The Research Of Deconvolutional Layer Acceleration Based On Classic CNN Accelerator

Fully Convolutional Neural Networks(FCN)are prevalent in computer vision applications,such as object detection,semantic/image segmentation,and the Generative Adversarial Networks(GANs)that can learn to automatically create labeled datasets in massive application domains such as speech,image,video and texts.In an FCN,traditional convolutional layers and deconvolutional layers contribute to the majority of the computation complexity.However,prior deep learning accelerator designs mostly focus on CNN optimization and the acceleration of deconvolutional layer gains much less attention form researchers.In current research,one common approach is to design independent compute-resources to handle deconvolutional layer,which leads to considerable hardware resource overhead.In this thesis,the accelerator design of the deconvolutional layer is intensely explored.The deconvolution layer on hardware accelerator is deeply explored,and the deconvolution acceleration is maximized by using the widely used convolution accelerator architecture.The acceleration of the full convolutional neural network is realized based on the convolution accelerator.We re-optimized the conventional CNN accelerator architecture of one-dimensional(dot production)processing unit array and regular two-dimensional processing unit array.For the one-dimensional array,a computational model is proposed to map the deconvolutional layer to the accelerator,which can achieve 1.6X ~ 3.9X speedup and reduce energy consumption by 41.7% to 72.3% in a set of representative applications.For two-dimensional array,by exploiting the locality in deconvolutional filters,this architecture reduces the consumption of onchip memory communication from 24.79 GB to 6.56 GB and improves the power efficiency significantly.Compared to prior baseline deconvolution acceleration scheme,the proposed accelerator achieves 1.3X ~ 2.7X speedup and reduces the energy consumption by 14.6% ~ 63.5% on a set of representation benchmark applications.This dissertation has further proposed a novel deconvolutional layer implementation with a software approach,which reorganizes the computation of deconvolutional layer and treats deconvolutional operations as the standard convolutional layer by splitting the original deconvolutional filters and transferring them into multiple small filters.This approach of accelerating the deconvolutional operation by using the existing CNN accelerator without adding any hardware modification.The proposed data flow is implemented on both one-dimensional array and two-dimensional array,achieving 2.4X ~ 4.3X performance acceleration and reducing energy consumption by 27.7% to 54.5% on a set of benchmarks.