Fast and Portable Vector DSP Simulation Through Automatic Vectorization

Vector DSPs are quite common in embedded SoCs used in compute-intensive domains such as imaging and wireless communication. To achieve short time-to-market, it is crucial to provide system architects and SW developers with fast and accurate instruction set simulators of such DSPs. To this end, a methodology for accelerating the simulation of vector instructions in vector DSPs is proposed. The acceleration is achieved by enabling automatic translation of the vector instructions in a given vector DSP binary into host SIMD instructions. The key advantage of the proposed methodology is its independence from the host architecture. Empirical evaluation, using a set of commercial vector DSPs, shows that the proposed methodology provides a 4x average reduction in simulation time of a vector instruction and a 2x average reduction in simulation time of a whole application.