[async] Filter list operations that are clearly redundant

[yuanming-hu]

Related issue = #742

... for faster async optimizations

Benchmark on 2D MGPCG:

Before:

• synchronize: 7.907
  • rebuild_graph in optimize_listgen: 2.225

After (33% faster):

• synchronize: 5.935
  • rebuild_graph in optimize_listgen: 1.668

[Click here for the format server]

---

[k-ye]

LGTM!

It reminds me of how the SFG itself works: if we batch the records for a certain number of kernel launches, then run this list filtering, we could potentially eliminate more redundant listgens?

[yuanming-hu]

It reminds me of how the SFG itself works: if we batch the records for a certain number of kernel launches, then run this list filtering, we could potentially eliminate more redundant listgens?

Not sure if I clearly get the question, but I believe current the early listgen filtering trick in this PR already eliminates as many redundant listgens as it can :-) We can probably chat more in the meeting tonight :-)