所提出的存取方案並非基於有序的資料指標，故運作過程中無須對記憶體模組進行交換排序，而是針對每個記憶體模組直接產生列指標，乃是對控制邏輯的進一步簡化；又基於所提出存取方案的資料分布特性，於實際運算架構中抑制連續 FFT 運算的資料分布相依性，而得以再次進一步簡化控制邏輯。總結本研究的核心成果，首先在於針對不定資料分布的研究提出方法，其次又基於所提出的存取方案得以在兩個層面簡化控制邏輯，說明所能帶來的新發展。 ;Under the condition of guaranteeing sufficient throughput rate, memory-based architecture with low area overhead would be preferable for executing Cooley-Tukey algorithm, thus memory access scheme is the object of this study while it is an essential issue of memory-based architecture. Data distribution is a primary consideration for access scheme, and a criterion of constant distribution is established in previous work, which means the mapping between data sequence and memory address is fixed throughout entire compute procedure. In contrast, this work attempts to develop an access scheme under a criterion of inconstant distribution and begins with developing a novel modeling method.
This work proposes an access scheme for arbitrary power-of-two radix FFT algorithm and a corresponding row index generator which is highly hardware efficient. The area overhead of generator grows linearly with word length of row index, thus it is suitable for long FFT length. On the other hand, the propagation delay of generator remains constant for arbitrary power-of-two radix, thus it is suitable for high radix FFT algorithm. Furthermore, the proposed access scheme is applied to Ping-pong Cache-memory architecture for evaluating its feasibility. At the same time, issues of eliminating non-ideal effects encountered by memorial unit are also discussed. Compared to previous work, this work extracts sufficient information of distributing data from signal flow graph, SFG, rather than ordered data indexes, hence permutation of memory modules for parallel accessing is not required; besides, data distribution dependency between FFT computations is eliminated. Thus complexity of control logic is reduced in two aspects.
In conclusion, the primary achievement of this work is proposing a novel modeling method for inconstant distribution and reducing complexity of control logic to demonstrate possible evolution which could be brought by inconstant distribution.
