Studies - Texture Prefetching

Several studies, including one done by Hakura and Gupta, have demonstrated the efficiency of caching with respect to texture mapping. However, the latencies of texture memory systems is often too high to for a simple blocking cache, and some sort of prefetching is required. Because the access patterns of texture mapping are constrained, we designed a specialized prefetching architecture that can take advantage of them. By separating the cache tags from the cache data in time and introducing a reorder buffer that holds prefetched cache blocks until they are needed by the cache, we are able to prefetch data without worrying about conflicts introduced by prefetching. A cycle- accurate simulator has shown that the architecture covers arbitrarily high latencies with over 97% efficiency.

The primary developers of this prefetching texture cache architecture are Homan Igehy, Matthew Eldridge, and Kekoa Proudfoot. Details of the work can be found in the HWWS '98 paper entitled Prefetching in a Texture Cache Architecture.


homan@graphics.stanford.edu