Passing views for copy!() to/from HYPREVector?

[johnomotani]

I wanted to be able to pass a view to copy!(), but at the moment this is not possible, because the arguments require a Vector{HYPRE_Complex} here

HYPRE.jl/src/HYPRE.jl

Lines 221 to 248 in faa61cc

	function Base.copy!(dst::Vector{HYPRE_Complex}, src::HYPREVector)
	ilower, iupper = Internals.get_proc_rows(src)
	nvalues = iupper - ilower + 1
	if length(dst) != nvalues
	throw(ArgumentError("length of dst and src does not match"))
	end
	indices = collect(HYPRE_BigInt, ilower:iupper)
	@check HYPRE_IJVectorGetValues(src, nvalues, indices, dst)
	return dst
	end
	function Base.copy!(dst::HYPREVector, src::Vector{HYPRE_Complex})
	ilower, iupper = Internals.get_proc_rows(dst)
	nvalues = iupper - ilower + 1
	if length(src) != nvalues
	throw(ArgumentError("length of dst and src does not match"))
	end
	# Re-initialize the vector
	@check HYPRE_IJVectorInitialize(dst)
	# Set all the values
	indices = collect(HYPRE_BigInt, ilower:iupper)
	@check HYPRE_IJVectorSetValues(dst, nvalues, indices, src)
	# TODO: It shouldn't be necessary to assemble here since we only set owned rows (?)
	# @check HYPRE_IJVectorAssemble(dst)
	# TODO: Necessary to recreate the ParVector? Running some examples it seems like it is
	# not needed.
	return dst
	end

Would it be possible to make this AbstractVector instead? If it's correct/legal, I can make a PR to make the change and add a test, but I don't know enough about ccall to know if this is sensible.

[johnomotani]

I'd be interested in the case when the views are associated with a contiguous piece of memory, I guess it might be necessary to restrict to that in order to pass a pointer to C? Maybe there is a way to restrict the type signature of a SubArray to enforce this?

[fredrikekre]

I think StridedArray may be the closest, but that allows non-1 strides so not sure it would work. There was some discussion in #16 (comment) but I can't seem to trigger the failure discussed there with non-contiguous views.

[johnomotani]

It looks like #16 was talking about PartitionedArrays specifically? I was trying to use the 'Direct assembly' method. Just trying to clarify because I don't know if it's (exactly?) the same problem.

[fredrikekre]

It was about passing views from PartitionedArrays. Anyway, I think you can relax the signature to StridedVector and then we add a runtime @assert stride(x, 1) == 1.