we were successfully running an OpenACC code on an AMD FirePro W9100 GPU with fairly good performances, but we noticed that in recent versions of the PGI compiler the radeon:hawaii target architecture cannot be specified anymore.
Using a generic -ta=radeon target with the latest compiler, we get a 50% performance drop on the same code.
Was the hawaii architecture support dropped?
Without specifying “hawaii” in an old compiler e.g. pgi 15.10, we get the same bad performances as using 16.5.
It turns out we have not been able to fully validate PGI OpenACC on Hawaii cards. Unfortunately, the hawaii option was inadvertently enabled in certain previous PGI releases. At this point, we don’t have any estimate on when it might be re-enabled.
if we compile with just “-ta=radeon” we can expect a binary that could be run also on an hawaii card, although with sub-optimal performances, or we have to expect a binary that may not work correctly?
Thanks and Best Regards,
I’m not sure how backward compatible the AMD cards are. It may work, but it’s not something we test or support.
The good news is that I just found out that we did re-enable the “-ta=radeon:hawaii” flag in the 16.9 release. The caveat being it hasn’t been as rigorously tested as we like so I can’t say it’s fully supported, only available.