High Performance CNN Accelerators Based on Hardware and Algorithm Co-Optimization

Abstract

Convolutional neural networks (CNNs) have been widely used in image classification and recognition due to their effectiveness; however, CNNs use a large volume of weight data that is difficult to store in on-chip memory of embedded designs. Pruning can compress the CNN model at a small accuracy loss; however, a pruned CNN model operates slower when implemented on a parallel architecture. In this paper, a hardware-oriented CNN compression strategy is proposed; a deep neural network (DNN) model is divided into “no-pruning layers ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$NP$ </tex-math></inline-formula> -layers)” and “pruning layers ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$P$ </tex-math></inline-formula> -layers)”. A <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$NP$ </tex-math></inline-formula> -layer has a regular weights distribution for parallel computing and high performance. A <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$P$ </tex-math></inline-formula> -layer is irregular due to pruning, but it generates a high compression ratio. Uniform and incremental quantization schemes are used to achieve a tradeoff between compression ratio and processing efficiency at a small loss in accuracy. A distributed convolutional architecture with several parallel finite impulse response (FIR) filters is further proposed for the regular model in the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$NP$ </tex-math></inline-formula> -layers. A shift-accumulator based processing element with an activation-driven data flow (ADF) is proposed for the irregular sparse model in the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$P$ </tex-math></inline-formula> -layers. Based on the proposed compression strategy and hardware architecture, a hardware/algorithm co-optimization (HACO) approach is proposed for implementing a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$NP-P$ </tex-math></inline-formula> hybrid compressed CNN model on FPGAs. For a hardware accelerator on a single FPGA chip without the use of off-chip memory, a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$27.5\times $ </tex-math></inline-formula> compression ratio is achieved with 0.44% top-5 accuracy loss for VGG-16. The implementation of the compressed VGG-16 model on a Xilinx VCU118 evaluation board processes 83.0 frames per second (FPS) for image applications, this is <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1.8\times $ </tex-math></inline-formula> superior than the state-of-the-art design found in the technical literature.