Efficient Support for All Levels of Parallelism for Complex Media Applications

Real-time execution of contemporary media applications (e.g., high resolution video encoding/conferencing/editing, face/image/speech recognition, and image synthesis like ray-tracing) needs a considerable amount of processing power that surpasses the capabilities of current superscalar processors. Further, high performance processors are often constrained by power/energy consumption, especially in the mobile systems where media applications have become increasingly popular. Fortunately, most media applications have a lot of data-level parallelism (DLP) which can potentially be exploited for energy-efficient high performance. Unfortunately, DLP in these complex applications is often interspersed with control. Therefore, architectures that support only DLP efficiently are unlikely to suffice. This work makes two broad contributions. First, it makes the case that the complexity of contemporary media applications requires energy-efficient support for multiple forms of parallelism, including ILP, TLP, and various forms of DLP such as sub-word SIMD instructions, vectors, and streams. Second, it shows that all of these forms of parallelism can be efficiently integrated within an evolutionary architecture with little additional hardware support compared to conventional paradigms. We propose such an architecture, called ALP, which is a superscalar based CMP/SMT architecture augmented with novel DLP support. Our evaluations show that our design decisions in ALP are effective. For our application suite, relative to a singlethread superscalar, ALP achieves speedups from 5X to 49X, energy reduction of up to 7.4X, and energy-delay product (EDP) reduction of 5X to 361X. We also find that no single type of parallelism gives the best possible speedup or energy efficiency for these applications. Support for all levels of parallelism is essential in obtaining the possible speedups and energy savings. Our results also support the claim that a superscalar based CMP/SMT architecture augmented with evolutionary DLP support is promising for obtaining high performance and energy efficiency in complex media applications. As part of this work, we have also performed a study to find the most energy efficient general purpose architecture for supporting mulitple threads. We make a design space search comparing the energy efficiency of various CMP, SMT and hybrid architectures and identify key factors that contribute to energy efficient implementation of thread support. In the future, we will improve ALP by enhancing the vector support and memory system, study the importance of ILP in the presence of DLP, and evaluate more applications to strengthen our conclusions.