Using Accelerate and simd

WWDC 2018

Composition of Accelerate Framework

• vDSP — Signal processing
• vImage — Image processing
• vForce — Vector transcendental functions
• BLAS, LAPACK, LinearAlgebra — Dense matrix computations
• Sparse BLAS, Sparse Solvers — Sparse matrix computations
• BNNS-Neural networks

Closely Related

• simd — Small vector and matrix computation for CPU
• Compression — Lossless data compression

Signal Processing Library

• Basic operations on arrays
  - Add, subtract, multiply, conversion, etc.
• Discrete Fourier/Cosine Transform
  - 1D DFT/ DCT/ FFT
  - 2D FFT
• Convolution and correlation

Demo

Signal Extraction from Noise Link

Essential Computations

• Analyze noisy signal with forward DCT
• Remove frequency components of amplitude less than threshold
• Reconstruct clean signal with inverse DCT

Demo

Halftone Descreening with 2D Fast Fourier Transform

Essential Computations

• Transform halftone image with 2D FFT
• Remove frequency components of the halftone screen
• Reconstruct continuous tone image

Frequency Removal

• zvmags
• vthrsc
• vclip
• zrvmul

The simd Module

• Simplified vector programming
• Small (fixed-size) vector and matrices
• Abstract architecture-specific types and intrinsics

simd

• Matrices of size 2, 3, or 4 and vector of 2, 3, 4, 8, 16, 32, or 64
• Arithmetic operators (+,-, *, /) work with both vectors and scalras
• Supports common vector math and geometry operations (dot, length, clamp)
• Quaternions

Rotating a Cube by Transforming Its Vertices Link

vImages

Image Processing Library

• Conversion
• Geometry
• Convolution
• Transform
• Morphology

Adjust Color Saturation

• The formula to adjust color saturation is
  Cb = ((Cb -128) * saturation) + 128
  Cr = ((Cr -128) * saturation) + 128

Workflow

More video effects

• Rotation

• Blur

• Dither

• Color quantization

LINPACK Benchmark

• How fast can you solve a system of equations?
• Actually 3 separate benchmarks
  - 100 x 100 system
  - 1000 x 1000 system
  - No hold barred