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

redesign matrix design

Browse Source
This commit is contained in:
Youmu Liu
2025-12-23 18:13:05 +08:00
parent b32c1cb4eb
commit b60ce56f1c
12 changed files with 493 additions and 225 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/k_bufop.c
-134
View File
@@ -1,134 +0,0 @@
#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
@@ -0,0 +1,113 @@
#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;
}
}