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

mat kernel turns to be generic

Browse Source
This commit is contained in:
Youmu Liu
2025-09-18 21:55:26 +08:00
parent 6e27c6219d
commit 8007eee846
13 changed files with 109 additions and 188 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/k_basicop.c
+21 -12
View File
@@ -1,6 +1,7 @@
#include "sn_intrinsic.h"
INTERNAL void SN_DECL kernel_add_sse(mat_kernel *left, mat_kernel *right, mat_kernel *ans)
INTERNAL void SN_DECL f32_add_sse(MAT_KERNEL(float) * left, MAT_KERNEL(float) * right,
MAT_KERNEL(float) * ans)
{
register __m128 c0 = _mm_add_ps(left->xmm[0], right->xmm[0]);
register __m128 c1 = _mm_add_ps(left->xmm[1], right->xmm[1]);
@@ -13,7 +14,8 @@ INTERNAL void SN_DECL kernel_add_sse(mat_kernel *left, mat_kernel *right, mat_ke
ans->xmm[3] = c3;
}
INTERNAL void SN_DECL kernel_add_avx(mat_kernel *left, mat_kernel *right, mat_kernel *ans)
INTERNAL void SN_DECL f32_add_avx(MAT_KERNEL(float) * left, MAT_KERNEL(float) * right,
MAT_KERNEL(float) * ans)
{
register __m256 c0 = _mm256_add_ps(left->ymm[0], right->ymm[0]);
register __m256 c1 = _mm256_add_ps(left->ymm[1], right->ymm[1]);
@@ -21,13 +23,15 @@ INTERNAL void SN_DECL kernel_add_avx(mat_kernel *left, mat_kernel *right, mat_ke
ans->ymm[1] = c1;
}
INTERNAL void SN_DECL kernel_add_512(mat_kernel *left, mat_kernel *right, mat_kernel *ans)
INTERNAL void SN_DECL f32_add_512(MAT_KERNEL(float) * left, MAT_KERNEL(float) * right,
MAT_KERNEL(float) * ans)
{
register __m512 c0 = _mm512_add_ps(left->zmm[0], right->zmm[0]);
ans->zmm[0] = c0;
}
INTERNAL void SN_DECL kernel_sub_sse(mat_kernel *left, mat_kernel *right, mat_kernel *ans)
INTERNAL void SN_DECL f32_sub_sse(MAT_KERNEL(float) * left, MAT_KERNEL(float) * right,
MAT_KERNEL(float) * ans)
{
register __m128 c0 = _mm_sub_ps(left->xmm[0], right->xmm[0]);
register __m128 c1 = _mm_sub_ps(left->xmm[1], right->xmm[1]);
@@ -40,7 +44,8 @@ INTERNAL void SN_DECL kernel_sub_sse(mat_kernel *left, mat_kernel *right, mat_ke
ans->xmm[3] = c3;
}
INTERNAL void SN_DECL kernel_sub_avx(mat_kernel *left, mat_kernel *right, mat_kernel *ans)
INTERNAL void SN_DECL f32_sub_avx(MAT_KERNEL(float) * left, MAT_KERNEL(float) * right,
MAT_KERNEL(float) * ans)
{
register __m256 c0 = _mm256_sub_ps(left->ymm[0], right->ymm[0]);
register __m256 c1 = _mm256_sub_ps(left->ymm[1], right->ymm[1]);
@@ -48,14 +53,15 @@ INTERNAL void SN_DECL kernel_sub_avx(mat_kernel *left, mat_kernel *right, mat_ke
ans->ymm[1] = c1;
}
INTERNAL void SN_DECL kernel_sub_512(mat_kernel *left, mat_kernel *right, mat_kernel *ans)
INTERNAL void SN_DECL f32_sub_512(MAT_KERNEL(float) * left, MAT_KERNEL(float) * right,
MAT_KERNEL(float) * ans)
{
register __m512 c0 = _mm512_sub_ps(left->zmm[0], right->zmm[0]);
ans->zmm[0] = c0;
}
// Matrix transpose implementations
INTERNAL void SN_DECL kernel_transpose_sse(mat_kernel *source, mat_kernel *target)
INTERNAL void SN_DECL kernel_transpose_sse(MAT_KERNEL(float) * source, MAT_KERNEL(float) * target)
{
register __m128 col0 = source->xmm[0];
register __m128 col1 = source->xmm[1];
@@ -72,7 +78,7 @@ INTERNAL void SN_DECL kernel_transpose_sse(mat_kernel *source, mat_kernel *targe
target->xmm[3] = _mm_movehl_ps(tmp3, tmp1);
}
INTERNAL void SN_DECL kernel_transpose_avx(mat_kernel *source, mat_kernel *target)
INTERNAL void SN_DECL kernel_transpose_avx(MAT_KERNEL(float) * source, MAT_KERNEL(float) * target)
{
register __m256 col_low = source->ymm[0];
register __m256 col_high = source->ymm[1];
@@ -87,7 +93,7 @@ INTERNAL void SN_DECL kernel_transpose_avx(mat_kernel *source, mat_kernel *targe
target->ymm[1] = _mm256_shuffle_ps(col_high, col_low, _MM_SHUFFLE(3, 1, 3, 1));
}
INTERNAL void SN_DECL kernel_transpose_512(mat_kernel *source, mat_kernel *target)
INTERNAL void SN_DECL kernel_transpose_512(MAT_KERNEL(float) * source, MAT_KERNEL(float) * target)
{
register __m512 matrix = source->zmm[0];
@@ -97,7 +103,8 @@ INTERNAL void SN_DECL kernel_transpose_512(mat_kernel *source, mat_kernel *targe
target->zmm[0] = _mm512_permutexvar_ps(transpose_indices, matrix);
}
INTERNAL void SN_DECL kernel_scalar_mul_sse(mat_kernel *matrix, float scalar, mat_kernel *ans)
INTERNAL void SN_DECL kernel_scalar_mul_sse(MAT_KERNEL(float) * matrix, float scalar,
MAT_KERNEL(float) * ans)
{
register __m128 temp = _mm_load_ss(&scalar);
register __m128 scalar_vec = _mm_shuffle_ps(temp, temp, _MM_SHUFFLE(0, 0, 0, 0));
@@ -113,7 +120,8 @@ INTERNAL void SN_DECL kernel_scalar_mul_sse(mat_kernel *matrix, float scalar, ma
ans->xmm[3] = c3;
}
INTERNAL void SN_DECL kernel_scalar_mul_avx(mat_kernel *matrix, float scalar, mat_kernel *ans)
INTERNAL void SN_DECL kernel_scalar_mul_avx(MAT_KERNEL(float) * matrix, float scalar,
MAT_KERNEL(float) * ans)
{
register __m256 scalar_vec = _mm256_broadcast_ss(&scalar);
@@ -124,7 +132,8 @@ INTERNAL void SN_DECL kernel_scalar_mul_avx(mat_kernel *matrix, float scalar, ma
ans->ymm[1] = c1;
}
INTERNAL void SN_DECL kernel_scalar_mul_512(mat_kernel *matrix, float scalar, mat_kernel *ans)
INTERNAL void SN_DECL kernel_scalar_mul_512(MAT_KERNEL(float) * matrix, float scalar,
MAT_KERNEL(float) * ans)
{
__m512 scalar_vec = _mm512_broadcastss_ps(_mm_load_ss(&scalar));
StgSharp.Native/sn-src/k_fma.c
+1 -1
View File
@@ -1,7 +1,7 @@
#include "sn_intrinsic.h"
// ans = left * right(T) + ans
INTERNAL void SN_DECL kernel_fma_sse(mat_kernel const *restrict left,
INTERNAL void SN_DECL f32_fma_sse(MAT_KERNEL(float) const *restrict left,
mat_kernel const *restrict right, mat_kernel *restrict ans)
{
register __m128 source0 = _mm_load_ps(&left->xmm[0]);
StgSharp.Native/sn-src/load_intrinsic.c
+6 -6
View File
@@ -66,10 +66,10 @@ void load_intrinsic_sse(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
context->kernel_add = (MATARITHMETIC)kernel_add_sse;
context->kernel_add = (MATARITHMETIC)f32_add_sse;
context->kernel_fma = (void(*)(void*, void*, void*))kernel_fma_sse;
context->kernel_scalar_mul = kernel_scalar_mul_sse;
context->kernel_sub = (MATARITHMETIC)kernel_sub_sse;
context->kernel_sub = (MATARITHMETIC)f32_sub_sse;
context->kernel_transpose = (MATTRANSPOSEPROC)kernel_transpose_sse;
context->normalize_3 = normalize;
}
@@ -79,10 +79,10 @@ void load_intrinsic_avx(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
context->kernel_add = (MATARITHMETIC)kernel_add_avx;
context->kernel_add = (MATARITHMETIC)f32_add_avx;
context->kernel_fma = (void(*)(void*, void*, void*))kernel_fma_sse;
context->kernel_scalar_mul = kernel_scalar_mul_avx;
context->kernel_sub = (MATARITHMETIC)kernel_sub_avx;
context->kernel_sub = (MATARITHMETIC)f32_sub_avx;
context->kernel_transpose = (MATTRANSPOSEPROC)kernel_transpose_avx;
context->normalize_3 = normalize;
}
@@ -92,10 +92,10 @@ void load_intrinsic_512(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
context->kernel_add = (MATARITHMETIC)kernel_add_512;
context->kernel_add = (MATARITHMETIC)f32_add_512;
context->kernel_fma = (void(*)(void*, void*, void*))kernel_fma_sse;
context->kernel_scalar_mul = kernel_scalar_mul_512;
context->kernel_sub = (MATARITHMETIC)kernel_sub_512;
context->kernel_sub = (MATARITHMETIC)f32_sub_512;
context->kernel_transpose = (MATTRANSPOSEPROC)kernel_transpose_512;
context->normalize_3 = normalize;
}