Updates to rocfft and hipfft
On Wed, Jun 26, 2024 at 03:36:05PM +0200, Cordell Bloor wrote:
> If you wanted to update a few libraries to ROCm 6.1.3, I would
> suggest rocfft and hipfft, as they're straightforward and the issues
> preventing the creation of hipfft autopkgtests were fixed upstream
> in ROCm 6.0.2. The addition of hipfft autopkgtests would be
> beneficial for the Debian ROCm CI.
Upstream git had only tags for ROCm 6.1.2 for the projects. I didn't
want to pick a version number that wasn't there so I went with that
instead.
I have updates to rocfft and hipfft available at
https://salsa.debian.org/kaol/rocfft and
https://salsa.debian.org/kaol/hipfft respectively.
They're pretty much self contained with no rdeps. I'll add myself to
Uploaders before finalizing them. I'm thinking just go with unstable
and not bother with experimental.
However, what's your opinion on testing? I ran tests but my 8GB GPU
couldn't do all of them. Would someone having more memory give them a
spin? Moreover, a new feature with these is support for multi-GPU
operation, I don't know if anyone has that available for testing.
hipfft didn't have tests enabled previously but the new version was
fine when I enabled them.
A funny detail about the multi-GPU support: at least for rocfft's
test, it made it try to run it on my GPU as well as the Ryzen 9
7900X's internal GPU. Do the developers just use Threadrippers for
everything? It got me a segfault but I'll return to reporting that
later. My case was obviously not something you'd expect to work.
Doing "export ROCR_VISIBLE_DEVICES=0" made the test just skip itself
as expected.
Doxygen stuff didn't quite work as expected out of the box and I had
to wrangle with that for both projects. I made one PR to upstream's
hipFFT git. Shortly put, TYPEDEF_HIDES_STRUCTS = YES means that only
typedef names are available and struct and enum names are hidden and
there was some confusion about that.
Reply to: