A low-cost high-speed twin-prefetching DSP-based shared-memory system for real-time image processing applications.

摘要

This dissertation introduces, investigates, and evaluates a low-cost high-speed twin-prefetching DSP-based bus-interconnected shared-memory system for real-time image processing applications. The proposed architecture can effectively support 32 DSPs in contrast to a maximum of 4 DSPs supported by existing DSP-based bus-interconnected systems. This significant enhancement is achieved by introducing two small programmable fast memories (Twins) between the processor and the shared bus interconnect. While one memory is transferring data from/to the shared memory, the other is supplying the core processor with data. The elimination of the traditional direct linkage of the shared bus and processor data bus makes feasible the utilization of a wider shared bus i.e., shared bus width becomes independent of the data bus width of the processors. The fast prefetching memories and the wider shared bus provide additional bus bandwidth into the system, which eliminates large memory latencies; such memory latencies constitute the major drawback for the performance of shared-memory multiprocessors. Furthermore, in contrast to existing DSP-based uniprocessor or multiprocessor systems the proposed architecture does not require all data to be placed on on-chip or off-chip expensive fast memory in order to reach or maintain peak performance. Further, it can maintain peak performance regardless of whether the processed image is small or large.; The performance of the proposed architecture has been extensively investigated executing computationally intensive applications such as real-time high-resolution image processing. The effect of a wide variety of hardware design parameters on performance has been examined. More specifically tables and graphs comprehensively analyze the performance of 1, 2, 4, 8, 16, 32 and 64 DSP-based systems, for a wide variety of shared data interconnect widths such as 32, 64, 128, 256 and 512. In addition, the effect of the wide variance of temporal and spatial locality (present in different applications) on the multiprocessor's execution time is investigated and analyzed. Finally, the prefetching cache-size was varied from a few kilobytes to 4 Mbytes and the corresponding effect on the execution time was investigated. Our performance analysis has clearly showed that the execution time converges to a shallow minimum i.e., it is not sensitive to the size of the prefetching cache. The significance of this observation is that near optimum performance can be achieved with a small (16 to 300 Kbytes) amount of prefetching cache.