Mahotas Internals

This section is of interest if you are trying to understand how mahotas works in order to fix something, extend it (patches are always welcome), or use some of its technology in your projects.


Mahotas should not suck.

This is my main development goal and, if I achieve it, this alone should put mahotas in the top ten to one percent of software packages.

Mahotas should have no bugs. None. Ever.

Of course, some creep in. So, we settle for the next best thing: Mahotas should have no **known bugs**. Whenever a bug is discovered, the top priority is to squash it.

Read the principles of mahotas

C++/Python Division

Mahotas is, for the most part, written in C++, but almost always, you call a Python function which checks types and then calls the internal function. This is slightly slower, but it is easier to develop this way (and, for all but the smallest image, it will not matter).

So each will have its associated _module.cpp.

C++ Templates

The main reason that mahotas is in C++ (and not in pure C) is to use templates. Almost all C++ functions are actually 2 functions:

  1. A py_function which uses the Python C/API to get arguments, &c. This is almost always pure C.

  2. A template function<dtype> which works for the dtype performing the actual operation.

So, for example, this is how erode is implemented. py_erode is generic:

PyObject* py_erode(PyObject* self, PyObject* args) {
    PyArrayObject* array;
    PyArrayObject* Bc;
    if (!PyArg_ParseTuple(args,"OO", &array, &Bc)) return NULL;
    PyArrayObject* res_a = (PyArrayObject*)PyArray_SimpleNew(array->nd,array->dimensions,PyArray_TYPE(array));
    if (!res_a) return NULL;
    PyArray_FILLWBYTE(res_a, 0);
#define HANDLE(type) \
    erode<type>(numpy::aligned_array<type>(res_a), numpy::aligned_array<type>(array), numpy::aligned_array<type>(Bc));\

#undef HANDLE

These functions normally contain a lot of boiler-plate code: read the arguments, perform some sanity checks, perhaps a bit of initialisation, and then, the switch on the input type with the help of the SAFE_SWITCH_ON_INTEGER_TYPES_OF() and friends, which call the right specialisation of the template that does the actual work. In this example erode implements (binary) erosion:

template<typename T>
void erode(numpy::aligned_array<T> res, numpy::aligned_array<T> array, numpy::aligned_array<T> Bc) {
    gil_release nogil;
    const unsigned N = res.size();
    typename numpy::aligned_array<T>::iterator iter = array.begin();
    filter_iterator<T> filter(res.raw_array(), Bc.raw_array());
    const unsigned N2 = filter.size();
    T* rpos =;

    for (int i = 0; i != N; ++i, ++rpos, filter.iterate_with(iter), ++iter) {
        for (int j = 0; j != N2; ++j) {
            T arr_val = false;
            filter.retrieve(iter, j, arr_val);
            if (filter[j] && !arr_val) goto skip_this_one;
        *rpos = true;
        skip_this_one: continue;

The template machinery is not that complicated and the functions using it are very simple and easy to read. The only downside is that there is some expansion of code size. Given the small size of these functions however, this is not a big issue.

In the snippet above, you can see some other C++ machinery:


This is a RAII object that release the GIL in its constructor and gets it back in its destructor. Normally, the template function will release the GIL after the Python-specific code is done.


This is a thin wrapper around PyArrayObject that knows its type and has iterators. Relying on these objects has the further advantage that in debug mode, it checks bounds for many memory accesses. While this is very costly for everyday usage, it can catch bugs faster than the alternatives.


This is taken from scipy.ndimage and it is useful to iterate over an image and use a centered filter around each pixel (it keeps track of all of the boundary conditions).

The inner loop is as direct an implementation of erosion as one would wish for: for each pixel in the image, look at its neighbours. If all are true, then set the corresponding output pixel to true (else, skip it as it has been initialised to zero).

Most of the functions follow this architecture.