• DocumentCode
    466835
  • Title

    Research on the RAW Dependency in Floating-point FFT Processors

  • Author

    Mou, Shengmei ; Yang, Xiaodong

  • Author_Institution
    Nat. Univ. of Defense Technol., Changsha
  • Volume
    1
  • fYear
    2007
  • fDate
    July 30 2007-Aug. 1 2007
  • Firstpage
    88
  • Lastpage
    92
  • Abstract
    The appearance of floating-point FFT processors and introduction of multiple butterfly units and high-radix structure increase the possibility of read-after-write (RAW) dependency between adjacent stages. The less the transform length is, the more butterfly units are used, the more cycles the butterfly units cost, the more possibility of RAW dependency lurks. In this paper, taking radix-2 time decimation FFT processors for example, we define the minimum slack cycles for each stage, which provides a quantitative method to identify RAW dependencies in FFT processors of variable transform length, and put forward some proposals to reduce or erase RAW dependencies. The minimum slack cycles also indicates the numbers of waiting cycles should be inserted into each stage, which simplifies the design of control path in FFT processors. As to the situation of multiple butterfly units and high-radix structure, same ideas can be used.
  • Keywords
    digital signal processing chips; fast Fourier transforms; floating point arithmetic; pipeline arithmetic; random-access storage; RAW dependency; fast Fourier transform; floating-point FFT processors; radix-2 time decimation FFT processors; read-after-write dependency; Artificial intelligence; Computer science; Costs; Delay; Digital signal processing; Discrete Fourier transforms; Distributed computing; Pipeline processing; Proposals; Software engineering;
  • fLanguage
    English
  • Publisher
    ieee
  • Conference_Titel
    Software Engineering, Artificial Intelligence, Networking, and Parallel/Distributed Computing, 2007. SNPD 2007. Eighth ACIS International Conference on
  • Conference_Location
    Qingdao
  • Print_ISBN
    978-0-7695-2909-7
  • Type

    conf

  • DOI
    10.1109/SNPD.2007.48
  • Filename
    4287480