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

mat kernel turns to be generic

Browse Source
This commit is contained in:
Youmu Liu
2025-09-18 21:55:26 +08:00
parent 6e27c6219d
commit 8007eee846
13 changed files with 109 additions and 188 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/sn_intrinsic.h
+30 -42
View File
@@ -5,7 +5,10 @@
#ifndef SSC_INTRIN
#define SSC_INTRIN
#define SSC_INTRIN_HEAD
#include "StgSharpNative.h"
#include "sn_matkernel.h"
#include <immintrin.h>
#include <smmintrin.h>
@@ -39,25 +42,6 @@ typedef union mat_kernel {
float m[4][4];
} mat_kernel;
typedef union column_2 {
__m256 stream;
__m128 column[2];
float m[4][2];
} __2_columnset;
typedef union column_3 {
__m256 part;
__m128 column[3];
float m[4][3];
} __3_columnset;
typedef union column_4 {
__m512 stream;
__m256 half[2];
__m128 column[4];
float m[4][4];
} __4_columnset;
typedef union scalar_pack {
uint64_t data[8];
} scalar_pack;
@@ -77,34 +61,38 @@ typedef uint64_t(SN_DECL *FACTORIALROC)(int n);
INTERNAL void normalize(__m128 *source, __m128 *target);
INTERNAL void SN_DECL kernel_add_sse(mat_kernel const *left, mat_kernel const *right,
mat_kernel *ans);
INTERNAL void SN_DECL kernel_add_avx(mat_kernel const *left, mat_kernel const *right,
mat_kernel *ans);
INTERNAL void SN_DECL kernel_add_512(mat_kernel const *left, mat_kernel const *right,
mat_kernel *ans);
INTERNAL void SN_DECL kernel_sub_sse(mat_kernel const *left, mat_kernel const *right,
mat_kernel *ans);
INTERNAL void SN_DECL kernel_sub_avx(mat_kernel const *left, mat_kernel const *right,
mat_kernel *ans);
INTERNAL void SN_DECL kernel_sub_512(mat_kernel const *left, mat_kernel const *right,
mat_kernel *ans);
INTERNAL void SN_DECL kernel_transpose_sse(mat_kernel const *source, mat_kernel *target);
INTERNAL void SN_DECL kernel_transpose_avx(mat_kernel const *source, mat_kernel *target);
INTERNAL void SN_DECL kernel_transpose_512(mat_kernel const *source, mat_kernel *target);
INTERNAL void SN_DECL kernel_scalar_mul_sse(mat_kernel const *matrix, float const scalar,
mat_kernel *ans);
INTERNAL void SN_DECL kernel_scalar_mul_avx(mat_kernel const *matrix, float const scalar,
mat_kernel *ans);
INTERNAL void SN_DECL kernel_scalar_mul_512(mat_kernel const *matrix, float const scalar,
mat_kernel *ans);
INTERNAL void SN_DECL f32_add_sse(MAT_KERNEL(float) const *left, MAT_KERNEL(float) const *right,
MAT_KERNEL(float) * ans);
INTERNAL void SN_DECL f32_add_avx(MAT_KERNEL(float) const *left, MAT_KERNEL(float) const *right,
MAT_KERNEL(float) * ans);
INTERNAL void SN_DECL f32_add_512(MAT_KERNEL(float) const *left, MAT_KERNEL(float) const *right,
MAT_KERNEL(float) * ans);
INTERNAL void SN_DECL f32_sub_sse(MAT_KERNEL(float) const *left, MAT_KERNEL(float) const *right,
MAT_KERNEL(float) * ans);
INTERNAL void SN_DECL f32_sub_avx(MAT_KERNEL(float) const *left, MAT_KERNEL(float) const *right,
MAT_KERNEL(float) * ans);
INTERNAL void SN_DECL f32_sub_512(MAT_KERNEL(float) const *left, MAT_KERNEL(float) const *right,
MAT_KERNEL(float) * ans);
INTERNAL void SN_DECL f32_transpose_sse(MAT_KERNEL(float) const *source,
MAT_KERNEL(float) * target);
INTERNAL void SN_DECL f32_transpose_avx(MAT_KERNEL(float) const *source,
MAT_KERNEL(float) * target);
INTERNAL void SN_DECL f32_transpose_512(MAT_KERNEL(float) const *source,
MAT_KERNEL(float) * target);
INTERNAL void SN_DECL f32_scalar_mul_sse(MAT_KERNEL(float) const *matrix, float const scalar,
MAT_KERNEL(float) * ans);
INTERNAL void SN_DECL f32_scalar_mul_avx(MAT_KERNEL(float) const *matrix, float const scalar,
MAT_KERNEL(float) * ans);
INTERNAL void SN_DECL f32_scalar_mul_512(MAT_KERNEL(float) const *matrix, float const scalar,
MAT_KERNEL(float) * ans);
#pragma endregion
#pragma region fma
INTERNAL void SN_DECL kernel_fma_sse(mat_kernel const *restrict left,
mat_kernel const *restrict right, mat_kernel *restrict ans);
INTERNAL void SN_DECL kernel_fma_sse(MAT_KERNEL(float) const *restrict left,
MAT_KERNEL(float) const *restrict right,
MAT_KERNEL(float) *restrict ans);
#pragma endregion
StgSharp.Native/sn-head/sn_matkernel.h
+19
View File
@@ -0,0 +1,19 @@
#ifndef SN_MATKERNEL
#define SN_MATKERNEL
#include <immintrin.h>
#define ELEMENTSIZE(T) sizeof(T)
#define XMMCOUNT(T) sizeof(T) / 16
#define YMMCOUNT(T) sizeof(T) / 32
#define ZMMCOUNT(T) sizeof(T) / 64
#define MAT_KERNEL(T) \
union { \
__m128 xmm[XMMCOUNT(T)]; \
__m256 ymm[YMMCOUNT(T)]; \
__m512 zmm[ZMMCOUNT(T)]; \
T m[16]; \
}
#endif