TAKT showcase: rav1e 0.8.0, AV1 encoding of Xiph test clips

Two builds of the rav1e AV1 encoder, version 0.8.0: the stock release and the same encoder after TAKT optimization. On five public Xiph.org test clips both write the same AV1 files, byte for byte, and the TAKT build is 1.38× faster at speed 10 and 1.30× faster at speed 6 (CPU cycles, one thread). Here you get both programs, a script that downloads the clips from Xiph.org, a script that runs the comparison, and our measurements.

What the benchmark does

Each clip is encoded with one command line, identical for both programs:

rav1e <clip>_first64.y4m --threads 1 --speed 10 --quantizer 100 --limit 64 --output <clip>.ivf
rav1e <clip>_first64.y4m --threads 1 --speed 6  --quantizer 100 --limit 64 --output <clip>.ivf

Speed 10 is rav1e's fastest preset, speed 6 is its default. The clips are the first 64 frames of original 8-bit 4:2:0 Y4M files from the Xiph.org Video Test Media collection, unchanged (no resize, no re-encode):

Clip	Frame size, rate	Content	Xiph.org source file
Elephants Dream	1280×720, 24 fps	3D animation	`elephants_dream_720p24.y4m.xz`
Park Joy	1280×720, 50 fps	people and foliage in a park	`park_joy_420_720p50.y4m`
Johnny	1280×720, 60 fps	talking head	`Johnny_1280x720_60.y4m`
Bus	352×288 (CIF), 30 fps	bus and street traffic	`bus_cif.y4m`
Stefan	352×240 (SIF), 29.97 fps	tennis player	`stefan_sif.y4m`

Measurement (TAKT rig)

AMD Threadripper PRO 5975WX (Zen 3), Linux. One dedicated core, the rest of its core complex kept idle; user-mode cycles and instructions from perf stat, plus wall time. 7 interleaved rounds: in every round each clip is encoded by both programs back to back, stock first in odd rounds and TAKT first in even rounds; we take the median per clip and program. The speedup for a preset is the sum of the five stock medians divided by the sum of the five TAKT medians. Exactly the files in bin/ were measured, with tools/run_bench.sh.

Preset	Stock, G cycles	TAKT, G cycles	Speedup, cycles	Stock, s	TAKT, s	Speedup, time
speed 10	139.8	101.6	1.38×	32.3	23.7	1.37×
speed 6	446.2	343.2	1.30×	101.9	78.2	1.30×

Per clip (speedup by cycles; by wall time in brackets):

Clip	speed 10	speed 6
Elephants Dream	1.31× (1.31×)	1.28× (1.27×)
Park Joy	1.38× (1.38×)	1.29× (1.30×)
Johnny	1.40× (1.38×)	1.37× (1.36×)
Bus	1.40× (1.38×)	1.29× (1.29×)
Stefan	1.45× (1.40×)	1.33× (1.33×)

The smallest per-clip speedup is 1.28× by cycles (Elephants Dream, speed 6). All samples, medians and the instruction counts are in measurements.json.

Equivalence

On this benchmark the TAKT build writes exactly the same bitstream as stock rav1e: in every round all ten encodes (5 clips × 2 presets) were compared byte for byte, 70 pairs in total, 70 identical. The SHA-256 of each output file is in tools/expected_ivf.sha256; tools/run_bench.sh checks your outputs against it. This showcase covers the configuration above (8-bit 4:2:0 input, one thread, speeds 10 and 6, quantizer 100); other rav1e options were not part of this measurement.

Running

python3 tools/fetch_clips.py              # downloads about 300 MB, writes clips/ (283 MB), checks SHA-256
tools/run_bench.sh                        # 5 rounds, both presets, both programs; about 20 min on our rig
CPU=3 tools/run_bench.sh 3                # pin every encode to core 3; 3 rounds
SPEEDS=10 tools/run_bench.sh 1            # a quick look: one round at speed 10 (about 1 min)

run_bench.sh prints every encode, a table of medians and speedups, and how many output pairs are identical; it exits with status 1 if any pair differs. Cycles are reported when perf can count cycles:u (otherwise wall time only). The raw data goes to runs.csv in the work directory it prints. Either program can also be run directly, as a regular rav1e command line:

bin/rav1e-0.8.0-takt clips/bus_cif_first64.y4m --threads 1 --speed 6 --quantizer 100 --limit 64 --output bus.ivf
bin/rav1e-0.8.0      clips/bus_cif_first64.y4m --threads 1 --speed 6 --quantizer 100 --limit 64 --output bus-stock.ivf
cmp bus.ivf bus-stock.ivf                 # identical

Requirements: Linux x86-64 with glibc 2.34 or newer, and a CPU with x86-64-v3 (AVX2, FMA, BMI2: Intel Haswell and newer, AMD Zen and newer). Python 3.8+ for fetch_clips.py (standard library only); perf is optional.

How the programs were built

Both the same way: rustc 1.96.0, cargo build --release --locked --bin rav1e with rav1e's own release profile (thin LTO) and default features (including its NASM assembly), -C target-cpu=x86-64-v3 -C debuginfo=0, symbols stripped. bin/rav1e-0.8.0 is built from the unmodified rav1e source at git tag v0.8.0 (commit 1fd67cd); you can build it yourself and compare. bin/rav1e-0.8.0-takt is built from the same source after TAKT optimization. The optimized source is not published: we deliver builds.

bin/rav1e-0.8.0: stock rav1e 0.8.0;
bin/rav1e-0.8.0-takt: rav1e 0.8.0 after TAKT optimization;
tools/fetch_clips.py: downloads the five clips from Xiph.org and checks their SHA-256;
tools/run_bench.sh: runs both programs, compares the outputs, prints the timings;
tools/expected_ivf.sha256: SHA-256 of the ten output files;
measurements.json, SHA256SUMS, LICENSE, THIRD_PARTY_LICENSES.txt.

Licenses and clip rights

rav1e is distributed under the BSD 2-Clause License and the Alliance for Open Media Patent License 1.0; both programs are distributed under the same terms (LICENSE), and the third-party code linked into them is listed in THIRD_PARTY_LICENSES.txt. Our scripts are under the MIT License.

The clips are not part of this package: tools/fetch_clips.py downloads them from Xiph.org, and they remain subject to the terms stated there. Elephants Dream is © Blender Foundation / Netherlands Media Art Institute, licensed under Creative Commons Attribution 2.5 (orange.blender.org). Park Joy is from the SVT MultiFormat test set (Sveriges Television; terms in the SVT document). Johnny is marked public domain in the Xiph.org catalog. Bus and Stefan are standard test sequences of the Xiph.org collection, which lists their original sources.