Understanding the General Concepts of Halide Programming Language

Write fast and maintainable code

Screenshot of the 3x3 box blur algorithm written in Halide. Source — halide-lang.org

Last time I presented how it’s possible to achieve high performance with simple and maintainable code with Halide. This time, I’ll share about the general concepts in Halide.

Introduction

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

In the last article of this series, I explained:

• What is Halide?
• What is the need for Halide? (Recap: fast code is often complex and hard to write and maintain)
• How Halide addresses the need.

In this article, I’ll dig deeper and share some general concepts in Halide — key data types and different schedule primitives. For different schedule primitives — an example of generated code and a demo of code execution is included.

This article is a summarized version of content available in the documentation and tutorials shared by the Halide development team.

I have tried to summarize some of the primary concepts for the readers and my own future reference. This seemed critical to share before talking in-depth about some of the topics.

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.

Halide concepts — the general structure

I’ll be covering this topic just at the surface, for learning more please checkout the tutorials at halide-lang.org/tutorials.

A Halide program typically have

• Some input(s) and one or more output(s).
• Some expressions or algorithms that that expresses the conversion of input(s) to output(s).
• Schedules that define how these expressions will be evaluated on a certain machine. We can also define multiple schedules each targeting a different subset of hardware.

In terms of these expressions or algorithms we have

• Func: A Func object…