transformations: Split varith into neighbour and own data across csl_stencil regions #3307
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
n-io
changed the title
50 tasks
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #3307 +/- ##
==========================================
- Coverage 90.06% 90.04% -0.02%
==========================================
Files 446 446
Lines 56382 56401 +19
Branches 5409 5417 +8
==========================================
+ Hits 50779 50788 +9
- Misses 4177 4183 +6
- Partials 1426 1430 +4 ☔ View full report in Codecov by Sentry. |
EdmundGoodman
pushed a commit
to EdmundGoodman/xdsl
that referenced
this pull request
Dec 6, 2024
…stencil regions (xdslproject#3307) Background: The `csl_stencil.apply` op does communicate-and-compute on a given stencil buffer. It holds two regions, one for processing chunks of neighbour data (of this one buffer only), and one region for processing everything else after the exchange is done. The `convert-stencil-to-csl-stencil` pass splits the computation of the `stencil.apply` op across these two regions. The split was done in two steps, first re-ordering arith ops in the `RestructureSymmetricReductionPattern`, and then calling the `get_ops_split` function on the re-shuffled arith ops. Intuitively, the re-order pass would identify chained reductions (`arith.addf`, `arith.mulf`) and restructure them such that all neighbour data which should end up in the first region is consumed first, and the chained arith ops become easily splittable. This PR replaces this logic by converting arith to varith, splitting the varith op into neighbour/other data in the `SplitVarithOpPattern` rewrite, and then proceeding with `get_ops_split` and everything else as before. At the end, varith is converted back to arith. Minor improvements: * Constants are now always duplicated and appear on both regions, which `dce` can clean up --------- Co-authored-by: n-io <[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.
Background: The
csl_stencil.applyop does communicate-and-compute on a given stencil buffer. It holds two regions, one for processing chunks of neighbour data (of this one buffer only), and one region for processing everything else after the exchange is done. Theconvert-stencil-to-csl-stencilpass splits the computation of thestencil.applyop across these two regions.The split was done in two steps, first re-ordering arith ops in the
RestructureSymmetricReductionPattern, and then calling theget_ops_splitfunction on the re-shuffled arith ops. Intuitively, the re-order pass would identify chained reductions (arith.addf,arith.mulf) and restructure them such that all neighbour data which should end up in the first region is consumed first, and the chained arith ops become easily splittable.This PR replaces this logic by converting arith to varith, splitting the varith op into neighbour/other data in the
SplitVarithOpPatternrewrite, and then proceeding withget_ops_splitand everything else as before. At the end, varith is converted back to arith.Minor improvements:
dcecan clean upNote:
This PR currently has a fix from transformations: (arith-to-varith) Support more cases #3330 merged, will revert changes toconvert-arith-to-varith.mlirandvarith_transformations.pyas they are not intended to be merged from this PR