Q_rsqrt NodeJS benchmark

Benchmark of fast inverse square root implementations for NodeJS.

Run

node index.js

Output

Single value inputs:
==========
addon/6: 51.186ms
llvm wasm/6: 13.694ms
wat wasm/6: 10.459ms
simple/6: 9.976ms
js/6: 311.792ms
js external/6: 11.276ms
==========


Array inputs:
==========
addon/1000x10: 0.12ms
wat/1000x10: 0.297ms
simple/1000x10: 0.56ms
==========
addon/1000x100: 0.429ms
wat/1000x100: 1.135ms
simple/1000x100: 0.502ms
==========
addon/1000x1000: 5.03ms
wat/1000x1000: 1.222ms
simple/1000x1000: 5.075ms
==========
addon/1000x10000: 42.844ms
wat/1000x10000: 7.926ms
simple/1000x10000: 48.996ms

/6 - means 1e6 iterations = 1 million

Variants

• addon - node-gyp compiled C addon that uses N-API

  node-gyp build

• llvm wasm - CLang compiled src/rsqrt.c to WASM. Command:

  clang --target=wasm32 -nostdlib -Wl,--no-entry -Wl,--export-all -o build/Release/rsqrt.wasm src/rsqrt.c

• wat wasm - wat2wasm compiled WebAssembly Text Format version of this alg. See src/rsqrt.wat

  npx wat2wasm src/rsqrt.wat

• simple

  const simple = (x) => 1 / Math.sqrt(x);

• js JS version from this gist
• js external JS version with buffers outside of function scope

Conclusion

• NodeJS optimizing jit-compiler is amazing
• WASM is the best way to perform computationally-hard operations over arrays for JS
• I suck at C