Exhaustive Evaluation of Fixed-Point Reciprocal Architectures for FPGAs
Navaneeth Kutty, Anish Murali, Iacovos Ioannou, Priyan Mariappan, Prabagarane Nagaradjane
原文摘要
Fixed-point reciprocals are used in normalization and division on field-programmable gate arrays (FPGAs). Designs that share an output format can differ in worst-case error and in their need for memory and multipliers. The effect of table width, internal precision and refinement was evaluated for reciprocal computation in the Q16.16 and Q8.24 formats. Every representable input in the interval from 0.5 to 2.5 was tested against an exact integer reference. Five arithmetic paths were reconstructed from register-transfer-level source. Guarded interpolation, Newton-Raphson refinement and residual correction were then evaluated alongside polynomial and digit-recurrence baselines. Correct rounding was verified for all 131,073 Q16.16 and 33,554,433 Q8.24 inputs by adding one residual check to a guarded interpolated seed and one refinement step. An implementation with product reuse was verified at one input per cycle with a latency of 10 cycles in both formats. The Q8.24 implementation used nine digital signal processing slices in 7 series synthesis. The exact radix-2 control used no such slices and had a latency of 28 cycles. The reconstructed refined paths reduced the maximum relative error by factors of 205 to 223 when eight fractional bits were added. The evaluated designs provide reproducible choices for fixed accuracy budgets. In particular, correct rounding can be obtained on the tested interval by using a compact seed and checking the integer residual before the result is returned.
cs.AR