CollOfScalar EquelleRuntimeCPU::solveForUpdate(const CollOfScalar& residual) const
{
Eigen::SparseMatrix<double, Eigen::RowMajor> matr = residual.derivative()[0];
CollOfScalar::V du = CollOfScalar::V::Zero(residual.size());
Opm::time::StopWatch clock;
clock.start();
// solve(n, # nonzero values ("val"), ptr to col indices
// ("col_ind"), ptr to row locations in val array ("row_ind")
// (these two may be swapped, not sure about the naming convention
// here...), array of actual values ("val") (I guess... '*sa'...),
// rhs, solution)
Opm::LinearSolverInterface::LinearSolverReport rep
= linsolver_.solve(matr.rows(), matr.nonZeros(),
matr.outerIndexPtr(), matr.innerIndexPtr(), matr.valuePtr(),
residual.value().data(), du.data());
if (verbose_ > 2) {
std::cout << " solveForUpdate: Linear solver took: " << clock.secsSinceLast() << " seconds." << std::endl;
}
if (!rep.converged) {
OPM_THROW(std::runtime_error, "Linear solver convergence failure.");
}
return du;
}
CollOfScalar EquelleRuntimeCPU::singlePrimaryVariable(const CollOfScalar& initial_values)
{
std::vector<int> block_pattern;
block_pattern.push_back(initial_values.size());
// Syntax below is: CollOfScalar::variable(block index, initialized from, block structure)
return CollOfScalar::variable(0, initial_values.value(), block_pattern);
}
void EquelleRuntimeCPU::output(const String& tag, const CollOfScalar& vals)
{
if (output_to_file_) {
int count = -1;
auto it = outputcount_.find(tag);
if (it == outputcount_.end()) {
count = 0;
outputcount_[tag] = 1; // should contain the count to be used next time for same tag.
} else {
count = outputcount_[tag];
++outputcount_[tag];
}
std::ostringstream fname;
fname << tag << "-" << std::setw(5) << std::setfill('0') << count << ".output";
std::ofstream file(fname.str().c_str());
if (!file) {
OPM_THROW(std::runtime_error, "Failed to open " << fname.str());
}
file.precision(16);
std::copy(vals.value().data(), vals.value().data() + vals.size(), std::ostream_iterator<double>(file, "\n"));
} else {
std::cout << tag << " =\n";
for (int i = 0; i < vals.size(); ++i) {
std::cout << std::setw(15) << std::left << ( vals.value()[i] ) << " ";
}
std::cout << std::endl;
}
}
CollOfScalar EquelleRuntimeCUDA::operatorOn(const CollOfScalar& data_in,
const CollOfIndices<codim>& from_set,
const CollOfIndices<codim>& to_set) {
if ( data_in.size() != from_set.size()) {
OPM_THROW(std::logic_error, "data_in (size " << data_in.size() << ") and from_set (size " << from_set.size() << ") have to be of the same size in On function for CollOfScalar.");
}
//if ( to_set.size() > from_set.size() ) {
// OPM_THROW(std::runtime_error, "To_set (size " << to_set.size() << ") has to be a subset of from_set (size " << from_set.size() << ")");
//}
return dev_grid_.operatorOn(data_in, from_set, to_set);
}
CollOfScalar EquelleRuntimeCUDA::operatorExtend(const CollOfScalar& data_in,
const CollOfIndices<codim>& from_set,
const CollOfIndices<codim>& to_set) const {
if (data_in.size() != from_set.size() ) {
OPM_THROW(std::runtime_error, "data_in (size " << data_in.size() << ") and from_set (size " << from_set.size() << ") have to be of the same size in Extend function.");
}
if (from_set.size() > to_set.size() ) {
OPM_THROW(std::runtime_error, "From_set (size " << from_set.size() << ") has to be a subset of to_set (size " << to_set.size() << ")");
}
return dev_grid_.operatorExtend(data_in, from_set, to_set);
}
/// This function is not provided by AutoDiffBlock, so we must add it here.
inline CollOfScalar sqrt(const CollOfScalar& x)
{
// d(sqrt(x))/dy = 1/(2*sqrt(x)) * dx/dy
const auto& xjac = x.derivative();
if (xjac.empty()) {
return CollOfScalar(sqrt(x.value()));
}
const int num_blocks = xjac.size();
std::vector<CollOfScalar::M> jac(num_blocks);
const auto sqrt_x = sqrt(x.value());
const CollOfScalar::M one_over_two_sqrt_x((0.5/sqrt_x).matrix().asDiagonal());
for (int block = 0; block < num_blocks; ++block) {
jac[block] = one_over_two_sqrt_x * xjac[block];
}
return CollOfScalar::ADB::function(sqrt_x, jac);
}
CollOfScalar EquelleRuntimeCPU::divergence(const CollOfScalar& face_fluxes) const
{
if (face_fluxes.size() == ops_.internal_faces.size()) {
// This is actually a hack, the compiler should know to emit interiorDivergence()
// eventually, but as a temporary measure we do this.
return interiorDivergence(face_fluxes);
}
return ops_.fulldiv * face_fluxes;//.matrix();
}
int compare(CollOfScalar scal, double sol[],
int sol_size,
std::string test)
{
// Test size:
if ( scal.size() != sol_size ) {
std::cout << "Error in valsOnGrid.cpp - testing " << test << "\n";
std::cout << "\tThe collection is of wrong size!\n";
std::cout << "\tSize is " << scal.size() << " but should be " << sol_size << "\n";
return 1;
}
// Testing indices
std::vector<double> host = scal.copyToHost();
std::cout << "CollOfScalar " << test << " is the following:\n";
bool correct = true;
for (int i = 0; i < host.size(); ++i) {
std::cout << host[i] << " ";
if (i < sol_size) {
//if (host[i] != sol[i]) {
if ( fabs(host[i] - sol[i]) > 1000*std::numeric_limits<double>::epsilon() ) {
std::cout << "(<- " << sol[i] << ") ";
correct = false;
}
}
}
if (correct) {
std::cout << "\n\tThis is correct\n";
} else {
std::cout << "\n\tThis is wrong\n";
std::cout << "Error in valsOnGrid.cpp - testing " << test << "\n";
std::cout << "\tThe indices in the collection is wrong\n";
return 1;
}
return 0;
}
/// This operator is not provided by AutoDiffBlock, so we must add it here.
inline CollOfScalar operator/(const Scalar& s, const CollOfScalar& x)
{
return CollOfScalar::V::Constant(x.size(), s) / x;
}
/// This operator is not provided by AutoDiffBlock, so we must add it here.
inline CollOfScalar operator/(const CollOfScalar& x, const Scalar& s)
{
return x / CollOfScalar::V::Constant(x.size(), s);
}
Scalar EquelleRuntimeCPU::maxReduce(const CollOfScalar& x) const
{
return x.value().maxCoeff();
}
/// This operator is not provided by AutoDiffBlock, so we must add it here.
inline CollOfScalar operator-(const CollOfScalar& x)
{
return CollOfScalar::V::Zero(x.size()) - x;
}
/// This operator is not provided by AutoDiffBlock, so we must add it here.
inline CollOfBool operator==(const CollOfScalar& x, const Scalar& s)
{
return x.value() == s;
}
/// This operator is not provided by AutoDiffBlock, so we must add it here.
inline CollOfBool operator==(const CollOfScalar& x, const CollOfScalar& y)
{
return x.value() == y.value();
}
/// This operator is not provided by AutoDiffBlock, so we must add it here.
inline CollOfBool operator==(const Scalar& s, const CollOfScalar& x)
{
return s == x.value();
}
CollOfScalar EquelleRuntimeCUDA::newtonSolve(const ResidualFunctor& rescomp,
const CollOfScalar& u_initialguess)
{
Opm::time::StopWatch clock;
clock.start();
// Set up Newton loop.
// Define the primary variable
CollOfScalar u = CollOfScalar(u_initialguess, true);
if (verbose_ > 2) {
output("Initial u", u);
output(" newtonSolve: norm (initial u)", twoNorm(u));
}
CollOfScalar residual = rescomp(u);
if (verbose_ > 2) {
output("Initial residual", residual);
output(" newtonSolve: norm (initial residual)", twoNorm(residual));
}
int iter = 0;
// Debugging output not specified in Equelle.
if (verbose_ > 1) {
std::cout << " newtonSolve: iter = " << iter << " (max = " << max_iter_
<< "), norm(residual) = " << twoNorm(residual)
<< " (tol = " << abs_res_tol_ << ")" << std::endl;
}
CollOfScalar du;
// Execute newton loop until residual is small or we have used too many iterations.
while ( (twoNorm(residual) > abs_res_tol_) && (iter < max_iter_) ) {
if ( solver_.getSolver() == CPU ) {
du = serialSolveForUpdate(residual);
}
else {
// Solve linear equations for du, apply update.
du = solver_.solve(residual.derivative(),
residual.value(),
verbose_);
}
// du is a constant, hence, u is still a primary variable with an identity
// matrix as its derivative.
u = u - du;
// Recompute residual.
residual = rescomp(u);
if (verbose_ > 2) {
// Debugging output not specified in Equelle.
output("u", u);
output(" newtonSolve: norm(u)", twoNorm(u));
output("residual", residual);
output(" newtonSolve: norm(residual)", twoNorm(residual));
}
++iter;
// Debugging output not specified in Equelle.
if (verbose_ > 1) {
std::cout << " newtonSolve: iter = " << iter << " (max = " << max_iter_
<< "), norm(residual) = " << twoNorm(residual)
<< " (tol = " << abs_res_tol_ << ")" << std::endl;
}
}
if (verbose_ > 0) {
if (twoNorm(residual) > abs_res_tol_) {
std::cout << "Newton solver failed to converge in " << max_iter_ << " iterations" << std::endl;
} else {
std::cout << "Newton solver converged in " << iter << " iterations" << std::endl;
}
}
if (verbose_ > 1) {
std::cout << "Newton solver took: " << clock.secsSinceLast() << " seconds." << std::endl;
}
return CollOfScalar(u.value());
}
Scalar EquelleRuntimeCPU::sumReduce(const CollOfScalar& x) const
{
return x.value().sum();
}
Scalar EquelleRuntimeCPU::prodReduce(const CollOfScalar& x) const
{
return x.value().prod();
}
double EquelleRuntimeCPU::twoNorm(const CollOfScalar& vals) const
{
return vals.value().matrix().norm();
}