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

Better L4 and array storage design

Browse Source
This commit is contained in:
Youmu Liu
2026-03-04 18:12:58 +08:00
parent 7e970352ab
commit dfa6f4ba92
103 changed files with 853 additions and 6837 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/sn_intrinsic.context.x86.h
+45 -8
View File
@@ -182,15 +182,52 @@ SN_MK_PROC_DECL_STD(double, avx, , fma);
SN_MK_PROC_DECL_STD(double, avx, _fma, fma);
SN_MK_PROC_DECL_STD(double, 512, , fma);
SN_MK_PROC_DECL_STD(float, sse, , tiled_fma);
SN_MK_PROC_DECL_STD(float, avx, , tiled_fma);
SN_MK_PROC_DECL_STD(float, avx, _fma, tiled_fma);
SN_MK_PROC_DECL_STD(float, 512, , tiled_fma);
SN_SK_PROC_DECL_STD(float, avx, , add);
SN_SK_PROC_DECL_STD(float, 512, , add);
SN_MK_PROC_DECL_STD(double, sse, , tiled_fma);
SN_MK_PROC_DECL_STD(double, avx, , tiled_fma);
SN_MK_PROC_DECL_STD(double, avx, _fma, tiled_fma);
SN_MK_PROC_DECL_STD(double, 512, , tiled_fma);
SN_SK_PROC_DECL_STD(double, sse, , add);
SN_SK_PROC_DECL_STD(double, avx, , add);
SN_SK_PROC_DECL_STD(double, 512, , add);
SN_SK_PROC_DECL_STD(float, sse, , sub);
SN_SK_PROC_DECL_STD(float, avx, , sub);
SN_SK_PROC_DECL_STD(float, 512, , sub);
SN_SK_PROC_DECL_STD(double, sse, , sub);
SN_SK_PROC_DECL_STD(double, avx, , sub);
SN_SK_PROC_DECL_STD(double, 512, , sub);
SN_SK_PROC_DECL_STD(float, sse, , transpose);
SN_SK_PROC_DECL_STD(float, avx, , transpose);
SN_SK_PROC_DECL_STD(float, 512, , transpose);
SN_SK_PROC_DECL_STD(float, sse, , scalar_mul);
SN_SK_PROC_DECL_STD(float, avx, , scalar_mul);
SN_SK_PROC_DECL_STD(float, 512, , scalar_mul);
SN_SK_PROC_DECL_STD(double, sse, , scalar_mul);
SN_SK_PROC_DECL_STD(double, avx, , scalar_mul);
SN_SK_PROC_DECL_STD(double, 512, , scalar_mul);
SN_SK_PROC_DECL_STD(float, sse, , fill);
SN_SK_PROC_DECL_STD(float, avx, , fill);
SN_SK_PROC_DECL_STD(float, 512, , fill);
SN_SK_PROC_DECL_STD(double, sse, , fill);
SN_SK_PROC_DECL_STD(double, avx, , fill);
SN_SK_PROC_DECL_STD(double, 512, , fill);
SN_SK_PROC_DECL_STD(float, sse, , quality);
SN_SK_PROC_DECL_STD(float, sse, , fma);
SN_SK_PROC_DECL_STD(float, avx, , fma);
SN_SK_PROC_DECL_STD(float, avx, _fma, fma);
SN_SK_PROC_DECL_STD(float, 512, , fma);
SN_SK_PROC_DECL_STD(double, sse, , fma);
SN_SK_PROC_DECL_STD(double, avx, , fma);
SN_SK_PROC_DECL_STD(double, avx, _fma, fma);
SN_SK_PROC_DECL_STD(double, 512, , fma);
#pragma endregion
StgSharp.Native/sn-head/sn_intrinsic.h
+21 -11
View File
@@ -73,6 +73,8 @@ static_assert(sizeof(MAT_TASK(float)) == 128, "matrix_parallel_task_package size
#endif
typedef void(SN_DECL *UNI_MK_PROC)(matrix_task *task);
typedef void(SN_DECL *UNI_SK_PROC)(mat_kernel const *left, mat_kernel const *right,
mat_kernel *restrict ans, uint64_t scalar);
typedef void(SN_DECL *VECTORNORMALIZEPROC)(__m128 *source, __m128 *target);
typedef void(SN_DECL *DOTPROC)(void *transpose, __m128 *vector, __m128 *ans);
@@ -85,25 +87,33 @@ typedef int(SN_DECL *INDEX_PAIR)(short const *stram, uint32_t target, int length
#pragma endregion
typedef struct mat_intrinsic {
UNI_MK_PROC buffer_add;
UNI_MK_PROC buffer_fill;
UNI_MK_PROC buffer_scalar_mul;
UNI_MK_PROC buffer_sub;
UNI_MK_PROC buffer_transpose;
UNI_MK_PROC ker_fma;
UNI_MK_PROC ker_tile_fma;
typedef struct mat_mk_intrinsic {
UNI_MK_PROC add;
UNI_MK_PROC fill;
UNI_MK_PROC scalar_mul;
UNI_MK_PROC sub;
UNI_MK_PROC transpose;
UNI_MK_PROC fma;
UNI_MK_PROC pivot;
} mat_mk_intrinsic;
} mat_intrinsic;
typedef struct mat_sk_intrinsic {
UNI_SK_PROC add;
UNI_SK_PROC fill;
UNI_SK_PROC scalar_mul;
UNI_SK_PROC sub;
UNI_SK_PROC transpose;
UNI_SK_PROC fma;
UNI_SK_PROC pivot;
} mat_sk_intrinsic;
// Mirrors C# IntrinsicContext field order so managed/unmanaged layouts stay in sync.
typedef struct sn_intrinsic {
HASH city_hash_simplify;
FACTORIALROC factorial_simd;
INDEX_PAIR index_pair;
VECTORNORMALIZEPROC vec_f32_normalize_3;
mat_intrinsic mat[4];
mat_mk_intrinsic mat_mk[0x0a];
mat_sk_intrinsic mat_sk[0x0a];
} sn_intrinsic;
typedef enum elment_mask { F32 = 0, F64 = 1, I32 = 2, I64 = 3 } element_mask;
StgSharp.Native/sn-head/sn_intrinsic.std.h
+9 -3
View File
@@ -10,12 +10,18 @@
#define mat_task_local ((_mat_task_local const *)task)
#define STD_MK_PROC_NAME(T_type, T_arch, T_feature, T_op) \
T_type##_mat_proc_##T_op##_##T_arch##_##T_feature
#define MK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op) \
T_type##_multi_kernel_proc_##T_op##_##T_arch##_##T_feature
#define SN_MK_PROC_DECL(T_type, P_name) INTERNAL void SN_DECL P_name(MAT_TASK(T_type) const *task)
#define SN_MK_PROC_DECL_STD(T_type, T_arch, T_feature, T_op) \
INTERNAL void SN_DECL STD_MK_PROC_NAME(T_type, T_arch, T_feature, \
INTERNAL void SN_DECL MK_PROC_NAME_STD(T_type, T_arch, T_feature, \
T_op)(MAT_TASK(T_type) const *task)
#define SK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op) \
T_type##_single_kernel_proc_##T_op##_##T_arch##_##T_feature
#define SN_SK_PROC_DECL_STD(T_type, T_arch, T_feature, T_op) \
INTERNAL void SN_DECL SK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op)( \
MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \
MAT_KERNEL(T_type) *restrict ans, uint64_t scalar)
#define _SN_STD__PROBE_0 0
#define _SN_STD__PROBE_1 1