linsolve on CUDA has an issue

[RibeiroAndre]

Hi!
Here's the issue:

using CUDA
using KrylovKit
linsolve(CuArray([1.0 2.0; 3.0 4.0]),CuArray([1.0,2.0]))
┌ Warning: Performing scalar operations on GPU arrays: This is very slow, consider disallowing these operations with `allowscalar(false)`
└ @ GPUArrays /pool/aerospace/Tools/julia-1.5.3/depot/packages/GPUArrays/WV76E/src/host/indexing.jl:43

This seems to fix it:

LinearAlgebra.mul!(x::CuArray{T,1},y::CuArray{T,1},α) where T = x .= α .* y;

Maybe something can be done about it in the code? I don't think this is an issue with LinearAlgebra as:

using LinearAlgebra
a=CuArray([1.0 2.0; 3.0 4.0]);
b=CuArray([1.0,2.0]);
c=CuArray([0.0,0.0]);
mul!(c,a,b)

works.

[Jutho]

Your second example is irrelevant right? There you are doing a matrix vector multiplication. What seems to be missing is a multiplication with a scalar using the mul! interface. This works for arrays in LinearAlgebra, but a method seems to be missing for CuArrays. I think this is a CUDA.jl issue.

[Jutho]

However, it would arguably be good to have dedicated tests for testing KrylovKit.jl - CUDA.jl interoperability.

[RibeiroAndre]

I thought the second example was showing that mul! works on CuArrays in general, so I don't know why KrylovKit requires making changes to mul! in order to work.
I can file this under CUDA.jl if you think that's more appropriate, but I don't know how to reproduce the issue without using KrylovKit.

[Jutho]

Because it's another method signature, you are illustrating

mul!(::CuArray, ::CuArray, ::CuArray)

the error is from

mul!(::CuArray, ::CuArray, ::Number)

[RibeiroAndre]

Ah, yes. Clearly. Thanks for pointing that out. I'll file a CUDA ticket and you can decide if you want to close this one.
Thanks!

[RibeiroAndre]

Works on Julia Version 1.6.2, CUDA toolkit 11.1.1.