#include "StgSharpGraphic.h"
#include "SSGinternal.h"
#include <stdio.h>
#include <stdalign.h>

#include <xmmintrin.h>



STGSHARPGRAPHIC_API __m128* InitM128(float x, float y, float z, float w)
{
    __m128* Ptr = _aligned_malloc(16, sizeof(__m128));
    
    if (NULL==Ptr)
    {
        return (__m128*) - 1;
    }
    else
    {
        Ptr->m128_f32[0] = x;
        Ptr->m128_f32[1] = y;
        Ptr->m128_f32[2] = z;
        Ptr->m128_f32[3] = w;
        return Ptr;
    }
}

STGSHARPGRAPHIC_API void DeinitM128(__m128* m128Ptr)
{
    free(m128Ptr);
}

STGSHARPGRAPHIC_API void __stdcall VecAddVec(__m128* ptr1,__m128* ptr2, __m128* ret)
{
    printf("%llu\n",(uint64_t)ptr1);
    printf("%llu\n",(uint64_t)ptr2);
    printf("%llu\n",(uint64_t)ret);

    printf("%f\t%f\t%f\t%f\t\n",
        ptr1->m128_f32[0], ptr1->m128_f32[1], ptr1->m128_f32[2], ptr1->m128_f32[3]
    );
    printf("%f\t%f\t%f\t%f\t\n",
        ptr2->m128_f32[0], ptr2->m128_f32[1], ptr2->m128_f32[2], ptr2->m128_f32[3]
        );

    *ret = _mm_add_ps(*ptr1, *ptr2);

    printf("%f\t%f\t%f\t%f\t\n",
        ret->m128_f32[0], ret->m128_f32[1], ret->m128_f32[2], ret->m128_f32[3]
    );

}

STGSHARPGRAPHIC_API  void VecSubVec(__m128* ptr1, __m128* ptr2, __m128* ret)
{
    printf("%llu\n", (uint64_t)ptr1);
    printf("%llu\n", (uint64_t)ptr2);
    printf("%llu\n", (uint64_t)ret);

    printf("%f\t%f\t%f\t%f\t\n",
        ptr1->m128_f32[0], ptr1->m128_f32[1], ptr1->m128_f32[2], ptr1->m128_f32[3]
    );
    printf("%f\t%f\t%f\t%f\t\n",
        ptr2->m128_f32[0], ptr2->m128_f32[1], ptr2->m128_f32[2], ptr2->m128_f32[3]
    );


    *ret = _mm_sub_ps(*ptr1, *ptr2);

    printf("%f\t%f\t%f\t%f\t\n",
        ret->m128_f32[0], ret->m128_f32[1], ret->m128_f32[2], ret->m128_f32[3]
    );

}

STGSHARPGRAPHIC_API void VecMulVec(__m128* ptr1, __m128* ptr2, __m128* ret)
{
    *ret = _mm_mul_ps(*ptr1, *ptr2);
}