gccとg++でcomplex.hを使った場合の速度を調べてみた

00_readme.md

CプログラムってC++コンパイラ(g++)でもコンパイルできるよね， その場合，生成されるコードってどのくらい違うのだろう？実行時間は？ という疑問に対して複素数扱うプログラムの場合で調べてみた，結果． 関連してC++のcomplexを使う場合についての実行時間も調べてみた．

結果，CプログラムをC++コンパイラ(g++)でコンパイルした場合， 関数名が違うくらいの，ほぼ同じコードが生成された． 一方で，なぜかg++で生成したバイナリの実行時間は gccで生成したバイナリの実行時間より長くなる傾向が見られた（なぜ？）

なお，デスクトップLinux環境上で適当に測定してるので，実行時間にゆらぎはある．

• 01_env - 使用環境
• complex_c_func.c - 評価対象のカーネル
• complex_c.c - 評価するためのガワ(complex_c_func.cをincludeしてる)
• complex_c_func_gcc_g++.diff gccとg++で生成されるアセンブラコードの比較
• complex_cpp.cpp - C++のcomplexを使用する場合
• diff.sh - complex_c_func.cをgccとg++でコンパイルして生成されるアセンブラを比較する
• test.sh - complex_c.cをgccとg++で，complex_cpp.cppをg++でコンパイルして実行時間を比較
• elapsed_time_comparison.txt - test.shの実行結果
• test2.sh - test.shに対して実行順序を変更
• gistfile1.txt - test2.shの実行結果
• test3.sh - test.shに対して実行順序を変更
• gistfile2.txt - test3.shの実行結果

01_env

Intel(R) Core(TM) i3-4160 CPU @ 3.60GHz
MemTotal:       16279924 kB
OS: Ubuntu 18.04.3 LTS
compiler: gcc (Ubuntu 7.4.0-1ubuntu1~18.04.1) 7.4.0, g++ (Ubuntu 7.4.0-1ubuntu1~18.04.1) 7.4.0

complex_c.c

#include <stdio.h>
#include <stdlib.h>
#include <complex.h>
#include <sys/time.h>

double time_in_sec(){
	struct timeval t;
	gettimeofday(&t, NULL);
	return ((double)t.tv_sec) + ((double)t.tv_usec) / 1000000.0;
}

#include "complex_c_func.c"

double _Complex dst;
int main(int argc, char **argv){
	if(argc < 3){
		printf("usage %s nums repetitions\n", argv[0]);
		return 0;
	}
	int n = atoi(argv[1]);
	int m = atoi(argv[2]);
        if(argc >= 4){
            srand(atoi(argv[3]));
        }
        double min = 0;
        double max = 0;
        double t = 0;
	for(int j = 0; j < m; j++){
		double _Complex* src0;
		double _Complex* src1;
		src0 = (double _Complex*)malloc(sizeof(double _Complex) * n);
		src1 = (double _Complex*)malloc(sizeof(double _Complex) * n);
		for(int i = 0; i < n; i++){
			double r;
			r = (double)rand()/RAND_MAX;
			src0[i] = r + cexp(I*3.14);
			r = (double)rand()/RAND_MAX;
			src1[i] = r + cexp(I*3.14);
		}
	
		double t0 = time_in_sec();
		func(src0, src1, &dst, n);
		free(src0);
		free(src1);
		double t1 = time_in_sec();
                double tt = t1 - t0;
        	t += tt;
                min = (min == 0) || (tt < min) ? tt : min;
                max = (max == 0) || (tt > max) ? tt : max;
#ifdef VERBOSE
		printf("%f + %f * i\n", creal(dst), cimag(dst));
		printf("%f\n", tt);
#endif
        }
	printf("%s: n=%d, avg=%f, max=%f, min=%f, (%g,%g)\n", argv[0], n, (t/m), max, min, creal(dst), cimag(dst));
}

complex_c_func.c

void func(double _Complex* src0, double _Complex* src1, double _Complex *dst, int n){
	double _Complex sum = 0;
	for(int i = 0; i < n; i++){
#ifdef VERBOSE
	printf("(%f+%fi) * (%f+%fi)\n", creal(src0[i]), cimag(src1[i]), creal(src1[i]), cimag(src1[i]));
#endif
		sum += src0[i] * src1[i];
	}
	*dst = sum;
}

complex_c_func_gcc_g++.diff

4,6c4,6
< 	.globl	func
< 	.type	func, @function
< func:
---
> 	.globl	_Z4funcPCdS0_S0_i
> 	.type	_Z4funcPCdS0_S0_i, @function
> _Z4funcPCdS0_S0_i:
75c75
< 	.size	func, .-func
---
> 	.size	_Z4funcPCdS0_S0_i, .-_Z4funcPCdS0_S0_i

complex_cpp.cpp

#include <iostream>
#include <iomanip>
#include <complex>
#include <cmath>
 
#include <sys/time.h>
#include <stdlib.h>

double time_in_sec(){
	struct timeval t;
	gettimeofday(&t, NULL);
	return ((double)t.tv_sec) + ((double)t.tv_usec) / 1000000.0;
}

void func(std::complex<double>* src0, std::complex<double>* src1, std::complex<double>& dst, int n){
	std::complex<double> sum = 0;
	for(int i = 0; i < n; i++){
#ifdef VERBOSE
        std:cout << src0[i] << " * " << src1[i] << std::endl;
#endif
		sum += src0[i] * src1[i];
	}
	dst = sum;
}

std::complex<double> dst;
int main(int argc, char **argv)
{
	if(argc < 3){
		printf("usage %s nums repetitions\n", argv[0]);
		return 0;
	}
	int n = atoi(argv[1]);
	int m = atoi(argv[2]);
        if(argc >= 4){
            srand(atoi(argv[3]));
        }
        double min = 0;
        double max = 0;
        double t = 0;
	for(int j = 0; j < m; j++){
            std::complex<double> *src0;
            std::complex<double> *src1;
            src0 = new std::complex<double>[n];
            src1 = new std::complex<double>[n];
            
            using namespace std::complex_literals;
            //std::cout << std::fixed << std::setprecision(1);
            
            for(int k = 0; k < n; k++){
		double r;
		r = (double)rand()/RAND_MAX;
		src0[k] = r + std::exp(1i*3.14);
		r = (double)rand()/RAND_MAX;
		src1[k] = r + std::exp(1i*3.14);
            }
            
            double t0 = time_in_sec();
            func(src0, src1, dst, n);
            double t1 = time_in_sec();
            double tt = t1 - t0;
            min = (min == 0) || (tt < min) ? tt : min;
            max = (max == 0) || (tt > max) ? tt : max;
            t += tt;
            delete[] src0;
            delete[] src1;
#ifdef VERBOSE
            std::cout << "" << dst << '\n';
            std::cout << "" << tt << '\n';
#endif
        }
        std::cout << argv[0] << ": "
                  << "n=" << n
                  << ", avg=" << (t/m)
                  << ", max=" << max
                  << ", min=" << min
                  << ", " << dst << std::endl;
}

diff.sh

gcc -O3 -S complex_c_func.c
mv complex_c_func.s complex_c_func_gcc.s
g++ -O3 -S complex_c_func.c
mv complex_c_func.s complex_c_func_g++.s
diff complex_c_func_gcc.s complex_c_func_g++.s > complex_c_func_gcc_g++.diff

elapsed_time_comparison.txt

miyo@grappa:% sh test.sh
./complex_c_gcc: n=500000, avg=0.001950, max=0.002074, min=0.001921, (124783,-795.622)
./complex_c_gcc: n=1000000, avg=0.003830, max=0.004015, min=0.003797, (249785,-1592.31)
./complex_c_gcc: n=5000000, avg=0.020895, max=0.021260, min=0.020601, (1.25049e+06,-7963.74)
./complex_c_gcc: n=10000000, avg=0.041484, max=0.042116, min=0.041181, (2.50075e+06,-15928.6)
./complex_c_gcc: n=50000000, avg=0.192739, max=0.193341, min=0.192106, (1.24998e+07,-79630.9)
./complex_c_gcc: n=100000000, avg=0.376987, max=0.377422, min=0.376400, (2.50008e+07,-159265)
./complex_c_gcc: n=500000000, avg=1.856979, max=1.882624, min=1.849500, (1.25001e+08,-796346)
./complex_c_g++: n=500000, avg=0.002454, max=0.002691, min=0.002411, (124783,-795.622)
./complex_c_g++: n=1000000, avg=0.004929, max=0.005160, min=0.004794, (249785,-1592.31)
./complex_c_g++: n=5000000, avg=0.025850, max=0.025954, min=0.025231, (1.25049e+06,-7963.74)
./complex_c_g++: n=10000000, avg=0.049112, max=0.049733, min=0.048610, (2.50075e+06,-15928.6)
./complex_c_g++: n=50000000, avg=0.231742, max=0.234580, min=0.231187, (1.24998e+07,-79630.9)
./complex_c_g++: n=100000000, avg=0.453577, max=0.455913, min=0.453133, (2.50008e+07,-159265)
./complex_c_g++: n=500000000, avg=2.234970, max=2.247742, min=2.229900, (1.25001e+08,-796346)
./complex_c++_g++: n=500000, avg=0.00175731, max=0.00213313, min=0.00170398, (124783,-795.622)
./complex_c++_g++: n=1000000, avg=0.00348706, max=0.00371003, min=0.00343704, (249785,-1592.31)
./complex_c++_g++: n=5000000, avg=0.0200314, max=0.0203891, min=0.01916, (1.25049e+06,-7963.74)
./complex_c++_g++: n=10000000, avg=0.0376434, max=0.0377338, min=0.0375719, (2.50075e+06,-15928.6)
./complex_c++_g++: n=50000000, avg=0.18886, max=0.189569, min=0.188623, (1.24998e+07,-79630.9)
./complex_c++_g++: n=100000000, avg=0.379331, max=0.379741, min=0.37898, (2.50008e+07,-159265)
./complex_c++_g++: n=500000000, avg=1.88151, max=1.88634, min=1.87581, (1.25001e+08,-796346)

gistfile1.txt

miyo@grappa:% sh test2.sh 
./complex_c++_g++: n=500000, avg=0.00173452, max=0.00189805, min=0.00170898, (124783,-795.622)
./complex_c_g++: n=500000, avg=0.002419, max=0.002482, min=0.002402, (124783,-795.622)
./complex_c_gcc: n=500000, avg=0.002082, max=0.002104, min=0.001978, (124783,-795.622)
./complex_c++_g++: n=1000000, avg=0.00348315, max=0.00380111, min=0.00342607, (249785,-1592.31)
./complex_c_g++: n=1000000, avg=0.004700, max=0.004824, min=0.004675, (249785,-1592.31)
./complex_c_gcc: n=1000000, avg=0.004192, max=0.004372, min=0.004151, (249785,-1592.31)
./complex_c++_g++: n=5000000, avg=0.0193249, max=0.0195231, min=0.0191698, (1.25049e+06,-7963.74)
./complex_c_g++: n=5000000, avg=0.024368, max=0.024399, min=0.024339, (1.25049e+06,-7963.74)
./complex_c_gcc: n=5000000, avg=0.020679, max=0.020736, min=0.020643, (1.25049e+06,-7963.74)
./complex_c++_g++: n=10000000, avg=0.0375857, max=0.0376141, min=0.0375359, (2.50075e+06,-15928.6)
./complex_c_g++: n=10000000, avg=0.048546, max=0.048622, min=0.048522, (2.50075e+06,-15928.6)
./complex_c_gcc: n=10000000, avg=0.041034, max=0.041125, min=0.041005, (2.50075e+06,-15928.6)
./complex_c++_g++: n=50000000, avg=0.189053, max=0.189582, min=0.188828, (1.24998e+07,-79630.9)
./complex_c_g++: n=50000000, avg=0.231149, max=0.231402, min=0.231043, (1.24998e+07,-79630.9)
./complex_c_gcc: n=50000000, avg=0.193406, max=0.193582, min=0.193232, (1.24998e+07,-79630.9)
./complex_c++_g++: n=100000000, avg=0.377249, max=0.380063, min=0.375594, (2.50008e+07,-159265)
./complex_c_g++: n=100000000, avg=0.454100, max=0.455543, min=0.453699, (2.50008e+07,-159265)
./complex_c_gcc: n=100000000, avg=0.378095, max=0.379139, min=0.377263, (2.50008e+07,-159265)
./complex_c++_g++: n=500000000, avg=1.88809, max=1.95081, min=1.8729, (1.25001e+08,-796346)
./complex_c_g++: n=500000000, avg=2.254588, max=2.269981, min=2.238011, (1.25001e+08,-796346)
./complex_c_gcc: n=500000000, avg=1.854065, max=1.858170, min=1.846826, (1.25001e+08,-796346)

gistfile2.txt

miyo@grappa:% sh test3.sh 
./complex_c_g++: n=500000, avg=0.002481, max=0.002888, min=0.002422, (124783,-795.622)
./complex_c++_g++: n=500000, avg=0.0017477, max=0.00199914, min=0.00170898, (124783,-795.622)
./complex_c_gcc: n=500000, avg=0.002112, max=0.002212, min=0.002078, (124783,-795.622)
./complex_c_g++: n=1000000, avg=0.005018, max=0.005115, min=0.004970, (249785,-1592.31)
./complex_c++_g++: n=1000000, avg=0.00345421, max=0.00377107, min=0.00340796, (249785,-1592.31)
./complex_c_gcc: n=1000000, avg=0.004191, max=0.004651, min=0.003915, (249785,-1592.31)
./complex_c_g++: n=5000000, avg=0.024869, max=0.025046, min=0.024569, (1.25049e+06,-7963.74)
./complex_c++_g++: n=5000000, avg=0.0197668, max=0.0200329, min=0.019223, (1.25049e+06,-7963.74)
./complex_c_gcc: n=5000000, avg=0.020821, max=0.020868, min=0.020795, (1.25049e+06,-7963.74)
./complex_c_g++: n=10000000, avg=0.048805, max=0.049189, min=0.048685, (2.50075e+06,-15928.6)
./complex_c++_g++: n=10000000, avg=0.0376932, max=0.037771, min=0.0376091, (2.50075e+06,-15928.6)
./complex_c_gcc: n=10000000, avg=0.041294, max=0.041545, min=0.041143, (2.50075e+06,-15928.6)
./complex_c_g++: n=50000000, avg=0.232907, max=0.233014, min=0.232796, (1.24998e+07,-79630.9)
./complex_c++_g++: n=50000000, avg=0.188717, max=0.188896, min=0.188546, (1.24998e+07,-79630.9)
./complex_c_gcc: n=50000000, avg=0.194860, max=0.194969, min=0.194785, (1.24998e+07,-79630.9)
./complex_c_g++: n=100000000, avg=0.455017, max=0.455631, min=0.453399, (2.50008e+07,-159265)
./complex_c++_g++: n=100000000, avg=0.377277, max=0.379055, min=0.376125, (2.50008e+07,-159265)
./complex_c_gcc: n=100000000, avg=0.377480, max=0.378683, min=0.376184, (2.50008e+07,-159265)
./complex_c_g++: n=500000000, avg=2.236948, max=2.256663, min=2.227983, (1.25001e+08,-796346)
./complex_c++_g++: n=500000000, avg=1.91338, max=1.94352, min=1.87593, (1.25001e+08,-796346)
./complex_c_gcc: n=500000000, avg=1.853870, max=1.856689, min=1.851566, (1.25001e+08,-796346)

test.sh

#!/bin/sh

gcc -O3 -o complex_c_gcc complex_c.c
for i in 500000 1000000 5000000 10000000 50000000 100000000 500000000
do
  ./complex_c_gcc $i 10
done

g++ -O3 -o complex_c_g++ complex_c.c
for i in 500000 1000000 5000000 10000000 50000000 100000000 500000000
do
  ./complex_c_g++ $i 10
done

g++ -O3 -std=c++14 -o complex_c++_g++ complex_cpp.cpp
for i in 500000 1000000 5000000 10000000 50000000 100000000 500000000
do
./complex_c++_g++ $i 10
done

test2.sh

#!/bin/sh
g++ -O3 -std=c++14 -o complex_c++_g++ complex_cpp.cpp
for i in 500000 1000000 5000000 10000000 50000000 100000000 500000000
do
  ./complex_c++_g++ $i 10
  ./complex_c_g++ $i 10
  ./complex_c_gcc $i 10
done

test3.sh

#!/bin/sh
g++ -O3 -std=c++14 -o complex_c++_g++ complex_cpp.cpp
for i in 500000 1000000 5000000 10000000 50000000 100000000 500000000
do
  ./complex_c_g++ $i 10
  ./complex_c++_g++ $i 10
  ./complex_c_gcc $i 10
done

https://gist.github.com/miyo/e8200ce82e1732a9cd4b2f21764f3b20#file-elapsed_time_comparison-txt をグラフにしてみた

test2.shのような順序で実行した場合

test3.shのような順序で実行した場合