Skip to content

Conversation

@m-brettell
Copy link
Contributor

Overhall of the seed extraction kernels:

Now extract all paths through the graph into path_store <- size can be counted in the CCA, so no more exceedable maximums for those allocations. This reduces both the allocation size and now only an int2 is passed in and out of memory while branching through the graph, rather than the large edgeState (80 bytes).

All paths are then fit to obtain a quality score for disambiguation rather than fitting only some at a time.
One bidding round selects only the highest quality seed starting at each edge. Then the same iterative bidding occurs to pick the best seeds as before. The change to the first round improves seed purity and high pT efficiency.

Reduces the seed extraction kernel times from ~25% to ~5% of kernel time.

@m-brettell m-brettell force-pushed the GBTS_new_seed_extraction branch from fb414a4 to 5da4446 Compare November 3, 2025 17:20
@sonarqubecloud
Copy link

sonarqubecloud bot commented Nov 3, 2025

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.

1 participant