Re: [Boost-users] converting to float128 from double

Subject: Re: [Boost-users] converting to float128 from double
From: John Maddock (jz.maddock_at_[hidden])
Date: 2016-05-03 13:27:39

On 03/05/2016 17:35, Cooper, Bridgette R D wrote:
> Hi John,
>
> Thanks for the response.
>
> Here are some more snippets from my code.

I don't see any obvious errors, but in any case you must have extracted 
that from something else but it doesn't compile!

Not least you can't implicitly return from a float128 to a double in

     return(bcoef.at(0))

So if that's compiling for you, then I suspect your declaration of bcoef.

If I re-write and simplify in a slightly more C++ idiom how does it 
compare to your fortran code now:

double stieltjes_gamma(const std::vector<double> & e8, const 
std::vector<double> & g8, int printlevel)
{
    assert(e8.size() == g8.size());  // precondition?
    std::vector<float128> epoint(e8.begin(), e8.end()), 
gpoint(g8.begin(), g8.end()), bcoef(e8.size());

    bcoef[0] = 0.0Q;
    for(int i = 0; i < e8.size(); i++)
    {
       bcoef[0] += gpoint[i];
    }
    std::cout << 
std::setprecision(std::numeric_limits<float128>::max_digits10) << 
bcoef.at(0) << std::endl;
    return static_cast<double>(bcoef[0]);
}

And finally... note that summing at higher precision can not actually 
protect you from cancellation error (in any language) since the error is 
inherent in the input values.

HTH, John.

> #include <boost/multiprecision/float128.hpp>
> #include <boost/math/cstdfloat/cstdfloat_types.hpp>
> #include <boost/math/cstdfloat/cstdfloat_limits.hpp>
> #include <boost/math/cstdfloat/cstdfloat_cmath.hpp>
> #include <boost/math/cstdfloat/cstdfloat_iostream.hpp>
> #include <boost/multiprecision/detail/float_string_cvt.hpp>
> #include <boost/multiprecision/detail/generic_interconvert.hpp>
> extern "C"{
> #include <quadmath.h>
> }
> #include <iostream>
> #include <sstream>
> #include <vector>
> #include <iterator>
>
>          using namespace boost::multi precision;
>
>
>   double stieltjes_gamma(int num, const std::vector<double> & e8, const std::vector<double> & g8, int printlevel )
> {
>
>         std::vector<float128> epoint(num), gpoint(num), bcoef(num);
>
>
>
>         for (int i=0; i < num; i++)
>                  {
>                  epoint.at(i)=(float128(e8.at(i)));
>                  gpoint.at(i)=(float128(g8.at(i)));
>                  }
>
>   bcoef.at(0)=0.0Q;
>          for (int i = 0; i < num; i++)
>                  {
>                  bcoef.at(0)+=gpoint.at(i)
>                  }
>         std::cout << std::setprecision(std::numeric_limits<float128>::max_digits10)  << bcoef.at(0)<< std::endl;
>      return(bcoef.at(0))
> }
>
> When I compare this to fortran code that does the same I don't see the same value for bcoef.at(0).
>
> 0.0345217724606016853085032291117978629  <- from above c++ code
> 3.452175962187407162907246287483402E-0002   <- Fortran correctly utilising quad precision on same input vector
>
> The input vectors in this case are 7976 so errors in precision are really quickly apparrant. If I print  the gpoint vector, it gets filled with the same garbage for additional precision from double  as is for the fortran.
>
> Thanks,
> ________________________________________
> From: Boost-users <boost-users-bounces_at_[hidden]> on behalf of John Maddock <jz.maddock_at_[hidden]>
> Sent: Tuesday, May 3, 2016 5:06:47 PM
> To: boost-users_at_[hidden]
> Subject: Re: [Boost-users] converting to float128 from double
>
> On 03/05/2016 13:44, Cooper, Bridgette R D wrote:
>> Hi,
>>
>>
>> I have a function that takes as an argument a vector of doubles and
>> tries to convert it to a vector of float128.
>>
>>
>> When I accumulate the values in the float128 vector it looks like it
>> does double additions and only casts up at the last step.
>>
>>
>> The initial vector is defined as conststd::vector<double>
>>
>> The float128 vector is defined as std::vector<float128> epoint(num)
>>
>>
>> And I try to do the casting (inside a loop) via:
>> epoint.at(i)=(float128(e8.at(i)));
>>
>>
>> If I do the accumulation on the original vector with no casting into a
>> float128, I get the same as accumulating the vector that should be of
>> float128 type.
>>
>>
>> How should I be doing the casting?
>>
> Sorry but you'll have to post way more code than that to see where the
> mistake is.
>
> John.
>
>> Thanks,
>>
>>
>>
>> _______________________________________________
>> Boost-users mailing list
>> Boost-users_at_[hidden]
>> http://lists.boost.org/mailman/listinfo.cgi/boost-users
> _______________________________________________
> Boost-users mailing list
> Boost-users_at_[hidden]
> http://lists.boost.org/mailman/listinfo.cgi/boost-users
> _______________________________________________
> Boost-users mailing list
> Boost-users_at_[hidden]
> http://lists.boost.org/mailman/listinfo.cgi/boost-users
>