Section: New Results

VMAD and APOLLO

The goal of the APOLLO project is to provide a set of annotations (pragmas) that the user can insert in the source code to perform advanced analyses and optimizations, for example dynamic speculative parallelization. It is based on the prototype VMAD developed previously by the team between 2009 and 2012.

APOLLO includes a modified LLVM compiler and a runtime system. The program binary files are first generated by our compiler to include necessary data, instrumentation instructions, parallel code skeletons, and callbacks to the runtime system which is implemented as a dynamic library. External modules associated to specific analyses and transformations are dynamically loaded when required at runtime.

APOLLO uses sampling and multi-versioning to limit the runtime overhead (profiling, analysis, and code generation). At runtime, targeted codes are launched by successive chunks that can be either original, instrumented or optimized/parallelized versions. After each chunk execution, decisions can be taken relatively to the current optimization strategy. APOLLO is handling advanced memory access profiling [26] , [17] through linear interpolation of the addresses, dynamic dependence analysis [18] , version selection [26] and speculative polyhedral parallelization [22] , [17] .

Alexandra Jimborean defended her PhD thesis on this topic in 2012 [25] .

In 2012, Aravind Sukumaran-Rajam started his PhD in our team to extend this work in order to handle more general programs which do not exhibit a pure polyhedral memory behavior. The investigated approach will explore approximative modelling of dependences still allowing advanced optimizing transformations of loop nests. A main issue concerns speculation verification when using approximative modelling.

Juan Manuel Martinez started his PhD in our team in 2013, with the goal of improving the flexibility of the parallel code generation phase inside Apollo. Indeed, although code skeletons are a good solution to fast dynamic parallel code generation, their shapes limit the kind of optimizing transformations that may be applied at runtime. Juan Manuel's work consists in defining elementary code skeletons that may be assembled at runtime to form a large panel of possible codes. These elementary skeletons will be defined as the objects forming the Apollo specific intermediate representation. Juan Manuel Martinez is a former master student of the University of Buenos Aires, Argentina (associate team EA-Ancome), and has already been working on VMAD to make the code generation support tiling. He defended his master thesis on this subject in October 2013 at the University of Buenos Aires.

Jean-Fran cois Dollinger will extend the framework to handle heterogeneous architectures (GPGPUs) in 2014. Willy Wolff, a master student from the University of Strasbourg, joined the APOLLO group in September 2013. His work is to implement just-in-time compilation in the APOLLO framework.