Designed for AI-on-chips, the ONNC Compiler is a comprehensive bundle of C++ libraries and tools tailored to enhance compiler development for deep learning accelerators. It efficiently transforms neural networks into machine instructions suitable for diverse SoC architectures, from single core systems to more complex layouts with multi-level memory hierarchies. The compiler allows seamless connectivity to leading deep learning frameworks such as PyTorch and TensorFlow. It enables the scaling of deep learning tasks across heterogeneous multicore AI SoCs by utilizing both single backend and multiple backend modes to optimize computing resources. Additionally, it supports intricate features like multiple view address maps, ensuring effective memory allocation and data movement across fragmented memory spaces. Known for performance optimization, the ONNC Compiler employs hardware/software co-optimization techniques to reduce data movement overhead, thereby improving system throughput and efficiency.
