Coordination on scan-chain optimization / possible scan_opt research contribution

[wellitsabhinav]

Hi OpenROAD DFT maintainers,

I am currently developing a research prototype called PACT around physical- and test-activity-aware scan-chain optimization, and I wanted to coordinate before taking the work further toward OpenROAD.

The current prototype treats scan ordering/architecture as a bi-objective optimization problem between a placement-aware scan-path cost and an ATPG-derived spatial shift-activity metric. The implementation is currently external to OpenROAD and uses sparse candidate neighborhoods, exact incremental physical-cost updates, exact activity evaluation, and bounded Pareto search.

I am aware that physical-aware and power-aware scan reordering are established research areas, so I am not assuming that the general idea itself is new. My current research question is around the combination of physical cost, cycle-resolved spatial shift activity, and scalable Pareto exploration.

Before taking this further, I wanted to ask:

1. Is anyone currently implementing or planning scan-chain reordering / scan_opt?
2. Is there ongoing UCSD or other work that substantially overlaps with this direction?
3. If a placement-aware scan optimizer were eventually contributed, what scope/interface would you prefer for the first implementation?
4. Would it make more sense to keep ATPG/activity-aware optimization external initially and first upstream only general physical scan-reordering infrastructure?

I am happy to share a short design note and current experimental results if useful. My main goal is to avoid duplicating ongoing work and to make sure the research architecture does not evolve in a direction incompatible with OpenROAD.

Thanks,
Abhinav Choudhary

[maliberty]

There are already two reordering PRs #10176 and #10666. A third one would need to be substantially better to be worth opening.

I don't know enough of what 'ATPG/activity-aware optimization" and "ATPG-derived spatial shift-activity metric" mean to comment on those.

[maliberty]

What is goal/concern that considering simultaneously switching scan addresses?

[wellitsabhinav]

The main concern I’m looking at is not simultaneous switching itself, since that is normal during scan shift. It is whether a lot of switching happens at the same time in the same physical region of the chip. If scan cells with high transition activity are placed close together and toggle in the same shift cycles, that could create more localized current demand and potentially make test-mode IR drop or power hotspots worse.

So the idea I’m exploring is whether scan reordering can consider both physical distance and ATPG-derived shift activity, instead of optimizing mainly for scan wirelength.

[maliberty]

I would be surprised if scan mode introduces more IR-drop than non-scan. It sounds more like a research area than a production capability at this point.

[wellitsabhinav]

Thanks for the earlier feedback!! I wanted to share an update, since I’ve been able to test the idea more concretely since this discussion.

I agree that it would be too strong to frame this as evidence that scan shifting necessarily creates a worse IR-drop problem than functional operation. I’ve therefore kept the current work focused on a narrower question:

Can scan-chain ordering trade a small amount of physical interconnect cost for lower overall and spatially concentrated scan-shift switching activity?

I’ve now evaluated this end-to-end on two larger previously unseen benchmark circuits. The candidate scan orders were physically implemented and routed, and then evaluated using the same ATPG patterns and activity measurements.

On one circuit, a candidate reduced the overall activity-related metric by about 0.3%, while reducing localized switching-hotspot metrics by roughly 2.4% and 1.0% at two spatial neighborhood scales. The routed scan-path cost increased by about 1%, with no routing violations and no loss in fault coverage.

The second circuit showed a more mixed result: some orderings reduced overall or localized activity while making another activity metric slightly worse. That was useful in itself, because it suggests this is genuinely a multi-objective trade-off rather than simply another way of minimizing scan wirelength.

The main observation so far is therefore:

The scan order with the lowest physical connection cost is not necessarily the one with the best scan-shift activity distribution, and there appear to be useful intermediate points between those objectives.

I’m now testing whether this continues to hold on additional, structurally different designs, and comparing it against reproducible open-source baselines such as:

• physical-only ordering,
• activity-only ordering,
• simple combined physical/activity heuristics,
• and OpenROAD’s physical scan-ordering flow.

I still see this primarily as a research question rather than a production scan_opt feature at this stage.

If the behavior generalizes, though, I’d be interested in your opinion on whether the eventual OpenROAD scan-reordering infrastructure could benefit from an extensible cost/objective interface, so that physical wire cost is not necessarily the only quantity available to a reordering algorithm.

[maliberty]

What is the practical impact of the 2.4% reduction? Does this translate in IR-drop or thermal hotspot improvement?