#pragma once

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

#include <xmmintrin.h>
#include <pmmintrin.h>

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);

    if (NULL != ptr)
    {
        *ptr = _mm_loadu_ps(&array);
        return ptr;
    }
    else
    {
        return -1;
    }
}

SSGC_API __m128* __cdecl load_cpy(__m128* ptr)
{
    __m128* ret = _aligned_malloc(sizeof(__m128*),16);

    if (NULL != ret)
    {
        *ret = *ptr;
    }

    printf();

    return ret;
}

SSGC_API __m128* __cdecl addps(__m128* ptr1, __m128* ptr2)
{
    
    __m128* ret = _aligned_malloc(sizeof(__m128*),16);

    if (NULL != ret)
    {
        *ret = _mm_add_ps(*ptr1, *ptr2);
        return ret;
    }
    else
    {
        return -1;
    }

}

SSGC_API __m128* __cdecl subps(__m128* ptr1, __m128* ptr2)
{
    __m128* ret = _aligned_malloc(sizeof(__m128*), 16);

    if (NULL != ret)
    {
        *ret = _mm_sub_ps(*ptr1, *ptr2);
        return ret;
    }
    else
    {
        return -1;
    }
}

SSGC_API __m128* __cdecl mulps_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_mul_ps(*ptr,valueVec);
        return ret;
    }
    else
    {
        return -1;
    }
}

SSGC_API __m128* __cdecl mulps(__m128* ptr1, __m128* ptr2)
{
    __m128* ret = _aligned_malloc(sizeof(__m128), 16);

    if (NULL != ret)
    {
        *ret = _mm_mul_ps(*ptr1, *ptr2);
        return ret;
    }
    else
    {
        return -1;
    }
}

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;
    }
}








SSGC_API void __cdecl deinitM128(__m128* ptr)
{
    free(ptr);
}