Research And Implementation Of Simulation Of Diamond Film Growth Based On CUDA

Parallel computing is becoming a trend with the development of computer. Because the traditional CPU serial computation can not meet the demand of the development, especially in the field of scientific computing which needs a lot of calculations. In previous researches, high-configured minicomputer or computer cluster is needed to complete most of the calculations. However, the use of them will have a high cost both in time and money, and some researchers even do not have opportunity or condition to use such equipment. This makes many complex problems difficult to resolve in scientific research.The emergence of GPU (Graphic Process Unit) makes it possible to solve the above problems. GPU has a powerful ability in floating-point computing and parallel computing, which is more suitable for large amounts of data in terms of floating-point operations than CPU. Using GPU platform to simulate the growth of diamond film not only can improve the efficiency, but also can reduce the economic cost effectively. CUDA (Compute Unified Device Architecture) is a general purpose parallel computing architecture based on GPU, which allows GPU to solve complex computational problems. It exploits like-C language to develop program which is relatively easy to master without the need to relearn new language syntax. The program can run with extra high performance in processor supporting CUDA. In this thesis, we research on the simulation of diamond film growth based on CUDA platform.Firstly, in this thesis, the CUDA technology is introduced, including the hardware and software architecture of CUDA, program structure, threads structure, and memory model, etc. Then we introduce the principles of two methods which are Kinetic Monte Carlo (KMC) and Molecular Dynamics (MD) for simulation the growth of diamond film, and both of them are implemented on CPU and CUDA platform respectively. For the characteristics of KMC, we optimized the access of global memory through adding data. Running efficiency of the program can be improved by the use of constant memory, and the computation efficiency of GPU is also improved by designing rational dimensions of grid and block. In MD method, the calculation of intermolecular forces takes a great proportion in the total computation, and the original model can not make full use of the GPU feature. To solve this problem, we optimize the model by use shared memory. Due to the limited capacity of shared memory, iterative method is used to read grid data, and the superposed way for calculatation the intermolecular force. Finally, we compare the three implementations through experiments. The results show that CUDA in the best case can get75times higher acceleration performance.