Asking about the runtime of power activity propagation

[Zhihui-SJTU]

@jjcherry56 @maliberty Will using CUDD in power activity propagation result in higher accuracy but worse runtime compared to not using CUDD? I noticed that using CUDD requires time to build and maintain the binary decision diagram (BDD). Additionally, I would like to ask why the Power::ensureActivities()function does not use bfs.visitParallel(levelize_->maxLevel(), &visitor);.

[jjcherry56]

Higher accuracy than what? There non-trivial boolean operations on the expressions required for propagating the activities that aren't really practical without using CUDD (boolean diff).

It is possible to use a parallel bfs for propagating the activities but it would require adding locks or a lot of careful design to avoid them, which would negate the advantages. Try it and you will see it is not as simple
as you suggest.

[Zhihui-SJTU]

Thank you for the explanation. To clarify my first question: "Higher accuracy than what?" – I meant higher accuracy but worse runtime compared to not using CUDD(like the one illustrated in this image:).

Regarding the parallel BFS, you are absolutely right. I have already tried simply changing the code to bfs.visitParallel(levelize_->maxLevel(), &visitor);and found that the reported power values changed significantly. As you pointed out, it likely requires a lot of careful design (with locks or lock-free techniques) to ensure correctness before any runtime improvement can be achieved. I will continue my attempts based on this understanding.