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

arrange of proj structure

Browse Source
This commit is contained in:
Youmu Liu
2026-02-27 17:09:13 +08:00
parent 15434ed053
commit a798ab07f1
173 changed files with 32304 additions and 33955 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/m.fma.x86.c
+18 -18
View File
@@ -244,33 +244,33 @@ DECLARE_KER_PROC_LEFT_RIGHT_ANS(double, avx, , fma)
register __m256d ccol;
/* Column 0 */
ccol = _mm256_load_pd((double *)&ans->f32_y[0]);
ccol = _mm256_load_pd((double *)&ans->f64_y[0]);
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a0, _mm256_permute4x64_pd(b0, _FMA_SELECT(0))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a1, _mm256_permute4x64_pd(b0, _FMA_SELECT(1))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a2, _mm256_permute4x64_pd(b0, _FMA_SELECT(2))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a3, _mm256_permute4x64_pd(b0, _FMA_SELECT(3))));
ans->f64_y[0] = ccol;
_mm256_store_pd((double *)&ans->f64_y[0], ccol);
/* Column 1 */
ccol = _mm256_load_pd((double *)&ans->f32_y[1]);
ccol = _mm256_load_pd((double *)&ans->f64_y[1]);
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a0, _mm256_permute4x64_pd(b1, _FMA_SELECT(0))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a1, _mm256_permute4x64_pd(b1, _FMA_SELECT(1))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a2, _mm256_permute4x64_pd(b1, _FMA_SELECT(2))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a3, _mm256_permute4x64_pd(b1, _FMA_SELECT(3))));
ans->f64_y[1] = ccol;
_mm256_store_pd((double *)&ans->f64_y[1], ccol);
/* Column 2 */
ccol = _mm256_load_pd((double *)&ans->f32_y[2]);
ccol = _mm256_load_pd((double *)&ans->f64_y[2]);
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a0, _mm256_permute4x64_pd(b2, _FMA_SELECT(0))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a1, _mm256_permute4x64_pd(b2, _FMA_SELECT(1))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a2, _mm256_permute4x64_pd(b2, _FMA_SELECT(2))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a3, _mm256_permute4x64_pd(b2, _FMA_SELECT(3))));
ans->f64_y[2] = ccol;
_mm256_store_pd((double *)&ans->f64_y[2], ccol);
/* Column 3 */
ccol = _mm256_load_pd((double *)&ans->f32_y[3]);
ccol = _mm256_load_pd((double *)&ans->f64_y[3]);
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a0, _mm256_permute4x64_pd(b3, _FMA_SELECT(0))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a1, _mm256_permute4x64_pd(b3, _FMA_SELECT(1))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a2, _mm256_permute4x64_pd(b3, _FMA_SELECT(2))));
ccol = _mm256_add_pd(ccol, _mm256_mul_pd(a3, _mm256_permute4x64_pd(b3, _FMA_SELECT(3))));
ans->f64_y[3] = ccol;
_mm256_store_pd((double *)&ans->f64_y[3], ccol);
}
DECLARE_KER_PROC_LEFT_RIGHT_ANS(double, avx, _fma, fma)
@@ -288,33 +288,33 @@ DECLARE_KER_PROC_LEFT_RIGHT_ANS(double, avx, _fma, fma)
register __m256d ccol;
/* Column 0 */
ccol = _mm256_load_pd((double *)&ans->f32_y[0]);
ccol = _mm256_load_pd((double *)&ans->f64_y[0]);
ccol = _mm256_fmadd_pd(a0, _mm256_permute4x64_pd(b0, _FMA_SELECT(0)), ccol);
ccol = _mm256_fmadd_pd(a1, _mm256_permute4x64_pd(b0, _FMA_SELECT(1)), ccol);
ccol = _mm256_fmadd_pd(a2, _mm256_permute4x64_pd(b0, _FMA_SELECT(2)), ccol);
ccol = _mm256_fmadd_pd(a3, _mm256_permute4x64_pd(b0, _FMA_SELECT(3)), ccol);
ans->f64_y[0] = ccol;
_mm256_store_pd((double *)&ans->f64_y[0], ccol);
/* Column 1 */
ccol = _mm256_load_pd((double *)&ans->f32_y[1]);
ccol = _mm256_load_pd((double *)&ans->f64_y[1]);
ccol = _mm256_fmadd_pd(a0, _mm256_permute4x64_pd(b1, _FMA_SELECT(0)), ccol);
ccol = _mm256_fmadd_pd(a1, _mm256_permute4x64_pd(b1, _FMA_SELECT(1)), ccol);
ccol = _mm256_fmadd_pd(a2, _mm256_permute4x64_pd(b1, _FMA_SELECT(2)), ccol);
ccol = _mm256_fmadd_pd(a3, _mm256_permute4x64_pd(b1, _FMA_SELECT(3)), ccol);
ans->f64_y[1] = ccol;
_mm256_store_pd((double *)&ans->f64_y[1], ccol);
/* Column 2 */
ccol = _mm256_load_pd((double *)&ans->f32_y[2]);
ccol = _mm256_load_pd((double *)&ans->f64_y[2]);
ccol = _mm256_fmadd_pd(a0, _mm256_permute4x64_pd(b2, _FMA_SELECT(0)), ccol);
ccol = _mm256_fmadd_pd(a1, _mm256_permute4x64_pd(b2, _FMA_SELECT(1)), ccol);
ccol = _mm256_fmadd_pd(a2, _mm256_permute4x64_pd(b2, _FMA_SELECT(2)), ccol);
ccol = _mm256_fmadd_pd(a3, _mm256_permute4x64_pd(b2, _FMA_SELECT(3)), ccol);
ans->f64_y[2] = ccol;
_mm256_store_pd((double *)&ans->f64_y[2], ccol);
/* Column 3 */
ccol = _mm256_load_pd((double *)&ans->f32_y[3]);
ccol = _mm256_load_pd((double *)&ans->f64_y[3]);
ccol = _mm256_fmadd_pd(a0, _mm256_permute4x64_pd(b3, _FMA_SELECT(0)), ccol);
ccol = _mm256_fmadd_pd(a1, _mm256_permute4x64_pd(b3, _FMA_SELECT(1)), ccol);
ccol = _mm256_fmadd_pd(a2, _mm256_permute4x64_pd(b3, _FMA_SELECT(2)), ccol);
ccol = _mm256_fmadd_pd(a3, _mm256_permute4x64_pd(b3, _FMA_SELECT(3)), ccol);
ans->f64_y[3] = ccol;
_mm256_store_pd((double *)&ans->f64_y[3], ccol);
}
DECLARE_KER_PROC_LEFT_RIGHT_ANS(double, 512, , fma)
@@ -359,8 +359,8 @@ DECLARE_KER_PROC_LEFT_RIGHT_ANS(double, 512, , fma)
c01 = _mm512_fmadd_pd(a30, b01_perm, c01);
c23 = _mm512_fmadd_pd(a12, b23_perm, c23);
ans->f64_z[0] = c01;
ans->f64_z[1] = c23;
_mm512_store_pd((double *)&ans->f64_z[0], c01);
_mm512_store_pd((double *)&ans->f64_z[1], c23);
}
#endif