Add Logging for Empty Traces
Summary: Recently, we have had users seen empty traces when the system is idle leading to confusion as to whether it was caused by a bug in kineto formatting or not. This diff adds further logging to ensure that the user is aware that kineto finds no valid trace events and is expecting an empty trace.
Differential Revision: D60311331
This pull request was exported from Phabricator. Differential Revision: D60311331
This pull request was exported from Phabricator. Differential Revision: D60311331
This pull request was exported from Phabricator. Differential Revision: D60311331
This pull request was exported from Phabricator. Differential Revision: D60311331
Should we also add the functions into roctracer's handleactivity functions?
@aaronenyeshi Sure, can we do that as a followup diff instead?
Should we also add the functions into roctracer's handleactivity functions?
@aaronenyeshi Sure, can we do that as a followup diff instead?
yes that sounds good
This pull request was exported from Phabricator. Differential Revision: D60311331
This pull request was exported from Phabricator. Differential Revision: D60311331
This pull request was exported from Phabricator. Differential Revision: D60311331
Should we also add the functions into roctracer's handleactivity functions?
@aaronenyeshi Sure, can we do that as a followup diff instead?
yes that sounds good
Looking at it now, I think this already covers the roctracer impl. handleGpuActivity with ItraceActivity as an arg is shared between CUPTI and Rocm, and the Rocm handleRuntimeActivity is also covered in this diff. We will see for sure when the next diff tests on all architectures though
This pull request has been merged in pytorch/kineto@da2f2682cabaf95d601fa2a9b7e0979f84fe7667.