Pybind11 vectorization vs numpy broadcasting

[SylvainCorlay]

What is the reasoning behind implementing a custom vectorization in pybind11 over the numpy API and helpers to create vectorized functions?

(See http://docs.scipy.org/doc/numpy-1.10.0/reference/c-api.ufunc.html and http://docs.scipy.org/doc/numpy-1.10.1/user/c-info.ufunc-tutorial.html).

[wjakob]

The vectorization in NumPy is just a nice syntax -- the actual loop executes in Python, so it's super-slow. The vectorization in Pybind11 is on the C++ side.

[SylvainCorlay]

Actually in both cases of the Python API (np.vectorize) and the C API (PyUFunc_FromFuncAndData), the loop is done in C.

The former case is slow because of the overhead of the python function calls (to the scalar function), but the latter case (the C API to create Ufuncs) should be fast.

[wjakob]

np.vectorize run on a function bound using Pybind11 would involve passing through pybind11 for each function invocation (lots of python API calls), hence the slowness. pybind11::vectorize is there to avoid that. PyUFunc_FromFuncAndData is an alternative, but this does not have some of the amenities of stateful lambda closures -- take a look at the examples, and I'm sure you will agree that the approach in pybind11 is much more powerful.

[SylvainCorlay]

PyUFunc_FromFuncAndData is an alternative, but this does not have some of the amenities of stateful lambda closures -- take a look at the examples, and I'm sure you will agree that the approach in pybind11 is much more powerful.

Yes, that is really cool. Thanks

[JohanMabille]

I have a problem with broadcasting in pybind11. Here is a code sample to reproduce the problem:

import test_pybind as tp
import numpy as np
x = np.array([1, 1, 1])
y = np.array([1, 1])[:, np.newaxis]
# Arrays are compatible for broadcasting
print(x + y)
# It also works with numpy ufuncs
print(np.logaddexp(x, y))
# But not with the pybind11 vectorize function
print(tp.vec_add(x, y))

[wjakob]

This would require a rewrite of pybind11::vectorize. It's not on my TODO list but patches are welcomed. (BTW: It's probably better if you create a separate ticket)

[JohanMabille]

Ok, I will open an new issue and probably start working on a patch tonight

[wjakob]

@jmabille Any news on this?

[JohanMabille]

@wjakob I have a fix on a private repo in my company, I'm waiting for an authorization to push it on my public fork

[wjakob]

great -- please keep me posted.

[ax3l]

@JohanMabille following the links here - did you contrib this or is py::vectorize still the way to go? :)

[JohanMabille]

For what I know py::vectorize is still the way to go; I remember contributing to this part of the code (i.e. numpy related) but it has been rewritten many times, so I don't know if the issue is still here.