Optimal FPGA mapping and retiming with efficient initial state computation

Authors
Citation
J. Cong et C. Wu, Optimal FPGA mapping and retiming with efficient initial state computation, IEEE COMP A, 18(11), 1999, pp. 1595-1607
Citations number
24
Categorie Soggetti
Eletrical & Eletronics Engineeing
Journal title
IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS
ISSN journal
02780070 → ACNP
Volume
18
Issue
11
Year of publication
1999
Pages
1595 - 1607
Database
ISI
SICI code
0278-0070(199911)18:11<1595:OFMARW>2.0.ZU;2-8
Abstract
For sequential circuits with given initial states, new equivalent initial s tates must be computed for retiming, which unfortunately is:NP-hard. In thi s paper, we present a novel polynomial time algorithm for optimal field pro grammable gate array (FPGA) napping with forward retiming to minimize the c lock period-with efficient initial state computation, It considers forward retiming, initial state computation and mapping simultaneously.. Our algori thm enables a new methodology of separating forward retiming from backward retiming, Since we guarantee to compute an optimal mapping with forward ret iming solution, backward retiming can be performed as a preprocessing to tr y to push hip-hops (FF's) to primary inputs with consideration of only init ial state computation, Thus, we can avoid the time-consuming iterations bet ween retiming for clock period minimization and initial state computation. This algorithm compares very favorably to both conventional approaches of m apping followed by separate retiming and recent approaches of combined mapp ing with retiming, but; without consideration of initial state computation [1], [2], Our results show that our algorithm can reduce the clock period: by 17.5% over conventional approaches of separate mapping with retiming. On the other hand, our algorithm can guarantee efficient initial state comput ation in the mapped and retimed circuits with only 2.8% increase in clock p eriod over the optimal mapping with general retiming solutions, while the i nitial state computation for the latter solutions may need prohibitively lo ng runtime for designs with over several hundred FF's.