GPU compatibility of data types

[youngdae]

I'm not sure if I'm not doing wrong, but user-defined data types, such as TronSparseMatrixCSC and TronDenseMatrix, seem to cause issues in passing them to a kernel function; compilation fails with complaints that it is a non-bitstype argument.

For example, the following call generates an error.

function kernel(p)
    @cuprintln(p.diag_vals) # It does not matter if I call @cuprintln or not. Compiling this will fail.
    return
end

d = ExaTron.TronSparseMatrixCSC(I, J, V, 8) # I, J, and V are CuArrays of types Int (I,J) and Float64 (V)
@cuda threads=128 blocks=1 kernel(d)

I was trying to implement a batch-solve of Tron, and I was thinking of doing this by calling dtron routines for each thread block in a kernel. However, dtron takes aforementioned user-defined data types, leading to failures. I guess all the internal routines using user-defined data types would cause issues in this case.

I'm a bit concerned that we might have to go back to the original implementation where bitstype arrays were used, which means dismantling Matrix into array components.

What do you think? Am I doing something wrong?

[frapac]

I don't think we could pass user types directly into CUDA's kernels. What we did in ExaPF is dismantling types in subcomponents before passing them to the kernels, e.g.
https://github.com/exanauts/ExaPF.jl/blob/master/src/models/polar/kernels.jl#L150-L154

In that sense, it is better to have struct of arrays rather than arrays of struct when we want to use GPU.
In ExaTron, I think we have the right structure for the matrix, and I would be in favor of dismantling the matrix into array components, as you said.