A fast and robust mixed-precision solver for the solution of sparse symmetric linear systems

摘要: On many current and emerging computing architectures, single-precision calculations are at least twice as fast double-precision calculations. In addition, the use of single precision may reduce pressure on memory bandwidth. The penalty for using solution linear systems is a potential loss accuracy in computed solutions. For sparse systems, mixed which iterative methods preconditioned by factorization can enable recovery high-precision solutions more quickly less than direct solver run arithmetic.In this article, we consider within solvers symmetric exploiting both reduction requirements performance gains. We develop practical algorithm to apply mixed-precision approach suggest parameters techniques minimize number solves required process. These experiments provide basis our new code HSL_MA79—a fast, robust, that included mathematical software library HSL.Numerical results wide range problems from applications presented.