AI-based estimation of embedded software execution cycles in host-compiled simulation

In Embedded and Cyber-Physical Systems (E&CPS) the functional and temporal interactions between the digital and physical parts are crucial. In these systems, most of the functionality is implemented by software deployed on heterogeneous, distributed platforms so its performance largely affects the whole E&CPS behavior. Their HW/SW co-design requires fast and accurate simulation tools capable of evaluating the performance of each platform configuration for the applications selected while minimizing software porting needs. Host-compiled simulation avoids porting, but has limitations to obtain accurate results due to the difficulty of extracting and modeling the microarchitectural details of the target platforms while maintaining simulation speed. To solve that, this work proposes replacing the traditional approach of generating code annotations with all processor internal details by the use of neural networks. Training the neural network for a specific processor enables considering its internal details when estimating the cost of executing each basic block of the software in the target processor.