Question

这是我尝试使用std::vector<double>和普通旧double*的循环。

对于1000万个元素，矢量版本始终在double*版本的大约80％的时间内运行;几乎任何N的值，矢量都明显更快。

在GCC STL源代码中窥视，我没有看到std::vector正在做的事情比double*成语正在进行的任何事情都要好（即用普通的new[]分配，operator[]取消引用偏移量）。 This question也会这样说。

为什么矢量版本更快的想法？

Compiler: GCC 4.6.1
Example compile line: g++ -Ofast -march=native -DNDEBUG \
                      -ftree-vectorizer-verbose=2 -o vector.bin \
                      vector.cpp -lrt
OS: CentOS 5
CPU: Opteron 8431
RAM: 128 GB

如果我使用icpc 11.1或在Xeon上运行，结果在质量上是相同的。此外，矢量化器转储表示只有std::vector构造函数中的填充操作被矢量化。

矢量版本：

#include <vector>
#include <iostream>
#include <boost/lexical_cast.hpp>
#include "util.h"
#include "rkck_params.h"

using namespace std;

int main( int argc, char* argv[] )
{
    const size_t N = boost::lexical_cast<size_t>( argv[ 1 ] );

    vector<double> y_old( N );
    vector<double> y_new( N );
    vector<double> y_err( N );
    vector<double> k0( N );
    vector<double> k1( N );
    vector<double> k2( N );
    vector<double> k3( N );
    vector<double> k4( N );
    vector<double> k5( N );

    const double h = 0.5;

    const timespec start = clock_tick();
    for ( size_t i = 0 ; i < N ; ++i )
    {
        y_new[ i ] =   y_old[ i ]
                     + h
                      *(
                           rkck::c[ 0 ]*k0[ i ]
                         + rkck::c[ 2 ]*k2[ i ]
                         + rkck::c[ 3 ]*k3[ i ]
                         + rkck::c[ 5 ]*k5[ i ]
                       );
        y_err[ i ] =  h
                     *(
                          rkck::cdiff[ 0 ]*k0[ i ]
                        + rkck::cdiff[ 2 ]*k2[ i ]
                        + rkck::cdiff[ 3 ]*k3[ i ]
                        + rkck::cdiff[ 4 ]*k4[ i ]
                        + rkck::cdiff[ 5 ]*k5[ i ]
                      );
    }
    const timespec stop = clock_tick();
    const double total_time = seconds( start, stop );

    // Output
    cout << "vector\t" << N << "\t" << total_time << endl;

    return 0;
}

double*版本：

#include <iostream>
#include <boost/lexical_cast.hpp>
#include "util.h"
#include "rkck_params.h"

using namespace std;

int main( int argc, char* argv[] )
{
    const size_t N = boost::lexical_cast<size_t>( argv[ 1 ] );

    double* y_old = new double[ N ];
    double* y_new = new double[ N ];
    double* y_err = new double[ N ];
    double* k0 = new double[ N ];
    double* k1 = new double[ N ];
    double* k2 = new double[ N ];
    double* k3 = new double[ N ];
    double* k4 = new double[ N ];
    double* k5 = new double[ N ];

    const double h = 0.5;

    const timespec start = clock_tick();
    for ( size_t i = 0 ; i < N ; ++i )
    {
        y_new[ i ]
            =   y_old[ i ]
              + h
               *(
                    rkck::c[ 0 ]*k0[ i ]
                  + rkck::c[ 2 ]*k2[ i ]
                  + rkck::c[ 3 ]*k3[ i ]
                  + rkck::c[ 5 ]*k5[ i ]
                );
        y_err[ i ]
            =  h
              *(
                   rkck::cdiff[ 0 ]*k0[ i ]
                 + rkck::cdiff[ 2 ]*k2[ i ]
                 + rkck::cdiff[ 3 ]*k3[ i ]
                 + rkck::cdiff[ 4 ]*k4[ i ]
                 + rkck::cdiff[ 5 ]*k5[ i ]
               );
    }
    const timespec stop = clock_tick();
    const double total_time = seconds( start, stop );

    delete [] y_old;
    delete [] y_new;
    delete [] y_err;
    delete [] k0;
    delete [] k1;
    delete [] k2;
    delete [] k3;
    delete [] k4;
    delete [] k5;

    // Output
    cout << "plain\t" << N << "\t" << total_time << endl;

    return 0;
}

rkck_params.h：

#ifndef RKCK_PARAMS_H
#define RKCK_PARAMS_H

namespace rkck
{

// C.f. $c_i$ in Ch. 16.2 of NR in C++, 2nd ed.
const double c[ 6 ]
    = { 37.0/378.0,
        0.0,
        250.0/621.0,
        125.0/594,
        0.0,
        512.0/1771.0 };

// C.f. $( c_i - c_i^* )$ in Ch. 16.2 of NR in C++, 2nd ed.
const double cdiff[ 6 ]
    = { c[ 0 ] - 2825.0/27648.0,
        c[ 1 ] - 0.0,
        c[ 2 ] - 18575.0/48384.0,
        c[ 3 ] - 13525.0/55296.0,
        c[ 4 ] - 277.0/14336.0,
        c[ 5 ] - 1.0/4.0 };

}

#endif

util.h：

#ifndef UTIL_H
#define UTIL_H

#include <time.h>
#include <utility>

inline timespec clock_tick()
{
    timespec tick;
    clock_gettime( CLOCK_REALTIME, &tick );
    return tick;
}

// \cite{www.guyrutenberg.com/2007/09/22/profiling-code-using-clock_gettime}
inline double seconds( const timespec& earlier, const timespec& later )
{
    double seconds_diff = -1.0;
    double nano_diff = -1.0;

    if ( later.tv_nsec < earlier.tv_nsec )
    {
        seconds_diff = later.tv_sec - earlier.tv_sec - 1;
        nano_diff = ( 1.0e9 + later.tv_nsec - earlier.tv_nsec )*1.0e-9;
    }
    else
    {
        seconds_diff = later.tv_sec - earlier.tv_sec;
        nano_diff = ( later.tv_nsec - earlier.tv_nsec )*1.0e-9;
    }

    return seconds_diff + nano_diff;
}

#endif

Answer 1

在矢量版本中，您的数据初始化为零。在new版本中它未初始化，因此可能会完成不同的工作。

您是否以不同的顺序多次运行？

vector <double>比double更快*：为什么？</double>

1 个答案: