From: Brian P. <br...@vm...> - 2009-12-09 16:09:04
|
Allen Akin wrote: > On Tue, Dec 08, 2009 at 06:54:40PM -0700, Brian Paul wrote: > | In the case of readpixPerf, we only filter for rgb visuals. Buf if > | the visual also has Z (or stencil) we do extra tests (and produce > | additional test results). So, the ``A'' config and ``B'' config list > | were loosely filtered, but the test results are tied to specific > | configs. The 'looseness' of the match() function causes us to compare > | results from disparate configs. > > I guess I would have separated the tests -- one for color buffers, one > for depth buffers, one for stencil buffers, and so on. Not only would > that have solved the match() problem, but it would have allowed other > tests to depend on exactly the subset of readpixPerf tests that they > need. > > But that might not be expedient at this point. So one thing that really > makes sense to me is to ensure that exact matches take precedence over > any partial match. That ought to cover all the most common cases. > > | Right. But when the XIDs match AND ALL the other fields match, the > | configs should be identical (and the results comparable), even if the > | results come from different machines/drivers. > > If ALL the other fields match, then the XID doesn't matter, right? The > configs should be identical from the GL's point of view? Yes, except for the case I described below (there's some other attribute that glean's not aware of). > | But in the case where there's many configs with identical fields, but > | different XIDs, the equal() test should consider the XID. > | > | With GLX, for example, we can have two visuals which have identical > | attributes but one is TrueColor and the other is DirectColor. The > | DrawingSurfaceConfig code doesn't consider visual classes. This is > | where matching the XID is good. > > If that really matters, then maybe Visual class should be added to the > set of filterable attributes. I don't think I ever encountered a system > where the distinction between DirectColor and TrueColor mattered to the > GL, but maybe there are examples nowadays. I'm happy to defer to your > judgement there. I don't have time for any major changes now. So here's a new patch which simply looks for an exact match before doing the error-based match algorithm. It fixes the readpixPerf problem. -Brian |