MATMUL broken with frontend optimization.

[Thomas Koenig]

Am 18.03.21 um 21:22 schrieb Steve Kargl:
> On Thu, Mar 18, 2021 at 07:24:21PM +0100, Thomas Koenig wrote:
>> I didn't finish the previous mail before hitting "send", so here
>> is the postscript...
>>
>>> OK, so I've had a bit of time to look at the actual test case.  I
>>> missed one very important detail before:  This is a vector-matrix
>>> operation.
>>>
>>> For this, we do not have a good library routine (Harald just
>>> removed it because of a bug in buffering), and -fexternal-blas
>>> does not work because we do not handle calls to anything but
>>> *GEMM.
>>
>> A vector-matrix multiplicatin would be a call to *GEMV, a worthy
>> goal, but out of scope so close to a release.
> 
> Agreed.
> 
>>> The idea is that, for a vector-matrix-multiplication, the
>>> compiler should have enough information about the information
>> about how to optimize for the relevant architecture, especially
>> if the user compilers with the right flags.
>>
>> So, the current idea is that, if we optimize, we can inline.
>>
>> What would a better heuristic be?
>>
> 
> Does _gfortran_matmul_r4 (and friends) work for vector-matrix
> products?

Yes.

> I haven't checked.  If so, how about disabling
> in-lining MATMUL for 11.1;

Absolutely not for the general case. This would cause a huge regression
in execution time for 2*2 matrices, and also for small matrix-vector
multiplications.

What we could do is only to enable the inlining for vector*matrix
at -O2 or higher. Again, this will mean a penalty for smaller loops,
but at less than -O2, people probably don't care too much.

If there is agreement on that, I will prepare a patch.

Regards

	Thomas