Strict typing for numpy arrays? Or conservative type-casting?

[ssrothman]

Hi pybind experts, I have what I suspect is a very naive question: is there a way to restrict the type-casting that is performed by pybind11 for numpy arrays? In particular, take a module defined like

#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>

namespace py = pybind11;

uint32_t xor_uint32(uint32_t a, uint32_t b) {
    return a ^ b;
}

py::array_t<uint32_t> xor_uint32_vec(const py::array_t<uint32_t>& a, const py::array_t<uint32_t>& b) {
    return py::vectorize(xor_uint32)(a, b);
}

PYBIND11_MODULE(test, m) {
    m.def("xor_uint32", &xor_uint32, "XOR two uint32_t values");
    m.def("xor_uint32_vec", &xor_uint32_vec, "XOR two arrays of uint32_t values");
}

(my actual application is obviously much more complicated than an xor)

is it possible to restrict the types that can be passed in from python? I am looking for two possible behaviors:

1. Totally strict typing - ie if I try to pass a numpy array with dtype anything other than np.uint32, I get an error
2. Conservative typing - ie only allow "safe" type-casting. For example, passing an array of np.uint16 would be successful, but passing an array of np.uint64 will fail.

In either scenario I would expect that passing non-integral times (eg np.float32 and np.float64 should fail. I would hope for an error message like this one from numpy:

>>> np.arange(10, dtype=np.int64) ^ np.random.random(10)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: ufunc 'bitwise_xor' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''

Instead, the current behavior is to simply cast the inputs to the expected data type of np.uint32 when the function is called. I'm particularly concerned about the "unsafe" casting - eg from floating-point types to integral types, or from larger integer types to smaller integer types. I'm also concerned about the implicit copy that must be taking place, as this is a performance-sensitive application.

Any help would be greatly appreciated!