Patent Application
20060269054
The present invention relates to apparatus operative to accelerate cryptographic co-processing peripherals and additionally but not exclusively to an accelerated processing apparatus for polynomial based and prime number field arithmetic, extending the range of computational fields of integers and width of serial input operands in modular arithmetic public key cryptographic coprocessors designed for elliptic curve and RSA type computations.
Security enhancements and performance accelerations for computational devices are described in Applicant's U.S. Pat. No. 5,742,530, hereinafter “P1”, U.S. Pat. Nos. 5,513,133, 5,448,639, 5,261,001; and 5,206,824 and published PCT patent application PCT/IL98/00148 (WO98/50851); and corresponding U.S. patent application Ser. No. 09/050958, hereinafter “P2”, Onyszchuk et al's U.S. Pat. No. 4,745,568; Omura et al's U.S. Pat. No. 4,5877,627, and applicant's U.S. patent application Ser. No. 09/480,102; the disclosures of which are hereby incorporated by reference. Applicant's U.S. Pat. No. 5,206,824 shows an early apparatus operative to implement polynomial based multiplication and squaring, which cannot perform operations in the prime number field, and is not designed for interleaving in polynomial based computations. An additional analysis is made of an approach to use the extension field in polynomial based arithmetic in Paar, C., F. Fleischmann and P. Soria-Rodriguez, “Fast Arithmetic for Public-Key Algorithms in Galois Fields with Composite Exponents”, IEEE Transactions on Computers, vol. 48, No. 10, October 1999, henceforth “Paar”. W. Wesley Peterson and E. J. Weldon Jr., in the second edition of “Error-Correcting Codes”, published by the MIT Press, Cambridge, Mass., 1972, pages 174-179, demonstrated circuits for performing division in the polynomial based residual number system GF(2q), hereinafter, “Peterson”. Peterson's circuit can only be used in a device where the multiplier is exactly the length of the modulus. Typically, that would demand a device that would be more than twice as long as present devices, and would not be cost effective for compact implementations. It could not be used in interleaved implementations, and could not be useful where l is longer than 1, as he has not provided an anticipatory device for determining the Y0 of a multibit character.
Whereas, Knuth [D. Knuth, The art of computer programming, vol. 2: Seminumerical algorithms, Addison-Wesley, Reading Mass., 1981] page 407, implies that using an ordinary division process on a single l bit character in polynomial based division, we can assume a method to anticipate the next character in the quotient, this invention discloses a method for anticipating the next character of a quotient deterministically using a logic configuration.
It is an aim of the present invention to provide a microelectronic specialized arithmetic unit operative to perform large number computations in the polynomial based and prime integer based number fields, using the same anticipating methods for simultaneously performing interleaved modular multiplication and reduction on varied radix multipliers.
A further aim of the invention also relates to a compact microelectronic specialized arithmetic logic unit, for performing modular and normal (natural, non-negative field of integers) multiplication, division, addition, subtraction and exponentiation over very large integers. When referring to modular multiplication and squaring using both Montgomery methods and a reversed format method for simplified polynomial based multiplication and squaring, reference is made to the specific parts of the device as a superscalar modular arithmetic coprocessor, or SMAP, or MAP or SuperMAP™, also as relates to enhancements existing in the applicant's U.S. patent pending Ser. No. 09/050,958 filed Mar. 31, 1998 and a continuation in part, Ser. No. 09/480,102 filed on Jan. 10, 2000.
Preferred embodiments of the invention described herein provide a modular computational operator for public key cryptographic applications on portable Smart Cards, typically identical in shape and size to the popular magnetic stripe credit and bank cards. Similar Smart Cards as per applicant's technology of U.S. Pat. Nos. 5,513,133 and 5,742,530, and applicants above mentioned pending applications are being used in the new generation of public key cryptographic devices for controlling access to computers, databases, and critical installations; to regulate and secure data flow in commercial, military and domestic transactions; to decrypt scrambled pay television programs, etc. and as terminal devices for similar applications. Typically, these devices are also incorporated in computer and fax terminals, door locks, vending machines, etc.
The preferred architecture is of an apparatus operative to be integrated to a multiplicity of microcontroller and digital signal processing devices, and also to reduced instruction set computational designs while the apparatus operates in parallel with the host's processing unit.
This embodiment preferably uses only one multiplying device which inherently serves the function of two or three multiplying devices, basically similar to the architecture described in applicant's U.S. Pat. No. 5,513,133 and further enhanced in U.S. patent application Ser. No. 09/050,958 and PCT application PCT/IL98/0048. Using present conventional microelectronic technologies, the apparatus of the present invention may be integrated with a controlling unit with memories onto a smart card microelectronic circuit.
The main difference between hardware implementations in the polynomial based field, and in the prime number field, is that polynomial based additions and subtractions are simple XOR logic operations, without carry signals propagating from LS to MS. Consequently, there is no interaction between adjacent cells in the hardware implementation, and subtraction and addition are identical procedures. The earliest public notice that the authors are aware of was a short lecture by Marco Bucci of the Fondazione Ugo Bordoni, at the Eurocrypt Conference Rump Session in Perugia, Italy, in 1994, though, even then, this configuration was well known to engineers practiced in the art.
Previous applicant's apparati, described in P1 and P2, were typically prepared to efficiently compute elliptic curve cryptographic protocols in the GF(p) field. For use in the GF(2q) field, in this invention, we show that as there is no interaction between adjacent binary bits in the polynomial field, computations can be processed efficiently, simultaneously performing reduction and multiplication on a supersealar multiplication device without introducing Montgomery functions and Montgomery parasites. Multiplication in GF(2q) is performed as the computation machine preferably starts from the most significant partial products. Reduction is performed by adding as many moduli as are necessary to reset MS ones to zeroes. As there is no carry out, in these additions, results are automatically modularly reduced. In this invention polynomial computations are performed using the same architecture, wherein, in GF(2q) the operands are fed in MS characters first, wherein all internal carry signals are forced to zero. GF(p) computations are preferably executed as in P1 and P2, wherein LS characters are processed first and MS characters last.
The architecture has been extended to allow for a potentially faster progression, in that serial multipliers are l bit wide characters, and at each clock, an I bit character is emitted from the carry save accumulator, the CSA. This has somewhat complicated the anticipation process (Y0), in that for single bit wide busses, an inversion of an odd number over a mod 2x base is also an odd number, and the least significant bit of the multiplicand (J0) was always a one. For both number fields, however the reduction process is identical, assuming the switch out of (suppression of) carries, if we remember that our only aim is to output a k character zero string, and we regard the Y0 function only as a zero forcing vector.
The present invention also seeks to provide an architecture for a digital device, which is a peripheral to a conventional digital processor, with computational, logical and architectural novel features relative to the processes described in U.S. Pat. No. 5,513,133.
A concurrent process and a hardware architecture are provided, to perform modular exponentiation without division preferably with the same number of operations as are typically performed with a classic multiplication/division device, wherein a classic device typically performs both a large scale multiplication and a division on each operation. A particular feature of a preferred embodiment of the present invention is the concurrency of larger scale anticipatory zero forcing functions, the extension of number fields, and the ability to integrate this type unit for safe communications.
The advantages realized by a preferred embodiment of this invention result from a synchronized sequence of serial processes. These processes are merged to simultaneously (in parallel) achieve three multiplication operations on n character operands, using one multiplexed k character serial/parallel multiplier in n effective clock cycles, where the left hand final k characters of the result reside in the output buffer of the multiplication device. This procedure accomplishes the equivalent of three multiplication computations in both fields, as described by Montgomery, for the prime number field and the equivalent of two multiplications and a division process in GF(2q).
By synchronizing loading of operands into the SuperMAP and on the fly detecting values of operands, and on the fly preloading and simultaneous addition of next to be used operands, the apparatus is operative to execute computations in a deterministic fashion. All multiplication and exponentiation circuitry is preferably added which on the fly preloads, three first k character variables for a next iteration squaring sequence. A detection device is preferably provided where only two of the three operands are chosen as next iteration multiplicands, eliminating k effective clock cycle wait states. Conditional branches are replaced with local detection and compensation devices, thereby providing a basis for a simple control mechanism. The basic operations herein described are typically executed in deterministic time in GF(p) using a device described in U.S. Pat. No. 5,513,133 to Gressel et al or devices as by STMicroelectronics in Rousset, France, under the trade name ST19-CF58.
The apparatus of the present invention has particularly lean demands on external volatile memory for most operations, as operands are loaded into and stored in the device for the total length of the operation. The apparatus preferably exploits the CPU onto which it is appended, to execute simple loads and unloads, and sequencing of commands to the apparatus, whilst the MAP performs its large number computations. Large numbers presently being implemented on smart card applications range from 128 bit to 2048 bit natural applications. The exponentiation processing time is virtually independent of the CPU, which controls it. In practice, architectural changes are typically unnecessary when appending the apparatus to any CPU. The hardware device is self-contained, and is preferably appended to any CPU bus.
In general, the present invention also relates to arithmetic processing of large integers. These large numbers are typically in the natural field of (non-negative) integers or in the Galois field of prime numbers, GF(p), composite prime moduli, and polynomial based numbers in GF(2q). More specifically, preferred embodiments of the present invention seek to provide devices that can implement modular arithmetic and exponentiation of large numbers. Such devices are suitable for performing the operations of Public Key Cryptographic authentication and encryption protocols, which, in the prime number field work over increasingly large operands and which cannot be executed efficiently with present generation modular arithmetic coprocessors, and cannot be executed securely in software implementations. Preferably, the same general architecture is used in elliptic curve implementations, on integers, which are orders of magnitude smaller. Using the novel reverse mode method of multiplication, polynomial arithmetic is advantageous as generating zeroes does not encumber computations with the parasitic 2−n factor.
The architecture offers a modular implementation of large operand integer arithmetic, while allowing for normal and smaller operand arithmetic enabled by widening the serial single character bus, i.e., use of a larger radix. Typically, this is useful for accelerating computations, for reducing silicon area for implementing the SuperMAP, and for generating a device of length compatible with popular Digital Signal Processors (DSP).
For modular multiplication in the prime and composite field of odd numbers, A and B are defined as the multiplicand and the multiplier, respectively, and N is defined as the modulus in modular arithmetic. N, is typically larger than A or B. N also denotes the composite register where the value of the modulus is stored. N, is, in some instances, typically smaller than A. A, B, and N are typically n characters long, where characters are typically one to 8 bits long. k the number of l bit characters in the size of the group defined by the size (number of cells) of the multiplying device. Similarly, in polynomial based GF(2q) computations, the modulus, N, is n bits long wherein the MS bit is a one (a monic), and the A, S and B operands are also, when properly reduced, n bits long. If a result of a GF(2q), computation is monic it is preferably “reduced” to a value with an MS zero, by XORing said result value with the modulus. In a preferred embodiment, as the first significant bit of a GF(2q) is formed in the reverse mode, the MAP can sense if the bit is a one, and perform the preferred reduction.
In the prime field, ≡, or in some instances =, is used to denote congruence of modular numbers, for example 16≡2 mod 7. 16 is termed “congruent” to 2 modulo 7 as 2 is the remainder when 16 is divided by 7. When Y mod N≡X mod N; both Y and X may be larger than N; however, for positive X and Y, the remainders are identical. Note also that the congruence of a negative integer Y, is Y+u·N, where N is the modulus, and if the congruence of Y is to be less than N, u is the smallest integer which gives a positive result.
In GF(2q) congruence is much simpler, as addition and subtraction are identical, and normal computations typically do not leave a substantial overflow. For N=1101 and A=1001, as the left hand MS bit of A is 1, we must reduce (“subtract”) N from A by using modulo 2 arithmetic, A XOR N=1001 XOR 1101=0100.
The Yen symbol, ¥, is used to denote congruence in a limited sense, especially useful in GF(p). During the processes described herein, a value is often either the desired value, or equal to the desired value plus the modulus. For example X¥2 mod 7. X can be equal to 2 or 9. X is defined to have limited congruence to 2 mod 7. When the Yen symbol is used as a superscript, as in B¥, then 0≦B¥<2 N, or stated differently, B¥ is either equal to the smallest positive B which is congruent to B¥, or is equal to the smallest positive congruent B plus N, the modulus. Other symbols, specific to this invention appear later in this summary.
When X=A mod N, X is defined as the remainder of A divided by N; e.g., 3=45 mod 7, and much simpler in GF(2q)−1111 mod 1001≡0110.
In number theory, the modular multiplicative inverse of X is written as X−1, which is defined by X·X−1 mod N=1. If X=3, and N=13, then X−1=9, i.e., the remainder of 3·9 divided by 13 is 1 in GF(p).
For both number fields, we typically choose to compute the multiplicative inverse of A using the exponential function, e.g., A−1 mod q≡Aq−2 mod q.
The acronyms MS and LS are used to signify “most significant” and “least significant”, respectively, when referencing bits, characters, and full operand values, as is conventional in digital nomenclature, but in the reversed mode polynomial base, operands are loaded MS data first and LS last, wherein the bit order of the data word is reversed when loaded.
Throughout this specification N designates both the value N, and the name of the shift register which stores N. An asterisk superscript on a value, denotes that the value, as stands, is potentially incomplete or subject to change. A is the value of the number which is to be exponentiated, and n is the bit length of the N operand. After initialization when A is “Montgomery P field normalized” to A* (A*=2nA—explained in P1) A* and N are typically constant values throughout the intermediate step in the exponentiation. In GF(2q) computations where computations might be performed with the normal unreversed positioning of bits we would be bound by this same protocol. However, using the reversed format, our computations generate most significant zeroes, which are disregarded, and do not represent a multiplication shift, as there is no carry out.
During the first iteration, after initialization of an exponentiation, B is equal to A*. B is also the name of the register wherein the accumulated value that finally equals the desired result of exponentiation resides. S or S* designates a temporary value; and S designates the register or registers in which all but the single MS bit of a number in GF(p) S is stored. (S* concatenated with this MS bit is identical to S.) S(i−1) denotes the value of S at the outset of the i'th iteration. In these polynomial computations there is no need to perform modular reduction on S.
Typically, Montgomery multiplication of X and Y in the prime number field is actually an execution of (X·Y·2−n) mod N, where n is typically the number of characters in a modulus. This is written, P(A·B)N, and denotes MM or multiplication in the P field. In the context of Montgomery mathematics, we refer to multiplication and squaring in the P field and in the polynomial based field as multiplication and squaring operations.
We will redefine this innovative extension of a Montgomery type arithmetic in the GF(2q) to mean a reversed format data order, wherein MS zero forcing does not change congruence, or initiate a burdensome parasitic factor. We may thus introduce a new set of symbols to accommodate the arithmetic extension, and to enable an architecture with wider serial multiplier buses. Such a more than one bit serial multiplier stream is preferable for enabling a natural integer superscalar multiplier. Such multiplier device may accept 32 bit multiplicands and 4 bit multipliers, into an apparatus that can perform modular arithmetic multiplications and reductions simultaneously.
Symbols in the Serial/Parallel Super Scalar Modular Multiplier Enhancement
According to a first aspect of the present invention, there is therefore provided a microelectronic apparatus for performing {circle around (×)} multiplication and squaring in both polynomial based GF(2q) and GF(p) field arithmetic, squaring and reduction using a serial fed radix 2l multiplier, B, with k character multiplicand segments, Ai, and a k character {circle around (⊕)} accumulator wherein reduction to a limited congruence is performed “on the fly”, in a systolic manner, with Ai, a multiplicand, times B, a multiplier, over a modulus, N, and a result being at most 2n+1 characters long, including the k first emitting disregarded zero characters, which are not saved, where k characters have no less bits than the modulus, wherein said operations are carried out in two phases, the apparatus comprising;
a first (B), and second (N) main memory register means, each register operative to hold at least n bit long operands, respectively operative to store a multiplier value designated B, and a modulus, denoted N, wherein the modulus is smaller than 2n;
a digital logic sensing detector, Y0, operative to anticipate “on the fly” when a modulus value is to be {circle around (⊕)} added to the value in the {circle around (⊕)} adder accumulator device such that all first k characters emitting from the device are forced to zero;
a modular multiplying device for at least k character input multiplicands, with only one, at least k characters long {circle around (⊕)} adder, {circle around (⊕)} summation device operative to accept k character multiplicands, the {circle around (×)} multiplication device operative to switch into the {circle around (⊕)} accumulator device, in turn, multiplicand values, and in turn to receive multiplier values from a B register, and an “on the fly” simultaneously generated anticipated value as a multiplier which is operative to force k first emitting zero output characters in the first phase, wherein at each effective machine cycle at least one designated multiplicand is {circle around (⊕)} added into the {circle around (⊕)} accumulation device;
the multiplicand values to be switched in turn into the {circle around (⊕)} accumulation device consisting of one or two of the following three multiplicands, the first multiplicand being an all-zero string value, a second value, being the multiplicand Ai, and a third value, the N0 segment of the modulus;
an anticipator to anticipate the l bit k character serial input Y0 multiplier values;
the apparatus being operable to input in turn multiplier values into the multiplying device said values being first the B operand in the first phase, and concurrently, the second multiplier value consisting of the Y0, “on the fly” anticipated k character string, to force first emitting zeroes in the output;
the apparatus further comprising an accumulation device, {circle around (⊕)}, operative to output values simultaneously as multiplicands are {circle around (⊕)} into the {circle around (⊕)} accumulation device;
an output transfer mechanism, in the second phase operative to output a final modular {circle around (×)} multiplication result from the {circle around (⊕)} accumulation device.
According to a preferred embodiment {circle around (⊕)} summations into the {circle around (⊕)} accumulation device are activated by each new serially loaded higher order multiplier characters.
Preferably, the multiplier characters are operative to cause no {circle around (⊕)} summation into the {circle around (⊕)} accumulation device if both the input B character and the corresponding input Y0 character are zeroes;
are operative to {circle around (⊕)} add in only the Ai multiplicand if the input B character is a one and the corresponding Y0 character is a zero;
are operative to {circle around (⊕)} add in only the N, modulus, if the B character is a zero, and the corresponding Y0 character is a one; and
are operative to {circle around (⊕)} add in the {circle around (⊕)} summation of the modulus, N, with the multiplicand Ai if both the B input character and the corresponding Y0 character are ones.
Preferably, the apparatus is operative to preload multiplicand values Ai and N, into two designated preload buffers, and to {circle around (⊕)} summate these values into a third multiplicand preload buffer, obviating the necessity of {circle around (⊕)} adding in each multiplicand value separately.
Preferably, the multiplier character values are arranged for input in serial single character form, and wherein the Y0 detect device is operative to anticipate only one character in a clocked turn.
In a preferred embodiment wherein the {circle around (⊕)} accumulation device performs modulo 2 computations, XOR addition/subtraction, wherein all carry bits in addition and subtraction components are disregarded, thereby precluding provisions for overflow and further limiting convergence in computations.
Preferably, carry inputs are disabled to zero, denoted, L=0, typically operative to perform polynomial based multiplication.
Preferably the apparatus is operative to provide non-carry arithmetic by omitting carry circuitry, such that an L equal to zero acting on an element in a circuit equation computing in GF(2q), L designates omitted circuitry and all adders and subtractors, designated {circle around (⊕)} to XOR, modulo 2 addition/subtraction elements.
A preferred embodiment is adapted such that the first k character segments emitted from the operational unit are zeroes, being controlled by the following four quantities in anticipating the next in turn Y0 character:
i. the l bit Sout bits of the result of the l bit by l bit mod 2l {circle around (×)} multiplication of the right-hand character of the Ai register times the Bd character of the B Stream, A0·{circle around (×)} Bd mod 2l;
ii. the first emitting carry out character from the {circle around (⊕)} accumulation device, L(CO0);
iii. the l bit Sout character from the second from the right character emitting cell of the {circle around (⊕)} accumulation device, SO1;
iv. the l bit J0 value, which is the negative multiplicative inverse of the right-hand character in the N0 modulus multiplicand register,
wherein values, A0{circle around (×)}Bd mod 2l, L(CO0), and SO1 are {circle around (⊕)} added character to character together and “on the fly” multiplied by the J0 character to output a valid Y0 zero-forcing anticipatory character to force an l bit egressing string of zeroes.
The apparatus is preferably operable to perform multiplication on polynomial based operands performed in a reverse mode, multiplying from right hand MS characters to left hand LS characters, operative to perform modular reduced {circle around (×)} multiplication without Montgomery type parasitic functions.
Preferably, the apparatus further comprises preload buffers which are serially fed and where multiplicand values are preloaded into the preload buffers on the fly from one or more memory devices.
The apparatus is preferably operative to {circle around (⊕)} summate into a multiplication stream a previous value, emitting from an additional n bit S register, via an l bit {circle around (⊕)} adder circuit such that first emitting output characters are zeroes when the Y0 detector is operative to detect the necessity of {circle around (⊕)} adding moduli to the {circle around (⊕)} summation in the {circle around (⊕)} accumulation device, wherein the Y0 detector operates to detect, utilizing the next in turn {circle around (⊕)} added characters A-{circle around (×)}Bd mod 2l, L(CO0), SO1, Sd and L(COz), the composite of {circle around (⊕)} added characters to be finite field {circle around (×)} multiplied on the fly by the l bit J0 value, where {circle around (⊕)} defines the addition and {circle around (×)} defines the multiplication as befits the finite field used in the process.
Preferably, for l=1, J0 is implicitly 1, and the J0 {circle around (×)} multiplication is implicit, without additional hardware.
Preferably, a comparator is operative to sense a finite field output from the {circle around (×)} modular multiplication device, whilst working in GF(p), where the first right hand emitting k zero characters are disregarded, where the output is larger than the modulus, N, thereby operative to control a modular reduction whence said value is output from the memory register to which the output stream from the multiplier device is destined, and thereby precluding allotting a second memory storage device for the smaller product values.
Preferably, for {circle around (×)} modular multiplication in the GF(2q), the apparatus is operative to multiply without an externally precomputed more than l bit zero-forcing factor.
A preferred embodiment is operative to compute a J0 constant by resetting either the A operand value or the B operand value to zero and setting the partial result value, S0, to 1.
According to a second aspect of the present invention there is provided a microelectronic apparatus for performing interleaved finite field modular multiplication of integers A and B, so as to generate an output stream of A times B modulus N wherein a number of characters in a modulus operand register, n, is larger than a segment length of k characters, wherein the {circle around (×)} multiplication process is performed in a plurality of interleaved iterations, wherein at each interleaved iteration with operands input into a {circle around (×)} multiplying device, consisting of N, the modulus, B, a multiplier, a previously computed partial result, S, and a k character string segment of A, a multiplicand, the segments progressing from the A0 string segment to the Am−1 string segment, wherein each iterative result is {circle around (⊕)} summated into a next in turn S, temporary result, in turn, wherein first emitting characters of iterative results are zeroes, the apparatus comprising:
first (B), second (S) and third (N) main memory registers, each register respectively operative to store a multiplier value, a partial result value and a modulus;
a modular multiplying device operative to {circle around (⊕)} summate into the {circle around (⊕)} accumulation device, in turn one or two of a plurality of multiplicand values, during each one of a plurality of phases of the iterative {circle around (×)} multiplication process, and in turn to receive as multipliers, in turn, inputs from:
said B register,
an “on the fly” anticipating value, Y0, being usable as a multiplier to force first emitting right-hand zero output characters in each iteration, and
said N, register;
the multiplicand parallel registers operative at least to receive in turn, values from the A, B, and N register sources, and in turn, also a multiplicand zero forcing Y0, value;
the apparatus further utilizing the Y0 detect device operative to generate a binary string operative to be a multiplier during the first phase and operative to be a multiplicand in the second phase;
the apparatus being operable to obtain multiplicand values suitable for switching into the {circle around (⊕)} accumulation device for the first phase consisting of a first zero value, a second value, Ai, which is a k character string segment of a multiplicand, A, and a third value N0, being the first emitting k characters of the modulus, N;
the apparatus further being operable to utilize a temporary result value, S, resulting from a previous iteration, operative to be {circle around (⊕)} summated with the value emanating from the {circle around (⊕)} accumulation device, to generate a partial result for the next-in-turn iteration;
the apparatus further being operable to utilize multiplicand values to be input, in turn, into the {circle around (⊕)} accumulation device for a second multiplication phase being, comprising firstly a first zero value, a second Ai operand, remaining in place from the first phase, and thirdly a Y0 value having been anticipated in the first phase;
multiplier values input into the multiplying device in the first phase being firstly an emitting string, B0, said multiplication device being operable to multiply said string segment concurrently {circle around (×)} multiplying with a second {circle around (×)} multiplier value consisting of the anticipated Y0 string which is simultaneously loaded character by character as it is generated into a preload multiplicand buffer for the second phase;
the two multiplier values operative to be input into the apparatus during a second phase being left hand n−k character values from the B operand, designated B, and the left hand n−k characters of the N modulus, designated N, respectively; and
wherein said apparatus further comprises a multiplying flush out device operative in a last phase to transfer the left hand segment of a result value remaining in the {circle around (⊕)} accumulation device into a result register.
Preferably, the apparatus is operable to perform {circle around (×)} multiplication on polynomial based operands is performed in a reverse mode, multiplying from MS characters to LS characters, operative to perform modular reduction without the Montgomery type parasitic functions, as in applicant's patent U.S. Pat. No. 5,742,530.
According to a third aspect of the present invention there is provided apparatus operative to anticipate a Y0 value using first emitting values of the multiplicand, and present inputs of the B multiplier, carry out values from the {circle around (⊕)} accumulation device, {circle around (⊕)} summation values from the {circle around (⊕)} accumulation device, the present values from the previously computed partial result, and carry out values from the {circle around (⊕)} adder which {circle around (⊕)} summates the result from the {circle around (⊕)} accumulation device with the previous partial result.
Preferably, the apparatus is adapted to insure that k first emitting values from the device are zero characters said adaptation comprising anticipation of a next in turn Y0 character using the following quantities:
i. the l bit Sout bits of the result of the l bit by l bit mod 2l {circle around (×)} multiplication of the right-hand character of the Ai register times the Bd character of the B Stream, A0·Bd mod 2l;
ii. the first emitting carry out character from the {circle around (⊕)} accumulation device, L(CO0);
iii. the l bit Sout character from the second from the right-hand character emitting cell of the {circle around (⊕)} accumulation device, SO1;
iv. the next in turn character value from the S stream, Sd;
v. the l bit carry out character from the Z output full adder, L(COz);
vi. the l bit J0 value, which is the negative multiplicative inverse of the right-hand character in the N0 modulus multiplicand register;
wherein values, A0·Bd mod 2l, L(CO0), SO1, Sd are {circle around (⊕)} added character to character together and “on the fly” {circle around (×)} multiplied by the J0 character to output a valid Y0 zero-forcing anticipatory character.
In a further embodiment there is also provided at least one sensor operative to compare the output result to N, the modulus, the mechanism operative to actuate a second subtractor on the output of the result register, thereby to output a modular reduced value which is of limited congruence to the output result value precluding the necessity to allot a second memory storage for a smaller result.
In a yet further embodiment, a value which is a {circle around (⊕)} summation of two multiplicands is loaded into a preload character buffer with at least a k characters memory means register concurrently whilst one of the first values is loaded into another preload buffer.
According to a fourth aspect of the present invention there is provided apparatus with one {circle around (⊕)} accumulation device, and an anticipating zero forcing mechanism operative to perform a series of interleaved {circle around (×)} modular multiplications and squarings and being adapted to perform concurrently the equivalent of three natural integer multiplication operations, such that a result is an exponentiation.
In an embodiment, next-in-turn used multiplicands are preloaded into a preload register buffer on the fly.
In a further embodiment, apparatus buffers and registers are operative to be loaded with values from external memory sources and said buffers and registers are operative to be unloaded into the external memory source during computations, such that the maximum size of the operands is dependent on available memory means.
In a yet further embodiment there is also provided memory register means, said memory means are typically serial single character in/serial single character out, parallel at least k characters in/parallel at least k characters out, serial single character in/parallel at least k characters out, and parallel k characters in/serial single character out.
Preferably the apparatus is operable to provide, during a final phase of a multiplication type iteration, the multiplier inputs are zero characters which are operative to flush out the left hand segment of the carry save {circle around (⊕)} accumulator memory.
Preferably, the apparatus is operable to preload next in turn multiplicands into preload memory buffers on the fly, prior to their being required in an iteration.
Preferably, the apparatus is operable to preload multiplicand values into preload buffers on the fly from central storage memory means.
The same device is operable to compute the k character Montgomery constant J0, related to the right hand k character segment of a modulus preferably by resetting both A and B to zero and setting S0=1, whilst subsequently performing a k bit multiplication. The result will reside in the Y0 register.
Modular Multiplication Sequences Using Montgomery Type Arithmetic
The k character carry save adder, the CSA, is the basis for the serial/parallel superscalar modular multiplication in both the polynomial field and in the prime number field. Polynomial GF(2q) based computations are executed preferably with all carry mechanisms switched off.
The Serial-Parallel Super Scalar Montgomery Multiplier computes Montgomery modular product in three phases, wherein in one preferred embodiment, the last phase may be a single clock dump of the whole left hand k character segment with a carry of the CSA (MS for normal multiplications, LS for reverse mode polynomial computations), and in a more compact embodiment, the last phase may be a k effective clock serial flush out of the contents of the CSA.
In previous P2 disclosures, the Y0 factor was computed bit by bit, consequently, only the right hand bit of J0 was consequential, and by definition a one, and a function of the right hand bit of the modulus. In this enhancement device, the device is character serial, and an l bit character Y0 is generated at each clock cycle. As in previous P1 disclosures, Y0 is the first phase zero forcing function which adds in the value of modulus a necessary number of times into the accumulated result, so that the relevant answer is congruent and never longer than mk+1 characters, and the first right hand emitted characters were all zeroes. X+QN≡X, as QN≡0 mod N.
Modular Multiplication Sequences
Prior to initiating computation, we assume that there are no previous temporary or random values in the device; that the operands N, B, and at least the first segment value of A are available in the registers of the device. S0=0; the partial product at initiation is typically zero. Typically, modular arithmetic is executed on operands consisting of two or more k character segments, typically in three distinct phases. For a normal full multiplication, where there are m segments of the modulus, there are typically m superscalar multiplication interleaved iterations whence each segment of the multiplicand is multiplied by the total multiplier, typically, B.
The process of the first phase, on the (i=0'th segment) of each interleaved superscalar multiplication is the generic superscalar multiplication accumulation interaction:
Si{circle around (⊕)}Ai·B0{circle around (⊕)}Y0·N0
(B0 and Y0 are serially character by character fed from the first segment of the operand into the multiplier, Ai and N0 are parallel single slice operands, Si is the partial product from a previous iteration/computation. Si=0 on the 0'th first iteration.)
The first phase process implements a {circle around (⊕)} summation of the two superscalar products with the right hand segment of the previous result. A k character string of zeroes has emitted from the multiplying device, and is disregarded; a partial first segment result resides in the device buffer, which is summated into the second phase result.
The first phase result consists typically of R, the contents of the CSA concatenated with the serial outputted right hand segment of all zeroes. (In GF(p) computations there is an additional LS carry-in bit to R.)
The process of the second phase is the generic superscalar multiplication accumulation interaction:
R{circle around (⊕)}Si{circle around (⊕)}Ai·B{circle around (⊕)}Y0·N
(Remember, an underlined variable, e.g., B, is the remaining left hand value of an operand. It is typically one or more segments, i.e., m−1 segments. B and N are serially fed character by character into the multiplier, and both Ai, remaining from the first phase, and Y0, which was a multiplier in the first phase and was loaded into the machine in the first phase to be a multiplicand in subsequent iterations, are parallel operands).
At the end of the second phase, typically consisting of m−1 iterations, the left hand segment of Si remains in the CSA—ready to be transferred—and the right hand slice (k character segment) has emanated from the device, typically into an S register. Note that multiplication in the prime number field has been performed in a conventional carry save summation method. Multiplication in the GF(2q) reversed format mode has progressed from most to least significant characters. The Y0 function has anticipated when a modulus value must be “added” into the accumulator. Except for the disabled carry bits in the device, the mechanical process is typically identical for the two number systems.
We disclose the method used for deriving the Yoj characters of the zero-forcing vector in finite fields.
Compute: J0j≡−N00−1 mod 2l.
All natural integers which are relatively prime to a modulus have multiplicative inverses in both number fields. N00 is odd and, therefore, has no factor of 2. All factors of mod 2l are 2, so that any number with a least significant 1, and a modulus whose only factor is 2, are relatively prime, and J0j always exists. Formally, for odd N00 and 2l, gcd(N00,2l)=1.
This single character of the function can be hardwire implemented with random logic, with simple circuitry, or with a simple look up table. There are only 2k−1 different values that must be derived in a look up table. In the reverse mode format, the polynomial modulus must be right justified, nominally odd. In typical exponentiational functions, both number fields, the right hand bit of the modulus bit is a one, nominally odd, and the multiplicative inverse of an odd number mod 2k must always be odd.
If l=1, the J00 multiplier is explicitly equal to 1, and need not be computed. The result of a character output forced by the Y0 function during the first phase is always 0, consequently the j'th character output, Zij, of the SuperMAP is:
0=(2lR{circle around (⊕)}Sij{circle around (⊕)}Ai0·B0j{circle around (⊕)}Y0j·N00) mod 2l=Zij; therefore,
(R{circle around (⊕)}Sij{circle around (⊕)}Ai0·B0j)≡−Y0j·N00; and
Y0j≡−N00−1(R{circle around (⊕)}Sij{circle around (⊕)}Ai0·B0j) mod r.
From the above equation, we learn that J00 is preferably the negative value of the modular multiplicative inverse of the right hand k character of the modulus for both number systems; noting that in modulo 2 arithmetic, positive and negative values are the same.
R is the summation of the value remaining in the CSA summated to the carry out bit from the final serial adder, 460, in
Formalizing the Super Scalar Modular Multiplication Method for Both Fields:
S0=0;
For i=0 to m−1 (interleave iterations)
For j=0 to k−1 (each character of first phase)
Y0j=(J00(R{circle around (⊕)}S0j{circle around (⊕)}Ai0{circle around (×)}B0j)) mod 2l
Zij=(R{circle around (⊕)}Sij{circle around (⊕)}Ai0{circle around (×)}B0j{circle around (⊕)}Y0j{circle around (×)}J0) mod 2l; and
R=[(2lR{circle around (⊕)}Sij{circle around (⊕)}Ai0{circle around (×)}B0j{circle around (⊕)}Y0j{circle around (×)}N0)]/2l
After k effective clock cycles, the first segment of the Z stream was all zeroes, and was disregarded; the relevant Y0, k character vector, is now prepared to be a multiplicand in the next phase, and the summated R value will be used in the next phase.
Second phase:
For j=k to n−1
Zij=(R{circle around (⊕)}Sij{circle around (⊕)}Ai0{circle around (×)}B0j{circle around (⊕)}Y0j{circle around (×)}N0j) mod 2l;
R=[2lR{circle around (⊕)}Sij{circle around (⊕)}Ai0·B0j{circle around (⊕)}Y0j·N0j]/2l
Implementation of the above algorithm with a character based serial-parallel multiplier is a simple extension of the above protocol:
(Quotient(x, y) is the integer division function without remainder. For example, if x=10101b, and y=10000b, then Quotient(x,y)=1).
S0=0
For i=0 to m−1 (Interleaved loop)
For j=0 to k−1
Y0j=(J00 Si0{circle around (⊕)}Ai0{circle around (×)}B0j{circle around (⊕)}L·Carry0{circle around (⊕)}Sum1{circle around (⊕)}Quotient(Si0{circle around (⊕)}Sum0, r))) mod r
For t=0 to k−1 ( Whole loop with 1 clock pulse)
Sum1=(Sumt+1{circle around (⊕)}L·Carryt{circle around (⊕)}Ait{circle around (×)}B0j{circle around (⊕)}Y0j{circle around (×)}N0y) mor r
Carryt=(Quotient((Sumt+1{circle around (⊕)}Carryt{circle around (⊕)}Ait{circle around (×)}B0j{circle around (⊕)}Y0j{circle around (×)}N0t), r)
(Output of multiplier device in this stage is ‘0’s)
Main Part
Carrya=0
For j=k to n−1
For t=0 to k−1 ( Whole loop with 1 clock pulse)
Sumt=(Sumt+1{circle around (⊕)}L·Carryt{circle around (⊕)}Ait{circle around (×)}Bj{circle around (⊕)}Y0t{circle around (×)}Nj) mod r
Carryt=Quotient((Sumt+1{circle around (⊕)}Carryt{circle around (⊕)}Ait{circle around (×)}Bj{circle around (⊕)}Y0t{circle around (×)}Nj), r)
Si-j-k=(Si,j−2k{circle around (⊕)}Sum0{circle around (⊕)}L·Carrya) mod r
Carrya=Quotient((Si,j−2k{circle around (⊕)}Sum0{circle around (⊕)}Carrya), r)
Flushing of the Multiplier
For j=n to (n+k−1)
For t=0 to k−1 ( Whole loop with 1 clock pulse)
Sum1=(Sumt+1{circle around (⊕)}L·Carryt) mod r
Carryt=Quotient((Sumt+1{circle around (⊕)}Carryt), r)
Si,j−k=(Si,j−2k{circle around (⊕)}Sum0{circle around (⊕)}L·Carrya) mod r
Carrya=Quotient((Si,j−2{circle around (⊕)}Sum0{circle around (⊕)}Carrya), r)
For a formal explanation with examples of the particular case where l=1 in the GF(p) field, see P1.
The above describes a microelectronic method and apparatus for performing interleaved finite field modular multiplication of integers A and B operative to generate an output stream of A times B modulus N wherein n is the number of characters in the modulus operand register and is larger than k, wherein the {circle around (×)} multiplication process is performed in iterations, wherein at each interleaved iteration with operands input into a {circle around (×)} multiplying device, consisting of N, the modulus, B, a multiplier, a previously computed partial result, S, and a k character string segment of A, a multiplicand, the segments progressing from the A0 string segment to the Am−1 string segment, wherein each iterative result is {circle around (⊕)} summated into a next in turn S, temporary result, in turn, wherein first emitting characters of iterative results are zeroes, the apparatus comprising:
Typically, there may be four serial l bit character registers feeding the multiplying device, first (B), second (S) and third (N) and preferably (A), configured to efficiently load the multiplier. For computations on long operands which typically are not accommodated in the MAP's internal registers, the CPU can load operands from its accessible memory.
Typically, these main memory registers store and output operands, respectively operative to store a multiplier value, a partial result value and a modulus, N.
The modular multiplying device operative to ED summate into the {circle around (⊕)} accumulation device, in turn one or two of a plurality of multiplicand values, in turn, during the phases of the iterative {circle around (×)} multiplication process, and in turn to receive as multipliers, in turn, inputs from a first value B register, second, from an “on the fly” anticipating value, Y0, as a multiplier to force first emitting right-hand zero output characters in each iteration, and third values from the modulus, N, register.
The multiplicand parallel registers are operative to receive in turn, values from the A, B, and N register sources, and in turn, also a multiplicand zero forcing Y0, value.
The zero forcing Y0 detect device is operative to generate a binary string operative to be a multiplier during the first phase of operation and is operative to be a multiplicand in the second phase of each iterative multiplication.
The multiplicand values to be switched into the {circle around (⊕)} accumulation device for the first phase consist can be on of four values, a first zero value, a second value, Ai, which is a k character string segment of a multiplicand, A, and a third value N0, being the first emitting k characters of the modulus, N. The N0 value is typically switched in at the start of a multiplication, if there is a fourth preload buffer as in
If a computation is typically on a single k character modulus, then there is no need for the S register, or for the temporary result value, S. If the operand is 2k characters or longer, then the manipulations must be iterative, with progressing Ai slices. For squaring operations slices of B are typically snared from the B stream on the fly and preloaded into the Ai preload buffer.
At the first iteration of a multiplication procedure, the temporary result is zero.
Subsequent temporary results from previous iterations, are operative to be {circle around (⊕)} summated with the value emanating from the {circle around (⊕)} accumulation device, to generate a partial result for the next in turn iteration;
The multiplicand values to be input, in turn, into the {circle around (⊕)} accumulation device for the second phase being, a first zero value, which is a pseudo register value, a second Ai operand, remaining in place from the first phase, and a third Y0 value having been anticipated in the first phase operative to continue multiplying the remaining characters of the N modulus.
The multiplier values input into the multiplying device in the first phase being a first emitting string, B0, being the first emitting string segment of the B operand, concurrently {circle around (×)} multiplying with the second {circle around (×)} multiplier value consisting of the anticipated Y0 string which is simultaneously loaded character by character as it is generated into a preload multiplicand buffer for the second phase;
The two multiplier values input into the apparatus during the second phase being the left hand n−k character values from the B operand, designated B, and the left hand n−k characters of the N modulus, designated N, respectively.
The third phase is a flush out of the device operative to transfer the left hand segment of a result value remaining in the {circle around (⊕)} accumulation device. This can either be a single clock data dump, or a simple serial unload, driven by zero characters fed in the multiplier inputs.
If the dump is a parallel dump, some means for comparing to decide if the result demands an additional reduction by the modulus.
One of the more innovative enhancements in the invention, is the reverse mode multiplication in the GF(2q). Because of the lack of interaction between adder cells in this arithmetic, it is possible to perform multiplication and reduction starting from the MS end of the product, thereby having a product that is the modular reduced answer, without a burdensome parasite, caused by disregarded zeroes which, are tantamount to performing a right shift in conventional Montgomery multiplication.
The second innovation that allows for automatic zero forcing is an extension of the Y0 function of patent application P2, which describes a device wherein only one bit was anticipated at a time. There the J00 bit, only, multiplied the single bit XORed values. Both the multiplicative inverse of an odd number and its negative value produce an odd number. This saved implementing a look up table or a random logic circuit to compute the J0 value for l=1. Note, J0 is a different quantity in non-alike number systems. We have shown in this extension how a Y0 value can be derived, for both relevant number fields.
The following describes the elements of the circuitry operative to anticipate the Y0 value using first emitting values of the multiplicand, and present inputs of the B multiplier, carry out values from the {circle around (⊕)} accumulation device, {circle around (⊕)} summation values from the {circle around (⊕)} accumulation device, the present values from the previously computed partial result, and carry out values from the {circle around (⊕)} adder which {circle around (⊕)} summates the result from the {circle around (⊕)} accumulation device with the previous partial result.
Stated differently, the six values operative to control the zero forcing function, are:
i. the l bit Sout bits of the result of the l bit by l bit mod 2l {circle around (×)} multiplication of the right-hand character of the Ai register times the Bd character of the B Stream, A0{circle around (×)}Bd mod 2l;
ii. the first emitting carry out character from the {circle around (⊕)} accumulation device, L(CO0);
iii. the l bit Sout character from the second from the right-hand character emitting cell of the {circle around (⊕)} accumulation device, SO1;
iv. the next in turn character value from the S stream, Sd;
v. the l bit carry out character from the Z output full adder, L(COz);
vi. the l bit J0 value, which is the negative multiplicative inverse of the right-hand character in the N0 modulus multiplicand register;
wherein values, A0{circle around (×)}Bd mod 2l, L(CO0), SO1, Sd are {circle around (⊕)} added character to character together and “on the fly” {circle around (×)} multiplied by the J0 character to output a valid Y0 zero-forcing anticipatory character to force an l bit egressing character string of zeroes. Just as in P1, in order to determine if an output must be modular reduced, a sensor operative to compare the output result to N, the modulus, the mechanism operative to actuate a second subtractor on the output of the result register, thereby to output a modular reduced value which is limited congruent to the output result value precluding the necessity to allot a second memory storage for a smaller result.
The single {circle around (⊕)} accumulation device, configured to perform multiplication, and an anticipating zero forcing mechanism together are operative to perform a series of interleaved {circle around (×)} modular multiplications and squarings. The total device performs the equivalent of three integer multiplications, as in a conventional Montgomery method, J0 is a k character device multiplying the first k character summation of B0{circle around (×)}Ai and Si, and in finally using the Y0 to multiply N.
Whilst the SuperMAP is computing the last iteration of a multiplication, the first slice of a next multiplication can be preloaded into a preload register buffer means on the fly. This value may be the result of a previous multiplication or a slice of a multiplicand residing in one of the register segments in the register bank of
The preloaded value which is a {circle around (⊕)} summation of two multiplicands is {circle around (⊕)} summated into a k character register, only, for GF(2q) computations. In GF(p) computations, provision must be made for an additional carry bit.
Especially for very long moduli, buffers and registers adjacent to the SuperMAP typically have insufficient memory resources. Means for loading operands directly into preload buffers is provided, operative to store operands in the CPU's memory map. For reverse format multiplication, bit order of input words from the CPU are typically reversed in the Data In and Data Out devices.
In the drawings:
Thick lines designate k character (kl bit) wide parallel bus lines. Thinner contiguous signal lines depict l bit wide lines. Most control lines are not depicted; those that are included are typically necessary to understand procedures and are typically drawn as dash-dot-dash lines
In the drawings:
Thick lines designate k character (kl bit) wide parallel bus lines. Thinner contiguous connecting signal lines depict l bit wide lines. Typically, control lines are not depicted; those that are preferably necessary to understand procedures, are typically drawn as dash-dot-dash lines
Single Multiplexers—Controlled Switching Elements which select one signal or character stream from a multiplicity of inputs of signals and direct it this chosen signal to a single output. Multiplexers are marked M1 to M13, and are intrinsic parts of larger elements.
The Multiplexer and pre-adder, 390, is an array of k l+1 multiplexers, and chooses which of the four k or k+1 character inputs are to be added into the CSA, 410.
The B (70) and 80), SA (130), SB (180), and N (200) and (210) are the four main serial main registers in a preferred embodiment. The SA is conceptually and practically redundant, but can considerably accelerate very long number computations, and save volatile memory resources, especially in the case where the length of the modulus is 2·k·m characters long.
Serial Adders and Serial Subtractors are logic elements that have two serial character inputs and one serial character output, and summate or perform subtraction on two long strings of characters. Components 90 and 500 are subtractors, 330, and 460 are serial adders. The propagation time from input to output is very small. Serial subtractors 90 and 500 typically reduce B* to B if B* is larger than or equal to N and/or S* to S if S* is larger than or equal to N. Serial Subtractor 480, is used, as part of a comparator component to detect if B* will be larger than or equal to N. Full Adder 330, adds the two character streams which feed the Load Buffer 340, with a value that is equal to the sum of the values in the 290 and 320 Load Buffers.
Fast Loaders and Unloaders, 10 and 20, and 30 and 40, respectively, are devices to accelerate the data flow from the CPU controller. Typically, these devices eliminate the necessity for other direct memory access components. 20 and 40 are for reversing the data word, as is necessary for reversing the data words for reverse format GF(2q) multiplications.
Data In, 50, is a parallel in serial out device, as the present ALU device is a serial fed systolic processor, and data is fed in, in parallel, and processed in serial.
Data Out, 60, is a serial in parallel out device, for outputting results from the coprocessor. The quotient generator is that part of
Flush Signals on Bd, 240; on S*d, 250; and on Nd, 260, are made to assure that the last k+1 characters can flush out the CSA. A second embodiment would reconcile the R data at the end of the second phase, and would perform a single parallel data dump to flush out the CSA.
Load Buffers R1, 290; R2, 320; and R3, 340 are serial in parallel out shift registers adapted to receive the three possible more than zero multiplicand combinations.
Latches L1, 360; L2, 370; and L3, 380; are made to receive the outputs from the load buffers, thereby allowing the load buffers, the temporal enablement to process the next phase of data before this data is preferably latched into L1, L2, and L3. Latch L0 is typically a “virtual” constant all zero input into 390, which typically is not implemented in latched logic.
Y0 Sense, 430, is the logic device, which determines the number of times the modulus is accumulated, in order that a k character string of LS zeros will exit at Z in {circle around (×)} multiplications.
One character delay devices 100, 220 and 230 are inserted in the respective data streams to accommodate for computation synchronization between the data preparation devices in
The k character delay shift register 470, synchronizes N and the subtractor subtracts N from the result after disregarding the right hand output zero character string for the larger than N comparison.
The Carry Save Accumulator is almost identical to a serial/parallel multiplier, excepting for the fact that three different larger than zero values can be summated, instead of the single value as conventionally is latched onto the input of the s/p multiplier. When used in polynomial based computations “all carry dependent” functions are disabled.
The Insert Last Carry, 440, is used to insert the (m·k l+1)'th bit of the S stream, as the S register is only m-k characters long.
The borrow/overflow detect, 490, typically detects if a result is larger than or equal to the modulus (from N), or in GF(p) computations. In polynomial based computations the overflow is detected if the first significant result bit is a one.
The control mechanism is not depicted, but is preferably understood to be a set of cascaded counting devices with finite state machines for specific functions with switches set for systolic data flow in both GF(p) and GF(2q).
For modular multiplication in the prime and composite prime field of numbers, we define A and B to be the multiplicand and the multiplier, and N to be the modulus which is typically larger than A or B. N also denotes the register where the value of the modulus is stored. N, may, in some instances, be smaller than A. We define A, B, and N as m·k=n character long operands. Each k character group will be called a segment, the size of the group defined by the size of the multiplying device. Then A, B, and N are each m characters long. For ease in following the step by step procedural explanations, assume that A, B, and N are 512 bits long, (n=512); assume that k is 64 characters long because of the present cost effective length of such a multiplier, and data manipulation speeds of simple CPUs; and m=8 is the number of segments in an operand and also the number of iterations in a squaring or multiplying loop with a 512 bit operand. All operands are positive integers. More generally, A, B, N, n, k and m may assume any suitable values.
In non-modular functions, the N and S registers can be used for temporary storage of other arithmetic operands.
We use the symbol, ≡, to denote congruence of modular numbers, for example 16≡2 mod 7, and we say 16 is congruent to 2 modulo 7 as 2 is the remainder when 16 is divided by 7. When we write Y mod N≡X mod N; both Y and X may be larger than N; however, for positive X and Y, the remainders will be identical. Note also that the congruence of a negative integer Y, is Y+u·N, where N is the modulus, and if the congruence of Y is to be less than N, U will be the smallest integer which will give a positive result.
We use the symbol, ¥, to denote congruence in a more limited sense. During the processes described herein, a value is often either the desired value, or equal to the desired value plus the modulus. For example X¥2 mod 7. X can be equal to 2 or 9. We say X has limited congruence to 2 mod 7. In the polynomial based field, the analog is a monic value, which we say is larger than N, and is reduced by XORing to the modulus. As in GF(2q), there is no overflow, this Yen value is typically disregarded.
When we write X=A mod N, we define X as the remainder of A divided by N; e.g., 3=45 mod 7.
In number theory the modular multiplicative inverse is a basic concept. For example, the modular multiplicative inverse of X is written as X−1, which is defined by X·X−1 mod N=1. If X=3, and N=13, then X−1 =9, i.e., the remainder of 3·9 divided by 13 is 1.
The acronyms MS and LS are used to signify most significant and least significant when referencing bits, characters, segments, and full operand values, as is conventional in digital nomenclature.
Throughout this specification N designates both the value N, and the name of the shift register which contains N. An asterisk superscript on a value, denotes that the value, as stands, is potentially incomplete or subject to change. A is the value of the number which is to be exponentiated, and n is the character length of the N operand. After initialization when A is “Montgomery normalized” to A* (A*=2n·A—to be explained later) A* and N are constant values throughout the intermediate step in the exponentiation. During the first iteration, after initialization of an exponentiation, B is equal to A*. B is also the name of the register wherein the accumulated value, which finally equals the desired result of exponentiation resides. S* designates a temporary value, and S, SA and SB designate, also, the register or registers in which all but the single MS bit of S is stored. (S* concatenated with this MS bit is identical to S.) S(i−1) denotes the value of S at the outset of the i'th iteration; So denotes the LS segment of an S(i)'th value.
We refer to the process in the GF(p) field (defined later) P(A·B)N as multiplication in the P field, or sometimes, simply, a multiplication operation.
As we have used the standard structure of a serial/parallel multiplier as the basis for constructing a double acting serial parallel multiplier, we differentiate between the summating part of the multiplier, which is based on carry save accumulation, (as opposed to a carry look ahead adder, or a ripple adder, the first of which is considerably more complicated and the second very slow), and call it a carry save adder or accumulator, and deal separately with the preloading mechanism and the multiplexer and latches, which allow us to simultaneously multiply A times B and C times D, while continuously summate both results with a previous result, S, e.g., A·B+C·D+S, converting this accumulator into a more versatile engine. Additional logic is added to this multiplier in order to provide for an anticipated sense operation necessary for modular reduction and serial summation necessary to provide for modular arithmetic and ordinary integer arithmetic on very large numbers.
Montgomery Modular Multiplication in GF(p)
The following description refers to Montgomery arithmetic in the GF(p) of numbers. The present device may be used for Montgomery arithmetic on polynomial based numbers in GF(2q), but would be degraded in performance, as computations would be in the P field, where all executable operands are multiplied by a factor of 2n.
In a classic approach for computing a modular multiplication, A·B mod N, the remainder of the product A·B is computed by a division process. Implementing a conventional division of large operands is more difficult to perform than serial/parallel multiplications.
Using Montgomery's modular reduction method, division is essentially replaced by multiplications using two precomputed constants. In the procedure demonstrated herein, there is only one precomputed constant, which is a function of the modulus. This constant is, or can be, computed using this ALU device.
A simplified presentation of the Montgomery process, as is used in this device is now provided, followed by a complete preferred description.
If we have an odd number (an LS bit one), e.g., 1010001 (=8110) we can always transform this odd number to an even number (a single LS bit of zero) by adding to it another fixing, compensating odd number, e.g., 1111 (=1510); as 1111+1010001=1100000 (9610). In this particular case, we have found a number that produced five LS zeros, because we knew in advance the whole string, 81, and could easily determine a binary number which we could add to 81, and would produce a new binary number that would have as many LS zeros as we might need. This fixing number must have a right hand one, else it has no effect on the progressive LS characters of a result.
If our process is a clocked serial/parallel carry save process, where it is desired to have a continuous number of LS zeros, and wherein at each clock cycle we only have to fix the next bit, at each clock it is sufficient to add the fix, if the next bit were to be a one or not to add the fix if the anticipated bit were to be a zero. However, in order not to cause interbit overflows (double carries), this fix is preferably summated previously with the multiplicand, to be added into the accumulator when the relevant multiplier bit is one, and the Y Sense also anticipates a one.
Now, as in modular arithmetic, we only are interested in the remainder of a value divided by the modulus, we know that we can add the modulus any number of times to a value, and still have a value that would have the same remainder. This means that we can add Y·N=Σ yi·ri·N to any integer, and still have the same remainder; Y being the number of times we add in the modulus, N, to produce the required kl right hand zeros. As described, the modulus that we add can only be odd. (Methods exist wherein even moduli are defined as ri times the odd number that results when i is the number of LS zeros in the even number.)
Montgomery interleaved reductions typically reduce storage requirements, and the cost effective size of the multiplication devices. This is especially useful when performing public key cryptographic functions where we multiply one large integer, e.g., n=1024 bit, by another same length large integer; a process that would ordinarily produce a double length integer.
We can add in Ns (the modulus) enough times to A·B=X or A·B+S=X during the process of multiplication (or squaring) so that we will have a number, Z, that has n LS zeros, and, at most, n+1 MS characters.
We can continue using such numbers, disregarding the LS n characters, if we remember that by disregarding these zeros, we have divided the desired result by ra.
When the LS n characters are disregarded, and we only use the most significant n (or n+1) characters, then we have effectively multiplied the result by r−n, the modular inverse of rn. If we would subsequently re-multiply this result by rn mod N (or rn) we would obtain a value congruent to the desired result (having the same remainder) as A·B+S mod N. As is seen, using MM, the result is preferably multiplied by r2n to overcome the r−n parasitic factor reintroduced by the MM.
Example:
A·B+S mod N=(12·11+10) mod 13=(1100·1011+1010)2 mod 10112. l=1, r=2
We will add in 2i N whenever a fix is necessary on one of the n LS bits.
And the result is 10001 00002 mod 13=17·24 mod 13.
As 17 is larger than 13 we subtract 13, and the result is:
17·24=4·24 mod 13.
formally 2−n(A·B+S) mod N=9 (12·11+10) mod 13≡4
In Montgomery arithmetic we utilize only the MS non-zero result, 4, and effectively remember that the real result has been divided by 2n; n zeros having been forced onto the MM result.
We have added in (8+2)·13=10·13 which effectively multiplied the result by 24 mod 13≡3. In effect, had we used the superfluous zeros, we can say that we have performed, A·B+Y·N+S−(12·11+10·13+10) in one process, which will be described possible on a preferred embodiment.
Check−(12·11+10) mod 13=12; 4·3=12.
In summary, the result of a Montgomery Multiplication is the desired result multiplied by 2−n.
To retrieve the previous result back into a desired result using the same multiplication method, we would have to Montgomery Multiply the previous result by 22n, which we will call H, as each MM leaves us with a parasitic factor of 2−n.
The Montgomery Multiply function P(A·B)V performs a multiplication modulo N of the A·B product into the P field. (In the above example, where we derived 4). The retrieval from the P field back into the normal modular field is performed by enacting P on the result of P(A·B)N using the precomputed constant H. Now, if P≡(A·B)N, it follows that P(P·H)N≡A·B mod N; thereby performing a normal modular multiplication in two P field multiplications.
Montgomery modular reduction averts a series of multiplication and division operations on operands that are n and 2n characters long, by performing a series of multiplications, additions, and subtractions on operands that are n or n+1 characters long. The entire process yields a result which is smaller than or equal to N. For given A, B and odd N there is always a Q, such that A·B+Q·N will result in a number whose n LS characters are zero, or:
P·2n=A·B+Q·N
This means that we have an expression that is 2n characters long (with a possible one bit overflow), whose n LS characters are zero.
Now, for radix r=2l; let I·rn≡1 mod N (I exists for all odd N). Multiplying both sides of the previous equation by I yields the following congruences: from the left side of the equation:
P·I·rn≡P mod N; (Remember that I·rn≡1 mod N)
and from the right side:
A·B·I+Q·N·I≡A·B·I mod N; (Remember that Q·N·I≡0 mod N)
therefore:
P≡A·B·I mod N.
This also means that a parasitic factor I=r−n mod N is introduced each time a P field multiplication is performed.
We define the P operator such that:
P≡A·B·I mod N≡P(A·B)N.
and we call this “multiplication of A times B in the P field”, or Montgomery Multiplication.
The retrieval from the P field can be computed by operating P on P·H, making:
P(P·H)N≡A·B mod N;
We can derive the value of H by substituting P in the previous congruence.
We find:
P(P·H)N≡(A·B·I)(H)(I) mod N;
(see that A·B·I←P; H←H; I←and any multiplication operation introduces a parasitic I)
If H is congruent to the multiple inverse of I2 then the congruence is valid, therefore:
H=I−2 mod N≡r2n mod N
(H is a function of N and we call it the H parameter)
In conventional Montgomery methods, to enact the P operator on A·B, the following process may be employed, using the precomputed constant J:
1) X=A·B
2) Y=(X·J) mod rn (only the n LS characters are necessary)
3) Z=X+Y·N
4) S=Z/rn (The requirement on J is that it forces Z to be divisible by rn)
5) P¥S mod N (N is to be subtracted from S, if S≧N)
Finally, at step 5):
P¥P(A·B)N,
[After the subtraction of N, if necessary:
P=P(A·B)N]
Following the above:
Y=A·B·J mod rn (using only the n LS characters);
and:
Z=A·B+(A·B·J mod rn)·N.
In order that Z be divisible by rn (the n LS characters of Z are preferably zero) and the following congruence will exist:
[A·B+(A·B·J mod rn)·N] mod rn=0
In order that this congruence will exist, N·J mod rn is congruent to −1 or:
J≡N−1 mod rn.
and we have found the constant J.
J, therefore, is a precomputed constant which is a function of N only. However, in a machine that outputs a MM result, character by character, provision should be made to add in Ns at each instance where the output character in the LS string would otherwise have been a zero, thereby obviating the necessity of precomputing J and subsequently computing Y=A·B·J mod rn, as Y can be detected character by character using hardwired logic. We have also described that this method can only work for odd Ns.
Therefore, as is apparent, the process described employs three multiplications, one summation, and a maximum of one subtraction, for the given A, B, N, and a precomputed constant to obtain P(A·B)N. Using this result, the same process and a precomputed constant, H, (a function of the module N) we are able to find A·B mod N. As A can also be equal to B, this basic operator can be used as a device to square or multiply in the modular arithmetic.
Interleaved Montgomery Modular Multiplication
The previous section describes a method for modular multiplication which involved multiplications of operands which were all n characters long, and results which required 2n+1 characters of storage space.
Using Montgomery's interleaved reduction as described in P1, it is possible to perform the multiplication operations with shorter operands, registers, and hardware multipliers; enabling the implementation of an electronic device with relatively few logic gates.
First we will describe how the device can work, if at each iteration of the interleave, we compute the number of times that N is added, using the J0 constant. Later, we describe how to interleave, using a hardwire derivation of Y0, which will eliminate the J0+ phase of each multiplication {(2) in the following example}, and enable us to integrate the functions of two separate serial/multipliers into the new single generic multiplier which can perform A·B+C·N+S at better than double speed using similar silicon resources.
Using a k character multiplier, it is convenient to define segments of k character length; there are m segments in n characters; i.e., m·k=n. J0 will be the LS segment of J.
Therefore:
J0≡−N0−1 mod rk (J0 exists as N is odd).
Note, the J and J0 constants are compensating numbers that when enacted on the unreduced output, tell us how many times to add the modulus, in order to have a predefined number of least significant zeros. We will later describe an additional advantage to the present serial device; since, as the next serial bit of output can be easily determined, we can always add the modulus (always odd) to the next intermediate result. This is the case if, without this addition, the output character, the LS serial bit exiting the CSA, would have been a “1”; thereby adding in the modulus to the previous even intermediate result, and thereby promising another LS zero in the output string. Remember, congruency is maintained, as no matter how many times the modulus is added to the result, the remainder is constant.
In the conventional use of Montgomery's interleaved reduction, P(A·B)N is enacted in m iterations as described in steps (1) to (5):
Initially S(0)=0 (the ¥ value of S at the outset of the first iteration).
For i=1,2 m:
1) X=S(i−1)+Ai−1·B (Ai−1 is the i−1'th character of A; S(i−1) is the value
of S at the outset of the i'th iteration.)
2) Y0 =X0·J0 mod rk (The LS k characters of the product of X0·J0)
(The process uses and computes the k LS characters only, e.g., the least significant 64 characters) In the preferred implementation, this step is obviated, because in a serial machine Y0 can be anticipated character by character.
3) Z=X+Y0·N
4) S(i)=Z/rk (The k LS characters of Z are always 0, therefore Z is always divisible by rk. This division is tantamount to a k character right shift as the LS k characters of Z are all zeros; or as will be seen in the circuit, the LS k characters of Z are simply disregarded.
5) S(i)=S(i) mod N (N is to be subtracted from those S(i)'s which are larger than N ). Finally, at the last iteration (after the subtraction of N, when necessary), C=S(m)=P(A·B)N.
To derive F=A·B mod N, the P field computation, P(C·H)N, is performed.
It is desired to know, in a preferred embodiment, that for all S(i)'s, S(i) is smaller than 2N. This also means, that the last result (S(m)) can always be reduced to a quantity less than N with, at most, one subtraction of N.
We observe that for operands which are used in the process:
S(i−1)<rn+1 (the temporary register can be one bit longer than the B or N register),
B<N<rn and Ai−1<rk.
By definition:
S(i)=Z/rk (The value of S at the end of the process, before a possible subtraction)
For all Z, Z(i)<rn+k+1
Xmax=Smax+Ai·B<rn+1−1+(rk−1)(rn−1)
Qmax=Y0N<(rk−1)(rn−1)
therefore:
Zmax<rk+n+1−rk+1+1<rk+n+1−1.
and as Zmax is divided by rk:
S(m)<rn+1−r1.
Because Nmin>rn−r, S(m)max is always less than 2·Nmin, and therefore, one subtraction is all that is necessary on a final result.
S(m)max−Nmin=(rn+1−r1−1)−(rn−1)=rn−4<Nmin.
Example of a Montgomery Interleaved Modular Multiplication:
The following computations in the hexadecimal format clarify the meaning of the interleaved method:
The expected result is F≡A·B mod N≡99b·5c3 mod a59≡375811 mod a59=22016.
Initially: S(0)=0
X=S(0)+A0·B=0+b·5c3=3f61 Step 1
Y0=X0·J0 mod rk=7 (Y0—hardwire anticipated in SuperMAP)
Z=X+Y0·N=3f61+7·a59=87d0
S(1)=Z/rk=87d
X=S(1)+A1·B=87d+95c3=3c58 Step 2
Y0=X0·J0 mod rk=8.7 mod 24=8 (Hardwire anticipated)
Z=X+Y0·N=3c58+52c8=8f20
S(2)=Z/rk=8f2
X=S(2)+A2·B=8f2+9·5c3=3ccd Step 3
Y0=d·7 mod 24=b (Hardwire anticipated)
Z=X+Y0·N=3ccd+b·a59=aea0
S(3)=Z/rk=aea,
as S(3)>N,
S(m)=S(3)−N=aea−a59=91
Therefore C=P(A·B)N=9116.
Retrieval from the P field is performed by computing P(C·H)N:
Again initially: S(0)=0
X=S(0)+C0·H=0+1·44b=44b Step 1
Y0=d (Hardwire anticipated in SuperMAP)
Z=X+Y0·N=44b+8685=8ad0
S(1)=Z/rk=8ad
X=S(1)+C1·H=8ad+9.44b=2f50 Step 2
Y0=0 (Hardwire anticipated in SuperMAP)
Z=X+Y0·N=2f50+0=2f50
S(2)=Z/rk=2f5
X=S(2)+C2·H=2f5+0·44b=2f5 Step 3
Y0=3 (Hardwire anticipated in SuperMAP)
Z=X+Y0·N=2f5+3·a59=2200
S(3)=Z/rk=22016
which is the expected value of 99b·5c3 mod a59.
If at each step we disregard k LS zeros, we are in essence multiplying the n MS characters by rk. Likewise, at each step, the i'th segment of the multiplier is also a number multiplied by rik, giving it the same rank as S(i).
It can also be noted that in another preferred embodiment, wherein it is of some potential value to know the J0 constant, if Ai·B+S=1; then Y0=−N0−1=J0
Exponentiation:
The following derivation of a sequence [D. Knuth, The art of computer programming, vol. 2: Seminumerical algorithms, Addison-Wesley, Reading Mass.,
hereinafter referred to as “Knuth”, explains a sequence of squares and multiplies, which implements a modular exponentiation.
After precomputing the Montgomery constant, H=22n, as this device can both square and multiply in the P field, we compute:
C=AE mod N.
Let E(j) denote the j bit in the binary representation of the exponent E, starting with the MS bit whose index is 1 and concluding with the LS bit whose index is q, we can exponentiate as follows for odd exponents:
B ¥ P(B·A)N E(0)=1; B is the last desired temporary result multiplied by 2″,
After the last iteration, the value B is ¥ to AE mod N, and C is the final value.
To clarify, we shall use the following example:
E=1011→E(1)=1; E(2)=0; E(3)=1; E(4)=1;
To find A1011 mod N; q=4
A*=P(A·H)N=A·I−2I=A·I−1 mod N
B=A*
FOR j=2 to q
B=P(B·B)N which produces: A2(I−1)2·I=A2·I−1
E(2)=0; B=A2·I−1
j=3 B=P(B·B)N=A2(I−1)2·I=A4·I−1
E(3)=1 B=P(B·A*)N=(A4·I−1)(A·I−1)·I=A5·I−1
j=4 B=P(B·B)N=A10·I−2·I=A10·I−1
As E(4) was odd, the last multiplication will be by A, to remove the parasitic I−1.
B=P(B·A)N=A10·I−1·A·I=A11
C=B
A method for computing the H parameter by a reciprocal process is described in U.S. Pat. No. 5,513,133.
Reference is now made to
When a cryptographic program, such as verifying an RSA signature is executed, it may require modular arithmetic functions such as modular exponentiation. The cryptographic program that calls the cryptographic function is preferably run on the CPU, 502.
Reference is now made to
The coprocessor 6075 is connected to a data bus with a CPU as in previous figures. A register bank, 6205, comprises a B register 6070, an A register 6130, an S register 6180, and an N register 6200. The outputs of each of the registers are connected to a serial data switch and serial process conditioner 6020, which in turn is connected to an operational unit, 6206, which carries out the modular arithmetic operations. Connected between the N register, 6200, and the operational unit, 6206, is a J0 generator, 552.
In the embodiment the J0 generator compiles an l bit primary zero forcing function for use in the modular arithmetic functions described above.
Reference is now made to
Reference is now made to
Reference is now made to
Reference is now made to
Reference is now made to
In
It is appreciated that various features of the invention, which are, for clarity, described in the contexts of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the invention, which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable subcombination.
It will be appreciated by persons skilled in the art, that the present invention is not limited to what has been particularly shown and described hereinabove. Rather, the scope of the present invention includes both combinations and subcombinations of the various features described hereinabove as well as variations and modifications thereof, which would occur to persons skilled in the art upon reading the foregoing description which are not in the prior art.
In the following claims, symbols such as have the meanings given in the preceding description.
| Number | Date | Country | Kind |
|---|---|---|---|
| 136151 | May 2000 | IL | national |
| 139674 | Nov 2000 | IL | national |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 09854853 | May 2001 | US |
| Child | 11497496 | Aug 2006 | US |
