About `minMax`

[Nyrio]

That function is indeed quite inefficient. The overhead of allocations and unnecessary device synchronization is large compared to the relatively fast execution time of the kernel.

A few tricks to get rid of this overhead with only minor changes:

• Use cv::cuda::findMinMax instead of cv::cuda::minMax
• Provide a GPU matrix for the output argument, and reuse the allocation for this matrix
• Copy the results to pinned memory (e.g cv::cuda::HostMem) and reuse that allocation too
• Only synchronize when you need to read the host outputs. You want to be as asynchronous as possible to avoid launch and synchronization latencies.

The code using these tips looks like this:

cv::cuda::Stream stream;
cv::cuda::GpuMat minmax_g(1, 2, CV_32FC1);
cv::cuda::HostMem minmax_h(1, 2, CV_32FC1);
for (uint32_t i{0}; i < 10; ++i) {
  float *minValue = (float *)minmax_h.data;
  float *maxValue = (float *)(minmax_h.data + sizeof(float));
  cv::cuda::findMinMax(m, minmax_g, cv::noArray(), stream);
  minmax_g.download(minmax_h, stream);
  stream.waitForCompletion();
  std::cout << *minValue << " " << *maxValue << std::endl;
}

Before:

After:

You can see that there are still launch and synchronization overheads. For the sake of comparison, without the stream sync it looks like this (but the host pointers only contain correct values after sync):

On a side note, the reduction kernels are suboptimal. The two one-block one-thread initialization kernels incur launch latencies, and the main kernel has room for improvement too.

[nikitablack]

@Nyrio thank you for the update. The findMinMax function indeed performs better. Somehow I missed it, maybe because it's not documented.

[Nyrio]

@nikitablack minMax is using findMinMax, the only difference is that we're skipping the internal allocations, synchronization, etc ;)

It's listed in the docs although there aren't many details: https://docs.opencv.org/3.4/d5/de6/group__cudaarithm__reduce.html#gae7f5f2aa9f65314470a76fccdff887f2