Analysis, optimization and execution of general purpose multimedia applications on subword vliw datapaths
Steven P. Levitan, Majd F. Sakr · 2003
In this thesis we evaluate the characteristics of multimedia applications and propose a Multiple Instruction Stream Multiple Data Stream (MIMD) subword Very Long Instruction Word (VLIW) datapath that overcomes the limitations of current architectures to effectively execute multimedia applications. The characteristics of multimedia applications differ from these of conventional technical applications, in that multimedia applications are highly parallel, computationally intensive, and use low precision (subword) data that can be streaming in nature. Conventional architectures with fixed-width datapaths cannot effectively perform the required computation. Current solutions employ media-centric single instruction stream multiple data stream (SIMD) based instruction set extensions. However, compilers cannot automatically target these instructions and have to either use hand-tuned assembly libraries or compiler intrinsics. These restrictions limit the full exploitation of the available parallelism in general purpose media applications implemented using high-level programming constructs, which limits performance gains. We alleviate these restrictions by allowing the compiler to target a flexible MIMD datapath with support for subword execution. Also, we enable the compiler to better exploit the parallelism within the media applications by introducing simple code transformations. To measure the effectiveness of our solution, we evaluate the performance of multimedia applications on the proposed subword MIMD VLIW datapath. This evaluation is performed on a set of multimedia benchmark kernels by analyzing compiler transformations and optimizations and then compiling and executing each kernel using the set of techniques that best target our architecture. The result is an architecture and an analysis methodology that exploits parallelism across a wide range of multimedia applications by providing better performance and enhanced applicability which in turn enables the required realism in multimedia applications running on general purpose processors.