Complete and incomplete sparse cholesky factorization

Specific Domains GPU
1

Say I have a CuSparseCSR matrix “A” and I want to take complete and incomplete cholesky factorizations of it, but crucially leave that factorization on the GPU, how can I do that?

Things I’ve tried:

  1. CuSparse.jl - deprecated
  2. CuArrays.jl - possibly deprecated, got this error trying to “add CuArrays” in the package manager
ERROR: Unsatisfiable requirements detected for package CUDAapi [3895d2a7]:
 CUDAapi [3895d2a7] log:
 ├─possible versions are: 0.5.0-4.0.0 or uninstalled
 └─restricted by julia compatibility requirements to versions: uninstalled — no versions left
  1. Preconditioners.jl - only for CPU matrices
  2. cholesky(A) - works but puts the factorization on the CPU

Is there a way to do this?

2

CUSPARSE etc are now part of CUDA.jl, so just use the CUDA.CUSPARSE submodule.

For this specific functionality, you’d need to look in the sparse support of CUSOLVER, cuSOLVER. I don’t think the library supports what you’re looking for.

3

Thanks, I’ll try that. It does seem odd though that Julia would have this incredibly important functionality like a GPU-based cholesky factorization and then get rid of it during updates

4

Based on your explanation here and this, I would expect to be able to use ic0 when using CUDA.CUSPARSE, but it says ic0 is undefined. Where did this functionality end up? I am also OK with a full factorization, my matrix is block diagonal anyways

5

I’m not sure what you’re suggesting here. If you do a quick search on ic0 in the CUDA.jl repository you can find many functions. Nothing got deleted from the CUSPARSE package/module, but I’m not sure why you’re expecting the README instructions from a 5 year old package to work on a current version of CUDA.jl’s CUSPARSE module.

6

I don’t expect it to work as is, but I did expect there to be some way to do it, especially if you’re saying that nothing got deleted from the module (and it was possible 5 years ago). I’m specifically looking for complete or incomplete sparse csr gpu-based cholesky factorization (which you said in your first message is impossible in Julia, but definitely used to be possible)

7

You seem to have misunderstood my message. I’m not familiar with functionality you’re looking for, so I pointed you towards the CUSOLVER documentation. As you know what you are looking for, I recommend you take a look at the CUSOLVER/CUSPARSE docs to figure out which API calls you need. Every CUDA API call is available in CUDA.jl, so it I can guarantee you we did not remove functionality (with the sole exception for APIs that are have been by NVIDIA). Whether there’s high-level functionality wrapping those APIs is another question.

1 Like