Dan Schult

Thanks for those explanations for the difference between `dot` and `matmul`. It seems like we have two possible approaches to implementing `dot`. Can you put those changes on a new...

Looks like the timing is better for the current version. (same for the dense version and 4 times slower for sparse version in my crude timings). I would go with...

I was able to get 2 of the failed tests to pass by restarting them (internet troubles) but there are still two left. The M1 test failure seems unrelated. But...

Ahh.... as in `og` for `original`... hmmm... maybe `old` instead of `original`? `old_shape_self`? Or `remaining` or `persistent` but those are long too and not as clear. The term `nonreduced` is...

Yes @tylerjereddy, the current relnote seems sensible enough. And the relnote in the highlight section is more specific about what "basic arithmetic" refers to. I guess we could copy the...

Thanks for this work and set of ideas. I think it is natural to have multiple approaches to nD. Here is one I have been exploring. The idea is to...

Thanks for finding and fixing the bug with `np.add.at`. :) > I think numpytools definitely outperforms coo_todense_nd, so this replacement seems like a sensible choice. I did have a question...

It is not surprising that CSR matmul is faster even with the conversion, because CSR was designed for sparse @ dense. Of course CSR @ CSR is not CSR @...

Did we switch the `toarray` method for 1D input to use coo_todense_nd? Also, what tool were you using for benchmarking/timing the methods? Is it a short script you could post...

I suspect the segfault will occur on all systems. [based on the originally posted code:] Something is wrong with your line `n = min(vars_per_ctr, nvars - start)`. It gives a...