| Summary: | 碩士 === 國立清華大學 === 資訊工程學系 === 95 === With increasing number of multimedia applications for embedded systems, greater computing powers are needed for embedded system to support applications. Therefore, the embedded systems now frequently contains one MPU (main process unit) to run regular tasks and one DSP (digital signal processor) to handle special tasks which need lots of computations. In addition, there are needs to boost performance of DSP accelerators, this frequently coming with needs to handle more diversified applications. Frequently seen extensions include the extension of the DSP processors with more instructions to speed up a particular set of applications, the increase of data memory and program memory banks, the increase of instruction address and address bus, the increase of memory width, the increase of data size for each register, the increase of register size, the increase of concurrent execution units, etc. The advantage of this is that DSP core has more special instructions such as accumulate operation or multiple memory back access to accelerate execution speed or increase data bandwidth, and therefore obtains better performance than uni-core architecture. However, if tool-chain can not use these hardware features e±ciently the advantage of the enhanced system can not be performed. In this thesis, we investigate the issues to enable toolkit flows for such DSP extensions via a case study with GCC flows on audio DSP processors. The toolkit flow thus consists of compilers, assembler, linkers, ELF format, and library designs. Experiments are performed to show that our flow is a workable solution in practice.
|
|---|
