Make pairwise support CuArray

[xukai92]

No description provided.

[nalimilan]

Thanks. Unfortunately, I'm not sure this change is appropriate: for example, if a is a sparse matrix this will return a sparse matrix, which isn't the best type. I'm not sure what's the right approach here, maybe you can ask to CuArrays.jl people?

[maleadt]

We don't have a similar but dense interface right now. Maybe collect(similar(...)) ought to fullfill that role? Currently, collect(CUDA array) yields an Array, but I think it would make sense to have that return a CuArray (e.g. to get a dense CuArray from a sparse CuArray).

[nalimilan]

Unfortunately collect() will make a copy, which would be absurd since values are uninitialized. Looks like we need another interface.

[nalimilan]

See also JuliaLang/julia#13731 (comment). There Matt suggested introducing an allocate function which would allow specifying much more precisely what properties are expected from the resulting object. Somebody should really revive that idea.

[xukai92]

I see.

Is the only edge case being sparse matrices? If so maybe we can implement something similar to result_type and deal with them explicitly?

[nalimilan]

Sparse arrays are just an example. A PooledArray{Float64} could also be passed, yet we wouldn't want to return a PooledArray.

Can you file an issue in Julia? If we can agree on an API, it could be added to Compat so that we can use it in current Julia versions.

[xukai92]

Opned JuliaLang/julia#33361.