| High performance computing provides powerful computation ability to solve many difficult scientific problems that human face, which will greatly promote economic development and improve living standards. At present, parallel computing is a major way to realize high performance computing. GPU is good choice, because of GPU’s graphics computing has natural inherent parallelism that there are massive parallel computing units within it, which makes GPU become a research focus in high performance computing for scholars engaged in research on parallel computing. General purpose GPU has been widely used in science and engineering technology since NVIDIA introduced CUDA in 2007.This dissertation focuses on the theme of video deinterlacing, image interpolation and two-dimensional surface simulation implemented on GPU. The three problems have the shortcoming of computing expensive, but there exists a high degree of parallelism in their calculations. Thus, we implemented the three problems above on a CPU-GPU heterogeneous platform. Finally, their execution efficiency was improved greatly through optimizing parallel scheme. The main jobs of this dissertation are following:1. The basic idea of the edge-directed adaptive intra-field deinterlacing method is to estimate the vacant pixel positions according to the correlation among image pixels. The proposed method can effectively improve video’s image quality and provide convenience for video scanning format conversion. The unknown pixels of the odd line or even line of deinterlaced image were estimated through prediction model constructed in the interlaced image. The proposed method has natural parallelism that each process of estimating unknown pixels is independent. Thus, the proposed was implemented on a CPU-GPU heterogeneous platform. We finally obtained a speedup of 94.6x through using shared memory, registers, asynchronous data transmission and multi-GPU to optimizing parallel program.2. The autoregressive modeling image interpolation scheme obtains reconstructed image and better vision effect than conventional methods. The proposed algorithm consists oftwo steps: the first step is to obtain the prediction model’s weights of sampled image through constructing prediction model by using known pixels of reconstructed version; the second step is to adjust the prediction model’s weights by importing a feedback mechanism that takes into account the mutual influence between the estimated missing pixels and known pixels. Due to the fact that each missing pixel was estimated independently, we implanted the proposed method on GPU in order to accelerate the process. We finally gained a speedup of 21.2x through carefully optimizations.3. Two-dimensional surface simulation is the foundation of the research of the electromagnetic scattering from the sea surface and the composite scattering from the electrically large targets. The two-dimensional surface simulation is based on linear wave theory and is implemented by linear superposition method in the dissertation. The smaller sampling interval, the truer surface is simulated, which results in the large computation required. However, each sampling point’s calculation is independently, which makes us take parallel strategy to implement more fine surface simulation. We implemented two-dimensional surface simulation on a CPU-GPU heterogeneous platform, which improves greatly the proposed scheme’s execution efficiency and finally got a speedup of 315.9x. |
