cuda2rust_sandpit

Simplest configuration and testing of Rust sparse matrix data structures for sparse matrix linear algebra on CUDA.

First, define a build.rs which configures the linking of cuda libs necessary for computation. Then C libraries are created under myclib which creates the most basic interface for controlling cuBlas and cuSparse linalg routines.

A very simple FFI is constructed in lib.rs which references functions in myclib, and build.rs manages the construction of the dynamic shared library using:

cc::Build::new()
        .file("myclib/mycfuncs.c")
        .file("myclib/test_cublas.c")
        .file("myclib/test_cusparse.c")
        .file("myclib/sparse_funcs.c")
        .cuda(true)
        .compile("mycfuncs");

The idea is to manage memory using Rust, bue still leverage the low level power of CUDA which is most conveniently accessed via a custom C library.

Why not use cuBlas FFI?

I don't see any point in using a monolithic FFI for CUDA libs when it is clear that CUDA interfaces will not be natively written in Rust any time soon. Further, there is no FFI for cuSparse which is arguably the more important CUDA library.

It is much easier to write a small implementation using the cuSparse header and then create a small FFI related to a Rust project, which is playing to the strengths of both languages.

Implementation Notes

A quick note on build.rs linking ordering. For cargo test to work the linked libraries need to go after cc::Build::new() otherwise a linking error will be generated, i.e.:

cc::Build::new()
    .file("myclib/mycfuncs.c")
    .file("myclib/test_cublas.c")
    .file("myclib/test_cusparse.c")
    .cuda(true)
    .compile("mycfuncs");
    
println!("cargo:rustc-link-lib=dylib=cublas");

Tests

Run lib tests with cargo test.