exchange BLAS.BlasFloat for Union{AbstractFloat, Complex{<:AbstractFl…

[MikaelSlevinsky]

…oat}}

There is an upcoming package for FFTs on generic number types https://github.com/JuliaApproximation/GenericFFT.jl. This will enable, in particular, convolutions with BigFloat and Complex{BigFloat} arrays. To make this Just Work^TM, we may simply remove the type restrictions referring to FFTW number types (quoted as BLAS.BlasFloats). Then when both packages are loaded, using DSP, GenericFFT, conv(rand(BigFloat, 10), rand(BigFloat, 20)) works analogously to finding BigFloat eigenvalues through GenericLinearAlgebra.

[MikaelSlevinsky]

Hi there, the tests pass locally for me and the change seems rather benign overall. Could a maintainer please take a look and approve running workflows?

[MikaelSlevinsky]

92/93! Yahoo! (I think that coverage is actually cumulative, so I'm not sure it's true that total coverage decreased.)

[MikaelSlevinsky]

Ping

[MikaelSlevinsky]

Hello, could a maintainer please take a look at this? Thank you

[galenlynch]

Yes, I'll look soon, sorry!

[MikaelSlevinsky]

Any other changes requested?

[martinholters]

This should probably be ok. But once we get to addressing #292, it would be unclear which types to use frequency domain or time domain convolution for if the applicability of fft depends on whether some other package (GenericFFT) has been loaded. Given how long #292 has been open, that shouldn't stop this PR offering an immediate improvement for now, though.

[martinholters]

I think this is fine. Any comments, @galenlynch? Merge and tag?

[MikaelSlevinsky]

This can probably be merged, right? FYI, this benign change reduces my package load time by about 40%.

[galenlynch]

Sorry I missed this, @martinholters. I'll try to finish #292. I got stuck figuring out how to best select between the many conv methods after implementing a bunch of different variants that work better in different situations. I think for now it may be better to accept occasionally choosing the wrong method in the name of being faster and more general, on average.