multiprecision/doc/performance.qbk
2022-06-07 11:45:28 +02:00

73 lines
2.9 KiB
Plaintext

[/
Copyright 2011 - 2020 John Maddock.
Copyright 2013 - 2019 Paul A. Bristow.
Copyright 2013 Christopher Kormanyos.
Distributed under the Boost Software License, Version 1.0.
(See accompanying file LICENSE_1_0.txt or copy at
http://www.boost.org/LICENSE_1_0.txt).
]
[section:perf Performance Comparison]
In the beginning of the project and throughout,
many performance analyses, counts of multiprecision-operations-per-second
and the like have been performed.
Some of these are already listed in the ensuing sections.
We will now provide some general notes on performance, valid
for all of the multiprecision backends, before the detailed
benchmarks of the following sections.
The header-only, library-independent Boost-licenses integer
and floating-point backends including
__cpp_int for multiprecision integers,
__cpp_bin_float and __cpp_dec_float for multiprecision floating-point types
are significantly slower than the world's fastest implementations
generally agreed to be found in GMP/MPIR, MPFR and MPC (which are based on GMP).
Complex types __cpp_complex that are synthesized from these types
share similar relative performances.
The backends which effectively wrap GMP/MPIR and MPFR
retain the superior performance of the low-level big-number engines.
When these are used (in association with at least some level of optmization)
they achieve and retain the expected low-level performances.
At low digit counts, however, it is noted that the performances of __cpp_int,
__cpp_bin_float and __cpp_dec_float can actually meet or exceed
those encountered for GMP/MPIR, MPFR, etc. The reason for this
is because stack allocation and/or the use of fast container
storage can actually out-perform the allocation mechanisms in
GMP/MPIR, which dominate run-time costs at low digit counts.
As digit counts rise above about 50 or so, however,
GMP/MPIR performance steadily increases,
and simultaneously increases beyond (in relation to)
the performances of the Boost-licensed, self-written backends.
At around a few hundred to several thousands of digits,
factors of about two through five are observed,
whereby GMP/MPIR-based calculations are (performance-wise)
supreior ones.
At a few thousand decimal digits, the upper end of
the Boost backends is reached. At the moment,
advanced big-number multiplication schemes in the
Boost-licensed, self-written backends is limited
to school multiplication and Karatsuba multiplication.
Higher-orders of Toom-Cook and FFT-based multiplication
are not (yet) implemented. So it is not yet
feasible to perform mega-digit calculations
with the Boost-licensed, self-written backends,
whereas these are readily possible with
the GMP/MPIR and MPRF based backends.
[include performance_overhead.qbk]
[include performance_real_world.qbk]
[include performance_integer_real_world.qbk]
[include performance_rational_real_world.qbk]
[include performance_float.qbk]
[include performance_integer.qbk]
[include performance_rational.qbk]
[endsect]