I’m sharing a project I created in academia. DLVM is a modern compiler infrastructure for deep learning systems, inspired obviously by LLVM and SIL. DLVM has a linear algebra IR, automatic differentiation, domain-specific optimizations and a CPU/GPU code generator. NNKit, a staged linear algebra DSL embedded in Swift (described in the paper), targets DLVM for optimizations and code generation.
The project is written entirely in Swift!
Here are some related publications:
1. DLVM: A modern compiler framework for neural network DSLs
- Will be presented at the ML Systems Workshop at NIPS 2017
2. DLVM: A modern compiler infrastructure for deep learning systems
- Pre-print: https://arxiv.org/abs/1711.03016