在发布模式下，在 Visual Studio (C++) 中测量循环内方法的运行时出错

Question

I have a method named function (see below) which has two inputs: an image and an integer which is an instruction to the function to do some operations on my image. Therefore, each number means a different operation.

for(int i=0;i<8;i++){

     auto start = std::chrono::high_resolution_clock::now();
     function(image, i);
     auto elapsed = std::chrono::high_resolution_clock::now() - start;
     long long microseconds =std::chrono::duration_cast<std::chrono::microseconds>(elapsed).count();
     cout << "i=" << i << ", time=" << microseconds << " micro seconds" << "  or "<<microseconds/(long double)1000000<<" seconds"<<"\n";

    }

Here is the output of the program:

i=0, time=357586 micro seconds  or 0.357586 seconds
i=1, time=15624 micro seconds  or 0.015624 seconds
i=2, time=15624 micro seconds  or 0.015624 seconds
i=3, time=15625 micro seconds  or 0.015625 seconds
i=4, time=15626 micro seconds  or 0.015626 seconds
i=5, time=15624 micro seconds  or 0.015624 seconds
i=6, time=15624 micro seconds  or 0.015624 seconds
i=7, time=15625 micro seconds  or 0.015625 seconds

However, if I change the order of instruction 1 and 0, I get the following:

i=1, time=348474 micro seconds  or 0.348474 seconds
i=0, time=15625 micro seconds  or 0.015625 seconds
i=2, time=15612 micro seconds  or 0.015612 seconds
i=3, time=15625 micro seconds  or 0.015625 seconds
i=4, time=15625 micro seconds  or 0.015625 seconds
i=5, time=15625 micro seconds  or 0.015625 seconds
i=6, time=15625 micro seconds  or 0.015625 seconds
i=7, time=15625 micro seconds  or 0.015625 seconds

In other words, first call of the function always takes more time irrespective of the type of operation. How can I fix this problem? My goal is to compare runtime of different operations and then discard those operations that are time-consuming.

Answer 1

Try declaring all of the variables globally instead of locally in the loop, which may be what is taking time. I imagine after it sets up the variables into registers, it re-uses them, so that's why you see the initial time spike. However it's also possible that the optimized time is because the function pointer is already set up in registers as well and it's just passing a new parameter into it each time. Heck for all I know it's because your CPU uses predictive pipelining, and halfway processed the function in the future iterations. It's difficult to say, but if you want to get a true running time, is it possible to just call your function with dummy parameters for the first iteration and discard the results?