Fixes for 50-series GPUs (CUDA backend) #152
Merged
Conversation
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
merlinND
commented
Jun 12, 2025
merlinND
force-pushed
the
fix-cuda-50-series
branch
from
3d8744e to
5ac6f14
Compare
Fix usage of the `.v2` variant. Prefer the `.f16x2` variant for `.add` since it is more broadly supported.
We probably do not need volatile stores in order to guarantee progress in decoupled lookback-style algorithms.
merlinND
force-pushed
the
fix-cuda-50-series
branch
from
3603c28 to
29113be
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
Reductions
Somewhat related to #151.
For f16 or f16x2, only.addis supported by theredinstruction.Mark other reduction operations as unsupported. Note this means that DrJit will now raise an exception in cases where it was previously "fine".TODO: actually raise an exception in the unsupported cases, right now it's triggering a confusing error message.Edit: @wjakob implemented the scatter reduce operations in a difference way that is actually supported (included in this PR).
Compress / prefix reduction
Symptom: on 50-series GPUs, any usage of
dr::compress()with an input of size larger than 4096 would hang with 100% GPU usage.After looking at this Cub implementation, it seems that our implementations of the decoupled-lookback pattern in the Compress and Prefix Reduction kernels needed a
__threadfence_block().Edit: @wjakob proposed a different implementation using
volatilestore / loads, which is more targeted.Testing
On a 5080 with driver 575.57.08:
test_graphvizbut that's unrelated and already failing onmaster)