missing numpy cuda math functions, e.g. matmul

[cristipurdel]

I have a kernel runing with guvectorize on 'cpu' and 'cuda'
Since there is no matmul support, I have to implement this outside of the kernel as a jitted function decorated depending if it is 'cpu' or 'cuda' used.
This swould be a relative "easy" way to generate functions like matmul, transpose. I do not how to add patches, but I could write the functions example.
Not sure if this is related to Docs 6 & 8 but it wold be usefull to have those functions and additional ones when using 'cuda' and basically extend http://numba.pydata.org/numba-doc/latest/cuda/cudapysupported.html#built-in-functions

[stuartarchibald]

Thanks for the request. The following is under development and may help?

1. There is no e.g. matmul support for CUDA directly from Numba.
2. The reason for 1. is in part that core linear algebra functionality really needs to come from bindings to e.g. the cuBLAS library so that performance optimal BLAS kernels are doing the necessary work.
3. Numba supports the __cuda_array_interface__ specification, which means it can interoperate with CuPy.
4. CuPy provides np.dot along with many more NumPy like functions.

Failing this, pyculib provides a cuBLAS wrapper.

[cristipurdel]

CuPy from what I saw is more like C, and I would prefer to use python :)
pyculib has not seen too much activity in the last year or so
I will use my approach with jitted function and I'll wait for the feature to be added

[stuartarchibald]

Just to check we are talking about the same thing. When I say CuPy I mean https://github.com/cupy/cupy

CuPy : NumPy-like API accelerated with CUDA

It is a Python package that has a NumPy like API and targets CUDA GPUs. No C or C-like behaviour involved unless you want to write custom kernels/do lower level things. As you asked for NumPy like functions, they support a lot https://docs-cupy.chainer.org/en/stable/overview.html, which was my reason for suggesting it. The specific matrix product related functions are here https://docs-cupy.chainer.org/en/stable/reference/linalg.html#matrix-and-vector-products.

Hope this helps?

[cristipurdel]

CuPy looks definitely interesting, but it would be nice if numba can integrate natively some of those functions
I would prefer to leave this issue open as a feature request.

[cristipurdel]

CUPy is definitely interesting, especially for numba interoperability:
cupy/cupy#1760
Also I made some test runs using cupy & tensorflow vs numpy
https://github.com/cristipurdel/python/blob/master/linalg_solve_np_cp_tf.py

[cristipurdel]

Any chance of making a wrapper for at least the BLAS functions since there is a certain degree of interoperability with cupy?

[cristipurdel]

Would it be possible to inline cupy/numpy functions in the guvectorize decorator?
I tried it but it gives errors.
If this would work, it would make guvectorize behave more like openacc/openmp :)