Research On Boundary Face Method Base On GPU-acceleration

Graphics processing unit (GPU) is a low-cost, low-power (watts per flop) and veryhigh performance alternative to conventional microprocessors. In addition to theirapplications in graphics processing, researches are more and more attracted by theirpotential application in numerical computing due to its powerful ability of parallelcomputing and floating-point calculation, and rapidly improved programmability. Theintroduction of CUDA (Compute Unified Device Architecture) provides a newdevelopment platform for high-performance parallel computing. Recently, CUDA hasbeen widely used in the field of finance, oil exploitation, astronomy, fluid dynamics,signal processing, image processing, etc. A large number of achievements have beenachieved.Boundary face method (BFM) not only inherited the advantages of boundaryelement method (BEM), such as dimensionality reduction of solved problems,high-accuracy stress calculation, it also has its own characteristics: the geometric dataat gaussian integration points, such as the physical coordinates, the Jacobians and thenormals, are calculated directly from the faces rather than from element interpolationapproximation, and thus avoiding geometric errors. However, the high computationalcost of the BFM prohibits its application to solving large-scale problems. Therefore,the research of boundary face method with GPU-acceleration based on CUDAplatform is of great significance.Based on an in-depth analysis of CUDA programming model and system structure,this thesis studies the regular integration and singular integration of boundary facemethod, and gives the parallel program and algorithm process of regular integrationand singular integration. An optimization program of parallel program is put forwardbased on the characteristics and optimization strategy of CUDA platform and thehardware performance of GPU. Comparative computations are made on both NVIDIAGTX680GPU and Intel(R) Core(TM) i7-3700K CPU. The result of numericalexamples shows that in the premise of high-accuracy calculation and same level ofaccuracy compared to CPU serial program, the speedup of regular integration parallelprogram is up to8.2.