Alexis Montoison

`cusparseHpruneCsr2csr` is for half precision but I don't know why it's not in [libcusparse.jl](https://github.com/JuliaGPU/CUDA.jl/blob/master/lib/cusparse/libcusparse.jl).

`BlasFloat` is `Union{BlasReal, BlasComplex}`. :) I don't use `where {T}` because other types are not supporter by the `geam` routine.

> > `BlasFloat` is `Union{BlasReal, BlasComplex}`. :) I don't use `where {T}` because other types are not supporter by the `geam` routine. > > Ok great! > > Regarding the...

> Ok, but I think that the problem is the conversion from/to COO matrix. > > In my functions I often call this conversion and this is why they fails....

@dpo It's for the last section of the JOSS paper.

Hi @KlausC ! The results of the first paper are not great. `CG` and `GMRES` should be only used to solve generic symmetric and square systems. For least squares problems...

Hi @albertomercurio! Good catch, I need to update my function `ktypeof` for a SubArray of a matrix. (https://github.com/JuliaSmoothOptimizers/Krylov.jl/blob/main/src/krylov_utils.jl#L221-L223). I will fix it tomorrow but you could use this temporary fix...

If you use `arnoldi` on GPUs, the Krylov basis `V` is a `CuMatrix` but the Hessenberg matrix is always stored as a `SparseMatrixCSC` for information.

You need to replace ```julia idxs = sortperm(abs.(μₖ)) ``` by ```julia idxs = sortperm(abs.(μ)) ``` at the end of your function `eigsolve`. For the `implicit restarted Arnoldi iterations`, what reference(s)...

Thanks @albertomercurio!