Open ericphanson opened 1 year ago
Is there a relatively small reproducer? I will be changing how memory management works but I'd like to figure out where this happens first.
No sorry. I gave up on SuiteSparseGraphBLAS shortly after that (also in my particular case it was a lot slower than SparseArrays, but that also could be my usage of it).
Ah sorry about that. Is Convex.jl using SparseArrays currently? I can take a look at why it was slower if you'd like
Yeah, though currently it is a bit of a dynamic mess of what exactly is being used where. In https://github.com/jump-dev/Convex.jl/pull/504 I worked on refactoring it to have a consistent pipeline, among other things. I experimented w/ switching back and forth, you can see the commented lines here: https://github.com/jump-dev/Convex.jl/pull/504/files#diff-e52f82517aa9bce265865136a070cfe938d823edd33d7b19ff6d0cde4a2ad584R132-R177 (though I think SparseSuiteGraphBLAS also needed the code elsewhere to be tweaked a little bit too).
Unfortunately I've kinda dropped that PR, will need to find some energy for Convex.jl again at some point to pick it up :)
Alrighty I'll give it a look. At worst it'll help me find better performance defaults!
running the tests of Convex.jl on commit https://github.com/jump-dev/Convex.jl/pull/504/commits/dc3f4f3827b5527e876ffb29204721ead7dd554f