A delay-optimal technology mapping algorithm is developed on a general model of FPGA with hard-wired non-homogeneous logic block architectures which is composed of different sizes of look-up tables (LUTs) hard-wired together. This architecture has the advantages of short delay of hard-wired connections and area-efficiency of non-homogeneous structure. The Xilinx XC4000 is one commercial example, where two 4-LUTs are hard-wired to one 3-LUT. In this paper, we present a two-dimensional labeling approach and a level-2 node cut algorithm to handle the hard-wired feature. The experimental results show that our algorithm generates favorable results for Xilinx XC4000 CLBs. Over a set of MCNC benchmarks, our algorithm produces results with 17% fewer CLB depth than that of FlowMap in similar CPU time on average, and with 4% fewer CLB depth than that of PDDMAP on average while PDDMAP needs 15 times more CPU time.

A new technology mapping algorithm is developed on a general model of FPGA with composite logic block architectures, consisting of different sizes of look-up tables (LUTs) and possibly different logic gates. In additions, the logic blocks may have hard-wired connections and limit accessible fanouts. Xilinx XC4000 is one example containing LUTs of different sizes and AT&T ORCA is another example containing both LUTs and logic gates. We use a multiple-fanout pattern graph library to model the composite logic block and a premapping technique to generate the subject graph dynamically. A new matching algorithm and a new covering algorithm are also developed for the subject graph covering. The experimental results show that our algorithm is an effective technology mapper for FPGAs with composite logic block architectures, especially for larger circuits. Over a set of MCNC benchmarks, our algorithm requires on the average 4.25% fewer CLBs than PPR, 6.79% fewer CLBs than TEMPT, and 2,79% fewer CLBs than ASYL when used as the XC4000 mapper. Over a set of larger benchmarks, our algorithm outperforms PPR by 13.70%. Very encouraging results were obtained when our algorithm is used as an ORCA mapper, while there was no prior published results.