Study On The Multilayer Periodic Structure Of Grating Diffraction Simulation Algorithm GPU Acceleration

With the increasing integration of the semiconductor, micro fabrication has reachedthe Nano scale. If there is a tiny flaw in the structure will cause great changes in theperformance. Using optical methods to measure the critical dimensions of the device arewidely used, while the simulation algorithms of these structures are usuallycomputation-intensive and time-consuming. The huge parallel computing power ofGPU has achieved good results in many areas. The article research purpose is takeadvantage of the GPU’s parallel computing power to accelerate the calculation of thesimulation algorithms.Most of these algorithms are starting from the Maxwell equations, using differentmethods to derive conclusions eventually. In the derivation process, a large number ofequations exist, thus resulting in the process of simulation, involving a large number ofmatrix operations, such as eigenvalue problem, LU decomposition, inverse matrix andmatrix multiplication. All these computational processes are computation-intensive andtime-consuming. In order to reduce the computing time and speed up the calculationprocess, the article starting from the hardware and software aspects to solve theseproblems. On the hardware side, the parallel computing power of GPU was used; on thesoftware side, the CUDA technology was used to implement parallel computing onGPU.At first of the thesis, the algorithm RCWA and SAM are introduced fromMaxwell’s equations. Then the development of GPU used in general-purpose computingand the CUDA models and acceleration principles on them are carefully described.Several common matrix operations used in them are introduced and several its CUDA Ccode was completed. The efficiency of each function to calculate the matrix of differentsizes was tested, and the GPU computation time and the CPU computing time wascompared. At the same time, the strengths and weaknesses of the GPU acceleratedcomputation in this application were analyzed. Finally, according to the characteristicsof the algorithms, a preferred aspect was selected. A new matrix library was compiledby several own functions and some matrix computation libraries. The GPU interface code and the CPU interface code are completed, and the new matrix library was used inthe RCWA algorithm and SAM algorithm. The both algorithm acceleration effect in thecase of one-dimensional and two dimensional cases was tested. The results show thatthe GPU acceleration of the RCWA algorithm in the case of two-dimensional has someeffect; the GPU acceleration of the SAM algorithm in the case of two-dimensional has agood effect. Finally, the CUDA code of the SAM algorithm is optimized, theacceleration effect is enhanced.