Estimating the Potential Speedup of Computer Vision Applications on Embedded Multiprocessors
This work provides a tool for developers to optimize parallelization on embedded systems, though it is incremental as it builds on existing simulation approaches.
The paper tackles the challenge of predicting parallel performance for computer vision applications on embedded multiprocessors by presenting a trace-driven simulation method, achieving less than 10% error and up to 20x faster execution compared to a reference simulator.
Computer vision applications constitute one of the key drivers for embedded multicore architectures. Although the number of available cores is increasing in new architectures, designing an application to maximize the utilization of the platform is still a challenge. In this sense, parallel performance prediction tools can aid developers in understanding the characteristics of an application and finding the most adequate parallelization strategy. In this work, we present a method for early parallel performance estimation on embedded multiprocessors from sequential application traces. We describe its implementation in Parana, a fast trace-driven simulator targeting OpenMP applications on the STMicroelectronics' STxP70 Application-Specific Multiprocessor (ASMP). Results for the FAST key point detector application show an error margin of less than 10% compared to the reference cycle-approximate simulator, with lower modeling effort and up to 20x faster execution time.