R-Car DNN Compiler

DNN compiler for automatically applying program optimization for deep learning model for R-Car V4H

Program optimization is needed in order to achieve highly efficient execution of deep learning models trained with the deep learning framework on the R-Car V4H and to realize real-time inference processing. Specifically, this involves program conversion for high-speed computation using CNNIP, an accelerator for deep learning that is equipped in R-Car, and memory optimization to maximize the utilization of the high-speed, small-capacity SRAM installed in R-Car. Manually performing this kind of optimization is extremely difficult and requires a great deal of man-hours, since it requires a deep understanding of the target hardware.

This tool provides the ability to generate a fast executable program by taking a trained deep learning model as input and automatically applying the optimizations for R-Car V4H. This tool was developed by adding a backend for R-Car V4H to the OSS Apache TVM, so performance optimization for R-Car V4H can be applied in the same way as compiling for a CPU and GPU.