Communication-Computation Overlap Strategy for Vector DSP Based on Phase Feature Optimization

As the scale of computation continues to expand, it is necessary to focus not only on the improvement of computing performance but also on the impact of communication overhead, which cannot be ignored. The communication-computation overlap (CC overlap) strategy, as a technology that effectively reduces the impact of transmission on performance, is becoming increasingly important. However, traditional CC overlap strategies lack a clear data volume threshold to guide their application and fail to hide the prefetching overhead. We propose a phase feature optimized CC overlap strategy in vector DSPs. It divides the process into three phases and introduces three optimizations, building an efficient coding framework. This boosts overlap efficiency and adaptability to different data scales. Experiments on the FT-M7004 platform show the scheme speeds up two DSPLIB algorithms by over $10 \%$ on average in small and large data volume scenarios compared to traditional methods.