[WIP] Add support for ROCm

[jpsamaroo]

This is nowhere near ready to go yet, but I wanted to get this posted since things are progressing well for AMDGPU support overall :slightly_smiling_face:

TODO:

• [x] Add synchronization to AMDGPUnative, and then use it here
• [x] Merge JuliaGPU/AMDGPUnative.jl#6 for math intrinsics support
• [ ] Implement some means to select the desired backend
• [ ] Enable scratch and shmem support, and test them
• [ ] Ensure CI passes on some version of Julia (requires LLVM 7.0+ and probably JuliaLang/julia#31970)
• [ ] (Optional) Add aliases for threads, blocks, etc. to AMDGPUnative

[vchuravy]

Related #63

[jpsamaroo]

Ok, this is now working for me, albeit without synchronization (I still have to add those intrinsics to AMDGPUnative). Slightly modified example from the GPUifyLoops docs:

using GPUifyLoops, AMDGPUnative, HSARuntime

function kernel(A)
    @loop for i in (1:size(A,1);
                    threadIdx().x)
        A[i] = 2*A[i]
    end
    # TODO: @synchronize
    return nothing
end

kernel(A::HSAArray) = @launch ROC() kernel(A, groupsize=length(A))

data = HSAArray(rand(Float32, 1024))
kernel(data)

[vchuravy]

We should probably define threadIdx etc. in GPUIfyLoops, users currently still have to manually do using CUDAnative to get them.