SYCL: Improve and simplify parallel_scan implementation #6064
Conversation
This file contains 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
masterleinad
force-pushed
the
sycl_improve_parallel_scan_new
branch
from
e3c9427
to
6cc246a
Compare
masterleinad
commented
Apr 18, 2023
masterleinad
force-pushed
the
sycl_improve_parallel_scan_new
branch
from
3354928
to
e07be73
Compare
masterleinad
changed the title
|
Requires #6065. |
dalg24
reviewed
Apr 25, 2023
masterleinad
force-pushed
the
sycl_improve_parallel_scan_new
branch
from
2144d9b
to
bdaa12c
Compare
crtrott
approved these changes
May 4, 2023
| (global_range + max_subgroup_size - 1) / max_subgroup_size; | ||
|
|
||
| const auto local_range = sg.get_local_range()[0]; | ||
| const int local_range = sg.get_local_range()[0]; |
There was a problem hiding this comment.
did SYCL change so that you need int now?
There was a problem hiding this comment.
No, it's not necessary to do that but I have seen that SYCL is very sensitive to 64-bit index calculations, and local indices surely will never require that.
There was a problem hiding this comment.
Just to be clear, auto deduces size_t, not int. I agree we don't need that much of an index range.
nliber
approved these changes
May 8, 2023
dalg24
reviewed
May 9, 2023
|
The SYCL build passed. Ignoring the rest. |
Closed
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Outdated suggestions cannot be applied.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.
This pull request changes the previous recursive
parallel_scanin theSYCLbackend to a two-pass one as we use forCudaandHIPandSYCLreductions. This simplifies the code (also making it more uniform) and reduces the memory footprint (since we only need to store intermediate results for all items and group scans but not recursive group scans).On the way, I made sure that all local operations operate on indices of type
intavoiding 64-bit index operations.A second improvement is switching to an auto-detection of the work group size as we do for reductions by querying a dummy kernel for the maximum group size.
Finally, this fixes a couple of unit tests that were failing with
SYCL+Cudasince #5707.