求解惑，double 与 float 的运算速度测试，为什么是这样？
重复执行4次一样的代码，只有第一次的时候，double与float有明显区别，后面三次区别就很小了

#include <iostream>
#include <math.h>
#include <chrono>


using namespace std;


class TicToc
{
public:
    TicToc(const std::string& name):m_name(name) { tic(); }

    void tic() {
        start = std::chrono::system_clock::now();
    }

    void toc() {
        auto millinon_second = std::chrono::duration_cast<std::chrono::milliseconds>
                (std::chrono::system_clock::now()-start);
        std::cout << m_name << " cost " << millinon_second.count() << " ms." << std::endl;
    }

private:
    std::string m_name;
    std::chrono::time_point<std::chrono::system_clock> start, end;
};



constexpr size_t N = 20000000;
double result_double[N];
float  result_float[N];

int main()
{
    float pi_float = 3.1415926f;
    double pi_double = 3.1415926;

    TicToc double_time("double multiply ");
    for (size_t i = 0; i < N; ++i) {
        result_double[i] = double(i) * pi_double;    
    }
    double_time.toc();

    TicToc float_time("float multiply ");
    for (size_t i = 0; i < N; ++i) {
        result_float[i] = float(i) * pi_float;    
    }
    float_time.toc();
    /////////////////////////////////////////

    TicToc double_time2("double multiply 2");
    for (size_t i = 0; i < N; ++i) {
        result_double[i] = double(i) * pi_double;    
    }
    double_time2.toc();

    TicToc float_time2("float multiply 2");
    for (size_t i = 0; i < N; ++i) {
        result_float[i] = float(i) * pi_float;    
    }
    float_time2.toc();
    /////////////////////////////////////////

    TicToc double_time3("double multiply 3");
    for (size_t i = 0; i < N; ++i) {
        result_double[i] = double(i) * pi_double;    
    }
    double_time3.toc();

    TicToc float_time3("float multiply 3");
    for (size_t i = 0; i < N; ++i) {
        result_float[i] = float(i) * pi_float;    
    }
    float_time3.toc();

    /////////////////////////////////////////

    TicToc double_time4("double multiply 4");
    for (size_t i = 0; i < N; ++i) {
        result_double[i] = double(i) * pi_double;    
    }
    double_time4.toc();

    TicToc float_time4("float multiply 4");
    for (size_t i = 0; i < N; ++i) {
        result_float[i] = float(i) * pi_float;    
    }
    float_time4.toc();

}



编译： g++ float_double.cpp -std=c++11 -O3
四次运行结果：
xxxx$ ./a.out
double multiply  cost 73 ms.
float multiply  cost 28 ms.
double multiply 2 cost 11 ms.
float multiply 2 cost 9 ms.
double multiply 3 cost 11 ms.
float multiply 3 cost 14 ms.
double multiply 4 cost 14 ms.
float multiply 4 cost 14 ms.

xxxx$ ./a.out
double multiply  cost 51 ms.
float multiply  cost 28 ms.
double multiply 2 cost 12 ms.
float multiply 2 cost 9 ms.
double multiply 3 cost 11 ms.
float multiply 3 cost 14 ms.
double multiply 4 cost 14 ms.
float multiply 4 cost 14 ms.

xxxx$ ./a.out
double multiply  cost 51 ms.
float multiply  cost 28 ms.
double multiply 2 cost 11 ms.
float multiply 2 cost 9 ms.
double multiply 3 cost 11 ms.
float multiply 3 cost 14 ms.
double multiply 4 cost 14 ms.
float multiply 4 cost 14 ms.

xxxx$ ./a.out
double multiply  cost 78 ms.
float multiply  cost 28 ms.
double multiply 2 cost 11 ms.
float multiply 2 cost 9 ms.
double multiply 3 cost 11 ms.
float multiply 3 cost 14 ms.
double multiply 4 cost 14 ms.
float multiply 4 cost 13 ms.



平台：Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz



【 在 foliver 的大作中提到: 】
: 和寄存器指令有关最开始。
: 最开始浮点数都是fpu计算，fpu寄存器都是80bits大小，不管float还是double，没有区别，指令也一样。
: 后来2000年前后，intel推广sse，引入xmm寄存器，遵从float/double的大小，使用不同的指令。这时的float才是真float。
: ...................
--
FROM 14.155.19.*

汇编不怎么会玩。。。以前翻过几页而已  我测试 float 跟 double 计算速度好像没区别 ？ 我记得什么资料上或者版上有看到过说现在的cpu，这两种没什么区别

【 在 z16166 的大作中提到: 】
: 大胆推测，小心求证
: 你的推测是运行时分配内存，就不求证了？而且这个很容易求证，同样是看汇编代码。熟手可以不用看。
: 楼上也提示了，可能是cache问题
: ...................
--
修改:confinement FROM 14.155.19.*
FROM 14.155.19.*

是下面这样吗？  测试结果几乎没有什么变化。。。编译需要加什么flag吗

#include <iostream>
#include <math.h>
#include <chrono>
#include <string.h>



using namespace std;


class TicToc
{
public:
    TicToc(const std::string& name):m_name(name) { tic(); }

    void tic() {
        start = std::chrono::system_clock::now();
    }

    void toc() {
        auto millinon_second = std::chrono::duration_cast<std::chrono::milliseconds>
                (std::chrono::system_clock::now()-start);
        std::cout << m_name << " cost " << millinon_second.count() << " ms." << std::endl;
    }

private:
    std::string m_name;
    std::chrono::time_point<std::chrono::system_clock> start, end;
};



constexpr size_t N = 20000000;
constexpr size_t Nf = 20000000 - 4;
constexpr size_t Nd = 20000000 - 2;
double result_double[N];
float  result_float[N];
double pi_double = 3.1415926;
float  pi_float = 3.1415926f;

int main()
{

memset(result_double, 0, sizeof(result_double));
memset(result_float, 0, sizeof(result_float));

TicToc float_time1("float multiply 1");
for (size_t i = 0; i < N; ++i) {
result_float[i] = float(i) * pi_float;
}
float_time1.toc();

TicToc double_time1("double multiply 1");
for (size_t i = 0; i < N; ++i) {
result_double[i] = double(i) * pi_double;
}
double_time1.toc();

/////////////////////////////////////////

memset(result_double, 0, sizeof(result_double));
memset(result_float, 0, sizeof(result_float));

TicToc float_time2("float multiply 2");
for (size_t i = 0; i < N; ++i) {
result_float[i] = float(i) * pi_float;
}
float_time2.toc();

TicToc double_time2("double multiply 2");
for (size_t i = 0; i < N; ++i) {
result_double[i] = double(i) * pi_double;
}
double_time2.toc();

/////////////////////////////////////////

memset(result_double, 0, sizeof(result_double));
memset(result_float, 0, sizeof(result_float));

TicToc float_time3("float multiply 3");
for (size_t i = 0; i < Nf; i += 4) {
result_float[i] = float(i) * pi_float;
result_float[i+1] = float(i+1) * pi_float;
result_float[i+2] = float(i+2) * pi_float;
result_float[i+3] = float(i+3) * pi_float;
}
float_time3.toc();

TicToc double_time3("double multiply 3");
for (size_t i = 0; i < Nd; i += 2) {
result_double[i] = double(i) * pi_double;
result_double[i+1] = double(i+1) * pi_double;
}
double_time3.toc();

/////////////////////////////////////////

memset(result_double, 0, sizeof(result_double));
memset(result_float, 0, sizeof(result_float));

TicToc float_time4("float multiply 4");
for (size_t i = 0; i < Nf; i += 4) {
result_float[i] = float(i) * pi_float;
result_float[i+1] = float(i+1) * pi_float;
result_float[i+2] = float(i+2) * pi_float;
result_float[i+3] = float(i+3) * pi_float;
}
float_time4.toc();

TicToc double_time4("double multiply 4");
for (size_t i = 0; i < Nd; i += 2) {
result_double[i] = double(i) * pi_double;
result_double[i+1] = double(i+1) * pi_double;
}
double_time4.toc();


}


【 在 foliver 的大作中提到: 】
: 直接把i++变成i+=4。循环里面连续计算4个，编译器就会合并计算。
--
修改:confinement FROM 14.155.19.*
FROM 14.155.19.*