2023 EuroLLVM - Tutorial: Controllable Transformations in MLIR

Introduction

This tutorial covers the use of the MLIR transform dialect, which allows developers to specify and control compiler transformations directly within the MLIR framework. By the end of this guide, you'll understand how to leverage the transform dialect for optimizing MLIR operations and how to adapt or extend its functionalities to suit your projects.

Step 1: Understanding MLIR Transform Dialect

• Familiarize yourself with the concept of the transform dialect, which allows the specification of compiler transformations.
• Recognize that transformations can be embedded within the original Intermediate Representation (IR) or supplied separately, much like pragmas or scheduling languages.
• Review the tutorial slides for visual aids and additional context: Tutorial Slides.

Step 2: Composing Transform Dialect Operations

• Start by identifying the available transform dialect operations in upstream MLIR.
• Create a sequence of operations to optimize a linear algebra operation. This involves:
  1. Selecting the appropriate transform operations.
  2. Composing them to form a transformation pipeline.
  3. Testing the sequence to ensure it produces efficient code.

Step 3: Defining New Transform Dialect Operations

• Learn how to define new operations within the transform dialect.
• Follow these steps:
  1. Determine the need for a new operation based on your specific transformation requirements.
  2. Implement the operation by writing the corresponding code in MLIR.
  3. Integrate the new operation into existing transformation logic.

Step 4: Adapting Existing Transformation Code

• Review existing transformation code that you may want to adapt to work with the transform dialect.
• Steps to adapt:
  1. Identify the transformation logic that needs modification.
  2. Refactor the code to use transform dialect operations instead of traditional methods.
  3. Test the adapted code to ensure it works correctly within the new framework.

Step 5: Setting Up Transform Dialect Infrastructure

• To use the transform dialect in a downstream out-of-tree project:
  1. Set up your MLIR environment to support custom dialects and transformations.
  2. Follow the guidelines for integrating the transform dialect into your project structure.
  3. Implement custom transformations and passes that leverage the transform dialect.

Conclusion

By mastering the MLIR transform dialect through this tutorial, you can effectively apply and extend compiler transformations in your MLIR projects. Start experimenting with composing operations, defining new ones, and adapting existing code to enhance your development process. As you become more familiar with these techniques, consider exploring more complex transformations and sharing your experiences with the MLIR community.