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

decide to use L4 to solve numa and pack problem for large compute.

Browse Source
This commit is contained in:
Youmu Liu
2026-03-02 17:33:03 +08:00
parent 67fbbde449
commit 7e3ed8b0fd
9 changed files with 192 additions and 86 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/m.bufop.x86.c
+39
View File
@@ -20,6 +20,7 @@ SN_MK_PROC_DECL_STD(float, sse, , fill)
SN_MK_PROC_DECL_STD(float, avx, , fill)
{
__mk_param_std(BUFFER, ANS_SCALAR, float);
register __m256 vec = _mm256_broadcast_ss(((float *)(scalar->data)));
for (size_t i = 0; i < count; i++) {
ans[i].f32_y[0] = vec;
@@ -29,6 +30,7 @@ SN_MK_PROC_DECL_STD(float, avx, , fill)
SN_MK_PROC_DECL_STD(float, 512, , fill)
{
__mk_param_std(BUFFER, ANS_SCALAR, float);
register __m512 vec = _mm512_broadcastss_ps(((__m128 *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
ans[i].f32_z[0] = vec;
@@ -37,6 +39,7 @@ SN_MK_PROC_DECL_STD(float, 512, , fill)
SN_MK_PROC_DECL_STD(double, sse, , fill)
{
__mk_param_std(BUFFER, ANS_SCALAR, float);
register __m128d vec = _mm_set1_pd(((double *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
ans[i].f64_x[0] = vec;
@@ -52,6 +55,7 @@ SN_MK_PROC_DECL_STD(double, sse, , fill)
SN_MK_PROC_DECL_STD(double, avx, , fill)
{
__mk_param_std(BUFFER, ANS_SCALAR, float);
register __m256d vec = _mm256_broadcast_sd(((double *)(scalar->data)));
for (size_t i = 0; i < count; i++) {
ans[i].f64_y[0] = vec;
@@ -63,6 +67,7 @@ SN_MK_PROC_DECL_STD(double, avx, , fill)
SN_MK_PROC_DECL_STD(double, 512, , fill)
{
__mk_param_std(BUFFER, ANS_SCALAR, float);
register __m512d vec = _mm512_broadcastsd_pd(_mm_loadu_pd(scalar));
for (size_t i = 0; i < count; i++) {
ans[i].f64_z[0] = vec;
@@ -89,6 +94,8 @@ SN_MK_PROC_DECL_STD(float, sse, , add)
SN_MK_PROC_DECL_STD(float, avx, , add)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
size_t s = sizeof(MAT_KERNEL(float));
register __m256 c0 = _mm256_add_ps(left[i].f32_y[0], right[i].f32_y[0]);
@@ -100,6 +107,8 @@ SN_MK_PROC_DECL_STD(float, avx, , add)
SN_MK_PROC_DECL_STD(float, 512, , add)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
register __m512 c0 = _mm512_add_ps(left[i].f32_z[0], right[i].f32_z[0]);
ans[i].f32_z[0] = c0;
@@ -107,6 +116,8 @@ SN_MK_PROC_DECL_STD(float, 512, , add)
}
SN_MK_PROC_DECL_STD(double, sse, , add)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
register __m128d c0 = _mm_add_pd(_mm_castps_pd(left[i].f32_x[0]),
_mm_castps_pd(right[i].f32_x[0]));
@@ -138,6 +149,8 @@ SN_MK_PROC_DECL_STD(double, sse, , add)
SN_MK_PROC_DECL_STD(double, avx, , add)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
size_t s = sizeof(MAT_KERNEL(float));
register __m256d c0 = _mm256_add_pd(left[i].f64_y[0], right[i].f64_y[0]);
@@ -153,6 +166,8 @@ SN_MK_PROC_DECL_STD(double, avx, , add)
SN_MK_PROC_DECL_STD(double, 512, , add)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
register __m512d c0 = _mm512_add_pd(left[i].f64_z[0], right[i].f64_z[0]);
register __m512d c1 = _mm512_add_pd(left[i].f64_z[1], right[i].f64_z[1]);
@@ -163,6 +178,8 @@ SN_MK_PROC_DECL_STD(double, 512, , add)
SN_MK_PROC_DECL_STD(float, sse, , sub)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
register __m128 c0 = _mm_sub_ps(left[i].f32_x[0], right[i].f32_x[0]);
register __m128 c1 = _mm_sub_ps(left[i].f32_x[1], right[i].f32_x[1]);
@@ -178,6 +195,8 @@ SN_MK_PROC_DECL_STD(float, sse, , sub)
SN_MK_PROC_DECL_STD(float, avx, , sub)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
register __m256 c0 = _mm256_sub_ps(left[i].f32_y[0], right[i].f32_y[0]);
register __m256 c1 = _mm256_sub_ps(left[i].f32_y[1], right[i].f32_y[1]);
@@ -188,6 +207,8 @@ SN_MK_PROC_DECL_STD(float, avx, , sub)
SN_MK_PROC_DECL_STD(float, 512, , sub)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
register __m512 c0 = _mm512_sub_ps(left[i].f32_z[0], right[i].f32_z[0]);
ans[i].f32_z[0] = c0;
@@ -196,6 +217,8 @@ SN_MK_PROC_DECL_STD(float, 512, , sub)
SN_MK_PROC_DECL_STD(double, sse, , sub)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
register __m128d c0 = _mm_sub_pd(left[i].f64_x[0], right[i].f64_x[0]);
register __m128d c1 = _mm_sub_pd(left[i].f64_x[1], right[i].f64_x[1]);
@@ -219,6 +242,8 @@ SN_MK_PROC_DECL_STD(double, sse, , sub)
SN_MK_PROC_DECL_STD(double, avx, , sub)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
register __m256 c0 = _mm256_sub_ps(left[i].f32_y[0], right[i].f32_y[0]);
register __m256 c1 = _mm256_sub_ps(left[i].f32_y[1], right[i].f32_y[1]);
@@ -229,6 +254,8 @@ SN_MK_PROC_DECL_STD(double, avx, , sub)
SN_MK_PROC_DECL_STD(double, 512, , sub)
{
__mk_param_std(BUFFER, LEFT_RIGHT_ANS, float);
for (size_t i = 0; i < count; i++) {
register __m512 c0 = _mm512_sub_ps(left[i].f32_z[0], right[i].f32_z[0]);
ans[i].f32_z[0] = c0;
@@ -237,6 +264,8 @@ SN_MK_PROC_DECL_STD(double, 512, , sub)
SN_MK_PROC_DECL_STD(float, sse, , scalar_mul)
{
__mk_param_std(BUFFER, RIGHT_ANS_SCALAR, float);
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].f32_x[0], vec);
@@ -253,6 +282,8 @@ SN_MK_PROC_DECL_STD(float, sse, , scalar_mul)
SN_MK_PROC_DECL_STD(float, avx, , scalar_mul)
{
__mk_param_std(BUFFER, RIGHT_ANS_SCALAR, float);
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].f32_y[0], vec);
@@ -265,6 +296,8 @@ SN_MK_PROC_DECL_STD(float, avx, , scalar_mul)
SN_MK_PROC_DECL_STD(float, 512, , scalar_mul)
{
__mk_param_std(BUFFER, RIGHT_ANS_SCALAR, float);
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].f32_z[0], vec);
@@ -275,6 +308,8 @@ SN_MK_PROC_DECL_STD(float, 512, , scalar_mul)
SN_MK_PROC_DECL_STD(double, sse, , scalar_mul)
{
__mk_param_std(BUFFER, RIGHT_ANS_SCALAR, double);
register __m128d vec = _mm_set1_pd(((double *)(scalar->data))[0]);
for (size_t i = 0; i < count; i++) {
register __m128d c0 = _mm_mul_pd(right[i].f64_x[0], vec);
@@ -299,6 +334,8 @@ SN_MK_PROC_DECL_STD(double, sse, , scalar_mul)
SN_MK_PROC_DECL_STD(double, avx, , scalar_mul)
{
__mk_param_std(BUFFER, RIGHT_ANS_SCALAR, double);
register __m256d vec = _mm256_broadcast_sd(((double *)(scalar->data)));
for (size_t i = 0; i < count; i++) {
register __m256d c0 = _mm256_mul_pd(right[i].f64_y[0], vec);
@@ -315,6 +352,8 @@ SN_MK_PROC_DECL_STD(double, avx, , scalar_mul)
SN_MK_PROC_DECL_STD(double, 512, , scalar_mul)
{
__mk_param_std(BUFFER, RIGHT_ANS_SCALAR, double);
register __m512d vec = _mm512_broadcastsd_pd(_mm_load_pd(scalar));
for (size_t i = 0; i < count; i++) {
register __m512d c0 = _mm512_mul_pd(right[i].f64_z[0], vec);