%kUbk
z8!<8c>-ZqSGrTIZr1md&<M%tm5V2Cg#Pr?ee1FPs*B+Vzp5aTUiJ)&={>#t{i-Vi_
z|9L{j-K_oYQNOzNI!QA!&3*Qx-YbpAXmaOdS3wu8&(|K%NI5`p)d%uR_atN0fjZ-y
zLeOfwUhd7SqIzliZIkP;XsfzsC#I}{;FIPsu+Ybb?aFV+$M%d#H#o$4bnOte@P0Wz
zwvGYX(S1i!O(Y$F$9#|Z8wJ_d-#szlIj5*vVG!}`*$sxK<0U?(aN!?@c&-?Xm5(pW
zsM6P!5&GygpsUh+=&PrlN3_R*G<e#`s?-xG(x=T{Gcj<y?Webo<ZRK!u6D&soi>v|
zm1^V+#;CE_xs12FtSyYRRC&^-N~<<raM!<(tS)k$9NSY4X!Ld<KQtd<D3^e?Kd~SW
zs+ev-@)$KlAnX9copU8xmI2bLq~FbDRkpu<#M@QmGT!YfO9i@R<e;KKIvV6#ez$H%
z>~3S-!P>B#&JtzPm%1V9OiO77tsAH1SEeWSvwv#KRvk5=(z8N{jBw)FuX9Z2yfSpB
z#ZYhj<)kFG`;6S+@46GWITRhtiG<xi3)Hr0O?pI0)vbo?L_$l4B5Woq*~F!wxFJY)
zkf_?oU~ghPhW1gX8JZA)uyEb*UZ7QJcM$l`KjMM6`q=)HAtuz58CjUrYtsn-lTK2G
zZ98sS*4<fPZx1N5OxMXyvZx8!DN$juO?L^f7@eufF_>^uACLN7YJoel7!co%B+_1S
zo3(3y5Z}i@;U(6UfgnMK?CM5&(3-4OP=;SCA{dtYSoIRGNt9raRWwY^C;M`GdmDGa
zAE3m%uNfo%p@#r?W=ps8Ap(=_#M;HDT8dc=Etx$U4RGagn#JL-Q{lU1qdc}<dmzG-
zB!l{z!-R$6##*sGMdZxcddVhdlsDDb?E|2nZ5hCp-|>&%PKh2jE@~6-nmAQZ8c4y@
z7pXlC#V>~Y+L}ulC~IZjx)=b3H{LN(w)z=p86xm7jDWF3NUzTv4o6?}88K&<Oxy2o
zPab@BTx+paODSL#KV3>Q1SLd|LrswY5A(LqXl2|f)>!JtikZ)AzZR6<7Igj{HY5|Q
zxPoicdiUGUVXq?<Jb)>r$<A*&>E}ncm}_Dbyoy$kK#o3SROk>7KVjH*<nnW~wdrzS
zbhOS$%77nZYN7TU6*-TltZ|DVYy6qsKx9lX3W_;x*C{o&RE#-Co9fYoNJY<Ry<Sbd
zp5rH;UV3%<M*ZG`Wrk>`$8+9ed1fF(Rj>Npc+z;YO{&oWJKz+^Uz;knA`=_0zS8(j
zSA$=x)h33rb}+_5m=Bd=3bZ(PSBoxhAqSwyryQW)GRycJOEHG}`n=~GitKj!Q#C7I
zYF4=2E2iQ$_3LgA^WGkgKzkU^xTjCuW^zjxuvq~GZ4Vy6YU;1^XUF%lWV0yUSxHtl
z*~>k+71Ned!NKV+{#^hKZQ8nP4He<-`q-KKYZ}vl{LHHY$(hjn^PN0WHSF_5<*Vj~
z$0tEu=op_0gfiRX=AGE$$)O=fI?%>XJSJbW7)2fEiUJdA>5tu7&zaSs((^Dh#Z<7+
z6^I|vNx0!gF$c=b9}F^zAE>|gq7>`!dL$D&{OUrM(#CQX*RwvbzV>WAACGU(Ft*`o
zH>lzjH_WZM`1`Q|!tY?_+-!BR+^LOlOAoxoDBO5A44`73DFa;V*s`F{a=cuAW2%Q6
zP%J&&<R>M8&%r3u%#UTuvs8@1P88ZPAXVg)PV31A9)FMC5|5<u6o~kYVN+;C!CPjf
zqKyXfSCeP-G^fdc1tQa8{bda|xSMIUppMIyNXhk;Ddd3)p1~@0dZv@sW-Ke_?;DaV
zH`7H4v>Wbhs(Q3GacZe*cO<sR+9orkQlBN)#=lyW@rO1yUBvN>E+B96+r+ehG^Xt)
zWuH~lZ}_?KnWu9<y8-=90?cr7hrl-+#d5(6gh}kS0TAN*I`i$uht5p4r9OQmsMOmq
zz*CneVz!{^XOpJ6#W*_;6yj9@w<7AG`Qr1Ki0MU;ZrsA63AT_;M`k?RZ{f!cJ`&J4
zIS0xLqu6YnB?8fRsFa<<Xt@5xCy_AUwMl8JAmmM8UIaz3n{+@5<K9PCy<2UR+@{c|
z{RCwo#AW1TC?X6Ue#UiUePt>Fkvt8m>2BG8Az%7pI8fZ&GDP&@?|t;G%X3y8K+)4x
z^Y|eB0C(`1Y)P2vHJ0`)DgL9|YE&Z$FFk?VM$Dap*qHX+#NqF9{T}4{QKgKt+3VCc
z3hGzfL0WJ@l?n@fHcrB|sbBq|7fbwTjLO}ysoTxYBq8}8N&S43!{wg!PYR3j9T`pz
zeJ2aU7Pi>UoT&;?Ww7m&Dq)j?6{mbqXL!@}POjNFbjRMt`@jgbzB?P&aGM+SJVlE!
z;%xxhzb1z-Kk4WJSYm1FDaQBbT7VTPf5^79>n?QMvqN)4462Ba8CPr{x_yCPRDSD2
zm(A;4S8&CI6Cj=kJUe{pS$*?42t&xj_H?qxPoP9neER)ZCxAC1SIG8W>U-*RB!!-B
zS%>Z*z}f%})5Y-VtWD%--Q*5P<3>C9u_-ZBbm<9#hz8tMgs%+nA?Qp`U>+%=p5As_
zb#2Vg|1gv#(vRsb%p%I$0mU-VZs9X<n&c~OuDAZWlW#=Ty0n*|trxJs5ctD#m_N+W
zf&;DeGzf91BMe2Bm`ll%U+%vRQaH(gWKtSM0a|}~l}r`4ZAa#QsRC?uFn_vfxsP6|
z&f$W4l_m|vYrv;LwrL4rm1q*=^EocSZR-aR8C!<Kpl?B;_~{Pc5VN}r7gZ2mZ-g|`
z_wvQ*%*dqf-I_AxIMc19cxanZ-LFk-E4TF2h~s-olv@1-S&y@|)495Rb}rF{JKB53
zJ!UNY5~aHxs#g0?v44#~|9L7x=Rj4X`#E4p;!i$Y2*EPeo*p?+$)UL1zoRL~>*Fsd
zLFA1Bd^U-LL(&&0MMM<IB&CkW-+8Ppb&4!_rGvEKM<ggBI5I?ZaGlOR{OSYtCx1^Z
z1vaOAs)<PbC3~J<4TpiVCrkl6h<G@`$&Q*4%`2gAo<uN!qB{YOYd-DSFZ%m*$@~@F
zL|kq@kZ0UBfZ-$X^i^eGpVE{bP7gGULD=XW0|BnKNoli$MdfLic*$4+;J~oU76Bi4
ziVUi_thfs?Dc6<Oa|vjC2cePAxmg7}WwvP5nxr{pxcEF!`2e#@|Eg7HEcQW8m(p^D
zN>)GXR>*&nz~YB->0vEs#?Ndhozd@=C?u<`c(HF0K-O*u97ppVH{!ER?sD_wTYqlu
z*jI+%<QMOnd?WZH{`T`1Qtmh8G&z^?%^==!fw@Bd!Bh>f`p~Y(R%IyIE@u`KkZktA
z|KSoJ8<-6R*tqN7bq7|lVv$Si7ou8ULw<u`E=&%+u<b1&6Aw{NmJMi9(qV3t%Aij6
zzVmdQP{Ktd6nSGO8qVZ0(!9Se6ROrV;}OVT#UWF)E2Ok007&TfH!nvIkeGD>$_C7d
zrld87FwI_(wqk~`b#!;JU3T^46Bx0n8LwB<Ar^Ul{No<>?!b(lq+rnSw$DkgEyIoT
z5odbeZ(@Jr;IA)0%Rcua8*-);Y3tD{0@qk#OFJtp=pP65@jG2~Tjd*6_7=;#kN9_(
zk=ZP758UTwfsZ}#S3dKJ`)St$M(O^2t?-WF=AIyN(!T-bIc?>zK#+JJt*ofcXNr=Y
zFb&~jUBF2;E?C6S1`5{zijw&lnn!S=Wad}Mha{1;IOJzu?^vr|Jv;Epg=%a?@YgK{
zs>uml2Vqb)C_>~MG)R~~k5Swa3#@ExF8-lMt8Q?cdOHRVKaM%ebECr^1wawa;nUlN
zm-3}Y+eA<qe0ipZB!>kc?KoEnTEPni9A0n1YFIW%69*na)@fuz1_@2nO%7&JD=)&B
z@IMVyme{_IN+)pqQ7(A@ax@!rTdBX`(YE#z`iWgzHUqNQ{p0UPt~D(F5^5HiEA)`!
zxMwR63z_xZ1d&Ce3yD9~c!#n;sj#d3jPtj(S!1(>U6?4;gerR584JLFfT|&X!r6Pf
zvZji33tFR6F@PLB-Jm<hP@`;pmOo#3dgKEr3Fw68(jm=71`oVvtMG<3ng%Eb(7TW*
z^YnyKzWzyS0`WS#NVf!^I+Tf#B~|uLL@@I<R1sjnT5?jdc#q;8d)F>ZssGhH;I3-y
z>3({ABh92#_Cele#{2a=*h@&0xMJa=+Xb?a@eJW&4}d|%GmCu+%MX<2*tYiVUbTF4
zbhAgo`HX%+0f)tH!HaEds_lVAEWFnBb2+zNuxoLZ#`6_o+wxzK*7{ku6(#>X5RaAL
z`cf{u3~qscTU>r6+}iqtd>^_ecM$-w6osa#5dXq>wIIW+Yj3(8vO7&ep!8Ixs}*L)
z?Ld(e-@sKen9IP#Q=BhIYE@g;@^3Wow4NZI2Gxjkz5M}|?8<gU{AwZj0`NLv8G3>O
za4EItST*fP?k<87UTlmfd0e)zCaAJ=pb>YnD+ebGhX@^M`5esqrE^L7Eun7*DqM2u
z*E}Yr%BSkY%O{3zGP1JqNIJ%>l(s=D-y}zhI3k>#mJ#RWj@JFA;R=E|{ou%_s7-AA
z)US1QTVpqWt+3LRX31ptZ%V1`NXUE4aZOEb=7%f^n64&39n7yEj{D{7B2u1%-QoCU
z4ahb|<v#RlV-sLj59G^z4pmD7q|#KDYAM6dvPTTPmvE$E%8Vzty#V+@SAw?pjO!EH
zA-Rut52{vcFR!&5*^`HXV<6{Kh>~Ki*!(Yx3VVB7*;&&#wW09Ye1>=J9Jldj*YS?5
zgwol!yh}9)Lki~uS&3y5d7$8qC!;?pgcpRkm8a!X$5W`Pdm@yUo?H@phnX*IAMF5z
zx07_hd)P2gk(o?s5gdBcDlBbMTVVuc&cd<qBZ71{Hsc$aCD&$dKMI7I-VRVx6LtR~
zN&o?I81Q!`tj>Q8iI6o553rm8Bb6gv*V6oWSSk<J6)4fz*n|*Q7K8Zl`S?^3ZiKkC
zYXgTU(c>achmu052W3E01JLi=fYtf5^Vs&?M+(if%=_Rdv^lWG+O~BCaGUHUjO-xY
zXXig2$jY;aS^)I19w~J~;4gZwM}W-H%bGzjTvXe?Mz5TP9}3}@Yd>-&YHyi)vf>hc
z9!EYe_rl+Hv4MliXIBjJQg=XY5r;kDE*4a|kM^Fb43-@`R&cRGl-&=>wT+ny2evb6
zaYcxtz~wjN!x29y(0<qVfSf;s3M%`yC@u&!7ucIp+1RIvuV+J)q+mBbC&^r}FPS~h
z&4)SD#n1q&0u@Urr+{zQZ^yWV(h6V<CIAiY`T*{Kq*Ln7^ofG*U9dw?Sn7Zce!c=Y
ztTZK$)5fc&5y!Jh7${|(QTvgC*RldXl>*ynj}KulJE<Fq$P~FlL2ltqwd*VE38;qI
zltQm2oxBWyIi-6)6wTY9T*YitX1J*7=cZ@Z7$xJ{fS?wx*e2Lf8Jg`y(w1*=>@Vb}
zaRdPG>z;@awy!mC=pAMrA1bRY;SA+uA;Sq9&_p6&o1*2YR}@#S&qJxOT>wBo8Mx2&
znt;)pd|?&iXx>eLh!VC5M@eO?1Zy8%M8jt4YboLqScQfCS_83IrS+NH`1RT9rR7f{
zMu6h!>5l8%1Vn!hu%ya59}9>g<U?}6?5tXs<Jo^Aw*hFtIh|5TE=QxCpG3QVyoWVn
z$%tsB6o4x1g97pxaCIU6$gl^XJdZ&UkZla88OC%%v=3=UcYrM7<YKZJG5!Y(Wyyp5
zMD`h>e-e%C@!#g_eRu)L$>V{xJIk-wZGgg}>-i~*?ij_~5Qn{czCGQqs~VVG4!q{e
z@ThFO@O47?tsiFCk{KC?_ZiA>$YVI#ipReI&=^SrJ`6Dr*vX@dQ$RmMVeMPM?o2v7
z23CvH{Ntcf1ZQ4rV3%H79VZz9Gj%$kO|UiB*p<es(CnMMJL0W@6{T}Pe}FBU?Gl0m
zGu>M{`TmkPVzEiuw8sz>;@Tsl)N{pDz64NO#Im`-cZ?SfR6oo&)vJSF0q8HNQ5zm_
zU$&Bz`b2l&8bG%b6|91}!?$Ie5xE&Atm(S1Cq>o07DfgFRT{_m$fa0d{l;#zPnWGV
zUQt*?67{oQ^8V`4^7351=OV^U_5>TL^HDqN3lrB5mZxnyPI`@{CSR^XQHiJ~<l$0)
z1G!a+5vc0cy!Ss<5#U$9>$`w{^1_6D)&DETIFx!Twhy@_QB@g=7MajJ_fcIC&jmAr
zQd8pXNX1h{DJY#%PZx&8J=GIODqWIbu_l<cbw>fxkS!pcupV{V&hX~8z7*DnnW`$$
z1cN5|`OD)}QY?eqX=vgRcu2VKw7xuxwTFhDi9jNij}M5Cbpt=pEZ{VX70>sIEG5Er
zK^26g8Cmab1AzYMOu3qn2ITfeG&~cSjT61n%!2?tX?8<NM;pz*74a4~Eh}L=Z`Bkt
zS*a&OoP){+s-&$b)u7IO$EY<3=I}xr4QuvGX&3{RX~OY#?@XVKj<n!6vfmQqH#~2K
zf^l_*9sPs`GfP=uyEhzZ#xLbE)+R5jkbDkAQdrEb%60%8I)Tq#wIY_e&|URX>N-dn
z`Q9r|(mQ5A)s%Mmg+zvDE~wBQ1&RH_)9&XYLiFH?!0ao<Lj3@3Ou65QE;t*Oa_QOe
zET_R*a#`4M1Bv*jqos*sR;A*Av$+lzL1qGD=B$r~hNO%e2E@S=CME9X&Bd$lK2&|4
zhGyuIi+|L-^WDC1lFiTT57v$JBW!|t-z|#JhYT^}Aqj^6)~EIatqdo@&bicG&^`Q_
z0S&P-GFJ!_#Q4V?vLkV?O3={g*t6Co+B3p6i}Z&LqbnVPe*R){g83ssJ63H2P%HL;
z-Zl1Yc}b=<G0S~=E`!HAPfBRlHP|#*mv;z$8$iYwJyV{dD-rz%D=9xBSCEsYS%42-
z0fo=-cx;q=b07nl>Jg_s7jDw@tNd*Q#-sA19qXf(F02u+Z9J0hbqFf@{N$3vokp^w
zckqv5j|T9`&=@_VKWDuQ%JI2@r|kjsx9;!;EZlcHaB708IbuiM*AVgbqL1BT2DBX;
z`=Eegy~CE0Ha_F`pc6z>TIrBu@Xoi5-ycFd4gd<ag}S;o-wpSdO9;Ih2OhjDu+Uqf
zYb1Z1S$=m|U`I=i%6(%@z8+Hf;<i_biIeZgyFova>Tp-K<t5d%!S*Y~ox%mH4UXTh
zS0O9ri(LutzrG6qSvXW5Qu_0q)S#HbQY#SFO)69@u%R{rS#wYTERSx*^%*q=<V?5g
z%4dZ7j>-TwkmP)FfG6%tj}~C~8O=5NfJOLJbfY0a7I25sXGx7zFr%A2=ze9J8(wR`
z8)OT(ZR`{>+9DRe<T#%UYWgytoE@~y_4$wnkam1FR;&tQnYGwoph1vepd|@g-n!k<
zUZC9ot1H;+1IFjH?}m3l8h&l5asRsh7!6#74e%CcgC-6}XIZ-U%A8)}ML4573x1+j
zsq~K5>~{>5pZ~Gfn`7d4Ng~vrS!`u4KX7*+K=(s8L&*rr*U?WMB#1#BvcM(LnjbNo
z=P*zyb{xu-0gXfBfS_VCRg<%t?FVdJwsfC&E5nn=(^v>fg?*KJ4Uc7Q+K+QjtyUeA
zqAV3Lz%Vuza3O1yDU<zEz~ydiN{3A*Z-^UY|KYPft1h=bQ$;5{pDI;ZehXX($)ibo
z6l00DT#aIexyN$JjU<b8kWHqR00*rP42vIkrl_%g{$c)pW&_#Xdsd?5!jF?@%U0J%
zw+IICd_4kE7}-GJ#DP&t!e~K0z+{=6G}bW@L9mABi)40#8hP^7(<|9P6|w4IfC@3&
z6IWl}LD+S+VcOG131{;{8WYA^7%VVwoJZS(gp$E##0D3+t}fv5iXRkg8L}uS=~B2@
z-iI73MqM-^9WrVHFEK81cI+wxPU$H_M$ZCBiMTp%L}tvLpmbJ#=TLsnk@%_gxvLbS
z>SqcO)%C}{-IEhJO|BJra@%~rpTlvlx1S8M@$<Ss7W(-fcX4Y6+#ov*P5}7jN+V#Q
zy`6C?zsy5Pgj)?6=|%;Oz3!jbx$}Yc)B+qUg=&DBYG+~F^ITPzoeEmU`MLAQI)HmJ
zh-eEvPmRB=1Xz-0VZ{0@n7-ZfdrlbMsl0MR8_wXhjSG1@_rGQ3*{P4GXj0RYe6E`$
z2|+5E5rJY@5H;A07SVc(V-y^fIIeSx4!lMf_zzeN#t`tf;LXX$|8@^>Z78|2vR)H<
zh2gnEZE}q8;8j3{5;7r55@F|LbNp?#1CsVc@>N9IED&UdMKF>i>O8|cHlsUF97{?H
z)zdT|@naPj5wO}wkU=t*fQMs-JhdbX<_Dj`0HeMpC7YpR5kY+7w%+}N$xV67Z=l15
zc}~@7-`DW+HKl)krVNPog90Mn8pc4jz%<s|foF)_W+qr80N7=q>V5Y0(N0W>2;8?c
zUwcC_CI(E6sigY`4f}~gx_%jk4Q6)-G%7NU_N}J^Mw-{8L#Zj<%Xb3^Wa_1-qc720
z`Es+H(fPm_WrTI{`}AGRy`{Ws``Ha&`3r3-Mc*LCx6g~OD65S=1Nm=YcNqZBOMne7
z7SajJ1Zt7n{L59)Qou2B`w~<%_RP2|*ZOJK+4*9E@ZAPNDzmGVR6scfG^|-Y3+e-v
ziRw6jm6!T?-uHjMs8~R&c%E5l44VDBL{N;51BN@Xg~EmRmozCRQeR$OU;sUQ>tqQt
zLLAB>TtBK6V@q&nc>8+L$uDTsd<^Wo#dZ%wH6d`0qRD8fno{hHooH9@5+@GI7EHUj
zR&$UhpykWoC>aK{@)rhV)=5)T2JCC10UM2v#9btT9tsBI+#~;p!7TtO&ni4RcK7~U
z%1+y*z+&Y&Z95_h0R@xMSjdJ(XxV^;A>j=4uWYeUz^3{!I{5arTP~grO9XD8AV7pF
z5*xCOV+R|r>WPFd4b#FhuAW9fd8>O&Cda}Zfi>o@8Ns3!=TK3&BZK*`i`uev?qWaa
z%P9cv<qyCmVFx7C9AGl<WGG!g5@{fcIhue`Qu6~qN6C2!G<&s6O^*ZS!mcyddP*@@
z!o6QjLn+*5xoD=E0c#kVUc)ECpwF>B&@mTapI`N9qJ273?IGLx3SjUHj??tZyvatS
ziyaezN{w^MY2Tzz#<;y=6metWLEpdXAyjE=FyN1PPhGpC`TS%>u+oNmD_<}}m&7&7
zRNrWdiB=NxH#u-iIfL!@V^=FdyWddoc+j4mO7}>Cn;=QmH`Vc!nS_Aq3!|HKISYh9
zajD-OK>glxLll*f8|$+wJTEr20(4zFfUAF_4mz7JuTW?jlOQ?3=rd=jFIwXx4o2U$
z{A(Zqn&AUhhj+8G&1~OA0@$6`K>oZ2np|?()Ilw9jUoF)yR;j$XIZEh)HfCD3yD2@
z0B-a7R^;p$(F74r^hfncP2)gK(H)h-C<`nz_EPde$E#E1G^kJG?{|PVYUl$Uz1xZY
z`V&COv><014_03-dX20nnf7q@-p~WcDC=5)WsEO0Oft+P7%2h0ir?RO(TJMkGOh;{
z`{$9Z?RND)(5=Tpnd=7(g-rFtzMr2ZaBT#&IbJQ@ZMt^}#Ntj+^<e-<k`>vt2wG{g
zb+5?0k^18Zh=@9B%4%$zb9LhRmUVlib7vTGZ~)up8t)7Uw;hZ&Rh#xS$Nkp8CTN^V
zXm$pL^Uj-~YdVJkabIJ=*xytN$RH#EQKE#`)gACoQ_B&&t`+_)`<5Oo-W!Uf@TJcW
zH@Qhq0=aJZ#cFbr4{=Tp%I6U?P&(PN0;P@<T_{Hvb?HAaV9FKnxXQf#vV8Qu?<T#6
zm^RREd8KZK;%6#YL7fy%ew_${Ksx|jyfRL&*i7e2o-&vc*P>g@)qqSdn!j152NkU*
zt!TqvgC^lVmZv;0{eiSfy?|_x93-LowKz#{r4vdsuYc*0&w}CdVgY6d_xyP8sA0Nc
z-ujSZpN$CyA0#aGs0~<8IT&z*ktX$$F7N9q@yM}m)CaXvFa)0f#_p-9gOg1X-_0RL
zgJb}#G11{Lm~P5QKN5zTfo}$&BLVvw_UC>b2xh{bFAckPpkav7?^_Jw*<?4a$;Izo
z#wd%`yGvUjVX-fTL(r<nhmdveDYGlBzH_+EoblozMG8#~^yX`t+on-w0A`d9$OG`9
zfK5I}P76yy^%Du`020i5x1EZ1T{-2@{U3}!pC1Fina9>9>+kTV#<?v4M;V4!HG3d2
zr|YF?@8<>0#@2-`164uAuq>qW03d6GjFl;w5mczcI?bH8q*r4ABAfhnBNrd25)<<w
z>Q1`s&L5KPTr5bo8(s|~-@oN##YEZf{*W_8yNbuD@Mg<PSJ1wo57x^BV4}=Y2D~?6
zxGmqWQwBbN6$}nER??3C{xt*_j>dWHL`2PiHDym<qq)%;&fYX&(ZvQ+0p>Vj96QSS
z-7Na^n4Exs=gFw}_9_>lShizPoh2SPRa873OzzKEYinyq{!f69**QtHSfAH=JIYdE
z*R{tm*2DPKZa@er<dy~xM^0Om(C63Gi>b5DG2M#T-1)=fPNk<e6gca~e~+Zb$Da`|
z8aWUtgG*wxewFTFv^MA2es~X@5q6#&KP!Taz<Heu1kSgp0L5>$$#z&b0R8PCmfu-w
z_hBH1g2rgMYL|zgPvXuJ<rE-;!^}5-J9z$1JE#)mtBUHixR#?qNR6M2^%#2546uom
zV;#I^@CL)mB9}45r8$@n{5zpw#5QA&aseqZmtm`AM|gjFB;5(DITHZo7J!a`F2=hW
zU^MHWp97fB62NbGq?M>*`V|`bZQYs;i_10)r=;#Vu5SjE3{AbrJMDHKvflvI_o#V3
zO1`7eF#%kD1BmpKI}_q9N#vo3D8487`P5rO?xi2^i`oM8&S<;=QrDx!sh%!+oLS|a
z?MWgxe=Gt2Q>yJ?#O)c8X4bE3PKcGcpxfy|v}>m|wgX~$#sgadhMP=4J{W?qwOioY
zG6iM&C>5ofd?gQD#+zI_J>JwOvj89z$l!+b<5dG}0Y?M8;j}bYb7R;xplW_;{^u@P
z0)```2a$8*&h9IjKw_9Y>Fojwb-4GpaS6Q+F6zRPe>4BLi!$*8<Z^FqxVSycMEwzu
zB@j<@c%RTf$%XT?EP*nBML%x5a(VWXMTB%KYmQI3`x<GIN;0(vT{bazP#ow1rz)?+
z^}Esdr4RS`?&D!8uYu!bKm-lVdV+p+&~j9VX0cscn$(O}`eE7q=gHKFf{LW%$#BWn
zi&Yleo{OQ>#(2+N@B+;{s_GN(6CY>-gOeEf^Da?PCEWBl&H;LK&EcYY&hwd9K+8m(
zG1Bgg#oHYP4It3P4Cc4!5oGmn`Jhf5W|+wUWUszVZz5?zZ|~*j=H@;Zc>>q2?gI5Y
zpYAPzI>!QFC<|O}cwmmFb=rs*4+Qd};RQpUbVk1<2ODxQoyPeO7v4buV_9q%HUDiP
z$f=ltD4Dc!B~r@z8l!Q!6sTQbt>)foL;`<{9Pl*tl$CiRI{g3y7#@0W3zTb0nt^2J
zjjQ2nnV=VdtoQQlb}uldn(znj+Rq2`Su^?wr4<^xj{va95XBjO5!YrFRs-M|m`sYq
zI>nm}*)5_~3WKHXd*urZtC0>S0MoUuzrw_N&)w@fPl>`rQ1wDAdqAPXfciuj2-aCf
zf}<h6rn#VaUnBx|3pN&%Vcig1z4K_NM5!K!(t0MszPEc(cbEi_I3pee(8;ILoO5yN
zh9*FhV=!$Epl`iVfrjSJ2OS_E#EK^*xGT?h2K|6t{zIF*ZsZhl=-Ef$+x-;{cD!K*
znHaaF6F|Lb!Z|JQHUwVqA_g|K-J|z8(0zECG3p)YFjEZGd{yKE$Q##o1FH|9MJXMq
zCrcptkmDpe$#WR8s#!oXEC6VJ&v&Tt6EVnfG4R$6ZU;bic0yc?qM7R@=K1x2_b<8t
z6VWq^OeIh+RBI>RaA&;Hq~|n4&1A#EN4^xDi`s`X*L!_8@p7ZDo<^B3mwBd<9S}{U
zY~m$mvZkOKHeO6yA_KXtd+qG--NaF}Rg`0(0?5qy&Z9aE@G}Xz^spOk0EtyBl!3<+
z2_C%gv#P<d(}cGI02J`*=})f#9-<4t*-w=nVWOy^G4h$c{}B2w00!v^tnORn@;=fG
zf)20bjfbGyq+t>5*ZDxaY-C7H^Ro`pIu&-Fou5cXAIxJkT1aWxf`n^4!fhRSvySUB
zAfhQ{t7D=C_X5vi!Fh8V=51HhI25k`!n)IW7=;)F{MYk%ae!1wY<h_Dl*$@V?iM=e
zLgBD<#AhsnHI8=ceU};Q`%>9RWgcMMDgSODI~8kFNz~FnyQ$vNVqy$*IGtI)!Nm(?
z$mfm40I2cGI^SFtTLSQmeBY=?5EvGd=Dh*Akzt>$r;*BJ5Eh?RKTE_r_k!2&v<7kr
zP6aOl<9^|%)F#6Hc3gTLF@JMoEz-w5?v_#5TZZ=r_(bbJe7^pCxTHv+@xk|%#+RtE
zq7FuEvqW#<OBXali@GZkcVC^IdpvT8oOU%K&^MyFlw`;+=W6>09X>;_-qsJi+Go3u
zyjK;>7kNED2|AMIIi3z}fE&!;uNGK%@~Syuh>hJIQ1s~vfOcX44b(Y;%vpF80V0|7
zOcKe~n=8YDx(>Xl46lT1&HkW5LOsKl2n4z7K+Uob)ReaE=m+Z5o3udDb{pe-TLaWp
z!hUB8;B3RpfxBn}v`S&0UyJ?(ysr*)<l(G9?q>y{@$`}9$S7EW&HSeDR|Ek_u#NGC
z0;&~D0fuMk3@Nfwuha^p&3#rYpy9ZvA5v0o=Gb=gYw-MQXq2&R>uchN$#`V&UoK!k
zjb4!%<U*ZGC1?gPavi(z)4=!m`3C789Lu9wset=nCOc4`5MD+D>OHam3C|#SHjF3T
za#lWBPpU*Vh!R(>z$I>AVG<BOfk}72>8qss9SplXRE8P(Jfq!hPdX`$s~MxNctP*>
zh?RV@KaAi=3*6B@9=~D;&m|HlO}`N1?;k|4R1U}bmJzo>p}eE3l{DSK+)4wMVxA;M
zx8VRW9q{ac0q2%(a|nccEWZMj8}=@Sep0z&1-#5=OvC|HOsj>W+4D0JXAsQ4vG@zN
zaosQU54a%8x$c&D(a^9np~BZ<;nE94MY{dT;nkdq5Z8?homW~<ea(AcCW(PrwQt@%
z57g<uO2cC{%sh&E@nI3MM2hJ&NU@jaeF1=R3EJ;KKix9LpuluF8NR+dKsuJ11n(V6
zn9RTQ%w}8vmzSfY%MJjqE-&KY0=x*U@YC!hzR@HP1NzA(z&7*^@EGRXTJx026t~Bz
z+`b$6yC_O5x0@rf_-4z2g3D?O#9D+-dG2LgCP8>cuO<r8h39-ppOm!9x>VidPKF!U
ziT~1o2Vj_b=C7|L#0Mj^GKbOu5lI3CL2U-;pk!#lk)KW*tX@#!RF3Os+v+z?QhD8Z
zxbS_WO3f<DVe|4yBs~|rE@<$zzwaV?|9A0+#q<`8>c7c@3-~&nlq;Ee^xPS~69E)-
zX*K>j-Jvof8L2k*wv|gcc<-qLT)BYst;&$qfg%_Ua>PcY{&Q~4dCISFjf`vvmw=gS
z1$cyP{#24y9aI%@8!DKH2-IopmTI$i)f`gfYrq<%Q?;uxa}%I;@rt6vf+tO$L2vQO
zb$x%~=!+qiGSKSCAk?rvX6!A*q*^PMYQl4fDfXde)luLHpkQOrrF~N6<G@KcFjN4+
zoxn?&o@8|`YzGoSobpNU^<^P%y!f(xLx2CA7{~dG&%MB8=MXU|Syrf+%Vl<dhu-zw
z2ng>7@0hw+Qjb3Lbz|srWyWivuD!XRA<>MwiK*#)A!$|g%UijqOjjO7Z}(%i6dZJZ
zX{yJFLzD6XM7kiaC~jY_e9V*)G=srHd2y>X8$jgT(R1}&N-PJmY97-yf%WmB&in!a
zTiLqxAJwzJc9xj}ZJ;h+tE{O^_VhK@4WpES7=-#bD3LYQ_KoA?lP>l#+QlNI?Xd~x
zAFbMte=gB=;NQx#1cVN<jdmO$hpQufE^`WVYBdt$P=Gb)WAUGkSGtX%68)!XNfmaE
zGkw3y`EPD`|K`n&Xqcv+-~7I;g0_gDbr-{ii@t93+q+fj5RssdSRe($+tIOLy%q8L
z=YzLM?elNlyL0&TJuZ37a}@iUtW_f)icX04rk~8u5|Eb8*9EO<Xi8``VA;yYZ~LDK
z$cU(;NC(7yZ8NWJWV|Gbbbs6=ES%iqO|jKJ0gtFd6f|CCM94>JD~1$6IUl8!xsN7>
z+k)gTRwxgq*weCW$o(q^L*%c8n|P+O^POpy3P~S2#CWBMF?iR^1q8Ppw$%0iVe7l&
zsc!%OOA<v$>nPD8L^?JZrGbnxvx#G`?93z@2${zSrDGp^@04-uWUo-fF|!H3*Sq_^
zOW)t;kNffH;V$oUuJ?7l#`E=jJ}*ST_ex$DxPXj+^iO5N5KWtLaVG2wA9s=iLq1Rm
zNI$OdZwVW^{IHKY^}BF`?JiCVOK>(W^N%wmlUAi_Rto){(AVyws#N^Z8tH)SFGJ8(
zp9^T535>j@yx=`E=>XE%!c67y4z8RoO<RBBdkw2HOtCjVqxpJ~W_6dhSf@LX$atH^
zi`gS*!&6jp(D14lwJLZF>{MhID&+uY8bGVQXDgg)PYgKl)t*v$D`h`lMJ{wXy)g9F
z+D1y!W92C$%11H*L=@LmCD#OKK?C0G)hdsD)Cb)gxkXk_m0gKobkVyry*_(M!{G*t
zz`p<XFH(Yiy$9*B_Q^>k=KUJ<U+5;;KgTnb;geE^uqv&S2r^AXY?N#@aS%GWnLuQ=
za6p*1!dRP6d9V{ky!Ka>@x#^4>ufFjgHF}T&Nn9$ubDDP#3#J3(`s8D^4M;*u1gv5
z`<!nzWZ9+dl1}V`+|OVV+=^Bvu<cS^y`o>^*ZN~_A2I@fCt6wt<+oVLc6@eDUtgX!
zqepLmxWF#70jd1lOAfC?hF?=+O3K$cNt%)F6?C3+|6^5IgDv$t(?Xo&+U#H3Z@=mL
zFGL}2cL=oI(q;Xc7ytX#ff0`~3_CJYO`#sofmDCaA@kz0!ZTva5+INzdTwVe@Y1)b
zjG_iCVA2%i5bzK?cC`%PXc-Ci)Z!JO-B3;U0Y$>>Q^NBDZ+I)9t8k}RPqelNqOWE7
z9S)^$X+BJY${ee=oE_Py0{SaFJOFR8{r>qs0>W5x3^6QuRxJ+%AH8>PF?{1If;cu{
z?`@S~M&<K?NBMn9{W-DxL+7Qtq35$ekmzb=c1@MMRJ>pkUqe9n^;m#0826J_5pAoe
zRTpo^-mO|OOZ?)lJr^^`$EdwuYlkl>(>v0LYTkF+mYhfbd%==2UqTe+vV>;}s{zm*
zIrv4Q?ZeBkCGGu%qFbx_2gY)3CCZ>N`3w+dm|mi!#81QSFBL42whqA24^sJ9DpTD|
zCLvlZsA;cFRpkQXw35wz)ThiNE0jY6BZ)x!e1UuM1yLEN=;ciTnCrjhT9g%ktwp=c
zxkxV(@t)(qg!hqDd6X8{q1tEKtk4u5rPXdW1kPouo5^NfyvRroxBn}I6(PHHa0cO5
z?j9RQc(zbFCLule-mw<ZjWie>HsecXgoKG{Ihw1lV>1HE&prYQ8|85>-5ae9IbL!0
z*XG<d4XihG=2o`=!CY*o`B3k%cai45l>o%zGaaxGmb7j0zbAapG<3nP#;(J?Hw#AK
z4q&lS>*jaQ@q$)yt_Z}Mc->T}Kf>CjJMOhGSO_mwwfLvRd#k3af8$dwOxm?vy3m~S
z<6A=?G6Xz{qj$cioRxBl0c6S3fTXq^ENc%-ZGkYIO-4mu_-F~R&1|GJ7g4ED{?-dz
zwut78BK2B7)F_+SkzO`kmGj>bD(L%8$nG$X^6C$4zs$hJXr7>#dnA|5wS%i3(jPaK
zNObI2Qq#UhJKEi(*)-%n$KB5-m0KM)|8bY5c$v!m8ILfzK6TywV>?Im2)p1g;q8Kr
z`me~b-xcuwX)l~yOfh>)ebKL2pm0arlL!u%Y2Z4D3(Pi<lC2J3;o#tC@=l<Cp|+`T
z#ORHz#J)=!`HkKU^eXDL1YdE%wk?+ISo|6kIyrDKdX*o>0wag>QRsON_ua~`FCl1P
zrqSl=iXk<rf;<zsnlIVV=$^;y6|99OKbKP<b1i@T^7=0;c&RhMHa&NVbO7E~49M$>
z6(0%;{BxF*PtQTGra@!Vji{+Rh|niYjkzf+^Km!;ZqS|HeaO?=Q}>B?SYLsahhUOJ
zI{igPTBQnY)RQa^AFJx$_!$!W>Q3zTIoy20vH6$9=0D9TL>iemAm3SP54XL&RdO1~
z=<)M`8Bm~2EOii6Co^X+&^rV~j*tBO0;Q=xb+f%3SH}9WD(mrEGgAvTt~(3LI&CPR
zru@5M?D3-l(90ciRY^9_I6`AnkPB?QZg|saApAM{<g^MD7YUO_Xa}GDT?Veo-xb~I
z3MX(;c=E?xr>|D{k{!t?AnU-dz<+%UefF(7JD#@x1w3Kzx6rRb|22xpp4SbhILxWp
z-7GkMGD_5~I0kYlpY528SzmrS6@$=cgtO4FUiBPUi@bZus89X8ZvD;QtB=3R!Wx3>
zd2-Fv1RR;F|Bg0d5LAp~n`YwFJD<W7yMGGWU5D+>l|Jx2`6XV1P29RSWB+{wjjL=v
zOg<K(O+$IPUm(fO?TUar5pb8HDw^uVsYY8H^BrG<<{7vM(f4@2+Y!(q9XFD|+0TdC
zQH{o-8^~n8j5)!ep$`t8z9`uVgyk^5pHw+L%InKgO59|*B3G_;yl&wFMU{0mH6;JE
zd!@DaU3NZ5QWMhP1L5UA#Xgb@uaQ6;%R}%*K;#Xfk}jwZX3;?46__<-Lpkc5KZ;EI
z0{ck3$f%spl50^@yc8BG)Y<^ttelQC>zmESm6LzzbkD$&-E`2~2?E2v4-|)7GyR<9
zbtX(Y+Srtmbg$CfNAi=r=-RpX(JVT62#v4!oYq1^{L@pt%*XGq#_Mi`o!~QH{R*52
zOPYpqbFV)yW*3Pd9g}qvmeJ*JXC~Uc2sVbypAca2j)h@jj6>woM0+hLj2Irc97PsR
z@Q47Iac&)C({LfHEfN6hI!i_}i5SPmi3dG@v23zoO4ii-H1F?s^Pc3qbVw!=>%ZhZ
zTo^+vQjrtm556Xk*N)POa{l6=s8LX!y5i(C$IvcRxOt?JD8w1jwKSHNZ^ZD=$^Pee
zewBfjV&p!xW6QyJG+@62Y62Ii0H>f@DuDQ;PUC(Lj7H62aMm-^KnDL95ea-3Vqa~$
zH41~{qN**g$HU&Ys8Hwxr8Bkn|C&P8sl!9e7(4%_#uPDILyrNbw(V0O=>*y6m>Qsz
z(=Zy(gw3icyl)>uY$~mMw0n;`Gz?DdOm++>lu8`5V4TpX+f(i?@|TNYpfKQpMWc0h
zcODdE0w*2LX;CofoQlj8L5ZP{z-xIMVlGA`Rvn+W=9Ssl^l4w5%kZN(f|WXR=o0H`
zT8!T(rZbCHuKqy=lJ7o=W7*~nag_1bFG!saAv@Q!<l&xLGV3A~V>=_b^4et<Ro*hb
zK%B~fN-<+;V0>`$=kh|@P5=JIiS|*U*qbzuc99-BtxHA}10OC^LmaWpVK58!_eNu<
zeR{YgW3KluQl?$I$;Ee(Ab<5fU8S{u)6=)epa1)@<bF@YAqprx{chlODK!xw_bFB!
zYV!dA;KyM8I0DnSuaRJpRqQxB0~;-7D_Eb5=8?EGEmw~HZJ;jBva4l(J<pe^Wv<ob
z*korBs^De7_17cxNsi-9To;ZD{eJ2racA1V9%z7XuK+;;*HwwKh^M`voDxz35wibl
zeXtfXTFum}?gz^01fo8@R>I66v5TbVb|qS?wr7?h^!QejOjmq9p13D$cY~uu&qk2$
z*p8=4ov9Q{2_}7I`QH~szB8Ob;4PG4!5@m*su-N&&Z^kH#GpKZhftEGUn|rgZVOBz
z8^lwy<Xf0VKU4wPex&WNiu<D~w>O8jn{t;liIEfQ@nsIVya_Ce(|wKw@5?zsO7Uxf
ze?5ySm_oPE@W%kga>vJEz3~NcF_CzB&>k2tc8F9DCNqZ8{sC2oFYV?QbM2Xcebj)=
zd1pdZwI9I9#u3ct6y^{5rNS>q-HMZs6|D$oRUDJqC|mHu;{5@C&IsUW&oix&4WYQY
z<N0yV$aM>k(i(H_`TgSIbI~TLX1qqQe8jn!UkiA1dykWwOXwmrEMQprA+xqVy{8PC
za89&f5&8lCC@`r#f$W{V$QE;MNU_q#nYj=XYaqvof^TriL3YX5=mR^R;%R>V(QstC
z<Liez7Ub{=PsEBO3a?VrX2f&9_-J3c<ivqk7|P1^;ANxu>=7D~8L$$z`@4&MSw$?#
z_U$H5IJTQxYa?pH8SC$CwXylb9Rczax+srx{MkdhNMFS6@-dqq+5;8ye|wxM(3=#o
z;N^Fk(*?4wx>d0|pp$Bno~+|HbU`BE2$HHeWMqr5U~~L<9AVR!zzpWAJzMQ*_%$D7
z<3L)afAMYk^O!r$GF=6*#P0np?KFR0Fs!xpal9+}&Yg8fh7@5;uK;byxqT_SfzuY5
zLS6QM(ts1%?WVG|F|=K5hA*mV0(Zg~n7WpKor9dD;P%~*nuJD@jsymu(dJmZ5q3(o
z0VfqUu|CU#5~_ak=bSzg&w|s-v8++KYsYi5v5)+6ncBh2b*d*P2er>O;A$MOJn$|n
zGc$0o91x}zl$f^1wCsYS_^q`4as)_WshHeJ!X|_|)ME|RO|bY&_i)6H`B>T8zt<4f
z!Qj$v$}T7b8c+WD;!HQlqro6^2JsR@X7gw*P`39SMxm-DZ;${8U<20cb|93Ffvt^0
zCKCM91?D}kUQ}abAYeq&X7XxPWdC^2qJbk`Yp`Q*g)x(ZOb#oT=FSHn7K8_%xz)TE
zc^m%*E~ZQo&vrTCCPX@J3FC7-<aMOY=qsmnVG5Mc1$5fZ@NprMe5c{Zc$m!Np$M}a
ze(%lo7}oHPE{Tfvz2pTJ?8LgiNaOsY&|@>oU*slsoEkV|GP|mIy{tGDy?(FyFe&h%
zzX17vD0XN?#j@21!^W-XHhu(-aHB<r?HLn@QV;D%12VKq_*Q<xnXR>#Yn3Ew`)CUX
zHCDC?8gI=%sYxV$<1GM^zuSrSB^Y}KByNP5&^2<mZ3Lm|BuB(wyto0l)~}d{96PAG
zC!0k#%=zzscy_M&37%;e94cKbzG4P8RU(9*tW7tg7N>r}?|=_|b4f7Y6z8(~y-r9@
zT77CGbLHfpt%JQ0XEjtlla|vg-J1{+^zeK9Q)eybiFXf|g>Amp^)FXF9M0f1{cH`o
z*JU%sgN&m6^<VrBN8YlY0r%@!N!d3uIhXvvp_dW3{Jy<>KPv-IOW3sJ`Nhsv(+&;U
zG;nN1!_H<NnQx6j<|YV9LL<*Kw8bul+eq&FaLC{c5Z7YI!6Pr`)9A$)RY%SrNS{uj
z>;ke-5u)auuzd){UkSp@zJMd}!WX7wwuU#FmOI2wW&Xmdw_2*Jr~2Gmm>Sl9LS4b&
z$o}k$U7oJAmDWwi_~ek+{LhA&kPz>QZadSNWjy|5hLPJ)W~OhVJ+TCaH>X$JuJRc?
zLS~0E(0ba8N87$c#zp!LD*)KnYCTp}!y!&}jawrkt1z1_0kOu%zAG~hMz0^@__emd
zSsSuOj4=O<tw}`LuQ0O$<(m*RmS9v+47g75F>R;QE}mSraK>0kg&b!}dkGP-R&d55
zCVI_*nv&7ub^F_^Pl7T3?QG;TGf6<iHmvhEp#PmQnCN$7!;s*JzLh7ML>MRdnSgUE
zGQawvb$=ClQCm=xzXj@1G{TYn`3-hFgT?W_=~ynk$H7u3@Bk^V%ZQ-T_hbFT^l}E)
zhJ{CX!*DX#SBah{1|8E7(yU84s%EkPc=0n(xLPATF#gt87{l6P=GsrOmEgTTwZ23|
z%w5Mk_=N2zzp58lPKR-7MgE)rvuL@z?A@?v3@2~<66=QJho0rRPa&Y{n!0PgXXO3w
zZvtmb@-h?vtPT>#5B~R#Xe9Y|V~VC><|eA4bpF;cC@~uR36+Y^`+gc0rCjp%Ym1AX
zD||3IqTZl9t_++_WQ^0740L2Utg}jXD>7&|odgxo_mF86>p^hwtJr{e_fp+mV6n6^
zf~xs|NLZZ%jLTx&)}}7D*NT7?_<1zFqeiIgL0+wKAV01Y@ibbJs8g@PCeKhTnntp3
zeWeH3uH~HApYs$l{`E?kt|q82?3Y(`bco+cH(b^#@|CJA_fCcc?~F}>r%Aaqxxd~-
zQ~f@MH2&h5%-@0L7S{LmjW4GSfy)9&`0P66t}|Os{JG^xGIzY!5pgp+lspLY>XHe)
zmDs~P2&!t9!N7jwa}>6dz?lPQ>!(l}<VVZz@zeQ;wjE)=S8RjOt7pdPjBr}%8-S%^
z#J(ZI9B!lL#sxdfv06I9^@G=Z)}V)(QNe3Y_*-C(+&VB2?vz2B8~}&B#v<aDkQjD@
zC(M|*Qe(@IRmVPJLYExSx<NPLbY??|(*>%(=h258`vx5PnnZroos~1=AdV}>b|Dhh
z?dX~T3M%CLKcW2&oDPCqzkZ5EMfG52+Pc>J$ho6`(#0+kwObHUK`B7z^Y7_oIzf?i
z(E{uw;J%2ZXiQ=uFd0jr5?~p+5W>nLA#*d4a7Yyfa{MQFgR<X0q(U-wi$yln8e^7E
zEvG<5Gz0bga>?VVj@HrmKE$3EVMC>A6yPCvs(+7nwHUIlrka7jxZu`H@3<GuuH&(3
z@7STNT7<9;8Ng*159TlBA=v~mc3&rG04#yoiiZZl5&*!W(a?D}*lNW{!NTzHYY-kL
zdgV>Pxjvcq`2k4AwSt*tT}?~|ZiU8`eR+Jq6>Pw_OsUkcrT0S{TO$jK@d?v&<f4fn
zzxzSS6AtR>V$VJMiY3RA(EJ^$cWj&?A<cmI@dJprbHwlzkI-_J6Da@vp=q7@qI*W{
z0j=6F`<f?bnYJ-1Ecb=l|7y>-;ugfkRCuvfVsjakw&S(o;;IjLk;!C?v`H+K5?fa(
zFmgyDeh_o}EV2cI$nOabpHk5-3}+A?l<ATJxZYz#-s;9=gxg=IQV}W3okx%J{W+CS
z9uXD@na-f*cNNq_IjPWd4f&|wky}MDT(<PoI}-{Y5kTwcCGDw#1M9U=bhsTuSkwIP
z-YgwEjz@n9p`hm=k;sqhNx!-c2))Dd`5dD;paQrAqXmx6rc%nIe87sojK2z#!bZIJ
z%?l$<unxI<kxoI4AcSw!$!c4VMpG*bErT!2{P=_@M5S8MoV(1X33)oaDwwZqqpDq@
z@q)ky(K(}b9%tX|v9%D}PdVr`ekgSSa_sbYGBL$q5m8UeE<p4Bq5k;i-nWra1YX}J
z$i1^v!_ersmSle_?wXW0OipO|!k$q&Vs25@U#|Q@-2zC<!d6@jpTnfmZ9|umRkcO|
z53~gwhahsP``OAC>GXnghX^n{z4^sMSyL}6rF*4K1E%1@d3D)I+wHuKCiRV^D|w<4
z(Aiu6>LOrV{5JB(^F>(=2=l@e-Fq}0u@lY}$|?dqY??Aokc$b1EEI-RJ5t*2aW2iZ
z!@u)`=S8MCzZ&RHieVXtT2xT}05fUg)gsABuzH@C-`TvsdcjQYX1sLor21Ze*zFzA
zCx37}>(yNT*}wN2D8r_>k6_N3aCAwYmbiO!L{516Yg3-2#g80$wFrWY(uD)|!1SlE
zKdZ0E9F2C$J30j~XqxI4stf%p6w@0ou}>#O4X3MHOwgpHG{A<NX1JWSZ9@lWg3_a`
z{=!=YoL)j^-bz38+L|q|MF!x_O|&4_CdACVyM{^ha$%VG>9n#mJ2a#tF<oL$90SXw
z;-+NSuu&~D9jV!7nCR3exntbMtxT~}=?}FYd^|$ib@J$0x0-OcZGu}B2(5;(Z4Az|
z*aiZdGFeu(GNuxUK?6%bjLHd@N=^GRC;4^jN$iQOE#Q<bo3WZzG|nLw@i0MDb;r{y
z0uHm;5-^qx$9}Fo_=k2+IjA>%S{%A&cZKdbW<&x3OWd6t^7F?n_Zh44*OnDtI$2MP
z90>Aa(K}lxPdgT9Ze|<<=YKaho&w|-g5_&;Y1lekDtGTAwW4!o04j%720Do?@x&89
z8B8;qLE)J4OO<!p%AEX4>R!q|6EKDuZ-`|x%K@leG?=?kF1L%xOxZ^-fm}v6RCxe&
zhG%)}SIQAx_a5pitwsmITIPI9(Y|A^1cmLs-?3t5&&h$p#1i&6r?+aJsj^YeHPN$b
z%=t36y=9Q$j7ZmYpy!A3-|NthcVXY7vJfi8L8oZubn^wANFfZwJrz`-X@)`x+FUtB
zZ5}ZG8$!G@ZfJX6>LnJ5BEZG^V$6uUi(yn$Yf?50F8Q6CM4jy)`avdXje{^Je7Hvq
zXO^JQt>ZfH&Cj_#yS>#iv6Why(|mvm8^botZmW^8j=eLMySlfhXV|u>Ws{=Ja5Z0a
zfxx74A@v~`F*IAexjZ4ts;_BQY>M|j0*53+F|63nDl{={E!F<nf$S|SJ^xVm(*W?Z
z=t9rAu`x=q!C(_6auE#y4V^4zdXs#|iShh;sM1=NQ{MSc<&KgS$G&2-N+oZyq+(_`
zu8Njc#__5cr^{k=e5A*^5T=1FKf8Q{hISs<S7a6cs6}r6DHwezyiR!+bIHf2W)mQn
zH7zu6FatU&lp^^KE=i`^>DshGWqp#$7$zZ*l{rTNn{hSY<onMM)EdMd3&)gBGUW+9
zZ^GKm{aOnk109-ah0ABYbv^0WjiDcBF&G?xn}y2#(r0`dKS)kt$jxMRZ)6@dTU|0>
z)_M3X)V+YReXcdnZrh7*xYlchaQL9K?!Jx1lr(v7Q5N-Ee9wLk#(w!Z*Tqz=xp5YA
ze-^2pQtx4cM%NmjuEf^4JZPo=KwpapLkv{dz3yQvNg=R%UZ?2x;^yy}4Xs+<C=61j
zbgDWv#u{VHm<u(|?@m%&2%1aT@sir^-<6FxE!`c~Pl(@&5u@p5=iHr%At=Ar2HxHZ
zI4D-yv;mj=2%}7f6>gf*phrmF@^l}YRyNZkduQ#XP;ylx{Xt6J6~ud|i<)YgWnl@x
zj|SeKxp*S5wHBsCx#U)Qj2J6?QV6G>{<`9UV{ba8u6HN9a<ttisUoxDn`o^(=9Fu;
z0R{<39^P_F5PQ54mv@7X3d7HSK((jj1z1w_@EW&PT0Py4Uotp?P7nP?FS`Au3~hY`
zc9`}DnMMGxIvE3&Zso3#*>)7gFEq_=KIbFD%%uFxyN0b`Lvbjoc2S_;2MCOi>zop{
zlWAMOUE_`^mxp^o#QNjR7ohZAe^p)oJxUaoaq6@#3a??VwUEEO9^F)25BklUa0-&A
zPFKqOUb9`K*)hAeSDTjQcP;!`UJ4p=8KuzI{vqi}c!N4N0~@e}?++(k7Jk2V2YU~T
z54xrnz;+<@qFQK_^4bX8LdC1lZn2!4mf6Xm-1$2DyfU8(Z3tah)w`@?Di==J6GB+!
zQ02WZL;O+B)0F}p>UmYpdGWQ?1s@gT(V6=(i7Bp`+1{Z|DGw22QpR@1c(qq(ONrav
zlTNEH)3<LvtLS7>Hcaij*%6tBUQ%inyz-dWDpZTU{sXmD#e>nCR!j}L)Kpk(*0osg
z%3S{r&c%D~v+~A_!+q6HsaMWtV=STKk&0LN<g9zOCx7HHD;<AnhC!L?gkfk<=eX4Q
zEahdj*wmSUk`0CC`r?C|kjb#Jm6V-LxYX$tcasJ?3q22N%7Th;d}c*#ey2jq)C9Ni
z@6vgf=S3vUeq-E)qyBxwN6yFjH6bm{7@?+t-Al11qO-jeY?w#Nu3HLYTn`@Sss!Cw
zi;)N++P~IX1pMw?ebPIbmHCpg*)W6dK-n*!HD-P=<9(p;Zn^8j*rKFQ`Rde4WJ~YD
zwdu1|L#qI4^b*5e)JbQFRILl4bHQTwqatBSvD^pQqp#Ucxzm^i%~Sn0v(d<3-i)+C
z^K)VDsR=E|fp4=$g%c>Q%*G29`KGJK&z7kkHBA~uz%ik$fFsYFc`H^}I*;4Q5EFq?
z^PW41HoXNWeFHQW1fwuEUrlTwnkX?Ip;}QdfcJxYdPIAn=C;^c2hy-@05r#ZL@}08
zp~c;R$8dU{Y+sVzaEbHVJXQ8N<#1X-SL&4GTv%faFZZ9gNAu{BEnrjof1m7Kq$6jz
zz^iRY4mU(6e}$G?QHgjpMJA=WK43WIK)ty_y(Wgx{8~g9f7vu$5tC)=%Eo@N+8zYf
z<)jt>Z<J9SSG~2@*sj+kaqV*5ITeQ)#Mmo~AN2@P?!`t?_QJ6Fy7@7a#6}s64jC37
zk23=$w;L|C_d+V~vsUviT^+qLwmv1poX*E?NMTun%eKW7!{E(2s(}of3@+OBS`GG5
zWDiR3`WR-vWm;j$$V#$8MJYhNz1f>Za8eG-(azJh0tE1<#J-11p;My-=djaP0N?|z
zvRJHrf3hOWwykD}T0#BmW19el-na#=n}5EK*#UO@?KjwT)!x~hq3Ow(7nY=&6g&Hq
z8}e}mZ91>mul)gn5>9wR+x>8l;;1N)17pxqd7MT($))M{!B1@l`tF6kwY5>Cy+{N%
zmy0v)UKdp`)cAW@XB_+4_%)T{71E$2PrT&D9(-Ay)KEWD#W>diV6nxX_nw~g5sC5U
zm|{`zI+5!s8?hpe7a@g;WAaoRl<$|+SEPn?K#FLUpE3&b?$wK`7G|duj7cd&^40;R
zfWeWWhS_H}ef9Hiec2voop<(l<39^@(K%eUwuA(>&RIbBYhBO`Fp^fD8k8PYt)uCT
zEJSP(g@MZI*69E@<+&zLjnLC|0igZK_mPNsUR4SgPr(oiW8_5dJks$8YIE!6{?#P-
z0m6E})c8+<n*9oT1Wq8}wrHUe9(s4iQ)Ul?zd@+#LQ<`=yK=m}-t>XKr0Hwab>|ev
z;-otWt07X!ly#uayOgUIZ;1#4<<!bwfMAxp;IQrL4Me0+al1w(_fQl1CE}|QOM&7d
zr?kLa19od_(Ex#xT1FMs5~lYxVB$6FMnU1VQR+rpRBM%bE`31bC}OS>pX_Rpa0i7+
zS4=f;g>w9ssltVgA1)qmsKy{m%#SUAQu$Q6l;zD_nI9uBp_}VJ(GQ$D8<+=Y2?Gd}
z&)w+j;-|?7NQD*vqs5&m_j$5t-Ss^@M2vA}U+FiJh4<e43w0_4h`+}Fh0pJ$VsfEK
zx{~$=K8;+D??Z3tdNMOS+uOTD#W9v#zQk*&{O1NmWIbtMcMvI70mzpnV4ur<H<oH4
zi4*O#yeUT|<87+d6II4CJDp=h-JC7Cv8n2P?d6%v8y~dJKRZJ3xibHOZqor2Xs#4`
z+cjH(6TLvWh@AkjHGSfg6e{#13QY|au)Ebd8*vrT(LqCj<@2Q;87)0P#T~spMMec=
zHR>)Pv{Z~GFW=Ch<2R0N?1GxYSYKfH3Ej_lK*3jS+J}jgh~#)hd(t*&e|(F4z$p>e
zqEUr+(hnE5UkcB3kDYq=UI4X+lqoQhbZX_Jw9BsB9w<&es+^~mlZj)Js3U)>3II`V
z?yAny{V8zb_o*|uvO6?MIqcn;*HfLlBm)^PC`zHC*C;SA$S%VjoVwgCs@Rt*1g-)t
z-O90<3vZ}U%Idy+DlqS=Wo35+s`l%Q&Pe68PEb+{hf=7FnE^zSuhF7NG&==lf*dY|
zQU$Yw7+DN{@62Kf8A#>($bc}6wOUk}U5!n2vA#gNpCgh`!@#`dw$IYuW13&am%v+R
zb$AF=N9BI@cbm^c%>H!TKHxF0e=PXrdjy0&N96S$8B2kW(b@JDyj@B!02(Vn_)-f{
zB#&>F-L1F&)kkQPnu{#lSgx!Q?3`gty*Kwt#4aZ<?W#75%?=oS7m3Mrz(xEmk4pTR
z&waeA3+8*G)k<A*AFM6F%Mj3_QbwD=(^217m3{KT8el6I==>kuie==qbnb${{g}bg
znt`jY=m`RfYNPk!UlpSnk^9TK+(N5F>?f6U98@unLfh6rU{?}XYSXF;vw#7Zg@46^
zd8CIqNtq4Je_jAEyq<yF^TVt_o#K`wTnSMO6?uzM7cdI0nAVTPRlIQm(NP(HxIIVc
z(SuFuotHyn!=9&24MK1BYocN08<eeS1>&bS$H(}(>Si~4aKVB6goD#PIf*SW{FtrS
zq_1_3yV}#}EGbUSW}M!uyV)J6=<BA4B|=f*2mrVZ;Zw(+-J`VumzEz~-UZF_62(2N
zTj&v$hr?)<yb=4<!8;sFx0=Xz)LKk5VxEWQ`QM#K%G})Zb4}3ex~>AR_mcw&<v;74
z{qzx~D&`=f`mR?}mg;i4=0of+vK<iJm>LCcn@221%0g~UYW8`c$y@2*RormSX*v~y
zo8W7T%o$#L*1Tp58rq-}b;lBrj&%cYVmGqvMwtPA%(+AfSI~3J00O*hCTu3Fhq0{{
zb_o5D-!f?)HBLLhrCI|`%DDh>NL-xxLmv9+wM8nE+vnI>E3P~&Yp2_JPF@5cgRHD^
z0`#J7fM|0Fb^e>F6apVf3$c6g#K?3V{!kbAp4?bO8^cfm;k5abr7gns2P}!)EN@eT
zvTGWfxsRj=aLdl9MmvT(Z>~QNNPqDW8#cWz&({=Nv?I$IFz=S$Z#3qHQTZLgAO+T)
zhZ`sJ4o<>MK02E=kMrVg&&P+7?!-Sv=7z-Hyi^?*6*HCq6gMtzz(y1zz%VIzA#XYX
z&gv31p>74`f_}|Qh7t)MdF;fFk)*+H80lOrte2tv9R{E%jKI5+?e}_wNh(6<)reD}
z9GN*1-uUteSLu+?^#(y0RrxnIFMoc?pTAVR?}i!32!A6ZFeN+yzu2xw(T{HDFW!5O
zseK6mdFJ-KM<)f{Tu1=)bO>gVcVIjA6Kr;hK%r9t7^}mE&y*+hK!FmK1d4*o?MwCz
zc4eIeN=Guvov-E-7G90ekVwtJ-{mN5H&O~sr(y31JO%pgLswFS#*YjfyWnz<0w8>|
zczHl3o{y1XhmA}m@9s@MF}-rrmRqYLzka0*ibHxM22y6qKff^WGJ+wGsi0|he!UO3
z#BpYsI3mcXM?ly{GxR+7vG_?~B8-6zSHFNNaO#ZI%w`a+5TKtgMr9&lZF+8g0d<BB
zhFHXx^)RCwV3`GQoL!;qsd==va|e*Y(lr(aG$L{9cPB_eduLe>Ttf%I0zwOLuYRac
zLR!~&e8eyQ8t?)9X@hu+A=h4lUIrfpevoO30+1UGSUrK1(p+${%hzIE7>q>cJK&}o
z)kcNGGJk@<IFmmlPTYKx{`VaZrMW=viI^B5K0)nyOf-W0Y@SeBcfhq_bqLA>wUN~f
z0;5Cg`Lg%*9ylY*hhS%PZ5S308k42~Q)y`Q*L0}owios{I0M_`EV4aH&HPY1zoXO{
zYlzt;$hEvs!ZjKdH^+*JNr|*mk>(#V2HfNXu>CGwHvc9u0O!Cij?@*YrJo$iZ0t0e
z+B^#(R<KEWfSO#S<_~TP3#bqS+azTOj3NHK)m@&~m|F`U=*LdHm{6~a%PHIdQf4u7
z>mk4e12C4>IdhleRxdO104I{#leRJ}4R&tTGF|y_l1lbo+_^@Q#{jLJut_}5vU79p
z4|*h{iYZ1KN~2<gi|*~jTikq;1csm`$e{(>>BqQ=;0fDyL1zi%;Lz;WP;?CrJ!@r&
zBd)+i*!??l$?gI4Uys^{Jjy!!_}QJ0dhw_dps8@PV#o^pdc{X~FJ3IUjtG)~-p)(B
zA*dK@oUsyd;d1nzC&$PQggxO3?k$rF_GbQr(oY!?_sz~a2j)hovHsg_5aD4jY>C`|
z^7}{d0WY9Q&|3Q_kALCT!W1`0@n_WGo4m7MR1Wd=_2N)C7g@CS3KuQ%FA%hGTs@3<
z-sk=}&`QgI{xYMB9B5>uh7|PTYzV0Ft)#m9_m7a{ew_rFUH_{8+*kD)wGlXV1(vP^
zM2Ce4SqzE(m-}pKNY>8H1;;qPnov_b8!S!9sDB3eJ%zBXD=^1<qf8-FxTk$Q`1dnY
z5GblcxVmuXGw+a-LN{S+WHwR+O=Zng8B~A~_{$UzxTHus)y29ZdL!7PaFOa!<j5yS
zYcUqcbg{tb+@{kH#qI1#`e)I0kt9n)?VvnMIr?Yf319UUMF@vRrxw@~Tf&sl5^U-p
zF1rI$(5c-Xvb$(IqCkB3T4E3R*WgNShu7I!;Spp}uhG-S@SULJz5}B{4azDWZY`?{
zz`!^OHYa6vzN#C@tNQM*cki!PMM(l8X9u`mIh;6DQK3<yQtzE5fg@IW?qR^0naiD&
z-M>5ZaAz^zDk5Y)rMGKe1bjUvUyAPabRwuJYH?u`Df`eqkrU3wM@Mxhcl>Hr<U8K5
zqOAVwJL=I~fO+RHVh)6-ft(i8iN?}V#3%+-hEEXAo-2+REol2g^!ZVmI!4cL4!uRV
zZSH#4Pzsm)=dfe4BObvZJz}`GQK@rL0qh$c>hGVFVfeGUOnZ^<`9)Swa>s|ExgbNr
zp`L3E0zG@63EDu#XQUZ!+MTIqxq)mKz2X{g|C?)PGkfHSmO?`b+c|RR0li=A%G?S*
zy(>-T)8_BIsq-pId<-w|$L+`?fsY~VD}wLmxt(GEFtaCozmk^>!p=F$*wpzNHKUks
zvRxpH@|Edwif_X+kuvwz+sow0l*#JLe3Jo~FYp~{qn+CaoG3x2%Qr38I{f0t%G@2L
zl0qOvr~QM6eGv8sayF1-;`yE#<Q2yXK{WU|{D^k;^@fp;#~|o-*VrNY*iqpi6=WG1
zr*r9KFGWORMbTOK)T>NPf$5|`ag5`AU6EzprEOUJ`c?YL2>u=C&J-#b+4PMn<G;?p
zInpU#(=JecK|Le{ZOHZ$Jj(P}Gyz21ks?N3pZuKU7e=#7IZKv`9PLIpMOoyVdH*v;
z#|?bW?cCx#p-(JP*<Z`$FxAb_=vT(&DmJ_0Yck!r2Mc<@_u}S`OefCtks=zz=y)*c
zLk`9aP%Na4bJ1EzaS!IjraIbA%TqnSKEPKmae3i!l6E;9EAGUvh4m&~Uz~W$;?{f5
zWY}%GD*i6|-X2N_&wU<n9o}s32CZRD7{&6!{=e4s+@oRj2)l_LiEkGPt6Ut+YQdCJ
zb_QsOl1>Q-#S3xy*Z{ShCiLe0Fn5WuW{gIPi&Des0RBw_dSmBLO1;<Njj7#?14E6s
z9Mww?q8!|~@6EO^&H~EB3?clr$<3+y<@}$}lCTP1<U&B-{{mn<LK(9Ap!we$|Mm0a
z+i`HQfgBbO&8T`gWjm6KL>3ykEn)17kW7$ym-IkAI$P%ST)OO6f%X%)d|ZKS=&$9#
zkOKT@8zsBO4WA*`ss0p`^T)y2qmSe7*6~ICb!d1XiFo^jz1;bF0#A|VhNX7eZMJx9
z1!wGr7mz#yR?GIVF!nGvD%yZ+A%1jFvs`7`P2zE{Z$XASGNhhwK&kK75?rs8Cd&Zn
zuK~-3uz4^RubyM?%esm;%dyZeG;AC9O#c1SJx9*b{5+d0P+OU|^YObp@1L`Vky|kc
ze_Mg$>@ou}qq$2GVQ%q**Q1fy#KU%ba2i{id>YIYxXV+vGe0=_oyn#4eYl2SU;jMm
z#ndG<`{^rfjh(twU58F9UNYIB?ux{p-st|rxPf>aPXjA|IL+>_eX)4Py3eAZ2zu!v
zxXNP8<qKkY$**P6bOl&yz=2tt@ufQrkEQKYA@XqY`|~Vu&wFTjiHkESQrt5%6Xe#|
z7ZslpyJmyC^Oo1O*PzIUZdhc;>;6PB7u%X}_MOC^GY&-%qbG2z$08GuL0UPBH8_kE
zBexzvA}VIb7ru9cV(8xoL$t=5MVe0DkJRlQK$aYJv$C;s|FUjCr-elfciz1r3$+}M
zaf9O|xE+Hr2x9j)-}3Bz+7S4A=lv62&a~pSgjGTNRS}jNYTUSj0iGpPjWdFYq9B%>
zz{!UnyD1*mO7YWBRiK)uJHansUgl{ek|eyyBZiCLcUqKn;BuLa)YS`k)kC<dm(oBf
zmu;6z2AALL_}XY-&oFzPar+(EV4E}TAxG{aoPi;ZJvb2ABaLONAK(`PR>diFSUk`k
z)jZcn4ah%8>6Yb{bbfAD8d%Iqjz`i2`k6SC7Kz~HJxm_;=`9eVyYIhP3TbVILn_4T
zg=<mT@jd@{WF0wU4KA}qz<C-Ffyup`S=P~W;OI`2Rc*L#gLpP=e5E^o@(|61LnM2q
z*wE^j_cA;Kx=7BTJkwfaJKh2QJ4GFhor)KI?d6(3pD`>^|G1a1->9|Bi0$_&V>*Il
zs}U_dVAuUwj|<EsB~#hHzlgNJ;T(8Idv^bo7ee-vh2WzijKw2`4+J50>sY&8o&pr5
z`qPuxu_&_0Q<vvGbdkLlOujhTVKxakolEYfNPAynmn>g_g-loUlxir7Wye=%x?@bb
z+F}?V|4#p}4ac$@^A>?hXFx?_3ki^O?%Yd}<$L@#UnNz+g6<a?1Hk;Fh4Mm=#G^?=
zmYCa@M1nhAT|07p|5^V^@XXC__Up3bdsl3j>5k>*)iLbO*0~+|GCNO)-KaGLyMM>~
ze}SZ%tP}68|N3nGq`(nk>4N!n0a%!>d{&e{j}d?tBbQ&RA%t~QwhOr0T&$a)C`d<`
zGT-+`W9aV-wIopMDAOTm`*BYw^I9@1j9wO)4CwXh5}F+H&}$(_nlm?=?0!E-oDF&O
zyOOGVzkdWTBJd0;;%A`>28Jsz1uTK9&RH-@2tN{W%@2<612Aa?u!?i<bc8&ipED@U
z<+JDQ;4!&6Jo+oj=#gFwwtt&G-Rt<2!66CLgJ<-X3CMnbnrml_k>rHt=If6H{+(-Y
zK)QS5O)qlwWccBkp$d;rV@eRsHQSdW&F~Lq%MztCNJ(G#iHnseD7NEqj<VsuO0qo^
zi&H%dF6`W{De%x{T;lpFRihtAsHY-EddJpc6^E*9uO*4Zj@<@D_8|edL1y4`?a9&G
zySL%{N}U)9x^-#-FyU{6Fb*Ofvdju}a4hLoLP+;oqY<Sj%MOK$`brRT*a9D63N0_U
z+e2w~t9wzj2PQF{cE4Jp|ITxu9i$SeThdQ=d={FzlO&LZZ4ojxI6=xzCcK0(9zQsN
zs=|P$E8jwc9|j8v%koAvbzYU@4w~n^jUt>QUvr0EhUP(>>=0MGy=VHYvmCffRWwvi
zTt5G289fg(!32tg-}Cf;p8!dyOv)6W(>1zy&AVF>?~fS(-X$lwsD-Dv_>rq{ps7cu
zKFx?%x72iOGgII`&yNm254We{!f*vLF2BN>1jRk6VGK%$7q|D{6b;+uP9xMTzH4OG
z5C41w2-`wP2PKp~VT!sY#W;!Fh$P#E&?N`J(Z=#qf|OJR8N)Ruvfb&#RjYO{T(E3x
z#6jL)jnU4%xXbhZgQ}qzsFNc>rZ;9lWI>cfQCA+?d^&v+OyLIY(xGE4SsRNQctSz<
z%AHE+i8v)m)bu?TMUD0?oC%JM`7gZS*BhZZ20O8ktBw3mAq_BuY#fl?;W|?rxW>la
zUAD`hJuq1q<Rk@<0e}M1TLcMTuqQlDjQs13FD7QzFHL+W?Av+&$L|khf!}|Hw(YMf
zh9;2K89L|^!0_7;8wYQaua$eK?JkwhJ??{UQwwP913)0$|4qnlKPa9D^YvLGjz+$|
zJYWP-_Rr$~^T*2&NNYY+U)}iv#F>~{VH+VJ%L?MX-jSNOi1~pyI?Q45p#OIO%ZWuY
z^B)=40%)E|eQ^<HYhU74bZpPTDg2M`KobN7`1OlZJBx%}r0;)qZOF(EVWC-qwt4tz
z@Nzi_7YFSI0HV?aSxi5;7r*M{XSm8FM{+xT0L9F4!8On1jma@2PVL;Fusq!_cli(<
zB)k9pKXv9LP#-Q6FIn~%UY_Z?bEfKH{Ko$5GEhH=9V{DTdAz`z(b^3n3^3p9b2RpN
zZcgllE+aJxu7TbF+;{-^T>}uarhr;^M3sL{@1)RBA?K@vn;-25&QZl${<s<kEbAH2
zO*_boEQP?0p;DsG&t?x})79IR$DpXwLbA^Y(nm<1+Wqc<4wSQ>TNlesbpF?8_t}LB
z%CD;Wr8s~iIz|_`wt}E|JZ<oJK}>$SX3-~q6n9N5;&Kb8eC^5l^Ute@(>lx>Z6Q)d
zAg<zPc%K(p!SS$!S;0(%{>ukof8%P|2ZV7!21O<C_uw^8nk5_%17hDd!^Q~Oy#5rK
z1Z40!!K?5!d>%2AA!by}fEUlcE@0?OY8@k#g9v&6xr}JV4KZ1V-s~8`{yUF7`FC!_
zIjj)es!9-RawzCCTFDZ#->bJH+9^C3gv(LOe_x^9?Rn;0T%sqQpMz%!?PhLj5NYv0
zHsR}CD7-7p@69QGi2MROGVDj7c>=G7BV@D0r3(nB0>&F==cO^lh{7C(l^nyV?tp<=
zes7I0A~Sl=U`;~$>iQSt+A@H!7f?xrubhqDz~cF7e^BT<6M|8KOhH{P_>p)J-FFLB
za}xz0;;nzWt^ac}D(RT2L#avr$+ghAYr4*Ve1J(N2wY4{0{H~ufZhkR%2&?M0V)&n
zcLAw^)HJ_>5&bqOV<)@vKF=e77Yy87O9Yb}=MZjt162#S3kAqyi{j%c+n6?#u}|Vp
zD>}lM-4fQ4a-fNi^o92$v>a?Z*DkQefx0~K8bdYkUZ-=2h~+a(wnTL}OA+;+v#iil
zaZ-{Lv&~c~lTEA05X-W>DK86}op}PuRe@!sb=FPbTHE+<H&8BP4MHzXZMM5_uE)0t
zfx(^@SXK^znmhEfSzT`FWJ+qZU*vnf3wd%x5G8J|$_osWI)b9i!YF_83sgqLar)ya
zXdTsz&O@HtzTl(NQ$QR1o~L-*@kTyVpcu(I)&lh9ifVH|4a`}bJIa=O8~jg+^mZiO
zt%i5fAOfSj=MGn0$j^CIYIN4u<1o0pIJy5nY-C46)Clv~6#-!TuN)LZ4%Zh|Wegvm
zjp%R(0v;Z8tF-bxK!I(+55TebFgic=6Z*;MP7CD=IeBn!F3T%+nr5n`;}LwM#$2b>
zD^mZi!P-rFBRy??l_3IYXX(O;to4^2i8G1#7}+-1+!dA?l&z0{p8VKg4%pZtG+GNq
zpdhdWvh4h8X<_nn!H?W0y13o+i-#5?o^NZMApJi%K|KFZT!MA};KQ+h$HtOpc83y_
zeWDsa9N;Wuy5Hh~+}XGwtE_7`6K&f(ua?)V5|-G>n+`YkDdw9wU9eU9<_MF3Nn~K@
zagQuTdRe+{1O(+aAV~0tkv_nFmXjQ7f|krXH3_{<sZ;{xAOaXO*wme#d3N@t<b|`#
z$XuK`bNw0rZB=5?7SPvh!2V{=$C`_;%v6TH;0+*fq6gLhIbN9t()H*HJG;YAC*@qb
z86=AgW1i>f+_bEey%~{>2|W!GXco(GH;urJL^E%nnQy@C{8r1(t=`j4KzF+4kr|!I
z(yby{0sjGEtU26JpJMd3r2#yxHeKO5?HBzsN}X#6V;>nHji<D{>D&TEc6LP~!hEnu
z^@m(i3>4_vS;a71b3;p^z5F8T9$W_5ViVR@0+&gyU1o&QraNVyra<-JRK@ld9+0RR
zV0_PG^4@O#WU=jCUsJ^qf%7^BiVb%wn*z8T9vdEaiTs1a7bCL-8P!d6d5_XZyCM_f
zCDT5p-w_5*#;;A~f~XH#06+cb+<0C)%S5N>W4s#qHvSmUI)km=y2bZAMT{ecGV(O5
z^?@!r1Ko2&lPT4D_+8|V#$pJs7VWCv*U(=KTw@;8po+*$SJhJ|<!?rJ-sli87_0x<
zUL5jH`r}GV0;mnw8-NrLL~q3doX`7;7|9p!&U`s%%hjE@8E;{5)90yQXWWO>u?yA8
zwudtiO^+6gS_WOIn%V0%53IMQ`n7&>uYYt@0*mjxWO5MBlkI$1Qs?G=QPul%2x)KY
zT1K)3ss?nj^&Bl}#1}zh>=V6gZ6{5O*){}q*=&WxM#G9Ow)0&$kpA#_!70i>Fhd?k
zx-+VIr}2}b1$ItD(6_Dt=h<mNk(%O(4QJXsLR+weD}81lEjl79&x)ZKp}ReKnNICi
zN+ce=K{s(mvS<xu%fjbU-+U;@#pAw{8@FxfdTw~Un8`MAJ3oJ@fJ#u>4jt6oMSl)s
zEA*+FW$n`G;P1x9e`rn$kHlSv&fzM>QU8WqD5GwOGe>s%D3M~Z&Y=(9p{o$<H~w31
z@80dRR<s2OO7y^8vtq<xlvP+Bm5Ck)QymQ}1vbUI*q1@s7S$Q&>C{lkFMzq{%_!IF
zf^j4;g*-@Fx}Vo<1~zSO+n5HjAv24z=$dn*h47b(deT6&(qAOHM~Qh1hVE|5=AXzS
z+Jk1o(dx=*HDU({Kzfa$*!C!mz4g<@K7h%Z-1kptKa4q{RJlSym#1MWn#bhTxGk_@
zog^HA!Zf#{WLW6;5S5CIjJDxsp0kl<Yp4QrlD!so1UjP?RN;NjYGrEDy1B5^sV=7v
zvhl|xpI%r%hFV~U=zy5XA#Q80k|#Z11f9KA4211ZPzUR8y+)FTB);he6U{Z<GK&Hp
z4)nwxfeyQu$=1Y=eR~lYA^eg17<0tp!tvu7Mf-Z5N$S3uPPdxWFXq1Q+bw)H#^YJc
z4~u4#jUM7ix0dzc-{q^pQ7AYYi(?zf{`pZLjVBy=-Z#6C*uC+S#gHAnRd-k^$u$Wr
zCy_Iz^ae4Fu!Z^JGRklMrPcgM`p^baXw*j^xP}Tse+c=u4wD{4-*g_nC{82O@}2ad
zTL-XlS=X4g(gHn`Rv`*)ruNpNoiH~i%lE2<q9c?CRn)>Gbt<bI!#;>9<=?{@8sjq1
z;Vp2bq1AOZeKlNa&6~>~PoS%gBu3l3xI}%@`L&-ciU5IOpyqm~NQYb4yeM=fR>g!z
zRAsYQ#nw<Oh#Py83ThAvhNXA0URyjyZlcQ9G?<G-8`xAC)=>w04GBbLRjqIEmtp}5
zT<H0B<1#-M4_DgGVf?`cf?`{(5O@K;BcC+)ki?UQe6)XR;wg3DfQP}W-8}6GTvUIW
z>8w0P1kdk(TP}Q&WJh`G`4wrDPu+I@7tl{&*p2eZI>SY$pU3K_PoSZcsRHWE4W1MO
zp;N&5r$FJ+B%#wujn%qtZLRZNa93@{N=jOn)0vRiY`L@EWAO@wS=Y>TeT-6SOk=w)
zhh|YVcDClL-%|ETMLUO1L_V89^~&d3q<fq&&G*S~FH;ChB^)!`?04!aQD$5}X*xA2
z#uGBI4wcEO7Mt=(R-RL+Qi3H6FdeQ}^R_)8mZwqz{Shm<BKI>-8!w}s)fvq#*Z{ZD
zgZfK2{hg3*<u1hFRkGmio#lyB)QDzou|qw3-4(gr)7yqa$XomrDOu0_kjz_|S*()w
zsvA=A9*!2@OT6Y|_EX;c$j`{doyFGwY12F#&&FjJ2Q|LY#Sso&e0hi2kVgFkulj{x
zMeMuy8l&9&UJr1JeAq6Jc?8O@=_X|$N<3G`FtWAZeZhZ?`)F5Mn}TvO?bV`zxwa>R
zW*BegKI>3cdJZKARct-sR9O6L<2onmfVf!QrcPbi&N+uFkq&2k(3!ReW+FXI_uWJ>
zq-+uz*!uW!(s9Gy)K0RnVL{-Y>x`+2Y1}xn?h1?|TRaTiw$|5OS0+2yN4)mk;AO@K
z=aFjXW7$iVlmRrpEt`R$3x9WQb-E-U7dTAG_Fzo^%{F#t?iATzDoz=v_^Z8hJA&qs
zkZqD2mHA+;{9G;l65v)!u_npp@>0NUWxJhfUTfTwGTar8J$jvM?2Fr6F<8kIhH>PT
z4&}$Ai`WFCrtUac>T)rc%$BbEh7Hh1C;rGlYrf6e-+RNkLnwZjHX?#xd{WHab#-`p
z5bjjMbo+SV4ly`#xu-X!@S5p+*KkzxE}njTZ@PUgBHKB%r}xc`1HbA}>o1jOVqTuQ
za2&>}$j#83z4&!a5P)}Ov}-2p8NN@`9Xa3}=^0_JFwxG3JB<``7t~4KP1~?t{ZYa4
zZlt9Ab$MUJ+tuc070Q9}C7fG|uD6?RY#;w5_9{=s;!Kt=Z>Q^z4LjYdEdhKg15p)k
z84SdJgItpQB=a2&w+qxRYyY>RpCZNY5CmRR_Bl0dlZhV$9K!MA5mf=q3R}))uEwxS
zIx}!T=<O9m#%NX1f!2HY6n))m0W=jy=Sf3FRhG31Lp)WF(<X;l>uItmfU~f$0O#3@
zK{Q6#s9{=g=See3wYOM!McFk+UUQ^e%!%<Nbqz!5WbwlrbEt#O^{H;fjH*Dn>k)o1
za=pUA;%>_=>J#7!(T${k_vt~HUajXj_&T4|q$atk>oak#{Tj5!oXE0B`iV1`ZAra+
zoY?RDWJ>Y*u;hcF>1b1RGX*vOAz9~LBV3BPceVqt<vi)=`h*TaE!mRDL;g>vN&Pof
zf@L~y$y^=s=vhw0ZQ)+vjN|7;_0~Jic{!Lg6R)^T8@*WitjZxQ+P@eh#XK#B%aGxZ
z8j&i$@O{kz`z+ptyDj&_x#sm-D@^U!9p!ccucy;iP-rKe!Q6knLjsv6qLmW~WQw{=
zOd}}yV$3%;9Y?5Z8Y)x4buygn-JxorRm327nnc&E!MtqaM?pL)t&J?h=7V~a(kG#J
zXqD0D#$*;(#*f~La+-;q@y@>Y_J(yvNNBVqDGswLhzO^K&bM#A@m<{1+JdvUFIN8Y
z^N`CYgkCV^@CvXj-osh4H!MMS@j%3R(Hc>v28Azmw5$L{{F`OrFj?%{#heYm+Mchu
zoKu|LyTIl^g1%u;EXP-WGTC6#y!So=D3?-vXQZt4<`u3Ry!~Q2w#EkkcG{zoaUhaH
z{e3>B@wLTzzg_mk0Y&3zQ>&Yi%MCq6z1j_tb}tvL6eWWF8k|OH%#yN*!v|1)p7BZL
zO93B?1x@c~OzKQulh9XS4Pmf#u*~oJhTD&u#Bt*2WT<3@<%5%R3Bk-wS7}dXnDhl;
zdBX1v?K`^YN^~!2y=Nd)&?NX~$P2AW-PD%1+)GH<5dR$`Jaf;9`<ST@@{#@5IB<}7
zo=$m{ezj9?mE7w`#_-13GHTvwM8PX5`%@H!@r*x`cei;rmJ_*kOx4LDS(*UtIG}V%
zv$0cAa~vA36MfN&`sLrRpxh%@#$$;trKJ*NR;6mW4_8#?j@B1@WX*)GZUa8)p81>}
z-)4CHIKDXcK0QZn%`A+POIc4P%f1;sxao@6Ed&)WZWz?C*{ahzzZ?px@-WuwgAS_*
zaScSA>+Tt*WVay`VnkQ$0_r0zuw>HAH$WXbV9iPI`^a4?+Hk${+*#vitYK|>PoYR;
zdRl4cA|iT=-FDfYy62d-_H{CIt9D`BTUj*U3NLFa7?;XaxX8xe{Ww@h6WX&LIl<M=
z$&4*K7Uh1qULbU>$cwejSf8cNTp2eR#Jp4%K4BAFqtvt*#=7{N*MH$iaH9%CdvPmT
z-Cp}f`zG^8R%TQ|eEicZuep<{Q4Q;xZ_9aO>11akoB1dDCGWN_oLx$GE0eL1jf8ts
z24v}j8%On+vpNbd{X@9;r(57ah7R8>^jrUp-=ELW7}yh%G&6Op9K71M;8r-(lFc>;
zV-e_jKRHS-%eEcUaoWS9(1+qS{Gi3Dg!>Ng+&J(J5Mz<U$-;@|9dULRy4P-+AJ10A
z`T|*V(Z%46sSH#{{Xk>UU|L|gP;`<g^!()Q&duPeqfOSTWDo8%uK?;KXW(?VkMKc>
zLJtHoLD$mkT87*=EciXPeoD%gUXG|Mln{$WJfA+dB-SJd5_1y#Ce2@fiC@Qr(OUo|
zMa@bD#_+M<5QAi1%l__g0hJoF)#{@kFLUCtp`->}(H5_;?S#cwQikM^y+1I|g)^v3
zXqiaJwLhd?{&v!aV7S@;xss<nftNBY>J+~}7psB0k56MmjBK!stPERK$;ZV2lTxXg
zucpB=UAD5%eQFZS`i<9X63Q`$#jIWxB_HBKD+)(EzB86WS#Qsb3bcyY;$dCfNbohu
z8oc+pZ~fKA`1Ut*H;W9Th(Z1d`bo1$sS4lEu{B3n_dB~Mhz4ZFOv?~HSs2h8ekxP`
z&S<2HCBrT<yrZ=L(~u{ehgvqb)Pj=!-&$6kG~{#&`t^QPs>c?}mBnCaCQW0UDx=kH
z-fydLbg?XqV=mX=xeH*cwoO9Q!Gg=+_E4Nr(dj5P?@g(SN_hd7l-D8rfWDXuIXqkj
zaCK|xz`U_ATrn`M<bH4LHKhR71$yo7nrH11$2c{=10#@#2u2BEvOh~4Fw?9`rh+Or
zjdCB{?2Ap?NR&|s4WyFZM|8gbX6h57@n|p`=y#Ily>-HDb<24v-@WBRj>ZZvkMG@B
z5e&m6t>VwIxFwgfH{5H|^Q4au`a7N9rK%E3y3^xjte|cYXMcI+ha^i)&tY^m$mPD|
zM24EEYsTArhL4?80wlu$xL0rj)c-=;!pVg-)%H1&mGn`YH*oQ|8fHvG{hiYgGrt}x
z>?#h<I(F)<jYCL|Dk}N~8}6ukLs6ftcD0UsF2f#MhkOdL<Fy;gojBDHha8LTz-ayH
zX<n|YOB-*VPVlvFZ+YrjUAqPfCbc(LPg<Tps~w@nk|?&%)~)H6l{p+3pv~55U63g5
zHC#|}No;f%G5;~ocWY|}muf*^663qIE&TrIpX5(Rvc8Y{58Au`x;XA8jPXkA6SA|w
zjP3_21>N^p;XPXy`zX&~H-IZzh@%9wZ)JA{VXU2aC>zay2_sa7c(QqMyn0AKA@r+j
zc)(^0Hsue>#ZY?aI;C~{CVIDuZ~X`ec8B7g{{$km_89^QxUA|z@ZF8j<WzU%^?O}+
zlM*a}2G|cX^uWNiw~f-?V^?6ZSK)@-pbVL*!)kmbz+gZO^T2cv=^bnjYYap&<tSey
zd5f;lrxKH|DHLTvtAGK`w+}k8bf~G~d{M9cdC3YlhEgpVp*0kBK`!v#fV)GYr4N}k
zqhHy#<cV78_N#)LNylX))o|-IGPkCJqC0#U(c<bHD-tkFryPCC=-xeHy9J-S7&IxX
zyxIVnCXO6G*abp=G@W`>$UtZPV!2e0Va2auj38<Gvrn?$gB=LQT6#k0#A&f*^}*aF
zlZxn=)0e7>*rwSP?TBKW*XTcVEAdTizrNmIK772--phnPm_UCCE!z-OW=<D6HE<Va
zCbQ4N)_@o#A+hX#py%Mw&whssiUOqpk6$yo$9!>XiuV7wOz)o+d7*DfSMjW3?uPNx
zKZP;S=z#;YP|n5nPh}_GODixIq3=4Ogjueqq>{OV)gfF<gYM|P?JFtf;9nXP)LR0*
za-D$S_c$+2z+4BwKW}74l`PIrF4Q>&7$N*#dnE3~)U{p#Z-{7_#_JNemwxhCXY`gq
zalNq;pBattY-!FV4t5nbM>a(@BBAC4dePt>TaGF&OMr#_)q`z0p^0-Xcd+jX^r45@
z)R3##wgsORDA}G__%wdKh6tq6(g&Q&SCi7x(SiD*mz!q`mjXTUCk$lv5%a>}{+tub
z=$WO5gwEC$6JY!FgAwNR?E_*k3IFJGjqE4LTEBSAEQ;6hc4>}+Q{5DomhW_!!GbsD
zvmGA??`G{{P;xx$Rt06+W!UxC&t+CEZ_P69rj0Vs;a<I696i116m>ReC?tP?Aftm?
zv;I&fcB6lzA_gt$71(p3n)jK$`RJiwf@$5!!S$~(|Btcnj%xzh+Fl~K2#SR!Dn*eR
z0R;hRf~X+969hsOjM97WQ32^9AkskyJs`aoQA9(pp-7kBdxvl0y}KKB@9Td5<rj1^
z<;*!}PI(?Pttq0QrNy;_@nBPaM`mR4#H9@$@!t3mZ<9MMr{~3OMYep_xr)@V&?nb^
zmL31c&drk&kIS~351%{_6sH7)xGI;Y?=<d$JE*Tw-XacDNwY5IB}Se(j`k*ZMn?Mc
zwujwer&LJMyaTR0FR?}(%neI*vYVIdLG{PdJN7wWbOE@&#U(cX`<8*VjtXD)$tSWK
z_2?F&JXC=3!m_^Y#mJ8uZmJ`_>w_5=eOOi%jyvA@g0XoqJ7dF`MWU;|syIH&-mwm7
zB`iTG<tu!#q~uHHt}Q5?+HKSkBYoev{eI7t23MgrUj}NpEJ4vjtl1b@W03kqeU-pT
z&hl2ilyO#Mvzrm#Df~+Ow#S^#Dk8ezz{oJ>nFZ@Rb#90H)?=6ZdHE)PREXE;(MG9%
zDGHXEY`R`Ysen;Q<G3(?N$hNzyu!N2vjVB2>O3Q>=4IB6jKVl%a?nkFK8^zc4yaxB
zv`&XkO1*sQJ9XzRhEPe|G;*h@z>fY@|EqJ9W{0IIt;^JFuE?*(u73OSQ4t3pv-7?p
z6*yZi1z*n28eSlD$*ECGsUz*iODF$8lt6VdGcl`6MT9jfKm7sn0kH+#;cjdO@|*!@
zd}`;auYbw99cp#>yp*>4yuo^Ai;S|fkY@RVdNN+~Sg1{8Pn-mRofyCqyE_AB)CNO~
zq4!f8Q@7XWdOMzdYcMtI%|l-lFny=CAw-1IXS&{*U0Qqd%UbsI^B9;5-X>gOI`i(u
z?3{KzZTtF@g)e@s>lDvTaf9~H8@u^a>{CWk)>A4Zy|W|_AHDsUzM6DDxt)(hOuW4?
ztn09H+OT6R#Rs3;|L>ds>jMZ2pS7ku*`F~dybt`~1Voz|W>S*z-SG-b4Y|z{9=sn9
z?}%X5PBBqFD7&5`Gj!&a<giA6&YN0Zmawy_0etv(<^{;mgzi&Xmq27dMyEMc?X^{u
zjkYdNs!nG64ehva8`Jzs5VEs6=mx0kdcQ|siNqw1d)Y)yK3(F>EV%Q5uQsz$6Pe^)
z9}`2rEhb)3J?vIf`t$=cHZ#wsu2d>)Z0-;I4p7FX0Ez(67N=hj_8dy=P)-4C@{S8w
zP`9!?9;fzLN*l59<)X!!i7<$JjW2*~j=b078{TtQmrx43CquF3Ri<yI@s5AUiLWQ+
zHNQUEyP29_{@ylYu}Lb*JL<qp_9kcl37E7{dd0j&w2f&@Wj4Q&JFg@4keOTWLW^_J
z<cFBbMA4^xRntPdiv>7(>!MEx;r<}i5s`T}@Puwf51~)O>yRnO_PaMc(ypzM0|VbG
z)T=|_8T}4CqqP#HfOCY0U?eAjjKh~o6<dx-;^Dkf8b!@PvjDEj&1Dnlu~K}=Cq^rA
zCL2^zczJePh;^-%8x_br)B~rxBo6tgcEipLr+C;8AI5MCiRqf;1V>5kc9&5)L<jM4
zgqcOfD{~CY$|oLA!yZ%_sXm>5+9AeTN2B(rU}m$%ypQf2ZQq8Y+9O$SMw}xxFS3wf
zj=?#UQUoAHcQkzrBFzCZB^j|`*T0>e9xzSc;hVyAGq{g67v}mz^@Kjb;^^^iUpexp
z_WEz2KOup*V9*%Y?uvpq9+fR(JfKg8CT8Y1sI~_20=+HTfK@U)D+!*i5pc2FH$@9N
zR3CCz`vdT?Gs?Y9YV!W_cjX1~*7Fe*m18aO-8jcYo|kpAsTcM-Vsq|?i#)Gbf8>)n
z_s2Uu$<=2T_&J>4rEFn7AH43}iZFUD<2LJ@uTUNJMAZEHx8J&2JkFqUU9#p|7mvYV
zlU)3jPj;GawkN+1$#lCTUv#QT+Vw)@+sMy~QYqOUsnf^QUY{_D?!R}ZWhf>8wcv%+
z2}e&EgYCMtw=y<uS{cy=i>d@!<@Prqb4Kux#HRt3&p_~yBp~0i!Kotojr!hWN_MrB
zfO1gqjOVoe4pggFPlR0BA)fdblyAY!L8jv;ju<F&|4)5tNKJ5eggj1&6<If6j?dvv
zhWYMz8=+;M>hs=xIuprChuM+-cq3&@VW(NX_iaS<6a5LcC%iMyv<-MnR9VZhL=A@f
zjkaaeC)5l@^(~$R%uFZ4<u61x_kZjhvVA^H)Nvs+PZRmw+k2wp^BP)4%32|`&y+Wi
zKi2HciA!8~uit%L1QTSGAmj0E!6d^X<4K>P^Ot0&ebc`($`ab=hWpX~_4yyXt`P)n
zQA;CgDL*DrKZtYg0c@)O#N_nPl>l~Jyq`D#kbOpsJJZhQo2f8f0$Okukbks%rWbQA
z$OP4jDyJ_)vvGiBM$lnF7wJeJLQUB`bQLLCVU&TKF)I66o#~|{fwVxrS}<)tcBVRN
zg)($5Sqi8;dK=+1!yKRJCt;kkwv`;3U*-2BWU5b3MAbGuh|HCadK!5kaNjSlNUd5L
z#A7REj+<Tq@6aT;38RQ*n#{xURRw)Xg75&VGA~BDW%y`OXTcSwyEdQ(AY1x0OKY+w
zo?&Pm#N|NE_SbLcyQA{KMfLTRf5j&s|89YoV=VCZlv5BQ(3-(SX`3R0v{*GMlDhbw
zSM!B+l~W{U#vpe3B6xyZ=6BhbYJiFt!>5^jU%#q5{jgB*jl$T=qFP{#A(H5qL;TYt
z7aKJ}#oT`*oUadnBiC`rz<cbE4aRHx<Wneo4jiN-d+^k83!q*>vD6AQ3Dri3;%P$Q
zpcZc+Se)~t@6Ss%-FY08Wfe!o2`qa!{x$Z0&Bg<M$lf!qt_9MuBYh6ulqH}?1`hJU
za6s$w{Y#*B8CRjC_Yz3WPXG*}37l<LA6{Mgdh+6$`76!KzT9kA$KoUQe?+@}ev}}?
z2k8@|KPvxWrhX>RPZ0wu0t<(q#WsKdClF9Wu%v=_gTW@C$|wOaWS{VaLQQd)44<DW
z5o@OR$hpln-6%%JMC9C&UjUa76vt^DW4&)q{0R(B7G8rd(-PVPRW<CL{;fiP7|3PH
zR^e=Nnrln5rFam7dC+jZ9IyM53qUi@tLv5A^F5ahvQ_2)|CX~h^VtO2b_leB^ZdU1
zsxjtqVw0@+)fWmFc{O?3C{27z1=Bj|CFI?VN7_%KBAh=N+#{INH&#&UW%_;%4>@@S
zRc@w?J^p8c^Z^&7yw7rjk4unC29s#U7qzls?$JThAsNa`;rsmTd9&0*#K*F)*!*>m
zEuxf*3XGA>eEWF!iXidvzdq)@_H1O1l)5ON9yjk86=<aqq*w@pK<Z=?;^Dfde{(hO
z%fIz|?a|8R|22+ZgoIi79&3kAok&1Qw)^`+Re?loNgN%H?&foek+UTYWdP){K|Otm
z`)&)tlK=O|bj4wgiwOZq{%URs0##)x)U4su=1;CLDaQ~C_tmZEuLc?|j$R_g{pAV#
z?{Uts--fvKqG3_)yy5(4=SybC_67=?ZSQ29e!GkmMd_G%1Jp?V`RL_*<n#Yy1iaM!
zXAjSwAWmr$I~$xbsE85Tjh2~h(R+hGz=u-9Pos0>af=EXd8P~Uv`6Lm&;S1C_%)<q
zyy@3w#8Wa-Y+v0qeJvg>5S9JFhluY>0IOwXlX<>Two!FJ97iajQ}fSZz<qi3U|7CC
z^IBytnAAk`j6hbre6&0c<T<F=Py+T=k=d*A(<{T0Rx6`@fByR4uGxRbWf~(5$fx^e
zD<&5B)ZQ~t<9I|s_QiI??<e{BP81XwO%r);rx+7B#DBl7e+=VKs-9n(eZvje_c4q+
z(T=a0@7uM7*W53j#?Z+P%SK1a<49PMkN<ku{O{R-4ug9MK1~{5#+!@;CM(S<5R(Cw
zCQ*U0BTpGn4Q8_12*;3VM5LhDQ(^$Q*wfbhX5}Js?sM$>m@J={SZN`emW9|cBdR2Y
ztlaRT)eYWs1^4$^vFf~EeD$@tzgSTabc6nVApmHmHKK@I0V+@Jztu-l`8}@tV+1SR
zmrbeyzGfc@q7%BtlNQ&x1iIKCAz~;{wpW%uoZ6u5z_3Q5g5Rx`J5?IRPOGVTs1;Nj
z=;av*R|m<obyKO|DCt}5-SLrclvXg_4VH-<Hx|U&z6S;<p=(R;1{Z(&wMvFQr=F{7
zo2;Qbr#w8_26O43PU3S3$yMu{Z<5b#?aPOpY`<KDz?nYcG}#F@DwC;}KH#l168=VY
zPo&?mJ{e0jqqFMRN)?z%|Eo)%kVbklpM1JO$L(Q^d9Kf?Exyn%Lu>=bTO9d<YFGrA
zBU!57Q8bnG-ndM9kM;UD`&oyj&=d}#e&<7nuSJ6`g7ju&r{{6YJFIm1eH2vZv!C_d
z-q$7B4OgiWo^um)VdUUbf0`Q5J*-=qP3CmyZPswL+r8=X&58bt^TGT@YV%utG4{q#
z=B>>bSY^bZVE7F$sU-*4#s2L$j$49Z(eho!DMUi`T~CqV*E}9^B7eD-cPA=LaJY!i
zAyp@NCKAV;w&9pM_0TFuclMU<o1PLQ=Sseo!JZCTq<kXZR9&M$g)yBupMB@)q{PLp
zvnl6N#1tYTA)%I~W{XeB0=G9<xArkOarqbW#ZeU3)8&iZQ*7s+sTqXIt0OTC$ekRS
z`m0Ekcm2KoM1z?z8Z~67<UXf?z^xRVp?BtidE4?*E(ag~yXeM=e1(sD_Qmsu?`-3!
zqA<o5HjFfHK4!S{@Xs5cD^`=|HM0XPZ^-MLH#M$vbSj>fA#saOiv8<3_Ls)_<I+;`
zppxcRb|hgBZ%v=q4qk+ZR=IXZ?C;b@c}iHFEvezIXilChh^pf!%_c`(xdIlsvk!f7
zL-mrm$P3kyO>h2p7K?V-3_B{m1IOeOZ!`~1mAJ>LYu*wK>)LrEWA-)@@~J~$AeVJM
zf*bfht#fUL-X;aP9IF~`&`*V~$<1c*^7@6-=j}zmT51g0)do$^52QSrJ1qrZVjecZ
z%*=x=W8i<x_3sx7Xa?nz#0<qIGyeTFb(Zr_bn{~D&OH%6Rb=LNLP;W_gp{3Ji8Uav
z6NFu6xh)6YqkLXc{DB34KQHX(ll&r|axgxR*Bk=!gzhr>MaJhNzKlIiX}#4tICW|I
zOY5=6{ay2w4A#o!I(LIkHS`KHKS6Az@ahTvDDX#?cn2%&$2=-)p)xTQ@39J$o0nw1
zu8X9R_G2yLd*@PcZk%zVU|T7fqKB|q_U|U{zLEr*ESxL<D;1Yo>XdYN54A(y#GUQ0
zBSn}up}^DdmtZdcAdh27dGJ)hD$svck|~lH@@d^mR9LJj==bBJ2bawzC+BdB3rnpp
z$Uh3*;TT>L__$kB<I36dI456TSUiWnDs{5J|7nTTh3MFG{G7cZ$3Yq1qrEl&8IULC
z$K;|p84=W3vr+lD-RdNc-LvNp5EO~0=Z2?e8Y^-dkG$U_0cWAYJ;dI^vodA{HWr)N
zp_B*(*t<2P0|zA$YLscI3|^3KbQ%OmUVzRv-tZbFCwWu7uged_$%nizofa_pVQLuL
zfp{LPqBr?t?sRio|1p~luCGpk8+q6JP%--<Uuc}MH#Sm=r{SRty6tyqGD<AdWyW_5
z?(5atnHLDK=-V~k_|x}%oetblzV&8ao<F|=M)e5=5%PG`O*>u<r4A!Zu@yr9LiFm+
z`^vYVES{YLLOJ)I7rwz;X-0OEssAd|YZYSkw}P6c6hi0I{hnt)UrP7J&^y0l&Cv?&
zV7ete^ogW;cQwOanLD_dIZz>W3Ke%Fg-705IlXAwq=SzY)9YMd6oqf;_`@dxg@7k<
zS8J0&Up#>wN>C;>&C|2EF>mOcrkt&OGnVt>6mN%R2QXX}8Q-tt_wM>2cVBx!j!eqz
z0kDaYQp292tm-U391V+1KDgL@IJ<J;31-1)H;jV;YW(BU$>xxcBI3^wQ-#dyDqQW&
zuiZkvq<pGWXp-l3$>G<y8EbBA&_0yt?0IBvy!g2tLla<80g(iAtFWl#G?11W@Ab0w
zdOxi`v8=(0yui=P(Q!TJ8moyU{+|RC>FGp2k^N?i0q?NaGQ^6--|p#{sIQ93d1!ql
z-EXrP->=`ite|)TM2Ek3*<T9J=Ua=ZaqZIn(Dpus{?^sLkD?p#@iwk$l@tP)7AMA?
zE0k~lKep|JZQfvb+v+=+-t6fMJNng_X{(ON6r6$x@Or6!M@`PX$3;@`RwDbr6OBoY
zbn5c+%p`d!mY3`0S~r@hlzbl*#v-G#Pmv=-7lw<A4#&9r3!`PcCs!}~AfIzS$(Z>V
zz0`UkZQzr`mu<bt18C~JgGzJyv)$7TLMa=+(w<M~P|X5ET&!*(u_u-j6eL$8joWEg
z@praH!Tzxu4U#ooos$S<3(Vc|oR<7#JKc5VJ83sEekZ@pJH5rK_DG^HC3t%yH7Nb6
zMe?mK`<arPb;<4D?--uy(qmoE96SB4c8n4j1&?U;13|~BiAVhm=Iyyip&X??6T`-0
zL16|n(;d3*<rakzj~rdYUH5N7*#i3tiOZ3dZ3iNf4&J7?7Vu45W1pZ+-yYi!QGp{b
z%A=Zy$%{0sD7R5y-yq{D_GK&W>c@z&^YBPZ=?({q;&E4+>H=IvcZ&SbL*5lry!Ld4
zo3}4&MMf`1t~tEoJY^W4r(|*6tKPLgajZ>Vp>Xj@`?hg$+s#kO2e8NjrFQLsa<CeS
z{>1qFgjo3z^yS6)Tr;{t;kdMszT3(1)uuY$AHGnp9eSAsUr)C{u2(%FgM8v=`Uw5=
zQV|Hs&yT2+Y?j_lN(idzs5`3_BF_{N{j;~}Zt|qOaVHMzO^MG@QqX?!##2K09O+4z
z#@SonZ}w6>iJ4igY))?3@=bYgW5_-Oe(e(=KI3<LvQwYD@g{q<a#~6)i{0e+-N+}_
z#-hAtNbiGq89C4N@5MVEyRGjmEDFTCd)90&J^7k}xjMFeHdJ<n(Twa%#pj|IHI}Cy
zED`oW{%w27NdpRmX2^ehp}UEr5VDCf$k~0?TJcVH=WI!fcUC+Pos3B8&=TDZV%@Zn
z3&UxK7gOKOyu%WUgsx>+*Ga!qygokpOp4Xek#%mjNJ-oW7QI3_qh`>4S9bUQlmxC}
z(UIA<p8JTtAy7TwBdD#}?ejdmxo_CVF>svGphEkZGMalbI-}mO{xT_*0IJHKC{k=O
z3L8mEnoNm$0AL;h5D#-5qOyZ;BzW}ZGBLwg@p%eDjdt`G*r^~Nh3&-1(3_*PYqI49
zvU1!4^xCeUhOuWXPhUJ1nSIEQY=3d;dtW2lYb>Z>GhMeq`<Mg;TSBq@RX7FXBs)Dh
ze9-OasD#H3`c2u9mCZLX{2UkALr2d=H#&!QcZ+a_vVPHR52}&4m>Ur9oAM;$v4Y6?
zlef%Blk-NjYuT%wwYPL#2zCDw71hkov1<30<NA|}Jp#X8i#mi&XAyf&RX{n#t2;G3
zvZ6NmTo}08FFYF)pULxj>hVtB`5gjDS7Zu51<O^7G{@Yp`j{M#v2?T8Rpk3;gAx(x
zp18{<5x0h&hnwqfXV0AB;3R`QKb7nMh5O1Ez2{NYQ9W`uS&gsde!d!|??M@vTc>Bx
znU?DbmoWI`?RTo5#FXg)^*ui)+=b>ZpKkGOHpkvx9cama$CtXW8mZ&GJ>W+12Puv6
zIaIVQm4}vC!BuJK{2kFUqoC2<uCfyOy9t*dRmsyBzv`9+Rc(%&#GDUnV8_oVuVwyF
z%&-^;{}et1o#c>aKX^DKm3Y~ckg11|K1Z1s6>gh6jU}#2@;1hpWl~0=OO~+XEGF|^
zn7CD|kafkn`!HHWX-@K)|H10|*QgR0(mh?aQOVZrLTgEm$q_lS-S*n5&mVaB|1;B=
z4?NJh#ll%?Kr(xTY>RL*KWC|Y48mZx>9-P?r+6wxsZ6a$nU4*+8ou9-sSzcXW%`g6
zbVJ^%Z9U`MrQo|v|L3~!e+YXp{JU^-F_w}=SWk#zOn)U6D>*)FZR-h8F^Gl`D>v%y
z<Ep{V5i;Z?WRsNaDZ#7vVnqKq0x%Rmn`L10l;QYqb(_&yb=@#Fe#vR{rBq!^YsNi1
z=Q_7)`x*9V3BW9~eJ#y&=hC0lZW0NaB(A+&M+w=Lp^#@Y_o~#nHHEveGU%9kY3C};
z3Tox_IE?I{%--0m5+4iYmoELurY8Ytfc&lu>u36BXL3%m7xvxm_BBhbK$n!bzLmTS
z{z$l<-%z}y%9dQmgnbBkrcm$6bW<yXH5HY+p)(=a@>Ln(Jtkzp%Cm7%ub-#WdJ~0=
zD!O1oAULfhzmed+1Zo{4{DBO)n$53!rK5+7x+-vZOKGuMVcnx~k~(IW!<5#a^Jpi}
zO6aXtc4eF!Z_laKC77HwZNhL=Ih_vSAxeysS7dPS9HdmuEqn|g8mMr-bysIGIY3oW
zJzHg~I_{?sN?iR=$zY-3iu7y<lMZ`GxM!&6qy_9>^?FVa93Cl?vJ-^i+fD8nO%E}}
z%nELr0_pjJVwZ4uSbx~3u=`|~=fte??}pFz+$6@bwgO5#=k05T$jgiwmoQmpkHu$A
zS7E>Rh_e1cMhe#dRnD{8x!FHBkl;Isgy$rpE2~eCEaSCg3(peEHVW+viml^U)W#Lj
zg+C8w=6+4Q&33Tjg34{pz5JoNF2~)_IAzm7|LcG8O}#FbsNB;q7*w7+a*`RE!&d9A
z(sdEE?kUDC@2BYzQkjGAFQuJ?ZKzFf6p)>H^wNZ+g4Ax%>jM2qpB&HZS!{@`QXTU*
zHSyIQ&DGl~-RcuiLQyAU9j@L>A~;_l=7rZRV-y!IjvJ#Hq+Gc>ot$6Jp299*ucIq}
zPTyFjM!aQHgF*8`*bp?6(>%RSvt{+!qe)IG9n77-1dlxp2+DKoJU)W6K9eSaI^D^J
zkhF%6<EVS8^=-c~mj<b3D~pzGsvw<y3SH-O325x+4(tH7S$4}6Dr(%V+9Xan1T)t&
z=2jw*Y4+C=DT41w`ozlb9N8jH$>GCvS_|ki=sxFr>AJ<2Jn7ry<5!R-<1%Rfp9{!<
zKMGMCK7-5{jkgQr_c&PBsq1ee92qa-&qMR>9-TQ6LUKh59cZmVvXS!+E({kDm!?Y#
zP}Nk$Yeb1ca>vgz=6;JR^JlXnqarF3zlR%<OfdhYnO^6BQC&%T8ID|>`@_{UQ>U<s
z*gIYA(~$L!O)gP|hX^O_=|ud!K^fCJ%CHUZmXxJMrBBVYRJq!{NC|k^WlVU4|MLQ^
zlWj3%tA)vhHFvpOXqwqZ6m%_Pwf+2C+^FzcNO>O8GE4l0;~)T9m^L$+6#hFkp>n!U
znJ}z<Q`<h^FU|fNgqr=jf}8NeWuzTXR?Pu*xg2cfH)+mFYFaql6@6J*5CT93WUuL#
zTwk6qjy!%hC3Q)F0y9dymp}Iq={Z_!Le8RRyS_TaPB?czPN3D&D&{T`ZVIRJu)zAX
z%s4x8mTV4@TW*^~u|$b3Un$6olf0Y2P0ZjtmG~JXxgs2=SrEQOMrHu#<eOQ=0H}J5
zDmrTtF8RoY{&Ny+7kY3v3KLch<<2WBT@qGZZgDxXMryD|bv^r_BWvWejM-Tav@UEf
zliby6G#^NeT}<7|!zP6(;5R5tc^H6B$$z))Q1Hc9WxbTfMBY8s6&au^%o(?V934<p
zU65Dg;;!H=<F6ei`g2hUB@ilI|JjDYBZ`8TWDr)|tmaCGRfFdH-(T$77p~RATKh?R
zc8wV#Vg*j2=I}5@ilD`vFA|h%A`wD9Xpk+EyH(YAgNu#mu(3)fGTk(F>)As?hA!t<
zypnX;TfCo+>@W8R`L^I1A_D0KQiY`|Q63Oh-o}icRaw8)IHwZ3nT>tnm%6DErwoMo
z%n946*2@3_!DU#*oyf!6&dWN0S995)S4GBbm48gyk0)C3uX`?$a1|0@$#sJ7cfxy!
zBBNUJS*hnV3(w^I;DMfLR}*eMQ^)X$D%cs_g{lD<)KPYCS0`JPZlR5v=2MHI^;0U0
zs{jI?i%f}Qz~36uG&H@^*U792H(h7+F-=N_78%>;HMLhS-tQ14Fcr$SL{cR*vG0x%
zeCH`aLN1D_8c4g8=vXjY9p-GxGn4!~M=-}GyANkEwBQ^$sG;}q1)w!zPU^uKc;@1W
znU9T;e3VMJr=b#T=Rh~SxjvrI%(ygO6m)3&f+7IXHh>558305@XfNz~fu5>zUk<gD
zt5u-7_8ur$q0G$6lYI|*4Zq1va2HyyeJ1$~=(?De0c$S{XvTf3Rbu(r+LtG1R&L(>
z!^h*k3}VjLdTl{2>YE*C4Grl5@amKe*Fq1Hd~yxySqDKS+&$1ieGYKtj=h-|Ym0Gr
zE(dh2ywjV2YV<1T{vCPf%MCBPika(*5pda<`wUtn<G$k=tP~1Na3FgR_>n$?^eCnH
z*2mN5K;x11DBxz3!J|I<f=Wf!h~0{AZ4=CAsf%V|AH1Rm4gt#-Yic#<p8qih$^^i7
zpnY%ttzmz+$FUW7ejO-$jrhI@T8YB<qNYHZvOb_CQCJWOU373s+ez0GK1Qp9Q_I$L
zPrP<;rIVPzdqK&bd2m=ACp$$4Ku-R=58#t5vt<-TeyX+a%7?rX*(6_K6N@R=Vl`Kp
z5uNZ2Y#|xuB?O8oP_3<#6r^3$1f=TdparMLPSFBpRWM7yXLk^BEv@X`q@<xCcND<H
z%=ZPnKq8H@TZ75m=Y<z5Ru%P^hClj(GA)~L)=G{Zpr1=iAGm086D}ex4kKNwAp1A=
z8C>R=DM0jV9ScCe*>nSJS62Yn0!yme;qnc5&8dDTC!5`AGkBQmOnU0u0OM1eI_o#$
z03g#7@o2u4&Plq)v3{3o?u6z@J~pG9R->at5gW@>+T}qEV!{vE2aHl<zt{cvI{YE$
z$H{#@0NwUv@B1WURPgL*7I{yi(-6b>4k+qj>1b_z4O%lE<XJLwr{_u^)_L+cRppJw
zcX_A~`r7I-z<&6-$V{CfbZWx19z_K?WLW20qpY)3bDAQCziR@w_yRh~I#&`JA_0)7
zDM2x}LQtpeWxm4U8A%;OP(~rES5|~<c>7p0@3$%7Kn605T!<p~;4UhAP`=)+!v^$e
z=p~M6{UJa(5_3%x;D29v_;NiDPYF=?q^)zeU*DhW4f8`W{Z6x$AK$)(#534UpN&ei
zoC%3#Se_nd$c@e(p3zSxe{0S_azPtx;n~WC>V7|H_PZYAOw5~LhKDm}QwJ?73@-)2
zUP&+ZPXJ*vThSGh?8*t8{}Rj?e{+c)w%I@@Y-_mk0idiZnHTpyjb_o*F)RF0aoxFF
zX|zr=xET*8L3yyVlq)*ZMPeH~MK55gHJb2(Lbht+TWn{ucvb3}csLE#9b0+CHT0v}
zTJS(p_lgO75o&(YuzSNa^mScP24NUk>(5xRuf%pu_i=*Z)tuEZLl&x~akqin+do1V
z>ASq{(~7$ZJq&)oM%v(L6Vlxjv2`kf+1_fkNpMJBue&=U%MQhNO$SsDg)-}aV%ane
z1dZTeZJGk@NuYD~e~deQvZmD#9I);9p+*ZEGXu0einr}T_y)wZ;MZ+e#{=N^Y#UPJ
zxi6HLo}X(t?^0dWnT`400w8(ZWK{S{^@$FbNnycD^w%HNylEhE30njx3=yPtmcjD&
z`Q_3#L>@;XUUWKfCp&-6j^-hscOzk=ZQ9*jjPC)rI5a0ccyZJ`un;lrx}fqD`hv<2
zBbo4*|EHb}qzvn3Hbr+t$~1>>(VJI|x)2v}&nH3`3`0AU=BZ)1%__Q%=<o+jRJvk2
zotvhonmV|%OW&QB%JG)e(_*929b+~Ce~74E7o}zF;1AZbBWfl2?syi*%mZK2F$m;u
zH!FIgjx<*3O45eUt+7~l1YSEug4>$vp4jMMM!xvu4e(Wy-(6b#usiC4p-Jk!8K4}v
z=utZUR*vS2SWAu$chlyw9Z>~<>-e@?kka3sg#0AtI^_5XMx5i@E&I$0JsPsU$nM_e
z8f-*p`l-2q*YO^~AnZdZ0pN@+^TeUE9M9e^m4rL9BNpvGtQ1)1SqIi0UiMO+>jvmX
zFL@efX~@#zC?Y5~rZm}Zw{%n$r!JnJX_(sZh12vTu?+MTTg2-w*s1aXa(NbJ_Obf!
zPeK*mkia%MZqi!7f4I(Kir+-coLU_VX7Hlh8Ezn}k0!5Q2$fFw!Xx<oyL;JUVjX$Y
zV?D<?pk-^5WXK0}w;DTZS)T9r{b(2M!!>!?i6D(X_C(Odi8z?2bY&o`e1v!1F$ZI`
zh;CPtI_3pD)A1k6Qm>GhaG0kavJ-yw$5%i9NTpk|X}O=fa&dgfPg<+XW_4WDLB&<5
zB`KT&dl@U~rFoAi*c$M`$mb204LToO^t@LaL218A3Zvb2b<%EGWmgRzw!PC`r%Y>b
zH?d{bXEav&wQyh|=yP{Y*P!9nedh>J9YS|Er(nSq>sH@ukya6XOwOYXWaT|GkZkFi
zBJ)beVaM|WGcHx$au;1fCpP#3ARASqyK*=-THS#|NrB)eT(;zsRAh^h8`ZUB&}$EW
z)Y{zG$@}qJUF;NQKs>=TK{vHz3fn?=`9JT=pAJsOUC-wxb`^?i=Z;#M4-!5?s!n>#
zJ_l71)1%lotTP3rJ4MASJmV7R%F*|0f@GKnki_8&erepM&V5PrZYOkt_3cKwB4*e3
z)!<J~F{`ObJC!lbPUsu-7nxn0x2iClmRrA#heA+(X@_b_;)IjV0bK9AS33ZJGj$-9
zCw~@_Agvd~D1JM4BFcmnM-HTV{AFcA>HZqpVIVJ%SRp&XBaiykC*jiDJB92|R=0`C
zY1{J)w}i^C>!AU=`>r`U+S=PXk?<v-kU`8|qS@7O|Je~Yk{D8j7)E1r!p6cI!kjIt
z7&82`1lfJqKZUu4eK(g+KV_&uxkRctRJ{tIEo9}RZW?{F1t?(=RHt0lrb4cjl%0vj
z49_$WDa8u&kA&o1G#w^-sbEsIh$nT3s!-1jE(9ey9*b`(Mh4TTH>M<${j;hVr$J@b
z>ll37=?X(G11j2ir~`k9E4GgX12z<Jj4k=k(1PWq3G)pCJ|WqWJC~|stS!G5n0qxJ
zM(&Kq=dMy6$}h9S2ydkZTg|>h#sW*yCjtp(Qih=N?E;REi=tsCf?%h7mac#M6b22y
zj5WZHUJNvTuIY%nuG+;6`!xO%zIBr8dwth^C0bs$^zX`Pr)IZqP>t4#sg5;_c2w;x
z0`*s%{B=^&qGWRl;Hhm@>Sq?OSKiI%M`-rU1Gs&e1!#ad9H^zYR6(CMn&zMF<(1!7
zOqtL(sgQ^?s+4`l*!&I9=Fn7ei=HbokRAfK!cjRq^YJW)%B-9^>{L2l5`VeHX^KQ-
z*<IUZZ6<`TNcQX+-T8y_b(`1h?syhy)!@w6(-4Ri8z?e4@4Y1SxhDvZO<Z{|om=L0
zB{S7cW(c$cT&>y~^}l!b#o=wH5Yqd7KSHPCEvNN@d7EqxRS{8JfQniv09fd=+-(>A
zLT^i@xV*r7r*ah3MLN89DM3Y~jO}bUm(U78%A_E{DZdKL>HqHD_tEQVG4SW?WDdu6
zH`V~*Owo_Q{PYIS`EB9ZtkAhPi;=T9F8Jyzy<i6We0U;{S<J#alBh}A*%Nh}MsB*j
z@=iL0@kHf9@WRR5KqrkIIlL!Cw8ZA>_8RWmKMXYb7h05kS~LYT<5a6I+NZM>)RNjR
z4vLM)EtX9QSdD)Ev9EgfahtRys)<?8jSfyrInYk6>#{x*ey!8F63E+iC*RQ@*f#O1
z?@IubBHDFgx7Ir(rRsA|v~;fX50i@RxNPp`GDISm*02tfN43}J3}Ar|-HbkUzo(Nw
zx?gGW8L-ew3(cGP>te`HUv*p7oT|ar`r-*4N)^$2lnddr&z)xWVSs}+*41nP&-2Lw
zAh(Ovi}$0p3k#e8KQ(2v*+djlqu^9mLj=`O@w$bdN0lP%h@V#Y+omOECITcD|J1`E
zJ`h5p*mmHcrhDE|J@S1k4f)=~)A?a?;fI_eoc=x!ji0}*_9V~XP5>YX{U>N|!CwOi
z-RbJ+jaRVX3+sl`v=8ov++3fkI-Fq+7Jc&$uq=#%LzXK5co`41*4-p4Ed+!0>3KiH
z{Pn);Hm)O$-AeUoVf6uYp{@+oHy6RPC3F?ux#qqv8qj!ZcX|9pyvM=zlaZcPR|HNT
zBecI(<tnblh$o^{q(K2Hi!aGu{!!do<sRIu<sYUe!&^mpahryAm;0<l03T&&PGG9p
zh)Vs!;@1%7P;UmVITh^V<UDD{(+r}w#z>zmratdGm}yXug^n*7YM&;2kY*qrO&_Lf
z9By0U+0y?qN)ner)NF=U7FOqVzpot<F+CmPw`j8)(2hu?R%W`b&+jfYV)<%I3_dpN
zJwDR5NsRY~D32XK`G95;z+Q$fRU+V~c4vz{4i9`2@A`0aY&eO~uG)iI*n2?Ng_c|8
zb~9`5PpM=8CE$?vl`iNa&!G3%PFs>K{EVH=Ss`8FTSCm^@e^x&bG5)&<L=8a*6%Ni
zDT@u{AgVtzxXgu+Cdqyh7<c^+Xk$aU?G<rvL3#LR@bU=l!TX`0`1nC%v>=a@lM$_s
z8v_GL`QgDX8)JtV;?Ame8jpwig6dPw7h%yqqc4A~o1b}Du##|h@FAZ$$fnWqAeI5U
zG?&>ut{SYno638)_aR+UYdh(bUizUNU-^%G(JS4Xh5$&&toDAhCPsAXx{VF0n;Bji
zC*8)K-`GyuXl>Q8s(ba?lZL|hksgM&nTCb3>(g>q%cr#Gn<DIs^8IrdoW6fgku>AD
z@x4ul`fzUz{c=-<X6UsRCE<jO?<a&=<eAh3tP;N7P&<@b8mW6lygCmBojZ?lu@1}e
z3?;cHIqnE~y}hE$L4pf2UoHqDmesPHXY1Y#C-U>pu4i98x^}i27B7<pmhqQ_v55py
z2t$Zn)Tl|Op4Q8KTrD?ayUbjMW7-b`tb^dOE=pi8I>SAy4-x9!Dtq~>XT0<?2@?{9
z9Kf*?K){;G2ps@29?MvJAz2~KUa3Sk#G|mM0G<Kn&EbNN`KR%qDDBFE*$*0x9L9N$
zJQSj6K!B)&hllNmmq<djR4Ev&DJt7ghZ7*1--R%A2|QTd*LCw9Iff4rbyj(Luns2d
zDL5f|b)D3Y=*U%|BTWuiCWw#NZ%+uLA!3r<1X`_TI88FxIyWEI65dZmn*lsI<#Nb?
zoiQ^=<BVWWQ_;c+ox4si;HsB7MheLsu`mfWh{7CLI0AZ+#<IGwD09<RvY(TY1INyW
z5B#Qy5h3&k#3_r_xs_cxmXAiw%g%xo3-a)@`Tcb<0f+iIt>pbnbn^`nH0Bx!``x4F
ztW(Llwr*-xi94J49W$1$@pv#2f}syT7RDEDaU_yx#{y0yThCnaNZx|%>QdCkDlIZ$
zB@;5Auj#-i!&X$el4l(>S_=gOX6V$AbbCM#+$lCqD)Y;fd<O|sX{#$RtWCuRa)hf+
zvq0I{?NpgAFowCWE3{4ZDs^0V5uQpN{A!iOd;2gA2#I=P&2Q`5EJ*RX``u094o8I1
z4qD@3IlF<L{7iT<OiC!Hz5_!v#=5dyG7(Ah4s*YpJ(%6TT;ZV<I9CZ!o12^+q`AY%
z#!(fpbR=nmD#Hs8jz^p=m~bnd$V4&oIm)izz9W8=<~aB4nZ*7W#8zTaSTWg)_~>^y
z*N3b!htd@#q&M$_(j(I2=?T2Q733ef<3HN-2gsKT96k_2Oh*r8&*;x?3pX%Ai`iW`
zK80=IkA_q25x?E6(YPb>WuEU!452sq8|s1Qb0~uBs2X;M*c5DsV6CyI$RcQHB$>@Z
z-ZG}>rk_+{P}9W629+N|MAAiaIIEO057U15U}*`j3=y$DkZ^<Z@>h<`vWe<EdcGFz
zB}Nv|T~c#c;D_N>FH)7_gyJtCd$s4G<&R>&#nL$!u3gpPXB-CxCkWlbBlE*=rZN=~
zf(92sb09md6UN-Rpa}nJMPwf{+7JP<Z{89P^J@lmo971KpZ`5A#Ld3j=_G7VOh{^E
zGLleJ{BzU`s5`=q#aMPPIdbyLKK3!)g#*M$%%vg&4PxuFJzCJ+;47>gmDar<6uSr1
zCwmi0dYoUhYZBh?Ba8m|11!`fy3#heqqHny#%7o1m}0*`ka>A3W-;fUsh}X@s`0wt
zJ>b*IsC_6bBB&cR$-zM6s$MGm(TY$EkA%(e%95KaW(Zdk*Y}+yo(|i+J@1D(`<30&
z`|Z|g<^`DcQ{BwlGzA?-gK0fu%t*@rB=Y__wf}5Dn8u^mK?`SXnW?$CuD!g}&x8?|
zi%go!Y6Ux4)6|#lA3#fOFsYH(axgP_(U0DnBlgJc{^u^9h3IPEs*ZTfXLeg-TdJ3*
z8lKA!F(T<RT;TKb5A;ceClx*0$@_SabR6+zSKdpU;w*$9T1cq1f0_kOX-_%w$wTd|
zOt@>tHPxU-4~ID-)Z&S(G@(;|<h%Xu7Di^xQ^_4=cdZhgD)PdQ2>3roQ+@Q>xeo2M
z>SCmcn_%SRrihIh%I>*<DuB!yW%E21yFVRmAJ`YUeJs{*v14X2wLF~qh@SZ8dfOl)
z?X~qVTOMe+O=g;l0X41ffpRY#?V!Q=A$n&%q|<_<0X8l4WVVQhcFzP^#S0Vq7n#wx
ze5ShKgS1I|6kF^4=s-<f+e-Nn_fW|MYV~1QUvza9gQhmMZOiyTFIITZz^`Rqp^n*^
z&@K?VRX>%FgvN=vgh6Ygkv&8nMPFU0AGvffdjrVBka1WB+E3)<$dQ+J@Gi{y2ya}i
zo3t%#SNye9Yj9Y4InTfUJ%Xlraoi-SX%V2-Pfdd1SX#A3E~Ia{6h`rs&#smzFEcvJ
zgTy4Ecl<vlM{a2}Qd{ZD&^vQBo7>OnbEb3^xX59iJc#|OJc-_1Wf86s_njH7k(K@y
zqcWuv!A*&TF2moi2S{q;2VCA^$gYY2$#EWb?l)~<yqdng5NU5kwQkN%MY#|0$7};4
zu~$H%yTP0s^{Y~ZLY6b!T(0_Nv6{Wj6H3B){@DiM8rW`f3SEB)V6_*;+Qqb|*D6^G
ztuH#)4Uxw~yKeEq22%bCsr_Tgzd-`YORkSLgu7$1I<{uA+K5qV1wrHrjO>|$kKHuy
zB&0*CTm26DE@T&&&PcK7SsbHZq_5j|eeh7_*9XJT^#R>%p1AU+mPMR=Rh2P)V2d{+
zo^zJwX~Y}}Dp}&Zkh439^QD~I+rfep(1TCM5XRBJS~_W`S53U#TZRyOMR<&vk)pK+
z{lrCN2JeTPK0ny)m|CSKw<<;#ag4C9+x9SV(V1?ym9OSigM`muLbvhfeMIe?GsJYz
zCv7o=M0erRY1d61+6E`t^TY0tokB^HZXMLba~$;beK;f1Xu!B@q;sf(^2Z2r(Ep3x
ziGa*bXH9Zc9qN~tQL#76SYQU4m0i=j#QbbXH?(~qI#V(G`+W(v&`;Yr2?m21C9OMN
zdBML{7c+Xp>L*G(t&IlcsAwKB&zt9sp8fngwLNYQ6GRf!TDrSy_fi$R%d?m^rVtS1
zR?PtWnF=N}_<vIv=M2+DF|L^>H4VKewdBgJLySYLbw}(Q;%0P*d1F1&4<6Jq5AD~V
zKiD2_c*2-Y5Zd}*60b_EiL;?q)4enomKTt^ePXSBVvLIRYjdLo%rsC!wT4+on)4t^
zSe`@kVjBIf_1NuZHCe`A+XJT8cWRh%x>}_ypRAt>B5EQK^hrINwG)fpTxYee47-#B
z!CaMaP2rLPw6-2OTh={;heon;|1xIltC|<u3-5<prMeBQ79b2ldja}Gnbc%jk_qE1
z!@I;SDz`PtCJR#+1@v@6c{mS*-4q@w{aVsVZ>l^(Vgy<IQjNT*wG_);?pfWKrNGx$
zFfpQz735aO={B}Z&Zc`69in+*57T0Du4VqqZdz0I05A^ERf(YfT;~pA-OoG4`eP%q
zB3|ASw;>*FUk;qhvLZ#M)8`q2H<nQh#l{ovJtzNRVhGQP>D-c=F1%*h4Q}K`Ylw#H
zrS@!b;Hw-YY)JOjZ%QuXIz9?5`~Wsrtgkzw%shT<BkM>u`ft)h%Jf>QX7gffTBGU?
zdqFMv*12pC4c)6Q_wjn6RhP(Pf6JvCu=7^JFPgv$YvF#?2e)DNzgEA%y{GC?+1=4x
zCT)C6@|~w#TYK14%&yHOrrl)vzUW;hHdIr=hUhpmTJ9~^W_w->mB+yLN;fx9(f_#G
zgqz{VwB#_~aX(Vc?}LT?c5IDRly0(X*4m-DfEwwKYnEX}OPh1dvNMdbj#P8i@5ejV
zw$%Pbij?o6?E5y${iK^tv5#8y*j=&mJTn1RqYhS228&0T6eRj*d-%=QY#PU&YpR9?
zXMTQr!}k|=5jc|fh-eRK*mF1<XqIoy35WFDC$$Od_2ZcawQ52Tvw}9mx(H%%NxGDh
z-$Z135&t5QVOqlu!g1sR9$d>-v&)t`Gs$kHuwYIG?cm>HZ0-;{8}sCY2i$(8i~pPX
zs%&lfze0uoYa`j;N<ing=<4dCFV_i^?!vEps7kr5M#Cx19oaw(QJ~bin;EF8Yq-8S
z=ZJpsYc+O;JxPL*-X&!vy<H=b8kUe_e~21=-)mGSM&*)omgw;D=X9)YomY25I3Kmg
zseymfI@oJVl=MS&(T8EBjO*503&L$U>S%-(4_lLyfLu+|dxYv<A+KVu)QNd+1;+#5
z$H#;JA;w3%(Q(SHREqi0rnd1}NjSX|Ni2qE(sXr}Dp)`H@)_c{U7t&_u1J)1RlBBd
zagg&$_Nbj}@S|To5IiO5NQ9?*#=?cL+GLn5NW=7*(;VDogO~!4P&K{$yqL-Z9-tzi
z!dtjrP&1Uz2!%e?_TB#VEYW9|#${<ad7v0Mo3CZ!ez6uXU0oN+>4er9wHlHw|M2@*
z(eZ0YI&|)a8fG{rzCQeIAg0lp@QwP9V@8dV?WvSIEKfWiW$>=7C)PIHv94R-;e=Cb
zyPcRz>)2Ob#?|Cwg=M+tWifMR`}=bLA~Q*z(w&tFj$s2#UYf!v@hN5Qs7lP*wSmV%
zs&pi}y48Bivj^cTa%iC>bvL1zh%k%rfnU3~_?_ID?#T}i4sS-zNsBS;5XXYHlQHTA
zH1y>KGlnkLFkTI-+;tUn@EfIBOkj1u=#fO*Klc<2NYRlCCoRhia7)A8-HnSO590|_
zn4wPeZgIIA+i&AnU42{QK#4=WUqi(D<6sh8Y$p7dJ_?hv6cMcsoJu(B=ZK4ml<}F3
z>P0!Qmls49gM{Pd!n{;O=bgOFN7_4K)Xw#DrwR4cKeQN=AQ3{bCxUuo+{-QUv9=LO
zQP;>*b_CQVe53UerUt@LH`A7NaO`-n`8qpJpMU7shE-ktis8R6`}lN^e;92Tk_wC|
z&_2Qnt@^xl87+2G8tUn}`01iwd=(;dv43W_n5G<)cFgs?D>_SU?nSQ0uSIAMLTi{S
z-LQXS*Z3OOo0_OTix;CVHgW371|Rxs$qG6a_V!l-M{)DQ=uA~?Q6R;H<i1wj{=ca*
znS{ci-Mmp3rLtFdp_LM+=~ooIY8|HMZNx?w`i7IH3Y`^#btcMQK(algI)i7)OO))7
zCe#1!^R}KOLR_#5;?<i@Zk6R}z3;Pzj)mJ+$}U$58-+8s%E7mTM{&>Ij<$Wpgjow9
zs<tlIx$PJUiPYx)@+C}s&e!1OaBUO)qcDYp%Amb>=k$ILzMT>dtI7G8Lp{X9U%{RE
zwpdP%zA`$_Ez43Iht_g@{wtdhCOvFU-+dpQ)L`Qn#EC+Nms3?!1hY4A6s(Y%=6tF`
zZxx<TJ*mQ(YbB18b=8=nB+QckZPA_R8ZzP8<H&5@1a1<}v1B))kxXjZoj5gNXhHjl
ztgO2O?|X)Z$e%&EU&5?K;ITFyrST>#)v{~9dJxY@J|d4Dhm0$r;qVx%@wgw`Gg1DA
z;6=h&9L?XsN2{Tv_LS3}F(oTo{jUE|%lfeK)O9z)n%{p2zwtw+fVbr-Z<X?7o9C(4
zS9O4`M})7HUMSyYw%a)d48!Sgw=iRq<=xVaoTBCq#IFQ0;VrSK_JT@XT`D84-B)9}
z^isAvkNRzxKw7&^TBo^1a!{c9Ug1IrhG9f|I^wh@Eng9)@H2C#7-!Vk8M;WVBbMRs
z+Jg#FvzZ%{+^C!}o|jPv;}|+CI%*xq3%61Ud>ijsU6rFjbAMUP2BvD9ctf(1M(=*}
zUln+Ne<gybrR1Ym&9xEv^q7KdU%rfm(4IR3@#sd4jkE<d4xPpRsjtu2?bp(}*}?&S
z?Ip&<s_N$8e^W%LRa;Zqlmi6aZcSn}WsiC&g1yRg?($I6(~J|4T_gG-avY*%5$Xah
z>>Fk7-0hr^vDW+hDhZ`Xf}iwSlH%=D$K+Y=uGSIZsj(>G`r~yu-#F|ntG44tS^GY!
z`H^F)?iDp<Y~RMrxQF(gE__=p6$jgM`!@$8RKkKDw@I=76{~^m{GFn_peDmvSkvn4
zaxI#6`P#W9+{`Hu1SwN4Rh>@hR4bQY11OTKv(w2!uwM$mITG&-HSX_QaS>e5+!uB|
zsm;QZ&#0=oxvv$e8v5STxg3LT)HGdvmlvlR+_b6ZerNdsY$!DHUw0M?cJ@Ovw*lj=
zwRXt-n(ez$+_-yP-4|!`!_-j5oSrc_@-%v|e+III4ikrIW;mQ|X8eb&6KH$(R8N1W
z9B7s3tibpjX}z9o&E{mqL#vQwBXp>$g;|2)t$>Z#l<82a;Dux^8q3Rl2*zM`E}E9n
zU&{Y)C~MNJq!;dJdG)feq+^%(FrNN~z)4qkQJBu|ygQp2S^-~Xv7u#;;DA&2ZIWbn
zUGs2@+1S6|NLW1i&v~<>YaR=~BNyBZUrD%>DFaD0;i?v3<lK)cxNn#JLvF0BL>RHW
zy_1LCr`CRo8ZVeKt9u%a)C*+!m2~|EL7yp6_}waTHUfR?eEnCMVN%R|!)+y10jyC2
z7E$`5%rS^pXBjuNinS-h7(r%D3Uk6wQ|bJR>yN$Ks#5QUh)gxrO}TB^(+F`f2_ZM(
z3!mr9(d3tN71yNQWP@@SR3+^_NF3*0A~)0jbkSxgjEmrK{^2SSfmqcCU;ByYnbHk*
zoIiqkVRwi{S8TjTYiDnh^cqY%jJk9%ByXMj#(mEH{UvmPs$moD`D}&x?Xii)=Ul(G
zA!L^=v+PG}n<~9}=yuslYP|c#CjEGB%t<2o3uC)_VbmUO3NS@jA4kC+iR)=ygmqtb
z=6JQp#xHdh)9c`x`M#d}xrp(6$|9~P3B-JF((C;Tu@4K9sLbo5jbKWg3pvu0ddapR
z7(m^lQJ%ldjz<5RCm$;DJGiu+F*&9W@q0T5ftoCzuOlh)r$%HHxAz^NY4xDmY-6_0
zfjM&)jGowomlrwkQ;qtk*2Mj46HsJl&L`Y%RAJ?WLleYpM|L5xH%-DuA^jLEuPjT!
z!OUkkwpeAMJ881bkEd1B=X~Z?FpFs3zqRm4q3(iPyUL)+X+?~rGvb_R<^JHMlKhFk
zQkTx9C?&`(I`6?UuIU(7iVOonu(Ot{VVhwac%SIkj_w2M5=}l`Q4tY4q86xBLb$Va
z>?m$p2;I3U4TslP+*wAx81Rv%l7f*X2Fw%F_vd(3bN@03`5OdDwM2@=QPFnTvm1!t
zvK@m=no!gB4yrrfta2VS$x0F1ySqGY;`W^_97C<WO(J?_fVpG8Y)<E3q4n4=^<foB
zck6VEIx9zJr{Mj%`2C#DyF(&&6?tz}1154%)zB(a47u~6#&M1U>^ZH61KX0@st*hE
z%A_oQbsazgwnr&-ZnTfRF5&}PBfI9~Odq7Kt_VdkmqL!Ah1yVia8)GiDJOzI>CHqT
zpooco)38rba(1q?7R^u_@msEAka8pNdNM_(eu1@=N%3%1>TYW<&mQ{a`AB9`pe#Lx
z;$hMb(=5O6e(<SFjGZw<*OXj$S%P$?DGJjmhEVL@)DrtZ2cJ(h)6F8^G+Z}PN-U&%
zjT}sV5p)<b3USTl=7UeD{aE^y-U5j?$ATn($DmXrkT1(F+sqxkvhAyoIGpk34=jK=
zV-Ji&W}fo$tM1e=SLtuYv<19mrbz}os#4H~uw2GV7yhYZ36zcmgkfL?5+^O{+zG|x
zk!Y%I4yx*kiiyj?1UvJ}sc0uE77VaOWEcVBy}n($eK)fhif5=_`IYL$o{M8Mqk)UF
zvYD<Z?wL^=KSch;ur2obru99PVPu3$->P9D%ZDt*X8|DQomHe5nEy+QkfD?iUM4DQ
zg)$fS(%gnLr2BuOs(yN!vr2v`sm&(Xc)vI=L^2bOTDLV`4RZ*$FLa5DIa1O|=w{$&
zu1pWfOp6W;i^OVowT+i8qIrnKB(2)Imff^1)Le%&wcSW#vO3mGJe=HOxS#zpu%cd6
zLopNQZ*KvBNXs+r2Fi<scjGx0q1;7H!3;wk21;I<;t(#|3})&fpKYIYo4y#1UkXww
zBsJZHdvrjh6rR-|w>^k?1`VaImQo8BXNQ8^D|eZK`uP#gm%_%Cw%H7k2flfd^s$1e
zM{?MI7bo>B&&9*(=W#qRS!lI{h24niIJ7QD?!}5+-KJGw@r$+3N$_g*q5z&pWfLS6
zIJzB^v56>gHN`W(5}Py#iehH09}7Y5a!y%b?5&t%d&H@$DY#QAGcF|*sPckBGy5W5
z>0Yr8*~T0h)MH_&N9>$N=(Bp6UuFw5SU(i)h{yr`<D&j>D6_P@ifytPYvAYqtT_pL
z2Ayt}pt&b$mAG-;OOp|jIdwCYBOVH$9SajWk|HM%@1M?aLLP~G^1^miLJPD((;k*t
z<xnNMl<Ox=Q~q^EKV7=mIn2bSFRCx?&T^HaMKeWD3MYyCd%a!$Lzu=!oFYL&{-1V`
zpWQO9`5Q@0rt#8u&XF+ZOiEe~hDgI_;yX?}EJG7#gL9yRi!6T9E$zmLVZGS*h}C`L
zLv^7e{Z{{&?h1)^hX2S)k&lMELuQIke@3bUl$@HjUNo<ZSy!yR@HHmXUO;yFhNyl~
zp=72R=x23S3*(_us*hegbjcWX+5Z=}MfU3O?wuf5m8O7-VH2fcQWJ%So`&vk>*G;A
zP;Pf>o3->{m*7!@9rq9PRN7|bb2%NWv-3Eyf>Sq_-;8cg`b3Hz&3gY`<ieyLsnLHS
z!7clde<(;Hx5Z^la)wgVmOq6?aF}qQjTzGSPS@3YsDNN@S5XbUuvqor?@$lr2`Kl9
z2=lMAAD$AvKuDK$YnIt%XqOCoNs8L^5nnbh)|oSMRI9p14p|r4*go1s#<WW(c4d1W
zULTcX(Jkup(yZQSz65#+f1Ru%4bx%@cq5wgxgPWO6IV_ULr#E(q?O6bluZ>*^n-bq
zI&Rh5YOawFwdYS1;5gtJTXsRT+MjUAjDr7`4q>6u(yVv6V|BLv{=GhZqM`nc3*Z{o
zWldv;!tBCiO?NZYR>SD98{s{7b~;T_u?6FfbqJqGKaAc)t;nxl$!8=IPC~@&<cET7
zQ$U91_1h8ke1^Q(w>!KINz}9-E-AUhMQAu*Os3f3w;i!dlgu<HI=t~Y(65?8ERWw+
z;i+?l%<TW2F33=SM7QozDy51})$AH*-Y8SYqLni*lzb5ue=s@ba<>X?t2~pwj{D}k
zmH{d`-;JwafH_gg?&=m$_YsHbCWvJB2`hP2vn_c<@yUG?$m6GaGPFO~-S~Hz|MStB
zc7bSU=z7Pi=<7oy0XO^>X?x*`)Ci{k$JcvDv*G{y<L_!!)oRgdtK{9GR?*r!-HIAD
zYlf<t5nBjSqxG&D@7AhK)v6IQR_v<QiXDW+C_xZJ1c~IA_dWOC@BQ5K``mN#pOcg0
z^?J_7^YMJVu*BVXUS`7Xn_Ww17JS|zd7Xu-+XdVku^>S6<V`q%#i6+taQs1A#l-AU
z33*8{7X!Fp9l*6`N(EWaLtaaxLabc=J?w7$mTmehEdTHG@I5iPpwcdKIZ~Fe+YD8l
zt?$#D2>i6yFwYj`p2!H-YMqBw#)<HnsZ<(wxmtk2smjj6g{jsm*#oXUh-HeN^jnKo
zwl-9ijw6jS5E$$j-e5uUc!vJ}FG=%;*cabwfg=5ulgmp98d9>WGRE|X9dSLKw-%Za
zo(O$_NTAE|t7V@a#CrvUwZh?b%e+>@W1gPJ*>mpEojtutg1MFl)^NKqaVfhLYjE2*
z&AW#{&s!watzVqdpQEDJp53rRS8R<HQ2)1+{%2OacYO9c7pzRFYuitTk3CV~id9<A
zA?NSi=Eh6Q{oJ(r%c*$B>E_d{`J37^^Vbp{mTQO5M})0Dt-jD65LVXAuF~_{xIVcx
z3t}9~Kug!#I!vxjJSLN-S)FjvPFET8>EiSM^W2*R_pMjyp}LN6;|*Gq@?_}c$9Rvs
zrZ)CmDD;YA;=cQI7Eb~A3Ne#Zh>=V@4sv~ky-=r#ZDL)|fw7BJxLww9%8v|?kdU9R
zo3h)Mga#yYGt_U5gQzT96=kvW*IrOcXj_6X(8LkLnnP-U4$XEjH6M%<B{?L($su6)
z<a=y4!i5y_xFnyv4-h<1qF@@9Fq1m~-EBXUGyjtz_&>|kyqQzJG8^$DoJUcPLiZog
z5^sgJcSelFIqj`>y}@W`W(bMea-)pW^P>4a>$o3X6|?cN_#hxGCAzTxja93~jJ5q<
zF^uIZ*@_x7;H)SOBk#vZ;=&Ua+tkR(j(z9E>cJJ4?U_Ta5pt~?O9mFGO^v=DJqhlX
z&;CiHdC(@E!jSP9FFp5uO2FDzk*RK4*!2LH5wdIu>e=Oj#!D(T!H?7)`s`@f`7LaP
zJEty&Y2@txd??pH8D8|3PqMQ+6(ZGc#%l8x&rVGv(%N;pFb+P6Q>a!%-e7`B&hswJ
zR7?h=q9)^k&F%SIPT8TbBF{$XMW#j58#p35H*vwE9iEaaFsJk1Zu_tPyl(O*rVA%1
zc^I`I>}1U|g6a%GDQ55jL}Ma7V1s;VKe+oBA8Vpvc2+vNn81-|LT9>>Adi{H4}n3~
zAg+p=gffYICD$%;4P~$(UNml!+?+t$H}}ku^Jp&ShqMsiV*(s0y?Tv%rL@cKX201(
z+S=rGcaCK1etc-Q53&@)`G(n{HtS=lz!4*pnZOW6_2O~@dhjuEJG8FIt}eczfO0cj
z6to^pSMuI*VIG_w@z>G9RYZryEcN8J=P#$nqDH)+5^p*vGcj|=N-uN=3(SwTSUHgd
zqK@DN8YE<YHyFF(g{CZPH<V1>Huyx_+LipIo_A0U;I^DVKsRYXZ$mol!M{14b3gz5
zTi;Pm@X~ST_FDC^lIsjvGh-}K)a{UO`B7oRVqnyw^U#DIsHftCZybMYrtDdy=^|KP
zi?e5fpxD|pYVC$-JV$%aADfMXH%EFP+p_b)sBEr$aS5VS{```Ykuho(5Zi0B?Y)EP
z(Gj}bjg(b?_cE;^;6zY)P5dGoq~G<(Pb=h*OAocY+Z??)G#X7^zV0?#6_XS5VC`gv
zz0S_z;cka)NZ2>1hLGKyiV79@nQ1y!7V>&rvw6#^2YV1;!<tdG?A_d33d*oouwGrt
zDUHUgx4CNb^$2WxERim^i*j5f-7%P{ZC~M##7K@^p#H+Eh<@C>F2D}GrkD>d4%i^%
zS~4E;SyQ+UU`r~o%rh$)>p+pAe|Kgjz=$Hj{pR<}gg&#L$NWO15Gj_8To*L5ts{zD
ze3U3V+*Gv~F!X%1J|g4tq?(2^yAWsVroJ9F**Sbhhm-v=q9K;wI_j1tnr4VmK~U=!
z0~o8~ZZpp_2+P_}P;Vp4(hh4{WMBTUfob{Vv!4IFKKK?r9vJ(Y<!#i%ht%N=)=>p?
zFI~&nZp-57dZvXF9TXPL6w2Kw6VIP-b*Xc()Or`z10q(<>1BBPbPzoSUa$My=wSz;
zsbX)(>f7Cl(r8LMn2QySY;*N6=JPIF=cGkZsZHGRG5q5qu;>iP6ybXGO3eNz#?)GX
zeG(HAZg8yi8_1tO@$L;g_4SdOov_PKtN-22{8vq?By~=1BdpXf>l1$>j{7k`u=#Uf
z7C(e1WYL+i-3|}EuNYP<+8X-hQ>DX|iqxxzD-_KT>ALWYGwdk9+jKWk#IPoW_)4m=
zh?p4^km_6c$?dSy`r^Kb+_x>!Owt#V=KXK4L$Dg7GUKRgBff~G^Tqi+ptOmufhVHW
z(cNLRK|z4sZyxewe@!dbOM7$=SS9fZ5B6AHjX+~IHa0iay!0F~$k-2DN`aoRGD@qf
z%xHexWX^ePTl4q(FK0Z9T907U_$b*_A=X~zi6JwbSr=wRCu1L=bl~wJwf&yl<|4SK
zO-N!V*C0f%%p~H%w>rGP@W1I%N<1Q<EdP2ks&yn^B`0B1IO5Lw?~>eJ4fT)gdW?Ds
z=U$L=sJtRGr7|-~g9wAvoI?_)Le|c*ws%r%GR2H+Jo?!65;rJMYaICH+@l#sqb)m!
zA<bX4DkLE^f|y@w&Xm0MR=}=Iqo^=8?L&r_hBS_Pb(tfU>+70t#tKZUeQ3cJ7rPS7
zP>9tT4Sr=;^x)e?F`a9$V|{wG#d9{GIfOn#x7@)5<=Yr-Za?Y)NiQVcEOI?M+XEu7
z=nnoPvbpB;W?Z@8Qow}QzcSwL9iB>+#OJHY|Bps;kbnEkuU~G+tZ%^rf{Co>;&8vJ
zya!zwjokg+8`Q37Q4FhYwMblP0+GM0`SXc3rByhu+f4oF89JFFA_&4w-SFs!9lgyM
zntb3RiB(VEAGwZwq9#iesQ}s}idJkWJM28H`OMlFaT;V7PORFMa1!!>n7xy*(HVKT
zyZiIT9yG?cVUf%?=;amPD*6Glh&MSqTND*+wkz4UtXHJ3?eRRxMy_HF7&`@;W|2Y2
zS9?IhOfyk)`LIoo^D;&eZZkN1?!NXTE2|yi0_M%XLgeGo_YS;^ZW+Ey`hQubgewB$
zp9&!O6s;jCu0NGWv}-RsYwg2~goL{a8N~EjR6DQTS}J%rCS!zYsKPlIG&)qjRP281
z2(qSzEe9-@tu~2DHNWk@V)PfqciMI4hUM{*a$&%*o@`_I7rOEEv+0M^y3=NBUk`SN
zhRR~OwC(W-%5H4tplgb_cXM-Hj&Ix|VpBTB7P~(4r=XRV^$Zxve|DSCL!HsOvCP6F
zr$S16Q+(^>^`;GWR$u;KL;U1>^H7oQv;zN(5XA7fm(n9FAf@@w5Qaq^eUx>7o<iCW
zK}U}jwARPTnz1!tMY*}|6B(D0c4Ei}PGRI`g)d;8)sddB;}=P3I@@oNEw@4kAYE~a
ziAFlatjyQxF5n9?)bPc}OSYgirMgy~KVd+o@dK=lK#q#^z)$iBySESsV9Jt`1?jm0
zSUHz*?)59^p2r%Q@|QJherqwjpqVlI_+T>=>Iw(uRHDkGF=0hT5Ocax>L$$egPRz?
zs!s0?aS3>F?SJu(BEAW^T2Wj&IR8h;GQJeyGvbAGdliIz(r<U>3c#S{&hM26#X4YT
zcbbd!@K5??-RuozyRWLoZpJ1ynDNUO_GgvhCUp3@#OwS1uB1l-B9tkOfoJ5^$j$t|
z!~$oV)se@Zz#o}9V?3ssyOdJb&c9bj_+<AQonaN06TU*k6$zR|-RRi^K5pBa6purQ
z*chg#**Y+F@%e!4S5RK#c1Ueaam^RsD&O`+ga!-o9C|JkTDm%sLP$3FKRE+}R)1nX
zCa*OwU0-G<7q9}OrjQG2*5|yOosE@(*4Z(lt*}jJ&-c3{bmscqxxaq01X%9ZB{Wb=
zCWfylE(E6^1qRm5QXdSAi;l99y5BP$oyFq3?JUA}ogE)C8DG}Mm8}&}np;;uy#&EY
z&H{6zMC4^*&|PQsa<Y{t16{BgiZXB_FC}diuo8=3b*6>r2{<U^cVm0I<_z?*a}`T)
zAfLajsBL_Y(pRb1OacxTnSbuUt)p#wfTls;s@g_G{ym=SZkQ*?ZdssJ+W2M(1{1CP
zj(rlco3OatCiE&u45!Iz<lu~gwfcg$o0QeA$Pb04DOHBD?3geo?6KL{&dT{v(Lteo
zV((D(1wC``IqTjH;r@&aoQ2xM%ZO-3@FN=`R*8$a?3mr|p&ou{iD1&EcJEjRkf%`0
zVmFYUD(-0YS4%7v*216hAy9N=<1i@&Ai<)K<b1eizja--0>uU&UpTgovn<jXTWD&x
zpi|~C$rfR3jHZR$__miFM&A&$z5gdl^}maXi1qUl=A9F1jiHZy?)IjC(ac8yw5Xaw
zq5CGI^`=UG`TR)cgE85g2TdQ&+B_Kx&(ZIkpw@c+1*R1hX?>x{^w2<gbi?#wC_aaD
zy~3vSY^AmrvJoN*p3(a?9?JVckVFt(HeY5J8z9sDQlKqQjk<ioIOmI5;xqVO1=K~+
zx#k3Zh|StBg5Ff%mm+}qgNR8o-eQI|F6U~5!awbPT||(!)x*ly@Djojbu;Hy5o6ob
zTpjsb6XD|T|F#BTVVl8hg;I@218L^6X10Hvc`OP%%K0Swn_^N^f8TnQ_PF^cVbivb
z!rHa_;eY-Y+4f!>Xus}S20U3*p+2C#qXz^5EJiL~>UnG10GD)}_W7l@EZx|A=QxWs
zcet+S-f3DpioBme{1PN#7DM*Aw)dx4ef=<ZuE(XYtCQ3M9Vyrz1Sn=lUY*wv6mnm}
z4o;A{k`vBWD6)y>H^=+7Edno{LBxPQa3tqSBcx^fj$h<Yq~7Dp9(vJ(p!LMIBmOnl
z{VUGQ(n+%m;KfSGCMy^JF~7M4uF7f5yK_xjq1P=jj<Pxsp@&t(Rw9BuzNUk~wYCR1
z<7w+1HI+&F(ViL7cw>W&;;UzB1E^4ay5o@z7OcEDcJ^jf^FCQ&*<Q?kZ~y0#M~YTT
zpq{I2U)Z<CGc8u1e>n}Ff!<`A#=@T*@m-@U{I9(9Kf)TT2n)FBT{A?v-Sf8X+2%zV
z@4KC8x-6+62wR&bS+vi&4;I@lbTx52++s(p(C%r#j;1NrjB7<gwy(LvuuLXGSv{Mz
z1=M`0)c^lPl8nXKy8Uxx`44=%+BDtWv5ys2r0rc?*B|8EFpgrA&GPzsfNQ;x3=xDM
zYb>RdU4YB|wl>3BVh={UoCJk@-0H=KO^W?W7ZFJ+f{C907Z`R=<ji-jZ*3fFaw*v<
z-N_z;iHm+G7TQ_NW$O#u$G(YXUZ%Gs^E=jQYrK{$Q8kNQu&BXS7{0jio=%xo#F)RH
zY`Kxy>2che&y#<HJp6OHe;Ybj4qin>UpJ-ud|Uy2@n<dGjK|)SsKR{{IC{YeBFiSX
z7TK0?Tg3VRhivLV0Mhlh7MW}uYjv9PA<Z>*;GF9&|3b&9|MTxnoNB#yc-F~(kbbX-
z!1a5`+<LCMS;+v8!js7Nyy^|(9CD*O^!t@^3v1_JwY?Jm2VRp%d!vKpF;O<Nc;04=
zvkY|`IL#ujDsk<Z6F7Qmto`L>L)$9iVt128izRwI!{wG}*B}*F^Jhzj-HjY{vPWBE
zL8M%%pnnk@`!Bp8LY;q`BcAbxylI#7XwI-Y)AB*uT!D7*_j?s}8vfS<+mIH9fBomy
zVIJG5NfUtW0P1mi$#?u1s+5o>O_P+CT6Wdm(urkGxzqkp^Gkiy0q@u?U6f`&s2T5Z
z&#qhQ?9=UYjQwzEm;7VK*SgbfZYp+UH+AF)_k$ITfm7VH55RZ8)Hnn-%Z((y(V&K{
zH6Wj&>oj(1F*usiS5!z6p8s)<f0x}mXvL1>Yj<+73qg%=W=!smqrK_Ti1N%(qL=8^
zk8VIwhjXcg+AF>TKN6OfOXI1$rPnq869|Z?IiqOd9S?V1bHif?AXR5aa3hAH^supw
zxJZvH$%?hS7){{>!`kv3mjx3vLYu(i0N39hwij=RTBH^5nifgzP^BPihZp)7eX)9l
zbtl$;KI=K&w3N*L0$>+fEz}ATtZ%<1XF`u~xt0O^N{`>oSp>d|gAc-H;CaxL#ie48
zwm!Y7m05r3ujwih$iX4W{$rfNh8)GGCRM9o_UnQ6!oo>w1Le}glVfrzRrqTLGqEXW
zuT>ciwCZ@qq_6Z^IR1yALf&6&wRI4teO7_ua@g+>);Ye`?*Yq+(Y>^8sTJ?Q=^E1{
zQ0rX#mkH|o92oB62x{9q)YD4U0=fsg06%^3A26kuO@Cn4_$HL}w^Wzke`Z>iX5vA|
z??8$J3-S9#IGJ)I*+&mr;9pb3HA0VwqXrviWN(VfkTYwn-57ZrKFgNun)yq0a>7R^
zRcC_O8be(=1^=<@LAJ1oV9=L|$1I*7C>4kT?7*T4DIYE*I@sx!{ziUKCopK;MHi%+
z3if2=P9!P^bdG?EnP#>d1I34Y9HQo3m{tJLV#_keq6-#`&MoG)FtL&3s@w@@X+pHX
zlV|d8qxr^b4z#y^)_9o;9m0kT*BC!*OKBxB6!{Al3WEp^fuZv(w<$RlVeCyI^`19a
zHExb9!*e||MJ4uTRPiDB<wm#8gOe)Gg`g(iGkMX<o6?cCWv@b*Ih(4_EmJiGms-0&
zWh?ymDDn68bJ|~<+CIjsyB@~G;_n<HhPX!7E9R}w)Q*Q4zj$`Riheb1Pj=ujE0s%G
zIRpoI!IoL-Eqc)0bK2#H&&XRAhR=~Ebj-%Yq@i(4m^S0&ZBSr5D^`F5tf>;=6=62K
zaIi`-W&BC51TIkr$*@?<|DM^Nz-!(`{h;&3QxJ52tVgO5R~gpJbf|5~fqbExH5sSp
zH<A)`Y84Bo`-v@|H*|zPwn!b;IV5c*YLcfR6=34|QO5Ld{}0($*)f++C$S>kz00Q-
z*)5q~n#~`zze+cTBzatbwBK{<BTd>AP>R<_(ryO!Tbzj(vONy}lOdX)m_sVhDnFpc
z(H5SW(rM!L&Kid6%{?re3E|H9@5z$^zpH!aga(--Yr)_n(D+w_Gi}=`XmhrAcQpk{
z&O{CETvmHV>#huy-JG1|>-Sw<G$~VFhW5H~X)MckO#s+d$k!lF$GaT!hoNV-_}J%q
zpAGgfe^`q(D^9wk{_``Gn^lp9lQ(t@%^8JprnbgsNecA)!}@)A)}kv;ucA&m;JUUE
zaMay0uh4MQw+BJ-==-Kuxp^8qzu9h~8LMG~y~kB7X^-#egGyU;)>i|kZ0)iyRWurX
zNCrquTZ3e{fgF>gdoHmJTDrfAooPTFZCXVdSh8`lUwSb3`Wu#td~{@)oa6@ngF>>X
zm3n6BeDlM)R;tgd)XmnPxoZcBjVLd=7Eyad6ODiHz5&`9P=64qIL4UULeo&Wj&7+g
z8it{im#sm60ixVk&h}^7Jq;OyR-XVQfq_oX{2=Jq8hZBPuk6Rn-z`Q=apgV9Hc-*|
zdwFYYO@g-6Nw4GAKX-n9!b>vFEngGEqThxUP@b%g0P3K^y?~++5a1D73z4~)YV9rZ
zEf_7ysQQ6z#3xD|r+9){`~w!w3vIdF_jLKHnB8-T$6yd}%WO8#c*~@VL)$>BvcUih
zALS;|%BLb%L`mzeAfAxD;KL|Y5?1Y@F4#_SN!~f`C3QOX80FQ>_U*V;4$wOYc^PJ<
z3FuArN|UfBJnDh`U=~4{<O}US;m@1Q*1mo9V2(V40Pk2s<3#c1q1rJz<gO?XbD{55
zPe1-SBY&ghyNy`mcUre1+ZjR2JQEyOGayWm>9uzQ>~@0b#mryCt1&IK&FHn7-`P`r
z`ZoJY`vUbe3N133t7|?z1o(rBoSTIDM;#z4;|i5l%+Oy^*1mhMESlQ7Lrs`Q(z7of
z5&L>=l#q)xBcXB{D#}3Y?>}X@@dX;i<*U?g`@pb1y1&tJjBlHI0tRmPYhMQrkDG*9
zBzg%+<Gjjqe~)z0dZT=2b(*O^9fcS-j6qw|$U(R1{cJC|b`IW@Mm6xuz?W$Zx;0(2
z&D`z%n11Al_Gjw!*J+c_-}|%WUNf~M9S5Y*(7DyhRD$(>PL(2n6^yh%8l3H8Tj-i@
zXy*hI$T^5O7LSEDXkUH(8ak_BjV2!Xl?~Tic#w$Ev&UP>c>cY-GvI26AfdMwWw1Vr
z#0$jMhxAK3YqoHBZ^K|fbFH%LN8%G}e7nwZD~o3s!Ot1Zy!saKCSL1Z`^#SxgFsV5
zaeSCpSyVKG{4P&&u{^4S%HD;JE(&$vhm*%`T4r=P@liV}4y8K5H?2uew0IN10?2q^
zB|jJitHY^;Xo?}@bJ2-fxn*kx$>Wpu3SP#GG+u~Fw<dgbDfi*d#ROs`?pzQ169j&I
z$$cVjlUHV^tunr1WOhkQWK%jzpT4s`#n5WX-mT`JJk%TGazg8PG$-*xV%Fbds6O<C
z`LC-@gu{~$RfJ}aXM4HQym$Oa_{3{G8Mvpq;2cr_UCGByQ3R=Gi=La2#KAw4_|7dM
z3eZd9dafz+{6}3kSQZ3g=}cayhpxV$cNc0!;u8}-P@ME;>ero$xFi|b3w+OP$D1s7
zNO9X&4o83ucoVPZh9-24Z?=eRUOOZ*GGT*yWq+#Bvvt0s>CAza?gv}tzbar=MaPVZ
zx4oXEbf?UlOw~@j0&6RyRk5(5C}k^=pIM;D1gV7rj$pNt;J+=-Vt<=@Q>w!OSd4Pc
z$D;aC)b_RQ>XBUo%B7>K#~I+7!7C+i3kq%z7f{LfJ?&KtS`J$qF<hivn)^|1ER%0O
z7qqjI3$bJo2EvW59PQ^9$(fk|XZJid!|<hzn|ZT^p+!&|X4^J135EBssC}JNfkNCi
zLB+#g9|R3SlKnwKX|8RF!Cw~Hi1_5Ytc9bt=k{jwHf#@ed{RdNRk8Ha*OlmcJ<B8U
zj>57}Ne-XoEoplU5YNc}-p)($*QF&Eg<fqp^y-x$N#XXr3olwr0+SK!azpvsY*_cS
zHC(}wU&6MQ2yk6wgzmAVkY0HvV*B#X()D<}OALnehP$=j%$u?fk4I8T?97~B<6aYN
zfa3dZ=WX0_UE9*Y=QByamQsdQ+Qr?<14hp(2tvq@`ue651T19sU@=wr3Z(`7^PnfJ
z2eN2b_$4LznJuA_z%|hsj}^a>5Nxs4f=dPgJ<~-R3(g2PJuDnG{55WYVS|(2;SZH5
z+M&F{|J1e;U*6!yX_=L!P4esdrnpyxnaxn@nhM{7CO!8l3*`DOdoKKzE&gXw_ziO8
zBB`k<7VoVFkjxPZw{Z0>UEx!>C}H1=uVo|>R-PBFjgaBOnFJL{uNG;0*5Bw09cP>$
z_5fCnPDU=Jti@u2@9Bb7T;HXsSxPguzgK)%WDdg%q0_y3bOU|Wt!IO|1;0aV!Jn!N
zX?)N85-}j1>ZP}_Xs4HZv*aU*e9G{5&mQ8`FJRDd^Wn+-{J!GYw0H~pm6*ViTbfL(
zo_@{ax?S-eY?IbkXdCDT(7u%07zJ<N;*eZW+U1DLPKJ663qXQ=_EuG9n8t=?A<|xs
z=~TaxMFgN>A|+E0(*AVX#;5F5T94KOiz)jT*Fh9Uvaf9?o&EDZT9bLzTm7N~ylJZo
zm~=~8yA6(M{i0bg9^z3dVEb?fjl)YcNJAtTVIua44!g}Ob0r5+F%`+<k<qUwz{FyO
z+;0b3F>v^(l+C=(x_#6l^+9<1M-Qy7@PLW!PcBkn&E^f14w+UKn#n|5ZBs#m7l%&w
zAb&N?uTu*a4s>3rhC_`ua&Kfi5f7*_4Xv`u2-dq<1)E(yrU1+LfxfPFIe5ZRK(?S3
z^++Ij7&#UV+~cP0(2SV&WT(%Bu<yXMGlbuMm%bd-Q=7GQ;KL%w%zxq&l*|mGlNHLB
zoZ$K26RD`2V#c#DiJ295t)p&YRoFqh5fqo_EG)UR!W55Q4j+8buZZu*N@`n74Kd}m
zF~QI+wjHD7@s%Z*NV{R$lQGB4ZVKvR1{0w4Y+C<46Iga}DBiCmG{Teg{Fp2e+TMfy
zo#4u--${Q~d_jW3dT<*(VOoHjEz@a#$7r2L+nd-RBZ)=UYA8lbq1PoXDqv&4428=r
z2wu2mb8YXyRT$D>J(c%S2i_Wd{9cmCDz08A(Q_!Ii3g~lr*n|X_V#1=_1u1}qmG;C
z!_HZR8$GD8y=cn<74~$Uw53uvfX9{&@{T1}Wy7iOs8=0v=SzimSPH&p#qT_B*;UDe
zpDCY>j)wg9gw)h$zOJ{QS>KUZ3%{?390S)sIOsgjOf+!9R<pj)CU6EQP|5U~G%*b?
zsDHnk$KQsReP7}8S`iPpgiILZOX_{nkLOZ&nYi%KU1-I`2D<dyNxQi2ncaP~oGEJd
z?FbZWY}6HVp734hdlAPE)__YHn7i~NFey-e4(_n4h4WGW#4~!>EU>_X_Xg^Y9l6^>
z1sYsD5A!qj!#7vywz}YjlOl+q=j&aohs`0}ki8*YtK$3U?z>MsnuGdwPlUAH+u3P&
zYIQNkzL1mD{Wt+CU967+JiGCx<7I>=kVL*aV=GK^d1JcHr#(x$ucyj*o%7sjupNUj
zqqnRpF27!OcCO?f!X`4_9N$<zu^hTA9t-?Qg`KeujM%K70`NCg)#~k=TUoz8(QXtA
zpKJd{D3&N&8Aim{ZxLeg*sHnDExgtj6H>sfw_9Gge`MM9e$bYQzC7@<AoevOd7|i1
zf2C1+?wT8y%r>Tp=6(NN%2s?T?Jkb4J%=5%*ccsOLEoiobfp(T0qyZQF#7{|fBr~~
z3uTpBU{7-Ec}vNUwb5{HUUW8+*kAqHSdn)SR+Xk^Bb|TfQcO5}=)r_LM=vsqZx{HR
z(`Vit9}!$ob|vDt$uKG=r@_Yqk(GC#Rg;!m_#OK&Wbu&Ew9zZg()$6Ogs&_zgWfuY
zm#%q>W;|m!7B%>UK%b1)gkAKnp4>UY5%T+XW3lUf2C->}#4v~$?#kZx{;qum#q6V)
zv>i&B&(Vc|8NzC45&ijE1xt?NZO5M?H$#sL+jJ-djYOT1?oX*(<XADT^cQ;bhFWBT
z8AGjY(L_I%{3<n&h#`-=Qp3@g!c*}V2R{JIi^Ufxc&WBOGem`C0I4kazXwk~P{@X|
znvQxM)h+j(_?l-l>gR?Xk_@wrhgI@)sZ-s%ut#@s88qCr4W}RjXnPZbu}_QgW)Net
z3pI~CdWo~2<(ZYPZ6e{t2TTVqhDd0UJZLA7%5rneOuI{~XL(J*B2~X%QaMy_N4_2M
z^w=uT>Yr>8-66R79q)=EG}f{_#BbKT(*kb++ig;17~Nz6N$0jR@g~S^g-*EA<{GqU
zE%*Q(a$pYwkrx4h{un>z;=XUYq(1nDQg=W7^tF0}Np3kZoNkKU%2;q7;xiR`f<FFn
zw|q<tIW#yze;?v0oSCa0Y{v?1CGS$vKHb{#T1BkP(AKMA$Whw~bK{-4VZE{(E|6OK
zmhZjjUlraZMNylZ)rUL3wj&bo^q^R|r|J6~S~v`M4`4rfGLI8Y&Fp44zTVimVgNgw
zs}3$r()Kzy`T!>-kBQYMvOtdr(vyfHT5xV6dSzw`D4vyYxKWahW!zNXAAjuy;><jJ
zNLNa>OA$zJXTtY~PBL128O)5C88@S^*Tg>RHS~S^`<(pNb=_a~={)oY*KT}zH+s(P
z&vc$I*JL(*@oo}4Vku?}CV{+V+mxl9i>8vZsX-`QVlNqHN3bR|+q>8iP(VWHLb10<
zGTCEfYWtR|2=ANhPGKthHD!Nx+#~b))^lH$hXbx|X2?0RQ>y3bJ<i>rPMbVT=wBOF
z$g^++8<7DR_q%b?CEYKhrm`ZlZs8P8I&FV{tv27yJbOXi)^0gBbAN{K@gO;7emRt^
zCm>~$0$roYnjblQX;1?DH=j0u?9L|CXw-qZ_&1`c`f~}-(ebGxoS^a?0yE*en*Ngu
z;VF%jt+LU$9F8Jb`{*ydRuEFYpFZ0>w4*U?eKYXr{;LKT?YU7*kw`;kvzCy$RRch0
zD6cB4uF}cw@~zCn`IcFIjUjn^MD-TI@8RvUaj9$%(+dJa?C)A1XRdCFcK@MS)I-iA
z;8eOE+_|!%AcQf$T3Oj03|))DWX?=F=Kt1047o1my6*yRZYTz+ef!m3A_?MXy_yeG
zq}Wnp0}4fVR4fB>o~ID&ML)!^%sQ4}APBNb)Dl%k8Jw9L4Uaq=l8MzydfH@<YQ2f<
zlB@O9LWt{I1P_%xYiGMyF{<ycc75pXZ9ik(XR7rVRu!*h&%{2P8-d+uR=t#vDkDGe
z4mLp*kmus9K4vdbnRUvy#INKZ7j_8m<WJobIBRrJ&@}gMl(1t<2vrZdtNMJ+zxrxL
z8}Q@1R#(jJAHLIpLHu~Y<Vo=}JziqI?zQu$YL(?S&>Z5Y@TCf;FEgFeb{j0$DB+EP
zNn0l!0dfAdNg3P$NCa4TBLqv~Xwe)c`hBXH<I907uRj>s5&NCCX|Q*fHtJM8^fC8i
zs1c#usub_BoAapgL5#0gv9!il1oo?%<K`r__@hdJBz5%GRf$W|cBj{Mb_+UQNDLE?
z@aQAhgbHFztspm$CaEyH52Li^QmCw#nwbmXXA`63bE2rxhz>Vz8O`BnS(t!V8)l2v
zNo?WUz@N^H#OPzULv~K6vdO+d7QFd7^{1d}Mon%sPlrblm8W{A_#RDFj_UGmZ2nFo
zPa7Zhegmv5{WqAcJtpyu-z!qDyG@gIV?nIZ@6|a20@g_>Sba6LnJb*FUJX5FRmwB5
zaOAp)gU=<1!a0H@R)T#&P=LP%J_VNGg~@3xC3cS^MKCv^=yqtzT4TLTXG|X$8xu`<
zLs&EnlzX4MmVAHag7>2&s%V$5jK`-VUYvn$LnhUX5Lmjby;Z$*XrMlAcoR1)=@eO9
zYA`Z+tgjGymwwJG65Zu~N5fOwVm(|%bb#E&Z|r~NS+3?v>`+vx3cs@!y39#jIH8SF
z8*d1vCRr%Dfm)7Au|}7B3EJA`2I3xf1E8F$xin#@J^rG)9qFm)U9u+QLs32CoUG<w
zsr~i6VF$|<Ax7n+ZP-5H_#Sy+)n=!wa`vt}TqurZvXB`P^$4s+JqeD{iDDiS5AiO3
zDqWY&4Vx~)WHjsa5Cybq2W)RwpTUoB(Fo5)4*p4N<zKaJRr8b?uobXh=8&`Cp5Ubu
z?4eZ*W`}TB=TwFuGUqnPd^Y+ZcpEEKA!1i8>C_}{nO<hLrf<F`rY`i6uN%=Bb680O
zcxz7O4(9Fs`tS&dRDQb=quMmfZy06MiS#-Rxn$3i?J3M(0ehOX_WR%x^bgIulPjOw
z*izd8q7h{sTsFm^OCRZ<D|9wIwspKCf9sBM{p9MbPQ80u7d7Q@wC^t!e4il=y)U6|
zZ$$SS&Vi<F(zSLK15Ast9nF?qLix}B&efwo{d{7DR#uAdo;*jaxHo<A9j*~}e-0It
z;<%jZ*Hv0q^PD3`cjn8<@=o2T;{YH+JlKBRfO*%yriQsOb#3f;x|zB4x8<jt(d6n`
zJ+Y5fS#mf~VL<F8rC5pmT`kU#STd#=G>r*iQ<|l;7Pyf!LLM8?!whnW8)-cm&0}Ti
zb6y9UL9C`KLwzptLWXbZ+!rau0$pI9URi;&hBX4f`%MaMX!p(Q<`woIk0R==`h$52
zv>JkWh-#O!$WU8#pxVj(cZYdM9xI9CkiQz25oM$6YTUnSJ@0JVbP_GpfY9U18>q?w
z$Z3tlx!eFXWr9H^-oyMd@g7Drc1T&{laf#?>NV&zFdE&Bb+`f@$~pPc3DL1oKDz0*
zJ%v#B*mxM>mMqz8gs7UeInLh2M-JwX-Bkjv<rsm91Y>l5?%3B%s@|o_VCIV>lu>TH
zLD{-02C*4<(UVlA?(+uPglD+N_~LG0bEw7Ho<aQjOw5{@!<#y6#*s##gi7nL+GCfX
zpQMyJ#dMoaALAoSqRq<TN<j0<M@?c<M4?w0aEVg&6;^S4<=~I@fr<Blbz@KKaa4^J
zhKdilD9hQ>^Ao0gG!-(+-CPKRL&6C2!j$=pqS@Xv9|r$?IhUIpeLs3VL^D)K6LRGF
zB&_X}JzD!@Ze(H?$AZ?p_U`~Map=j7KSXWZkFG;cc%Bu733oJaD{coT9vxiu;CyjP
zGmsD&Q9f|~s*AX&+o+ZclozTraI(OyZ1f=C`#fEEBdxcW>Ja9NJ%2NcV+?zFVz<p8
zJ<-zb_|aTig`bR|jOeX!rAEZ<vvUJW-DCpK{3XpNLvLCgwO3ztU+&jcEa4Sa<{|05
z7_!k&aXTF~(eO>3D6k^R=f&6iTLB4)^uu;eF$9qaXQrP<%Wda`qYbI}Ri_o7<>omZ
z_36c%A-qA%;4@oq#};nPSKm!}7C-(izPI*sxH!}f%f0euVG3B{nF(-JsyW-wJVjpj
zD8clNHW1E}UJ3o}nr>>HF6S|-C7T9WZfX=sM&@Gs@CAJ~)hQUTPs6ta!t~45NvHI(
zJgqGgjJK*E2SUL)PJc4qVyD>5+2Y$MhV>~MjSrctnL{@lTTNdCeH%S`KY1K#0=!^y
z>ugEK>_^90=_&~=f8Wi=s7$qm>bR9b`)aSX6*+eM4Pp%0{t@bYm<)7rj_<Swmr4Gf
zGi(w6PCQcRJIzt83aoZco7?_XZ19)*Hs0^W-IuGl>jrg4MyfoM9EF)|;+u=7c1*Y7
z^UWSGPA+`R+l)uVX;MXJU{JLDyO0Ms2tLQ?I5%o((zadTgt!sTIP{G2yNWzdOigz7
zqs>3;Z`w9szkle-;7J9Z4nMKJxO-gn2lN`_F7*C^<abd-hi%2p#55IUTb<RfTN>WN
zadmTsQ3Sr~=MB_nFW5CJHwyNfb)u|Px(xh>BWM^0Ps6imVOXc)G%aK2Ay7{|Z(wW1
zdX)XucuESnqIaPuZ2u>D2wzqcBtTUqIrCBaZX&dMgJnIi@tdboTdzqH8h`@r1rO5$
zN3}MiW&529_0uNA)y~E>vWZ$c+uyPF6UHo@OwBIC(f;lcZgJy5+zgemWzX<?H;S<0
zq+7LiF8js_1N3N<Wt$gna|ha6uBbS-nKoE1X0MbPiPef`I!X3x4_`s-pHL4xWy`11
z?P?2iRaSn)jx%h%FfVdPK$~xA4rQ+a7deX~HV^~!YU9?#5Yb|-HZS_>iO4biaD+cu
z6SU{#{v|j()TTQ=TBp+`Zve;xVh2Ud&EMV*Cs}yoK%UF|YYKbx-?qKMABiG-E5ZJ2
zMK=^|L}ucai<R6Ew?dyw;~x`+uy23D`nt}<Sn>|zj()=cOwDb_z+;C4$J99^^>n}G
z@Rgc?Ia2NXi+1iUeV8H?*T!uagXjk3B@J#QzbFhAk=Pf!{&3oI()|ilD4+Y4GoYlO
z=WDPX^7{m{-q}(4oflUFSjFOM9t79mo}++_6zbaaK6UjhCHPH`BUZW2Tk-Z}-QJme
zYQI;S$Q7I(R8`Oqh(Bj2%PG<|p^H5uOZpol^UH~V*^VoEgT^pU6v23Dik@6HUB||r
z_pgKAX6v&pv<`b{TgURc+(v)>Xx`DZd4Jdq4v1}2RO*i)`LvaBeWs;2S=V-wJN#8o
zAKj(+!2~m7o?6NJnNg(N3TN*vkR=@eS;el@2(+>N?6%$mG@jaFb|;Z$xUyqA-~?`X
zC(M}$ZcO(6>UoQ2lO1$|@UgdZdN^?!kO*nWN6*ZW6cK(i>E5*;pCeN;_wCxzMYQVb
z42{NY4%l&9q`FW?ts_U{sXFnl{?|RCT0$x#RuP2f6<DR?k=b6R3`G-enBqj*k-F(T
ze0D%<^-Ad-E5YsW?d)_86g2A(g-+cv@e;Ynq^gCf=VlD)1LLLJ2jZ%nv`f&4U?8+>
zhHpk$h*PIaep`4ua$je7@=b1}fUumQy+igh_4l^IywMdM444UBN)Y?`XQG4GffkBG
zHZr)AtFK`Jy6W6|v=Y54xc!JJeZZE*mEiaMKqB;<ZJJPwUmnu@g(|OMq*0{GQ?3Ys
z?+qO>jzNtlP>zAzhL>6;kF+%}YqPHlz));n)S2~PGw!cKd9_b&-Pq#7xx(TGylVOO
zd&7?6w?({1_iP(QHdya<hBb+X{;@rN!or#WQn)w${`2NeY4BF?Q2Eq6a~7kMB7c9&
z!v9dJ8$5<IByOX8lr#m}cyq`)eLwWlKCVM%-m^tLcc@!Y9x@N~Qs0$%fY5=z>**x6
z_O2?-3(YrCP8^&}W+91XJ@ZN13d8G}nEX1Xt2FaiBdeO(kivKgD4Jqyn@kCBxi6k7
z*e#I2ig22eYeN)&qnzqKCm{>H+VRWHw(lDrVp#-d>#byq^$bhbVxBtNb&9|wSPdc<
z5QqK)pxgv+J@fiOHUwo?Y-62mzdr6L0!dRs43Ut}s&$=)t^u~XZhS54G#+R|=X>9T
zdX*0b;!mhVkX=dH_H08DTc=5A&4zssvdOjy^g0y1?K3PINpn20eIEuyJ~jUd>sk2h
z<lT;WJ#<9F10lzdy=pkvL`+L|z4aZzfzWWnZGDsc#!y-1eeWASLnRgu57U*hExoOd
zyzDPesDD`9UE(hft(d-|9J72+Yq@p4mD-RG_9EM_Fzh`$`7b|O-OnnPB!%X+6$N<|
zjfML=HFRl>0-xm2kVLsgQW}(#Sf_5kORTaCZ$Npi70%V5$Ldd1B-_P)B5^d3U>$y)
zPp8xseWjq4Oo6HW;Q8?F>)xOc?W=5Qo$Plq_OD$Dii5|jZJR<b%?fFW9}En*aPE_9
zjXiI!d?>8QkH(s;+0D0wYC`GEEAI+OE+y^KvpfkH7qGUt?FE)e`3XXCYV4=w(TO*-
zbnJ4J;cQ%;tQ~bi8>@&K^lkAx1&kp?V2}|5e$#r!D*$^ZeUy!BdgtOHR=u@L9Sv2{
zMC#*WzjM_1+S+)rdt?9V&7)C#?(naWKVc$fU|QQ#>%DC{`6oNny8Gu}d=|3jow=KM
zx_}YX`kF_D`=tS-qEA-WHe_1(ml>^dNQ41GJc0cVi$S?WkxwJ6X1E;BADp!1?3u+I
zU^Is~oIWVM<*<R=AwHefEqNneGqozQ?Vhn2_|SOr;KX&qz8^DLtKf<9J#ci4rAXt)
zgHCU?tP>RWxSspzd<v*PC&VfGFT!PJ`3yaAe1@V~^zoJ7M%a@ZM&MkhPOhEJs&E1R
z0eQ$95Vv?&;I^dd*{u^`Iq!DrDHJ=MOA+$fFveg(zUtU(vJ4_@>h^s5I{9oPU#!{>
zzRraC=)S<Ah?y0q&ag!#V_8iqKio12>7BIomH1j>-{!jgPiMRB$X&ptpE%u!;z;M7
zfK+IEzv2G6`VWo53Fr0DCFds7DanfRf^nIJZUY<VCK2_Kwb$-NP4Gq%ihumzj_TF*
z<Ig>JRcKWzjd)ouQJtBQd4@)6ppx|1!url=zp{rs(?GS>vXe8z@|KTin)Jhluva|0
z$`Ks%r!JKn6hsofdSMUW8bpZ-!MzrbPwg|)=hX_*g0;vibgH)jBGLf&Vr^8%(&Xkp
z`8Q~E@hKXN;>^MA`2)IIGaEbYwq*`(9L%#Wd27#Y+i~PDFNNGq`K~`x&duPiKQ?KP
z?!F9jE5L8h6y^B|4(`<VDoHzK9Z#n*LQk!%cBm5FS;{~*ym^tBiAX=aC9w@bx!}#=
z<A&@N-~2kI;oW=WS(A2~snYOli^aJI*a~u7Lz;eiU~~U6?wzX+>|!adfFfV-JG3>s
z`{+P9>%__W<I8=j2%um%ha3l{{sM{@rPFfqv^>FJIlPN<(eHTZ;{e=R_v)Yi;V%Z7
zDBc4w^@^~{P<}U@r8_9f^mUIqbI01}kJ~4393p%-+$do_EpzBkRVP&~aRR7zAK}Ee
z!C;zK6w!t!Ffyp)llPfMwC#!d3R^k<zvMuLcAy?we6vv#sL?1)r0vYZYWrrYJDSqM
z^TS<ZZ(YAZvte#=PCj$bcgTArc>M0R)`4)`*WhF7?C-GQ*WN0E8tH%eefDrX0li`S
zM}gZ|#t-<@g97)*TYR91;x~kNnT{nR;nKs!f(o8ZQcG!(MGoIO%%k_NU>K$-hagCe
zO6zJe_B{PDyMsLBwXM4vkZ=>0+9Xn6tbIju3x&POCw0J9C=@x?K{TzVy;t4;uwXKE
z2Kg=|v*tim+FE9}F>B|eG5SIl!2jhPEt%e?&8)LAU1*N6jz65i%!>KI?WvvcLE`wm
zKrZ&7Q;7<>rB}J+*bsw!UGMMAtk^E~4C)Q7`8n^Clnkb)wSrbBoe2`4Q@st^wx<wm
z6|Laea0SPJ7w1E!Ltll^HOGxB@kXzMy}SX$SAHK~ti*^mR7+OA(a`tw7u@br9_|6>
z(MQ&%3k0OGfssC6Kk^1%bOT0!HGWev_jt}uCVe$Q|DyFSQ0Cv$W#-@4C>weAc6J6_
z-!XTj-2y|pu$fR1sC*bu^}H2ZU}RKBlE-DV%X=o*jB-+Rfde)0N&RdE<f2avQtnho
zU+sTs0nDDB^YZEoX#q2hOf2KwYb%HyT)&JKUb|w&47%8&?Qfn||D}<T*j00m24Y`v
zy*Tk>bAfTB&$2wW+NB?4L^-v$O60)dX<avfCxjuhv}WacE?viXKc(u#6rZ>T*yi1-
z#D{}UeY%uKNW&}k-EW3CAfdB5xYJ;=s>W%Sevf#E16WOJgB!nY>~nuK5ImY)l}6T&
z`I4xwo*{_t)`%p?AUMI;chMx3_Z~<EtDNj+|5d|JUph*p?_BZAq>gx?FQ{}e?4E3M
zjA~KaE@(D{UpcD(=nhwPA~hfcTC=df4r|G(BYG&)si7cVY^kkAg=tgZj6<sabl_?T
z*6sK<Rm#aezYb6Gl~kD;R(skqWc^hQu^Z8MabBr4S#nu7Ha?$?FUTs5rkf72)tx;&
zc<wF4{UhOP6?qNNA-AAVbzy*y6y56hXKmI#<Y2CR3eL2ByqCUx+dF)6)Bf*rE2ORf
z&hUVGOFP*?ODcuNMu@f7k5xmHal>IWKW_p5+S;E7XKr)0+$%oR`c?2!&cj;k({`1t
z$97xz*MUvb$+h75T=Q=DEr#xaQq~FkTrC}z3mP~%if&&W{bfIO^II#Jmz5jliYG-_
zqVd0ip0#tJolSej+mlM0um&wx62O<%1FN#_ky7b6VGoY3x!RXZ8sXwaJWuAs*7jpa
z=|`K4`=&U!7ia$)bg?@&=50ZzpH=n847t?(cUQI^^$36K>mIQjB)k(?nEf`Wz8;MB
z%RwvuPP~(f)RZRMR)igl%=^ISg81I?bz$dqhWj;7<4$?pOCMvEDGi>jo}qW3o5)a+
z$sNOBLBKF0k>4xR-eQGALDFC1DrZBN>y)|Tx)0fE+kSxT2V6Pq_4ujZLEiJYN1pLR
zmrt!8r$-e5d2;xe6W=*c*S)_J)R^?VWMRMRB%Ra##1^dgKDJWKF<`1L`?-vq&a58+
z<&cTzAiB99XsTl0@&=xBgLPb}3XGxOG`P!-+W!7TCn9%bBki>7(dda?9ckEL(Lxeu
zYhz$7>0>pxVL*Gd0zc?=eJSV;Pjhyh`kG~0J#%h=^Zk+6nZg7WX?yRwT<t6VbPwE?
zh#w<!suZ#kl5_)XlM(IGVbhruszP2I2Dcx$n64ekrf@WuLM;8T8o_qlC}>H%qvM@^
zKXZy+SU0Q2i90NO#yXPbZ1uT=d(eck{sFXa`uw>~?08qRyYmbD=PK3L`!bVtA}BtK
zH{=E1)z^mev(y|O+$+^zS2+D5nI2PYZ9MM2rv%9_{65v*X~c)K-TGFVkGx%e+OQ*Q
zN{Dvuqk^Ll^bcEEbveUOqp34>-J(42_Nc^hU!WNTVqNDvA5$d{)r6lt?UX|tigxop
zy>TWdO$k$fBD9Bn<7#7w#l{G5>2caO{r68I?mnM~X-&${HVv-T*$n}*E&TZ3L*71i
zRBL2Es+Q-w;zjEgfdRv^SDg)=|KJy9p3O243i^YfIP_8Z9rl_sO>kSgM46!M7fE=G
z25LW(Aq#$aC{G&6<ty&g_7>9=*^dNvKOY`MURyFSSh3{HrcMN2<U&LhI)Jt#JXNsK
zahsU~pg`$youkvqJF}HbYuZzM?&OU8<FlwIl$w~Inxg(vkP~HH+?V3i;8)A7T;hvH
z&F!RsY5omK%E<wm*5r0L;O07`?i}ihCP&?-%1tQ1Dcz77ZXXl$4ssuR@unl@)(Db0
zAv;Xi9Z801ZJgOM%|^!1mZ|rp(ZQm28Fcs@>BBMbx-<Sj+K2QwG2(+`-5_9FkaqvQ
z|C!drr0P4=7EG6vw-O6*vV<B@gBlAW=I`3Hytk^vXlslgt~yDtpZA%|N%LTVx<qB3
zxo0o7ei4!Svra_GZIeqLF+_C(@h$Pa*^!Fe29;(1*(O8_np*Q-8$ihz|J66!LEil}
z-!$tndoRTJ)zm(_HI%2Ek7Iv21MR{4?y!@{kiJKzD`qIm#ltUa^0{goU2zq?Q}!qP
zN641(q{h^?hMDo0$o#Vx+~M5IFRmSZR>M=Lf=_62$wnhv6Hf*BX&;eqjd}cxueu)k
z$Iv^Z8VRPKbdVJII`tov_cFU?mDUJA`R7X3RU4yJ#Jbnm*I?7(!K_Qk&C=#?o9lx?
zrVK6fX~PkNOkd=tZZW{m*Z8Pr{*Mb;p4LchM+3w<%x_HD=*?tPL-vT{G}mWmwCAp-
zT63sJ?X1l3U3WDrq?BT-9%;Pc+iE%2wu`Q+K-NI&K6ke2rLB`aN7F;=<CT`{!rF*m
z4FhbqzKa&W5qdyb>$zdhSpH6pOmQqx9`}1tOc~T&d?s-WhZR3CpW*vfQ)oC{Oq(V&
zXWQ8Je4D>nSVvTab7Cyqa09JZ-G#S8h{&0pCEY5ZVe<!`pfKFml)I_he~&^1!ar4Q
zsPOSqZ*9&>4Tey5h8D}wF&Q|IQP`Ek)VrwK_q2J!?M61#^c{R=e9%4Y>%ScsvYT=4
zho#@a{l30j^s`<bsRk{VE#Az$E|v`d^5?dmz1_D?_}prDDXAXv-Ma5C>ufx(_RanP
z+?fhT#E>hI+TunE`8B=+TQ3SzR=sBcTT?MG-tZey#+eZ-Ue4v51sx6-pu!hsd&Bv(
zxjzhQvu7nG`&HzV2j(2kUvjfaKOI&=3h}?4ObJhDJT5_iw}fc-n_^m$4+QeTLu38N
z8Gdy)9M%G_r3OXHBU&<{>6rpy176~<a}D22SgUf7PptbGujKNbgL91|MY)`9AteeS
z--&*VVpy_=i`~OW;*tCSrY{Tko037cpG@}21eTh^HBz&4Q>(9<E%OnsEIoXV-Mhbj
z+C1~g90L$hV@Q4fviQ8@+Liz=anM|zGcCfeL*6U5_+hou`{d^Sv`1ZDYZui{mMaLL
z@N%yKVXmQCpP4o#;L}`1)MmbtexvVDwOb+tyLqjd6=OjA?S$sm4@y%U1e>UOt`dj2
z)(|tzB1u1JqI`A@PVWnP(6fEMX^nr|wt;_KIqUGtCb!FNg%SvWuPQ&4M+YxoB+rQG
zin1nd)lYdOG+(&T-@LAz02$@0;O|!$GHJ{8uKvb&6(3$R@~t8UlgSr}-hVR6xbIZi
zPx#U5875EkOAJ`|^DAx^|LEG}D??VZDh_^BN(?~6(VE1nIjJSP%Gd1C=TVlk+^F!u
z=+PgfzvOtq#JU_7F2hYYviotD?_+z!(rrR6CV9X@Ri)oD${7+xxCgeWGTx8qUgutf
zM1Cp6q=#VDbhMDa+7q|k&V693)vh>qxhR#*Pkz?-;e5Q69IVQ6<&WT;I8o8OxFjub
z=Rczxn2Nso#hDcz#7g|mJ&PZf>&aDd?KZ~|St)ZzuVFdgv(J&{*;X#0d{7rFqd}dP
zogW-LH!gV7{N<?_U#AszEZ0iV!u3l=D|eTA;Gjl8LX`dfHTh$ojwe+T$#?lhkb`+|
z;~jxet@1NAkt^|okd9yTAcyOjd~Z;kGXkkmA-+9}xBXyQJRDuk^EXjb6ORUha<6;~
z-tt~#K89Y?1k}O1|3Av!GpxyM`vaw;ltcwZq{M;(3W6Xll%S}HNLPBVA}t7^w}^nK
zfP&P}Q9$V}p@sko(mMeHB!JXVLht0pnRCwEng5w{=YC^9B-!uUYp>tx>%B_ig`SaT
z-w3r)+NG&wigUV{00Tgxdz}2(1|)gizZ}HuE@_#A^R_~F?vWUI=M<}uY$@)N6dO-Q
z>r1j8LlW+zcSK{HOsC@Is3#O%bD((iIDkFDglV_O?d-0AG^L2I(^?_wR;{bUgzSEC
zyG#sjCQ8)MgO>FmeVu)n-EX(*ytHRr4Z#v>y7?r5x%>tJMoMQEcB@~jo==9-51*W%
z3xpFfRh7UgKj|EAbv*0w;$056z>lW0O3hMDPYdqW+-~1FlRD5wM0zQ3cCqqosFMV}
z=om^>1xf5)p20Km!KPMak8iEIpy{|U5$2FtcutcxB~fr;g32W<mJln-?dD7tf*P5e
z=t9}OB>FQ5u_hg9#~_bBzM8{GO)3_%9(>&x@GT5ely~%$p7qjTsDP$^&PG-NJ}OMy
z9XopQb)=q?6b@xc20T}g?QjblTG$)Rk|~=jYKcAQQnhdx=9arTUO!V;EbGVh;5quC
zPLo&^8Oi$8EJ`WLZn+!UlCQrC+1c#kjUDqw=#Hn=t!!E;Oz-(Z==q_E`(C<Ss}jHv
zaSM+m?3J{XC5hMS+{z=%7@3e+eQ@U$TN}Vsw+o{H5nLMyqB=5ecXAe*0~ESQ&nmnS
z@+#O$92zHGnYRYUREm3bdWj`7r@GXp7P4s2cFuTqzB$^7+F>+mbRVubSvTuI#9S`3
zj8YJ<tq$jZO`~Q0>TG4o_AyAhiu0Y!E!1y0#B3EJOf0}zr@4B<kDWL)JHuYQbo|*o
zp7c~rraewYF+MZus>_o>?>4J1DOVu&ZK;|z8ZLd#U;7s49g_SGo?J791@P?HVqYjr
zg%`QRk(H#)Vul1sw%>h99iN)}#0*iqp(8(?*?>mWUH<V$_1l-~^&z+_ojPqi$y;;3
zPBG=ft`$?hzWpcN73NTd*V1Q9PfR$zvy*+_H8PjgY``C%4nZ_LCv<i`k%-#cGH1)}
z1}XJU-@aZ|Fh2YJO_#cWY>L^Vs*F&|h|&hbo+72NOxn?hi(-T}ZSxLxQfP;NIYEgs
zig#9YqWiOZi3`_qm0Dsv&uzsh<i+}Dbv}7ZPHGd81@0VS{m0OjiSEN}oZJw){xtVS
z!Od@(7y&^-?TkLu>|jcu;7TVTKgcbl1Rb`PN;)7Q>rWTLUj%A)p+XggkEuFXX%n3c
zw*k4h+-d&zdU#aXF2qAt6}ER~2R>N=Wm*DG;JfIZ%7C7ffoRz*%?fuWOVLgT`SXWl
zcWBD}dNE+CdeZV%b=WoEp(Y~I)-(Fp#_O^iTWt$cxX9ZOf<2$6@^P(PVuzZjMKXVy
z{ju&<`FOh1*5TM*>2v_yBY7McR4e6zOXB<ZVoN-YV)>$5zGEkczM_}q=)&&$8uQ6-
zyJ#losp%R$OJ>RMcsd63Nx34YJ07RKM044>z3$Q7upa*ou`tgcT#nH#i)hWRTif0R
z*&il5IGik{trG9}w+!0{@<gJ+Esj`_W>k4>fVVNy>Bbr3q5%&%=4|h=bH4Pa`wttm
zmg<?jc<N{-m=$O_opb_G!9<=PwbvFB%4*%g_?W|lJ>$>iE_?P-GwzY|$-+*)X-{yx
zLLX~A=8b1O2{7^zq2hCV1y&ps((-(zn3Y&cHWjZiC?a}(DO326*@Svz83h$wmjkTA
zwYz)vz8@Ehx|d<2T_pyvenis~-C_FduW7P!6%#ADC4a!YE(<aG%I{5-sstwVAbQld
z&z(^n_bDAO=!{8P51b#+kM(#s=!}BteYd+W^>uM+qANcx#d>1=<B?;mLY+#@7fPDA
z``J3rYCG|ruHWf6)dP^rdg^_B_fhmyw5eRlb(CZ5Els?T;mvABy{*w+u@4bV6$Ksh
zB1MhY+Y|__i&2v>(dfvlNn8?cmU_JR9&=Sxe41R##MrxkJKDmm^<`{cVJc@I{xI`G
zUE)xqVyQ{9D8xuHC#-=_<2>cEb^^8T9@QEUZPG6HznDlsN<4Mh7`x`grTz<>p|d5C
zS=grhZRs*@`p~SP9k3m|r8l+7UI50QOykx!Eds|KYY^l&Z3*&7V5eSC#$iIw#%3Fd
zW|-5IzB_1iq|q{88{?Z&ai?>j&WSmu_}07fET=o80}DST+KHw)uo$^^_35I6NAeIX
zCqbd?(&oX~t4yU*ejSGk!}#zQKvg~d3i+_!DvgIlVJl`1ZM$;0!!ED&UGUL8mufR8
zRzp~=y2?to!t{IRmbTmLkJ9ie##~)~z5A5REXyJ;O&x%pocDF?Upf7%ZfVe3JuGMY
zItBS9lzHAY4Z^@422~Z2KCZ0imZkx)DZWS%vGA6-j0hmTZU$GsxdgOpsK`5Vt9H}T
zOS?W+`zF&p_@p3IcCbz}@?%qlM-l~okTNQj?8Z5p)6j9`^?Hdl^;xy%BhAs}o#qCv
zkJ@wKsc~nT<N<@Ouri`z$-=$I^<VhP8)fv-R6^6NuBe!)IfR(P$IYK#40bLT7sfPr
zr#l^kA1Ei=Sc?Mma3V1L{n{A$_<JpmxC32!q<!C#i<B&ag}B`RIjH{Y^&``C)x<BU
zOMF{9#f(BS)Zce6`g?Gh$x$Dq-UOP$)q(B|_(#w0!FzSmkDoeL9e*q~CK`J<yPw!S
z3$L=s9C_C5WEy($z7!}DR%-K5sQ#Toxe|N|>aP@5J6R4HD*@W^5;c9~I~GL;`0m3k
z1dGpC6unX=oNan87|WkYvzpa!TI0;XlT{+nqXO^fmi;1D(SMkMF~d}ieKGl@|7P=m
zLS5c<54qDs6Wue08Ic$<Ww}~PYPz>-yH2f6z57(boLjlJtsYx31qBKq$t4}>=;^WV
zp@x#&{$`JikVH4+Q={d$@9RQcFEnn1`@gE2%{9`(Z+}8&YWdy_<P8zRsf)no__W)&
z_0F(8fD4?W&s1bjxX!ez*~>ihP{w#Jt3VbZ2|Kg@$iL)%crukbZ`*gj&(SKCwy|ix
zSm(_hwF$^c6k|%r_o1s|w2u&>O0JcL-6{=lG&{l@5tElh>*Wm+CO+Zs4R@w$gu-Ux
zi-p+TUhHv_?u73#Qf$65&BfGvwU2Yp@vQ{M)(`*!HjBs{sRc@I?A}N<M`KL7B<L}o
z=JRs!*AB5(HzSWSY2z9}@v8p1*7fX;*1h_Mqa0iWn?Z@nCq>@*$C(i_Dq-uFo1Yfi
z*EN!#%p~{wahbAgVD1%<F3d|STRCss7_D%cvyoMD>bAba|9($_+gGf7`NEkFvY51|
zBTf_d%i9o-g0a-`+E%nA%m|2{VcJ#PmaGIG&iFV!7c8c2I^x*<QYD--)FVEsaQ4Sx
ztwzQ*i;}YQF<1tJJPEW^$&%b-0G$8sE&PD&xsEbtZ-)DuCqb{YyZ6jYR%e6`9NX;f
z#mR8*SEOn7fTbE43Ld!?AD$K<`ah3_4D4m9{W8{*j+eBzK2tjE%y81Q+{DMEt09()
zB%w;)4ze&yvrnfMD)SQV<EUS@l_9RA+B?R1LN7hFU8HM0=c^ws^8UZMsLk}Qyy@wh
z0+Xp1Gsmg=6SbV~>sw+1-ZxH*Ag_Ihp75w#tP}>W3Qf~$^6SmSdxJgDOrMWyza5n9
zH;nZ@5!foaa_m5wt@xCp;U|;p1ZJMaKjHUy9T;@`;Mqvnq$QMQDfma<aqBmZP^1AC
z{CN!jIv(j>>1;eBtXezlewi}Q47GFHMra4Pdv#a@YcUF%tWg_V&gs{C+q9*{v%a&&
zR9RoW%AOq5k{wFpoh~kc+<5q|4<x=9)a)=3O{j3NAsN<7<^#Lob4|<^MqX!^PPo85
zgRn)wCjQAX=tXXpuqIc+)}p&KTUUrGfk9v{yCKxn(<w0E{Ho+RH^kG^lXnHc;52uD
zt5fW@=ahmU(He5GWI)(LK8*7@1)yvEkDleZ9CcY095iRI{AN&`7bH0jxq;@rOsw7w
zB@Vs=(bwwwXw1Xf+%$!l2AGnYUt$gBR7$>l#(1-!I!cD@S^WWVQTglsLvHFyM=TxQ
zo=#oku&;3*@;kf6E3?;@=U$mTo9b#ks2CKOANx}7OYgcfZWe?^Mjzjws$hkopK<*T
zwK*n^xxo}0715WmwA-HGa;dPn0<c&|^Dl;Hl%lIUrxBaTra`fZVEMkEkcYd<q}04m
zBNGY~|48d11zTMUF=NxUfs~FIrX>vmy(c6ABuOVi2{>=l@T&()fCx@>U>9WQXA9;Y
z+r|}=h8I7};P$6Xpf!XpPwJH1WD%4mVI^WobBXGD3cVQvh_8roG>~vX`UD_U-F6&;
z@oiwduJFE=Hh`Io{dUr}!e-yR<G3<n0;r0KOP6s)YAAFBfE|HxzP#Tu6-DUPulBoD
zusQLM)x(DeGznK&Ha4Ks1Ls#>zk9)IxaLDp^NxtXwx>+FX=j@UgkvpI6gX^{;sjlr
zwoC1KTSB%=7ht1uY6eZt`YpBI4b@prz_OQ&7{-TJZ%-6?`|HNz2&Z<hNrMueNgBE#
zI8T(^T4F2;J8PlZvjWFscu5@d)?HuDy^Ju|f|?aLRW9R(G+PudYz+H2hAYEA=wc|G
z1srCyc4PLc=L&y4pkrv2;30b?LwC!>m1YVIzNQ&xC+&@?6TGd=^l2CE<iXv3$3OQz
zOPZaB9`B@FGkynUuH=dgMl~v16mQ#IcLacfu`j#RKMjArho<goaiAzt*Yajda_HA0
zSP^wv#$si_Ss|RFDDnQgD<^k?5#}CA2N4G?@Zy{K!0(0EERRdBUb3KUOJU|aGx|}W
zRkMpmzxde%X5Ek(%*p7O;t!><_CFA<?F~`_q0bfOWsAPlp2&zPl(BHqXDXx#)RK(O
zzUlCiSi|DZ4s4D~C#5_M5q?c(I;L3^WB0urn`|w@3P#~#P$4QtjG=vIU%&rGrB-?{
zjRN0Js88nb7568~wh{I~i3HQBl)D>b`G>rp{3gww*6NJX{xoXt38+}M=kqdIwN>C2
zbnA3`9muLq*CO`8(WApD8Ieyf)x@IN{-K<r`v$znpLJO&Ya`0f(!Zf`xw`H|y0bSc
zI_%5x_1m>#n)be(enKKi&5Pf?`dl<O;x3b1aLG-8RO6P;voX#^0=xeEDTP%*5xU*!
zDtHH|W6zjqT1oTZ`HG#HE<?9Ha(Sb2dqXf9JgPbOHcKL|uQg-dXi*{YHJHq58?+-c
z-FbMY0PrI4@t2LmqJK-r$xKD>6OwnT5zgvV^kDgq+#2Hnpp-<QIs<=6%96ktJY9XL
zu<Y4~hX3M#D!n_^zWD$ndrEPCN}vCE%W4ZX)NN3uVOOze&x_?J63%tu&NsxU0+R<M
zD|!TT%C^t<^!nU-{Df=O4H5p;VV{To+Dn?W^q7N25p%ADXOQz5?IjihgedEyj2`Am
zWfHg&Wci!i{;qX@<NLi|;Qhda8uTbGY-A!mQtIFqPibHdW{d~CbEePF=!Y22^M{7^
z?~h3NFe#Vjc(KlSgB4bO;h1v*7;w%>Y}hLe_n09<JXAn5^(?9EZ7L4Uyw|YqHKCo=
zkMe%DElW^<-rOC`{$+>%-~F}p%xfVZ^Lf5<WHzUgEQe2`>Cv#CotrKf&;7Cue*22M
z9!BJWS0{T)0fZh;wXr<ZYCY}mES^~2rJsrMdV+Bq1va>iQ%{!!FZ3q}knD%16rjTS
zx2`#KJo@)f{6^^S3-d=eL{IZRLVpy#&DN$g1>R(fGNYl)dJv*O+V4j*Uw}!m4m^Er
zqiHT&6tz$n8_h+ZiGiLaF<v{q*nbeZ13;Awq!lwQJ?cJ!7m^QFW!OB}7rg#TUsk4~
zLip|LrRob0j^2^VF3|f4+%hFAr_?`5V<0w^64c7-b<ym$ib70}Ve}sNf(8l91Uuv$
z<1#hLB1=-FgH^$+^l=IQnMCG8h+Wf>ncdr-=0!E%c(&@}4~}WGw*I`u{kt8+7P++D
zi@G)BUUTE4r%gMU`vfxR?S1%%Cm5@i=@(+Sdx^NmvreK~&D#v%{pXb|H|;A^i;bni
zn<;`3aBchmQGs#t=l+(RN)ymJ=}S0V-GVYv0QIlb{D*azY4+jw>?yg*5;6{F_KWp@
zRf0w`_5)xUrIc~70p_ui!?}@TE*MH(M4fW~4;Msv9b#sZVZ0gZQiISc-qhmX^eXRP
z7=RF3r%T)xbbf7Q9xXt+)4X=^qr=O|Iw*jf#U*U6@#QH*0#m*s%7qTK5y$3sVqxoq
z0Ks@qm(=pzCIC#+nH~PuoBYR9w`K-KcD<c6D}EuK>kMe^b*~IafW<Qo1ztUtO_<o^
z2Vft(zC*{QIYch;{&{in)3e{nWif?&=T31vnP534&BsG!KRP63s3+x3((Rsiqy?{A
zvMIqU7G_Ndz{H`n^Cs#d%6H&T?KiHI5{|il5-nc;TlW9Ub9p_1%b)avZpJ=(?+nlb
zogvdq7{y9}&$v{ks>p6nxAY)CcvKdCW8w}I8Zk8hwNv`T^uk=v%)1ZnUHF&2EqQA0
zt&5Uaq7R6Cq8t&aR|E@@4JThXtp5RaOrr)vt)?~LY75b6Iy=f1;0a3|9r)Gr9*Ob$
z;R|c~pD9U6nAqZfd9dFktm7K}q0MmxA=nJxB{<YMX}FG?d?;<K3Yi#jPM%pecUT2@
zmtC>Xlr}btly=5h*;2~t?0!EoIdwSw!;upnAw|JhDSVO;Kwk-Ba-b~DJ;91-$~=Xj
zWQwA1&fl(Qh)X>hR{`3_&}@`UfPz|xiSMV`PXBjAqVO@Hjj8ry!$I+UY21k`Gf@A_
z%a0%)JRwW%&OA)r>S1~63*R%29mgm>14ot#2n`&X&1`(>|72ZQ5TCj)C4^PpV-Now
zZc6kp#P&tf^<iKDozzS0w<J!XgPE6g7P9*EzY4E$2A^^Fb9OIV)$ixKCJTu2AMKv0
z6fp+exbp`f{1v#&-&`jI8-dvU^Cga*0ay*46}jDKd45Nsd4M62ru*qx-BPfwuwYDg
z@gS0c8OeVPG&m5;GLS5g`>`|&mCqEEO({k!tA;xH*IqAv6l1%u{nRIXv1c=|Z9lQa
z7Zb3@(6^Q@+2mS@4%*LAJbDnb#@0~lYzwKU+1?8vv2sB)yt(Co)gu|!!{oY}vk!fK
zXT>{J`dV9O>eb*Zf;OpoQKpXMIc4_}JDy@L|ARmLw8h6x?-(56aRSg%M-RRu=00bQ
z!ieV-vT%C;{{4UB-yh^=7t(;wt}=kmJC8(XAex?rWzE2*26ed$!`|<z3IJbH+H=(q
z)LTutcIE4;b!Tp>qbZlihuS-Ci#!bzhvYbLLOS*Hal0MZlD!)O{-U)g!=YQoyyZz!
zBAX5_S%kp!@?~5O45v`+%Ua&KJRMxCy=pPFQxUfu<6>04EX=2`k>mL<K3qf69Ov5m
z4|4)umZclFw@YLm^R3nVz_YXg6J9qzcnBBijZ8YYa((Ve$(KA#9%nTfzXKvoeYv6Z
zfcHPD<S!(C`GP4s@8K<7qp$js0uylV!taS*qRR!WZkXueE4!4Q{{GfozMk2zhzrhq
zw*$%7ot&pe^})u04f{)BUERQj>*)n2YLvX$P|vVMO)!f^1-haBleOIJ8xfz}Fmfpa
zFoZ*~@x-(0@LeRq?zMkqUQ_iZw%YeD&K-)F^BTofnMk((j;x}$&g#cHy8sUwzi1YI
zNhN<LZfp_ADVuMBIL!Jfu4<YtSHBH?|0t?>D+>42&R4q`&-ok=3;&@e|3aI9=zVrC
z;PmFDNi*;}&s<7;;=nI#`>`SO;a0uPb|}(B`K5PFg}6TNjWFT{LSZ<{8Q3n#UnCtT
ztS75%u=HHyYRh`*-7sy{m{~;Ltc&4H+VH2Nh#r&+PLls5KYVK?XsY8HxP5lg#dX*j
z`|0YxZcS@ZP#WI^lL8e!q&>sfu5D}Hqb~qpT4M$@CD#>x+t)F7ezQQ(?WR4*i`I_O
zhv(cvLV^#P6k@!RIsat~{gsekQG(AHj_?D;AKmLnCkiQ$&qE%&gH8vtdBn|)uw|(`
zg%0bsIZa!yvVjUh#4XI8vYJdwUr|%V(14VkY@$n+Gu`)+>g%r9OEpSTMiDHKOZHWs
zo&xTwQMG{YG3>uzkC-Ra^+Ico!)I%+k$DeG?^X?+uRfIB1uxMAoNl7@|A23Zf?Dl@
zN4NGl7*ibhVWon_w`swkF;FB`;O@D9f%Jb6);#(;p$KaAZ?*YGXEpWo$0@t=!MwWF
z4`e`(U=7V~Ytx!F_<lu914_RpahlYR4lc<P@T1SCPYaF@226+(UO@dhl$-+MsuKJc
zZNaZaLRd2jb=vT1)&~07e0=XHic^zG=$?Iw=D4GlnB2}@kiV#;W(}q41G-9{I+Wc6
zD_0QrWLyJ3x(3Qx1=$d*QcdT!oYh!79AUEHVm~$=%PV<7W=W(l-L0sCLrUPuoL0Yn
zGVv9c(k)a7w)pM!o7Yx=1=-bmMGU44c8zuo-E9HPzcfqzttEScu`55Rf4nG^`^Nv#
zd@*X&F1i09KF*P^hf_*h(k&?TknH;1uTplxaVuLXkTcy%obktM+>xbNz=-vJJsLp6
z^VzcjXu+6U)URP<(BJFUp0B?H!QM9btZq^&tr}Oz#ui1RMF?akCCCB&9u_*C<Zp0-
ziWr|Xd-dX-Pd?8{aYe>lbq@Twu?s#U+t$Z)iyrl@r9`6j+Yh_0mFJ9GcnEflMWIpf
zt_k&Rnll>IpmpEKs-p0N#1|Xjky^eLx<zwuqIU$e4nk>NKy==3u4#ZRLZNkmFGb*h
z0eOrEgeJw&^-Mhz2izxbVZnUw{zXY2ruiqHfnb0Foye=t=2yMwmsnTIRvh{#+o9(w
zU5pv=VNs2$=3h<iYP-MzTe}Rk7k&Zo0q4F4NsZ*S&ng5g`Omfa<yik|p^}SHdDn_6
zQC}8CN+9Gmh>Z7Iq!z-Ei^)Tu0+F#3rpASWb~R{LZ3IwI4PT({I92bozU@4-7a9rM
ze*PKzw&k`Bpy-Tjga&_GeNw4QQJcpbg&M)RYM3TV^K3flL9xhua<8kUYzT8W(Y)6^
zJ0YNknRXb}ipmE>AWa^>^o|Rs#CUoP;7$*>rF%=NX+K)&%%uSYH>?BoIh>v1Kzsu-
z2J6BH+Rt!@eB3lJnUrXoJ&P%qpwT$<$uj~ra>@%GexDv~$dytJY)hV7RKT~EP75sS
zlO;3}fPT;&L91uv!v<v)otW(^#3utHw%4$YbBK7r*$!TKtQU-wPkY6cG2of9s1+w)
zB!AMf%#506+NhBcLnNNnbh&{V=_H`bf%EQ$C$rxaG%E7c&xM#mG%w0Z-e)H%UrCLU
zZz&MJ-<)jm?{MQE+U?)|fS@MtVcNLGKni#CIayx6gwf8)#U`w&uIzQ}^Ru;)Hn@p`
z*p(?LcVUhW-UEorXs(V}eX%ZF6DT-*)$gfJS7|V;5{Vc%StfkQ?Wp0`{;>a&cHbK|
z58Vf05NG9f-tkMTE7x2uhIxg3Re3|B^MDew-HOQr^953+<{W<b)(AcSaav3Hr%kb7
z2mP^m)2>;A*xX&5pX`tvek>U<Q_%S|9CUKF$`mgkv@LTv`yl&EWLMQ4qf-;~K=>Xg
z#B#ghldrmIq4lXk#<?Jv`}3hWkrbM2qCn}gm`7%rg0L@b#$LT>h7H7GM2@F)NSJ6G
zysMY~lM~5P*dC*sT$XcY=P-X3`~ek)Xz&4>41y((*`Y@A93EjxlXszz7d2=7#Bu(V
z8h;q+2~<|&LI0Q8Tf~56V7{dmVe28`)})`~+aIU@PR9y}R<RGqPwrma>;p6{&c2sJ
zf0?Eo996e!K>CR)Xr%PMA+i@ItB33vdQ+mg$JzGXCGKTYeDf`nK4&6^>mlwA`6^fo
zbD1wZ=hQq}Fk!t4){;xMcF`7ynTR2~+wIwP+|?tOt=V%HTZ}^72jb9ZRBZfriVPj+
zDY{nr)b1i}ah#uGHbCab^66Q{POJP4hX;!MjwKc0a&)N&)hV=xEMH(S-Lw6z&qh(I
zcW)d%k0G~TKl-eew56L%X&JV^3=G~w#NBHwLzEwVEeBVt#uTxAL9*%>3wu9J(fVFL
zM2S^w*g3V7npdmi)R7m$!C}H#{W6lb*r7K`kG+^H(^6?|WUGb0`?nHz?}+laKT$uI
zv)}yg``Y5O?cJ#RaQlY@p!1<c8{+@v^!^0~1VmNXy{=4?;5m`iNJ{JMbKe>ymqQ#(
zw}Y-dnaZgql5WGp!LE30nH4)lh0a&vi7{nG3LZ%tqYBzw)M{f26sLaWIk1!aliBJh
zbakQj>^Tlm9tDmd;y#<Z@2R7`>~}62^U1jd9--3cU%q<E+2PQiFN_baJ>x&P5+1iU
z*58V}TvA@5_>^Fzwh#PT?y;%DH~+B$-BO_2<tqMKMEu4~AX%t2)1_x&_&O!z+ZP6m
z{igfOn<)JB^FeTeKVjjiA<y6)&C!CCUc7yuNFg(7s=dUxSy$E=W*9GtwIBS#(Jr$x
z0KT<$G6$2qIeg@%EKL<a?|Em0k);lu%FNhgWZ^%*16J*IwGqjdGL5gxE>bA;iR<dP
zNk3%yT`K5czr{|EzjWeH0Q#rJho^+r+j_h&kRZAt;FAAE(<QIVLo(%jw<qn@s=XDS
z1j-I{-Ql4}k6WT$ykI`*5cV)4BSO;2)6VNM#i;9%#e}8I?o2~J59gEfo_pgTkMKTk
zX{;9NrrKa$W^AzgFo$!R3a+pOOFc>vIh9s#eVg(cC_T7MWQb)qzxLu$^ls_ZV5svh
zOYRpNzo2oX-Ustkn{!TfwlWmlZ($bz=yh9cxsvs%FWUvT#I%_3;}X8Ts2u`e@lp%u
z*14*5F6?c}82sEihDW}hY~%}$%x~|A)FgLkcZ2e4ndJf(NEhC9Dow)7Ju6Q>LR2Um
zzy7vKrqq=%iV1r@d^-PKEorWOPEpE4Vff>LltrEMIja~&_*2ec?DeR;{z(_jZ3(5K
zY4qNa@G$F`#%yr7dDj)*(8OYzZ<#?GmFv?ybNyX2f9tcoreQONt1i?7#r<>5aGE@Y
zIlVtp%n@Lr?w2<R5c4EpJDxhlQ2>Aqv7ZMNJFTg4;uET>431?oy3{uIVU-FYrFxRE
z{5dbj4H2~@tKH{pQuWJF`J%RdEiK%fL2=|5E!ZW14t&|~qC>SyV{V`m$h=`I$wdaw
zr-Sp&DEAa|-Fk*QP$}*RS$-#^ObJ&(nApSex>B4e9?OBdAW5#T{g#2OvHXOp4<sXU
z#1cKD2O<vbB>SB71L(d)E)4I{mt#NoZ-*7ry1zh=+F^H0NQ@ULFKp(5`2qUgdKaSJ
zMHZXR`cN!nV?6d#ENqkQ=8(hHvP-M>`5UpKTHHZpcQP*(^iev%2_>_zQe$kp!qoBn
zEht?rn`G#If|cvoFk&IRCBJ2$Z#r9V;m6Vw^GmwNI?zB#cbWfT$ivSL$D<F5=LkdR
zLl>d)?>&ZS@S1`yCZlM;yNx?`_oQkfS=v2(o~?2j&Mvb$_21c-6|E<bf2<2@;V+LD
zsX&s8^-*iFemwEY7$?oq+sV`-A)Q}08>Ee4+O@RL!<_BSa=v?-!{)8KudQ)cf*$cm
zpf7lYIeV~h#_9c*Kp#JQUN|hT#b3s`h(Ev-<mF7Uyv|LWNX3pu@&~>vVJ`37a;8IW
z`3gWr1?GT(ue4-EQ099sn(Ht?%87#};-rn{x}y<-?NN<y$D@Vg7UIh-;FmXkpHwz3
z^hVu`Gbs5A;mtHP;>M;xPvtY0^lF)NCXHWE0Dwt$cSrbHL80x0hr?`=aC!^51g!nR
zJTQB4+F5Q>*grv@siX5bkbdVagi<d5S0nrnYbqg~X&2fOHiK%CyCTQHLkGSl^$Mx;
z;+396Ue*1*17m%zH{#~H&7$P|$46=we4zKE^=C21i$C?5&jkfI-ei(#j<86Vy6o4J
ze+3)XQz-I_Tb}gDOYMw!Fu8hU#9H4r?IwLH_NP&{Lbk~BKL0E?K1YcQBOK0**qJC1
zxv0(3z1FdF0f3<o(~2901YA+}(FNTQq<7)<twpz`LA8B5+*!kRJ&75>mp-%JPe)mg
z5YKu;7D&DX4b4rN_)2vWysMWzoK=T<?t;k6Is&68ZJUSa=5Qr<<V(+BtXok;S(C!I
zN6Rq>%G(hmCG^d6V8{d1n22)_ylZ;ejc(!T-#Pi27$N$|*&?8i-zgBnm}%(;bPTER
z>Q`c~#|Tb|obB6wZJ9b<S3`QG7X#r*WdetoVtYDe#3TH0v7Rqyfin4a!<}7xO<=HN
z1@>io!1e&@gluA3stEBbPa))a-_bpPsf~heZInkPxGNg{0@e1Zh`Qe*jMt`FDJ}`b
zaNP6Vb(|U9u+9BCB8BE{*;kH6WddC#Z(D$EF55N}^f-&nLkRzB%Ykisl7$ky&{CwI
zpgX2oUlnEjUr?k8hs~F{D5gsOO0Vze5~Qq&iX*w6B|j;KU=U5#0@Pn{nP}~O3S1}u
zXx)m4N+xp*bgL~EyX`&gX`Q`X@^Pd>kYsU}Xgnn3$@D1ED5)cg&2X+zJp^&P=yco%
z8N}`2<@8UoUC4#xBZZG*6-|Ss0oR7*cw$mIV6kE|5o78z%{oagb!gVEFfnITOO2x4
zulhgIjk89LVFX{wEEp%NT(%+9$0$08a_y}SV3q}CtSjvJ{DXbs<*Z(Oxr`)}W5e^q
zz94M4)Vt2>WpRJfWPe(G&?OLW&o|X~KB%57m%_0NbX`F?{fe8w&kD4>1J=|~E~r=G
zSdJ>(IgK5;%V|EIXkKLr=1Zf;2Jh8}SVAsYjz$#UQE|LW++#@Fo5E!#L5c<J%`Q%O
zTqw9Z`cy5^>STr9oi7{B27=_HGWOopCwp-YftGdB_x#J;7zJE@DunFzny1Jn6)!`_
zYgJGkdg%c~jk%vWUk?jjYuxr_;PEg9m+?&3D~_8{IX4Y2>6mNj7O)M0F79RpL4t9)
zgs9$|Ax#!QlG=<9+1mBttOta}ILd|4MPDM2PIG)>U4#%dv;FeFClP*~1%^%s6E^DE
zD8rVIUvRdg_q)=8`r`cZKA1!0IpyXH(?7sXepT9J66i`cq9m!D^ZJ~xp=sESB7PKT
zF6xtS@6r!!p?|z^O0&xts0yo$oB}m~6#1jQ3j@`9+~qohv|FT(jJszDb~izkPT8)z
z?Fof3$l<T|ZobsUuLtJE)vFx_*;Xw>uX^j^qHgbjYsn|(?y8iH`%@}5Bf)q>Sz7mO
z>1?86PFv|IE)bJ6$mevWiTbNV&20yzo#nTanqIPwR7&2egT&1pj3jKnc`x!ri2;*P
z3;u<I{3|c+^^seDgt!01d;!r0b}#Tz3d4qaaX?N@l%DyW=<J5YsP0FPD1@7^pLYt&
z%QK-u11i`mq+Q(!`E*Xc7fvcS1Z&C$yfdSzxr1|`sHH7@<-vo$dju%&20DD$ZrAlr
zDO1?RpD&+IFSk3W=<IHsBhVBV4Gp?hgXWE2unfVfV$`mz-R)^G8nT$jRmwJ6mr9)P
zI40>q-+|ZEL!B&czK#%~kNYGm0*`7?fl}Tro(s7r4U$88hQcz^mUFX)KzE0S^e7hZ
z8yaRd8J;nz0Gcl1{}AYZM^Z8sRpI<)a;)y|i}~C?{0Pn?!WLKbsTst0isF24I#Lgr
zQTdFNS}Oo0U+sa}2NFLu701hec+A64k<$6Hi<U>bt^;(G&xi=}PwbGC3dRB4bLZ1_
zLG_P|LHSH2Yk3$kcQwYn36fPM>sUa@!E5q4G@T1op!DS0uw+z;S@4#7V@sx@7G^_f
zZ*w!`Tb=FyN|yX5jY4tb%Q*X$El|8xM!~h~-h@t{kBBw{a5IEj{LUOAjpvwtiE3}1
z?($<QqYnzsz_gIGCKFBW;!w|~BE^=kGeSkebjPZU?Muuh0-84n-VOd%*W`Hq$#G%$
zbTCpDFoyYye(GT*d$ubY4PpVQUh1!5dW!z0bo@bSwz{bLhPlZvrx5x~`Rbl#%*Vj#
zGJ+`-anAYbid_hVR!?Ffpj!GI{d4+>j>h^s{KvDf2UCCW<L?#AH$6=Er^SbT@J%Xi
zmZH#t1}QTU+70&~&h#b6XZ7D?f?nliNY8iip|9+Y%npCkQ4HH3EoSrMxfwL2FM=B;
zS?*Ff0?^c4Z)pP5k9nb#H@*J+tNZ?kivQz3*WziIMU$-vmvU1cTrF(UY|(O@XL_9V
z^#YlBU`w4u4lblG2`>mrd$-*`R?}2M0CKdISnZzJZ3<~3EEg;6lI5**G^oj8+Ytny
zzJlCVXxeFtHThey(jS}rRT>ixndA~Lt+1p~1`HXQ>-@kaycWHSlwPKE;fGCqJP=-6
z5QTiBK#uCIS2r-rzL)(#ws%J+Lcumi^+je3c&}|T-><@ycJU5X%n^Z0l;a=WE6P)a
zrem%R$edYWOcC*H!kWv!R&%63TzlyfdT?#p2$3PARw=y1oD=LKgH_j$SCb0l>T%^l
zBr^#$76$;F1Idh}m;b|V`V+n9Gifw3CwK|uBOyTL#_Ls@J#yM>(L0?|oJEWf!NcsE
zR-{f{JalWL54~oXylMj}K6WJ-^>Q`&okDOkbu@`aQlYl!i&-|s$brzmPp|xmrRG%h
zl9$Pvl^pKh4ru{!yE~3kg2iV{QcnV#94sy(5HwsE=XCKECX%qh4*j0gAvGG{5}MP%
zeF{;@w4{a|M~a)P&0A|+TFZIwG_I2J>EHEAe<JL!3rdLsa}4svyzz}LB;-sW<E~8K
z@HZ`UiE5uKa&$Og2F`g^crpvhba=K+!D82QmvdKes=(spYc(PIL;MK1|0sXcwxIC9
z-X8DnVYJ%Q)^7Q?GV&jU`0ajynDR9egC}`=7G8k30p0N7Z<-h>b{FcD=OQ5Z&4bw`
z2LjbD>lm<tZQC#|D%7(vlrZODHEjgv=#$4kBXfL2WpG?vnfpGDL@z9PYzfc*y~6y{
z#+Xy=J$q9KXnoj}@dg<^TcCF@foj`*-|NP_yZ`qSJ1D<5UY?^z0Ap70d|SXe)tVbY
zu}7=s^0($R8e;Ef1jcFSlJwa&Nuf!c3MKjc|0OHGoi~dP9QK7QQOQN&EE}4&J9K#W
zsrl`c!euQHbxOR{g+vaVp#N@fmut1kND3IHn++ab(le*8h6)bgzJ3E;8khP!_nUL_
zPmt(@faZjcd_3}v&g|u>8aSCzk`m;~O)11hSnSN2n`i_}HA!JmS)kBT{O;`X9`sZ#
zuP%IeNsFM9k93SzpH#NhK$Q;tBf1_t(;sp#U^5Pj^zc6N?1!M#lC{r4>e%|1>Rl!7
z)s$%)?_Y6XOeu^bGgJhx>c(lj1U6=i_PI0!gi4~qh6B%f(P}{ijRVhAPh3OKEFzZQ
z{*PSbpM`~;y(fHIR~q;xYxeHc@oG!$)|(U4--bSczAx;zx%Tnblz`_qH(~|@gTm66
zl)<mZi_k2x7$})z>1nwIuBG9O=%Csh>H_hJaFQo8L~8JR+b*RK?XTkTr#*jLh@NAT
zi;3Oc8x9PLMEFjg@;dj;ww`8saKZ4#v|U~rFrayx{aHP|?-hG~j4we(RAkdiRV&xy
z+C=lZ*9*BF@z4)KWq^or+KOxAQ|P08Dj~-I5fc9s7j9`{P^O-6=5%RSKi@T{;#1#j
z(`YoT22nvSlIkD3iG_E}@GRgB5%_+@V;;C&(=ka*{|tD<gD=j%x47I@Xzj}`lGyz2
zz{~5@bNS~HqppE+r~lrK{L@LCx1yIc_s73lAIi&UCn`B^3uq=62B1ANrA-_@A3oXA
zC7qP!SqUAL@jBZc7MxZc0Pdh>R5Zzs)R;k;a45xlzD%PQN!oqj5FX!idnZ2U|5bwW
zr+xmqFqae{APQ<4K@^(1uAs$tIY)+vuWQPhm`Nb>c-JaARo)UCagzd-fHnnNFa9z5
znSvq8xD1uFuP!y08-#or0_#~kgR0$M)b7j7{ZoK&)Wwh&_Jw!(=ZE)Xx#=A4I4-f1
zoW2`jq&UpPJ@R;BB!rs`O;0jns(M<zhdpYmmvDsv0bbW}ax4>4X@M-j;A-}qa`0@R
z*p+`s=ky);{rPXaP(k)Ntid-k;Kp6Y&9glVKH~HGzD@H*&d7WVw;{vrB6#DJzFa8y
zZFk&6WKrx*RZj($fb6PgJ8#zpzi(dqYd+;qm*+nVMS4ki)+BXLiJL4X6cjc*e?7ff
zDyXX&^={>>ttt^@Rq)+FI)S9SSZ9gc^ztiox2lE-9MU4W(@Y*0{nZlqE5H6FE`E8W
zBdEkY_XeA=*wy}j&jQGX&_aN+Qos76W9_$nG*>;}bzQeL%7fT+&U6k^YDHRvQ`Uie
z{x@ICe+?qLc?ZA#`$+qrEEG}n*@WN~Cu)97{T=0mH1s&5?yKP+H;U!d&MEKsK0xn{
z^i3t_4@!x_vOV0YCsZe{LdE<}Z@+S~aO$Z#{|`lg67n;I+sq#+TU?v>+9F4!(ijB;
zDu#aulB|>V*>X3O9~6#^4E6cs@7<|n@YC5x5XgMv)RM2R{*OrapQt!RpZ$sktOqNl
z#2rjvOkGqdC14w?D_NN6+is_GS0_Y2IPro(IfwTqUM~3l1EBilDdq|g>_T?vA>#Q~
zJ>7tVy-E91TxgoNz27Kf{SJWjRpozqg8#u^O8J?&Gq^(jkoypl0zOehMF!5rT%-(j
zJ(sa4lb&sM!AJZH=ig>V{<;t5o~Ah~*!$wxN$|<lV3&sdQ(Q?j%v#HjHM?Cf-2c={
zqg+4pGn8Dl6#_5^kGw0xF~KtXogoKv$iBmB36ncB`#-$s|AX~9ig4ER(-{O@A(+Mg
zV2g_6G^|f%lKYUE^YTLKiH(0K%D-u7!OS3X_m&BP>^W#ZPxd|HbY%3lVmrMrN;X?5
zNe_6AqWY&6FNGlqmQ1xlYpGOZhLkX=Mt*aH*IT;g>ZSE37q^XiKJfhG>#JqXbe&|~
zeCWE%Jg?|?0u2OPTh*Q2l^c1!t*rmmv-elD=&x-0KZo?Y0V0}smK4)i6x83K!Ag8s
zd*+bNb%Jp3)y?f&eQ$MdxC_So{HMP}x%BU8^2CQkFg)qvYiZCEa<rCFdW00dg-@x}
z&pG}%aTl57EEH}h0Oo=t_4Zh60kiO5)`fU)*kn<0S6~d=|F2(~Jjr#r@id8Q-mC8H
zcCg<Aj?))>uTBO-`Q>U<RR7@_{q`}h$MCDK{VCZ>o+9+1_hm=w5`MYK`>&y(Lt0bO
z5V|#|dHM^pJU=VvzAlb7a~JGyNbMaV559e^+gDp^K=v;_IH#;gjmp4P)(O{qj>25;
zzef@LX+?^*Xxb-4q2xFmW{geJa@4x_6wfxt4`LT&TJ*5s?Pl+Li81AK>^wJahmY*|
ze2J<3dwytXVHllp7!AVdo;{zdOD7fdQ`(RxgUPao%%&3eUkL!bc7mB1WoC=CCVWC{
z47Xfd)&UNmJq`Wn!_hX5M%acnP0B#fkKmKay6|cHE+@BVc_YpX1(pWCYBpX9ib@?F
zG04%)aMvq~hvFseeWlv3{@6<wxE0X2JpX7UpqmFWzBn2qk#~F(UokK(Ub-r88uRHM
zTuaCN@}ZB_07F{ne>KaVFdc?fcIl=RYQSmD?RS~RSQN}nII=TSn$pvNo((69Z#=$=
z3u@S|S@v&Q*t){kwOzj9PhyR|VJ==t{Q&1-OAmJI=2|}c`RLctP(o^iqfJk#o;9Bv
zhrGupDUk<uQj3qNQcfFipskY9z#~|s$X%!t!10>uzu4eCci@niS-VjGw*+D6+1;Qm
zZE0px=4B0r=-XeG#U7we=5}5pxh6b9>ZAqxbIzPwEpsfglzqP^4~5j6mkg_K)F1ZC
zw^q@26w;M!z)rfZsrFpr<>I10$BP)W7S>A7q7iY@K7tRGlH_*PV?3xc=4k8RpW;fU
z35YY`-?kQ=qcT-Gq8?g88fV$=ocpo=Ur+3}G%Ifo=E0|3&B5-)ea>nlisu0*J}Ldo
zxfGJlU<Q{l7c<=ZRIKjNvWR$mB*(ISsdh$He!5-J3B70OgVJC+2wiHTq_R>C*})@<
z%kE$MHAZCxtjUE7InXxuCj>9;K~hg-l?rzxjd-r#<=yGffXLR#{2Q<=MI*R_3TZgN
zPByFMMHi#|H~Ug|V>|<;om*}68AH~dibcz$Lh1AgfO%)O8g8p4=$B30oI&H=1mkx8
z=q~tQOq=G<G;{rV3&k{kbiA)L8d#y+yW4Y5itBIv_%&zbWDCYR0JmfJPGhcGy+8i0
zNdMmPR5s>A9n<y?Up8;$CS&Ni+EXi`;pPGRUT9vPt%wo+Ey}1^2}ok_c(*a0rtSjI
zMVI_z&$1-e7FTKABW}gQTkYEgCCigfNW<JW7+Lcc+s9Is##hJel-HGonq(NpiXN1T
z_=jWJ7b`S?%J>k5{_vYNZ?7E-*xq)moD1%q@7J$b09H>h9-A-Rg+`Xpz2o3&gXQAU
zQQy!btJT%9T#F?V13u&XA2$D9S_QL%AeI3#hct)eA~XzodiSIys7=kIe_@RlI+oM)
zQZzBYcCs20e5S>Fii&nqd>~GiYfS2$YQ0P4%DoZgB=B=^Pr&Uymr7pBSh-+MH4yK!
zIJFIBhVSZ^Ek{Gu8Ve~YT&wXa74aVG8aDlH1p%MKx4TgmH-62a)t`_9O$D6m3JGjd
zBINk)&_vb$R#^lOR5VC)MWuK<86oJf{NPsPtA2uR@o^`?%ZxQ;sh`hqC##@i+5mZ+
zQLtHBDgB}7<s3^3Or)}Nuqvj^o~zF!3ir^6HIf?D%PJcCnBQRpIpgK(?E!wKYjfo6
zfXX@edQ`=VW+W<!_{t`(xW%DTTUyc-?fMTz^|_XAX`;v%&Z#r(2p&vD$y3JPHcXc|
zA>KoI50R94Jw?Nm>B%v9n+;3-O}(DJLksbYZ_s*Y$k)8`cN`ovdye1Yl$~rTr4_MO
z+eT`mH(ymmZ0@By`<8x7Yf&7i(2QJWRM|K@5=-EE(B@M&=|gwJwZ@r#W#`B^td=*<
z|C?L&bY<Zqqi8Q2LzzpV%1mbG`f{j`P+*eW!65oU>~$r1WT;_LzSO_EPySl!@SBpD
z&kX`Lx~I+$ufc~?NZ$JN*UA<)`>1~=bcd`_tom`zI1>%0mI{bm9c3yA3VR;LU^pi%
zsvnjIoPRrQ%9z&G26z0#Q)pLq_W&P0HD9J7o+M^;ig)0fYD$X1ke||yvSuHvyuz`;
z=Y@pzOFm8<#pWYYaP??y^2BxL(X}2V>VxL(@^>v3Cf$s?;E&I7uR}^BrOt?SvA0j?
z%^1v5L5(X$fE(_}!6D&svF<RgjC?&cKchmS1L&cOCeHm_hbLDRWMWhZ;X(WOY93GF
zeQ_mJ;^Z@#XVv!k)1*S%@j|@Qmn*d^h`Cot;m3FPdLIcIfk4?NOy3i+g?M{!ftH|U
zp%Y<B3t<*jsj=m%4~^KKyDiP;c}w{+aej`7{d%__u?gQ^$XJwxEc;GeCvJ~vMzt<)
zGR(K6-pS}w&w;LGf51!5Y&*xPiuTO~_pMkn;Q&}eJ5)-0U@)zP$@#*q7nP=p(}gag
z*h@uWur%D7O&$PfgWnv!;RI-&9Cg6|E_|_clu1N?FMShCyHrOh1g_f(=liJ%I1iCm
zGUlbD0W2MB;FYudtrv@2n#%*lQ{50V@-^>@q*vVgsz5Bwlhs@^?X)xaQ)yiLB{;!P
zD`8tYl~H2gb?%lqj0d?-2!Ui`N8I^#C?c8R(y-S&Tw`I=qd*hadwS!hK6Bfx;#;mM
zuV%?K&u;*$VPC(@$YP{|^J@f?S%zX*(-V>)blD)<_B$n8uajwK7ka*?jM&s9&Pi5s
zj(LFPV1<S|bfRVo)D<w_gmK|<#+mOY{jRS`G?WNa;ev^t@^f&ah-Av^g~%qkY4BG+
z-&P-;)@m8!!B3izIUn46-21$E&VcNPd^soMiceL_!{(Few_E);Ml_8^ik67{lY(m5
z2j5Hyz%%(o1$-3i*7@S)ri^EvC)NfZe>)duk|F)lg--#(-oIukAGP5;Tlok_ndXyC
zX++&#n4CB2(g?JZFLd}|JDzuj_vCh~(+<=5AKYbgOj@+W6}%8aw`+t_aR%7(!(8W2
zq44ml=ZNXmfJ^qijc1geG)|Y5Z;g@h6DdIkFNS7l;!oYZC}hFUcj*+ec=e@)+1XR~
zK3q*_k!`s6Ld5IN&qtaf4D5IAv0oO=XP~?H!CyyGi~W7EQi9^}of8fh`MmH_UmB-=
zCYyF`c5HSW9(Ql3KlR@ld=X#*^GV&`ckf=ZN?cW{LpMA8<Z41699cTbcK<xzZjA6Z
zJKv4Mb{Z02xi5LNbUPL=)m1Ech=-vb22oMd(49DSQSm>2m@m_yRIvbLBybv%x);Xp
zFM#IP1^Iap(t%k;NXA-B%=(4>_ocu=BV~S*6{QL9t$6#TeGb5lRWTvdvbGXVIuY=3
zJ1N)bKqBWV&rc&&wVkrKV~6F9V)vygzZVS*xX`USr0lNeaB=_-%aX#ec~`ZhyuiXf
zKtbKXuV)-qFKTkD16vxbXF|p)8S#5Q93A40?ijk--=k3u^JjE!S!^P?*7)#B7A8EF
zy;v<_U#}BG8&h&lzQc6jxX*F;sMcewp`)DEWSTIP!cRn6qj7_^dlA<1TQA*@fAB`T
zUM%I!`>bn~Hc{iBX6O-g4Uw8<8XLDl&T8GPJ}+x}O&qT}nv=!=B7ze+R}YRKO?q!_
z-u9?LhaUGdJg-=i@7Y20E)3H%RyBO0yztF`<>OxJ^fpW7_HbvdgO33|&o0nEZ9c$`
zvA@q<H)3_<?L>b|Kv};=PDW7`p3T;wj<&EJ4eH?aG+yx?yb&|LW=ekCRa5BD>D9SC
zN_0TGi2?lVJO-;BeBQMr)jfziQn<~#d7}xvDlhpsyWJF_%yO_K<MJ(tSh_P*ic>q$
zac0<!etX2tEJ^Rdf4KPDA0A&Y2)E6E7qd*QI?<|pfvJ9%ylwl!?0mJiD#rO2Fheyl
z2!3B}7}YXk>sGWwS8ef8nO{+c6S1Q@_MEJuHcEkSmEgx{R8z!D46^pjJ(%Hx$j;rv
zzenOV!0Dnz&D2TE);4Bj$0OAX)lnmh!AmG$M3}m+Goqv0_s7v-{%U$}M6|}^eRVro
zD)H5dO=?fZL~xJi<WK5WvB9vP%B)-pwhaLAJSpE|Ja}$%p{zDuZNFn7)1LpNu|mt<
z%%r|X=jMP+%aP@1zzuKL_3_qyWm`1XzX&zWzuZ4^T-^9JdO|5oO;0GnY@2r%KVnvt
zrxQJWZs`Eue|%^$o^@{10O{Q;Q74a<SS8@_S65A-mQiX4oGSQ!M2ZR5T0n7Gx3@TN
z(N1sV*R24MVY|0usDs;xufxdl&{QT6E`Qft9<o)UTx%}!^3Rg8WDSpTs|6HtBixHW
zr#+Li{b3}wOdvW$2aMW#7S*>h&tr5U)N9#4u`%^(+piYQA2HrvO&k+R1b`2<<A9G9
z5Vi6K4m`Krm(m?TjoUHKJe#tY7htYrYjHhB$_pU~bZNJ9j^ylX9`!ZswsaamoP5ac
z7PiL@O)nT7TK2tc_XV5su4<^SMAynKV6)P;7mv0YtUoU*&SviW>DKaB*}T}YPONS&
z?%ffWzqmI-FUx6$7qLF7iYBKn?35?YR0ug6sOF4+(qO#R;%nC{h|TvP$GZ2taCl)^
z`eD$)DPQlMxLbB$3ewhaT4K}F-gt8Y?^eI+p7nG#zK*fkZQ<xJ_-tgao4<MAPIWuq
z5!mrbXH7P$#q%&DLF6YL$GDyRj#F!Fv@G9^x;g8uQ0#*`&|{5!?9q5OYfbOr2N|qq
z(F=5)9#|2--)-%%-uIK(8krkg7{2DePeYU=*;Y!%*P-IZom1{rvkm*xvsA92&1CJ=
z|8&J`;xYCz<8!Q4D%n7dee+FZYip|+U$+`Maj9!-DEY868q3dvS<LZDO<g)!At5rA
z+Ug(O!gu8U#DV*~zs2kU##espVBzqUMR0vvd9Y*s$1#zGphMN@PD7B<%(t$zt)ol9
zMpuvA7tThn84Ns&&#tQIFz@QN)9BXl(b(d3FvLHah)aH7dwD_Wz#Fz++pL09NvyZ&
zqfWAU&OQ9K-4|ONyg@l*lr5f3*UhY{aOcO(Y4eT+Zx}hjQOz_h-=pAjyY?WlYHRru
zcG&0n*tQ#S<Qc1fd=+CeagQx;i5wHNrPp;Rwli<d0g;ehP2!D+TjOC2<^7O#J8$r+
z9<i8q$YtQ50U%wWhH4Km;|UC!`kshS@)H>8t@wc5^(sV6jv8-I7<$yKRUpm`Pt!M(
zf3B>F3^4Qkf%(7A6!DNSE)$=f*5@<-9@4cL-ugB=V5%go4QG-Eyc>9P!CV<q<^F7a
z-nuUM!-%!cAi_=3p9-hpxy7bcqykWPjxRTbZd-Sf(PHu~4MVP3j5gV6KNc(2;#=#l
zrKw9=rIa?RjDv|_%!t|N@kR442w&Qx{nk&bfcC_W0-M5E%9)}xoD&W$`-<APVwm~s
z<vyk9S38Nxz8;w_kQ9~rf7*NZxFqlXf4s8BwidRvavjiGD%&hAEzJ|usug))TTSzX
zmZlk=1x*pixpG?v3zlc4mPzFSPk2B9n@b~)L5fO<O*|kVDIf^~->dHX^Z4EC-aY^P
z{WJd)xZt|puh;YWejT2#>jFPBzs~A>MrivAvNll8s~<DyB=;pv#rfmI_iI<R9~iLh
z#{gbbLIDe(VsE@yCe^3P$a32{aeW!HR6Szio}&c221$4o@`T|&8d>Qz(U6=3P(yhv
zXrjlV@L7btUc1TrWdM4r_GysgANr6xy^#*`*p>zzVPP2XPevGBBf$(+`A_Kjj)6ow
z_!DpVw<7E`B;Ea3h{%1q|NQQ^wo=VUufFxx42lsPA(`+|#Fh(k=wZr@#ZY^&DlDnt
zJ|UruvTIQ%CPFhUHC}{VStT~yyT#;>EW}I>rd7=+gZ*0^oIm$NcH<t49;WZ)Y$mU^
zh5%1-mmHh~TDt*pUqW^ezT>9%G(b%~>#cYUHRhDbr0&9^=+Yvp6!Hnfi;F8Wsp5MS
z-`sN3v&{52y8_Jq^WR|OnDh~vawtIGOADTHkGoSHU&K2+A${I$&SRRC*&VC5hFZJU
zwo6C3&SxprwwvB{NglYGkuT~^6ZIm!E3a?KFF&D+I3o%s><PQe8e^!R1FOrrMwp&?
zG6vq@NtMEzZuPcIT<%5}*6n1J=BVXT<DrBiu^kES<rs>O+IL@`0w26WL*MC<z-#YR
za6CvwvnC%qcY$f$3DLH=*qzNna3*%yih{$pL{;nGicqT1t6XviGJl5r&=e1_#smUn
z124j~&b5Xmmk|;q^Bj@2=IBVS>>6!5RZkvv=~A!z#SA&)v?CT$x_Ng>Vv;@hBh|Wm
z<j9XjGr!|i!}1Wcg{{TSD35S7{Xp)Jp=_sy)#S^S)OmM)Nj!U-{{n7m6XWU`0c|LV
z&fP(QE1foG<B+z;$Ci13xTXQ8K)zj!oe0H7tam#2YwUibE$0mBh!wDc&9J>&s(O91
z0$+?0v$2p9X`+(eM!6R>y0#V{t-pOG2FD+|nV1c*DQve4eK#>Or(lg<tDhF1GJx9Z
zGM?zxC~L{HzV~|SY4vWKc8<A@p!tF0TiN+)Ccm=f7#mWZ>UKl{GTzbkcg^EObdn*i
z%Yo#6rfRR<c=;ywO!M5Ife$d~Z18Ej>wntFk{r?v?+@uLA=el5B_{4jFBnBdPVx0h
zqE3%Kb|+Jz4M#_AZpoKbs%Hz&v%i7_zEX<6K!6Up+-td^Zj{L~+Q1fhk)ZR*N$TU2
z&Sy6-UG?r(*{n{AGa_^~9eJ`p%F)tTZroyBe7Hf~HkTDX-0=t)t5|5im?tr*aRYzK
z#)J#~&ER?0jE)$>;`djz<(2TX>b(o&hl-Bh%GjOkfg%^i_w>6~+|^8e-uBd-IeIVW
zD2`S%fbK0WRLAx*v1MXyM0b&rg37{EB<M7rvoGb&T?MW4xk<0Z{IlhixS~*6gj3)b
z<RfdOs9Z^2+oIW}NKl}{tqTcd7V8Xj&*&`0<x3_buDnjWeXokRBN}-H=c^Xb87Ju%
z3J|<ooxa1oQ03w}OZAo4dnhZ*y5s)sr7^lym_Pmu`qx~F%ieF%XiViIkLWemVQGqG
zffnHk@BZaK7~by$n<w{$Iq(tAO{o@6la^WL16PBy!27!fg1EaX6(Nu6SlJgVaWi%C
zJ&PYkN<(6F^S)~|r&M+A`WCmVnY3-Sle>j{3@hVf_lackO!EBg+4@uZ4ErMAWRK8>
zSVDB^LC`JKuJX;gz%VW_<3x<$I=Tv$b7FKYoS!L>;bB7uN7gDC=MFratE7N)y&I$q
zaH&Q9m42Lzg#bl1vrPPXF56Z|!wJK=3y0+mpJ<Z{A}n>#(4ajI$fJzPrfB<t?5?^B
z&9&jC6~@9gH7u&qF>#9Jlod>ih%mH})n-r9X!J@&y)UAD*vUa37Rgs<eiZYIr*y#Y
zRT5WS#zbBbc+tiI)({#h5Lx{=gMm|5Nxqn66=W<Db2%rO1l@bTSK=ZP?uM0{IPf&3
zhCVO8Q;t{rFe$IT0^?;T6@<3n2aTLO^Smqlnp!CoYUOTdCaIrMc})U{d&HfYGU>G<
zzlNs@kVA7r;}5MtwO(wW?yC{?!SlMl#sY9@G8>63Rm_fQ8%(9h9!v#);=?vap+aU{
z#jUH_5QOU-3FwSzc)7CTKQg2Li#)1qRuXJM@(aBZgat$WP_dm%n2Slij}spvURU{^
z8Hrg(QN7w_Xmp+&O#gvLBCN>|u&SkhXi_PhJV>71XZ&S>K#D4-oe2+Y$d-t*Hdorq
zX)cArU=X=|I6JSB*f+S#p`v$?s)rS}*e)F9q8l=D*f}bfDQ$i7^?JIyDT6|iH?vFm
zd{+|R(paoI{(ZFZG#A7V^#a}3a_+?IMgn3z7?f(c$ex%N=6~5d-0qyB9xsou7RG|^
zj8IyfHVTFfzlZq;)b!2wA~$y{K28V!#%DFPcYWx3zG3f1z9yfbxh|nAFVuDWfc#m%
zToDj#m9d^MSfuXkgz3=%e*j!U%{bY36bJ-yodG1I*Dxz@fko>c7~}4W;YTW75d5n*
zFj+bdKVF{-3T)`mJaLZEw})y*h-mKenQHm!Jh-JFQq8<AumB9E1~(>|b+%Y^IoFk9
z-+#VW-yt1scM!z%YHrU=uP5WPI`*ASxN@eld%Lyn`H#!_XA~7@8VXv{YX2L2D;=};
zgYbr}3+K3*Df(wm9mufX1(4|t&YW0F$a&mLYxfs7-wDJ$fxN2Ja}XB$dj`H*)0o7j
zdADcQBXhUh%lI@@=U<A3#L2fAeCP5?GOUG~T5@&$nbPRX?*s}I=MbQ%(n;)B*{*Aq
zi3s$bC`x8ge+}NGQtyI+k5Ok!J7%C7)b+q^c>|bMy2+y~?{?j}=PZ-GNaQ@Q$2@JF
z{&hZH7Z6@*5=e!P#k<Hu1WfYKFzHgI)2NG0?O2ZQrZ`9zp|>mT(>!p#M%quC?Is__
z{!~6iw}%wUlSIgC9=5sV)NCL`)N8ba1a7*hxGnd%gUBK<7n~hF3#B6`%x|g+JE94N
z3x->VK>@mhz%^kk0xr$d19_d=VtB`^%i>OQ;#89g*E`EmrPVZ+4o@xTo2#-psUxP0
ztHZGL=5XaK<2mqRTdl&$Qcv_&Tr`hx8O?uGYY6;36)?mlJtlbb?$P1Ivpe7H*h5Fu
zc5iLfK!lm@WgRzC*qdEqdQP{t(sUkG@Z5GthP_!hiupQm<komNaZOJwHXRlQvu`?3
zV=SEEdkWT1*2N-^)9xN5!49#p+=n7?OIMlXf^=0T7_xsl1uHPbwOuO*Cj_}RA?$Vg
z2}jKNtmG@L%ZKDwNt><BP?Otia|{ihV5G~@CU<%<E@UL$HoLvR$~}k9a?%vy>_l!7
zQ#;v@!{#E7$sc0!RQl00PWTQ5s9?6NW()}X0t^0y|FQ2N>3`ZS-&9IkNh_SXlk|I&
z^dFxB1;a^hmP6&H*@@@iiL)haLhYn(E^OpNVE+1ai1n`6R%|Au>)Lz&Sz-A<9{E@N
z?Z|FD<@^Zl&C0`URZ2Nzl6vud21XRwe_&u8saVRk%<=p<y~rx06&rzY1^@Q~|NE5?
zr*%`b6jy7BXuT|CB8z0`?lLesyM6RGZ$dtvNg-G_o%{7&*$2x_DqRl_?fz56=TGO;
zT)iLUpvZ0cTykOJ6@rU9LlA{I^qVBuQ=xjyOob!wC+8e@pVkXfbKPC_Z~yyyHrvIi
zWpmCP4H?{0OmcH<I$Ps6?dW{tIPm|7-8U<fN?KRdDg(gInrm*?@$3GAX8)a@$edT0
zb$TomiIoTt%5|AeCO*drzgOH*UottLiGbfN``_pAnj5O_uhiOH%sMtwfQeIfaHM2`
zRWrjjZ`~)o(wlj3V<l``2lUU0&|8es`f2*+h)11d2NIZ7oU4j`Q0z4H9i0eX<d#e-
z4dUyEKK$S34p=>e_x_e_2i&>2iV~BeU<ARg=}=AC^sI9n9Iv}8{;GM}a!byCyV!re
zf?nSk;sUc@@Z_fRd89FN8MOJ=nd}ULoh0Qs&bRPF=tqBl0?ZYYJ?}tdY01Fx6Kgnf
za#0Y!dpQ!33MS?j+I;iY<^Ja{18)uPTuE?@NcnQ^PDw_wg||7d02MRk3=CK`Qqn;@
zld%8SC6-_CZlGg?7RtGg0Ks(0s4OqrJlU1$WeN|Hxz2=h`=?L;e(__h;?fJ{u~twk
z$f8W-i}y}ARO2w6oOs;=8hk9K#Q{9^$r2ZEHXUEn7K|7nV@By2w`ho%8?veCcy__!
zE(y)aXNiNOR}pL@M%Wdp;oOIOjg{tDPm$~%!zXDmucQMPo>07$@f+4Re%aHKK@oYl
ztW8ardo2DQsBv6ALybkGmJ_{gZ+!Mv@%2woT=}kSV1-%t5FECUz<t<voD7A6pqqpA
zAIo_p^T|Yx`lG{deg4lZlP*M!wxq1mZ!WVyxNJ>Lo%_Co4L*NI%CBwXkrYtP&{UDY
z;(rp#o3Bo~+1iva>S`V9;qr30<?|+d%?)gZO!iBO&PXny5WZVt`9cBg5ravKpGjJ$
zKx7<LXTABGoG0ERc^FZww^n-f?bPlqUzq<ay>p^p6-y&PC8JXu&UTUsNnW800}=Bu
zY^W9A`{Tbq`}uw8m&NP#F>~W3CMCw3iw6=oVL0i-p@*hqLl9pVqv(l3{{0Gee^EvB
zmTlueU$a=Y>i7d+&w47saWha4fl+SCa+b2+jTJiaclu*<BPAIn5!S?l=ezvsZHw3F
zG=a_R8O@U1k+q>Q1to-A!@XTjYUjn*fU0TwbEuV&?@Ja~{`?kH-(mq=OKZRK{RV3)
z6ms5u_X0gk6vo8nM8%y;_-bxY9Jmzl%<bahP%kW?B5#+MjGfNb&cQXs)1r&|uetce
zpoH5?IX-RezFoArWHQYXvPVvPcP2r=_n$;|m_daGGH36)B^;Oe)h$Yt92%%Y@WF-i
zWEv6yJ~rTQitI2v_ycg<VCh%dC0G~v;gl#rF2cw~N%B^yDQVJD-~SMOe>0UAL5^wO
zE1?E2<$GVGCyw_UPS4GhWKbil?K~<A%6-hdGWnLmneCkZrJ%?Bb~HStpXTLqVYp>?
z)0vuYv90qBWD6^i*G$1uPD5Y8e_}(7O!x{QY=TyOjdKhe=2J|nGbWM^6(^TkI-k*(
zEpri2y{-~+RKFj177mQl!Z^7pIHyjLNh({?Yv`?K=BAB%7Es)W&$6t5MblZ=_6P4?
z94w(+(|adSmNdLs<N1|_UCw(Yt<rk&DNYJ`Xj+t^|B_4S^#C7T3b3i^tB;Jvb9~9*
zdxYI7C+uDCrT$dMbd01RE%{zc9Q^K~(M8U7axWBeLZ3QEeBBe$+FTM$W*&5*{rx7j
zKV<Eo9S;Ph0KTrRC}F3*I?Xw$$DrZk!)pCAG+h5+$>aAWmue8Od`yVil|v@S$I;md
zuNDw|HGi$hdWkJ|>#h$6p-RnBZ%>y-EmM4-bC*2u4FbLj7e04{!&(BT%hG?^ZzM=*
zmKH}{UZ(obT@jCpUy7Lsfvddc@3+=v>CP*Xy#+cFxXQ3K#V27w+$ioq5cMzhOQ3W4
ziq$J;uFq?LWXj#$87859VhH5c`&yZK_&flGjW762hd`F#<*;<Y#3SZ*$yR9=$;{fg
zYX@;RFZk$0znV8_u=AMD;~&_(#5Zh6516=UIBrk)LtT+)9H?5H%=3nG7{S?aw{o<R
z=(&Wi%%`RY91dCh9Zl?iX+{D+Ana*EakWH*pj0Bdu*9ymb#1!9QgO`X-f%D^?JMr`
z!3c0Cya(elI=rBRE<wsG-$@so(aw~ZJct1EQZg!nL<fdkt(Q8;dPPI;;v{(pN%zii
zU5oQixU&QbFSKrr%kb2_H*lB|0|B9$4N}g{C9Imx1cHSRUmdjv`knD{*r#<Fl|Z;*
zwZy66=s-Jo+H6f_<uk*>sN~yB|DMZ#8%~GNHtZv!^kw8&>#@=4b8MgUGQK2$tY4B(
zeMSRQ#)(CkwODtPN;&aMuKYg<ykP(6^McVWg)826n~>O#8{<(bCz6Gw&Xg0@&-f~J
zOK!`l{$iK^h?T0^RxR!kgP}>w-VVoF^H<vDqH;XIn-d;+(h4xCKZ)HqD!-o4n#EuK
z(*{+mSCtwOqKniQeg7R2mAxCdRZ4_4`6L`m7!toa1zW-tZ(qaThA97h_^qIE39A18
zbup_7-YrBsIU}ivw@+j6J0MbQ)tQ2f-;{PAqsRKpG<=&`?&`4c;tM>3(IKA=h_Iwm
z_BPZLfg5N{7Gd0X0ypUOb8Ozd0?$ffo$@_Cc38eS@dR$${t(~y13v??zvOC$vMiYz
z9YHWYq*$jX^eI+L@ut1j=g98{elZzO7TY7(cEqH|KjqD%+25*l2iX?oYMnua>oBbm
zW<g!y$ccghO>c3zR&hQk2`;4k<#y6)`{6737P`sJ?&g8GcRD<^S%U5Yg0p_4$ZdzD
z?Nti)+f5zO#!yt|`r%^dp?ckvNhp1G#gVX_G5*vjXU2?<WaL`ye_g!5se`Q{Cm;$P
zYqCMBza8CS{eNZXvxJ7Vje$LLR}rAj4w?p|Ro4NxMsV{xA?o=#M#(J6P?h!X<(_hD
z@9oAAxM4w{aQ?T(U%Z7PcKO{}*vfZz*zGd>AfCfjS-lOGklqdFN5$jnb+(YBf*+@s
zP2J$sl1!~Z);9g>nz?)%Y;L&}>VGYNhn<72I!Z+s$8<(TZ3aj}mcAnE3EIU42y$}B
z;L~`gZ*aw}(MpoRy$R1n+G^H4usDh;YJp3jo_5)8+ZN&y9);a?A^L2>1S1~1tY~h!
zu$(VkTs@_Ft1y%f`67-d(|>`mxTiUQmL7;e4yY7ro|K!IaB?`NsO5{uAor9WzeF2j
zBKxMVF-%_Mko=UgW2X~OuDf3Z9(nL+1%Aqqy$!vv_vvL#*cza%8#D@TdynPF%ODy0
zx>jGR*PI#Fd~XO}LC39jv!og?0ecYzIrB@PrNk7bPv9!#AZV7C7M-dyIRWHoHkrN;
z0?v$Q`2Md=eg05-Mann&@6zl%38A^hbERXVuRCSV)46q+f{DJwa`%WfX(mffe392b
zM*G;%*GHw?8%{3mZFvNK2SXf7DpHudP(SnKHVK3{_stUBx^%)v41$+Mm(7ZmNyK#I
zxRt7y@X7`C<h?e>4qE8upyZ-ftHde8_3zj{76>rH=K=UPO=k2kn_Ydno>g?IQcC-H
zM)fT0uHrqy-M9-MrwhMng$5>f(DD<l-(rfSORX0Zt~Z;4Qokv+LlS-o{B+@N*q!MG
zG9(CR4==)pZeZmC^0LJr`Z0P8i^9+@+&(46Jo3(`34Wd4kJZlyJUexB{^q=UEYF-q
z`PzSe<4L%%xfb>Um!4u5ay450oOl(F6;!X#yQq9KZIA0$3jK|?WB&6Z{Z5=MlkYee
zIFP^!rQDoR%4dbexG1-Vr{XKsr!uH`JbK}F{u)@h&}*gke!mF`{OgFL1RCi3z|i;1
zBNMELxBW^JIn-3btNJa<^PP4TOe8)#&nyakWP5I~7<YCEmpmj-`tU~EO!taTLquY%
zqjLCdPh%b&z1PwO_QBjzR&e-aa9X`NW<KnoK~LkLw49t2Z1^VEDFH*|XS6j;@@pOt
zV`qvuRI}uQ=81PA_55vSkkmy(Nd^0KV}CGhkkC8x<@8Ij6(^GO6VS}bzsUAv{i0f)
zWU()K7^#3c=Rwvo5W@^bj#Wi~l^s%xedRMpFJVtt>xz6KVX0|z_h0`L+`uCvE!hQK
zx_~C62QF(N00BBlDNftE;LSxs^#OAc2vBT<G$C@T-U~$LA7_(#XkILm>ZFB-;fp3O
zdz!)as=GJayCT!l0y&*>;@70Bt>cM<V`P*A6UE9=p|v$)2IrW%uRJY?YY-hxHO22+
z=oY7OQp+@*BB*=T3~iFT;A<V!@U76@Cb%fgalvP<TP){PBS7xq=oVkm-saOaX=f*l
z*vEIFa%qa`khw@b5=uV4$dRqjR4SzV!p`EmN}{)HCG_;wk0?ydx|4{-R}{`z;=#GW
zu;<C){oZT$?dv!{7*}W(F{}tc-=8c!t`OKk`2kf2*nD`}$>bbQ1^t#N{M?MuieMwb
z9b+eJm?bo2Bto?A8X`4V<zM96Invx2<D}S9_g7hP)#X(qIb#~WQ9mnon{lqgE>qQz
z=Os}V`;(eWGuD*#-{IcIv$G4)rAkB_W6Yj8y-W0vo?=3bgq6YaiOp<~CH|gkapzro
zw!hE?(r!?eg4VbW+N*F!O5JI&FmoNzQ5h<9bATUh$6}>Y874{!oA+pE91myAgmR+E
zyi4P(8F1k>0UoZbHLaJ<QOGd&B#V!bSsFV(7}350d>+g0^&YYBN?vFtFWAO<RFXZV
zKkmQ;@iNt+JI^V!N-Ot)Mq!qN2sAV{w=s{6v=0z)`(Sz({m?FWOrmjzVhy&{8h3+<
zV&AwDoJo+kqm!a=t%ZrKkYUvBp-OxNt-<tew7W@+skag3F&6LIsx0CpC^EngdVIVn
zIadT&7p8ghbKL_z&{)OMH50!wzV*4HW6DZ+*|+g@&Gj?*9A<|#^%Z#ty=Kydb!|v;
zII~!>Negk^kwsp23Ajh}C5~z$4-Pz}R6oCIYVRKNv0w+{i$n4)>AHzscl3K5r>>7N
zRa2*kv`yEZD!j}niC4^%-7>dF48S96V&;vj9fF88uaIqy`jLj>X$=QlcWsg@QR&53
z{bo#|bf%~`9Q_8y;aRbHe^Kd4_u3jKFKl~tz1mdJF@11x_JyQ6bgWC2yCX|r?ou0*
zG<$sy=0c4=7}jjx30)(-wt0Y=eA3W;DkW*IwoOZTP+{VdyfM4kNdKS_^_iZLHL)4d
zpkucs2(&h-0Rf~+TrzigdQbAG*pwWM@TY6Tx!70)0k#nY26rtUCsm?jNLTtZE24Ei
zT>fl*NpKb(NZJRIpDW;D1b<}DmkshJk+a0H!i03h5`K(fi09OyaxT^z&*#Eycx6!h
zEg|R+jdy!!(pq#HB+nha_JP0=)Ntw^*t;98n1(|Ju#^DOYO9J)XaG5e|FAmvSSfwy
z>L8~W;<*Xq7Yh~Qs<7C5n(@O)w`%x2So5n@$sr}dRDzecr<NmLZ3(&8h@R^OGH^!o
zDf@~MMGzl7qalGk>r;h`-u8yLhh>7sOVfH7N@T}bJYy{gbX78aaV_8}FJkMq%kaH}
z!|CKUK`!Gg7x_(cL5v{%;Rrue__;&0@x|b-h+%#|tosa9+VDZNJMFMyn*b1JY?uDq
zV^|X`<KD5+vm$aw{VaYWC{ikQbIcof&+n{mme<~;(VyPZI*s#V&*OU~YM$n)ou#l6
zHt9Xz&Im=RcqH^piJ1}P?A?v)jy-{Qb)AM;xXWJN^27zz%O4#SV7OA2G=DqQ#-8l}
z3Ji8FJSe(se&dSbbKkB-2Wx3+F<z8AP@|avq@h{S#eeBJmb+k{GAbh+jozfaFk8!M
zSxyPPq?RY1ND$;LewnQ(Qf-^_CC}@QC5(sFhl2vMyff#GSwq*_=7Pb};C(wpXHL#I
zqBq7t99_}e#OWk={+srtG~~E)Gbc*uvPRBB=9%wmPB>{uE0rtnakTMksd5EkVxuu-
z7!VlL@TOZ235==3`|>Orq|08uc8nOl+Zo1bdm_CnixTLT#2q<6iWh)yU7_F8PfPmJ
zs~N)cGB*bZ`&w=AWVD1IiWRU1+DOa_7|F~nm1wI(9xXLVapjGD*VdG_hi6#(FbaF&
zSN_>vO|P{Ljl2M^tJmKKSWpH~jNU&ujdA+nBIJ@m7(u)qT5p?o$$?bKmV}Md9YmpZ
z`oS=iaO>4jN(a*Aie?&+_9n5rTV8WQ!=!mZOG{HN#T#9rMY#nij<vaU0BkU{oSy@h
zLk9DGGpW@kMB!x7BzAE1Fd_Q9e((9<xvH=cX7<FI`?KhNP2OQ_art2EMVV{u3)uy<
zuLv$dpG&pyLOTFBw1e&`770tCPI<(Ml3v+siSKJWdUx)4J92%UKNK7eF%~{{BEh24
zRE1=Qu&Po@Jk_HOwU%~(vTeL{?>#y=A)|T$_2T(P+Y;@qlr5?P{qM+L+~PA&kZV9P
zqCHdrw9;pqGu}~7PjX-V9yYT7$sln{%UO1{_GOCOq#*0618+MNZ|mZh{97MuCohGj
z`6E(6<+_aSt%z@>o3`7z1?_e)mhaF!_x9@9Tkd-LEL!LPP9ViWMSN7HZ$b3l&ZrSa
zSfpk3^ZV<;Uk72Mo_Cg7^hzrpPVMpzrqqK!!0ZZ(cs8DReb&vq@wep`zhLc-833Ks
zKv%MWOiqD+eSM-9)~)}|vstbdR2}o8wHa$~#nN;rV@_^4@xTPmmg1o`ueoP%u^FZ*
zc86a5m%S*;k%RRe$+(&7S^CcWYYM+3y^^Zp<$RW}itYpn2{@r&etmu+^Dt!o<-86|
zW50tJN$cRhD@i4y$#%_)u9%~G+&ni_r4Rdtx!q<s;u|ggw7;IkWz`xf=Xrf2IK}sz
zlLe<*v1!4tfSIzy(j0`~75<#dC_~0<mL(;6d#y_k+P+K2+t56*W6RCAQCO9-lHg2D
zxKTpcU*NR%Wqlp-bLY<_Ax}1w4)=rZM|An6q}K%{oP~eELs}sr*}nOhTM7jkXRYxt
zzj;ONk>|W?A!=^EIv(mtsUQ8sNSY6?*D`1~G1sTxueFKD7Qt)XYzfhYoVsjaLX2sD
zzk8g7w6ycqs#Aa_;IA~G5E8@OT_MVEo{+WYoZ%u*2JLgUbhh$~y=pQZ{P+@N(f@>?
z=8t0>2+1##Rhngm3-zKz%s4E=Xq6m0OQs|0I_5$ZzxCP!!XGLB(h0j~h5P$jY%tjJ
zSiN0k#$%zNf&qUEWJi{@I}Hj=MxcsS%v+QJY5c^R;wzI}^fZX`>vYq6u^i_K=u2{=
zPPj++Tmvk&eWu!(T6f!v|H=Hk-qZBx2Bw)RHAX!Q^W~;DGw-+KT4*VSB0HtVoU2Ig
z^{BU|c+@wz>>Hjoos><b9tFRAA?_1@j$k)LWp5((MchrFWPuhV>~$vF-o#7|eqDv2
zbjl)R^~e8Rr-@jN`SgeB$z{tvU5hw$;KW*bKS-c(L(dc%F@3lD4_qbqN2Pr>QXeVY
zv-<Ko<J$MAlh%;Wv}Ez`C6x1i74?AdYQ2K&X6?K^QfQQMdRun!iWa^m(9_w{iL{yG
zEqR#hFMztTyGF8XcyJ^V?yOWtwr!d)I^wlzTq^g#)6h%L?yw*yw3ox4W-O3vWA3^{
z*3@L&zRap8$4urzEWUt$iq7w!CMCL*hr=SswFa2%J-=#VLh4-Gso{ml33@|CaMt{0
z>g9<Jb&<EjJf+@il85)APo+viPL%O|;fR428%an6?J_tjIxOhapdaY0&!!Sea#y_I
zF#knM^Z<82Lz;v(c*WRTl52Yu(RuSy<Ig1-v{1E6Ooih<J*T$0NUgZ3j0FWaK9~sV
zz4O$bwK+a3`WZT)j#6f7(?gNq%&_@&6sS_qQg$N*Ekh#DP<d+aHO(A`ah9c-Y^>Cr
z$?3o)CBe#){I6=%stde&un%lE&|w|`en|h~0xSE>1Cv_~$5OA_JqYIftnBeyn5%<0
zQ}N&MIDWIP90yXSHIdx1edq)1!V61S^AnEBU8Z%wXET#?gfSB@H@qivNk42nFocZ|
zKYrWD=|RxA^I&0SMMWYs2z6h(HqC>(oTnZzt9keh!r~-qYg6o8X&5~<(FIdAW{-gD
zU*!V@^(2IJRRl8W37b0cnJ6f(1eTnNNXR;c*XfS{d5g&o{651Zr<;`0e}H>tx@l;t
z9}zv1`~1`U{q#$VG*9Ao$^Pe2*juOW=7Uc_s=n=5CGEL6j{+oKmgtRN7dz!9Kk>tz
z6v5p!4rDM8iw)UHZXKY8rV^ST0G_Ydy7+4(AMy*Jxb^f&dO22#0PR8LZ=tlnhkj%{
zt|A}i8-EQ|El!L474;FZ^jD(Zh<-VFvo$+c<$gP3oBl;cusN46T0R7Rt`2NZzjG5N
z%#VrCf1&gam~_R}Y%|)Xn)(eS-<hdilM-(uE71f>RpN|@?g|5vjV#|YI!ucIz4U25
z;sfVCaZe>Wx@*D%%@so1Eaezx*s(`btk6n;Cx>FkS~2011LxVv!%_x6jdq`)NI}(%
z==9r%O8arD1VEP`wKG)&J6}JxDO+c<Uvo^0Qu8@<(#>;P{VfePSbDCX&&`j)C3P+m
z(atD}p4fM+-{O<piZco!p!9NQomtWBdCNHa7sdIA)oxqKuqs6crx|LqM(e&;s*6*l
zb10UbCoGEDdl&AChtbauSR7@4mg1SfE-|TO7pt~OT*+io(RZSN7yFS<7~uNX-CbYz
zNm!w1qdM66@#Rd_`An$BLu<3yuGq~r1v&RL%s<o1Y@!QkanEk-xm2=_gC)Vv8hK%P
z{v7;M1FT>@{vJ?A^LruN8fDRf<SieIPbsvZMq5m}27MkXszzoB7l4(5S4y$0fla!k
z)>9r_kQRAVyLWJ86Wfr}b1k(PvL)Yl(tB**y&>dK(x^@>exGxjbFBa2Eb9{Z>%Hjv
zs8FM$SOd(NOhp4xb%RG`j+>417m_9wHgIQCLd~~cd}6}gYYtLd$6B}M#wmZ)8^aDM
zpHGd?9%sV4uEo|Ja8>Y?bBOK^T_f+Xk_iHCs|!QpYR?iz7_$aholZ9TvM$<1q&w5%
z<j{PoUm>h8A%~JcR>bRz-+Ws@e}?1U=YOqHPl;b}{2xq|344|EJY{Og+(Yp(&N{N0
z)#d3i@_Jv+BHKILtQ!@7vU)(nEaugV0hr=OstPFifiu+X9@NSCk)WhzXX>q=-rgI5
zB29YG=W#Pz7L^E#dKhjEbnfx(Se*#NWt*l72Fk~m%M_+~cD^J^y*x~l_`K`q?U&|Z
z<%(seEq<#PnUErdR;$OF0?#o{uTchXtY>NpSpk=@YX1|EucxYA8*A^uNbBXi%&s@V
zyc8X90>4WL#PXTmps-=&`+M<&a9T`WdwjDbL&aYs4}M%{Vsd#p-tVcr*VE;bflcfU
z(G*zW>q#>eS~Ll@>8FK{1L8Jj2O!-9vV<R)bXK+*D&r=XhU6z}v&9AkD9EYC<)~wE
zTet7vtV1S&Ina@c4`;{)?U$T+oh22T{W>AOM>|xynR=T~pkAJ))cmR|dvTKX#8aS^
zC~r=`=Ql)fr?PzOW5{(C8QnT7<pTvm&by*Hr;FY=yn&e+m8F<}9_rq3bDHNj#7|Ha
zdMh>!KwN8Qez>iFg}^DKPsgheI+<)e@9Q#LSLI-Lj_a)<&9I)#3fA!QA$2WbT(q%I
zoWL0)KaIRAIp5pR+f2^1hMW&7MP0wjprE8YbkbeCJ8rO4&`s#o%~w^LG+v%guK86k
zk79^+iN?2+BeB=%@@`{Z<;h*dZopo~_U~cJ{9uPD&FvJM%z6n6xE;9v8>7!QmcC8P
zucO<jkJ96B#f}{8Nui#<zi0l%SzWAGyNk6$G+Ntto}L=#QWpz<IFQ-LbNPIAQ1bJ-
z_*|l*Y9{uQB-w#ES`||m1x&=@+l#zP1MEdg|2-$U{HH)(n00jGJJ2OUVI2}w)Uk;T
z#t$;@RsOPGSh>;SM@7Mc;)CeqP7Dfr>#R(99g)dDLwGK#!*j<Yz>k099AxJpeYWhb
z!|F2leT2;0#9+}lR8OtFb3ii$l7&WhdJ4?<ILa>aVn28Xn;Y&j%0*tr#)WFtkN>zd
zkW$;wYvrCvJXtmAQKHu!q)<Jzn9)!-jTAXdttHy4B=~0!@AXgbmh|C^8^7$R?ll8O
z3EiUj(XJt`G%d;rya@uA>|)MO3KMlDv7o4SzRO6N@DWg+c#3;G|0<@}eIPO5zT>%8
zddC!X9ok`UwC<#Ovu?3_urbHWBAf4ug4gn9bw>mZR!Gmv+^AT{FOVyI98J_&QPXC)
z^s_|QYY%8v4QfLPZJd26axwx0h+BU-h3}20YZ{%_;B~MBdRPUPkCDw5j38_K6|=;G
zEBI1idht#52^W+7R!Q+qjxzXY1rew|y?6oT$D5Ds7iwy{FJ6*g<6zR>ZOUHG$9Za#
zLqh#+)-;+IT?53g_!pSAQwCcytX5ad(^K=O`}8pPAtQC59CngB19>HxH{R$1f1*AW
z1H!i}jkR4_d%=#C$*pt&+ff43?boJq>a-VLRdddF*&uzc1$R_RGbDiP7DCFB^v_Q4
zH4nAlk*iBHym^_T%frn=0eMC$4%6;uP7W?_oRu0XyO8LB4GdvM=~dj&;=Z-tW=bx2
zuqQgK_=6!?#2!N`(W%yC1G~7s_y})#CioLTZk~~29mJ2O0o9KKK(R7Nb5r~kr!J=G
z;w%=H%s72<fzr*G4K(aiP8LXTO*@Egk(6S7-AMk$dQGO|3_R$@g4EK&gJEz5`ux>e
z)-EuZ>7djf_^}$?)zwKJ=@?6?h-jtj{015XEO$vf(t98=>#pu+`VEuHhQ~L)Du@VU
z75__%->z<e_{z@kBpTk+(=7^JnK26)ldWPVZZA_*kC~pp?^GruwKg0o_UjGB<SUc?
zxF%)qvC_RGoi_O2=z3a#O4r-S_NMY<HuOFajEkO**5*+(ZQ#rjy@DJcjO<xVl?K-y
zXi(_BG*Y<8Pa9Gx?9U_&nkkQE^!itEs<gxLuM||RwjWPz{q5B)<*amH)*Wtg+N~5-
zl~A>O_BMLZ*P<j#H%09(#eTbiQE7sx?O%SC2juzz-ARw<fLxn(x<)bk3BL9DFsV+i
z3_AIBmZz>U$YLMcYGG1bc2K8tz|Zb7W_Y@e3K_?_a>C%nRsA5=7>k;*FO#b8Bx5bQ
z7o{X8=vdp!7!8$EH%yfmT{Ag~fBal2hqY%4s1b%!oVsg|e>-mrLi+4D7n=i~PhyWH
z1$Qk5kR{H9ZJqeDu!?BEDM!c$g8^k3Id=j;fd&^~p0YCR$g_P@eW5+ziTT_B@@!fv
zgpE7e8y!X5KM_%8@_F#Fb{M5|aWgp!Ih=ZR%k^@4G3nZvA$2j?l7=w2J8zdTP-vzj
z6}zxQ1=`*o?b`ZGr#?=uj}2!jptb78JwRDf%MzO)cBYzY+@k~GOy7nN<Th{ky+(7=
z8~sR&8JJgZai{W|FJU`+%3V*Lg%e7dm)SXmc^Ha%TK5NzTwNhYdd?4|o=2Tu)2<n{
zIAgFN<ACuE*P58(B8^ry+mSfs?yN%zu~`TLdP-C7Nu>`g=f@~Z!rV|)<!ga^774iH
zxwluULhUYg;Hb1I$n4&n@&+xd50~Ob8O{>11p}H9_{IUaD=%^<PGQ?DR1Hr2B^g{?
zVT5+fnEVcFZIQY3j4euBVV?3^UrBSr()OoJ`RxSySU;7|j2>!8Y4hZhfwJoB@gVlb
z0ZL({#fcst5HMcQ^i1~TyKK{5ZE>#?Y4+jFfi_2A(w5-6X-6bqgkS;LWM3p`52d)!
zsmXgzUm}M@MF~Hb%$YV<-P*9Gv0nEbpfC{<QlIx2BN8z70W^<;eE8lbe=gr0A0ou-
z1MC$qG#^G<WF^K;bof7mRVwI8GfoaUUgxd*TK6h!1lt%rrMu;+8JO13Pm6uYsDrKy
zpVta~{2PTn%0cg+v{pF*q_a%omv036Oz6J^`dagsRpL;D#=(ntVRRK|Jt>7AyopkX
z8BJycff8KKx++pVmAx{u+ZP)ZV=)5FWm9OfcTdq3UBOIfXCG@Aca)*;s$q2kTl>f{
z`csoi_ko>clnoVn=R2wea;8Yut^Tp)l|N>y_6Em_T;ylg`>;6Yis&*mn1WKqdc=_G
zv`ng$(ti<g)p+E*swDbpl!#ZVQM4$W$Ljgro@`&Q-bfeYqjCYMOUFKt&p$EgL1Wor
zb~T7Vy+^h5R~aQ5!)mg5ZKv|wn^N>dQHJ9RSFaFUOnp%j34G}7Cid1U!TYM^bL+{l
z&y0X}tCf4;F4_`m4f%NZp37EOmmSWZ0ezgz-r5dto}Zerh|3fs$zXH3Z8qkmx=wr$
z0Xmp;BFM#cW|`cV9F54jKe0N{vRhv_9$t=0WM;dLzT%Xa+(%b*-VIeYiEDvYP)-F*
z=?;`GB)39AYNWTz{<`c<G(F)7gC7by;Zp`KN5Psrxw^Y1^vywnv1Zrjp%4}8TtYhF
z);T05vvS=9@Emp-(brdTd9$Vdz)z;u79RDJu<nV?-<mx^fh`h3N|{O8jO9!sKS5C{
z$bTT}#9bvsaCMM8w%^<Z@d9~%Xz2e-Q2r{cSv&aN3Sg8ZLqeVTA=G*0zxK$k6;53o
zcU$Mwi)5K8APsC1xw=mM50P1jo!PEYMo#cD6zuiEi-&Ge#Pa>;;1$F>lsZGNCp<-~
zEx~J(Q5wX@fQTnV#6&Ioi<tV&4+9d>{j0Q27j@F_V~ZD$&n=3h5yYJ-#BSv`2qGRB
z(;>P_TFd{k(Z4L683;5|HFK+5J6|h&(kUNJiHHOQAPhGtt6exW9P?Vt{7$-IIrc9U
ziXpKZ8hucUhsnoD<oWP-X+>|&&~v42(SoaGj#p1Xn#)4i97~+Jl-%wAF5_>m@SnHd
zQUd>_TmJKzf0pGRr2K<|e^Brb3jRUCKPdPI1^=Mn9~At9f`3r(4+{Q4!GEA2r1srK
aV$2ip-0sZfdzJxzh{N88D!&Z*{{I2&Ho|WJ

literal 0
HcmV?d00001

diff --git a/doc/design/fluid.md b/doc/design/fluid.md
new file mode 100644
index 0000000000..b3d039d6db
--- /dev/null
+++ b/doc/design/fluid.md
@@ -0,0 +1,122 @@
+# Design Doc: PaddlePaddle Fluid
+
+## Why Fluid
+
+When Baidu developed PaddlePaddle in 2013, the only well-known open source deep learning system was Caffe.  However, when it open-sourced PaddlePaddle in 2016, there had been many other choices over there.  We were facing a challenge -- why would we open source yet another one?
+
+Fluid is the answer.  Fluid is similar to PyTorch and TensorFlow Eager Execution, which describes the "process" of training or inference a model, but not the model itself.  Indeed, in PyTorch, Eager Execution, and Fluid, there is no such a concept of the model at all. I will explain in this article, Fluid is currently more extreme in this idea than PyTorch and Eager Execution, and we are pushing Fluid towards a compiler and even a new programming language for deep learning
+
+## The Evolution of Deep Learning Systems
+
+Deep learning infrastructure is one of the fastest involving technology.  Within only four years, there have been three generations of technologies invented. 
+
+| Since around | model = sequence of layers | model = graph of operators | No model |
+|--|--|--|--|
+| 2013 | Caffe, Theano, Torch, PaddlePaddle | | |
+| 2015 | | TensorFlow, MxNet, Caffe2, ONNX, n-graph | |
+| 2016 | | | PyTorch, TensorFlow Eager Execution, PaddlePaddle Fluid |
+
+From the above table, we see that the technology is evolving towards the removal of the concept of the model.  To better understand the reason, let us compare the *programming paradigms*, or, the ways we program deep learning applications using these systems.
+
+## Deep Learning Programming Paradigms
+
+With any system listed as the first or second generation, e.g., Caffe or TensorFlow, an AI application training program looks like the following:
+
+```python
+x = layer.data("image")
+l = layer.data("label")
+f = layer.fc(x, W)
+s = layer.softmax(f)
+c = layer.mse(l, s)
+
+for i in xrange(1000): # train for 1000 iterations
+    m = read_minibatch()
+    forward({input=x, data=m}, minimize=c)
+    backward(...)
+    
+print W # print the trained model parameters.
+```
+
+The above program includes two parts:
+
+1. the first part describes the model, and
+2. the second part describes the training process (or inference process).
+
+This paradigm has a well-known problem that limits programmers' productivity.  Suppose that we made some mistakes at configuring the model in the first part of the program, when we run the program, it wouldn't prompt error messages until the execution enters the second part, when the invocation to `forward` or `backward` raise errors.   It is difficult for the programmer to realize and locate that there is a mistake many lines away from where the error appears.
+
+This problem of hard to debug a program is the primary reason that programmers prefer PyTorch than elder systems.  Using PyTorch, we would write the above program like the following
+
+```python
+W = tensor(...)
+
+for i in xrange(1000): # train for 1000 iterations
+    m = read_minibatch()
+    x = m["image"]
+    l = m["label"]
+    f = layer.fc(x, W)
+    s = layer.softmax(f)
+    c = layer.mse(l, s)
+    backward()
+    
+print W # print the trained model parameters.
+```
+
+We can see that the main difference is the moving of the model configuration, the first part, into the train loop.  This change would allow that mistakes in model configuration reported where they appear.  This change also represents the model, or its forward pass, by the process in the training loop.
+
+## Describe Arbitrary Models for the Future
+
+Describing the process instead of the model also brings Fluid the flexibility to define models not yet invented.
+
+As we can program the process, we can write an RNN as a loop, instead of an RNN layer or operator.  A PyTorch example could look like
+
+```python
+for i in xrange(1000):
+    m = read_minibatch()
+    x = m["sentence"]
+    for t in xrange x.len():
+        h[t] = the_step(x[t])
+```        
+
+With Fluid, the training loop and the RNN in the above program are not Python loop, but a "loop structure" provided by Fluid and implemented in C++:
+
+```python
+train_loop = layers.While(cond)
+with train_loop.block():
+  m = read_minibatch()
+  x = m["sentence"]
+  rnn = layers.While(...)
+  with rnn.block():
+    h[t] = the_step(input[t])
+```    
+
+A real Fluid example is [here](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44).
+
+From these examples, you can see that Fluid programs look similar to their PyTorch equivalent, except that Fluid's loop structure, wrapped with Python's `with` statement, could run much faster than Python's loop.
+
+We have more examples of the [`if-then-else`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/if_else_op.md) structure of Fluid.
+
+## Turing Completeness
+
+In computability theory, a system of data-manipulation rules, such as a programming language, is said to be Turing complete if it can be used to simulate any Turing machine.  For a programming language, if it provides if-then-else and loop, it is Turing complete.  From above examples, Fluid seems Turing complete; however, I would like to point out is a slight difference between the if-then-else of Fluid and that in a programming language is that the former runs both of its branches.  It splits the input minibatch into two -- one for the true condition and one for the false.  I am not sure if this is equivalent to the if-then-else that makes programming languages Turing-complete.  I talked with [Yuang Yu](https://research.google.com/pubs/104812.html), but I need to figure out more.
+
+## The Execution of a Fluid Program
+
+There are two ways to run a Fluid program.  When we run an example program, it creates a protobuf message [`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/paddle/framework/framework.proto#L145) that describes the process and conceptually likes an [abstract syntax tree](https://en.wikipedia.org/wiki/Abstract_syntax_tree).
+
+We have a C++ class [`Executor`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h), which runs a `ProgramDesc` like that an interpreter runs a Python program.
+
+We are moving towards a compiler, which we will explain in more details later in this article.
+
+## Backward Compatibility
+
+Given all advantages from the removal of the concept *model*, hardware manufacturers might still prefer the existence of the concept model, so they could build their hardware reads and runs a trained model for inference.  For example, Nervana, a startup company acquired by Intel, has been working on an XPU that reads models in the format known as [n-graph](https://github.com/NervanaSystems/ngraph).  Similarly, [Movidius](https://www.movidius.com/) is producing a mobile deep learning chip that reads and runs graphs of operators too.  The well-known [ONNX](https://github.com/onnx/onnx) is also a file format of graphs of operators.
+
+For Fluid, we can write a converter that extracts parts in the `ProgramDesc` protobuf message, converts them into a graph of operators, and exports into the ONNX or n-graph format.
+
+## Towards a Deep Learning Language and the Compiler
+
+We can change the if-then-else and loop structure a little bit in the above Fluid example programs so to make it a new programming language, different from Python.
+
+Even if we don't invent a new language, as long as we get the `ProgramDesc` message filled in, we can write a transpiler, which translates each invocation to an operator into a C++ call to a kernel function of that operator. For example, a transpiler that weaves the CUDA kernels outputs an NVIDIA-friendly C++ program, which can be built using `nvcc`.  Another transpiler could generate MKL-friendly code that should be built using `icc` from Intel.  More interestingly, we can translate a Fluid program into its distributed version of two `ProgramDesc` messages, one for running on the trainer process, and the other one for the parameter server.  For more details of the last example, let us check the [concurrent programming design](concurrent_programming.md).  The following figure explains this two-stage process:
+
+![](fluid-compiler.png)

From 4d93f0fc1a5708c9d10cb17d161a7cc7f0391297 Mon Sep 17 00:00:00 2001
From: Abhinav Arora <abhinavarora28@gmail.com>
Date: Thu, 14 Dec 2017 07:01:05 +0530
Subject: [PATCH 52/94] Polish the support new device doc (#6594)

---
 doc/design/support_new_device.md | 40 ++++++++++++++++----------------
 1 file changed, 20 insertions(+), 20 deletions(-)

diff --git a/doc/design/support_new_device.md b/doc/design/support_new_device.md
index 92443e4392..fd23dc211a 100644
--- a/doc/design/support_new_device.md
+++ b/doc/design/support_new_device.md
@@ -1,33 +1,33 @@
-# Design Doc: Support new Device/Library
+# Design Doc: Supporting new Device/Library
 
 ## Background
 
-Deep learning has a high demand for computing resources. New high-performance device and computing library are coming constantly. The deep learning framework has to integrate these high-performance device and computing library flexibly.
+Deep learning has a high demand for computing resources. New high-performance devices and computing libraries are appearing very frequently. Deep learning frameworks have to integrate these high-performance devices and computing libraries flexibly and efficiently.
 
-On the one hand, hardware and computing library are not usually one-to-one coresponding relations. For example, in Intel CPU, there are Eigen and MKL computing library. And in Nvidia GPU, there are Eigen and cuDNN computing library. We have to implement specific kernels for an operator for each computing library.
+On one hand, hardware and computing libraries usually do not have a one-to-one correspondence. For example,Intel CPUs support Eigen and MKL computing libraries while Nvidia GPUs support Eigen and cuDNN computing libraries. We have to implement operator specific kernels for each computing library.
 
-On the other hand, users usually do not want to care about the low-level hardware and computing library when writing a neural network configuration. In Fluid, `Layer` is exposed in `Python`, and `Operator` is exposed in `C++`. Both `Layer` and `Operator` are independent on hardwares.
+On the other hand, users usually do not want to care about the low-level hardware and computing libraries when writing a neural network configuration. In Fluid, `Layer` is exposed in `Python`, and `Operator` is exposed in `C++`. Both `Layer` and `Operator` are hardware independent.
 
 So, how to support a new Device/Library in Fluid becomes a challenge.
 
 
 ## Basic: Integrate A New Device/Library
 
-For a general overview of fluid, please refer to [overview doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/read_source.md).
+For a general overview of fluid, please refer to the [overview doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/read_source.md).
 
-There are mainly there parts we have to consider in integrating a new device/library:
+There are mainly three parts that we have to consider while integrating a new device/library:
 
 - Place and DeviceContext: indicates the device id and manages hardware resources
 
 - Memory and Tensor: malloc/free data on certain device
 
-- Math Functor and OpKernel: implement computing unit on certain device/library
+- Math Functor and OpKernel: implement computing unit on certain devices/libraries
 
 ### Place and DeviceContext
 
 
 #### Place
-Fluid use class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent specific device and computing library. There are inheritance relationships between different kinds of `Place`.
+Fluid uses class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent different devices and computing libraries. There are inheritance relationships between different kinds of `Place`.
 
 ```
         |   CPUPlace   --> MKLDNNPlace
@@ -43,7 +43,7 @@ typedef boost::variant<CUDAPlace, CPUPlace, FPGAPlace> Place;
 
 #### DeviceContext
 
-Fluid use class [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L30) to manage the resources in certain hardware, such as CUDA stream in `CDUADeviceContext`. There are also inheritance relationships between different kinds of `DeviceContext`.
+Fluid uses class [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L30) to manage the resources in different hardwares, such as CUDA stream in `CDUADeviceContext`. There are also inheritance relationships between different kinds of `DeviceContext`.
 
 
 ```
@@ -52,7 +52,7 @@ DeviceContext ---->  CUDADeviceContext  --> CUDNNDeviceContext
                 \->  FPGADeviceContext
 ```
 
-A example of Nvidia GPU is as follows:
+An example of Nvidia GPU is as follows:
 
 - DeviceContext
 
@@ -93,7 +93,7 @@ class CUDNNDeviceContext : public CUDADeviceContext {
 
 #### memory module
 
-Fluid provide following [memory interfaces](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/memory/memory.h#L36):
+Fluid provides the following [memory interfaces](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/memory/memory.h#L36):
 
 ```
 template <typename Place>
@@ -106,12 +106,12 @@ template <typename Place>
 size_t Used(Place place);
 ```
 
-To implementing these interfaces, we have to implement MemoryAllocator for specific Device
+To implementing these interfaces, we have to implement MemoryAllocator for different Devices
 
 
 #### Tensor
 
-[Tensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h#L36) holds data with some shape in certain Place.
+[Tensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h#L36) holds data with some shape in a specific Place.
 
 ```cpp
 class Tensor {
@@ -168,7 +168,7 @@ t.mutable_data(place);
 
 ### Math Functor and OpKernel
 
-Fluid implements computing unit based on different DeviceContext. Some computing unit is shared between operators. These common part will be put in operators/math directory as basic Functors.
+Fluid implements computing units based on different DeviceContexts. Some computing units are shared between operators. This common part will be put in operators/math directory as basic Functors.
 
 Let's take [MaxOutFunctor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/math/maxouting.h#L27) as an example:
 
@@ -183,7 +183,7 @@ class MaxOutFunctor {
 };
 ```
 
-CPU implement in .cc file
+CPU implemention is in .cc file
 
 ```
 template <typename T>
@@ -197,7 +197,7 @@ class MaxOutFunctor<platform::CPUDeviceContext, T> {
 };
 ```
 
-CUDA implement in .cu file
+CUDA implemention is in .cu file
 
 ```
 template <typename T>
@@ -212,11 +212,11 @@ class MaxOutFunctor<platform::CUDADeviceContext, T> {
 ```
 
 
-We get computing handle from concrete DeviceContext, and make compution on tensors.
+We get computing handle from a concrete DeviceContext, and make compution on tensors.
 
-The implement of `OpKernel` is similar to math functors, the extra thing we need to do is registering the OpKernel to global map.
+The implemention of `OpKernel` is similar to math functors, the extra thing we need to do is to register the OpKernel in a global map.
 
-Fluid provides different register interface in op_registry.h
+Fluid provides different register interfaces in op_registry.h
 
 
 Let's take [Crop](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/crop_op.cc#L134) operator as an example:
@@ -240,7 +240,7 @@ REGISTER_OP_CUDA_KERNEL(
 
 ## Advanced topics: How to switch between different Device/Library
 
-Generally, we will impelement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not sutibale in a specific Device. For example, crf operator can be only run at CPU, whereas most other operators can be run at GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library.
+Generally, we will impelement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not sutibale on a specific Device. For example, crf operator can only run on CPU, whereas most other operators can run at GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library.
 
 
 We will discuss how to implement an efficient OpKernel switch policy. 

From c5afc3fe7cf1f1f93399cb9d6c92ed6f86a5fc74 Mon Sep 17 00:00:00 2001
From: kavyasrinet <kavyasrinet@baidu.com>
Date: Wed, 13 Dec 2017 18:40:30 -0800
Subject: [PATCH 53/94] Updating the design doc of Fluid (#6597)

---
 doc/design/fluid.md | 54 ++++++++++++++++++++++-----------------------
 1 file changed, 27 insertions(+), 27 deletions(-)

diff --git a/doc/design/fluid.md b/doc/design/fluid.md
index b3d039d6db..585dc8ef39 100644
--- a/doc/design/fluid.md
+++ b/doc/design/fluid.md
@@ -2,25 +2,25 @@
 
 ## Why Fluid
 
-When Baidu developed PaddlePaddle in 2013, the only well-known open source deep learning system was Caffe.  However, when it open-sourced PaddlePaddle in 2016, there had been many other choices over there.  We were facing a challenge -- why would we open source yet another one?
+When Baidu developed PaddlePaddle in 2013, the only well-known open source deep learning system at the time was Caffe.  However, when PaddlePaddle was open-sourced in 2016, many other choices were available. There was a challenge -- what is the need for open sourcing yet another deep learning framework?
 
-Fluid is the answer.  Fluid is similar to PyTorch and TensorFlow Eager Execution, which describes the "process" of training or inference a model, but not the model itself.  Indeed, in PyTorch, Eager Execution, and Fluid, there is no such a concept of the model at all. I will explain in this article, Fluid is currently more extreme in this idea than PyTorch and Eager Execution, and we are pushing Fluid towards a compiler and even a new programming language for deep learning
+Fluid is the answer.  Fluid is similar to PyTorch and TensorFlow Eager Execution, which describes the "process" of training or inference using the concept of a model.  In fact in PyTorch, TensorFlow Eager Execution and Fluid, there is no  concept of a model at all. The details are covered in the sections below. Fluid is currently more extreme in the above mentioned idea than PyTorch and Eager Execution, and we are trying to push Fluid towards the directions of a compiler and a new programming language for deep learning.
 
 ## The Evolution of Deep Learning Systems
 
-Deep learning infrastructure is one of the fastest involving technology.  Within only four years, there have been three generations of technologies invented. 
+Deep learning infrastructure is one of the fastest evolving technologies. Within four years, there have already been three generations of technologies invented.
 
-| Since around | model = sequence of layers | model = graph of operators | No model |
+| Existed since | model as sequence of layers | model as graph of operators | No model |
 |--|--|--|--|
 | 2013 | Caffe, Theano, Torch, PaddlePaddle | | |
 | 2015 | | TensorFlow, MxNet, Caffe2, ONNX, n-graph | |
 | 2016 | | | PyTorch, TensorFlow Eager Execution, PaddlePaddle Fluid |
 
-From the above table, we see that the technology is evolving towards the removal of the concept of the model.  To better understand the reason, let us compare the *programming paradigms*, or, the ways we program deep learning applications using these systems.
+From the above table, we see that the deep learning technology is evolving towards getting rid of the concept of a model.  To understand the reasons behind this direction, a comparison of the *programming paradigms* or the ways to program deep learning applications using these systems, would be helpful. The following section goes over these.
 
 ## Deep Learning Programming Paradigms
 
-With any system listed as the first or second generation, e.g., Caffe or TensorFlow, an AI application training program looks like the following:
+With the systems listed as the first or second generation, e.g., Caffe or TensorFlow, an AI application training program looks like the following:
 
 ```python
 x = layer.data("image")
@@ -33,18 +33,18 @@ for i in xrange(1000): # train for 1000 iterations
     m = read_minibatch()
     forward({input=x, data=m}, minimize=c)
     backward(...)
-    
+
 print W # print the trained model parameters.
 ```
 
 The above program includes two parts:
 
-1. the first part describes the model, and
-2. the second part describes the training process (or inference process).
+1. The first part describes the model, and
+2. The second part describes the training process (or inference process) for the model.
 
-This paradigm has a well-known problem that limits programmers' productivity.  Suppose that we made some mistakes at configuring the model in the first part of the program, when we run the program, it wouldn't prompt error messages until the execution enters the second part, when the invocation to `forward` or `backward` raise errors.   It is difficult for the programmer to realize and locate that there is a mistake many lines away from where the error appears.
+This paradigm has a well-known problem that limits the productivity of programmers. If the programmer made a mistake in configuring the model, the error messages wouldn't show up until the second part is executed and `forward` and `backward` propagations are performed. This makes it difficult for the programmer to debug and locate a mistake that is located blocks away from the actual error prompt.
 
-This problem of hard to debug a program is the primary reason that programmers prefer PyTorch than elder systems.  Using PyTorch, we would write the above program like the following
+This problem of being hard to debug and re-iterate fast on a program is the primary reason that programmers, in general,  prefer PyTorch over the older systems.  Using PyTorch, we would write the above program as following:
 
 ```python
 W = tensor(...)
@@ -57,17 +57,17 @@ for i in xrange(1000): # train for 1000 iterations
     s = layer.softmax(f)
     c = layer.mse(l, s)
     backward()
-    
+
 print W # print the trained model parameters.
 ```
 
-We can see that the main difference is the moving of the model configuration, the first part, into the train loop.  This change would allow that mistakes in model configuration reported where they appear.  This change also represents the model, or its forward pass, by the process in the training loop.
+We can see that the main difference is the moving the model configuration part (the first step) into the training loop.  This change would allow the mistakes in model configuration to be reported where they actually appear in the programming block.  This change also represents the model better, or its forward pass, by keeping the configuration process in the training loop.
 
 ## Describe Arbitrary Models for the Future
 
-Describing the process instead of the model also brings Fluid the flexibility to define models not yet invented.
+Describing the process instead of the model also brings Fluid, the flexibility to define different non-standard models that haven't been invented yet.
 
-As we can program the process, we can write an RNN as a loop, instead of an RNN layer or operator.  A PyTorch example could look like
+As we write out the program for the process, we can write an RNN as a loop, instead of an RNN as a layer or as an operator.  A PyTorch example would look like the following:
 
 ```python
 for i in xrange(1000):
@@ -77,7 +77,7 @@ for i in xrange(1000):
         h[t] = the_step(x[t])
 ```        
 
-With Fluid, the training loop and the RNN in the above program are not Python loop, but a "loop structure" provided by Fluid and implemented in C++:
+With Fluid, the training loop and the RNN in the above program are not really Python loops, but just a "loop structure" provided by Fluid and implemented in C++ as the following:
 
 ```python
 train_loop = layers.While(cond)
@@ -89,34 +89,34 @@ with train_loop.block():
     h[t] = the_step(input[t])
 ```    
 
-A real Fluid example is [here](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44).
+An actual Fluid example is described  [here](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44).
 
-From these examples, you can see that Fluid programs look similar to their PyTorch equivalent, except that Fluid's loop structure, wrapped with Python's `with` statement, could run much faster than Python's loop.
+From the example, the Fluid programs look very similar to their PyTorch equivalent programs, except that Fluid's loop structure, wrapped with Python's `with` statement, could run much faster than just a Python loop.
 
 We have more examples of the [`if-then-else`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/if_else_op.md) structure of Fluid.
 
 ## Turing Completeness
 
-In computability theory, a system of data-manipulation rules, such as a programming language, is said to be Turing complete if it can be used to simulate any Turing machine.  For a programming language, if it provides if-then-else and loop, it is Turing complete.  From above examples, Fluid seems Turing complete; however, I would like to point out is a slight difference between the if-then-else of Fluid and that in a programming language is that the former runs both of its branches.  It splits the input minibatch into two -- one for the true condition and one for the false.  I am not sure if this is equivalent to the if-then-else that makes programming languages Turing-complete.  I talked with [Yuang Yu](https://research.google.com/pubs/104812.html), but I need to figure out more.
+In computability theory, a system of data-manipulation rules, such as a programming language, is said to be Turing complete if it can be used to simulate any Turing machine.  For a programming language, if it provides if-then-else and loop, it is Turing complete.  From the above examples, Fluid seems to be Turing complete; however, it is noteworthy to notice that there  is a slight difference between the `if-then-else` of Fluid and that of a programming language. The difference being that the former runs both of its branches and splits the input mini-batch into two -- one for the True condition and another for the False condition. This hasn't been researched in depth if this is equivalent to the `if-then-else` in programming languages that makes them Turing-complete.  Based on a conversation with [Yuang Yu](https://research.google.com/pubs/104812.html), it seems to be the case but this needs to be looked into in-depth.
 
 ## The Execution of a Fluid Program
 
-There are two ways to run a Fluid program.  When we run an example program, it creates a protobuf message [`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/paddle/framework/framework.proto#L145) that describes the process and conceptually likes an [abstract syntax tree](https://en.wikipedia.org/wiki/Abstract_syntax_tree).
+There are two ways to execute a Fluid program.  When a program is executed, it creates a protobuf message [`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/paddle/framework/framework.proto#L145) that describes the process and is conceptually like an [abstract syntax tree](https://en.wikipedia.org/wiki/Abstract_syntax_tree).
 
-We have a C++ class [`Executor`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h), which runs a `ProgramDesc` like that an interpreter runs a Python program.
+There is a C++ class [`Executor`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h), which runs a `ProgramDesc`, similar to how an interpreter runs a Python program.
 
-We are moving towards a compiler, which we will explain in more details later in this article.
+Fluid is moving towards the direction of a compiler, which is explain in more detail later in this article.
 
-## Backward Compatibility
+## Backward Compatibility of Fluid
 
-Given all advantages from the removal of the concept *model*, hardware manufacturers might still prefer the existence of the concept model, so they could build their hardware reads and runs a trained model for inference.  For example, Nervana, a startup company acquired by Intel, has been working on an XPU that reads models in the format known as [n-graph](https://github.com/NervanaSystems/ngraph).  Similarly, [Movidius](https://www.movidius.com/) is producing a mobile deep learning chip that reads and runs graphs of operators too.  The well-known [ONNX](https://github.com/onnx/onnx) is also a file format of graphs of operators.
+Given all the advantages from the removal of the concept of a *model*, hardware manufacturers might still prefer the existence of the concept of a model, so it would be easier for them to support multiple frameworks all at once and could run a trained model during inference.  For example, Nervana, a startup company acquired by Intel, has been working on an XPU that reads the models in the format known as [n-graph](https://github.com/NervanaSystems/ngraph).  Similarly, [Movidius](https://www.movidius.com/) is producing a mobile deep learning chip that reads and runs graphs of operators.  The well-known [ONNX](https://github.com/onnx/onnx) is also a file format of graphs of operators.
 
-For Fluid, we can write a converter that extracts parts in the `ProgramDesc` protobuf message, converts them into a graph of operators, and exports into the ONNX or n-graph format.
+For Fluid, we can write a converter that extracts the parts in the `ProgramDesc` protobuf message, converts them into a graph of operators, and exports the graph into the ONNX or n-graph format.
 
 ## Towards a Deep Learning Language and the Compiler
 
-We can change the if-then-else and loop structure a little bit in the above Fluid example programs so to make it a new programming language, different from Python.
+We can change the `if-then-else` and loop structure a little bit in the above Fluid example programs, to make it into a new programming language, different than Python.
 
-Even if we don't invent a new language, as long as we get the `ProgramDesc` message filled in, we can write a transpiler, which translates each invocation to an operator into a C++ call to a kernel function of that operator. For example, a transpiler that weaves the CUDA kernels outputs an NVIDIA-friendly C++ program, which can be built using `nvcc`.  Another transpiler could generate MKL-friendly code that should be built using `icc` from Intel.  More interestingly, we can translate a Fluid program into its distributed version of two `ProgramDesc` messages, one for running on the trainer process, and the other one for the parameter server.  For more details of the last example, let us check the [concurrent programming design](concurrent_programming.md).  The following figure explains this two-stage process:
+Even if we do not invent a new language, as long as we get the `ProgramDesc` message filled in, we can write a transpiler, which translates each invocation to an operator, into a C++ call to a kernel function of that operator. For example, a transpiler that weaves the CUDA kernels outputs an NVIDIA-friendly C++ program, which can be built using `nvcc`.  Another transpiler could generate MKL-friendly code that should be built using `icc` from Intel.  More interestingly, we can translate a Fluid program into its distributed version of two `ProgramDesc` messages, one for running on the trainer process, and the other one for the parameter server.  For more details of the last example, the [concurrent programming design](concurrent_programming.md) document would be a good pointer.  The following figure explains the proposed two-stage process:
 
 ![](fluid-compiler.png)

From 97c3de0cfbf150670317e289e7e8e58651d907cb Mon Sep 17 00:00:00 2001
From: chengduoZH <zhaochengduo@163.com>
Date: Thu, 14 Dec 2017 11:21:41 +0800
Subject: [PATCH 54/94] follow comments

---
 paddle/operators/conv_transpose_op.h | 1 -
 1 file changed, 1 deletion(-)

diff --git a/paddle/operators/conv_transpose_op.h b/paddle/operators/conv_transpose_op.h
index 0aae000dc4..1171b0435f 100644
--- a/paddle/operators/conv_transpose_op.h
+++ b/paddle/operators/conv_transpose_op.h
@@ -225,7 +225,6 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
 
       if (input_grad) {
         input_grad->mutable_data<T>(context.GetPlace());
-        // set_zero is unnecessary, math::matmul will reset input_grad.
       }
       if (filter_grad) {  // filter size (m, c, k_h, k_w)
         filter_grad->mutable_data<T>(context.GetPlace());

From 685d1e3b330ff6605935ed4a33825bdeb97ced07 Mon Sep 17 00:00:00 2001
From: Yibing Liu <liuyibing01@baidu.com>
Date: Thu, 14 Dec 2017 04:27:05 +0000
Subject: [PATCH 55/94] Enable reshape_op to support dimension inference

---
 paddle/operators/reshape_op.cc | 27 ++++++++++++++++++---------
 1 file changed, 18 insertions(+), 9 deletions(-)

diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc
index 39bf2118d6..306dfa8069 100644
--- a/paddle/operators/reshape_op.cc
+++ b/paddle/operators/reshape_op.cc
@@ -34,21 +34,27 @@ class ReshapeOp : public framework::OperatorWithKernel {
     auto shape = ctx->Attrs().Get<std::vector<int>>("shape");
     PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty.");
     auto x_dims = ctx->GetInputDim("X");
-    // TODO(qiao) change batch_size
-    for (size_t i = 1; i < shape.size(); ++i) {
-      PADDLE_ENFORCE(shape[i] > 0,
-                     "Each dimension of Attr(shape) "
-                     "must be positive except the first one.");
-    }
-    if (shape[0] < 0) {
-      shape[0] = x_dims[0];
+
+    std::vector<size_t> neg_dims_idx;
+    for (size_t i = 0; i < shape.size(); ++i) {
+      PADDLE_ENFORCE(shape[i] > 0 || shape[i] == -1,
+                     "Each dimension of Attr(shape) must be positive or -1.");
+      if (shape[i] == -1) {
+        neg_dims_idx.push_back(i);
+        PADDLE_ENFORCE(neg_dims_idx.size() <= 1,
+                     "Only one dimension of Attr(shape) can be -1.");
+      }
     }
+
     // capacity check
     int64_t capacity =
         std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>());
     int64_t in_size = framework::product(x_dims);
-    PADDLE_ENFORCE_EQ(capacity, in_size,
+    if (neg_dims_idx.size() == 1) {
+       shape[neg_dims_idx[0]] = in_size / (-capacity);
+       PADDLE_ENFORCE(shape[neg_dims_idx[0]] > 0,
                       "The size of Input(X) mismatches with Attr(shape).");
+    }
     // resize output
     std::vector<int64_t> shape_int64(shape.size(), 0);
     std::transform(shape.begin(), shape.end(), shape_int64.begin(),
@@ -88,6 +94,9 @@ the tensor X into a 1-D tensor:
 
     [1, 2, 3, 4]
 
+One dimension in the target shape can be set -1, and the real dimension 
+will be infered from the original shape of Input(X) and other
+dimensions in the target shape.
 )DOC");
   }
 };

From 0a75ed6f5bcad53b3622ee89d48a5613cc4ef396 Mon Sep 17 00:00:00 2001
From: Yibing Liu <liuyibing01@baidu.com>
Date: Thu, 14 Dec 2017 05:06:19 +0000
Subject: [PATCH 56/94] Add unit test for dimension inference in reshape_op

---
 paddle/operators/reshape_op.cc                  | 14 +++++++++-----
 python/paddle/v2/fluid/tests/test_reshape_op.py | 14 ++++++++++++++
 2 files changed, 23 insertions(+), 5 deletions(-)

diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc
index 306dfa8069..164f3104eb 100644
--- a/paddle/operators/reshape_op.cc
+++ b/paddle/operators/reshape_op.cc
@@ -42,19 +42,23 @@ class ReshapeOp : public framework::OperatorWithKernel {
       if (shape[i] == -1) {
         neg_dims_idx.push_back(i);
         PADDLE_ENFORCE(neg_dims_idx.size() <= 1,
-                     "Only one dimension of Attr(shape) can be -1.");
+                       "Only one dimension of Attr(shape) can be unknown.");
       }
     }
 
-    // capacity check
     int64_t capacity =
         std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>());
     int64_t in_size = framework::product(x_dims);
     if (neg_dims_idx.size() == 1) {
-       shape[neg_dims_idx[0]] = in_size / (-capacity);
-       PADDLE_ENFORCE(shape[neg_dims_idx[0]] > 0,
-                      "The size of Input(X) mismatches with Attr(shape).");
+      // dim infer
+      shape[neg_dims_idx[0]] = in_size / (-capacity);
+      // recalculate capacity
+      capacity = std::accumulate(shape.begin(), shape.end(), 1,
+                                 std::multiplies<int>());
     }
+    // capacity check
+    PADDLE_ENFORCE(capacity == in_size,
+                   "The size of Input(X) mismatches with Attr(shape).");
     // resize output
     std::vector<int64_t> shape_int64(shape.size(), 0);
     std::transform(shape.begin(), shape.end(), shape_int64.begin(),
diff --git a/python/paddle/v2/fluid/tests/test_reshape_op.py b/python/paddle/v2/fluid/tests/test_reshape_op.py
index 16bb6bb2af..18ee3aece6 100644
--- a/python/paddle/v2/fluid/tests/test_reshape_op.py
+++ b/python/paddle/v2/fluid/tests/test_reshape_op.py
@@ -17,5 +17,19 @@ class TestReshapeOp(OpTest):
         self.check_grad(["X"], "Out")
 
 
+class TestReshapeOpDimInfer(OpTest):
+    def setUp(self):
+        self.op_type = "reshape"
+        self.inputs = {'X': np.random.random((10, 20)).astype("float32")}
+        self.attrs = {'shape': [4, -1, 5]}
+        self.outputs = {'Out': self.inputs['X'].reshape(self.attrs['shape'])}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
 if __name__ == '__main__':
     unittest.main()

From 9b1675048561e2708af40948b192f24a199bf70a Mon Sep 17 00:00:00 2001
From: dzhwinter <dzhwinter@gmail.com>
Date: Wed, 13 Dec 2017 22:44:33 -0800
Subject: [PATCH 57/94] "remove AllReduce2 comments"

---
 doc/design/paddle_nccl.md | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/doc/design/paddle_nccl.md b/doc/design/paddle_nccl.md
index dba451b9b3..5e28144b95 100644
--- a/doc/design/paddle_nccl.md
+++ b/doc/design/paddle_nccl.md
@@ -53,8 +53,13 @@ Need to notice that AllReduce operator force GPUs synchronized at that point. Th
 
 As it shown in the picture, when each GPU compute the gradient of `W`, followed with a `AllReduce` operator, accumulate the `dW` to full batch of data, then run the optimize process individually and apply the gradient to its `W`.
 
-- **AllReduce2**
-If we use the NCCL2 AllReduce primitive, every GPU optimized full batch of data, wasted (n-1) GPU compute resources. In addition, AllReduce will only utilize the communicate resource during synchronization, then update the gradient will be a seperated phase. In fact, we can amortize the update gradient time cost into the communicating phase.
-- Every parameter has its root card. That card will call **Reduce** operator and collect the gradients from GPUs.
-- The whole model's parameter will be hashed to different root card, ensure the load balance between GPUs.
-Then we have another version AllReduce operator. Other part keep the same with before.
+- **AllReduce**
+  Need to note that our AllReduce operator is a ring-base AllReduce implementation. If we use the NCCL2 AllReduce primitive, every GPU optimized full batch of data, wasted (n-1) GPU compute resources. In addition, NCCL2 built-in AllReduce will only utilize the communicating resource during synchronization, then update the gradient will be a subsequent phase. In fact, we can amortize the update gradient time cost into the communicating phase. The process is
+1. Every parameter has its root card. That card will responsible for aggregating the gradients from GPUs.
+2. The whole model's parameter will be hashed to different root card, ensure the load balance between GPUs.
+3. Logically neighberhood card will start send parameter to the next one. After one round, the parameter main card will aggregate the full gradients.
+4. Then the root card will optimize the parameter.
+5. This parameter card will send its optimized result to its neighberhood, then the neighberhood will send parameter to its next one.
+6. Finish the sychronization round.
+
+The total time cost will be 2 * (n-1) * per-parameter-send-time, we reach the goal of amortize the upgrade time into communicating phase.

From dafd449c68c23c642ba117a55135c823c6594772 Mon Sep 17 00:00:00 2001
From: fengjiayi <fengjiayi@baidu.com>
Date: Thu, 14 Dec 2017 15:43:14 +0800
Subject: [PATCH 58/94] Unify `step_block` and `block` to `sub_block`

---
 paddle/framework/backward.cc             | 4 ++--
 paddle/operators/conditional_block_op.cc | 8 ++++----
 paddle/operators/recurrent_op.cc         | 2 +-
 paddle/operators/while_op.cc             | 2 +-
 python/paddle/v2/fluid/layers.py         | 6 +++---
 5 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc
index a17036c652..faf6e60cbd 100644
--- a/paddle/framework/backward.cc
+++ b/paddle/framework/backward.cc
@@ -430,14 +430,14 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
     std::vector<std::unique_ptr<OpDescBind>> op_grads;
 
     if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") {
-      int step_block_idx = (*it)->GetBlockAttr("step_block");
+      int step_block_idx = (*it)->GetBlockAttr("sub_block");
       BlockDescBind* backward_block = CreateStepBlock(
           program_desc, no_grad_vars, grad_to_var, step_block_idx);
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block});
     } else if ((*it)->Type() == "conditional_block") {
       BlockDescBind* backward_block =
           CreateStepBlock(program_desc, no_grad_vars, grad_to_var,
-                          (*it)->GetBlockAttr("block"));
+                          (*it)->GetBlockAttr("sub_block"));
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block});
     } else {
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var);
diff --git a/paddle/operators/conditional_block_op.cc b/paddle/operators/conditional_block_op.cc
index 03c58a7eab..6f2ef9174e 100644
--- a/paddle/operators/conditional_block_op.cc
+++ b/paddle/operators/conditional_block_op.cc
@@ -65,7 +65,7 @@ class ConditionalBlockOp : public ConditionalOp {
       scopes->front() = &scope.NewScope();
       auto &cur_scope = *scopes->front();
 
-      auto *block = Attr<framework::BlockDescBind *>("block");
+      auto *block = Attr<framework::BlockDescBind *>("sub_block");
       framework::Executor exec(dev_ctx);
       exec.Run(*block->Program(), &cur_scope, block->ID(), false);
     }
@@ -88,7 +88,7 @@ class ConditionalBlockOpProtoMaker : public framework::OpProtoAndCheckerMaker {
               "unify the conditional block, rnn and while op, the type of "
               "scope is std::vector<Scope*>");
     AddAttr<framework::BlockDescBind *>(
-        "block", "The step block of conditional block operator");
+        "sub_block", "The step block of conditional block operator");
     AddComment(R"DOC(Conditional block operator
 
 Run the sub-block if X is not empty. Params is the other inputs and Out is the
@@ -117,7 +117,7 @@ class ConditionalBlockGradOp : public ConditionalOp {
       auto &scopes = scope_var->Get<std::vector<framework::Scope *>>();
       framework::Scope &cur_scope = *scopes[0];
 
-      auto *block = Attr<framework::BlockDescBind *>("block");
+      auto *block = Attr<framework::BlockDescBind *>("sub_block");
       framework::Executor exec(dev_ctx);
       exec.Run(*block->Program(), &cur_scope, block->ID(), false);
 
@@ -181,7 +181,7 @@ class ConditionalBlockGradMaker : public framework::SingleGradOpDescMaker {
     grad_op->SetInput("Scope", Output("Scope"));
     grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
     grad_op->SetOutput(framework::GradVarName("Params"), InputGrad("Params"));
-    grad_op->SetBlockAttr("block", *this->grad_block_[0]);
+    grad_op->SetBlockAttr("sub_block", *this->grad_block_[0]);
     return std::unique_ptr<framework::OpDescBind>(grad_op);
   }
 };
diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc
index 29f9163643..232d926f7b 100644
--- a/paddle/operators/recurrent_op.cc
+++ b/paddle/operators/recurrent_op.cc
@@ -25,7 +25,7 @@ constexpr char kOutputs[] = "outputs";
 constexpr char kStepScopes[] = "step_scopes";
 constexpr char kExStates[] = "ex_states";
 constexpr char kStates[] = "states";
-constexpr char kStepBlock[] = "step_block";
+constexpr char kStepBlock[] = "sub_block";
 constexpr char kReverse[] = "reverse";
 constexpr char kIsTrain[] = "is_train";
 #define GRAD_SUFFIX "@GRAD"
diff --git a/paddle/operators/while_op.cc b/paddle/operators/while_op.cc
index b8e44bcc5a..9a092a570f 100644
--- a/paddle/operators/while_op.cc
+++ b/paddle/operators/while_op.cc
@@ -25,7 +25,7 @@ namespace operators {
 using StepScopeVar = std::vector<framework::Scope *>;
 using LoDTensor = framework::LoDTensor;
 
-constexpr char kStepBlock[] = "step_block";
+constexpr char kStepBlock[] = "sub_block";
 constexpr char kCondition[] = "Condition";
 constexpr char kStepScopes[] = "StepScopes";
 constexpr char kParameters[] = "X";
diff --git a/python/paddle/v2/fluid/layers.py b/python/paddle/v2/fluid/layers.py
index f67d6d08c7..2781017ec4 100644
--- a/python/paddle/v2/fluid/layers.py
+++ b/python/paddle/v2/fluid/layers.py
@@ -1130,7 +1130,7 @@ class StaticRNN(object):
             attrs={
                 'ex_states': pre_memories,
                 'states': memories,
-                'step_block': rnn_block
+                'sub_block': rnn_block
             })
 
 
@@ -1207,7 +1207,7 @@ class While(object):
             },
             outputs={'Out': out_vars,
                      'StepScopes': [step_scope]},
-            attrs={'step_block': while_block})
+            attrs={'sub_block': while_block})
 
 
 def lstm(x,
@@ -1671,7 +1671,7 @@ class ConditionalBlock(object):
             },
             outputs={'Out': out_list,
                      'Scope': [step_scope]},
-            attrs={'block': inside_block})
+            attrs={'sub_block': inside_block})
 
 
 class IfElseBlockGuard(object):

From 0e9b393b340990cf581ec9f6e5f33af74912c0b6 Mon Sep 17 00:00:00 2001
From: dzhwinter <dongzhihong01@baidu.com>
Date: Thu, 14 Dec 2017 16:21:58 +0800
Subject: [PATCH 59/94] "derived cudnnDevice context" (#6585)

* "derived cudnnDevice context"

* "leave remove cudnn handle from CUDADeviceContext"

* "fix math function error"
---
 paddle/operators/math/math_function.cu |  7 +++++++
 paddle/platform/device_context.cc      | 16 ++++++++++++++++
 paddle/platform/device_context.h       | 16 ++++++++++++++++
 paddle/platform/device_context_test.cc | 16 ++++++++++++++++
 paddle/platform/place.h                |  7 ++++++-
 5 files changed, 61 insertions(+), 1 deletion(-)

diff --git a/paddle/operators/math/math_function.cu b/paddle/operators/math/math_function.cu
index 1b560a7e2d..e33070c40f 100644
--- a/paddle/operators/math/math_function.cu
+++ b/paddle/operators/math/math_function.cu
@@ -273,6 +273,13 @@ void set_constant_with_place<platform::GPUPlace>(
                            TensorSetConstantGPU(context, tensor, value));
 }
 
+template <>
+void set_constant_with_place<platform::CudnnPlace>(
+    const platform::DeviceContext& context, framework::Tensor* tensor,
+    float value) {
+  set_constant_with_place<platform::GPUPlace>(context, tensor, value);
+}
+
 template struct RowwiseAdd<platform::CUDADeviceContext, float>;
 template struct RowwiseAdd<platform::CUDADeviceContext, double>;
 template struct ColwiseSum<platform::CUDADeviceContext, float>;
diff --git a/paddle/platform/device_context.cc b/paddle/platform/device_context.cc
index 2c7f964216..1c72b50559 100644
--- a/paddle/platform/device_context.cc
+++ b/paddle/platform/device_context.cc
@@ -125,6 +125,22 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_handle_; }
 
 cudaStream_t CUDADeviceContext::stream() const { return stream_; }
 
+CudnnDeviceContext::CudnnDeviceContext(CudnnPlace place)
+    : CUDADeviceContext(place), place_(place) {
+  PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
+  PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream()));
+}
+
+CudnnDeviceContext::~CudnnDeviceContext() {
+  SetDeviceId(place_.device);
+  Wait();
+  PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_));
+}
+
+Place CudnnDeviceContext::GetPlace() const { return CudnnPlace(); }
+
+cudnnHandle_t CudnnDeviceContext::cudnn_handle() const { return cudnn_handle_; }
+
 #endif
 
 }  // namespace platform
diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h
index 596d9d0bba..f67194993d 100644
--- a/paddle/platform/device_context.h
+++ b/paddle/platform/device_context.h
@@ -86,6 +86,22 @@ class CUDADeviceContext : public DeviceContext {
   cublasHandle_t cublas_handle_;
 };
 
+class CudnnDeviceContext : public CUDADeviceContext {
+ public:
+  explicit CudnnDeviceContext(CudnnPlace place);
+  virtual ~CudnnDeviceContext();
+
+  /*! \brief  Return place in the device context. */
+  Place GetPlace() const final;
+
+  /*! \brief  Return cudnn  handle in the device context. */
+  cudnnHandle_t cudnn_handle() const;
+
+ private:
+  cudnnHandle_t cudnn_handle_;
+  CudnnPlace place_;
+};
+
 #endif
 
 }  // namespace platform
diff --git a/paddle/platform/device_context_test.cc b/paddle/platform/device_context_test.cc
index 4893cd92f6..be3b2af5af 100644
--- a/paddle/platform/device_context_test.cc
+++ b/paddle/platform/device_context_test.cc
@@ -46,3 +46,19 @@ TEST(Device, CUDADeviceContext) {
     delete device_context;
   }
 }
+
+TEST(Device, CudnnDeviceContext) {
+  using paddle::platform::CudnnDeviceContext;
+  using paddle::platform::CudnnPlace;
+  if (paddle::platform::dynload::HasCUDNN()) {
+    int count = paddle::platform::GetCUDADeviceCount();
+    for (int i = 0; i < count; ++i) {
+      CudnnDeviceContext* device_context =
+          new CudnnDeviceContext(CudnnPlace(i));
+      cudnnHandle_t cudnn_handle = device_context->cudnn_handle();
+      ASSERT_NE(nullptr, cudnn_handle);
+      ASSERT_NE(nullptr, device_context->stream());
+      delete device_context;
+    }
+  }
+}
diff --git a/paddle/platform/place.h b/paddle/platform/place.h
index 5370360a7d..f0dcec8f52 100644
--- a/paddle/platform/place.h
+++ b/paddle/platform/place.h
@@ -43,6 +43,11 @@ struct GPUPlace {
   int device;
 };
 
+struct CudnnPlace : public GPUPlace {
+  CudnnPlace() : GPUPlace() {}
+  explicit CudnnPlace(int d) : GPUPlace(d) {}
+};
+
 struct IsGPUPlace : public boost::static_visitor<bool> {
   bool operator()(const CPUPlace &) const { return false; }
   bool operator()(const GPUPlace &gpu) const { return true; }
@@ -52,7 +57,7 @@ struct IsGPUPlace : public boost::static_visitor<bool> {
 // should be less equal than 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT)
 #define NUM_PLACE_TYPE_LIMIT_IN_BIT 4
 
-typedef boost::variant<GPUPlace, CPUPlace> Place;
+typedef boost::variant<CudnnPlace, GPUPlace, CPUPlace> Place;
 
 // static check number of place types is less equal than
 // 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT)

From e11a561c128a894ce03bd07bd3985f55f74220f8 Mon Sep 17 00:00:00 2001
From: fengjiayi <fengjiayi@baidu.com>
Date: Thu, 14 Dec 2017 17:06:25 +0800
Subject: [PATCH 60/94] update

---
 paddle/pybind/pybind.cc | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index c16d3e0cbe..9ea4e70a26 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -282,6 +282,16 @@ All parameter, weight, gradient are variables in Paddle.
     }
     return ret_values;
   });
+  m.def("get_grad_op_desc",
+        [](const OpDescBind &op_desc,
+           const std::unordered_set<std::string> &no_grad_set,
+           std::unordered_map<std::string, std::string> &grad_to_var,
+           const std::vector<BlockDescBind *> &grad_sub_block) {
+          return framework::OpInfoMap::Instance()
+              .Get(op_desc.Type())
+              .GradOpMaker()(op_desc, no_grad_set, &grad_to_var,
+                             grad_sub_block);
+        });
   m.def("prune", [](const ProgramDescBind &origin,
                     const std::vector<std::array<size_t, 2>> &targets) {
     ProgramDescBind prog_with_targets(origin);

From 1059796cbfa48273b8c056fee23f8e405dc24bb8 Mon Sep 17 00:00:00 2001
From: Luo Tao <luotao02@baidu.com>
Date: Thu, 14 Dec 2017 17:16:47 +0800
Subject: [PATCH 61/94] update the link of docs icon

---
 README.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index bbb2d49858..ceeb6d9e51 100644
--- a/README.md
+++ b/README.md
@@ -2,8 +2,8 @@
 
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://doc.paddlepaddle.org/develop/doc/)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://doc.paddlepaddle.org/develop/doc_cn/)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html)
 [![Coverage Status](https://coveralls.io/repos/github/PaddlePaddle/Paddle/badge.svg?branch=develop)](https://coveralls.io/github/PaddlePaddle/Paddle?branch=develop)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)

From 991826317a4192cd4fb3e8505a4a7c60c0fe830c Mon Sep 17 00:00:00 2001
From: Leding Li <lileding@gmail.com>
Date: Thu, 14 Dec 2017 17:23:39 +0800
Subject: [PATCH 62/94] Add extern "C" to paddle_error_string (#6611)

---
 paddle/capi/error.cpp | 2 +-
 paddle/capi/error.h   | 8 ++++++++
 2 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/paddle/capi/error.cpp b/paddle/capi/error.cpp
index 169b65f921..96ce31b45f 100644
--- a/paddle/capi/error.cpp
+++ b/paddle/capi/error.cpp
@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "error.h"
 
-const char* paddle_error_string(paddle_error err) {
+extern "C" const char* paddle_error_string(paddle_error err) {
   switch (err) {
     case kPD_NULLPTR:
       return "nullptr error";
diff --git a/paddle/capi/error.h b/paddle/capi/error.h
index 9d9d0ed63a..2da9e0a3ef 100644
--- a/paddle/capi/error.h
+++ b/paddle/capi/error.h
@@ -29,9 +29,17 @@ typedef enum {
   kPD_UNDEFINED_ERROR = -1,
 } paddle_error;
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 /**
  * Error string for Paddle API.
  */
 PD_API const char* paddle_error_string(paddle_error err);
 
+#ifdef __cplusplus
+}
+#endif
+
 #endif

From 044a13d02262f1cf84ee685a0575cf2ec28e5623 Mon Sep 17 00:00:00 2001
From: fengjiayi <fengjiayi@baidu.com>
Date: Thu, 14 Dec 2017 17:50:56 +0800
Subject: [PATCH 63/94] expose GradOpMaker to Python

---
 paddle/pybind/pybind.cc | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index 9ea4e70a26..1faf24bcb8 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -282,15 +282,22 @@ All parameter, weight, gradient are variables in Paddle.
     }
     return ret_values;
   });
-  m.def("get_grad_op_desc",
+  m.def("get_grad_op_descs",
         [](const OpDescBind &op_desc,
            const std::unordered_set<std::string> &no_grad_set,
            std::unordered_map<std::string, std::string> &grad_to_var,
            const std::vector<BlockDescBind *> &grad_sub_block) {
-          return framework::OpInfoMap::Instance()
-              .Get(op_desc.Type())
-              .GradOpMaker()(op_desc, no_grad_set, &grad_to_var,
-                             grad_sub_block);
+          std::vector<std::unique_ptr<OpDescBind>> grad_op_descs =
+              framework::OpInfoMap::Instance()
+                  .Get(op_desc.Type())
+                  .GradOpMaker()(op_desc, no_grad_set, &grad_to_var,
+                                 grad_sub_block);
+          std::vector<OpDescBind *> grad_op_desc_ptrs(grad_op_descs.size());
+          std::transform(
+              grad_op_descs.begin(), grad_op_descs.end(),
+              grad_op_desc_ptrs.begin(),
+              [](std::unique_ptr<OpDescBind> &p) { return p.release(); });
+          return grad_op_desc_ptrs;
         });
   m.def("prune", [](const ProgramDescBind &origin,
                     const std::vector<std::array<size_t, 2>> &targets) {

From e0698e33a869f72282ae9f7d78a33c5a4ac2ef00 Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Thu, 14 Dec 2017 17:57:49 +0800
Subject: [PATCH 64/94] Make layers as a python module (#6564)

* Make cast op support bool

Also add `elemwise_sub/mul/abs/clip` layers

* Make fuild.layers as a module

* Move layers as a module

* Split layers.py into layers module

* Fix CI

* Fix CI
---
 python/paddle/v2/fluid/layers/__init__.py     |   17 +
 .../{layers.py => layers/control_flow.py}     | 1023 +----------------
 python/paddle/v2/fluid/layers/io.py           |   57 +
 python/paddle/v2/fluid/layers/nn.py           |  785 +++++++++++++
 python/paddle/v2/fluid/layers/ops.py          |    9 +
 python/paddle/v2/fluid/layers/tensor.py       |  130 +++
 .../book/test_image_classification_train.py   |    4 +-
 .../book/test_understand_sentiment_lstm.py    |   48 +-
 python/setup.py.in                            |    1 +
 9 files changed, 1057 insertions(+), 1017 deletions(-)
 create mode 100644 python/paddle/v2/fluid/layers/__init__.py
 rename python/paddle/v2/fluid/{layers.py => layers/control_flow.py} (51%)
 create mode 100644 python/paddle/v2/fluid/layers/io.py
 create mode 100644 python/paddle/v2/fluid/layers/nn.py
 create mode 100644 python/paddle/v2/fluid/layers/ops.py
 create mode 100644 python/paddle/v2/fluid/layers/tensor.py

diff --git a/python/paddle/v2/fluid/layers/__init__.py b/python/paddle/v2/fluid/layers/__init__.py
new file mode 100644
index 0000000000..249f570e13
--- /dev/null
+++ b/python/paddle/v2/fluid/layers/__init__.py
@@ -0,0 +1,17 @@
+import ops
+from ops import *
+import nn
+from nn import *
+import io
+from io import *
+import tensor
+from tensor import *
+import control_flow
+from control_flow import *
+
+__all__ = []
+__all__ += nn.__all__
+__all__ += io.__all__
+__all__ += tensor.__all__
+__all__ += control_flow.__all__
+__all__ += ops.__all__
diff --git a/python/paddle/v2/fluid/layers.py b/python/paddle/v2/fluid/layers/control_flow.py
similarity index 51%
rename from python/paddle/v2/fluid/layers.py
rename to python/paddle/v2/fluid/layers/control_flow.py
index 2781017ec4..5af6c78977 100644
--- a/python/paddle/v2/fluid/layers.py
+++ b/python/paddle/v2/fluid/layers/control_flow.py
@@ -1,424 +1,18 @@
+from ..layer_helper import LayerHelper, unique_name
+from ..framework import Program, Variable, Operator
+from .. import core
+from tensor import assign, fill_constant
 import contextlib
 
-import proto.framework_pb2 as framework_pb2
-import core
-from framework import OpProtoHolder, Variable, Program, Operator
-from initializer import Constant, Normal, Xavier, Initializer
-from paddle.v2.fluid.layer_helper import LayerHelper, unique_name
-from registry import register_layer
-from param_attr import ParamAttr
-
 __all__ = [
-    'fc', 'data', 'cross_entropy', 'conv2d', 'pool2d', 'embedding', 'concat',
-    'StaticRNN', 'cast', 'sequence_conv', 'sequence_pool', 'sums', 'cos_sim',
-    'batch_norm', 'accuracy', 'split_lod_tensor', 'While'
-]
-
-_REGISTER_LAYER_FROM_OPS = [
-    'mean', 'mul', 'dropout', 'reshape', 'sigmoid', 'scale', 'transpose',
-    'sigmoid_cross_entropy_with_logits', 'elementwise_add', 'elementwise_div',
-    'elementwise_sub', 'elementwise_mul', 'clip', 'abs'
+    'split_lod_tensor', 'merge_lod_tensor', 'BlockGuard', 'StaticRNNGuard',
+    'StaticRNNMemoryLink', 'WhileGuard', 'While', 'lod_rank_table',
+    'max_sequence_len', 'topk', 'lod_tensor_to_array', 'array_to_lod_tensor',
+    'increment', 'array_write', 'create_array', 'less_than', 'array_read',
+    'shrink_memory', 'array_length', 'IfElse', 'DynamicRNN', 'ConditionalBlock',
+    'StaticRNN'
 ]
 
-for _OP in set(_REGISTER_LAYER_FROM_OPS):
-    globals()[_OP] = register_layer(_OP)
-    __all__.append(_OP)
-
-
-def fc(input,
-       size,
-       num_flatten_dims=1,
-       param_attr=None,
-       bias_attr=None,
-       act=None,
-       name=None,
-       main_program=None,
-       startup_program=None):
-    """
-    Fully Connected Layer.
-
-    Args:
-       input: The input tensor to the function
-       size: The size of the layer
-       num_flatten_dims: Number of columns in input
-       param_attr: The parameters/weights to the FC Layer
-       param_initializer: Initializer used for the weight/parameter. If None, XavierInitializer() is used
-       bias_attr: The bias parameter for the FC layer
-       bias_initializer: Initializer used for the bias. If None, then ConstantInitializer() is used
-       act: Activation to be applied to the output of FC layer
-       name: Name/alias of the function
-       main_program: Name of the main program that calls this
-       startup_program: Name of the startup program
-
-    This function can take in multiple inputs and performs the Fully Connected
-    function (linear transformation) on top of each of them.
-    So for input x, the output will be : Wx + b. Where W is the parameter,
-    b the bias and x is the input.
-
-    The function also applies an activation (non-linearity) on top of the
-    output, if activation is passed in the input.
-
-    All the input variables of this function are passed in as local variables
-    to the LayerHelper constructor.
-
-    """
-    helper = LayerHelper('fc', **locals())
-
-    dtype = helper.input_dtype()
-
-    mul_results = []
-    for input_var, param_attr in helper.iter_inputs_and_params():
-        input_shape = input_var.shape
-        param_shape = [
-            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
-        ] + [size]
-        w = helper.create_parameter(
-            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
-        tmp = helper.create_tmp_variable(dtype)
-        helper.append_op(
-            type="mul",
-            inputs={
-                "X": input_var,
-                "Y": w,
-            },
-            outputs={"Out": tmp},
-            attrs={'x_num_col_dims': num_flatten_dims,
-                   'y_num_col_dims': 1})
-        mul_results.append(tmp)
-
-    # sum
-    if len(mul_results) == 1:
-        pre_bias = mul_results[0]
-    else:
-        pre_bias = helper.create_tmp_variable(dtype)
-        helper.append_op(
-            type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias})
-    # add bias
-    pre_activation = helper.append_bias_op(pre_bias)
-    # add activation
-    return helper.append_activation(pre_activation)
-
-
-def embedding(input,
-              size,
-              is_sparse=False,
-              param_attr=None,
-              dtype='float32',
-              main_program=None,
-              startup_program=None):
-    """
-    Embedding Layer.
-
-    Args:
-       param_initializer:
-       input: The input to the function
-       size: The size of the layer
-       is_sparse: A flag that decleares whether the input is sparse
-       param_attr: Parameters for this layer
-       dtype: The type of data : float32, float_16, int etc
-       main_program: Name of the main program that calls this
-       startup_program: Name of the startup program
-
-    This function can take in the input (which is a vector of IDs) and
-    performs a lookup in the lookup_table using these IDs, to result into
-    the embedding of each ID in the input.
-
-    All the input variables of this function are passed in as local variables
-    to the LayerHelper constructor.
-
-    """
-
-    helper = LayerHelper('embedding', **locals())
-    w = helper.create_parameter(
-        attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False)
-    tmp = helper.create_tmp_variable(dtype)
-    helper.append_op(
-        type='lookup_table',
-        inputs={'Ids': input,
-                'W': w},
-        outputs={'Out': tmp},
-        attrs={'is_sparse': is_sparse})
-    return tmp
-
-
-# TODO(qijun): expose H0 and C0
-def dynamic_lstm(input,
-                 size,
-                 param_attr=None,
-                 bias_attr=None,
-                 use_peepholes=True,
-                 is_reverse=False,
-                 gate_activation='sigmoid',
-                 cell_activation='tanh',
-                 candidate_activation='tanh',
-                 dtype='float32',
-                 main_program=None,
-                 startup_program=None):
-    helper = LayerHelper('lstm', **locals())
-    size = size / 4
-    weight = helper.create_parameter(
-        attr=helper.param_attr, shape=[size, 4 * size], dtype=dtype)
-    bias_size = [1, 7 * size]
-    if not use_peepholes:
-        bias_size[1] = 4 * size
-    bias = helper.create_parameter(
-        attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
-
-    hidden = helper.create_tmp_variable(dtype)
-    cell = helper.create_tmp_variable(dtype)
-    batch_gate = helper.create_tmp_variable(dtype)
-    batch_cell_pre_act = helper.create_tmp_variable(dtype)
-
-    helper.append_op(
-        type='lstm',
-        inputs={'Input': input,
-                'Weight': weight,
-                'Bias': bias},
-        outputs={
-            'Hidden': hidden,
-            'Cell': cell,
-            'BatchGate': batch_gate,
-            'BatchCellPreAct': batch_cell_pre_act
-        },
-        attrs={
-            'use_peepholes': use_peepholes,
-            'is_reverse': is_reverse,
-            'gate_activation': gate_activation,
-            'cell_activation': cell_activation,
-            'candidate_activation': candidate_activation
-        })
-    return hidden, cell
-
-
-def gru_unit(input,
-             hidden,
-             size,
-             weight=None,
-             bias=None,
-             activation='tanh',
-             gate_activation='sigmoid',
-             main_program=None,
-             startup_program=None):
-    """
-    GRUUnit Operator implements partial calculations of the GRU unit as following:
-
-    $$
-    update \ gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\
-    reset \ gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r)  \\
-    output \ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\
-    output: h_t = dot((1 - u_t), h_{t-1}) + dot(u_t, {h}_t)
-    $$
-
-    which is same as one time step of GRU Operator.
-
-    @note To implement the complete GRU unit, fully-connected operator must be
-    used before to feed xu, xr and xc as the Input of GRUUnit operator.
-
-    TODO(ChunweiYan) add more document here
-    """
-    activation_dict = dict(
-        identity=0,
-        sigmoid=1,
-        tanh=2,
-        relu=3, )
-    activation = activation_dict[activation]
-    gate_activation = activation_dict[gate_activation]
-
-    helper = LayerHelper('gru_unit', **locals())
-    dtype = helper.input_dtype()
-    size = size / 3
-
-    # create weight
-    if weight is None:
-        weight = helper.create_parameter(
-            attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
-
-    # create bias
-    if bias is None:
-        bias_size = [1, 3 * size]
-        bias = helper.create_parameter(
-            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
-
-    gate = helper.create_tmp_variable(dtype)
-    reset_hidden_pre = helper.create_tmp_variable(dtype)
-    updated_hidden = helper.create_tmp_variable(dtype)
-
-    helper.append_op(
-        type='gru_unit',
-        inputs={'Input': input,
-                'HiddenPrev': hidden,
-                'Weight': weight},
-        outputs={
-            'Gate': gate,
-            'ResetHiddenPrev': reset_hidden_pre,
-            'Hidden': updated_hidden,
-        },
-        attrs={
-            'activation': 0,
-            'gate_activation': 1,
-        })
-
-    return updated_hidden, reset_hidden_pre, gate
-
-
-def data(name,
-         shape,
-         append_batch_size=True,
-         dtype='float32',
-         lod_level=0,
-         type=core.VarDesc.VarType.LOD_TENSOR,
-         main_program=None,
-         startup_program=None,
-         stop_gradient=True):
-    """
-    Data Layer.
-
-    Args:
-       name: The name/alias of the function
-       shape: Tuple declaring the shape.
-       append_batch_size: Whether or not to append the data as a batch.
-       dtype: The type of data : float32, float_16, int etc
-       type: The output type. By default it is LOD_TENSOR.
-       lod_level(int): The LoD Level. 0 means the input data is not a sequence.
-       main_program: Name of the main program that calls this
-       startup_program: Name of the startup program
-       stop_gradient: A boolean that mentions whether gradient should flow.
-
-    This function takes in input and based on whether data has
-    to be returned back as a minibatch, it creates the global variable using
-    the helper functions. The global variables can be accessed by all the
-    following operations and layers in the graph.
-
-    All the input variables of this function are passed in as local variables
-    to the LayerHelper constructor.
-
-    """
-    helper = LayerHelper('data', **locals())
-    shape = list(shape)
-    for i in xrange(len(shape)):
-        if shape[i] is None:
-            shape[i] = -1
-            append_batch_size = False
-        elif shape[i] < 0:
-            append_batch_size = False
-
-    if append_batch_size:
-        shape = [-1] + shape  # append batch size as -1
-
-    return helper.create_global_variable(
-        name=name,
-        shape=shape,
-        dtype=dtype,
-        type=type,
-        stop_gradient=stop_gradient,
-        lod_level=lod_level)
-
-
-def create_tensor(dtype, name=None, main_program=None, startup_program=None):
-    helper = LayerHelper("create_tensor", **locals())
-    return helper.create_variable(name=helper.name, dtype=dtype)
-
-
-def cast(x, dtype, main_program=None):
-    """
-    This function takes in the input with input_dtype
-    and casts it to the output_dtype as the output.
-    """
-    helper = LayerHelper('cast', **locals())
-    out = helper.create_tmp_variable(dtype=dtype)
-    helper.append_op(
-        type='cast',
-        inputs={'X': [x]},
-        outputs={'Out': [out]},
-        attrs={'in_dtype': x.dtype,
-               'out_dtype': out.dtype})
-    return out
-
-
-def concat(input, axis, main_program=None, startup_program=None):
-    """
-    This function concats the input along the axis mentioned
-    and returns that as the output.
-    """
-    helper = LayerHelper('concat', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype())
-    helper.append_op(
-        type='concat',
-        inputs={'X': input},
-        outputs={'Out': [out]},
-        attrs={'axis': axis})
-    return out
-
-
-def sums(input, out=None, main_program=None, startup_program=None):
-    """
-    This function takes in the input and performs the sum operation on it
-    and returns that as the output.
-    """
-    helper = LayerHelper('sum', **locals())
-    if out is None:
-        out = helper.create_tmp_variable(dtype=helper.input_dtype())
-    helper.append_op(type='sum', inputs={'X': input}, outputs={'Out': out})
-    return out
-
-
-def linear_chain_crf(input,
-                     label,
-                     param_attr=None,
-                     main_program=None,
-                     startup_program=None):
-    helper = LayerHelper('linear_chain_crf', **locals())
-    size = input.shape[1]
-    transition = helper.create_parameter(
-        attr=helper.param_attr,
-        shape=[size + 2, size],
-        dtype=helper.input_dtype())
-    alpha = helper.create_tmp_variable(dtype=helper.input_dtype())
-    emission_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
-    transition_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
-    log_likelihood = helper.create_tmp_variable(dtype=helper.input_dtype())
-    helper.append_op(
-        type='linear_chain_crf',
-        inputs={"Emission": [input],
-                "Transition": transition,
-                "Label": label},
-        outputs={
-            "Alpha": [alpha],
-            "EmissionExps": [emission_exps],
-            "TransitionExps": transition_exps,
-            "LogLikelihood": log_likelihood
-        })
-
-    return log_likelihood
-
-
-def crf_decoding(input,
-                 param_attr,
-                 label=None,
-                 main_program=None,
-                 startup_program=None):
-    helper = LayerHelper('crf_decoding', **locals())
-    transition = helper.get_parameter(param_attr.name)
-    viterbi_path = helper.create_tmp_variable(dtype=helper.input_dtype())
-    helper.append_op(
-        type='crf_decoding',
-        inputs={"Emission": [input],
-                "Transition": transition,
-                "Label": label},
-        outputs={"ViterbiPath": [viterbi_path]})
-
-    return viterbi_path
-
-
-def assign(input, output, main_program=None, startup_program=None):
-    helper = LayerHelper('assign', **locals())
-    helper.append_op(
-        type='scale',
-        inputs={'X': [input]},
-        outputs={'Out': [output]},
-        attrs={'scale': 1.0})
-    return output
-
 
 def split_lod_tensor(input,
                      mask,
@@ -460,404 +54,6 @@ def merge_lod_tensor(in_true,
     return out
 
 
-def cos_sim(X, Y, **kwargs):
-    """
-    This function performs the cosine similarity between two tensors
-    X and Y and returns that as the output.
-    """
-    helper = LayerHelper('cos_sim', **kwargs)
-    out = helper.create_tmp_variable(dtype=X.dtype)
-    xnorm = helper.create_tmp_variable(dtype=X.dtype)
-    ynorm = helper.create_tmp_variable(dtype=X.dtype)
-    helper.append_op(
-        type='cos_sim',
-        inputs={'X': [X],
-                'Y': [Y]},
-        outputs={'Out': [out],
-                 'XNorm': [xnorm],
-                 'YNorm': [ynorm]})
-    return out
-
-
-def cross_entropy(input, label, **kwargs):
-    """
-    This function computes cross_entropy using the input and label.
-    """
-    helper = LayerHelper('cross_entropy', **kwargs)
-    out = helper.create_tmp_variable(dtype=input.dtype)
-    helper.append_op(
-        type='cross_entropy',
-        inputs={'X': [input],
-                'Label': [label]},
-        outputs={'Y': [out]},
-        attrs=kwargs)
-    return out
-
-
-def square_error_cost(input, label, **kwargs):
-    """
-    This functions returns the squared error cost using the input and label.
-    The output is appending the op to do the above.
-    """
-    helper = LayerHelper('square_error_cost', **kwargs)
-    minus_out = helper.create_tmp_variable(dtype=input.dtype)
-    helper.append_op(
-        type='elementwise_sub',
-        inputs={'X': [input],
-                'Y': [label]},
-        outputs={'Out': [minus_out]})
-
-    square_out = helper.create_tmp_variable(dtype=input.dtype)
-    helper.append_op(
-        type='square', inputs={'X': [minus_out]}, outputs={'Y': [square_out]})
-    return square_out
-
-
-def accuracy(input, label, k=1, correct=None, total=None, **kwargs):
-    """
-    This function computes the accuracy using the input and label.
-    The output is the top_k inputs and their indices.
-    """
-    helper = LayerHelper("accuracy", **kwargs)
-    topk_out = helper.create_tmp_variable(dtype=input.dtype)
-    topk_indices = helper.create_tmp_variable(dtype="int64")
-    helper.append_op(
-        type="top_k",
-        inputs={"X": [input]},
-        outputs={"Out": [topk_out],
-                 "Indices": [topk_indices]},
-        attrs={"k": k})
-    acc_out = helper.create_tmp_variable(dtype="float32")
-    if correct is None:
-        correct = helper.create_tmp_variable(dtype="int64")
-    if total is None:
-        total = helper.create_tmp_variable(dtype="int64")
-    helper.append_op(
-        type="accuracy",
-        inputs={
-            "Out": [topk_out],
-            "Indices": [topk_indices],
-            "Label": [label]
-        },
-        outputs={
-            "Accuracy": [acc_out],
-            "Correct": [correct],
-            "Total": [total],
-        })
-    return acc_out
-
-
-def chunk_eval(input,
-               label,
-               chunk_scheme,
-               num_chunk_types,
-               excluded_chunk_types=None,
-               **kwargs):
-    """
-    This function computes the accuracy using the input and label.
-    The output is the top_k inputs and their indices.
-    """
-    helper = LayerHelper("chunk_eval", **kwargs)
-
-    # prepare output
-    precision = helper.create_tmp_variable(dtype="float32")
-    recall = helper.create_tmp_variable(dtype="float32")
-    f1_score = helper.create_tmp_variable(dtype="float32")
-
-    helper.append_op(
-        type="chunk_eval",
-        inputs={"Inference": [input],
-                "Label": [label]},
-        outputs={
-            "Precision": [precision],
-            "Recall": [recall],
-            "F1-Score": [f1_score]
-        },
-        attrs={
-            "num_chunk_types": num_chunk_types,
-            'chunk_scheme': chunk_scheme,
-            'excluded_chunk_types': excluded_chunk_types or []
-        })
-    return precision, recall, f1_score
-
-
-def sequence_conv(input,
-                  num_filters,
-                  filter_size=3,
-                  filter_stride=1,
-                  padding=None,
-                  bias_attr=None,
-                  param_attr=None,
-                  act=None,
-                  main_program=None,
-                  startup_program=None):
-    """
-    This function creates the op for sequence_conv, using the inputs and
-    other convolutional configurations for the filters and stride as given
-    in the input parameters to the function.
-    """
-
-    # FIXME(dzh) : want to unify the argument of python layer
-    # function. So we ignore some unecessary attributes.
-    # such as, padding_trainable, context_start.
-
-    helper = LayerHelper('sequence_conv', **locals())
-    dtype = helper.input_dtype()
-    filter_shape = [filter_size * input.shape[1], num_filters]
-    filter_param = helper.create_parameter(
-        attr=helper.param_attr, shape=filter_shape, dtype=dtype)
-    pre_bias = helper.create_tmp_variable(dtype)
-
-    helper.append_op(
-        type='sequence_conv',
-        inputs={
-            'X': [input],
-            'Filter': [filter_param],
-        },
-        outputs={"Out": pre_bias},
-        attrs={
-            'contextStride': filter_stride,
-            'contextStart': -int(filter_size / 2),
-            'contextLength': filter_size
-        })
-    pre_act = helper.append_bias_op(pre_bias)
-    return helper.append_activation(pre_act)
-
-
-def conv2d(input,
-           num_filters,
-           filter_size,
-           stride=None,
-           padding=None,
-           groups=None,
-           param_attr=None,
-           bias_attr=None,
-           act=None,
-           name=None,
-           main_program=None,
-           startup_program=None):
-    """
-    This function creates the op for a 2-dimensional Convolution.
-    This is performed using the parameters of filters(size, dimensionality etc)
-    , stride and other configurations for a Convolution operation.
-    This funciton can also append an activation on top of the
-    conv-2d output, if mentioned in the input parameters.
-    """
-
-    if stride is None:
-        stride = [1, 1]
-    helper = LayerHelper('conv2d', **locals())
-    dtype = helper.input_dtype()
-
-    num_channels = input.shape[1]
-    if groups is None:
-        num_filter_channels = num_channels
-    else:
-        if num_channels % groups != 0:
-            raise ValueError("num_channels must be divisible by groups.")
-        num_filter_channels = num_channels / groups
-
-    if isinstance(filter_size, int):
-        filter_size = [filter_size, filter_size]
-    if isinstance(stride, int):
-        stride = [stride, stride]
-    if isinstance(padding, int):
-        padding = [padding, padding]
-
-    input_shape = input.shape
-    filter_shape = [num_filters, num_filter_channels] + filter_size
-
-    def _get_default_param_initializer():
-        std = (2.0 / (filter_size[0]**2 * num_channels))**0.5
-        return Normal(0.0, std, 0)
-
-    filter_param = helper.create_parameter(
-        attr=helper.param_attr,
-        shape=filter_shape,
-        dtype=dtype,
-        default_initializer=_get_default_param_initializer())
-
-    pre_bias = helper.create_tmp_variable(dtype)
-
-    helper.append_op(
-        type='conv2d_cudnn',
-        inputs={
-            'Input': input,
-            'Filter': filter_param,
-        },
-        outputs={"Output": pre_bias},
-        attrs={'strides': stride,
-               'paddings': padding,
-               'groups': groups})
-
-    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
-
-    return helper.append_activation(pre_act)
-
-
-def sequence_pool(input, pool_type, **kwargs):
-    """
-    This function add the operator for sequence pooling.
-    This is applied on top of the input using pool_type mentioned
-    in the parameters.
-    """
-    helper = LayerHelper('sequence_pool', input=input, **kwargs)
-    dtype = helper.input_dtype()
-    pool_out = helper.create_tmp_variable(dtype)
-    max_index = helper.create_tmp_variable(dtype)
-
-    helper.append_op(
-        type="sequence_pool",
-        inputs={"X": input},
-        outputs={"Out": pool_out,
-                 "MaxIndex": max_index},
-        attrs={"pooltype": pool_type.upper()})
-
-    return pool_out
-
-
-def pool2d(input,
-           pool_size,
-           pool_type,
-           pool_stride=None,
-           pool_padding=None,
-           global_pooling=False,
-           main_program=None,
-           startup_program=None):
-    """
-    This function adds the operator for pooling in 2 dimensions, using the
-    pooling configurations mentioned in input parameters.
-    """
-    if pool_padding is None:
-        pool_padding = [0, 0]
-    if pool_stride is None:
-        pool_stride = [1, 1]
-    if pool_type not in ["max", "avg"]:
-        raise ValueError(
-            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
-            str(pool_type))
-    if isinstance(pool_size, int):
-        pool_size = [pool_size, pool_size]
-    if isinstance(pool_stride, int):
-        pool_stride = [pool_stride, pool_stride]
-    if isinstance(pool_padding, int):
-        pool_padding = [pool_padding, pool_padding]
-
-    helper = LayerHelper('pool2d', **locals())
-    dtype = helper.input_dtype()
-    pool_out = helper.create_tmp_variable(dtype)
-
-    helper.append_op(
-        type="pool2d",
-        inputs={"X": input},
-        outputs={"Out": pool_out},
-        attrs={
-            "pooling_type": pool_type,
-            "ksize": pool_size,
-            "global_pooling": global_pooling,
-            "strides": pool_stride,
-            "paddings": pool_padding
-        })
-
-    return pool_out
-
-
-def batch_norm(input,
-               act=None,
-               is_test=False,
-               momentum=0.9,
-               epsilon=1e-05,
-               param_attr=None,
-               bias_attr=None,
-               data_layout='NCHW',
-               main_program=None,
-               startup_program=None):
-    """
-    This function helps create an operator to implement
-    the BatchNorm layer using the configurations from the input parameters.
-    """
-    helper = LayerHelper('batch_norm', **locals())
-    dtype = helper.input_dtype()
-
-    input_shape = input.shape
-    if data_layout == 'NCHW':
-        channel_num = input_shape[1]
-    else:
-        if data_layout == 'NHWC':
-            channel_num = input_shape[-1]
-        else:
-            raise ValueError("unsupported data layout:" + data_layout)
-
-    param_shape = [channel_num]
-
-    # create parameter
-    scale = helper.create_parameter(
-        attr=helper.param_attr,
-        shape=param_shape,
-        dtype=dtype,
-        default_initializer=Constant(1.0))
-
-    bias = helper.create_parameter(
-        attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=True)
-
-    mean = helper.create_global_variable(
-        dtype=input.dtype, shape=param_shape, persistable=True)
-    helper.set_variable_initializer(var=mean, initializer=Constant(0.0))
-
-    variance = helper.create_global_variable(
-        dtype=input.dtype, shape=param_shape, persistable=True)
-    helper.set_variable_initializer(var=variance, initializer=Constant(1.0))
-
-    # create output
-    # mean and mean_out share the same memory
-    mean_out = mean
-    # variance and variance out share the same memory
-    variance_out = variance
-    saved_mean = helper.create_tmp_variable(dtype)
-    saved_variance = helper.create_tmp_variable(dtype)
-
-    batch_norm_out = helper.create_tmp_variable(dtype)
-
-    helper.append_op(
-        type="batch_norm",
-        inputs={
-            "X": input,
-            "Scale": scale,
-            "Bias": bias,
-            "Mean": mean,
-            "Variance": variance
-        },
-        outputs={
-            "Y": batch_norm_out,
-            "MeanOut": mean_out,
-            "VarianceOut": variance_out,
-            "SavedMean": saved_mean,
-            "SavedVariance": saved_variance
-        },
-        attrs={"momentum": momentum,
-               "epsilon": epsilon,
-               "is_test": is_test})
-
-    return helper.append_activation(batch_norm_out)
-
-
-def beam_search_decode(ids, scores, main_program=None, startup_program=None):
-    helper = LayerHelper('beam_search_decode', **locals())
-    sentence_ids = helper.create_tmp_variable(dtype=ids.dtype)
-    sentence_scores = helper.create_tmp_variable(dtype=ids.dtype)
-
-    helper.append_op(
-        type="beam_search_decode",
-        inputs={"Ids": ids,
-                "Scores": scores},
-        outputs={
-            "SentenceIds": sentence_ids,
-            "SentenceScores": sentence_scores
-        })
-
-    return sentence_ids, sentence_scores
-
-
 class BlockGuard(object):
     """
     BlockGuard class.
@@ -1210,50 +406,6 @@ class While(object):
             attrs={'sub_block': while_block})
 
 
-def lstm(x,
-         c_pre_init,
-         hidden_dim,
-         forget_bias=None,
-         main_program=None,
-         startup_program=None):
-    """
-    This function helps create an operator for the LSTM (Long Short Term
-    Memory) cell that can be used inside an RNN.
-    """
-    helper = LayerHelper('lstm_unit', **locals())
-    rnn = StaticRNN()
-    with rnn.step():
-        c_pre = rnn.memory(init=c_pre_init)
-        x_t = rnn.step_input(x)
-
-        before_fc = concat(
-            input=[x_t, c_pre],
-            axis=1,
-            main_program=main_program,
-            startup_program=startup_program)
-        after_fc = fc(input=before_fc,
-                      size=hidden_dim * 4,
-                      main_program=main_program,
-                      startup_program=startup_program)
-
-        dtype = x.dtype
-        c = helper.create_tmp_variable(dtype)
-        h = helper.create_tmp_variable(dtype)
-
-        helper.append_op(
-            type='lstm_unit',
-            inputs={"X": after_fc,
-                    "C_prev": c_pre},
-            outputs={"C": c,
-                     "H": h},
-            attrs={"forget_bias": forget_bias})
-
-        rnn.update_memory(c_pre, c)
-        rnn.output(h)
-
-    return rnn()
-
-
 def lod_rank_table(x, level=0, main_program=None):
     """
     This function creates an operator for creating a LOD_RANK_TABLE
@@ -1331,72 +483,6 @@ def array_to_lod_tensor(x, table, main_program=None, startup_program=None):
     return tmp
 
 
-def fill_constant(shape,
-                  dtype,
-                  value,
-                  out=None,
-                  main_program=None,
-                  startup_program=None):
-    """
-    This function creates a tensor , with shape as mentioned in the input and
-    specified dtype and fills this up with a constant value that
-    comes in the input. It also sets the stop_gradient to be True.
-    """
-    helper = LayerHelper("fill_constant", **locals())
-    if out is None:
-        out = helper.create_tmp_variable(dtype=dtype)
-    helper.append_op(
-        type='fill_constant',
-        inputs={},
-        outputs={'Out': [out]},
-        attrs={'shape': shape,
-               'dtype': out.dtype,
-               'value': float(value)})
-    out.stop_gradient = True
-    return out
-
-
-def fill_constant_batch_size_like(input,
-                                  shape,
-                                  dtype,
-                                  value,
-                                  input_dim_idx=0,
-                                  output_dim_idx=0,
-                                  main_program=None,
-                                  startup_program=None):
-    helper = LayerHelper("fill_constant_batch_size_like", **locals())
-    out = helper.create_tmp_variable(dtype=dtype)
-    helper.append_op(
-        type='fill_constant_batch_size_like',
-        inputs={'Input': input},
-        outputs={'Out': [out]},
-        attrs={
-            'shape': shape,
-            'dtype': out.dtype,
-            'value': float(value),
-            'input_dim_idx': input_dim_idx,
-            'output_dim_idx': output_dim_idx
-        })
-    out.stop_gradient = True
-    return out
-
-
-def ones(shape, dtype, main_program=None):
-    """
-    This function performs the same function as fill_constant() declared above
-    with the constant value being 1.0.
-    """
-    return fill_constant(value=1.0, **locals())
-
-
-def zeros(shape, dtype, main_program=None):
-    """
-    This function performs the same function as fill_constant() declared above
-    with the constant value being 0.0.
-    """
-    return fill_constant(value=0.0, **locals())
-
-
 def increment(x,
               value=1.0,
               in_place=True,
@@ -1508,95 +594,6 @@ def array_length(array, main_program=None):
     return tmp
 
 
-def conv2d_transpose(input,
-                     num_filters,
-                     output_size=None,
-                     filter_size=None,
-                     padding=None,
-                     stride=None,
-                     param_attr=None,
-                     main_program=None,
-                     startup_program=None):
-    """
-    The transpose of conv2d layer.
-
-    This layer is also known as deconvolution layer.
-
-    Args:
-        input(Variable): The input image with [N, C, H, W] format.
-        num_filters(int): The number of filter. It is as same as the output
-            image channel.
-        output_size(int|tuple|None): The output image size. If output size is a
-            tuple, it must contain two integers, (image_H, image_W). This
-            parameter only works when filter_size is None.
-        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
-            it must contain two integers, (filter_size_H, filter_size_W).
-            Otherwise, the filter will be a square.  None if use output size to
-            calculate filter_size
-        padding(int|tuple): The padding size. If padding is a tuple, it must
-            contain two integers, (padding_H, padding_W). Otherwise, the
-            padding_H = padding_W = padding.
-        stride(int|tuple): The stride size. If stride is a tuple, it must
-            contain two integers, (stride_H, stride_W). Otherwise, the
-            stride_H = stride_W = stride.
-        param_attr: Parameter Attribute.
-        main_program(Program): the main program
-        startup_program(Program): the startup program
-
-    Returns:
-        Variable: Output image.
-    """
-    helper = LayerHelper("conv2d_transpose", **locals())
-    if not isinstance(input, Variable):
-        raise TypeError("Input of conv2d_transpose must be Variable")
-    input_channel = input.shape[1]
-
-    op_attr = dict()
-
-    if isinstance(padding, int):
-        op_attr['paddings'] = [padding, padding]
-    elif padding is not None:
-        op_attr['paddings'] = padding
-
-    if isinstance(stride, int):
-        op_attr['strides'] = stride
-    elif stride is not None:
-        op_attr['strides'] = stride
-
-    if filter_size is None:
-        if output_size is None:
-            raise ValueError("output_size must be set when filter_size is None")
-        if isinstance(output_size, int):
-            output_size = [output_size, output_size]
-
-        padding = op_attr.get('paddings', [0, 0])
-        stride = op_attr.get('strides', [1, 1])
-
-        h_in = input.shape[2]
-        w_in = input.shape[3]
-        filter_size_h = output_size[0] - \
-            (h_in - 1) * stride[0] + 2 * padding[0]
-        filter_size_w = output_size[1] - \
-            (w_in - 1) * stride[1] + 2 * padding[1]
-        filter_size = [filter_size_h, filter_size_w]
-    elif isinstance(filter_size, int):
-        filter_size = [filter_size, filter_size]
-
-    filter_shape = [input_channel, num_filters] + filter_size
-    img_filter = helper.create_parameter(
-        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)
-
-    out = helper.create_tmp_variable(dtype=input.dtype)
-    helper.append_op(
-        type='conv2d_transpose',
-        inputs={'Input': [input],
-                'Filter': [img_filter]},
-        outputs={'Output': out},
-        attrs=op_attr)
-
-    return out
-
-
 class ConditionalBlockGuard(BlockGuard):
     def __init__(self, block):
         if not isinstance(block, ConditionalBlock):
diff --git a/python/paddle/v2/fluid/layers/io.py b/python/paddle/v2/fluid/layers/io.py
new file mode 100644
index 0000000000..f03d8e3c3e
--- /dev/null
+++ b/python/paddle/v2/fluid/layers/io.py
@@ -0,0 +1,57 @@
+from .. import core
+from ..layer_helper import LayerHelper
+
+__all__ = ['data']
+
+
+def data(name,
+         shape,
+         append_batch_size=True,
+         dtype='float32',
+         lod_level=0,
+         type=core.VarDesc.VarType.LOD_TENSOR,
+         main_program=None,
+         startup_program=None,
+         stop_gradient=True):
+    """
+    Data Layer.
+
+    Args:
+       name: The name/alias of the function
+       shape: Tuple declaring the shape.
+       append_batch_size: Whether or not to append the data as a batch.
+       dtype: The type of data : float32, float_16, int etc
+       type: The output type. By default it is LOD_TENSOR.
+       lod_level(int): The LoD Level. 0 means the input data is not a sequence.
+       main_program: Name of the main program that calls this
+       startup_program: Name of the startup program
+       stop_gradient: A boolean that mentions whether gradient should flow.
+
+    This function takes in input and based on whether data has
+    to be returned back as a minibatch, it creates the global variable using
+    the helper functions. The global variables can be accessed by all the
+    following operations and layers in the graph.
+
+    All the input variables of this function are passed in as local variables
+    to the LayerHelper constructor.
+
+    """
+    helper = LayerHelper('data', **locals())
+    shape = list(shape)
+    for i in xrange(len(shape)):
+        if shape[i] is None:
+            shape[i] = -1
+            append_batch_size = False
+        elif shape[i] < 0:
+            append_batch_size = False
+
+    if append_batch_size:
+        shape = [-1] + shape  # append batch size as -1
+
+    return helper.create_global_variable(
+        name=name,
+        shape=shape,
+        dtype=dtype,
+        type=type,
+        stop_gradient=stop_gradient,
+        lod_level=lod_level)
diff --git a/python/paddle/v2/fluid/layers/nn.py b/python/paddle/v2/fluid/layers/nn.py
new file mode 100644
index 0000000000..f231f38b3e
--- /dev/null
+++ b/python/paddle/v2/fluid/layers/nn.py
@@ -0,0 +1,785 @@
+"""
+All layers just related to the neural network.
+"""
+
+from ..layer_helper import LayerHelper
+from ..initializer import Normal, Constant
+from ..framework import Variable
+
+__all__ = [
+    'fc', 'embedding', 'dynamic_lstm', 'gru_unit', 'linear_chain_crf',
+    'crf_decoding', 'cos_sim', 'cross_entropy', 'square_error_cost', 'accuracy',
+    'chunk_eval', 'sequence_conv', 'conv2d', 'sequence_pool', 'pool2d',
+    'batch_norm', 'beam_search_decode', 'conv2d_transpose'
+]
+
+
+def fc(input,
+       size,
+       num_flatten_dims=1,
+       param_attr=None,
+       bias_attr=None,
+       act=None,
+       name=None,
+       main_program=None,
+       startup_program=None):
+    """
+    Fully Connected Layer.
+
+    Args:
+       input: The input tensor to the function
+       size: The size of the layer
+       num_flatten_dims: Number of columns in input
+       param_attr: The parameters/weights to the FC Layer
+       param_initializer: Initializer used for the weight/parameter. If None, XavierInitializer() is used
+       bias_attr: The bias parameter for the FC layer
+       bias_initializer: Initializer used for the bias. If None, then ConstantInitializer() is used
+       act: Activation to be applied to the output of FC layer
+       name: Name/alias of the function
+       main_program: Name of the main program that calls this
+       startup_program: Name of the startup program
+
+    This function can take in multiple inputs and performs the Fully Connected
+    function (linear transformation) on top of each of them.
+    So for input x, the output will be : Wx + b. Where W is the parameter,
+    b the bias and x is the input.
+
+    The function also applies an activation (non-linearity) on top of the
+    output, if activation is passed in the input.
+
+    All the input variables of this function are passed in as local variables
+    to the LayerHelper constructor.
+
+    """
+    helper = LayerHelper('fc', **locals())
+
+    dtype = helper.input_dtype()
+
+    mul_results = []
+    for input_var, param_attr in helper.iter_inputs_and_params():
+        input_shape = input_var.shape
+        param_shape = [
+            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
+        ] + [size]
+        w = helper.create_parameter(
+            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
+        tmp = helper.create_tmp_variable(dtype)
+        helper.append_op(
+            type="mul",
+            inputs={
+                "X": input_var,
+                "Y": w,
+            },
+            outputs={"Out": tmp},
+            attrs={'x_num_col_dims': num_flatten_dims,
+                   'y_num_col_dims': 1})
+        mul_results.append(tmp)
+
+    # sum
+    if len(mul_results) == 1:
+        pre_bias = mul_results[0]
+    else:
+        pre_bias = helper.create_tmp_variable(dtype)
+        helper.append_op(
+            type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias})
+    # add bias
+    pre_activation = helper.append_bias_op(pre_bias)
+    # add activation
+    return helper.append_activation(pre_activation)
+
+
+def embedding(input,
+              size,
+              is_sparse=False,
+              param_attr=None,
+              dtype='float32',
+              main_program=None,
+              startup_program=None):
+    """
+    Embedding Layer.
+
+    Args:
+       param_initializer:
+       input: The input to the function
+       size: The size of the layer
+       is_sparse: A flag that decleares whether the input is sparse
+       param_attr: Parameters for this layer
+       dtype: The type of data : float32, float_16, int etc
+       main_program: Name of the main program that calls this
+       startup_program: Name of the startup program
+
+    This function can take in the input (which is a vector of IDs) and
+    performs a lookup in the lookup_table using these IDs, to result into
+    the embedding of each ID in the input.
+
+    All the input variables of this function are passed in as local variables
+    to the LayerHelper constructor.
+
+    """
+
+    helper = LayerHelper('embedding', **locals())
+    w = helper.create_parameter(
+        attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False)
+    tmp = helper.create_tmp_variable(dtype)
+    helper.append_op(
+        type='lookup_table',
+        inputs={'Ids': input,
+                'W': w},
+        outputs={'Out': tmp},
+        attrs={'is_sparse': is_sparse})
+    return tmp
+
+
+# TODO(qijun): expose H0 and C0
+def dynamic_lstm(input,
+                 size,
+                 param_attr=None,
+                 bias_attr=None,
+                 use_peepholes=True,
+                 is_reverse=False,
+                 gate_activation='sigmoid',
+                 cell_activation='tanh',
+                 candidate_activation='tanh',
+                 dtype='float32',
+                 main_program=None,
+                 startup_program=None):
+    helper = LayerHelper('lstm', **locals())
+    size = size / 4
+    weight = helper.create_parameter(
+        attr=helper.param_attr, shape=[size, 4 * size], dtype=dtype)
+    bias_size = [1, 7 * size]
+    if not use_peepholes:
+        bias_size[1] = 4 * size
+    bias = helper.create_parameter(
+        attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
+
+    hidden = helper.create_tmp_variable(dtype)
+    cell = helper.create_tmp_variable(dtype)
+    batch_gate = helper.create_tmp_variable(dtype)
+    batch_cell_pre_act = helper.create_tmp_variable(dtype)
+
+    helper.append_op(
+        type='lstm',
+        inputs={'Input': input,
+                'Weight': weight,
+                'Bias': bias},
+        outputs={
+            'Hidden': hidden,
+            'Cell': cell,
+            'BatchGate': batch_gate,
+            'BatchCellPreAct': batch_cell_pre_act
+        },
+        attrs={
+            'use_peepholes': use_peepholes,
+            'is_reverse': is_reverse,
+            'gate_activation': gate_activation,
+            'cell_activation': cell_activation,
+            'candidate_activation': candidate_activation
+        })
+    return hidden, cell
+
+
+def gru_unit(input,
+             hidden,
+             size,
+             weight=None,
+             bias=None,
+             activation='tanh',
+             gate_activation='sigmoid',
+             main_program=None,
+             startup_program=None):
+    """
+    GRUUnit Operator implements partial calculations of the GRU unit as following:
+
+    $$
+    update \ gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\
+    reset \ gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r)  \\
+    output \ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\
+    output: h_t = dot((1 - u_t), h_{t-1}) + dot(u_t, {h}_t)
+    $$
+
+    which is same as one time step of GRU Operator.
+
+    @note To implement the complete GRU unit, fully-connected operator must be
+    used before to feed xu, xr and xc as the Input of GRUUnit operator.
+
+    TODO(ChunweiYan) add more document here
+    """
+    activation_dict = dict(
+        identity=0,
+        sigmoid=1,
+        tanh=2,
+        relu=3, )
+    activation = activation_dict[activation]
+    gate_activation = activation_dict[gate_activation]
+
+    helper = LayerHelper('gru_unit', **locals())
+    dtype = helper.input_dtype()
+    size = size / 3
+
+    # create weight
+    if weight is None:
+        weight = helper.create_parameter(
+            attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
+
+    # create bias
+    if bias is None:
+        bias_size = [1, 3 * size]
+        bias = helper.create_parameter(
+            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
+
+    gate = helper.create_tmp_variable(dtype)
+    reset_hidden_pre = helper.create_tmp_variable(dtype)
+    updated_hidden = helper.create_tmp_variable(dtype)
+
+    helper.append_op(
+        type='gru_unit',
+        inputs={'Input': input,
+                'HiddenPrev': hidden,
+                'Weight': weight},
+        outputs={
+            'Gate': gate,
+            'ResetHiddenPrev': reset_hidden_pre,
+            'Hidden': updated_hidden,
+        },
+        attrs={
+            'activation': 0,
+            'gate_activation': 1,
+        })
+
+    return updated_hidden, reset_hidden_pre, gate
+
+
+def linear_chain_crf(input,
+                     label,
+                     param_attr=None,
+                     main_program=None,
+                     startup_program=None):
+    helper = LayerHelper('linear_chain_crf', **locals())
+    size = input.shape[1]
+    transition = helper.create_parameter(
+        attr=helper.param_attr,
+        shape=[size + 2, size],
+        dtype=helper.input_dtype())
+    alpha = helper.create_tmp_variable(dtype=helper.input_dtype())
+    emission_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
+    transition_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
+    log_likelihood = helper.create_tmp_variable(dtype=helper.input_dtype())
+    helper.append_op(
+        type='linear_chain_crf',
+        inputs={"Emission": [input],
+                "Transition": transition,
+                "Label": label},
+        outputs={
+            "Alpha": [alpha],
+            "EmissionExps": [emission_exps],
+            "TransitionExps": transition_exps,
+            "LogLikelihood": log_likelihood
+        })
+
+    return log_likelihood
+
+
+def crf_decoding(input,
+                 param_attr,
+                 label=None,
+                 main_program=None,
+                 startup_program=None):
+    helper = LayerHelper('crf_decoding', **locals())
+    transition = helper.get_parameter(param_attr.name)
+    viterbi_path = helper.create_tmp_variable(dtype=helper.input_dtype())
+    helper.append_op(
+        type='crf_decoding',
+        inputs={"Emission": [input],
+                "Transition": transition,
+                "Label": label},
+        outputs={"ViterbiPath": [viterbi_path]})
+
+    return viterbi_path
+
+
+def cos_sim(X, Y, **kwargs):
+    """
+    This function performs the cosine similarity between two tensors
+    X and Y and returns that as the output.
+    """
+    helper = LayerHelper('cos_sim', **kwargs)
+    out = helper.create_tmp_variable(dtype=X.dtype)
+    xnorm = helper.create_tmp_variable(dtype=X.dtype)
+    ynorm = helper.create_tmp_variable(dtype=X.dtype)
+    helper.append_op(
+        type='cos_sim',
+        inputs={'X': [X],
+                'Y': [Y]},
+        outputs={'Out': [out],
+                 'XNorm': [xnorm],
+                 'YNorm': [ynorm]})
+    return out
+
+
+def cross_entropy(input, label, **kwargs):
+    """
+    This function computes cross_entropy using the input and label.
+    """
+    helper = LayerHelper('cross_entropy', **kwargs)
+    out = helper.create_tmp_variable(dtype=input.dtype)
+    helper.append_op(
+        type='cross_entropy',
+        inputs={'X': [input],
+                'Label': [label]},
+        outputs={'Y': [out]},
+        attrs=kwargs)
+    return out
+
+
+def square_error_cost(input, label, **kwargs):
+    """
+    This functions returns the squared error cost using the input and label.
+    The output is appending the op to do the above.
+    """
+    helper = LayerHelper('square_error_cost', **kwargs)
+    minus_out = helper.create_tmp_variable(dtype=input.dtype)
+    helper.append_op(
+        type='elementwise_sub',
+        inputs={'X': [input],
+                'Y': [label]},
+        outputs={'Out': [minus_out]})
+
+    square_out = helper.create_tmp_variable(dtype=input.dtype)
+    helper.append_op(
+        type='square', inputs={'X': [minus_out]}, outputs={'Y': [square_out]})
+    return square_out
+
+
+def accuracy(input, label, k=1, correct=None, total=None, **kwargs):
+    """
+    This function computes the accuracy using the input and label.
+    The output is the top_k inputs and their indices.
+    """
+    helper = LayerHelper("accuracy", **kwargs)
+    topk_out = helper.create_tmp_variable(dtype=input.dtype)
+    topk_indices = helper.create_tmp_variable(dtype="int64")
+    helper.append_op(
+        type="top_k",
+        inputs={"X": [input]},
+        outputs={"Out": [topk_out],
+                 "Indices": [topk_indices]},
+        attrs={"k": k})
+    acc_out = helper.create_tmp_variable(dtype="float32")
+    if correct is None:
+        correct = helper.create_tmp_variable(dtype="int64")
+    if total is None:
+        total = helper.create_tmp_variable(dtype="int64")
+    helper.append_op(
+        type="accuracy",
+        inputs={
+            "Out": [topk_out],
+            "Indices": [topk_indices],
+            "Label": [label]
+        },
+        outputs={
+            "Accuracy": [acc_out],
+            "Correct": [correct],
+            "Total": [total],
+        })
+    return acc_out
+
+
+def chunk_eval(input,
+               label,
+               chunk_scheme,
+               num_chunk_types,
+               excluded_chunk_types=None,
+               **kwargs):
+    """
+    This function computes the accuracy using the input and label.
+    The output is the top_k inputs and their indices.
+    """
+    helper = LayerHelper("chunk_eval", **kwargs)
+
+    # prepare output
+    precision = helper.create_tmp_variable(dtype="float32")
+    recall = helper.create_tmp_variable(dtype="float32")
+    f1_score = helper.create_tmp_variable(dtype="float32")
+
+    helper.append_op(
+        type="chunk_eval",
+        inputs={"Inference": [input],
+                "Label": [label]},
+        outputs={
+            "Precision": [precision],
+            "Recall": [recall],
+            "F1-Score": [f1_score]
+        },
+        attrs={
+            "num_chunk_types": num_chunk_types,
+            'chunk_scheme': chunk_scheme,
+            'excluded_chunk_types': excluded_chunk_types or []
+        })
+    return precision, recall, f1_score
+
+
+def sequence_conv(input,
+                  num_filters,
+                  filter_size=3,
+                  filter_stride=1,
+                  padding=None,
+                  bias_attr=None,
+                  param_attr=None,
+                  act=None,
+                  main_program=None,
+                  startup_program=None):
+    """
+    This function creates the op for sequence_conv, using the inputs and
+    other convolutional configurations for the filters and stride as given
+    in the input parameters to the function.
+    """
+
+    # FIXME(dzh) : want to unify the argument of python layer
+    # function. So we ignore some unecessary attributes.
+    # such as, padding_trainable, context_start.
+
+    helper = LayerHelper('sequence_conv', **locals())
+    dtype = helper.input_dtype()
+    filter_shape = [filter_size * input.shape[1], num_filters]
+    filter_param = helper.create_parameter(
+        attr=helper.param_attr, shape=filter_shape, dtype=dtype)
+    pre_bias = helper.create_tmp_variable(dtype)
+
+    helper.append_op(
+        type='sequence_conv',
+        inputs={
+            'X': [input],
+            'Filter': [filter_param],
+        },
+        outputs={"Out": pre_bias},
+        attrs={
+            'contextStride': filter_stride,
+            'contextStart': -int(filter_size / 2),
+            'contextLength': filter_size
+        })
+    pre_act = helper.append_bias_op(pre_bias)
+    return helper.append_activation(pre_act)
+
+
+def conv2d(input,
+           num_filters,
+           filter_size,
+           stride=None,
+           padding=None,
+           groups=None,
+           param_attr=None,
+           bias_attr=None,
+           act=None,
+           name=None,
+           main_program=None,
+           startup_program=None):
+    """
+    This function creates the op for a 2-dimensional Convolution.
+    This is performed using the parameters of filters(size, dimensionality etc)
+    , stride and other configurations for a Convolution operation.
+    This funciton can also append an activation on top of the
+    conv-2d output, if mentioned in the input parameters.
+    """
+
+    if stride is None:
+        stride = [1, 1]
+    helper = LayerHelper('conv2d', **locals())
+    dtype = helper.input_dtype()
+
+    num_channels = input.shape[1]
+    if groups is None:
+        num_filter_channels = num_channels
+    else:
+        if num_channels % groups != 0:
+            raise ValueError("num_channels must be divisible by groups.")
+        num_filter_channels = num_channels / groups
+
+    if isinstance(filter_size, int):
+        filter_size = [filter_size, filter_size]
+    if isinstance(stride, int):
+        stride = [stride, stride]
+    if isinstance(padding, int):
+        padding = [padding, padding]
+
+    input_shape = input.shape
+    filter_shape = [num_filters, num_filter_channels] + filter_size
+
+    def _get_default_param_initializer():
+        std = (2.0 / (filter_size[0]**2 * num_channels))**0.5
+        return Normal(0.0, std, 0)
+
+    filter_param = helper.create_parameter(
+        attr=helper.param_attr,
+        shape=filter_shape,
+        dtype=dtype,
+        default_initializer=_get_default_param_initializer())
+
+    pre_bias = helper.create_tmp_variable(dtype)
+
+    helper.append_op(
+        type='conv2d_cudnn',
+        inputs={
+            'Input': input,
+            'Filter': filter_param,
+        },
+        outputs={"Output": pre_bias},
+        attrs={'strides': stride,
+               'paddings': padding,
+               'groups': groups})
+
+    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
+
+    return helper.append_activation(pre_act)
+
+
+def sequence_pool(input, pool_type, **kwargs):
+    """
+    This function add the operator for sequence pooling.
+    This is applied on top of the input using pool_type mentioned
+    in the parameters.
+    """
+    helper = LayerHelper('sequence_pool', input=input, **kwargs)
+    dtype = helper.input_dtype()
+    pool_out = helper.create_tmp_variable(dtype)
+    max_index = helper.create_tmp_variable(dtype)
+
+    helper.append_op(
+        type="sequence_pool",
+        inputs={"X": input},
+        outputs={"Out": pool_out,
+                 "MaxIndex": max_index},
+        attrs={"pooltype": pool_type.upper()})
+
+    return pool_out
+
+
+def pool2d(input,
+           pool_size,
+           pool_type,
+           pool_stride=None,
+           pool_padding=None,
+           global_pooling=False,
+           main_program=None,
+           startup_program=None):
+    """
+    This function adds the operator for pooling in 2 dimensions, using the
+    pooling configurations mentioned in input parameters.
+    """
+    if pool_padding is None:
+        pool_padding = [0, 0]
+    if pool_stride is None:
+        pool_stride = [1, 1]
+    if pool_type not in ["max", "avg"]:
+        raise ValueError(
+            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
+            str(pool_type))
+    if isinstance(pool_size, int):
+        pool_size = [pool_size, pool_size]
+    if isinstance(pool_stride, int):
+        pool_stride = [pool_stride, pool_stride]
+    if isinstance(pool_padding, int):
+        pool_padding = [pool_padding, pool_padding]
+
+    helper = LayerHelper('pool2d', **locals())
+    dtype = helper.input_dtype()
+    pool_out = helper.create_tmp_variable(dtype)
+
+    helper.append_op(
+        type="pool2d",
+        inputs={"X": input},
+        outputs={"Out": pool_out},
+        attrs={
+            "pooling_type": pool_type,
+            "ksize": pool_size,
+            "global_pooling": global_pooling,
+            "strides": pool_stride,
+            "paddings": pool_padding
+        })
+
+    return pool_out
+
+
+def batch_norm(input,
+               act=None,
+               is_test=False,
+               momentum=0.9,
+               epsilon=1e-05,
+               param_attr=None,
+               bias_attr=None,
+               data_layout='NCHW',
+               main_program=None,
+               startup_program=None):
+    """
+    This function helps create an operator to implement
+    the BatchNorm layer using the configurations from the input parameters.
+    """
+    helper = LayerHelper('batch_norm', **locals())
+    dtype = helper.input_dtype()
+
+    input_shape = input.shape
+    if data_layout == 'NCHW':
+        channel_num = input_shape[1]
+    else:
+        if data_layout == 'NHWC':
+            channel_num = input_shape[-1]
+        else:
+            raise ValueError("unsupported data layout:" + data_layout)
+
+    param_shape = [channel_num]
+
+    # create parameter
+    scale = helper.create_parameter(
+        attr=helper.param_attr,
+        shape=param_shape,
+        dtype=dtype,
+        default_initializer=Constant(1.0))
+
+    bias = helper.create_parameter(
+        attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=True)
+
+    mean = helper.create_global_variable(
+        dtype=input.dtype, shape=param_shape, persistable=True)
+    helper.set_variable_initializer(var=mean, initializer=Constant(0.0))
+
+    variance = helper.create_global_variable(
+        dtype=input.dtype, shape=param_shape, persistable=True)
+    helper.set_variable_initializer(var=variance, initializer=Constant(1.0))
+
+    # create output
+    # mean and mean_out share the same memory
+    mean_out = mean
+    # variance and variance out share the same memory
+    variance_out = variance
+    saved_mean = helper.create_tmp_variable(dtype)
+    saved_variance = helper.create_tmp_variable(dtype)
+
+    batch_norm_out = helper.create_tmp_variable(dtype)
+
+    helper.append_op(
+        type="batch_norm",
+        inputs={
+            "X": input,
+            "Scale": scale,
+            "Bias": bias,
+            "Mean": mean,
+            "Variance": variance
+        },
+        outputs={
+            "Y": batch_norm_out,
+            "MeanOut": mean_out,
+            "VarianceOut": variance_out,
+            "SavedMean": saved_mean,
+            "SavedVariance": saved_variance
+        },
+        attrs={"momentum": momentum,
+               "epsilon": epsilon,
+               "is_test": is_test})
+
+    return helper.append_activation(batch_norm_out)
+
+
+def beam_search_decode(ids, scores, main_program=None, startup_program=None):
+    helper = LayerHelper('beam_search_decode', **locals())
+    sentence_ids = helper.create_tmp_variable(dtype=ids.dtype)
+    sentence_scores = helper.create_tmp_variable(dtype=ids.dtype)
+
+    helper.append_op(
+        type="beam_search_decode",
+        inputs={"Ids": ids,
+                "Scores": scores},
+        outputs={
+            "SentenceIds": sentence_ids,
+            "SentenceScores": sentence_scores
+        })
+
+    return sentence_ids, sentence_scores
+
+
+def conv2d_transpose(input,
+                     num_filters,
+                     output_size=None,
+                     filter_size=None,
+                     padding=None,
+                     stride=None,
+                     param_attr=None,
+                     main_program=None,
+                     startup_program=None):
+    """
+    The transpose of conv2d layer.
+
+    This layer is also known as deconvolution layer.
+
+    Args:
+        input(Variable): The input image with [N, C, H, W] format.
+        num_filters(int): The number of filter. It is as same as the output
+            image channel.
+        output_size(int|tuple|None): The output image size. If output size is a
+            tuple, it must contain two integers, (image_H, image_W). This
+            parameter only works when filter_size is None.
+        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
+            it must contain two integers, (filter_size_H, filter_size_W).
+            Otherwise, the filter will be a square.  None if use output size to
+            calculate filter_size
+        padding(int|tuple): The padding size. If padding is a tuple, it must
+            contain two integers, (padding_H, padding_W). Otherwise, the
+            padding_H = padding_W = padding.
+        stride(int|tuple): The stride size. If stride is a tuple, it must
+            contain two integers, (stride_H, stride_W). Otherwise, the
+            stride_H = stride_W = stride.
+        param_attr: Parameter Attribute.
+        main_program(Program): the main program
+        startup_program(Program): the startup program
+
+    Returns:
+        Variable: Output image.
+    """
+    helper = LayerHelper("conv2d_transpose", **locals())
+    if not isinstance(input, Variable):
+        raise TypeError("Input of conv2d_transpose must be Variable")
+    input_channel = input.shape[1]
+
+    op_attr = dict()
+
+    if isinstance(padding, int):
+        op_attr['paddings'] = [padding, padding]
+    elif padding is not None:
+        op_attr['paddings'] = padding
+
+    if isinstance(stride, int):
+        op_attr['strides'] = stride
+    elif stride is not None:
+        op_attr['strides'] = stride
+
+    if filter_size is None:
+        if output_size is None:
+            raise ValueError("output_size must be set when filter_size is None")
+        if isinstance(output_size, int):
+            output_size = [output_size, output_size]
+
+        padding = op_attr.get('paddings', [0, 0])
+        stride = op_attr.get('strides', [1, 1])
+
+        h_in = input.shape[2]
+        w_in = input.shape[3]
+        filter_size_h = output_size[0] - \
+                        (h_in - 1) * stride[0] + 2 * padding[0]
+        filter_size_w = output_size[1] - \
+                        (w_in - 1) * stride[1] + 2 * padding[1]
+        filter_size = [filter_size_h, filter_size_w]
+    elif isinstance(filter_size, int):
+        filter_size = [filter_size, filter_size]
+
+    filter_shape = [input_channel, num_filters] + filter_size
+    img_filter = helper.create_parameter(
+        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)
+
+    out = helper.create_tmp_variable(dtype=input.dtype)
+    helper.append_op(
+        type='conv2d_transpose',
+        inputs={'Input': [input],
+                'Filter': [img_filter]},
+        outputs={'Output': out},
+        attrs=op_attr)
+
+    return out
diff --git a/python/paddle/v2/fluid/layers/ops.py b/python/paddle/v2/fluid/layers/ops.py
new file mode 100644
index 0000000000..fa312ace60
--- /dev/null
+++ b/python/paddle/v2/fluid/layers/ops.py
@@ -0,0 +1,9 @@
+from ..registry import register_layer
+__all__ = [
+    'mean', 'mul', 'dropout', 'reshape', 'sigmoid', 'scale', 'transpose',
+    'sigmoid_cross_entropy_with_logits', 'elementwise_add', 'elementwise_div',
+    'elementwise_sub', 'elementwise_mul', 'clip', 'abs'
+]
+
+for _OP in set(__all__):
+    globals()[_OP] = register_layer(_OP)
diff --git a/python/paddle/v2/fluid/layers/tensor.py b/python/paddle/v2/fluid/layers/tensor.py
new file mode 100644
index 0000000000..a839ed897d
--- /dev/null
+++ b/python/paddle/v2/fluid/layers/tensor.py
@@ -0,0 +1,130 @@
+from ..layer_helper import LayerHelper
+
+__all__ = [
+    'create_tensor', 'cast', 'concat', 'sums', 'assign',
+    'fill_constant_batch_size_like', 'fill_constant', 'ones', 'zeros'
+]
+
+
+def create_tensor(dtype, name=None, main_program=None, startup_program=None):
+    helper = LayerHelper("create_tensor", **locals())
+    return helper.create_variable(name=helper.name, dtype=dtype)
+
+
+def cast(x, dtype, main_program=None):
+    """
+    This function takes in the input with input_dtype
+    and casts it to the output_dtype as the output.
+    """
+    helper = LayerHelper('cast', **locals())
+    out = helper.create_tmp_variable(dtype=dtype)
+    helper.append_op(
+        type='cast',
+        inputs={'X': [x]},
+        outputs={'Out': [out]},
+        attrs={'in_dtype': x.dtype,
+               'out_dtype': out.dtype})
+    return out
+
+
+def concat(input, axis, main_program=None, startup_program=None):
+    """
+    This function concats the input along the axis mentioned
+    and returns that as the output.
+    """
+    helper = LayerHelper('concat', **locals())
+    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    helper.append_op(
+        type='concat',
+        inputs={'X': input},
+        outputs={'Out': [out]},
+        attrs={'axis': axis})
+    return out
+
+
+def sums(input, out=None, main_program=None, startup_program=None):
+    """
+    This function takes in the input and performs the sum operation on it
+    and returns that as the output.
+    """
+    helper = LayerHelper('sum', **locals())
+    if out is None:
+        out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    helper.append_op(type='sum', inputs={'X': input}, outputs={'Out': out})
+    return out
+
+
+def assign(input, output, main_program=None, startup_program=None):
+    helper = LayerHelper('assign', **locals())
+    helper.append_op(
+        type='scale',
+        inputs={'X': [input]},
+        outputs={'Out': [output]},
+        attrs={'scale': 1.0})
+    return output
+
+
+def fill_constant(shape,
+                  dtype,
+                  value,
+                  out=None,
+                  main_program=None,
+                  startup_program=None):
+    """
+    This function creates a tensor , with shape as mentioned in the input and
+    specified dtype and fills this up with a constant value that
+    comes in the input. It also sets the stop_gradient to be True.
+    """
+    helper = LayerHelper("fill_constant", **locals())
+    if out is None:
+        out = helper.create_tmp_variable(dtype=dtype)
+    helper.append_op(
+        type='fill_constant',
+        inputs={},
+        outputs={'Out': [out]},
+        attrs={'shape': shape,
+               'dtype': out.dtype,
+               'value': float(value)})
+    out.stop_gradient = True
+    return out
+
+
+def fill_constant_batch_size_like(input,
+                                  shape,
+                                  dtype,
+                                  value,
+                                  input_dim_idx=0,
+                                  output_dim_idx=0,
+                                  main_program=None,
+                                  startup_program=None):
+    helper = LayerHelper("fill_constant_batch_size_like", **locals())
+    out = helper.create_tmp_variable(dtype=dtype)
+    helper.append_op(
+        type='fill_constant_batch_size_like',
+        inputs={'Input': input},
+        outputs={'Out': [out]},
+        attrs={
+            'shape': shape,
+            'dtype': out.dtype,
+            'value': float(value),
+            'input_dim_idx': input_dim_idx,
+            'output_dim_idx': output_dim_idx
+        })
+    out.stop_gradient = True
+    return out
+
+
+def ones(shape, dtype, main_program=None):
+    """
+    This function performs the same function as fill_constant() declared above
+    with the constant value being 1.0.
+    """
+    return fill_constant(value=1.0, **locals())
+
+
+def zeros(shape, dtype, main_program=None):
+    """
+    This function performs the same function as fill_constant() declared above
+    with the constant value being 0.0.
+    """
+    return fill_constant(value=0.0, **locals())
diff --git a/python/paddle/v2/fluid/tests/book/test_image_classification_train.py b/python/paddle/v2/fluid/tests/book/test_image_classification_train.py
index 4e71b6f345..3d336ffe95 100644
--- a/python/paddle/v2/fluid/tests/book/test_image_classification_train.py
+++ b/python/paddle/v2/fluid/tests/book/test_image_classification_train.py
@@ -1,9 +1,9 @@
 from __future__ import print_function
 
-import numpy as np
+import sys
+
 import paddle.v2 as paddle
 import paddle.v2.fluid as fluid
-import sys
 
 
 def resnet_cifar10(input, depth=32):
diff --git a/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py b/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py
index 80f8599679..c0b051f862 100644
--- a/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py
+++ b/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py
@@ -1,6 +1,51 @@
 import numpy as np
 import paddle.v2 as paddle
 import paddle.v2.fluid as fluid
+from paddle.v2.fluid.layer_helper import LayerHelper
+
+
+def lstm(x,
+         c_pre_init,
+         hidden_dim,
+         forget_bias=None,
+         main_program=None,
+         startup_program=None):
+    """
+    This function helps create an operator for the LSTM (Long Short Term
+    Memory) cell that can be used inside an RNN.
+    """
+    helper = LayerHelper('lstm_unit', **locals())
+    rnn = fluid.layers.StaticRNN()
+    with rnn.step():
+        c_pre = rnn.memory(init=c_pre_init)
+        x_t = rnn.step_input(x)
+
+        before_fc = fluid.layers.concat(
+            input=[x_t, c_pre],
+            axis=1,
+            main_program=main_program,
+            startup_program=startup_program)
+        after_fc = fluid.layers.fc(input=before_fc,
+                                   size=hidden_dim * 4,
+                                   main_program=main_program,
+                                   startup_program=startup_program)
+
+        dtype = x.dtype
+        c = helper.create_tmp_variable(dtype)
+        h = helper.create_tmp_variable(dtype)
+
+        helper.append_op(
+            type='lstm_unit',
+            inputs={"X": after_fc,
+                    "C_prev": c_pre},
+            outputs={"C": c,
+                     "H": h},
+            attrs={"forget_bias": forget_bias})
+
+        rnn.update_memory(c_pre, c)
+        rnn.output(h)
+
+    return rnn()
 
 
 def lstm_net(dict_dim, class_dim=2, emb_dim=32, seq_len=80, batch_size=50):
@@ -23,8 +68,7 @@ def lstm_net(dict_dim, class_dim=2, emb_dim=32, seq_len=80, batch_size=50):
     c_pre_init = fluid.layers.fill_constant(
         dtype=emb.dtype, shape=[batch_size, emb_dim], value=0.0)
     c_pre_init.stop_gradient = False
-    layer_1_out = fluid.layers.lstm(
-        emb, c_pre_init=c_pre_init, hidden_dim=emb_dim)
+    layer_1_out = lstm(emb, c_pre_init=c_pre_init, hidden_dim=emb_dim)
     layer_1_out = fluid.layers.transpose(x=layer_1_out, axis=[1, 0, 2])
 
     prediction = fluid.layers.fc(input=layer_1_out,
diff --git a/python/setup.py.in b/python/setup.py.in
index 9ccb4dc176..8396fb44cf 100644
--- a/python/setup.py.in
+++ b/python/setup.py.in
@@ -68,6 +68,7 @@ packages=['paddle',
           'paddle.v2.plot',
           'paddle.v2.fluid',
           'paddle.v2.fluid.proto',
+          'paddle.v2.fluid.layers',
           'py_paddle']
 
 with open('@PADDLE_SOURCE_DIR@/python/requirements.txt') as f:

From a2dfabb46a5dc2ac11d57b91c3fe8ed749c17c82 Mon Sep 17 00:00:00 2001
From: dzhwinter <dzhwinter@gmail.com>
Date: Thu, 14 Dec 2017 03:07:21 -0800
Subject: [PATCH 65/94] "fix based on comments"

---
 doc/design/paddle_nccl.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/doc/design/paddle_nccl.md b/doc/design/paddle_nccl.md
index 5e28144b95..7f0d8e14e4 100644
--- a/doc/design/paddle_nccl.md
+++ b/doc/design/paddle_nccl.md
@@ -7,7 +7,7 @@ This Design Doc refers to the NCCL feature in  paddle.  We propose an approach t
 
 ## Motivation
 
-NCCL is a NVIDIA library support Multi-GPU communicating and optimized for NVIDIA GPUs, it provides routines such as all-gather, all-reduce, broadcast, reduce, reduce-scatter, that can achieve high bandwidth over PCIe and NVLink high-speed interconnect. [NCCL](https://developer.nvidia.com/nccl). With NCCL library, we can easily accelerate the training in parallel. 
+[NCCL](https://developer.nvidia.com/nccl) is a NVIDIA library support Multi-GPU communicating and optimized for NVIDIA GPUs, it provides routines such as all-gather, all-reduce, broadcast, reduce, reduce-scatter, that can achieve high bandwidth over PCIe and NVLink high-speed interconnect. With NCCL library, we can easily accelerate the training in parallel. 
 
 - Pros
 1. easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library.

From a02a68dc6daecbd3d5b17b57db03e0b3f916646e Mon Sep 17 00:00:00 2001
From: dzhwinter <dzhwinter@gmail.com>
Date: Thu, 14 Dec 2017 06:09:54 -0800
Subject: [PATCH 66/94] "fixed based on comment"

---
 doc/design/paddle_nccl.md | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/doc/design/paddle_nccl.md b/doc/design/paddle_nccl.md
index 7f0d8e14e4..c7dac70998 100644
--- a/doc/design/paddle_nccl.md
+++ b/doc/design/paddle_nccl.md
@@ -28,9 +28,9 @@ Besides, it needs interfaces to synchronize model update with each different GPU
 
 As mentioned above, we wrap the NCCL routines as several kinds of operators. Need to note that NCCL need to create Communicator between gpu at the beginning, so there is a NCCLInit operator created.
 
-### Graph Converter
+### Transpiler
 
-To be compatible with [parameter server design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md), the graph converter converts the user defined operation graph into sub-graphs to be executed on different devices.
+To be compatible with [parameter server design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md), the transpiler compiles the user defined operation graph into sub-graphs to be executed on different devices.
 
 1. The user-defined model will be a single device program
 
@@ -40,7 +40,7 @@ To be compatible with [parameter server design doc](https://github.com/PaddlePad
 
    <img src="images/multigpu_before_convert.png" width="300"/>
 
-After convert, the graph as shows
+After compiling, the graph as shows
 
 <img src="images/multigpu_allreduce.png" width="1000"/>
 

From e0c33176460ffe322c8ff6da024628741423d020 Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Thu, 14 Dec 2017 23:14:18 +0800
Subject: [PATCH 67/94] add MKLDNNPlace

---
 paddle/platform/place.cc      |  6 ++++++
 paddle/platform/place.h       | 19 ++++++++++++++++++-
 paddle/platform/place_test.cc |  6 ++++++
 3 files changed, 30 insertions(+), 1 deletion(-)

diff --git a/paddle/platform/place.cc b/paddle/platform/place.cc
index 856e54df89..16f1774285 100644
--- a/paddle/platform/place.cc
+++ b/paddle/platform/place.cc
@@ -23,6 +23,7 @@ class PlacePrinter : public boost::static_visitor<> {
  public:
   explicit PlacePrinter(std::ostream &os) : os_(os) {}
   void operator()(const CPUPlace &) { os_ << "CPUPlace"; }
+  void operator()(const MKLDNNPlace &) { os_ << "MKLDNNPlace"; }
   void operator()(const GPUPlace &p) { os_ << "GPUPlace(" << p.device << ")"; }
 
  private:
@@ -38,6 +39,7 @@ const Place &get_place() { return the_default_place; }
 
 const GPUPlace default_gpu() { return GPUPlace(0); }
 const CPUPlace default_cpu() { return CPUPlace(); }
+const MKLDNNPlace default_mkldnn() { return MKLDNNPlace(); }
 
 bool is_gpu_place(const Place &p) {
   return boost::apply_visitor(IsGPUPlace(), p);
@@ -46,6 +48,10 @@ bool is_cpu_place(const Place &p) {
   return !boost::apply_visitor(IsGPUPlace(), p);
 }
 
+bool is_mkldnn_place(const Place &p) {
+  return boost::apply_visitor(IsMKLDNNPlace(), p);
+}
+
 bool places_are_same_class(const Place &p1, const Place &p2) {
   return p1.which() == p2.which();
 }
diff --git a/paddle/platform/place.h b/paddle/platform/place.h
index 5370360a7d..e745b2e839 100644
--- a/paddle/platform/place.h
+++ b/paddle/platform/place.h
@@ -31,6 +31,14 @@ struct CPUPlace {
   inline bool operator!=(const CPUPlace &) const { return false; }
 };
 
+struct MKLDNNPlace : public CPUPlace {
+  MKLDNNPlace() {}
+
+  // needed for variant equality comparison
+  inline bool operator==(const MKLDNNPlace &) const { return true; }
+  inline bool operator!=(const MKLDNNPlace &) const { return false; }
+};
+
 struct GPUPlace {
   GPUPlace() : GPUPlace(0) {}
   explicit GPUPlace(int d) : device(d) {}
@@ -45,14 +53,21 @@ struct GPUPlace {
 
 struct IsGPUPlace : public boost::static_visitor<bool> {
   bool operator()(const CPUPlace &) const { return false; }
+  bool operator()(const MKLDNNPlace &) const { return false; }
   bool operator()(const GPUPlace &gpu) const { return true; }
 };
 
+struct IsMKLDNNPlace : public boost::static_visitor<bool> {
+  bool operator()(const MKLDNNPlace &) const { return true; }
+  bool operator()(const CPUPlace &) const { return false; }
+  bool operator()(const GPUPlace &) const { return false; }
+};
+
 // Define the max number of Place in bit length. i.e., the max number of places
 // should be less equal than 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT)
 #define NUM_PLACE_TYPE_LIMIT_IN_BIT 4
 
-typedef boost::variant<GPUPlace, CPUPlace> Place;
+typedef boost::variant<GPUPlace, CPUPlace, MKLDNNPlace> Place;
 
 // static check number of place types is less equal than
 // 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT)
@@ -65,9 +80,11 @@ const Place &get_place();
 
 const GPUPlace default_gpu();
 const CPUPlace default_cpu();
+const MKLDNNPlace default_mkldnn();
 
 bool is_gpu_place(const Place &);
 bool is_cpu_place(const Place &);
+bool is_mkldnn_place(const Place &);
 bool places_are_same_class(const Place &, const Place &);
 
 std::ostream &operator<<(std::ostream &, const Place &);
diff --git a/paddle/platform/place_test.cc b/paddle/platform/place_test.cc
index 33e2e5a439..184af12c23 100644
--- a/paddle/platform/place_test.cc
+++ b/paddle/platform/place_test.cc
@@ -21,9 +21,15 @@ TEST(Place, Default) {
   EXPECT_TRUE(paddle::platform::is_gpu_place(paddle::platform::get_place()));
   EXPECT_TRUE(paddle::platform::is_gpu_place(paddle::platform::default_gpu()));
   EXPECT_TRUE(paddle::platform::is_cpu_place(paddle::platform::default_cpu()));
+  EXPECT_TRUE(
+      paddle::platform::is_mkldnn_place(paddle::platform::default_mkldnn()));
 
   paddle::platform::set_place(paddle::platform::CPUPlace());
   EXPECT_TRUE(paddle::platform::is_cpu_place(paddle::platform::get_place()));
+
+  paddle::platform::set_place(paddle::platform::MKLDNNPlace());
+  EXPECT_FALSE(paddle::platform::is_cpu_place(paddle::platform::get_place()));
+  EXPECT_TRUE(paddle::platform::is_mkldnn_place(paddle::platform::get_place()));
 }
 
 TEST(Place, Print) {

From f271210595953fee610dd049909d7e3a40f58285 Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Fri, 15 Dec 2017 15:15:18 +0800
Subject: [PATCH 68/94] fix undefined issue when with_gpu

---
 paddle/operators/math/math_function.cc | 8 ++++++++
 paddle/platform/place.h                | 2 +-
 2 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/paddle/operators/math/math_function.cc b/paddle/operators/math/math_function.cc
index 2b35e4532a..a05810d778 100644
--- a/paddle/operators/math/math_function.cc
+++ b/paddle/operators/math/math_function.cc
@@ -277,6 +277,14 @@ void set_constant_with_place<platform::CPUPlace>(
                            TensorSetConstantCPU(tensor, value));
 }
 
+template <>
+void set_constant_with_place<platform::MKLDNNPlace>(
+    const platform::DeviceContext& context, framework::Tensor* tensor,
+    float value) {
+  framework::VisitDataType(framework::ToDataType(tensor->type()),
+                           TensorSetConstantCPU(tensor, value));
+}
+
 struct TensorSetConstantWithPlace : public boost::static_visitor<void> {
   TensorSetConstantWithPlace(const platform::DeviceContext& context,
                              framework::Tensor* tensor, float value)
diff --git a/paddle/platform/place.h b/paddle/platform/place.h
index e745b2e839..70acc5b2d4 100644
--- a/paddle/platform/place.h
+++ b/paddle/platform/place.h
@@ -31,7 +31,7 @@ struct CPUPlace {
   inline bool operator!=(const CPUPlace &) const { return false; }
 };
 
-struct MKLDNNPlace : public CPUPlace {
+struct MKLDNNPlace {
   MKLDNNPlace() {}
 
   // needed for variant equality comparison

From a92f057ed1d52f8eed341212e11a76153512a1cf Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Fri, 15 Dec 2017 16:46:27 +0800
Subject: [PATCH 69/94] fix conflict of Place

---
 paddle/platform/place.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/paddle/platform/place.h b/paddle/platform/place.h
index 5a1ce52800..4526945792 100644
--- a/paddle/platform/place.h
+++ b/paddle/platform/place.h
@@ -72,7 +72,7 @@ struct IsMKLDNNPlace : public boost::static_visitor<bool> {
 // should be less equal than 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT)
 #define NUM_PLACE_TYPE_LIMIT_IN_BIT 4
 
-typedef boost::variant<CudnnPlace, GPUPlace, CPUPlace> Place;
+typedef boost::variant<CudnnPlace, GPUPlace, CPUPlace, MKLDNNPlace> Place;
 
 // static check number of place types is less equal than
 // 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT)

From bf269d67b3d1e8d88797022e4ea542f34c06df47 Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Fri, 15 Dec 2017 16:48:00 +0800
Subject: [PATCH 70/94] fix place_test on MKLDNNPlace

---
 paddle/platform/place.cc | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/paddle/platform/place.cc b/paddle/platform/place.cc
index 16f1774285..25fe8d21b4 100644
--- a/paddle/platform/place.cc
+++ b/paddle/platform/place.cc
@@ -45,7 +45,7 @@ bool is_gpu_place(const Place &p) {
   return boost::apply_visitor(IsGPUPlace(), p);
 }
 bool is_cpu_place(const Place &p) {
-  return !boost::apply_visitor(IsGPUPlace(), p);
+  return !is_gpu_place(p) && !is_mkldnn_place(p);
 }
 
 bool is_mkldnn_place(const Place &p) {

From a243edf44bbfad597cc16de1e8c21943c909898a Mon Sep 17 00:00:00 2001
From: Xi Chen <putcncx@gmail.com>
Date: Thu, 14 Dec 2017 14:31:49 -0800
Subject: [PATCH 71/94] add cross compiling doc in en

---
 doc/mobile/cross_compiling_for_ios_en.md | 120 +++++++++++++++++++++++
 1 file changed, 120 insertions(+)
 create mode 100644 doc/mobile/cross_compiling_for_ios_en.md

diff --git a/doc/mobile/cross_compiling_for_ios_en.md b/doc/mobile/cross_compiling_for_ios_en.md
new file mode 100644
index 0000000000..5428f501e8
--- /dev/null
+++ b/doc/mobile/cross_compiling_for_ios_en.md
@@ -0,0 +1,120 @@
+# iOS Compiling Guide
+
+This tutorial will walk you through cross compiling the PaddlePaddle library for iOS from the source in MacOS.
+
+## Preparation
+
+Apple provides Xcode for cross-compiling and IDE for iOS development. Download from App store or [here](https://developer.apple.com/cn/xcode/). To verify your installation, run command as follows
+
+```bash
+$ xcodebuild -version
+Xcode 9.0
+Build version 9A235
+```
+
+## Cross-compiling configurations
+
+PaddlePaddle provides cross-compiling toolchain configuration documentation [cmake/cross_compiling/ios.cmake](https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/cross_compiling/ios.cmake), which has some default settings for frequently used compilers.
+
+There are some mandatory environment variables need to be set before cross compiling PaddlePaddle for iOS:
+
+- `CMAKE_SYSTEM_NAME`, CMake compiling target platform name, has to be `iOS`. PaddlePaddle CMake will compile all the third party dependencies and enforce some parameters (`WITH_C_API=ON`、`WITH_GPU=OFF`、`WITH_AVX=OFF`、`WITH_PYTHON=OFF`、`WITH_RDMA=OFF`) when this variable is set with value `iOS`.
+
+- `WITH_C_API`, Whether to compile inference C-API library, has to be `ON`, since C-API is the only supported interface for inferencing in iOS.
+- `WITH_SWIG_PY`, has to be `ON`. It's not supported to inference or train via swig in iOS.
+
+Optional environment variables for iOS are：
+
+- `IOS_PLATFORM`, either `OS` (default) or `SIMULATOR`.
+  - `OS`, build targets ARM-based physical devices like iPhone or iPad.
+  - `SIMULATOR`, build targets x86 architecture simulators.
+- `IOS_ARCH`, target architecture. By default, all architecture types will be compiled. If you need to specify the architecture to compile for, please find valid values for different `IOS_PLATFORM` settings from the table below:
+
+    <table class="docutils">
+    <colgroup>
+      <col width="35%" />
+      <col width="65%" />
+    </colgroup>
+    <thead valign="bottom">
+      <tr class="row-odd">
+      <th class="head">IOS_PLATFORM</th>
+      <th class="head">IOS_ARCH</th>
+    </tr>
+    </thead>
+    <tbody valign="top">
+      <tr class="row-even">
+      <td>OS</td>
+      <td>armv7, armv7s, arm64 </td>
+    </tr>
+    <tr class="row-odd">
+      <td>SIMULATOR</td>
+      <td>i386, x86_64 </td>
+    </tr>
+    </tbody>
+    </table>
+
+- `IOS_DEPLOYMENT_TARGET`, minimum iOS version to deployment, `7.0` by default.
+- `IOS_ENABLE_BITCODE`, whether to enable [Bitcode](https://developer.apple.com/library/content/documentation/IDEs/Conceptual/AppDistributionGuide/AppThinning/AppThinning.html#//apple_ref/doc/uid/TP40012582-CH35-SW3), values can be `ON/OFF`, `ON` by default.
+- `IOS_USE_VECLIB_FOR_BLAS`, whether to use [vecLib](https://developer.apple.com/documentation/accelerate/veclib) framework for BLAS computing. values can be `ON/OFF`, `OFF` by default.
+- `IOS_DEVELOPMENT_ROOT`, the path to `Developer` directory, can be explicitly set with your `/path/to/platform/Developer`. If left blank, PaddlePaddle will automatically pick the Xcode corresponding `platform`'s `Developer` directory based on your `IOS_PLATFORM` value.
+- `IOS_SDK_ROOT`, the path to `SDK` root, can be explicitly set with your  `/path/to/platform/Developer/SDKs/SDK`. if left black, PaddlePaddle will pick the latest SDK in the directory of `IOS_DEVELOPMENT_ROOT`.
+
+other settings：
+
+- `USE_EIGEN_FOR_BLAS`, whether to use Eigen for matrix computing. effective when `IOS_USE_VECLIB_FOR_BLAS=OFF`. Values can be `ON/OFF`, `OFF` by default.
+- `HOST_C/CXX_COMPILER`, host C/C++ compiler. Uses value from environment variable `CC/CXX` by default or `cc/c++` if `CC/CXX` doesn't exist.
+
+some typical cmake configurations:
+
+```bash
+cmake -DCMAKE_SYSTEM_NAME=iOS \
+      -DIOS_PLATFORM=OS \
+      -DIOS_ARCH="armv7;arm64" \
+      -DIOS_ENABLE_BITCODE=ON \
+      -DIOS_USE_VECLIB_FOR_BLAS=ON \
+      -DCMAKE_INSTALL_PREFIX=your/path/to/install \
+      -DWITH_C_API=ON \
+      -DWITH_TESTING=OFF \
+      -DWITH_SWIG_PY=OFF \
+      ..
+```
+
+```bash
+cmake -DCMAKE_SYSTEM_NAME=iOS \
+      -DIOS_PLATFORM=SIMULATOR \
+      -DIOS_ARCH="x86_64" \
+      -DIOS_USE_VECLIB_FOR_BLAS=ON \
+      -DCMAKE_INSTALL_PREFIX=your/path/to/install \
+      -DWITH_C_API=ON \
+      -DWITH_TESTING=OFF \
+      -DWITH_SWIG_PY=OFF \
+      ..
+```
+
+You can set other compiling parameters for your own need. I.E. if you are trying to minimize the library size, set `CMAKE_BUILD_TYPE` with `MinSizeRel`; or if the performance is your concern, set `CMAKE_BUILD_TYPE` with `Release`. You can even manipulate the PaddlePaddle compiling procedure by manually set `CMAKE_C/CXX_FLAGS` values.
+
+**TIPS for a better performance**:
+
+- set `CMAKE_BUILD_TYPE` with `Release`
+- set `IOS_USE_VECLIB_FOR_BLAS` with `ON`
+
+## Compile and install
+
+After CMake, run following commands, PaddlePaddle will download the compile 3rd party dependencies, compile and install PaddlePaddle inference library.
+
+```
+$ make
+$ make install
+```
+
+Please Note: if you compiled PaddlePaddle in the source directory for other platforms, do remove `third_party` and `build` directory within the source with `rm -rf` to ensure that all the 3rd party libraries dependencies and PaddlePaddle is newly compiled with current CMake configuration.
+
+`your/path/to/install` directory will have following directories after `compile` and `install`:
+
+- `include`, contains all the C-API header files.
+- `lib`, contains PaddlePaddle C-API static library.
+- `third_party` contains all the 3rd party libraries.
+
+Please note: if PaddlePaddle library need to support both physical devices and simulators, you will need to compile correspondingly, then merge fat library with `lipo`
+
+Now you will have PaddlePaddle library compiled and installed, the fat library can be used in deep learning related iOS APPs. Please refer to C-API documentation for usage guides.

From d2b6d2ef8a42feed15ac5ce4cd43877436f8abe1 Mon Sep 17 00:00:00 2001
From: Xi Chen <putcncx@gmail.com>
Date: Thu, 14 Dec 2017 14:35:44 -0800
Subject: [PATCH 72/94] minor tweaks.

---
 doc/mobile/cross_compiling_for_ios_en.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/doc/mobile/cross_compiling_for_ios_en.md b/doc/mobile/cross_compiling_for_ios_en.md
index 5428f501e8..e2a1f6fc5f 100644
--- a/doc/mobile/cross_compiling_for_ios_en.md
+++ b/doc/mobile/cross_compiling_for_ios_en.md
@@ -1,4 +1,4 @@
-# iOS Compiling Guide
+# PaddlePaddle Compiling Guide for iOS
 
 This tutorial will walk you through cross compiling the PaddlePaddle library for iOS from the source in MacOS.
 

From f5dc9eadaafc74e606aef849145693ece2e743e9 Mon Sep 17 00:00:00 2001
From: zhushuang02 <zhushuang02@baidu.com>
Date: Fri, 15 Dec 2017 10:22:59 +0800
Subject: [PATCH 73/94] Add refer code to get started

---
 doc/getstarted/concepts/src/infer.py        | 14 ++++++++++++++
 doc/getstarted/concepts/src/train.py        |  5 +++++
 doc/getstarted/concepts/use_concepts_cn.rst |  5 +++++
 3 files changed, 24 insertions(+)
 create mode 100644 doc/getstarted/concepts/src/infer.py

diff --git a/doc/getstarted/concepts/src/infer.py b/doc/getstarted/concepts/src/infer.py
new file mode 100644
index 0000000000..780d831da8
--- /dev/null
+++ b/doc/getstarted/concepts/src/infer.py
@@ -0,0 +1,14 @@
+import paddle.v2 as paddle
+import numpy as np
+
+paddle.init(use_gpu=False)
+x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(2))
+y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
+
+# loading the model which generated by training
+with open('hello_params_pass_90.tar', 'r') as f:
+    parameters = paddle.parameters.Parameters.from_tar(f)
+
+i = [[[1, 2]]]
+
+print paddle.infer(output_layer=y_predict, parameters=parameters, input=i)
diff --git a/doc/getstarted/concepts/src/train.py b/doc/getstarted/concepts/src/train.py
index 8aceb23406..4bccbfca3c 100644
--- a/doc/getstarted/concepts/src/train.py
+++ b/doc/getstarted/concepts/src/train.py
@@ -26,6 +26,11 @@ def event_handler(event):
         if event.batch_id % 1 == 0:
             print "Pass %d, Batch %d, Cost %f" % (event.pass_id, event.batch_id,
                                                   event.cost)
+    # product model every 10 pass
+    if isinstance(event, paddle.event.EndPass):
+        if event.pass_id % 10 == 0:
+            with open('params_pass_%d.tar' % event.pass_id, 'w') as f:
+                trainer.save_parameter_to_tar(f)
 
 
 # define training dataset reader
diff --git a/doc/getstarted/concepts/use_concepts_cn.rst b/doc/getstarted/concepts/use_concepts_cn.rst
index c243083794..7ff0b29ac3 100644
--- a/doc/getstarted/concepts/use_concepts_cn.rst
+++ b/doc/getstarted/concepts/use_concepts_cn.rst
@@ -147,4 +147,9 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和
 ..  literalinclude:: src/train.py
     :linenos:
 
+使用以上训练好的模型进行预测的例子：
+
+..  literalinclude:: src/infer.py
+    :linenos:
+
 有关线性回归的实际应用，可以参考PaddlePaddle book的 `第一章节 <http://book.paddlepaddle.org/index.html>`_。

From 7fe3c5789d4d65cd082550d2144108aabdfbda4b Mon Sep 17 00:00:00 2001
From: Liu Yiqun <liuyiqun01@baidu.com>
Date: Fri, 15 Dec 2017 02:40:27 +0000
Subject: [PATCH 74/94] Fix typo in cross-compiling documentation of iOS and
 add it to the index.

---
 doc/mobile/cross_compiling_for_ios_cn.md | 4 ++--
 doc/mobile/cross_compiling_for_ios_en.md | 8 ++++----
 doc/mobile/index_en.rst                  | 1 +
 3 files changed, 7 insertions(+), 6 deletions(-)

diff --git a/doc/mobile/cross_compiling_for_ios_cn.md b/doc/mobile/cross_compiling_for_ios_cn.md
index 9da48e7f21..d5196d9a4c 100644
--- a/doc/mobile/cross_compiling_for_ios_cn.md
+++ b/doc/mobile/cross_compiling_for_ios_cn.md
@@ -18,11 +18,11 @@ PaddlePaddle为交叉编译提供了工具链配置文档[cmake/cross_compiling/
 
 - `CMAKE_SYSTEM_NAME`，CMake编译的目标平台，必须设置为`iOS`。在设置`CMAKE_SYSTEM_NAME=iOS`后，PaddlePaddle的CMake系统会自动编译所有的第三方依赖库，并且强制设置一些PaddlePaddle参数的值（`WITH_C_API=ON`、`WITH_GPU=OFF`、`WITH_AVX=OFF`、`WITH_PYTHON=OFF`、`WITH_RDMA=OFF`）。
 - `WITH_C_API`，是否编译C-API预测库，必须设置为ON。在iOS平台上只支持使用C-API来预测。
-- `WITH_SWIG_PY`，必须设置为ON。在iOS平台上不支持通过swig调用来训练或者预测。
+- `WITH_SWIG_PY`，必须设置为`OFF`。在iOS平台上不支持通过swig调用来训练或者预测。
 
 iOS平台可选配置参数：
 
-- `IOS_PLATFORM`，可设置为`OS/SIMULATOR`，默认值为`OS`。
+- `IOS_PLATFORM`，可设置为`OS`（默认值）或`SIMULATOR`。
   - `OS`，构建目标为`arm`架构的iPhone或者iPad等物理设备。
   - `SIMULATOR`，构建目标为`x86`架构的模拟器平台。
 - `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示，默认编译所有架构：
diff --git a/doc/mobile/cross_compiling_for_ios_en.md b/doc/mobile/cross_compiling_for_ios_en.md
index e2a1f6fc5f..aa390cd61f 100644
--- a/doc/mobile/cross_compiling_for_ios_en.md
+++ b/doc/mobile/cross_compiling_for_ios_en.md
@@ -18,12 +18,12 @@ PaddlePaddle provides cross-compiling toolchain configuration documentation [cma
 
 There are some mandatory environment variables need to be set before cross compiling PaddlePaddle for iOS:
 
-- `CMAKE_SYSTEM_NAME`, CMake compiling target platform name, has to be `iOS`. PaddlePaddle CMake will compile all the third party dependencies and enforce some parameters (`WITH_C_API=ON`、`WITH_GPU=OFF`、`WITH_AVX=OFF`、`WITH_PYTHON=OFF`、`WITH_RDMA=OFF`) when this variable is set with value `iOS`.
+- `CMAKE_SYSTEM_NAME`, CMake compiling target platform name, has to be `iOS`. PaddlePaddle CMake will compile all the third party dependencies and enforce some parameters (`WITH_C_API=ON`, `WITH_GPU=OFF`, `WITH_AVX=OFF`, `WITH_PYTHON=OFF`,`WITH_RDMA=OFF`) when this variable is set with value `iOS`.
 
 - `WITH_C_API`, Whether to compile inference C-API library, has to be `ON`, since C-API is the only supported interface for inferencing in iOS.
-- `WITH_SWIG_PY`, has to be `ON`. It's not supported to inference or train via swig in iOS.
+- `WITH_SWIG_PY`, has to be `OFF`. It's not supported to inference or train via swig in iOS.
 
-Optional environment variables for iOS are：
+Optional environment variables for iOS are:
 
 - `IOS_PLATFORM`, either `OS` (default) or `SIMULATOR`.
   - `OS`, build targets ARM-based physical devices like iPhone or iPad.
@@ -115,6 +115,6 @@ Please Note: if you compiled PaddlePaddle in the source directory for other plat
 - `lib`, contains PaddlePaddle C-API static library.
 - `third_party` contains all the 3rd party libraries.
 
-Please note: if PaddlePaddle library need to support both physical devices and simulators, you will need to compile correspondingly, then merge fat library with `lipo`
+Please note: if PaddlePaddle library need to support both physical devices and simulators, you will need to compile correspondingly, then merge fat library with `lipo`.
 
 Now you will have PaddlePaddle library compiled and installed, the fat library can be used in deep learning related iOS APPs. Please refer to C-API documentation for usage guides.
diff --git a/doc/mobile/index_en.rst b/doc/mobile/index_en.rst
index 3c08d73671..ef421dacad 100644
--- a/doc/mobile/index_en.rst
+++ b/doc/mobile/index_en.rst
@@ -5,4 +5,5 @@ MOBILE
   :maxdepth: 1
 
   cross_compiling_for_android_en.md
+  cross_compiling_for_ios_en.md
   cross_compiling_for_raspberry_en.md

From d7b67f2b7460bd68d4db41c63dbc3c92d4b95497 Mon Sep 17 00:00:00 2001
From: Yancey <yanxu05@baidu.com>
Date: Fri, 15 Dec 2017 13:16:44 +0800
Subject: [PATCH 75/94] fix pipe_reader on multi passes (#6627)

fix pipe reader on multi passes
---
 python/paddle/v2/reader/decorator.py          |  4 +--
 .../paddle/v2/reader/tests/decorator_test.py  | 30 +++++++++++++++++++
 2 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/python/paddle/v2/reader/decorator.py b/python/paddle/v2/reader/decorator.py
index 7e457f987d..27c82c95f7 100644
--- a/python/paddle/v2/reader/decorator.py
+++ b/python/paddle/v2/reader/decorator.py
@@ -390,8 +390,6 @@ def pipe_reader(left_cmd,
     if not callable(parser):
         raise TypeError("parser must be a callable object")
 
-    process = subprocess.Popen(
-        left_cmd.split(" "), bufsize=bufsize, stdout=subprocess.PIPE)
     # TODO(typhoonzero): add a thread to read stderr
 
     # Always init a decompress object is better than
@@ -400,6 +398,8 @@ def pipe_reader(left_cmd,
         32 + zlib.MAX_WBITS)  # offset 32 to skip the header
 
     def reader():
+        process = subprocess.Popen(
+            left_cmd.split(" "), bufsize=bufsize, stdout=subprocess.PIPE)
         remained = ""
         while True:
             buff = process.stdout.read(bufsize)
diff --git a/python/paddle/v2/reader/tests/decorator_test.py b/python/paddle/v2/reader/tests/decorator_test.py
index 5a92951b10..06e14796da 100644
--- a/python/paddle/v2/reader/tests/decorator_test.py
+++ b/python/paddle/v2/reader/tests/decorator_test.py
@@ -145,5 +145,35 @@ class TestXmap(unittest.TestCase):
                             self.assertEqual(e, mapper(idx))
 
 
+class TestPipeReader(unittest.TestCase):
+    def test_pipe_reader(self):
+        def simple_parser(lines):
+            return lines
+
+        import tempfile
+
+        records = [str(i) for i in xrange(5)]
+        temp = tempfile.NamedTemporaryFile()
+        try:
+            with open(temp.name, 'w') as f:
+                for r in records:
+                    f.write('%s\n' % r)
+
+            cmd = "cat %s" % temp.name
+            reader = paddle.v2.reader.pipe_reader(
+                cmd, simple_parser, bufsize=128)
+            for i in xrange(4):
+                result = []
+                for r in reader():
+                    result.append(r)
+
+                for idx, e in enumerate(records):
+                    print e, result[idx]
+                    self.assertEqual(e, result[idx])
+        finally:
+            # delete the temporary file
+            temp.close()
+
+
 if __name__ == '__main__':
     unittest.main()

From 537a0d4c0e22dae88b7c917928bda2f976cb6700 Mon Sep 17 00:00:00 2001
From: Coke <memorycancel@gmail.com>
Date: Fri, 15 Dec 2017 13:47:11 +0800
Subject: [PATCH 76/94] Update infer.py

---
 doc/getstarted/concepts/src/infer.py | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/doc/getstarted/concepts/src/infer.py b/doc/getstarted/concepts/src/infer.py
index 780d831da8..4cc58dfee0 100644
--- a/doc/getstarted/concepts/src/infer.py
+++ b/doc/getstarted/concepts/src/infer.py
@@ -6,9 +6,13 @@ x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(2))
 y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
 
 # loading the model which generated by training
-with open('hello_params_pass_90.tar', 'r') as f:
+with open('params_pass_90.tar', 'r') as f:
     parameters = paddle.parameters.Parameters.from_tar(f)
 
-i = [[[1, 2]]]
-
+# Input multiple sets of data，Output the infer result in a array.
+i = [[[1, 2]], [[3, 4]], [[5, 6]]]
 print paddle.infer(output_layer=y_predict, parameters=parameters, input=i)
+# Will print:
+# [[ -3.24491572]
+#  [ -6.94668722]
+#  [-10.64845848]]

From d62552dc99583396ddd639fee58242ff9aae368f Mon Sep 17 00:00:00 2001
From: Coke <memorycancel@gmail.com>
Date: Fri, 15 Dec 2017 13:52:12 +0800
Subject: [PATCH 77/94] Update use_concepts_cn.rst

---
 doc/getstarted/concepts/use_concepts_cn.rst | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/doc/getstarted/concepts/use_concepts_cn.rst b/doc/getstarted/concepts/use_concepts_cn.rst
index 7ff0b29ac3..e695ff283e 100644
--- a/doc/getstarted/concepts/use_concepts_cn.rst
+++ b/doc/getstarted/concepts/use_concepts_cn.rst
@@ -147,7 +147,7 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和
 ..  literalinclude:: src/train.py
     :linenos:
 
-使用以上训练好的模型进行预测的例子：
+使用以上训练好的模型进行预测，取其中一个模型params_pass_90.tar，输入需要预测的向量组，然后打印输出：
 
 ..  literalinclude:: src/infer.py
     :linenos:

From d5cab4f07c3e63970642cd428a9a33b284ddf4f1 Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Fri, 15 Dec 2017 14:22:30 +0800
Subject: [PATCH 78/94] Fix compile on CUDA9.1 & MacOS (#6642)

---
 paddle/math/float16.h           |  4 ++--
 paddle/platform/dynload/nccl.cc |  5 +++++
 paddle/platform/dynload/nccl.h  | 24 ++++++++++++------------
 paddle/platform/nccl_test.cu    |  4 ++--
 paddle/platform/variant.h       | 13 +++++++++++++
 python/.gitignore               |  1 +
 6 files changed, 35 insertions(+), 16 deletions(-)

diff --git a/paddle/math/float16.h b/paddle/math/float16.h
index 76ad3a0123..efebbce504 100644
--- a/paddle/math/float16.h
+++ b/paddle/math/float16.h
@@ -79,7 +79,7 @@ public:
 #ifdef PADDLE_CUDA_FP16
   HOSTDEVICE inline explicit float16(const half& h) {
 #if CUDA_VERSION >= 9000
-    x = reinterpret_cast<__half_raw*>(&h)->x;
+    x = reinterpret_cast<__half_raw*>(const_cast<half*>(&h))->x;
 #else
     x = h.x;
 #endif  // CUDA_VERSION >= 9000
@@ -145,7 +145,7 @@ public:
 #ifdef PADDLE_CUDA_FP16
   HOSTDEVICE inline float16& operator=(const half& rhs) {
 #if CUDA_VERSION >= 9000
-    x = reinterpret_cast<__half_raw*>(&rhs)->x;
+    x = reinterpret_cast<__half_raw*>(const_cast<half*>(&rhs))->x;
 #else
     x = rhs.x;
 #endif
diff --git a/paddle/platform/dynload/nccl.cc b/paddle/platform/dynload/nccl.cc
index 8f92b8d94d..91168f37ef 100644
--- a/paddle/platform/dynload/nccl.cc
+++ b/paddle/platform/dynload/nccl.cc
@@ -25,6 +25,11 @@ void *nccl_dso_handle;
 
 NCCL_RAND_ROUTINE_EACH(DEFINE_WRAP);
 
+void LoadNCCLDSO() {
+  platform::call_once(nccl_dso_flag,
+                      [] { GetNCCLDsoHandle(&nccl_dso_handle); });
+}
+
 }  // namespace dynload
 }  // namespace platform
 }  // namespace paddle
diff --git a/paddle/platform/dynload/nccl.h b/paddle/platform/dynload/nccl.h
index 981b2ab258..11007c1031 100644
--- a/paddle/platform/dynload/nccl.h
+++ b/paddle/platform/dynload/nccl.h
@@ -28,18 +28,18 @@ extern std::once_flag nccl_dso_flag;
 extern void* nccl_dso_handle;
 
 #ifdef PADDLE_USE_DSO
-#define DECLARE_DYNAMIC_LOAD_NCCL_WRAP(__name)                         \
-  struct DynLoad__##__name {                                           \
-    template <typename... Args>                                        \
-    auto operator()(Args... args) -> decltype(__name(args...)) {       \
-      using nccl_func = decltype(__name(args...)) (*)(Args...);        \
-      platform::call_once(nccl_dso_flag,                               \
-                          paddle::platform::dynload::GetNCCLDsoHandle, \
-                          &nccl_dso_handle);                           \
-      void* p_##__name = dlsym(nccl_dso_handle, #__name);              \
-      return reinterpret_cast<nccl_func>(p_##__name)(args...);         \
-    }                                                                  \
-  };                                                                   \
+extern void LoadNCCLDSO();
+
+#define DECLARE_DYNAMIC_LOAD_NCCL_WRAP(__name)                   \
+  struct DynLoad__##__name {                                     \
+    template <typename... Args>                                  \
+    auto operator()(Args... args) -> decltype(__name(args...)) { \
+      using nccl_func = decltype(__name(args...)) (*)(Args...);  \
+      paddle::platform::dynload::LoadNCCLDSO();                  \
+      void* p_##__name = dlsym(nccl_dso_handle, #__name);        \
+      return reinterpret_cast<nccl_func>(p_##__name)(args...);   \
+    }                                                            \
+  };                                                             \
   extern DynLoad__##__name __name
 #else
 #define DECLARE_DYNAMIC_LOAD_NCCL_WRAP(__name) \
diff --git a/paddle/platform/nccl_test.cu b/paddle/platform/nccl_test.cu
index c99dae68be..94ab360a19 100644
--- a/paddle/platform/nccl_test.cu
+++ b/paddle/platform/nccl_test.cu
@@ -31,7 +31,7 @@ namespace platform {
 TEST(NCCL, init) {
   std::vector<ncclComm_t> comms;
   comms.resize(dev_count);
-  PADDLE_ENFORCE(dynload::ncclCommInitAll(comms.data(), dev_count, nullptr));
+  dynload::ncclCommInitAll(comms.data(), dev_count, nullptr);
   for (int i = 0; i < dev_count; ++i) {
     dynload::ncclCommDestroy(comms[i]);
   }
@@ -62,7 +62,7 @@ TEST(NCCL, all_reduce) {
   std::vector<ncclComm_t> comms;
   comms.resize(dev_count);
   VLOG(1) << "Initializing ncclComm";
-  PADDLE_ENFORCE(dynload::ncclCommInitAll(comms.data(), dev_count, nullptr));
+  dynload::ncclCommInitAll(comms.data(), dev_count, nullptr);
   VLOG(1) << "ncclComm initialized";
   VLOG(1) << "Creating thread data";
   std::vector<std::unique_ptr<PerThreadData<double>>> data;
diff --git a/paddle/platform/variant.h b/paddle/platform/variant.h
index 619897ca19..284b4c42ac 100644
--- a/paddle/platform/variant.h
+++ b/paddle/platform/variant.h
@@ -14,6 +14,19 @@
 
 #pragma once
 
+#ifdef __CUDACC__
+#ifdef __CUDACC_VER_MAJOR__
+// CUDA 9 define `__CUDACC_VER__` as a warning message, manually define
+// __CUDACC_VER__ instead.
+#undef __CUDACC_VER__
+
+#define __CUDACC_VER__                                         \
+  (__CUDACC_VER_MAJOR__ * 10000 + __CUDACC_VER_MINOR__ * 100 + \
+   __CUDACC_VER_BUILD__)
+#endif
+
+#endif
+
 #include <boost/config.hpp>
 
 #ifdef PADDLE_WITH_CUDA
diff --git a/python/.gitignore b/python/.gitignore
index cc7d0ece4a..1ba1d4c9b0 100644
--- a/python/.gitignore
+++ b/python/.gitignore
@@ -2,6 +2,7 @@
 build
 dist
 paddle.egg-info
+paddlepaddle_gpu.egg-info
 .idea
 paddle/proto/*.py
 paddle/proto/*.pyc

From d37ed6cb13fc63aa2c81ece8de005bfbea4916be Mon Sep 17 00:00:00 2001
From: Yibing Liu <liuyibing01@baidu.com>
Date: Fri, 15 Dec 2017 07:41:54 +0000
Subject: [PATCH 79/94] polish code in reshape_op

---
 paddle/operators/reshape_op.cc | 18 ++++++++++--------
 1 file changed, 10 insertions(+), 8 deletions(-)

diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc
index 5baf459536..d82d828747 100644
--- a/paddle/operators/reshape_op.cc
+++ b/paddle/operators/reshape_op.cc
@@ -36,10 +36,13 @@ class ReshapeOp : public framework::OperatorWithKernel {
     auto x_dims = ctx->GetInputDim("X");
 
     std::vector<size_t> neg_dims_idx;
+    // set some dimension to -1 if it is unknown
+    const int unknown_size = -1;
     for (size_t i = 0; i < shape.size(); ++i) {
-      PADDLE_ENFORCE(shape[i] > 0 || shape[i] == -1,
-                     "Each dimension of Attr(shape) must be positive or -1.");
-      if (shape[i] == -1) {
+      PADDLE_ENFORCE(shape[i] > 0 || shape[i] == unknown_size,
+                     "Each dimension of Attr(shape) must be positive or %d.",
+                     unknown_size);
+      if (shape[i] == unknown_size) {
         neg_dims_idx.push_back(i);
         PADDLE_ENFORCE(neg_dims_idx.size() <= 1,
                        "Only one dimension of Attr(shape) can be unknown.");
@@ -53,8 +56,7 @@ class ReshapeOp : public framework::OperatorWithKernel {
       // dim infer
       shape[neg_dims_idx[0]] = in_size / (-capacity);
       // recalculate capacity
-      capacity = std::accumulate(shape.begin(), shape.end(), 1,
-                                 std::multiplies<int>());
+      capacity = shape[neg_dims_idx[0]] * (-capacity);
     }
     // capacity check
     PADDLE_ENFORCE(capacity == in_size,
@@ -98,9 +100,9 @@ the tensor X into a 2-D tensor:
 
     [[1, 2, 3, 4]]
 
-One dimension in the target shape can be set -1, and the real dimension 
-will be infered from the original shape of Input(X) and other
-dimensions in the target shape.
+One dimension in the target shape can be set -1, representing that its
+size is unknown. In this case, the real dimension will be infered from 
+the original shape of Input(X) and other dimensions in the target shape.
 )DOC");
   }
 };

From acaef9a1302a7907f9d42dffb7a75616b483b658 Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Thu, 14 Dec 2017 20:58:10 +0800
Subject: [PATCH 80/94] rename mkldnn doc

---
 doc/design/{mkldnn => mkl}/image/engine.png    | Bin
 doc/design/{mkldnn => mkl}/image/gradients.png | Bin
 doc/design/{mkldnn => mkl}/image/layers.png    | Bin
 doc/design/{mkldnn => mkl}/image/matrix.png    | Bin
 doc/design/{mkldnn => mkl}/image/overview.png  | Bin
 doc/design/{mkldnn/README.MD => mkl/mkldnn.md} |   1 -
 6 files changed, 1 deletion(-)
 rename doc/design/{mkldnn => mkl}/image/engine.png (100%)
 rename doc/design/{mkldnn => mkl}/image/gradients.png (100%)
 rename doc/design/{mkldnn => mkl}/image/layers.png (100%)
 rename doc/design/{mkldnn => mkl}/image/matrix.png (100%)
 rename doc/design/{mkldnn => mkl}/image/overview.png (100%)
 rename doc/design/{mkldnn/README.MD => mkl/mkldnn.md} (99%)

diff --git a/doc/design/mkldnn/image/engine.png b/doc/design/mkl/image/engine.png
similarity index 100%
rename from doc/design/mkldnn/image/engine.png
rename to doc/design/mkl/image/engine.png
diff --git a/doc/design/mkldnn/image/gradients.png b/doc/design/mkl/image/gradients.png
similarity index 100%
rename from doc/design/mkldnn/image/gradients.png
rename to doc/design/mkl/image/gradients.png
diff --git a/doc/design/mkldnn/image/layers.png b/doc/design/mkl/image/layers.png
similarity index 100%
rename from doc/design/mkldnn/image/layers.png
rename to doc/design/mkl/image/layers.png
diff --git a/doc/design/mkldnn/image/matrix.png b/doc/design/mkl/image/matrix.png
similarity index 100%
rename from doc/design/mkldnn/image/matrix.png
rename to doc/design/mkl/image/matrix.png
diff --git a/doc/design/mkldnn/image/overview.png b/doc/design/mkl/image/overview.png
similarity index 100%
rename from doc/design/mkldnn/image/overview.png
rename to doc/design/mkl/image/overview.png
diff --git a/doc/design/mkldnn/README.MD b/doc/design/mkl/mkldnn.md
similarity index 99%
rename from doc/design/mkldnn/README.MD
rename to doc/design/mkl/mkldnn.md
index 61d453de24..e2fe1e6b26 100644
--- a/doc/design/mkldnn/README.MD
+++ b/doc/design/mkl/mkldnn.md
@@ -208,4 +208,3 @@ if use_mkldnn
 但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。
 4. MKL-DNN的高性能格式与PaddlePaddle原有的`NCHW`不同(PaddlePaddle中的cuDNN部分使用的也是`NCHW`，所以不存在这个问题)。
 所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。
-

From 1670ec081bd35c36e63298de2b31f0b88007888b Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Thu, 14 Dec 2017 21:03:13 +0800
Subject: [PATCH 81/94] add mkl packed desgin doc

---
 doc/design/mkl/mkl_packed.md | 91 ++++++++++++++++++++++++++++++++++++
 1 file changed, 91 insertions(+)
 create mode 100644 doc/design/mkl/mkl_packed.md

diff --git a/doc/design/mkl/mkl_packed.md b/doc/design/mkl/mkl_packed.md
new file mode 100644
index 0000000000..a9fb56d709
--- /dev/null
+++ b/doc/design/mkl/mkl_packed.md
@@ -0,0 +1,91 @@
+# Intel® MKL Packed Optimization on PaddlePaddle: Design Doc
+
+
+## Contents
+
+- [Overview](#overview)
+- [Key Points](#key-points) 
+   - [Background](#background)
+   - [Solution](#solution)
+- [Actions](#actions)
+    - [CMake](#cmake)
+	- [Layers](#layers)
+	- [Unit Tests](#unit-tests)
+	- [Python API](#python-api)
+	- [Benchmarking](#benchmarking)
+
+
+## Overview
+我们计划将 Intel® MKL 中引入的 GEMM Packed APIs\[[1](#references)\] 集成到 PaddlePaddle 中，充分发挥英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
+现阶段的优化主要针对 Recurrent Neural Network(以下简称RNN)相关层（包括`RecurrentLayer`, `GatedRecurrentLayer`和`LstmLayer`）， 以及 PaddlePaddle V1 API。
+
+## Key Points
+
+### Background
+为了达到最佳性能， Intel® MKL 中的 cblas_?gemm 会在计算前将原数据转换为更适合英特尔平台的Packed格式， 这一数据格式的转换操作 (Packing)，在问题本身的计算量比较小的时候显得相对来说较为耗时。
+在现有的某些情况下（例如RNN），多次调用 cblas_?gemm 时会使用相同的原数据，每次调用时对原数据的重复Packing便成为了冗余。
+
+为了最大程度减少多次调用 cblas_?gemm 在Packing上的耗时，Intel® MKL 引入了以下四个API:
+   * cblas_?gemm_alloc
+   * cblas_?gemm_pack 
+   * cblas_?gemm_compute
+   * cblas_?gemm_free
+
+通过使用这些API，我们可以先完成对原数据的Packing操作，再把已转换为Packed格式的数据传递给那些复用同一数据的gemm_compute函数，从而避免了Packing冗余。
+
+### Solution
+在RNN的case下，同一次 forward/backward 过程中所有time state共享同一个weight矩阵。当只做 inference 时，各次 forward 之间也都使用相同的weight矩阵，没有必要在每次forward中每个time state的计算时对weight进行重复的packing操作。
+
+我们通过使用新引入的GEMM Packed APIs，在layer init时先完成对weight的packing操作，然后在 forward/backward 时复用已pack过后的weight，并在每次weight更新后重新Packing。
+
+* 优化前，对于sequence length = `T` 的model, `N` 次iteration执行的Packing次数为：   
+  - `inference`: `N * T`  
+  - `training`: `2 * N * T`
+* 优化后，对于sequence length = `T` 的model, `N` 次iteration执行的Packing次数减少至：
+  - `inference`: `1`    
+  - `training`: `2 * N`
+
+## Actions
+
+添加的相关文件和目录结构如下：
+
+```txt
+PaddlePaddle/Paddle
+├── ...
+└── paddle/
+    ├── ...
+    └── gserver/
+        ├── ...
+        ├── layers/
+        │   ├── ...
+        │   ├── MKLPackedRecurrentLayer.*
+        |   ├── MKLPackedGatedRecurrentLayer.*
+        |   ├── MKLPackedLstmLayer.*
+        |   └── MKLPackedGemm.h
+        └── tests/
+            ├── ...
+            └── test_MKLPacked.cpp
+```
+
+### CMake
+在对应的`CMakeLists.txt`中根据`WITH_MKL`是否打开，来决定是否开启MKL Packed相关功能。
+
+### Layers
+所有的`MKLPacked*Layer`都继承于PaddlePaddle的基类`Layer`, 并添加头文件 `MKLPackedGemm.h`，该文件中实现的对相关GEMM Packed APIs做了封装。
+
+### Unit Tests
+我们会添加`test_MKLPacked.cpp`用于MKL Packed优化后layer的测试。
+对于每一个新加的RNN layer，我们会对比如下2个方面：
+1. 对比优化后layer自身，sequence mode（`rnn_use_batch=false`）与batch mode(`rnn_use_batch=true`)的结果。
+2. 对比优化后layer与相对应的PaddlePaddle原有layer, 在batch mode下的结果。
+
+### Python API
+TBD
+
+### Benchmarking
+会添加相应的脚本用于测试和对比在使用MKL Packed recurrent layers 前后的网络性能。
+
+## References 
+1. [Introducing the new Packed APIs for GEMM](https://software.intel.com/en-us/articles/introducing-the-new-packed-apis-for-gemm)
+
+

From 3f2fa0a1f65f6ca7d6881a4bac93d190b2032b78 Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Sat, 16 Dec 2017 13:50:37 +0800
Subject: [PATCH 82/94] follow comments and refine doc

---
 doc/design/mkl/mkl_packed.md | 32 ++++++++++++++++++--------------
 1 file changed, 18 insertions(+), 14 deletions(-)

diff --git a/doc/design/mkl/mkl_packed.md b/doc/design/mkl/mkl_packed.md
index a9fb56d709..55d13d777c 100644
--- a/doc/design/mkl/mkl_packed.md
+++ b/doc/design/mkl/mkl_packed.md
@@ -1,4 +1,4 @@
-# Intel® MKL Packed Optimization on PaddlePaddle: Design Doc
+# Intel® MKL Packed on PaddlePaddle: Design Doc
 
 
 ## Contents
@@ -17,13 +17,17 @@
 
 ## Overview
 我们计划将 Intel® MKL 中引入的 GEMM Packed APIs\[[1](#references)\] 集成到 PaddlePaddle 中，充分发挥英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
-现阶段的优化主要针对 Recurrent Neural Network(以下简称RNN)相关层（包括`RecurrentLayer`, `GatedRecurrentLayer`和`LstmLayer`）， 以及 PaddlePaddle V1 API。
+现阶段的优化主要针对 Recurrent Neural Network（以下简称RNN）相关层（包括`RecurrentLayer`, `GatedRecurrentLayer`和`LstmLayer`）， 以及 PaddlePaddle V1 API。
 
 ## Key Points
 
 ### Background
-为了达到最佳性能， Intel® MKL 中的 cblas_?gemm 会在计算前将原数据转换为更适合英特尔平台的Packed格式， 这一数据格式的转换操作 (Packing)，在问题本身的计算量比较小的时候显得相对来说较为耗时。
-在现有的某些情况下（例如RNN），多次调用 cblas_?gemm 时会使用相同的原数据，每次调用时对原数据的重复Packing便成为了冗余。
+目前PaddlePaddle采用了 Intel® MKL库的cblas_?gemm函数，这个函数本身会在计算前将原数据转换为更适合英特尔平台的内部格式。
+
+1. 转换耗时 \
+这一数据格式的转换操作（Packing），在问题本身的计算量比较小的时候，显得相对来说较为耗时。例如在DeepSpeech2 \[[2](#references)\] 的Vanilla RNN部分中，矩阵大小是`batch_size * 2048`。
+2. 转换冗余 \
+由于在现有的某些情况下（例如RNN），多次调用 cblas_?gemm 会使用相同的原数据，因此，每次调用时对原数据的重复Packing便成为了冗余。
 
 为了最大程度减少多次调用 cblas_?gemm 在Packing上的耗时，Intel® MKL 引入了以下四个API:
    * cblas_?gemm_alloc
@@ -34,16 +38,16 @@
 通过使用这些API，我们可以先完成对原数据的Packing操作，再把已转换为Packed格式的数据传递给那些复用同一数据的gemm_compute函数，从而避免了Packing冗余。
 
 ### Solution
-在RNN的case下，同一次 forward/backward 过程中所有time state共享同一个weight矩阵。当只做 inference 时，各次 forward 之间也都使用相同的weight矩阵，没有必要在每次forward中每个time state的计算时对weight进行重复的packing操作。
+在RNN的情况下，同一次**前向/后向**（forward/backward）过程中所有**时间步**（time step）共享同一个**权重**（weight）。当只做**预测**（inference）时，各次**前向**之间也都使用了相同的**权重**，没有必要在每次**前向**中每个**时间步**的计算时对**权重**进行重复的Packing操作。
 
-我们通过使用新引入的GEMM Packed APIs，在layer init时先完成对weight的packing操作，然后在 forward/backward 时复用已pack过后的weight，并在每次weight更新后重新Packing。
+我们通过使用新引入的GEMM Packed APIs，在层**初始化**的时时候，先完成对**权重**的Packing操作，然后在**前向/后向**时复用已经转换过的**权重**，并在每次**权重**更新后，对新的**权重**进行转换用于下次迭代。
 
-* 优化前，对于sequence length = `T` 的model, `N` 次iteration执行的Packing次数为：   
-  - `inference`: `N * T`  
-  - `training`: `2 * N * T`
-* 优化后，对于sequence length = `T` 的model, `N` 次iteration执行的Packing次数减少至：
-  - `inference`: `1`    
-  - `training`: `2 * N`
+* 优化前，对于序列长度（sequence length）为`T`的网络模型（model）, `N`次迭代执行的转换次数为：
+  - `inference`： `N * T`  
+  - `training`： `2 * N * T`
+* 优化后，对于同样设置的网络模型，其转换次数减少至：
+  - `inference`： `1`    
+  - `training`： `2 * N`
 
 ## Actions
 
@@ -71,7 +75,7 @@ PaddlePaddle/Paddle
 在对应的`CMakeLists.txt`中根据`WITH_MKL`是否打开，来决定是否开启MKL Packed相关功能。
 
 ### Layers
-所有的`MKLPacked*Layer`都继承于PaddlePaddle的基类`Layer`, 并添加头文件 `MKLPackedGemm.h`，该文件中实现的对相关GEMM Packed APIs做了封装。
+所有的`MKLPacked*Layer`都继承于PaddlePaddle的基类`Layer`, 并添加头文件 `MKLPackedGemm.h`，该文件对相关GEMM Packed APIs做了封装。
 
 ### Unit Tests
 我们会添加`test_MKLPacked.cpp`用于MKL Packed优化后layer的测试。
@@ -87,5 +91,5 @@ TBD
 
 ## References 
 1. [Introducing the new Packed APIs for GEMM](https://software.intel.com/en-us/articles/introducing-the-new-packed-apis-for-gemm)
-
+2. [DeepSpeech2 on PaddlePaddle](https://github.com/PaddlePaddle/DeepSpeech#deepspeech2-on-paddlepaddle)
 

From c1a6870706bf593e6f8e394d5cbcbf20f9fd1f23 Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Sat, 16 Dec 2017 15:06:17 +0800
Subject: [PATCH 83/94] follow comments and refine doc

---
 doc/design/mkl/mkl_packed.md | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/doc/design/mkl/mkl_packed.md b/doc/design/mkl/mkl_packed.md
index 55d13d777c..c07f7d0cbe 100644
--- a/doc/design/mkl/mkl_packed.md
+++ b/doc/design/mkl/mkl_packed.md
@@ -22,7 +22,7 @@
 ## Key Points
 
 ### Background
-目前PaddlePaddle采用了 Intel® MKL库的cblas_?gemm函数，这个函数本身会在计算前将原数据转换为更适合英特尔平台的内部格式。
+目前PaddlePaddle采用了 Intel® MKL库的[cblas_?gemm](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm)函数，这个函数本身会在计算前将原数据转换为更适合英特尔平台的内部格式。
 
 1. 转换耗时 \
 这一数据格式的转换操作（Packing），在问题本身的计算量比较小的时候，显得相对来说较为耗时。例如在DeepSpeech2 \[[2](#references)\] 的Vanilla RNN部分中，矩阵大小是`batch_size * 2048`。
@@ -38,9 +38,9 @@
 通过使用这些API，我们可以先完成对原数据的Packing操作，再把已转换为Packed格式的数据传递给那些复用同一数据的gemm_compute函数，从而避免了Packing冗余。
 
 ### Solution
-在RNN的情况下，同一次**前向/后向**（forward/backward）过程中所有**时间步**（time step）共享同一个**权重**（weight）。当只做**预测**（inference）时，各次**前向**之间也都使用了相同的**权重**，没有必要在每次**前向**中每个**时间步**的计算时对**权重**进行重复的Packing操作。
+在RNN的情况下，同一次前向、后向（forward/backward）过程中所有时间步（time step）共享同一个权重（weight）。当只做推断（inference）时，各次前向之间也都使用了相同的权重，没有必要在每次前向中每个时间步的计算时对权重进行重复的Packing操作。
 
-我们通过使用新引入的GEMM Packed APIs，在层**初始化**的时时候，先完成对**权重**的Packing操作，然后在**前向/后向**时复用已经转换过的**权重**，并在每次**权重**更新后，对新的**权重**进行转换用于下次迭代。
+我们通过使用新引入的GEMM Packed APIs，在层初始化的时候，先完成对权重的Packing操作，然后在前向，后向时复用已经转换过的权重，并在每次权重更新后，对新的权重进行转换用于下次迭代。
 
 * 优化前，对于序列长度（sequence length）为`T`的网络模型（model）, `N`次迭代执行的转换次数为：
   - `inference`： `N * T`  

From 0fce0fe6983f4f167b873465fc90cff08fc31bd9 Mon Sep 17 00:00:00 2001
From: dangqingqing <dangqingqing@baidu.com>
Date: Fri, 15 Dec 2017 15:48:58 +0800
Subject: [PATCH 84/94] Reduce memory usage in conv layer and RoI layer for
 mobile inference.

---
 paddle/function/GemmConvOp.cpp         |  5 +++++
 paddle/gserver/layers/ROIPoolLayer.cpp | 27 +++++++++++++++++---------
 2 files changed, 23 insertions(+), 9 deletions(-)

diff --git a/paddle/function/GemmConvOp.cpp b/paddle/function/GemmConvOp.cpp
index 8d34eee886..ffbf366fa9 100644
--- a/paddle/function/GemmConvOp.cpp
+++ b/paddle/function/GemmConvOp.cpp
@@ -233,6 +233,11 @@ public:
       inputGrad += inputChannels * inputHeight * inputWidth;
       outputGrad += outputChannels * outputHeight * outputWidth;
     }
+#ifdef PADDLE_MOBILE_INFERENCE
+    if (Device == DEVICE_TYPE_CPU) {
+      delete memory_;
+    }
+#endif
   }
 };
 
diff --git a/paddle/gserver/layers/ROIPoolLayer.cpp b/paddle/gserver/layers/ROIPoolLayer.cpp
index 2c8256b91c..7d7c30b4d8 100644
--- a/paddle/gserver/layers/ROIPoolLayer.cpp
+++ b/paddle/gserver/layers/ROIPoolLayer.cpp
@@ -84,12 +84,15 @@ void ROIPoolLayer::forward(PassType passType) {
   size_t poolChannelOffset = pooledHeight_ * pooledWidth_;
 
   real* outputData = outputValue->getData();
-  Matrix::resizeOrCreate(maxIdxs_,
-                         numROIs,
-                         channels_ * pooledHeight_ * pooledWidth_,
-                         false,
-                         false);
-  real* argmaxData = maxIdxs_->getData();
+  real* argmaxData = nullptr;
+  if (passType != PASS_TEST) {
+    Matrix::resizeOrCreate(maxIdxs_,
+                           numROIs,
+                           channels_ * pooledHeight_ * pooledWidth_,
+                           false,
+                           false);
+    argmaxData = maxIdxs_->getData();
+  }
 
   for (size_t n = 0; n < numROIs; ++n) {
     // the first five elememts of each RoI should be:
@@ -128,14 +131,18 @@ void ROIPoolLayer::forward(PassType passType) {
           bool isEmpty = (hend <= hstart) || (wend <= wstart);
           size_t poolIndex = ph * pooledWidth_ + pw;
           outputData[poolIndex] = isEmpty ? 0 : -FLT_MAX;
-          argmaxData[poolIndex] = -1;
+          if (argmaxData) {
+            argmaxData[poolIndex] = -1;
+          }
 
           for (size_t h = hstart; h < hend; ++h) {
             for (size_t w = wstart; w < wend; ++w) {
               size_t index = h * width_ + w;
               if (batchData[index] > outputData[poolIndex]) {
                 outputData[poolIndex] = batchData[index];
-                argmaxData[poolIndex] = index;
+                if (argmaxData) {
+                  argmaxData[poolIndex] = index;
+                }
               }
             }
           }
@@ -143,7 +150,9 @@ void ROIPoolLayer::forward(PassType passType) {
       }
       batchData += channelOffset;
       outputData += poolChannelOffset;
-      argmaxData += poolChannelOffset;
+      if (argmaxData) {
+        argmaxData += poolChannelOffset;
+      }
     }
     bottomROIs += roiOffset;
   }

From 349609207e0d6b6a2ff1548b8d0773c7857989f0 Mon Sep 17 00:00:00 2001
From: dangqingqing <dangqingqing@baidu.com>
Date: Fri, 15 Dec 2017 16:22:19 +0800
Subject: [PATCH 85/94] Fix the error function/GemmConvOp.

---
 paddle/function/GemmConvOp.cpp | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/paddle/function/GemmConvOp.cpp b/paddle/function/GemmConvOp.cpp
index ffbf366fa9..3387cae396 100644
--- a/paddle/function/GemmConvOp.cpp
+++ b/paddle/function/GemmConvOp.cpp
@@ -126,6 +126,11 @@ public:
       inputData += inputChannels * inputHeight * inputWidth;
       outputData += outputChannels * outputHeight * outputWidth;
     }
+#ifdef PADDLE_MOBILE_INFERENCE
+    if (Device == DEVICE_TYPE_CPU) {
+      delete memory_;
+    }
+#endif
   }
 };
 
@@ -233,11 +238,6 @@ public:
       inputGrad += inputChannels * inputHeight * inputWidth;
       outputGrad += outputChannels * outputHeight * outputWidth;
     }
-#ifdef PADDLE_MOBILE_INFERENCE
-    if (Device == DEVICE_TYPE_CPU) {
-      delete memory_;
-    }
-#endif
   }
 };
 

From 1b0c7d7c7afd4c5f09eb43358dd4851ba1735c3f Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Fri, 15 Dec 2017 16:24:27 +0800
Subject: [PATCH 86/94] Simplize system_allocator and fix GPU_INFO (#6653)

---
 paddle/memory/detail/system_allocator.cc | 50 ++++++------------------
 paddle/platform/gpu_info.cc              | 19 ++++-----
 2 files changed, 23 insertions(+), 46 deletions(-)

diff --git a/paddle/memory/detail/system_allocator.cc b/paddle/memory/detail/system_allocator.cc
index 6a815a1b57..509250debc 100644
--- a/paddle/memory/detail/system_allocator.cc
+++ b/paddle/memory/detail/system_allocator.cc
@@ -19,6 +19,7 @@ limitations under the License. */
 
 #include <stdlib.h>    // for malloc and free
 #include <sys/mman.h>  // for mlock and munlock
+#include <algorithm>   // for std::max
 
 #include "gflags/gflags.h"
 
@@ -28,7 +29,7 @@ limitations under the License. */
 // of memory available to the system for paging.  So, by default, we
 // should set false to use_pinned_memory.
 DEFINE_bool(use_pinned_memory, true, "If set, allocate cpu pinned memory.");
-
+DECLARE_double(fraction_of_gpu_memory_to_use);
 namespace paddle {
 namespace memory {
 namespace detail {
@@ -77,45 +78,20 @@ void* GPUAllocator::Alloc(size_t& index, size_t size) {
   // CUDA documentation doesn't explain if cudaMalloc returns nullptr
   // if size is 0.  We just make sure it does.
   if (size <= 0) return nullptr;
-
-  size_t available = 0;
-  size_t capacity = 0;
-  paddle::platform::GpuMemoryUsage(available, capacity);
-
-  // Reserve memory for page tables, etc.
-  size_t reserving = 0.05 * capacity + paddle::platform::GpuMinChunkSize();
-  size_t usable = available > reserving ? available - reserving : 0;
-
-  // If remaining size no less than expected size, using general
-  // cudaMalloc to allocate GPU memory.
-  void* p = 0;
-  if (size <= usable) {
-    cudaError_t result = cudaMalloc(&p, size);
-    if (result == cudaSuccess) {
-      index = 0;
-      gpu_alloc_size_ += size;
-      return p;
-    }
-  }
-
-  // If remaining size less than expected size or cudaMalloc failed,
-  // cudaMallocHost will be considered as a fallback allocator.
-  //
-  // NOTE: here, we use GpuMaxAllocSize() as the maximum memory size
-  // of host fallback allocation. Allocates too much would reduce
-  // the amount of memory available to the underlying system for paging.
-  usable = paddle::platform::GpuMaxAllocSize() - fallback_alloc_size_;
-
-  if (size > usable) return nullptr;
-
-  cudaError_t result = cudaMallocHost(&p, size);
+  void* p;
+  cudaError_t result = cudaMalloc(&p, size);
   if (result == cudaSuccess) {
-    index = 1;
-    fallback_alloc_size_ += size;
+    index = 0;
+    gpu_alloc_size_ += size;
     return p;
+  } else {
+    LOG(WARNING)
+        << "Cannot malloc " << size / 1024.0 / 1024.0
+        << " MB GPU memory. Please shrink FLAGS_fraction_of_gpu_memory_to_use "
+           "environment variable to a lower value. Current value is "
+        << FLAGS_fraction_of_gpu_memory_to_use;
+    return nullptr;
   }
-
-  return nullptr;
 }
 
 void GPUAllocator::Free(void* p, size_t size, size_t index) {
diff --git a/paddle/platform/gpu_info.cc b/paddle/platform/gpu_info.cc
index 4fa2eaed31..541eca5f39 100644
--- a/paddle/platform/gpu_info.cc
+++ b/paddle/platform/gpu_info.cc
@@ -73,19 +73,20 @@ size_t GpuMaxChunkSize() {
   size_t available = 0;
 
   GpuMemoryUsage(available, total);
-
-  // Reserving the rest memory for page tables, etc.
-  size_t reserving = 0.05 * total;
-
+  VLOG(10) << "GPU Usage " << available / 1024 / 1024 << "M/"
+           << total / 1024 / 1024 << "M";
+  size_t reserving = static_cast<size_t>(0.05 * total);
   // If available less than minimum chunk size, no usable memory exists.
   available =
-      std::max(std::max(available, GpuMinChunkSize()) - GpuMinChunkSize(),
-               reserving) -
-      reserving;
+      std::min(std::max(available, GpuMinChunkSize()) - GpuMinChunkSize(),
+               total - reserving);
+
+  // Reserving the rest memory for page tables, etc.
 
-  size_t allocating = FLAGS_fraction_of_gpu_memory_to_use * total;
+  size_t allocating = static_cast<size_t>(FLAGS_fraction_of_gpu_memory_to_use *
+                                          (total - reserving));
 
-  PADDLE_ENFORCE_LT(allocating, available);
+  PADDLE_ENFORCE_LE(allocating, available);
 
   return allocating;
 }

From e1049c478697329b504ac682125210112fc13295 Mon Sep 17 00:00:00 2001
From: Luo Tao <luotao02@baidu.com>
Date: Fri, 15 Dec 2017 17:18:28 +0800
Subject: [PATCH 87/94] add fluid python api in chinese website

---
 doc/api/index_cn.rst | 1 +
 1 file changed, 1 insertion(+)

diff --git a/doc/api/index_cn.rst b/doc/api/index_cn.rst
index 9be0b370ee..84f9097a6c 100644
--- a/doc/api/index_cn.rst
+++ b/doc/api/index_cn.rst
@@ -7,3 +7,4 @@ API
     模型配置 <v2/model_configs.rst>
     数据访问 <v2/data.rst>
     训练与应用 <v2/run_logic.rst>
+    v2/fluid.rst

From 84cb542c13905f5113a6e91d86129ec7a1a90e06 Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Fri, 15 Dec 2017 17:23:03 +0800
Subject: [PATCH 88/94] use intel openmp to speedup seq2batch when WITH_MKL

---
 paddle/gserver/layers/SequenceToBatch.cpp | 25 +++++++++++++++++------
 1 file changed, 19 insertions(+), 6 deletions(-)

diff --git a/paddle/gserver/layers/SequenceToBatch.cpp b/paddle/gserver/layers/SequenceToBatch.cpp
index 5fa7b6f488..7892d475a6 100644
--- a/paddle/gserver/layers/SequenceToBatch.cpp
+++ b/paddle/gserver/layers/SequenceToBatch.cpp
@@ -171,12 +171,25 @@ void SequenceToBatch::sequence2BatchCopy(Matrix &batch,
     hl_sequence2batch_copy(
         batchData, seqData, idxData, seqWidth, batchCount, seq2batch);
   } else {
-    for (int i = 0; i < batchCount; ++i) {
-      if (seq2batch) {
-        memcpy(batch.rowBuf(i),
-               sequence.rowBuf(idxData[i]),
-               seqWidth * sizeof(real));
-      } else {
+    if (seq2batch) {
+      const int blockMemSize = 8 * 1024;
+      const int blockSize = blockMemSize / sizeof(real);
+#ifdef PADDLE_USE_MKLML
+#pragma omp parallel for collapse(2)
+#endif
+      for (int i = 0; i < batchCount; ++i) {
+        for (int j = 0; j < seqWidth; j += blockSize) {
+          memcpy(batch.rowBuf(i) + j,
+                 sequence.rowBuf(idxData[i]) + j,
+                 (j + blockSize > seqWidth) ? (seqWidth - j) * sizeof(real)
+                                            : blockMemSize);
+        }
+      }
+    } else {
+#ifdef PADDLE_USE_MKLML
+#pragma omp parallel for
+#endif
+      for (int i = 0; i < batchCount; ++i) {
         memcpy(sequence.rowBuf(idxData[i]),
                batch.rowBuf(i),
                seqWidth * sizeof(real));

From 9c27c13e469f4c2f11f00c6e2570927e1d640e3c Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Sat, 16 Dec 2017 21:16:01 +0800
Subject: [PATCH 89/94] follow comments using macro to separate the original
 implements

---
 paddle/gserver/layers/SequenceToBatch.cpp | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/paddle/gserver/layers/SequenceToBatch.cpp b/paddle/gserver/layers/SequenceToBatch.cpp
index 7892d475a6..6b769378d2 100644
--- a/paddle/gserver/layers/SequenceToBatch.cpp
+++ b/paddle/gserver/layers/SequenceToBatch.cpp
@@ -172,11 +172,10 @@ void SequenceToBatch::sequence2BatchCopy(Matrix &batch,
         batchData, seqData, idxData, seqWidth, batchCount, seq2batch);
   } else {
     if (seq2batch) {
+#ifdef PADDLE_USE_MKLML
       const int blockMemSize = 8 * 1024;
       const int blockSize = blockMemSize / sizeof(real);
-#ifdef PADDLE_USE_MKLML
 #pragma omp parallel for collapse(2)
-#endif
       for (int i = 0; i < batchCount; ++i) {
         for (int j = 0; j < seqWidth; j += blockSize) {
           memcpy(batch.rowBuf(i) + j,
@@ -185,6 +184,13 @@ void SequenceToBatch::sequence2BatchCopy(Matrix &batch,
                                             : blockMemSize);
         }
       }
+#else
+      for (int i = 0; i < batchCount; ++i) {
+        memcpy(batch.rowBuf(i),
+               sequence.rowBuf(idxData[i]),
+               seqWidth * sizeof(real));
+      }
+#endif
     } else {
 #ifdef PADDLE_USE_MKLML
 #pragma omp parallel for

From 77cf7d4f9e42d4c153aad648bbe5d7049ab2ff29 Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Sat, 16 Dec 2017 18:49:50 +0800
Subject: [PATCH 90/94] Fix read_source.md (#6673)

Optimizer is not a part of fluid. The `paddle::platform`
was missing.
---
 doc/howto/read_source.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/doc/howto/read_source.md b/doc/howto/read_source.md
index 383acb0c82..e4211abb3b 100644
--- a/doc/howto/read_source.md
+++ b/doc/howto/read_source.md
@@ -6,10 +6,10 @@ Core: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework
 
 Operator: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators
 
-Optimizer: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/optimizer
-
 Memory: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory
 
+Platform: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform
+
 # Compile Time
 
 The following **defines** the NN. The definition goes into this [protocol buffer](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto).

From 1eac27630047e59cb461846022bada5bf0f490c1 Mon Sep 17 00:00:00 2001
From: sweetsky0901 <work@yq01-idl-gpu-online20.yq01.baidu.com>
Date: Sun, 17 Dec 2017 22:09:29 +0800
Subject: [PATCH 91/94] add spp avg

---
 paddle/operators/spp_op.cc                  |  5 +++
 paddle/operators/spp_op.h                   | 34 ++++++++++++++++-----
 python/paddle/v2/fluid/tests/test_spp_op.py | 23 +++++++++++---
 3 files changed, 51 insertions(+), 11 deletions(-)

diff --git a/paddle/operators/spp_op.cc b/paddle/operators/spp_op.cc
index c4bd4f5ab3..b1807b6261 100644
--- a/paddle/operators/spp_op.cc
+++ b/paddle/operators/spp_op.cc
@@ -30,6 +30,11 @@ class SppOpMaker : public framework::OpProtoAndCheckerMaker {
               "N * M."
               "M = C * H * W");
     AddAttr<int>("pyramid_height", "(int), multi level pooling");
+    AddAttr<std::string>(
+        "pooling_type",
+        "(string), pooling type, can be \"max\" for max-pooling "
+        "and \"avg\" for average-pooling.")
+        .InEnum({"max", "avg"});
     AddComment(R"DOC(
         "With spatial pyramid pooling, the input image can
         be of any sizes. This not only allows arbitrary aspect
diff --git a/paddle/operators/spp_op.h b/paddle/operators/spp_op.h
index 16510cb826..f35b305d02 100644
--- a/paddle/operators/spp_op.h
+++ b/paddle/operators/spp_op.h
@@ -27,6 +27,8 @@ class SppKernel : public framework::OpKernel<T> {
     const framework::Tensor* in_x = context.Input<framework::Tensor>("X");
     auto* out = context.Output<framework::Tensor>("Out");
     int pyramid_height = context.template Attr<int>("pyramid_height");
+    std::string pooling_type =
+        context.template Attr<std::string>("pooling_type");
     out->mutable_data<T>(context.GetPlace());
     auto out_stride = framework::stride(out->dims());
     int input_h = in_x->dims()[2];
@@ -48,10 +50,17 @@ class SppKernel : public framework::OpKernel<T> {
       framework::DDim output_shape(framework::make_ddim(output_shape_vec));
       out_level.mutable_data<T>(output_shape, context.GetPlace());
       // pooling
-      math::Pool2dFunctor<DeviceContext, math::MaxPool<T>, T> pool_forward;
-      math::MaxPool<T> max_process;
-      pool_forward(context.template device_context<DeviceContext>(), *in_x,
-                   kernel_size, strides, paddings, max_process, &out_level);
+      if (pooling_type == "max") {
+        math::Pool2dFunctor<DeviceContext, math::MaxPool<T>, T> pool_forward;
+        math::MaxPool<T> max_process;
+        pool_forward(context.template device_context<DeviceContext>(), *in_x,
+                     kernel_size, strides, paddings, max_process, &out_level);
+      } else if (pooling_type == "avg") {
+        math::Pool2dFunctor<DeviceContext, math::AvgPool<T>, T> pool_forward;
+        math::AvgPool<T> avg_process;
+        pool_forward(context.template device_context<DeviceContext>(), *in_x,
+                     kernel_size, strides, paddings, avg_process, &out_level);
+      }
       // flatten pooling output shape
       int output_flatten_w = in_x->dims()[1] * bins * bins;
       std::vector<int64_t> output_flatten_shape_vec(
@@ -79,6 +88,8 @@ class SppGradKernel : public framework::OpKernel<T> {
     framework::Tensor* in_x_grad =
         context.Output<framework::Tensor>(framework::GradVarName("X"));
     int pyramid_height = context.template Attr<int>("pyramid_height");
+    std::string pooling_type =
+        context.template Attr<std::string>("pooling_type");
     auto& device_ctx = context.template device_context<DeviceContext>();
     math::SetConstant<DeviceContext, T> zero;
     in_x_grad->mutable_data<T>(context.GetPlace());
@@ -130,10 +141,19 @@ class SppGradKernel : public framework::OpKernel<T> {
       outgrad_level.ShareDataWith(outgrad_level);
       outgrad_level.Resize(out_shape);
       // pooling backward
-      math::MaxPool2dGradFunctor<DeviceContext, T> pool2d_backward;
-      pool2d_backward(context.template device_context<DeviceContext>(), *in_x,
+      if (pooling_type == "max") {
+        math::MaxPool2dGradFunctor<DeviceContext, T> pool2d_backward;
+        pool2d_backward(context.template device_context<DeviceContext>(), *in_x,
+                        *&out_level, *&outgrad_level, kernel_size, strides,
+                        paddings, in_x_grad);
+      } else if (pooling_type == "avg") {
+        math::Pool2dGradFunctor<DeviceContext, math::AvgPoolGrad<T>, T>
+            pool_backward;
+        math::AvgPoolGrad<T> avg_process;
+        pool_backward(context.template device_context<DeviceContext>(), *in_x,
                       *&out_level, *&outgrad_level, kernel_size, strides,
-                      paddings, in_x_grad);
+                      paddings, avg_process, in_x_grad);
+      }
     }
   }
 };
diff --git a/python/paddle/v2/fluid/tests/test_spp_op.py b/python/paddle/v2/fluid/tests/test_spp_op.py
index b57f4a795d..007723f0e3 100644
--- a/python/paddle/v2/fluid/tests/test_spp_op.py
+++ b/python/paddle/v2/fluid/tests/test_spp_op.py
@@ -2,6 +2,7 @@ import unittest
 import numpy as np
 from op_test import OpTest
 from test_pool2d_op import max_pool2D_forward_naive
+from test_pool2d_op import avg_pool2D_forward_naive
 
 
 class TestSppOp(OpTest):
@@ -24,8 +25,8 @@ class TestSppOp(OpTest):
                                      bins.astype("double")).astype("int32")
             padding[1] = (
                 (kernel_size[1] * bins - wsize + 1) / 2).astype("int32")
-            out_level = max_pool2D_forward_naive(input, kernel_size,
-                                                 kernel_size, padding)
+            out_level = self.pool2D_forward_naive(input, kernel_size,
+                                                  kernel_size, padding)
             out_level_flatten.append(
                 out_level.reshape(nsize, bins * bins * csize))
             if i == 0:
@@ -34,7 +35,10 @@ class TestSppOp(OpTest):
                 output = np.concatenate((output, out_level_flatten[i]), 1)
         # output = np.concatenate(out_level_flatten.tolist(), 0);
         self.inputs = {'X': input.astype('float32'), }
-        self.attrs = {'pyramid_height': self.pyramid_height}
+        self.attrs = {
+            'pyramid_height': self.pyramid_height,
+            'pooling_type': self.pool_type
+        }
 
         self.outputs = {'Out': output.astype('float32')}
 
@@ -42,11 +46,22 @@ class TestSppOp(OpTest):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Out', max_relative_error=0.05)
+        if self.pool_type != "avg":
+            self.check_grad(['X'], 'Out', max_relative_error=0.05)
 
     def init_test_case(self):
         self.shape = [3, 2, 4, 4]
         self.pyramid_height = 3
+        self.pool2D_forward_naive = max_pool2D_forward_naive
+        self.pool_type = "max"
+
+
+class TestCase2(TestSppOp):
+    def init_test_case(self):
+        self.shape = [3, 2, 4, 4]
+        self.pyramid_height = 3
+        self.pool2D_forward_naive = avg_pool2D_forward_naive
+        self.pool_type = "avg"
 
 
 if __name__ == '__main__':

From 93a2d9c59d46589d3fefaaee09a3003f45fda9e5 Mon Sep 17 00:00:00 2001
From: QI JUN <qijun1994@hotmail.com>
Date: Mon, 18 Dec 2017 11:53:22 +0800
Subject: [PATCH 92/94] add more place test and rename Cudnn to CUDNN (#6621)

* add more place_test and rename Cudnn to CUDNN

* fix ci
---
 paddle/operators/math/math_function.cu |  2 +-
 paddle/platform/device_context.cc      |  8 ++++----
 paddle/platform/device_context.h       |  8 ++++----
 paddle/platform/device_context_test.cc | 10 +++++-----
 paddle/platform/place.h                |  8 ++++----
 paddle/platform/place_test.cc          |  6 ++++++
 6 files changed, 24 insertions(+), 18 deletions(-)

diff --git a/paddle/operators/math/math_function.cu b/paddle/operators/math/math_function.cu
index e33070c40f..7852bb53a9 100644
--- a/paddle/operators/math/math_function.cu
+++ b/paddle/operators/math/math_function.cu
@@ -274,7 +274,7 @@ void set_constant_with_place<platform::GPUPlace>(
 }
 
 template <>
-void set_constant_with_place<platform::CudnnPlace>(
+void set_constant_with_place<platform::CUDNNPlace>(
     const platform::DeviceContext& context, framework::Tensor* tensor,
     float value) {
   set_constant_with_place<platform::GPUPlace>(context, tensor, value);
diff --git a/paddle/platform/device_context.cc b/paddle/platform/device_context.cc
index 1c72b50559..8cdc5f4340 100644
--- a/paddle/platform/device_context.cc
+++ b/paddle/platform/device_context.cc
@@ -125,21 +125,21 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_handle_; }
 
 cudaStream_t CUDADeviceContext::stream() const { return stream_; }
 
-CudnnDeviceContext::CudnnDeviceContext(CudnnPlace place)
+CUDNNDeviceContext::CUDNNDeviceContext(CUDNNPlace place)
     : CUDADeviceContext(place), place_(place) {
   PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
   PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream()));
 }
 
-CudnnDeviceContext::~CudnnDeviceContext() {
+CUDNNDeviceContext::~CUDNNDeviceContext() {
   SetDeviceId(place_.device);
   Wait();
   PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_));
 }
 
-Place CudnnDeviceContext::GetPlace() const { return CudnnPlace(); }
+Place CUDNNDeviceContext::GetPlace() const { return CUDNNPlace(); }
 
-cudnnHandle_t CudnnDeviceContext::cudnn_handle() const { return cudnn_handle_; }
+cudnnHandle_t CUDNNDeviceContext::cudnn_handle() const { return cudnn_handle_; }
 
 #endif
 
diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h
index f67194993d..56813a1d5b 100644
--- a/paddle/platform/device_context.h
+++ b/paddle/platform/device_context.h
@@ -86,10 +86,10 @@ class CUDADeviceContext : public DeviceContext {
   cublasHandle_t cublas_handle_;
 };
 
-class CudnnDeviceContext : public CUDADeviceContext {
+class CUDNNDeviceContext : public CUDADeviceContext {
  public:
-  explicit CudnnDeviceContext(CudnnPlace place);
-  virtual ~CudnnDeviceContext();
+  explicit CUDNNDeviceContext(CUDNNPlace place);
+  virtual ~CUDNNDeviceContext();
 
   /*! \brief  Return place in the device context. */
   Place GetPlace() const final;
@@ -99,7 +99,7 @@ class CudnnDeviceContext : public CUDADeviceContext {
 
  private:
   cudnnHandle_t cudnn_handle_;
-  CudnnPlace place_;
+  CUDNNPlace place_;
 };
 
 #endif
diff --git a/paddle/platform/device_context_test.cc b/paddle/platform/device_context_test.cc
index be3b2af5af..109c13a881 100644
--- a/paddle/platform/device_context_test.cc
+++ b/paddle/platform/device_context_test.cc
@@ -47,14 +47,14 @@ TEST(Device, CUDADeviceContext) {
   }
 }
 
-TEST(Device, CudnnDeviceContext) {
-  using paddle::platform::CudnnDeviceContext;
-  using paddle::platform::CudnnPlace;
+TEST(Device, CUDNNDeviceContext) {
+  using paddle::platform::CUDNNDeviceContext;
+  using paddle::platform::CUDNNPlace;
   if (paddle::platform::dynload::HasCUDNN()) {
     int count = paddle::platform::GetCUDADeviceCount();
     for (int i = 0; i < count; ++i) {
-      CudnnDeviceContext* device_context =
-          new CudnnDeviceContext(CudnnPlace(i));
+      CUDNNDeviceContext* device_context =
+          new CUDNNDeviceContext(CUDNNPlace(i));
       cudnnHandle_t cudnn_handle = device_context->cudnn_handle();
       ASSERT_NE(nullptr, cudnn_handle);
       ASSERT_NE(nullptr, device_context->stream());
diff --git a/paddle/platform/place.h b/paddle/platform/place.h
index 4526945792..ca98920d41 100644
--- a/paddle/platform/place.h
+++ b/paddle/platform/place.h
@@ -51,9 +51,9 @@ struct GPUPlace {
   int device;
 };
 
-struct CudnnPlace : public GPUPlace {
-  CudnnPlace() : GPUPlace() {}
-  explicit CudnnPlace(int d) : GPUPlace(d) {}
+struct CUDNNPlace : public GPUPlace {
+  CUDNNPlace() : GPUPlace() {}
+  explicit CUDNNPlace(int d) : GPUPlace(d) {}
 };
 
 struct IsGPUPlace : public boost::static_visitor<bool> {
@@ -72,7 +72,7 @@ struct IsMKLDNNPlace : public boost::static_visitor<bool> {
 // should be less equal than 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT)
 #define NUM_PLACE_TYPE_LIMIT_IN_BIT 4
 
-typedef boost::variant<CudnnPlace, GPUPlace, CPUPlace, MKLDNNPlace> Place;
+typedef boost::variant<CUDNNPlace, GPUPlace, CPUPlace, MKLDNNPlace> Place;
 
 // static check number of place types is less equal than
 // 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT)
diff --git a/paddle/platform/place_test.cc b/paddle/platform/place_test.cc
index 184af12c23..c536b59ed8 100644
--- a/paddle/platform/place_test.cc
+++ b/paddle/platform/place_test.cc
@@ -5,16 +5,22 @@
 TEST(Place, Equality) {
   paddle::platform::CPUPlace cpu;
   paddle::platform::GPUPlace g0(0), g1(1), gg0(0);
+  paddle::platform::CUDNNPlace d0(0), d1(1), dd0(0);
 
   EXPECT_EQ(cpu, cpu);
   EXPECT_EQ(g0, g0);
   EXPECT_EQ(g1, g1);
   EXPECT_EQ(g0, gg0);
+  EXPECT_EQ(d0, dd0);
 
   EXPECT_NE(g0, g1);
+  EXPECT_NE(d0, d1);
 
   EXPECT_TRUE(paddle::platform::places_are_same_class(g0, gg0));
   EXPECT_FALSE(paddle::platform::places_are_same_class(g0, cpu));
+
+  EXPECT_TRUE(paddle::platform::is_gpu_place(d0));
+  EXPECT_FALSE(paddle::platform::places_are_same_class(g0, d0));
 }
 
 TEST(Place, Default) {

From 964f01e3e880d4c835489ce5fad8215a964c188d Mon Sep 17 00:00:00 2001
From: qiaolongfei <qiaolongfei@baidu.com>
Date: Mon, 18 Dec 2017 11:57:20 +0800
Subject: [PATCH 93/94] fix simple_gru2 doc

---
 python/paddle/trainer_config_helpers/networks.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/python/paddle/trainer_config_helpers/networks.py b/python/paddle/trainer_config_helpers/networks.py
index 9776ae1805..8bfe56d795 100644
--- a/python/paddle/trainer_config_helpers/networks.py
+++ b/python/paddle/trainer_config_helpers/networks.py
@@ -1119,8 +1119,9 @@ def simple_gru2(input,
     :param gru_bias_attr: bias parameter attribute of gru layer, 
                           False means no bias, None means default bias.
     :type gru_bias_attr: ParameterAttribute|False|None
-    :param gru_layer_attr: Extra attribute of the gru layer.
-    :type gru_layer_attr: ExtraLayerAttribute
+    :param gru_param_attr: param parameter attribute of gru layer,
+                          None means default param.
+    :type gru_param_attr: ParameterAttribute|None
     :return: the gru group.
     :rtype: LayerOutput
     """

From 24fda392207e1fd608e66ea376a79f9f9716c3ca Mon Sep 17 00:00:00 2001
From: dzhwinter <dongzhihong01@baidu.com>
Date: Mon, 18 Dec 2017 13:01:41 +0800
Subject: [PATCH 94/94] Feature/global context (#6537)

* "add DeviceContextPool"

* "add devicecontextpool in pybind"

* "add comments in python side "

* "fix static link error"

* "fix CI error"

* "add executor.py"

* "fix CI error"

* "add with gpu macro"

* "remove comment out codes"

* "add TODO items"

* "update init devices"
---
 paddle/framework/CMakeLists.txt    |  3 ++
 paddle/framework/ddim_test.cc      | 13 +++++
 paddle/framework/executor.cc       | 33 ++----------
 paddle/framework/executor.h        | 84 ++++++++++++++++++++++++++++--
 paddle/framework/init.cc           | 80 ++++++++++++++++++++++++++++
 paddle/framework/init.h            | 28 ++++++++++
 paddle/framework/init_test.cc      | 27 ++++++++++
 paddle/pybind/CMakeLists.txt       |  2 +-
 paddle/pybind/pybind.cc            | 23 ++------
 python/paddle/v2/fluid/executor.py |  7 +++
 10 files changed, 248 insertions(+), 52 deletions(-)
 create mode 100644 paddle/framework/init.cc
 create mode 100644 paddle/framework/init.h
 create mode 100644 paddle/framework/init_test.cc

diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt
index 4b0eff3adb..206e298eb2 100644
--- a/paddle/framework/CMakeLists.txt
+++ b/paddle/framework/CMakeLists.txt
@@ -58,3 +58,6 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
         proto_desc)
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
+
+cc_library(init SRCS init.cc DEPS gflags executor place stringpiece)
+cc_test(init_test SRCS init_test.cc DEPS init)
diff --git a/paddle/framework/ddim_test.cc b/paddle/framework/ddim_test.cc
index 756232b1b5..bd5ea09d7d 100644
--- a/paddle/framework/ddim_test.cc
+++ b/paddle/framework/ddim_test.cc
@@ -1,3 +1,16 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
 #include <sstream>
 #include <vector>
 
diff --git a/paddle/framework/executor.cc b/paddle/framework/executor.cc
index 83aa927c29..a8b8a6f8e8 100644
--- a/paddle/framework/executor.cc
+++ b/paddle/framework/executor.cc
@@ -33,32 +33,12 @@ namespace framework {
 const std::string kFeedOpType = "feed";
 const std::string kFetchOpType = "fetch";
 
-Executor::Executor(const std::vector<platform::Place>& places) : own_(true) {
-  PADDLE_ENFORCE_GT(places.size(), 0);
-  device_contexts_.resize(places.size());
-  for (size_t i = 0; i < places.size(); i++) {
-    if (platform::is_cpu_place(places[i])) {
-      device_contexts_[i] = new platform::CPUDeviceContext(
-          boost::get<platform::CPUPlace>(places[i]));
-    } else if (platform::is_gpu_place(places[i])) {
-#ifdef PADDLE_WITH_CUDA
-      device_contexts_[i] = new platform::CUDADeviceContext(
-          boost::get<platform::GPUPlace>(places[i]));
-#else
-      PADDLE_THROW(
-          "'GPUPlace' is not supported, Please re-compile with WITH_GPU "
-          "option");
-#endif
-    }
-  }
-}
+DeviceContextPool* DeviceContextPool::pool = nullptr;
 
-Executor::~Executor() {
-  if (own_) {
-    for (auto& device_context : device_contexts_) {
-      delete device_context;
-    }
-  }
+Executor::Executor(const std::vector<platform::Place>& places) {
+  DeviceContextPool& pool = DeviceContextPool::Get();
+  auto borrowed_contexts = pool.Borrow(places);
+  device_contexts_.swap(borrowed_contexts);
 }
 
 static void CreateTensor(Variable* var, VarDesc::VarType var_type) {
@@ -132,8 +112,5 @@ void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
   }
 }
 
-Executor::Executor(const platform::DeviceContext& device)
-    : device_contexts_({&device}), own_(false) {}
-
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/executor.h b/paddle/framework/executor.h
index b745f4f647..073e04729b 100644
--- a/paddle/framework/executor.h
+++ b/paddle/framework/executor.h
@@ -14,19 +14,98 @@ limitations under the License. */
 
 #pragma once
 
+#include <map>
+#include <unordered_map>
+
 #include "paddle/framework/op_info.h"
 #include "paddle/framework/program_desc.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace framework {
 
+class DeviceContextPool {
+ public:
+  static DeviceContextPool& Get() {
+    PADDLE_ENFORCE_NOT_NULL(pool, "Need to Create DeviceContextPool first!");
+    return *pool;
+  }
+
+  static DeviceContextPool& Create(const std::vector<platform::Place>& places) {
+    if (pool == nullptr) {
+      pool = new DeviceContextPool(places);
+    }
+    return *pool;
+  }
+
+  std::vector<const platform::DeviceContext*> Borrow(
+      const std::vector<platform::Place>& places) {
+    PADDLE_ENFORCE_GT(places.size(), 0);
+    PADDLE_ENFORCE_LE(places.size(), device_contexts_.size());
+    std::vector<const platform::DeviceContext*> borrowed_contexts;
+    for (auto& place : places) {
+      auto range = device_contexts_.equal_range(place);
+      if (range.first == range.second) {
+        PADDLE_THROW(
+            "'Place' is not supported, Please re-compile with WITH_GPU "
+            "option");
+      }
+      // TODO(dzhwinter) : assign the first found device. Will enhanced later.
+      // device load balancer maybe useful here.
+      borrowed_contexts.emplace_back(range.first->second);
+    }
+    return borrowed_contexts;
+  }
+
+  explicit DeviceContextPool(const std::vector<platform::Place>& places) {
+    PADDLE_ENFORCE_GT(places.size(), 0);
+    for (size_t i = 0; i < places.size(); i++) {
+      if (platform::is_cpu_place(places[i])) {
+        device_contexts_.emplace(
+            places[i], new platform::CPUDeviceContext(
+                           boost::get<platform::CPUPlace>(places[i])));
+      } else if (platform::is_gpu_place(places[i])) {
+#ifdef PADDLE_WITH_CUDA
+        device_contexts_.emplace(
+            places[i], new platform::CUDADeviceContext(
+                           boost::get<platform::GPUPlace>(places[i])));
+#else
+        PADDLE_THROW(
+            "'GPUPlace' is not supported, Please re-compile with WITH_GPU "
+            "option");
+#endif
+      }
+    }
+  }
+
+  ~DeviceContextPool() {}
+
+ private:
+  static DeviceContextPool* pool;
+  struct Hash {
+    std::hash<int> hash_;
+    size_t operator()(const platform::Place& place) const {
+      return hash_(place.which());
+    }
+  };
+  std::unordered_multimap<const platform::Place, const platform::DeviceContext*,
+                          Hash>
+      device_contexts_;
+  DISABLE_COPY_AND_ASSIGN(DeviceContextPool);
+};
+
 class Executor {
  public:
+  // TODO(dzhwinter) : Do not rely on this function, it will be removed
+  explicit Executor(const platform::DeviceContext& device)
+      : Executor(std::vector<platform::Place>({device.GetPlace()})) {}
+
+  explicit Executor(const platform::Place& place)
+      : Executor(std::vector<platform::Place>({place})) {}
+
   explicit Executor(const std::vector<platform::Place>& places);
-  explicit Executor(const platform::DeviceContext& devices);
-  ~Executor();
 
   /* @Brief
    * Runtime evaluation of the given ProgramDesc under certain Scope
@@ -39,7 +118,6 @@ class Executor {
 
  private:
   std::vector<const platform::DeviceContext*> device_contexts_;
-  bool own_;
 };
 
 }  // namespace framework
diff --git a/paddle/framework/init.cc b/paddle/framework/init.cc
new file mode 100644
index 0000000000..1c4476f4b3
--- /dev/null
+++ b/paddle/framework/init.cc
@@ -0,0 +1,80 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+#include <algorithm>
+#include <string>
+
+#include "paddle/framework/executor.h"
+#include "paddle/framework/init.h"
+#include "paddle/platform/place.h"
+#include "paddle/string/piece.h"
+
+namespace paddle {
+namespace framework {
+
+std::once_flag gflags_init_flag;
+
+// TODO(qijun) move init gflags to init.cc
+void InitGflags(std::vector<std::string> &argv) {
+  std::call_once(gflags_init_flag, [&]() {
+    int argc = argv.size();
+    char **arr = new char *[argv.size()];
+    std::string line;
+    for (size_t i = 0; i < argv.size(); i++) {
+      arr[i] = &argv[i][0];
+      line += argv[i];
+      line += ' ';
+    }
+    google::ParseCommandLineFlags(&argc, &arr, true);
+    VLOG(1) << "Init commandline: " << line;
+  });
+}
+
+bool InitDevices(const std::vector<std::string> &devices) {
+  // device format
+  // CPU
+  // GPU:1
+  // TODO(dzhwinter) : add device format annotation for users.
+  std::vector<platform::Place> places;
+  for (auto &device : devices) {
+    auto p = string::Piece(device);
+    if (string::Find(p, ':', 0) == string::Piece::npos) {
+      places.emplace_back(platform::CPUPlace());
+    } else if (string::HasPrefix(p, "GPU")) {
+#ifdef PADDLE_WITH_CUDA
+      auto pos = string::RFind(p, ':', string::Piece::npos);
+      auto number = device.substr(pos + 1);
+      places.emplace_back(platform::GPUPlace(std::stoi(number)));
+#else
+      LOG(WARNING)
+          << "'GPU' is not supported, Please re-compile with WITH_GPU option";
+#endif
+    } else {
+      return false;
+    }
+  }
+
+  if (std::find_if(places.begin(), places.end(),
+                   [&](const platform::Place &place) {
+                     return platform::is_cpu_place(place);
+                   }) == places.end()) {
+    places.emplace_back(platform::CPUPlace());
+    LOG(WARNING) << "Not specified any device, use CPU by Default.";
+  }
+  DeviceContextPool::Create(places);
+  return true;
+  return true;
+}
+
+}  // namespace framework
+}  // namespace paddle
diff --git a/paddle/framework/init.h b/paddle/framework/init.h
new file mode 100644
index 0000000000..1715cd81e6
--- /dev/null
+++ b/paddle/framework/init.h
@@ -0,0 +1,28 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+#pragma once
+#include <mutex>
+
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+
+namespace paddle {
+namespace framework {
+
+void InitGflags(std::vector<std::string> &argv);
+
+bool InitDevices(const std::vector<std::string> &devices);
+
+}  // namespace framework
+}  // namespace paddle
diff --git a/paddle/framework/init_test.cc b/paddle/framework/init_test.cc
new file mode 100644
index 0000000000..f65e881a76
--- /dev/null
+++ b/paddle/framework/init_test.cc
@@ -0,0 +1,27 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+#include "gtest/gtest.h"
+
+#include "paddle/framework/init.h"
+
+TEST(Init, InitDevices) {
+  using paddle::framework::InitDevices;
+  std::vector<std::string> ds1 = {"CPU"};
+  ASSERT_EQ(InitDevices(ds1), true);
+
+#ifdef PADDLE_WITH_CUDA
+  std::vector<std::string> ds2 = {"CPU", "GPU:0", "GPU:1"};
+  ASSERT_EQ(InitDevices(ds2), true);
+#endif
+}
diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt
index fd55f410d3..1fb69de90d 100644
--- a/paddle/pybind/CMakeLists.txt
+++ b/paddle/pybind/CMakeLists.txt
@@ -1,7 +1,7 @@
 if(WITH_PYTHON)
   cc_library(paddle_pybind SHARED
     SRCS pybind.cc exception.cc protobuf.cc
-    DEPS pybind python backward proto_desc paddle_memory executor prune
+    DEPS pybind python backward proto_desc paddle_memory executor prune init
     ${GLOB_OP_LIB})
 endif(WITH_PYTHON)
 
diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index 1faf24bcb8..4248db34c6 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -16,11 +16,11 @@ limitations under the License. */
 
 #include <mutex>  // for call_once
 #include <unordered_map>
-#include "gflags/gflags.h"
 #include "paddle/framework/backward.h"
 #include "paddle/framework/executor.h"
 #include "paddle/framework/feed_fetch_method.h"
 #include "paddle/framework/framework.pb.h"
+#include "paddle/framework/init.h"
 #include "paddle/framework/lod_rank_table.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/lod_tensor_array.h"
@@ -51,24 +51,6 @@ static size_t UniqueIntegerGenerator(const std::string &prefix) {
   return generators[prefix].fetch_add(1);
 }
 
-std::once_flag gflags_init_flag;
-
-// TODO(qijun) move init gflags to init.cc
-void InitGflags(std::vector<std::string> &argv) {
-  std::call_once(gflags_init_flag, [&]() {
-    int argc = argv.size();
-    char **arr = new char *[argv.size()];
-    std::string line;
-    for (size_t i = 0; i < argv.size(); i++) {
-      arr[i] = &argv[i][0];
-      line += argv[i];
-      line += ' ';
-    }
-    google::ParseCommandLineFlags(&argc, &arr, true);
-    VLOG(1) << "Init commandline: " << line;
-  });
-}
-
 bool IsCompileGPU() {
 #ifndef PADDLE_WITH_CUDA
   return false;
@@ -438,7 +420,8 @@ All parameter, weight, gradient are variables in Paddle.
       .def("run", &Executor::Run);
 
   m.def("unique_integer", UniqueIntegerGenerator);
-  m.def("init_gflags", InitGflags);
+  m.def("init_gflags", framework::InitGflags);
+  m.def("init_devices", &framework::InitDevices);
 
   m.def("is_compile_gpu", IsCompileGPU);
   m.def("set_feed_variable", framework::SetFeedVariable);
diff --git a/python/paddle/v2/fluid/executor.py b/python/paddle/v2/fluid/executor.py
index bdc82eede9..9a99b045dc 100644
--- a/python/paddle/v2/fluid/executor.py
+++ b/python/paddle/v2/fluid/executor.py
@@ -46,6 +46,13 @@ class Executor(object):
             p.set_place(each)
             act_places.append(p)
 
+        # TODO(dzhwinter) : consider that our fluid tests all written in 
+        # GPUPlace(gpu_id), this will be changed in next PR.
+        if core.is_compile_gpu():
+            core.init_devices(["CPU", "GPU:0"])
+        else:
+            core.init_devices(["CPU"])
+
         self.executor = core.Executor(act_places)
         self.places = places