17

I have to draw a buffer that holds a couple thousand vertices. I am using a vbo to store the data.

I know I will have to update the VBO many times - but only in small parts at a time.

So I am wondering what the best method to doing so is:

  1. Split VBO up into smaller VBOs (that hold like 300 verts) and then update individual VBOs with 1 call?
  2. One big VBO and use lots of glBufferSubData() calls?
  3. Use glMapBuffer() and one big VBO?
Brian Tompsett - 汤莱恩
  • 5,753
  • 72
  • 57
  • 129
Pubby
  • 51,882
  • 13
  • 139
  • 180

2 Answers2

10

There is another option, which is a bit like option 3 - use one big VBO (probably with GL_STREAM_DRAW mode) that is reset each frame (by calling glBufferData with a NULL buffer pointer and the same size each time) then glMapBuffer-ed right away. The buffer is left mapped as it is filled in, then unmapped just before drawing. Repeat.

The call to glBufferData tells OpenGL that the old buffer contents aren't needed, so the glMapBuffer doesn't have to potentially wait to ensure the GPU is finished with by the GPU.

This approach seems to be the one officially sanctioned by the vertex_buffer_object extension. See the "Vertex arrays using a mapped buffer object" example:

http://www.opengl.org/registry/specs/ARB/vertex_buffer_object.txt

This suggests that OpenGL (or the driver?) will be watching for this sort of behaviour, and (when spotted) arrange things so that it is performed efficiently.

2
  1. Doesn't sound like a good idea: it forces you to draw it in several calls while changing the bound buffer between each draw call.
  2. Might do the trick if your buffer is huge.
  3. The whole buffer will certainly be uploaded to the GPU. This will certainly be as efficient as one glBufferData, but you can do it asynchronously.

If think that glBufferData or glMapBuffer are the better solution if your buffer is small. 100000 * sizeof(float) * 3 ~= 1MB. There should be no problem with that.

tibur
  • 11,531
  • 2
  • 37
  • 39
  • Side note: I already know I have to draw it in several calls - whether or not I use many VBOs or not. I also know that I will only be drawing part of the vertices at a time. Is there lots of overhead to changing bound VBO buffer? Also, is mapBuffer slower than bufferdata? – Pubby Jan 31 '11 at 18:44
  • The advantage of mapBuffer is that you can copy your data inside the buffer while GL is doing something else. On the other hand, the real perf of buffer mapping is really implementation dependant. If you can avoid changing your VBO multiple time, I would advise you to do it. – tibur Jan 31 '11 at 18:53