eps(::AbstractFloat) inside a kernel doesn't compile on Julia 1.5.1

[mwarusz]

See the title. Works in a CUDA kernel and in KA on Julia 1.4.1.

MWE

using CUDA
using KernelAbstractions

function cuda_kernel(A)
  @inbounds A[1] = eps(A[1])
  nothing
end

@kernel function ka_kernel(A)
  @inbounds A[1] = eps(A[1])
end

let
  A = CuArray(ones(1))
  @cuda cuda_kernel(A)
  @show Array(A)

  fill!(A, 1)
  wait(ka_kernel(CUDADevice(), 1, 1)(A))
  @show Array(A)
end

[jakebolewski]

The actual issue here is with exponent, specifically with the construction of @noinline closures to raise specific errors:

https://github.com/JuliaLang/julia/blob/release-1.5/base/math.jl#L783

https://github.com/JuliaGPU/KernelAbstractions.jl/pull/130/files is a fix but there might be a better way fix the fundamental issue wrt to properly inlining the @noinline closures with Cassette / KA.

cc @vchuravy

[vchuravy]

@jakebolewski I must have been to tired the last time I read your message, and didn't read the part about @noinline.

@noinline is something we explicitly honor (see https://github.com/JuliaGPU/KernelAbstractions.jl/blob/8c7052d23c32d6950ca9abb326af144cce8dae1c/src/compiler/pass.jl#L15-L19) mostly to not blow up code size on the CPU.

[jakebolewski]

I noticed that, although I would think that behavior might have to be back end dependent for the GPU?

[vchuravy]

Well it depends, even for the GPU it is probably good to outline these error calls. But yes we can make this context dependent.