Last update: Jan 3, 2022

Library for arithmetics with double-double precision

Masahide Kashiwagi

1. Introduction

This library provides a computation with the double-double precision (corresponding to the quadruple precision) using Knuth's twosum algorithm and Dekker's twoproduct algorithm, as the software provided by David H. Bailey in High-Precision Software Directory .

This library (dd.hpp) has:

Moreover, this library is able to perform the interval arithmetic with the double-double precision with the following functions in rdd.hpp:

2. Files

(requiring conv-dd.hpp, fpu53.hpp, convert.hpp, and constants.hpp)
(requiring hwround.hpp, rdd-hwround.hpp, and rdd-nohwround.hpp)

3. How to use

See test-dd.hpp for the dd arithmetic. See test-idd.hpp for the interval arithmetic with the dd precision.

4. Algorithm

See section 1-4 in double-double arithmeric and double-double interval arithmeric (Version 2018/6/9, in Japanese) . Or see simple memos.

5. Interval arithmetic with dd precision

The interval arithmetic with the dd precision is performed by changing a rounding mode accurately. See section 5 in double-double arithmeric and double-double interval arithmeric (Version 2018/6/9, in Japanese) .

6. Compile Option

It may be possible to speed up with the compile option, see 20. Changing Rounding Mode and Compile Options for details.

7. Problem of 32bit mode of Intel's CPU

Intel's CPUs has the following two floating point units:

The computational precision in FPU is 80bit (total) and 64bit (significand), which is different from the IEEE754 standard of 64bit (total) and 53bit (significand). This causes deviance in dd-precision arithmetics.

On the other hand, SSE2 conforms the IEEE754 standard and Intel's CPUs in the 64bit mode employs not FPU but SSE2. Therefore, in the 64bit mode, such deviance does not occur. Even in the use of 32bit CPUs, "-msse2 -mfpmath=sse" provides a safe computation.

Fortunately, in the FPU Control Register, there is the flag that specifies a computational precision. If one can change this flag from 64bit to 53bit, the deviance due to FPU is prevented. Although this can be done with including fpu53.h in this library, some compiler may not support this change.

(Modified in version 0.4.54.) In order for dd operations to work properly, it must be fully compilant with IEEE754. For example, the 32bit environment of Intel CPU may not comply with IEEE754, so kv library cannot be compiled in such an environment. Actually, after #include <cfloat>, the macro FLT_EVAL_METHOD is checked, and if it is not 0, it is an error and cannot be compiled. If you really want to use kv in 32bit Intel CPU environment, you can use -msse2 -mfpmath=sse compile option to use SSE2, which will set the above macro to 0 and allow you to compile. We decided not to support compilers that do not define this macro, or CPUs that do not have SSE2, because they are too old.

8. Problem of frexp

By calling the frexp function like
    double x, y;
    int i;
    y = frexp(x, &i);
for input x, we can get y, i such that 0.5 ≤ |y| < 1、x=y×2i . The function is often used to get the exponent of x. For dd type, we implement frexp of dd which can be used as
    kv::dd x, y;
    int i;
    y = frexp(x, &i);

However, considering the input dd number such that the magnitude of high and low is much different, for example, x=1+2-1074, the output of frexp should be y=0.5+2-1075, i=1, but it is impossible because 2-1075 is not representable by double precision format. Namely, frexp for double is error-free, but, be careful to the fact that frexp for dd may include error in very special case. Note that the exponent i is always reliable.

9. Others

(Added on Nov. 12, 2014) We implemented a mathematical function (approximation) when using dd.hpp alone. Verified mathematical functions for interval<dd> can be used originally.

stringtodd, ddtostring contained in conv-dd.hpp performs mutual conversion between dd type and decimal character string. Since conv-dd.hpp can be used alone, if you are troubled with dd input/output, please use it.

We implemented infinite similar to IEEE 754 Std. The internal representation is (±∞, 0).