Is it possible to add cuDSS as linear solver for SCS ?

[Franc-Z]

According to https://docs.nvidia.com/cuda/cudss/, cuDSS is a high performance linear solver for large scale problem. Is it possible to add cuDSS as linear solver for SCS ?

[kalmarek]

should be pretty straightforward: these are the functions that need to be implemented:
https://www.cvxgrp.org/scs/linear_solver/index.html#implementing-a-new-linear-solver

Just have a look at what the indirect solver is doing.

(I was even thinking of it in my spare time, but I don't currently have running access to CUDA-12 enabled gpu)

[kalmarek]

@Franc-Z if you feel adventurous and own a cuda-12 gpu you may try this one:
kalmarek#2

most probably it will crash (I have no option to test it), but should not eat your cat 😺 If you try, please report back!

[kalmarek]

ok, I rented a gpu ;) for the afternoon and got this working (i.e. it's passing all tests, yay!).
To compile You will need to have

• libopenblas, liblapack
• cudss-0.5.0
• cuda-runtime
• cuda-blas
• nvcc

1. Define CUDA_PATH (defaults to /usr/local/cuda)
2. Define CUDSS_PATH (mandatory, no default)
3. run make cudss

This is my first CUDA code, honestly ;)

@bodono I could also open a pull directly here, if you want to review!

[Franc-Z]

@kalmarek Thanks！