[opt] Add pass eliminate_immutable_local_vars #6926
Merged
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
✅ Deploy Preview for docsite-preview ready!
To edit notification comments on pull requests, go to your Netlify site settings. |
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
strongoier
requested review from
ailzhang,
jim19930609,
lin-hitonami and
bobcao3
bobcao3
reviewed
Dec 20, 2022
bobcao3
reviewed
Dec 20, 2022
quadpixels
pushed a commit
to quadpixels/taichi
that referenced
this pull request
May 13, 2023
Issue: taichi-dev#6933 ### Brief Summary There are many redundant copies of local vars in the initial IR: ``` <[Tensor (3, 3) f32]> $128 = [$103, $106, $109, $112, $115, $118, $121, $124, $127] $129 : local store [$100 <- $128] <[Tensor (3, 3) f32]> $130 = alloca $131 = local load [$100] $132 : local store [$130 <- $131] <[Tensor (3, 3) f32]> $133 = alloca $134 = local load [$130] $135 : local store [$133 <- $134] <[Tensor (3, 3) f32]> $136 = alloca $137 = local load [$133] $138 : local store [$136 <- $137] // In fact, `$128` can be used wherever `$136` is loaded. ``` These can come from many places; one of the main sources is the pass-by-value convention of `ti.func`. The consequence is that the number of instructions is unnecessarily large, which significantly slows down compilation. My solution here is to identify and eliminate such redundant instructions in the first place so all later passes can take a much smaller number of instructions as input. These redundant local vars are essentially immutable ones - they are assigned only once and only loaded after the assignment. In this PR, I add an optimization pass `eliminate_immutable_local_vars` as the first pass. (P.S. The type check processes of `MatrixExpression` and `LocalLoadStmt` are fixed by the way to make the pass work properly.) Let's study the effects in two cases: taichi-dev#6933 and [voxel-rt2](https://github.com/taichi-dev/voxel-rt2/blob/main/example7.py). First, let's compare the number of instructions after `scalarization` pass (which happens immediately after the first pass). | Kernel | Before this PR | After this PR | Rate of decrease | | ------ | ------ | ------ | ------ | | `test` (taichi-dev#6933) | 45859 | 26452 | 42% | | `spatial_GRIS` (voxel-rt2) | 48519 | 17713 | 63% | Then, let's compare the total time of `compile()`. | Case | Before this PR | After this PR | Rate of decrease | | ------ | ------ | ------ | ------ | | taichi-dev#6933 | 20.622s | 8.550s | 59% | | voxel-rt2 | 27.676s | 9.495s | 66% | Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
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.
Issue: #6933
Brief Summary
There are many redundant copies of local vars in the initial IR:
These can come from many places; one of the main sources is the pass-by-value convention of
ti.func. The consequence is that the number of instructions is unnecessarily large, which significantly slows down compilation.My solution here is to identify and eliminate such redundant instructions in the first place so all later passes can take a much smaller number of instructions as input. These redundant local vars are essentially immutable ones - they are assigned only once and only loaded after the assignment. In this PR, I add an optimization pass
eliminate_immutable_local_varsas the first pass.(P.S. The type check processes of
MatrixExpressionandLocalLoadStmtare fixed by the way to make the pass work properly.)Let's study the effects in two cases: #6933 and voxel-rt2.
First, let's compare the number of instructions after
scalarizationpass (which happens immediately after the first pass).test(#6933)spatial_GRIS(voxel-rt2)Then, let's compare the total time of
compile().