Severin Dicks
Severin Dicks
@leofang as mentioned, this PR is important for the single-cell community as the old implementation is a significant bottleneck and responsible for many oom errors due to CSR to CSC...
@leofang @asi1024 After extensive testing and reviewing SciPy’s implementation, I found that my earlier version failed to handle duplicate indices such as [5, 5, 3, 2, 1, 1]. SciPy simply...
@leofang @asi1024 Thank you for your patience with this one. All `cupyx.scipy.sparse` tests are now passing on my system, and I’ve tried to keep the kernel formatting as close as...
Yes. from my tests (and understanding) the limiting factor is that `cusparse.csr2dense` only outputs F-Arrays. The conversion to C-Arrays is pretty costly. In addition to that checking for `canonical_format` can...
@viclafargue switched it
@viclafargue Thank you for the fast fix can you rename `repulsion_strength` into `gamma`. That the name in UMAP's implementation and would make it easier for people to use it as...
@flying-sheep #1363 does that
Yes, we already have a good mask for sparse scaling. Boolean arrays are very effective for indicating where computations should be performed, as they eliminate the need for copying and...
Can not do this with a smarter numba kernel? at least for the sparse case. I think this should remove all the issues?
Todo: - [ ] Protect Kernels against NullPointers - [ ] Figure out if pointers or nb:arrays are the better option for passing arg - [ ] add keywords -...