use correct wavefrontsize for device-side rng #822
+19
−14
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
Using the wrong wavefrontsize here technically could cause non-deterministic results even with a fixed seed. This is because mutiple threads inside the same wavefront are writing different values to the same memory location non-atomically. In CUDA, which value is actually stored is undefined and I am assuming ROCm makes similar assumptions, though I didn't find it documented anywhere. Testing with RDNA3.5 in wavefront64 mode though, I wasn't able to produce non-deterministic results from this, so it might not be an issue in practice. Still, I believe it's good practice to not rely on this being deterministic. One issue that remains is that we are allocating 32 counters and keys, even though 16 would be sufficient for wavefront64 targets. I didn't find a good way to emit these globals in a target-specific way, any suggestions welcome! closes JuliaGPU#819
test rng with wavefrontsize64 on RDNA as well
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Outdated suggestions cannot be applied.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.
Using the wrong wavefrontsize here technically could cause non-deterministic results even with a fixed seed. This is because mutiple threads inside the same wavefront are writing different values to the same memory location non-atomically. In CUDA, which value is actually stored is undefined and I am assuming ROCm makes similar assumptions, though I didn't find it documented anywhere. Testing with RDNA3.5 in wavefront64 mode though, I wasn't able to produce non-deterministic results from this, so it might not be an issue in practice. Still, I believe it's good practice to not rely on this being deterministic.
One issue that remains is that we are allocating 32 counters and keys, even though 16 would be sufficient for wavefront64 targets. I didn't find a good way to emit these globals in a target-specific way, any suggestions welcome!
closes #819