Hi Ilia, On Friday, November 14, 2014 15:48:46 Ilia Mirkin wrote: > I don't think I've tested this since v2, but an interesting > observation is that on nvidia hw, if I force clip_halfz to 0, the v2 > test passed, but if I force it to 1, then some of the things fail. > Could there be another clipping stage, e.g. viewport, which "fixes" > things? If so, perhaps you can try to disable/alter it somehow? If > not, any other ideas as to why the test might be working on nvidia hw?
Probably something get's clipped in window space or the range representable by the depth buffer: The clip space to window space transform translates [clip_space_z_min, clip_space_z_max] -> [0, 1] independent of the clip spaces z bounds configured from clip control. If your driver now clips the windows space depth values, you get that implicitly right. Well, I have omitted one thing here. The [0, 1] range is only true if the glDepthRange setting covers the whole [0, 1] range, which is probably the most often exercised case. If this driver clips at [min(gl_DepthRange.near,gl_DepthRange.far), max(gl_DepthRange.near,gl_DepthRange.far)] then you are probably again fine - at least for the clip control cases. To get that exactly right in the bigger picture, I have to dig out where exactly depth clamping happens then if enabled. IIRC this area with non [0, 1] depth ranges is where both closed source drivers also disagreed with mesa even with non clip control cases - that appeared for non standard glDepthRange settings and the tests I put together for the old original nvidia depth buffer extension. The glDepthRange values exceeding [0, 1] were unique to that nvidia depth buffer extension, but the ones not filling the [0, 1] range should be valid without any of the extensions in question. But since this is something that failed also independent of the requested extension, I tried to keep that separate. But I could see what happens if we set glDepthRange to cover a smaller range than [0, 1]. Greetings Mathias _______________________________________________ Piglit mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/piglit
