Re: [Gnu3dkit-discuss] G3DRenderer - Object

[Philippe C . D . Robert]

On Sunday, March 16, 2003, at 10:15  Uhr, Brent Gulanowski wrote:
> > Do you mean an aggregation of arbitrary, geometric primitives by 
> > 'Object'? The idea here is to use a compile action on the respective 
> > subtree which would generate an optimised (internal!) represenation 
> > for the renderer of choice. This is hence not an application specific 
> > optimisation, but application specific optimisations on the scene 
> > graph level will still be required to maximise rendering >> performances.
>
> I meant the Object in the Renderman sense (beginObject: etc.). Whether 
> that is created using a compile action, I cannot tell from the current 
> interface.

The problem here is what do we want to expose and do we have to expose 
to 3DKit developers. Maybe some parts of this interface should only be 
useable by the 3DKit itself?

> On Sunday, March 16, 2003, at 02:33  PM, Philippe C.D. Robert wrote:
> > Talking MVC, a scene graph is a data type and hence it provides the 
> > model only, actions such as the G3DRenderAction are the controllers 
> > which operate on this data, and the camera/view duo is the view which 
> > is used to visualise the data.
>
> In MVC as I know it, a model is more than just data, it is also 
> behaviour. A model object stores its own data and acts upon it. 
> Controllers communicate with the model objects -- they do not operate 
> on the data. If they do, it violates data hiding.

Data access happens via method invocation, so there is no data hiding 
violation. Here (as I see it) the graph is really only a data 
structure, otherwise you will run into severe problems afterwards.

> > In other words, the scene graph only contains the data to be used by 
> > the renderers to visualise the respective scene, this includes the 
> > geometry but also transformations, shaders and likewise data. 
> > Therefore you can for example render the same scene using different 
> > renderers to generate different results (in any aspects) without 
> > touching the scene data at all. So the model is indeed completely 
> > separated from the view.
>
> This sounds somewhat contradictory. On the one had, the scene graph is 
> little more than a container -- in which case, why have classes at 
> all? On the other

Why wouldn't you use classes for data structures (containers)? This 
sounds weird to me.

>  hand, you claim that the renderer never touches the data. How is that 
> possible? Something must be able to access it, while guaranteeing that 
> the data is preserved and internally consistent. The renderer cannot 
> take on that responsibility. In which case, the scene classes provide 
> mutator and accessor methods, and nothing else? If a shape holds in it 
> a set of polygon (attribute)s, say ten thousand, is this really going 
> to be copied into a dictionary, and then read back out?

No, this is not what I meant. A render action traverses the graph and 
manages the internal gfx states, now whenever it reaches let's say a 
shape having a geometry node attached, it sets up all the states 
required for drawing as defined by the shape and then tells the 
geometry object to draw itself (by passing the correct renderer to be 
used by the geometry object). So the render action controls the 
rendering process but the geometry objects draw themselves. Does this 
make more sense to you?

> You could help me in my confusion by describing the steps necessary to 
> access a scene and draw something in a frame buffer, by sketching the 
> call tree if possible, because I really cannot visualize it.

I'll try...

// setup a scene graph
scene = ....;

// we use OGL here as an example
openGLRenderer = [[G3DOGLRenderer alloc] init];

// setup a render action w/ the OGL renderer
ra = [[G3DRenderAction alloc] initWithRenderer:openGLRenderer];

// init and set the camera object
myCamera = [[G3DCamera alloc] initWithView:myView];
[ra setCamera:myCamera];

Now to render a scene the drawRect: method (for example) of myView 
tells the render action to traverse the scene graph and render it by 
calling

[ra apply:scene];

Rendering is done by sending the appropriate commands - based on the 
graph nodes state - to its renderer (state handling, transformations, 
...) using the renderer API, ie. like

[openGLRenderer worldBegin];
...
[openGLRenderer tranformBegin];
...
[geometry drawGeometry: openGLRenderer];
...
[openGLRenderer tranformEnd];
...
[openGLRenderer worldEnd];

> I want to draw a sharp distinction between what I consider the real 
> scene data -- the things in the scene and their relation to one 
> another, including scene level attributes -- and the -representations- 
> of the things in the scene. You have me questioning whether such a 
> distinction is possible, but I am confident that it *is* possible, and 
> that the distinction is very important and could make a big difference 
> in the design and performance of the software. If I can find that 
> difference, I want to pull out all the representation data and let the 
> renderer itself manage it, with help from data/file controllers. This 
> will make the scene smaller and lighter, and leave the RenderKit to 
> manage it more intelligently. It will make it easier to use different 
> renderers to present the same scene, and even allow other kinds of 
> views to pose as a renderer to present the scene data in a completely 
> different fashion. But this is dependent on things that I cannot know 
> by looking at the interface.

Can you please give me a concrete example showing what you want to 
achieve and what is not possible with the current design approach?

-Phil
--
Philippe C.D. Robert
http://www.nice.ch/~phip