public static unsafe Mat4x4 TransposeReadOnlySpan(Mat4x4 source)
{
Mat4x4 result = new Mat4x4();
Mat4x4 *srcP = &source;
Mat4x4 *resP = &result;
float *src = (float *)srcP;
float *res = (float *)resP;
var avx1 = Avx.LoadVector256(src);
var avx2 = Avx.LoadVector256(&src[8]);
Vector256 <int> permuter;
fixed(int *p = &pmuter.GetPinnableReference())
{
permuter = Avx.LoadVector256(p);
}
var tmp1 = Avx2.PermuteVar8x32(avx1, permuter);
var tmp2 = Avx2.PermuteVar8x32(avx2, permuter);
var res1 = Avx.Shuffle(tmp1, tmp2, 0b10_00_10_00);
var res2 = Avx.Shuffle(tmp1, tmp2, 0b11_01_11_01);
Avx.Store(&res[0], res1);
Avx.Store(&res[8], res2);
return(result);
}
public static unsafe Mat4x4 TransposeLoadArray(Mat4x4 source)
{
Mat4x4 result = new Mat4x4();
Mat4x4 *srcP = &source;
Mat4x4 *resP = &result;
float *src = (float *)srcP;
float *res = (float *)resP;
var avx1 = Avx.LoadVector256(src);
var avx2 = Avx.LoadVector256(&src[8]);
Vector256 <int> permuter = Avx.LoadVector256(tranpose_permuter);
var tmp1 = Avx2.PermuteVar8x32(avx1, permuter);
var tmp2 = Avx2.PermuteVar8x32(avx2, permuter);
var res1 = Avx.Shuffle(tmp1, tmp2, 0b10_00_10_00);
var res2 = Avx.Shuffle(tmp1, tmp2, 0b11_01_11_01);
Avx.Store(&res[0], res1);
Avx.Store(&res[8], res2);
return(result);
}