Istvan, In addition to Thomas's recommendation, you might also have a look at http://wiki.python.org/moin/PythonSpeed/PerformanceTips ... and more specifically: http://wiki.python.org/moin/PythonSpeed/PerformanceTips#Loops
It is well-known among experienced Python programmers that for loops are very expensive -- I would almost guarantee that there are no mature Python applications that would naively execute a for loop over 10,000 elements. If it's not possible to replace the for loop with a map or a list comprehension, Thomas's suggestion is then applicable. Cheers, Steve On 10/09/2011 04:01 AM, Thomas Paviot wrote:
Hi Istvan, It's not a surprise that the python code is slower than the equivalent C++ code. Have a look at the file src/wrapper/SWIG_files/linux_darwin/gp_wrap.cpp (the C++ file generated with swig): you can see that creating a gp_Pnt from coordinates requires that the _wrap_new_gp_Pnt function is called, then _wrap_gp_Pnt_SetCoord__SWIG_1. That means that calling gp_Pnt(0,0,0) from python runs about 50 lines of C++ code. The wrapper code size/speed is independant from the C++ function wrapped: you won't see any difference if you call OCE/OCC methods that require high computation (boolean ops, STEP import/export etc.). It's not the case when calling "atomic" methods (creating a gp_Pnt from coords for instance). In my opinion, if you look for performance in your python code, the best way to proceed is to: - identify loops that may create a bottleneck - implement these loops in C++ functions - wrap these functions with SWIG - call the functions from python This will solve performance (as well as memory) issues. You can easily extend pythonocc with your own SWIG modules. You can for instance copy/paste the src/wrapper/Visualization folder to src/wrapper/MyFunctions. Modify the C++ and .i files, and add this new python module to the setup.py script so that it is compiled with the usual python setup.py build install. Thomas 2011/10/8 István Csanády <istvancsan...@gmail.com <mailto:istvancsan...@gmail.com>> Hi, I made some tests with pythonOCC and with the pure C++ OCE. According to my measurements the 'same' code in python is about 10 times slower than the C++ code. For example the following python code: for i in xrange(10000): p = gp_Pnt(0,0,0) face = BRepBuilderAPI_MakeFace(gp_Sphere(gp_Ax3(p,gp_Dir(1.,0.,0.)),50.)).Face() compared to the C++ code: for (int i = 0; i<10000; ++i) { gp_Pnt p(0,0,0); gp_Sphere sphere(gp_Ax3(p,gp_Dir(1.,0.,0.)),50.); TopoDS_Face face = BRepBuilderAPI_MakeFace(sphere).Face(); } The python version is about 8 times slower if I turn off the gc and about 10 times if I turn it on. Does anybody have any perfomance tips to make the pythonOCC code faster? István _______________________________________________ Pythonocc-users mailing list Pythonocc-users@gna.org <mailto:Pythonocc-users@gna.org> https://mail.gna.org/listinfo/pythonocc-users _______________________________________________ Pythonocc-users mailing list Pythonocc-users@gna.org https://mail.gna.org/listinfo/pythonocc-users
_______________________________________________ Pythonocc-users mailing list Pythonocc-users@gna.org https://mail.gna.org/listinfo/pythonocc-users
