Re: [boost] Looking for some "real world" extended precision integer arithmetic tests

Subject: Re: [boost] Looking for some "real world" extended precision integer arithmetic tests
From: John Maddock (boost.regex_at_[hidden])
Date: 2012-02-21 13:31:58

> John,
>
> Benchmark description follows (test code and results in the attachments):

Andrii,

I finally got around to running your benchmark myself and see rather 
different results:

 boost::polygon::detail::extended_int<128>
Voronoi Benchmark Test (time in seconds):
| Number of points | Number of tests | Time per one test |
|               10 |           10000 |          0.000053 |
|              100 |            1000 |          0.000861 |
|             1000 |             100 |          0.009630 |
|            10000 |              10 |          0.104400 |
|           100000 |               1 |          1.064000 |

 mp_number<fixed_int<128, true> >
Voronoi Benchmark Test (time in seconds):
| Number of points | Number of tests | Time per one test |
|               10 |           10000 |          0.000062 |
|              100 |            1000 |          0.000874 |
|             1000 |             100 |          0.009530 |
|            10000 |              10 |          0.105400 |
|           100000 |               1 |          1.089000 |

 mp_number<fixed_int<256, true> >
Voronoi Benchmark Test (time in seconds):
| Number of points | Number of tests | Time per one test |
|               10 |           10000 |          0.000054 |
|              100 |            1000 |          0.000868 |
|             1000 |             100 |          0.009470 |
|            10000 |              10 |          0.105200 |
|           100000 |               1 |          1.075000 |

 mp_number<fixed_int<512, true> >
Voronoi Benchmark Test (time in seconds):
| Number of points | Number of tests | Time per one test |
|               10 |           10000 |          0.000056 |
|              100 |            1000 |          0.000865 |
|             1000 |             100 |          0.009500 |
|            10000 |              10 |          0.105500 |
|           100000 |               1 |          1.071000 |

 mp_number<fixed_int<1024, true> >
Voronoi Benchmark Test (time in seconds):
| Number of points | Number of tests | Time per one test |
|               10 |           10000 |          0.000069 |
|              100 |            1000 |          0.001027 |
|             1000 |             100 |          0.011230 |
|            10000 |              10 |          0.122700 |
|           100000 |               1 |          1.250000 |

 mp_number<cpp_int_backend<> >
Voronoi Benchmark Test (time in seconds):
| Number of points | Number of tests | Time per one test |
|               10 |           10000 |          0.000055 |
|              100 |            1000 |          0.000866 |
|             1000 |             100 |          0.009630 |
|            10000 |              10 |          0.107000 |
|           100000 |               1 |          1.077000 |

 mp_number<mpz_int >
Voronoi Benchmark Test (time in seconds):
| Number of points | Number of tests | Time per one test |
|               10 |           10000 |          0.000064 |
|              100 |            1000 |          0.001076 |
|             1000 |             100 |          0.013260 |
|            10000 |              10 |          0.139300 |
|           100000 |               1 |          1.375000 |

 mp_number<tommath_int>
Voronoi Benchmark Test (time in seconds):
| Number of points | Number of tests | Time per one test |
|               10 |           10000 |          0.000063 |
|              100 |            1000 |          0.000897 |
|             1000 |             100 |          0.009770 |
|            10000 |              10 |          0.108800 |
|           100000 |               1 |          1.097000 |

So basically the times are all about the same, except for very large 
fixed_int's, or gmp - and even those aren't *much* worse.  There's also an 
experimental "cpp_int" in there which is an all C++ arbitrary precision type 
that uses the "small value optimization" to avoid memory allocation when 
possible.

This is with the latest gtl sandbox code BTW.  Maybe you've changed 
something to make the code much less dependent on the integer type?  I 
assume that the only code I need to change to test a new type is the line 
that defines big_int_type in voronoi_user_traits?

BTW I tried to run the test on Linux as well, but I couldn't get your code 
to compile there....

Cheers, John.