Blender Git Loki

merge troonk revusion 49612

* Fix: smoke collision velocity got messed up if domain was moved during simulation.
* Fixed possible crash when changing smoke collision type.
* Fix: Viewport preview of high resolution smoke was drawn with too low density.

Merge from trunk r49109-r49601

Two fixes for domain boundary conditions.
* Domain vertical velocity was accelerating over time. This issue has been in trunk since smoke was first introduced, but was especially troublesome for fire simulations due to higher temperatures and therefore higher buoyancy lift.
* Domain sides were partly considered obstacles regardless of the active "Border Collisions" setting. Now all boundaries behave identically. This should be especially useful when using new gsoc features like domain rotation or changing gravity direction.

Merge with trunk r49600

Isomap Unwrapper
================
* Update graph distance as well if needed.

Isomap Unwrapper
=================
* Fix, pass squares of distances to eigen solver
* Use a heap instead of a stack to get the next vertex to iterate
through in Dijkstra graph search algorithm. This is needed because we
may follow a non-optimal path and not get the nearest path between
vertices. This introduces the O(n*log(n)) performance (was more like
simply O(n) till now, but unfortunately, it was incorrect).

Isomap Unwrapper
=================
* First part of optimizations found in paper:

"Texture Mapping using Surface Flattening via Multi-Dimensional Scaling"
by Gil Zigelman, Ron Kimmel and Nahum Kiryati

(Multi-Dimensional Scaling is actually isomap with another name).

Implemented Dijkstra graph search algorithm that drops distance build
times from n^4 to n*(n*logn). The paper also describes a better distance
calulation algorithm so it seems like we shall soon see ISOMAP in its
full glory after all ;)

By adding some padding to the VBOs to make them 32byte aligned, I got a noticeable performance boost out of an ATI card. Nvidia doesn't seem to care as much. We could eventually do a version check since this does make use of a bit of extra vram (especially for things like mobile that have limited resources).

Using glMapBuffer to directly write data into VBOs. This makes updating the VBOs not nearly as slow, which provides a huge speed up to scenes with a lot of updating (e.g., skinned meshes). By the way, this code might have to fixed up for Android/GLES.

Patch from Daniel Stokes to fix bugs when using VBOs and a mesh gets split across different material buckets.

Comment cleanup

Isomap Unwrapper
==================
* First working version, however:

1) It's extremely slow (finding geodesic distances between all vertices
on mesh is O^4 expensive!) It can be optimized away, however there is
problem number...
2) The result is not that good when only using mesh connectivity from
edges. I may have to look at an alternative to calculate distances.

However, reading some more papers it looks like this is quite an
expensive algorithm. Coupled with the fact that it can't support pinned
UVs this will probably as far as this will go since there is no time to
investigate other methods. I will try to see if result can be improved a
bit though.

Isomap Unwrapper
==================
Add more debug prints, do math more according to the paper. Still not
having a good result :/

Some glLight functions remained in KX_Light.cpp

svn merge ^/trunk/blender -r49085:49508

whitespace changes

added gpuInitializeLighting/gpuShutdownLighting to blenderplayer

missed adding gpu_lighting_gl11.c

GPUlighting - thin layer to wrap OpenGL's Phong-Blinn lighting state functions