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

daily update on 20230804 by nitload youmu

Browse Source
This commit is contained in:
Nitload
2023-08-04 22:51:24 +08:00
parent aae7e0292c
commit b3785e35af
55 changed files with 1363 additions and 791 deletions
Download Patch File Download Diff File Expand all files Collapse all files
源代码/StgShqrpGraphic/linear.c
+54 -6
View File
@@ -6,8 +6,9 @@
#include <stdalign.h>
#include <xmmintrin.h>
#include <pmmintrin.h>
STGSHARPGRAPHIC_API __m128* __cdecl load(float x, float y, float z, float w)
SSGC_API __m128* __cdecl load(float x, float y, float z, float w)
{
float array[4] = {x,y,z,w};
__m128* ptr = _aligned_malloc(sizeof(__m128),16);
@@ -23,7 +24,7 @@ STGSHARPGRAPHIC_API __m128* __cdecl load(float x, float y, float z, float w)
}
}
STGSHARPGRAPHIC_API __m128* __cdecl addps(__m128* ptr1, __m128* ptr2)
SSGC_API __m128* __cdecl addps(__m128* ptr1, __m128* ptr2)
{
__m128* ret = _aligned_malloc(sizeof(__m128*),16);
@@ -40,7 +41,7 @@ STGSHARPGRAPHIC_API __m128* __cdecl addps(__m128* ptr1, __m128* ptr2)
}
STGSHARPGRAPHIC_API __m128* __cdecl subps(__m128* ptr1, __m128* ptr2)
SSGC_API __m128* __cdecl subps(__m128* ptr1, __m128* ptr2)
{
__m128* ret = _aligned_malloc(sizeof(__m128*), 16);
@@ -55,7 +56,7 @@ STGSHARPGRAPHIC_API __m128* __cdecl subps(__m128* ptr1, __m128* ptr2)
}
}
STGSHARPGRAPHIC_API __m128* __cdecl mulps_const(__m128* ptr, float value)
SSGC_API __m128* __cdecl mulps_const(__m128* ptr, float value)
{
float c[4] = { value, value, value, value };
__m128* ret = _aligned_malloc(sizeof(__m128),16);
@@ -73,7 +74,7 @@ STGSHARPGRAPHIC_API __m128* __cdecl mulps_const(__m128* ptr, float value)
}
}
STGSHARPGRAPHIC_API __m128* __cdecl mulps(__m128* ptr1, __m128* ptr2)
SSGC_API __m128* __cdecl mulps(__m128* ptr1, __m128* ptr2)
{
__m128* ret = _aligned_malloc(sizeof(__m128), 16);
@@ -88,9 +89,56 @@ STGSHARPGRAPHIC_API __m128* __cdecl mulps(__m128* ptr1, __m128* ptr2)
}
}
SSGC_API float __cdecl mulps_sum(__m128* ptr1, __m128* ptr2)
{
__m128 ret = _mm_mul_ps(*ptr1,*ptr2);
ret = _mm_hadd_ps(ret,ret);
ret = _mm_hadd_ps(ret,ret);
return *(float*)&ret;
}
SSGC_API __m128* __cdecl divps_const(__m128* ptr, float value)
{
float c[4] = { value, value, value, value };
__m128* ret = _aligned_malloc(sizeof(__m128), 16);
__m128 valueVec = _mm_loadu_ps(&c);
if (NULL != ret)
{
*ret = _mm_div_ps(*ptr, valueVec);
return ret;
}
else
{
return -1;
}
}
SSGC_API __m128* __cdecl divps(__m128* ptr1, __m128* ptr2)
{
__m128* ret = _aligned_malloc(sizeof(__m128), 16);
if (NULL != ret)
{
*ret = _mm_div_ps(*ptr1, *ptr2);
return ret;
}
else
{
return -1;
}
}
STGSHARPGRAPHIC_API void __cdecl deinitM128(__m128* ptr)
SSGC_API void __cdecl deinitM128(__m128* ptr)
{
free(ptr);
}