Correct reduce-then-scan -O0 workaround behavior with different host and device optimization levels #2133
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
mmichel11
requested review from
akukanov,
dmitriy-sobolev,
SergeyKopienko,
danhoeflinger and
adamfidel
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce_then_scan.h
Outdated
Show resolved
Hide resolved
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce_then_scan.h
Outdated
Show resolved
Hide resolved
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce_then_scan.h
Outdated
Show resolved
Hide resolved
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce_then_scan.h
Outdated
Show resolved
Hide resolved
include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_reduce_then_scan.h
Outdated
Show resolved
Hide resolved
…ismatch Signed-off-by: Matthew Michel <[email protected]>
…ithin kernel Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
The query of the nd-range in the function call operator caused a perf regression on NVGPUs. Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
- Adds a function to check device vendor IDs and return candidate sub-group sizes to the host compiler. - Use _ONEDPL_DETECT_SPIRV_COMPILATION along with _ONEDPL_DETECT_COMPILER_OPTIMIZATIONS_ENABLED to choose the sub-group size during device compilation. Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
…orkaround_sg_sz Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
mmichel11
force-pushed
the
dev/mmichel11/remove_sg_sz_template_rts
branch
from
3ccd466 to
b0a7eaf
Compare
danhoeflinger
approved these changes
Apr 4, 2025
|
Will run through internal CI once more to confirm there are no issues and will then merge. |
timmiesmith
pushed a commit
that referenced
this pull request
Jun 9, 2025
…and device optimization levels (#2133) The host and device compiler may use different optimization levels causing the value of _ONEDPL_DETECT_COMPILER_OPTIMIZATIONS_ENABLED to differ between host and device code. This patch corrects this issue in the reduce-then-scan sub-group size workaround to ensure correctness and avoid missing kernel runtime exceptions. --------- Signed-off-by: Matthew Michel <[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.
#2046 introduced a workaround for a hardware bug that prevents sub-group sizes of 32 from being used with
-O0compilation on certain devices based on the detection of optimization via the__OPTIMIZE__macro. However, this approach does not work when the host and device compiler use different optimization levels, specifically when only one of the two phases uses-O0compilation. The integral sub-group size constant is embedded in the kernel name causing the host to call a kernel not compiled by the device which results in "Kernel not found" errors.To support these cases where the host and device compiler optimization levels may differ, we need to make the following changes:
__num_sub_groups_local,__num_sub_groups_global, etc.) on the device itself as the true sub-group size can only be determined when on the device.