A complete industrial multi-target graphical tool-chain for parallel implementations of signal/image applications
Résumé
Today, applications developed in an industrial environment need to shift from a sequential implementation to one or several different parallel ones, under stringent productivity and performance constraints. As this operation is generally complex and moreover needs particular skills which are not part of the traditional software engineering cultural background, there is a clear need for tools to assist it. A parallel programming tool-chain must be able to address various parallel target hardware platforms. This may be due to obsolescence of components, to different environmental constraints, or different requirements of applications that make some candidate targets more efficient than others. So the tool-chain must be easily adaptable to heterogeneous compositions of FPGAs, general purpose processors, DSPs, GPUs, etc. while hiding as much as possible the cumbersome details of these architectures to the user. However, with the current state of the art, except for very simple cases, it seems difficult to find sufficiently robust completely automated tool-chains tackling the often tricky architecture-specific trade-offs done for mapping), especially when real time and performance are needed. Thus, a man-in-the-loop approach is preferred, where the user can experiment different variants through fast prototyping, performance analysis, and rapid design space exploration. This implies providing the user with sufficiently abstracted views of both application and architecture, giving visibility on the information that is relevant to parallelization and mapping, as well as the commands enabling him/her to drive the main mapping decisions. We present here such models of both applications and architectures, and two complementary tools handling them and covering the entire design flow. The first tool based on the source-to-source compiler PIPS1, generates application models from application code, and the second tool, SpearDE2, supports the flow from models to code generation. We address the signal and image processing domains. [...]