mirror of
https://github.com/rn10950/RetroZilla.git
synced 2024-11-16 20:40:11 +01:00
318 lines
14 KiB
GLSL
318 lines
14 KiB
GLSL
***** BEGIN LICENSE BLOCK *****
|
|
Version: MPL 1.1/GPL 2.0/LGPL 2.1
|
|
|
|
The contents of this file are subject to the Mozilla Public License Version
|
|
1.1 (the "License"); you may not use this file except in compliance with
|
|
the License. You may obtain a copy of the License at
|
|
http://www.mozilla.org/MPL/
|
|
|
|
Software distributed under the License is distributed on an "AS IS" basis,
|
|
WITHOUT WARRANTY OF ANY KIND, either express or implied. See the License
|
|
for the specific language governing rights and limitations under the
|
|
License.
|
|
|
|
The Original Code is the elliptic curve math library.
|
|
|
|
The Initial Developer of the Original Code is Sun Microsystems, Inc.
|
|
Portions created by Sun Microsystems, Inc. are Copyright (C) 2003
|
|
Sun Microsystems, Inc. All Rights Reserved.
|
|
|
|
Contributor(s):
|
|
Stephen Fung <fungstep@hotmail.com> and
|
|
Nils Gura <nils.gura@sun.com>, Sun Microsystems Laboratories
|
|
|
|
Alternatively, the contents of this file may be used under the terms of
|
|
either the GNU General Public License Version 2 or later (the "GPL"), or
|
|
the GNU Lesser General Public License Version 2.1 or later (the "LGPL"),
|
|
in which case the provisions of the GPL or the LGPL are applicable instead
|
|
of those above. If you wish to allow use of your version of this file only
|
|
under the terms of either the GPL or the LGPL, and not to allow others to
|
|
use your version of this file under the terms of the MPL, indicate your
|
|
decision by deleting the provisions above and replace them with the notice
|
|
and other provisions required by the GPL or the LGPL. If you do not delete
|
|
the provisions above, a recipient may use your version of this file under
|
|
the terms of any one of the MPL, the GPL or the LGPL.
|
|
|
|
***** END LICENSE BLOCK *****
|
|
|
|
The ECL exposes routines for constructing and converting curve
|
|
parameters for internal use.
|
|
|
|
The floating point code of the ECL provides algorithms for performing
|
|
elliptic-curve point multiplications in floating point.
|
|
|
|
The point multiplication algorithms perform calculations almost
|
|
exclusively in floating point for efficiency, but have the same
|
|
(integer) interface as the ECL for compatibility and to be easily
|
|
wired-in to the ECL. Please see README file (not this README.FP file)
|
|
for information on wiring-in.
|
|
|
|
This has been implemented for 3 curves as specified in [1]:
|
|
secp160r1
|
|
secp192r1
|
|
secp224r1
|
|
|
|
RATIONALE
|
|
=========
|
|
Calculations are done in the floating-point unit (FPU) since it
|
|
gives better performance on the UltraSPARC III chips. This is
|
|
because the FPU allows for faster multiplication than the integer unit.
|
|
The integer unit has a longer multiplication instruction latency, and
|
|
does not allow full pipelining, as described in [2].
|
|
Since performance is an important selling feature of Elliptic Curve
|
|
Cryptography (ECC), this implementation was created.
|
|
|
|
DATA REPRESENTATION
|
|
===================
|
|
Data is primarily represented in an array of double-precision floating
|
|
point numbers. Generally, each array element has 24 bits of precision
|
|
(i.e. be x * 2^y, where x is an integer of at most 24 bits, y some positive
|
|
integer), although the actual implementation details are more complicated.
|
|
|
|
e.g. a way to store an 80 bit number might be:
|
|
double p[4] = { 632613 * 2^0, 329841 * 2^24, 9961 * 2^48, 51 * 2^64 };
|
|
See section ARITHMETIC OPERATIONS for more details.
|
|
|
|
This implementation assumes that the floating-point unit rounding mode
|
|
is round-to-even as specified in IEEE 754
|
|
(as opposed to chopping, rounding up, or rounding down).
|
|
When subtracting integers represented as arrays of floating point
|
|
numbers, some coefficients (array elements) may become negative.
|
|
This effectively gives an extra bit of precision that is important
|
|
for correctness in some cases.
|
|
|
|
The described number presentation limits the size of integers to 1023 bits.
|
|
This is due to an upper bound of 1024 for the exponent of a double precision
|
|
floating point number as specified in IEEE-754.
|
|
However, this is acceptable for ECC key sizes of the foreseeable future.
|
|
|
|
DATA STRUCTURES
|
|
===============
|
|
For more information on coordinate representations, see [3].
|
|
|
|
ecfp_aff_pt
|
|
-----------
|
|
Affine EC Point Representation. This is the basic
|
|
representation (x, y) of an elliptic curve point.
|
|
|
|
ecfp_jac_pt
|
|
-----------
|
|
Jacobian EC Point. This stores a point as (X, Y, Z), where
|
|
the affine point corresponds to (X/Z^2, Y/Z^3). This allows
|
|
for fewer inversions in calculations.
|
|
|
|
ecfp_chud_pt
|
|
------------
|
|
Chudnovsky Jacobian Point. This representation stores a point
|
|
as (X, Y, Z, Z^2, Z^3), the same as a Jacobian representation
|
|
but also storing Z^2 and Z^3 for faster point additions.
|
|
|
|
ecfp_jm_pt
|
|
----------
|
|
Modified Jacobian Point. This representation stores a point
|
|
as (X, Y, Z, a*Z^4), the same as Jacobian representation but
|
|
also storing a*Z^4 for faster point doublings. Here "a" represents
|
|
the linear coefficient of x defining the curve.
|
|
|
|
EC_group_fp
|
|
-----------
|
|
Stores information on the elliptic curve group for floating
|
|
point calculations. Contains curve specific information, as
|
|
well as function pointers to routines, allowing different
|
|
optimizations to be easily wired in.
|
|
This should be made accessible from an ECGroup for the floating
|
|
point implementations of point multiplication.
|
|
|
|
POINT MULTIPLICATION ALGORITHMS
|
|
===============================
|
|
Elliptic Curve Point multiplication can be done at a higher level orthogonal
|
|
to the implementation of point additions and point doublings. There
|
|
are a variety of algorithms that can be used.
|
|
|
|
The following algorithms have been implemented:
|
|
|
|
4-bit Window (Jacobian Coordinates)
|
|
Double & Add (Jacobian & Affine Coordinates)
|
|
5-bit Non-Adjacent Form (Modified Jacobian & Chudnovsky Jacobian)
|
|
|
|
Currently, the fastest algorithm for multiplying a generic point
|
|
is the 5-bit Non-Adjacent Form.
|
|
|
|
See comments in ecp_fp.c for more details and references.
|
|
|
|
SOURCE / HEADER FILES
|
|
=====================
|
|
|
|
ecp_fp.c
|
|
--------
|
|
Main source file for floating point calculations. Contains routines
|
|
to convert from floating-point to integer (mp_int format), point
|
|
multiplication algorithms, and several other routines.
|
|
|
|
ecp_fp.h
|
|
--------
|
|
Main header file. Contains most constants used and function prototypes.
|
|
|
|
ecp_fp[160, 192, 224].c
|
|
-----------------------
|
|
Source files for specific curves. Contains curve specific code such
|
|
as specialized reduction based on the field defining prime. Contains
|
|
code wiring-in different algorithms and optimizations.
|
|
|
|
ecp_fpinc.c
|
|
-----------
|
|
Source file that is included by ecp_fp[160, 192, 224].c. This generates
|
|
functions with different preprocessor-defined names and loop iterations,
|
|
allowing for static linking and strong compiler optimizations without
|
|
code duplication.
|
|
|
|
TESTING
|
|
=======
|
|
The test suite can be found in ecl/tests/ecp_fpt. This tests and gets
|
|
timings of the different algorithms for the curves implemented.
|
|
|
|
ARITHMETIC OPERATIONS
|
|
---------------------
|
|
The primary operations in ECC over the prime fields are modular arithmetic:
|
|
i.e. n * m (mod p) and n + m (mod p). In this implementation, multiplication,
|
|
addition, and reduction are implemented as separate functions. This
|
|
enables computation of formulae with fewer reductions, e.g.
|
|
(a * b) + (c * d) (mod p) rather than:
|
|
((a * b) (mod p)) + ((c * d) (mod p)) (mod p)
|
|
This takes advantage of the fact that the double precision mantissa in
|
|
floating point can hold numbers up to 2^53, i.e. it has some leeway to
|
|
store larger intermediate numbers. See further detail in the section on
|
|
FLOATING POINT PRECISION.
|
|
|
|
Multiplication
|
|
--------------
|
|
Multiplication is implemented in a standard polynomial multiplication
|
|
fashion. The terms in opposite factors are pairwise multiplied and
|
|
added together appropriately. Note that the result requires twice
|
|
as many doubles for storage, as the bit size of the product is twice
|
|
that of the multiplicands.
|
|
e.g. suppose we have double n[3], m[3], r[6], and want to calculate r = n * m
|
|
r[0] = n[0] * m[0]
|
|
r[1] = n[0] * m[1] + n[1] * m[0]
|
|
r[2] = n[0] * m[2] + n[1] * m[1] + n[2] * m[0]
|
|
r[3] = n[1] * m[2] + n[2] * m[1]
|
|
r[4] = n[2] * m[2]
|
|
r[5] = 0 (This is used later to hold spillover from r[4], see tidying in
|
|
the reduction section.)
|
|
|
|
Addition
|
|
--------
|
|
Addition is done term by term. The only caveat is to be careful with
|
|
the number of terms that need to be added. When adding results of
|
|
multiplication (before reduction), twice as many terms need to be added
|
|
together. This is done in the addLong function.
|
|
e.g. for double n[4], m[4], r[4]: r = n + m
|
|
r[0] = n[0] + m[0]
|
|
r[1] = n[1] + m[1]
|
|
r[2] = n[2] + m[2]
|
|
r[3] = n[3] + m[3]
|
|
|
|
Modular Reduction
|
|
-----------------
|
|
For the curves implemented, reduction is possible by fast reduction
|
|
for Generalized Mersenne Primes, as described in [4]. For the
|
|
floating point implementation, a significant step of the reduction
|
|
process is tidying: that is, the propagation of carry bits from
|
|
low-order to high-order coefficients to reduce the precision of each
|
|
coefficient to 24 bits.
|
|
This is done by adding and then subtracting
|
|
ecfp_alpha, a large floating point number that induces precision roundoff.
|
|
See [5] for more details on tidying using floating point arithmetic.
|
|
e.g. suppose we have r = 961838 * 2^24 + 519308
|
|
then if we set alpha = 3 * 2^51 * 2^24,
|
|
FP(FP(r + alpha) - alpha) = 961838 * 2^24, because the precision for
|
|
the intermediate results is limited. Our values of alpha are chosen
|
|
to truncate to a desired number of bits.
|
|
|
|
The reduction is then performed as in [4], adding multiples of prime p.
|
|
e.g. suppose we are working over a polynomial of 10^2. Take the number
|
|
2 * 10^8 + 11 * 10^6 + 53 * 10^4 + 23 * 10^2 + 95, stored in 5 elements
|
|
for coefficients of 10^0, 10^2, ..., 10^8.
|
|
We wish to reduce modulo p = 10^6 - 2 * 10^4 + 1
|
|
We can subtract off from the higher terms
|
|
(2 * 10^8 + 11 * 10^6 + 53 * 10^4 + 23 * 10^2 + 95) - (2 * 10^2) * (10^6 - 2 * 10^4 + 1)
|
|
= 15 * 10^6 + 53 * 10^4 + 21 * 10^2 + 95
|
|
= 15 * 10^6 + 53 * 10^4 + 21 * 10^2 + 95 - (15) * (10^6 - 2 * 10^4 + 1)
|
|
= 83 * 10^4 + 21 * 10^2 + 80
|
|
|
|
Integrated Example
|
|
------------------
|
|
This example shows how multiplication, addition, tidying, and reduction
|
|
work together in our modular arithmetic. This is simplified from the
|
|
actual implementation, but should convey the main concepts.
|
|
Working over polynomials of 10^2 and with p as in the prior example,
|
|
Let a = 16 * 10^4 + 53 * 10^2 + 33
|
|
let b = 81 * 10^4 + 31 * 10^2 + 49
|
|
let c = 22 * 10^4 + 0 * 10^2 + 95
|
|
And suppose we want to compute a * b + c mod p.
|
|
We first do a multiplication: then a * b =
|
|
0 * 10^10 + 1296 * 10^8 + 4789 * 10^6 + 5100 * 10^4 + 3620 * 10^2 + 1617
|
|
Then we add in c before doing reduction, allowing us to get a * b + c =
|
|
0 * 10^10 + 1296 * 10^8 + 4789 * 10^6 + 5122 * 10^4 + 3620 * 10^2 + 1712
|
|
We then perform a tidying on the upper half of the terms:
|
|
0 * 10^10 + 1296 * 10^8 + 4789 * 10^6
|
|
0 * 10^10 + (1296 + 47) * 10^8 + 89 * 10^6
|
|
0 * 10^10 + 1343 * 10^8 + 89 * 10^6
|
|
13 * 10^10 + 43 * 10^8 + 89 * 10^6
|
|
which then gives us
|
|
13 * 10^10 + 43 * 10^8 + 89 * 10^6 + 5122 * 10^4 + 3620 * 10^2 + 1712
|
|
we then reduce modulo p similar to the reduction example above:
|
|
13 * 10^10 + 43 * 10^8 + 89 * 10^6 + 5122 * 10^4 + 3620 * 10^2 + 1712
|
|
- (13 * 10^4 * p)
|
|
69 * 10^8 + 89 * 10^6 + 5109 * 10^4 + 3620 * 10^2 + 1712
|
|
- (69 * 10^2 * p)
|
|
227 * 10^6 + 5109 * 10^4 + 3551 * 10^2 + 1712
|
|
- (227 * p)
|
|
5563 * 10^4 + 3551 * 10^2 + 1485
|
|
finally, we do tidying to get the precision of each term down to 2 digits
|
|
5563 * 10^4 + 3565 * 10^2 + 85
|
|
5598 * 10^4 + 65 * 10^2 + 85
|
|
55 * 10^6 + 98 * 10^4 + 65 * 10^2 + 85
|
|
and perform another reduction step
|
|
- (55 * p)
|
|
208 * 10^4 + 65 * 10^2 + 30
|
|
There may be a small number of further reductions that could be done at
|
|
this point, but this is typically done only at the end when converting
|
|
from floating point to an integer unit representation.
|
|
|
|
FLOATING POINT PRECISION
|
|
========================
|
|
This section discusses the precision of floating point numbers, which
|
|
one writing new formulae or a larger bit size should be aware of. The
|
|
danger is that an intermediate result may be required to store a
|
|
mantissa larger than 53 bits, which would cause error by rounding off.
|
|
|
|
Note that the tidying with IEEE rounding mode set to round-to-even
|
|
allows negative numbers, which actually reduces the size of the double
|
|
mantissa to 23 bits - since it rounds the mantissa to the nearest number
|
|
modulo 2^24, i.e. roughly between -2^23 and 2^23.
|
|
A multiplication increases the bit size to 2^46 * n, where n is the number
|
|
of doubles to store a number. For the 224 bit curve, n = 10. This gives
|
|
doubles of size 5 * 2^47. Adding two of these doubles gives a result
|
|
of size 5 * 2^48, which is less than 2^53, so this is safe.
|
|
Similar analysis can be done for other formulae to ensure numbers remain
|
|
below 2^53.
|
|
|
|
Extended-Precision Floating Point
|
|
---------------------------------
|
|
Some platforms, notably x86 Linux, may use an extended-precision floating
|
|
point representation that has a 64-bit mantissa. [6] Although this
|
|
implementation is optimized for the IEEE standard 53-bit mantissa,
|
|
it should work with the 64-bit mantissa. A check is done at run-time
|
|
in the function ec_set_fp_precision that detects if the precision is
|
|
greater than 53 bits, and runs code for the 64-bit mantissa accordingly.
|
|
|
|
REFERENCES
|
|
==========
|
|
[1] Certicom Corp., "SEC 2: Recommended Elliptic Curve Domain Parameters", Sept. 20, 2000. www.secg.org
|
|
[2] Sun Microsystems Inc. UltraSPARC III Cu User's Manual, Version 1.0, May 2002, Table 4.4
|
|
[3] H. Cohen, A. Miyaji, and T. Ono, "Efficient Elliptic Curve Exponentiation Using Mixed Coordinates".
|
|
[4] Henk C.A. van Tilborg, Generalized Mersenne Prime. http://www.win.tue.nl/~henkvt/GenMersenne.pdf
|
|
[5] Daniel J. Bernstein, Floating-Point Arithmetic and Message Authentication, Journal of Cryptology, March 2000, Section 2.
|
|
[6] Daniel J. Bernstein, Floating-Point Arithmetic and Message Authentication, Journal of Cryptology, March 2000, Section 2 Notes.
|