This paper describes a compiling strategy for SVM that exploits parallelism and minimises overheads when using an SPMD execution model. Our strategy integrates compiler techniques developed separately for data and control parallelism paradigms and integrates them into one approach. This strategy has been implemented in a prototype compiler, called MARS, currently running on the KSR-1 architecture. The technique is intended to be generic for all SVM systems. Initial results are encouraging.
|Title of host publication||Languages, Compilers and Run-Time Systems for Scalable Computers|
|Editors||Boleslaw K. Szymanski, Balaram Sinharoy|
|Place of Publication||Boston, MA|
|Number of pages||13|
|Publication status||Published - 1996|