Skip to content

[WIP, do not review] Polling-based asynchronous verbose mode for execution time-tracking#5102

Draft
avmanerikar wants to merge 8 commits intomainfrom
amanerik/main/async-verbose-mode-polling
Draft

[WIP, do not review] Polling-based asynchronous verbose mode for execution time-tracking#5102
avmanerikar wants to merge 8 commits intomainfrom
amanerik/main/async-verbose-mode-polling

Conversation

@avmanerikar
Copy link
Copy Markdown
Contributor

@avmanerikar avmanerikar commented Apr 30, 2026

Description

This PR is an alternative approach to the callback-based asynchronous verbose mode implemented in PR #4187 - the method relies on a periodic event polling to record and log profiling info. The motivation to using this approach is to have a unified implementation for the asynchronous mode that does not rely on runtime-specific callback APIs.

The implementation addresses points that the callback-based approach fails to triage:

  • Event-polling does not rely on runtime-specific callback API (required for supporting L0 runtime).
  • The callback method does not triage MFDNN-14331 for SYCL graphs.
Failing reproducer for SYCL Graph:
DNNL_VERBOSE=1 .\tests\benchdnn\benchdnn.exe --eltwise --engine=gpu --execution-mode=graph 1
  • Extension to CPU Threadpool is difficult with callback-based method.

This prototype is implemented for OpenCL GPU runtime.

@avmanerikar avmanerikar requested review from a team as code owners April 30, 2026 18:19
@avmanerikar avmanerikar marked this pull request as draft April 30, 2026 18:19
@github-actions github-actions Bot added platform:gpu-intel Codeowner: @oneapi-src/onednn-gpu-intel component:common labels Apr 30, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

component:common platform:gpu-intel Codeowner: @oneapi-src/onednn-gpu-intel

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant