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

mat queue find solution, native supports clang

Browse Source
This commit is contained in:
Youmu Liu
2026-03-27 18:14:54 +08:00
parent 06e3aa119b
commit 61303635ad
16 changed files with 373 additions and 647 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/StgSharpNative.h
+1 -1
View File
@@ -55,7 +55,7 @@ typedef struct Image {
typedef char *(*imageLoader)(char const *filename, int *x, int *y, int *channels_in_file,
int desired_channels);
SN_API char infolog[512];
extern SN_API char infolog[512];
SN_API int SN_DECL glCheckShaderStat(GladGLContext *context, uint64_t shaderHandle, int key,
char **logRef);
SN_API void SN_DECL initGL(int majorVersion, int minorVersion);
StgSharp.Native/sn-head/sn_intrinsic.context.x86.h
+1 -1
View File
@@ -132,7 +132,7 @@
#pragma region matix function
INTERNAL void f32_normalize(VEC(float) * source, VEC(float) * target);
INTERNAL void f32_normalize(sn_vec_f32 * source, sn_vec_f32 * target);
SN_MK_PROC_DECL_STD(float, sse, , add);
SN_MK_PROC_DECL_STD(float, avx, , add);
StgSharp.Native/sn-head/sn_intrinsic.h
+19 -20
View File
@@ -47,36 +47,35 @@ typedef union scalar_pack {
uint64_t data[8];
} scalar_pack;
/* Mirrors StgSharp.Mathematics.Numeric.MatrixParallelTaskPackage<T> (128 bytes). */
#define MAT_TASK(T) \
struct { \
MAT_KERNEL(T) * mat_0; /* Offset 0: 1st matrix kernel base */ \
MAT_KERNEL(T) * mat_1; /* Offset 8: 2nd matrix kernel base */ \
MAT_KERNEL(T) * mat_2; /* Offset 16: 3rd matrix kernel base */ \
void *scalar; /* Offset 24: scalar pack */ \
int32_t x; /* Offset 32: row index of the result tile */ \
int32_t y; /* Offset 36: column index of the result tile */ \
int32_t z; /* Offset 40: column index of the result tile */ \
int32_t common_x; /* Offset 44: row index of the result tile */ \
int32_t common_y; /* Offset 48: column index of the result tile */ \
int32_t common_z; /* Offset 52: column index of the result tile */ \
uint32_t mask; /* Offset 56: mask for task schedualing, not useful in C side */ \
uint32_t reverse; /* Offset 60: reserved for future use */ \
scalar_pack padding; /* Offset 64: padding to make the struct size 128 bytes */ \
}
typedef struct matrix_task matrix_task;
typedef struct matrix_task {
void *mat_0;
void *mat_1;
void *mat_2;
void *scalar;
int32_t x;
int32_t y;
int32_t z;
int32_t common_x;
int32_t common_y;
int32_t common_z;
uint32_t mask;
uint32_t reverse;
scalar_pack padding;
} matrix_task;
#define MAT_TASK(T) matrix_task
#if defined(__cplusplus)
static_assert(sizeof(matrix_task) == 128, "matrix_parallel_task_package size mismatch");
#elif defined(__STDC_VERSION__)
static_assert(sizeof(MAT_TASK(float)) == 128, "matrix_parallel_task_package size mismatch");
static_assert(sizeof(matrix_task) == 128, "matrix_parallel_task_package size mismatch");
#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 *VECTORNORMALIZEPROC)(sn_vec_f32 *source, sn_vec_f32 *target);
typedef void(SN_DECL *DOTPROC)(void *transpose, __m128 *vector, __m128 *ans);
typedef uint64_t(SN_DECL *FACTORIALROC)(int n);
StgSharp.Native/sn-head/sn_intrinsic.matkernel.h
+17 -4
View File
@@ -12,8 +12,8 @@
#define YMMCOUNT(T) (sizeof(T) / 2)
#define ZMMCOUNT(T) (sizeof(T) / 4)
#define MAT_KERNEL(T) \
union /* MAT_KERNEL_##T## /**/ { \
#define SN_DEFINE_MAT_KERNEL_TYPE(name, T) \
typedef union name { \
__m128 f32_x[XMMCOUNT(T)]; \
__m256 f32_y[YMMCOUNT(T)]; \
__m512 f32_z[ZMMCOUNT(T)]; \
@@ -22,7 +22,18 @@
__m256d f64_y[YMMCOUNT(T)]; \
__m512d f64_z[ZMMCOUNT(T)]; \
T m[4][4]; \
}
} name
SN_DEFINE_MAT_KERNEL_TYPE(sn_mat_kernel_float, float);
SN_DEFINE_MAT_KERNEL_TYPE(sn_mat_kernel_double, double);
SN_DEFINE_MAT_KERNEL_TYPE(sn_mat_kernel_uint64_t, uint64_t);
#define SN_MAT_KERNEL_TYPE(T) SN_MAT_KERNEL_TYPE_##T
#define SN_MAT_KERNEL_TYPE_float sn_mat_kernel_float
#define SN_MAT_KERNEL_TYPE_double sn_mat_kernel_double
#define SN_MAT_KERNEL_TYPE_uint64_t sn_mat_kernel_uint64_t
#define MAT_KERNEL(T) SN_MAT_KERNEL_TYPE(T)
#define VEC(T) \
union { \
@@ -36,8 +47,10 @@
T v[4]; \
}
typedef VEC(float) sn_vec_f32;
/* Provide a named alias for the zero kernel type to ensure consistent linkage in C++. */
typedef MAT_KERNEL(uint64_t) sn_zero_kernel_u64;
typedef sn_mat_kernel_uint64_t sn_zero_kernel_u64;
/* Helpers (all compile-time): */
#define SN_LANES(T_VEC, T) (sizeof(T_VEC) / sizeof(T))
StgSharp.Native/sn-head/sn_intrinsic.std.h
+22 -21
View File
@@ -8,16 +8,18 @@
#define static_assert _Static_assert
#endif
#define mat_task_local ((_mat_task_local const *)task)
#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 GET_MK_PROC_STD(T_type, T_arch, T_feature, T_op) \
((UNI_MK_PROC)MK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op))
#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 FORCEINLINE 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 GET_SK_PROC_STD(T_type, T_arch, T_feature, T_op) \
((UNI_SK_PROC)SK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op))
#define SN_SK_PROC_DECL_STD(T_type, T_arch, T_feature, T_op) \
INTERNAL FORCEINLINE 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, \
@@ -37,28 +39,28 @@
#define _SN_MATPROC_ANS_SCALAR 1
#define _SN_MK_EXPAND_RIGHT_ANS \
SN_MAT_KERNEL_LOCAL const *right = mat_task_local->mat_1; \
SN_MAT_KERNEL_LOCAL *restrict ans = mat_task_local->mat_2;
SN_MAT_KERNEL_LOCAL const *right = (SN_MAT_KERNEL_LOCAL const *)task->mat_1; \
SN_MAT_KERNEL_LOCAL *restrict ans = (SN_MAT_KERNEL_LOCAL *)task->mat_2;
#define _SN_MK_EXPAND_LEFT_RIGHT_ANS \
SN_MAT_KERNEL_LOCAL const *left = mat_task_local->mat_0; \
SN_MAT_KERNEL_LOCAL const *right = mat_task_local->mat_1; \
SN_MAT_KERNEL_LOCAL *restrict ans = mat_task_local->mat_2;
SN_MAT_KERNEL_LOCAL const *left = (SN_MAT_KERNEL_LOCAL const *)task->mat_0; \
SN_MAT_KERNEL_LOCAL const *right = (SN_MAT_KERNEL_LOCAL const *)task->mat_1; \
SN_MAT_KERNEL_LOCAL *restrict ans = (SN_MAT_KERNEL_LOCAL *)task->mat_2;
#define _SN_MK_EXPAND_RIGHT_ANS_SCALAR \
SN_MAT_KERNEL_LOCAL const *right = mat_task_local->mat_1; \
SN_MAT_KERNEL_LOCAL *restrict ans = mat_task_local->mat_2; \
scalar_pack *restrict scalar = mat_task_local->scalar;
SN_MAT_KERNEL_LOCAL const *right = (SN_MAT_KERNEL_LOCAL const *)task->mat_1; \
SN_MAT_KERNEL_LOCAL *restrict ans = (SN_MAT_KERNEL_LOCAL *)task->mat_2; \
scalar_pack *restrict scalar = (scalar_pack *restrict)task->scalar;
#define _SN_MK_EXPAND_LEFT_RIGHT_ANS_SCALAR \
SN_MAT_KERNEL_LOCAL const *left = mat_task_local->mat_0; \
SN_MAT_KERNEL_LOCAL const *right = mat_task_local->mat_1; \
SN_MAT_KERNEL_LOCAL *restrict ans = mat_task_local->mat_2; \
scalar_pack *restrict scalar = mat_task_local->scalar;
SN_MAT_KERNEL_LOCAL const *left = (SN_MAT_KERNEL_LOCAL const *)task->mat_0; \
SN_MAT_KERNEL_LOCAL const *right = (SN_MAT_KERNEL_LOCAL const *)task->mat_1; \
SN_MAT_KERNEL_LOCAL *restrict ans = (SN_MAT_KERNEL_LOCAL *)task->mat_2; \
scalar_pack *restrict scalar = (scalar_pack *restrict)task->scalar;
#define _SN_MK_EXPAND_ANS_SCALAR \
SN_MAT_KERNEL_LOCAL *restrict ans = mat_task_local->mat_2; \
scalar_pack *restrict scalar = &(mat_task_local->padding);
SN_MAT_KERNEL_LOCAL *restrict ans = (SN_MAT_KERNEL_LOCAL *)task->mat_2; \
scalar_pack *restrict scalar = (scalar_pack *restrict)&(task->padding);
#define _SN_PARAM_STYLE(T) _SN_MATPROC_##T
#define _SN_MK_EXPAND_PARAM(style) _SN_MK_EXPAND_##style
@@ -93,10 +95,9 @@
#pragma endregion
#define __mk_param_std(mem, param, T) \
typedef MAT_KERNEL(T) SN_MAT_KERNEL_LOCAL; \
typedef MAT_TASK(T) _mat_task_local; \
_mat_task_local *_ = (_SN_PARAM_STYLE(param), (_SN_MEM_LAYOUT(mem)), task); \
_SN_MK_EXPAND_PARAM(param); \
#define __mk_param_std(mem, param, T) \
typedef MAT_KERNEL(T) SN_MAT_KERNEL_LOCAL; \
matrix_task const *_ = (_SN_PARAM_STYLE(param), (_SN_MEM_LAYOUT(mem)), task); \
_SN_MK_EXPAND_PARAM(param); \
_SN_MK_EXPAND_MEM(mem)
#endif // !SN_INTRINSIC_PARAM_EXPAND