Writing Fast and Maintainable Code With Halide —The Pilot Episode

Writing high-performance code often comes at a cost of reduced portability and simplicity —Halide fixed that and we use it often at Google.

Image by Marc-Olivier Jodoin on Unsplash.

Introduction

Halide is an open-source programming language designed to make it easier to write and maintain high-performance image processing or array processing code on modern machines. Halide currently targets

• Different CPUs
• Different Operating Systems
• Diff compute APIs like CUDA, OpenCL, OpenGL, etc.

Rather than being a standalone programming language, Halide is embedded in C++. This means you write C++ code that builds an in-memory representation of a Halide pipeline using Halide’s C++ API. You can then compile this representation to an object file, or JIT-compile it and run it in the same process. You can read more about it at halide-lang.org.

Halide is used very often in Google, for example in the famous HDR+ algorithm that powers the camera app on the flagship Pixel devices. Beyond that, Halide pipelines are run at scale within Google Photos, Youtube, and elsewhere [3].

HDR image (left) & Standard Image (right). Captured with Camera from Google on a low end Android device. Image captured by the Author.

I am a TL in Camera from Google at Google and we use Halide language extensively for implementing our features like HDR or Night Mode. Migrating to Halide from standard C++ or OpenCV has helped us significantly improve performance of some of these algorithms on the ARM chipsets that power most of Android devices.

This has been critical in landing these features on low end devices and making good photography accessible to all.

Important Disclaimer: Any opinion called out in this article are my own and don’t reflect opinion or stance of the organizations I work with.

Need: Why should I care about yet another programming language?

Scale — Let’s look at the scale first