Compiling techniques for first-order linear recurrences on a vectorcomputer

摘要

The authors discuss high-performance methods for solvingfirst-order linear recurrences on a vector computer, investigateautomatic transformations, and develop the compiling techniques forfirst-order linear recurrence problems. The results show that theimproved vector code generated by the vectorizing compiler run at therate of 150 MFLOPS (million floating operations per second), in the caseof the moderate loop lengths (>1000), and at the rate of over 200MFLOPS, in the case of long loop lengths (>10000), on the HITAC S-820supercomputer. Overall performance improvements of 69% in the 14Lawrence Livermore Loops and 25% in the 24 Lawrence Livermore Loops, asmeasured by the harmonic mean, are attained