Benchmarks of MATRIX Blockset for REX System

摘要

This paper presents performance benchmarks of several high-level LAPACK subroutines converted to C language in the author’s previous work. The tested algorithms cover singular value decomposition (DGESDD), eigenvalues (DGEES, DGEEV) and Sylvester equation (DTR-SYL modified for general matrices). Benchmarks compare: original LAPACK functions compiled by GFortran, C language versions compiled by GCC and by Visual Studio 2015 (VS2015) and C language versions embedded to REX function blocks compiled by VS2015 for Windows and by GCC for Raspbian. All tests are provided on the PC with Windows and on the Raspberry Pi 3 Model B+,The computation speed on Raspberry Pi is similar in all cases but on PC the speed for matrices of orders 250 and higher is up to 1.5 times faster compared to the speed of original code compiled by GFortran.