CSE 246: Computer Arithmetic Algorithms and Hardware Design Instructor: Prof. Chung-Kuan Cheng Winter 2004 Lecture 2
CSE 2462 Topics: Redundant Number Systems Residue Number Systems Mixed Radix Number Systems (as they pertain to conversions of Residue Number Systems)
CSE 2463 Redundant Number Systems Examine number systems that can be characterized by the 3-tuple (r, , ) r : radix, [, ] : set of digits The redundancy of such a system is said to be non-redundant if - = r redundant if - > r minimally redundant if - = r + 1 maximally redundant if - =2r - 1 over redundant if - > 2r - 1
CSE 2464 Redundant Number Systems Ex: (2,0,2) Redundant - furthermore, minimally redundant Addition example: (6) (7) [3 not in radix] = 4*2 + 2*2 +1 = 13 =
CSE 2465 Redundant Number Systems More addition examples:
CSE 2466 Redundant Number Systems What has redundancy bought us? From the examples, it seems carry propagation has been somewhat reduced Intuitively - redundancy acts as a buffer against carry propagation However, carry propagation was not eliminated in (2,0,2) Need to formalize the degree to which a particular degree of redundancy eliminates carry propagation
CSE 2467 Redundant Number Systems To formalize: x i+1 x i where x i + y i = r*t i+1 + w i + y i+1 y i t i+1 w i t i+2 w i+1 t i+1 + w i+1 we must bound t i+1 + w i+1 for all i to eliminate carry propagation
CSE 2468 Redundant Number Systems As x i and y i are in [, ] and x i + y i = r * t i+1 + w i+1 12 r*t i+1 + w i 2 suppose there exist and such that 2 t i+1 we need 3 - w i+1 - to avoid propagation. Since this must be true for all i, whether the subscript is i or i+1 is not important. From 2 & 3 we have 4r* + - r*t i+1 + w i+1 r* + - From 1 & 4 we have 52 r* + -
CSE 2469 Redundant Number Systems 52 r* + - gives the following bounds for & to insure no propagation /(r-1) /(r-1) For (2,0,2): /(r-1) -> 2/(2-1) /(r-1) -> 0/(2-1) which, from 3 means all w i must be in the range [0,0] and there exist inputs requiring carry propagation
CSE Redundant Number Systems How then to ensure that no carry propagation is needed? Change (r, , ) Consider (3,0,5) /(r-1) -> 5/(3-1) /(r-1) -> 0/(3-1) yields : 3 (first integer after 5/2) : 0 From 3, this gives a range on w i of [0,2] Thus, when presented with multiple possibilities for representing the sum of two digits, always choose terms in [0,2] and there will be no carry propagation
CSE Redundant Number Systems Ex: for (3,0,5) = 8 could use 15, but choose from [0,2] 22 yielding 22 and no propagation
CSE Redundant Number Systems Advantages Constant time addition/subtraction! Disadvantages More space Comparison no unique forms/representations conversion to canonical form as expensive as summation Uses: excellent for intermediate results of addition Multiplication DSP loops that have no comparisons
CSE Residue Number Systems Define a Residue Number System as follows For any given integer x, x=(x 1 |x 2 |...|x k )RNS(P 1 |P 2 |...|P k ) where x i = x mod P i and i,j P i is relatively prime to P j EX: 84 = (0|4|0)RNS(7|5|3) 1 = (1|1|1)RNS(7|5|3) 2 = (2|2|2)RNS(7|5|3) 3 = (3|3|0)RNS(7|5|3) Residue numbers are not positional. Residue numbers have unique representations mod i.e. for (7|5|3) there are unique representations of
CSE Residue Number Systems Need a conversion system to/from binary What benefit to doing operations in RNS? Binary # RNS ops +,-,* Binary # RNS #
CSE Residue Number Systems Addition (subtraction is similar) x+y = ((x 1 +y 1 ) p 1 | (x 2 +y 2 ) p 2 |... | (x k +y k ) p k )RNS(P 1 |P 2 |...|P k ) where x i = (x) p i and y i = (y) p i Multiplication x*y = ((x 1 *y 1 ) p 1 | (x 2 *y 2 ) p 2 |... | (x k *y k ) p k )RNS(P 1 |P 2 |...|P k ) where x i = (x) p i and y i = (y) p i Division ? Hard. What does a fraction look like in RNS?
CSE Residue Number Systems Advantages: Parallel processing of +,*,- on smaller numbers Adding more primes without increasing range allows for use of fields to assist in fault tolerance (Ex: go from (7|5|3) to (7|5|3|2) where last field is used for parity) Disadvantages: No division Comparison non-trivial Conversion costs How to do conversion?
CSE Residue Number Systems Conversion: RNS -> Binary # Given (x 1 |x 2 |...|x k )RNS(P 1 |P 2 |..|P k ) Binary number x = mod where i = inv One may consider, for the purposes of computation, all i as being a pre-defined part of each particular RNS system See: Chinese Remainder Theorem
CSE Residue Number Systems Conversion -> Bin# example: 84 = (0|4|0)RNS(7|5|3) For (7|5|3): ( 1 * 5*3/7) 7 = 1 1 : 1 ( 2 * 7*3/5) 5 = 1 2 : 1 ( 3 * 7*5/3) 3 = 1 3 : 2 x = (0* 1 *5*3*7/7 + 4* 2 *7*3*5/5 + 0* 3 *7*5*3/3) 105 x = (0 + 4* 2 *21 + 0) 105 x = (4*1*21) 105 =
CSE Residue Number Systems Conversion -> Bin # example: 1 = (1|1|1)RNS(7|5|3) ( 1 * 5*3/7) 7 = 1 1 : 1 ( 2 * 7*3/5) 5 = 1 2 : 1 ( 3 * 7*5/3) 3 = 1 3 : 2 x = (1* 1 *5*3*7/7 + 1* 2 *7*3*5/5 + 1* 3 *7*5*3/3) 105 x = (1* 1 *15 + 1* 2 *21 + 1* 3 *35) 105 x = (1*1*15+1*1*21+1*2*35) 105 =(106) 105 = 1 105
CSE Residue Number Systems Minimize: total length of primes (determining range) max length of any prime (determining HW cost/delay per unit) # of primes (determining # of parallel units) Popular choices: 2 r -1 (2,3 particularly popular) Thm: 2 a -1 & 2 b -1 are relatively prime iff a & b are relatively prime Binary # RNS ops +,-,* Binary # RNS # Conversion -> Chinese remainder theorem How to choose primes? What about Bin -> RNS?
CSE Residue Number Systems Binary # to RNS conversion Uses a lookup table Observation: for n digit binary number y (y n-1, y n-2,...,y 0 ) p i = [(2 n-1 ) p i *y n-1 + (2 n-2 ) p i *y n (2 0 ) p i *y 0 ] p i Use a table to store (2 i ) p i Example for (7|5|3) (1011)=(x 1 |x 2 |x 3 )RNS(7|5|3) x 1 = (1+2+1) 7 = 4 x 2 = (3+2+1) 5 = 1 x 3 = (2+2+1) 3 = 2
CSE Residue Number Systems Comparison Could convert back and forth to/from binary. Another approach: convert to a mixed radix system, as numbers in a mixed radix system are comparable.
CSE Mixed Radix Number Systems We shall describe a Mixed Radix System as follows: x =(Z k-1 |Z k-2 |...Z 0 )MRS(P k-1 |P k-2 |...|P 1 ) x = Z k-1 P k-1 P k-2...P 1 + Z k-2 P k-2 P k-1...P Z 1 P 1 +Z 0
CSE Mixed Radix Number Systems Conversion from RNS to MRS Given x = (x k-1 |x k-2 |...|x 0 )RNS(P k-1 |P k-2 |...|P 0 ) We want x=(Z k-1 |Z k-2 |...|Z 0 )MRS(P k-2 |P k-3 |...|P 0 ) Observation: The MRS digit Z 0 is in units of 1, so fewer primes needed in MRS than in RNS
CSE Mixed Radix Number Systems Question: what is the relationship between x 0 and Z 0 ? Can it be found by simple inspection? Yes. 1x 0 = Z 0. Why? x 0 is the residue left from x mod P 0 - all other terms are multiples of P 0 -> x 0 = Z 0
CSE Mixed Radix Number Systems This yields 2x-x 0 = (x ’ k-1 |x ’ k-2 |...|x ’ 1 |-)RNS(P k-1 |P k-2 |...|P 1 |-) = (Z k-1 |Z k-2 |...|Z 1 |0)MRS(P k-2 |P k-3 |...|P 0 ) where x ’ i = (x i -x 0 ) p i Note that this is the only change in MRS
CSE Mixed Radix Number Systems Which leads to 3(x-x 0 )/P 0 = (x ” k-1 |x ” k-2 |...|x ” 1 |-)RNS(P k-1 |P k-2 |...|P 1 |-) = (Z k-1 |Z k-2 |...|Z 1 ) MRS (P k-2 |P k-1 |...|P 1 ) Z 1 = x ” 1 and so forth. (Deduction, division, repeat) However, it was earlier noted that division in RNS is hard - yet here we are doing division. The trick? In this case, we know that we will always get integer results.
CSE Back to RNS Division in RNS x ” i =x i * (P 0 -1 ) p i where (P j -1 ) p i is the multiplicative inverse of P j with respect to P i Ex:(3 -1 ) 7 = 5 --> (3*(3 -1 ) 7 ) 7 = 1 (3 -1 ) 5 = 2 --> (3*(3 -1 ) 5 ) 5 = 1
CSE Example: RNS -> MRS Y=(1|3|2)RNS(7|5|3) = (Z 2 |Z 1 |Z 0 )MRS(5|3) That is, Y = Z 2 *5*3+Z 1 *3+Z 0 by 1 (see slide #25) Z 0 = x 0 = 2 from 2 (see slide #26) we have y-x 0 = y-2 =(x ’ 2 |x ’ 1 |0)RNS(7|5|3) = (Z 2 |Z 1 |0)MRS(5|3) x ’ 2 = (x 2 -x 0 ) p 2 = (1-2) 7 = 6 x ’ 1 = (3-2) 5 = 1 y-2 = (6|1|0)RNS(7|5|3)=(Z 2 |Z 1 |0)MRS(5|3)
CSE Example continued. 3 (See slide #27) then gives (y-x 0 )/P 0 = (y-2)/3 = (x ” 2 |x ” 1 |-)RNS(7|5|3) = (Z 2 |Z 1 )MRS(5) then Q: How does one derive x ” 2 ? A: “ It ’ s hard. ” One has to try values one by one up to the modulus remembering that (3 -1 ) 7 = 5, (3 -1 ) 5 = 2 x ” 2 = (5*6) 7 = 2, x ” 1 =(1*2) 5 = 2 (y-2)/3 = (2|2|-)RNS(7|5|3) = (Z 2 |Z 1 )MRS(5)
CSE Ex. Cont. Apply 1 again, Z 1 =x ” 1 =2 gives by 2 (y-x 0 )/P 0 - x ” 1 = (x ”’ 2 |0|-)RNS(7|5|3) = (Z 2 |0)MRS(5) x ”’ 2 =(x ” 2 -2) 7 = (2-2) 7 = 0 (0|0|-)RNS(7|5|3)=(Z 0 |0)MRS(5) By 3 x ”” 2 = (x ”’ 2 *(5 -1 ) 7 ) 7 = 0 = Z 2 Yields final result (0|2|2)MRS(5|3)
CSE Ex. Concluded. Check correctness (0|2|2)MRS(5|3) = 0*5*3+2*3+2 = = 1, 8 5 = 3, 8 3 = 2 -> (1|3|2)RNS(7|5|3) Correct! Closing remark / Moral of the examples - “ Inversion is key ”