Blender Git Loki

Render Branch: some hacked code to test per tile subdivision, and
how it would interact with GI, it has no catmull-clark, grid cracks,
slow performance, no shadow, sss, raytrace, etc. Example file:

http://users.pandora.be/blendix/snow_subdiv.blend
http://users.pandora.be/blendix/snow_subdiv.png

Render Branch: fix for lowres raytracing of instanced objects, and some
cache test code that's still disabled.

Render Branch: increase max number of AO samples from 32*32 to 128*128.

Render Branch: remove some test code left in accidentially, should fix
compile on windows.

Render Cache: irradiance cache:

* increase truncated SVD epsilon to avoid avoid large color changes when
interpolating between near colinear points.
* added neighbour clamping. it doesn't seem to help much and make things
a bit slower at the moment, but hopefully can be used to get better
distribution later.

Main remaining causes of artifacts seem to be:

* even 32x32 (1024) samples can still be noisy leading to low frequency
noise after interpolation
* borders between tiles can have discontinuities. this is mostly visible
in combination with the above problem, i.e. if the result is splotchy
it will have a different splotches in the next tile.
* very small features may be missed still, leading to missing shadows in
some cases or even stippled patterns instead if they are detected part
of the time, neighbour clamping helps somewhat here.

Rende Branch: irradiance caching, fix part of the black pixels with
a division by zero check in the TNT SVD code.

Render Branch: tweak irradiance cache clamping for reconstruction to
avoid corner samples leaking into open areas.

Render Branch:

Should have mentioned a few more things in my previous commit log:
* Indirect light now also includes lamps.
* Adaptive sampling is currently not working.
* Teaser image (spot the artifacts)

http://users.pandora.be/blendix/indirect_teaser.png

Render Branch: Irradiance Cache

Implementation borrows some code from Pixie though has changed quite
a bit. This is work in progress and has artifacts and hardcoded
parameters still.

I tried irradiance gradients and found them problematic, they seem
to reduce part of the artifacts but for others have no effect, and
can even make things worse. There's a clamping trick that helps a
bit here, but I couldn't get that reliable either.

Currently it is using least squares reconstruction to smooth out the
results, it's too smooth now, still trying to find out the best way
to weight samples without introducing problems, but I'm finding this
much more promising than gradients.

Render Branch: quick hack to support 4x4 matrix SVD decomposition
in C, by using the TNT matrix library from the IK solver.

Render Branch: math functions

* add len_squared_v3v3
* add mul_v4_m4v4
* add box_minmax_bounds_m4
* rename disc_to_sh to vec_fac_to_sh

Render Branch: redrawing while render would miss a few scanlines each time
with antialiasing enabled.

Render Branch: disable ray trace epsilon for testing, not using it
gives artifacts with overlapping faces, but it also seems to lead
to self intersections.

[#20116] calculation problem in bvh_importer.py
from Martin K (martinbk)

tagging for 2.51

bugfix in 2.4x also [#20480] crash/lock-up for certain aspect ratios

COLLADA branch:
* fix bug in normals import code - could crash on meshes with ngons
* disabled creation of empties on armature import (used for debugging)
* fix memleak (temporary FCurves were not freed)

COLLADA branch: remove signed/unsigned comparison warnings.

COLLADA branch: merge from trunk -r 26429:26848.

COLLADA branch: import mesh smoothing info. Normal values from DAE are not assigned to vertices but are used to flag faces as smooth or flat. This trick works.