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.


git-svn-id: svn://svn.icculus.org/gtkradiant/GtkRadiant/trunk@378 8a3a26a2-13c4-0310-b231-cf6edde360e5
This commit is contained in:
rambetter 2010-12-29 05:20:32 +00:00
parent f36b40ae62
commit f487ea7c54
4 changed files with 71 additions and 7 deletions

View File

@ -46,7 +46,7 @@ And we get rid of the recpirocal length ilength altogether. Even the
slightest math errors are magnified in successive calls to linear algebra
functions.
The change described above was commmitted to GtkRadiant trunk as revision r363.
The change described above was commmitted to GtkRadiant trunk as revision 363.
POSSIBLE SIDE EFFECTS:
@ -162,3 +162,22 @@ It seems that FixBrokenSurface() should be fixed to completely fix the case
where there are two close points, and should report the surface as fixed.
This might be a destabilizing change however, so if this is indeed fixed, it
may make sense to activate the fix only if a certain flag is set.
MORE NOTES:
===========
As stated above, the accuracy after revision 363 is:
(67.000229 -1021.998657 0.000000)
(88.000175 -891.999146 -767.997437)
(133.999146 -1014.998779 0.000000)
A further change was committed for a related problem in revision 377. After
this change:
(66.99955750 -1022.00262451 0.00000000)
(87.99969482 -892.00170898 -768.00524902)
(133.99958801 -1015.00195312 0.00000000)
The results look similar with respect to the amount of error present.

View File

@ -16,7 +16,7 @@ SOLUTION TO PROBLEM:
It was discovered that BaseWindingForPlane() in polylib.c did some sloppy
mathematics with significant loss of precision. Those problems have been
addressed in commit revision 371.
addressed in commits to revisions 371 and 377.
POSSIBLE SIDE EFFECTS:
@ -67,17 +67,27 @@ is this:
As you can see, all points but one have an increase in accuracy. This is
still not accurate enough in my opinion, but is a step in the right direction.
After the fix committed in revision 377, which is a further attempt to address
BaseWindingForPlane(), we get the following accuracy:
(6784.00000000 16241.00000000 -1722.00000000)
(6144.00000000 16083.00000000 -1443.00000000)
(6144.00000000 16122.00000000 -1424.00000000)
It's just a fluke for this particular case, but obviouly revision 377 looks
favorably upon this regression test, because there is zero percent error.
MORE NOTES:
===========
I attempted to improve upon revision 371 by streamlining the code in
BaseWindingForPlane() some more. Those attempts were committed as r375.
After revision 375:
BaseWindingForPlane() some more. Those attempts were committed as revision
375. After revision 375:
(6784.09375000 16241.01757812 -1722.04687500)
(6144.00000000 16082.99414062 -1443.00390625)
(6144.00000000 16122.00000000 -1424.00097656)
Revision 375 has since been reverted (undone) because of the loss in
accuracy.
accuracy. Revision 377 is a fix for those failed attempts.

View File

@ -14,4 +14,36 @@ compile for any Q3 mod.
SOLUTION TO PROBLEM:
====================
None yet. Probably due to sloppy math code.
More work has been done to BaseWindingForPlane() to make it more accurate.
This function is in polylib.c. The changes to fix this regression test were
committed in revision 377; however, those changes are not "good enough".
IN-DEPTH DISCUSSION:
====================
This is the problem triangle:
In ParseRawBrush() for brush 0
Side 0:
(6144.000000 16122.000000 -2048.000000)
(6144.000000 16083.000000 -2048.000000)
(6784.000000 16241.000000 -2048.000000)
Computed winding before fix:
(6784.16406250 16241.04101562 -2048.00000000)
(6144.00000000 16122.00976562 -2048.00000000)
(6144.00000000 16083.00000000 -2048.00000000)
Obviously the 6784.16406250 is beyond epsilon error.
After revision 377:
(6783.85937500 16240.96484375 -2048.00000000)
(6144.00000000 16121.99218750 -2048.00000000)
(6144.00000000 16083.00000000 -2048.00000000)
Even though this fixes the regression test, the error in 6783.85937500 is
still greater than epsilon (but fortunately in the opposite direction). So
I don't consider this test case to be fixed quite yet.

View File

@ -15,4 +15,7 @@ SOLUTION TO PROBLEM:
====================
None yet. The problem is likely caused by sloppy math operations (significant
loss of precision).
loss of precision). This bug pops in and out of existence with every other
commit at the moment. The problem is likely caused by the operations in the
brush winding computation (where the planes are intersected with each other).
I have not gotten around to addressing that code yet.