conv corner case

[KristofferC]

Moved from JuliaLang/julia#20539

conv( randn(5), randn(0) ), say
should be, i'd think, should be a vector of length 0
right now it's an error

conv(randn(0), randn(0)) is trickier, but a vector length 0 is not unreasonable

[wheeheee]

Doesn't error now, but will have a length of 4.

[martinholters]

Yep, length(conv(u, v)) == length(u) + length(v) - 1 is a nice property which we can, do, and I'd say should, preserve even if one of u and v is empty. For length(u)==length(v)==0, the best we can do is probably returning an empty vector, which indeed we do.

However, conv(randn(5), randn(0); algorithm=:fft) still throws, which it shouldn't.

[wheeheee]

Oh, hadn't thought to check that...
Should this just be special-cased to use :direct?

[martinholters]

Don't know, haven't looked into it. The error originates in optimalfftfiltlength, so fixing that might be the minimal bugfix, provided the FFT based filtering then works.

[wheeheee]

That doesn't look right. Besides, it will be more performant with :direct anyway, allocating a zeroed array.

[martinholters]

Right, if I'm not mistaken, :direct will reduce to zeroing the output and and an empty loop, so that's good. OTOH, I've just tried patching optimalfftfiltlength so that it doesn't throw, but then unsafe_conv_kern_os! or _conv_kern_fft! throw (depending on which :fft_ variant is chosen), so that route is not as simple.