The verified-dse programme searches ABC logic-synthesis pass sequences for open-source chip flows (Yosys, ABC and OpenROAD on SKY130). Its rule is simple. Time one point. If routing a point costs more than 3 times scoring its mapped netlist, score every candidate on the mapped netlist and route only the best 8. Otherwise, route randomly chosen candidates for the same wall-clock.
The rule had one known failure: multipliers at a relaxed clock, with 6 to 13% regret. The last step looked for a statistic that flags those cells. Its pre-registered guard failed on fresh data (previous post).
Why the failure happens
The same 8 candidates that fail at the relaxed clock succeed at the tight one, so the ranking alone does not explain the failure. What changes between the clocks is the routed result. At a relaxed clock OpenROAD repairs timing up to the period and stops there. On every relaxed multiplier in the 52 held-out cells, the median routed delay is between 0.998 and 1.000 of the clock period. So routed area x delay comes out as area x period, and it ranks candidates by area alone. The mapped score is mapped area x mapped delay, and it still rewards mapped speed the routed design never keeps.
The fix
Score a mapped point the way the router will: mapped area x max(period, 1.25 x mapped delay). The 1.25 allows for wire delay the mapped netlist cannot see. It was chosen from six values on the 52 held-out cells.
The blind test
The test ran on seven designs no earlier table had touched: 14- and 26-bit multipliers, 16- and 24-bit squarers, a 32-port crossbar and two GF(2) matrices. We committed the design list, the analysis code and six predictions before the flow ran (1,680 Yosys and OpenROAD runs, about an hour on 4 cores).
- Mean regret fell from 2.73% to 0.83%. The clipped score was better in 5 cells and worse in none. The design-bootstrap 95% interval of the difference is -3.63 to -0.41 points.
- Every relaxed cell that runs the ladder reached 0.00%. The 26-bit multiplier fell from 9.58% and the 24-bit squarer from 11.30%. Tight cells were unchanged.
- All six predictions held. That includes the mechanism: routed delay is 0.993 to 1.000 of the period on every new relaxed multiplier and squarer.
Where it can hurt
The clip assumes routing lands near the clock. A 128-bit barrel shifter routes at 0.40 of its relaxed period. On a table we looked at while exploring (so not a test), the clip took that cell from 0.00% to 19.2% regret. With whole design families held out, the selection shows no gain, because all of the gain comes from one family. The evidence for the clip is the blind table and the mechanism. The next step is a check the flow already provides: the first routed point's delay as a share of the period, which is about 1 where the clip is right.
Code, tables and tests: experiments/clip_score/.