Research And Design Of Speedup Methods Based On Accelerative Algorithms And Sparse Convolutional Neural Networks

Convolutional neural networks(CNNs)have greatly improved the accuracy of image recognition.But CNNs are computationally intensive,and its computational cost is hard to accept.In order to speed up execution,people have proposed several solutions.The solution mainly starts from two aspects.One is accelerating the operation of convolution layers by using accelerative algorithms.However,each algorithm has its advantages and disadvantages,and there is no one algorithm that can handle all situations.Another solution is designing custom hardware accelerators to solve the complicated computation.However,the current hardware accelerators mostly use the traditional convolution algorithm.It is lack of support for sparse neural network,and lose the rise space from an algorithmic perspective.In order to solve these problems,we propose two solutions.In this paper,we examine the performance of various algorithms in GPU environment.By building a customized CNNs model,we have fully explored the impact of the data structure on the performance of algorithms,including inference/training speed,memory and power.In addition to the algorithms,we also focus on how their implementations in GPU environment affects their performance.We traced the kernel functions of these implementations to further generalize the characteristics of these algorithms.Finally,we summarize the applicable conditions for each algorithm,and design a strategy to assigns the appropriate implementation for different convolutional layers in CNNs.With using our strategy,we can make Alex Net 1.2x to 2.8x faster than other strategies in GPU environment.This work has very important meaning for understanding these algorithms and may provide insights for further optimizations of the architecture of GPUs and accelerators.we also redesigned a convolutional neural network accelerator based on the Winograd-Sparse algorithm,which proved to effectively reduce the computational complexity of convolutional neural networks and also be well adapted to sparse neural networks.Through the combination of hardware and the algorithm,we can achieve considerable computational efficiency while reducing hardware resources.Experiments show that our accelerator design improves the computation speed by nearly 4.15 times compared with the traditional algorithm.From the perspective of multiplier utilization,we have increased the utilization rate by up to 9 times compared with other existing designs.