/*
Copyright (C) 1997, 1998, 1999, 2000 Victor Shoup
This program is free software; you can redistribute it and/or
modify it under the terms of the GNU General Public License
as published by the Free Software Foundation; either version 2
of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
*****************************************************
The quad_float package is derived from the doubledouble package of
Keith Briggs. However, the version employed in NTL has been extensively
modified. Below, I attach the copyright notice from the original
doubledouble package, which is currently available at
http://www.labs.bt.com/people/briggsk2
*****************************************************
Copyright (C) 1997 Keith Martin Briggs
Except where otherwise indicated,
this program is free software; you can redistribute it and/or
modify it under the terms of the GNU General Public License
as published by the Free Software Foundation; either version 2
of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
*/
#include <NTL/quad_float.h>
#include <NTL/RR.h>
#include <float.h>
#include <NTL/new.h>
NTL_START_IMPL
#if (defined(__GNUC__) && NTL_EXT_DOUBLE && (defined(__i386__) || defined(__i486__) || defined(__i586__)))
#if (!defined(NTL_X86_FIX) && !defined(NTL_NO_X86_FIX))
#define NTL_X86_FIX
#endif
#endif
#if (NTL_EXT_DOUBLE && !defined(NTL_X86_FIX))
#define DOUBLE volatile double
#else
#define DOUBLE double
#endif
#ifdef NTL_X86_FIX
#define START_FIX \
volatile unsigned short __old_cw, __new_cw; \
asm volatile ("fnstcw %0":"=m" (__old_cw)); \
__new_cw = (__old_cw & ~0x300) | 0x200; \
asm volatile ("fldcw %0": :"m" (__new_cw));
#define END_FIX asm volatile ("fldcw %0": :"m" (__old_cw));
#else
#define START_FIX
#define END_FIX
#endif
static
void normalize(quad_float& z, const double& xhi, const double& xlo)
{
START_FIX
DOUBLE u, v;
u = xhi + xlo;
v = xhi - u;
v = v + xlo;
z.hi = u;
z.lo = v;
END_FIX
}
#if (NTL_BITS_PER_LONG >= NTL_DOUBLE_PRECISION)
quad_float to_quad_float(long n)
{
START_FIX
DOUBLE xhi, xlo;
DOUBLE u, v;
xhi = double(n);
// Because we are assuming 2's compliment integer
// arithmetic, the following prevents long(xhi) from overflowing.
if (n > 0)
xlo = double(n+long(-xhi));
else
xlo = double(n-long(xhi));
// renormalize...just to be safe
u = xhi + xlo;
v = xhi - u;
v = v + xlo;
END_FIX
return quad_float(u, v);
}
quad_float to_quad_float(unsigned long n)
{
START_FIX
DOUBLE xhi, xlo, t;
DOUBLE u, v;
const double bnd = double(1L << (NTL_BITS_PER_LONG-2))*4.0;
xhi = double(n);
if (xhi >= bnd)
t = xhi - bnd;
else
t = xhi;
// we use the "to_long" function here to be as portable as possible.
long llo = to_long(n - (unsigned long)(t));
xlo = double(llo);
// renormalize...just to be safe
u = xhi + xlo;
v = xhi - u;
v = v + xlo;
END_FIX
return quad_float(u, v);
}
#endif
long quad_float::oprec = 10;
void quad_float::SetOutputPrecision(long p)
{
if (p < 1) p = 1;
if (NTL_OVERFLOW(p, 1, 0))
Error("quad_float: output precision too big");
oprec = p;
}
quad_float operator +(const quad_float& x, const quad_float& y ) {
START_FIX
DOUBLE H, h, T, t, S, s, e, f;
DOUBLE t1;
S = x.hi + y.hi;
T = x.lo + y.lo;
e = S - x.hi;
f = T - x.lo;
t1 = S-e;
t1 = x.hi-t1;
s = y.hi-e;
s = s + t1;
t1 = T-f;
t1 = x.lo-t1;
t = y.lo-f;
t = t + t1;
s = s + T;
H = S + s;
h = S - H;
h = h + s;
h = h + t;
e = H + h;
f = H - e;
f = f + h;
END_FIX
return quad_float(e, f);
}
quad_float& operator +=(quad_float& x, const quad_float& y ) {
START_FIX
DOUBLE H, h, T, t, S, s, e, f;
DOUBLE t1;
S = x.hi + y.hi;
T = x.lo + y.lo;
e = S - x.hi;
f = T - x.lo;
t1 = S-e;
t1 = x.hi-t1;
s = y.hi-e;
s = s + t1;
t1 = T-f;
t1 = x.lo-t1;
t = y.lo-f;
t = t + t1;
s = s + T;
H = S + s;
h = S - H;
h = h + s;
h = h + t;
e = H + h;
f = H - e;
f = f + h;
x.hi = e;
x.lo = f;
END_FIX
return x;
}
quad_float operator -(const quad_float& x, const quad_float& y ) {
START_FIX
DOUBLE H, h, T, t, S, s, e, f;
DOUBLE t1, yhi, ylo;
yhi = -y.hi;
ylo = -y.lo;
S = x.hi + yhi;
T = x.lo + ylo;
e = S - x.hi;
f = T - x.lo;
t1 = S-e;
t1 = x.hi-t1;
s = yhi-e;
s = s + t1;
t1 = T-f;
t1 = x.lo-t1;
t = ylo-f;
t = t + t1;
s = s + T;
H = S + s;
h = S - H;
h = h + s;
h = h + t;
e = H + h;
f = H - e;
f = f + h;
END_FIX
return quad_float(e, f);
}
quad_float& operator -=(quad_float& x, const quad_float& y ) {
START_FIX
DOUBLE H, h, T, t, S, s, e, f;
DOUBLE t1, yhi, ylo;
yhi = -y.hi;
ylo = -y.lo;
S = x.hi + yhi;
T = x.lo + ylo;
e = S - x.hi;
f = T - x.lo;
t1 = S-e;
t1 = x.hi-t1;
s = yhi-e;
s = s + t1;
t1 = T-f;
t1 = x.lo-t1;
t = ylo-f;
t = t + t1;
s = s + T;
H = S + s;
h = S - H;
h = h + s;
h = h + t;
e = H + h;
f = H - e;
f = f + h;
x.hi = e;
x.lo = f;
END_FIX
return x;
}
quad_float operator -(const quad_float& x)
{
START_FIX
DOUBLE xhi, xlo, u, v;
xhi = -x.hi;
xlo = -x.lo;
// it is a good idea to renormalize here, just in case
// the rounding rule depends on sign, and thus we will
// maintain the "normal form" for quad_float's.
u = xhi + xlo;
v = xhi - u;
v = v + xlo;
END_FIX
return quad_float(u, v);
}
quad_float operator *(const quad_float& x,const quad_float& y ) {
START_FIX
DOUBLE hx, tx, hy, ty, C, c;
DOUBLE t1, t2;
C = NTL_QUAD_FLOAT_SPLIT*x.hi;
hx = C-x.hi;
c = NTL_QUAD_FLOAT_SPLIT*y.hi;
hx = C-hx;
tx = x.hi-hx;
hy = c-y.hi;
C = x.hi*y.hi;
hy = c-hy;
ty = y.hi-hy;
// c = ((((hx*hy-C)+hx*ty)+tx*hy)+tx*ty)+(x.hi*y.lo+x.lo*y.hi);
t1 = hx*hy;
t1 = t1-C;
t2 = hx*ty;
t1 = t1+t2;
t2 = tx*hy;
t1 = t1+t2;
t2 = tx*ty;
c = t1+t2;
t1 = x.hi*y.lo;
t2 = x.lo*y.hi;
t1 = t1+t2;
c = c + t1;
hx = C+c;
tx = C-hx;
tx = tx+c;
END_FIX
return quad_float(hx, tx);
}
quad_float& operator *=(quad_float& x,const quad_float& y ) {
START_FIX
DOUBLE hx, tx, hy, ty, C, c;
DOUBLE t1, t2;
C = NTL_QUAD_FLOAT_SPLIT*x.hi;
hx = C-x.hi;
c = NTL_QUAD_FLOAT_SPLIT*y.hi;
hx = C-hx;
tx = x.hi-hx;
hy = c-y.hi;
C = x.hi*y.hi;
hy = c-hy;
ty = y.hi-hy;
// c = ((((hx*hy-C)+hx*ty)+tx*hy)+tx*ty)+(x.hi*y.lo+x.lo*y.hi);
t1 = hx*hy;
t1 = t1-C;
t2 = hx*ty;
t1 = t1+t2;
t2 = tx*hy;
t1 = t1+t2;
t2 = tx*ty;
c = t1+t2;
t1 = x.hi*y.lo;
t2 = x.lo*y.hi;
t1 = t1+t2;
c = c + t1;
hx = C+c;
tx = C-hx;
tx = tx+c;
x.hi = hx;
x.lo = tx;
END_FIX
return x;
}
quad_float operator /(const quad_float& x, const quad_float& y ) {
START_FIX
DOUBLE hc, tc, hy, ty, C, c, U, u;
DOUBLE t1;
C = x.hi/y.hi;
c = NTL_QUAD_FLOAT_SPLIT*C;
hc = c-C;
u = NTL_QUAD_FLOAT_SPLIT*y.hi;
hc = c-hc;
tc = C-hc;
hy = u-y.hi;
U = C * y.hi;
hy = u-hy;
ty = y.hi-hy;
// u = (((hc*hy-U)+hc*ty)+tc*hy)+tc*ty;
u = hc*hy;
u = u-U;
t1 = hc*ty;
u = u+t1;
t1 = tc*hy;
u = u+t1;
t1 = tc*ty;
u = u+t1;
// c = ((((x.hi-U)-u)+x.lo)-C*y.lo)/y.hi;
c = x.hi-U;
c = c-u;
c = c+x.lo;
t1 = C*y.lo;
c = c - t1;
c = c/y.hi;
hy = C+c;
ty = C-hy;
ty = ty+c;
END_FIX
return quad_float(hy, ty);
}
quad_float& operator /=(quad_float& x, const quad_float& y ) {
START_FIX
DOUBLE hc, tc, hy, ty, C, c, U, u;
DOUBLE t1;
C = x.hi/y.hi;
c = NTL_QUAD_FLOAT_SPLIT*C;
hc = c-C;
u = NTL_QUAD_FLOAT_SPLIT*y.hi;
hc = c-hc;
tc = C-hc;
hy = u-y.hi;
U = C * y.hi;
hy = u-hy;
ty = y.hi-hy;
// u = (((hc*hy-U)+hc*ty)+tc*hy)+tc*ty;
u = hc*hy;
u = u-U;
t1 = hc*ty;
u = u+t1;
t1 = tc*hy;
u = u+t1;
t1 = tc*ty;
u = u+t1;
// c = ((((x.hi-U)-u)+x.lo)-C*y.lo)/y.hi;
c = x.hi-U;
c = c-u;
c = c+x.lo;
t1 = C*y.lo;
c = c - t1;
c = c/y.hi;
hy = C+c;
ty = C-hy;
ty = ty+c;
x.hi = hy;
x.lo = ty;
END_FIX
return x;
}
quad_float sqrt(const quad_float& y) {
if (y.hi < 0.0)
Error("Quad: attempto to take square root of negative number");
if (y.hi == 0.0) return quad_float(0.0,0.0);
double c;
c = sqrt(y.hi);
ForceToMem(&c); // This is fairly paranoid, but it doesn't cost too much.
START_FIX
DOUBLE p,q,hx,tx,u,uu,cc;
DOUBLE t1;
p = NTL_QUAD_FLOAT_SPLIT*c;
hx = (c-p);
hx = hx+p;
tx = c-hx;
p = hx*hx;
q = hx*tx;
q = q+q;
u = p+q;
uu = p-u;
uu = uu+q;
t1 = tx*tx;
uu = uu+t1;
cc = y.hi-u;
cc = cc-uu;
cc = cc+y.lo;
t1 = c+c;
cc = cc/t1;
hx = c+cc;
tx = c-hx;
tx = tx+cc;
END_FIX
return quad_float(hx, tx);
}
void power(quad_float& z, const quad_float& a, long e)
{
quad_float res, u;
unsigned long k;
if (e < 0)
k = -((unsigned long) e);
else
k = e;
res = 1.0;
u = a;
while (k) {
if (k & 1)
res = res * u;
k = k >> 1;
if (k)
u = u * u;
}
if (e < 0)
z = 1.0/res;
else
z = res;
}
void power2(quad_float& z, long e)
{
z.hi = _ntl_ldexp(1.0, e);
z.lo = 0;
}
long to_long(const quad_float& x)
{
double fhi, flo;
fhi = floor(x.hi);
if (fhi == x.hi)
flo = floor(x.lo);
else
flo = 0;
// the following code helps to prevent unnecessary integer overflow,
// and guarantees that to_long(to_quad_float(a)) == a, for all long a,
// provided long's are not too wide.
if (fhi > 0)
return long(flo) - long(-fhi);
else
return long(fhi) + long(flo);
}
// This version of ZZ to quad_float coversion relies on the
// precise rounding rules implemented by the ZZ to double conversion.
void conv(quad_float& z, const ZZ& a)
{
double xhi, xlo;
conv(xhi, a);
if (!IsFinite(&xhi)) {
z.hi = xhi;
z.lo = 0;
return;
}
static ZZ t;
conv(t, xhi);
sub(t, a, t);
conv(xlo, t);
normalize(z, xhi, xlo);
// The following is just paranoia.
if (fabs(z.hi) < NTL_FDOUBLE_PRECISION && z.lo != 0)
Error("internal error: ZZ to quad_float conversion");
}
void conv(ZZ& z, const quad_float& x)
{
static ZZ t1, t2, t3;
double fhi, flo;
fhi = floor(x.hi);
if (fhi == x.hi) {
flo = floor(x.lo);
conv(t1, fhi);
conv(t2, flo);
add(z, t1, t2);
}
else
conv(z, fhi);
}
ostream& operator<<(ostream& s, const quad_float& a)
{
quad_float aa = a;
if (!IsFinite(&aa)) {
s << "NaN";
return s;
}
long old_p = RR::precision();
long old_op = RR::OutputPrecision();
RR::SetPrecision(long(3.33*quad_float::oprec) + 10);
RR::SetOutputPrecision(quad_float::oprec);
static RR t;
conv(t, a);
s << t;
RR::SetPrecision(old_p);
RR::SetOutputPrecision(old_op);
return s;
}
istream& operator>>(istream& s, quad_float& x)
{
long old_p = RR::precision();
RR::SetPrecision(4*NTL_DOUBLE_PRECISION);
static RR t;
s >> t;
conv(x, t);
RR::SetPrecision(old_p);
return s;
}
void random(quad_float& x)
{
long old_p = RR::precision();
RR::SetPrecision(4*NTL_DOUBLE_PRECISION);
static RR t;
random(t);
conv(x, t);
RR::SetPrecision(old_p);
}
quad_float random_quad_float()
{
quad_float x;
random(x);
return x;
}
long IsFinite(quad_float *x)
{
return IsFinite(&x->hi) && IsFinite(&x->lo);
}
long PrecisionOK()
{
START_FIX
long k;
DOUBLE l1 = (double)1;
DOUBLE lh = 1/(double)2;
DOUBLE epsilon;
DOUBLE fudge, oldfudge;
epsilon = l1;
fudge = l1+l1;
k = 0;
do {
k++;
epsilon = epsilon * lh;
oldfudge = fudge;
fudge = l1 + epsilon;
} while (fudge > l1 && fudge < oldfudge);
END_FIX
return k == NTL_DOUBLE_PRECISION;
}
quad_float floor(const quad_float& x)
{
double fhi = floor(x.hi);
if (fhi != x.hi)
return quad_float(fhi, 0.0);
else {
double flo = floor(x.lo);
quad_float z;
normalize(z, fhi, flo);
return z;
}
}
quad_float ceil(const quad_float& x) {
return -floor(-x);
}
quad_float trunc(const quad_float& x) {
if (x>=0.0) return floor(x); else return -floor(-x);
}
long compare(const quad_float& x, const quad_float& y)
{
if (x.hi > y.hi)
return 1;
else if (x.hi < y.hi)
return -1;
else if (x.lo > y.lo)
return 1;
else if (x.lo < y.lo)
return -1;
else
return 0;
}
quad_float fabs(const quad_float& x)
{ if (x.hi>=0.0) return x; else return -x; }
quad_float to_quad_float(const char *s)
{
quad_float x;
long old_p = RR::precision();
RR::SetPrecision(4*NTL_DOUBLE_PRECISION);
static RR t;
conv(t, s);
conv(x, t);
RR::SetPrecision(old_p);
return x;
}
quad_float ldexp(const quad_float& x, long exp) { // x*2^exp
double xhi, xlo;
quad_float z;
xhi = _ntl_ldexp(x.hi, exp);
xlo = _ntl_ldexp(x.lo, exp);
normalize(z, xhi, xlo);
return z;
}
quad_float exp(const quad_float& x) { // New version 97 Aug 05
/*
! Calculate a quadruple-precision exponential
! Method:
! x x.log2(e) nint[x.log2(e)] + frac[x.log2(e)]
! e = 2 = 2
!
! iy fy
! = 2 . 2
! Then
! fy y.loge(2)
! 2 = e
!
! Now y.loge(2) will be less than 0.3466 in absolute value.
! This is halved and a Pade aproximation is used to approximate e^x over
! the region (-0.1733, +0.1733). This approximation is then squared.
*/
if (x.hi<DBL_MIN_10_EXP*2.302585092994045684017991)
return to_quad_float(0.0);
if (x.hi>DBL_MAX_10_EXP*2.302585092994045684017991) {
Error("exp(quad_float): overflow");
}
// changed this from "const" to "static" in v5.3, since "const"
// causes the initialization to be performed with *every* invocation.
static quad_float Log2 =
to_quad_float("0.6931471805599453094172321214581765680755");
quad_float y,temp,ysq,sum1,sum2;
long iy;
y=x/Log2;
temp = floor(y+0.5);
iy = to_long(temp);
y=(y-temp)*Log2;
y=ldexp(y,-1L);
ysq=y*y;
sum1=y*((((ysq+3960.0)*ysq+2162160.0)*ysq+302702400.0)*ysq+8821612800.0);
sum2=(((90.0*ysq+110880.0)*ysq+30270240.0)*ysq+2075673600.0)*ysq+17643225600.0;
/*
! sum2 + sum1 2.sum1
! Now approximation = ----------- = 1 + ----------- = 1 + 2.temp
! sum2 - sum1 sum2 - sum1
!
! Then (1 + 2.temp)^2 = 4.temp.(1 + temp) + 1
*/
temp=sum1/(sum2-sum1);
y=temp*(temp+1);
y=ldexp(y,2L);
return ldexp(y+1,iy);
}
quad_float log(const quad_float& t) { // Newton method. See Bailey, MPFUN
if (t.hi <= 0.0) {
Error("log(quad_float): argument must be positive");
}
double s1 = log(t.hi);
ForceToMem(&s1); // Again, this is fairly paranoid.
quad_float s;
s = s1;
quad_float e;
e=exp(s);
return s+(t-e)/e; // Newton step
}
long operator> (const quad_float& x, const quad_float& y) {
return (x.hi> y.hi) || (x.hi==y.hi && x.lo> y.lo); }
long operator>=(const quad_float& x, const quad_float& y) {
return (x.hi>y.hi) || (x.hi==y.hi && x.lo>=y.lo); }
long operator< (const quad_float& x, const quad_float& y) {
return (x.hi< y.hi) || (x.hi==y.hi && x.lo< y.lo); }
long operator<=(const quad_float& x, const quad_float& y) {
return (x.hi<y.hi) || (x.hi==y.hi && x.lo<=y.lo); }
long operator==(const quad_float& x, const quad_float& y)
{ return x.hi==y.hi && x.lo==y.lo; }
long operator!=(const quad_float& x, const quad_float& y)
{ return x.hi!=y.hi || x.lo!=y.lo; }
NTL_END_IMPL
syntax highlighted by Code2HTML, v. 0.9.1