DIGIT-SERIAL ARITHMETIC

Size: px

Start display at page:

Download "DIGIT-SERIAL ARITHMETIC"

Barnard Gardner
6 years ago
Views:

1 DIGIT-SERIAL ARITHMETIC 1 Modes of operation:lsdf and MSDF Algorithm and implementation models LSDF arithmetic MSDF: Online arithmetic

2 TIMING PARAMETERS 2 radix-r number system: conventional and redundant Serial signal: a numerical input or output with one digit per clock cycle For an n digit signal, the clock cycles are numbered from 1 to n Timing characteristics of a serial operation determined by two parameters: initial delay δ: additional number of operand digits required to determine the first result digit execution time T n ; the time between the first digit of the operand and the last digit of the result T n = δ + n + 1

3 3 cycle: input compute output δ = 0 T 12 = (a) cycle: input compute output δ =3 T 12 = (b) Figure 9.1: Timing characteristics of serial operation with n = 12. (a) With δ = 0. (b) With δ = 3.

4 LSDF AND MSDF 4 1. Least-significant digit first (LSDF) mode (right-to-left mode) x = n 1 x ir i i=0 2. Most-significant digit first (MSDF) mode (left-to-right mode) also known as online arithmetic initial delay called online delay x = n i=1 x ir i

5 ALGORITHM MODEL 5 Operands x and y, result z: n radix-r digits In cycle j the result digit z j+1 is computed Cycles labeled δ,..., 0, 1,..., n In cycle j receive the operand digits x j+1+δ and y j+1+δ, and output z j x[j], y[j] and z[j] are the numerical values of the corresponding signals when their representation consists of the first j + δ digits for the operands, and j digits for the result. In iteration j x[j + 1] = (x[j], x j+1+δ ) y[j + 1] = (y[j], y j+1+δ ) z j+1 = F(w[j], x[j], x j+1+δ, y[j], y j+1+δ, z[j]) z[j + 1] = (z[j, ], z j+1 ) w[j + 1] = G(w[j], x[j], x j+1+δ, y[j], y j+1+δ, z[j], z j+1 )

6 Cycle j j+1 6 Input x j+1+δ y j+1+δ x[j+1] x j+2+δ y j+2+δ x[j+2] Compute z j+1 z j+2 Output z j z j+1 (a) x j+1+δ y j+1+δ X; Y x[j] y[j] z j+1 F; G z z j j+1 Z z[j] w[j+1] W w[j] (residual) z j (b) digit-serial digit-parallel

7 7 Table 9.1: Initial delay (δ) Operation LSDF MSDF Addition 0 2 (r = 2) 1 (r 4) Multiplication 0 3 (r = 2) 2 (r = 4) (only MS half) n Division 2n 4 Square root 2n 4 Max/min n 0

8 8 a) LSDF mode n-digit addition: Cycle: LSD MSD Inputs: x x x x x x x x x Output: x x x x x x x x x n by n --> 2n multiplication: Inputs: Output: LSD MSD x x x x x x x x x x x x x x x x x x x x x x x x x x x MS half b) MSDF mode Cycle: n-digit operation: MSD LSD Inputs: x x x x x x x x x Output: --- online delay = 2 x x x x x x x x x Figure 9.3: LSDF and MSDF modes.

9 COMPOSITE ALGORITHM: EXAMPLE 9 Givens method for matrix triangularization Rotation factors: c = The online delay of the network x x2 + y 2 s = y x2 + y 2 rot = δ1 + δ2 + δ3 + δ4 = 13 Execution time (latency): D rot = rot + n + 4

10 10 x h y h Operation: δ1 (on-line delay) OLSQ a i i = h-δ1 1 OLSQ b i δ1 Squaring δ2 OLADD Addition (p-bit shift register) f k k=i δ2 1 =p =p δ3 OLSQRT Square root δ4 OLDIV g p p=k δ3 1 OLDIV δ4 Division c q q=p δ4 1 s q (a) x,y a,b f g c, s δ1 δ2 δ3 δ4 (b) Figure 9.4: Online computation of rotation factors: (a) Network. (b) Timing diagram.

11 LSDF ARITHMETIC: ADDITION/SUBTRACTION 11 The cycle delay t LSDFadd k = t CPA(k) + t FF The total time for n-bit addition T LSDFadd n = ( n k + 1)t LSDFadd k The cost: one k-bit CPA, one flip-flop, and one k-bit output register

12 Operand X: Operand Y: x z i Result Z: i y k z i z k k-bit k k i-1 k CPA result digit register SUB 12 (a) c carry/borrow FF (initialize to 0 if ADD 1 if SUB) Operand Y: Operand X: Result Z: y i.0 x i,0 z i,0 z z i-1,0 y i,1 x i,1 y i,2 x i,2 4-bit ADDER z i,1 z i,2 z z z i-1,1 z i-1,2 y i,3 x i,3 z i,3 z z i-1,3 SUB carry/borrow FF (initialize to 0 if ADD 1 if SUB) c

13 LSDF MULTIPLICATION 13 For radix-2 and 2 s complement representation: 1. Serial-serial (LSDF-SS) multiplier, both operands used in digit-serial form. 2. Serial-parallel ( LSDF-SP) multiplier, one operand first converted to parallel form Operation cannot be completed during the input of operands

14 SERIAL-SERIAL MULTIPLIER 14 Define internal state (residual) w[j] = 2 (j+1) (x[j] y[j] p[j]) where x[j] = j i=0 x i 2 i and similarly for y[j] and p[j]. Both operands used in serial form; the recurrence is w[j + 1] = 2 (j+2) (x[j + 1] y[j + 1] p[j + 1]) = 2 (j+2) ((x[j] + x j+1 2 j+1 )(y[j] + y j+1 2 j+1 ) (p[j] + p j+1 2 j+1 )) = 2 1 (w[j] + y[j + 1]x j+1 + x[j]y j+1 p j+1 ) This can be expressed as v[j] = w[j] + y[j + 1]x j+1 + x[j]y j+1 and w[j + 1] = 2 1 v[j] p j+1 = v[j]mod 2

15 Position Cycle y 0 x 0 1 x 0 y 1 y 1 x 1 y 0 x 1 2 x 1 y 2 x 0 y 2 y 2 x 2 y 1 x 2 y 0 x 2 3 x 2 y 3 x 1 y 3 x 0 y 3 y 3 x 3 y 2 x 3 y 1 x 3 y 0 x 3 4 x 3 y 4 x 2 y 4 x 1 y 4 x 0 y 4 y 4 x 4 y 3 x 4 y 2 x 4 y 1 x 4 y 0 x 4 15

16 16 (shift-register for load control in left-append registers not shown) x j+2 x j+1 LA-Reg X y j+2 LA-Reg Y y j+1 x j+1 n w[j+1]: x[j] n SELECTOR n x[j] n shifted WC Reg WC [4:2] ADDER n x j+1 n+1 y j+1 n shifted WS Reg WS y[j+1] n SELECTOR n sign(x) sign(y) FA y[j+1] cycle n LS bits p j+1 carry-out w[j]: n n SA MS bits (register control signals not shown) SA - Serial adder Figure 9.6: Serial-serial 2 s complement radix-2 multiplier.

17 cont. 17 The total execution time T SSMULT = 2nt cyc The delay of the critical path in a cycle t cyc = t SEL + t 4 2CSA + t FF Cost: one n-bit [4:2] adder, 5 n-bit registers, and gates to form multiples

18 SERIAL-PARALLEL MULTIPLIER 18 One of the operands is a constant, One possibility: perform operation in 3n cycles Phase 1: Serial input and conversion of one operand to parallel form; Phase 2: Serial-parallel processing and output of the LS half of the product. Phase 3: Serial output of the MS half of the product. The critical path in a cycle t cyc = t SEL + t CSA + t FF The delay of the LSDF-SP multiplier T SPrnd = 3n t cyc

19 (shifted -in in Phase 1 or constant) x j (used serially in Phase 2) y j n n shifted WC Reg WC Shift-Reg X x SELECTOR n n [3:2] ADDER n+1 w[j+1] n x { 0, x, x } shifted WS Reg WS sign(y) HA cycle n in Phase 2 (shifted -out in Phase 2) LS bits carry-out 19 n w[j] n SA (shifted -out in Phase 3) MS bits (register control signals not shown) SA - Serial adder Phase 1: shift-in operand X (n cycles) Phase 2: serial-parallel carry-save multiplication (n cycles) shifted sum and carry bit-vectors loaded bit-parallel Phase 3: MS bits obtained using bit-serial adder SA operating on bits shifted out of WC and WS shift-registers (n cycles) Figure 9.7: 3-phase serial multiplier.

20 MSDF: ONLINE ARITHMETIC 20 Online arithmetic algorithms operate in a digit-serial MSDF mode To compute the first digit of the result, δ + 1 digits of the input operands needed Thereafter, for each new digit of the operands, an extra digit of the result obtained The online delay δ typically a small integer, e.g., 1 to 4. Cycle Input x 1 x 2 x 3 x 4 x 5 Compute z 1 z 2 z 3 Output z 1 z 2 δ = 2 Figure 9.8: Timing in online arithmetic.

21 REDUNDANT REPRESENTATION 21 The left-to-right mode of computation requires redundancy Both symmetric { a,..., a} and asymmetric {b,..., c}) digit-sets used in online arithmetic Over-redundant digit-sets also useful Examples of radix-4 redundant digit sets: {-1,0,1,2,3} (asymmetric, minimally-redundant), {-2,-1,0,1,2} (symmetric, minimally-redundant), and {-5,-4,-3,-2,-1,0,1,2,3,4,5} (symmetric, over-redundant) Heterogeneous representations to optimize the implementation Conversion to conventional representation: use on-the-fly conversion

22 ADDITION/SUBTRACTION 22 The online addition/subtraction algorithm: the serialization of a redundant addition (carry-save or signed-digit) Radix r > 2 and (t j+1, w j+2 ) = where x j, y j, z j { a,..., a}. (0, x j+2 + y j+2 ) if x j+2 + y j+2 a 1 (1, x j+2 + y j+2 r) if x j+2 + y j+2 a ( 1, x j+2 + y j+2 + r) if x j+2 + y j+2 a z j+1 = w j+1 + t j+1

23 EXAMPLE OF ONLINE ADDITION (r = 4, a = 3) 23 Operands x = ( ) The result z = ( ). y = ( ) j x j+2 y j+2 t j+1 w j+2 w j+1 z j+1 z j * 1 0* * latches initialized to 0.

24 RADIX r > 2 ONLINE ADDER 24 x j y j x j+1 y j+1 x j+2 y j+2 x j+2 y j+2 TW TW TW TW t j-1 w j t j w j+1 t j+1 w j+2 t j+2 t j+1 w j+2 latch w j+1 SUM SUM SUM SUM z j+1 z j z j+1 z j+2 z j (a) (b) Figure 9.9: (a) A segment of radix-r > 2 signed-digit parallel adder. (b) Radix-r > 2 online adder. All latches cleared at start.

25 RADIX-2 ONLINE ADDER 25 Digit-parallel radix-2 signed-digit adder converted into a radix-2 online adder with online delay δ = 2 x + x - y + y - j+1 j+1 j+1 j+1 x + x - j+2 j+2 y + y - j+2 j+2 x + y + y - j+3 x- j+3 j+3 j+3 x + j+3 x- j+3 y + y - j+3 j+3 FA FA FA FA h j g j+1 h j+1 g j+2 h j+2 g j+3 hj+3 h j+2 g j+3 g j+2 FA FA FA FA t j w j+1 t j+1 w j+2 t j+2 w j+3 t j+3 w j+2 latch z + z + z - z + z - j+1 z - j+1 j+2 j+2 j+3 j+3 z - j+1 = t j+1 z + = w j+1 j+1 (output latches) (a) (b) z - j z + j Figure 9.10: (a) A segment of radix-2 signed-digit parallel adder. (b) Online adder.

26 cont. 26 The cycle time is t cyc = 2t FA + t FF The operation time T OLADD 2 = (2 + n + 1)t cyc The cost 2 FAs and 5 FFs. To reduce the cycle time, pipeline the two stages: reduces the cycle time by one t FA ; increases online delay to δ = 3

27 EXAMPLE OF RADIX-2 ONLINE ADDITION 27 x = ( ) y = ( ) z = ( ) j x j+3 y j+3 x + j+3x j+3 y j+3y + j+3 h j+2 g j+3 g j+2 t j+1 w j+2 z j+1z + j+1 z j * * g latches initialized to 00.

28 METHOD FOR DEVELOPING ONLINE ALGORITHMS 28 Part 1: development of the residual recurrence w[j + 1] = G(w[j], x[j], x j+1+δ, y[j], y j+1+δ, z[j], z j+1 ) for δ j n 1 where x[j] = j+δ x ir i, y[j] = j+δ i=1 y ir i, z[j] = j z ir i i=1 i=1 are the online forms of the operands and the result Part 2: the result digit selection z j+1 = F(w[j], x[j], x j+1+δ, y[j], y j+1+δ, z[j])

29 Part 1: RESIDUAL AND ITS RECURRENCE 29 Step 1. Describe the online operation by the error bound after j digits Step 2 Transform expression to use only multiplication by r (shift), addition/subtraction, multiplication by a single digit f(x[j], y[j]) z[j] < r j B < G(f(x[j], y[j]) z[j]) < B where G is the required transformation and B and B are the transformed bounds Example: division error expression x[j]/y[j] z[j] < r j transformed into x[j] z[j] y[j] < r j y[j]

30 cont. 30 Step 3 Define a scaled residual with the bound w[j] = r j (G(f(x[j], y[j]) z[j])) ω = r j B < w[j] < r j B = ω and initial condition w[ δ] = 0. ω and ω are the actual bounds determined in Step 6 Step 4 Determine a recurrence on w[j] w[j+1] = rw[j]+r j+1 (G(f(x[j+1], y[j+1]) z[j+1]) G(f(x[j], y[j]) z[j])) Step 5 Decompose recurrence so that H 1 is independent of z j+1 w[j + 1] = rw[j] + H 1 + H 2 (z j+1 ) = v[j] + H 2 (z j+1 )

31 cont. 31 Step 6 Determine the bounds of w[j + 1] in terms of H 1 and H 2 ω = rω + max(h1) + H2(a) resulting in Similarly, ω = max(h 1) + H 2 (a) r 1 ω = min(h 1) + H 2 ( a) r 1

32 Part 2a: SELECTION FUNCTION WITH SELECTION CONSTANTS 32 z j+1 = k if m k v[j] < m k+1 where v[j] is an estimate of v[j] obtained by truncating the redundant representation of v[j] to t fractional bits. Selection constants need to satisfy max( L k ) m k min(ûk 1) where [ L k, Ûk] is the selection interval of the estimate v[j] Step 7 Determine [ L k, Ûk] First, determine [L k, U k ] for v[j] ω = U k + H 2 (k) ω = L k + H 2 (k) Substituting ω and ω the selection intervals for v[j] is, U k = max(h 1)+H 2 (a) r 1 H 2 (k) L k = min(h 1)+H 2 ( a) r 1 H 2 (k)

33 cont. 33 Now restrict the intervals because of the use of the estimate v[j] e min v[j] v[j] e max producing the error-restricted selection interval [L k, Uk] with Uk = U k e max L k = L k + e min The errors are For carry-save representation e max = 2 t+1 ulp and e min = 0. For signed-digit representation e max = 2 t ulp and e min = (2 t ulp). U k 1 = Uk t t = L k t L k where x t and x t indicate x values truncated to t fractional bits.

34 34 L k U k-1 2 -t - possible choices for m k 2 -t+1 v[j] L k U * k-1 U k-1 L * k (the ticks on the v[j] line represent the estimate v[j]) Figure 9.11: The choices of selection constant m k.

35 cont. 35 Step 8 Determine t and δ. To determine m k, we need min(ûk 1) max( L k ) 0 This relation between t and δ is used to choose suitable values. Step 9 Determine the selection constants m k and the range of v[j] as rω + min(h 1 ) e max t v[j] rω + max(h 1 ) + e min t

36 Part 2b: SELECTION BY ROUNDING 36 In algorithms using a higher radix (r > 4) w[j + 1] = rw[j] + H 1 + H 2 (z j+1 ) = v[j] + H 2 (z j+1 ) In the rounding method, the result digit is obtained as z j+1 = v[j] with v[j] < r 1 2 to avoid over-redundant output digit. w[j + 1] = v[j] + H 2 ( v[j] ) For CS form of t fractional bits, the estimate error e max = 2 t+1 ulp When ˆv[j] = m k 2 t it must be possible to choose z j+1 = k 1 m k 2 t + e max = 2k t Ûk 1

37 GENERIC FORM OF EXECUTION AND IMPLEMENTATION. 37 Execution: n + δ iterations of the recurrence, each one clock cycle Iterations (cycles) labeled from δ to n 1 One digit of each input introduced during cycles δ to n 1 δ and digits value 0 thereafter Result digits 0 for cycles δ to 1 and z 1 is produced in cycle 0 Result digit z j is output in cycle j (one extra cycle to output z n )

38 cont. 38 The actions in cycle j: Input x j+1+δ and y j+1+δ. Update x[j+1] = (x[j], x j+1+δ ) and y[j+1] = (y[j], y j+1+δ ) by appending the input digits. Compute v[j] = rw[j] + H 1 Determine z j+1 using the selection function. Update z[j + 1] = (z[j], z j+1+δ ) by appending the result digits. Compute the next residual w[j + 1] = v[j] + H 2 (z j+1 ) Output result digit z j

39 IMPLEMENTATION 39 Similar structure of algorithms all implemented with same basic components, such as (i) registers to store operands, results, and residual vectors; (ii) multiplication of vector by digit; (iii) append units to append a new digit to a vector; (iv) Two-operand and multioperand redundant adders, such as signed digit adders, [3:2] carry-save adders and their generalization to [4:2] and [5:2] adders; (v) converters from redundant representations (i.e., signed digit and carry save) to conventional representations; (vi) carry-propagate adders of limited precision (3 to 6 bits) to produce estimates of the residual functions; and (vii) digit-selection schemes to obtain output digits.

40 cont. 40 Online algorithm implementation similar to implementation of digit-recurrence algorithms Algorithms and implementations developed for most of basic arithmetic operations and for certain composite operations Larger set of operations possible than with LSDF approach

41 DIGIT-SLICE ORGANIZATION 41 x j+1+δ y j+1+δ * 1*** 2 m z j+1 z j digit slice ** * paths for appending input digits ** left-shifted bits of the residual *** the width of the MS slice depends on the selection function Figure 9.12: A typical digit-slice organization of online arithmetic unit

42 ONLINE MULTIPLICATION 42 Online forms The error bound at cycle j The residual x[j] = j+δ x ir i, y[j] = j+δ i=1 with the bound w[j] < ω The residual recurrence y ir i, p[j] = j p ir i i=1 i=1 x[j] y[j] p[j] < r j w[j] = r j (x[j] y[j] p[j]) w[j + 1] = rw[j] + (x[j]y j+1+δ + y[j + 1]x j+1+δ )r δ p j+1 = v[j] p j+1

43 SELECTION FUNCTION 43 Decomposition Bound Selection intervals H 1 = (x[j]y j+1+δ + y[j + 1]x j+1+δ )r δ H 2 = p j+1 ω = ω = ω = ρ(1 2r δ ) U k = ρ(1 2r δ ) + k L k = ρ(1 2r δ ) + k With carry-save representation for w[j] and v[j], the grid-restricted intervals are U k = ρ(1 2r δ ) + k 2 t t L k = ρ(1 2r δ ) + k t The expression to determine t and δ: resulting in ρ(1 2r δ ) + k 1 2 t t ρ(1 2r δ ) + k t 0 ρ(1 2r δ ) t 2 1 (1 + 2 t )

44 cont. 44 Several examples of relations between r, ρ, t, and δ Radix ρ t δ / /3 2 3

45 RADIX-2 ONLINE MULTIPLICATION 45 δ = 3 and t = 2 Selection constants m k s obtained from where L k m k Ûk 1 U k = k = k L k = k 2 = k Since Ûk 1 = k 2 1 and L k = k 3 2 2, m k = k 2 1 is acceptable. The selection constants are Range of v[j] is m 0 = 2 1, m 1 = v[j] 7/4 The selection function SELM( v[j]) is p j+1 = SELM( v[j]) = 1 if 1/2 v[j] 7/4 0 if 1/2 v[j] 1/4 1 if 2 v[j] 3/4

46 IMPLEMENTATION OF SELECTION FUNCTION 46 Estimate ˆv represented by (v 1, v 0, v 1, v 2 ) Product digit p j+1 = (pp, pn) with the code p j+1 pp pn Switching expressions: pp = v 1(v 0 v 1 ) pn = v 1 (v 0 v 1)

47 ˆv v 1 v 0 v 1 v 2 p j+1 7/ / / / / / / / / / / /

48 48 1. [Initialize] x[ 3] = y[ 3] = w[ 3] = 0 for j = 3, 2, 1 x[j + 1] CA(x[j], x j+4 ); y[j + 1] CA(y[j], y j+4 ) v[j] = 2w[j] + (x[j]y j+4 + y[j + 1]x j+4 )2 3 w[j + 1] v[j] end for 2. [Recurrence] for j = 0...n 1 x[j + 1] CA(x[j], x j+4 ); y[j + 1] CA(y[j], y j+4 ) v[j] = 2w[j] + (x[j]y j+4 + y[j + 1]x j+4 )2 3 p j+1 = SELM( v[j]); w[j + 1] v[j] p j+1 P out p j+1 end for Figure 9.13: Radix-2 online multiplication algorithm.

49 predecessor on-line unit (shift-register for load control in right-append registers not shown) predecessor on-line unit 49 x j+5 LX CA-Reg X y j+5 LY CA-Reg Y n x[j] n x[j] x j+4 y j+4 n y[j+1] n y[j+1] y j+4 SELECTOR x j+4 SELECTOR n n n+2 [4:2] ADDER v[j1] n+2 c x =1 if x j+4 < 0 c y =1 if y j+4 < 0 V 4 4 SELM 3 4 v n-2 n-2 p j+1 M wired shift left Pout 3 2w[j+1] p j Reg WS Reg WC v[j] vs -1 vs 0. vs 1 vs 2 vs 3 vs vc -1 vc 0. vc 1 vc 2 vc 3 vc n+2 n+2 2w[j] V block produces estimate of v M block performs subtraction of p j+1 (register control signals not shown) (a) v -1 v 0. v 1 v estimate of v[j] 2 v * o v 1. v 2 vs vs w[j+1] vc 3 vc 4... v * o = v 0 XOR p j+1 (b) Figure 9.14: (a) Implementation of radix-2 online multiplier. (b) Calculation of 2w[j + 1].

50 EXAMPLE OF RADIX-2 ONLINE MULTIPLICATION 50 Operands: x = ( ) y = ( ) j x j+4 y j+4 x[j + 1] y[j + 1] v[j] p j+1 w[j + 1]

51 cont. 51 Computed product: p = ( ) The exact double precision product p = ( ) The absolute error wrt to the exact product truncated to 8 bits: p p tr = 2 8 Note: p[8] + w[8]2 8 = p

52 ONLINE DIVISION 52 Online forms x[j] = j+δ x ir i, y[j] = j+δ i=1 y ir i, q[j] = j q ir i i=1 i=1 Error bound at cycle j x[j] q[j]d[j] < d[j]r j Residual w[j] = r j (x[j] q[j]d[j]) w[j] < ω d[j] Residual recurrence w[j + 1] = rw[j] + x j+1+δ r δ q[j]d j+1+δ r δ d[j + 1]q j+1 = v[j] d[j + 1]q j+1

53 RADIX-2 ONLINE DIVISION 53 δ = 4 and t = 3 Selection intervals and selection constants minû0 = Û0[d[j + 1] = 1/2] = = 2 2 max L 1 = L 1 [d[j + 1] = 1] = = 2 3 resulting in m 1 = 2 2 minû 1 = Û 1[d[j + 1] = 1] = = 2 2 max L 0 = L 0 [d[j + 1] = 1/2] = = so that m 0 = 2 2. q j+1 = SELD( v[j]) = 1 if 1/4 v[j] 15/8 0 if 1/4 v[j] 1/8 1 if 2 v[j] 1/2

54 RADIX-2 ONLINE DIVISION: ALGORITHM [Initialize] x[ 4] = d[ 4] = w[ 4] = q[0] = 0 for j = 4,..., 1 d[j + 1] CA(d[j], d j+5 ) v[j] = 2w[j] + x j w[j + 1] v[j] end for 2. [Recurrence] for j = 0...n 1 d[j + 1] CA(d[j], d j+5 ) v[j] = 2w[j] + x j q[j]d j q j+1 = SELD( v[j]); w[j + 1] v[j] q j+1 d[j + 1] q[j + 1] CA(q[j], q j+1 ) Q out q j+1 end for

55 (shift-register for load control in right-append registers not shown) predecessor on-line unit 55 q j+1 CA-Reg Q d j+6 LD CA-Reg D x j+6 q s x j+5 LX (single digit) predecessor on-line unit n q[j] n d j+5 d j+5 SELECTOR U u all 6 bits wide n 5 q[j] v[j] [3:2] ADDER ws q j+1 wc n d[j+1] n d[j+1] SELECTOR n c d = 1 if (d j+5 > 0 and q>0) or (d j+5 < 0 and q<0) 5 SELD 4 V v [3:2] ADDER c q = 1 if (q j+1 > 0 and d>0) or (q j+1 < 0 and d<0) q j+1 wired left shift Qout 2w[j+1] q j (register control signals not shown) Reg WS Reg WC n+2 n+2 2w[j] ws wc Figure 9.16: Block diagram of radix-2 online divider.

56 REDUCTION OF DIGIT-SLICES 56 Selection valid if (t fractional bits) p + h = n + δ p 2h + δ t p = 2n + δ + t 3 Total number of bit-slices: ib + p, ib - no. integer bits For example, the number of bit-slices for 32-bit radix-2 online multiplication is = = 25 3 compared to 34 in implementation without slice reduction.

57 57 δ not implemented ib t n p (a) after left shift: p x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x e x x x x x e e x x x x x e e x x x x e e e (b) Figure 9.17: Reduction of bit-slices in implementation.

58 MULTI-OPERATION AND COMPOSITE ONLINE ALGORITHMS 58 To reduce the overall online delay of a group of operations - combine several operations into a single multi-operation online algorithm Example: x 2 + y 2 + z 2 Inputs in [1/2,1), output in [1/4, 3) Online delay δ ss = 0 when the output digit is over-redundant. Online delay (3+2+2=7) of the corresponding network

59 ALGORITHM FOR SUM OF SQUARES [Initialize] w[0] = x[0] = y[0] = z[0] = 0 2. [Recurrence] for j = 0...n 1 v[j] = 2w[j] + (2x[j] + x j 2 j )x j + (2y[j] + y j 2 j )y j + (2z[j] + z j 2 j )z j w[j + 1] csfract(v[j]) s j+1 csint(v[j]) x[j + 1] (x[j], x j+1 ); y[j + 1] (y[j], y j+1 ); z[j + 1] (z[j], z j+1 ) S out s j+1 end for Figure 9.18: Radix-2 online sum of squares algorithm.

60 serial parallel x j+1 APPEND y j+1 APPEND z j+1 APPEND x[j+1] w[j+1] x[j] y[j] z[j] 60 2w[j] WS WC MUL/ APPEND MUL/ APPEND MUL/ APPEND [5:2] ADDER s j+1 CPA Sout s j in {0,...,8} w[j+1] APPEND implements x[j+1]=x[j]+x j+1 2 -j-1 δ ss = 0 x. x x x x x x x x x x. x x x x x x x x x x. x x x x x x x x x x. x x x x x x x x x x. x x x x x x x x x csint csfrac max(csint) = 8 (a) 2w[j] MUL/APPEND implements 2x[j]x j+1 +x 2 j+1 2-j-1 (2x[j]x j+1 +x 2 j+1 2-j-1 ) (2y[j]y j+1 +y 2 j+1 2-j-1 ) (2z[j]z j+1 +z 2 j+1 2-j-1 ) Note: the fractional portion of the 5-2 CSA produces at most three carries

61 COMPOSITE ALGORITHM 61 d = (x 2 + y 2 + z 2 ) Overall online delay of 5 A network of standard online modules: online delay of 11

62 62 y j+5 x serial j+5 z j+5 parallel APPEND APPEND APPEND x[j+1] w[j+1] x[j] y[j] z[j] Operation: WS WC MUL/ APPEND MUL/ APPEND MUL/ APPEND Sum of squares 2w[j] [5:2] ADDER CPA s j+6 w[j+1] Sout s j+5 {0,...8} R[j+1] On-the-Fly converter CONVERT d j+1 RS RC R[j] d d APPEND u=-(2 d[j]d j+1 +d 2 j+1 2-j-1 ) Square root dsel [3:2] ADDER d j+1 {-1,0,1} Dout R[j+1] d j Figure 9.20: Composite scheme for computing d = (x 2 + y 2 + z 2 ).

63 ONLINE IMPLEMENTATION OF RECURSIVE FILTER 63 IIR filter y(k) = a 1 y(k 1) + a 2 y(k 2) + bx(k) Conventional parallel arithmetic time to obtain y[k]: T CONV = 6t module. t module 6t FA rate of filter computation: R CONV 1/(4 6t FA ) LSDF serial arithmetic time to obtain y[k]: T LSDF = nt FA. rate of filter computation: R LSDF 1/(n t FA ) Online arithmetic Multioperation modules of type vu + w, online delay of 4 cycle time t M 3t FA Throughput independent of working precision but not the number of online units Rate: R OL = 1/( iter t M ) 1/(12t FA )

64 64 x[k] MUL ADD y[k] a 1 b y[k-1] ADD MUL (a) MUL a 2 y[k-2] coefficients x or y M1 M2 M3 M4 M5 y (Multiply) (Multiply-add) ([4:2] adder) ([4:2] adder) (CPA) CS form (b) CYCLE: k k+1 k+2 k+3 k+4 k+5 Module: M1 bx[k] R a 2 y[k-2] R a 1 y[k-1] R bx[k+1] R M2 M3 (bx[k] R ) +bx[k] L (a 2 y[k-2] R )+ a 2 y[k-2] L (a 1 y[k-1] R )+ a 1 y[k-1] L (bx[k]) + (a 2 y[k-2]) M4 (bx[k] + a 2 y[k-2])+ (a 1 y[k-1]) M5 y[k] (c) Figure 9.21: Conventional implementation of second-order IIR filter: (a) Filter. (b) 5-stage pipeline. (c) Timing diagram.

65 y[k-2] y[k-1] 65 MODULE M δ=2 δ=3 δ=3 x[k] b vu vu+w a 2 a 1 c d vu+w y[k] (a) x[k-2] M y[k-2] x[i] PARALLEL/SERIAL CONVERTER + DEMUX x[k-1] x[k] x[k+1] M M M y[k-1] y[k] y[k+1] SERIAL/PARALLEL CONVERTER + MUX y[j] Array for n=16 and iter =4 (b) y[k-2] y[k-1] from other modules x[k] c d y[k] (c) Figure 9.22: Online implementation of second-order IIR filter.

DIVISION BY DIGIT RECURRENCE

DIVISION BY DIGIT RECURRENCE 1 SEVERAL DIVISION METHODS: DIGIT-RECURRENCE METHOD studied in this chapter MULTIPLICATIVE METHOD (Chapter 7) VARIOUS APPROXIMATION METHODS (power series expansion), SPECIAL