Open william-silversmith opened 2 years ago
Got the first pass not crashing. Seems slower at the moment. Will have to figure out why. Looks like this speeds the first pass up, but slows down relabeling. Probably has something to do with making the renumbering array bigger. That can probably be handled with offsets.
Managed to increase perf from 120 MVx/sec to ~280 MVx/sec with 8 cores. Something isn't right, it should scale much closer to 1:1.
Some timing info from a 1GVx volume colored in with 8 copies of connectomics.npy.
Some things to observe:
Adds parallel flag.
The relabel pass was easy, but pretty marginal in terms of the performance impact. The changes that need to be made to the core equivalence pass are kind of outrageous. It might be better to maintain two separate pieces of code for single and parallel just so there's a sane version.