[SYCL] Fix a bug when using no device split and reqd_work_group_size #16236
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
sarnex
reviewed
Dec 2, 2024
| @@ -152,7 +152,8 @@ std::optional<T> getKernelSingleEltMetadata(const Function &Func, | |||
|
|
|||
| PropSetRegTy computeModuleProperties(const Module &M, | |||
| const EntryPointSet &EntryPoints, | |||
| const GlobalBinImageProps &GlobProps) { | |||
| const GlobalBinImageProps &GlobProps, | |||
| module_split::IRSplitMode SplitMode) { | |||
There was a problem hiding this comment.
i would prefer we didnt add another argument to this function, can we instead use module metadata like we do for for sycl/esimd split (saveSplitInformationAsMetadata)?
There was a problem hiding this comment.
Done, I just added a way to remove properties from the PropertySetRegistry and removed it in sycl-post-link.cpp
sarnex
reviewed
Dec 2, 2024
| // remove the required work group size metadata in this case. | ||
| if (SplitMode == module_split::SPLIT_NONE) | ||
| PropSet.remove(PropSetRegTy::SYCL_DEVICE_REQUIREMENTS, | ||
| "reqd_work_group_size_uint64_t"); |
There was a problem hiding this comment.
should we make reqd_work_group_size_uint64_t a constant somewhere so we dont have any typo problems in the future?
maarquitos14
reviewed
Dec 3, 2024
| // When the split mode is none, the required work group size will be added | ||
| // to the whole module, which will make the runtime unable to | ||
| // launch the other kernels in the module that have different | ||
| // required work group sizes or no requried work group sizes. So we need to |
There was a problem hiding this comment.
Suggested change
| // required work group sizes or no requried work group sizes. So we need to | |
| // required work group sizes or no required work group sizes. So we need to |
maarquitos14
approved these changes
Dec 4, 2024
sarnex
approved these changes
Dec 4, 2024
| // This test checks that with -fsycl-device-code-split=off, kernels | ||
| // with different reqd_work_group_size dimensions can be launched. | ||
|
|
||
| // RUN: %{build} -fsycl -fsycl-device-code-split=off -o %t.out |
There was a problem hiding this comment.
nit: %{build} includes -fsycl
sommerlukas
pushed a commit
that referenced
this pull request
Mar 14, 2025
Incorporates recent changes to `sycl-post-link` into the RTC-specific version: - #16729 - #16236 - #17211 --------- Signed-off-by: Julian Oppermann <[email protected]>
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.
There was a bug (#13523) where a kernel couldn't be launched when
-fsycl-device-code-split=offwas used and multiple kernels with different required work group sizes were present. This issue was fixed by ensuring that the required work group size metadata is not attached to the device image when multiple required work group sizes are detected in a single module.However, there was a similar but related case that was not fixed by that PR, which is now demonstrated in the new test no-split-reqd-wg-size-2.cpp. This issue occurs when there is a single kernel with a required work group size and another kernel without one. In this case, the module doesn't contain multiple required work group sizes, so the required work group size metadata is still attached. As a result of the metadata being attached, the runtime cannot launch the kernel without a required work group size.
This PR removes the logic of ensuring metadata is not attached when there are multiple required work group sizes, and instead adds logic that ensures the metadata is not attached when the split mode is
SPLIT_NONE. This covers the old cases from the previous PR and the new case in this PR.