onnx_optim

The onnx_optim library provides lightweight, non-owning data structures used by ONNX graph optimisation passes. The central types are OptimDim, OptimShape and OptimTensor, which together describe a tensor whose shape may be fully known, fully symbolic, or any mix in between.

Unlike a fully featured tensor type, OptimTensor never allocates memory: it stores a raw void * pointer to a buffer owned by the caller, the element TensorType and an OptimShape. This makes it suitable for use during shape-inference and rewrite passes where the underlying constant data is already materialised elsewhere (initialisers, attribute payloads, etc.) and only needs to be referenced.

• optim_tensor.h
• optim_sequence.h
• expressions.h
• shapes