This approach unifies much of the code handling the clean up of OpenGL graphics data, avoids lots of local mutexes and static variables that were previously required,
and enables the clean up scheme to be easily extended by users providing their own GraphicsObjectManager subclasses.
git-svn-id: http://svn.openscenegraph.org/osg/OpenSceneGraph/trunk@15130 16af8721-9629-0410-8352-f15c8da7e697
algorithm consisting of two consequent phases :
- first phase is a GLSL shader performing object culling and LOD picking ( a culling shader ).
Every culled object is represented as GL_POINT in the input osg::Geometry.
The output of the culling shader is a set of object LODs that need to be rendered.
The output is stored in texture buffer objects. No pixel is drawn to the screen
because GL_RASTERIZER_DISCARD mode is used.
- second phase draws osg::Geometry containing merged LODs using glDrawArraysIndirect()
function. Information about quantity of instances to render, its positions and other
parameters is sourced from texture buffer objects filled in the first phase.
The example uses various OpenGL 4.2 features such as texture buffer objects,
atomic counters, image units and functions defined in GL_ARB_shader_image_load_store
extension to achieve its goal and thus will not work on graphic cards with older OpenGL
versions.
The example was tested on Linux and Windows with NVidia 570 and 580 cards.
The tests on AMD cards were not conducted ( due to lack of it ).
The tests were performed using OSG revision 14088.
The main advantages of this rendering method :
- instanced rendering capable of drawing thousands of different objects with
almost no CPU intervention ( cull and draw times are close to 0 ms ).
- input objects may be sourced from any OSG graph ( for example - information about
object points may be stored in a PagedLOD graph. This way we may cover the whole
countries with trees, buildings and other objects ).
Furthermore if we create osgDB plugins that generate data on the fly, we may
generate information for every grass blade for that country.
- every object may have its own parameters and thus may be distinct from other objects
of the same type.
- relatively low memory footprint ( single object information is stored in a few
vertex attributes ).
- no GPU->CPU roundtrip typical for such methods ( method uses atomic counters
and glDrawArraysIndirect() function instead of OpenGL queries. This way
information about quantity of rendered objects never goes back to CPU.
The typical GPU->CPU roundtrip cost is about 2 ms ).
- this example also shows how to render dynamic objects ( objects that may change
its position ) with moving parts ( like car wheels or airplane propellers ) .
The obvious extension to that dynamic method would be the animated crowd rendering.
- rendered objects may be easily replaced ( there is no need to process the whole
OSG graphs, because these graphs store only positional information ).
The main disadvantages of a method :
- the maximum quantity of objects to render must be known beforehand
( because texture buffer objects holding data between phases have constant size ).
- OSG statistics are flawed ( they don't know anymore how many objects are drawn ).
- osgUtil::Intersection does not work
Example application may be used to make some performance tests, so below you
will find some extended parameter description :
--skip-dynamic - skip rendering of dynamic objects if you only want to
observe static object statistics
--skip-static - the same for static objects
--dynamic-area-size - size of the area for dynamic rendering. Default = 1000 meters
( square 1000m x 1000m ). Along with density defines
how many dynamic objects is there in the example.
--static-area-size - the same for static objects. Default = 2000 meters
( square 2000m x 2000m ).
Example application defines some parameters (density, LOD ranges, object's triangle count).
You may manipulate its values using below described modifiers:
--density-modifier - density modifier in percent. Default = 100%.
Density ( along with LOD ranges ) defines maximum
quantity of rendered objects. registerType() function
accepts maximum density ( in objects per square kilometer )
as its parameter.
--lod-modifier - defines the LOD ranges. Default = 100%.
--triangle-modifier - defines the number of triangles in finally rendered objects.
Default = 100 %.
--instances-per-cell - for static rendering the application builds OSG graph using
InstanceCell class ( this class is a modified version of Cell class
from osgforest example - it builds simple quadtree from a list
of static instances ). This parameter defines maximum number
of instances in a single osg::Group in quadtree.
If, for example, you modify it to value=100, you will see
really big cull time in OSG statistics ( because resulting
tree generated by InstanceCell will be very deep ).
Default value = 4096 .
--export-objects - write object geometries and quadtree of instances to osgt files
for later analysis.
--use-multi-draw - use glMultiDrawArraysIndirect() instead of glDrawArraysIndirect() in a
draw shader. Thanks to this we may render all ( different ) objects
using only one draw call. Requires OpenGL version 4.3 and some more
work from me, because now it does not work ( probably I implemented
it wrong, or Windows NVidia driver has errors, because it hangs
the apllication at the moment ).
This application is inspired by Daniel Rákos work : "GPU based dynamic geometry LOD" that
may be found under this address : http://rastergrid.com/blog/2010/10/gpu-based-dynamic-geometry-lod/
There are however some differences :
- Daniel Rákos uses GL queries to count objects to render, while this example
uses atomic counters ( no GPU->CPU roundtrip )
- this example does not use transform feedback buffers to store intermediate data
( it uses texture buffer objects instead ).
- I use only the vertex shader to cull objects, whereas Daniel Rákos uses vertex shader
and geometry shader ( because only geometry shader can send more than one primitive
to transform feedback buffers ).
- objects in the example are drawn using glDrawArraysIndirect() function,
instead of glDrawElementsInstanced().
Finally there are some things to consider/discuss :
- the whole algorithm exploits nice OpenGL feature that any GL buffer
may be bound as any type of buffer ( in our example a buffer is once bound
as a texture buffer object, and later is bound as GL_DRAW_INDIRECT_BUFFER ).
osg::TextureBuffer class has one handy method to do that trick ( bindBufferAs() ),
and new primitive sets use osg::TextureBuffer as input.
For now I added new primitive sets to example ( DrawArraysIndirect and
MultiDrawArraysIndirect defined in examples/osggpucull/DrawIndirectPrimitiveSet.h ),
but if Robert will accept its current implementations ( I mean - primitive
sets that have osg::TextureBuffer in constructor ), I may add it to
osg/include/PrimitiveSet header.
- I used BufferTemplate class writen and published by Aurelien in submission forum
some time ago. For some reason this class never got into osg/include, but is
really needed during creation of UBOs, TBOs, and possibly SSBOs in the future.
I added std::vector specialization to that template class.
- I needed to create similar osg::Geometries with variable number of vertices
( to create different LODs in my example ). For this reason I've written
some code allowing me to create osg::Geometries from osg::Shape descendants.
This code may be found in ShapeToGeometry.* files. Examples of use are in
osggpucull.cpp . The question is : should this code stay in example, or should
it be moved to osgUtil ?
- this remark is important for NVidia cards on Linux and Windows : if
you have "Sync to VBlank" turned ON in nvidia-settings and you want to see
real GPU times in OSG statistics window, you must set the power management
settings to "Prefer maximum performance", because when "Adaptive mode" is used,
the graphic card's clock may be slowed down by the driver during program execution
( On Linux when OpenGL application starts in adaptive mode, clock should work
as fast as possible, but after one minute of program execution, the clock slows down ).
This happens when GPU time in OSG statistics window is shorter than 3 ms.
"
git-svn-id: http://svn.openscenegraph.org/osg/OpenSceneGraph/trunk@14531 16af8721-9629-0410-8352-f15c8da7e697
Provided are lua, python and V8 (for javascript) plugins that just open up enough of a link to the respective libs to run a script, there is no scene graph <-> script communication in current implementation.
"The idea of this new OpenGL feature is :
- set RestartIndex = "n"
- draw elements strip
-> when the index is "n", the strip is "stopped" and restarted
It's very usefull for drawing tiles with a single strip and a "restart" at the end of each row.
The idea a an OSG StateAttribute is :
Usually we use to build geometry from code, because software modelers rarely support it (and 3d file formats doesn't support it) :
-RootNode <= "PrimitiveRestartIndex=0" // So now, we know that our restart index is 0 for all drawables under this node
|
- Drawable 1 : triangles => as usual
|
- Drawable 2 : triangles strip => as usual
|
- Drawable 3 : triangles strip + "GL_PRIMITIVE_RESTART" mode = ON => use the restart index
|
- Drawable 4 : triangles strip + "GL_PRIMITIVE_RESTART" mode = ON => use the restart index
|
- Drawable 5 : triangles strip => as usual
With a StateAttribute, it's easy for the developper to say "0 will be my restart index for all this object" and then activate the mode only on some nodes.
The main problem is if you set and restart index value which is not included in the vertex array (for exemple set restart index = 100 but you have only 50 vertex). There is no problem with OpenGL, but some OSG algorithms will try to access the vertex[100] and will segfault.
To solve this, I think there is two ways :
1/ add restart index in osg::PrimitiveSet and use this value in all algorithms. It's a lot of work, maybe dangerous, and it concern only a few situations : developpers who use this extension should be aware of advanced OpenGL (and OSG) data management
2/ use a StateAttribute, and choose a "correct" restart index. In my applications, I always use "0" as a restart index and duplicate the first vertex (vertex[0] = vertex[1]). So there is no difference for OpenGL and all OSG algorithms works properly.
"
New methods osg::Geometry::containsDeprecatedData() and osg::Geometry::fixDeprecatedData() provide a means for converting geometries that still use the array indices and BIND_PER_PRIMITIVE across to complient
versions.
Cleaned up the rest of the OSG where use of array indices and BIND_PER_PRIMITIVE were accessed or used.
GeometryNew is only temporary and will be renamed to Geometry on the completion of refactoring work and feedback from community.
Ported osggeometry across to use GeometryNew.
TextureBuffer objects may use osg::Texture::bindToImageUnit(), so GLSL shaders are able to use not only texelFetch() function , but also functions defined in GL_ARB_shader_image_load_store extension : imageLoad(), imageStore(), imageAtomicAdd() etc."
second email: "After a while I found that osg::Texture::applyTexParameters() used with TextureBuffer may cause some OpenGL errors ( applying texture filters and wraps to TextureBuffer makes no sense ) so I fixed it."
SETUP_EXE, in order to have a unique entry point to build the
libraries. With this changes the android integration will be less
painful (currently is a big IF(ANDROID) for each CMakeLists.txt) and
more maintainable in the future. I hope next submissions will be for
supporting android from my colleague Jorge.
"
A few things remain to do:
* The binding between a uniform block in a shader program and a buffer indexed target number is fixed, like a vertex attribute binding. This is too restrictive because that binding can be changed without relinking the program. This mapping should be done by name in the same way that uniform values are handled i.e., like a pseudo state attribute;
* There's no direct way yet to query for the offset of uniforms in uniform block, so only the std140 layout is really usable. A helper class that implemented the std140 rules would be quite helpful for setting up uniform blocks without having to link a program first;
* There's no direct support for querying parameters such as the maximum block length, minimum offset alignment, etc. Having that information available outside of the draw thread would make certain instancing techniques easier to implement."
changed extensions from .c to .cpp and got compiling as C files as part of the osg core library.
Updated and cleaned up the rest of the OSG to use the new internal GLU.
Texture2DMultismaple as name suggests provides means to directly access subsamples of rendered FBO target. (GLSL 1.5 texelFetch call).
Recently I was working on deferred renderer with OSG, during that I noticed there is no support for multisampled textures (GL_ARB_texture_multisample extension). After consultations with Paul Martz and Wojtek Lewandowski I added Texture2DMultisample class and made few necessary changes around osg::FrameBufferObject, osg::Texture and osgUtil::RenderStage classes."
and from follow email:
"Fixed. According to ARB_texture_multisample extension specification multisample textures don't need TexParameters since they can only be fetched with texelFetch."
which utilises a global recursive mutex that is dedicated to manage Observer and ObserverSet.
The new global mutex for observers avoids problems with deadlocks that were occurring previously when
an osg::Refenced object was being deleted at the same time as on osg::ObserverNodePath.
almost all Linux distributions. Although it is 100% compatible with ld,
by default it gives an error if a library has unresolved symbols at link
time, that is, it has set -Wl,--no-undefined by default. Debian folks
have found that libosg.so and libosgDB.so use some functions belonging
to libdl.so {dlsym,dlopen,dlclose,dlerror} without linking to it.
My changes link those two libraries to libdl.so explicitly in the same
way it is already done for libm.so and librt.so."
settimeofday), OSG animations will freeze your application because
osg::Timer uses gettimeofday internally on non-Win32 platforms. This
is wrong and should be replace with times(2) or clock_gettime(2).
The attached patch fixes the issue in a binary-compatible way by using
clock_gettime when it's available, and falling back to gettimeofday
when it's not."
