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

matrix mul matrix is full function

Browse Source
This commit is contained in:
Nitload
2025-12-16 23:17:30 +08:00
parent 6344242b3c 88ed5a95c5
commit b13adc1a1b
26 changed files with 1049 additions and 861 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/k_basicop.c
+17
View File
@@ -1,6 +1,9 @@
#include "sn_intrinsic.h"
<<<<<<< HEAD
<<<<<<< HEAD
=======
>>>>>>> stgsharp-dev/giga
DECLARE_PNL_PROC_ANS(float, sse, clear)
{
register __m128 zero = _mm_setzero_ps();
@@ -13,15 +16,22 @@ DECLARE_PNL_PROC_ANS(float, sse, clear)
DECLARE_PNL_PROC_ANS(float, avx, clear)
{
register __m256 zero = _mm256_setzero_ps();
<<<<<<< HEAD
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;
}
>>>>>>> stgsharp-dev/giga
}
DECLARE_PNL_PROC_ANS(float, 512, clear)
{
register __m512 zero = _mm512_setzero_ps();
<<<<<<< HEAD
PANEL_COL_VEC((*ans), 0, 0) = zero;
PANEL_COL_VEC((*ans), 1, 0) = zero;
PANEL_COL_VEC((*ans), 2, 0) = zero;
@@ -30,6 +40,13 @@ DECLARE_PNL_PROC_ANS(float, 512, clear)
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
for (size_t i = 0; i < 16; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
>>>>>>> stgsharp-dev/giga
// Matrix transpose implementations
DECLARE_KER_PROC_RIGHT_ANS(float, sse, transpose)
{