Repository navigation
[OpenMP][offload] Grid size selection convergence - #5000
Merged
Merged
Conversation
ro-i
force-pushed
the
amd/dev/ro-i/ol-grid-conv
branch
from
October 16, 2026 20:05
fcffb20 to
876335e
Compare
Large reductions (> 32 B per team) get one block per CU, scalar ones 16 waves per CU, on every AMDGPU device.
…c code Default to 6 blocks/CU on AMDGPU, and trade threads for blocks while the distribute loop has iterations for them, as the AMDGPU plugin does.
Only big-jump-loop kernels keep AMDGPU-specific heuristics. Remove the AMDGPU knobs for the replaced ones: occupancy-based team counts, low trip count blocks/threads, ADJUST_TEAMS, and the xteam reduction knobs.
Give big-jump-loop kernels one thread per iteration, like No-Loop ones, capped at 1M blocks as before. This removes the AMDGPU launch geometry hook and the LIBOMPTARGET_AMDGPU_BIG_JUMP_LOOP_MAX_TOTAL_TEAMS envar.
ro-i
force-pushed
the
amd/dev/ro-i/ol-grid-conv
branch
from
October 20, 2026 14:04
876335e to
28f91ca
Compare
Author
|
This is the downstream part for converging grid size selection up- and downstream. For upstream, corresponding patches are planned and going to be posted as soon as this PR has landed downstream |
jhuber6
approved these changes
Oct 20, 2026
Collaborator
|
!PSDB |
Collaborator
|
also landing ROCm/aomp#2633 (review) |
ronlieb
approved these changes
Oct 21, 2026
ronlieb
added a commit
that referenced
this pull request
Oct 23, 2026
- restore downstream LIBOMPTARGET_AMDGPU_LOW_TRIPCOUNT and LIBOMPTARGET_BLOCKS_FOR_LOW_TRIP_COUNT - restore handling of short xteam reductions that don't use the whole device to be handled like regular SPMD loops - restore handling of no-loop and big-jump-loop kernels that don't use the whole device so that they get less threads and more blocks, like regular SPMD kernels --------- Co-authored-by: theRonShark <rlieberm@amd.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.