Exploiting multiple granularities of parallelism in a compiler

Parallelism can be exploited at different granularities, and many different schemes have been proposed for exploiting parallelism at each of these levels. The Stardent Titan compiler and architecture exploit parallelism at all granularities, driving the parallelism detection off a single dependence graph. The Titan architecture and compilers are described, with particular emphasis on the exploitation of parallelism at all levels. A Titan comprises one to four processing units. All processing units access a shared memory through a very high-speed bus. Synchronization is provided through a test-and-set primitive. Each Titan processing unit contains an integer processor and a completely independent floating-point processor.<>