panvpx Benchmark

A standalone Gradle subproject that measures VP8 and VP9 encoder/decoder throughput using the panvpx library.

Running

# Show usage
./gradlew :benchmark:run

# Full benchmark — all codecs, encoder + decoder
./gradlew :benchmark:run --args="--codec=both --mode=all"

Options

Option	Default	Description
--codec=vp8|vp9|both	both	Codec(s) to benchmark
--mode=encoder|decoder|all	all	Mode(s) to benchmark, comma-separated
--width=N	1920	Frame width in pixels
--height=N	1080	Frame height in pixels
--bitrate=N	4000	Target bitrate in kbps
--threads=N	1	Encoder and decoder thread count
--warmup=N	3	Warmup frames (excluded from results)
--frames=N	25	Measurement frames
--preset=quality|realtime	quality	Encoding deadline: quality = 1 s, realtime = 1 µs
--cpu-used=N	0	CPU speed/quality trade-off (VP8: 0–16; VP9 quality: 0–5; VP9 realtime: 5–8)
--row-mt	off	Enable VP9 row-based multithreading
--tile-columns=N	0 (1 column)	VP9 tile columns — actual count: 1, 2, 4, 8, 16, 32, or 64
--token-partitions=N	0 (1 partition)	VP8 token partitions — actual count: 1, 2, 4, or 8
--input=/path/to/file.yuv	—	Raw I420 YUV input file
--synthetic	off	Use generated frames instead of the built-in clip

Input frames

Without --input or --synthetic, the benchmark uses a built-in VP9 1080p25 clip (bundled as a classpath resource). This is the recommended baseline for reproducible comparisons.

--synthetic generates I420 frames on the fly using a three-wave interference pattern with per-frame temporal variation. Synthetic frames are more compressor-resistant than simple gradients — no single motion vector can compensate all three waves simultaneously — making them a reasonable stress test when no real content is available.

--input=/path/to/file.yuv reads a raw I420 YUV file. --width and --height must match the file's actual dimensions.

Output

Each run prints a summary header then a results table:

panvpx Benchmark — 1920x1080 @ 4000 kbps  |  1 threads  |  quality  |  25 frames  |  3 warmup  |  built-in clip (1080p25)

Running Encoder/VP8 (3 warmup + 25 measure frames)…
Running Decoder/VP8 (25 measure frames, setup encodes first)…
…

Mode          Codec   Throughput    Out (MB/s)   Avg frame   Runtime
Encoder       VP8     123.4 fps     4.56 MB/s    12345 B     0.203 s
Decoder       VP8     567.8 fps     4.56 (in)    12345 B     0.044 s
…

Column	Encoder	Decoder
Throughput	Encode fps (includes flush)	Decode fps
Out / In (MB/s)	Compressed output rate	Compressed input rate fed to decoder
Avg frame	Average compressed packet size	Average compressed packet size
Runtime	Wall time for all measurement frames + flush	Wall time for all measurement frames

Examples

# VP9 encoder, 8 threads, realtime preset, fastest cpu-used, tile columns
./gradlew :benchmark:run --args="--codec=vp9 --mode=encoder --threads=8 --row-mt --tile-columns=8 --cpu-used=6 --preset=realtime"

# VP8 encoder with 4 token partitions, quality preset
./gradlew :benchmark:run --args="--codec=vp8 --mode=encoder --token-partitions=4 --preset=quality"

# Decoder-only benchmark, 720p, 4 threads, external file
./gradlew :benchmark:run --args="--codec=vp9 --mode=decoder --width=1280 --height=720 --threads=4 --input=/tmp/video.yuv"

# Quick encoder + decoder comparison with more frames
./gradlew :benchmark:run --args="--mode=encoder,decoder --warmup=10 --frames=50"

# Synthetic frames stress test, VP9 only
./gradlew :benchmark:run --args="--codec=vp9 --synthetic --frames=50"

Measurement methodology

• Warmup: the first --warmup frames are encoded/decoded and discarded. For VP9 in quality mode the 25-frame lookahead is fully drained during warmup so its latency does not inflate the measurement.
• Flush included: the timer runs from the first encode() call through the end of the flush loop. VP9 quality mode emits most packets during flush, so omitting it would produce misleadingly high fps numbers.
• Single-threaded by default: --threads=1 isolates single-core throughput. Increase with --threads=N to measure multi-threaded scaling.