MLIR first impression

Overview

• Multi-level IR compiler framework
• Provides IR for high-level programming languages
• Core language + dialects
  • You can define your own dialect!

Core concepts

• Operations
  • Operations are statically and fully typed.
  • e.g. module (builtin.module,) function (func.func)
• Regions
  • SSA CFG region
  • Graph region
• Blocks
  • A list of operations (instructions)
  • Or, a set of operations (nodes)
• Attributes
  • Given to operations as name-value pairs
  • Attributes are also typed.
• Types

Dialects

• Defines custom operations, types, and attributes.
• e.g. builtin, llvm, scf, gpu, async

Passes

• Transformation
  • e.g. canonicalization, SCE, DCE, function inlining, SCCP, etc.
• Conversion
  • Converts a dialect to another
  • e.g. scf to llvm
• Others
  • Dialect-specific passes

Smell of machine learning

• bfloat16 in the builtin dialect.
• Linear algebra-related dialects
  • e.g. linalg, tensor, sparse_tensor, quant
• Tensor shape inference
  • e.g. tensor<$1x$2xf32> * tensor<$2x$3xf32> = tensor<$1x$3xf32>

References

• MLIR documentation