我正在使用.NET 4的System.Numerics.BigInteger structure。
我需要计算非常大数字的平方(x 2 ) - 数百万的十进制数字。
如果x
是BigInteger
,那么时间复杂度是多少:
x*x;
或
BigInteger.Pow(x,2);
如何使用.NET 4 BigInteger以最快的方式倍增这么大的数字?是否有Schönhage–Strassen algorithm的实施?
答案 0 :(得分:7)
这取决于你的数字有多大。正如维基百科页面告诉你的那样:
在实践中,Schönhage-Strassen算法开始优于旧方法,如Karatsuba和Toom-Cook乘法,数量超过2 2 15 至2 2 < sup> 17 (10,000到40,000十进制数字)。
System.Numerics.BigInteger
使用Karatsuba algorithm或标准的教科书乘法,具体取决于数字的大小。 Karatsuba的时间复杂度为O( n log 2 3 )。但如果你的数字小于上面引用的数字,那么你可能不会看到实施Schönhage-Strassen的速度加快。
对于Pow()
,这个数字在计算过程中至少将数字平方一次(并且只需执行num * num
即可 - 所以我认为这也不会更有效。
答案 1 :(得分:3)
这里有一个Schönhage-Strassen实施:
答案 2 :(得分:2)
一种非常简单的实现方法是基于FFT。由于将两个数相乘可以执行系数的卷积,然后通过消除进位的通过,您应该能够通过FFT方法(n =位数)在O(n log n)运算中进行卷积。
数字食谱有一章就此。对于如此大的数字来说,这肯定比分割和征服方法(如Karatsuba)更快。
答案 3 :(得分:2)
System.Numerics.BigInteger
不使用[Karatsuba
算法]使用O(n 0.5 )并使用标准的教科书乘法O(n 2 )。通过此代码,您可以在1.4毫秒内实现多个两个30,000位(大约9000个十进制数字)。
public void benchMark()
{
Xint U, V,Temp;
int n = 30000;
while (n > 29000)
{
U = RND(n << 1);
//_______________________
sw.Restart();
Temp = U * U;
sw.Stop();
label7.Text = Convert.ToString("Micro " + sw.Elapsed.TotalMilliseconds + " ms");
//_______________________
}
n>>=1;
}
public static Xint MTP(Xint U, Xint V)
{
return MTP(U, V, Xint.Max(U.Sign * U, V.Sign * V).ToByteArray().Length << 3);
}
public static Xint MTP(Xint U, Xint V, int n)
{
if (n <= 3000) return U * V;
if (n <= 6000) return TC2(U, V, n);
if (n <= 10000) return TC3(U, V, n);
if (n <= 40000) return TC4(U, V, n);
return TC2P(U, V, n);
}
private static Xint MTPr(Xint U, Xint V, int n)
{
if (n <= 3000) return U * V;
if (n <= 6000) return TC2(U, V, n);
if (n <= 10000) return TC3(U, V, n);
return TC4(U, V, n);
}
private static Xint TC2(Xint U1, Xint V1, int n)
{
n >>= 1;
Xint Mask = (Xint.One << n) - 1;
Xint U0 = U1 & Mask; U1 >>= n;
Xint V0 = V1 & Mask; V1 >>= n;
Xint P0 = MTPr(U0, V0, n);
Xint P2 = MTPr(U1, V1, n);
return ((P2 << n) + (MTPr(U0 + U1, V0 + V1, n) - (P0 + P2)) << n) + P0;
}
private static Xint TC3(Xint U2, Xint V2, int n)
{
n = (int)((long)(n) * 0x55555556 >> 32); // n /= 3;
Xint Mask = (Xint.One << n) - 1;
Xint U0 = U2 & Mask; U2 >>= n;
Xint U1 = U2 & Mask; U2 >>= n;
Xint V0 = V2 & Mask; V2 >>= n;
Xint V1 = V2 & Mask; V2 >>= n;
Xint W0 = MTPr(U0, V0, n);
Xint W4 = MTPr(U2, V2, n);
Xint P3 = MTPr((((U2 << 1) + U1) << 1) + U0, (((V2 << 1) + V1 << 1)) + V0, n);
U2 += U0;
V2 += V0;
Xint P2 = MTPr(U2 - U1, V2 - V1, n);
Xint P1 = MTPr(U2 + U1, V2 + V1, n);
Xint W2 = (P1 + P2 >> 1) - (W0 + W4);
Xint W3 = W0 - P1;
W3 = ((W3 + P3 - P2 >> 1) + W3) / 3 - (W4 << 1);
Xint W1 = P1 - (W4 + W3 + W2 + W0);
return ((((W4 << n) + W3 << n) + W2 << n) + W1 << n) + W0;
}
private static Xint TC4(Xint U3, Xint V3, int n)
{
n >>= 2;
Xint Mask = (Xint.One << n) - 1;
Xint U0 = U3 & Mask; U3 >>= n;
Xint U1 = U3 & Mask; U3 >>= n;
Xint U2 = U3 & Mask; U3 >>= n;
Xint V0 = V3 & Mask; V3 >>= n;
Xint V1 = V3 & Mask; V3 >>= n;
Xint V2 = V3 & Mask; V3 >>= n;
Xint W0 = MTPr(U0, V0, n); // 0
U0 += U2; U1 += U3;
V0 += V2; V1 += V3;
Xint P1 = MTPr(U0 + U1, V0 + V1, n); // 1
Xint P2 = MTPr(U0 - U1, V0 - V1, n); // -1
U0 += 3 * U2; U1 += 3 * U3;
V0 += 3 * V2; V1 += 3 * V3;
Xint P3 = MTPr(U0 + (U1 << 1), V0 + (V1 << 1), n); // 2
Xint P4 = MTPr(U0 - (U1 << 1), V0 - (V1 << 1), n); // -2
Xint P5 = MTPr(U0 + 12 * U2 + ((U1 + 12 * U3) << 2),
V0 + 12 * V2 + ((V1 + 12 * V3) << 2), n); // 4
Xint W6 = MTPr(U3, V3, n); // inf
Xint W1 = P1 + P2;
Xint W4 = (((((P3 + P4) >> 1) - (W1 << 1)) / 3 + W0) >> 2) - 5 * W6;
Xint W2 = (W1 >> 1) - (W6 + W4 + W0);
P1 = P1 - P2;
P4 = P4 - P3;
Xint W5 = ((P1 >> 1) + (5 * P4 + P5 - W0 >> 4) - ((((W6 << 4) + W4) << 4) + W2)) / 45;
W1 = ((P4 >> 2) + (P1 << 1)) / 3 + (W5 << 2);
Xint W3 = (P1 >> 1) - (W1 + W5);
return ((((((W6 << n) + W5 << n) + W4 << n) + W3 << n) + W2 << n) + W1 << n) + W0;
}
private static Xint TC2P(Xint A, Xint B, int n)
{
n >>= 1;
Xint Mask = (Xint.One << n) - 1;
Xint[] U = new Xint[3];
U[0] = A & Mask; A >>= n; U[2] = A; U[1] = U[0] + A;
Xint[] V = new Xint[3];
V[0] = B & Mask; B >>= n; V[2] = B; V[1] = V[0] + B;
Xint[] P = new Xint[3];
Parallel.For(0, 3, (int i) => P[i] = MTPr(U[i], V[i], n));
return ((P[2] << n) + P[1] - (P[0] + P[2]) << n) + P[0];
}
private static long current_n;
private static Xint Product(int n)
{
if (n > 2) return MTP(Product(n - (n >>= 1)), Product(n));
if (n > 1) return (current_n += 2) * (current_n += 2);
return current_n += 2;
}
public static Xint[] DQR(Xint A, Xint B)
{
int n = bL(B);
int m = bL(A) - n;
if (m <= 6000) return SmallDivRem(A, B);
int signA = A.Sign; A *= signA;
int signB = B.Sign; B *= signB;
Xint[] QR = new Xint[2];
if (m <= n) QR = D21(A, B, n);
else
{
Xint Mask = (Xint.One << n) - 1;
m /= n;
Xint[] U = new Xint[m];
int i = 0;
for (; i < m; i++)
{
U[i] = A & Mask;
A >>= n;
}
QR = D21(A, B, n);
A = QR[0];
for (i--; i >= 0; i--)
{
QR = D21(QR[1] << n | U[i], B, n);
A = A << n | QR[0];
}
QR[0] = A;
}
QR[0] *= signA * signB;
QR[1] *= signA;
return QR;
}
private static Xint[] SmallDivRem(Xint A, Xint B)
{
Xint[] QR = new Xint[2];
QR[0] = Xint.DivRem(A, B, out QR[1]);
return QR;
}
private static Xint[] D21(Xint A, Xint B, int n)
{
if (n <= 6000) return SmallDivRem(A, B);
int s = n & 1;
A <<= s;
B <<= s;
n++;
n >>= 1;
Xint Mask = (Xint.One << n) - 1;
Xint B1 = B >> n;
Xint B2 = B & Mask;
Xint[] QR1 = D32(A >> (n << 1), A >> n & Mask, B, B1, B2, n);
Xint[] QR2 = D32(QR1[1], A & Mask, B, B1, B2, n);
QR2[0] |= QR1[0] << n;
QR2[1] >>= s;
return QR2;
}
private static Xint[] D32(Xint A12, Xint A3, Xint B, Xint B1, Xint B2, int n)
{
Xint[] QR = new Xint[2];
if ((A12 >> n) != B1) QR = D21(A12, B1, n);
else
{
QR[0] = (Xint.One << n) - 1;
QR[1] = A12 + B1 - (B1 << n);
}
QR[1] = (QR[1] << n | A3) - MTP(QR[0], B2, n);
while (QR[1] < 0)
{
QR[0] -= 1;
QR[1] += B;
}
return QR;
}
public static Xint SQ(Xint U)
{
return SQ(U, U.Sign * U.ToByteArray().Length << 3);
}
public static Xint SQ(Xint U, int n)
{
if (n <= 700) return U * U;
if (n <= 3000) return Xint.Pow(U, 2);
if (n <= 6000) return SQ2(U, n);
if (n <= 10000) return SQ3(U, n);
if (n <= 40000) return SQ4(U, n);
return SQ2P(U, n);
}
private static Xint SQr(Xint U, int n)
{
if (n <= 3000) return Xint.Pow(U, 2);
if (n <= 6000) return SQ2(U, n);
if (n <= 10000) return SQ3(U, n);
return SQ4(U, n);
}
private static Xint SQ2(Xint U1, int n)
{
n >>= 1;
Xint U0 = U1 & ((Xint.One << n) - 1); U1 >>= n;
Xint P0 = SQr(U0, n);
Xint P2 = SQr(U1, n);
return ((P2 << n) + (SQr(U0 + U1, n) - (P0 + P2)) << n) + P0;
}
private static Xint SQ3(Xint U2, int n)
{
n = (int)((long)(n) * 0x55555556 >> 32);
Xint Mask = (Xint.One << n) - 1;
Xint U0 = U2 & Mask; U2 >>= n;
Xint U1 = U2 & Mask; U2 >>= n;
Xint W0 = SQr(U0, n);
Xint W4 = SQr(U2, n);
Xint P3 = SQr((((U2 << 1) + U1) << 1) + U0, n);
U2 += U0;
Xint P2 = SQr(U2 - U1, n);
Xint P1 = SQr(U2 + U1, n);
Xint W2 = (P1 + P2 >> 1) - (W0 + W4);
Xint W3 = W0 - P1;
W3 = ((W3 + P3 - P2 >> 1) + W3) / 3 - (W4 << 1);
Xint W1 = P1 - (W4 + W3 + W2 + W0);
return ((((W4 << n) + W3 << n) + W2 << n) + W1 << n) + W0;
}
private static Xint SQ4(Xint U3, int n)
{
n >>= 2;
Xint Mask = (Xint.One << n) - 1;
Xint U0 = U3 & Mask; U3 >>= n;
Xint U1 = U3 & Mask; U3 >>= n;
Xint U2 = U3 & Mask; U3 >>= n;
Xint W0 = SQr(U0, n); // 0
U0 += U2;
U1 += U3;
Xint P1 = SQr(U0 + U1, n); // 1
Xint P2 = SQr(U0 - U1, n); // -1
U0 += 3 * U2;
U1 += 3 * U3;
Xint P3 = SQr(U0 + (U1 << 1), n); // 2
Xint P4 = SQr(U0 - (U1 << 1), n); // -2
Xint P5 = SQr(U0 + 12 * U2 + ((U1 + 12 * U3) << 2), n); // 4
Xint W6 = SQr(U3, n); // inf
Xint W1 = P1 + P2;
Xint W4 = (((((P3 + P4) >> 1) - (W1 << 1)) / 3 + W0) >> 2) - 5 * W6;
Xint W2 = (W1 >> 1) - (W6 + W4 + W0);
P1 = P1 - P2;
P4 = P4 - P3;
Xint W5 = ((P1 >> 1) + (5 * P4 + P5 - W0 >> 4) - ((((W6 << 4) + W4) << 4) + W2)) / 45;
W1 = ((P4 >> 2) + (P1 << 1)) / 3 + (W5 << 2);
Xint W3 = (P1 >> 1) - (W1 + W5);
return ((((((W6 << n) + W5 << n) + W4 << n) + W3 << n) + W2 << n) + W1 << n) + W0;
}
private static Xint SQ2P(Xint A, int n)
{
n >>= 1;
Xint[] U = new Xint[3];
U[0] = A & ((Xint.One << n) - 1); A >>= n; U[2] = A; U[1] = U[0] + A;
Xint[] P = new Xint[3];
Parallel.For(0, 3, (int i) => P[i] = SQr(U[i], n));
return ((P[2] << n) + P[1] - (P[0] + P[2]) << n) + P[0];
}
public static Xint POW(Xint X, int y)
{
if (y > 1) return ((y & 1) == 1) ? SQ(POW(X, y >> 1)) * X : SQ(POW(X, y >> 1));
if (y == 1) return X;
return 1;
}
public static Xint[] SR(Xint A)
{
return SR(A, bL(A));
}
public static Xint[] SR(Xint A, int n)
{
if (n < 53) return SR52(A);
int m = n >> 2;
Xint Mask = (Xint.One << m) - 1;
Xint A0 = A & Mask; A >>= m;
Xint A1 = A & Mask; A >>= m;
Xint[] R = SR(A, n - (m << 1));
Xint[] D = DQR(R[1] << m | A1, R[0] << 1);
R[0] = (R[0] << m) + D[0];
R[1] = (D[1] << m | A0) - SQ(D[0], m);
if (R[1] < 0)
{
R[0] -= 1;
R[1] += (R[0] << 1) | 1;
}
return R;
}
private static Xint[] SR52(Xint A)
{
double a = (double)A;
long q = (long)Math.Sqrt(a);
long r = (long)(a) - q * q;
Xint[] QR = { q, r };
return QR;
}
public static Xint SRO(Xint A)
{
return SRO(A, bL(A));
}
public static Xint SRO(Xint A, int n)
{
if (n < 53) return (int)Math.Sqrt((double)A);
Xint[] R = SROr(A, n, 1);
return R[0];
}
private static Xint[] SROr(Xint A, int n, int rc) // rc=recursion counter
{
if (n < 53) return SR52(A);
int m = n >> 2;
Xint Mask = (Xint.One << m) - 1;
Xint A0 = A & Mask; A >>= m;
Xint A1 = A & Mask; A >>= m;
Xint[] R = SROr(A, n - (m << 1), rc + 1);
Xint[] D = DQR((R[1] << m) | A1, R[0] << 1);
R[0] = (R[0] << m) + D[0];
rc--;
if (rc != 0)
{
R[1] = (D[1] << m | A0) - SQ(D[0], m);
if (R[1] < 0)
{
R[0] -= 1;
R[1] += (R[0] << 1) | 1;
}
return R;
}
n = (bL(D[0]) << 1) - bL(D[1] << m | A0);
if (n < 0) return R;
if (n > 1)
{
R[0] -= 1;
return R;
}
int shift = (bL(D[0]) - 31) << 1;
long d0 = (int)(D[0] >> (shift >> 1));
long d = (long)((D[1] >> (shift - m)) | (A0 >> shift)) - d0 * d0;
if (d < 0)
{
R[0] -= 1;
return R;
}
if (d > d0 << 1) return R;
R[0] -= (1 - (((D[1] << m) | A0) - SQ(D[0], m)).Sign) >> 1;
return R;
}
public static int bL(Xint U)
{
byte[] bytes = (U.Sign * U).ToByteArray();
int i = bytes.Length - 1;
return (i << 3) + bitLengthMostSignificantByte(bytes[i]);
}
private static int bitLengthMostSignificantByte(byte b)
{
return b < 08 ? b < 02 ? b < 01 ? 0 : 1 :
b < 04 ? 2 : 3 :
b < 32 ? b < 16 ? 4 : 5 :
b < 64 ? 6 : 7;
}
public static int fL2(int i)
{
return
i < 1 << 15 ? i < 1 << 07 ? i < 1 << 03 ? i < 1 << 01 ? i < 1 << 00 ? -1 : 00 :
i < 1 << 02 ? 01 : 02 :
i < 1 << 05 ? i < 1 << 04 ? 03 : 04 :
i < 1 << 06 ? 05 : 06 :
i < 1 << 11 ? i < 1 << 09 ? i < 1 << 08 ? 07 : 08 :
i < 1 << 10 ? 09 : 10 :
i < 1 << 13 ? i < 1 << 12 ? 11 : 12 :
i < 1 << 14 ? 13 : 14 :
i < 1 << 23 ? i < 1 << 19 ? i < 1 << 17 ? i < 1 << 16 ? 15 : 16 :
i < 1 << 18 ? 17 : 18 :
i < 1 << 21 ? i < 1 << 20 ? 19 : 20 :
i < 1 << 22 ? 21 : 22 :
i < 1 << 27 ? i < 1 << 25 ? i < 1 << 24 ? 23 : 24 :
i < 1 << 26 ? 25 : 26 :
i < 1 << 29 ? i < 1 << 28 ? 27 : 28 :
i < 1 << 30 ? 29 : 30;
}
private static int seed;
public static Xint RND(int n)
{
if (n < 2) return n;
if (seed == int.MaxValue) seed = 0; else seed++;
Random rand = new Random(seed);
byte[] bytes = new byte[(n + 15) >> 3];
rand.NextBytes(bytes);
int i = bytes.Length - 1;
bytes[i] = 0;
n = (i << 3) - n;
i--;
bytes[i] >>= n;
bytes[i] |= (byte)(128 >> n);
return new Xint(bytes);
}
//++++++++++++++++++++++++++++++++++++++++++++++++++++++++
} }
答案 4 :(得分:1)
您可以将C# wrapper用于GNU MP Bignum Library,这可能会尽可能快。对于纯C#实现,您可以尝试IntX。
最快的乘法算法实际上是Fürer's algorithm,但我没有找到任何实现。