<div class="gmail_quote">2011/4/19 Michel Dänzer <<a href="mailto:michel@daenzer.net">michel@daenzer.net</a>> <blockquote class="gmail_quote" style="margin: 0pt 0pt 0pt 0.8ex; border-left: 1px solid rgb(204, 204, 204); padding-left: 1ex;"> <div class="im">On Die, 2011-04-19 at 12:00 +0200, Christoph Bumiller wrote: > On 16.04.2011 18:50, Christoph Bumiller wrote: > > I hope the new version is correct, the commit message describes why I > > did the first change, and the second change is described in a comment. > > > > Note that the MAD for inversion uses Height - 1 instead of Height. > > > > With this, piglit glsl-arb-fragment-coord-conventions and > > fbo-depth-sample-compare pass on nvc0. > > > > I was assuming that integer pixel centers for size 100 range from 0 to > > 99 and half-integer pixel centers from 0.5 to 99.5. > > > Attached a better version of the patch, potentially saving an > instruction and avoiding precision issues with NEAREST filtering that > made piglit's blending-in-shader fail. > > The only other location I found STATE_FB_WPOS_Y_TRANSFORM used was > r600 classic, which should also work more correctly now since (at > least according to gallium caps) it uses half-integer pixel center, > and for H=100, 0.5 * -1 + 99 obviously isn't the desired 99.5. </div>FWIW, this fixes compiz alpha-blurred contents being off by one vertically on r300g. Thanks! </blockquote><div> </div></div>If that's true, it should be a candidate for 7.10 as well. Marek