[/==============================================================================
    Copyright (C) 2001-2010 Joel de Guzman
    Copyright (C) 2001-2010 Hartmut Kaiser

    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:numeric_performance Performance of Numeric Generators]

[section:int_performance Comparing the performance of a single int_ generator]

These performance measurements are centered around default formatting of a 
single `int` integer number using different libraries and methods. 
The overall execution times for those examples are compared below. We compare
using `sprintf`, C++ iostreams, __boost_format__, and __karma__.

For the full source code of the performance test please see here:
[@../../optimization/karma/int_generator.cpp int_generator.cpp]. All the 
measurements have been done by executing `1e7` iterations for each 
formatting type (NUMITERATIONS is set to `1e7` in the code shown below).

[import ../../optimization/karma/int_generator.cpp]

Code used to measure the performance for `ltoa`:

[karma_int_performance_ltoa]

Code used to measure the performance for standard C++ iostreams:

[karma_int_performance_iostreams]

Code used to measure the performance for __boost_format__:

[karma_int_performance_format]

Code used to measure the performance for __karma__ using a plain character buffer:

[karma_int_performance_plain]

The following table shows the overall performance results collected 
while using different compilers. All times are in seconds measured for `1e7`
iterations (platform: Windows7, Intel Core Duo(tm) Processor, 2.8GHz, 4GByte RAM). 
For a more readable comparison of the results see this 
[link spirit.karma.int_performance figure].

[table Performance comparison for a single int (all times in [s], `1e7` iterations)
    [[Library]              [gcc 4.4.0 (32 bit)] [VC++ 10 (32 bit)] [Intel 11.1 (32 bit)] [gcc 4.4.0 (64 bit)] [VC++ 10 (64 bit)] [Intel 11.1 (64 bit)]]
    [[ltoa]                 [1.542]              [0.895]            [0.884]               [1.163]              [1.099]            [0.906]]
    [[iostreams]            [6.548]              [13.727]           [11.898]              [3.464]              [8.316]            [8.115]]
    [[__boost_format__]     [16.998]             [21.813]           [20.477]              [17.464]             [14.662]           [13.646]]
    [[__karma__ int_]       [1.421]              [0.744]            [0.697]               [1.072]              [0.953]            [0.606]]
]

[fig int_performance.png..Performance comparison for a single int..spirit.karma.int_performance]

[endsect]

[/////////////////////////////////////////////////////////////////////////////]
[section:double_performance Comparing the performance of a single double_ generator]

These performance measurements are centered around default formatting of a 
single `double` floating point number using different libraries and methods. 
The overall execution times for those examples are compared below. We compare
using `sprintf`, C++ iostreams, __boost_format__, and __karma__.

For the full source code of the performance test please see here:
[@../../optimization/karma/double_performance.cpp double_performance.cpp]. All the 
measurements have been done by executing `1e6` iterations for each 
formatting type (NUMITERATIONS is set to `1e6` in the code shown below).

[import ../../optimization/karma/double_performance.cpp]

Code used to measure the performance for `sprintf`:

[karma_double_performance_printf]

Code used to measure the performance for standard C++ iostreams:

[karma_double_performance_iostreams]

Code used to measure the performance for __boost_format__:

[karma_double_performance_format]

The following code shows the common definitions used by all __karma__ performance
measurements as listed below:

[karma_double_performance_definitions]

Code used to measure the performance for __karma__ using a plain character buffer:

[karma_double_performance_plain]

The following table shows the overall performance results collected 
while using different compilers. All times are in seconds measured for `1e6`
iterations (platform: Windows7, Intel Core Duo(tm) Processor, 2.8GHz, 4GByte RAM). 
For a more readable comparison of the results see this 
[link spirit.karma.double_performance figure].

[table Performance comparison for a single double (all times in [s], `1e6` iterations)
    [[Library]              [gcc 4.4.0 (32 bit)] [VC++ 10 (32 bit)] [Intel 11.1 (32 bit)] [gcc 4.4.0 (64 bit)] [VC++ 10 (64 bit)] [Intel 11.1 (64 bit)]]
    [[sprintf]              [0.755]              [0.965]            [0.880]               [0.713]              [0.807]            [0.694]]
    [[iostreams]            [2.316]              [2.624]            [1.964]               [1.634]              [1.468]            [1.354]]
    [[__boost_format__]     [3.188]              [3.737]            [2.878]               [3.217]              [2.672]            [2.011]]
    [[__karma__ double_]    [0.813]              [0.561]            [0.368]               [0.426]              [0.260]            [0.218]]
]

[fig double_performance.png..Performance comparison for a single double..spirit.karma.double_performance]

[endsect]

[////////////////////////////////////////////////////////////////////////////]
[section:format_performance Comparing the performance of a sequence of several generators]

These performance measurements are centered around formatting of a sequence of
different items, including 2 `double` floating point numbers using different 
libraries and methods. The overall execution times for those examples are 
compared below. We compare using `sprintf`, C++ iostreams, __boost_format__, 
and __karma__.

For the full source code of the performance test please see here:
[@../../optimization/karma/format_performance.cpp format_performance.cpp]. All the 
measurements have been done by doing `1e6` iterations for each formatting 
type (NUMITERATIONS is set to `1e6`).

[import ../../optimization/karma/format_performance.cpp]

Code used to measure the performance for sprintf:

[karma_format_performance_printf]

Code used to measure the performance for standard iostreams:

[karma_format_performance_iostreams]

Code used to measure the performance for __boost_format__:

[karma_format_performance_format]

The following code shows the common definitions used by all __karma__ 
performance measurements as listed below:

[karma_format_performance_definitions]

Code used to measure the performance for __karma__ using a plain character 
buffer:

[karma_format_performance_plain]

The following table shows the overall performance results collected 
while using different compilers. All times are in seconds measured for `1e6`
iterations (platform: Windows7, Intel Core Duo(tm) Processor, 2.8GHz, 4GByte RAM). 
For a more readable comparison of the results see this 
[link spirit.karma.format_performance figure].

[table Performance comparison for a sequence of several items (all times in [s], `1e6` iterations)
    [[Library]              [gcc 4.4.0 (32 bit)] [VC++ 10 (32 bit)] [Intel 11.1 (32 bit)] [gcc 4.4.0 (64 bit)] [VC++ 10 (64 bit)] [Intel 11.1 (64 bit)]]
    [[sprintf]              [1.725]              [1.892]            [1.903]               [1.469]              [1.608]            [1.493]]
    [[iostreams]            [4.827]              [5.287]            [4.444]               [3.112]              [3.319]            [2.877]]
    [[__boost_format__]     [5.881]              [7.089]            [5.801]               [5.455]              [5.254]            [4.164]]
    [[__karma__]            [1.942]              [1.242]            [0.999]               [1.334]              [0.758]            [0.686]]
]

[fig format_performance.png..Performance comparison for a sequence of several items..spirit.karma.format_performance]

[endsect]

[endsect]