Method to improve DSP kernel's performance/power ratio

摘要

For instruction clusters for which no significant performance penalty is incurred, such as execution of hardware loops, a processor automatically and dynamically switches to a pipelined two-cycle access to an associated associative cache rather than a single-cycle access. An access involving more than one cycle uses less power because only the hit way within the cache memory is accessed rather than all ways within the indexed cache line. To maintain performance, the single-cycle cache access is utilized in all remaining instructions. In addition, where instruction clusters within a hardware loop fit entirely within a pre-fetch buffer, the cache sub-system is idled for any remaining iterations of the hardware loop to further reduce power consumption.