Fixes dpctl.tensor.round on CUDA devices#1700
Merged
ndgrigorian merged 1 commit intomasterfrom May 29, 2024
Merged
Conversation
…vidia hardware When compiled for CUDA, `std::rint` would incorrectly round values halfway between two integers toward 0, rather than to the nearest even number as required per array API. `sycl::rint` avoids such issues by not relying on the current rounding mode
|
Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. 🤞 |
|
Array API standard conformance tests for dpctl=0.18.0dev0=py310h15de555_29 ran successfully. |
oleksandr-pavlyk
approved these changes
May 29, 2024
Contributor
oleksandr-pavlyk
left a comment
oleksandr-pavlyk
left a comment
There was a problem hiding this comment.
It might be a good idea to scan the code base for remaining uses of std namespace transcendental functions and replace those one by one too
Collaborator
Collaborator
Author
I agree. I'll do this as a separate PR, but I think it's a good idea too. |
Contributor
|
I have opened gh-1701 for build break with LLVM SYCL compiler, it is unrelated to changes in this PR. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
When compiled for CUDA,
std::rintwould incorrectly round values halfway between two integers towards 0 (i.e.,1.5 -> 1.). The array API specification requires that these values be rounded to the nearest even integer instead.To resolve this,
std::rinthas been replaced withsycl::rint, which does not rely on the current floating-point rounding mode (see SYCL specification).As was pointed out at the time of implementation the floating-point rounding mode can vary between devices.