DIVISION BY DIGIT RECURRENCE

Transcription

1 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 METHODS SUCH AS CORDIC (Chapter 11) AND CONTINUED PRODUCT METHODS. IMPLEMENTATIONS: SEQUENTIAL COMBINATIONAL 1. PIPELINED 2. NONPIPELINED COMBINATIONAL/SEQUENTIAL

2 DEFINITION AND NOTATION 2 DIVIDEND x DIVISOR d QUOTIENT q REMAINDER rem rem < d ulp INTEGER QUOTIENT: ulp = 1, x = q d + rem FRACTIONAL QUOTIENT: ulp = r n TWO TYPES OF DIVISION OPERATION: and sign(rem) = sign(x) 1. INTEGER DIVISION, WITH INTEGER OPERANDS AND RESULT USUALLY REQUIRES AN EXACT REMAINDER 2. FRACTIONAL DIVISION TO AVOID QUOTIENT OVERFLOW: x < d QUOTIENT ROUNDED, WHICH CAN RESULT IN A NEGATIVE REMAIN- DER

3 FRACTIONAL DIVISION - ASSUMPTIONS 3 OPERANDS/RESULT IN SIGN-AND-MAGNITUDE FORMAT = CONSIDER MAGNITUDES ONLY 1/2 d < 1; x < d; 0 < q < 1 FOR SIMPLER SELECTION: q j D a = { a, a+1,..., 1, 0, 1,...,a 1,a} REDUNDANCY FACTOR ρ = a r 1, 1 2 < ρ 1

4 RECURRENCE 4 QUOTIENT AFTER j STEPS: q[j] = q[0] + j i=1 q i r i FINAL QUOTIENT: q = q[n] = q[0] + n i=1 q i r i FINAL QUOTIENT ERROR BOUND: 0 ǫ q = x d q < r n ERROR AT STEP j: ǫ[j] = x d q[j] ǫ[n] + n FOR CONVERGENCE: ǫ[j] ρr j i=j+1 max(q i)r i = ǫ[n] + FOR SIMPLER SELECTION, ALLOW NEGATIVE ERRORS a r 1 (r j r n ) ǫ[j] = x d q[j] ρr j ELIMINATE DIVISION FROM ERROR BOUND: x dq[j] ρdr j DEFINE RESIDUAL (PARTIAL REMAINDER): w[j] = r j (x dq[j]) RESIDUAL RECURRENCE: w[j + 1] = rw[j] dq j+1 w[0] = x

5 cont. 5 BOUND ON w[j]: w[j] ρd SELECT QUOTIENT DIGIT TO KEEP w[j + 1] BOUNDED: q j+1 = SEL(rw[j], d) REDUNDANCY IN QUOTIENT-DIGIT SET ALLOWS q j+1 = SEL(ŷ, d) WHERE ŷ IS TRUNCATED rw[j], etc.

6 IMPLEMENTATION OF DIVISION STEP 6 1. ONE DIGIT ARITHMETIC LEFT-SHIFT OF w[j] TO PRODUCE rw[j]; 2. DETERMINATION OF THE QUOTIENT DIGIT q j+1 BY THE QUOTIENT-DIGIT SELECTION FUNCTION; 3. GENERATION OF THE DIVISOR MULTIPLE d q j+1 ; and 4. SUBTRACTION OF dq j+1 from rw[j]. 5. UPDATE OF THE QUOTIENT q[j] TO q[j + 1] BY THE ON-THE-FLY CONVERSION q[j + 1] = CONV (q[j],q j+1 )

7 7 w[j] - present residual ARITHMETIC LEFT SHIFT d - divisor rw[j] - shifted residual y d SEL q j+1 selection dq j+1 subtraction rw[j] q j+1 DIVISOR MULTIPLE GENERATION SUBTRACTION dq j+1 quotient update (on-the-fly) recurrence step time (b) w[j+1] - next residual QUOTIENT CONVERSION q j+1 (a) q[j+1] Figure 5.1: (a) COMPONENTS OF A DIVISION STEP. (b) TIMING.

8 MAIN FACTORS AFFECTING COST AND EXECUTION TIME 8 RADIX r QUOTIENT-DIGIT SET 1. CANONICAL: 0 q j r 1 2. REDUNDANT: q j D a = { a, a + 1,..., 1, 0, 1,...,a 1, a} REDUNDANCY FACTOR: ρ = a r 1, ρ > 1 2 REPRESENTATION OF RESIDUAL: 1. NONREDUNDANT (e.g., 2 s complement) 2. REDUNDANT: carry-save, signed-digit QUOTIENT-DIGIT SELECTION FUNCTION

9 INITIALIZATION AND TERMINATION 9 Initialization: w[0] = x dq[0] and w[0] ρd. Options: Make q[0] = 0 and For ρ = 1 we make w[0] = x/2. For 1/2 < ρ < 1 we make w[0] = x/4 Compensated in the termination step Make q[0] = 1 and w[0] = x d. Applicable for ρ < 1 because q > 1 +ρ not allowed. Termination: QUOTIENT: q = q[n] if w[n] 0 q[n] r N if w[n] < 0 N number of iterations If dividend shifted in initialization - shift quotient (extra iteration)

10 ON-THE-FLY QUOTIENT CONVERSION 10 j MS DIGITS OF CONVERTED QUOTIENT UPDATE Q[j] = j i=1 q ir i Q[j + 1] = Q[j] + q j+1 r (j+1) SINCE q j+1 CAN BE NEGATIVE: Q[j + 1] = Q[j] + q j+1 r (j+1) if q j+1 0 Q[j] r j + (r q j+1 )r (j+1) if q j+1 < 0 DISADVANTAGE: SUBTRACTION Q[j] r j REQUIRES THE PROPAGATION OF A BORROW SLOW

11 ON-THE-FLY CONVERSION (cont.) 11 DEFINE ANOTHER FORM QM[j] QM[j] = Q[j] r j NEW CONVERSION ALGORITHM IS Q[j + 1] = Q[j] + q j+1 r (j+1) if q j+1 0 QM[j] + (r q j+1 )r (j+1) if q j+1 < 0 SUBTRACTION REPLACED BY LOADING THE FORM QM[j] UPDATE FORM QM[j] AS FOLLOWS: QM[j + 1] = Q[j + 1] r (j+1) = ALL ADDITIONS ARE CONCATENATIONS Q[j] + (q j+1 1)r (j+1) if q j+1 > 0 QM[j] + ((r 1) q j+1 )r (j+1) if q j+1 0

12 ON-THE-FLY CONVERSION ALGORITHM 12 Q[j + 1] = QM[j + 1] = (Q[j],q j+1 ) if q j+1 0 (QM[j], (r q j+1 )) if q j+1 < 0 (Q[j],q j+1 1) if q j+1 > 0 (QM[j], ((r 1) q j+1 )) if q j+1 0 INITIAL CONDITIONS Q[0] = QM[0] = 0 (for a positive quotient)

13 EXAMPLE OF RADIX-2 CONVERSION 13 j q j Q[j] QM[j]

14 IMPLEMENTATION OF THE CONVERSION 14 Q 2-1 MUX QM load select q Q REG Q QM 2-1 MUX Q in load select LOAD & SHIFT-IN CONTROL q j+1 QM REG QM in QM Q QM Figure 5.2: Implementation of on-the-fly conversion. shift Q with insert (Q in ) if C shiftq = 1 shift QM with insert (Q in ) if C loadq = 1 shift QM with insert (QM in ) if C shiftqm = 1 shift Q with insert (QM in ) if C loadqm = 1

15 cont. 15 Q in = QM in = q j+1 if q j+1 0 r q j+1 if q j+1 < 0 q j+1 1 if q j+1 > 0 (r 1) q j+1 if q j+1 0 REGISTER CONTROL SIGNALS: C loadq = C shiftq and C loadqm = C shiftqm

16 EXAMPLE OF RADIX-4 CONVERSION 16 q j+1 Q in C shiftq Q[j + 1] QM in C shiftqm QM[j + 1] (Q[j], 3) 2 0 (Q[j], 2) (Q[j], 2) 1 0 (Q[j], 1) (Q[j], 1) 0 0 (Q[j], 0) (Q[j], 0) 3 1 (QM[j], 3) (QM[j], 3) 2 1 (QM[j], 2) (QM[j], 2) 1 1 (QM[j], 1) (QM[j], 1) 0 1 (QM[j], 0)

17 x 17 INITIAL STEP w[0] d q[j+1] RESIDUAL REG w[j] RECURRENCE STEP w[j+1] w[n] TERMINAL STEP x INITIAL STEP w[0] RESIDUAL REG w[ik] d (a) q x INITIAL STEP w[0] REC. STEP w[1] REC. STEP w[2] q 1 q 2 COMBINATIONAL ON-THE-FLY CONVERTER & TERMINAL STEP d q ik+1 REC. STEP w[ik+1] q ik+2 REC. STEP w[ik+2] q (i+1)k REC. STEP w[(i+1)k] (c) w[n] COMBINATIONAL ON-THE-FLY CONVERTER & TERMINAL STEP q q N REC. STEP w[n] (b) q Figure 5.3: DIVISION IMPLEMENTATION: (a) TOTALLY SEQUENTIAL. (b) TOTALLY COMBINATIONAL. (c) COMBINED IMPLEMENTATION.

18 EXAMPLES OF ALGORITHMS AND IMPLEMENTATIONS 18 CELLS: delay function of the load; delay and area in terms of 2-NAND. DEGREE OF OPTIMIZATION: the same modules have been used in all designs. INTERCONNECTIONS NOT INCLUDED: not considered the delay, area nor load of interconnections. EXECUTION TIME AND THE AREA FOR 53-BIT OPERANDS AND RE- SULT INCLUDED DELAY AND AREA OF REGISTERS

19 SCHEMES COMPARED 19 r2 Scheme Radix-2 with carry-save residual. r4 Scheme Radix-4 with a = 2 and carry-save residual. r8 Scheme Radix-8 with a = 7 and carry-save residual. r16over Scheme Radix-16 with two overlapped radix-4 stages. r512 Scheme Radix-512 with a = 511, carry-save residual, scaling, and quotientdigit selection by rounding.

20 20 2-input MUX 3-input MUX: Delay: 1.6 Area: input MUX: Delay: 1.8 Area: 5.6 [3:2] MODULE (FULL ADDER) g a b c p a b c HA g p HA* Delay: 1.4 Area: 3.0 (a) vci+1 vsi Delays: BUFFER Delay: 1.8 Area:2.6 (b) REGISTER CELL SET D Q CK Delay: 4.0 Area: 4.0 vci+1 vsi (c) a,b to vsi: 4.2 c to vsi: 2.2 a,b to vci+1 : 3.8 c to vci+1: 2.0 Area: 6.7 [4:2] MODULE a b c d a b b c c a a b c d t i+1 [4:2] Module t i vc i vc i+1 vs i Delays: to ti+1: 1.6 t i+1 d 0 1 MUX t i to vci+1: 5.6 vc i to vsi: 6.0 Area: 15.5 vc i+1 (d) vs i Figure 5.4: Basic modules: (a) Multiplexers. (b) Buffer and register cell. (c) Full-adder. (d) [4:2] module.

21 RADIX-2 DIVISION WITH CS RESIDUAL 21 REDUNDANT RESIDUAL w[j] = (W C[j], W S[j]) 1. [Initialize] WS[0] x/2; WC[0] 0; q 0 0;Q[-1]=0 2. [Recurrence] for j = 0... n + 1 (n + 2 iterations because of initialization and guard bit) q j+1 SEL(ŷ); (WC[j + 1], WS[j + 1]) CSADD(2WC[j], 2WS[j], q j+1 d); Q[j] CONV ERT(Q[j 1], q j ) end for 3. [T erminate] If w[n + 2] < 0 then q = 2(CONV ERT(Q[n + 1], q n+2 1)) else q = 2(CONV ERT(Q[n + 1], q n+2 ))

22 RADIX-2 ALGORITHM (cont.) 22 n is the precision in bits, SEL is the quotient-digit selection function: 1 if 0 ŷ 3/2 q j+1 = SEL(ŷ) = 0 if ŷ = 1/2 1 if 5/2 ŷ 1 The estimate ŷ has four bits (three integer bits and one fractional bit) of the shifted residual in carry-save form, CSADD is carry-save addition q j+1 d is in 2 s complement form, and CON V ERT on-the-fly quotient conversion function

23 Dividend x = ( ), divisor d = ( ), scaled residual 2w[0] = 2(x/2) = x, q computed = q/2 23 2W S[0] = WC[0] = ŷ[0] = 0.5 q 1 = 1 q 1 d = W S[1] = WC[1] = ŷ[0] = 1 q 2 = 1 q 2 d = W S[2] = WC[2] = ŷ[1] = 0.5 q 3 = 1 q 3 d = W S[3] = WC[3] = ŷ[2] = 0 q 4 = 1 q 4 d = W S[4] = WC[4] = ŷ[4] = 1.5 q 5 = 1 q 5 d = * for 2 s complement of q j+1 d q = 2( ) =.1101 Figure 5.7: Example of radix-2 division with residual in carry-save form.

24 Divisor d WC[0] = 0 WC[j+1] WS[0] = x /2 WS[j+1] 24 q j+1 buffers 2 Reg. D Reg. WC Reg. WS d d 2WC[j] 4 2WS[j] 2-1 MUX {d, 0, d} c b [3:2] SZ; Convert q a q - WS[j+1] WC[j+1] qj+1 4 qsel qsel buff MUX HA* REG [3:2] adder HA qj+1 q + q critical path Modules defined in Fig. 5.4 Figure 5.8: IMPLEMENTATION OF RADIX-2 SCHEME.

25 DELAY AND AREA OF RADIX-2 STAGE 25 element delay area q-digit selection buffers MUX CSA registers (3) Convert & Round (NC) 1360 Cycle time 16.2 Total area 2585 NC denotes a delay not in the critical path

26 RADIX-4 DIVISION WITH CS RESIDUAL QUOTIENT DIGIT SET {-2,-1,0,1,2} 2. ρ < 1 initialize WS[0] x/4 3. THE NEXT RESIDUAL (WC[j + 1], WS[j + 1]) CSADD(4WC[j], 4WS[j], q j+1 d) 4. QUOTIENT-DIGIT SELECTION DEPENDS ON ESTIMATES OF SHIFTED RESIDUAL AND DIVISOR described in terms of SELECTION CONSTANTS m k (i) q j+1 = k if m k (i) ŷ < m k+1 (i) 5. FINAL QUOTIENT = 4 x OBTAINED QUOTIENT

27 SELECTION CONSTANTS 27 q j+1 = k if m k (i) ŷ < m k+1 (i) i = 16 ˆd and ˆd divisor truncated to the 4th fractional bit and ŷ is 4w[j] truncated to the 4th fractional bit. i m 2 (i) m 1 (i) m 0 (i) m 1 (i) : real value = shown value/16

28 Dividend x = ( ), divisor d = ( ) (i = 16(0.1100) 2 = 12) scaled residual 4w[0] = 4(x/4) = x, q computed = q/4 28 4WS[0] + = WC[0] + = ŷ[0] = 10/16 q 1 = 1 q 1 d + = W S[1] = W C[1] = WS[1] + = WC[1] + = ŷ[1] = 6/16 q 2 = 0 q 2 d + = W S[2] = W C[2] = WS[2] + = WC[2] + = ŷ[2] = 22/16 q 3 = 2 q 3 d + = w[3] = * least-significant 1 for 2 s complement of q j+1 d + only one integer bit used in the recurrence, because of the range of w[j + 1]. q[3] = = Figure 5.9: Example of radix-4 division with residual in carry-save form.(on-the-fly conversion and termination not shown)

29 DELAY AND AREA OF RADIX-4 STAGE 29 element delay area q-digit selection buffers MUX CSA registers (3) Convert & Round (NC) 1360 Cycle time 20.6 Total area 2840 NC denotes a delay not in the critical path

30 Divisor d WC[0]=0 WC[j+1] WS[0] = x /4 WS[j+1] 30 q j+1 buffers 4 Reg. D Reg. WC Reg. WS d d 4WC[j] 7 4WS[j] 4-1 MUX 7 qsel 3 {2d, d, 0, d, 2d} c b a [3:2] SZ; Convert (q j+1 < 0) WS[j+1] WC[j+1] 4 {d} 4 q j+1 q [3:2] adder (q2 +, q2 -, q1 +, q1 - ) qj+1 = 0 coded as (0, 0, 0, 0) 2.2 qsel buff MUX HA* REG HA critical path Modules defined in Fig. 5.4 Figure 5.10: IMPLEMENTATION OF RADIX-4 SCHEME.

31 RADIX-8 DIVISION WITH CS RESIDUAL 31 QUOTIENT DIGIT SET {-7,..., 7} DECOMPOSED INTO AND q H = { 8, 4, 0, 4, 8} q L = { 2, 1, 0, 1, 2} element delay area q-digit selection (q h ) buffers MUXes CSAh CSAl registers (3) Convert & Round (NC) 1360 Cycle time 26.2 Total area 3960

32 Divisor d WC[0]=0 WC[j+1] WS[0]=x/2 WS[j+1] 32 q H j+1 4 q L j+1 4 buffers buffers Reg. D Reg. WC Reg. WS d d 8WC[j] 8 8WS[j] 4-1 MUX {8d, 4d, 0, 4d, 8d} d 4-1 MUX d [3:2] (q H j+1 < 0) 8 qsel 4 4 q q H j+1 q L j+1 : j+1 (decoded as in Fig. 5.10) 3 {d} 4 {2d, d, 0, d, 2d} [3:2] SZ; Convert q (q L j+1 < 0) WS[j+1] WC[j+1] (q H j+1,q L j+1) q H j qsel buff MUX HA* HA REG HA 4.2 qsel buff MUX HA* q L j+1 Modules defined in Fig. 5.4 critical path Figure 5.11: Implementation of radix-8 scheme.

33 RADIX-16 DIVISION WITH TWO RADIX-4 OVERLAPPED STAGES 33 element delay area CSA q-digit selection MUX 1.4 buffers MUXes CSA1 (NC) 360 CSA registers (3) Convert & Round (NC) 1360 Cycle time 26.6 Total area 4390

34 34 Divisor d WC[0]=0 WC[j+2] WS[0]=x/4 WS[j+2] q j+1 4 buffers Reg. D Reg. WC Reg. WS d d 4WC[j] 4WS[j] 4-1 MUX q j+2 4 buffers {2d, d, 0, d, 2d} d 4-1 MUX {2d, d, 0, d, 2d} [3:2] d WS[j+1] WS[j+2] [3:2] (q j+1 < 0) WC[j+1] 4W[j+1] (q j+2 < 0) WC[j+2] (a) 4WS[j] 4WC[j] x x x. x x x x x x x x x x. x x x x x x x {4W[j]} 7 {4W[j]} 7 7 most 8 least {2d} 7 {d} 7 {d} 7 {2d} 7 8 least 7 most {d} qsel [3:2] [3:2] [3:2] [3:2] qsel qsel qsel qsel qsel 5-1 MUX Conditional truncated q j+1 q j+2 (b) residuals Figure 5.12: Implementation of radix-16 with radix-4 stages.

35 35 CSA2 q j+1 qsel buff MUX CSA HA REG 4 q j+2 [3:2] 8 qsel MUX buff MUX HA* 2.2 Modules defined in Fig. 5.4 critical path Figure 5.13: Critical path in radix-16 scheme.

36 RADIX-512 DIVISION WITH PRESCALING AND SELECTION BY ROUNDING 36 Cycle 1 : COMPUTE M 1/d; compare d and x and set g Cycle 2 : COMPUTE z = Md (in c-s form); v = 2 g x Cycle 3 : INITIALIZE w[0] = Mv (in c-s form); ASSIMILATE z; Cycles 4 to 9 : ITERATE q j+1 = round(ŷ); w[j + 1] = 512w[j] q j+1 z Cycle 10 : CORRECTING AND ROUNDING.

37 RADIX-512 DIVISON (cont.) 37 RECTANGULAR MULTIPLIER-ACCUMULATOR DELAY-AREA element delay area M-module (NC) 1800 MUX 1.4 recoder buffer 1.8 MUX levels of 4-2 CSA registers(3) Convert & Round (NC) 1360 Cycle time 27 Total area 9980

38 Module M 15 {d} 15 x d (Initialization not shown) 38 scale/iterate MUX M MUX z q j+1 Round & Recode M ; q j+1 MULTIPLIER - ACCUMULATOR Md; Mx; 512w -qz = w[j] WC[j+1] WS[j+1] W WC[j] WS[j] 2(10+2) CPA WC[j+1] WS[j+1] SZ; CONVERT q j+1 q Figure 5.14: Implementation of radix-512 scheme.

39 OVERALL COMPARISONS 39 Scheme r2 r4 r8 r16 (overlapped) r512 Cycle-time factor Number of cycles Speedup Area factor Correction: two cycles for radix-2, one cycle for other cases.

40 QUOTIENT-DIGIT SELECTION FUNCTION 40 Quotient-digit set: Redundancy factor: q j+1 D a = { a, a + 1,..., 1, 0, 1,...,a 1,a} ρ = a r 1, 1 2 < ρ 1 TWO FUNDAMENTAL CONDITIONS FOR q-selection CONTAINMENT must guarantee bounded residual CONTINUITY there must exist a valid choice of q j+1 in the range of shifted residual

41 CONTAINMENT AND SELECTION INTERVALS 41 RESIDUAL RECURRENCE w[j + 1] = rw[j] dq j+1 w[j] ρd ρ = a/(r 1) a q j a SELECTION INTERVALS If rw[j] [L k,u k ] then q j+1 = k makes w[j + 1] bounded L k rw[j] U k ρd w[j + 1] = rw[j] k d ρd EXPRESSIONS FOR SELECTION INTERVALS U k = (k + ρ)d L k = (k ρ)d

42 ROBERTSON S AND P-D DIAGRAMS 42 w[j+1] ρd -a k a -rρd -ad -d d rw[5] kd ad rρd rw[j] w[6] -ρd L k U k (a) Figure 5.16: ROBERTSON S DIAGRAM

43 rw[j] U a =rρd 43 k>0 selection region for q j+1 =k U k L k U 0 0 d L 0 U k k<0 L k (b) L -a =-rρd Figure 5.16: P-D DIAGRAM (for d > 0).

44 CONTINUITY CONDITION, OVERLAP, AND Q-DIGIT SELECTION 44 q j+1 = SEL(w[j],d) SEL represented by the set {s k }, a k a, q j+1 = k if s k rw[j] s k+1 ulp s k defined as the minimum value of rw[j] for which q j+1 = k s k s are functions of the divisor d CONTAINMENT: L k s k U k CONTINUITY: q j+1 = k 1 for rw[j] = s k ulp U k 1 U k U k 1 + ulp L k s k U k 1 + ulp or L k s k U k 1 OVERLAP RESULTING IN U k 1 L k = (k 1 + ρ)d (k ρ)d = (2ρ 1)d ρ 2 1 REDUNDANCY IN q-digit SET OVERLAP BETWEEN SELECTION INTERVALS - SIMPLER SELECTION

45 45 w[j+1] q j+1 =k-1 q j+1 =k rw[j] L k-1 k-1 L k k or k-1 Uk-1 overlap k U k (a) Figure 5.17: Overlap between selection intervals.

46 46 rw[j] U k-1 s k L k q j+1 =k overlap (k or k-1) q j+1 =k-1 1/2 normalized divisor range (b) 1 d Figure 5.17: Selection function.

47 SELECTION USING CONSTANTS 47 USE CONSTANTS m k, INDEPENDENT OF DIVISOR max(l k ) m k min(u k 1 ) + ulp max and min for the range 2 1 d < 1 For k > 0 which requires For k 0 which requires (k ρ) m k (k 1 + ρ)2 1 + ulp ρ k (k ρ)2 1 k 1 + ρ ρ ( k) + 2 3

48 48 rw[j] U k-1 m k L k k>0 1/2 1 k<0 d m k U k-1 L k Figure 5.18: BOUNDS ON m k.

49 RADIX-2 DIVISION WITH NON-REDUNDANT RESIDUAL 49 EXTENSION OF NON-RESTORING DIVISION: {-1,0,1} SRT ALLOWS SKIPPING OVER ZEROS SELECTION CONSTANTS: U 1 = 2d L 1 = 0 U 0 = d 1/2 L 0 = d 1/2 U 1 = 0 L 1 = 2d 0 m 1 1/2, 1/2 m 0 0 Choose: m 1 = 1/2 and m 0 = 1/2 THE QUOTIENT-DIGIT SELECTION FUNCTION 1 if 1/2 2w[j] q j+1 = 0 if 1/2 2w[j] < 1/2 1 if 2w[j] < 1/2

50 w[j+1] 50 d d -1/2 1/2 2w[j] -d d 2d -d L -1 U -1 L 1 U 1 L 0 U (a)

51 2w[j] 51 2 U 1 q j+1 =1 1 U 0 1/2 m 1 =1/2 U -1 d 1/2 L 1 1 q j+1 =0-1/2 m 0 =-1/2 L 0-1 L -1 q j+1 =-1-2

52 STAIRCASE SELECTION FUNCTION 52 FOR r > 2, m k DEPENDS ON DIVISOR DIVIDE RANGE OF DIVISOR INTO INTERVALS [d i,d i+1 ) with d 0 = 1 2, d i+1 = d i + 2 δ δ MS FRACTIONAL BITS OF DIVISOR REPRESENT THE INTERVAL FOR EACH INTERVAL, THERE IS A SET of selection constants m k (i) for d [d i, d i+1 ), q j+1 = k if m k (i) rw[j] m k+1 (i) ulp

53 DEFINITION OF m k (i) 53 rw[j] m k+1 (i) q j+1 =k+1 ulp q j+1 =k m k (i) d 2 -δ d i d i+1 Figure 5.20: DEFINITION OF m k (i).

54 RANGE CONDITION FOR SELECTION CONSTANTS 54 max(l k (d i ), L k (d i+1 )) m k (i) min(u k 1 (d i ), U k 1 (d i+1 )) + ulp rw[j] min(u k-1 (d i ), U k-1 (d i+1 )) region for m k (i) U k-1 L k max(l k (d i ), L k (d i+1 )) d 2 -δ d i d i+1 Figure 5.21: SELECTION CONSTANT REGION.

55 LOW-PRECISION SELECTION CONSTANTS 55 m k (i) = A k (i)2 c WHERE A k (i) IS INTEGER CAN USE TRUNCATED RESIDUAL IN COMPARISONS WITH SELECTION CONSTANTS RESIDUAL MUST BE IN 2 s COMPLEMENT SELECTION CONDITIONS for k > 0 L k (d i + 2 δ ) A k (i)2 c U k 1 (d i ) for k 0 L k (d i ) A k (i)2 c U k 1 (d i + 2 δ ) + ulp (5.1)

56 USE OF SELECTION CONSTANTS 56 rw[j] U k-1 2 -c m k (4) L k m k (2,3) m k (1) m k (0) 2 -δ d d 0 d 2 d 3 d 4 d 5 d 1 Figure 5.22: QUOTIENT-DIGIT SELECTION WITH SELECTION CONSTANTS.

57 SELECTION WITH TRUNCATED RESIDUAL AND DIVISOR log 2 r + c {rw [j ]} c log 2 r + c + δ SELECTION FUNCTION δ 1 {d } δ q j+1 Figure 5.23: SELECTION WITH TRUNCATED RESIDUAL AND DIVISOR.

58 LOWER BOUND ON δ 58 CONSIDER CASE k > 0 (similar argument for k 0); FROM CONTINUITY CONDITION = U k 1 (d i ) L k (d i + 2 δ ) 0 (ρ + k 1)d i ( ρ + k)(d i + 2 δ ) 0 Worst case: d 1/2 and k a: (2ρ 1)d i (k ρ)2 δ 2 δ 2ρ 1 2(a ρ) = 2ρ 1 2ρ(r 2) MINIMUM VALUE OF δ CAN RESULT IN A LARGE VALUE OF c OPTIMIZE THE VALUES OF δ AND c TOGETHER

59 RADIX-4 DIVISION WITH NONREDUNDANT RESIDUAL 59 Known as Robertson s division q j { 2, 1, 0, 1, 2} U k = ( k)d BOUND ON δ L k = ( k)d 2 δ 2ρ 1 2(a ρ) = 1 8

60 60 4w[j] U U 1 L 2 2d/3 w[j+1] 1 1 U d/3 4d/3 2d/3 5d/3 8d/3 4w[j] d=1/2 0 1 L 1 U -1 d -2d/3 (a) -1-1 L 0 U L -1 L -2 Figure 5.24: ROBERTSON S AND PD DIAGRAMS FOR RADIX-4 AND a = 2. (b)

61 [d i,d i+1 ) + [8, 9) [9, 10) [10, 11) [11, 12) L 2 (d i+1 ), U 1 (d i ) # 36, 40 40, 45 44, 50 48, 55 m 2 (i) L 1 (d i+1 ), U 0 (d i ) # 9,16 10, 18 11, 20 12, 22 m 1 (i) L 0 (d i ), U 1 (d i+1 ) # -16, -9-18, , , -12 m 0 (i) L 1 (d i ),U 2 (d i+1 ) # -40, , , , -48 m 1 (i) [d i,d i+1 ) + [12, 13) [13, 14) [14, 15) [15, 16) L 2 (d i+1 ), U 1 (d i ) # 52, 60 56, 65 60, 70 64, 75 m 2 (i) L 1 (d i+1 ), U 0 (d i ) # 13, 24 14, 26 15, 28 16, 30 m 1 (i) L 0 (d i ), U 1 (d i+1 ) # -24, , , , -16 m 0 (i) L 1 (d i ),U 2 (d i+1 ) # -60, , , , -64 m 1 (i)

62 QUOTIENT-DIGIT SELECTION FOR RADIX-4 DIVISION; NON-REDUNDANT RESIDUAL 62 {4w[j]} 3 U 1 m 2 (7) m 2 (4,5,6) 12/8 11/8 10/8 9/8 L 2 q j+1 =2 {4w[j]} 3 xxx.xxx 6 {d} 4 0.1xxx 3 m 2 (2,3) 1 SELECTION m 2 (1) m 2 (0) 7/8 6/8 q j+1 =1 3 5/8 q j+1 1/2 10/16 12/16 14/16 1 {d} 4 (b) 9/16 11/16 13/16 15/16 (a) Figure 5.25: QUOTIENT-DIGIT SELECTION: (a) FRAGMENT OF THE P-D DIAGRAM. (b) IMPLEMENTATION.

63 USE OF REDUNDANT ADDER 63 SO FAR: COMPUTE rw[j] IN FULL PRECISION, TRUNCATE, AND COM- PARE WITH LOW-PRECISION CONSTANTS FULL-PRECISION ADDITION: significant portion of the cycle time OVERLAP BETWEEN SELECTION INTERVALS = COULD USE AN ESTIMATE OF rw[j] ERROR IN ESTIMATE: ǫ min y ŷ ǫ max BASIC CONSTRAINT: if we choose q j+1 = k for an estimate ŷ then y [ŷ + ǫ min, ŷ + ǫ max ] L k = L k ǫ min U k = U k ǫ max

64 CONSTRAINTS ON SELECTION CONSTANTS 64 max(l k(d i ),L k(d i+1 )) m k (i) min(u k 1(d i ),U k 1(d i+1 )) U k-1 U* k-1 region for m k (i) ε max L* k ε min L k d i d i+1 Figure 5.26: CONSTRAINTS FOR SELECTION BASED ON ESTIMATES.

65 MINIMUM OVERLAP AND RANGE OF ESTIMATE 65 OVERLAP min(uk 1(d i ),Uk 1(d i+1 )) max(l k(d i ), L k(d i+1 )) 0 RANGE rw[j] rρd < rρ (for d < 1) rρ ǫ max < ŷ < rρ ǫ min y=rw[j] ε max -ε min -rρ y (estimate) rρ Figure 5.27: RANGE OF ESTIMATE

66 REDUNDANT ADDER 66 -q j+1 d rw[j] (redundant) (a) REDUNDANT ADDER w[j+1] (redundant) ws wc x x - x x + x. x. +. x x + t x x + x x + x x + x x + x x + x x + x x + rw[j] x + - x + - x. x t x + - x + - x + - x + - x + - x + - x + - rw[j] y - y^ y + y. y +. + y y + (estimate) 0 < ε < 2 -t+1 -ulp + (truncation error) y + - y^ y. y y y (estimate) y + - -(2 -t -ulp) < ε < (truncation error) 2 -t -ulp (b) (c) Figure 5.28: USE OF REDUNDANT ADDER: (a) Redundant adder. (b) Carry-save case. (c) Signed-digit case.

67 CARRY-SAVE ADDER 67 ERRORS ǫ min = 0 ǫ max = 2 t+1 ulp RESTRICTED SELECTION INTERVAL U k = U k 2 t+1 + ulp L k = L k U k 1 = Uk t t = U k 1 2 t t L k = L k t = L k t

68 y U k t Case B ^ U k-1 U* k-1 (represented by ) 2 -t+1 Case A ^ L k L k (represented by ) L* k d d d i i+1 i+2 Case A: U * k-1 is on the grid; U k-1 = U * k-1 +2-t on the grid Case B: U * k-1 is off the grid; U k-1 > U * k-1 on the grid Figure 5.29: Û and L for residual in carry-save form.

69 LOWER BOUND FOR t and δ 69 (for positive k) U k 1 (d i ) L k (d i+1 ) 0 U k 1 (d i ) 2 t L k (d i+1 ) 0 RANGE 2ρ 1 2 (a ρ)2 δ 2 t rρ 2 t t ŷ rρ ulp t z t = 2 t 2 t z

70 RADIX-2 DIVISION WITH CARRY-SAVE ADDER δ 2 t max( L k (d i ), L k (d i+1 )) m k (i) min(ûk 1(d i ),Ûk 1(d i+1 )) L 1 (1) = 0 U 0 (1/2) = 0 L 0 (1/2) = 1/2 U 1 (1) = 1/2 ( L 1 (1) = 0) m 1 (Û0(1/2) = 0) ( L 0 (1/2) = 1/2) m 0 (Û 1(1) = 1/2) This results in the selection constants m 1 = 0 and m 0 = 1/2

71 SELECTION FUNCTION ŷ 2 ulp ŷ 3/2 q j+1 = 1 if 0 ŷ 3/2 0 if ŷ = 1/2 1 if 5/2 ŷ 1

72 RADIX-2 WITH CSA: PD PLOT 72 2w[j] 2 3/2 U 1 y[j] ^ x x x. x x x x. x 1 U y[j] ^ U^ 0 1/2 0-1/2 L ^ m(0)=-1/ /2 q j+1 =1 q j+1 =0 q j+1 =-1 m(1)=0 ^ L 1 U^ U* 0 L, 1 U -1 U* -1 L 0, L* 0 d SELECTION FUNCTION q j+1 (b) -2-5/2 d=1/2 (a) d=1-ulp L -1 Figure 5.30: RADIX-2 DIVISION WITH CARRY-SAVE ADDER: (a) P-D PLOT. (b) SELECTION FUNCTION.

73 IMPLEMENTATION 73 q j+1 = (q s,q m ) where and q m = (p 1 p 0 p 1 ) (5.2) q s = p 2 (g 1 + p 1 g 0 + p 1 p 0 g 1 ) p i = c i s i g i = c i s i (c 2, c 1, c 0, c 1 ) (s 2,s 1, s 0,s 1 ) q j+1 = 0 (q s, q m ) = (1, 0)

74 RADIX-4 ALGORITHM WITH CARRY-SAVE ADDER δ 2 t 2 t = /3 1/16 4 ŷ 8/3 ulp ŷ 42 16

75 [d i, d i+1 ) + [8, 9) [9, 10) [10, 11) [11, 12) 75 L 2 (d i+1 ),Û1(d i ) + 12, 12 14, 14 15, 15 16, 17 m 2 (i) L 1 (d i+1 ),Û0(d i ) + 3, 4 4, 5 4, 5 4, 6 m 1 (i) L 0 (d i ),Û 1(d i+1 ) + -5, -4-6, -5-6, -5-7, -5 m 0 (i) L 1 (d i ),Û 2(d i+1 ) + -13, , , , -17 m 1 (i) [d i, d i+1 ) + [12, 13) [13, 14) [14, 15) [15, 16) L 2 (d i+1 ),Û1(d i ) + 18, 19 19, 20 20, 22 22, 24 m 2 (i) L 1 (d i+1 ),Û0(d i ) + 4, 7 5, 7 5, 8 6, 9 m 1 (i) L 0 (d i ),Û 1(d i+1 ) + -8, -6-8, -6-9, -6-10, -7 m 0 (i) L 1 (d i ),Û 2(d i+1 ) + -20, , , , -23 m 1 (i) : real value = shown value/16; L k = L k 4, Ûk = U k

76 76 y[j] ^ U 1 U^ 1 12/8 L^ 2 11/8 L 2 10/8 9/8 y^ x x x.x x x x x x x d^ 1 SELECTION FUNCTION 7/8 q j+1 (b) 6/8 5/8 1/2 9/16 10/16 11/16 12/16 13/16 14/16 15/16 1 ^ d m 2 (i) region of choice for m 2 (i) (a) Figure 5.31: SELECTION FUNCTION FOR RADIX-4 SCHEME WITH CARRY-SAVE ADDER: (a) Fragment of P-D diagram. (b) Implementation.