Implement 3D async copy

[pxl-th]

No description provided.

[luraess]

Suggesting this minor change to be able to pass a tuple as [src/dst]Pos

[luraess]

Having an issue with the following MWE for dim=1

using AMDGPU

dim = 2  # works for 2 and 3, but not 1

P = zeros(Int32, 2, 3, 4)
P .= [iz*1e2 + iy*1e1 + ix for ix=1:size(P,1), iy=1:size(P,2), iz=1:size(P,3)];
P = ROCArray(P)

if dim==1
    ranges = [2:2, 1:size(P,2), 1:size(P,3)]
    buf = zeros(Int32, size(P,2), size(P,3))
elseif dim==2
    ranges = [1:size(P,1), 3:3, 1:size(P,3)]
    buf = zeros(Int32, size(P,1), size(P,3))
elseif dim==3
    ranges = [1:size(P,1), 1:size(P,2), 3:3]
    buf = zeros(Int32, size(P,1), size(P,2))
end

# device to host
AMDGPU.Mem.unsafe_copy3d!(
        pointer(buf), AMDGPU.Mem.HostBuffer,
        pointer(P), typeof(P.buf),
        length(ranges[1]), length(ranges[2]), length(ranges[3]);
        srcPos=AMDGPU.ROCDim3(ranges[1][1], ranges[2][1], ranges[3][1]),
        srcPitch=sizeof(Int32)*size(P,1), srcHeight=size(P,2),
        dstPitch=sizeof(Int32)*length(ranges[1]), dstHeight=length(ranges[2])
);

if dim==1
    @assert buf[:] ≈ Array(P[2,:,:])[:]
elseif dim==2
    @assert buf[:] ≈ Array(P[:,3,:])[:]
elseif dim==3
    @assert buf[:] ≈ Array(P[:,:,3])[:]
end

# host to device
P2 = similar(P)

AMDGPU.Mem.unsafe_copy3d!(
        pointer(P2), typeof(P2.buf),
        pointer(buf), AMDGPU.Mem.HostBuffer,
        length(ranges[1]), length(ranges[2]), length(ranges[3]);
        dstPos=AMDGPU.ROCDim3(ranges[1][1], ranges[2][1], ranges[3][1]),
        srcPitch=sizeof(Int32)*length(ranges[1]), srcHeight=length(ranges[2]),
        dstPitch=sizeof(Int32)*size(P,1), dstHeight=size(P,2)
);

if dim==1
    @assert Array(P2[2,:,:])[:] ≈ Array(P[2,:,:])[:]
elseif dim==2
    @assert Array(P2[:,3,:])[:] ≈ Array(P[:,3,:])[:]
elseif dim==3
    @assert Array(P2[:,:,3])[:] ≈ Array(P[:,:,3])[:]
end