From: Charles L. <ch...@cn...> - 2009-08-20 03:28:02
|
I've just examined some changes when it comes to putting the burden on the GPU and not. Reducing the quantity of certain calls, like glLoadMatrix(...) did not seem to affect the performance at all when the GPU was under load, however had a significant improvement when I shifted the CPU back to being the bottleneck. 472 -> 472 772 -> 868 In order to reduce the overhead, I made the system only load the matrix and update the shader if the specific node had children. Otherwise, it just keeps going. This indicates to me that some (possibly many) of the OpenGL functions we thought may have had something to do with stealing cycles from the GPU, do not, that it is possible there may be another level of pipelining going on here that we don't understand. Additionally, I feel as though we shouldn't strive excessively hard to speed up the CPU if that's not the bottleneck, but the fact is, that in the end, we should strive to make it as fast as possible because there are always portions where it won't be the GPU that's bottlenecked, and will in fact be the GPU waiting on the CPU. The two systems cannot be totally decoupled. I feel we can learn a lot by examining the performance changes both under GPU bottleneck (Window at 640x480) and CPU bottleneck (Window at 50x50 or less). Sadly, I'm not sure how much we can push this, since we've seen examples where doing more work can result in a higher framerate. Charles Josh wrote: > Attached is a count of opengl API calls for a scene where most things > are culled. > I will be figuring out how to make this more efficient. > ------------------------------------------------------------------------ > > ------------------------------------------------------------------------------ > Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day > trial. Simplify your report design, integration and deployment - and focus on > what you do best, core application coding. Discover what's new with > Crystal Reports now. http://p.sf.net/sfu/bobj-july > ------------------------------------------------------------------------ > > _______________________________________________ > Hgengine-devs mailing list > Hge...@li... > https://lists.sourceforge.net/lists/listinfo/hgengine-devs |