Performance consern regardless `zenkit::Polygon`
Closed this issue · 6 comments
I've noticed a performance regress on load-time after merging new version of phoenix(zenkit). While haven't done a real profiling yet, pausing application in debug almost always shows this callstack:
For now - just raising awareness.
And while at it, can you remind me why zenkit::Polygon is a thing?
Asking about it, because currently I'm looking into more efficient mesh packing, and one of promising direction is instantiated list of 62 triangle strips - basically looking into what are the options are.
Hio, I added the Polygon for compatibility with the BSP-Tree. Projects like GothicVR want to use it for performance improvements themselves. Yes, I know that it's a bit slow with all the allocations happening so I do mean to write my own allocator at some point to speed it up (or to just set an upper limit on vertices per polygon).
write my own allocator
Alternative might be to put features/vertices in a contiguous array of data and reference it with offsets in Polygon or with std::span
@Try I've improved performance by ~25% on my machine. Please try it out the patch and let me know if this is to your liking :)
Testing, time that World::World(...) takes
Baseline:
// debug
newworld.zen load time: 47.480000
newworld.zen load time: 47.005001
newworld.zen load time: 47.536999
oldworld.zen load time: 11.428000
oldworld.zen load time: 13.113000
oldworld.zen load time: 12.029000
// rel-debug
newworld.zen load time: 52.644001
newworld.zen load time: 52.631001
newworld.zen load time: 52.368999
oldworld.zen load time: 12.678000
oldworld.zen load time: 12.822000
oldworld.zen load time: 12.899000
With new changes:
// debug
newworld.zen load time: 12.190000
newworld.zen load time: 12.774000
newworld.zen load time: 16.729000
oldworld.zen load time: 4.434000
oldworld.zen load time: 4.440000
oldworld.zen load time: 4.618000
// rel-debug
newworld.zen load time: 16.590000
newworld.zen load time: 13.988000
newworld.zen load time: 14.205000
oldworld.zen load time: 7.482000
oldworld.zen load time: 7.746000
oldworld.zen load time: 7.606000
Aside from mingw curiosities, when release build runs slower than debug, can say that it pretty substantial improvement.
More than 3x - great :)
Out of curiosity: what compiler version are you using? It seems to produce very low performance code :/
MinGW 11.2.0 - the one that comes, as the default, in package with QtCreator on windows.
The code probably is not important in this case. Most likely, c++ runtime, that comes with mingw has non optimized allocator for windows