Regular gcc benchmark runs for sparse-matrix vector multiplication?

[Harald Anlauf]

Hi Toon,

On 12/15/18 16:49, Toon Moene wrote:
> It is probably helpful to ask this question on the GNU Fortran mailing
> list (copied here), as more Fortran developers are listening there ...
> 
> Two questions off-hand:
> 
> 1. Do you have a small test case that shows the problem ?

not yet.  I'd need to sample from the data (matrices) used,
in case it turns out due to different tuning in gcc-9.

I was hoping that one of the codes on
  https://www.gnu.org/software/gcc/benchmarks/
already covers this kind of pattern.

> 2. Is this single- or multi-threaded code ?

The CSC times vector code is just the vanilla Fortran version, and
the CSR times vector uses simple OpenMP, so this is nothing fancy.

Thanks,
Harald

> On 12/13/18 10:31 PM, Harald Anlauf wrote:
>> Hello,
>>
>> is there a public gcc benchmark that checks for performance
>> regressions in particular for sparse-matrix vector multiplication?
>>
>> I've noticed (and confirmed by perf measurements) a significant
>> performance drop on x86_64/Skylake from gfortran-8 to gfortran-9
>> (current trunk version) for a code that uses both CSC and CSR
>> storage.  gfortran-8 performance is the same as gfortran-7.
>>
>> Thanks,
>> Harald
>>
> 
> Kind regards,
>