wasm-perf
Explore how to achieve maximum performance in WebAssembly
MIT License
WebAssembly Performance
Explore how to achieve maximum performance in WebAssembly.
Full post: https://jeromewu.github.io/improving-performance-using-webassembly-simd-intrinsics/
Setup
We benchmark the performance by measuring time to complete 10,000 times of two 64 x 64 matrix multiplication with various implementation and flags:
Implementation:
- mul_mats.js: JavaScript version
- mul_mats.c: C version
- mul_mats_intrin.c: C version with SIMD intrinsics
Flags:
-
-O3: optimize code and enable vectorization -
-msimd128: enable SIMD 128 optimization
Performance
Environment:
- Hardware
- CPU: 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz
- RAM: 16 GB
- Software
- Linux Kernel: 5.19.1-3-MANJARO
- NodeJS: v16.16.0
- Emscripten: 3.1.18
- GCC: 12.2.0
Here we use mul_mats.js as baseline to compare the speed / time.
| Implementation | Flags | Time |
|---|---|---|
| mul_mats.js | 5.768s (+-0%) | |
| mul_mats-wasm.js | 6.865s (+19%) | |
| mul_mats-wasm-simd.js | -msimd128 | 6.964s (+21%) |
| mul_mats_intrin-wasm-simd.js | -msimd128 | 4.718s (-18%) |
| mul_mats-wasm-O3.js | -O3 | 2.012s (-65%) |
| mul_mats-wasm-O3-simd.js | -O3 -msimd128 | 0.278s (-95.2%) |
| mul_mats_intrin-wasm-simd-O3.js | -O3 -msimd128 | 0.245s (-95.8%) |
| mul_mats (gcc) | 7.697s (+33%) | |
| mul_mats-O3 (gcc) | -O3 | 0.401s (-93%) |
| mul_mats-O3-avx (gcc) | -O3 -mavx2 | 0.119s (-98%) |
Use SIMD intrinsics with -O3 and -msimd128 flags can be 95.8% faster than pure
JavaScript implementation. 🎉
Try it
To build binaries, you need to install Docker 19.03+ and run:
make
You should find all binaries in dist/ folder.
To run all of them, simply hit:
make run-all
Some of the execution might failed you are NOT using Linux, check Makefile to see how to run a specific case.