DSP Design Lecture 2. Fredrik Edman.

Size: px

Start display at page:

Download "DSP Design Lecture 2. Fredrik Edman."

Sharon Poole
5 years ago
Views:

1 DSP Design Lecture Number representation, scaling, quantization and round-off Noise Fredrik Edman

2 Representation of Numbers

3 Numbers is a way to use symbols to describe and model the world! However, numbers may be represented in many different ways and there are considerations to make when it comes to digital signal processing!

4 Definition: Radix In a mathematical numeral system, the radix or base is the number of unique digits, including zero, used to represent numbers in a positional numeral system. Example: In the decimal system (the most common system in use today) the radix is 10, because it uses the ten digits from 0 through 9. The number 47 may then be written as: , Base 10

5 Definition: Radix Each position have its on weight depending on the current base = 1*100 + *10 + 5*1 = 1*10 + * *10 0 Representing fractions is a simple extension of this idea = *10 + 5*1 + 4* *0.01 = * * * *10 -

6 l i= k 1 Unsigned Number Representation General description of a fixed radix (base) systems (fixed point positional number system) r i a i a {0,1,,... r 1} r The digits in a radix system: = k 1 k l k 1 k l = r a + r a ra + ra + r a r a described in a fixed point positional number system: a a a a. a a i i l Whole part Fractional part digit set

7 What about numbers in DSP? In DSP, numbers are encoded by means of binary digits or bits {0,1}. For binary digits the base =. All (base 10) numbers can be represented by groups of bits, for example = 1* 6 + 0* 5 + 1* 4 + 0* 3 + 1* + 1* 1 + 0* 0 or =

8 Binary numbers (base=) Representing unsigned integers MSB = Most Significant Bit LSB = Least Significant Bit N bits (0) (1) () (3) (4) (5) (6) (7) In base 10 N words Ex. 101 = 5 10

9 l i= k 1 10 i a i Example: Unsigned Number = { a {0,1,,... 9} in radix 10} k a k l k ak a a a a l = l i= k 1 i a i = { a {0, 1} in radix } k a k l k ai a a a a l = + + +

10 Signed Numbers and Fractions But we also need to represent numbers like and in binary. Devote 1 bit to indicate sign: = = -35. Fractions can be expressed like this = 1* * *0.15 = 1* * - + 1* -3 = However there are also other possibilities!

11 Number Representation Fixed point & Floating point

12 Fixed point & Floating point Digital signal processing number representation can be separated into two categories - fixed point and floating point. These designations refer to the format used to store and manipulate numeric representations of data. Fixed-point DSPs are designed to represent and manipulate integers positive and negative whole numbers. Floating-point DSPs represent and manipulate rational numbers.

13 Fixed point In fixed point we only have a single value (n) In fixed-point representation, a specific radix point (decimal point in English. ) is chosen so there is a fixed number of bits to the right (fractional bits) and a fixed number of bits to the left (integer bits) of the radix point.

14 Floating point In floating point a value is represented by mantissa (m) determining the resolution/precision exponent (e) determining the dynamic range m b e In floating-point representation, the placement of the decimal point can float relative to the significant digits of the number.

15 Floating vs. Fixed point - Design Considerations With fixed-point notation, the gaps between adjacent numbers are always equal a value of one, whereas in floating-point notation, gaps between adjacent numbers are not uniformly spaced (ANSI/IEEE Std. 754 standard format), with large gaps between large numbers and small gaps between small numbers. Floating point gives higher dynamic range but often high cost in: energy area calculation time For energy efficient implementations fixed point is preferred, however fixed-point numbers are more susceptible to common numeric computational inaccuracies. When choosing number representation you also need to consider - Cost - Performance - Ease of implementation/development - Cell library?

16 There are different fixed point numbering systems! Sign Magnitude One s Complement Two s Complement Etc.

17 Signed Magnitude Unsigned numbers with a sign-bit (MSB = sign bit) Signed Magnitude Two Zeros + Low Power? + Easy to convert to Negative (flip bit) 100

18 One s Complement Signed numbers by inverting (Complement) One's Complement Two Zeros + Easy to convert to Negative (invert)

19 Two s Complement Most widely used fixed point numbering system Two's Complement Complement + LSB + One Zero + Easy Add/Sub - Not so easy to convert to Neg Add operation for negative numbers is identical to the add operation for positive numbers, so no additional logic is required to handle the negative case.

20 Table of number representations Wikipedia

21 Considerations Much to consider when choosing between fixed point or floating point representaion. Type of application, development time, cost, tools, etc. will influence your choice! Choosing different fixed point numbering systems will also influence your design!

22 Quantization Ch in Phari Overview of Effects of Finite Register Length in DSP (download from homepage)

23 What is Quantization? Quantization, in mathematics and in DSP, is the process of mapping a large set of input values to a (countable) smaller set. The operations rounding and truncation are typical examples of quantization processes. Quantization is involved to some degree in nearly all digital signal processing, as the process of representing a signal in digital form ordinarily involves rounding.

24 Quantization Mainly two types: Signal Quantization Coefficient Quantization

25 Signal Quantization x s [n] sampled waveform x a (t) T = sample period 3T T T 0 T T 3T 4T time samples, n x a (t) S&H x s [n] x[n] ˆ xˆ B [n] Quantizer Coder T A/D

26 Signal Quantization

27 Different quantization techniques can be applied Level Linear Quantization: Same distance between quantization levels indepentent of input voltage, i.e. quantization error is the same for all input signal. Input Voltage What happens with the SNR? Higher SNR for high signal levels.

28 Different quantization techniques can be applied Level Non-Linear Quantization: Unequal distance between quantization, e.g. closer levels at lower signal levels. A fixed SNR can be achieved or higher SNR in most important areas. Input Voltage More complex quantization process

29 Quantization effects Round-off Noise Affect the output as a random disturbance Limit Cycle Oscillations Undesired periodic components due to non-linear behavior in the feedback (rounding or overflow)

30 Coefficient Quantization IIR filter x(n) + b 0 + y(n) + + a 1 a n-1 a n Z -1 Z -1 Z -1 b 1 b m-1 b m + + Filter becomes unstable! Non-ideal transfer function!

31 Rounding & Truncation Rounding/Truncation is always there! Especially necessary in recursive systems Q x(n) y(n) Without quantization - infinite wordlength Multiplication n+m output bits Addition n+1 output bits

32 Truncation & Rounding Level X+1 Level X+1 Level X Level X Truncation All values approximated in the same direction Max error = 1LSB Rounding Values approximated up or down Max error = 1/ LSB

33 Different rounding schemes y round down (towards ) round up (towards + ) round towards zero round away from zero round to nearest

34 Rounding & Truncation No energy added to the system Often used in recursive algorithms Truncation Rounding Truncation towards zero In In In DC error All values goes towards -infinity 100 Rounded to even 100 Add LSB before truncation if negative

35 Scaling Scaling is about adjusting the signal range to fit the hardware. Unchanged transfer function (scaled coefficients might move the pole-zeros) However you might loose in precision!

36 An example where scaling is needed x(n) y(n) x(n) 1 x( n) = ± 3 4 y(n) 1 1 x, y < Overflow 1 y(n) 1

37 Scaling to avoid overflow x(n) f(n) y(n) Safe scaling if x(n) 1 β f(n) β y(n) β = i= 0 f ( i) ( l 1 scaling in the book) Where f(i) is the unit sample response

38 Example 1: Safe Scaling i= 0 i= 0 ( 05) β = f(i) =. = = = (. ) (. ) (. ) i Geometric series u(n) x(n) u(n) x(n) Original filter with overflow 0.5

39 Example : Safe Scaling = 7 x( n) and y( n) 7 give safe scaling β = i 0 f(i) = = 3.5 x(n) / y(n) h(n) Increased roundoff noise Internal scaling might improve 7/ (Linear phase FIR)

40 Scaling Safe scaling is pessimistic Alternative is scaling with i= In practice: Scaling with β = ±n Easy to do - a shift β = 0 ( f () i ) (l scaling in the book) Increased internal wordlength an alternative However, overflow may occur!!

41 Summary Scaling Techniques l 1 - scaling β = f ( i), Safe scaling i= 0 l - scaling β = δ i= 0 f i possible overflow ( ), ( ) = unit sample response, ( ) = Variance white noise input i= 0 f i f i Safe scaling but not guaranteed δ sets the probability for an overflow Typically one overflow every 10 6 sample is accepted in audio [Wanhammar]

42 Limit Cycles

43 Limit Cycles Oscillations Limit cycle oscillations can be described as the undesirable periodic components at the output due to the fact that quantization is a nonlienear operation.

44 Limit Cycles Oscillations Example: zero input oscillations in nd order IIR Q b Q b 1 X(n) y(n) b = = ; b = =

45 Limit Cycles Zero Input Example: zero input oscillations Rounding after multiplication Truncation after multiplication Source: Lars Wanhammar, DSP Integrated circuits

46 Limit Cycles Zero input oscillations Not accepted in some applications such as in audio applications. Very difficult problem Real Part In general, no solutions for structures > nd order Can be limited by increased internal wordlength Can in some nd order structures be eliminated by pole positioning nd order Wave Digital Filters are free from parasitic oscillations Imaginary Part Poles close to the unity circle Matlab: zplane(1,[ ])

47 We may also get Overflow Oscillations 3-bit two s complement sum Correct sum Oscillations are limited by saturation 3-bit saturated sum Correct sum Overflow change the sign

48 Limit Cycles due to overflow Zero Input Two s Complement Arithmetic Saturated Arithmetic Source: Lars Wanhammar, DSP Integrated circuits

49 Saturation Arithmetic From Adder C out-msb 0 = NOF C out-msb C in-msb C in-msb Signbit 1 = POF Saturated Output Overflow if C out-msb differs from C in-msb

50 Simple Noise Analysis

51 Quantization Analysis Using real rounding, truncation, and overflow Gives exact result Tricky - need integer representation Using noise models Floating point representation can still be used Suitable for Matlab, C/C++...

52 Wordlength (W) = 8 bits (W-1) 15 bits a 8 bits Rounding Model a Q e(n) u(n) 8 bits y(n) u(n) y(n)

53 Wordlength (W) = 8 bits 15 bits a 8 bits Rounding Model a Q u(n) 8 bits y(n) u(n) ( ) u n y(n) e(n) ( ) en ( ) = u n un ( ) Modeled with added noise as an input error

54 Analysis of Round-off Noise We need to analyze the round-off noise to determine its effect on the output signal! If the noise variance is not negligible in comparison with the output signal level the wordlength must be increased or some low-noise structure must be used!

55 Roundoff Noise If the quantization error probability is uniformly distributed in the interval e( n) where = ( W 1) W is the number of bits after the rounding P e (e), ½ LSB Probability density function 1 e

56 Statistical model of e(n) The statistical model is based on the following assumptions: 1. e(n) is a sample sequence of a stationary random process.. e(n) is uncorrelated with x(n) 3. The error is a white noise process 4. The probability distribution of the error process is uniform over the range of the quantization error. It is easy to find situtations where this is not valid, e.g. if x(n) is a square wave. However, when x(n) is complicated, e.g. speech or music, the assumptions are realistic.

57 Roundoff Noise / ( ) ( ) / Mean value e = E[ e( n)] = 0 Variance = E[ e e ] = e e P ( e) de = [ e = 0] / 3 / W 1 e 1 /8 /8 e de = = = = / / P e (e) e ½ LSB 1 e

58 Roundoff Noise / ( ) ( ) Variance = E[ e e ] = e e P ( e) de = [ e = 0] / Mean value e = E[ e( n)] = 0 / 3 / W 1 e 1 /8 /8 e de = = = = / / P e (e) e ½ LSB 1 e

59 Roundoff Noise / ( ) ( ) Variance = E[ e e ] = e e P ( e) de = [ e = 0] / Mean value e = E[ e( n)] = 0 / 3 / W 1 e 1 /8 /8 e de = = = = / / P e (e) e ½ LSB 1 e

60 Roundoff Noise / ( ) ( ) Variance = E[ e e ] = e e P ( e) de = [ e = 0] / Mean value e = E[ e( n)] = 0 / 3 / W 1 e 1 /8 /8 e de = = = = / / P e (e) e ½ LSB 1 e

61 Example: Roundoff Noise In the case of rounding (mean=0) the variance and the average power are the same, i.e. if a value is rounded the quantization noise becomes: σ e = W 3 If we scale down one bit: ( W 1) W = = σ e

62 Analysis of Round-off Noise Not only the noise gain at the output needs to be determined! We also need to determine the SNR at the output to get the full picture!

63 Signal to Noise Ratio (SNR) Signal power (variance) SNR = σ x e 3 10log = 10log σ W x σ Roundoff error power (variance)

64 Signal to Noise Ratio (SNR) One extra bit reduces quantization error by a factor 4 4 σ SNR = 10log σ e = 6. 0 e db Good to remember: 6 db increase in SNR per bit

65 Signal to Noise Ratio (SNR) Example: Full scale sinus wave rounded to 8 bits SNR 3 A = 10log 50 db; -1 A 1 8 =

66 Roundoff Noise: Addition E[( e ) 1+ e ] = Ee [ 1 + ee 1+ e] = 1 = Ee [ ] + E[ ee 1] + Ee [ ] = zero if u1and u independent u (n) 1 Ee = Ee [ ] + [ ] u 1 (n) e (n) e 1 (n) y(n)

67 A larger IIR-filter with several quantization noise sources (I) Oppenheim and Schafer, Discrete-time Signal Processing, Prentice Hall

68 A larger IIR-filter with several quantization noise sources (II) Oppenheim and Schafer, Discrete-time Signal Processing, Prentice Hall

69 A larger IIR-filter with several quantization noise sources (III) Assumption: noise sources are independent of (a) the input and (b) each other can be added σ = σ + σ + σ + σ + σ e e0 e1 e e3 e4 Oppenheim and Schafer, Discrete-time Signal Processing, Prentice Hall

70 End of Lecture

Numbering Systems. Computational Platforms. Scaling and Round-off Noise. Special Purpose. here that is dedicated architecture

Numbering Systems. Computational Platforms. Scaling and Round-off Noise. Special Purpose. here that is dedicated architecture Computational Platforms Numbering Systems Basic Building Blocks Scaling and Round-off Noise Computational Platforms Viktor Öwall viktor.owall@eit.lth.seowall@eit lth Standard Processors or Special Purpose