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

mat proc kernel std tamplate is ok, merging old decl

Browse Source
This commit is contained in:
Nitload
2026-03-02 00:24:15 +08:00
parent e198f22712
commit 67fbbde449
8 changed files with 279 additions and 290 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/m.basic.x86.c
+5 -52
View File
@@ -3,59 +3,12 @@
#if SN_IS_ARCH(SN_ARCH_X86_64)
#include "sn_intrinsic.h"
#include "sn_intrincontext.x86.h"
#include "sn_intrinsic.context.x86.h"
#include "sn_intrinsic.std.h"
STD_MK_PROC_DECL(float, sse, , clear)
{
register __m128 zero = _mm_setzero_ps();
for (size_t i = 0; i < 4; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
STD_MK_PROC_DECL(float, avx, , clear)
{
register __m256 zero = _mm256_setzero_ps();
for (size_t i = 0; i < 8; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
STD_MK_PROC_DECL(float, 512, , clear)
{
register __m512 zero = _mm512_setzero_ps();
for (size_t i = 0; i < 16; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
STD_MK_PROC_DECL(double, sse, , clear)
{
register __m128 zero = _mm_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;
}
STD_MK_PROC_DECL(double, avx, , clear)
{
register __m256 zero = _mm256_setzero_ps();
for (size_t i = 0; i < 8; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
STD_MK_PROC_DECL(double, 512, , clear)
{
register __m512 zero = _mm512_setzero_ps();
for (size_t i = 0; i < 16; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
// Matrix transpose implementations
STD_MK_PROC_DECL(float, sse, , transpose)
SN_MK_PROC_DECL_STD(float, sse, , transpose)
{
register __m128 col0 = right->f32_x[0];
register __m128 col1 = right->f32_x[1];
@@ -72,7 +25,7 @@ STD_MK_PROC_DECL(float, sse, , transpose)
ans->f32_x[3] = _mm_movehl_ps(tmp3, tmp1);
}
STD_MK_PROC_DECL(float, avx, , transpose)
SN_MK_PROC_DECL_STD(float, avx, , transpose)
{
register __m256 col_low = right->f32_y[0];
register __m256 col_high = right->f32_y[1];
@@ -87,7 +40,7 @@ STD_MK_PROC_DECL(float, avx, , transpose)
ans->f32_y[1] = _mm256_shuffle_ps(col_high, col_low, _MM_SHUFFLE(3, 1, 3, 1));
}
STD_MK_PROC_DECL(float, 512, , transpose)
SN_MK_PROC_DECL_STD(float, 512, , transpose)
{
register __m512 matrix = right->f32_z[0];