Profile-guided optimizations (PGO)

[Kobzol]

I was wondering whether you have considered using PGO for optimizing rav1d. Both for the straightforward "let's generate more optimized artifacts" approach, but also for the less obvious "let's do PGO, see what it has optimized and try to backport these optimizations to the original source code".

rav1d contains a lot of assembly that most likely won't be touched by LLVM's PGO pipeline, but out of curiosity, I tried to apply https://github.com/Kobzol/cargo-pgo to rav1d, and it still seems to produce a quite nice ~2% speedup:

hyperfine --warmup 0 --runs 1 "./rav1d -q -i Chimera-AV1-8bit-1280x720-3363kbps.ivf -o /dev/null --threads=1" "./rav1d-pgo -q -i Chimera-AV1-8bit-1280x720-3363kbps.ivf -o /dev/null --threads=1"
Benchmark 1: ./rav1d -q -i Chimera-AV1-8bit-1280x720-3363kbps.ivf -o /dev/null --threads=1
  Time (abs ≡):        50.679 s               [User: 50.599 s, System: 0.077 s]
 
Benchmark 2: ./rav1d-pgo -q -i Chimera-AV1-8bit-1280x720-3363kbps.ivf -o /dev/null --threads=1
  Time (abs ≡):        49.915 s               [User: 49.840 s, System: 0.074 s]
 
Summary
  ./rav1d-pgo -q -i Chimera-AV1-8bit-1280x720-3363kbps.ivf -o /dev/null --threads=1 ran
    1.02 times faster than ./rav1d -q -i Chimera-AV1-8bit-1280x720-3363kbps.ivf -o /dev/null --threads=1

Anyway, I don't have more than that, just was curious what do you think of this approach.

[kkysen]

In general, PGO would be part of the application using rav1d (e.x. chromium), and would be done for both rav1d and dav1d, so we haven't been exploring it very much. The tests also might not necessarily be super similar to what you'd encounter in real-world usage (we can use the couple videos we've been benchmarking on and it'll help with those, but I'm not sure if other videos will be different).

let's do PGO, see what it has optimized and try to backport these optimizations to the original source code

This could be a very good idea, though. Is there a simple way to set up PGO for Rust these days? How did you do it?

[Kobzol]

I'm using https://github.com/Kobzol/cargo-pgo. You need to install the llvm-tools component with rustup (rustup component add llvm-tools-preview), install cargo-pgo (cargo install cargo-pgo), and then do something like this:

1. Run cargo pgo build to create a PGO instrumented build
2. Run ./target/<target-triple>/release/rav1d ... to gather profiles (it needs to run for the full duration, if you cancel it in the middle, profiles might not be generated)
3. Run cargo pgo optimize to create a PGO optimized build