profiling

[joevenzon]

Interesting! It's starting to sound like it might be operating system related. I'm using GCC 4.4.1 and an AMD X4 on Linux.

Does codeanalyst's line-by-line profiling information give you any ideas about optimizing this function? It's pretty cache unfriendly at the moment due to iterating over maps and lists, and it does a fair amount of data copying, but maybe there's something else that's an obvious reason for its slowness that I'm missing.

I'm thinking about how to rewrite the scenegraph system and I'm undecided about where to go with it. Right now it's a sort of scenegraph lite, with only two types of nodes. Should I make it a more conventional scenegraph system with nodes for texture changes, shader changes, etc? Should I try to make it only re-traverse branches of the tree that have changed since last time (but increasing the overhead of changing a node)? Should I add a hierarchical space partitioning system so culling is faster? If so, how to integrate that into the scenegraph and minimize render state changes while keeping the culling process as fast as possible?