The verified-dse programme searches over ABC logic-synthesis pass sequences in an open chip flow (Yosys, ABC and OpenROAD on SKY130). Its working rule ranks every candidate by a cheap, clipped score from the mapped netlist and pays for place-and-route only on the best 8. On most designs that beats spending the same time routing random candidates. On some wide shifters it loses badly.
For seven steps we tried to predict those shifters from the RTL: the shift operator, the share of multiplexers, and how many operations or gates sit in front of the shift amount. Each rule was placed on designs we had already seen, and each failed or moved on the next blind test. The last one found the same family winning at 28 gates and losing at 36.
So this step stopped guessing and measured. The idea: route 4 randomly drawn candidates in full, check whether the cheap score puts them in the same order as the routed result does (Kendall tau-b), and trust the cheap ranking only if it does. Otherwise spend the rest of the budget on random candidates. Every rule spends the same wall-clock. We wrote the rule and six predictions into the repository before computing anything, then replayed it on all 90 cells of the last seven steps. No new flow run was needed.
| rule | mean regret over 90 cells |
|---|---|
| best of "ranking" or "random" per cell (oracle, a floor) | 2.42% |
| gate-depth guard (placed on most of these cells) | 3.91% |
| cheap ranking, no guard | 4.77% |
| spot-check guard, 4 random probes | 5.14% |
| equal-cost random routing | 5.44% |
Regret is how far the best routed candidate a rule finds sits above the best in the pool. Three of the six predictions held; the three that mattered failed. The guard is worse than no guard, and it hardly ever backs off where it should: it falls back to random in 12% of draws on the 17 cells where the ranking loses, and in 7% on the 40 where it wins.
The reason showed up when we looked afterwards (so treat it as a description, not a test). Over all of a cell's candidates, the cheap score's order agrees with the routed order about as well where the ranking fails (tau-b +0.45) as where it works (+0.53). The difference is at the top. Among the cheap score's own 16 best candidates, the tau-b is -0.18 on failing cells and +0.23 on winning ones. The ranking is fine in general and wrong exactly where we use it. Four random probes almost never land in that top slice, so they cannot see the problem.
That matches an older result in the same programme: an unclipped version of this check failed the same way in milestone 111. Two negative results close off random spot checks. The next step is to check the top: route a few of the ranking's own best candidates next to a few random ones, and trust the ranking only if its picks win. We will register that rule before replaying it.
Code, the full per-cell table and the log: experiments/pilot_rank/, LOG.md (milestone 126). The results come from one tool version, one process corner and replayed cells we had seen before; a blind test is still to come.