A Compiler for Deep Neural Network Accelerators to Generate Optimized Code for a Wide Range of Data Parameters from a Hand-crafted Computation Kernel

This paper presents the design and implementation of a compiler for a deep neural network accelerator that provides high performance and energy efficiency. The compiler allows deep learning frameworks, such as TensorFlow, to exploit the accelerator hardware by automatically creating data transfer code and outer loops around highly-tuned hand-crafted inner-loops for a wide range of neural network parameters. In other words, our compiler significantly reduces the development effort for deep learning libraries without sacrificing their performance. We have evaluated our prototype compiler to show that it can generate code for five most-critical deep learning operators with a comparative performance obtained from hand-tuned code.