I do not have a radeon, so I cannot test this.
Have any of you tried to use one object buffer(ob) per vertex property. So 1 ob for vertices, 1 ob for normals, 1 ob for texCoords. What kind of performance/allocation issue have you run into?
I know it is probably better to allocate one big and stuff all the property in, but I guess it also depends on the memory allocator.
For example, if I have vertices, normals and texCoords, I would create 3 ob in a row. And when I done with them I would delete all 3 in row. If a first fit allocator is used with static on one side and dynamic on the other side of an agp memory block, this would not cause more fragmentation than allocating 1 ob.
So my question is what kind of memory allocator is a driver likely to use to allocate the memory?