High CPU overhead on very complex scenes

[rytone]

Right now I am working on a project where I need to render very complex lines (~2000 points per line, all straight segments per line, usually 3 lines) every frame and have been trying out Pathfinder for it. While I am aware this is probably not an intended use case, I think this may still be valuable for improving the performance of Pathfinder.

The code I am working with can be found here. (see line 296 for some of the main rendering code)

First of all, in order to even get realtime performance, I had to break each line into smaller PathObjects instead of one single PathObject per line. I ended up with rendering taking roughly 8-11ms per frame (using RayonExecutor, measured very non-scientifically), which doesn't really leave enough time for the other stuff I want to do. One of the things I found was by not using OutlineStrokeToFill, although that obviously renders incorrectly, the render time reduced to about 2-4ms. #119 mentions that the stroke algorithm could be done on the GPU, and that might increase performance greatly in this scenario.

Something else I noticed under this scenario was that RayonExecutor doesn't seem to scale well with added threads. Even though my CPU has 8 threads, using RayonExecutor was only about 70% faster than SequentialExecutor (again measured very roughly, about 16-17ms per frame using SequentialExecutor). Additionally, the CPU usage reported by htop was about double that of using SequentialExecutor

Finally, I decided to compare performance of Pathfinder to Blend2D, a purely CPU based rasterizer. My code for testing it is on the blend2d branch of the same repo. Using it, only 3-5ms were spent drawing the lines. However, the caveat is that I have later steps that I want to perform on the GPU, and copying the rendered image from the CPU to the GPU takes a significant amount of time (at least with my naive implementation), bringing the total frame time back up to only a little faster than Pathfinder using RayonExecutor. Despite that, the CPU usage reported while using Blend2D is roughly half of Pathfinder using SequentialExecutor.

[pcwalton]

Thanks for trying it out; this feedback is very valuable!

Yes, I think you described the biggest potential performance improvement: rendering strokes with a stroke shader instead of using fills. Another potential win will simply be to make tiling faster; @nical has a work-in-progress branch to improve things significantly. Finally, moving tiling to GPU would help a lot.

[pcwalton]

You may wish to try again. With @Veedrac's optimizations and the new tiling code that just landed, I've measured an 83% improvement in CPU time on some test cases.

[pcwalton]

Closing for now as there have been many CPU time improvements, including an entirely new backend, since this was filed. Feel free to reopen if more issues are found.