Re: MATMUL broken with frontend optimization.

[Steve Kargl <sgk@troutmask.apl.washington.edu>]

On Fri, Mar 19, 2021 at 07:19:16AM +0100, Thomas Koenig wrote:
> 
> Hi Steve,
> 
> > On my old core2 cpu, a quick test with N=1000 and NxN matrix
> > suggest a cross over near N=1000 for REAL(4).  This cpu doesn't
> > have any AVX* instruction, so YMMV.  Program follows .sig
> 
> Looking at your data with AVX (which I think we can mostly count
> on now),
> 
> - The library is always faster for matmul(vector,matrix) for any n >=100
> - For matmul(matrix,vector) there is no appreciable difference
> 
> So, putting in the same inline limits for matmul(vector,matrix)
> that we have for matmul(matrix,matrix), and leaving
> mamul(matrix,vector) alone, seems like a reasonable thing to do.
> 
> I'll work on a patch.
> 

Thanks for working on this in such short notice.  I agree the core2
is old, and gfortran should look toward the future so using the same
inline threshold seems right.

I saw your other email with the concept patch.  I'll look through
it tomorrow when I'm a little more cogent (it's after midnight here).

-- 
Steve