KOKKOS_FORCEINLINE_FUNCTION
static void
unrollIndex(IdxType &i, IdxType &j, IdxType &k,
const DimType dim0,
const DimType dim1,
const IterType iter) {
IdxType tmp;
unrollIndex(tmp, k, dim0*dim1, iter);
unrollIndex( i, j, dim0, tmp);
}
KOKKOS_INLINE_FUNCTION
void operator()(const size_type iter) const {
size_type cl, bf, pt;
size_type leftRank(_leftInput.rank()), rightRank(_rightInput.rank());
if (_hasField)
unrollIndex( cl, bf, pt,
_output.dimension(0),
_output.dimension(1),
_output.dimension(2),
iter );
else
unrollIndex( cl, pt,
_output.dimension(0),
_output.dimension(1),
iter);
auto result = ( _hasField ? Kokkos::subview(_output, cl, bf, pt) :
/**/ Kokkos::subview(_output, cl, pt));
const auto left = (_leftInput.dimension(1) == 1) ? Kokkos::subview(_leftInput, cl, 0, Kokkos::ALL(), Kokkos::ALL()) :
/**/ Kokkos::subview(_leftInput, cl, pt, Kokkos::ALL(), Kokkos::ALL());
const auto right = (rightRank == leftRank + int(_hasField)) ?
( _hasField ? Kokkos::subview(_rightInput, cl, bf, pt, Kokkos::ALL(), Kokkos::ALL()) :
/**/ Kokkos::subview(_rightInput, cl, pt, Kokkos::ALL(), Kokkos::ALL())) :
( _hasField ? Kokkos::subview(_rightInput, bf, pt, Kokkos::ALL(), Kokkos::ALL()) :
/**/ Kokkos::subview(_rightInput, pt, Kokkos::ALL(), Kokkos::ALL()));
const size_type iend = left.dimension(0);
const size_type jend = left.dimension(1);
value_type tmp(0);
for(size_type i = 0; i < iend; ++i)
for(size_type j = 0; j < jend; ++j)
tmp += left(i, j)*right(i, j);
result() = tmp;
}
KOKKOS_INLINE_FUNCTION
void operator()(const size_type iter) const {
size_type cell, pt;
unrollIndex( cell, pt,
_physPoints.dimension(0),
_physPoints.dimension(1),
iter );
/**/ auto phys = Kokkos::subdynrankview( _physPoints, cell, pt, Kokkos::ALL());
const auto dofs = Kokkos::subdynrankview( _worksetCells, cell, Kokkos::ALL(), Kokkos::ALL());
const auto valRank = _basisVals.rank();
const auto val = ( valRank == 2 ? Kokkos::subdynrankview( _basisVals, Kokkos::ALL(), pt) :
/**/ Kokkos::subdynrankview( _basisVals, cell, Kokkos::ALL(), pt));
const auto dim = phys.dimension(0);
const auto cardinality = val.dimension(0);
for (size_type i=0;i<dim;++i) {
phys(i) = 0;
for (size_type bf=0;bf<cardinality;++bf)
phys(i) += dofs(bf, i)*val(bf);
}
}