Boost Python is slow?

[Tyomich on the AIR]

Hello everyone! I'm currently embedding Python and have a problem. I have a struct (point) which looks like

struct TPoint
{
    int x, y;
};

And I want it to be usable from Python. I made it in two ways: the first one was writing all the warpping by myself according to Python/C API, and the second was wrap it with Boost.Python. So it looks like

class_<TPoint>("TPoint")
    .def_readwrite("x", &TPoint::x)
    .def_readwrite("y", &TPoint::y);

It wraps ok but here is a problem: I create a huge (1000000 items) list of objects of that type:

def cr_fig(n):
    res = []
    while n>0:
        res.append(TPoint())
        n -= 1
    return res

And then when I use a simple function which just increments the x member of every object, this function takes about 4 seconds to process the whole list (1000000 items) of Boost-wrapped objects and less than a second to process objects of my own wrapping! How this can be possible? Is Boost.Python much more slower than Python itself?

Regards, Artyom Chirkov.

P.S. I couldn't find out how to write to Boost.Python community, so I wrote here. If anyone knows how to do that, please let me know!
_______________________________________________
Boost-users mailing list
Boost-users@...
http://lists.boost.org/mailman/listinfo.cgi/boost-users

[Lars Viklund]

On Thu, Jul 09, 2009 at 01:00:47PM +0400, Tyomich on the AIR wrote:
> P.S. I couldn't find out how to write to Boost.Python community, so I wrote here. If anyone knows how to do that, please let me know!

http://mail.python.org/mailman/listinfo/cplusplus-sig
They deal in all kinds of Python interop but has a tendency towards BPL.

--
Lars Viklund | zao@... | 070-310 47 07
_______________________________________________
Boost-users mailing list
Boost-users@...
http://lists.boost.org/mailman/listinfo.cgi/boost-users

[Ravi-41]

On Thursday 09 July 2009 05:00:47 Tyomich on the AIR wrote:
> And then when I use a simple function which just increments the x member of
> every object, this function takes about 4 seconds to process the whole list
> (1000000 items) of Boost-wrapped objects and less than a second to process
> objects of my own wrapping! How this can be possible? Is Boost.Python much
> more slower than Python itself?

Boost.Python is a little slower than wrapping directly using Python C-API,
especially in very simple cases like yours. (It is a variant of the
abstraction penalty.) However, it is hard to believe that it is 4x slower.
Would you mind posting complete compilable examples so that we can take a look
at it?

Regards,
Ravi

_______________________________________________
Boost-users mailing list
Boost-users@...
http://lists.boost.org/mailman/listinfo.cgi/boost-users

[Tyomich on the AIR]

> Boost.Python is a little slower than wrapping directly using Python C-API,
> especially in very simple cases like yours. (It is a variant of the
> abstraction penalty.) However, it is hard to believe that it is 4x slower.
> Would you mind posting complete compilable examples so that we can take a look
> at it?
> Regards,
> Ravi

Sure! Here they are: http://www.filefactory.com/file/ahda0d8/n/_export_Boost_vs_pure_Python_7z
I tried to get/set and increase the member of a struct and here are my results:

Getting boost.python point 1000000 times:
0.30395876051
Getting Python point 1000000 times:
0.096010029312

Setting boost.python point 1000000 times:
0.373980616451
Setting Python point 1000000 times:
0.109051436823

Increasing boost.python point 1000000 times:
0.708493801882
Increasing Python point 1000000 times:
0.20538183019

The numbers are time in seconds taken to accomplish calculations. As you can see, Boost.Python is apparently 3-4 times slower than its equivalent in pure Python.
_______________________________________________
Boost-users mailing list
Boost-users@...
http://lists.boost.org/mailman/listinfo.cgi/boost-users