Combining FPGA prototyping and high-level simulation approaches for Design Space Exploration of MPSoCs

Modern embedded systems are parallel, component-based, heterogeneous and finely tuned on the basis of the workload that must be executed on them. To improve design reuse, Application Specific Instruction-set Processors (ASIPs) are often employed as building blocks in such systems, as a solution capable of satisfying the required functional and physical constraints (e.g. throughput, latency, power or energy consumption etc.), while providing, at the same time, high flexibility and adaptability. Composing a multi-processor architecture including ASIPs and mapping parallel applications onto it is a design activity that require an extensive Design Space Exploration process (DSE), to result in cost-effective systems. The work described here aims at defining novel methodologies for the application-driven customizations of such highly heterogeneous embedded systems. The issue is tackled at different levels, integrating different tools. High-level event-based simulation is a widely used technique that offers speed and flexibility as main points of strength, but needs, as a preliminary input and periodically during the iteration process, calibration data that must be acquired by means of more accurate evaluation methods. Typically, this calibration is performed using instruction-level cycleaccurate simulators that, however, turn out to be very slow, especially when complete multiprocessor systems must be evaluated or when the grain of the calibration is too fine, while FPGA approaches have shown to performbetter for this particular applications. FPGA-based emulation techniques have been proposed in the recent past as an alternative solution to the software-based simulation approach, but some further steps are needed before they can be effectively exploitedwithin architectural design space exploration. Firstly, some kind of technology-awareness must be introduced, to enable the translation of the emulation results into a pre-estimation of a prospective ASIC implementation of the design. Moreover, when performing architectural DSE, a significant number of different candidate design points has to be evaluated and compared. In this case, if no countermeasures are taken, the advantages achievable with FPGAs, in terms of emulation speed, are counterbalanced by the overhead introduced by the time needed to go through the physical synthesis and implementation flow. Developed FPGA-based prototyping platform overcomes such limitations, enabling the use of FPGA-based prototyping for micro-architectural design space exploration of ASIP processors. In this approach, to increase the emulation speed-up, two different methods are proposed: the first is based on automatic instantiation of additional hardware modules, able to reconfigure at runtime the prototype, while the second leverages manipulation of application binary code, compiled for a custom VLIW ASIP architecture, that is transformed into code executable on a different configuration. This allows to prototype a whole set of ASIP solutions after one single FPGA implementation flow, mitigating the afore-mentioned overhead.A short overview on the tools used throughout the work will also be offered, covering basic aspects of Intel-Silicon Hive ASIP development toolchain, SESAME framework general description, along with a review of state-of-art simulation and prototyping techniques for complex multi-processor systems. Each proposed approach will be validated through a real-world use case, confirming the validity of this solution