like the ones proposed here h_ttps://arxiv.org/pdf/2505.05896 h_ttps://github.com/mkauers/matrix-multiplication
thanks!
1 Like
You could make a package with those, do performance tests and highly precise numerical errors testing. Then unable running test on those with Sciml cases for instance.
Then only you can start a thread here explaining exactly what’s inside and how it compare to existing ones, and finally people will decide if it’s worth it. Note that even if it’s not your package will still be there and people will be able to use it.
This is the best choice because forking a repo and making a full PR end finally have it regected is hard to handle while knowing it’s in the ecosystem once and for all even though maybe not standard is less depressing.
3 Likes
The applicability of those algo is weaker than Strassen.
That’s why I insisted on the numerical error testing that’s what the papers never mention and where the bottleneck lives
1 Like