m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add gru refer code and remove redundant avx code

Browse Source 
test=develop
local_add_cudnn_lstm
tensor-tang 6 years ago
parent f913860873
commit 3562051302
6 changed files with 163 additions and 428 deletions
Show Stats Download Patch File Download Diff File

67
paddle/fluid/operators/fused/fusion_gru_op.cc
Unescape View File

@ -183,24 +183,27 @@ class FusionGRUKernel : public framework::OpKernel<T> {
const int total_T = x_dims[0]; \
const int D3 = wh_dims[1]
#define INIT_OTHER_DEFINES \
auto* h0 = ctx.Input<Tensor>("H0"); \
auto* wx = ctx.Input<Tensor>("WeightX"); \
auto* bias = ctx.Input<Tensor>("Bias"); \
auto* hidden_out = ctx.Output<LoDTensor>("Hidden"); \
bool is_reverse = ctx.Attr<bool>("is_reverse"); \
const int M = x_dims[1]; \
const int D = wh_dims[0]; \
const int D2 = D * 2; \
const auto& ker = math::jitkernel::KernelPool::Instance() \
.template Get<math::jitkernel::GRUKernel<T>, \
const std::string&, const std::string&>( \
ctx.Attr<std::string>("gate_activation"), \
ctx.Attr<std::string>("activation"), D); \
const T* x_data = x->data<T>(); \
const T* wx_data = wx->data<T>(); \
const T* wh_data = wh->data<T>(); \
auto place = ctx.GetPlace(); \
#define INIT_OTHER_DEFINES \
auto* h0 = ctx.Input<Tensor>("H0"); \
auto* wx = ctx.Input<Tensor>("WeightX"); \
auto* bias = ctx.Input<Tensor>("Bias"); \
auto* hidden_out = ctx.Output<LoDTensor>("Hidden"); \
bool is_reverse = ctx.Attr<bool>("is_reverse"); \
const int M = x_dims[1]; \
const int D = wh_dims[0]; \
const int D2 = D * 2; \
const math::jitkernel::gru_attr_t attr( \
D, ctx.Attr<std::string>("gate_activation"), \
ctx.Attr<std::string>("activation")); \
math::jitkernel::gru_t one_step; \
const auto& ker = \
math::jitkernel::KernelPool::Instance() \
.template Get<math::jitkernel::GRUKernel<T>, \
const math::jitkernel::gru_attr_t&>(attr); \
const T* x_data = x->data<T>(); \
const T* wx_data = wx->data<T>(); \
const T* wh_data = wh->data<T>(); \
auto place = ctx.GetPlace(); \
T* xx_data = xx->mutable_data<T>(place)
void SeqCompute(const framework::ExecutionContext& ctx) const {
@ -237,7 +240,9 @@ class FusionGRUKernel : public framework::OpKernel<T> {
if (h0_data) {
prev_hidden_data = h0_data + bid * D;
} else {
ker->ComputeH1(xx_data, hidden_out_data);
one_step.gates = xx_data;
one_step.ht = hidden_out_data;
ker->ComputeH1(&one_step, &attr);
prev_hidden_data = hidden_out_data;
tstart = 1;
move_step();
@ -247,12 +252,15 @@ class FusionGRUKernel : public framework::OpKernel<T> {
blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D2, D, static_cast<T>(1),
prev_hidden_data, D, wh_data, D2, static_cast<T>(1), xx_data,
D3);
ker->ComputeHtPart1(xx_data, prev_hidden_data, hidden_out_data);
one_step.gates = xx_data;
one_step.ht_1 = prev_hidden_data;
one_step.ht = hidden_out_data;
ker->ComputeHtPart1(&one_step, &attr);
// gemm rt * Ws
blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D, D, static_cast<T>(1),
hidden_out_data, D, wh_state_data, D, static_cast<T>(1),
xx_data + D2, D3);
ker->ComputeHtPart2(xx_data, prev_hidden_data, hidden_out_data);
ker->ComputeHtPart2(&one_step, &attr);
// save prev
prev_hidden_data = hidden_out_data;
move_step();
@ -314,7 +322,9 @@ class FusionGRUKernel : public framework::OpKernel<T> {
T* cur_out_data = batched_out_data;
// W: {W_update, W_reset; W_state}
for (int i = 0; i < max_bs; ++i) {
ker->ComputeH1(cur_in_data, cur_out_data);
one_step.gates = cur_in_data;
one_step.ht = cur_out_data;
ker->ComputeH1(&one_step, &attr);
// add offset
cur_in_data += D3;
cur_out_data += D;
@ -339,8 +349,11 @@ class FusionGRUKernel : public framework::OpKernel<T> {
T* cur_out_data = batched_out_data;
T* cur_prev_hidden_data = prev_hidden_data;
for (int i = 0; i < cur_bs; ++i) {
ker->ComputeHtPart1(cur_batched_data, cur_prev_hidden_data,
cur_out_data);
one_step.gates = cur_batched_data;
one_step.ht_1 = cur_prev_hidden_data;
one_step.ht = cur_out_data;
ker->ComputeHtPart1(&one_step, &attr);
cur_batched_data += D3;
cur_prev_hidden_data += D;
cur_out_data += D;
@ -354,8 +367,10 @@ class FusionGRUKernel : public framework::OpKernel<T> {
cur_prev_hidden_data = prev_hidden_data;
for (int i = 0; i < cur_bs; ++i) {
ker->ComputeHtPart2(cur_batched_data, cur_prev_hidden_data,
cur_out_data);
one_step.gates = cur_batched_data;
one_step.ht_1 = cur_prev_hidden_data;
one_step.ht = cur_out_data;
ker->ComputeHtPart2(&one_step, &attr);
cur_batched_data += D3;
cur_prev_hidden_data += D;
cur_out_data += D;

8
paddle/fluid/operators/math/jit_kernel.h
Unescape View File

@ -122,18 +122,18 @@ class VTanhKernel : public VActKernel<T> {};
template <typename T>
class LSTMKernel : public Kernel {
public:
void (*ComputeCtHt)(lstm_t *, const lstm_attr_t *);
// compute c1 and h1 without c0 or h0
void (*ComputeC1H1)(lstm_t *, const lstm_attr_t *);
void (*ComputeCtHt)(lstm_t *, const lstm_attr_t *);
};
template <typename T>
class GRUKernel : public Kernel {
public:
// compute h1 without h0
virtual void ComputeH1(T *gates, T *ht) const = 0;
virtual void ComputeHtPart1(T *gates, const T *ht_1, T *ht) const = 0;
virtual void ComputeHtPart2(T *gates, const T *ht_1, T *ht) const = 0;
void (*ComputeH1)(gru_t *, const gru_attr_t *);
void (*ComputeHtPart1)(gru_t *, const gru_attr_t *);
void (*ComputeHtPart2)(gru_t *, const gru_attr_t *);
};
template <typename T>

152
paddle/fluid/operators/math/jit_kernel_exp.cc
Unescape View File

@ -25,10 +25,6 @@ limitations under the License. */
#include "paddle/fluid/platform/dynload/mklml.h"
#endif
#ifdef __AVX__
#include <immintrin.h>
#endif
namespace paddle {
namespace operators {
namespace math {
@ -235,154 +231,6 @@ REGISTER_JITKERNEL(vexp, VExpKernel);
REGISTER_JITKERNEL(vsigmoid, VSigmoidKernel);
REGISTER_JITKERNEL(vtanh, VTanhKernel);
namespace detail {
#ifdef __AVX__
#define ALIGN32 __attribute__((aligned(32)))
#define _PS256_CONST(Name, Val) \
static const float _ps256_##Name[8] ALIGN32 = {Val, Val, Val, Val, \
Val, Val, Val, Val}
#define _PI256_CONST(Name, Val) \
static const int _pi256_##Name[8] ALIGN32 = {Val, Val, Val, Val, \
Val, Val, Val, Val}
_PI256_CONST(0x7f, 0x7f);
_PS256_CONST(one, 1.f);
_PS256_CONST(0p5, 0.5f);
_PS256_CONST(exp_hi, 88.3762626647949f);
_PS256_CONST(exp_lo, -88.3762626647949f);
_PS256_CONST(cephes_LOG2EF, 1.44269504088896341);
_PS256_CONST(cephes_exp_C1, 0.693359375);
_PS256_CONST(cephes_exp_C2, -2.12194440e-4);
_PS256_CONST(cephes_exp_p0, 1.9875691500E-4);
_PS256_CONST(cephes_exp_p1, 1.3981999507E-3);
_PS256_CONST(cephes_exp_p2, 8.3334519073E-3);
_PS256_CONST(cephes_exp_p3, 4.1665795894E-2);
_PS256_CONST(cephes_exp_p4, 1.6666665459E-1);
_PS256_CONST(cephes_exp_p5, 5.0000001201E-1);
typedef union imm_xmm_union {
__m256i imm;
__m128i xmm[2];
} imm_xmm_union;
#define COPY_IMM_TO_XMM(imm_, xmm0_, xmm1_) \
{ \
imm_xmm_union u ALIGN32; \
u.imm = imm_; \
xmm0_ = u.xmm[0]; \
xmm1_ = u.xmm[1]; \
}
#define COPY_XMM_TO_IMM(xmm0_, xmm1_, imm_) \
{ \
imm_xmm_union u ALIGN32; \
u.xmm[0] = xmm0_; \
u.xmm[1] = xmm1_; \
imm_ = u.imm; \
}
#define AVX2_BITOP_USING_SSE2(fn) \
static inline __m256i avx2_mm256_##fn(__m256i x, int y) { \
/* use SSE2 to perform the bitop AVX2 */ \
__m128i x1, x2; \
__m256i ret; \
COPY_IMM_TO_XMM(x, x1, x2); \
x1 = _mm_##fn(x1, y); \
x2 = _mm_##fn(x2, y); \
COPY_XMM_TO_IMM(x1, x2, ret); \
return ret; \
}
#define AVX2_INTOP_USING_SSE2(fn) \
static inline __m256i avx2_mm256_add_epi32(__m256i x, __m256i y) { \
/* use SSE2 to perform the AVX2 integer operation */ \
__m128i x1, x2; \
__m128i y1, y2; \
__m256i ret; \
COPY_IMM_TO_XMM(x, x1, x2); \
COPY_IMM_TO_XMM(y, y1, y2); \
x1 = _mm_##fn(x1, y1); \
x2 = _mm_##fn(x2, y2); \
COPY_XMM_TO_IMM(x1, x2, ret); \
return ret; \
}
AVX2_BITOP_USING_SSE2(slli_epi32);
AVX2_INTOP_USING_SSE2(add_epi32);
#define AVXEXP_BASE \
__m256 tmp = _mm256_setzero_ps(), fx; \
__m256 one = *reinterpret_cast<const __m256*>(_ps256_one); \
__m256i imm0; \
x = _mm256_min_ps(x, *reinterpret_cast<const __m256*>(_ps256_exp_hi)); \
x = _mm256_max_ps(x, *reinterpret_cast<const __m256*>(_ps256_exp_lo)); \
/* express exp(x) as exp(g + n*log(2)) */ \
fx = _mm256_mul_ps(x, \
*reinterpret_cast<const __m256*>(_ps256_cephes_LOG2EF)); \
fx = _mm256_add_ps(fx, *reinterpret_cast<const __m256*>(_ps256_0p5)); \
tmp = _mm256_floor_ps(fx); \
/* if greater, substract 1 */ \
__m256 mask = _mm256_cmp_ps(tmp, fx, _CMP_GT_OS); \
mask = _mm256_and_ps(mask, one); \
fx = _mm256_sub_ps(tmp, mask); \
tmp = _mm256_mul_ps(fx, \
*reinterpret_cast<const __m256*>(_ps256_cephes_exp_C1)); \
__m256 z = _mm256_mul_ps( \
fx, *reinterpret_cast<const __m256*>(_ps256_cephes_exp_C2)); \
x = _mm256_sub_ps(x, tmp); \
x = _mm256_sub_ps(x, z); \
z = _mm256_mul_ps(x, x); \
__m256 y = *reinterpret_cast<const __m256*>(_ps256_cephes_exp_p0); \
y = _mm256_mul_ps(y, x); \
y = _mm256_add_ps(y, \
*reinterpret_cast<const __m256*>(_ps256_cephes_exp_p1)); \
y = _mm256_mul_ps(y, x); \
y = _mm256_add_ps(y, \
*reinterpret_cast<const __m256*>(_ps256_cephes_exp_p2)); \
y = _mm256_mul_ps(y, x); \
y = _mm256_add_ps(y, \
*reinterpret_cast<const __m256*>(_ps256_cephes_exp_p3)); \
y = _mm256_mul_ps(y, x); \
y = _mm256_add_ps(y, \
*reinterpret_cast<const __m256*>(_ps256_cephes_exp_p4)); \
y = _mm256_mul_ps(y, x); \
y = _mm256_add_ps(y, \
*reinterpret_cast<const __m256*>(_ps256_cephes_exp_p5)); \
y = _mm256_mul_ps(y, z); \
y = _mm256_add_ps(y, x); \
y = _mm256_add_ps(y, one); \
/* build 2^n */ \
imm0 = _mm256_cvttps_epi32(fx)
__m256 ExpAVX(__m256 x) {
AVXEXP_BASE;
// two AVX2 instructions using SSE2
imm0 = avx2_mm256_add_epi32(imm0,
*reinterpret_cast<const __m256i*>(_pi256_0x7f));
imm0 = avx2_mm256_slli_epi32(imm0, 23);
__m256 pow2n = _mm256_castsi256_ps(imm0);
y = _mm256_mul_ps(y, pow2n);
return y;
}
#endif
#ifdef __AVX2__
__m256 ExpAVX2(__m256 x) {
AVXEXP_BASE;
// two AVX2 instructions
imm0 = _mm256_add_epi32(imm0, *reinterpret_cast<const __m256i*>(_pi256_0x7f));
imm0 = _mm256_slli_epi32(imm0, 23);
__m256 pow2n = _mm256_castsi256_ps(imm0);
y = _mm256_mul_ps(y, pow2n);
return y;
}
#endif
} // namespace detail
} // namespace jitkernel
} // namespace math
} // namespace operators

30
paddle/fluid/operators/math/jit_kernel_impl.h
Unescape View File

@ -38,20 +38,34 @@ typedef struct {
void* checked{nullptr};
} lstm_t;
typedef struct lstm_attr_s {
bool use_peephole;
typedef struct {
void* gates; // gates: {W_update, W_reset; W_state}
const void* ht_1;
void* ht;
} gru_t;
struct rnn_attr_s {
int d;
std::string act_gate, act_cand, act_cell;
std::string act_gate, act_cand;
rnn_attr_s() = default;
rnn_attr_s(int _d, const std::string& _act_gate, const std::string& _act_cand)
: d(_d), act_gate(_act_gate), act_cand(_act_cand) {}
};
struct lstm_attr_s : public rnn_attr_s {
bool use_peephole;
std::string act_cell;
lstm_attr_s() = default;
lstm_attr_s(int _d, const std::string& _act_gate,
const std::string& _act_cand, const std::string& _act_cell,
bool _use_peephole = false)
: use_peephole(_use_peephole),
d(_d),
act_gate(_act_gate),
act_cand(_act_cand),
: rnn_attr_s(_d, _act_gate, _act_cand),
use_peephole(_use_peephole),
act_cell(_act_cell) {}
} lstm_attr_t;
};
typedef struct rnn_attr_s gru_attr_t;
typedef struct lstm_attr_s lstm_attr_t;
} // namespace jitkernel
} // namespace math

40
paddle/fluid/operators/math/jit_kernel_refer.h
Unescape View File

@ -185,6 +185,46 @@ void LSTMC1H1(lstm_t* step, const lstm_attr_t* attr) {
VMul(gates + d2, gates + d3, ht, d);
}
// compute h1 without h0
template <typename T>
void GRUH1(gru_t* step, const gru_attr_t* attr) {
T* gates = reinterpret_cast<T*>(step->gates);
T* ht = reinterpret_cast<T*>(step->ht);
auto act_gate = getActFunc<T>(attr->act_gate);
auto act_cand = getActFunc<T>(attr->act_cand);
int d = attr->d;
int d2 = d * 2;
act_gate(gates, gates, d);
act_cand(gates + d2, gates + d2, d);
VMul(gates, gates + d2, ht, d);
}
template <typename T>
void GRUHtPart1(gru_t* step, const gru_attr_t* attr) {
// W: {W_update, W_reset; W_state}
T* gates = reinterpret_cast<T*>(step->gates);
T* ht = reinterpret_cast<T*>(step->ht);
const T* ht_1 = reinterpret_cast<const T*>(step->ht_1);
auto act_gate = getActFunc<T>(attr->act_gate);
act_gate(gates, gates, attr->d * 2);
VMul(ht_1, gates + attr->d, ht, attr->d);
}
template <typename T>
void GRUHtPart2(gru_t* step, const gru_attr_t* attr) {
T* gates = reinterpret_cast<T*>(step->gates);
T* ht = reinterpret_cast<T*>(step->ht);
const T* ht_1 = reinterpret_cast<const T*>(step->ht_1);
auto act_cand = getActFunc<T>(attr->act_cand);
int d = attr->d;
T* y = gates + d * 2;
act_cand(y, y, d);
// out = zt*ht~ + (1-zt)*ht_1
for (int i = 0; i < d; ++i) {
ht[i] = gates[i] * y[i] + (static_cast<T>(1) - gates[i]) * ht_1[i];
}
}
} // namespace refer
} // namespace jitkernel
} // namespace math

294
paddle/fluid/operators/math/jit_kernel_rnn.cc
View File

File diff suppressed because it is too large Load Diff
Write Preview
Loading…
Cancel
Save