[Synth] Refactor and improve cut rewriter infrastructure #8909
Conversation
This commit refactors the cut rewriter and implements correct priority sort. Previously `depth` of a cut represents a logic depth *within* a cut but this is not correct. What the original paper (ICCAD 2007) proposed was to use arrival time of cut outputs instead. For tha this commit changes to perform cut matchings for each cut, and use worst arrival time as a priority of cuts. LUT6 depth for 16-bit multiplexier improved from 17 to 7, which is on per with Yosys/ABC.
uenoku
force-pushed
the
dev/hidetou/cut-improvement
branch
from
72a193c to
08c4d99
Compare
cowardsa
left a comment
cowardsa
left a comment
There was a problem hiding this comment.
LGTM - only comment on slightly improving documentation for heuristics chosen - perhaps this could also be reflected in dialect or synthesis documentation to make it clear what features are implemented and where the ideas come from?
| compareDelayAndArea(options.strategy, currentMatch->getArea(), | ||
| currentMatch->getArrivalTimes(), | ||
| bestCut->getMatchedPattern()->getArea(), | ||
| bestCut->getMatchedPattern()->getArrivalTimes())) { |
There was a problem hiding this comment.
Is the sorting criteria used above identical to the compareDelayAndArea?
There was a problem hiding this comment.
Great point! I modified the sorting function to use compareDelayAndArea, and change the second loop just to return the first matched one.
| auto resultNumber = cast<mlir::OpResult>(input); | ||
| inputArrivalTimes.push_back( | ||
| matchedPattern->getArrivalTime(resultNumber.getResultNumber())); | ||
| matchedPattern.getArrivalTime(resultNumber.getResultNumber())); |
There was a problem hiding this comment.
Nit: mixture of result and input is a bit confusing here?
uenoku
force-pushed
the
dev/hidetou/cut-improvement
branch
from
80981ee to
79ae9d7
Compare
uenoku
force-pushed
the
dev/hidetou/cut-improvement
branch
from
79ae9d7 to
a242f39
Compare
Great point, I'll write down a rational document for Synth dialect and cut rewriter in a follow up. |
This commit refactors the cut rewriter and implements correct priority sort metric for depth. Previously
depthof a cut represents a logic depth within a cut but this is not correct. What the original paper (ICCAD 2007) proposed was to use arrival time of cut outputs instead.For that this commit changes to perform cut matching for each cut, and use worst arrival time as a priority of cuts.
LUT6 depth for 16-bit multiplier improved from 17 to 7, which is on per with Yosys/ABC.