Re: [patch, fortran] Fix PR 69154, inline matmul with WHERE

[Thomas Koenig <tkoenig at netcologne dot de>]

Am 12.01.2016 um 00:17 schrieb Toon Moene:
> On 01/12/2016 12:01 AM, Thomas Koenig wrote:

> > > Am I correct in assuming that your inlining of MATMUL only treats the
> > > form MATMUL(A, B) and not MATMUL(A, TRANSPOSE(B)) ?
> > >
> > > Our code has about half of one and half of the other ...
> >
> > This is, currently, correct, as Steve pointed out.
>
> Yep, I see.
>
> Still, one example of our MATMUL(A, TRANSPOSE(B)) occurrences has B
> being a 65*65 square matrix, with A being (roughly) 22,000 * 65.
>
> I surely going to see what happens, performance wise, if I just change
> the expression to the following lines (TMP being a 65*65 local array):
>
> TMP = TRANSPOSE(B)
> ... MATMUL(A, TMP)

If you are going to do some tests, you could also test (assuming
you're on a 64-bit system)


  integer(kind=8) :: i, j, k

  do i=0, size(b,1)-1
     do k=0, size(a,2)-1
        do j=0, size(a,1)-1
           c(j+1,i+1) = c(j+1,i+1) + a(j+1,k+1) * b(i+1, k+1)
        end do
     end do
  end do

because this is what the unrolling will produce for
c = matmul(a,transpose(b)).

Regards

	Thomas