Feature single precision solve type

This commit is contained in:
Mattijs Janssens
2019-11-19 11:10:07 +00:00
committed by Andrew Heather
parent f2eb3e1cee
commit 2d080ff331
12 changed files with 293 additions and 38 deletions

View File

@ -24,6 +24,10 @@ License
Description
Simple field tests
Test use of Kahan/Neumaier to extend precision for when running SPDP
mode. Conclusion is that it is easier/quicker to run these summation
loops as double precision (i.e. solveScalar).
\*---------------------------------------------------------------------------*/
#include "primitiveFields.H"
@ -32,6 +36,116 @@ Description
using namespace Foam;
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * //
template<class Type, class CombineOp, class ResultType>
void sumNeumaier
(
const UList<Type>& vals,
const CombineOp& cop,
ResultType& result
)
{
// Neumaier version of Kahan
ResultType sum = Zero;
ResultType c = Zero;
for (const Type& vali : vals)
{
ResultType val;
cop(val, vali);
const ResultType t = sum + val;
for
(
direction cmpt = 0;
cmpt < pTraits<ResultType>::nComponents;
cmpt++
)
{
if (mag(sum[cmpt]) >= mag(val[cmpt]))
{
// If sum is bigger, low-order digits of input[i] are lost.
c[cmpt] += (sum[cmpt] - t[cmpt]) + val[cmpt];
}
else
{
// Else low-order digits of sum are lost.
c[cmpt] += (val[cmpt] - t[cmpt]) + sum[cmpt];
}
}
sum = t;
}
result = sum + c;
}
template<class CombineOp, class ResultType>
void sumNeumaier
(
const UList<scalar>& vals,
const CombineOp& cop,
ResultType& result
)
{
// Neumaier version of Kahan
ResultType sum = Zero;
ResultType c = Zero;
for (const scalar vali : vals)
{
ResultType val;
cop(val, vali);
const ResultType t = sum + val;
if (mag(sum) >= mag(val))
{
// If sum is bigger, low-order digits of input[i] are lost.
c += (sum - t) + val;
}
else
{
// Else low-order digits of sum are lost.
c += (val - t) + sum;
}
sum = t;
}
result = sum + c;
}
template<class Type>
Type mySum(const UList<Type>& f)
{
typedef typename Foam::typeOfSolve<Type>::type solveType;
solveType Sum = Zero;
if (f.size())
{
sumNeumaier(f, eqOp<solveType>(), Sum);
}
return Type(Sum);
}
//- The sumSqr always adds only positive numbers. Here there can never be any
// cancellation of truncation errors.
template<class Type>
typename outerProduct1<Type>::type mySumSqr(const UList<Type>& f)
{
typedef typename outerProduct1<solveScalar>::type prodType;
prodType result = Zero;
if (f.size())
{
sumNeumaier(f, eqSqrOp<prodType>(), result);
}
return result;
}
// Main program:
int main(int argc, char *argv[])
@ -63,6 +177,71 @@ int main(int argc, char *argv[])
Info<< "negated: " << lfield << nl;
// Summation (compile in SPDP)
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~
Pout.precision(16);
Sout.precision(16);
const scalar SMALLS(1e-6);
const vector SMALLV(SMALLS, SMALLS, SMALLS);
const scalar GREATS(1e6);
const vector GREATV(GREATS, GREATS, GREATS);
// scalarField summation
{
scalarField sfield(10, SMALLS);
sfield[8] = GREATS;
sfield[9] = -sfield[8];
Info<< "scalarField:" << sfield.size() << nl
<< " sum :" << sum(sfield) << nl
<< " corrected:" << mySum(sfield) << endl;
}
// vectorField summation
{
vectorField vfield(10, SMALLV);
vfield[8] = GREATV;
vfield[9] = -vfield[8];
Info<< "vectorField:" << vfield.size() << nl
<< " sum :" << sum(vfield) << nl
<< " corrected:" << mySum(vfield) << endl;
}
// sphericalTensorField summation
{
sphericalTensorField tfield(10, SMALLS);
tfield[8] = GREATS;
tfield[9] = -tfield[8];
Info<< "sphericalTensorField:" << tfield.size() << nl
<< " sum :" << sum(tfield) << nl
<< " corrected:" << mySum(tfield) << endl;
}
// symmTensorField summation
{
symmTensorField tfield(10, SMALLS*symmTensor::I);
tfield[8] = GREATS*symmTensor::I;
tfield[9] = -tfield[8];
Info<< "symmTensorField:" << tfield.size() << nl
<< " sum :" << sum(tfield) << nl
<< " corrected:" << mySum(tfield) << endl;
}
// tensorField summation
{
tensorField tfield(10, SMALLS*tensor::I);
tfield[8] = GREATS*tensor::I;
tfield[9] = -tfield[8];
Info<< "tensorField:" << tfield.size() << nl
<< " sum :" << sum(tfield) << nl
<< " corrected:" << mySum(tfield) << endl;
}
return 0;
}