SIMD_Matrix.cpp

#include<intrin.h>
#include<Windows.h>

typedef struct matrix4X4d
{
  double matrix[4][4];
  
}Matrix4x4D;

void M4x4MultM4x4 (Matrix4x4D *m1, Matrix4x4D *m2,
					  Matrix4x4D *m3)
{
	__m128 otherRow0;
	__m128 otherRow1;
	__m128 otherRow2;
	__m128 otherRow3;
	__m128 newRow0;
	__m128 newRow1;
	__m128 newRow2;
	__m128 newRow3;

    otherRow0 = _mm_loadu_ps(m2->matrix[0]);
    otherRow1 = _mm_loadu_ps(m2->matrix[1]);
    otherRow2 = _mm_loadu_ps(m2->matrix[2]);
    otherRow3 = _mm_loadu_ps(m2->matrix[3]);

    newRow0 = _mm_mul_ps(otherRow0, _mm_set1_ps(m1->matrix[0][0]));
    newRow0 = _mm_add_ps(newRow0, _mm_mul_ps(otherRow1, _mm_set1_ps(m1->matrix[0][1])));
    newRow0 = _mm_add_ps(newRow0, _mm_mul_ps(otherRow2, _mm_set1_ps(m1->matrix[0][2])));
    newRow0 = _mm_add_ps(newRow0, _mm_mul_ps(otherRow3, _mm_set1_ps(m1->matrix[0][3])));

    newRow1 = _mm_mul_ps(otherRow0, _mm_set1_ps(m1->matrix[1][0]));
    newRow1 = _mm_add_ps(newRow1, _mm_mul_ps(otherRow1, _mm_set1_ps(m1->matrix[1][1])));
    newRow1 = _mm_add_ps(newRow1, _mm_mul_ps(otherRow2, _mm_set1_ps(m1->matrix[1][2])));
    newRow1 = _mm_add_ps(newRow1, _mm_mul_ps(otherRow3, _mm_set1_ps(m1->matrix[1][3])));

    newRow2 = _mm_mul_ps(otherRow0, _mm_set1_ps(m1->matrix[2][0]));
    newRow2 = _mm_add_ps(newRow2, _mm_mul_ps(otherRow1, _mm_set1_ps(m1->matrix[2][1])));
    newRow2 = _mm_add_ps(newRow2, _mm_mul_ps(otherRow2, _mm_set1_ps(m1->matrix[2][2])));
    newRow2 = _mm_add_ps(newRow2, _mm_mul_ps(otherRow3, _mm_set1_ps(m1->matrix[2][3])));

    newRow3 = _mm_mul_ps(otherRow0, _mm_set1_ps(m1->matrix[3][0]));
    newRow3 = _mm_add_ps(newRow3, _mm_mul_ps(otherRow1, _mm_set1_ps(m1->matrix[3][1])));
    newRow3 = _mm_add_ps(newRow3, _mm_mul_ps(otherRow2, _mm_set1_ps(m1->matrix[3][2])));
    newRow3 = _mm_add_ps(newRow3, _mm_mul_ps(otherRow3, _mm_set1_ps(m1->matrix[3][3])));

    _mm_storeu_ps(m3->matrix[0], newRow0);
    _mm_storeu_ps(m3->matrix[1], newRow1);
    _mm_storeu_ps(m3->matrix[2], newRow2);
    _mm_storeu_ps(m3->matrix[3], newRow3);
}

int main()
{
  //Initialize and call multiplication for matrix routine
  return 0;
}