Hello petsc4py users, I need some advice on petsc4py and how it can coexist with other approaches to get high performance with Python. Here's the situation:
- I am working on a reimplementation of a graph laplacian solver that uses a multigrid approach. [1] - The original sequential implementation is in MATLAB, so Python was a natural choice because it makes it easy to translate concepts from MATLAB. The project is not small and translating everything to C++ would have been difficult and tedious. Numpy/Scipy also provide many equivalents for built-in MATLAB functions. - Unlike the original implementation, the new version should be parallelized. Here, PETSc (via petsc4py) should provide the parallel data structures and some operations. - There are probably functions for which pure Python does not deliver the necessary speed (and PETSc probably does not provide the operations needed). I am researching how to rewrite such performance-critical parts in C/C++, and embed them in the Python code - using methods like scipy.weave or ctypes. [2] - My question: Is it possible to share PETSc objects between the Python code and the embedded C/C++ code? For example, pass a PETSc.Mat via Python, operate on it with C/C++ and return the matrix to Python? In general: Do you think my approach makes sense? I am new to PETSc and have not done high performance Python before, so I hope to learn from your answers, hints and criticism. Chris [1] https://code.google.com/p/lamg/ [2] http://www.scipy.org/PerformancePython#head-cafc55bbf8fd74071b2c2ebcfb6f24ed1989d540
