Section: New Results

VMAD

Participants : Alexandra Jimborean, Philippe Clauss, Jean-François Dollinger, Aravind Sukumaran-Rajam, Juan Manuel Martinez Caamaño, Vincent Loechner.

The goal of the VMAD 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.

VMAD contains 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 templates, and callbacks to the runtime system. External modules associated to specific analyses and transformations are dynamically loaded when required at runtime. Dynamic information, such as memory locations of the modules entries, are patched at startup in the loaded executable.

VMAD 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. VMAD is handling advanced memory access profiling [17] through linear interpolation of the addresses, dynamic dependence analysis, version selection [17] and speculative polyhedral parallelization [19] , [16] .

Alexandra Jimborean defended her PhD thesis on this topic in 2012 [12] . In 2012, Aravind Sukumaran-Rajam started a PhD in our team to continue this work, especially on extending the dependence analysis to make it handle more general programs, keeping it fast and accurate. Jean-François Dollinger will extend the framework to handle heterogeneous architectures (GPGPUs). Juan Manuel Martinez Caamaño, a master student of University of Buenos Aires (associate team EA-Ancome) is also working on VMAD to make the code generation support tiling.