Skip to content

Latest commit

 

History

History
50 lines (33 loc) · 1.49 KB

ffmpeg.md

File metadata and controls

50 lines (33 loc) · 1.49 KB
Raw

A benchmark with ffmpeg

What's this

A simple benchmark with ffmpeg, intended to measure computation performance with a real application.

For each runtimes, it converts a video frame with ffmpeg.wasm and measures the time taken.

Result

Result

Observations

  • toywasm performs reasonably well for an interpreter. It's slower than the fastest interpreters though.

  • toywasm with annotations disabled is very slow.

  • toywasm's memory consumption is reasonable as well. (See "maximum resident set size" and "peak memory footprint". I don't know what exactly these metrics reported by macOS's time(1) mean though. Also, for some reasons, these numbers seem a bit unstable and vary among runs. If you know better metrics to show memory footprint, preferably available on macOS, please let us know.)

  • toywasm's SIMD overhead is not huge for this particular workload. (Note: this workload doesn't actually involve SIMD at all.)

  • toywasm is far slower than JIT-based engines as expected.

  • Some of runtimes involve surprisingly large RSS. I'm not sure why.

  • Many of JIT-based engines use multiple threads to hide its compilation time. (You can see their "user" time is often longer than "real" because of this.)