Address-based data processing over N-ary trees

Abstract

The existing trends and growing tendency to high-performance computing strongly require traditional computational algorithms to be revised in such a way that would permit optimal hardware implementations to be found. The latter can be done if algorithms can efficiently be mapped to hardware-specific functional blocks that require different style of thinking and manipulation by different operations and memory models. Besides, concurrency (pipelining and parallelism) can be widely applied. This paper focuses on data processing algorithms that use values of incoming data items as memory addresses with one-bit flags indicating presence or absence of data. The main problem of similar known methods is an appearance of huge number of memory cells with no data items (empty holes), which, nevertheless, have to be processed involving lots of unnecessary time. It is proposed to apply N-ary tree technique that enables data items to be stored and found very fast through executing special traversal procedure. Such trees completely differ from known tree-walk tables because they have pre-established configuration that does not require explicit addresses of subsequent (child) nodes. As a result the size of memory needed to keep data items is decreased in number of times. Finally, significantly more complicated problems can be solved faster and with smaller hardware resources. The presented results of numerous experiments clearly demonstrate advantages of the proposed method compared to known implementations.