Memory Dependence Prediction in Multimedia Applications

We identify that a set of multimedia applications exhibit highly regular read-after-read (RAR) and read-after-write (RAW) memory dependence streams. We exploit this regularity to predict both RAW and RAR memory dependences. We also study how two previously proposed memory dependence prediction-based memory latency reduction techniques perform for this multimedia workload. In the first technique, a load can obtain a value by simply naming a preceding load (or store) with which a RAR (or RAW) dependence is predicted. The second technique speculatively converts a series of LOAD1-USE1,...,LOADN-USEN (or DEF-STORE-LOAD-USE) chains into a single LOAD1-USE1...USEN (or DEF-USE) producer/consumer graph. We show that via memory dependence prediction it is possible to correctly predict 33.3% of all loads on the average. Moreover, the two memory dependence prediction based techniques result on average performance improvements of 2.6% over a highly-aggressive, out-of-order, superscalar processor. The actual range of performance improvements is 0% to 8.5%. When cache latency is increased from 2 to 3 cycles, performance improves by 3.75% on average, with the range being 0% to 16.35%.