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/dllmain.c
+1 -1
View File
@@ -1,5 +1,5 @@
#include "StgSharpNative.h"
#include "sn_matkernel.h"
#include "sn_intrinsic.matkernel.h"
#include <stdint.h>
#ifdef _WIN32
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];
StgSharp.Native/sn-src/m.bufop.x86.c
+28 -25
View File
@@ -3,9 +3,10 @@
#if SN_IS_ARCH(SN_ARCH_X86_64)
#include "sn_intrinsic.h"
#include "sn_intrincontext.x86.h"
#include "sn_intrinsic.std.h"
#include "sn_intrinsic.context.x86.h"
DECLARE_BUF_PROC_ANS_SCALAR(float, sse, ,fill)
STD_MK_RPOC_DECL(float, sse, , fill)
{
register __m128 vec = _mm_set_ps1(((float *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
@@ -16,7 +17,7 @@ DECLARE_BUF_PROC_ANS_SCALAR(float, sse, ,fill)
}
}
DECLARE_BUF_PROC_ANS_SCALAR(float, avx, ,fill)
DECLARE_BUF_PROC_ANS_SCALAR(float, avx, , fill)
{
register __m256 vec = _mm256_broadcast_ss(((float *)(scalar->data)));
for (size_t i = 0; i < count; i++) {
@@ -25,7 +26,7 @@ DECLARE_BUF_PROC_ANS_SCALAR(float, avx, ,fill)
}
}
DECLARE_BUF_PROC_ANS_SCALAR(float, 512, ,fill)
DECLARE_BUF_PROC_ANS_SCALAR(float, 512, , fill)
{
register __m512 vec = _mm512_broadcastss_ps(((__m128 *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
@@ -33,7 +34,7 @@ DECLARE_BUF_PROC_ANS_SCALAR(float, 512, ,fill)
}
}
DECLARE_BUF_PROC_ANS_SCALAR(double, sse, ,fill)
DECLARE_BUF_PROC_ANS_SCALAR(double, sse, , fill)
{
register __m128d vec = _mm_set1_pd(((double *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
@@ -48,7 +49,7 @@ DECLARE_BUF_PROC_ANS_SCALAR(double, sse, ,fill)
}
}
DECLARE_BUF_PROC_ANS_SCALAR(double, avx, ,fill)
DECLARE_BUF_PROC_ANS_SCALAR(double, avx, , fill)
{
register __m256d vec = _mm256_broadcast_sd(((double *)(scalar->data)));
for (size_t i = 0; i < count; i++) {
@@ -59,7 +60,7 @@ DECLARE_BUF_PROC_ANS_SCALAR(double, avx, ,fill)
}
}
DECLARE_BUF_PROC_ANS_SCALAR(double, 512, ,fill)
DECLARE_BUF_PROC_ANS_SCALAR(double, 512, , fill)
{
register __m512d vec = _mm512_broadcastsd_pd(_mm_loadu_pd(scalar));
for (size_t i = 0; i < count; i++) {
@@ -68,8 +69,10 @@ DECLARE_BUF_PROC_ANS_SCALAR(double, 512, ,fill)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, ,add)
STD_MK_PROC_DECL(float, sse, , add)
{
__mk_param(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
register __m128 c0 = _mm_add_ps(left[i].f32_x[0], right[i].f32_x[0]);
register __m128 c1 = _mm_add_ps(left[i].f32_x[1], right[i].f32_x[1]);
@@ -83,7 +86,7 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, ,add)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, ,add)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, , add)
{
for (size_t i = 0; i < count; i++) {
size_t s = sizeof(MAT_KERNEL(float));
@@ -94,14 +97,14 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, ,add)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, ,add)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, , add)
{
for (size_t i = 0; i < count; i++) {
register __m512 c0 = _mm512_add_ps(left[i].f32_z[0], right[i].f32_z[0]);
ans[i].f32_z[0] = c0;
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, sse, ,add)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, sse, , add)
{
for (size_t i = 0; i < count; i++) {
register __m128d c0 = _mm_add_pd(_mm_castps_pd(left[i].f32_x[0]),
@@ -132,7 +135,7 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, sse, ,add)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, avx, ,add)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, avx, , add)
{
for (size_t i = 0; i < count; i++) {
size_t s = sizeof(MAT_KERNEL(float));
@@ -147,7 +150,7 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, avx, ,add)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, 512, ,add)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, 512, , add)
{
for (size_t i = 0; i < count; i++) {
register __m512d c0 = _mm512_add_pd(left[i].f64_z[0], right[i].f64_z[0]);
@@ -157,7 +160,7 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, 512, ,add)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, ,sub)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, , sub)
{
for (size_t i = 0; i < count; i++) {
register __m128 c0 = _mm_sub_ps(left[i].f32_x[0], right[i].f32_x[0]);
@@ -172,7 +175,7 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, ,sub)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, ,sub)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, , sub)
{
for (size_t i = 0; i < count; i++) {
register __m256 c0 = _mm256_sub_ps(left[i].f32_y[0], right[i].f32_y[0]);
@@ -182,7 +185,7 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, ,sub)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, ,sub)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, , sub)
{
for (size_t i = 0; i < count; i++) {
register __m512 c0 = _mm512_sub_ps(left[i].f32_z[0], right[i].f32_z[0]);
@@ -190,7 +193,7 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, ,sub)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, sse, ,sub)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, sse, , sub)
{
for (size_t i = 0; i < count; i++) {
register __m128d c0 = _mm_sub_pd(left[i].f64_x[0], right[i].f64_x[0]);
@@ -213,7 +216,7 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, sse, ,sub)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, avx, ,sub)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, avx, , sub)
{
for (size_t i = 0; i < count; i++) {
register __m256 c0 = _mm256_sub_ps(left[i].f32_y[0], right[i].f32_y[0]);
@@ -223,7 +226,7 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, avx, ,sub)
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, 512, ,sub)
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, 512, , sub)
{
for (size_t i = 0; i < count; i++) {
register __m512 c0 = _mm512_sub_ps(left[i].f32_z[0], right[i].f32_z[0]);
@@ -231,7 +234,7 @@ DECLARE_BUF_PROC_LEFT_RIGHT_ANS(double, 512, ,sub)
}
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, sse, ,scalar_mul)
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, sse, , scalar_mul)
{
register __m128 vec = _mm_set_ps1(((float *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
@@ -247,7 +250,7 @@ DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, sse, ,scalar_mul)
}
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, avx, ,scalar_mul)
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, avx, , scalar_mul)
{
register __m256 vec = _mm256_broadcast_ss(((float *)(scalar->data)));
for (size_t i = 0; i < count; i++) {
@@ -259,7 +262,7 @@ DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, avx, ,scalar_mul)
}
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, 512, ,scalar_mul)
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, 512, , scalar_mul)
{
register __m512 vec = _mm512_broadcastss_ps(((__m128 *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
@@ -269,7 +272,7 @@ DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, 512, ,scalar_mul)
}
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, sse, ,scalar_mul)
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, sse, , scalar_mul)
{
register __m128d vec = _mm_set1_pd(((double *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
@@ -293,7 +296,7 @@ DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, sse, ,scalar_mul)
}
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, avx, ,scalar_mul)
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, avx, , scalar_mul)
{
register __m256d vec = _mm256_broadcast_sd(((double *)(scalar->data)));
for (size_t i = 0; i < count; i++) {
@@ -309,7 +312,7 @@ DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, avx, ,scalar_mul)
}
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, 512, ,scalar_mul)
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(double, 512, , scalar_mul)
{
register __m512d vec = _mm512_broadcastsd_pd(_mm_load_pd(scalar));
for (size_t i = 0; i < count; i++) {