Add rocprofiler-sdk support 2 #1128
Conversation
|
@mwootton , ie. conditionally compile rocTracer file or rocProf files and keep both files in the repo. cc @jeffdaily |
|
Added support to conditionally compile the older roctracer implementation on rocm versions < 6.4. |
|
@sraikund16 Could you take a look at this an confirm it functions properly on your internal builds? |
|
@sraikund16 has imported this pull request. If you are a Meta employee, you can view this in D82773951. |
|
@sraikund16 has imported this pull request. If you are a Meta employee, you can view this in D82773951. |
|
@mwootton has updated the pull request. You must reimport the pull request before landing. |
|
@mwootton has updated the pull request. You must reimport the pull request before landing. |
|
@sraikund16 has imported this pull request. If you are a Meta employee, you can view this in D82773951. |
Summary: Supersedes pytorch#1050 Convert to using rocprofiler-sdk instead of roctracer for collecting hip api calls and AMD gpu activity. Reuses most existing roctracer infrastructure with a name for name replacement. Simultaneous support for both roctracer and rocprofiler-sdk was deemed impractical. This would require a whole new set of #ifdefs, a major refactor of the roctracer code, and additional build support. Even then, only one could be active at a time (and you wouldn't want both active). In homage to the abandoned refactor, RocLogger.cpp/h were created to contain the rocprofbase classes and the api filter. Roctracer has no established end date. Rocprofiler-sdk is in rocm_3.1 forward. This will create a dependency where (newest kineto on old rocm) and (old kineto on newest rocm) could fail to build with AMD gpu support. That window is already over 1 year wide. Pull Request resolved: pytorch#1128 Reviewed By: malfet, aaronenyeshi Differential Revision: D82773951 Pulled By: sraikund16
|
@mwootton has updated the pull request. You must reimport the pull request before landing. |
5 participants
Supersedes #1050
Convert to using rocprofiler-sdk instead of roctracer for collecting hip api calls and AMD gpu activity.
Reuses most existing roctracer infrastructure with a name for name replacement. Simultaneous support for both roctracer and rocprofiler-sdk was deemed impractical. This would require a whole new set of #ifdefs, a major refactor of the roctracer code, and additional build support. Even then, only one could be active at a time (and you wouldn't want both active).
In homage to the abandoned refactor, RocLogger.cpp/h were created to contain the rocprofbase classes and the api filter.
Roctracer has no established end date.
Rocprofiler-sdk is in rocm_3.1 forward.
This will create a dependency where (newest kineto on old rocm) and (old kineto on newest rocm) could fail to build with AMD gpu support. That window is already over 1 year wide.