I ran a very simple test program (attached) on the Middlebury teddy data. When using cv::gpu::StereoConstantSpaceBP, the output most of the time contains artifacts and generally looks buggy (see attachment). I have made a few observations that lead me to think that it has to do with improper initialization:

1. Sometimes it just works, and once it does, it usually does for the rest of the session.
2. Yesterday, when I run StereoBeliefPropagation [[StereoBeliefPropagation]] first (in the same program), then StereoConstantSpaceBP, [[StereoConstantSpaceBP]], the latter seemed to work just fine. However, I cannot reproduce this behavior today. (StereoBeliefPropagation itself runs just fine)
3. When some other GPU code is run before, it also seems to work (this was only in one program, and I did not try to reproduce it)

I am running Fedora 15 with a GeForce [[GeForce]] GTX 470 (nvidia driver 280.13) and Cuda 4.0 (gcc 4.4)

However, the same behavior could be observed on two other machines, one of which is equipped with a GeForce [[GeForce]] 9600 GT.

I do not have a good enough overview of the OpenCV [[OpenCV]] GPU code to further pin down the error.