MEMORY LATENCY TOLERANCE IN BLOCK PROCESSING PIPELINES

摘要

Memory latency tolerance methods and apparatus for maintaining an overall level of performance in block processing pipelines that prefetch reference data into a search window. In a general memory latency tolerance method, search window processing in the pipeline may be monitored. If status of search window processing changes in a way that affects pipeline throughput, then pipeline processing may be modified. The modification may be performed according to no stall methods, stall recovery methods, and/or stall prevention methods. In no stall methods, a block may be processed using the data present in the search window without waiting for the missing reference data. In stall recovery methods, the pipeline is allowed to stall, and processing is modified for subsequent blocks to speed up the pipeline and catch up in throughput. In stall prevention methods, processing is adjusted in advance of the pipeline encountering a stall condition.

主权项

1. An apparatus, comprising:
a block processing pipeline comprising a plurality of stages each configured to perform one or more operations on a block of pixels from a frame passing through the pipeline; and a memory configured to store a search window including a plurality of columns of blocks of pixels from a reference frame, wherein the search window covers two or more overlapping search regions from the reference frame, each search region including a plurality of blocks; wherein the block processing pipeline is configured to process blocks of pixels from the frame according to corresponding search regions in the search window, wherein, to process the blocks, the block processing pipeline is configured to:
generate requests to prefetch specified columns of blocks of pixels from the reference frame into the search window, wherein the requested columns of blocks include pixels for processing upcoming blocks from the frame at one or more stages of the pipeline;monitor the search window to detect conditions in which one or more blocks from the reference frame are not prefetched into the search window for at least one upcoming block from the frame; andin response to detecting a condition in which one or more blocks from the reference frame are not prefetched into the search window for at least one upcoming block, reduce the search region for one or more blocks from the frame when processing the one or more blocks in the pipeline.