High level architecture

Front End

This part of the design produces LLVM 8.0 IR (Internal Representation) without regard to accelerator specific optimization, which are handled in the back-end support for each device individually.

ONNX support

While, ONNX has two official ONNX variants;

1. The neural-network-only ONNX and
2. it's classical Machine Learning extension, ONNX-ML.

DNNC supports neural-network-only ONNX with support for tensors as input and output types (no support for sequences and maps)

Rererence

1. ONNX support
2. LLVM docs