GGlow

Graph lowering (Glow) experimental implementation using MLIR

Requirements

• Install the latest version of bazelisk: details
• Install the python3.10 (preferebly using conda): help
• Have Clang compiler tools installed

Dependencies Added

Need to resolve them to be done automatically from build.

• "/usr/local/lib/libmlir_c_runner_utils.so"
• "/usr/local/lib/libmlir_runner_utils.so"

Building Steps

• Clone the repo and go inside

git clone https://github.com/gokulkrishna98/GGlow.git
cd GGlow

• Clone the submodules

git submodule init
git submodule update

• Run the bazel build command
  • For building GlowDialect

  bazel build lib/Dialect/GGlow:GGlowDialect
  

  • For building gglow

  bazel build src:gglow
  

  • For building all of mlir binaries from llvm-project

  bazel build @llvm-project//mlir/...:all
  

• You can find the exectuables from bazil-bin path

Current Goal

• Read the pytorch model: get graph and weights.
• Develop GGlow dialect to represent this model.
• Do simple lowering executing on my Intel cpu.

Format Stype

Using WebKit stype of format using clang-format tool

clang-format --style=WebKit -i gglow.cpp

Current TODO

• Lower Conv2d to LLVM dialect
• Generate resblock high level operation
• Convert pytorch model to IR in GGLow Dialect
• Lower to LLVM

References and Learning topics

• Torch-MLIR: https://github.com/llvm/torch-mlir/blob/main/docs/architecture.md