增加OpenMP中用于进度报告的共享循环计数器

Increment shared loop counter in OpenMP for progress reporting

本文关键字：共享 循环 计数器 报告 OpenMP 用于 增加      更新时间：2023-10-16

我想跟踪由长时间运行的光线跟踪过程处理的总像素和光线。如果我每次迭代都更新共享变量，那么由于同步的原因，这个过程会明显减慢。我想跟踪进度，最后仍然能得到准确的计数结果。有没有一种方法可以使用OpenMP for循环来实现这一点？

以下是有问题的循环的一些代码：

void Raytracer::trace(RenderTarget& renderTarget, const Scene& scene, std::atomic<int>& sharedPixelCount, std::atomic<int>& sharedRayCount)
{
    int width = renderTarget.getWidth();
    int height = renderTarget.getHeight();
    int totalPixelCount = width * height;
    #pragma omp parallel for schedule(dynamic, 4096)
    for (int i = 0; i < totalPixelCount; ++i)
    {
        int x = i % width;
        int y = i / width;
        Ray rayToScene = scene.camera.getRay(x, y);
        shootRay(rayToScene, scene, sharedRayCount); // will increment sharedRayCount
        renderTarget.setPixel(x, y, rayToScene.color.clamped());
        ++sharedPixelCount;
    }
}

既然动态调度的并行for循环的块大小为4096，为什么不将其用作计数器更新的粒度呢？

例如，以下内容可能会起作用。我没有测试这个代码，您可能需要为totalPixelCount%4096!=0添加一些记账。

与前面的答案不同，这不会向循环添加分支，而不是循环本身所暗示的分支，许多处理器都为其优化了指令。它也不需要任何额外的变量或算术运算。

void Raytracer::trace(RenderTarget& renderTarget, const Scene& scene, std::atomic<int>& sharedPixelCount, std::atomic<int>& sharedRayCount)
{
    int width = renderTarget.getWidth();
    int height = renderTarget.getHeight();
    int totalPixelCount = width * height;
    #pragma omp parallel for schedule(dynamic, 1)
    for (int j = 0; j < totalPixelCount; j+=4096)
    {
      for (int i = j; i < (i+4096); ++i)
      {
        int x = i % width;
        int y = i / width;
        Ray rayToScene = scene.camera.getRay(x, y);
        shootRay(rayToScene, scene, sharedRayCount);
        renderTarget.setPixel(x, y, rayToScene.color.clamped());
      }
      sharedPixelCount += 4096;
    }
}

现在还不清楚为什么sharedPixelCount需要在这个循环内部更新，因为它在循环体中没有被引用。如果这是正确的，我建议如下。

void Raytracer::trace(RenderTarget& renderTarget, const Scene& scene, std::atomic<int>& sharedPixelCount, std::atomic<int>& sharedRayCount)
{
    int width = renderTarget.getWidth();
    int height = renderTarget.getHeight();
    int totalPixelCount = width * height;
    int reducePixelCount = 0;
    #pragma omp parallel for schedule(dynamic, 4096) 
                         reduction(+:reducePixelCount) 
                         shared(reducePixelCount)
    for (int i = 0; i < totalPixelCount; ++i)
    {
        int x = i % width;
        int y = i / width;
        Ray rayToScene = scene.camera.getRay(x, y);
        shootRay(rayToScene, scene, sharedRayCount);
        renderTarget.setPixel(x, y, rayToScene.color.clamped());
        ++reducePixelCount; /* thread-local operation, not atomic */
    }
    /* The interoperability of C++11 atomics and OpenMP is not defined yet,
     * so this should just be avoided until OpenMP 5 at the earliest. 
     * It is sufficient to reduce over a non-atomic type and 
     * do the assignment here. */
    sharedPixelCount = reducePixelCount;
}

下面是一个如何做到这一点的示例：

void Raytracer::trace(RenderTarget& renderTarget, const Scene& scene, std::atomic<int>& sharedPixelCount, std::atomic<int>& sharedRayCount)
{
    int width = renderTarget.getWidth();
    int height = renderTarget.getHeight();
    int totalPixelCount = width * height;
    int rayCount = 0;
    int previousRayCount = 0;
    #pragma omp parallel for schedule(dynamic, 1000) reduction(+:rayCount) firstprivate(previousRayCount)
    for (int i = 0; i < totalPixelCount; ++i)
    {
        int x = i % width;
        int y = i / width;
        Ray rayToScene = scene.camera.getRay(x, y);
        shootRay(rayToScene, scene, rayCount);
        renderTarget.setPixel(x, y, rayToScene.color.clamped());
        if ((i + 1) % 100 == 0)
        {
            sharedPixelCount += 100;
            sharedRayCount += (rayCount - previousRayCount);
            previousRayCount = rayCount;
        }
    }
    sharedPixelCount = totalPixelCount;
    sharedRayCount = rayCount;
}

当循环运行时，它不会100%准确，但误差可以忽略不计。最后将报告精确值。

相关文章：