| As increasingly richness of computing resource,the subtasks are assigned to the corresponding computing resources according to the characteristics of the subtasks.Once the pattern of the heterogeneous systems is launched,it has been thought much by the scholars and the practical operators.From the networked and board-level heterogeneous system to heterogeneous system on chip,the dynamics fusing of heterogeneous system has been reinforced.In order to standardize the heterogeneous system and promote the commonality of heterogeneous platform,HSA Foundation Launches HSA 1.0 Specification with Multi-Vendor Architecture Support.Since 2015,AMD have released various CPU-GPU heterogeneous systems on chip which are supporting on HSA standard,this is known as APU,it provide a platform for further spreading heterogeneous computing.At the same time,we also found that the HSA standard and its practice platform research is still in the inception phase,but we also need to explore the characteristics of the benchmark,to select the required hardware resources,it will demonstrate the advantages of such heterogeneous system architecture.The characteristics of convolution neural network algorithm are prominent,among them,the matrix computation is large and the neuron independence is strong on the single layer network,these features are ideal for speeding up using GPU.In addition,the task scheduling between the layers requires the CPU to complete,so CPU-GPU heterogeneous system is very suitable for accelerating convolution neural network.However,in the system-level and board-level CPU-GPU heterogeneous platform,convolution neural network algorithm needs to frequently exchange data between the CPU and GPU,reducing the efficiency of the system.Based on the APU platform supporting HSA standard,this paper optimizes the convolution neural network algorithm,reduces the consumption of data transmission,and can use the HSA architecture to perform task scheduling flexibly.Combined with the technical features of HSA Specification,this article will revolve around the training process and task model of convolution neural network.In this paper,the main work and research results are as follows.1.Base on the idea of HSA heterogeneous programming model,this paper research the training process of convolution neural network,according to the characteristics of the independent neurons on single layer neural network,this paper turns convolution operation and pooling operation into matrix manipulation.And this paper research the key problems of network training process under the HSA and put forward the training optimization scheme which combines a real-time update weights and batch update weights.Experiments have shown that the optimized design scheme can better improve the execution efficiency of the algorithm.2.This paper implements convolution neural network using HSA heterogeneous programming model,in consideration of that the CPU resource utilization is low,so this paper defines the task model on the basis of increasing CPU resource utilization and classifies the convolution neural network for task model.This paper studies the key issues in the process of task assignment,and systematically summarizes and puts forward the design scheme of task assignment.Compared with the low resource utilization scheme,the design scheme can obtain a higher speedup ratio,which is an efficient and feasible optimization method. |
