David Widmann
David Widmann
> worse than the function itself? Not sure what function you mean here and what you expect to be worse/better. The main issue is that your benchmarking is flawed, variables...
Yes, I try to avoid `using XX` in packages nowadays and rather use `import XX` or `using XX: f, g, h` to make such relations clearer (still convenient to use...
Probably some ChainRules rule for `getindex` that can't cope with Symbols but is hit in our AxisArrays tests.
> But if you have to transpose A' you have to move all of the memory around, which is probably already equivalent to the existing reduce(hcat, ::AbstractVector) It's defined here:...
> they definitely appear completely spurious. Just from looking at the failing tests and logs, I would assume that the test failures are real: They compare the allocations in the...
I often use temporary environments. Changing Project.toml and/or Manifest.toml for benchmarking is not a good idea IMO.
Well, that's exactly what I expected, so I'm not surprised. (BTW I would directly benchmark the functions of interest to avoid measuring `invoke` as well.)
You could just run the same benchmark on the master branch and on the PR.
That's also safer in so far as you don't have to reason about which methods are called - maybe there's some specialization for `AbstractVector{
I would have chosen `KernelTensorSum`, for consistency with `KernelTensorProduct`, similar to how we use the names `KernelSum`/`KernelProduct`. But I don't have a strong opinion.