Some time ago, I discovered that in a thirty year old floating-point (FP) library the value of the mathematical expression 1 – x^{2} was determined by calculating 2·(1-x)-(1-x)^{2} for 0.5 ≤ x ≤ 1. At a first glance, this method appeared rather peculiar to me, but then I realised that the Lemma of Sterbenz (see, e.g. page 45 of this slide collection on FP calculations) guarantees that 1-x is calculated in FP without any round-off error in the specified range. The multiplication by 2 does not introduce errors as well. Yet, there are still two sources of errors: firstly, in case (1-x) has more than 26 significant bits, (1-x)^{2} is not exactly representable in FP, and secondly, finally taking the difference might cause a round-off error as well.

In my tests though, this old formula outperformed all other versions on average, and so I shared it with developer friends. Surprisingly, some of them responded with a statement like "The results produced by this formula do not differ from those obtained using (1-x)·(1+x)." In the end, the reason for this paradox could be identified: all my tests were executed on SSE registers, calculations on which strictly adhere to the double-precision floating-point format, whereas the calculations not showing any improvement had harnessed the intermediate extended double-precision format of the x87 stack. Modern x87-compatible FPUs provide both of these two very different kinds of intermediate FP number storage.

The SSE have been invented in favour of fast parallel data processing, whereas the x87 double-extended stack provides both 11 additional mantissa bits and 4 additional exponent bits to intermediate calculations in order to try to maintain an overall accuracy of 53 bits and to reduce the probability of intermediate under-/over-flows as well. The decision whether a piece of software uses the x87 or the SSE instruction sets has to be made up at compile time, because the generated byte code is different. Contrary, both the rounding-mode and the exception-handling properties may be influenced at runtime by changing the so called floating-point control word (FPUCW). I could not find an article on Wikipedia, so please see e.g. this collection of or this introduction to the FPUCW world. I also have come across these imho nice summaries that are written from C-programmers points of view: The pitfalls of verifying floating-point computations as well as SciLab is not naive. Unfortunately, values of the FPUCW having an identical meaning differ across platforms. For example, a value of the MS-FPUCW of hex 8001F corresponds to a GNU-GCC-FPUCW value of hex 37F … Fortunately, the ability to change the rounding mode during runtime enables programmers to harness interval arithmetic in order to compute reliable FP error estimates of their code for both the x87 and SSE instructions sets. The rounding mode can be chosen from the set UP, DOWN, CHOP, and TO NEAREST – TIES TO EVEN, the latter usually being the default setting. Only if the x87 instruction has been selected at compile time, the precision of the calculations may be changed during runtime as well.

The 11 additional significand bits of the x87 extDP format correspond to roughly 3 decimal figures, and thus are a valuable tool in order to minimise the effects of cancellation in intermediate calculations. The 4 additional exponent bits allow for an intermediate product of up to 16 huge (tiny) doubles without causing an over-(under-)flow. Yet, harnessing the extDP format bears one dirty little secret: when the result of an intermediate calculation is transferred from the x87 stack firstly into an extDP x87 register and finally into the DP main memory, the intermediate value is subject to a double-rounding procedure. In combination with the standard rounding mode "TO NEAREST – TIES TO EVEN" this can lead to a slightly wrong result in two situations: an intermediate calculation rounds up/down to an extDP value representing an exact DP tie that is again rounded up/down to a DP value. The following tables give an example:

decimal number | DP hex | comment | ||
---|---|---|---|---|

+18446744073709551616.0 | 43F0000000000000 | exact DP Num1 | ||

+6143.0 | 40B7FF0000000000 | exact DP Num2 | ||

+18446744073709557759.0 | Num1 + Num2, not representable in DP |
|||

+18446744073709555712.0 | 43F0000000000001 | DP_lo: nearest DP | ||

+18446744073709559808.0 | 43F0000000000002 | DP_hi: next larger DP | ||

-2047.0 | error of DP_lo | |||

+2049.0 | error of DP_hi |

When adding Num1 to Num2 on the x87 extDP stack, the hex representation of the final DP sum in memory will be 43F0000000000002 instead of the closest DP 43F0000000000001. This can be understood by analysing the binary representation of the numbers involved. In order to clearly show the rounding effects, some separators are used: a "V" indicates the border between any two DP half-bytes, each of which corresponds to a single character in the hex representation, an "X" shows where the DP bits end and the extDP bits begin, and finally a "Y" marks the end of the extDP bit sequence.

Num1 |
---|

+IOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO.O |

Num2 |

+IOIIIIIIIIIII.O |

Num1 + Num2 |

+IOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOIOIIIIIIIIIII.O |

same with separators |

+IVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOIXOIIIIIIIIIIYI.O |

1st step: round (up) to extDP |

+IVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOIXIOOOOOOOOOOYO.O |

2nd step: round DP tie (up) to even |

+IVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOOOVOOIOXOOOOOOOOOOOYO.O |

In this example, the addition creates a bit sequence, the mantissa of which exceeds the 64 bits of the extDP format. Thus, in a first step, the result is rounded to 64 bits. These extDP bits happen to represent a so-called tie (a number exactly half-way between the two adjacent DP numbers) of the DP format. In this example, the tie-breaking rule TO EVEN finally causes another rounding up procedure, so that the value stored in memory does not represent the DP number closest to the mathematical result of the summation, but the next larger one. Yet, it is still one of the two DP numbers bracketing the true value.

Only two different bit sequences among the 2^13 possible ones lead to this kind of error:

leadingbitsIXOIIIIIIIIIIYItrailingbits ++

leadingbitsOXIOOOOOOOOOOYOtrailingbits —

so that the odds ratio of this error should be expected to be close to 1/2^12 = 1/4096 ≈ 2.44E-4 . Of course, if the bit sequence of an intermediate result does not exceed the 64 bits of the extDP format, there will be only a single rounding step yielding a correct result. When analysing the results of 5.75E5 renormalisation steps of double-double numbers in VBA7, the sum of which is determined to exceed the 64 bit extDP limit, and that were introduced here at the end of this post, I found 1480 errors and thus an odds ratio of about 2.57E-4. Due to the relatively large sample size, I guess the deviation from the theoretical value is caused by non-randomness of the bit sequences involved.

Double-rounding is unavoidable when using the extDP registers of the x87 FPU, but it only slightly increases the representation error of about 0.5 ULPs by at maximum 0.05&percent;. Moreover, this only happens with an odds ratio of about 1/4096. Thus, this disadvantage of rarely occurring false double rounding is usually more than counterbalanced by the gain of 11 bits of intermediate precision in every calculation. However, there is a special kind of multi-precision software that cannot be run on the x87 stack due to the double rounding, i.e. all the double-double as well as quad-double software that use chained DP variables in order to store multi-precision values (see, e.g. this implementation). The double rounding on the x87 stack can lead to violations of the basic assumption of this kind of software, i.e., that the leading DP value always represents the DP value closest to the multi-precision value stored in a chain of doubles. Therefore, arbitrary precision tools being determined to use the x87 stack, for example Xnumbers, need to harness completely different algorithms for internal number storage. The numerical trick using a double-double representation of π/2 introduced in the aforementioned post works reliably, because the firstly calculated intermediate difference does not involve more than 64 significant bits.

The FP instruction set actually used by third-party numerical software is often unknown and needs to be determined by using test cases. In order to avoid problems with the digits chopping feature of most spreadsheet software, examples need to be created that use exact DP values corresponding to less than 14 significant decimal digits. The previously given example can be reformulated as sum of +1.8446744E+19 ≡ 43EFFFFFFDDAD230 and +73709557759 ≡ 4231296E97FF0000. In order to check the DP result for correctness without any debugging capabilities like those provided by Xnumbers for MS XL, e.g., both DP numbers that were summed up before now need to be successively subtracted from the sum, the bigger one at first, and the smaller one afterwards. It is important to assure that all intermediate differences are rounded to DP, which can most easily be achieved by assigning them to different DP variables, or different cells in a spreadsheet. Contrary to the first addition, the two final subtractions do not introduce any FP error, because they are subject to the lemma of Sterbenz. A final result of -2047 corresponds to calculations that have not been executed on the x87, whereas a value of +2049 indicates the opposite. Here are my PY3- and CAS-Maxima scripts:

Python 3 | CAS-Maxima | |
---|---|---|

num1 = +1.8446744E+19 | +1.8446744E+19; | |

num2 = +73709557759.0 | +73709557759.0; | |

sum12 = (num1+num2) | %o1+%o2; | |

sum12sub1 = (sum12-num1) | %-%o1; | |

sum12sub1sub2=(sum12sub1-num2) | %-%o2; | |

print(sum12sub1sub2) |

The following table summarises the results I found in some spreadsheet and scripting software running on MS WIN7x64prof. On Linux, all SW listed in the table that is available on that OS as well returns -2047.

spreadsheet SW | result | scripting SW | result | |||
---|---|---|---|---|---|---|

GNUmeric 1.12.9 | +2049 | PY2.7.3×64 | -2047 | |||

GScalc 11.4.2 | -2047 | PY3.2.3×64 | -2047 | |||

KSspread 9.1.0 | +2049 | R2.15.2×64 | -2047 | |||

LOcalc 4.1.4.2 | -2047 | R3.0.0×32 | +2049 | |||

LT SYM 3.0.1 | -2047 | R3.0.0×64 | -2047 | |||

MS XL14/VBA7 | +2049 | FreeMat 4.2 | -2047 | |||

SM PM 2012 | -2047 | Maxima x32 | +2049 |

As you can see from the table, the situation is completely messed up. It seems to me as if on the Win32 platform the x87 with its extDP format was used by default, whereas on the Win64 platform the SSE part of the FPU was involved. Of course, any software can override the default setting, and MS XL64 as well as VBA7 indeed do so. Generally speaking, specifying the FPUCW as well as the kind of FPU during compilation is a very good idea in case of numerical libraries that were developed and optimised for only one of the two scenarios. In case these two setting are omitted and the unknown defaults are used at runtime, small deviations are to be expected only in case the software was designed not to harness the extDP precision of the x87 stack. Otherwise, SSE lacking the additional 11 bits of precision will very likely cause things like polynomial approximations to special functions to fail at runtime. In particular, source code of libraries originally designed for the x87 must not be re-compiled for SSE without closely inspecting the results. Libraries that were built before the advent of the x87 can be a very valuable source of inspiration for changes in the code that are likely to become necessary.