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

use individual panel define as compute unit, kernel is now storage and loading unit

Browse Source
This commit is contained in:
Youmu Liu
2025-11-12 21:13:00 +08:00
parent bbe69f0384
commit 2dfb4ae86a
33 changed files with 522 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/k_fma.c
+30 -3
View File
@@ -1,5 +1,9 @@
#include "sn_intrinsic.h"
INTERNAL void SN_DECL f32_fma_sse(MAT_PANEL_SSE(int) * left)
{
}
// ans = left * right + ans
// right is not transposed
INTERNAL void SN_DECL f32_fma_sse(MAT_KERNEL(float) const *restrict left,
@@ -41,9 +45,32 @@ 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
*/
register __m256 source0 = _mm256_load_ps(&left->ymm[0]);
register __m256 source1 = _mm256_load_ps(&left->ymm[1]);
register __m256 ans0 = _mm256_load_ps(&ans->ymm[0]);
register __m256 ans1 = _mm256_load_ps(&ans->ymm[1]);
register __m128 bs0, bs1;
register __m256 broadcast;
for (size_t i = 0; i < 4; i++) {
bs0 = _mm_load_ps(&right->xmm[i]);
__m128 lo = _mm_shuffle_ps(bs0, bs0, 0x00);
__m128 hi = _mm_shuffle_ps(bs0, bs0, 0x55);
broadcast = _mm256_castps128_ps256(lo);
broadcast = _mm256_insertf128_ps(broadcast, bs0, 1);
ans0 = _mm256_add_ps(ans0, _mm256_mul_ps(source0, broadcast));
lo = _mm_shuffle_ps(bs0, bs0, 0xAA);
hi = _mm_shuffle_ps(bs0, bs0, 0xFF);
broadcast = _mm256_castps128_ps256(lo);
broadcast = _mm256_insertf128_ps(broadcast, bs0, 1);
ans1 = _mm256_add_ps(ans1, _mm256_mul_ps(source1, broadcast));
}
_mm256_store_ps(&ans->ymm[0], ans0);
_mm256_store_ps(&ans->ymm[1], ans1);
}
// ans = left * right + ans