Support CUDSS.jl

[amontoison]

@geoffroyleconte
I started to support CUDSS.jl in RipQP.jl but I have some difficulties to understand how to implement a K2CUDSSParams.

The KKT matrix must be stored as CuSparseMatrixCSR{T, Cint} and I don't know yet if we need to implement some GPU kernels to update the blocks.

I also don't know if we need to re-implement functions like init_pad! and solver_name.

cc @dpo

[geoffroyleconte]

Hello @amontoison !
If you want to add another solver on CPU you only have to implement these structures and functions:
https://jso.dev/RipQP.jl/stable/switch_solv/#Advanced:-write-your-own-solver
However for GPUs, there may be additionnal functions to re-implement (dispatch for CUDA arrays) (see gpu_utils.jl file). You may also need to add additional functions, if you see errors messages with missing functions and scalar indexing, don't hesitate to send me the error message and I can try to redirect you to the part that needs to be modified.

The init_pad! function was used to separate memory allocation and initialization (for exemple separate vector allocations and computation of the symbolic analysis and first numerical factorization) but is not used (this was I feature I wanted to add for in-place solves of QPs but I did not have the time to implement it). You can see that it just returns the PreallocatedData object by default here:

RipQP.jl/src/iterations/solvers/linearsolvers.jl

Line 7 in 2b697a9

function init_pad!(pad::PreallocatedData)

and I don't think that I specialized this functions for any of the subsolvers I implemented.