Acessing CUDA unified memory after OpenCV GPU operation

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

I had a question regarding CUDA unified memory:

Say I have a pointer and allocate with unified memory as follows:

unsigned char* ptr;

cudaMallocManaged(&ptr);

Then I set OpenCV Mat and GpuMat to the same unified memory location:

cv::Mat cpuMat(size, ptr);
cv::cuda::GpuMat gpuMat(size, ptr);

At this point I can modify the pointer manually, like ptr[0] = 1;
This changes the memory in both GPU and CPU, and the data is accessible/changeable from both CPU and CPU.

However, I do some openCV processing such as follows:

cv::GpuMat gpuMatOut(size);

cv::cuda::remap(gpuMat, gpuMatOut, map1, map2);

After running an OpenCV operation such as cv::cuda::remap, the data is no longer accessible in cpuMat, even though the memory location is still the same. The only way to access the memory again is to perform a copy operation such as:

gpuMat.download(cpuMat);

Does anyone know why the unified memory is no longer CPU accessible after the OpenCV operation?

Thanks!