军事知识服务平台
APR-BP:an adaptive memory partitioning and data reuse method for efficient back propagation
摘要 The access latency of off-chip memory is substantially higher than that of on-chip computation,especially during back propagation(BP).The limited on-chip buffers are strained by the frequent reuse and write-back of activations,weights,and gradients,which necessitates repeated off-chip da-ta transmission,and creates a severe performance bottleneck during training.To mitigate this,an adaptive memory partitioning and data reuse method for BP(APR-BP)is proposed.Considering the significant variation in access frequency and cross-layer reuse of activations,a priority-aware dynam-ic memory partitioning(PADP)strategy is introduced to characterize data reuse importance.By PADP,activations with higher reuse potential are preferentially allocated to contiguous on-chip mem-ory space,and effectively reduces redundant data movement and improves data reuse efficiency.Ai-ming at single fixed reuse pattern cannot simultaneously balance on-chip utilization and data access efficiency,a unified model is introduced to evaluate the benefit of three reuse patterns-input,weight and output.The optimal pattern for each stage and layer is selected through benefit-aware optimiza-tion.Experimental results demonstrate significant performance gains across three compute platforms,field-programmable gate array(FPGA),graphics processing unit(GPU),and embedded system-on-chip(SoC).Experimental results show that,compared to the baseline methods,the average data re-use rate is improved by approximately 43.0%,the total double data rate synchronous dynamic ran-dom access memory(DDR)transfer volume is reduced by about41.5%,and on-chip buffer utiliza-tion is increased by an average of 38.5%with the proposed APR-BP.Furthermore,the execution latency of BP for various networks is accelerated by an average of 1.47×.
年,卷(期): 2026, 32(3)
机标分类号: S TP333 TN918.91
在线出版日期: 2026年9月28日