cuSPARSE bindings for Sliced ELLpack (SELL) format

[gdalle]

Is your feature request related to a problem? Please describe.

An intern I supervise (@AntoineBut) is currently working on a cross-platform sparse matrix library with KernelAbstractions.jl (https://github.com/AntoineBut/GPUGraphs.jl). Part of this work involves benchmarking against cuSPARSE, but only some matrix formats can be accessed through the CUDA.jl interface. In particular, the SELL format (https://docs.nvidia.com/cuda/cusparse/#sliced-ellpack-sell), described by NVIDIA as the state of the art in many situations, is not available from Julia.

Describe the solution you'd like

Would it be possible to implement a CuSparseMatrixSELL in CUDA.CUSPARSE, the same way that CSC and CSR formats already exist?

Describe alternatives you've considered

Calling directly into the C API is not user-friendly for people who only know Julia.

Additional context

@amontoison this might be for you

[amontoison]

@gdalle What kind of operations you need on this sparse format?
I agree that it is a relevant format to interface in CUSPARSE.

[gdalle]

Thanks for following up!
At the moment, we'd mostly need

• construction utilities
• matrix-vector products
• possibly matrix-matrix products

The context is graph traversal algorithms, where the sparse matrix represents the adjacency structure. The multiplications involved there use non-standard algebras (like max-plus) but to check that our custom KA kernels are not too terrible, we benchmark multiplication with the normal (plus-times) algebra against cuSPARSE. In those experiments, the hand rolled SELL format performs best, hence the need to compare it with its cuSPARSE counterpart.