Support for atomic `max` on `Float`

[pxl-th]

For feature-parity it'd be good to support atomic max on Float.
Currently MWE below fails:

using AMDGPU
using ROCKernels
using KernelAbstractions
using KernelAbstractions: @atomic

@kernel function f(target, source, indices)
    i = @index(Global)
    idx = indices[i]
    v = source[i]
    @atomic max(target[idx], v)
end

function main()
    source = rand(Float32, 1024)
    indices = rand(1:32, 1024)
    target = zeros(Float32, 32)
    @assert length(unique(indices)) < length(indices)

    dsource, dindices, dtarget = ROCArray.((source, indices, target))

    for i in 1:1024
        idx = indices[i]
        target[idx] = max(target[idx], source[i])
    end

    wait(f(AMDGPU.default_device())(dtarget, dsource, dindices; ndrange=1024))
    @assert Array(dtarget) == target
end

Error

ERROR: InvalidIRError: compiling kernel gpu_f(KernelAbstractions.CompilerMetadata{KernelAbstractions.NDIteration.DynamicSize, KernelAbstractions.NDIteration.DynamicCheck, Nothing, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}, KernelAbstractions.NDIteration.NDRange{1, KernelAbstractions.NDIteration.DynamicSize, KernelAbstractions.NDIteration.DynamicSize, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}}}, AMDGPU.Device.ROCDeviceVector{Float32, 1}, AMDGPU.Device.ROCDeviceVector{Float32, 1}, AMDGPU.Device.ROCDeviceVector{Int64, 1}) resulted in invalid LLVM IR
Reason: unsupported call to an unknown function (call to ijl_get_nth_field_checked)
Stacktrace:
 [1] indexed_iterate
   @ ./namedtuple.jl:140
 [2] multiple call sites
   @ unknown:0
Reason: unsupported call to an unknown function (call to jl_f_tuple)
Stacktrace:
 [1] indexed_iterate
   @ ./namedtuple.jl:140
 [2] multiple call sites
   @ unknown:0
Reason: unsupported call to an unknown function (call to ijl_get_nth_field_checked)
Stacktrace:
 [1] atomic_pointermodify
   @ ~/.julia/packages/LLVM/9gCXO/src/interop/atomics.jl:395
 [2] modify!
   @ ~/.julia/packages/UnsafeAtomicsLLVM/i4GMj/src/internal.jl:18
 [3] modify!
   @ ~/.julia/packages/Atomix/F9VIX/src/core.jl:33
 [4] macro expansion
   @ ~/code/Nerf.jl/src/Nerf.jl:155
 [5] gpu_f
   @ ~/.julia/packages/KernelAbstractions/C8flJ/src/macros.jl:81
 [6] gpu_f
   @ ./none:0
Reason: unsupported dynamic function invocation (call to atomic_pointerreplace)
Stacktrace:
 [1] atomic_pointermodify
   @ ~/.julia/packages/LLVM/9gCXO/src/interop/atomics.jl:395
 [2] modify!
   @ ~/.julia/packages/UnsafeAtomicsLLVM/i4GMj/src/internal.jl:18
 [3] modify!
   @ ~/.julia/packages/Atomix/F9VIX/src/core.jl:33
 [4] macro expansion
   @ ~/code/Nerf.jl/src/Nerf.jl:155
 [5] gpu_f
   @ ~/.julia/packages/KernelAbstractions/C8flJ/src/macros.jl:81
 [6] gpu_f
   @ ./none:0

As a temporary workaround, we can reinterpret Float32 as UInt32:

wait(f(DEVICE)(
    reinterpret(UInt32, dtarget),
    reinterpret(UInt32, dsource), dindices; ndrange=1024))

[jpsamaroo]

atomicrmw fmax/fmin was added in https://reviews.llvm.org/D127041, which I think came after LLVM 14, so we'll probably have to wait for LLVM 15 support in Julia to get this as a single GCN instruction.

In the meantime, we should add a fallback to a CAS loop in https://github.com/JuliaConcurrent/UnsafeAtomics.jl.