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

new mat proc std

Browse Source
This commit is contained in:
Youmu Liu
2026-02-28 18:08:26 +08:00
parent 55cf3934bb
commit 6e07e67835
153 changed files with 91360 additions and 4547 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/m.basic.x86.c
+9 -9
View File
@@ -5,7 +5,7 @@
#include "sn_intrinsic.h"
#include "sn_intrincontext.x86.h"
DECLARE_PNL_PROC_ANS(float, sse, , clear)
STD_MK_PROC_DECL(float, sse, , clear)
{
register __m128 zero = _mm_setzero_ps();
for (size_t i = 0; i < 4; i++) {
@@ -13,7 +13,7 @@ DECLARE_PNL_PROC_ANS(float, sse, , clear)
}
}
DECLARE_PNL_PROC_ANS(float, avx, , clear)
STD_MK_PROC_DECL(float, avx, , clear)
{
register __m256 zero = _mm256_setzero_ps();
for (size_t i = 0; i < 8; i++) {
@@ -21,7 +21,7 @@ DECLARE_PNL_PROC_ANS(float, avx, , clear)
}
}
DECLARE_PNL_PROC_ANS(float, 512, , clear)
STD_MK_PROC_DECL(float, 512, , clear)
{
register __m512 zero = _mm512_setzero_ps();
for (size_t i = 0; i < 16; i++) {
@@ -29,7 +29,7 @@ DECLARE_PNL_PROC_ANS(float, 512, , clear)
}
}
DECLARE_PNL_PROC_ANS(double, sse, , clear)
STD_MK_PROC_DECL(double, sse, , clear)
{
register __m128 zero = _mm_setzero_ps();
PANEL_COL_VEC((*ans), 0, 0) = zero;
@@ -38,7 +38,7 @@ DECLARE_PNL_PROC_ANS(double, sse, , clear)
PANEL_COL_VEC((*ans), 3, 0) = zero;
}
DECLARE_PNL_PROC_ANS(double, avx, , clear)
STD_MK_PROC_DECL(double, avx, , clear)
{
register __m256 zero = _mm256_setzero_ps();
for (size_t i = 0; i < 8; i++) {
@@ -46,7 +46,7 @@ DECLARE_PNL_PROC_ANS(double, avx, , clear)
}
}
DECLARE_PNL_PROC_ANS(double, 512, , clear)
STD_MK_PROC_DECL(double, 512, , clear)
{
register __m512 zero = _mm512_setzero_ps();
for (size_t i = 0; i < 16; i++) {
@@ -55,7 +55,7 @@ DECLARE_PNL_PROC_ANS(double, 512, , clear)
}
// Matrix transpose implementations
DECLARE_KER_PROC_RIGHT_ANS(float, sse, ,transpose)
STD_MK_PROC_DECL(float, sse, , transpose)
{
register __m128 col0 = right->f32_x[0];
register __m128 col1 = right->f32_x[1];
@@ -72,7 +72,7 @@ DECLARE_KER_PROC_RIGHT_ANS(float, sse, ,transpose)
ans->f32_x[3] = _mm_movehl_ps(tmp3, tmp1);
}
DECLARE_KER_PROC_RIGHT_ANS(float, avx, ,transpose)
STD_MK_PROC_DECL(float, avx, , transpose)
{
register __m256 col_low = right->f32_y[0];
register __m256 col_high = right->f32_y[1];
@@ -87,7 +87,7 @@ DECLARE_KER_PROC_RIGHT_ANS(float, avx, ,transpose)
ans->f32_y[1] = _mm256_shuffle_ps(col_high, col_low, _MM_SHUFFLE(3, 1, 3, 1));
}
DECLARE_KER_PROC_RIGHT_ANS(float, 512, ,transpose)
STD_MK_PROC_DECL(float, 512, , transpose)
{
register __m512 matrix = right->f32_z[0];