Open Tenchi2xh opened 2 months ago
Update, after implementing an algorithm that reduces the number of lookups, the scales tipped to the other side:
The optimization alone makes Codon 35x faster, making the same render go from 11 minutes to a mere 19 seconds
The flame graph still looks the same with or without the optimization, so maybe something else is at play (or using dtrace
messes up with Codon?)
Hi, love the project!
I recently started implementing a ray tracer as an exercise for trying out Codon. After a while, I was curious to try and make the code work with vanilla Python and PyPy, and then found out that my renders are about twice as fast with PyPy compared to Codon.
After trying a few optimizations to no avail, I decided to try and profile the execution of the Codon-made binary:
It appears that more than half the time is spent on some internal
gc.alloc_atomic
, and also starting threads? (I have zero@par
in the whole codebase).I noticed that when using
time
, theuser
time is often twice as much as thereal
time (something in a thread is doing something). And in turn thereal
time is still twice as much as PyPy's.My suspicion is that creating a lot of
Vec3
classes all the time is somehow bogging down the GC. Maybe I have a basic misconception of how to use Codon?Here is an interactive version of the flame graph (unzip and open the SVG in a browser), and the code is available here: https://github.com/Tenchi2xh/RTOW-Codon (check out commit
379d5d0
, the master branch now has other types of optimizations). The main entry point isrtow/__main__.py
but it's easier to run it from therun.sh
script (a preprocessor has to remove python-specific stuff). To run it faster, just reducesamples_per_pixel
andmax_depth
on lines 52-53 (it runs even slower in the profiler)(Sorry to link to a whole repo, it's not a big codebase, but big enough to make it hard to produce a minimal reproducible example for a Github issue)
I am using the latest dev build of Codon, downloaded from a CI build