Hi,
I traced the code in the function and I think the subpixel calculation is wrong (offset by 1).
Here's the detail:
xrange_min is 0, width is 480, minD is 0, maxX2 is 480.
So line 194 (https://github.com/opencv/opencv/blob/master/modules/calib3d/src/stereosgbm.cpp#L194) width-1-maxX2 is actually -1.
So in line 204, the x in the loop starts at -1. This is fine for 'buffer' since it has adjusted at line 194, but for prow2 this is not right.
For example, the first buffer array (e.g. buffer[-1]) should be min of (prow2[0], prow2[0], (prow2[0]+prow2[1]/2), but it's now the min of (prow2[0], prow2[-1], (prow2[-1]+prow2[0]/2). All the rest is the same thing.