/tf_mlir_study

Overview

The repository would host questions and answers to:

  1. TensorFlow and its backend to MLIR.
  2. MLIR proper and how it handles StructuredOps.
  3. MIOpen and how to get it integrated with MLIR proper.

TensorFlow and MLIR : infrastructure

  • TF + MLIR build: how to enable it?
  • How does MLIR get pulled in TF build?
  • MLIR-related tools for TF dialect?
  • How to run TF dialect-related tests?
  • We focused on compiler/jit before, what's the importance of compiler/aot with MLIR?
  • How to enable compiler/xla/service/mlir_gpu backend in XLA?

TensorFlow and MLIR: operators / transoformations

  • How does HLO Dot gets lowered to LHLO Dot and further down to MLIR dialects such as Linalg or Affine?
  • How does HLO Conv gets lowered to LHLO Conv and further down to MLIR dialects such as Linalg or Affine?
  • How does decalarative optimization patterns of MLIR work in TensorFlow level?

MLIR

  • How does StructuredOps work?
  • Specifically, how to build an end-to-end example to depict some of the operations?
    • Fill
    • Copy
    • Matmul
    • Conv
  • How could these ops be hooked with HIP/MIOpen?
  • How does declarative patterns work?
    • Tiling
    • Fusion + tiling

MIOpen

  • For implicit GEMM kernels, what parameters are tunable, and what parameters are to be supplied by clients?
  • How to make implicit GEMM convolution kernels do regular GEMM operations?
  • For tunable parameters, how/where to model tuning process?