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

Gemm/Fma kernel on avx512 use new solution, with less permute and shuff instruction

Browse Source
This commit is contained in:
Youmu Liu
2026-06-04 16:43:17 +08:00
parent 6c38bcfade
commit 2690cf9341
3 changed files with 44 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-template/sn_template.fma.x86.h
+16 -18
View File
@@ -2,25 +2,23 @@
#if SN_IS_ARCH(SN_ARCH_X86_64)
/*
* requested ZMMs:
*
*/
#define SN_MK_FMA_F32_512_CYCLE(i)\
do{\
\
}while(0)
#define SN_MK_FMA_F32_512_CYCLE(i) \
do { \
c0 = _mm512_fmadd_ps(c0, a0, _mm512_permute_ps(b0, _FMA_SELECT(i))); \
c1 = _mm512_fmadd_ps(c1, a1, _mm512_permute_ps(b1, _FMA_SELECT(i))); \
c2 = _mm512_fmadd_ps(c2, a2, _mm512_permute_ps(b2, _FMA_SELECT(i))); \
c3 = _mm512_fmadd_ps(c3, a3, _mm512_permute_ps(b3, _FMA_SELECT(i))); \
} while (0)
#define SN_MK_FMA_F32_512_PHASE(i) \
do { \
ak = _mm512_broadcast_f32x4(left_##i->x[0]); \
c_0 = _mm512_fmadd_ps(ak, _mm512_permute_ps(b_##i, _FMA_SELECT(0)), c_##i); \
ak = _mm512_broadcast_f32x4(left_##i->x[0]); \
c_1 = _mm512_fmadd_ps(ak, _mm512_permute_ps(b_##i, _FMA_SELECT(0)), c_##i); \
ak = _mm512_broadcast_f32x4(left_##i->x[0]); \
c_2 = _mm512_fmadd_ps(ak, _mm512_permute_ps(b_##i, _FMA_SELECT(0)), c_##i); \
ak = _mm512_broadcast_f32x4(left_##i->x[0]); \
c_3 = _mm512_fmadd_ps(ak, _mm512_permute_ps(b_##i, _FMA_SELECT(0)), c_##i); \
#define SN_FMA_PACK_COLUMN(col, target) \
do { \
register __m256 low = _mm512_castps512_ps256(col); \
register __m256 high = _mm512_extractf32x8_ps(col, 1); \
register __m256 half = _mm256_add_ps(high, low); \
register __m128 high128 = _mm256_extractf32x4_ps(half, 1); \
register __m128 low128 = _mm256_castps256_ps128(half); \
register __m128 half128 = _mm_add_ps(high128, low128); \
_mm_store_ps(target, half128); \
} while (0)
// double