HyPPO：用于硬件高效设计的混合分段多项式近似和优化.pdf

1、HyPPO:Hybrid Piece-wise Polynomial Approximation and Optimization for Hardware Efficient DesignsLAKSHMI SAI NIHARIKA VULCHI,VALIPIREDDY PRANATHI,MAHATI BASAVARAJU AND MADHAV RAOIIIT Bangalore,India30th Asia and South Pacific Design Automation ConferenceASP-DAC 2025ABSTRACT01BACKGROUND INFORMATION02P

2、ROPOSED METHODS0304TABLE OF CONTENTSEXPERIMENTAL RESULTS05CONCLUSIONABSTRACT Hybrid Piece-wise Polynomial Approximation(PW-Hybrid)improves hardware implementation of non-linear functions by combining linear(PWL)and quadratic(PWQ)methods to reduce approximation errors.Particle Swarm Optimization(PSO)

3、fine-tunes quantized bit-widths for polynomial coefficients,improving hardware efficiency.The PSO optimized hardware design is evolved for different non-linear functions.Significant reductions in hardware resource usage and critical path delay achieved,evaluated using Cadence 45nm gpdk library.BACKG

4、RO U ND INFORMATIONIn Piece-wise Polynomial(PW-Poly)approximation,the domain of a function is divided into segments,and a polynomial is employed to define the function in each segmentApproximation of non-linear functions Piece wise linear Segment defined by a linear function,greater number of segmen

5、ts,requires more memory to store the coefficientsPiece wise quadratic Segment defined by a quadratic polynomial,computational hardware is highThe hardware identifies the segment containing input x,retrieves the corresponding polynomial coefficients from a LUT,and uses them for approximation.The coef

6、ficients are used in addition and multiplication units to compute the polynomial and generate the approximated output for the target function.Polynomial coefficients are computed through two steps:fitting,which minimizes the maximum absolute error(MAE),and segmentation,which optimally divides the ta