If you need a testament to the quality of one’s OpenGL wrapper, I am making a next-generation game engine so I naturally want to take care on the quality of my OpenGL and Direct3D wrappers.
With very careful use of critical sections I have full multi-threaded capabilities, such as ways to bind textures so that they cannot be rebound by another thread until the current thread is done with its texture(s).
I wrote a new shader language (pardon the incomplete documentation; it is documented as I have time) to wrap around the OpenGL shader system and the DirectX shader systems.

I would not expect you to write a new shader language to do what you want to do, so I will try to explain how I would go about your situation. This will be a mix of what I do and what I would do.

My vertex and index buffer wrappers are as fully flexible as their API-specific buffers.
When creating a vertex buffer you will pass the attributes you want it to have and the slots for those attributes.
In HLSL, these match exactly to semantics, so I would pass POSITION0, NORMAL0, NORMAL1, etc.

Offsets for each of the attributes are calculated based on the combination of attributes. When the user is done filling the vertex buffer, he or she calls Finalize(), at which point it is transferred to OpenGL or Direct3D or wherever.
Read here for an easy way to have one class managing OpenGL resources and another acting as the generic interface for that class, which never calls OpenGL functions directly.


Just as OpenGL and DirectX do, I support multiple streams with my wrappers.
In preparing for render, one or more vertex buffers will be selected and assigned to various streams.
When CFnd::Render() is called to make the final render (with an optional index buffer as a parameter), the array of active vertex buffers are notified that they are about to be used for rendering. It does not matter how many are active.
Each one connects its vertex attributes to the attributes in the currently active shader.



/**
* Prepare for rendering.
*/
LSVOID LSE_CALL COpenGlVertexBuffer::PrepareToRenderApi() {
assert( CFnd::GetShader() );
CCriticalSection::CLockerS lsLockBindBuffer( m_csCrit );
COpenGl::glBindBufferLSG( GL_ARRAY_BUFFER_ARB, m_uiVboId ); // Even if m_uiVboId is 0, that is perfect.
// Set vertex attribute locations.
for ( LSUINT32 I = m_vVertexAttribPointers.Length(); I--; ) {
const LSG_VERTEX_ATTRIB_POINTER * pvapThis = &m_vVertexAttribPointers;
GLint iAttrib = CFnd::GetShader()->GetAttributeLocs()[pvapThis->ui16LocClass].iLoc[pvapThis->ui16LocUsage];
COpenGl::glEnableVertexAttribArrayLSG( iAttrib );
COpenGl::glVertexAttribPointerLSG( iAttrib,
pvapThis->ui8Elements,
pvapThis->ui32Type,
pvapThis->ui8Normalized,
m_siStride,
reinterpret_cast<LSUINT8 *>(pvapThis->pvPointer) + m_ui32OffsetApi );
CFnd::AssertError();
}
}



Unfortunately, this is where you will have to deviate. My shader language fully parses input shader files, which are in a syntax I created.
When translating to GLSL, my parsers also extract information from the shader file, such as which variable (by name) is meant to connect to the vertex attribute “NORMAL1”, etc. Using this information the shader builds a table of attribute locations that the vertex buffer(s), as you can see here, uses to connect its attributes with those in the shader.


You will have to find a way to quickly and dynamically make this connection. It needs to be dynamic so that any vertex buffer will work with any shader as long as the attribute names match, in any way, the combination of vertex buffers that are active.

If you use predefined names for the variables in your shader it could work similarly to the way mine does, but you would have the restriction that those names exactly must be used. For example:
g_vPos3_0
g_vNormal3_0
g_vNormal3_1


Any of these names that are not accessed by the shader will be stripped, so you can search for all of them and build a table of only the ones that are actually there. That table could be used the way I have done with mine.

Uniforms can be done the same way. If they always use the same names you can simply keep a table of their locations and update only the ones that are actually used by the shader.
I do this in the CFnd::Render() function.


From the sounds of your framework this would be suitable, since you are only planning to use a subset of its functionality for effects that are well defined in nature and functionality.
I considered a similar thing long ago but my method for generating dynamic shaders is to have the model fill a small structure with flags that describe what properties it wants from the shader and submit that to a shader manager for the model system. The shader manager will either increase the reference count on the existing shader that matches the request or make a new one.
When a new shader is made, the same flags that were submitted are translated into macro definitions which enable some inputs and some pieces of code.
This way I only have to write one shader and it gets permutated automatically and on-the-fly. And shared, to reduce the cost of shader swapping.
This, along with the on-the-fly mix-and-match capabilities of the vertex buffers, handles changing meshes just as easily as static ones.

Hope that covers most of your bases.


L. Spiro

What trouble do you encounter on my site? I do need to know if others are having troubles registering or posting.

Also this topic gives me an idea for my next blog. I am taking inspiration for the topics of my blogs based on what questions people ask here so this is also just as useful for me as it is hopefully for you.
I will explain my rendering system in greater detail soon, but maybe I can help you with your immediate problem here before that.


L. Spiro

A critical section in your main render loop code? REALLY? I mean... REALLY?

(Also, 'C' prefixes on classes make my brain vomit... yay for unreadable mess \o/)

Actually the locker class does nothing to the critical section unless the user of the engine builds with a macro set that indicates that resources might be loaded on another thread while rendering takes place.
You may also want to consider the low overhead of ring-3 locking, which consists of only a few instructions on x86 and x64.
I currently do not load resources in a multi-threaded fashion so I am free to enable and disable the locker functionality simply to check for any discrepancies in performance. There are none.
There was a topic recently in which a guy asked why he is always told what not to do, but never why. For example, why must he not allocate memory in constructors? A lot of advice floating around about performance and safety issues were once valid, but as technology advances not many people ever re-evaluate which parts of advice have changed.
Of course I am open to suggestions on any better way to make OpenGL play nicely in a fully multi-threaded environment.

Additionally, classes without the “C” prefix (programmers often enclose single characters in single quotes when writing; most curious) make me vomit and create an unreadable mess.
But my opinion on code “prettiness” is no more—or less—valid than yours.

There's just no way to add new topics or respond to existing ones.

The buffer idea is pretty much what I had in mind, but it means that updating the buffer essentially becomes the same thing, just as a member instead of directly (unless you had methods to the new buffer class, I guess).

I do cache my resources, so if, say, you request a shader with whatever features, it will return a cached one or create it if it doesn't exist.
My shader objects are currently a thin wrapper that has the shader ID in them, as well as all the attributes and uniforms in a hash-map, so matching uniform names dynamically shouldn't be hard (I am not explicitly giving attributes specific indices (slots?), but rather let the driver do whatever it wants).
I didn't mean there is an issue with updating uniforms (or giving them values to begin with), I just don't see a nice way to do it.

I would go and make a general mesh class, but I am currently only having simple 2D textured quads (really, having a 2D grid world is so easy to maintain), so I am not sure how to associate textures with the buffer data.

I would like to also talk about skeletal animations, but I'd rather do it in your site since it isn't really related to this subject.

Are you talking about the forums?
Other people have made new threads and replied after first registering. Perhaps your account is not activated?
You can PM me your user name and where you want to post and we can continue from PM. I have a feeling this topic is going to become more about coding practices and styles rather than OpenGL wrappers.

But I can add one more on-topic thing.
Textures don’t need to be associated with any buffers. They should be entirely decoupled from shaders and vertex buffers.
You can add one sampler in your shader per texture slot (typically there are 8), each with a fixed name, and the first time the shader becomes active it assigns them to their texture units.
After that you need a way only to syncronize between which slots have what meaning. That is, this slot is for diffuse textures, this one for normal maps, etc. And then finally use the same “slot map” (slots here are texture units in OpenGL) to put whatever diffuse texture in this slot and whatever normal-mapping texture in that slot.

This would be handled by your model class. The way I handle this is that my model class may have any number of layered textures, and on the base layer there may also be normal maps, specular maps, etc.
So I keep a slot index counter and run through every layer on the model. For each layer I check for a diffuse texture. On the base layer I also check for other types of textures. Each texture I find, in my search order, is assigned the current slot index and the slot index increases.
The locations of textures (diffuse texture in slot 0, normal map in slot 1, 2nd-layer diffuse texture in slot 2, etc.) as well as which set of texture coordinates they use (diffuse 0 and normal map 0 will take slots 0 and 1, but will both access texture coordinates 0) is part of the information I send to the shader manager to get a permutation for that configuration.
My model class keeps this information as well, and uses it to properly assign the textures to their slots when rendering.


L. Spiro


[EDIT]
I have fixed your forum permissions so that you can post.
[/EDIT]

I remember that (OpenGL contexts are per-thread) now. How much I have forgotten.
Indeed I will have to make a single-threaded queue. Thank you for the reminder.

I feel so out-of-shape in some parts of OpenGL.


L. Spiro

(Also, 'C' prefixes on classes make my brain vomit... yay for unreadable mess \o/)

totally agreed, it is just unreadable... that little snippet reminds me of win32 programming with LPVOIDs and other stuff...