opm-simulators/tests/test_block.cpp

368 lines
10 KiB
C++

/*
Copyright 2013 SINTEF ICT, Applied Mathematics.
Copyright 2016 IRIS AS
This file is part of the Open Porous Media project (OPM).
OPM is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
OPM is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with OPM. If not, see <http://www.gnu.org/licenses/>.
*/
#include <config.h>
#define BOOST_TEST_MODULE AutoDiffBlockTest
#include <opm/autodiff/AutoDiffBlock.hpp>
#include <boost/test/unit_test.hpp>
#include <Eigen/Eigen>
#include <Eigen/Sparse>
using namespace Opm;
namespace {
template <typename Scalar>
bool
operator ==(const Eigen::SparseMatrix<Scalar>& A,
const Eigen::SparseMatrix<Scalar>& B)
{
// Two SparseMatrices are equal if
// 0) They have the same ordering (enforced by equal types)
// 1) They have the same outer and inner dimensions
// 2) They have the same number of non-zero elements
// 3) They have the same sparsity structure
// 4) The non-zero elements are equal
// 1) Outer and inner dimensions
bool eq = (A.outerSize() == B.outerSize());
eq = eq && (A.innerSize() == B.innerSize());
// 2) Equal number of non-zero elements
eq = eq && (A.nonZeros() == B.nonZeros());
for (typename Eigen::SparseMatrix<Scalar>::Index
k0 = 0, kend = A.outerSize(); eq && (k0 < kend); ++k0) {
for (typename Eigen::SparseMatrix<Scalar>::InnerIterator
iA(A, k0), iB(B, k0); eq && (iA && iB); ++iA, ++iB) {
// 3) Sparsity structure
eq = (iA.row() == iB.row()) && (iA.col() == iB.col());
// 4) Equal non-zero elements
eq = eq && (iA.value() == iB.value());
}
}
return eq;
// Note: Investigate implementing this operator as
// return A.cwiseNotEqual(B).count() == 0;
}
bool operator==(const AutoDiffMatrix& lhs,
const AutoDiffMatrix& rhs)
{
Eigen::SparseMatrix<double> lhs_s, rhs_s;
lhs.toSparse(lhs_s);
rhs.toSparse(rhs_s);
return lhs_s == rhs_s;
}
void checkClose(const AutoDiffBlock<double>& lhs, const AutoDiffBlock<double>& rhs, double tolerance) {
BOOST_CHECK(lhs.value().isApprox(rhs.value(), tolerance));
auto lhs_d = lhs.derivative();
auto rhs_d = rhs.derivative();
Eigen::SparseMatrix<double> lhs_s, rhs_s;
//If lhs has no derivatives, make sure all rhs derivatives are zero
if (lhs_d.size() == 0) {
for (size_t i=0; i<rhs_d.size(); ++i) {
rhs.derivative()[i].toSparse(rhs_s);
BOOST_CHECK_EQUAL(rhs_s.nonZeros(), 0);
}
}
//If rhs has no derivatives, make sure all lhs derivatives are zero
else if (rhs_d.size() == 0) {
for (size_t i=0; i<lhs_d.size(); ++i) {
lhs.derivative()[i].toSparse(lhs_s);
BOOST_CHECK_EQUAL(lhs_s.nonZeros(), 0);
}
}
//Else, check that derivatives are close to each other
else {
BOOST_CHECK_EQUAL(lhs_d.size(), rhs_d.size());
for (size_t i=0; i<lhs_d.size(); ++i) {
lhs.derivative()[i].toSparse(lhs_s);
rhs.derivative()[i].toSparse(rhs_s);
BOOST_CHECK(lhs_s.isApprox(rhs_s, tolerance));
}
}
}
}
BOOST_AUTO_TEST_CASE(ConstantInitialisation)
{
typedef AutoDiffBlock<double> ADB;
ADB::V v(3);
v << 0.2, 1.2, 13.4;
ADB a = ADB::constant(v);
BOOST_REQUIRE(a.value().matrix() == v.matrix());
const std::vector<ADB::M>& J = a.derivative();
for (std::vector<ADB::M>::const_iterator
b = J.begin(), e = J.end(); b != e; ++b) {
BOOST_REQUIRE(b->nonZeros() == 0);
}
}
BOOST_AUTO_TEST_CASE(VariableInitialisation)
{
typedef AutoDiffBlock<double> ADB;
std::vector<int> blocksizes = { 3, 1, 2 };
ADB::V v(3);
v << 1.0, 2.2, 3.4;
enum { FirstVar = 0, SecondVar = 1, ThirdVar = 2 };
ADB x = ADB::variable(FirstVar, v, blocksizes);
BOOST_REQUIRE(x.value().matrix() == v.matrix());
const std::vector<ADB::M>& J = x.derivative();
BOOST_REQUIRE(J[0].nonZeros() == v.size());
const ADB::M& d = J[0];
for (int i=0; i<d.cols(); ++i) {
for (int j=0; j<d.rows(); ++j) {
if (i==j) {
BOOST_REQUIRE_EQUAL(d.coeff(j, i), 1.0);
}
else {
BOOST_REQUIRE_EQUAL(d.coeff(j, i), 0.0);
}
}
}
for (std::vector<ADB::M>::const_iterator
b = J.begin() + 1, e = J.end(); b != e; ++b) {
BOOST_REQUIRE(b->nonZeros() == 0);
}
}
BOOST_AUTO_TEST_CASE(FunctionInitialisation)
{
typedef AutoDiffBlock<double> ADB;
std::vector<int> blocksizes = { 3, 1, 2 };
std::vector<int>::size_type num_blocks = blocksizes.size();
enum { FirstVar = 0, SecondVar = 1, ThirdVar = 2 };
ADB::V v(3);
v << 1.0, 2.2, 3.4;
std::vector<ADB::M> jacs(num_blocks);
for (std::vector<int>::size_type j = 0; j < num_blocks; ++j) {
Eigen::SparseMatrix<double> sm(blocksizes[FirstVar], blocksizes[j]);
sm.insert(0,0) = -1.0;
jacs[j] = ADB::M(sm);
}
ADB::V v_copy(v);
std::vector<ADB::M> jacs_copy(jacs);
ADB f = ADB::function(std::move(v_copy), std::move(jacs_copy));
BOOST_REQUIRE(f.value().matrix() == v.matrix());
const std::vector<ADB::M>& J = f.derivative();
for (std::vector<ADB::M>::const_iterator
bf = J.begin(), ef = J.end(), bj = jacs.begin();
bf != ef; ++bf, ++bj) {
BOOST_CHECK(*bf == *bj);
}
}
BOOST_AUTO_TEST_CASE(Addition)
{
typedef AutoDiffBlock<double> ADB;
std::vector<int> blocksizes = { 3, 1, 2 };
ADB::V va(3);
va << 0.2, 1.2, 13.4;
ADB::V vx(3);
vx << 1.0, 2.2, 3.4;
enum { FirstVar = 0, SecondVar = 1, ThirdVar = 2 };
ADB a = ADB::constant(va, blocksizes);
ADB x = ADB::variable(FirstVar, vx, blocksizes);
ADB xpx = x + x;
BOOST_CHECK_EQUAL(xpx.value().cwiseNotEqual(2 * x.value()).count(), 0);
const std::vector<ADB::M>& J1x = x .derivative();
const std::vector<ADB::M>& J2x = xpx.derivative();
BOOST_CHECK_EQUAL(J1x.size(), J2x.size());
for (std::vector<ADB::M>::const_iterator
j1b = J1x.begin(), j1e = J1x.end(), j2b = J2x.begin();
j1b != j1e; ++j1b, ++j2b) {
BOOST_CHECK(*j2b == ADB::M((*j1b) * 2));
}
ADB::V r = 2*x.value() + a.value();
ADB xpxpa = x + x + a;
BOOST_CHECK_EQUAL(xpxpa.value().cwiseNotEqual(r).count(), 0);
const std::vector<ADB::M>& J3 = xpxpa.derivative();
for (std::vector<ADB::M>::const_iterator
j1b = J1x.begin(), j1e = J1x.end(), j3b = J3.begin();
j1b != j1e; ++j1b, ++j3b) {
BOOST_CHECK(*j3b == ADB::M((*j1b) * 2));
}
}
BOOST_AUTO_TEST_CASE(AssignAddSubtractOperators)
{
typedef AutoDiffBlock<double> ADB;
// Basic testing of += and -=.
ADB::V vx(3);
vx << 0.2, 1.2, 13.4;
ADB::V vy(3);
vy << 1.0, 2.2, 3.4;
std::vector<ADB::V> vals{ vx, vy };
std::vector<ADB> vars = ADB::variables(vals);
const ADB x = vars[0];
const ADB y = vars[1];
ADB z = x;
z += y;
ADB sum = x + y;
const double tolerance = 1e-14;
checkClose(z, sum, tolerance);
z -= y;
checkClose(z, x, tolerance);
// Testing the case when the left hand side has empty() jacobian.
ADB yconst = ADB::constant(vy);
z = yconst;
z -= x;
ADB diff = yconst - x;
checkClose(z, diff, tolerance);
z += x;
checkClose(z, yconst, tolerance);
}
BOOST_AUTO_TEST_CASE(Pow)
{
typedef AutoDiffBlock<double> ADB;
// Basic testing of derivatives
ADB::V vx(3);
vx << 0.2, 1.2, 13.4;
ADB::V vy(3);
vy << 2.0, 3.0, 0.5;
std::vector<ADB::V> vals{ vx, vy };
std::vector<ADB> vars = ADB::variables(vals);
const ADB x = vars[0];
const ADB y = vars[1];
const double tolerance = 1e-14;
// test exp = double
const ADB xx = x * x;
ADB xxpow2 = Opm::pow(x,2.0);
checkClose(xxpow2, xx, tolerance);
const ADB xy = x * y;
const ADB xxyy = xy * xy;
ADB xypow2 = Opm::pow(xy,2.0);
checkClose(xypow2, xxyy, tolerance);
const ADB xxx = x * x * x;
ADB xpow3 = Opm::pow(x,3.0);
checkClose(xpow3, xxx, tolerance);
ADB xpowhalf = Opm::pow(x,0.5);
ADB::V x_sqrt(3);
x_sqrt << 0.447214 , 1.095445 , 3.6606;
for (int i = 0 ; i < 3; ++i){
BOOST_CHECK_CLOSE(xpowhalf.value()[i], x_sqrt[i], 1e-4);
}
// test exp = ADB::V
ADB xpowyval = Opm::pow(x,y.value());
// each of the component of y is tested in the test above
// we compare with the results from the above tests.
ADB::V pick1(3);
pick1 << 1,0,0;
ADB::V pick2(3);
pick2 << 0,1,0;
ADB::V pick3(3);
pick3 << 0,0,1;
ADB compare = pick1 * xx + pick2 * xxx + pick3 * xpowhalf;
checkClose(xpowyval, compare, tolerance);
// test exponent = ADB::V and base = ADB
ADB xvalpowy = Opm::pow(x.value(),y);
// the value should be equal to xpowyval
// the first jacobian should be trivial
// the second jacobian is hand calculated
// log(0.2)*0.2^2.0, log(1.2) * 1.2^3.0, log(13.4) * 13.4^0.5
ADB::V jac2(3);
jac2 << -0.0643775165 , 0.315051650 , 9.50019208855;
for (int i = 0 ; i < 3; ++i){
BOOST_CHECK_CLOSE(xvalpowy.value()[i], xpowyval.value()[i], tolerance);
BOOST_CHECK_CLOSE(xvalpowy.derivative()[0].coeff(i,i), 0.0, tolerance);
BOOST_CHECK_CLOSE(xvalpowy.derivative()[1].coeff(i,i), jac2[i], 1e-4);
}
// test exp = ADB
ADB xpowy = Opm::pow(x,y);
// the first jacobian should be equal to the xpowyval
// the second jacobian should be equal to the xvalpowy
for (int i = 0 ; i < 3; ++i){
BOOST_CHECK_CLOSE(xpowy.value()[i], xpowyval.value()[i], tolerance);
BOOST_CHECK_CLOSE(xpowy.derivative()[0].coeff(i,i), xpowyval.derivative()[0].coeff(i,i), tolerance);
BOOST_CHECK_CLOSE(xpowy.derivative()[1].coeff(i,i), xvalpowy.derivative()[1].coeff(i,i), tolerance);
}
}