Issue with vectorized functions on GPU about tullio.jl HOT 3 CLOSED

This is about mapslices-like operations, and I take it that you're not really looking for maximum(z; dims=1) but want other things.

That these work at all in Tullio is a bit of an accident, I thought a bit about this long ago in #20 . Mostly Tullio wants arrays of numbers only, things for which zero(T) makes sense.

I don't think there's any nice way of doing these on GPU arrays really, with or without this package. Something like map(f, eachcol(z)) iterates on the CPU & launches one GPU operation per slice, which will be slow. Tullio instead tries to generate a kernel, but such kernels should get numbers not whole arrays, and this is what the error is trying to say.

What you probably want to do is to alter f to act on the whole array, not ones slice at a time.

from tullio.jl.

Thank you for the detailed (and fast!) reply! Yes, my ideal workflow would be to vmap a function along a batch dimension for a CUDA array which seems extremely difficult/inefficient for CuArrays in Julia.

The longer story is that I want to have a user specify a symbolic function f(x, p) for symbolic vectors x, p, and to then evaluate the function for in parallel for batches of parameters P. If you have any more thoughts on how to do this, they'd be much appreciated, but this seems well outside the package scope -- but thanks for your work on this, Tullio is really great.

from tullio.jl.

Yes I think many people would like something like vmap. I've sent an invitation to something which may interest you...

from tullio.jl.