Extend accumulate!

[vchuravy]

CuArrays has accumulate!, but it's limited: does not support anything but vectors, does not support the init keyword, and is slow (should use the shmem/shfl optims from https://github.com/JuliaGPU/CUDAnative.jl/blob/master/examples/scan.jl)

Old post:

@dpsanders and I just run into the situation where we wanted to do a cumsum on a CuArray.

CUDAnative has it as a example, but we should probably add the functionality to CuArrays https://github.com/JuliaGPU/CUDAnative.jl/blob/master/examples/scan.jl

[maleadt]

Maybe as a higher-order scan(f, ..; inclusive=true) function?

[dpsanders]

I have a simple and naive version of cumsum working. I can turn this into a version of accumulate!. Should I add this as a PR here?

[maleadt]

Yes that would be nice! Is it based on the CUDAnative example?

[ali-ramadhan]

Happy to help work on this and contribute an efficient scan implementation.

I'm particularly interested in an efficient non-destructive mean(c, dims) where c is a non-contiguous view as it currently devolves into scalar operations.

The scan example seems to overwrite the array. I guess to make a non-destructive version you would take the work-efficient and accumulate into a temporary array?

Currently mean is pretty fast but I couldn't figure out how it's implemented.

[maleadt]

Fixed by #447