Clement LONGEAC <[email protected]> writes: > Source: clblast > Version: 1.6.3-1 > Severity: normal > X-Debbugs-Cc: [email protected] > > Dear Maintainer, > > *** Reporter, please consider answering these questions, where appropriate *** > > I am in an internship at Synchrotron-Soleil , my directors are Frederic- > Emmanuel PICCA and Emmanuel FARHI. I implemented Rocm and Pocl autopkgtest for > architecture amd64 and arm64. I implemented autopktests for rocm and pocl > using > Opencl on the package clblaston local. The aim is to have an overview of code > compatibility with various AMD graphics cards, the codes on all the AMD boards > available for CI rocm for GPU and Pocl for CPU. > > I implemented the POCL and ROCm tests in the control file. After running > sbuild, it turns out that the tests involving the CPU, i.e. POCL, pass. > However, the tests for the GPU part had accuracy problems, mainly due to the > nature of the tests. The reason why these tests don't pass is that the card > doesn't support, or supports poorly, FP16 calculations on ROCm. These > libraries > support these calculations, but this may not be enough to make them pass in > terms of precision. We therefore need to increase the tolerance to see if this > would be sufficient (rocBLAS, MIOpen and ROCm-Lib). This may be because the > card has a kernel limitation. Apparently these are common problems with FP16 > on > ROCm, and intrinsic to the card you have. It's symptomatic if you don't have > professional cards. > > Common problems with FP16 on ROCm, due to accuracy issues. CLBlast tests > sometimes fail because tolerances are too tight for FP16. ROCm may use > implicit > FP16 to FP32 conversions (for compatibility), which may explain the numerical > differences. Incomplete hardware support may be the cause. Some cards (such as > the RX 6400) support FP16, but with limited extensions (e.g. no FP16_MMA on > RDNA 2 consumer). The card we have would have partial FP16 support. This would > explain our problem and any other problems we might have in the future.
Dear Clement, Thank you for your work and this bug report! I would highly recommend you share this information with the CLBlast upstream [1]. They often take a while to respond, but usually attend to important matters like this in due time. Also, I'd be very happy to integrate your ROCm-based tests into the CLBlast package – could you provide them, for example as a git merge request on Salsa [2]? (I currently don't have access to any ROCm hardware, so it may be a while until I get around to testing things myself.) [1] https://github.com/CNugteren/CLBlast [2] https://salsa.debian.org/gspr/clblast Best, Gard
signature.asc
Description: PGP signature
