What's the differences between mathematical operations provided by cupy and nvmath?

[sijieh23]

Hi,

Both cupy and nvmath provide the same (on the face of it) mathematical operations, such as cupyx.scipy.fft and nvmath.fft. What are the differences between these operations? Is cupy version faster than nvmath, or vice versa? Or nvmath is more flexible and can work with different types of arrays (numpy.ndarray and cupy.ndarray), while cupy's operations only work with cupy.ndarry?

Thanks.

[nv-kriehl]

@leofang Can you help provide some insight here? Thank you :)

[leofang]

Hello @sijieh23 thanks for reaching out!

The missions of CuPy and nvmath-python are different. CuPy is a high-level NumPy-like library that covers a wide spectrum of NumPy/SciPy/JIT compiler functionalities, whereas nvmath-python is meant to be a lower-level engine that covers all of NVIDIA GPU/CPU math libraries, which, in turns of functionalities, translates to a subset of CuPy feature set. As far as performance is concerned, since the math library team knows the best how to use the library APIs to get the maximal performance, yes it is expected that nvmath-python routines should be more performant given our expert (or, "insider") knowledge.

However, it is by no means a competition against existing libraries. On the contrary, we are keen to see both used in conjunction. Eventually, nvmath-python will become a dependency of CuPy so that CuPy and all of Python GPU libraries can focus on implementing new algorithms and optimizing performance, and not worry about exposing math primitives to Python (which is not trivial). As you've also noticed, nvmath-python tries to support as many Python array libraries as possible, including both CPU and GPU arrays, and it is therefore not an array library on its own (i.e. there's no nvmath.ndarray or nvmath.tensor). The nvmath-python team is working toward this goal.

[sijieh23]

Thank you @leofang for your reply.

Then at the current stage, what would you suggest using these different libraries? For example, suppose I'm solving a PDE using Fourier transform. Would you suggest using cupyx.scipy.fft, or nvmath.fft with cupy arrays?

And maybe off the topic: I noticed that there is another library, cupynumeric, which supports parallelization (if I understand it correctly). How does it relate to the other twos (cupy & nvmath)?