youmuliu/stgsharp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

rewrite of f32 fma, and other optmization of simd kernel ops

Browse Source
This commit is contained in:
Youmu Liu
2025-11-10 20:40:19 +08:00
parent 1dee467220
commit 678a846e54
18 changed files with 302 additions and 156 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/k_fma.c
+39 -25
View File
@@ -1,6 +1,7 @@
#include "sn_intrinsic.h"
// ans = left * right + ans
// right is not transposed
INTERNAL void SN_DECL f32_fma_sse(MAT_KERNEL(float) const *restrict left,
MAT_KERNEL(float) const *restrict right,
MAT_KERNEL(float) *restrict ans)
@@ -10,34 +11,47 @@ INTERNAL void SN_DECL f32_fma_sse(MAT_KERNEL(float) const *restrict left,
register __m128 source2 = _mm_load_ps(&left->xmm[2]);
register __m128 source3 = _mm_load_ps(&left->xmm[3]);
register __m128 ans0 = _mm_load_ps(&ans->xmm[0]);
register __m128 ans1 = _mm_load_ps(&ans->xmm[1]);
register __m128 ans2 = _mm_load_ps(&ans->xmm[2]);
register __m128 ans3 = _mm_load_ps(&ans->xmm[3]);
register __m128 broadcast;
register __m128 ans0 = _mm_setzero_ps();
register __m128 ans1 = _mm_setzero_ps();
register __m128 ans2 = _mm_setzero_ps();
register __m128 ans3 = _mm_setzero_ps();
for (size_t i = 0; i < 3; i++) {
for (size_t i = 0; i < 4; i++) {
broadcast = _mm_load_ps(&right->xmm[i]);
_mm_add_ps(ans0, _mm_mul_ps(source0, broadcast));
_mm_add_ps(ans1, _mm_mul_ps(source1, broadcast));
_mm_add_ps(ans2, _mm_mul_ps(source2, broadcast));
_mm_add_ps(ans3, _mm_mul_ps(source3, broadcast));
for (size_t t = 0; t < 2; t++) {
broadcast = _mm_shuffle_ps(broadcast, broadcast, _MM_SHUFFLE(2, 1, 0, 3));
_mm_add_ps(ans0, _mm_mul_ps(source0, broadcast));
_mm_add_ps(ans1, _mm_mul_ps(source1, broadcast));
_mm_add_ps(ans2, _mm_mul_ps(source2, broadcast));
_mm_add_ps(ans3, _mm_mul_ps(source3, broadcast));
}
ans0 = _mm_add_ps(ans0,
_mm_mul_ps(source0, _mm_shuffle_ps(broadcast, broadcast, 0x00)));
ans1 = _mm_add_ps(ans1,
_mm_mul_ps(source1, _mm_shuffle_ps(broadcast, broadcast, 0x55)));
ans2 = _mm_add_ps(ans2,
_mm_mul_ps(source2, _mm_shuffle_ps(broadcast, broadcast, 0xAA)));
ans3 = _mm_add_ps(ans3,
_mm_mul_ps(source3, _mm_shuffle_ps(broadcast, broadcast, 0xFF)));
}
source0 = _mm_load_ps(&ans->xmm[0]);
source1 = _mm_load_ps(&ans->xmm[1]);
source2 = _mm_load_ps(&ans->xmm[2]);
source3 = _mm_load_ps(&ans->xmm[3]);
_mm_store_ps(&ans->xmm[0], _mm_add_ps(source0, ans0));
_mm_store_ps(&ans->xmm[1], _mm_add_ps(source1, ans1));
_mm_store_ps(&ans->xmm[2], _mm_add_ps(source2, ans2));
_mm_store_ps(&ans->xmm[3], _mm_add_ps(source3, ans3));
_mm_store_ps(&ans->xmm[0], ans0);
_mm_store_ps(&ans->xmm[1], ans1);
_mm_store_ps(&ans->xmm[2], ans2);
_mm_store_ps(&ans->xmm[3], ans3);
}
// ans = left * right + ans
INTERNAL void SN_DECL f32_fma_avx(MAT_KERNEL(float) const *restrict left,
MAT_KERNEL(float) const *restrict right,
MAT_KERNEL(float) *restrict ans)
{
/*
* to be implemented
*/
}
// ans = left * right + ans
INTERNAL void SN_DECL f32_fma_512(MAT_KERNEL(float) const *restrict left,
MAT_KERNEL(float) const *restrict right,
MAT_KERNEL(float) *restrict ans)
{
/*
* to be implemented
*/
}