1 | // Package bigfft implements multiplication of big.Int using FFT.
|
---|
2 | //
|
---|
3 | // The implementation is based on the Schönhage-Strassen method
|
---|
4 | // using integer FFT modulo 2^n+1.
|
---|
5 | package bigfft
|
---|
6 |
|
---|
7 | import (
|
---|
8 | "math/big"
|
---|
9 | "unsafe"
|
---|
10 | )
|
---|
11 |
|
---|
12 | const _W = int(unsafe.Sizeof(big.Word(0)) * 8)
|
---|
13 |
|
---|
14 | type nat []big.Word
|
---|
15 |
|
---|
16 | func (n nat) String() string {
|
---|
17 | v := new(big.Int)
|
---|
18 | v.SetBits(n)
|
---|
19 | return v.String()
|
---|
20 | }
|
---|
21 |
|
---|
22 | // fftThreshold is the size (in words) above which FFT is used over
|
---|
23 | // Karatsuba from math/big.
|
---|
24 | //
|
---|
25 | // TestCalibrate seems to indicate a threshold of 60kbits on 32-bit
|
---|
26 | // arches and 110kbits on 64-bit arches.
|
---|
27 | var fftThreshold = 1800
|
---|
28 |
|
---|
29 | // Mul computes the product x*y and returns z.
|
---|
30 | // It can be used instead of the Mul method of
|
---|
31 | // *big.Int from math/big package.
|
---|
32 | func Mul(x, y *big.Int) *big.Int {
|
---|
33 | xwords := len(x.Bits())
|
---|
34 | ywords := len(y.Bits())
|
---|
35 | if xwords > fftThreshold && ywords > fftThreshold {
|
---|
36 | return mulFFT(x, y)
|
---|
37 | }
|
---|
38 | return new(big.Int).Mul(x, y)
|
---|
39 | }
|
---|
40 |
|
---|
41 | func mulFFT(x, y *big.Int) *big.Int {
|
---|
42 | var xb, yb nat = x.Bits(), y.Bits()
|
---|
43 | zb := fftmul(xb, yb)
|
---|
44 | z := new(big.Int)
|
---|
45 | z.SetBits(zb)
|
---|
46 | if x.Sign()*y.Sign() < 0 {
|
---|
47 | z.Neg(z)
|
---|
48 | }
|
---|
49 | return z
|
---|
50 | }
|
---|
51 |
|
---|
52 | // A FFT size of K=1<<k is adequate when K is about 2*sqrt(N) where
|
---|
53 | // N = x.Bitlen() + y.Bitlen().
|
---|
54 |
|
---|
55 | func fftmul(x, y nat) nat {
|
---|
56 | k, m := fftSize(x, y)
|
---|
57 | xp := polyFromNat(x, k, m)
|
---|
58 | yp := polyFromNat(y, k, m)
|
---|
59 | rp := xp.Mul(&yp)
|
---|
60 | return rp.Int()
|
---|
61 | }
|
---|
62 |
|
---|
63 | // fftSizeThreshold[i] is the maximal size (in bits) where we should use
|
---|
64 | // fft size i.
|
---|
65 | var fftSizeThreshold = [...]int64{0, 0, 0,
|
---|
66 | 4 << 10, 8 << 10, 16 << 10, // 5
|
---|
67 | 32 << 10, 64 << 10, 1 << 18, 1 << 20, 3 << 20, // 10
|
---|
68 | 8 << 20, 30 << 20, 100 << 20, 300 << 20, 600 << 20,
|
---|
69 | }
|
---|
70 |
|
---|
71 | // returns the FFT length k, m the number of words per chunk
|
---|
72 | // such that m << k is larger than the number of words
|
---|
73 | // in x*y.
|
---|
74 | func fftSize(x, y nat) (k uint, m int) {
|
---|
75 | words := len(x) + len(y)
|
---|
76 | bits := int64(words) * int64(_W)
|
---|
77 | k = uint(len(fftSizeThreshold))
|
---|
78 | for i := range fftSizeThreshold {
|
---|
79 | if fftSizeThreshold[i] > bits {
|
---|
80 | k = uint(i)
|
---|
81 | break
|
---|
82 | }
|
---|
83 | }
|
---|
84 | // The 1<<k chunks of m words must have N bits so that
|
---|
85 | // 2^N-1 is larger than x*y. That is, m<<k > words
|
---|
86 | m = words>>k + 1
|
---|
87 | return
|
---|
88 | }
|
---|
89 |
|
---|
90 | // valueSize returns the length (in words) to use for polynomial
|
---|
91 | // coefficients, to compute a correct product of polynomials P*Q
|
---|
92 | // where deg(P*Q) < K (== 1<<k) and where coefficients of P and Q are
|
---|
93 | // less than b^m (== 1 << (m*_W)).
|
---|
94 | // The chosen length (in bits) must be a multiple of 1 << (k-extra).
|
---|
95 | func valueSize(k uint, m int, extra uint) int {
|
---|
96 | // The coefficients of P*Q are less than b^(2m)*K
|
---|
97 | // so we need W * valueSize >= 2*m*W+K
|
---|
98 | n := 2*m*_W + int(k) // necessary bits
|
---|
99 | K := 1 << (k - extra)
|
---|
100 | if K < _W {
|
---|
101 | K = _W
|
---|
102 | }
|
---|
103 | n = ((n / K) + 1) * K // round to a multiple of K
|
---|
104 | return n / _W
|
---|
105 | }
|
---|
106 |
|
---|
107 | // poly represents an integer via a polynomial in Z[x]/(x^K+1)
|
---|
108 | // where K is the FFT length and b^m is the computation basis 1<<(m*_W).
|
---|
109 | // If P = a[0] + a[1] x + ... a[n] x^(K-1), the associated natural number
|
---|
110 | // is P(b^m).
|
---|
111 | type poly struct {
|
---|
112 | k uint // k is such that K = 1<<k.
|
---|
113 | m int // the m such that P(b^m) is the original number.
|
---|
114 | a []nat // a slice of at most K m-word coefficients.
|
---|
115 | }
|
---|
116 |
|
---|
117 | // polyFromNat slices the number x into a polynomial
|
---|
118 | // with 1<<k coefficients made of m words.
|
---|
119 | func polyFromNat(x nat, k uint, m int) poly {
|
---|
120 | p := poly{k: k, m: m}
|
---|
121 | length := len(x)/m + 1
|
---|
122 | p.a = make([]nat, length)
|
---|
123 | for i := range p.a {
|
---|
124 | if len(x) < m {
|
---|
125 | p.a[i] = make(nat, m)
|
---|
126 | copy(p.a[i], x)
|
---|
127 | break
|
---|
128 | }
|
---|
129 | p.a[i] = x[:m]
|
---|
130 | x = x[m:]
|
---|
131 | }
|
---|
132 | return p
|
---|
133 | }
|
---|
134 |
|
---|
135 | // Int evaluates back a poly to its integer value.
|
---|
136 | func (p *poly) Int() nat {
|
---|
137 | length := len(p.a)*p.m + 1
|
---|
138 | if na := len(p.a); na > 0 {
|
---|
139 | length += len(p.a[na-1])
|
---|
140 | }
|
---|
141 | n := make(nat, length)
|
---|
142 | m := p.m
|
---|
143 | np := n
|
---|
144 | for i := range p.a {
|
---|
145 | l := len(p.a[i])
|
---|
146 | c := addVV(np[:l], np[:l], p.a[i])
|
---|
147 | if np[l] < ^big.Word(0) {
|
---|
148 | np[l] += c
|
---|
149 | } else {
|
---|
150 | addVW(np[l:], np[l:], c)
|
---|
151 | }
|
---|
152 | np = np[m:]
|
---|
153 | }
|
---|
154 | n = trim(n)
|
---|
155 | return n
|
---|
156 | }
|
---|
157 |
|
---|
158 | func trim(n nat) nat {
|
---|
159 | for i := range n {
|
---|
160 | if n[len(n)-1-i] != 0 {
|
---|
161 | return n[:len(n)-i]
|
---|
162 | }
|
---|
163 | }
|
---|
164 | return nil
|
---|
165 | }
|
---|
166 |
|
---|
167 | // Mul multiplies p and q modulo X^K-1, where K = 1<<p.k.
|
---|
168 | // The product is done via a Fourier transform.
|
---|
169 | func (p *poly) Mul(q *poly) poly {
|
---|
170 | // extra=2 because:
|
---|
171 | // * some power of 2 is a K-th root of unity when n is a multiple of K/2.
|
---|
172 | // * 2 itself is a square (see fermat.ShiftHalf)
|
---|
173 | n := valueSize(p.k, p.m, 2)
|
---|
174 |
|
---|
175 | pv, qv := p.Transform(n), q.Transform(n)
|
---|
176 | rv := pv.Mul(&qv)
|
---|
177 | r := rv.InvTransform()
|
---|
178 | r.m = p.m
|
---|
179 | return r
|
---|
180 | }
|
---|
181 |
|
---|
182 | // A polValues represents the value of a poly at the powers of a
|
---|
183 | // K-th root of unity θ=2^(l/2) in Z/(b^n+1)Z, where b^n = 2^(K/4*l).
|
---|
184 | type polValues struct {
|
---|
185 | k uint // k is such that K = 1<<k.
|
---|
186 | n int // the length of coefficients, n*_W a multiple of K/4.
|
---|
187 | values []fermat // a slice of K (n+1)-word values
|
---|
188 | }
|
---|
189 |
|
---|
190 | // Transform evaluates p at θ^i for i = 0...K-1, where
|
---|
191 | // θ is a K-th primitive root of unity in Z/(b^n+1)Z.
|
---|
192 | func (p *poly) Transform(n int) polValues {
|
---|
193 | k := p.k
|
---|
194 | inputbits := make([]big.Word, (n+1)<<k)
|
---|
195 | input := make([]fermat, 1<<k)
|
---|
196 | // Now computed q(ω^i) for i = 0 ... K-1
|
---|
197 | valbits := make([]big.Word, (n+1)<<k)
|
---|
198 | values := make([]fermat, 1<<k)
|
---|
199 | for i := range values {
|
---|
200 | input[i] = inputbits[i*(n+1) : (i+1)*(n+1)]
|
---|
201 | if i < len(p.a) {
|
---|
202 | copy(input[i], p.a[i])
|
---|
203 | }
|
---|
204 | values[i] = fermat(valbits[i*(n+1) : (i+1)*(n+1)])
|
---|
205 | }
|
---|
206 | fourier(values, input, false, n, k)
|
---|
207 | return polValues{k, n, values}
|
---|
208 | }
|
---|
209 |
|
---|
210 | // InvTransform reconstructs p (modulo X^K - 1) from its
|
---|
211 | // values at θ^i for i = 0..K-1.
|
---|
212 | func (v *polValues) InvTransform() poly {
|
---|
213 | k, n := v.k, v.n
|
---|
214 |
|
---|
215 | // Perform an inverse Fourier transform to recover p.
|
---|
216 | pbits := make([]big.Word, (n+1)<<k)
|
---|
217 | p := make([]fermat, 1<<k)
|
---|
218 | for i := range p {
|
---|
219 | p[i] = fermat(pbits[i*(n+1) : (i+1)*(n+1)])
|
---|
220 | }
|
---|
221 | fourier(p, v.values, true, n, k)
|
---|
222 | // Divide by K, and untwist q to recover p.
|
---|
223 | u := make(fermat, n+1)
|
---|
224 | a := make([]nat, 1<<k)
|
---|
225 | for i := range p {
|
---|
226 | u.Shift(p[i], -int(k))
|
---|
227 | copy(p[i], u)
|
---|
228 | a[i] = nat(p[i])
|
---|
229 | }
|
---|
230 | return poly{k: k, m: 0, a: a}
|
---|
231 | }
|
---|
232 |
|
---|
233 | // NTransform evaluates p at θω^i for i = 0...K-1, where
|
---|
234 | // θ is a (2K)-th primitive root of unity in Z/(b^n+1)Z
|
---|
235 | // and ω = θ².
|
---|
236 | func (p *poly) NTransform(n int) polValues {
|
---|
237 | k := p.k
|
---|
238 | if len(p.a) >= 1<<k {
|
---|
239 | panic("Transform: len(p.a) >= 1<<k")
|
---|
240 | }
|
---|
241 | // θ is represented as a shift.
|
---|
242 | θshift := (n * _W) >> k
|
---|
243 | // p(x) = a_0 + a_1 x + ... + a_{K-1} x^(K-1)
|
---|
244 | // p(θx) = q(x) where
|
---|
245 | // q(x) = a_0 + θa_1 x + ... + θ^(K-1) a_{K-1} x^(K-1)
|
---|
246 | //
|
---|
247 | // Twist p by θ to obtain q.
|
---|
248 | tbits := make([]big.Word, (n+1)<<k)
|
---|
249 | twisted := make([]fermat, 1<<k)
|
---|
250 | src := make(fermat, n+1)
|
---|
251 | for i := range twisted {
|
---|
252 | twisted[i] = fermat(tbits[i*(n+1) : (i+1)*(n+1)])
|
---|
253 | if i < len(p.a) {
|
---|
254 | for i := range src {
|
---|
255 | src[i] = 0
|
---|
256 | }
|
---|
257 | copy(src, p.a[i])
|
---|
258 | twisted[i].Shift(src, θshift*i)
|
---|
259 | }
|
---|
260 | }
|
---|
261 |
|
---|
262 | // Now computed q(ω^i) for i = 0 ... K-1
|
---|
263 | valbits := make([]big.Word, (n+1)<<k)
|
---|
264 | values := make([]fermat, 1<<k)
|
---|
265 | for i := range values {
|
---|
266 | values[i] = fermat(valbits[i*(n+1) : (i+1)*(n+1)])
|
---|
267 | }
|
---|
268 | fourier(values, twisted, false, n, k)
|
---|
269 | return polValues{k, n, values}
|
---|
270 | }
|
---|
271 |
|
---|
272 | // InvTransform reconstructs a polynomial from its values at
|
---|
273 | // roots of x^K+1. The m field of the returned polynomial
|
---|
274 | // is unspecified.
|
---|
275 | func (v *polValues) InvNTransform() poly {
|
---|
276 | k := v.k
|
---|
277 | n := v.n
|
---|
278 | θshift := (n * _W) >> k
|
---|
279 |
|
---|
280 | // Perform an inverse Fourier transform to recover q.
|
---|
281 | qbits := make([]big.Word, (n+1)<<k)
|
---|
282 | q := make([]fermat, 1<<k)
|
---|
283 | for i := range q {
|
---|
284 | q[i] = fermat(qbits[i*(n+1) : (i+1)*(n+1)])
|
---|
285 | }
|
---|
286 | fourier(q, v.values, true, n, k)
|
---|
287 |
|
---|
288 | // Divide by K, and untwist q to recover p.
|
---|
289 | u := make(fermat, n+1)
|
---|
290 | a := make([]nat, 1<<k)
|
---|
291 | for i := range q {
|
---|
292 | u.Shift(q[i], -int(k)-i*θshift)
|
---|
293 | copy(q[i], u)
|
---|
294 | a[i] = nat(q[i])
|
---|
295 | }
|
---|
296 | return poly{k: k, m: 0, a: a}
|
---|
297 | }
|
---|
298 |
|
---|
299 | // fourier performs an unnormalized Fourier transform
|
---|
300 | // of src, a length 1<<k vector of numbers modulo b^n+1
|
---|
301 | // where b = 1<<_W.
|
---|
302 | func fourier(dst []fermat, src []fermat, backward bool, n int, k uint) {
|
---|
303 | var rec func(dst, src []fermat, size uint)
|
---|
304 | tmp := make(fermat, n+1) // pre-allocate temporary variables.
|
---|
305 | tmp2 := make(fermat, n+1) // pre-allocate temporary variables.
|
---|
306 |
|
---|
307 | // The recursion function of the FFT.
|
---|
308 | // The root of unity used in the transform is ω=1<<(ω2shift/2).
|
---|
309 | // The source array may use shifted indices (i.e. the i-th
|
---|
310 | // element is src[i << idxShift]).
|
---|
311 | rec = func(dst, src []fermat, size uint) {
|
---|
312 | idxShift := k - size
|
---|
313 | ω2shift := (4 * n * _W) >> size
|
---|
314 | if backward {
|
---|
315 | ω2shift = -ω2shift
|
---|
316 | }
|
---|
317 |
|
---|
318 | // Easy cases.
|
---|
319 | if len(src[0]) != n+1 || len(dst[0]) != n+1 {
|
---|
320 | panic("len(src[0]) != n+1 || len(dst[0]) != n+1")
|
---|
321 | }
|
---|
322 | switch size {
|
---|
323 | case 0:
|
---|
324 | copy(dst[0], src[0])
|
---|
325 | return
|
---|
326 | case 1:
|
---|
327 | dst[0].Add(src[0], src[1<<idxShift]) // dst[0] = src[0] + src[1]
|
---|
328 | dst[1].Sub(src[0], src[1<<idxShift]) // dst[1] = src[0] - src[1]
|
---|
329 | return
|
---|
330 | }
|
---|
331 |
|
---|
332 | // Let P(x) = src[0] + src[1<<idxShift] * x + ... + src[K-1 << idxShift] * x^(K-1)
|
---|
333 | // The P(x) = Q1(x²) + x*Q2(x²)
|
---|
334 | // where Q1's coefficients are src with indices shifted by 1
|
---|
335 | // where Q2's coefficients are src[1<<idxShift:] with indices shifted by 1
|
---|
336 |
|
---|
337 | // Split destination vectors in halves.
|
---|
338 | dst1 := dst[:1<<(size-1)]
|
---|
339 | dst2 := dst[1<<(size-1):]
|
---|
340 | // Transform Q1 and Q2 in the halves.
|
---|
341 | rec(dst1, src, size-1)
|
---|
342 | rec(dst2, src[1<<idxShift:], size-1)
|
---|
343 |
|
---|
344 | // Reconstruct P's transform from transforms of Q1 and Q2.
|
---|
345 | // dst[i] is dst1[i] + ω^i * dst2[i]
|
---|
346 | // dst[i + 1<<(k-1)] is dst1[i] + ω^(i+K/2) * dst2[i]
|
---|
347 | //
|
---|
348 | for i := range dst1 {
|
---|
349 | tmp.ShiftHalf(dst2[i], i*ω2shift, tmp2) // ω^i * dst2[i]
|
---|
350 | dst2[i].Sub(dst1[i], tmp)
|
---|
351 | dst1[i].Add(dst1[i], tmp)
|
---|
352 | }
|
---|
353 | }
|
---|
354 | rec(dst, src, k)
|
---|
355 | }
|
---|
356 |
|
---|
357 | // Mul returns the pointwise product of p and q.
|
---|
358 | func (p *polValues) Mul(q *polValues) (r polValues) {
|
---|
359 | n := p.n
|
---|
360 | r.k, r.n = p.k, p.n
|
---|
361 | r.values = make([]fermat, len(p.values))
|
---|
362 | bits := make([]big.Word, len(p.values)*(n+1))
|
---|
363 | buf := make(fermat, 8*n)
|
---|
364 | for i := range r.values {
|
---|
365 | r.values[i] = bits[i*(n+1) : (i+1)*(n+1)]
|
---|
366 | z := buf.Mul(p.values[i], q.values[i])
|
---|
367 | copy(r.values[i], z)
|
---|
368 | }
|
---|
369 | return
|
---|
370 | }
|
---|