Overview of components

This document explains components of Chainer compiler by iterating the root directory of the project.

chainer_compiler/ch2o

CH2O is a Python library which translates Python code to an extended ONNX model.

chainer_compiler/elichika

Elichika is yet another Python-to-ONNX compiler being developed to replace CH2O.

common

C++ functions which are used by other components.

compiler

Compiler directory contains a bunch of things such as

• Library functions/classes to load/modify/store ONNX graph
• Auto-differentiation (gradient.cc and gradient_ops.cc)
• Constant propagation
• Naive code generators which uses NVRTC/TVM
• Generate code for ChainerX VM, a virtual machine based on ChainerX

but the most important file in this directory is gen_node.py, which maintains the list of supported extended ONNX operations.

runtime

The implementation of ChainerX VM, a Python/ONNX-free virtual machine based on ChainerX. Most operations are/should be simple wrappers of ChainerX's routines. However, there are some operations which complement ChainerX (e.g., NVRTC and cuDNN RNN).

Again, the operations supported by ChainerX VM are managed by chxvm_defs.py, which acts as an IDL of ChainerX VM ops.

chainer_compiler, chainer_compiler_cc

This provides a thin wrapper interface of the compiler and the runtime.

your_model = YourModel()
compiled_model = chainer_compiler.compile(your_model)

You need to specify -DCHAINER_COMPILER_ENABLE_PYTHON to use this functionality. See MNIST code for an example.

tools

run_onnx is a tool which compiles and runs an extended ONNX model. Run

$ ./build/tools/run_onnx --help

to see the list of flags.

scripts

Random scripts which is used for development.