Vertex Specification Best Practices
|This article needs proper formatting or improved descriptions. Further details can be found on the talk page.|
See VBO for general details.
Size of a VBO/IBO
- How small or how large should a VBO be?
You can make it as small as you like but it is better to put many objects into one VBO and attempt to reduce the number of calls you make to glBindBuffer and glVertexPointer and other GL functions.
You can also make it as large as you want but keep in mind that if it is too large, it might not be stored in VRAM or perhaps the driver won't allocate your VBO and give you a GL_OUT_OF_MEMORY.
1MB to 4MB is a nice size according to one nVidia document. The driver can do memory management more easily. It should be the same case for all other implementations as well like ATI/AMD, Intel, SiS.
Formatting VBO Data
VBOs are quite flexible in how you use them. For instance, there are a number of ways you can represent vertex attribute data in VBOs:
- (VVVV) (NNNN) (CCCC)
- One option for using them would be, for each batch (draw call) allocate a separate VBO per vertex attribute. This is certainly possible. If you have vertex, normal, and color as vertex attributes, pictorially this is: (VVVV) (NNNN) (CCCC)
- Another approach is to store the vertex attribute blocks in a batch, one right after the other, in the same block and stuff them all in the same VBO. When specifying the vertex attributes via
gl*Pointer()calls you'd pass byte offsets into the VBO to the ptr parameters. Pictorially, this is: (VVVVNNNNCCCC).
- Yet another approach is to interleave the vertex attributes for each vertex in a batch, and then store each of these interleaved vertex blocks sequentially, again combining all the vertex attributes into a single buffer. As before, you'd pass byte offsets into the VBO to the
gl*Pointer()ptr parameters, but you'd also use the stride parameter to ensure each vertex attribute array access only touched elements for that attribute array. Pictorially, this option is: (VNCVNCVNCVNC)
Now this is just a single batch. There's also nothing stopping you from storing the vertex attribute data for multiple batches inside a single VBO or set of VBOs.
The optimal layout depends on the specific GPU and driver (plus OpenGL implementation), but we can apply a little common sense to help steer our choices. Firstly, one should always vectorize data, and the size of each vector should ideally have four or eight 4-byte (i.e. float or int) components, since that makes them cache-friendly and memory-request-friendly, and also allows modern GPUs to more easily facilitate instruction level parallelism. That means that your 3D position, normal and color information, for example, should be padded into 4D vectors. This works well with most vector operations, given the last component is 0. If it is not zero, certain operations, such as the cross product, will yield wrong results. If you use 3-component vectors instead, some GPU drivers will automatically do the padding for you, others will simply suffer from reduced throughput and increased cache misses.
Secondly comes the big question of whether or not to interleave your data. That really depends on how the rendering pipeline inside each GPU is setup. Some are optimized for one format and some for the other. However, keep in mind that by default, global memory access on the GPU is sequential. Modern GPUs also have eight streaming multiprocessors. That means that they can only execute eight different instructions at the same time. However, each processor can concurrently run up to over 100 shaders, each performing the same task on a different vertex or pixel. Given the current instruction does something with a normal, all 100 shaders will send four memory fetches to the memory scheduler. If the currently requested vertex information is not ready yet, the scheduler will perform suboptimal reordering to guarantee as few memory fetches as possible and then send each request batch to memory. One request being able to deliver 32, 64, 128 or 256 bytes of sequential data. If the shaders are working on neighboring vertices and the normals are not interleaved with other data, a whole bunch of normals can be retrieved at once (up to 256/16 = 16 normals). If it is interleaved with other data, a lot of the data coming back from memory would be useless for computing the normal which results into much lower throughput. However, there are other things, like shared memory and the geometry- and texture-caches that all play a large role in the actual efficiency of the data layout. The best approach would be running benchmarks, and then sharing results with the community, by posting or linking them to this wiki.
Vertex, normals, texcoords
- Should you create a separate VBO for each? Would you lose performance?
If your data is static, then make 1 VBO for best performance. Be sure to interleave your vertex attribute data in the VBO and make the data block for each vertex a multiple of 32 bytes for good cache line coherence. See the other VBO page because it explains these details.
If one of the vertex attributes is dynamic, such as the vertex positions, you could store this in separate VBO.
By dynamic, we mean that you will be updating the VBO every frame. Perhaps you want to compute the new vertices on the CPU. Perhaps you are doing some kind of water simulation. etc.
No, you don't lose much performance if you use separate VBOs. It would be on the order of 5% but your testing might show otherwise.
//Binding the vertex glBindBuffer(GL_ARRAY_BUFFER, vertexVBOID); glVertexPointer(3, GL_FLOAT, sizeof(float)*3, NULL); //Vertex start position address //Bind normal and texcoord glBindBuffer(GL_ARRAY_BUFFER, otherVBOID); glNormalPointer(GL_FLOAT, sizeof(float)*6, NULL); //Normal start position address glTexCoordPointer(2, GL_FLOAT, sizeof(float)*6, sizeof(float*3); //Texcoord start position address
- If the contents of your VBO will be dynamic, should you call glBufferData or glBufferSubData (or glMapBuffer)?
If you will be updating a small section, use glBufferSubData. If you will update the entire VBO, use glBufferData (this information reportedly comes from a nVidia document). However, another approach reputed to work well when updating an entire buffer is to call glBufferData with a NULL pointer, and then glBufferSubData with the new contents. The NULL pointer to glBufferData lets the driver know you don't care about the previous contents so it's free to substitute a totally different buffer, and that helps the driver pipeline uploads more efficiently.
Another thing you can do is double buffered VBO. This means you make 2 VBOs. On frame N, you update VBO 2 and you render with VBO 1. On frame N+1, you update VBO 1 and you render from VBO 2. This also gives a nice boost in performance for nVidia and ATI/AMD.
Vertex Layout Specification
A lot of new code gets written this way
glBindBuffer(GL_ARRAY_BUFFER, vboID); glEnableClientState(GL_VERTEX_ARRAY); glVertexPointer(3, GL_FLOAT, sizeof(TVertex_VNTWI), info->posOffset); glTexCoordPointer(2, GL_FLOAT, sizeof(TVertex_VNTWI), info->texOffset); glEnableClientState(GL_TEXTURE_COORD_ARRAY); glNormalPointer(GL_FLOAT, sizeof(TVertex_VNTWI), info->nmlOffset); glEnableClientState(GL_NORMAL_ARRAY); /////////////// int weightPosition = glGetAttribLocation(programID, "blendWeights"); glVertexAttribPointer(weightPosition, 4, GL_FLOAT, GL_FALSE, sizeof(TVertex_VNTWI), info->weightOffset); glEnableVertexAttribArray(weightPosition); /////////////// int indexPosition = glGetAttribLocation(programID, "blendIndices"); glVertexAttribPointer(indexPosition, 4, GL_UNSIGNED_BYTE, GL_FALSE, sizeof(TVertex_VNTWI), info->indexOffset); glEnableVertexAttribArray(indexPosition); /////////////// glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, iboID); glDrawElements(GL_TRIANGLES, numIndices, GL_UNSIGNED_SHORT, 0);
and in the shader, would be using gl_Vertex, gl_Normal and gl_MultiTexCoord0. It is better to use generic vertex attributes for your vertex, normal and texcoord as well, since it is the modern way of specifying your vertex layout. You are already using it for your blendWeights and blendIndices.
In GL 3.1+ core contexts, you are forced to use your own vertex attributes with calls to glVertexAttribPointer.