hello_halide
Objectives
The objective of this project is to demonstrate a realistic production-ready implementation of an image processing algorithm using Halide. This includes:
- Use state-of-the-art auto-scheduler for optimal performance
- Create fully compiled code with no runtime dependencies
- Check Halide implementation correctness against baseline C++ implementation
- Compare performance between Halide and baseline C++ implementation
Setup
I used a Mac. The code should be easy to port to any Linux by adjusting the cmake file.
- Install XCode command-line tools (for clang/llvm)
- Install Halide and cmake:
brew install halide cmake
Build and run
mkdir build && cd build
cmake ..
make
./contrastAdjustTest
Example output
./contrastAdjustTest
Output pixel value range: [0:255], time: 0.008957 s
Halide Output pixel value range: [0:255], time: 0.000332 s
In this case, we achieved 27x speedup over optimized C++ code
Algorithm
The algorithm is a basic contrast adjustment of a grayscale image. For a given image, the pixel values are linearly scaled to maximize the dynamic range of [0-255] pixel values.
Code organization
Halide code is split into two generated functions: one to calculate the minimum and maximum of the pixel values, and another to apply the scaling.
Build process
We first compile the Halide code into a generator. Then we run the generator to create a static library for a specific platform target. Finally, we compile a test app that links in the Halide generated binary. This test app we can run to benchmark performance and check correctness.
Checking runtime dependencies
otool -L contrastAdjustTest