120 likes | 135 Views
This paper presents a complexity-aware quantization algorithm for FIR filters to simplify implementations. It includes successive approximation and common subexpression elimination. The algorithm distributes an addition budget among filter coefficients, ensuring efficient design. The study explores improved scaling factors with reduced complexity, offering better results with fewer candidates. The proposed method enables designers to balance quantization errors for simplified and adaptable implementations.
E N D
Area-Effective FIR Filter Design for Multiplier-less Implementation Tay-Jyi Lin, Tsung-Hsun Yang, and Chein-Wei Jen Department of Electronics Engineering National Chiao Tung University, Taiwan {tjlin, thyang, cwjen}@twins.ee.nctu.edu.tw
In this paper • We propose a complexity-aware quantization algorithm of FIR filters, which enables designers to explicitly trade quantization error for simpler implementations • The proposed algorithm precisely distributes a pre-defined addition budget among the filter coefficients with successive approximation and common subexpression elimination
Outline • Preliminary • Quantization by Successive Coefficient Approximation • Common Subexpression Elimination • Complexity-Aware Coefficient Quantization • Simulation Result • Conclusion
Quantization by Successive Approximation* * D. Li, Y. C. Lim, Y. Lian, and J. Song, “A polynomial-time algorithm for designing FIR filters with power-of-two coefficients,” IEEE Trans. Signal Processing, vol.50, pp.1935-1941, Aug 2002
Constant Multiplications • Consider a 4-tap FIR filter with the coefficients: h0=0.0111011, h1=0.0101110, h2=1.0110011, and h3=0.0100110 Common Subexpression across Coefficients (CSAC)
CSAC CSWC Common Subexpression Elimination • Tabular form
Steepest-descent CSE Heuristic* * M. Mehendale, S. D. Sherlekar, VLSI Synthesis of DSP Kernels - Algorithmic and Architectural Transformations, Kluwer Academic Publishers, 2001
Outline • Preliminary • Complexity-Aware Coefficient Quantization • Simulation Result • Conclusion
Complexity-Aware Quantization Complexity-aware allocation of non-zero terms (with CSE) Improved SF Exploration (next page)
Improved SF Exploration • Instead of the fixed 2-w stepping from the lower bound, the next SF is calculated as denotes the magnitude of a coefficient denotes the distance to its next quantization level as the SF increases, which depends on the approximation scheme (e.g. rounding to the nearest value, toward 0, or toward -∞, etc).
Simulation Result CSE Improved SF Search For 16-bit wordlength and ±3dB acceptable gain, the improved SF exploration has 14,986 to 20,429 candidates depending on the coefficients, instead of 45,875 for all.
Conclusion • Successive approximation with appropriate scaling can significantly reduce the addition complexity • The proposed algorithm controls the CSE to incur the minimum additions during the successive approximation • The improved SF exploration finds better or identical (but never worse) results with only 1/3 candidates • The proposed complexity-aware quantization algorithm allows designers to explicitly trade quantization error for simpler implementations, which can also be easily • modified for goals other than small area (e.g. low power, etc), or • adapted to other implementation styles (e.g. FIR code generation for programmable processors)