Acceleration of algebraic multigrid solver on APU

点击翻译按钮获取中文摘要

We studied efficient implementation technique of algebraic multigrid method on APU which has CPU cores and GPU cores in one chip. In order to minimize the communication between CPU cores and GPU cores, and to exploit GPU cores’ performance at most, we tested the mixed precision solver that iterates algebraic multigrid method with single precision on GPU cores, and calculate the residual vector with double precision on CPU cores. As a result, mixed precision method works well for algebraic multigrid method. For typical Poisson problems with convergence criterion of relative residual norm of 1.0E-7, precision correction needs only 3 times in our numerical test and there are little CPU-GPU communication. This method can be applied to Ivy Bridge and Haswell.