BGE performance, 3rd round: culling and rasterizer.
authorBenoit Bolsee <benoit.bolsee@online.be>
Thu, 7 May 2009 09:13:01 +0000 (09:13 +0000)
committerBenoit Bolsee <benoit.bolsee@online.be>
Thu, 7 May 2009 09:13:01 +0000 (09:13 +0000)
commit42557f90bd16771a5c6437dbfb3952527df7fb1a
tree783a84eef7c05adbea3015ecd09ec7e2b18a7cfa
parent779bf435ef2ba87fbcee6a28b053d97a551b8eb5
BGE performance, 3rd round: culling and rasterizer.

This commit extend the technique of dynamic linked list to the mesh
slots so as to eliminate dumb scan or map lookup. It provides massive
performance improvement in the culling and in the rasterizer when
the majority of objects are static.

Other improvements:
- Compute the opengl matrix only for objects that are visible.
- Simplify hash function for GEN_HasedPtr
- Scan light list instead of general object list to render shadows
- Remove redundant opengl calls to set specularity, shinyness and diffuse
  between each mesh slots.
- Cache GPU material to avoid frequent call to GPU_material_from_blender
- Only set once the fixed elements of mesh slot
- Use more inline function

The following table shows the performance increase between 2.48, 1st round
and this round of improvement. The test was done with a scene containing
40000 objects, of which 1000 are in the view frustrum approximately. The
object are simple textured cube to make sure the GPU is not the bottleneck.
As some of the rasterizer processing time has moved under culling, I present
the sum of scenegraph(includes culling)+rasterizer time

Scenegraph+rasterizer(ms)       2.48      1st round       3rd round

All objects static,            323.0           86.0             7.2
all visible, 1000 in
the view frustrum

All objects static,            219.0           49.7             N/A(*)
all invisible.

All objects moving,            323.0          105.6            34.7
all visible, 1000 in
the view frustrum

Scene destruction              40min          40min              4s

(*) : this time is not representative because the frame rate was at 60fps.
      In that case, the GPU holds down the GE by frame sync. By design, the
      overhead of the rasterizer is 0 when the the objects are invisible.

This table shows a global speed up between 9x and 45x compared to 2.48a
for scenegraph, culling and rasterizer overhead. The speed up goes much
higher when objects are invisible.

An additional 2-4x speed up is possible in the scenegraph by upgrading
the Moto library to use Eigen2 BLAS library instead of C++ classes but
the scenegraph is already so fast that it is not a priority right now.

Next speed up in logic: many things to do there...
34 files changed:
projectfiles_vc9/gameengine/rasterizer/RAS_rasterizer.vcproj
projectfiles_vc9/gameengine/rasterizer/openglrasterizer/RAS_openglrasterizer.vcproj
source/gameengine/Ketsji/BL_BlenderShader.cpp
source/gameengine/Ketsji/BL_BlenderShader.h
source/gameengine/Ketsji/KX_BlenderMaterial.cpp
source/gameengine/Ketsji/KX_BlenderMaterial.h
source/gameengine/Ketsji/KX_GameObject.cpp
source/gameengine/Ketsji/KX_GameObject.h
source/gameengine/Ketsji/KX_KetsjiEngine.cpp
source/gameengine/Ketsji/KX_Light.cpp
source/gameengine/Ketsji/KX_PolygonMaterial.cpp
source/gameengine/Ketsji/KX_PythonInit.cpp
source/gameengine/Ketsji/KX_Scene.cpp
source/gameengine/Rasterizer/CMakeLists.txt
source/gameengine/Rasterizer/Makefile
source/gameengine/Rasterizer/RAS_BucketManager.cpp
source/gameengine/Rasterizer/RAS_BucketManager.h
source/gameengine/Rasterizer/RAS_IPolygonMaterial.cpp
source/gameengine/Rasterizer/RAS_IPolygonMaterial.h
source/gameengine/Rasterizer/RAS_MaterialBucket.cpp
source/gameengine/Rasterizer/RAS_MaterialBucket.h
source/gameengine/Rasterizer/RAS_MeshObject.cpp
source/gameengine/Rasterizer/RAS_MeshObject.h
source/gameengine/Rasterizer/RAS_OpenGLRasterizer/CMakeLists.txt
source/gameengine/Rasterizer/RAS_OpenGLRasterizer/Makefile
source/gameengine/Rasterizer/RAS_OpenGLRasterizer/RAS_ListRasterizer.cpp
source/gameengine/Rasterizer/RAS_OpenGLRasterizer/SConscript
source/gameengine/Rasterizer/SConscript
source/gameengine/SceneGraph/SG_DList.h
source/gameengine/SceneGraph/SG_QList.h
source/gameengine/SceneGraph/SG_Spatial.cpp
source/gameengine/SceneGraph/SG_Spatial.h
source/gameengine/VideoTexture/ImageRender.cpp
source/kernel/gen_system/GEN_HashedPtr.cpp