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

Gemm/Fma kernel on avx2 use new solution, with less permute and shuff instruction

Browse Source
This commit is contained in:
Youmu Liu
2026-06-04 21:05:38 +08:00
parent 2690cf9341
commit 94b2d759ae
2 changed files with 46 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-template/sn_template.fma.x86.h
+17 -1
View File
@@ -10,7 +10,15 @@
c3 = _mm512_fmadd_ps(c3, a3, _mm512_permute_ps(b3, _FMA_SELECT(i))); \
} while (0)
#define SN_FMA_PACK_COLUMN(col, target) \
#define SN_MK_FMA_F32_AVX2_CYCLE(i) \
do { \
c0 = _mm256_fmadd_ps(c0, a0, _mm256_permute_ps(b0, _FMA_SELECT(i))); \
c1 = _mm256_fmadd_ps(c1, a1, _mm256_permute_ps(b1, _FMA_SELECT(i))); \
c2 = _mm256_fmadd_ps(c2, a2, _mm256_permute_ps(b2, _FMA_SELECT(i))); \
c3 = _mm256_fmadd_ps(c3, a3, _mm256_permute_ps(b3, _FMA_SELECT(i))); \
} while (0)
#define SN_FMA_512_PACK_COLUMN(col, target) \
do { \
register __m256 low = _mm512_castps512_ps256(col); \
register __m256 high = _mm512_extractf32x8_ps(col, 1); \
@@ -21,6 +29,14 @@
_mm_store_ps(target, half128); \
} while (0)
#define SN_FMA_AVX2_PACK_COLUMN(col, target) \
do { \
register __m128 high128 = _mm256_extractf32x4_ps(col, 1); \
register __m128 low128 = _mm256_castps256_ps128(col); \
register __m128 half128 = _mm_add_ps(high128, low128); \
_mm_store_ps(target, half128); \
} while (0)
// double
#define SN_MK_FMA_F64_SSE_PHASE(i) \
do { \