(Keep) Making the Web Faster with SIMD.js

Sajjad Taheri

Little About Me

• From Iran
• CS Phd Student at UC Irvine
• Working on high performance communication and computation for the Web platform

  • WebRTC

  • Computer Vision for the web

    • RGB/Depth Image processing

    • CV and SIMD = a perfect match!

SIMD

• Single Instruction Multiple Data
• A class of parallel computer architecture

• do an arithmetic operation on multiple data points in parallel

SIMD

• Many variants: SSE, AVX, FMA
  • they differ at
    • Vector operations they support
      • SSE4 has richer instructions than SSE2
    • Vector width = do more operations in one cycle        
      • AVX vectors have twice the size of SSE's
• SIMD.js considers SSE2 as the minimum SIMD support
  • Vectors are 128 bits wide

SIMD.js

• JavaScript API to expose SIMD to web apps
• Browsers provide efficient implementation for each underlying hardware
  • SIMD offers so much and many web apps can benefit from it
    • 3D Graphics, Video processing
  • Its already available on most processors
    • E.g. first X86 CPUs to support SSE2 :
      • Pentium 4 (2001), AMD K8 (2003)

SIMD.js

Firefox (Nightly) was the first browser to implement

A SIMD.js Demo

Boolean Vectors

• Latest addition to SIMD.js API
  •  New data types
    • Bool8x16, Bool16x8, Bool32x4
  • Operators
    • logical, comparison, selection, ...
• Boolean vectors are mainly used for vector comparison and conditional assignments
• Been mimiced by integer vectors before

Program will be cleaner

Performance will be better

SpiderMOneky Boolean Vectors Implementation

• Add boolean vectors and their operators to the interpreter
• JIT Compiler (Ion Monkey)
  • Stack/register allocator
  • Generate efficient machine code for operators (currently only Bool32x4 and for X86/64)

GL-Matrix

GL-Matrix (SIMD)

• Mat4 operations are getting SIMDized using the JavaScript API
• Operations include:
  • Multiply, rotate, adjoint, inverse, translate, scale.. 
• They're the most heavily used
• SIMD vectors can be better utilized to achieve the higher parallelism

GL-Matrix SIMD Benchmark

Small benchmark measures each function performance

GL-Matrix SIMD Performance

• Overhead in interpreter mode
  • efficient machine code only after JIT compilation
• Less improvement for small functions

Numbers are extracted from 500k iterations on Nightly

Thanks

Questions?