10.5.6 : The performances

The problem of this method is the matrices have to have a number of columns which is a multiple of the number of float in a vectorial register size. So you can try by yourself to find a solution to have an intrinsics method available for any matrix size.