/**************************************************************************\

SUMMARY:

Thus, with standard IEEE floating point, you should get the equivalent
of about 106 bits of precision (but actually just a bit less).

The interface allows you to treat quad_floats more or less as if they were
"ordinary" floating point types.

See below for more implementation details.

\**************************************************************************/

#include <NTL/ZZ.h>

public:

explicit quad_float(double a);  // promotion constructor

static void SetOutputPrecision(long p);
// This sets the number of decimal digits to be output.  Default is
// 10.

static long OutputPrecision();
// returns current output precision.

};

/**************************************************************************\

Arithmetic Operations

\**************************************************************************/

// PROMOTIONS: operators +, -, *, / promote double to quad_float
// on (x, y).

void operator++(quad_float& a, int); // postfix

void operator--(quad_float& a, int); // postfix

/**************************************************************************\

Comparison

\**************************************************************************/

long sign(const quad_float& x);  // sign of x, -1, 0, +1
long compare(const quad_float& x, const quad_float& y); // sign of x - y

// PROMOTIONS: operators >, ..., != and function compare
// promote double to quad_float on (x, y).

/**************************************************************************\

Input/Output
Input Syntax:

<number>: [ "-" ] <unsigned-number>
<unsigned-number>: <dotted-number> [ <e-part> ] | <e-part>
<dotted-number>: <digits> | <digits> "." <digits> | "." <digits> | <digits> "."
<digits>: <digit> <digits> | <digit>
<digit>: "0" | ... | "9"
<e-part>: ( "E" | "e" ) [ "+" | "-" ] <digits>

Examples of valid input:

17 1.5 0.5 .5  5.  -.5 e10 e-10 e+10 1.5e10 .5e10 .5E10

Note that the number of decimal digits of precision that are used
for output can be set to any number p >= 1 by calling
The default value of p is 10.
The current value of p is returned by a call to quad_float::OutputPrecision().

\**************************************************************************/

istream& operator >> (istream& s, quad_float& x);
ostream& operator << (ostream& s, const quad_float& x);

/**************************************************************************\

Miscellaneous

\**************************************************************************/

void power(quad_float& x, const quad_float& a, long e); // x = a^e

void power2(quad_float& x, long e); // x = 2^e

long IsFinite(quad_float *x); // checks if x is "finite"
// pointer is used for compatability with
// IsFinite(double*)

// generate a random quad_float x with 0 <= x <= 1

/***********************************************************************\

IMPLEMENTATION DETAILS

A quad_float x is represented as a pair of doubles, x.hi and x.lo,
such that the number represented by x is x.hi + x.lo, where

|x.lo| <= 0.5*ulp(x.hi),  (*)

and ulp(y) means "unit in the last place of y".

For the software to work correctly, IEEE Standard Arithmetic is sufficient.
That includes just about every modern computer; the only exception I'm
aware of is Intel x86 platforms running Linux (but you can still
use this platform--see below).

Also sufficient is any platform that implements arithmetic with correct
rounding, i.e., given double floating point numbers a and b, a op b
is computed exactly and then rounded to the nearest double.
The tie-breaking rule is not important.

This is a rather weird representation;  although it gives one
essentially twice the precision of an ordinary double, it is
not really the equivalent of quadratic precision (despite the name).
For example, the number 1 + 2^{-200} can be represented exactly as
a quad_float.  Also, there is no real notion of "machine precision".

useful rules, even if the underlying floating point arithmetic is IEEE
compliant.  Generally, when an overflow/underflow occurs, the resulting value
is unpredicatble, although typically when overflow occurs in computing a value
x, the result is non-finite (i.e., IsFinite(&x) == 0).  Note, however, that
some care is taken to ensure that the ZZ to quad_float conversion routine
produces a non-finite value upon overflow.

THE INTEL x86/x87 PROBLEM

[The following discussion was written before the advent of SSE2 instructions,
back when all floating point on x86 was done using the x87 FPU instruction set
and registers.  By now, it is mostly of historical interest, as modern x86 CPUs
(since SSE2) use a new set of instructions and registers that avoid all of
these problems, and by default, it seems that all modern C++ compilers (by
default) avoid the x87 altogether.  However, there are new problems (see THE
FMA PROBLEM, below).]

Although just about every modern processor implements the IEEE floating point
standard, there still can be problems on processors that support IEEE extended
double precision.  The only processor I know of that supports this is the x86.

While extended double precision may sound like a nice thing, it is not.  Normal
double precision has 53 bits of precision.  Extended has 64.  On x86s, the FP
registers have 53 or 64 bits of precision---this can be set at run-time by
modifying the cpu "control word" (something that can be done only in assembly
code).  However, doubles stored in memory always have only 53 bits.  Compilers
may move values between memory and registers whenever they want, which can
effectively change the value of a floating point number even though at the
C/C++ level, nothing has happened that should have changed the value.  Is that
sick, or what?  Actually, the new C99 standard seems to outlaw such
"spontaneous" value changes; however, this behavior is not necessarily
universally implemented.

This is a real headache, and if one is not just a bit careful, the quad_float
code will break.  This breaking is not at all subtle, and the program QuadTest
will catch the problem if it exists.

You should not need to worry about any of this, because NTL automatically
detects and works around these problems as best it can, as described below.  It
shouldn't make a mistake, but if it does, you will catch it in the QuadTest
program.  If things don't work quite right, you might try setting NTL_FIX_X86
or NTL_NO_FIX_X86 flags in ntl_config.h, but this should not be necessary.

Here are the details about how NTL fixes the problem.

The first and best way is to have the default setting of the control word be 53
bits.  However, you are at the mercy of your platform (compiler, OS, run-time
libraries).  Windows does this, and so the problem simply does not arise here,
and NTL neither detects nor fixes the problem.  Linux, however, does not do
this, which really sucks.  Can we talk these Linux people into changing this?

The second way to fix the problem is by having NTL fiddle with control word
itself.  If you compile NTL using a GNU compiler on an x86, this should happen
automatically.  On the one hand, this is not a general, portable solution,
since it will only work if you use a GNU compiler, or at least one that
supports GNU 'asm' syntax.  On the other hand, almost everybody who compiles
C++ on x86/Linux platforms uses GNU compilers (although there are some
commercial compilers out there that I don't know too much about).

The third way to fix the problem is to 'force' all intermediate floating point
results into memory.  This is not an 'ideal' fix, since it is not fully
equivalent to 53-bit precision (because of double rounding), but it works
(although to be honest, I've never seen a full proof of correctness in this
case).  NTL's quad_float code does this by storing intermediate results in
local variables declared to be 'volatile'.  This is the solution to the problem
that NTL uses if it detects the problem and can't fix it using the GNU 'asm'
hack mentioned above.  This solution should work on any platform that
faithfully implements 'volatile' according to the ANSI C standard.

THE FMA PROBLEM

Some CPUs come equipped with a fused-multiply-add (FMA) instruction, which
computes x + a*b with just a single rounding.  While this generally is faster
and more precise than performing this using two instructions and two roundings,
FMA instructions can break the logic of quad_float.

To mitigate this problem, NTL tries to detect whether the compiler emits FMA
instructions when it builds the make_desc.h file. The macro NTL_FMA_DETECTED
gets set to 1 if it detects this to be the case. Based on the setting of this
macro, code in quad_float.cpp will strategically replace certain
multiplications a*b with a*b + z, where z is an external variable that will
always be zero (but hopefully the compiler will not figure this out).  It is a
bit of a hack, but it generally works.

THE FLOATING POINT REASSOCIATION PROBLEM

The C++ standard says that compilers must issue instructions
that respect the grouping of floating point operations.
So the compiler is not allowed to compile (a+b)+c as a+(b+c).
Most compilers (at least by default) repect this rule.

One exception is the Intel icc compiler.  Because of this, the quad_float.cpp
file includes a pragma that will force the Intel compiler to respect the
standard.

Note that gcc respects this rule by default, unless you pass -ffast-math as a
compilation flag.  As long as quad_float.cpp is not compiled with this flag,
you should be OK.  The quad_float.cpp file contains a preprocessor check to
detect if it is compiled under gcc with -ffast-math, and if so, reports an
error.

This is the only file in NTL which requires this rule.  Also, it should be OK
to compile client code with -ffast-math, if that is desired.

BACKGROUND INFO

The code NTL uses algorithms designed by Knuth, Kahan, Dekker, and
Linnainmaa.  The original transcription to C++ was done by Douglas
Priest.  Enhancements and bug fixes were done by Keith Briggs ---
see http://keithbriggs.info/doubledouble.html.  The NTL version is a
stripped down version of Briggs' code, with some bug fixes and
portability improvements.

Here is a brief annotated bibliography (compiled by Priest) of papers
dealing with DP and similar techniques, arranged chronologically.

Kahan, W., Further Remarks on Reducing Truncation Errors,
{\it Comm.\ ACM\/} {\bf 8} (1965), 40.

M{\o}ller, O., Quasi Double Precision in Floating-Point Addition,
{\it BIT\/} {\bf 5} (1965), 37--50.

The two papers that first presented the idea of recovering the
roundoff of a sum.

Dekker, T., A Floating-Point Technique for Extending the Available
Precision, {\it Numer.\ Math.} {\bf 18} (1971), 224--242.

The classic reference for DP algorithms for sum, product, quotient,
and square root.

Pichat, M., Correction d'une Somme en Arithmetique \a Virgule
Flottante, {\it Numer.\ Math.} {\bf 19} (1972), 400--406.

An iterative algorithm for computing a protracted sum to working
precision by repeatedly applying the sum-and-roundoff method.

Linnainmaa, S., Analysis of Some Known Methods of Improving the Accuracy
of Floating-Point Sums, {\it BIT\/} {\bf 14} (1974), 167--202.

Comparison of Kahan and M{\o}ller algorithms with variations given
by Knuth.

Bohlender, G., Floating-Point Computation of Functions with Maximum
Accuracy, {\it IEEE Trans.\ Comput.} {\bf C-26} (1977), 621--632.

Extended the analysis of Pichat's algorithm to compute a multi-word
representation of the exact sum of n working precision numbers.
This is the algorithm Kahan has called "distillation".

Linnainmaa, S., Software for Doubled-Precision Floating-Point Computations,
{\it ACM Trans.\ Math.\ Soft.} {\bf 7} (1981), 272--283.

Generalized the hypotheses of Dekker and showed how to take advantage
of extended precision where available.

Leuprecht, H., and W.~Oberaigner, Parallel Algorithms for the Rounding-Exact
Summation of Floating-Point Numbers, {\it Computing} {\bf 28} (1982), 89--104.

Variations of distillation appropriate for parallel and vector
architectures.

Kahan, W., Paradoxes in Concepts of Accuracy, lecture notes from Joint
Seminar on Issues and Directions in Scientific Computation, Berkeley, 1989.

Gives the more accurate DP sum I've shown above, discusses some
examples.

Priest, D., Algorithms for Arbitrary Precision Floating Point Arithmetic,
in P.~Kornerup and D.~Matula, Eds., {\it Proc.\ 10th Symposium on Com-
puter Arithmetic}, IEEE Computer Society Press, Los Alamitos, Calif., 1991.

Extends from DP to arbitrary precision; gives portable algorithms and
general proofs.

Sorensen, D., and P.~Tang, On the Orthogonality of Eigenvectors Computed
by Divide-and-Conquer Techniques, {\it SIAM J.\ Num.\ Anal.} {\bf 28}
(1991), 1752--1775.

Uses some DP arithmetic to retain orthogonality of eigenvectors
computed by a parallel divide-and-conquer scheme.

Priest, D., On Properties of Floating Point Arithmetics: Numerical Stability
and the Cost of Accurate Computations, Ph.D. dissertation, University
of California at Berkeley, 1992.

More examples, organizes proofs in terms of common properties of fp
`