Skip to content

Realm: pass indirection only on appropriate side of the transfer graph

apryakhin requested to merge apriakhin/fix-cuda-gather into master

Fix the bug uncovered by looking at the path planner output in this issue:

In find_fastest_path we should not pass in indirection (and thus pickup cuda-dma scatter/gather channel)..for scatter when we attempting to get the first hop since scatter should always be done close to the destination (e.g. either direct or last hop) for gather when attempting to get the last hop since gather should always be done close to the source (e.g. either direct or first hop).

Edited by apryakhin

Merge request reports