A novel dynamic SIMD-chain

Traditional data parallel SIMD-architectures are limited to executing applications that are largely free of data-dependent control constructs. However, real-world applications consist of more than highly parallel inner loops, many loops are non-vectorizable because they contain loop-carried data dependencies from one iteration to the next, the majority of the resources sit idle in such situations. This paper proposes a novel dynamic SIMD-chain that efficiently accelerate loop depended code segments on SIMDs. Experimental results show that, it can achieve about 2.4x average performance improvement, at a cost of only 0.97% area overhead.