Using Cthulhu with GPUifyLoops on the GPU

[mwarusz]

First I tried something similar to this MWE

using GPUifyLoops, Cthulhu, CuArrays, CUDAnative

function kernel!(a, b)
  @inbounds @loop for i in (1:length(a);
                            (blockIdx().x - 1) * blockDim().x + threadIdx().x)
    a[i] = b[i]
  end
  nothing
end

a = CuArray(rand(Float32, 10 ^ 3))
b = similar(a)

threads = 256
blocks = ceil(Int, size(a, 1) / threads)
@descend @launch(CUDA(), threads = threads, blocks = blocks, kernel!(a, b))

and then I tried adding @descend just before the kernel call inside GPUifyLoops. Both approaches failed to allow me to step into kernel!. The second one got me much closer so it seems like there may be two different issues ? Should the above MWE work ?

[vchuravy]

This should definitely work,

using GPUifyLoops, Cthulhu, CuArrays, CUDAnative

function kernel!(a, b)
  @inbounds @loop for i in (1:length(a);
                            (blockIdx().x - 1) * blockDim().x + threadIdx().x)
    a[i] = b[i]
  end
  nothing
end

a = CuArray(rand(Float32, 10 ^ 3))
b = similar(a)

threads = 256
blocks = ceil(Int, size(a, 1) / threads)
f(a,b) = @launch(CUDA(), threads = threads, blocks = blocks, kernel!(a, b))
@descend f(a,b)

but won't since Cthulhu doesn't see the kwargs call to cufunction. Working on a fix