[Rd] double in summary.c : isum

Prof Brian Ripley ripley at stats.ox.ac.uk
Mon Mar 25 10:20:46 CET 2013


On 24/03/2013 15:01, Duncan Murdoch wrote:
> On 13-03-23 10:20 AM, Matthew Dowle wrote:
>> On 23.03.2013 12:01, Prof Brian Ripley wrote:
>>> On 20/03/2013 12:56, Matthew Dowle wrote:
>>>>
>>>> Hi,
>>>>
>>>> Please consider the following :
>>>>
>>>>> x = as.integer(2^30-1)
>>>> [1] 1073741823
>>>>> sum(c(rep(x, 10000000), rep(-x,9999999)))
>>>> [1] 1073741824
>>>>
>>>> Tested on 2.15.2 and a recent R-devel (r62132).
>>>>
>>>> I'm wondering if s in isum could be LDOUBLE instead of double, like
>>>> rsum, to fix this edge case?
>>>
>>> No, because there is no guarantee that LDOUBLE differs from double
>>> (and platform on which it does not).
>>
>> That's a reason for not using LDOUBLE at all isn't it? Yet src/main/*.c
>> has 19 lines using LDOUBLE e.g. arithmetic.c and cum.c as well as
>> summary.c.
>>
>> I'd assumed LDOUBLE was being used by R to benefit from long double (or
>> equivalent) on platforms that support it (which is all modern Unix, Mac
>> and Windows as far as I know). I do realise that the edge case wouldn't

Actually, you don't know.  Really only on almost all Intel ix86: most 
other current CPUs do not have it in hardware.  C99/C11 require long 
double, but does not require the accuracy that you are thinking of and 
it can be implemented in software.

Note that even on ix86 this is something that can be switched on or off 
in the CPU: last time I looked (years ago) it was off by default in 
Microsoft compilers.

All C99 requires is that long double is at least as precise as double. 
C11 recommends in §F.2

Recommended practice
2 The long double type should match an IEC 60559 extended format.

Notice the 'an': there are two such formats, and both are in use on R 
platforms.  But then some OS/compiler suppliers have never paid any heed 
to ISO standards.

>> be fixed on platforms where LDOUBLE is defined as double.
>
> I think the problem is that there are two opposing targets in R:  we
> want things to be as accurate as possible, and we want them to be
> consistent across platforms. Sometimes one goal wins, sometimes the
> other.  Inconsistencies across platforms give false positives in tests
> that tend to make us miss true bugs.  Some people think we should never
> use LDOUBLE because of that.  In other cases, the extra accuracy is so
> helpful that it's worth it.  So I think you'd need to argue that the
> case you found is something where the benefit outweighs the costs. Since
> almost all integer sums are done exactly with the current code, is it
> really worth introducing inconsistencies in the rare inexact cases?

But as I said lower down, a 64-bit integer accumulator would be helpful, 
C99/C11 requires one at least that large and it is implemented in 
hardware on all known R platforms.  So there is a way to do this pretty 
consistently across platforms.


>
> Duncan Murdoch
>
>
>>
>> What have I misunderstood?
>>
>>>
>>> Users really need to take responsibility for the numerical stability
>>> of calcuations they attempt.  Expecting to sum 20 million large
>>> numbers exactly is unrealistic.
>>
>> Trying to take responsibility, but you said no. Changing from double to
>> LDOUBLE would mean that something that wasn't realistic, was then
>> realistic (on platforms that support long double).
>>
>> And it would bring open source R into line with TERR, which gets the
>> answer right, on 64bit Windows at least. But I'm not sure I should be as
>> confident in TERR as I am in open source R because I can't see its
>> source code.
>>
>>
>>> There are cases where 64-bit integer accumulators would be
>>> beneficial, and this is one.  Unfortunately C11 does not require them
>>> but some optional moves in that direction are planned.
>>>
>>>>
>>>> https://svn.r-project.org/R/trunk/src/main/summary.c
>>>>
>>>> Thanks,
>>>> Matthew
>>>>
>>>> ______________________________________________
>>>> R-devel at r-project.org mailing list
>>>> https://stat.ethz.ch/mailman/listinfo/r-devel
>>
>> ______________________________________________
>> R-devel at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-devel
>>
>


-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595



More information about the R-devel mailing list