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

new LU solvability for kernel,on dev, quality is briefly logic clear

Browse Source
This commit is contained in:
Youmu Liu
2026-01-21 18:29:11 +08:00
parent cd36d0962b
commit cb62bd48f3
1 changed files with 37 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/m.solve.x86.c
+37 -37
View File
@@ -90,7 +90,7 @@ DECLARE_KER_PROC_LEFT_RIGHT_ANS_SCALAR(float, sse, , try_plu)
}
}
#define DOT_SSE(a, b, ans) \
#define DOT_F32_SSE(a, b, ans) \
do { \
__m128 _p = _mm_mul_ps((a), (b)); /* p0 p1 p2 p3 */ \
__m128 _t = _mm_shuffle_ps(_p, _p, _MM_SHUFFLE(2, 3, 0, 1)); \
@@ -99,33 +99,41 @@ DECLARE_KER_PROC_LEFT_RIGHT_ANS_SCALAR(float, sse, , try_plu)
ans = _mm_cvt_ss(_mm_add_ss(_p, _t)); \
} while (0)
#define FIND_MAX_F32_SSE(a, b) \
do { \
__m128 _t = _mm_movehl_ps(a, a); \
a = _mm_max_ps(a, _t); \
_t = _mm_shuffle_ps(a, a, _MM_SHUFFLE(2, 3, 0, 1)); \
b = _mm_max_ps(a, _t); \
} while (0)
/*
* layout of scalarpack:
*
*/
DECLARE_KER_PROC_RIGHT_ANS(float, sse, , quality)
{
static const __m128 abs_mask_source = { .m128_i32 = { 0x80000000, 0x80000000, 0x80000000,
0x80000000 } };
static const __m128 tiny_source = { .m128_f32 = { 1e-6f, 1e-6f, 1e-6f, 1e-6f } };
register __m128 tiny = _mm_load_ps(tiny_source.m128_f32);
register __m128 abs_mask = _mm_load_ps(abs_mask_source.m128_i32);
register __m128 col0 = _mm_andnot_ps(abs_mask, _mm_load_ps((float *)&right->f32_x[0]));
register __m128 col1 = _mm_andnot_ps(abs_mask, _mm_load_ps((float *)&right->f32_x[1]));
register __m128 col2 = _mm_andnot_ps(abs_mask, _mm_load_ps((float *)&right->f32_x[2]));
register __m128 col3 = _mm_andnot_ps(abs_mask, _mm_load_ps((float *)&right->f32_x[3]));
register __m128 col0 = _mm_load_ps((float *)&right->f32_x[0]);
register __m128 col1 = _mm_load_ps((float *)&right->f32_x[1]);
register __m128 col2 = _mm_load_ps((float *)&right->f32_x[2]);
register __m128 col3 = _mm_load_ps((float *)&right->f32_x[3]);
// norm
register __m128 row_sum = _mm_add_ps(col0, col1);
row_sum = _mm_add_ps(row_sum, col2);
row_sum = _mm_add_ps(row_sum, col3);
__m128 self_dot_src = { 0 };
// try in reigister mix later
DOT_F32_SSE(col0, col0, self_dot_src.m128_f32[0]);
DOT_F32_SSE(col1, col1, self_dot_src.m128_f32[1]);
DOT_F32_SSE(col2, col2, self_dot_src.m128_f32[2]);
DOT_F32_SSE(col3, col3, self_dot_src.m128_f32[3]);
// max(0,2) and max(1,3)
register __m128 max = _mm_max_ps(row_sum, _mm_movehl_ps(row_sum, row_sum));
max = _mm_max_ps(max, _mm_shuffle_ps(max, max, _MM_SHUFFLE(2, 3, 0, 1)));
float norm = _mm_cvtss_f32(max);
register __m128 self_dot = _mm_load_ps(self_dot_src.m128_f32);
register __m128 max;
FIND_MAX_F32_SSE(self_dot, max);
//float norm = _mm_cvtss_f32(max);
// min on cross
register __m128 col_shuff_1 = _mm_shuffle_ps(col1, col1, _MM_SHUFFLE(0, 3, 2, 1));
@@ -136,27 +144,18 @@ DECLARE_KER_PROC_RIGHT_ANS(float, sse, , quality)
min = _mm_min_ps(min, min0);
float quality = _mm_cvt_ss(_mm_div_ps(min, _mm_add_ps(max, tiny)));
ans->f32_x[0].m128_f32[2] = quality;
ans->f32_x[0].m128_f32[0] = quality;
if (quality < quality_threshold) {
return;
}
__m128 self_dot_src = { 0 };
// try in reigister mix later
DOT_SSE(col0, col0, self_dot_src.m128_f32[0]);
DOT_SSE(col1, col1, self_dot_src.m128_f32[1]);
DOT_SSE(col2, col2, self_dot_src.m128_f32[2]);
DOT_SSE(col3, col3, self_dot_src.m128_f32[3]);
register __m128 self_dot = _mm_load_ps(self_dot_src.m128_f32);
__m128 cross_dot_src[2];
DOT_SSE(col0, col1, cross_dot_src[0].m128_f32[0]);
DOT_SSE(col0, col2, cross_dot_src[0].m128_f32[1]);
DOT_SSE(col0, col3, cross_dot_src[0].m128_f32[2]);
DOT_SSE(col1, col2, cross_dot_src[0].m128_f32[3]);
DOT_SSE(col1, col3, cross_dot_src[1].m128_f32[0]);
DOT_SSE(col2, col3, cross_dot_src[1].m128_f32[1]);
DOT_F32_SSE(col0, col1, cross_dot_src[0].m128_f32[0]);
DOT_F32_SSE(col0, col2, cross_dot_src[0].m128_f32[1]);
DOT_F32_SSE(col0, col3, cross_dot_src[0].m128_f32[2]);
DOT_F32_SSE(col1, col2, cross_dot_src[0].m128_f32[3]);
DOT_F32_SSE(col1, col3, cross_dot_src[1].m128_f32[0]);
DOT_F32_SSE(col2, col3, cross_dot_src[1].m128_f32[1]);
register __m128 cross_dot_0 = _mm_load_ps(cross_dot_src[0].m128_f32);
register __m128 cross_dot_1 = _mm_load_ps(cross_dot_src[1].m128_f32);
@@ -164,17 +163,18 @@ DECLARE_KER_PROC_RIGHT_ANS(float, sse, , quality)
register __m128 base_0 = _mm_shuffle_ps(self_dot, self_dot, _MM_SHUFFLE(1, 0, 0, 0));
register __m128 base_1 = _mm_shuffle_ps(self_dot, self_dot, _MM_SHUFFLE(3, 3, 2, 1));
register __m128 relation =
register __m128 relation_0 =
_mm_div_ps(cross_dot_0, _mm_add_ps(_mm_mul_ps(base_0, base_1), tiny));
_mm_store_ps(ans->f32_x[1].m128_f32, relation);
base_0 = _mm_shuffle_ps(self_dot, self_dot, _MM_SHUFFLE(2, 3, 2, 1));
base_1 = _mm_shuffle_ps(self_dot, self_dot, _MM_SHUFFLE(3, 3, 3, 3));
relation = _mm_div_ps(cross_dot_0, _mm_add_ps(_mm_mul_ps(base_0, base_1), tiny));
register __m128 relation_1 =
_mm_div_ps(cross_dot_0, _mm_add_ps(_mm_mul_ps(base_0, base_1), tiny));
_mm_store_ps(ans->f32_x[2].m128_f32, relation);
relation_0 = _mm_max_ps(relation_0, relation_1);
FIND_MAX_F32_SSE(relation_0, relation_0);
ans->f32_x[0].m128_f32[1] = _mm_cvtss_f32(relation_0);
}
#endif