基于嵌入式移动设备上的3D应用低功耗研究

发布时间：2018-07-28 15:56

【摘要】：随着PC机上很多新颖的3D应用被移植到移动设备上,而移动设备是电池供电的,因此其能量消耗已经成为一个迫切需要解决的问题。目前存在两种体系架构：传统的CPU架构和比较流行的CPU-GPU异构架构。然而,已有的低功耗技术对于在这两种体系架构下的3D应用节能帮助却很少,因此需要进一步研究嵌入式移动设备上的3D应用低功耗。对于在传统的CPU架构下,为了使用低功耗技术DVFS,目前大多数的做法是通过当前3D应用图形帧的信息来预测未来帧的任务负载。然而,由于提取信息粒度较大且不完整而不能准确的预测任务负载。结合已有研究,如果提取信息粒度划分到函数级别,将会大大提高任务负载预测的精确度。而对于流行的CPU-GPU异构架构,由于低功耗模型不够完善,很少有软件层面的研究。当有大容量数据需要处理的情况下,PCI-E带宽会成为CPU和GPU计算的瓶颈。因此,为了更好的完善这一块,如果结合GPU的数据处理特性,对任务负载进行合理的分配到CPU和GPU上且并行运行,将会最大化的降低整个系统的功耗。针对所研究的问题,本文需要做的主要工作和完成的成果如下： 1.在CPU架构下,对具体的3D应用函数调用层次及主要耗时函数进行提取分析。 2.为了消除频繁调节CPU频率而带来的额外开销,提出了改进版本的贪心制。 3.对于CPU-GPU异构架构,为了合理分配任务负载到CPU和GPU上,提出了结合GPU数据处理特性的数据分块方案。 4.为了简化讨论,提出了功耗的最优化模型转化为时间的最优化模型并且有严谨的推理和分析过程。
[Abstract]:With many novel 3D applications of PC being transplanted to mobile devices, mobile devices are powered by batteries, so its energy consumption has become an urgent problem to be solved. At present, there are two architectures: the traditional CPU architecture and the more popular CPU-GPU heterogeneous architecture. However, the existing low power technologies are of little help to the energy saving of 3D applications under these two architectures. Therefore, it is necessary to further study the low power consumption of 3D applications on embedded mobile devices. In the traditional CPU architecture, in order to use low-power technology DVFS, most of the current methods are to predict the workload of future frames through the information of current 3D application graphics frames. However, due to the large granularity and incomplete information extraction, it can not accurately predict the task load. Combined with the previous research, if the granularity of information is divided into function level, the precision of workload prediction will be greatly improved. However, for the popular CPU-GPU heterogeneous architecture, due to the low power model is not perfect, there is little research on the software level. The PCI-E bandwidth becomes the bottleneck of CPU and GPU computing when there is a large amount of data to be processed. Therefore, in order to perfect this block better, if combining the data processing characteristic of GPU, the task load can be allocated to CPU and GPU reasonably and run in parallel, the power consumption of the whole system will be maximized. The main work and achievements of this paper are as follows: 1. In the CPU framework, the specific 3D application function call hierarchy and the main time-consuming functions are extracted and analyzed. 2. In order to eliminate the extra cost of adjusting CPU frequency frequently, an improved version of greedy system is proposed. For CPU-GPU heterogeneous architecture, in order to reasonably distribute the workload to CPU and GPU, a data block scheme combining the characteristics of GPU data processing is proposed. 4. In order to simplify the discussion, the optimization model of power consumption is transformed into the optimization model of time, and the reasoning and analysis process is rigorous.
【学位授予单位】：浙江工业大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41;TP368.1

【相似文献】