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

daily update on 20241228 by nitload youmu, better dot in checking

Browse Source
This commit is contained in:
Nitload
2024-12-29 04:30:40 +08:00
parent 592ad55d57
commit 63c3cd17e2
45 changed files with 7193 additions and 4884 deletions
Download Patch File Download Diff File Expand all files Collapse all files
内核源代码/StgSharpC/mat4.c
+13 -24
View File
@@ -6,11 +6,9 @@
INTERNAL void __cdecl transpose42(__4_columnset const *source, __2_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 t2 = _mm_shuffle_ps(ALIGN(source->column[2]),
ALIGN(source->column[3]),
__m128 t2 = _mm_shuffle_ps(ALIGN(source->column[2]), ALIGN(source->column[3]),
_MM_SHUFFLE(1, 0, 1, 0));
target->column[0] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(2, 0, 2, 0));
@@ -19,17 +17,13 @@ INTERNAL void __cdecl transpose42(__4_columnset const *source, __2_columnset *ta
INTERNAL void __cdecl transpose43(__4_columnset const *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));
__m128 t2 = _mm_shuffle_ps(ALIGN(source->column[2]),
ALIGN(source->column[3]),
__m128 t2 = _mm_shuffle_ps(ALIGN(source->column[2]), ALIGN(source->column[3]),
_MM_SHUFFLE(1, 0, 1, 0));
__m128 t3 = _mm_shuffle_ps(ALIGN(source->column[2]),
ALIGN(source->column[3]),
__m128 t3 = _mm_shuffle_ps(ALIGN(source->column[2]), ALIGN(source->column[3]),
_MM_SHUFFLE(3, 2, 3, 2));
target->column[0] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(2, 0, 2, 0));
@@ -39,17 +33,13 @@ INTERNAL void __cdecl transpose43(__4_columnset const *source, __3_columnset *ta
INTERNAL void __cdecl transpose44(__4_columnset const *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));
__m128 t2 = _mm_shuffle_ps(ALIGN(source->column[2]),
ALIGN(source->column[3]),
__m128 t2 = _mm_shuffle_ps(ALIGN(source->column[2]), ALIGN(source->column[3]),
_MM_SHUFFLE(1, 0, 1, 0));
__m128 t3 = _mm_shuffle_ps(ALIGN(source->column[2]),
ALIGN(source->column[3]),
__m128 t3 = _mm_shuffle_ps(ALIGN(source->column[2]), ALIGN(source->column[3]),
_MM_SHUFFLE(3, 2, 3, 2));
target->column[0] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(2, 0, 2, 0));
@@ -65,10 +55,9 @@ INTERNAL float __cdecl det_mat4(__4_columnset const *mat_ptr, __4_columnset cons
__m128 mc2 = ALIGN(mat_ptr->column[2]);
__m128 mc3 = ALIGN(mat_ptr->column[3]);
__m128 cogeneration = _mm_set_ps(det_mat3_internal(mc0, mc1, mc2),
det_mat3_internal(mc0, mc1, mc3),
det_mat3_internal(mc0, mc2, mc3),
det_mat3_internal(mc1, mc2, mc3));
__m128 cogeneration =
_mm_set_ps(det_mat3_internal(mc0, mc1, mc2), det_mat3_internal(mc0, mc1, mc3),
det_mat3_internal(mc0, mc2, mc3), det_mat3_internal(mc1, mc2, mc3));
__m128 retvec = _mm_mul_ps(ALIGN(transpose->column[3]), cogeneration);
retvec = _mm_hsub_ps(retvec, retvec);
retvec = _mm_hadd_ps(retvec, retvec);