2. Introduction
• Image processing is a topic
where optimization matters.
• But optimization for multiple
platforms is hard and
expensive.
• We want tools for obtaining
high-performance code
regardless of the platform.
• Halide is a tool that aims to
solve this problem.
3. Data-parallel and IP languages
• Data-parallel languages:
• CUDA and OpenCL propose a SIMD
programming model for multi-core CPUs
and GPUs.
• Implementations can be very efficient at
the cost of losing portability.
• IP languages:
• MATLAB and other suites release kernel
languages.
• But individual kernels are not enough.
6. The Halide language
• Halide is a functional programming language (“à
la Haskell”) for IP.
• It makes a distinction between the algorithm and
the schedule:
• The algorithm is what should be done.
• The schedule is how it should be done.
• Optimization is achieved by:
• Using LLVM for generating simple code.
• Using architecture-specific compilers for
generating vectorized and parallel code.
7. Evaluation of Halide
• Halide’s execution time is measured against:
• ImageMagick.
• MATLAB.
• Mathematica.
• OpenCV 2.
• Two test images:
• A normal sized image (512x512).
• A big sized image (6400x4800).
• For two methods:
• RGB to grayscale.
• Gaussian blur.
11. Gaussian blur results
Normal sized image Time (ms) Time / Halide Time (%)
Halide 2.674 1.000
ImageMagick 304.000 113.687
MATLAB 2.834 1.059
Mathematica 117.003 43.755
OpenCV 2 1.076 0.402
Big sized image Time (ms) Time / Halide Time (%)
Halide 219.274 1.000
ImageMagick 12265 55.935
MATLAB 277.388 1.265
Mathematica 199203 908.466
OpenCV 2 191.875 0.875
12. Conclusions
• Halide beats each other tool
(except OpenCV 2).
• There’s a lot of room for
improvement.
• Programming with Halide is
hard-to-learn, easy-to-
master.