Method for vectorization computing of solution of large-scale trigonometric linear system of equations for GPDSP

Abstract

The invention discloses a method for vectorization computing of a solution of a large-scale trigonometric linear system of equations for a GPDSP. The method comprises the steps that the solution of the large-scale trigonometric linear system of equations is partitioned into irregular triangular matrix partition and vector multiplication and regularized matrix partition and vector multiplication in a partitioning computing mode; a CPU core in the GPDSP runs an operating system, is in charge of logical partitioning of data and dispatching of tasks and computes the irregular triangular matrix partition and vector multiplication with a small number of loads; DSP cores in the GPDSP are in charge of the regularized matrix partition and vector multiplication, obtain new computing tasks from the CPU core uninterruptedly and report running results to the CPU core, and the CPU core achieves dynamic load balance of the computing tasks among all the DSP cores by means of dynamic dispatching. By means of the method, the principle is simple, operation is convenient, the general purpose computation capacity of the CPU core in the GPDSP and the powerful vector processing array parallel computation and high-bandwidth vector data loading capacity of the DSP cores in the GPDSP can be fully utilized, and the computing memory access ratio of the DSP cores is increased remarkably.

Description

Technical field [0001] The present invention mainly relates to a General-Purpose Digital Signal Processor (GPDSP for short), and specifically refers to a method for solving vectorized calculations of large-scale triangular linear equations suitable for GPDSP. Background technique [0002] Triangular linear equations are widely used in high-performance computing and scientific engineering computing. For example, the solution of large-scale dense linear equations usually involves the LU decomposition or QR decomposition of the coefficient matrix first, and then the solution of the triangular linear equations, and finally the solution of the final linear equations is obtained by solving the triangular linear equations. High Performance Linpack (HPL) is the most important test benchmark for TOP500. HPL solves dense linear algebra equations to test and evaluate the floating-point performance of high-performance computer systems. After HPL completes matrix LU decomposition, A lower tr...

AI Technical Summary

Problems solved by technology

The advantage of the CPU core in GPDSP lies in control and management, and its computing power accounts for a very small proportion in GPDSP, while the vector processing array of the DSP core has powerful vectorized computing power and high-bandwidth vector data loading capability, which is the main component of GPDSP computing performance. Contributors, therefore, the above acceleration method for solving triangular linear equations using CPU and GPU is not suitable for GPDSP, and cannot make full use of the vector array storage and memory access mode of DSP core and the architectural features of vector processing array concurrent vector processing, and it is difficult to make full use of GPDSP The advantages of vector computing

Images