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...
Realtime modifiers applied on mesh objects will be supported in
the game engine with the following limitations:
- Only real time modifiers are supported (basically all of them!)
- Virtual modifiers resulting from parenting are not supported:
armature, curve, lattice. You can still use these modifiers
(armature is really not recommended) but in non parent mode.
The BGE has it's own parenting capability for armature.
- Modifiers are computed on the host (using blender modifier
stack).
- Modifiers are statically evaluated: any possible time dependency
in the modifiers is not supported (don't know enough about
modifiers to be more specific).
- Modifiers are reevaluated if the underlying mesh is deformed
due to shape action or armature action. Beware that this is
very CPU intensive; modifiers should really be used for static
objects only.
- Physics is still based on the original mesh: if you have a
mirror modifier, the physic shape will be limited to one half
of the resulting object. Therefore, the modifiers should
preferably be used on graphic objects.
- Scripts have no access to the modified mesh.
- Modifiers that are based on objects interaction (boolean,..)
will not be dependent on the objects position in the GE.
What you see in the 3D view is what you get in the GE regardless
on the object position, velocity, etc.
Besides that, the feature is compatible with all the BGE features
that affect meshes: armature action, shape action, relace mesh,
VideoTexture, add object, dupligroup.
Known problems:
- This feature is a bit hacky: the BGE uses the derived mesh draw
functions to display the object. This drawing method is a
bit slow and is not 100% compatible with the BGE. There may
be some problems in multi-texture mode: the multi-texture
coordinates are not sent to the GPU.
Texface and GLSL on the other hand should be fully supported.
- Culling is still based on the extend of the original mesh.
If you have a modifer that extends the size of the mesh,
the object may disappear while still in the view frustrum.
- Derived mesh is not shared between replicas.
The derived mesh is allocated and computed for each object
with modifiers, regardless if they are static replicas.
- Display list are not created on objects with modifiers.
I should be able to fix the above problems before release.
However, the feature is already useful for game development.
Once you are ready to release the game, you can apply the modifiers
to get back display list support and mesh sharing capability.
MSVC, scons, Cmake, makefile updated.
Enjoy
/benoit
Separate getting a normal attribute and getting __dict__, was having to do too a check for __dict__ on each class (multiple times per getattro call from python) when its not used that often.
- comments to PyObjectPlus.h
- remove unused/commented junk.
- renamed PyDestructor to py_base_dealloc for consistency
- all the PyTypeObject's were still using the sizeof() their class, can use sizeof(PyObjectPlus_Proxy) now which is smaller too.
This changes how the BGE classes and Python work together, which hasnt changed since blender went opensource.
The main difference is PyObjectPlus - the base class for most game engine classes, no longer inherit from PyObject, and cannot be cast to a PyObject.
This has the advantage that the BGE does not have to keep 2 reference counts valid for C++ and Python.
Previously C++ classes would never be freed while python held a reference, however this reference could be problematic eg: a GameObject that isnt in a scene anymore should not be used by python, doing so could even crash blender in some cases.
Instead PyObjectPlus has a member "PyObject *m_proxy" which is lazily initialized when python needs it. m_proxy reference counts are managed by python, though it should never be freed while the C++ class exists since it holds a reference to avoid making and freeing it all the time.
When the C++ class is free'd it sets the m_proxy reference to NULL, If python accesses this variable it will raise a RuntimeError, (check the isValid attribute to see if its valid without raising an error).
- This replaces the m_zombie bool and IsZombie() tests added recently.
In python return values that used to be..
return value->AddRef();
Are now
return value->GetProxy();
or...
return value->NewProxy(true); // true means python owns this C++ value which will be deleted when the PyObject is freed
- The armature weakref list was being incref'd twice then decrefed twice (incref and decref were used incorrectly), now only once. My 'fix' broke this.
- In bpy_pydriver_create_dict the 2 refs added from running PyDict_SetItemString twice were undone when clearing the dictionary (added comment)
- changed Py_XDECREF to Py_DECREF int BPY_pyconstraint_update and BPY_pyconstraint_target, Py_XDECREF checs for NULL value which would have crashed blender before it got to Py_XDECREF anyway.
- after every error is reported (PyErr_Print), remove sys.last_traceback and clear the error, I found this fixed certain crashes (usually when starting the game engine or exiting blender), so best do this all the time.
- header_text.c, CcdPhysicsEnvironment.cpp, KX_CameraActuator.cpp - remove some warnings.
Use each types dictionary to store attributes PyAttributeDef's so it uses pythons hash lookup (which it was already doing for methods) rather then doing a string lookup on the array each time.
This also means attributes can be found in the type without having to do a dir() on the instance.
KX_PolygonMaterial and KX_BlenderMaterial - Added a print function (would raise a python error on printing)
* Crashes *
KX_GameObject SetParent - Disallowed setting a parent to its self, caused a recursion crash.
KX_MeshProxy "materials" attribute was segfaulting because of my recent change - I was wrong, you do need to check material types (no idea why since they are both PyObject * at the base)
KX_VisibilityActuator - Wasn't initialized with PyType_Ready() making it crash on access (own fault)
* Crashes because of missing NULL checks *
KX_PolygonMaterial's "gl_texture" attribute wasnt checking for a valid m_tface
KX_GameObject - added checks for GetPhysicsController()
KX_RayCast::RayTest - didnt check for a valid physics_environment
KX_SceneActuator's getCamera python function wasnt checking if there was a camera.
- Initialize python types with PyType_Ready, which adds methods to the type dictionary.
- use Pythons get/setattro (uses a python string for the attribute rather then char*). Using basic C strings seems nice but internally python converts them to python strings and discards them for most functions that accept char arrays.
- Method lookups use the PyTypes dictionary (should be faster then Py_FindMethod)
- Renamed __getattr -> py_base_getattro, _getattr -> py_getattro, __repr -> py_base_repr, py_delattro, py_getattro_self etc.
From here is possible to put all the parent classes methods into each python types dictionary to avoid nested lookups (api has 4 levels of lookups in some places), tested this but its not ready yet.
Simple tests for getting a method within a loop show this to be between 0.5 and 3.2x faster then using Py_FindMethod()
Added the method into the PyType so python knows about the methods (its supposed to work this way).
This means in the future the api can use PyType_Ready() to store the methods in the types dictionary.
Python3 removes Py_FindMethod and we should not be using it anyway since its not that efficient.
Python dir(ob) for game types now includes attributes names,
* Use "__dict__" rather then "__methods__" attribute to be Python 3.0 compatible
* Added _getattr_dict() for getting the method and attribute names from a PyObject, rather then building it in the macro.
* Added place holder *::Attribute array, needed for the _getattr_up macro.
Use 'const char *' rather then the C++ 'STR_String' type for the attribute identifier of python attributes.
Each attribute and method access from python was allocating and freeing the string.
A simple test with getting an attribute a loop shows this speeds up attribute lookups a bit over 2x.
the features that are needed to run the game. Compile tested with
scons, make, but not cmake, that seems to have an issue not related
to these changes. The changes include:
* GLSL support in the viewport and game engine, enable in the game
menu in textured draw mode.
* Synced and merged part of the duplicated blender and gameengine/
gameplayer drawing code.
* Further refactoring of game engine drawing code, especially mesh
storage changed a lot.
* Optimizations in game engine armatures to avoid recomputations.
* A python function to get the framerate estimate in game.
* An option take object color into account in materials.
* An option to restrict shadow casters to a lamp's layers.
* Increase from 10 to 18 texture slots for materials, lamps, word.
An extra texture slot shows up once the last slot is used.
* Memory limit for undo, not enabled by default yet because it
needs the .B.blend to be changed.
* Multiple undo for image painting.
* An offset for dupligroups, so not all objects in a group have to
be at the origin.
=======================================
Alpha blending + sorting was revised, to fix bugs and get it
to work more predictable.
* A new per texture face "Sort" setting defines if the face
is alpha sorted or not, instead of abusing the "ZTransp"
setting as it did before.
* Existing files are converted to hopefully match the old
behavior as much as possible with a version patch.
* On new meshes the Sort flag is disabled by the default, to
avoid unexpected and hard to find slowdowns.
* Alpha sorting for faces was incredibly slow. Sorting faces
in a mesh with 600 faces lowered the framerate from 200 to
70 fps in my test.. the sorting there case goes about 15x
faster now, but it is still advised to use Clip Alpha if
possible instead of regular Alpha.
* There still various limitations in the alpha sorting code,
I've added some comments to the code about this.
Some docs at the bottom of the page:
http://www.blender.org/development/current-projects/changes-since-246/realtime-glsl-materials/
Merged some fixes from the apricot branch, most important
change is that tangents are now exactly the same as the rest
of Blender, instead of being computed in the game engine with a
different algorithm.
Also, the subversion was bumped to 1.
All data layers, including MVert/MEdge/MFace, are now managed as custom
data layers. The pointers like Mesh.mvert, Mesh.dvert or Mesh.mcol are
still used of course, but allocating, copying or freeing these arrays
should be done through the CustomData API.
Work in progress documentation on this is here:
http://mediawiki.blender.org/index.php/BlenderDev/BlenderArchitecture/CustomData
Replaced TFace by MTFace:
This is the same struct, except that it does not contain color, that now
always stays separated in MCol. This was not a good design decision to
begin with, and it is needed for adding multiple color layers later. Note
that this does mean older Blender versions will not be able to read UV
coordinates from the next release, due to an SDNA limitation.
Removed DispListMesh:
This now fully replaced by DerivedMesh. To provide access to arrays of
vertices, edges and faces, like DispListMesh does. The semantics of the
DerivedMesh.getVertArray() and similar functions were changed to return
a pointer to an array if one exists, or otherwise allocate a temporary
one. On releasing the DerivedMesh, this temporary array will be removed
automatically.
Removed ssDM and meshDM DerivedMesh backends:
The ssDM backend was for DispListMesh, so that became obsolete automatically.
The meshDM backend was replaced by the custom data backend, that now figures
out which layers need to be modified, and only duplicates those.
This changes code in many places, and overall removes 2514 lines of code.
So, there's a good chance this might break some stuff, although I've been
testing it for a few days now. The good news is, adding multiple color and
uv layers should now become easy.
