Automatic parallelization with OSCAR API Analyzer: a cross-platform performance evaluation
CeciliaGonzalez-Alvarez, Kanehagi Youhei, Takemoto Kosei, Yohei Kishimoto, Muto Kohei, Hiroki Mikami, Akihiro Hayashi, Keiji Richard Kimura, Kasahara Hironori · Applied Reconfigurable Computing · 2012
To satisfy the demands of auto parallelizing compilers in the diverse industry of multicores, we have devel- oped the OSCAR API Analyzer. It allows programs automatically parallelized by the OSCAR compiler with OSCAR API directives to target many different platforms using just sequential compilers. We have evaluated the execution performance of the parallelization of Fortran SPEC benchmarks (tomcatv, swim2000, mgrid2000) and media C bench- marks (AAC encoder, Optical flow, MPEG2 encoder, MPEG2 decoder, Face detect) on five HPC servers and four embedded multicores. Speedups on servers were up to 18x for 32 cores (swim2000 on Hitachi SR16000), whereas on embedded systems, AAC encoder speedup was up to 47x on TilePro64, for 64 homogeneous cores, and up to 32.65x for the optical flow on the heterogeneous multicore RP-X, using 8 cores and 4 accelerators.