You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
106 lines
1.7 KiB
106 lines
1.7 KiB
|
|
#ifdef __aarch64__
|
|
.text
|
|
.align 5
|
|
//.p2align 5,,15
|
|
.global MatrixSub
|
|
#ifndef __APPLE__
|
|
.type MatrixSub, %function
|
|
#endif
|
|
|
|
|
|
|
|
//void MatrixSub(const float* matDataA, const float* matDataB, float* matDataC,
|
|
// size_t aStride, size_t bStride, size_t cStride, size_t width, size_t height)
|
|
|
|
//Auto: x0: matDataA, x1:matDataB, x2:matDatac,
|
|
//x3:aStride, x4:bStride, x5:cStride, x6:width, x7:height
|
|
|
|
MatrixSub:
|
|
mov x12, #4 //sizeof(float)
|
|
mul x3, x12, x3
|
|
mul x4, x12, x4
|
|
mul x5, x12, x5
|
|
|
|
loopH:
|
|
mov x8, x0
|
|
mov x9, x1
|
|
mov x10, x2
|
|
|
|
mov x11, x6
|
|
|
|
loop16LineIn:
|
|
cmp x11, #4
|
|
blt L8
|
|
sub x11, x11, #4
|
|
ld1 {v0.4s, v1.4s}, [x0], #32
|
|
ld1 {v2.4s, v3.4s}, [x1], #32
|
|
|
|
fsub v4.4s, v0.4s, v2.4s
|
|
fsub v5.4s, v1.4s, v3.4s
|
|
|
|
ld1 {v6.4s, v7.4s}, [x0], #32
|
|
ld1 {v8.4s, v9.4s}, [x1], #32
|
|
|
|
cmp x11, #4
|
|
blt loop16LineOut
|
|
|
|
loop16:
|
|
st1 {v4.4s, v5.4s}, [x2], #32
|
|
fsub v10.4s, v6.4s, v8.4s
|
|
fsub v11.4s, v7.4s, v9.4s
|
|
ld1 {v0.4s, v1.4s}, [x0], #32
|
|
ld1 {v2.4s, v3.4s}, [x1], #32
|
|
|
|
st1 {v10.4s, v11.4s}, [x2], #32
|
|
fsub v4.4s, v0.4s, v2.4s
|
|
fsub v5.4s, v1.4s, v3.4s
|
|
ld1 {v6.4s, v7.4s}, [x0], #32
|
|
ld1 {v8.4s, v9.4s}, [x1], #32
|
|
|
|
sub x11, x11, #4
|
|
cmp x11, #4
|
|
bge loop16
|
|
|
|
loop16LineOut:
|
|
st1 {v4.4s, v5.4s}, [x2], #32
|
|
fsub v10.4s, v6.4s, v8.4s
|
|
fsub v11.4s, v7.4s, v9.4s
|
|
st1 {v10.4s, v11.4s}, [x2], #32
|
|
|
|
L8:
|
|
cmp x11, #2
|
|
blt L4
|
|
|
|
ld1 {v0.4s, v1.4s}, [x0], #32
|
|
ld1 {v2.4s, v3.4s}, [x1], #32
|
|
|
|
fsub v4.4s, v0.4s, v2.4s
|
|
fsub v5.4s, v1.4s, v3.4s
|
|
|
|
sub x11, x11, #2
|
|
st1 {v4.4s, v5.4s}, [x2], #32
|
|
|
|
|
|
cmp x11, #0
|
|
beq loop16EndLine
|
|
|
|
L4:
|
|
ld1 {v0.4s}, [x0], #16
|
|
ld1 {v1.4s}, [x1], #16
|
|
fsub v0.4s, v0.4s, v1.4s
|
|
sub x11, x11, #1
|
|
st1 {v0.4s}, [x2], #16
|
|
|
|
|
|
loop16EndLine:
|
|
add x0, x8, x3
|
|
add x1, x9, x4
|
|
add x2, x10, x5
|
|
|
|
subs x7, x7, #1
|
|
bne loopH
|
|
|
|
ret
|
|
#endif
|