Update Alpaka and add support for ROCm #8284
Conversation
Update the version of alpaka to the HEAD of the develop branch as of 2023.02.01,
corresponding to the commit a68c866cc6c .
Major changes:
- remove the functions to pin/unpin an existing buffer; host memory
buffers pinnd/mapped for a particular accelerator platform can be
obtained with `allocMappedBuf(...)`;
- implement accelerator tags, to list all available acceerators;
- drop the Boost.Fiber back-end;
- add support for newer CUDA and ROCm backends;
- add optional support for mdspan
Other changes:
- add deduction guide for `Vec`;
- add elementwise_min and max functions;
- make CUDA/ROCm mangled kernel names as short as possible;
- add math hyperbolic functions;
- refactor `ConcurrentExecPool` and `QueueGenericThreadsNonBlocking`.
- implement trait constants.
|
A new Pull Request was created by @fwyzard (Andrea Bocci) for branch IB/CMSSW_13_0_X/master. @cmsbuild, @smuzaffar, @aandvalenzuela, @iarspider can you please review it and eventually sign? Thanks. |
|
please test with cms-sw/cmssw-config#91, cms-sw/cmssw#40678 |
| <flags CXXFLAGS="-DALPAKA_ACC_GPU_HIP_ENABLED -DALPAKA_HOST_ONLY"/> | ||
| <flags GENREFLEX_CPPFLAGS="-DALPAKA_ACC_GPU_HIP_ENABLED -DALPAKA_HOST_ONLY"/> | ||
| <!-- device comiplation should run without ALPAKA_HOST_ONLY defined --> | ||
| <flags ROCM_FLAGS="-DALPAKA_ACC_GPU_HIP_ENABLED -UALPAKA_HOST_ONLY"/> |
There was a problem hiding this comment.
As CXXFLAGS already has -DALPAKA_ACC_GPU_HIP_ENABLED so it is not needed here
There was a problem hiding this comment.
Yes, you're right.
However, I've copied this from the CUDA case, and I would prefer to clean them all up together in a separate PR.
There was a problem hiding this comment.
-UALPAKA_HOST_ONLY should work too but if you prefer then you can use
<flags REM_ROCM_HOST_CXXFLAGS="-DALPAKA_HOST_ONLY"/>
There was a problem hiding this comment.
Following what you told me the other day, I was going to move the flags to CPPDEFINES - but I guess we do not have a REM_CPPDEFINES or REM_ROCM_HOST_CPPDEFINES functionality ?
There was a problem hiding this comment.
So I think we can stick to CXXFLAGS and use REM_CUDA_HOST_CXXFLAGS/REM_ROCM_HOST_CXXFLAGS .
There was a problem hiding this comment.
yes that is correct, currently there is no REM_ROCM_HOST_CPPDEFINES , so lets stick to CXXFLAGS
|
-1 Failed Tests: RelVals-INPUT The following merge commits were also included on top of IB + this PR after doing git cms-merge-topic:
You can see more details here: RelVals-INPUTThe relvals timed out after 4 hours. Comparison SummarySummary:
|
|
@fwyzard , input test failure are not related to this change. Do you want to add more changes here otherwise I am happy to include it for next IB. |
|
@smuzaffar sorry, I missed the message earlier. This PR needs to go in together with cms-sw/cmssw#40678; it would be great if both could enter in time for tonight's IB. @perrotta @rappoccio FYI. |
|
+1 Summary: https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-d1ec0d/30366/summary.html The following merge commits were also included on top of IB + this PR after doing git cms-merge-topic:
You can see more details here: Comparison SummarySummary:
|
|
+1
|
|
merge |
4 participants
Update the version of alpaka to the HEAD of the develop branch as of 2023.02.01, corresponding to the commit alpaka-group/alpaka@a68c866cc6c .
Major changes:
allocMappedBuf(...);Other changes:
Vec;ConcurrentExecPoolandQueueGenericThreadsNonBlocking.Add support for the HIP/ROCm alpaka back-end.