Automatic Parallelization of Sequential Specifications for Symmetric MPSoCs

This paper presents an embedded system design toolchain for automatic generation of parallel code runnable on symmetric multiprocessor systems from an initial sequential specification written using the C language. We show how the initial C specification is translated in a modified system dependence graph with feedback edges (FSDG) composing the intermediate representation which is manipulated by the algorithm. Then we describe how this graph is partitioned and optimized: at the end of the process each partition (cluster of nodes) represents a different task. The parallel C code produced is such that the tasks can be dynamically scheduled on the target architecture; this is obtained thanks to the introduction of start conditions for each task. We present the experimental results obtained by applying our flow on the sequential code of the ADPCM and JPEG algorithms and by running the parallel specification, produced by the toolchain, on the target platform: with respect to the sequential specification, speedups up to 70% and 42% were obtained for the two bebchmarks respectively.