Yuxiang Wang <yw...@virginia.edu> wrote: > 4) I wanted to say that it seems to me, as the project gradually > scales up, Cython is easier to deal with, especially when I am using a > lot of numpy arrays. If it is even higher dimensional data, it would > be verbose while it is really succinct to use Cython.
The easiest way to speed up NumPy code is to use Numba which is an LLVM based JIT compiler. Numba will often give you performance comparable to C for free. All you have to do is to add the @numba.jit decorator to your Python function and possibly include some type hints. If all you want is to speed up NumPy code, just use Numba and it will take care of it in at least 9 of 10 cases. Numexpr is also a JIT compiler which can speed up Numpy code, but it does not give as dramatic results as Numba. Cython is easier to work with than ctypes, particularly when the problem scales up. If you use typed memoryviews in Cython you can also avoid having to work with pointer arithmetics. Cython is mainly a competitior to using the Python C API manually for C extension modules to Python. Cython also allows you to wrap external C and C++ code, and e.g. use Python and C++ objects together. The drawback is that you need to learn the Cython language as well as Python and C and C++ and know how they differ. Cython also have many of the same hidden dangers as C++, due to the possibility of exceptions being raised between C statements. But because Cython is the most flexible tool for writing C extensions to Python you will in the long run do yourself a favor by learning to use it. ctypes is good when you have a DLL, possibly form a foreign source, and you just want to use it without any build step. CFFI is easier to work with than ctypes and has the same usecase. It can parse C headers and does not require you to define the C API with Python statements like ctypes do. Generally I would say it is alway better to use CFFI than ctypes. ctypes is also currently an orphan, albeit in the Python standard library, while CFFI is actively maintained. Numba will also JIT compile ctypes and CFFI calls to remove the extra overhead. This is good to know if you need to call a C function in a tight loop. In that case Numba can JIT compile away the Python as well as the ctypes/CFFI overhead. Fortran 90/95 is also underrated. It is easier to work with than C, and gives similar results performance wise. You can call Fortran with f2py, ctypes, CFFI, or Cython (use fwrap). Generally I would say that it is better for a student to learn C than Fortran if you have to choose, because C is also useful for other things than numerical computing. But if you want fast and robust numerical code, it is easier to get good results with Fortran than C or Cython. Sturla _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
