August ROCm Package Testing Results
Howdy folks,
I did another run of manual tests on the ROCm packages and I have
updated the Debian Supported GPU List [1] using the results. It is,
perhaps, slightly out-of-date already since I was using the previous
version of rocrand. Nevertheless, I doubt much changed. The full set of
logs from both my July and August tests can be found on my website [2].
I really encourage you to take a look at the supported GPU list on the
wiki. The state of ROCm on Debian is truly special. The only place we
fall behind in hardware support is in RDNA 3, but I think we can save
that for after pytorch-rocm. With both LLVM 17 and ROCm 5.7 just around
the corner, a month or two from now will be pretty good time for a HIP
update.
However, there does seem to be an issue with the use of gfx1030 as a
fallback architecture for gfx1031 and gfx1032 in rocSPARSE. I think that
may have been a problem introduced by the update to rocPRIM 5.5.1, but I
haven't had a chance to dig into it yet. In any case, it is a serious
problem but it only affects prune_dense2csr_by_percentage.conversion and
I suspect that the classic export HSA_OVERRIDE_GFX_VERSION=10.3.0 trick
is an effective workaround for this issue. The problem probably is
caused by rocPRIM code that is not expecting the compile-time
architecture target name to differ from the reported run-time name of
the architecture.
A notable missing entry in my test strategy is gfx1011, which was not
tested because installing Navi 12 hardware seems to cause the amdgpu
driver to crash on startup. I know that Navi 12 hardware can work with
ROCm, because I've used it on AWS g4ad instances running Ubuntu. I'll
have to dig into this more.
Sincerely,
Cory Bloor
[1]:
https://salsa.debian.org/rocm-team/community/team-project/-/wikis/Supported-GPU-list
[2]: https://slerp.xyz/rocm/logs/
Reply to: