数值计算软件包自适应性能优化若干关键技术及评价标准研究

为了能够充分使用计算机资源，使软件运行能够尽可能地接近计算机峰值性能，研究人员 一直在努力。一个思路是为计算机开发优秀的编译器，并使用编译器相关技术对软件进行 性能优化；作为补充，另一个思路是开发能够共用的核心软件包，通过提高核心软件包中 程序的性能提高调用核心库的软件系能。 但科学计算中针对特定的计算机平台和特定的用户问题进行性能调优仍然是一个困难的问 题。速度和可移植性是数值计算软件开发中的一对矛盾。自适应性能优化技术正是为了 解 决数值计算软件中的可移植性问题和进行自动性能优化而提出的，科学家们希望数值计 算 软件能够动态地获悉计算环境的变化以及待求解问题的特征，根据需要改变自身以适应 这些变化以及多种复杂的问题情况，并且在多种求解方法中进行决策以选择最优的解决方 法来求解问题。 本文调研了使用自适应性能优化技术的几个著名软件包：~ATLAS，~SPARSITY，~OSKI以及数字 信号处理领域的快速傅立叶变换软件包FFTW，在此基础上，重点分析和比较了自适应性能 优化的关键技术，着重介绍了矩阵乘计算、MPI通信操作、快速傅立叶变换中涉及的经验搜 索、算法选择和自动代码生成等技术在这些软件包中的应用。 在深入调研的基础上，结合实际应用需求，本文还提出了自适应性能优化过程的新的评价 指标，试图权衡优化效果和优化时间开销，并在不同的实验平台上针对ATLAS的优化过程进 行的实验和过程评价，实验表明，综合优化效果和优化开销能够有效地发现ATLAS自适应优 化过程的特征。将其应用到实际开发和调优过程中，能够在不损失性能的前提下，节省优化时间。 论文最后对HPCC测试软件包在IBM刀片机群上进行了对比测试，发现了测试平台存在的性能 瓶颈，并消除了该瓶颈。表明HPCC软件包确实可以有效的发现被测平台存在的性能瓶颈问 题。

心理旋转中的整体优先效应及其机制

In the history of psychology research, more attention had been focused on the relation between local processing and global processing. For the global information and the local information, which is processed earlier? And which is processed faster? Precedence of the global over the local level in visual perception has been well established by Navon with compound stimuli, and Navon’s original study gave rise to many publications, including replications, generalization to other kinds of stimuli (nonverbal material, digits), populations (infants, children, brain-damaged subjects), and tasks (lateral visual hemifield presentation, copy drawing, memory recognition, and recall), and triggered some debate about the conditions in which global precedence is and is not observed (number, size, sparsity, and goodness of the stimuli, exposure duration, etc.). However, whether there is a global advantage or precedence in other cognitive processes was less tested. Most researches had suggested that there was a functional equivalency between visual perception and visual image processing. However, it’s still unknown whether there will be a global advantage on mental rotation. In the present study, we combined the mental rotation task with the compound stimuli to explore whether the global or local advantage also existed at the mental imagery transformation stages. In two pilot studies, the perceptual global precedence was found to be present in a normal/mirror-image judgment task when the stimuli exposure time was short; while the stimuli exposure time was prolonged (stimuli kept available till subjects’ response) the perceptual global precedence was showed to be eliminated. In all of the subsequent experiments, stimili would be presented till subjects’ response. Then mental rotation was added in normal/mirror-image judgment (some of the stimuli were rotated to certain angles from upright) in normal experiments, experiment 1 and 2 observed a global advantage on mental rotation both with a focused-attention design (Experiment 1) and divided-attention design (Experiment 2). Subjects’ reaction times were increased with rotation angles, and the accuracy was decreased with rotation angles, suggesting that subject need a mental rotation to make a normal/mirror judgment. The most important results were that subjects’ response to global rotation was faster than that to local rotation. The analysis of slope of rotation further indicated that, to some extend, the speed of global rotation was faster than that of local rotation. These results suggest a global advantage on mental rotation. Experiment 3 took advantage of the high temporal resolution of event-related potentials to explore the temporal pattern of global advantage on mental rotation. Event-related potential results indicated the parietal P300 amplitude was inversely related to the character orientation, and the local rotation task delayed the onset of the mental-rotation-related negativity at parietal electrodes. None clear effect was found for occipital N150. All these results suggested that the global rotation was not only processed faster than local rotation, but also occurred earlier than local rotation. Experiments 4 and 5 took the effect size of global advantage as the main dependent variable, and visual angle and exposure duration of the stimuli as independent variables, to examine the relationship between perceptual global precedence and global advantage on mental rotation. Results indicated that visual angle and exposure duration did not influence the effect size of global advantage on mental rotation. The global advantage on mental rotation and the perceptual global advantage seemed to be independent but their effects could be accumulated at some condition. These findings not only contribute to revealing a new processing property of mental rotation, but also deepen our understanding of the problem of global/local processing and shed light on the debate on locus of global precedence.