Typical accuracy function using onecold with a OneHotMatrix fails to compile on GPU

[KristofferC]

The following crashes when being compiled to GPU:

julia> x = rand(Float32, 10, 3) |> gpu;

julia> y = Flux.onehotbatch(1:3, 1:10) |> gpu;

julia> accuracy(x, y) = Flux.onecold(x) .== Flux.onecold(y);

julia> accuracy(x, y)
ERROR: GPU compilation of #23(CuArrays.CuKernelState, CUDAnative.CuDeviceArray{Bool,1,CUDAnative.AS.Global}, Base.Broadcast.Broadcasted{Nothing,Tuple{Base.OneTo{Int64}},typeof(==),Tuple{Base.Broadcast.Extruded{CUDAnative.CuDeviceArray{Int64,1,CUDAnative.AS.Global},Tuple{Bool},Tuple{Int64}},Base.Broadcast.Extruded{Array{Int64,1},Tuple{Bool},Tuple{Int64}}}}) failed
KernelError: passing and using non-bitstype argument

Argument 4 to your kernel function is of type Base.Broadcast.Broadcasted{Nothing,Tuple{Base.OneTo{Int64}},typeof(==),Tuple{Base.Broadcast.Extruded{CUDAnative.CuDeviceArray{Int64,1,CUDAnative.AS.Global},Tuple{Bool},Tuple{Int64}},Base.Broadcast.Extruded{Array{Int64,1},Tuple{Bool},Tuple{Int64}}}}.
That type is not isbits, and such arguments are only allowed when they are unused by the kernel.

[zsz00]

i meet the same problem..
run the mnist.jl demo use GPU . the code run on Flux 0.7.1 is OK, but error in 0.7.2

Julia Version 1.1.0
Commit 80516ca202 (2019-01-21 21:24 UTC)
Platform Info:
OS: Linux (x86_64-pc-linux-gnu)
CPU: Intel(R) Xeon(R) CPU E3-1230 v5 @ 3.40GHz
WORD_SIZE: 64
LIBM: libopenlibm
LLVM: libLLVM-6.0.1 (ORCJIT, skylake)

CUDN 10 + CUDNN 7.4

[KristofferC]

The PR got reverted, so would be good to reopen.

[darsnack]

This is fixed with all the latest changes to one-hot-related stuff (see #1448)

julia> x = rand(Float32, 10, 3) |> gpu;

julia> y = Flux.onehotbatch(1:3, 1:10) |> gpu;

julia> accuracy(x, y) = Flux.onecold(x) .== Flux.onecold(y);

julia> accuracy(x, y)
3-element CuArray{Bool,1}:
 1
 0
 0