Serious vertex batching
Author
I've been creating some minimal profiling algorithms to detect general startup prefrences for my graphics application.
I started tweaking w/ the number of verticies per batch sent to the pipeline. Here's some of my results:
10-1k verts - 11mil-16mil draws / sec
2k-4k verts - 24mil - 32mil draws / sec
>4k verts - 11mil - 16mil draws / sec
All my tests are done with varying levels of LOD shaders, no texturing, and lighting disabled.
It seems with such a large bump in that 2k-4k vert range, that it might be worth it to start tossing together smaller batches together @ runtime just to get a nice 2k-4k batch.
Thoughts?
~Main
Sounds about right. There is more to it than the number of verts per batch though, I can't remember the document (was a performance paper on one of the IHV's website) but they kept going on about the number of batches per frame, as well. I can't remember the exact figures, but with about 330 draw calls per frame you will be 100% CPU limited on a 1GHz processor.
-Mezz
-Mezz
Author
100% CPU lmited, or 100% GPU limited?
Not quite sure why just calling draws would bind the CPU speed. Aren't we just tossing the render data down to the GPU for processing?
~Main
Not quite sure why just calling draws would bind the CPU speed. Aren't we just tossing the render data down to the GPU for processing?
~Main
follow th link,
or in brief : the CPU starts to spend too much time just sending stuff to the video card.
If you plan to measure the perf of batch size, make sure that when you use small batches you don't send the same triangles to be drawn over and over again, because in this case, the GPU will always be using its higher speed cache.
the best way is to generate a certain fixed amount of different polys, an render that with more or less batches and then measure the speed, because in this case you will stress the cache in the same manner.
or in brief : the CPU starts to spend too much time just sending stuff to the video card.
If you plan to measure the perf of batch size, make sure that when you use small batches you don't send the same triangles to be drawn over and over again, because in this case, the GPU will always be using its higher speed cache.
the best way is to generate a certain fixed amount of different polys, an render that with more or less batches and then measure the speed, because in this case you will stress the cache in the same manner.
The document Mezz was refering to is a great document, its called batchbatchbatch.pdf from the nvidia site. I'd classify it a must read document.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
