如何确保一个 OpenMP 并行部分不会影响另一个并行部分？

Question

我写了一个基本程序，通过进入while循环并将循环内的时间与启动前的时间进行比较，将FPS降低到60 FPS左右，然后将FPS输出到控制台。

在同一个程序中，我有另一个代码块，可以使用类似的方法将FPS降低到10 FPS左右。

我想做的是使用OpenMP并行执行两个代码块，这样10 FPS块就不会减慢或干扰60FPS块。我的最终目标是让60 FPS的块以60 FPS的速度运行，即使程序中有一个循环，否则会将其降低到10 FPS。

我尝试过使用几种不同的OpenMP指令组合来实现这一点，但无论我使用什么指令，程序似乎总是以10 FPS的速度运行。

有可能将其与OpenMP一起使用吗？

这是代码：(编辑：我已经删除了dreamcrash提到的比赛条件，并修复了我注意到的其他一些错误(

#include <iostream>
#include <windows.h>
#include <stdint.h>
#include <omp.h>
int main() {

// Get the QueryPerformance frequency to measure the FPS
LARGE_INTEGER PerfCountFrequencyResult;
QueryPerformanceFrequency(&PerfCountFrequencyResult);
int64_t PerfCountFrequency = PerfCountFrequencyResult.QuadPart;

// 10 FPS for first thread
double Thread1TargetSecondsPerFrame = 1.0f / (double)10;

// 60 FPS for second thread
double Thread2TargetSecondsPerFrame = 1.0f / (double)60;

// Get Current time
LARGE_INTEGER LastCounter;
QueryPerformanceCounter(&LastCounter);


while(1){
#pragma omp parallel sections
{
// THREAD 1 (very slow 10FPS)
#pragma omp section
{
// Remove race condition
double SecondsElapsedForFrame;

// Get the seconds elapsed since last loop
LARGE_INTEGER CurrentCounter;
QueryPerformanceCounter(&CurrentCounter);
SecondsElapsedForFrame = ((float)(CurrentCounter.QuadPart - LastCounter.QuadPart) / (float)PerfCountFrequency);

// DELAY The time until it's 10FPS
if(SecondsElapsedForFrame < Thread1TargetSecondsPerFrame) {
while(SecondsElapsedForFrame < Thread1TargetSecondsPerFrame) {
// Get the seconds elapsed since last loop
QueryPerformanceCounter(&CurrentCounter);
SecondsElapsedForFrame = ((float)(CurrentCounter.QuadPart - LastCounter.QuadPart) / (float)PerfCountFrequency);
}           
}

}

// THEAD 2 (should not be interrupted by THREAD 1)
#pragma omp section
{
// Remove race condition
double SecondsElapsedForFrame;

// Get the seconds elapsed since last loop
LARGE_INTEGER CurrentCounter;
QueryPerformanceCounter(&CurrentCounter);
SecondsElapsedForFrame = ((float)(CurrentCounter.QuadPart - LastCounter.QuadPart) / (float)PerfCountFrequency);

// DELAY the time until it's 60FPS
if(SecondsElapsedForFrame < Thread2TargetSecondsPerFrame) {
while(SecondsElapsedForFrame < Thread2TargetSecondsPerFrame) {
// Get the seconds elapsed since last loop
QueryPerformanceCounter(&CurrentCounter);
SecondsElapsedForFrame = ((float)(CurrentCounter.QuadPart - LastCounter.QuadPart) / (float)PerfCountFrequency);
}                   
}

// Get the seconds elapsed and output the FPS
QueryPerformanceCounter(&CurrentCounter);       
double SecondsPerFrame = ((float)(CurrentCounter.QuadPart - LastCounter.QuadPart) / (float)PerfCountFrequency);
float FramesPerSecond = 1.0f/ SecondsPerFrame;
std::cout << "FPS:" << FramesPerSecond << "n";


// Get new time for beginning of loop
QueryPerformanceCounter(&LastCounter);  
}

}
}
return(0);
}

Answer 1

首先是正确性的代码，然后是performance。

您的OpenMP程序是格式错误的，因为它包含SecondsElapsedForFrame上的竞赛条件。事实上，变量在两个线程之间的并行区域中是隐式共享的，并且都写入其中。这也适用于LastCounter。根据您的需要，您可能希望将它们制成private、firstprivate或lastprivate。