The dEQP-GLES3.functional.fbo.blit.rect.nearest_consistency_* tests fail.
Setting the "address rounding" bits in SAMPLER_STATE, even for GL_NEAREST filtering, appears to fix these tests, but breaks others (e.g. dEQP-GLES3.functional.shaders.texture_functions.textureprojlod.usampler3d_vertex).
fixes all the tests, but it's immensely sketchy.
I ended up pushing that patch despite its sketchiness.
Author: Kenneth Graunke <email@example.com>
Date: Mon Mar 7 23:54:53 2016 -0800
i965: Set address rounding bits for GL_NEAREST filtering as well.
I had to revert the patch as it caused dEQP regressions on Sandybridge (apparently my CI build doesn't test those).
This was fixed by mesa 935ce14