【作者】 解利军；

【导师】 郑耀；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2007， 博士

【摘要】 现代大规模科学计算的参数多、精度高、尺度大，普通的显示手段很难展现其计算的全景。使用先进的显示设备构建新型的可视科学计算环境来协助科学家对科学计算进行有效的驾驭、分析和理解，是当前计算机领域和计算科学领域的一个研究热点。在众多显示手段中，显示墙由于能达到超高分辨率和超大显示面积而备受青睐。传统的显示墙使用紧耦合的图形流水线驱动高端投影仪拼接实现，由于所用的软硬件都由专业公司整体生产，构建成本高，构建后软硬件的升级、变更、维护和扩展都比较困难，实用性较差，只有少量政府科研机构有能力构建和使用。随着计算机硬件技术的飞速发展，使用通用的低端硬件来构建显示墙逐渐成为可能。本文研究了完全使用通用硬件来构建低成本、可伸缩立体显示墙的方法，并基于此构建了一个面向多学科模拟的可视计算环境。这种立体显示墙使用微机集群驱动一组普通中低档投影仪构成。投影仪阵列中每两个投影仪形成一个立体投影对，投影到屏幕上的同一区域，实现被动立体。投影仪位置不需要物理上精确定位，通过计算机视觉的方法实现投影图像几何和色彩上的无缝拼接。由于完全使用标准硬件，构建的方法和支撑软件都有良好的可伸缩性，所以用户可以根据需要非常迅速地搭建多种分辨率和显示尺寸的显示墙。本文研究的重点包括：1、系统研究了立体显示墙的构建方法，包括其系统构架、硬件选择、几何校正、颜色校正、立体实现、软件平台等各个方面。其中有特点的工作包括：■提出了两次投影、在线消除投影仪非线性因素的几何校正算法。该方法简化了几何校正过程，并提高了校正精度。■提出了利用双通道显卡，将校正算法透明地应用到显示墙系统中的集成方法。该方法在后台运行，不影响前台程序的正常运行，并且利用了显卡硬件的计算能力，校正速度较高。2、研究了驱动立体显示墙的并行绘制方法。分析了各种并行绘制方法的特点，指出sort-first方法是最合适当前集群硬件条件的并行绘制方法。研究了提高sort-first并行绘制效率的方法，主要包括：■提出和实现了流缓冲和动态列表构建。由于帧间具有相似性，将指令缓冲可以降低带宽负载，进一步将缓冲的指令编译为动态显示列表则可以提高指令执行效率。■提出了基于Hilbert曲线的静态负载平衡方法。利用Hilbert曲线在空间上良好的解聚集特点，能在多数情况下提高负载平衡率，能在较低分辨率下提高绘制速度。■提出了基于等分网格快速分组的动态负载平衡方法。该方法是一种近似的离散对象调度方法，综合考虑了几何对象边界覆盖和像素传输耗时，能在几何对象数目较少时实现绘制加速。3、设计和实现了问题求解环境EEMAS的主要部分。EEMAS与立体显示墙一起组成了一个可视的科学计算环境，用于支撑多学科的大规模科学计算。主要工作包括：■使用黑板框架设计和实现了EEMAS的主体框架，各个功能模块以可插拔的方式动态组合，协同解决多学科问题。■集成了许多有用模块，特别是基于偏微分方程应用类的支撑工具。开发了几何建模前处理模块，并改造了ParaView可视化模块。■验证了EEMAS在立体显示墙上的使用。应用本文所述方法，已在浙江大学和某航天研究院部署了两套立体显示墙，并结合问题求解环境EEMAS，进行了实际的科学计算模拟。实践表明，通过协助科研人员快速整合资源和可视驾驭计算，该环境能比较高效地实现多学科、大规模计算模拟。更多还原

【Abstract】 Modern large-scale scientific computing features multi-parameters, high resolutions and multi-scale, and standard display methods are not able to display the panoramic scene of these computational results. Visual Computing Environment (VCE) based on new styled display devices can help scientists steer, analyze and understand the scientific computing, and thus is a topic receiving many recent attentions.Display wall is preferred among various display devices for its super high display resolution and large display size. The traditional display wall is built on tight-coupled graphic pipeline and high-end projectors. The hardware and supported software are all produced by professional companies, and thus very costly. Also its updating, alteration, maintenance and extending is very difficult. Therefore, the traditional display wall is only accessible to high-end government laboratories.With the dramatic advances of computing technology, it is becoming feasible to build display walls completely using commodity hardware. This thesis studies the construction of a scalable stereo display wall using only standard commodity hardware and its integration into a multi-disciplinary VCE. The stereo display wall is tiled by a PC cluster-driven projector array. Each pair of projectors projects to the same area of the screen to implement passive stereo. The projectors can be aligned casually, and the precise alignment and color match is achieved by software method assisted with a digital camera. Both the hardware construction method and enabling software has good scalability, and users can build a display wall of many sizes and resolution readily under certain limits.The main contributions of this thesis include:1. The construction of a stereo display wall has been systematically studied, including system architecture, selection of hardware, geometry calibration, color calibration, stereo implementation, software platform, etc. Two improved methods are proposed. A two-pass projection geometry calibration method with on-line nonlinear elements elimination. This method simplifies the calibration process and improves the precision. A transparent integrating method for geometry and color calibration of applications. This method runs a calibration daemon on graphic card without interrupting the front ground application at all.2. Three technologies are studied to improve the speed of parallel rendering which drives the stereo display wall: Stream cache and dynamic display list construction. Because of the similarity between continuous frames, caching the data stream in the servers can reduce the amount of the data that have to be transferred, and constructing display lists on these cached streams can speedup instructions execution. Static load balance method based on Hilbert curve. Hilbert curve can decluster space evenly. Utilizing this characteristic to distribute the load of rendering can improve the load balance in most situations. Mesh-based fast clustering dynamic load balance method. This method adds object group boundary recognition to the GRID method to reduce the geometry overlay. It can improve the rendering speed when the number of geometry objects is not too large.3. A problem solving environment EEMAS running with the display wall is designed and implemented to compose a complete visual scientific computing environment. EEMAS is designed for large-scale, multi-disciplinary scientific computation. The framework of EEMAS uses the blackboard architecture. All kinds of function modules can plug in/out dynamically and collaborate with each other to solve a certain problem. A lots of modules are integrated, especially the enabling tools for problems based on Partial Differential Equations (PDE). A new geometry handling pre-processing module and a modified version of visualization software Paraview are developed.With these methods and software, two visual computing environment prototypes have been developed in Zhejiang University and a national aerospace research institute. Practice shows that such an environment can improve efficiency of scientists and engineers by helping them visually steer the computing process in a semi-immersive way.更多还原