Introduction

The Torq compiler is based on MLIR, a framework designed to support compilation for heterogeneous hardware. The compilation is implemented as a sequence of passes that lower an high-level representation to an IR expressed using dialects that can be customized to represent target-specific operations. MLIR itself is part of the LLVM ecosystem and, being completely generic, does not provide any specific support for features commonly needed in machine learning application. To avoid developing all these features from scratch, Torq compiler is built as a plugin on top of IREE, an MLIR-based end-to-end compiler and runtime specialized for lowering ML models to a variety of architectures including CPUs, GPUs and custom hardware.

This guide is intended for developers working on the Torq compiler and runtime. It covers the main workflows and extension points for adding new functionality, optimizing compute patterns, and integrating custom kernels. Each section provides step-by-step instructions, code examples, and references to relevant source files to help you get started quickly.

Topics covered include:

• How to add support for new operators and convert them to Torq kernels.

• Guidelines for developing efficient kernels using Torq EasyKernel API.

• Instructions for creating and registering new compiler passes.

• Methods for implementing and testing compute optimizations.