Can cuBLAS be supported?

[kimurayu45z]

Can cuBLAS be supported in this library?
If it is able, it is very useful.

[kimurayu45z]

https://github.com/rust-cuda/cuda-sys
I found this one.

I propose that:

• Creating "cublas-src" repository
• In the "cublas-src" repository, write the code of wrapper for "cublas-sys" with BLAS interface.
• In the "blas-src" repository, make the crate enable to choose "cublas-src"

I can contribute. How about it?

[IvanUkhov]

Hi, thank you for the note! Yes, absolutely. It is very much encouraged to keep on extending the list of sources as long as they conform to the standard BLAS/LAPACK interface. Please take a look at the architecture:

https://github.com/blas-lapack-rs/blas-lapack-rs.github.io/wiki#sources

Can you please clarify why cublas-sys is needed? Do you mean the set of functions is larger than what is in the standard?

[kimurayu45z]

Sorry it was my misunderstanding. cublas-sys is not needed.
However, the interface of cuBLAS seems to be slightly different from it of BLAS, so I think we need the wrapper for using cuda with BLAS interface, instead of cublas-sys.

[IvanUkhov]

It’s been a while. I think the main concern is that, if the APIs are drastically different, it would not fit into the current architecture, which is to have a small number of sys crates interfacing with different sources allowing one to build on top. Everything that is specific to a particular implementation should probably have its own sys create, as we discuss here.

[Pr0methean]

It might be more realistic to support NVBLAS, a library that Nvidia built on top of cuBLAS to be a drop-in replacement for BLAS. It also has heuristics about whether the GPU acceleration is worth the time it'd take to transfer the input to GPU memory and the output back.