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

Mat mul mat is now on full function

Browse Source
This commit is contained in:
Youmu Liu
2025-12-16 16:15:30 +08:00
parent 2c7277f3f9
commit 88ed5a95c5
14 changed files with 357 additions and 488 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/k_basicop.c
+6 -8
View File
@@ -12,19 +12,17 @@ DECLARE_PNL_PROC_ANS(float, sse, clear)
DECLARE_PNL_PROC_ANS(float, avx, clear)
{
register __m256 zero = _mm256_setzero_ps();
PANEL_COL_VEC((*ans), 0, 0) = zero;
PANEL_COL_VEC((*ans), 1, 0) = zero;
PANEL_COL_VEC((*ans), 2, 0) = zero;
PANEL_COL_VEC((*ans), 3, 0) = zero;
for (size_t i = 0; i < 8; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
DECLARE_PNL_PROC_ANS(float, 512, clear)
{
register __m512 zero = _mm512_setzero_ps();
PANEL_COL_VEC((*ans), 0, 0) = zero;
PANEL_COL_VEC((*ans), 1, 0) = zero;
PANEL_COL_VEC((*ans), 2, 0) = zero;
PANEL_COL_VEC((*ans), 3, 0) = zero;
for (size_t i = 0; i < 16; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
// Matrix transpose implementations