[Synth][CutRewriter] Migrate CutRewriter to use Flat IR#9805
Merged
Conversation
uenoku
commented
Mar 2, 2026
Comment on lines
+940
to
+941
| } else if (numFanins == 2) { | ||
| // Two-way merge (common case for AND gates) |
Member
Author
There was a problem hiding this comment.
Profile showed this merge sort is one of the critical path so these parts are manually written instead of relying on STL.
uenoku
changed the title
uenoku
force-pushed
the
dev/uenoku/cut-change
branch
from
41c4577 to
644a4ec
Compare
uenoku
force-pushed
the
dev/uenoku/cut-change
branch
from
644a4ec to
774d609
Compare
fabianschuiki
approved these changes
Mar 3, 2026
Contributor
fabianschuiki
left a comment
fabianschuiki
left a comment
There was a problem hiding this comment.
This is really really cool! Amazing speedup, and great to see that the cut rewriting is in the same ballpark as ABC! 🥳
uenoku
added a commit
that referenced
this pull request
Mar 3, 2026
…work (#9804) This commit introduces a data structure which reduces CutRewriter run time quite a bit (see benchmark result in #9805). It is well known that MLIR is too slow when naively used for specific logic synthesis transformation like Cut rewriting, and previous implementation was impractical when cut input size > 4. For these tasks, a cache-friendly flat IR data structure (like those in ABC or mockturtle) is much more efficient. This PR changes CutRewriter to first create a read-only, in-memory IR to be used by the CutRewriter. It allows 2-input AND, 3-input Majority, and 2-input XOR to coexist in the same data structure, which lets us represent AIG, MIG, XOR-AND, and XOR-MAJ together. This is slight different from ABC (AIG), Mockurtle's template based implementation (AIG/MIG/XOR-MAJ/XOR-AND). This PR introduces only data structure, test changes are included in #9805 AI-assisted-by: Claude Sonnet 4.6, Opus 4.6
Switch cut enumeration and truth table handling to LogicNetwork indices, update downstream mappers to consume the new cut API, and refresh Synth tests for the new ordering/operation support without area-recovery changes.
uenoku
force-pushed
the
dev/uenoku/cut-change
branch
from
6be9729 to
4662d46
Compare
Contributor
|
Amazing stuff @uenoku - sorry I was too slow to review haha - we've been running circt-synth on some really large testcases and my collaborators have complained about speed so hopefully this might offer some nice benefits right away |
Member
Author
|
No worry! Thank you for taking a look! Do your collaborators have performance issues around LUT mapping? If so this PR will immediately improve the compilation time. Otherwise I'm happy to investigate regression. |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Switch cut enumeration and truth table handling to LogicNetwork indices, update downstream mappers to the new cut API, and refresh Synth tests for the updated ordering and operations. PR staked on #9804.
This is NFCI except for a small bug fix in NPNClass utils; only operand orders are reordered.
The following is a runtime comparison of LUT mapping for
hyptest case (0.2M AIG nodes, largest test case of non MtM category) in EPFL Combinational Benchmark Suite. LUT mapping is a good benchmark to check cut enumeration performance since only cut enumeration is generally performed. Note that this is not a perfectly fair comparison to ABC because ABC performs more optimizations like area recovery or applying several strategies for cut selection. When k=6 is generally same, for k > 6 we are missing fast computation for truth table (since we cannot represent in single 64 bit integer) so it's reasonable. At least it got 8~10 times faster from the previous implementation.before:
after:
AI-assisted-by: Claude Sonnet 4.6, Opus 4.6