With additional tests, we found out the problem comes out from the barycentric
coordinates. On one of the triangle, one barycentric coordinate is very
slightly below the expected value. It's not just on some pixels, but on all the
pixels of the triangle.

We also tested on an r600 card (hd3650), and the bug is the same. However for
SI and R600, the handling of the barycentric interpolation seems to have
evolved a lot.

Could it be some state that has been forgotten in the doc, that would specify
barycentric coordinates rounding ?