Solving grid equations using the alternating-triangular method on a graphics accelerator

The paper describes a parallel-pipeline implementation of solving grid equations using the modified alternating-triangular iterative method (MATM), obtained by numerically solving the equations of mathematical physics. The greatest computational costs at using this method are on the stages of solving a system of linear algebraic equations (SLAE) with lower triangular and upper non-triangular matrices. An algorithm for solving the SLAE with a lower triangular matrix on a graphics accelerator using NVIDIA CUDA technology is presented. To implement the parallel-pipeline method, a three-dimensional decomposition of the computational domain was used. It is divided into blocks along the $y$ coordinate, the number of which corresponds to the number of GPU streaming multiprocessors involved in the calculations. In turn, the blocks are divided into fragments according to two spatial coordinates — $x$ and $z$. The presented graph model describes the relationship between adjacent fragments of the computational grid and the pipeline calculation process. Based on the results of computational experiments, a regression model was obtained that describes the dependence of the time for calculation one MATM step on the GPU, the acceleration and efficiency for SLAE solution with a lower triangular matrix by the parallel-pipeline method on the GPU were calculated using the different number of streaming multiprocessors.