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

keep working on PLN facotrize

Browse Source
This commit is contained in:
Youmu Liu
2025-12-24 18:09:16 +08:00
parent b60ce56f1c
commit 3b3a90a2af
12 changed files with 428 additions and 305 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/StgSharp.Native.vcxproj
+1 -1
View File
@@ -151,7 +151,7 @@
<ClCompile Include="sn-src\k_panel.c" />
<ClCompile Include="sn-src\k_basicop.c" />
<ClCompile Include="sn-src\dllmain.c" />
<ClCompile Include="sn-src\k_kerop.c" />
<ClCompile Include="sn-src\k_bufop.c" />
<ClCompile Include="sn-src\k_pivot.c" />
<ClCompile Include="sn-src\str_factorial.c" />
<ClCompile Include="sn-src\k_fma.c" />
StgSharp.Native/StgSharp.Native.vcxproj.filters
+1 -1
View File
@@ -48,7 +48,7 @@
<ClCompile Include="sn-src\load_intrinsic.c">
<Filter>源文件</Filter>
</ClCompile>
<ClCompile Include="sn-src\k_kerop.c">
<ClCompile Include="sn-src\k_bufop.c">
<Filter>源文件</Filter>
</ClCompile>
<ClCompile Include="sn-src\k_panel.c">
StgSharp.Native/sn-head/sn_intrinsic.h
+46 -27
View File
@@ -96,6 +96,40 @@ static_assert(sizeof(matrix_task) == 128, "matrix_parallel_task_package size mis
_Static_assert(sizeof(MAT_TASK(float)) == 128, "matrix_parallel_task_package size mismatch");
#endif
#pragma region matrix buffer operation function type
typedef void(SN_DECL *BUF_PROC_AS)(mat_kernel *ans, scalar_pack *scalar, int count);
typedef void(SN_DECL *BUF_PROC_RAS)(mat_kernel *right, mat_kernel *ans, scalar_pack *scalar,
int count);
typedef void(SN_DECL *BUF_PROC_LRAS)(mat_kernel *left, mat_kernel *right, mat_kernel *ans,
scalar_pack *scalar, int count);
typedef void(SN_DECL *BUF_PROC_RA)(mat_kernel *right, mat_kernel *ans, int count);
typedef void(SN_DECL *BUF_PROC_LRA)(mat_kernel *left, mat_kernel *right, mat_kernel *ans,
int count);
#define BUF_PROC(T_type, T_arch, T_op_name) T_type##_buf_##T_op_name##_##T_arch
#define DECLARE_BUF_PROC_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_op_name)( \
MAT_KERNEL(T_type) *restrict ans, scalar_pack const *scalar, size_t count)
#define DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_op_name)( \
MAT_KERNEL(T_type) const *right, MAT_KERNEL(T_type) *restrict ans, \
scalar_pack const *scalar, int count)
#define DECLARE_BUF_PROC_LEFT_RIGHT_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, 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, size_t count)
#define DECLARE_BUF_PROC_RIGHT_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_op_name)( \
MAT_KERNEL(T_type) const *right, MAT_KERNEL(T_type) *restrict ans, size_t count)
#define DECLARE_BUF_PROC_LEFT_RIGHT_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL BUF_PROC(T_type, T_arch, T_op_name)( \
MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \
MAT_KERNEL(T_type) *restrict ans, size_t count)
#pragma endregion
#pragma region matrix panel function type
typedef void(SN_DECL *PANEL_PROC)(mat_panel *panel, mat_kernel *kernel, int col_length,
@@ -142,9 +176,6 @@ typedef void(SN_DECL *KER_PROC_LAS)(mat_kernel *matrix, mat_kernel *ans, scalar_
#define KER_PROC(T_type, T_arch, T_op_name) T_type##_kernel_##T_op_name##_##T_arch
#define DECLARE_KER_PROC_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL KER_PROC(T_type, T_arch, T_op_name)( \
MAT_KERNEL(T_type) *restrict ans, scalar_pack const *scalar)
#define DECLARE_KER_PROC_LEFT_RIGHT_ANS_SCALAR(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL KER_PROC(T_type, T_arch, T_op_name)( \
MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \
@@ -185,25 +216,25 @@ typedef uint64_t(SN_DECL *FACTORIALROC)(int n);
INTERNAL void f32_normalize(VEC(float) * source, VEC(float) * target);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, sse, add);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, avx, add);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, 512, add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, add);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, add);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, sse, sub);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, avx, sub);
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, 512, sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, sub);
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, sub);
DECLARE_KER_PROC_RIGHT_ANS(float, sse, transpose);
DECLARE_KER_PROC_RIGHT_ANS(float, avx, transpose);
DECLARE_KER_PROC_RIGHT_ANS(float, 512, transpose);
DECLARE_KER_PROC_RIGHT_ANS_SCALAR(float, sse, scalar_mul);
DECLARE_KER_PROC_RIGHT_ANS_SCALAR(float, avx, scalar_mul);
DECLARE_KER_PROC_RIGHT_ANS_SCALAR(float, 512, scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, sse, scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, avx, scalar_mul);
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, 512, scalar_mul);
DECLARE_KER_PROC_ANS_SCALAR(float, sse, fill);
DECLARE_KER_PROC_ANS_SCALAR(float, avx, fill);
DECLARE_KER_PROC_ANS_SCALAR(float, 512, fill);
DECLARE_BUF_PROC_ANS_SCALAR(float, sse, fill);
DECLARE_BUF_PROC_ANS_SCALAR(float, avx, fill);
DECLARE_BUF_PROC_ANS_SCALAR(float, 512, fill);
#pragma endregion
@@ -249,18 +280,6 @@ INTERNAL int SN_DECL index_pair_sse(short const *str, uint32_t target, int lengt
#pragma endregion
typedef struct sn_mat_enumerator {
int CurrentColumn;
int CurrentRow;
int KernelColumnLength;
int KernelRowLength;
} sn_mat_enumerator;
typedef struct sn_mat_boarder {
int (*SN_DECL is_in_buffer)(int col, int row, sn_mat_enumerator *enumerator);
long (*SN_DECL get_index_offset)(int col, int row, sn_mat_enumerator *wnumerator);
} sn_mat_boarder;
// Mirrors C# IntrinsicContext field order so managed/unmanaged layouts stay in sync.
typedef struct sn_intrinsic {
HASH city_hash_simplify;
StgSharp.Native/sn-src/k_bufop.c
+134
View File
@@ -0,0 +1,134 @@
#include "sn_intrinsic.h"
DECLARE_BUF_PROC_ANS_SCALAR(float, sse, fill)
{
register __m128 vec = _mm_set_ps1(((float *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
ans[i].xmm[0] = vec;
ans[i].xmm[1] = vec;
ans[i].xmm[2] = vec;
ans[i].xmm[3] = vec;
}
}
DECLARE_BUF_PROC_ANS_SCALAR(float, avx, fill)
{
register __m256 vec = _mm256_broadcast_ss(((float *)(scalar->data)));
for (size_t i = 0; i < count; i++) {
ans[i].ymm[0] = vec;
ans[i].ymm[1] = vec;
}
}
DECLARE_BUF_PROC_ANS_SCALAR(float, 512, fill)
{
register __m512 vec = _mm512_broadcastss_ps(((__m128 *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
ans[i].zmm[0] = vec;
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, add)
{
for (size_t i = 0; i < count; i++) {
register __m128 c0 = _mm_add_ps(left[i].xmm[0], right[i].xmm[0]);
register __m128 c1 = _mm_add_ps(left[i].xmm[1], right[i].xmm[1]);
register __m128 c2 = _mm_add_ps(left[i].xmm[2], right[i].xmm[2]);
register __m128 c3 = _mm_add_ps(left[i].xmm[3], right[i].xmm[3]);
ans[i].xmm[0] = c0;
ans[i].xmm[1] = c1;
ans[i].xmm[2] = c2;
ans[i].xmm[3] = c3;
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, add)
{
for (size_t i = 0; i < count; i++) {
size_t s = sizeof(MAT_KERNEL(float));
register __m256 c0 = _mm256_add_ps(left[i].ymm[0], right[i].ymm[0]);
register __m256 c1 = _mm256_add_ps(left[i].ymm[1], right[i].ymm[1]);
ans[i].ymm[0] = c0;
ans[i].ymm[1] = c1;
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, add)
{
for (size_t i = 0; i < count; i++) {
register __m512 c0 = _mm512_add_ps(left[i].zmm[0], right[i].zmm[0]);
ans[i].zmm[0] = c0;
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, sse, sub)
{
for (size_t i = 0; i < count; i++) {
register __m128 c0 = _mm_sub_ps(left[i].xmm[0], right[i].xmm[0]);
register __m128 c1 = _mm_sub_ps(left[i].xmm[1], right[i].xmm[1]);
register __m128 c2 = _mm_sub_ps(left[i].xmm[2], right[i].xmm[2]);
register __m128 c3 = _mm_sub_ps(left[i].xmm[3], right[i].xmm[3]);
ans[i].xmm[0] = c0;
ans[i].xmm[1] = c1;
ans[i].xmm[2] = c2;
ans[i].xmm[3] = c3;
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, avx, sub)
{
for (size_t i = 0; i < count; i++) {
register __m256 c0 = _mm256_sub_ps(left[i].ymm[0], right[i].ymm[0]);
register __m256 c1 = _mm256_sub_ps(left[i].ymm[1], right[i].ymm[1]);
ans[i].ymm[0] = c0;
ans[i].ymm[1] = c1;
}
}
DECLARE_BUF_PROC_LEFT_RIGHT_ANS(float, 512, sub)
{
for (size_t i = 0; i < count; i++) {
register __m512 c0 = _mm512_sub_ps(left[i].zmm[0], right[i].zmm[0]);
ans[i].zmm[0] = c0;
}
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, sse, scalar_mul)
{
register __m128 vec = _mm_set_ps1(((float *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
register __m128 c0 = _mm_mul_ps(right[i].xmm[0], vec);
register __m128 c1 = _mm_mul_ps(right[i].xmm[1], vec);
register __m128 c2 = _mm_mul_ps(right[i].xmm[2], vec);
register __m128 c3 = _mm_mul_ps(right[i].xmm[3], vec);
ans[i].xmm[0] = c0;
ans[i].xmm[1] = c1;
ans[i].xmm[2] = c2;
ans[i].xmm[3] = c3;
}
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, avx, scalar_mul)
{
register __m256 vec = _mm256_broadcast_ss(((float *)(scalar->data)));
for (size_t i = 0; i < count; i++) {
register __m256 c0 = _mm256_mul_ps(right[i].ymm[0], vec);
register __m256 c1 = _mm256_mul_ps(right[i].ymm[1], vec);
ans[i].ymm[0] = c0;
ans[i].ymm[1] = c1;
}
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, 512, scalar_mul)
{
register __m512 vec = _mm512_broadcastss_ps(((__m128 *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
register __m512 c0 = _mm512_mul_ps(right[i].zmm[0], vec);
ans[i].zmm[0] = c0;
}
}
StgSharp.Native/sn-src/k_kerop.c
-113
View File
@@ -1,113 +0,0 @@
#include "sn_intrinsic.h"
DECLARE_KER_PROC_ANS_SCALAR(float, sse, fill)
{
register __m128 vec = _mm_set_ps1(((float *)(scalar->data))[0]);
ans->xmm[0] = vec;
ans->xmm[1] = vec;
ans->xmm[2] = vec;
ans->xmm[3] = vec;
}
DECLARE_KER_PROC_ANS_SCALAR(float, avx, fill)
{
register __m256 vec = _mm256_broadcast_ss(((float *)(scalar->data)));
ans->ymm[0] = vec;
ans->ymm[1] = vec;
}
DECLARE_KER_PROC_ANS_SCALAR(float, 512, fill)
{
register __m512 vec = _mm512_broadcastss_ps(((__m128 *)(scalar->data))[0]);
ans[i].zmm[0] = vec;
}
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, sse, add)
{
register __m128 c0 = _mm_add_ps(left->xmm[0], right->xmm[0]);
register __m128 c1 = _mm_add_ps(left->xmm[1], right->xmm[1]);
register __m128 c2 = _mm_add_ps(left->xmm[2], right->xmm[2]);
register __m128 c3 = _mm_add_ps(left->xmm[3], right->xmm[3]);
ans->xmm[0] = c0;
ans->xmm[1] = c1;
ans->xmm[2] = c2;
ans->xmm[3] = c3;
}
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, avx, add)
{
register __m256 c0 = _mm256_add_ps(left->ymm[0], right->ymm[0]);
register __m256 c1 = _mm256_add_ps(left->ymm[1], right->ymm[1]);
ans->ymm[0] = c0;
ans->ymm[1] = c1;
}
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, 512, add)
{
register __m512 c0 = _mm512_add_ps(left->zmm[0], right->zmm[0]);
ans->zmm[0] = c0;
}
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, sse, sub)
{
register __m128 c0 = _mm_sub_ps(left->xmm[0], right->xmm[0]);
register __m128 c1 = _mm_sub_ps(left->xmm[1], right->xmm[1]);
register __m128 c2 = _mm_sub_ps(left->xmm[2], right->xmm[2]);
register __m128 c3 = _mm_sub_ps(left->xmm[3], right->xmm[3]);
ans->xmm[0] = c0;
ans->xmm[1] = c1;
ans->xmm[2] = c2;
ans->xmm[3] = c3;
}
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, avx, sub)
{
register __m256 c0 = _mm256_sub_ps(left->ymm[0], right->ymm[0]);
register __m256 c1 = _mm256_sub_ps(left->ymm[1], right->ymm[1]);
ans->ymm[0] = c0;
ans->ymm[1] = c1;
}
DECLARE_KER_PROC_LEFT_RIGHT_ANS(float, 512, sub)
{
register __m512 c0 = _mm512_sub_ps(left->zmm[0], right->zmm[0]);
ans->zmm[0] = c0;
}
DECLARE_KER_PROC_RIGHT_ANS_SCALAR(float, sse, scalar_mul)
{
register __m128 vec = _mm_set_ps1(((float *)(scalar->data))[0]);
register __m128 c0 = _mm_mul_ps(right->xmm[0], vec);
register __m128 c1 = _mm_mul_ps(right->xmm[1], vec);
register __m128 c2 = _mm_mul_ps(right->xmm[2], vec);
register __m128 c3 = _mm_mul_ps(right->xmm[3], vec);
ans->xmm[0] = c0;
ans->xmm[1] = c1;
ans->xmm[2] = c2;
ans->xmm[3] = c3;
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, avx, scalar_mul)
{
register __m256 vec = _mm256_broadcast_ss(((float *)(scalar->data)));
for (size_t i = 0; i < count; i++) {
register __m256 c0 = _mm256_mul_ps(right[i].ymm[0], vec);
register __m256 c1 = _mm256_mul_ps(right[i].ymm[1], vec);
ans[i].ymm[0] = c0;
ans[i].ymm[1] = c1;
}
}
DECLARE_BUF_PROC_RIGHT_ANS_SCALAR(float, 512, scalar_mul)
{
register __m512 vec = _mm512_broadcastss_ps(((__m128 *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
register __m512 c0 = _mm512_mul_ps(right[i].zmm[0], vec);
ans[i].zmm[0] = c0;
}
}