Cache performance of video computation workloads

Many video computation workloads execute on general-purpose computers including workstations, desktop systems and personal computers. The microprocessors that form the central processing unit of these systems typically expend a significant part of their real estate for on-chip caches and hence it is important to understand cache behavior of video computation workloads. Caches generally improve application performance by capturing temporal and spatial locality in instruction and data accesses. Modern processors run at speeds in the GHz range, and contain 2 or 3 levels of caches. The access time of main memory units in these systems are in the order of 100 cycles. Unless the caches can result in good hit ratios, the memory access times are prohibitive. We study the cache performance of video workloads. The workloads studied include JPEG, MPEG-2 and H263. The impact of cache sizes, block sizes and associativity on cache performance is discussed. It is observed that caches are used very well by these video computation workloads.