Hi Cory, Cordell Bloor, on 2022-01-21: > On 2022-01-17 1:59 p.m., Étienne Mollier wrote: > > maxzor, on 2022-01-16: > > > It might prove useful to keep your RX560 available for testing as well! > > > There are numerous users on Polaris, that are sad at seing official > > > support go away for ROCm. > > > As far as I know, loosing official support does not mean that the GPUs > > > are not > > > "enabled" in the stack anymore, and that you cannot make them work. > > I don't plan to throw hardware away, I utterly dislike wasting. > > But if I recall correctly discussion with Cordell, I expect to > > have problems with at least rocblas on the polaris card. I can > > sure see how far I can go with the RX560 for hardware coverage > > of course. > > I was referring to the incorrect GEMM results in rocBLAS on gfx803 > <https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1218>. > Unfortunately, we do not test unsupported platforms, so breakages go > unnoticed when they're introduced. I suspect the underlying problems are > straightforward to resolve, but the process of debugging them is not. > > I think the best approach would be find a version of rocBLAS that passes all > tests on gfx803. I'm not sure which version that might be, but perhaps > rocBLAS for ROCm 3.5. Then search for the commit that introduced the test > failures. I'd do it myself, but it may be a long while before I have time. > > ROCm does not officially support gfx803, but rocBLAS would gladly accept > patches that fixed problems on the platform. And I would gladly provide > advice to anyone who wanted to track and fix ROCm bugs on gfx803. Thanks for the pointer, will see if I manage to get somewhere around that issue when I get to working on rocblas. Have a nice day, :) -- Étienne Mollier <emollier@emlwks999.eu> Fingerprint: 8f91 b227 c7d6 f2b1 948c 8236 793c f67e 8f0d 11da Sent from /dev/pts/6, please excuse my verbosity. On air: Gazpacho - 117
Attachment:
signature.asc
Description: PGP signature