I would like to report a likely bug with the CUDA NPPI MinIndx functions. This is the best place I could find to post it:
When many pixels are the same, it ends up returning a min value, an x and a y, but the value does not correspond to the value at that pixel. I think it must be taking the x from one pixel, and the y from another because they all share the same value and probably there is a race condition.