Section: New Results
Process network models with explicit data size handling
Participants : Amin Oueslati, Robert de Simone.
We concluded our activities in the definition of a process network, inspired from established formalisms such as Ptolemy's SDF, StreaMIT, and Thales Array-OL task graph languages. Our next formalisms described accurately how regular data structures (2-dimensional arrays or matrices mostly) get assembled or deassembled in actual data-flow computations for streaming intensive data/signal processing. This allows to allocate these computations to similar dedicated architectures (GPUs, TPUs) while making all kinds of parallelism (data-, task-, streaming) explicit. The resulting forms of specification are intently very close to representations that may be expressed in OpenMP or MPI, and cover the important class of Deep Networks filter stream models, which have raised tremendous interest lately in Artificial Intelligence.