Speed up HoistPlaintextOps #485
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
j2kun
force-pushed
the
hoist-plaintext-speed
branch
from
61a93ff
to
afc1d4a
Compare
j2kun
force-pushed
the
hoist-plaintext-speed
branch
2 times, most recently
from
ec62a43
to
b2f7a4c
Compare
j2kun
force-pushed
the
hoist-plaintext-speed
branch
from
b2f7a4c
to
977eac6
Compare
asraa
approved these changes
Mar 6, 2024
Comment on lines
+577
to
582
| for (Operation &op : opRange) { | ||
| if (canHoist(op)) { | ||
| opsToHoist.push_back(&op); | ||
| hoistedAny = true; | ||
| } | ||
| } |
There was a problem hiding this comment.
nit: you could consider llvm::make_filter_range here
j2kun
added
the
pull_ready
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Outdated suggestions cannot be applied.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.
Fixes #479
The two main problems:
HoistPlaintextOpsusedstd::find_ifto scan the entire generic body, only to extract a single op, leading to a quadratic scan when there were a linear number of ops to hoist. When unrolling an affine.for, each iteration gets an affine.apply op, and they can all be lifted, so it nearly always causes quadratic behavior.extractOpBeforeGeneric, the newly-created single-op generic was cloning all of the operands from the original generic op as its operands, even if it had zero secret operands. When lifting ~1k ops, this became a problem because the output of each single-op generic is added as an operand to the originalgenericit was lifted from, so we got into a situation like:After that, we have a quadratic number of unused args, so we apply a quadratic number of EraseUnusedGenericArg patterns. By more finely tracking the subset of operands needed in the new generic, we can reduce this to (# ops to lift * # operands per lifted op), which is minimal. The only tricky bit is that, if the op being extracted is a loop nest, you have to walk the op's subtree to find all the generic arguments it depends on.
The test I added ran in 15s with the bug still in place, and after these fixes it runs in 0.5s.