Backward Graph Construction and Lowering in DL Compiler for Model Training on AI Accelerators
Hyunjeong Kwon, Young-Su Kwon, Jinho Han · 2022 19th International SoC Design Conference (ISOCC) · 2022
A deep learning (DL) compiler is required to acceler ate model inference and training on AI accelerators. In this work, we propose a novel approach to constructing a backward graph from a PyTorch model, and lowering it to machine codes. The backward graph is constructed using information from PyTorch's autograd engine. The newly proposed lexer and parser convert the generated graph into an abstract syntax tree (AST). The AST is converted to GIR, an intermediate representation within the MLIR framework. IR lowering is then applied to the GIR, producing an LLVM IR for the LLVM backend. Among operators, those that can be accelerated using a DL accelerator are processed by the accelerator. This is achieved through the LLVM IR call function, which calls the accelerator's backend. In the experiment, the proposed compiler estimated the training loss with an average error of 1.46% within 6.7 seconds while processing 100 epochs.