The following benchmark suites are included, some of which required modifications to work with WASM/WASI:

• polybench: Polybench; 90 benchmarks (30 benchmarks x 3 sizes)
• mibench: Mibench; 35 benchmarks (18 benchmarks having small and large sizes except one)
• cortex: UCSD CortexSuite, excluding the San Diego Vision Benchmark suite; 24 benchmarks (8 benchmarks x 3 sizes)
• vision: San Diego Vision Benchmark Suite; 27 benchmarks (9 benchmarks x 3 sizes)
• libsodium: Libsodium Benchmarks, available pre-compiled here; 70 benchmarks (15 benchmark types, various configurations per type)
• apps: benchmarks made from common real-world applications:
  • python: the CPython interpreter compiled to WASM running 12 python scripts taken from the Codon benchmark suite. Note that WASM/WASI-compiled python currently does not support pip-installed modules; complicated imports are also somewhat flaky.
  • ffmpeg: FFmpeg standalone executable compiled to WASM, supporting zlib and x264 with 6 different operations.
  • lua: Lua interpreter, running 5 different scripts; requires WALI instead of only WASI to run.
  • sqlite3: SQLite running 5 different queries; requires WALI.

1. Install the WASI SDK:

   wget https://github.com/WebAssembly/wasi-sdk/releases/download/wasi-sdk-20/wasi-sdk-20.0-linux.tar.gz
   tar xvf wasi-sdk-20.0-linux.tar.gz
   sudo cp -r wasi-sdk-20.0 /opt/wasi-sdk

2. Build benchmarks:

   • make: build all benchmarks; saved to ./wasm. Any necessary data files are saved to ./data. Since paths are hard-coded into benchmarks, benchmarks must be executed with ./benchmarks as a WASI pre-opened directory (or equivalently, wasm and data in WASI working directory).
   • MODE=native make: build all benchmarks natively using gcc for comparison; saved to ./native.
   • make {benchmark}: make benchmark suites individually.
   • make clean: remove ./wasm and ./data.