Studies on automatic parallelization for heterogeneous and homogeneous multicore processors
Akihiro Hayashi · Institutional Repositories DataBase (IRDB) · 2012
There has been a growing interest in heterogeneous and homogeneous multicore processors due to their excellent characteristic, higher performance and lower power consumption than single core processors. In order to exploit the capability of heterogeneous and homonogeneous multicore processor, an automatic parallelization is greatly important since parallel programming by hand is time consuming task. This thesis proposes an innovative OSCAR (Optimally Scheduled Advanced Multicore Processors) heterogeneous automatic parallelizing compiler, which realizes an automatic parallelization and a power reduction for various heterogeneous and homogeneous multicores. Performance evaluation and power evaluation on 15 core heterogeneous multicore RP-X using media applications attains speedups of 32.6 times with eight SH-4A cores and four FE-GA cores, 18.8 times with two SH-4A cores and one FE-GA core, 5.4 times with eight SH-4A cores against sequential execution by a single SH4A core and 70% of power reduction for the optical flow. Performance evaluation on several SMP servers using dose calculation for heavy particle radiotherapy for cancer treatment attains good speedups of 9.0 times with 12 processor cores on Hitachi HA8000/RS220 system based on the Intel Xeon Processor and 50.0 times with the 64 processor cores on Hitachi SR16000 system