Closed migueldiascosta closed 4 years ago
I understand --enable-amd-trans
is meant for single thread only, but then it should be automatically disabled in the threaded ones, no? And the pure OpenMP tests do pass, what fails is the hybrid OpenMP+MPI one...
Ok, if I completely disable both OpenMP and MPI the tests pass - it would be better if --enable-amd-trans
was simply ignored in those cases, no?
Otherwise, at least in the EasyBuild context I suppose we can tweak the fftw easyblock to only allow --enable-amd-trans
for the sequential cases, or build a completely separate sequential module
(Performance difference is marginal for the default gearshifft
tests, I suppose they are simply too small)
added sequential variant to https://github.com/easybuilders/easybuild-easyconfigs/pull/8783
Hi Miguel Dias Costa, Thank you for reporting this issue.
The option "--enable-amd-trans" uses a new transpose method and adds a minor performance improvement. But this optional feature has a known issue in case of threaded/Hybrid execution modes. We suggest not to enable this option when you are running in cases of threaded/Hybrid tests.
For now, for threaded/Hybrid benchmark tests, please use amd-fftw with "--enable-amd-opt" option only (which is the main switch responsible for major performance improvements). Additionally, "--enable-amd-opt" is supported only for float and double libraries as of now.
I understand that "--enable-amd-trans" option should have been disabled for multi-threaded case automatically. Sorry for the inconvenience. We will plan to take care of this in the next release.
this is for
float
datatype using both--enable-amd-opt
and--enable-amd-trans
(no problem without--enable-amd-trans
)