Conections with TransmuteDims.jl, TensorCast.jl and CUDA.jl v5.

[tomsmierz]

In the pull request (link) in TransmuteDims.jl, which is part of the TensorCast.jl package, about updating dependency for Strided.jl from v1 to v2 there was mentioned some code problems, which prevent this update (most probably problems lies with UnsafeStridedView). This dependency issue prevents compatibility of TensorCast.jl with CUDA.jl v4 and v5.

The author of TransmuteDims.jl ( mcabbott) mentioned that these problems and required changes may be obvious to authors of Strided.jl. If you have time, can you look into it?

[Jutho]

Hi, I am certainly willing to look into this, if one could point me to what exactly I should be looking for. It will anyway be interesting, because other than dropping the support for the UnsafeStridedViews, there should not have been breaking changes.

[mcabbott]

That would be wonderful, sorry I've been snowed under & haven't got around to trying. This file is the all that touches Strided.jl.

Edit: when I make a quick attempt, this line is the problem, as I'm not sure what should replace UnsafeStridedView(pointer(dst), size(dst), strides(dst), 0, identity).

[tomsmierz]

Any updates?

[tomsmierz]

Asking again if there are any updates. It is an important issue, required to move to CUDA 5.

[Jutho]

Thanks for the reminder, I keep forgetting about this one. UnsafeStridedView is completely removed from Strided, because the overhead of dealing with structs that point to Arrays should have been resolved in Julia.

Hence, the whole block

    if isbitstype(T)
        _A = UnsafeStridedView(pointer(A), sz, st, 0, identity)
        _B = UnsafeStridedView(pointer(dst), size(dst), strides(dst), 0, identity)
        _densecopy_strided!(_B, _A)
    else
        # This path may be slower than the default one. Not really tested, either.
        _densecopy_strided!(dst, StridedView(A, sz, st, 0, identity))
    end

should just be replaced with

        _densecopy_strided!(dst, StridedView(A, sz, st, 0, identity))

Also, the _densecopy_strided! implementation:

@inline function _densecopy_strided!(dst, src)
    T = eltype(dst)
    LinearAlgebra.axpby!(one(T), src, zero(T), dst)
end

can actually just be replaced with Base.copy!.

[tomsmierz]

hey @mcabbott, can you implement these changes?

[mcabbott]

Thanks for the info above.

I was also trying a bit here: mcabbott/TransmuteDims.jl#43 and got tests passing, but perhaps there's a bit more to do.