Monodimensional linear interpolation performances

[remy-abergel]

Monodimensional linear interpolation performances

[ev-br]

Also, RegularGridInterpolator is for ndim> 1 data. Using it for 1D is bound to incur extra overhead as compared to cupy.interp or 1D interpolators from cupyx.scipy.interpolate namespace.

[takagi]

NumPy has its dispatch mechanism and when numpy.interp take a CuPy ndarray, it dispatches the input to cupy.interp. So what you're seeing is eventually the result of calling cupy.interp.

[kmaehashi]

I (accidentally) found that numpy.interp can take cupy arrays as inputs and run directly on the GPU device.

As @takagi suggested it is executed by CuPy, using mechanism described in NEP 18.

Also note that time.perf_counter() cannot be used for benchmarks that involves asynchronous GPU execution. (see https://docs.cupy.dev/en/latest/user_guide/performance.html#benchmarking for details)

I can observe significantly faster computation using numpy.interp instead of cupyx.scipy.interpolate.RegularGridInterpolator and cupyx.scipy.interpolate.interpn.

There is a functional difference between them - cupy.interp only supports 1-D but the latter supports N-D.

[remy-abergel]

Thank you very much for those very complete explanations.
Best wishes