Skip to content

GPU: disable delayed output when sorting TPC clusters on the host - #150

Open
ktf wants to merge 1 commit into
davidrohr:devfrom
ktf:pr15318-metal-fix
Open

ktf wants to merge 1 commit into
davidrohr:devfrom
ktf:pr15318-metal-fix

Conversation

@ktf

@ktf ktf commented Nov 21, 2026

Copy link
Copy Markdown

With tpcSortClusters set, the clusterizer copies the clusters to the host and SortClusters() sorts them there and uploads them back over the GPU clusters. With delayedOutput, that GPU->host copy was only queued in mOutputQueue, so the host sorted and uploaded an unfilled buffer and the tracking found no tracks. Disable delayedOutput for tpcSortClusters, as already done for deterministicGPUReconstruction and debugLevel > 3.

With tpcSortClusters set, the clusterizer copies the clusters to the host
and SortClusters() sorts them there and uploads them back over the GPU
clusters. With delayedOutput, that GPU->host copy was only queued in
mOutputQueue, so the host sorted and uploaded an unfilled buffer and the
tracking found no tracks. Disable delayedOutput for tpcSortClusters, as
already done for deterministicGPUReconstruction and debugLevel > 3.
@ktf
ktf requested a review from davidrohr as a code owner November 21, 2026 14:06
@ktf

ktf commented Nov 21, 2026 •

Copy link
Copy Markdown
Author

@davidrohr this is the other part of the reproducibility fix (only one commit, actually, the other one was the enum you already picked up).

@davidrohr

Copy link
Copy Markdown
Owner

The fix is indeed correct, but this option is used only for the standalone async QC (which I never tried with GPU). Thanks for spotting. There is already a flag || GetProcessingSettings().deterministicGPUReconstruction, so it cannot affect the deterministic mode at all. Can you open a PR to O2, I merged the PR already.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants