Adding copyto for non-contiguous matrices and vectors

[evelyne-ringoot]

No description provided.

[maleadt]

Woah, that's a lot of code... Can you explain your use case?

Also, why don't the existing methods in GPUArrays work for you?

• https://github.com/JuliaGPU/GPUArrays.jl/blob/4bb112c55c581e51e21d07a7df2dc2df7a8ca20e/src/host/abstractarray.jl#L51-L81
• https://github.com/JuliaGPU/GPUArrays.jl/blob/4bb112c55c581e51e21d07a7df2dc2df7a8ca20e/src/host/abstractarray.jl#L151-L214

[evelyne-ringoot]

Linking #1829 which is one of the use cases

I get the following scalar indexing error:

julia> a= CUDA.randn(10,10)
10×10 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:

julia> aview=view(a,2:2:6,3:2:7)
3×3 view(::CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, 2:2:6, 3:2:7) with eltype Float32:

julia> b= ones(Float32,8,8)
8×8 Matrix{Float32}:

julia> bview=view(b,2:2:6,3:2:7)
3×3 view(::Matrix{Float32}, 2:2:6, 3:2:7) with eltype Float32:

julia> copyto!(aview,bview)
┌ Warning: Performing scalar indexing on task Task (runnable) @0x000001dfa8d20010.
│ Invocation of setindex! resulted in scalar indexing of a GPU array.
│ This is typically caused by calling an iterating implementation of a method.
│ Such implementations *do not* execute on the GPU, but very slowly on the CPU,
│ and therefore are only permitted from the REPL for prototyping purposes.
│ If you did intend to index this array, annotate the caller with @allowscalar.
└ @ GPUArraysCore C:\Users\evely\.julia\packages\GPUArraysCore\HaQcr\src\GPUArraysCore.jl:106
3×3 view(::CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, 2:2:6, 3:2:7) with eltype Float32:
 1.0  1.0  1.0
 1.0  1.0  1.0
 1.0  1.0  1.0

julia> @which copyto!(aview,bview)
copyto!(dest::AbstractArray, src::AbstractArray)
     @ Base abstractarray.jl:1066

julia> typeof(aview)
SubArray{Float32, 2, CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}, Tuple{StepRange{Int64, Int64}, StepRange{Int64, Int64}}, false}

julia> typeof(bview)
SubArray{Float32, 2, Matrix{Float32}, Tuple{StepRange{Int64, Int64}, StepRange{Int64, Int64}}, false}

julia> 

Possibly https://github.com/JuliaGPU/GPUArrays.jl/blob/4bb112c55c581e51e21d07a7df2dc2df7a8ca20e/src/host/abstractarray.jl#L51-L81 cannot handle non-contiguous memory and https://github.com/JuliaGPU/GPUArrays.jl/blob/4bb112c55c581e51e21d07a7df2dc2df7a8ca20e/src/host/abstractarray.jl#L151-L214 does not handle a mix of Subarrays on GPU and CPU? Perhaps this could also be resolved by changing line 179 in the latter from 'AbstractGPUArray' to 'Union{SubArray{<:Any, <:Any, <:AbstractGPUArray}, AbstractGPUArray}', but this might increase compilation time? Might also need to include Adjoints and Transponse in the union, to be determined.

[evelyne-ringoot]

@maleadt Is this one ready? Also linking #1830 which is a more light-weight attempt to do the same, but does not cover all possible types of views I believe