Combining Target-independent Analysis with Dynamic Profiling to Build the Performance Model of a DSP

Fast and accurate performance estimation is a key aspect of heterogeneous embedded systems design flow, since cycle-accurate simulators, when exist, are usually too slow to be used during design space exploration. Performance estimation techniques are usually based on combination of estimation of the single processing elements which compose the system. Architectural characteristics of Digital Signal Processors (DSP), such as the presence of Single Instruction Multiple Data operations or of special hardware units to control loop executions, introduce peculiar aspects in the performance estimation problem. In this paper we present a methodology to estimate the performance of a function on a given dataset on a DSP. Estimation is performed combining the host profiling data with the function GNU GCC GIMPLE representation. Starting from the results of this analysis, we build a performance model of a DSP by exploiting the Linear Regression Technique. Use of GIMPLE representation allows to take directly into account the target-independent optimizations performed by the DSP compiler. We validate our approach by building a performance model of the MagicV DSP and by testing the model on a set of significative benchmarks.