TensorIR

TensorIR is one of the core abstractions in the Apache TVM stack, used to represent and optimize primitive tensor functions.

The TensorIR codebase consists of two modules (split from the former tir):

• tirx — Core IR definitions and lowering (PrimFunc, Buffer, SBlock, expressions, statements, lowering passes).

• s_tir (Schedulable TIR) — Schedule primitives, MetaSchedule, DLight, and tensor intrinsics.

In TVMScript, both modules are accessed via from tvm.script import tirx as T.

• Tensor Program Abstraction
  • Key Elements of Tensor Programs
  • Extra Structure in TensorIR
• Understand TensorIR Abstraction
  • Function Parameters and Buffers
  • Loop Iterations
  • Computational Block
  • Block Axis Properties
  • Why Extra Information in Block
  • Sugars for Block Axes Binding
• TensorIR Creation
  • Create TensorIR using TVMScript
  • Create TensorIR using Tensor Expression
• Transformation
  • Initialization Schedule
  • Loop Tiling
  • Leverage Localities
  • Rewrite Reduction
  • Trace the Transformation
• DLight: Rule-Based GPU Scheduling
  • Prepare a Model
  • Basic Usage: ApplyDefaultSchedule
  • Rule Catalog
  • Diagnosing Schedule Quality
  • DLight vs MetaSchedule
  • Writing a Custom Rule
  • Summary
• MetaSchedule: Search-Based Auto-Tuning
  • Architecture Overview
  • Prepare a Model
  • User-Facing Entry Points
  • Inspecting Tunable Tasks
  • Selective Operator Tuning
  • Database
  • Cross-Model Database Reuse
  • Key Parameters Reference
  • Summary