NVIDIA/Fuser

Pointwise heuristics incorrect memory access estimation

Open

#4,773 opened on Jul 11, 2025

View on GitHub
 (0 comments) (0 reactions) (0 assignees)C++ (80 forks)github user discovery
good first issue

Repository metrics

Stars
 (392 stars)
PR merge metrics
 (PR metrics pending)

Description

Hmm... I don't think the original code of this part makes sense. The unit of cur_transfer_size and right_transfer_size were in the unit of byte ^ N, where N is the number of dimensions on the left/right of the break point. My PR will change it to bit ^ N, so in this sense, my PR is NOT a no-op. But I believe we should not block my PR because of this. Instead, we should fix it in a later PR.

Originally posted by @zasdfgbnm in https://github.com/NVIDIA/Fuser/pull/4748#discussion_r2199798284

Contributor guide