William Silversmith
William Silversmith
This design is a little more tricky than one might think. If you make a separate DisjointSet per thread, it incurs a large memory overhead unless we change the design....
I think I have a good method for doing parallel, but it's so fast already... 70 MVx/sec/core isn't bad. c.f. #12
These are great approaches for 6 connected, which despite being the simplest version, is also often the slowest because you can't exploit connections between the arms of the stencil like...
This is a really interesting use case. For a single threaded application, a single linear scan of a uint32 512x512x64 volume takes about 26 msec. uint16 takes 18 msec and...
I was looking at the code again. I think it should be possible to simultaneously process blocks in a possibly lock free manner by setting `next_label` to different offsets. The...
I'm a bit jazzed about this parallel implementation concept, but I suspect it won't help your use case. What are the minimum requirements for your users? Just for cc3d, it...
I'm a bit time constrained at the moment, but I'm going to try to implement the parallel version of this. It might take a few days or weeks before you...
Dang. I was curious and did some reading and realized that I had been thinking (for some reason in contradiction to what I wrote in the README) that this library...
@matthew-walters I'm not sure if you're still experimenting, but the six connected algorithm has gotten significantly faster in recent releases. I'd be interested to hear how it's doing for you....
~In a small test on random data I saw 28msec (36 FPS) and on some of my structured connectomics data I saw 10 msec (100 FPS) for that size image.~...