Vectorized Evaluation of Boundary Integral Operators

It has become more or less standard in scientific codes to utilize shared- and distributed-memory parallelism achieved by OpenMP and MPI and thus to use the computational power of all available cores. However, in recent years the theoretical peak performance of CPUs has also been rising due to the capabilities of vector processing units able to perform simultaneous computations on vectors of data. This concept, known as SIMD, becomes increasingly important, yet it is still quite neglected. This paper deals with the intra-node optimized code utilizing the SIMD registers solving the boundary integral equations, and with numerical experiments on two modern Intel's processors, namely Xeon Phi 7250 and Xeon 8160.