Proof-of-concept: parallelize argmin #9066
Merged
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
This was referenced Mar 11, 2025
manuelcandales
approved these changes
Mar 12, 2025
kedarnath03
pushed a commit
to kedarnath03/executorch
that referenced
this pull request
Jun 25, 2025
I attempted to port `at::parallel_reduce` to ExecuTorch and use that in reduce_util.h, but it turned out to be much trickier than expected. (In brief: parallel reduction requires two steps: 1) split the input range into chunks and reduce over them (easily done like parallel_for), and then 2) combine the sub-results from chunks. The reduction function accepted by reduce_over_dim is not well-suited to step (2).) Instead, I ported the parallelization strategy used by binary_kernel_reduce_lastdim: just parallelize over the *non*-reduced dimensions of the tensor. I don't understand why this strategy isn't generally applicable and we aren't otherwise capable of parallelizing reductions, so I haven't gated it to the case where we are reducing over a contiguous last dimension. I will send a follow-up that packages up this strategy nicely and uses it in our reduction portable ops. ghstack-source-id: e4efdab ghstack-comment-id: 2707761383 Pull Request resolved: pytorch/executorch#9066
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Outdated suggestions cannot be applied.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.
I attempted to port
at::parallel_reduceto ExecuTorch and use thatin reduce_util.h, but it turned out to be much trickier than expected.
(In brief: parallel reduction requires two steps: 1) split the input
range into chunks and reduce over them (easily done like
parallel_for), and then 2) combine the sub-results from chunks. The
reduction function accepted by reduce_over_dim is not well-suited to
step (2).)
Instead, I ported the parallelization strategy used by
binary_kernel_reduce_lastdim: just parallelize over the non-reduced
dimensions of the tensor. I don't understand why this strategy isn't
generally applicable and we aren't otherwise capable of parallelizing
reductions, so I haven't gated it to the case where we are reducing
over a contiguous last dimension.
I will send a follow-up that packages up this strategy nicely and uses
it in our reduction portable ops.