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

rewrite of f32 fma, and other optmization of simd kernel ops

Browse Source
This commit is contained in:
Youmu Liu
2025-11-10 20:40:19 +08:00
parent 1dee467220
commit 678a846e54
18 changed files with 302 additions and 156 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/sn_intrinsic.h
+29 -24
View File
@@ -31,12 +31,7 @@ typedef union uint64_u {
uint32_t u32[2];
} uint64_u;
typedef union mat_kernel {
__m128 xmm[4];
__m256 ymm[2];
__m512 zmm[1];
float m[4][4];
} mat_kernel;
typedef struct mat_kernel mat_kernel;
typedef union scalar_pack {
uint64_t data[8];
@@ -44,13 +39,10 @@ typedef union scalar_pack {
typedef void(SN_DECL *VECTORNORMALIZEPROC)(__m128 *source, __m128 *target);
typedef void(SN_DECL *DOTPROC)(void *transpose, __m128 *vector, __m128 *ans);
typedef void(SN_DECL *MATRIXTRANSPOSEPROC)(void *source, void *target);
typedef void(SN_DECL *MATRIXDETPROC)(void *mat, void *transpose);
typedef void(SN_DECL *MATARITHMETIC)(void *left, void *right, void *ans);
typedef void(SN_DECL *SCALERMATARITHMETIC)(void *left, void *right, void *ans);
typedef void(SN_DECL *MATFMAPROC)(mat_kernel *left, mat_kernel *right, mat_kernel *ans);
typedef void(SN_DECL *SCALARMATARITHMETIC)(mat_kernel *matrix, float scalar, mat_kernel *ans);
typedef void(SN_DECL *MATTRANSPOSEPROC)(mat_kernel *source, mat_kernel *target);
typedef void(SN_DECL *MAT_PROC_AS)(mat_kernel *ans, scalar_pack *scalar);
typedef void(SN_DECL *MAT_PROC_LA)(mat_kernel *left, mat_kernel *ans);
typedef void(SN_DECL *MAT_PROC_LRA)(mat_kernel *left, mat_kernel *right, mat_kernel *ans);
typedef void(SN_DECL *MAT_PROC_LSA)(mat_kernel *matrix, scalar_pack *scalar, mat_kernel *ans);
typedef uint64_t(SN_DECL *FACTORIALROC)(int n);
#pragma region matix function
@@ -75,12 +67,18 @@ INTERNAL void SN_DECL f32_transpose_avx(MAT_KERNEL(float) const *source,
MAT_KERNEL(float) *restrict target);
INTERNAL void SN_DECL f32_transpose_512(MAT_KERNEL(float) const *source,
MAT_KERNEL(float) *restrict target);
INTERNAL void SN_DECL f32_scalar_mul_sse(MAT_KERNEL(float) const *matrix, float const scalar,
MAT_KERNEL(float) *restrict ans);
INTERNAL void SN_DECL f32_scalar_mul_avx(MAT_KERNEL(float) const *matrix, float const scalar,
MAT_KERNEL(float) *restrict ans);
INTERNAL void SN_DECL f32_scalar_mul_512(MAT_KERNEL(float) const *matrix, float const scalar,
MAT_KERNEL(float) *restrict ans);
INTERNAL void SN_DECL f32_scalar_mul_sse(MAT_KERNEL(float) const *matrix,
MAT_KERNEL(float) *restrict ans,
scalar_pack const *scalar);
INTERNAL void SN_DECL f32_scalar_mul_avx(MAT_KERNEL(float) const *matrix,
MAT_KERNEL(float) *restrict ans,
scalar_pack const *scalar);
INTERNAL void SN_DECL f32_scalar_mul_512(MAT_KERNEL(float) const *matrix,
MAT_KERNEL(float) *restrict ans,
scalar_pack const *scalar);
INTERNAL void SN_DECL f32_fill_sse(MAT_KERNEL(float) *restrict dest, scalar_pack *scaler);
INTERNAL void SN_DECL f32_fill_avx(MAT_KERNEL(float) *restrict dest, scalar_pack *scaler);
INTERNAL void SN_DECL f32_fill_512(MAT_KERNEL(float) *restrict dest, scalar_pack *scaler);
#pragma endregion
#pragma region fma
@@ -88,6 +86,12 @@ INTERNAL void SN_DECL f32_scalar_mul_512(MAT_KERNEL(float) const *matrix, float
INTERNAL void SN_DECL f32_fma_sse(MAT_KERNEL(float) const *restrict left,
MAT_KERNEL(float) const *restrict right,
MAT_KERNEL(float) *restrict ans);
INTERNAL void SN_DECL f32_fma_avx(MAT_KERNEL(float) const *restrict left,
MAT_KERNEL(float) const *restrict right,
MAT_KERNEL(float) *restrict ans);
INTERNAL void SN_DECL f32_fma_512(MAT_KERNEL(float) const *restrict left,
MAT_KERNEL(float) const *restrict right,
MAT_KERNEL(float) *restrict ans);
#pragma endregion
@@ -119,11 +123,12 @@ typedef struct sn_intrinsic {
int (*city_hash_simplify)(byte *str, int length);
FACTORIALROC factorial_simd;
int (*index_pair)(short const *str, uint32_t target, int length);
MATARITHMETIC f32_add;
void (*f32_fma)(void *left, void *right, void *ans);
SCALARMATARITHMETIC f32_scalar_mul;
MATARITHMETIC f32_sub;
MATTRANSPOSEPROC f32_transpose;
MAT_PROC_LRA f32_add;
MAT_PROC_AS f32_fill;
MAT_PROC_LRA f32_fma;
MAT_PROC_LRA f32_scalar_mul;
MAT_PROC_LRA f32_sub;
MAT_PROC_LA f32_transpose;
VECTORNORMALIZEPROC f32_normalize_3;
} sn_intrinsic;