GPU Map without reduction on multiple arrays indices

Specific Domains GPU
1

Hello,

I would like to compute a multi-dimensional array (4-D) A of a 2 components function f on a large multi-dimensional domain (X,Y,T) with GPU. So A will have size: 2, NX, NY, NT.
X, Y and T are 1D arrays

I tried to use the code of @SimonDanisch in An Introduction to GPU Programming in Julia - Nextjournal
with a simple map kernel.
This is what I wrote:

using GPUArrays, CuArrays
using SharedArrays
using DistributedArrays
using BenchmarkTools
# Overloading the Julia Base map! function for GPUArrays
function Base.map!(f::Function, A::GPUArray, X::GPUArray, Y::GPUArray, T::GPUArray)
    # our function that will run on the gpu
    function kernel(state, f, A, X, Y, T)
        # If launch parameters aren't specified, linear_index gets the index
        # into the Array passed as second argument to gpu_call (`A`)
        I = CartesianIndex(state)
    		if I[2]*I[3]*I[4] <= length(A)/2
          @inbounds A[1:2,I[2],I[3],I[4]] = f(X[I[2]], Y[I[3]], T[I[4]])
        end
        return
    end
    # call kernel on the gpu
    gpu_call(kernel, A, (f, A, X, Y, T))
end

# on the GPU:
NX, NY, NT = 10, 15, 13
X, Y, T= rand(NX), rand(NY), rand(NT)
a = zeros(2, NX, NY, NT)

kernel(x,y,t) = [x+y, y-t]

xgpu, ygpu, tgpu, agpu = cu(X), cu(Y),cu(T), cu(a)
gpu_t = @belapsed begin
  map!($kernel, $agpu, $xgpu, $ygpu, $tgpu)
  GPUArrays.synchronize($agpu)
end

I have the following error:

`GPU compilation of kernel(CuArrays.CuKernelState, typeof(kernel), CUDAnative.CuDeviceArray{Float32,4,CUDAnative.AS.Global}, CUDAnative.CuDeviceArray{Float32,1,CUDAnative.AS.Global}, CUDAnative.CuDeviceArray{Float32,1,CUDAnative.AS.Global}, CUDAnative.CuDeviceArray{Float32,1,CUDAnative.AS.Global}) failed
KernelError: kernel returns a value of type Union{}

Make sure your kernel function ends in return, return nothing or nothing.
If the returned value is of type Union{}, your Julia code probably throws an exception.
Inspect the code with @device_code_warntype for more details.

I don’t know how to use the macro @device_code_warntype to debug the code. I am working on a cluster, I wrote JULIA_DEBUG=CUDAnative in the terminal before to launch the julia REPL but I got ‘UndefVarError: @device_code_warntype not defined’ when I run the above code

2

You need to import CUDAnative to have access to those functions and macros.

How did you write that code? The call to CartesianIndex(state) cannot get inferred by Julia, and I find nothing of the like in Simon’s tutorial. Instead, he uses GPUArrays.linear_index.

Also note that I would not recommend using either GPUArrays or extending map! for such an operation, it seems like you would be much easier off trying to model this as either a use of map or broadcast (but it’s hard to predict whether that’s possible without more details of the problem) or just use a regular kernel without the GPUArrays abstractions on top of it. See https://juliagpu.gitlab.io/CuArrays.jl/tutorials/generated/intro/