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

adjust for clang-gcc

Browse Source
This commit is contained in:
Youmu Liu
2026-04-03 18:40:11 +08:00
parent 2f74e5e8a9
commit d702b3224c
19 changed files with 217 additions and 126 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/mk.fma.x86.c
+44 -35
View File
@@ -1,5 +1,5 @@
#include "sn_target.h"
/* clang-format off */
#if SN_IS_ARCH(SN_ARCH_X86_64)
#include "sn_intrinsic.h"
@@ -16,7 +16,8 @@ SN_MK_PROC_DECL_STD(float, sse, , fma)
{
__mk_param_std(TILE, LEFT_RIGHT_ANS, float);
SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
SN_MAT_KERNEL_LOCAL *restrict ans_cur =
ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
register __m128 c0 = _mm_load_ps((float *)&ans_cur->f32_x[0]);
register __m128 c1 = _mm_load_ps((float *)&ans_cur->f32_x[1]);
@@ -72,7 +73,8 @@ SN_MK_PROC_DECL_STD(float, avx, , fma)
{
__mk_param_std(TILE, LEFT_RIGHT_ANS, float);
SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
SN_MAT_KERNEL_LOCAL *restrict ans_cur =
ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
register __m256 c01 = _mm256_load_ps((float *)&ans_cur->f32_y[0]);
register __m256 c23 = _mm256_load_ps((float *)&ans_cur->f32_y[1]);
@@ -105,7 +107,8 @@ SN_MK_PROC_DECL_STD(float, avx, _fma, fma)
{
__mk_param_std(TILE, LEFT_RIGHT_ANS, float);
SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
SN_MAT_KERNEL_LOCAL *restrict ans_cur =
ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
register __m256 c01 = _mm256_load_ps((float *)&ans_cur->f32_y[0]);
register __m256 c23 = _mm256_load_ps((float *)&ans_cur->f32_y[1]);
@@ -136,8 +139,11 @@ SN_MK_PROC_DECL_STD(float, avx, _fma, fma)
SN_MK_PROC_DECL_STD(float, 512, , fma)
{
__mk_param_std(TILE, LEFT_RIGHT_ANS, float);
SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y);
SN_MAT_KERNEL_LOCAL *restrict ans_cur = ans + GET_INDEX(common_y, x, y);
register __m512 c = _mm512_load_ps((float *)ans_cur);
register __m512 c_1 = c;
register __m512 c_2 = c;
register __m512 c_3 = c;
for (size_t z = 0; z < common_z; z++) {
SN_MAT_KERNEL_LOCAL const *right_cur = right + GET_INDEX(common_z, x, z);
@@ -170,31 +176,31 @@ SN_MK_PROC_DECL_STD(float, 512, , fma)
#define DOUBLE_KERNEL_FMA_CYCLE(i) \
do { \
/* k = 0 */ \
register __m128d b_lo = _mm_load_pd(right_cur + (2 * i + 0)); \
register __m128d b_hi = _mm_load_pd(right_cur + (2 * i + 1)); \
register __m128d a_lo = _mm_load_pd(left_cur + (0 * 2 + 0)); \
register __m128d a_hi = _mm_load_pd(left_cur + (0 * 2 + 1)); \
register __m128d b_lo = _mm_load_pd(right_values + (2 * i + 0)); \
register __m128d b_hi = _mm_load_pd(right_values + (2 * i + 1)); \
register __m128d a_lo = _mm_load_pd(left_values + (0 * 2 + 0)); \
register __m128d a_hi = _mm_load_pd(left_values + (0 * 2 + 1)); \
register __m128d b = _mm_shuffle_pd(b_lo, b_lo, 0x0); \
c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \
c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \
\
/* group = 1 */ \
a_lo = _mm_load_pd(left_cur + (1 * 2 + 0)); \
a_hi = _mm_load_pd(left_cur + (1 * 2 + 1)); \
a_lo = _mm_load_pd(left_values + (1 * 2 + 0)); \
a_hi = _mm_load_pd(left_values + (1 * 2 + 1)); \
b = _mm_shuffle_pd(b_lo, b_lo, 0x3); \
c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \
c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \
\
/* groupp = 2 */ \
a_lo = _mm_load_pd(left_cur + (2 * 2 + 0)); \
a_hi = _mm_load_pd(left_cur + (2 * 2 + 1)); \
a_lo = _mm_load_pd(left_values + (2 * 2 + 0)); \
a_hi = _mm_load_pd(left_values + (2 * 2 + 1)); \
b = _mm_shuffle_pd(b_hi, b_hi, 0x0); \
c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \
c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \
\
/* froup = 3 */ \
a_lo = _mm_load_pd(left_cur + (3 * 2 + 0)); \
a_hi = _mm_load_pd(left_cur + (3 * 2 + 1)); \
a_lo = _mm_load_pd(left_values + (3 * 2 + 0)); \
a_hi = _mm_load_pd(left_values + (3 * 2 + 1)); \
b = _mm_shuffle_pd(b_hi, b_hi, 0x3); \
c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \
c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \
@@ -204,20 +210,23 @@ SN_MK_PROC_DECL_STD(double, sse, , fma)
{
__mk_param_std(TILE, LEFT_RIGHT_ANS, double);
SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
SN_MAT_KERNEL_LOCAL *restrict ans_cur =
ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
register __m128d c0_lo = _mm_load_pd(ans_cur->f64_x + (0 * 2 + 0));
register __m128d c0_hi = _mm_load_pd(ans_cur->f64_x + (0 * 2 + 1));
register __m128d c1_lo = _mm_load_pd(ans_cur->f64_x + (1 * 2 + 0));
register __m128d c1_hi = _mm_load_pd(ans_cur->f64_x + (1 * 2 + 1));
register __m128d c2_lo = _mm_load_pd(ans_cur->f64_x + (2 * 2 + 0));
register __m128d c2_hi = _mm_load_pd(ans_cur->f64_x + (2 * 2 + 1));
register __m128d c3_lo = _mm_load_pd(ans_cur->f64_x + (3 * 2 + 0));
register __m128d c3_hi = _mm_load_pd(ans_cur->f64_x + (3 * 2 + 1));
register __m128d c0_lo = ans_cur->f64_x[0];
register __m128d c0_hi = ans_cur->f64_x[1];
register __m128d c1_lo = ans_cur->f64_x[2];
register __m128d c1_hi = ans_cur->f64_x[3];
register __m128d c2_lo = ans_cur->f64_x[4];
register __m128d c2_hi = ans_cur->f64_x[5];
register __m128d c3_lo = ans_cur->f64_x[6];
register __m128d c3_hi = ans_cur->f64_x[7];
for (size_t z = 0; z < common_z; z++) {
SN_MAT_KERNEL_LOCAL const *right_cur = right + GET_INDEX(common_z, x, z);
SN_MAT_KERNEL_LOCAL const *left_cur = left + GET_INDEX(common_y, z, y);
const double *right_values = &right_cur->m[0][0];
const double *left_values = &left_cur->m[0][0];
DOUBLE_KERNEL_FMA_CYCLE(0);
@@ -228,21 +237,21 @@ SN_MK_PROC_DECL_STD(double, sse, , fma)
DOUBLE_KERNEL_FMA_CYCLE(3);
}
_mm_store_pd(ans_cur->f64_x + (0 * 2 + 0), c0_lo);
_mm_store_pd(ans_cur->f64_x + (0 * 2 + 1), c0_hi);
_mm_store_pd(ans_cur->f64_x + (1 * 2 + 0), c1_lo);
_mm_store_pd(ans_cur->f64_x + (1 * 2 + 1), c1_hi);
_mm_store_pd(ans_cur->f64_x + (2 * 2 + 0), c2_lo);
_mm_store_pd(ans_cur->f64_x + (2 * 2 + 1), c2_hi);
_mm_store_pd(ans_cur->f64_x + (3 * 2 + 0), c3_lo);
_mm_store_pd(ans_cur->f64_x + (3 * 2 + 1), c3_hi);
ans_cur->f64_x[0] = c0_lo;
ans_cur->f64_x[1] = c0_hi;
ans_cur->f64_x[2] = c1_lo;
ans_cur->f64_x[3] = c1_hi;
ans_cur->f64_x[4] = c2_lo;
ans_cur->f64_x[5] = c2_hi;
ans_cur->f64_x[6] = c3_lo;
ans_cur->f64_x[7] = c3_hi;
}
SN_MK_PROC_DECL_STD(double, avx, , fma)
{
__mk_param_std(TILE, LEFT_RIGHT_ANS, double);
SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
SN_MAT_KERNEL_LOCAL *restrict ans_cur = ans + GET_INDEX(common_y, x, y); // tKernelAddressUnsafe
register __m256d c0 = _mm256_load_pd((double *)&ans_cur->f64_y[0]);
register __m256d c1 = _mm256_load_pd((double *)&ans_cur->f64_y[1]);
@@ -314,7 +323,7 @@ SN_MK_PROC_DECL_STD(double, avx, _fma, fma)
{
__mk_param_std(TILE, LEFT_RIGHT_ANS, double);
SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
SN_MAT_KERNEL_LOCAL *restrict ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
register __m256d c0 = _mm256_load_pd((double *)&ans_cur->f64_y[0]);
register __m256d c1 = _mm256_load_pd((double *)&ans_cur->f64_y[1]);
@@ -370,7 +379,7 @@ SN_MK_PROC_DECL_STD(double, 512, , fma)
{
__mk_param_std(TILE, LEFT_RIGHT_ANS, double);
SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y);
SN_MAT_KERNEL_LOCAL *restrict ans_cur = ans + GET_INDEX(common_y, x, y);
register __m512d c01 = _mm512_load_pd((double *)&ans_cur->f64_z[0]);
register __m512d c23 = _mm512_load_pd((double *)&ans_cur->f64_z[1]);