Transcript Residue Number systems - IEEE

Residue Number systems
P.V. Ananda Mohan
FNAE, Fellow IEEE
[email protected]
IEEE CAS Chapter
8th March 2008
Bangalore
Why RNS
• Using several processors in parallel, some
operations can be faster.
Input Binary Number
Binary to
RNS
converter
r1
Binary to
RNS
converter
r2
Mod m1
Processor
Mod m2
Processor
Instruction
O1
O2
RNS to Binary Converter
Result
Binary to
RNS
converter
rj
Mod mj
Processor
Oj
Points to be considered
• Choice of moduli set
• Computation time and area requirements for the
following blocks:
• Binary to RNS conversion
• RNS to Binary conversion
• Multiplication
• Scaling
• Base extension
• Sign detection
• Comparison
Binary to RNS conversion
• (a) Conventional method: division to get
residue throwing away quotient
• --Very time consuming.
• Example (1000 0001 1010) mod 13?
• 2074 mod 13 = 7.
• (b) Iterative reduction mod mi
• (Capocelli and Giancarlo)
• Start with LSBs. Store residues of powers of two in
memory go on accumulating till end mod 13:
• 1,2,4,8,3,6,12,11,9,5,10,7
• Example (1000 0001 1010) mod 13?
• Last three bits you can skip.
• 2+23 mod 13 = 2+8 = 10
• 10+24 mod 13 = 10+3=0 and so on
• Hardware needed : a modulo adder, Memory containing
residues of Powers of 2 mod 13.
• (c) Use periodic properties of moduli
• For example consider modulus 18.
• Residues of powers of two are
(1,2,4,8,3,6), (12,11,9,5,10,7),(1,2,4,8..)
etc
• Note the periodic property
• (1,2,4,8,3,6), (-1,-2,-4,-8,-3,-6),
(1,2,4,8,3,6), (-1,-2,-4,-8,-3,-6)
Consider mod 89
• Residues of successive powers of two are
1,2,4,8,16,32,64,39,78,67,45,
1,2,4,8,16,32,64,39,78,67,45,
• Thus period (or order) is 11
• i.e. 211 mod 89=1
• Implementation: Group input bits based on
period or half period.
• If based on period, add all words with same
period mod 211 and have one Binary to RNS
converter of Capocelli and Giancarlo.
• If based on half-period add all odd fields and
add all even fields, Compute odd-even and use
Capocelli and Giancarlo method
•
•
•
•
•
•
Example
2074 mod 13= (100000 011010) mod 13
= (26-32) mod 13 = -6 mod 13 = 7.
2074 mod 7 = (100 000 011 010) mod7
= (4+0+3+2) mod 7=2
Use for full period case, Adders with end
around carry (EAC) and for half period
case, two adders with EAC
1
0
0
0
0
0
0
0
1
1
0
1
0
0
• Delay is (2+3+2)D
100
000
011
----111Sum
0000 Carry
010
------101 Sum
0100 Carry
-----1001
1
-----010
Modulo adders and subtractors
• (X+Y) mod mi = (X+Y) or (X+Y-mi)
• (X-Y) mod mi = (X-Y) or (X-Y+mi)
X
Y
n bit Adder
Two’s complement of mi or
(2n-mi)
(X+Y)
Sign
bit
(n +1) bit Adder
Delay = nDFA+(n+1)DFA+DMUX
Area = nAFA+(n+1)AFA+n D2:1MUX
2:1 MUX
Cascade of Adders
select
(X+Y) mod mi
Faster Adder Implementations
X
Y
Two’s complement of
mi or (2n-mi)
n bit Adder
Sign
bit
(n +1) bit Adder
(X+Y)
2:1 MUX
select
Delay = (n+2)DFA+DMUX
(X+Y) mod mi
Area = nAFA+2(n+1)AFA+n D2:1MUX
• Subtractor is same bur two’s compliment
of input to be added.
Modulo Multipliers
X
Y
Multiplier
mi
XY
Divider
Quotient
Throw it.
•
•
•
•
Reminder
Area Multiplier+divider
Delay Multiplier+divider
Divider can be restoring or non-restoring.
Word length of the processor 2n bits
Brickell’s Algorithm based Modulo
Multipliers
• Maximum word length (n+1) bits for taking
one bit at a time.
• Higher radix feasible.
• Area intensive
• Other methods exist such as using
Redundant Arithmetic, non-overlapping
multibit recoding
• 13.15 mod 23
• We do not want to do in a straight forward
manner .
• Write b = 13 in binary form:
• b3b2b1b0 =1101
• Do repeatedly starting from MSB:
• Old= (2.Old + bi.A) mod 23
EXAMPLE
• b3b2b1b0 =1101; A =15, mi = 23
P= (2.0 + 1.15) mod 23 = 15
P=(2.15 + 1.15) mod 23 = 22
P=(2.22 + 0.15) mod 23 = 21
P=(2.21+ 1.15) mod 23 = 11
Maximum value of P <3(23) i.e. 3mi
Modulo subtraction is by two comparisons:
Is P>N? or Is P>2n?
Answer is either P, P-mi, P-2mi; choose based on sign of P-mi,
P-2mi.
• Example 45 mod 23, anwers are 45,45-23=22,45-46=-1; since P2mi is negative and P-mi is positive, P-mi is the correct result.
• Multiple precision arithmetic to be used in PC based
implementations
•
•
•
•
•
•
•
•
Architecture for Modmul
Old
2Old
A
bi
TC of 2mi
LSB of
Zero
TC of mi
Adder
(n+2) bit adder
3:1 Mux
Latch
Latch
Adder
ModMUL
• Computation time= n[(n+2)DFA+DMux]
• Area = 3(n+2)AFA+A3:1MUX+nAAND
Modmul for IDEA
• IDEA (International Data Encryption
Algorithm) uses (xy) mod (216+1) as a
programmable S-Box (Substitution Box),
where x and y are 16 bit words.
• Ideal for DSPs
• Get P=xy a 32 bit word.
• Subtract MSB 16 bit word from LSB 16 bit
word. If negative, add (216+1)
RNS to Binary Conversion
•
•
•
•
•
•
•
•
•
•
•
CRT based
MRC based
CRT: RNS {m1,m2,m3} Residues {x1,x2,x3}
Define Mi=M/mi and M=m1m2m3
Decoded Binary number X
= [M1{(1/M1) mod m1}x1+ {M2 (1/M2) mod m2}x2+ M3{(1/M3) mod
m3}x3]mod M
e.g. {3,5,7} M=105, M1=35,M2=21,M3=15
(1/35) mod 3 = 2, (1/21) mod 5=1, (1/15) mod 7=1.
X= [70x1+21x2+15x3] mod 105
Consider (1,2,3), X = (70+42+45) mod 105 = 157 mod 105 = 52
Generally, Mi are large, Mi{(1/Mi) mod mi} are stored,involves
multiplication of these large numbers by xi in parallel and adding.
CRT Implementation
[M2(1/M2) mod m2]
X1
[M1(1/M1) mod m1]
X2
Multiplier
[M3(1/M3) mod m3]
X3
Multiplier
Multiplier
Mod M adder
X
• Modulo M adder may involve n subtractions for a n
moduli system
• Delay = D + D
MRC
m1
r1
- r3
(r1-r3) mod m1 = p
XA
UA
-UB
(UA-UB) mod m1 =r
XC
UC
•
•
•
•
m2
r2
- r3
(r2-r3) mod m2 =q
XB
UB
m3
r3
Example RNS {7,8,9}
7 8 9
1 2 3
-3 -3
5 7
x4 x1
6 7
-7
6
x1
6
X = 6.72+7.9+3 = 498
Note XA= (1/m3) mod m1 and
XB= (1/m3) mod m2, XC= (1/m2) mod m1
UC, UB and r3 are known as MRC digits.
X = UCm2m3+UBm3+r3 is always less than M.
MRC versus CRT
• MRC is sequential but avoids reduction modulo a large
number needed in CRT .
• MRC needs storage of multiplicative inverses, Modulo
subtraction and modulo multiplication, final addition of n
numbers for a n moduli RNS,
• Multiplicative inverses can be powers of two small
numbers such as 6 or 9 for powers of two related moduli
sets.
• Moduli set with all MIs of value unity also suggested e.g
{3,7,22}, Only modulo subtractions will do for evaluating
MRC digits; But multipliers are cumbersome.
• Generally need ROMs.
Architecture for XY mod 17
x3
x2
x1
x0
Write MSBs bi as (1- bi′)
y3
y2
y1
y0
y0x3 yox2 y0x1 yox0
1011
y1x3 y1x2 y1x1 y1x0 (y1x3)′ added 1
1101
y2x3 y2x2 y2x1 y2x0 (y2x3)′ (y2x2)′ added 3
1011
y3x3 y3x2 y3x1 y3x0 (y3x3)′ (y3x2)′ (y3x1)′ added 7
00001
101101
′
1011010
Modulo 17 adder
Adding 4 words in a CSA
1011
0001
1101
0111
10010 Added 1
1010
1111
00101 Added 1
0100 add 4 (correction
0111 term in a modulo
17 adder)
Scaling
• Division by a number
• E.g. RNS given {3,5,7}. Divide 99 (0,4,1)
by 11 (2,1,4).
• If division is exact, multiply 99 by
multiplicative inverse of 11.
• (1/11) = (2,1,2) =86 (Note (1/11) mod 3 = 2
etc.
• (99/11) = (0,4,1)x(2,1,2)= (0,1,4) =9
Scaling by arbitrary number when
division is not exact
•
•
•
•
•
•
•
•
•
•
•
•
•
Example 1 : 100/13 in RNS {3,5,7}
100 = (1,0,2}
Direct method by multiplying with (1/13) will not work.
100 = 1,0,2
(1/13) = 1,2,6
100/13 = 1,0,5 = 40 wrong.
First you need to find residue of 100mod 13 = 9.
Subtract from 100 to get (100-9)=91
100 = 1,0,2
9 = 0,4,2
91 = 1,1,0
(1/13) = 1,2,6
91/13 = 1,2,0 = 7.
Scaling by one modulus
•
•
•
•
•
•
•
•
•
Divide 100/7
100 = 1,0,2
Subtract residue 100mod 7 first =2
100
= 1, 0, 2
2
= 2, 2, 2
98
= 2, 3, 0
x(1/7) = x1 x3
= 2 4
Now you need to do base extension to get RNS number again
(2,4,0)
• Scaling by another modulus aso feasible in the same way.
• Note that MRC does this.
Scaled Residue /Montgomery’s Modular Multiplication
• Example: To evaluate (5.6) mod 13 = 4.
• Prescaling by 16: 5 = (5.16) mod 13 = 2, (6.16) mod 13
=5
• Montgomery step = [(5.16)(6.16)/16] mod 13 = (2.5/16)
mod 13 = (10/3) mod 13 = (10.9) mod 13 = 12.
• Result is obtained by post scaling: (12/16) mod 13 =
(12/3) mod 13 = 4.
• Prescaling is Binary to RNS conversion: Successive
multiplication by 2 and modulo reduction , (5.2) mod 13=
10, (10.2) mod 13 = (7.2) mod 13= 1, (1.2) mod 13 = 2.
• Post scaling is another Montgomery step.
• Montgomery step avoids modulo reduction. Only
conditional addition. If LSB is 1 add modulus, ignore
LSB.
• Example (2.5/16) mod 13.
• Four steps are needed.
• Each step a partial product is added and result scaled by
two.
• 2 = 0010 (binary)
• Computation of (0010)x5/16:
• Formula: (old value+ bix5)/2
• Old value =0.
• (0+0.5)/2= 0
• (0+1x5)/2 = (5+13)/2 = 9 since LSB of current result in
brackets is 1.
• (9+0.5)/2 = (9+13)/2 = 11
• (11+0.5)/2 = (11+13)/2 = 12.
• Addition of two numbers using a (n+1)-bit CPA, n AND
gates, n Flip-flops
•
Higher Radix Montgomery’s
Technique
•
•
•
•
•
•
•
•
•
•
•
Higher Radix possible.
16 or 8 or 4 bits at a time can be considered.
Example considering 4 bits at a time:
Consider [(10001100)/16] mod 23
Find (-1/23) mod 16=(-1/7) mod16 = 9
((-1/mi)
mod 2k)
Find 10001100 mod 16 = four LSBs= 12
(X mod 2k)
Find (12x9) mod 16 = 12
α= [(-X/mi)
mod 2k]
Find 10001100+12(23) = 11010 0000
(X+ αmi)
Ignore last 4 bits to get 26.
(X+ αmi)/2k
Need a multiplier mod 16 to get the multiple to be added.
Then addition of shifted versions of modulus (in this case
of radix 16, four shifted versions) using a CASA tree
followed by CPA.
Popular Powers-of-two related
moduli set
• (2n-1, 2n, 2n+1)
• Dynamic range <3n bits.
• Example 16 bit DSP needs n = 6; RNS
{63,64,65}
• RNS to binary conversion using CRT can
be done very fast.
• .

 1

B  
n n
 2 2 1




m1


n n
n n
n
x3 2 2  1 mod2 2  12  1

m1
 1 
 1 
x 2 2  1  
 x 2 12  1  

n
n
n
n
 2 12 1 
 2 2 1 
n
n
n
1
2
m1
n
The various multiplicative inverses used above are as follows:
 1  n

 mod2 1  2
 2 n2 n 1 
n 1


1

 mod2   1
 2 n 12 n 1 
n
 1 

 mod2  1  2
 2n2n 1 
n
n 1
1
• The beauty is these are powers of two related
facilitating easy implementation.
B  2n1 x1 2n 2n  1  x2 2n 12n  1  (2n1  1) x3 2n 2n 1mod2n 2n 12n  1
Subtract x2 from both sides
 n1 n n

2n
n n
(B  x2)  2 x1 2 2  1  x2 2   (2n1  1) x 2 2 1 mod 2n 2n 12n  1
3


Divide by 2n to get 2n MSBs of the result as
( B  x 2)
n
2
 2n1 x1 2n  1  x2 2n   (2n1  1) x3 2n  1mod 22n  1
• Example {7,8,9}
• [(32+4)x1-8x2+(36-1)x3] mod 63 yields 6
MSBs
Realization
• Andraros and Ahmad : Four 2n-bit words to be added
using two levels of Adders of rotated bits.
• Piestrak suggested using CSA two level with CPA using
end around carry for adding four 2n-bit words
• Delay - (4n+2) DFA, Area = (6n) AFA
• Suggested Low delay version (2n+2) DFA+DMUX also, 2n
A2:1MUXes needed.
• Dhurkadas (NPOL, Cochin) suggested simplification to
three 2n-bit inputs to be added
• Delay – (4n+2) DFA, Area = (4n) AFA
• Bhardwaj, Premkumar, Srikanthan [1998] suggested
using n-bit adders e.g Carry select adders n-bit
• Wang et al [2002] 2n-bit as well as n-bit adders three
converters.
{7,8,9} example (x1,x2,x3)
x1, x2 3 bit, x3 4 bit
x12x11x10, x22x21x20, x33x32x31x30
( B  x 2)
n
2
( B  x 2)
2
n
 2n 1 x1 2n  1  x2 2n   (2n 1  1) x3 2n  1mod 22 n  1


 (22 n 1  2n 1) x1  x2 2n   (22 n 1  2n 1  1) x3 mod 22 n  1
•[(32+4)x1-8x2+(36-1)x3] mod 63 :
x10 x12 x11 x10 x12 x11
x22′ x21′ x20′ 1 1 1
X3x x32 x31 x3x x32 x31
1
1 x33 ′ x32 ′ x31′ x30 ′
X3x= x30+x33 since either x30 or
x33 exist
Dhurkadas Simplified as
x10 x12 x11 x10 x12 x11
x22′ x21′ x20′ y x31′ x30 ′
X3x x32 x31 x30 x32 x31
Y= (x33+x32)′
Other three, Four and Five moduli
sets
• {2n,2n-1,2n-1-1} Hiasat and Abdel-Aty-Zohdy, Wang,
Wang, Swamy and Ahmad: not better than popular
moduli set, multipliers etc are simpler
• {2n,2n-1,2n+1-1} Ananda Mohan better in area or time,
multipliers are simpler
• {2n,22n-1,22n+1} Ananda Mohan better than Cao et al
four moduli set, one large modulus
• {2n,2n-1,2n+1, 2n+1-1 } Vinod and Premkumar
• {2n,2n-1,2n+1, 2n+1-1 } Bhardwaj, Srikanthan, Ananda
Mohan and Premkumar Area and Time intensive
• {2n,2n-1,2n+1, 22n+1} Cao et al better than other four
moduli sets but one modulus bigger in size.
• {2n-3,2n-1,2n+1,2n+3} Sheu et al uses ROM not attractive
• {2n-1-1, 2n-1,2n,2n+1,2n+1-1} Cao et al 2007 Increases
cardinality to 5, DR of 5n bits but RNS to Binary
conversion is slower/area consuming
Comparison of various converters for three
moduli sets
Converter
Moduli set
FA
HA
AND
/OR
XOR
/XNOR
Other
Delay
[8]
M2
6n-1
3n-7
----
-----
(n-1) MUX
4nDFA
[5]
M1
6n+1
----
n+3
n+1
2n MUX
(n+2)DFA+DMUX
[3,4]
M1
4n
---
2
---
----
(4n+1)DFA
[6] CI
M1
4n
1
-----
1
2 MUX
(4n+1)DFA
[6] CII
M1
6n
1
1
1
(2n+2) MUX
(n+1)DFA
[6] CIII
M1
4n
1
(2n+2)
(2n-1)
(2n+2) MUX
(n+1)DFA
Converter I
M4
4n+3
---
n
n
-----
(6n+5)DFA
Converter II
M4
14n+21
2n+3
---
---
(2n+1) 3:1MUX
(2n+7)DFA
Converter III
M4
12n+19
2n+2
---
---
10(2n+1)AROM
(2n+1) 2:1MUX
(2n+7)DFA
[9]
M3
37n+14
--
--
--
--
(14n+8)DFA
[12,13] 4-stage CE
M3
n2/2+11n+4
1
--
--
2 MUX
(11n+l+8)DFA
• M2 {2k,2k-1,2k-1-1}, M1{2k-1,2k,2k+1},
• M4{2k,2k-1,2k+1-1}, M3{2k-1,2k,2k+1,2k+1-1}
Base Extension
• Needed in scaling or division.
• Uses MRC fist to divide followed by base
extension.
• CRT can be used but is cumbersome.
Example: {3,5,7} 52= (1,2,3) Scale by 7
3 5 7
1 2 3
-3 -3
1 4
x1 x3
1 2
2 First Base Extension step
-2
2
X2
1 +(1x5)mod 7 Base Extension step
0
RSA using RNS/ECC
• Needs computation of PQ mod N
• e.g 1023 mod 37 = (1016)(104)(102)(101) mod 37
• Successive squaring mod 37 and Multiplications mod 37 of selected
results.
• Needs (XY) mod N ass basic step where X,Y,N are 1024 bit
numbers.
• RNS can be used.
• Montgomery technique has been used to find (X′Y′/M) mod N where
M is the product of Moduli in RNS.
• Needs two RNS dynamic ranges M and M′ which are mutually
Prime and a redundant modulus
• Determine q such that (X′Y′+qN) is a multiple of M.
• Extend q to RNS with Dynamic range M′.
• Find r = (X′Y′+qN)/M in second RNS
• Do base extension to First RNS
Sign Detection and Comparison
• Is difficult
• Needed to go to Binary number to detect
sign
• Comparison is also difficult Needed to go
to Binary numbers or sequential
techniques such as comparing Mixed
Radix Digits.
Applications
• FIR Filters (ensure that RNS dynamic
range is larger than that of the filter)
• Digital Frequency Synthesis
• Video Filters
• 2-D filters
• NTTs (Number Theoretic Transforms)
• Cryptography
Applications of RNS
•
•
•
•
•
•
•
[5] Freking, W.L., and Parhi, K.K., "Low-power FIR digital filters using residue
arithmetic, " in Conf. Record 31st Asil. Conf. Signals, Syst. and Comput.
(ACSSC 1997), vol. 1, Pacific Grove, CA USA [1997], 739-43.
[6] D'Amora, A. et al., "Reducing power dissipation in complex digital filters by
using the quadratic residue number system, " in Conf. Record 34th Asil. Conf.
Signals, Syst. Comput. (ACSSC 2000), vol. 2, Pacific Grove, CA USA [2000],
879-83.
[7] Cardarilli, G.C. et al., "Low-power implementation of polyphase filters in
Quadratic Residue Number system," in Proc. IEEE Int. Symp. Circuits Syst.
(ISCAS 2004), vol. 2, Vancouver, BC, Canada [2004], 725-728.
[8] Shanbag, N.R., and Siferd, R.E., A single-chip pipelined 2-D FIR filter using
residue Arithmetic, IEEE JSSC -26[1991], 796-805.
[9] Tuukka Toivonen., and Janne Heikkilä., Video Filtering With Fermat Number
Theoretic Transforms Using Residue Number System, IEEE CSVT-16[2006],
128-135.
[10] Schwemmlein, J., and Posch, K.C., Reinhard Posch. RNS-modulo
reduction upon a restricted base value set and its applicability to RSA
cryptography, Computer & Security [1998], 17, 637-650.
[11]Hanae Nozaki., Masahiko Motoyama., Atsushi Shimbo., and Shinichi
Kawamura., Implementation of RSA algorithm based on RNS Montgomery
multiplication, In C. Paar (ed). Cryptographic Hardware and Embedded
Systems – CHES, Springer-Verlag, Berlin, Germany [2001], 364-376.
• [12] Jean-Claude Bajard., Laurent Stephane Didier., Peter Kornerup.,
An RNS Montgomery modular multiplication Algorithm, IEEE C-47
[1998], 766-776.
• [13] Jean-Claude Bajard., and Laurent Imbert., A Full RNS
Implementation of RSA, IEEE C-53[2004],769-774.
• [14] Schinianakis, D.M., Kakarountas. A.P., and Stouraitis. T., A New
Approach to Elliptic Curve Cryptography: an RNS Architecture, IEEE
MELECON, May 16-19, Benalmádena (Málaga), Spain [2006], 12411245.
• [15] Lie-Liang Yang., and Lajos Hanzo., A Residue Number System
Based Parallel Communication Scheme Using Orthogonal Signaling:
Part I—System Outline, IEEE VT-51[2002],1534-1546.
• [16] Chaves, R., and Sousa, L., “RDSP: A RISC DSP based on
residue number system,” in Proc. Euro. Symp. Digital System
Design: Architectures, Methods, and Tools, Antalya, Turkey [2003],
128-135.
• [17] Wei, W. et al., "RNS application for digital image processing," in
4th IEEE Int. Workshop Syst.-on-Chip for Real Time Applications,
Banff, Alta., Canada [2004],77-80.
Conclusion
• Very mature today
• Can be used in place of Custom DSP
blocks
• Research on newer moduli sets with high
cardinality and Faster Reverse
Conversion is of interest