Research On High-Performance Computing Supporting Technologies For Large-Scale3D Terrain Construction

This thesis focuses on high-performance computing supporting technologies for large-scale3D terrain construction. The main research contents are divided into four parts, specific contentsand contributions include:The first part: design and implement a GPU accelerating large-scale3D terrain visualizationframework named MEGA, which is used for rapidly construction of3D terrain model off-line.MEGA has efficiency data organization as well as excellent render quality of the rendering. InMEGA:1. Proposes a terrain multi-resolution model FMRM, supporting efficiency large-scale datastorage and scheduling. The two level storage structure of FMRM not only achieve unifypartition and storage of full resolution terrain data, but also create the index to terrain datastorage model and construct corresponding multi-resolution logic structure, which based onexpression of the surface characteristics. FMRM solves the problems including data storageredundancy and scheduling inefficiency existing in the traditional terrain Pyramid model.2. Presents SFNE, a nested geometric error metric algorithm which fuses terrain surfacefeature quantization. Introduces terrain surface roughness to quantify the terrain geometricfeatures and built the node error dependency between different terrain resolution levels, controlof terrain model hierarchical multi-resolution based on the premise of geometric errors monotonetransfer.Terrain model simplification process is constraint by calculating of fidelity to avoiddistortion. Experiment results show that SFNE achieves relative ideal balance between terrainsimplification speed and rendering quality.3. Proposes a visual smoothing method based on vertex transition weight. On the basis ofGeomorphing, overall considering smooth transition in different terrain resolution levels as wellas resolution rank of adjacent terrain block boundary, real-time assign the time and spacetransition weights to model vertex in order to ensure smooth transition between differentresolution levels and different terrain blocks during terrain rendering.4. Designs and implements SRCaster, a fast terrain rendering algorithm based on GPURayCasting. SRCaster performs light projection to terrain nodesin screen selected by nodeevaluation function, and then a discrete traversal is called to these nodes. The traversaltermination condition is the obtaining of the point, which determined by light and terrain DEMintersection. Compared with other ray-casting terrain rendering methods, SRCaster simplifiedintersection test in all terrain blocks, and also convenient for graphics hardware acceleration byusing the correlation between adjacent pixels. Experiments show that SRCaster has significantperformance advantage in real-time rendering. The second part: study of high performance coding method to large-scale terrain data basedon CPU/GPU heterogeneous synergistic acceleration, so that to meet the demand of real-timeconstruction of large-scale3D terrain.1. Proposes a terrain multi-resolution model construction method based on lifting waveletmechanism. Create the terrain data wavelet transformation model to map the mesh refinementand simplification. Establishes multi-resolution quad-tree to DEM and Orthophoto texturerespectively by the lifting wavelet transform, and stores the quantified wavelet coefficientsinquad-treenode, to ensure the correctness of inverse wavelet transform.2. Designs and implements terrain data parallel hybrid entropy coding based on CUDA.Implements parallel run-length encoder to compress the sparse quantization wavelet coefficients,and use the parallel Hoffman encoder to further compress the run-length encoded data. Hybridentropy coding terrain data is organized into a plurality of data flow submitting to the decodingand rendering process. Experiment results show that this method has obvious advantages interms of the compression ratio, PSNR as well as coding and decoding throughput.3. Implements SPM V algorithm and its’optimization strategy based on CUDA, solve theproblem of terrain visualization performance decrease, which is caused by massive memoryintensive computation of decoder output terrain data stream. From the point of data distributioninvestigates the optimization strategy of CUDA-based SPM V algorithm for load balancing,including thread partitioning optimization, data reuse optimization, data access optimization, etc.Experiments show that the proposed algorithm and optimization strategy can significantlyimprove the terrain rendering and real-time interactive processing performance.The third part: In order to enhance the visual fidelity of terrain visualization scene, thisthesis investigates the realistic natural landscape simulation. Take the three-dimensional dynamiccloud as an example, study the modeling method of ground natural landscape based onmulti-core CPU acceleration.1. Designs and implements a3D dynamic cloud simulation method based on cellularautomata, achieves realistic cloud motion and cloud evolution, growth, dissipation simulation,constructs the multiple forward scattering lights model to enhance the visual reality of cloudmodel. The multi-core and multi-thread acceleration and optimization technology for cloudsimulations adopted, cloud modeling computing (CPU) and rendering (GPU) is separated, inorder to avoid the overall performance decreasein classical GPU based cloud modeling andrendering method.2. Presents a multi-thread accelerating cellular automata multiple neighbor unit status queryalgorithm M3NSQ, solves the problem of cloud simulation congestion, which caused by continueconcurrent queries of cell and neighbors’states in cloud modeling space. The algorithm is dividedinto two stages including query preprocessing and query execution, particularly design key data structure as the query table, grid index, query cache and nearest neighbor object cache, etc.Introduces query update ordering and temporal locality optimization method, adopts reasonabledata partition and thread scheduling to achieve a high performance parallel query processing.The fourth part: Designs and implements the VWBuilder framework to support large-scale3D terrain construction and application. VWBuilder integrates the research content of this thesis,and provide developer interface, which can provide pre-built off-line large-scale3D terrainmodel for third-party applications, and also be used as terrain construction module integrated inother applications to provide real-time online large-scale3D terrain building functions. Severalapplication examples based on VWBuilder is given, verify the effectiveness of the researchresults of this thesis.