并行并行运行代码的结果不同>>

Question

运行代码时，我得到不同的结果。我想同时在不同线程中更改向量的不同单元格。当线程数多于单元数时发生错误（是的，我知道在这种情况下并行化不是最佳选择，但我想了解为什么结果不同）。

不起作用

w[i] -= speed * derivative[i]; 
有效

w[i] = w[i] - speed * derivative[i]; 
所有代码

#include <iostream>
#include <vector>
#include <numeric>
#include <string>
#include <iomanip>
#include <omp.h>
#include "windows.h"
using namespace std;

class neuron
{
public:
neuron() //constructor with starting filling for example
{
    w.resize(3);
    w[0] = 0.7700484068199707;
    w[1] = 0.3427844318371112;
    w[2] = 0.6822299620512033;
    derivative.resize(3);
    derivative[0] = 0.232342234223;
    derivative[1] = 0.23232232234223;
    derivative[2] = 0.23234334231113;
}

void correction_of_scales(const double& speed)
{
#pragma omp parallel for num_threads(8) //threads more than cells
    for (size_t i = 0; i < w.size(); ++i)
    {
#pragma omp atomic
        w[i] -= speed * derivative[i]; //here is here is happening what the magic
       //w[i] = w[i] - speed * derivative[i]; - and here it works without errors
    }
}

vector <double> derivative;
vector <double> w; 
};

int main(void)
{
neuron nn;
for (int i = 0; i < 100; i++) //a hundred times in a row change the vector
    nn.correction_of_scales(0.01);

for (int i = 0; i < nn.w.size(); i++)
    cout << scientific << setprecision(15) << nn.w[i] << endl;

system("pause");
}
这是您顺序运行时出现的内容

5.377061725969665e-01  
1.104621094948808e-01  
4.498866197400767e-01
但是并行运行时会发生什么

-6.24004998518318e-01  
1.104621094948808e-01  
4.498866197400767e-01

运行代码时，我得到不同的结果。我想同时在不同线程中更改向量的不同单元格。当线程数多于单元数时发生错误（是的，我知道...

Answer 1

这是Visual Studio的例外developercommunity.visualstudio.com