Challenges to Combining General-Purpose and Multimedia Processors

M ultimedia workloads have always held an important role in embedded applications , such as video cards or set-top boxes, but these workloads are becoming increasingly common in general-purpose computing as well. Over the past three years the major vendors of general-purpose processors (GPPs) have announced extensions to their instruction set architectures that supposedly enhance the performance of multimedia workloads. These include MAX-2 extensions to Hewlett-Packard PA-RISC, 1 MMX for Intel's x 86, 2,3 UltraSparc's VIS, 4 and MDMX extensions to MIPS V. 5 Processors targeted to embedded multimedia appli-cations—the so-called multimedia processors (MMPs)—have employed similar semantics. These processors include Philip's TriMedia (TM-1), 6 Samsung's Multi-Media Signal Processor (MSP), 7 and Chromatic's Mpact. 8 In addition, some MMPs incorporate multimedia-specific semantics, including vector instructions (MSP and Mpact); a very-long-instruction-word architecture (TM-1 and Mpact); and hardwired special-purpose hardware such as video and audio ports and bitstream codecs. 9 Most of the new instruction semantics for both GPPs and MMPs are based on a subword execution model. This model uses the entire width of a processor data path (32 or 64 bits), even when processing the small native data types found in signal processing (8-or 12-bit pixel, 8-or 16-bit audio). For example, if the word size of a machine is 64 bits, the adder can be used to implement eight 8-bit additions in parallel (by, for example , disconnecting the carry chain in the adder at every eighth position). The same trick can be used to operate in parallel on four 16-bit data types or on two 32-bit data types (for more detail see the sidebar, " Instruction Set Architecture Semantics for Media Processing. ") Merging these new multimedia instructions with existing GPPs poses several challenges. Also, some doubt remains as to whether multimedia extensions are a real development or just a competition-induced fad in the GPP industry. If it is indeed a development, how must current processor microarchitectures change in reaction? And if they change, can GPPs and MMPs apply application-specific integrated circuit (ASIC) solutions to the same problems? Multimedia extensions pack multiple operations into one word's worth of work. This is often referred to as subword execution (as explained in the sidebar). Limited compiler support for targeting subword instructions complicates the use of multimedia benchmarks , even when suitable benchmarks exist. Today, much digital signal processing code is handwritten in assembly to ensure the highest possible efficiency, because such code …