Date: Thu Dec 30 21:03:13 2010
New Revision: 390
Modified:
GtkRadiant/trunk/libs/mathlib.h
GtkRadiant/trunk/libs/mathlib/mathlib.c
GtkRadiant/trunk/tools/quake3/common/polylib.c
Log:
Undoing commits r363 and r371 as it pertains to polylib.c, mathlib.c,
and mathlib.h (the regression tests have not been removed).
Trunk is now restored to a state that it was in before I started
trying to fix the math accuracy errors in q3map2. Commits r363 and
r371 were "correct" and did improve math accuracy significantly, but
unfortunately the underlying cause of math accuracy issues is something
else, which is being addressed in branch Rambetter-math-fix-experiments
currently. I'm taking the BSD approach here, which is "we not going to
partially fix the problem. it's all or nothing". Otherwise it's just
too risky in my opinion. I don't like playing Whack-A-Mole.
Someday, we might merge Rambetter-math-fix-experiments branch to trunk.
Sorry about all these needless commits to trunk.
New Revision: 379
Undoing revision 377 (reverting just those files modified by that
commit). I have disovered the fundamental problem to the math error problems,
and although this commit (r377) is "correct", it fails to address the
fundamental problem. Therefore, I'd rather leave the code in a state that
has the exact same behavior as before until I get a chance to address the
fundamental issue.
Date: Tue Dec 28 23:20:32 2010
New Revision: 378
Updating notes on regression tests. They are "mostly sort of fixed". Won't
be fixed for good until I look at the last bit of code that has not been
examined yet, which is the plane intersection code. I want the errors to be
much less than they are now, even though the disappearing_sliver* tests are
now working.
Continuing work on BaseWindingForPlane() in polylib.c. In fact I'm pursuing
the approach that was committed in r375 (but was then backed out). I can't
believe my eyes, but I seem to be getting 0.000000% error in some of my
regression tests. The trick is to scale by a power of 2 and never do a
VectorNormalize().
Following up on r371; found another optimization for math rounding errors
in BaseWindingForPlane(). This allows me to get rid of an extra call to
VectorSetLength() which contain floating point multiplications. Behavior
of BaseWindingForPlane() has been verified against base_winding regression
test (comparing output w/ logging patch).
Date: Tue Dec 28 04:02:11 2010
New Revision: 371
Rewriting BaseWindingForPlane() in polylib.c from the ground up. The behavior is totally unchanged, and I verified this by running extensive tests. The only difference is that the math precision is much much better now. Performance should be better as well (but that is not tested). This is a major milestone because it fixes two regression tests: disappearing_sliver2 and sparkly_seam.
Improvements to math precision is ongoing and more improvements can probably be made even after this patch.
I will update the README.txt files in the regression tests in a separate commit. This commit only includes the actual fixed code.
