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

daily update on 20241226 by nitload youmu

Browse Source
This commit is contained in:
Nitload
2024-12-26 03:09:46 +08:00
parent 7c8d462ca5
commit 592ad55d57
94 changed files with 3767 additions and 1484 deletions
Download Patch File Download Diff File Expand all files Collapse all files
内核源代码/StgSharpC/mat2.c
+15 -28
View File
@@ -1,45 +1,32 @@
#include "StgSharpC.h"
#include "ssc_intrinsic.h"
INTERNAL void __cdecl transpose23(__2_columnset *source,
__3_columnset *target)
INTERNAL void __cdecl transpose23(__2_columnset *source, __3_columnset *target)
{
__m128 t0 = _mm_shuffle_ps(ALIGN(source->column[0]),
ALIGN(source->column[1]),
__m128 t0 = _mm_shuffle_ps(ALIGN(source->column[0]), ALIGN(source->column[1]),
_MM_SHUFFLE(1, 0, 1, 0));
__m128 t1 = _mm_shuffle_ps(ALIGN(source->column[0]),
ALIGN(source->column[1]),
__m128 t1 = _mm_shuffle_ps(ALIGN(source->column[0]), ALIGN(source->column[1]),
_MM_SHUFFLE(3, 2, 3, 2));
target->column[0] =
_mm_shuffle_ps(t0, zeroVec, _MM_SHUFFLE(2, 0, 2, 0));
target->column[1] =
_mm_shuffle_ps(t0, zeroVec, _MM_SHUFFLE(3, 1, 3, 1));
target->column[2] =
_mm_shuffle_ps(t1, zeroVec, _MM_SHUFFLE(2, 0, 2, 0));
target->column[0] = _mm_shuffle_ps(t0, zeroVec, _MM_SHUFFLE(2, 0, 2, 0));
target->column[1] = _mm_shuffle_ps(t0, zeroVec, _MM_SHUFFLE(3, 1, 3, 1));
target->column[2] = _mm_shuffle_ps(t1, zeroVec, _MM_SHUFFLE(2, 0, 2, 0));
}
INTERNAL void __cdecl transpose24(__2_columnset *source,
__4_columnset *target)
INTERNAL void __cdecl transpose24(__2_columnset *source, __4_columnset *target)
{
__m128 t0 = _mm_shuffle_ps(ALIGN(source->column[0]),
ALIGN(source->column[1]),
__m128 t0 = _mm_shuffle_ps(ALIGN(source->column[0]), ALIGN(source->column[1]),
_MM_SHUFFLE(1, 0, 1, 0));
__m128 t1 = _mm_shuffle_ps(ALIGN(source->column[0]),
ALIGN(source->column[1]),
__m128 t1 = _mm_shuffle_ps(ALIGN(source->column[0]), ALIGN(source->column[1]),
_MM_SHUFFLE(3, 2, 3, 2));
target->column[0] =
_mm_shuffle_ps(t0, zeroVec, _MM_SHUFFLE(2, 0, 2, 0));
target->column[1] =
_mm_shuffle_ps(t0, zeroVec, _MM_SHUFFLE(3, 1, 3, 1));
target->column[2] =
_mm_shuffle_ps(t1, zeroVec, _MM_SHUFFLE(2, 0, 2, 0));
target->column[3] =
_mm_shuffle_ps(t1, zeroVec, _MM_SHUFFLE(3, 1, 3, 1));
target->column[0] = _mm_shuffle_ps(t0, zeroVec, _MM_SHUFFLE(2, 0, 2, 0));
target->column[1] = _mm_shuffle_ps(t0, zeroVec, _MM_SHUFFLE(3, 1, 3, 1));
target->column[2] = _mm_shuffle_ps(t1, zeroVec, _MM_SHUFFLE(2, 0, 2, 0));
target->column[3] = _mm_shuffle_ps(t1, zeroVec, _MM_SHUFFLE(3, 1, 3, 1));
}
INTERNAL void __cdecl addmatrix2_sse(__2_columnset* left, __2_columnset* right, __2_columnset* ans)
INTERNAL void __cdecl addmatrix2_sse(__2_columnset *left, __2_columnset *right, __2_columnset *ans)
{
__m128 temp0 = _mm_add_ps(ALIGN(left->column[0]), ALIGN(right->column[0]));
__m128 temp1 = _mm_add_ps(ALIGN(left->column[1]), ALIGN(right->column[1]));
@@ -48,7 +35,7 @@ INTERNAL void __cdecl addmatrix2_sse(__2_columnset* left, __2_columnset* right,
_mm_storeu_ps(&ans->m[0][1], temp1);
}
INTERNAL void __cdecl addmatrix2_avx(__2_columnset* left, __2_columnset* right, __2_columnset* ans)
INTERNAL void __cdecl addmatrix2_avx(__2_columnset *left, __2_columnset *right, __2_columnset *ans)
{
__m256 temp = _mm256_add_ps(ALIGN256(left->column[0]), ALIGN256(right->column[0]));
_mm256_storeu_ps(&ans->m[0][0], temp);