A Performance-driven Neural Network Compiler for Multi-core Computing-In-Memory Accelerator

Nowadays, artificial neural networks(ANN) are being widely used in daily life, such as target detection, speech recognition, etc. Computing-In-Memory(CIM) has attracted much attention because of its low power consumption and high speed. In this paper, we propose a performance-driven neural network compiler for multi-core CIM-based accelerator. We provide a cycle accurate runtime simulator to evaluate the performance for neural network models deployed on target CIM accelerator, and generate executable codes accordingly. In addition, we propose two optimization heuristics to further improve the resulting inference performance for compiled ANN models. Experimental results show that the inference performance can be 3.5 times faster compared with baseline results on average.