Combining dataflow programming and polyhedral optimization, a case study

Nowadays, parallel computers have become ubiquitous and currentprocessors contain several execution cores, anywhere from a couple tohundreds. This multi-core tendency is due to constraints preventingthe increase of clock frequencies, such as heat generation and powerconsumption. A variety of low-level tools exist to program these chipsefficiently, but they are considered hard to program, to maintain, andto debug, because they may exhibit non-deterministic behaviors. Thisproject focuses on adding an abstraction level in order to have asmuch performance as possible while not dealing with low-levelmechanisms. The approach is based on data flow programming, whichallows the programmer to specify only the operations to perform andtheir dependencies, without actually scheduling them. This projectcombines this paradigm with the Polyhedral Model, which allowsautomatic parallelization and optimization of loop nests, in order tomake the programming easier by delegating work to the compilers andstatic analyzers