HDictus
HDictus
I think there is also merit in leaving the functions themselves straightforward and letting the user manage any pre and post processing they want to do. Unless these are best...
huh, I thought using GPU acceleration in julia required using GPUArrays, which would mean they'd dispatch differently. But I don't see that in NNlib.jl.
A common base function makes sense if there are multiple conv implementations.
I like the former better, I think you're right about losing time re-planning ffts and such The advantage of right associativity is that the notation is more consistent with the...
If no one else is working on this rn, I'll implement it.
I'm not familiar with the overlap-save algorithm, but from a quick search it seems it breaks down a long convolution into a bunch of smaller ones? does it work for...
I'll work on some other things for a while and I'll get back to this after Galen finishes what he's doing with overlap-save, since It looks like implementing and then...
for the moment I have an implementation where it only worries about overlap-save when given two inputs, and otherwise only does fft convolution. Unfortunately, this is slower than the naive...
Yeah I was overthinking it, after a break I took another look and it's pretty straightforward as you say. I'm just having a little trouble figuring out optimalfftfiltlength in one...
Hi Pablo, I've been enjoying this feature. Can I ask why you haven't made a pull request into the main Concord repo with this?