tile fma added

This commit is contained in:
Youmu Liu
2026-01-09 18:13:46 +08:00
parent 4ff1d856a7
commit 6827259891
12 changed files with 736 additions and 265 deletions
+3
View File
@@ -8,16 +8,19 @@
#define _CRT_SECURE_NO_WARNINGS 1
#define WIN32_LEAN_AND_MEAN
#define SN_DECL __cdecl
#define SN_MAT_DECL __cdecl __forceinline
#define SN_API __declspec(dllexport)
#define FORCEINLINE __forceinline
#include <windows.h>
#elif defined(__clang__) || defined(__GNUC__)
#define SN_DECL __attribute__((cdecl))
#define SN_MAT_DECL __attribute__((cdecl)) __attribute__((always_inline))
#define SN_API __attribute__((dllexport))
#define FORCEINLINE __attribute__((always_inline))
#else
#define SN_API
#define SN_DECL
#define SN_MAT_DECL
#define FORCEINLINE
#endif
@@ -159,6 +159,16 @@ DECLARE_KER_PROC_LEFT_RIGHT_ANS(double, avx, , fma);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(double, avx, _fma, fma);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(double, 512, , fma);
DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(float, sse, , fma);
DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(float, avx, , fma);
DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(float, avx, _fma, fma);
DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(float, 512, , fma);
DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(double, sse, , fma);
DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(double, avx, , fma);
DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(double, avx, _fma, fma);
DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(double, 512, , fma);
#pragma endregion
#pragma region scaler_simd
+34 -1
View File
@@ -188,6 +188,38 @@ typedef void(SN_DECL *KER_PROC_LAS)(mat_kernel *matrix, mat_kernel *ans, scalar_
#pragma endregion
#pragma region matrix kernel tile function type
typedef void(SN_DECL *KERTILE_PROC_AS)(mat_kernel *ans, scalar_pack *scalar, __m128 *enumeration);
typedef void(SN_DECL *KERTILE_PROC_RA)(mat_kernel *left, mat_kernel *ans, __m128 *enumeration);
typedef void(SN_DECL *KERTILE_PROC_LRA)(mat_kernel *left, mat_kernel *right, mat_kernel *ans,
__m128 *enumeration);
typedef void(SN_DECL *KERTILE_PROC_LAS)(mat_kernel *matrix, mat_kernel *ans, scalar_pack *scalar,
__m128 *enumeration);
#define KERTILE_PROC(T_type, T_arch, T_feature, T_op_name) \
T_type##_kernel_tile_##T_op_name##_##T_arch##T_feature
#define DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS_SCALAR(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL KERTILE_PROC(T_type, T_arch, T_feature, T_op_name)( \
MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \
MAT_KERNEL(T_type) *restrict ans, scalar_pack const *scalar, \
__m128 const *enumeration)
#define DECLARE_KERTILE_PROC_RIGHT_ANS_SCALAR(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL KERTILE_PROC(T_type, T_arch, T_feature, T_op_name)( \
MAT_KERNEL(T_type) const *right, MAT_KERNEL(T_type) *restrict ans, \
scalar_pack const *scalar, __m128 const *enumeration)
#define DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL KERTILE_PROC(T_type, T_arch, T_feature, T_op_name)( \
MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \
MAT_KERNEL(T_type) *restrict ans, __m128 const *enumeration)
#define DECLARE_KERTILE_PROC_RIGHT_ANS(T_type, T_arch, T_feature, T_op_name) \
INTERNAL void SN_DECL KERTILE_PROC(T_type, T_arch, T_feature, T_op_name)( \
MAT_KERNEL(T_type) const *right, MAT_KERNEL(T_type) *restrict ans, \
__m128 const *enumeration)
#pragma endregion
#pragma region special matrix kernel operation
typedef void(SN_DECL *PIVOTPROC)(mat_kernel const *source, size_t col_begin, size_t col_length,
@@ -216,7 +248,8 @@ typedef struct mat_intrinsic {
KER_PROC_RA buffer_transpose;
PANEL_PROC build_panel;
PANEL_PROC clear_panel;
PNL_PROC_LRA panel_fma;
PNL_PROC_LRA ker_fma;
KERTILE_PROC_LRA ker_tile_fma;
PIVOTPROC pivot;
PANEL_PROC store_panel;