Tracking issue for performance

memorysafety / rav1d

An AV1 decoder in Rust.

BSD 2-Clause "Simplified" License

245 stars 15 forks source link

Tracking issue for performance #1294

Open rinon opened 1 month ago

rinon commented 1 month ago

This issue is intended to aggregate, track progress, and discuss performance optimization for rav1d.

Unless otherwise noted, the following conditions apply to these measurements:

all C compilation is done with Clang 18 to match the LLVM backend used by current rustc
Test input is 8-bit Chimera
--threads 8

rinon commented 1 month ago

CPU	Test	Time (s)
7700X	dav1d 2355eeb	5.286
7700X	rav1d 412cd4c	5.766 (9%)
7700X	dav1d 2355eeb 10-bit	13.538
7700X	rav1d 412cd4c 10-bit	14.32 (5.8%)
i7-1260p	dav1d 2355eeb	16.147
i7-1260p	rav1d 412cd4c89	17.287 (7%)
i7-12700K	dav1d 2355eeb	6.663
i7-12700K	rav1d 412cd4c89	7.075 (6%)
M2 MacBook (AAarch64)	dav1d 2355eeb	8.958
M2 MacBook (AAarch64)	dav1d w/out backports 412cd4c89	9.106
M2 MacBook (AAarch64)	rav1d 412cd4c89	9.818 (9.6% upstream, 7.8% w/out some backports) **
Pixel 8 (Tensor G3)	dav1d 2355eeb	34.529
Pixel 8 (Tensor G3)	rav1d 412cd4c89	38.504 (11.5%)***

** Some AArch64 relevant backports are not yet completed *** Using NDK 27 RC 1; used hyperfine --warmup 3 ... to lower variance

rinon commented 1 month ago

Latest results (raw times are a bit faster across the board because I'm benchmarking with a quieter OS environment, I re-did baselines for consistency):

8-bit Chimera:	CPU	Test
7700X	dav1d 2355eeb	5.148
7700X	dav1d 2355eeb (full LTO w/ LLD)	5.172 **
7700X	rav1d b80f92261	5.572 (8.2%)
7700X	rav1d #1320	5.492 (6.7%)

10-bit Chimera:	CPU	Test
7700X	dav1d 2355eeb 10-bit	13.204
7700X	rav1d b80f92261 10-bit	14.035 (6.3%)
7700X	rav1d #1320	13.995 (6.0%)

** Full LTO for the C code seems to make performance slightly worse, if anything. I'm surprised by this but the measurements are consistent on my machine.

rinon commented 1 month ago