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

small size buffer op ok

Browse Source
This commit is contained in:
Youmu Liu
2025-11-14 18:28:15 +08:00
parent 1417785726
commit a8c522ab23
19 changed files with 578 additions and 1048 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/k_basicop.c
-117
View File
@@ -1,87 +1,5 @@
#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_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;
}
// Matrix transpose implementations
DECLARE_KER_PROC_RIGHT_ANS(float, sse, transpose)
{
@@ -124,38 +42,3 @@ DECLARE_KER_PROC_RIGHT_ANS(float, 512, transpose)
ans->zmm[0] = _mm512_permutexvar_ps(transpose_indices, matrix);
}
DECLARE_KER_PROC_RIGHT_ANS_SCALAR(float, sse, scalar_mul)
{
register __m128 scalar_vec = ((__m128 *)(scalar->data))[0];
register __m128 c0 = _mm_mul_ps(right->xmm[0], scalar_vec);
register __m128 c1 = _mm_mul_ps(right->xmm[1], scalar_vec);
register __m128 c2 = _mm_mul_ps(right->xmm[2], scalar_vec);
register __m128 c3 = _mm_mul_ps(right->xmm[3], scalar_vec);
ans->xmm[0] = c0;
ans->xmm[1] = c1;
ans->xmm[2] = c2;
ans->xmm[3] = c3;
}
DECLARE_KER_PROC_RIGHT_ANS_SCALAR(float, avx, scalar_mul)
{
register __m256 scalar_vec = ((__m256 *)(scalar->data))[0];
register __m256 c0 = _mm256_mul_ps(right->ymm[0], scalar_vec);
register __m256 c1 = _mm256_mul_ps(right->ymm[1], scalar_vec);
ans->ymm[0] = c0;
ans->ymm[1] = c1;
}
DECLARE_KER_PROC_RIGHT_ANS_SCALAR(float, 512, scalar_mul)
{
register __m512 scalar_vec = ((__m512 *)(scalar->data))[0];
register __m512 c0 = _mm512_mul_ps(right->zmm[0], scalar_vec);
ans->zmm[0] = c0;
}
StgSharp.Native/sn-src/load_intrinsic.c
+12 -9
View File
@@ -65,10 +65,11 @@ 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->f32_kernel_add = KER_PROC(float, sse, add);
context->f32_buffer_add = BUF_PROC(float, sse, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
context->f32_kernel_scalar_mul = KER_PROC(float, sse, scalar_mul);
context->f32_kernel_sub = KER_PROC(float, sse, sub);
context->f32_buffer_fill = BUF_PROC(float, sse, fill);
context->f32_kernel_scalar_mul = BUF_PROC(float, sse, scalar_mul);
context->f32_kernel_sub = BUF_PROC(float, sse, sub);
context->f32_kernel_transpose = KER_PROC(float, sse, transpose);
context->f32_normalize_3 = f32_normalize;
}
@@ -78,10 +79,11 @@ 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->f32_kernel_add = KER_PROC(float, avx, add);
context->f32_buffer_add = BUF_PROC(float, avx, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
context->f32_kernel_scalar_mul = KER_PROC(float, avx, scalar_mul);
context->f32_kernel_sub = KER_PROC(float, avx, sub);
context->f32_buffer_fill = BUF_PROC(float, avx, fill);
context->f32_kernel_scalar_mul = BUF_PROC(float, avx, scalar_mul);
context->f32_kernel_sub = BUF_PROC(float, avx, sub);
context->f32_kernel_transpose = KER_PROC(float, avx, transpose);
context->f32_normalize_3 = f32_normalize;
}
@@ -91,10 +93,11 @@ 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->f32_kernel_add = KER_PROC(float, 512, add);
context->f32_buffer_add = BUF_PROC(float, 512, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
context->f32_kernel_scalar_mul = KER_PROC(float, 512, scalar_mul);
context->f32_kernel_sub = KER_PROC(float, 512, sub);
context->f32_buffer_fill = BUF_PROC(float, 512, fill);
context->f32_kernel_scalar_mul = BUF_PROC(float, 512, scalar_mul);
context->f32_kernel_sub = BUF_PROC(float, 512, sub);
context->f32_kernel_transpose = KER_PROC(float, 512, transpose);
context->f32_normalize_3 = f32_normalize;
}