inline bool operator !=(const CG<Base>& left, const CG<Base>& right) {
if (left.isParameter() && right.isParameter()) {
return left.getValue() != right.getValue();
} else if (left.isParameter() || right.isParameter()) {
return true;
} else {
return left.getOperationNode() != right.getOperationNode();
}
}
int main(int argc, char* argv[]) {
init_paralution();
info_paralution();
LocalVector<double> x;
LocalVector<double> rhs;
LocalStencil<double> stencil(Laplace2D);
stencil.SetGrid(100); // 100x100
x.Allocate("x", stencil.get_nrow());
rhs.Allocate("rhs", stencil.get_nrow());
// Linear Solver
CG<LocalStencil<double>, LocalVector<double>, double > ls;
rhs.Ones();
x.Zeros();
ls.SetOperator(stencil);
ls.Build();
stencil.info();
double tick, tack;
tick = paralution_time();
ls.Solve(rhs, &x);
tack = paralution_time();
std::cout << "Solver execution:" << (tack-tick)/1000000 << " sec" << std::endl;
ls.Clear();
stop_paralution();
return 0;
}
extern "C" __declspec(dllexport) void CPUsolveCSRDouble(int *row_offset, int *col, double *val,
const int nnz, const int N, double *_rhs, double *_x) {
std::string name = "s";
LocalVector<double> x;
LocalVector<double> rhs;
LocalMatrix<double> mat;
x.Allocate(name, N);
x.Zeros();
rhs.Allocate(name, N);
mat.AllocateCSR(name, nnz, N, N);
mat.CopyFromCSR(row_offset, col, val);
rhs.CopyFromData(_rhs);
CG<LocalMatrix<double>, LocalVector<double>, double> ls;
Jacobi<LocalMatrix<double>, LocalVector<double>, double> p;
ls.SetOperator(mat);
ls.SetPreconditioner(p);
ls.Build();
ls.Solve(rhs, &x);
x.CopyToData(_x);
mat.Clear();
x.Clear();
rhs.Clear();
ls.Clear();
}
inline CG<Base>& CG<Base>::operator*=(const CG<Base> &right) {
if (isParameter() && right.isParameter()) {
*value_ *= *right.value_;
} else {
CodeHandler<Base>* handler;
if (isParameter()) {
if (isIdenticalZero()) {
return *this; // nothing to do (does not consider that right might be infinity)
} else if (isIdenticalOne()) {
*this = right;
return *this;
}
handler = right.getCodeHandler();
} else if (right.isParameter()) {
if (right.isIdenticalZero()) {
makeParameter(Base(0.0)); // does not consider that left might be infinity
return *this;
} else if (right.isIdenticalOne()) {
return *this; // nothing to do
}
handler = getCodeHandler();
} else {
CPPADCG_ASSERT_UNKNOWN(getCodeHandler() == right.getCodeHandler());
handler = getCodeHandler();
}
std::unique_ptr<Base> value;
if (isValueDefined() && right.isValueDefined()) {
value.reset(new Base(getValue() * right.getValue()));
}
makeVariable(*handler, new OperationNode<Base>(CGOpCode::Mul,{argument(), right.argument()}), value);
}
return *this;
}
extern "C" __declspec(dllexport) void solveCSRSingle(int *row_offset, int *col, float *val,
const int nnz, const int N, float *_rhs, float *_x) {
std::string name = "s";
LocalVector<float> x;
LocalVector<float> rhs;
LocalMatrix<float> mat;
x.Allocate(name, N);
x.Zeros();
rhs.Allocate(name, N);
mat.AllocateCSR(name, nnz, N, N);
mat.CopyFromCSR(row_offset, col, val);
rhs.CopyFromData(_rhs);
// mat.Check();
/* rhs.SetDataPtr(&_rhs, name, N);
x.SetDataPtr(&_x, name, N);
mat.SetDataPtrCSR(&row_offset, &col, &val, name, nnz, N, N);
*/
mat.MoveToAccelerator();
x.MoveToAccelerator();
rhs.MoveToAccelerator();
CG<LocalMatrix<float>, LocalVector<float>, float> ls;
MultiColoredILU<LocalMatrix<float>, LocalVector<float>, float> p;
ls.SetOperator(mat);
ls.SetPreconditioner(p);
ls.Build();
ls.Solve(rhs, &x);
mat.MoveToHost();
x.MoveToHost();
rhs.MoveToHost();
/*
mat.LeaveDataPtrCSR(&row_offset, &col, &val);
rhs.LeaveDataPtr(&_rhs);
x.LeaveDataPtr(&_x);
*/
x.CopyToData(_x);
mat.Clear();
x.Clear();
rhs.Clear();
ls.Clear();
}
inline CG<Base>& CG<Base>::operator+=(const CG<Base> &right) {
if (isParameter() && right.isParameter()) {
*value_ += *right.value_;
} else {
CodeHandler<Base>* handler;
if (isParameter()) {
if (isIdenticalZero()) {
*this = right;
return *this;
}
handler = right.getCodeHandler();
} else if (right.isParameter()) {
if (right.isIdenticalZero()) {
return *this; // nothing to do
}
handler = getCodeHandler();
} else {
CPPADCG_ASSERT_UNKNOWN(getCodeHandler() == right.getCodeHandler());
handler = getCodeHandler();
}
std::unique_ptr<Base> value;
if (isValueDefined() && right.isValueDefined()) {
value.reset(new Base(getValue() + right.getValue()));
}
makeVariable(*handler, new OperationNode<Base>(CGOpCode::Add,{argument(), right.argument()}), value);
}
return *this;
}
int main(int argc, char* argv[]) {
if (argc == 1) {
std::cerr << argv[0] << " <matrix> [Num threads]" << std::endl;
exit(1);
}
init_paralution();
if (argc > 2) {
set_omp_threads_paralution(atoi(argv[2]));
}
info_paralution();
LocalVector<double> x;
LocalVector<double> rhs;
LocalMatrix<double> mat;
mat.ReadFileMTX(std::string(argv[1]));
// Compute and apply (R)CMK ordering
LocalVector<int> cmk;
// mat.CMK(&cmk);
mat.RCMK(&cmk);
mat.Permute(cmk);
mat.MoveToAccelerator();
x.MoveToAccelerator();
rhs.MoveToAccelerator();
x.Allocate("x", mat.get_nrow());
rhs.Allocate("rhs", mat.get_nrow());
// Linear Solver
CG<LocalMatrix<double>, LocalVector<double>, double > ls;
// Preconditioner
ILU<LocalMatrix<double>, LocalVector<double>, double > p;
double tick, tack;
rhs.Ones();
x.Zeros();
ls.SetOperator(mat);
ls.SetPreconditioner(p);
ls.Build();
mat.info();
tick = paralution_time();
ls.Solve(rhs, &x);
tack = paralution_time();
std::cout << "Solver execution:" << (tack-tick)/1000000 << " sec" << std::endl;
// Revert CMK ordering on solution vector
x.PermuteBackward(cmk);
stop_paralution();
return 0;
}
void AMG<OperatorType, VectorType, ValueType>::Build(void) {
if (this->build_ == true)
this->Clear();
assert(this->build_ == false);
this->BuildHierarchy();
this->build_ = true;
this->d_level_ = new VectorType*[this->levels_];
this->r_level_ = new VectorType*[this->levels_];
this->t_level_ = new VectorType*[this->levels_];
this->s_level_ = new VectorType*[this->levels_];
for (int i=0; i<this->levels_; ++i) {
this->r_level_[i] = new VectorType;
this->t_level_[i] = new VectorType;
this->s_level_[i] = new VectorType;
this->r_level_[i]->CloneBackend(*this->op_);
this->t_level_[i]->CloneBackend(*this->op_);
this->s_level_[i]->CloneBackend(*this->op_);
if (i > 0) {
this->d_level_[i] = new VectorType;
this->d_level_[i]->CloneBackend(*this->op_);
}
}
// Allocate temporary vectors for cycles
for (int level=0; level<this->levels_; ++level) {
if (level > 0) {
this->d_level_[level]->Allocate("defect correction", this->op_level_[level-1]->get_nrow());
this->r_level_[level]->Allocate("residual", this->op_level_[level-1]->get_nrow());
this->t_level_[level]->Allocate("temporary", this->op_level_[level-1]->get_nrow());
this->s_level_[level]->Allocate("temporary", this->op_level_[level-1]->get_nrow());
} else {
this->r_level_[level]->Allocate("residual", this->op_->get_nrow());
this->t_level_[level]->Allocate("temporary", this->op_->get_nrow());
this->s_level_[level]->Allocate("temporary", this->op_->get_nrow());
}
}
// Setup and build smoothers
if (this->set_sm_ == false) {
// Smoother for each level
FixedPoint<OperatorType, VectorType, ValueType > **sm = NULL;
sm = new FixedPoint<OperatorType, VectorType, ValueType >* [this->levels_-1];
MultiColoredGS<OperatorType, VectorType, ValueType > **gs = NULL;
gs = new MultiColoredGS<OperatorType, VectorType, ValueType >* [this->levels_-1];
this->smoother_level_ = new IterativeLinearSolver<OperatorType, VectorType, ValueType>*[this->levels_-1];
this->sm_default_ = new Solver<OperatorType, VectorType, ValueType>*[this->levels_-1];
for (int i=0; i<this->levels_-1; ++i) {
sm[i] = new FixedPoint<OperatorType, VectorType, ValueType >;
gs[i] = new MultiColoredGS<OperatorType, VectorType, ValueType >;
gs[i]->SetPrecondMatrixFormat(this->sm_format_);
// relxation
sm[i]->SetRelaxation(1.3);
sm[i]->SetPreconditioner(*gs[i]);
// be quite
sm[i]->Verbose(0);
this->smoother_level_[i] = sm[i];
// pass pointer to class so we can free it when clearing
this->sm_default_[i] = gs[i];
}
delete[] sm;
delete[] gs;
}
for (int i=0; i<this->levels_-1; ++i) {
if (i > 0)
this->smoother_level_[i]->SetOperator(*this->op_level_[i-1]);
else
this->smoother_level_[i]->SetOperator(*this->op_);
this->smoother_level_[i]->Build();
}
// Setup and build coarse grid solver
if (this->set_s_ == false) {
// Coarse Grid Solver
CG<OperatorType, VectorType, ValueType > *cgs
= new CG<OperatorType, VectorType, ValueType >;
// be quite
cgs->Verbose(0);
this->solver_coarse_ = cgs;
}
this->solver_coarse_->SetOperator(*this->op_level_[this->levels_-2]);
this->solver_coarse_->Build();
// Convert operator to op_format
if (this->op_format_ != CSR)
for (int i=0; i<this->levels_-1;++i)
this->op_level_[i]->ConvertTo(this->op_format_);
}
int main(int argc, char* argv[]) {
if (argc == 1) {
std::cerr << argv[0] << " <matrix> [Num threads]" << std::endl;
exit(1);
}
init_paralution();
if (argc > 2) {
set_omp_threads_paralution(atoi(argv[2]));
}
info_paralution();
LocalVector<double> b, b_old, *b_k, *b_k1, *b_tmp;
LocalMatrix<double> mat;
mat.ReadFileMTX(std::string(argv[1]));
// Gershgorin spectrum approximation
double glambda_min, glambda_max;
// Power method spectrum approximation
double plambda_min, plambda_max;
// Maximum number of iteration for the power method
int iter_max = 10000;
double tick, tack;
// Gershgorin approximation of the eigenvalues
mat.Gershgorin(glambda_min, glambda_max);
std::cout << "Gershgorin : Lambda min = " << glambda_min
<< "; Lambda max = " << glambda_max << std::endl;
mat.MoveToAccelerator();
b.MoveToAccelerator();
b_old.MoveToAccelerator();
b.Allocate("b_k+1", mat.get_nrow());
b_k1 = &b;
b_old.Allocate("b_k", mat.get_nrow());
b_k = &b_old;
b_k->Ones();
mat.info();
tick = paralution_time();
// compute lambda max
for (int i=0; i<=iter_max; ++i) {
mat.Apply(*b_k, b_k1);
// std::cout << b_k1->Dot(*b_k) << std::endl;
b_k1->Scale(double(1.0)/b_k1->Norm());
b_tmp = b_k1;
b_k1 = b_k;
b_k = b_tmp;
}
// get lambda max (Rayleigh quotient)
mat.Apply(*b_k, b_k1);
plambda_max = b_k1->Dot(*b_k) ;
tack = paralution_time();
std::cout << "Power method (lambda max) execution:" << (tack-tick)/1000000 << " sec" << std::endl;
mat.AddScalarDiagonal(double(-1.0)*plambda_max);
b_k->Ones();
tick = paralution_time();
// compute lambda min
for (int i=0; i<=iter_max; ++i) {
mat.Apply(*b_k, b_k1);
// std::cout << b_k1->Dot(*b_k) + plambda_max << std::endl;
b_k1->Scale(double(1.0)/b_k1->Norm());
b_tmp = b_k1;
b_k1 = b_k;
b_k = b_tmp;
}
// get lambda min (Rayleigh quotient)
mat.Apply(*b_k, b_k1);
plambda_min = (b_k1->Dot(*b_k) + plambda_max);
// back to the original matrix
mat.AddScalarDiagonal(plambda_max);
tack = paralution_time();
std::cout << "Power method (lambda min) execution:" << (tack-tick)/1000000 << " sec" << std::endl;
std::cout << "Power method Lambda min = " << plambda_min
<< "; Lambda max = " << plambda_max
<< "; iter=2x" << iter_max << std::endl;
LocalVector<double> x;
LocalVector<double> rhs;
x.CloneBackend(mat);
rhs.CloneBackend(mat);
x.Allocate("x", mat.get_nrow());
rhs.Allocate("rhs", mat.get_nrow());
// Chebyshev iteration
Chebyshev<LocalMatrix<double>, LocalVector<double>, double > ls;
rhs.Ones();
x.Zeros();
ls.SetOperator(mat);
ls.Set(plambda_min, plambda_max);
ls.Build();
tick = paralution_time();
ls.Solve(rhs, &x);
tack = paralution_time();
std::cout << "Solver execution:" << (tack-tick)/1000000 << " sec" << std::endl;
// PCG + Chebyshev polynomial
CG<LocalMatrix<double>, LocalVector<double>, double > cg;
AIChebyshev<LocalMatrix<double>, LocalVector<double>, double > p;
// damping factor
plambda_min = plambda_max / 7;
p.Set(3, plambda_min, plambda_max);
rhs.Ones();
x.Zeros();
cg.SetOperator(mat);
cg.SetPreconditioner(p);
cg.Build();
tick = paralution_time();
cg.Solve(rhs, &x);
tack = paralution_time();
std::cout << "Solver execution:" << (tack-tick)/1000000 << " sec" << std::endl;
stop_paralution();
return 0;
}
