Accelerating Vector Operations by Utilizing Reconfigurable Coprocessor Architectures

To enhance performance of digital signal processing tasks while keeping the flexibility of programmable solutions is a clear motivation for coprocessors implemented as reconfigurable hardware blocks. This paper investigates the applicability of such coprocessors targeting digital signal processing multi-media applications, initially in the field of speech and audio. A tightly coupled coprocessor architecture with reconfigurable datapath and a local memory system is presented. The coprocessor interacts with the main processor through asynchronous FIFOs. Three computational models that provide support for functionality of different granularities to be accelerated are investigated. A speedup in the range of 2 to 46 compared to processor execution is achieved for vector operations and larger kernels such as autocorrelation, block filtering and fast Fourier transform.