http://stackoverflow.com/questions/2290007/beginner-extending-c-with-python-specifically-numpy

If I have any success, I’ll post a solution with a link to this blog post. Thanks

I have included Travis’s code for automatic handling of array creation from malloc-ed memory (without copying the data), in a custom version of Bill Spotz’s numpy.i.
When the numpy array is deleted on the python side, the malloc block is also automatically de-allocated to prevent any memory leak.

I have written an entry in the scipy cookbook (explanations + complete example):
http://www.scipy.org/Cookbook/SWIG_Memory_Deallocation

Regards,
Egor

Thanks for the info. I was able to use SWIG to make a typemap that reads the C structure that contains the arrays, and uses PyArray_SimpleNewFromData to build the PyArrays and send them back to a dictionary/list in Python using PyArray_Return, which works.

Thanks again.

This is a C-extension example. Once you return the object created to Python it can be used just like any other NumPy array (you can set and get data or manipulate however you like).

There are many ways to get access to memory allocated in a C program (including SWIG or f2py or using this approach). The approach described here is really for being able to manage deallocation of the data. If you know that your data will never change, then you can just use PyArray_SimpleNewFromData.

I know this is a bit old, but I have a question about this…

I have an issue where I have a C program that generates multiple arrays that can each be several million cells long, just 1-D though, so like 1×4,000,000. Each array is a malloc’d array of type double. The program as is stores the data in a structure with other information about the data set (name, size, etc).

I need to get this data into a Python workspace. I can do it with SWIG using memory copies, but that is undesirable because it greatly slows down the program and will cause more memory to be used.

I would like to be able to return all the data as Numpy arrays, and this method seems like it could almost work (maybe in conjunction with SWIG to fill up a dictionary with the structure info), but not quite…

For this method, how do you use the “arr” array you create? Is it manipulatable (arr[i][j] = x)? or do you manipulate the data beforehand by using mymem?

Thanks,
Mike

PyCObject_FromVoidPtrAndDesc [2]

can be used instead of creating a new type.

[1] http://www.mail-archive.com/numpy-discussion@scipy.org/msg12442.html

[2] http://docs.python.org/c-api/cobject.html

Otherwise, you need to set up this indirection. You might also be able to handle the indirection with Cython though.

I think this pattern should actually be wrapped up and placed in NumPy itself, but I haven’t gotten the chance to do it.