copyto!() between cpu and gpu with subarrays

[btmit]

Copying a view from the CPU to existing memory on the GPU produces a scalar index error. Copy with the same way CPU-to-CPU and GPU-to-GPU still works.

using CUDA
CUDA.allowscalar(false)

a = zeros(ComplexF32, 4,3,4)
b = CUDA.zeros(ComplexF32, 4,3)

copyto!(b, view(a,:,:,1))

ERROR: scalar setindex! is disallowed

If I try to do the copy with a broadcast

b .= view(a,:,:,1)

I get

ERROR: GPU compilation of kernel broadcast_kernel(CUDA.CuKernelContext, CuDeviceArray{Complex{Float32},2,1}, Base.Broadcast.Broadcasted{Nothing,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}},typeof(identity),Tuple{Base.Broadcast.Extruded{SubArray{Complex{Float32},2,Array{Complex{Float32},3},Tuple{Base.Slice{Base.OneTo{Int64}},Base.Slice{Base.OneTo{Int64}},Int64},true},Tuple{Bool,Bool},Tuple{Int64,Int64}}}}, Int64) failed
KernelError: passing and using non-bitstype argument

[maleadt]

b .= view(a,:,:,1)

This is not supported, you can only have GPU arrays participate in a broadcast. So no CPU array, let alone a CPU view.

copyto!(b, view(a,:,:,1))

This is on purpose, too. We currently only support copying from Array, because there's no easy way to dispatch on dense CPU arrays. There's DenseArray, but that isn't limited to A<:Array:

julia> AbstractGPUArray <: DenseArray
true

And introducing a DenseCPUArray alias like

CUDA.jl/src/array.jl

Lines 148 to 164 in 62ae0c3

	# dense arrays: stored contiguously in memory
	DenseReinterpretCuArray{T,N,A<:Union{CuArray,ContiguousSubCuArray}} = Base.ReinterpretArray{T,N,S,A} where S
	DenseReshapedCuArray{T,N,A<:Union{CuArray,ContiguousSubCuArray,DenseReinterpretCuArray}} = Base.ReshapedArray{T,N,A}
	DenseSubCuArray{T,N,A<:Union{CuArray,DenseReshapedCuArray,DenseReinterpretCuArray}} = Base.FastContiguousSubArray{T,N,A}
	DenseCuArray{T,N} = Union{CuArray{T,N}, DenseSubCuArray{T,N}, DenseReshapedCuArray{T,N}, DenseReinterpretCuArray{T,N}}
	DenseCuVector{T} = DenseCuArray{T,1}
	DenseCuMatrix{T} = DenseCuArray{T,2}
	DenseCuVecOrMat{T} = Union{DenseCuVector{T}, DenseCuMatrix{T}}
	# strided arrays
	StridedSubCuArray{T,N,A<:Union{CuArray,DenseReshapedCuArray,DenseReinterpretCuArray},
	I<:Tuple{Vararg{Union{Base.RangeIndex, Base.ReshapedUnitRange,
	Base.AbstractCartesianIndex}}}} = SubArray{T,N,A,I}
	StridedCuArray{T,N} = Union{CuArray{T,N}, StridedSubCuArray{T,N}, DenseReshapedCuArray{T,N}, DenseReinterpretCuArray{T,N}}
	StridedCuVector{T} = StridedCuArray{T,1}
	StridedCuMatrix{T} = StridedCuArray{T,2}
	StridedCuVecOrMat{T} = Union{StridedCuVector{T}, StridedCuMatrix{T}}

would be too hard on the compiler (the current DenseCuArray already has significantly regressed precompilation and load times).

TLDR:

julia> CuArray(view(a,:,:,1))
4×3 CuArray{Complex{Float32},2}:
 0.0+0.0im  0.0+0.0im  0.0+0.0im
 0.0+0.0im  0.0+0.0im  0.0+0.0im
 0.0+0.0im  0.0+0.0im  0.0+0.0im
 0.0+0.0im  0.0+0.0im  0.0+0.0im

[btmit]

Thank you for the information and background.

Is there any way then to copy from a slice of an array on the CPU to pre-allocated memory on the GPU?

[maleadt]

You can always unsafe_copyto!(pointer(dst), pointer(src), n), but that's a litttle risky of course.

[ejmeitz]

Is there a way to copy from GPU into a CPU view?