|  | Home | Libraries | People | FAQ | More | 
The first set of tests measure the times taken to execute the multiprecision part of the Voronoi-diagram builder from Boost.Polygon. The tests mainly create a large number of temporaries "just in case" multiprecision arithmetic is required, for comparison, also included in the tests is Boost.Polygon's own partial-multiprecision integer type which was custom written for this specific task:
| Integer Type | Relative Performance (Actual time in parenthesis) | 
|---|---|
| polygon::detail::extended_int | 1(0.138831s) | 
| int256_t | 1.19247(0.165551s) | 
| int512_t | 1.23301(0.17118s) | 
| int1024_t | 1.21463(0.168628s) | 
| checked_int256_t | 1.31711(0.182855s) | 
| checked_int512_t | 1.57413(0.218538s) | 
| checked_int1024_t | 1.36992(0.190187s) | 
| cpp_int | 1.63244(0.226632s) | 
| mpz_int | 5.42511(0.753172s) | 
| tom_int | 29.0793(4.03709s) | 
Note how for this use case, any dynamic allocation is a performance killer.
The next tests measure the time taken to generate 1000 128-bit random numbers and test for primality using the Miller Rabin test. This is primarily a test of modular-exponentiation since that is the rate limiting step:
| Integer Type | Relative Performance (Actual time in parenthesis) | 
|---|---|
| cpp_int | 5.25827(0.379597s) | 
| cpp_int (no Expression templates) | 5.15675(0.372268s) | 
| cpp_int (128-bit cache) | 5.10882(0.368808s) | 
| cpp_int (256-bit cache) | 5.50623(0.397497s) | 
| cpp_int (512-bit cache) | 4.82257(0.348144s) | 
| cpp_int (1024-bit cache) | 5.00053(0.360991s) | 
| int1024_t | 4.37589(0.315897s) | 
| checked_int1024_t | 4.52396(0.326587s) | 
| mpz_int | 1(0.0721905s) | 
| mpz_int (no Expression templates) | 1.0248(0.0739806s) | 
| tom_int | 2.60673(0.188181s) | 
| tom_int (no Expression templates) | 2.64997(0.191303s) | 
        It's interesting to note that expression templates have little effect here
        - perhaps because the actual expressions involved are relatively trivial
        in this case - so the time taken for multiplication and division tends to
        dominate. Also note how increasing the internal cache size used by cpp_int is quite effective in this case
        in cutting out memory allocations altogether - cutting about a third off
        the total runtime. Finally the much quicker times from GMP and tommath are
        down to their much better modular-exponentiation algorithms (GMP's is about
        5x faster). That's an issue which needs to be addressed in a future release
        for cpp_int.
      
Test code was compiled with Microsoft Visual Studio 2010 with all optimisations turned on (/Ox), and used MPIR-2.3.0 and MPFR-3.0.0. The tests were run on 32-bit Windows Vista machine.