ECE4270 Fundamentals of DSP Lecture 20. Fixed-Point Arithmetic in FIR and IIR Filters (part I) Overview of Lecture. Overflow. FIR Digital Filter

Size: px

Start display at page:

Download "ECE4270 Fundamentals of DSP Lecture 20. Fixed-Point Arithmetic in FIR and IIR Filters (part I) Overview of Lecture. Overflow. FIR Digital Filter"

Winfred McGee
5 years ago
Views:

ECE4270 Fundamentals of DSP Lecture 20 Fixed-Point Arithmetic in FIR and IIR Filters (part I) School of ECE Center for Signal and Information Processing Georgia Institute of Technology Overview of

1 ECE4270 Fundamentals of DSP Lecture 20 Fixed-Point Arithmetic in FIR and IIR Filters (part I) School of ECE Center for Signal and Information Processing Georgia Institute of Technology Overview of Lecture Two s-complement arithmetic Overflow Issues in FIR implementation: Quantizing filter coefficients Roundoff noise and Scaling Introduction to IIR Filter Structures and Quantization Coefficient Quantization effects in IIR filters Overflow Overflow can occur in two ways: Accumulator overflows due to additions Taking 16 bits out of the wrong part of a product Overflow can be prevented by: ore precision - use larger accumulator Saturation arithmetic - clip accumulator at largest value Shift products to the right to discard low-order bits. This results in loss of precision. y[n] h[k]x[n k] FIR Digital Filter accumulation (((h[0]x[n] h[1]x[n 1]) ) h[]x[n ]) To implement this filter, we must do multiply followed by accumulation (AC). 1

2 Linear oise odel Error is uncorrelated with the input. Error is uniformly distributed over the interval ( / 2) e[n] ( / 2). Error is stationary white noise, (i.e. flat spectrum) P e ( ) e , Roundoff oise in FIR Filters Using the AC instruction with quantized coefficients and quantized input, we can compute (Assume unquantized y[n] h[k]x[n k] input and coefficients) This sum of products can be computed with 32-bit precision; i.e., with no quantization of the partial sums. The result is usually quantized to 15 bits + sign. y ˆ [n] Q 15 y[n] y[n] e[n] The resulting noise power in the output is therefore e 2 2 / /12 FIR Digital Filter with Quantized Output Absolute Scaling an FIR Digital Filter x[n] h[0] h[1] h[2] h[ 1] h[] 32-bit accumulation y[n] Q 15 y ˆ [n] Q 15 y[n] Q 15 h[k]x[n k] y[n] e[n] x[n] h[0] h[1] h[2] h[ 1] h[] y[n] In this case, noise is added directly to the output. e[n] y ˆ [n] y ˆ [n] y[n] h[k]x[n k] y[n] h[k]x[n k] h[k] x[n k] y[n] max x[n] h[k] 1 no overflow 2

3 Sinusoidal Scaling for an FIR Filter Assume that the input is a sinusoid x[n] cos( 0 n) Then the output is y[n] H e j 0 cos 0 n H e j 0 Therefore, if we want y[n] 1, then we must guarantee that He j 1 for all This scaling is appropriate for most narrowband input signals. Conclusions on FIR Coefficient quantization can modify the zero locations and therefore the frequency response. This is usually not severe for linear phase filters Using the AC instruction, we can avoid roundoff (quantization) until the very end of the computation Scaling must be used to avoid overflow IIR Filter Structures and Quantization IIR filters are more complicated with regard to the effects of quantization. any different equivalent structures Coefficient sensitivity may be high Feedback is inherent in IIR structures Possibility of instability Roundoff noise is shaped by filter Some analysis is possible, but given the power of software tools such as ATLAB, an empirical approach is often most effective. Fixed-Point Implementation Issues We need to represent coefficient and signal values by integers in a fixed range. Quantization errors in coefficients imply shifts of poles and zeros (even instability). For a given word-length, the quantization error is fixed in size. Therefore, signal values should be maintained as large as possible to maximize SR. If signal values get too large, additions can overflow (or clip), thereby creating large errors. Thus, fixed-point implementations require careful attention to scaling the signal values. 3

4 Cascade Form b H(z) 0k b 1k z 1 b 2k z 2 s 1 a 1k z 1 a 2k z 2 k1 w k [n] a 1k w k [n 1] a 2k w k [n 2] y k1 [n] y k [n] b 0k w k [n] b 1k w k [n 1] b 2k w k [n 2] Cascade Implementation y 0 [n] x[n], y[n] y s [n] Direct Form II (flow graph) w[n] a k w[n k] x[n] y[n] b k w[n k] k1 Equivalent Direct Form The cascade form groups zeros and poles in pairs (second-order factors). These can be multiplied out to obtain single numerator and denominator polynomials in the direct forms I and II. b H(z) 0k b 1k z 1 b 2k z 2 s 1 a 1k z 1 a 2k z 2 k1 b k z k 1 a k z k k1 4

5 16-Bit Quantized Coefficients Quantized umerator Coefficients ˆ b k b k Quantized Denominator Coefficients ˆ a k a k o Quantization 5

6 16-Bit Quantized Direct Form 16-Bit Quantized Cascade Form Direct Form I IIR Filter Linear oise odel H(z) 1 b k z k a k z k k1 B(z) A(z) y[n] a k y[n k] b k x[n k] k1 Each noise source has power e 2 2 2B /12 The noise sources are independent so their powers add. 6

7 Linear System with a White oise Input x[n] xx [m] x 2 [m] LTI System h[n],h(e j ) y [n] yy [m] x 2 c hh [m] xx (e j ) x 2 yy (e j ) x 2 C hh (e j ) yy [m] xx [m]c hh [m] x 2 [m] c hh [m] x 2 c hh [m] c hh [m] h[m k] h [k] h[m]h [m] k yy (e j ) xx (e j )C hh (e j ) x 2 C hh (e j ) C hh (e j ) H(e j )H (e j ) H(e j ) 2 7

DSP Configurations. responded with: thus the system function for this filter would be

DSP Configurations. responded with: thus the system function for this filter would be DSP Configurations In this lecture we discuss the different physical (or software) configurations that can be used to actually realize or implement DSP functions. Recall that the general form of a DSP