//=============================================================================
int Epetra_MsrMatrix::Multiply(bool TransA,
const Epetra_MultiVector& X,
Epetra_MultiVector& Y) const
{
(void)TransA;
int NumVectors = X.NumVectors();
if (NumVectors!=Y.NumVectors()) EPETRA_CHK_ERR(-1); // X and Y must have same number of vectors
double ** xptrs;
double ** yptrs;
X.ExtractView(&xptrs);
Y.ExtractView(&yptrs);
if (RowMatrixImporter()!=0) {
if (ImportVector_!=0) {
if (ImportVector_->NumVectors()!=NumVectors) { delete ImportVector_; ImportVector_= 0;}
}
if (ImportVector_==0) ImportVector_ = new Epetra_MultiVector(RowMatrixColMap(),NumVectors); // Create import vector if needed
ImportVector_->Import(X, *RowMatrixImporter(), Insert);
ImportVector_->ExtractView(&xptrs);
}
for (int i=0; i<NumVectors; i++)
Amat_->matvec(xptrs[i], yptrs[i], Amat_, proc_config_);
double flops = NumGlobalNonzeros();
flops *= 2.0;
flops *= (double) NumVectors;
UpdateFlops(flops);
return(0);
}
int Epetra_PETScAIJMatrix::Multiply(bool TransA,
const Epetra_MultiVector& X,
Epetra_MultiVector& Y) const
{
(void)TransA;
int NumVectors = X.NumVectors();
if (NumVectors!=Y.NumVectors()) EPETRA_CHK_ERR(-1); // X and Y must have same number of vectors
double ** xptrs;
double ** yptrs;
X.ExtractView(&xptrs);
Y.ExtractView(&yptrs);
if (RowMatrixImporter()!=0) {
if (ImportVector_!=0) {
if (ImportVector_->NumVectors()!=NumVectors) { delete ImportVector_; ImportVector_= 0;}
}
if (ImportVector_==0) ImportVector_ = new Epetra_MultiVector(RowMatrixColMap(),NumVectors);
ImportVector_->Import(X, *RowMatrixImporter(), Insert);
ImportVector_->ExtractView(&xptrs);
}
double *vals=0;
int length;
Vec petscX, petscY;
int ierr;
for (int i=0; i<NumVectors; i++) {
# ifdef HAVE_MPI
ierr=VecCreateMPIWithArray(Comm_->Comm(),X.MyLength(),X.GlobalLength(),xptrs[i],&petscX); CHKERRQ(ierr);
ierr=VecCreateMPIWithArray(Comm_->Comm(),Y.MyLength(),Y.GlobalLength(),yptrs[i],&petscY); CHKERRQ(ierr);
# else //FIXME untested
ierr=VecCreateSeqWithArray(Comm_->Comm(),X.MyLength(),X.GlobalLength(),xptrs[i],&petscX); CHKERRQ(ierr);
ierr=VecCreateSeqWithArray(Comm_->Comm(),Y.MyLength(),Y.GlobalLength(),yptrs[i],&petscY); CHKERRQ(ierr);
# endif
ierr = MatMult(Amat_,petscX,petscY);CHKERRQ(ierr);
ierr = VecGetArray(petscY,&vals);CHKERRQ(ierr);
ierr = VecGetLocalSize(petscY,&length);CHKERRQ(ierr);
for (int j=0; j<length; j++) yptrs[i][j] = vals[j];
ierr = VecRestoreArray(petscY,&vals);CHKERRQ(ierr);
}
VecDestroy(petscX); VecDestroy(petscY);
double flops = NumGlobalNonzeros();
flops *= 2.0;
flops *= (double) NumVectors;
UpdateFlops(flops);
return(0);
} //Multiply()
Epetra_OskiMultiVector::Epetra_OskiMultiVector(const Epetra_MultiVector& Source)
: Epetra_MultiVector(Source),
Epetra_View_(&Source),
Copy_Created_(false) {
double* A;
double** Aptr;
int LDA;
int* LDAptr;
LDAptr = new int[1];
Aptr = new double*[1];
if(Source.ConstantStride() || (Source.NumVectors() == 1)) {
if(Source.ExtractView(Aptr, LDAptr))
std::cerr << "Extract view failed\n";
else
Oski_View_ = oski_CreateMultiVecView(*Aptr, Source.MyLength(), Source.NumVectors(), LAYOUT_COLMAJ, *LDAptr);
}
else {
Copy_Created_ = true;
LDA = Source.MyLength();
A = new double[LDA*Source.NumVectors()];
if(Source.ExtractCopy(A, LDA))
std::cerr << "Extract copy failed\n";
else
Oski_View_ = oski_CreateMultiVecView(A, Source.MyLength(), Source.NumVectors(), LAYOUT_COLMAJ, LDA);
}
delete [] LDAptr;
delete [] Aptr;
}
//EpetraMultiVector_To_TpetraMultiVector: copies Epetra_MultiVector to non-const Tpetra_MultiVector
Teuchos::RCP<Tpetra_MultiVector> Petra::EpetraMultiVector_To_TpetraMultiVector(const Epetra_MultiVector& epetraMV_,
const Teuchos::RCP<const Teuchos::Comm<int> >& commT_)
{
//get map from epetraMV_ and convert to Tpetra::Map
auto mapT = EpetraMap_To_TpetraMap(epetraMV_.Map(), commT_);
//copy values from epetraMV_
int numVectors = epetraMV_.NumVectors();
int Length;
ST *values;
epetraMV_.ExtractView(&values, &Length);
Teuchos::ArrayView<ST> valuesAV = Teuchos::arrayView(values, Length*numVectors);
//create Tpetra_MultiVector copy of epetraMV_
Teuchos::RCP<Tpetra_MultiVector> tpetraMV_ = Teuchos::rcp(new Tpetra_MultiVector(mapT, valuesAV, Length, numVectors));
return tpetraMV_;
}
// Convert a Epetra_MultiVector with assumed block structure dictated by the
// vector space into a Thyra::MultiVectorBase object.
// const Teuchos::RCP<const Thyra::MultiVectorBase<double> > blockEpetraToThyra(const Epetra_MultiVector & e,const Teuchos::RCP<const Thyra::VectorSpaceBase<double> > & vs)
void blockEpetraToThyra(const Epetra_MultiVector & epetraX,const Teuchos::Ptr<Thyra::MultiVectorBase<double> > & thyraX)
{
TEUCHOS_ASSERT(thyraX->range()->dim()==epetraX.GlobalLength());
// extract local information from the Epetra_MultiVector
int leadingDim=0,numVectors=0,localDim=0;
double * epetraData=0;
epetraX.ExtractView(&epetraData,&leadingDim);
numVectors = epetraX.NumVectors();
blockEpetraToThyra(numVectors,epetraData,leadingDim,thyraX.ptr(),localDim);
TEUCHOS_ASSERT(localDim==epetraX.MyLength());
}
// Convert a Thyra::MultiVectorBase object to a Epetra_MultiVector object with
// the map defined by the Epetra_Map.
// const Teuchos::RCP<const Epetra_MultiVector>
// blockThyraToEpetra(const Teuchos::RCP<const Thyra::MultiVectorBase<double> > & tX,const RCP<const Epetra_Map> & map)
void blockThyraToEpetra(const Teuchos::RCP<const Thyra::MultiVectorBase<double> > & thyraX,Epetra_MultiVector & epetraX)
{
// build an Epetra_MultiVector object
int numVectors = thyraX->domain()->dim();
// make sure the number of vectors are the same
TEUCHOS_ASSERT(numVectors==epetraX.NumVectors());
TEUCHOS_ASSERT(thyraX->range()->dim()==epetraX.GlobalLength());
// extract local information from the Epetra_MultiVector
int leadingDim=0,localDim=0;
double * epetraData=0;
epetraX.ExtractView(&epetraData,&leadingDim);
// perform recursive copy
blockThyraToEpetra(numVectors,epetraData,leadingDim,thyraX,localDim);
// sanity check
TEUCHOS_ASSERT(localDim==epetraX.Map().NumMyElements());
}
//==============================================================================
int Ifpack_PointRelaxation::
ApplyInverseGS_RowMatrix(const Epetra_MultiVector& X, Epetra_MultiVector& Y) const
{
int NumVectors = X.NumVectors();
int Length = Matrix().MaxNumEntries();
std::vector<int> Indices(Length);
std::vector<double> Values(Length);
Teuchos::RefCountPtr< Epetra_MultiVector > Y2;
if (IsParallel_)
Y2 = Teuchos::rcp( new Epetra_MultiVector(Importer_->TargetMap(), NumVectors) );
else
Y2 = Teuchos::rcp( &Y, false );
// extract views (for nicer and faster code)
double** y_ptr, ** y2_ptr, ** x_ptr, *d_ptr;
X.ExtractView(&x_ptr);
Y.ExtractView(&y_ptr);
Y2->ExtractView(&y2_ptr);
Diagonal_->ExtractView(&d_ptr);
for (int j = 0; j < NumSweeps_ ; j++) {
// data exchange is here, once per sweep
if (IsParallel_)
IFPACK_CHK_ERR(Y2->Import(Y,*Importer_,Insert));
// FIXME: do I really need this code below?
if (NumVectors == 1) {
double* y0_ptr = y_ptr[0];
double* y20_ptr = y2_ptr[0];
double* x0_ptr = x_ptr[0];
if(!DoBackwardGS_){
/* Forward Mode */
for (int ii = 0 ; ii < NumLocalSmoothingIndices_ ; ++ii) {
int i = (!LocalSmoothingIndices_)? ii : LocalSmoothingIndices_[ii];
int NumEntries;
int col;
IFPACK_CHK_ERR(Matrix_->ExtractMyRowCopy(i, Length,NumEntries,
&Values[0], &Indices[0]));
double dtemp = 0.0;
for (int k = 0 ; k < NumEntries ; ++k) {
col = Indices[k];
dtemp += Values[k] * y20_ptr[col];
}
y20_ptr[i] += DampingFactor_ * d_ptr[i] * (x0_ptr[i] - dtemp);
}
}
else {
/* Backward Mode */
for (int ii = NumLocalSmoothingIndices_ - 1 ; ii > -1 ; --ii) {
int i = (!LocalSmoothingIndices_)? ii : LocalSmoothingIndices_[ii];
int NumEntries;
int col;
(void) col; // Forestall compiler warning for unused variable.
IFPACK_CHK_ERR(Matrix_->ExtractMyRowCopy(i, Length,NumEntries,
&Values[0], &Indices[0]));
double dtemp = 0.0;
for (int k = 0 ; k < NumEntries ; ++k) {
col = Indices[k];
dtemp += Values[k] * y20_ptr[i];
}
y20_ptr[i] += DampingFactor_ * d_ptr[i] * (x0_ptr[i] - dtemp);
}
}
// using Export() sounded quite expensive
if (IsParallel_)
for (int i = 0 ; i < NumMyRows_ ; ++i)
y0_ptr[i] = y20_ptr[i];
}
else {
if(!DoBackwardGS_){
/* Forward Mode */
for (int i = 0 ; i < NumMyRows_ ; ++i) {
int NumEntries;
int col;
IFPACK_CHK_ERR(Matrix_->ExtractMyRowCopy(i, Length,NumEntries,
&Values[0], &Indices[0]));
for (int m = 0 ; m < NumVectors ; ++m) {
double dtemp = 0.0;
for (int k = 0 ; k < NumEntries ; ++k) {
col = Indices[k];
dtemp += Values[k] * y2_ptr[m][col];
}
y2_ptr[m][i] += DampingFactor_ * d_ptr[i] * (x_ptr[m][i] - dtemp);
}
}
}
else {
/* Backward Mode */
for (int i = NumMyRows_ - 1 ; i > -1 ; --i) {
int NumEntries;
int col;
IFPACK_CHK_ERR(Matrix_->ExtractMyRowCopy(i, Length,NumEntries,
&Values[0], &Indices[0]));
for (int m = 0 ; m < NumVectors ; ++m) {
double dtemp = 0.0;
for (int k = 0 ; k < NumEntries ; ++k) {
col = Indices[k];
dtemp += Values[k] * y2_ptr[m][col];
}
y2_ptr[m][i] += DampingFactor_ * d_ptr[i] * (x_ptr[m][i] - dtemp);
}
}
}
// using Export() sounded quite expensive
if (IsParallel_)
for (int m = 0 ; m < NumVectors ; ++m)
for (int ii = 0 ; ii < NumLocalSmoothingIndices_ ; ++ii) {
int i = (!LocalSmoothingIndices_)? ii : LocalSmoothingIndices_[ii];
y_ptr[m][i] = y2_ptr[m][i];
}
}
}
#ifdef IFPACK_FLOPCOUNTERS
ApplyInverseFlops_ += NumVectors * (4 * NumGlobalRows_ + 2 * NumGlobalNonzeros_);
#endif
return(0);
} //ApplyInverseGS_RowMatrix()
int Ifpack_SORa::ApplyInverse(const Epetra_MultiVector& X, Epetra_MultiVector& Y) const{
if(!IsComputed_) return -1;
Time_.ResetStartTime();
bool initial_guess_is_zero=false;
const int lclNumRows = W_->NumMyRows();
const int NumVectors = X.NumVectors();
Epetra_MultiVector Temp(A_->RowMatrixRowMap(),NumVectors);
double omega=GetOmega();
// need to create an auxiliary vector, Xcopy
Teuchos::RCP<const Epetra_MultiVector> Xcopy;
if (X.Pointers()[0] == Y.Pointers()[0]){
Xcopy = Teuchos::rcp( new Epetra_MultiVector(X) );
// Since the user didn't give us anything better, our initial guess is zero.
Y.Scale(0.0);
initial_guess_is_zero=true;
}
else
Xcopy = Teuchos::rcp( &X, false );
Teuchos::RCP< Epetra_MultiVector > T2;
// Note: Assuming that the matrix has an importer. Ifpack_PointRelaxation doesn't do this, but given that
// I have a CrsMatrix, I'm probably OK.
// Note: This is the lazy man's version sacrificing a few extra flops for avoiding if statements to determine
// if things are on or off processor.
// Note: T2 must be zero'd out
if (IsParallel_ && W_->Importer()) T2 = Teuchos::rcp( new Epetra_MultiVector(W_->Importer()->TargetMap(),NumVectors,true));
else T2 = Teuchos::rcp( new Epetra_MultiVector(A_->RowMatrixRowMap(),NumVectors,true));
// Pointer grabs
int* rowptr,*colind;
double *values;
double **t_ptr,** y_ptr, ** t2_ptr, **x_ptr,*d_ptr;
T2->ExtractView(&t2_ptr);
Y.ExtractView(&y_ptr);
Temp.ExtractView(&t_ptr);
Xcopy->ExtractView(&x_ptr);
Wdiag_->ExtractView(&d_ptr);
IFPACK_CHK_ERR(W_->ExtractCrsDataPointers(rowptr,colind,values));
for(int i=0; i<NumSweeps_; i++){
// Calculate b-Ax
if(!initial_guess_is_zero || i > 0) {
A_->Apply(Y,Temp);
Temp.Update(1.0,*Xcopy,-1.0);
}
else
Temp.Update(1.0,*Xcopy,0.0);
// Note: The off-processor entries of T2 never get touched (they're always zero) and the other entries are updated
// in this sweep before they are used, so we don't need to reset T2 to zero here.
// Do backsolve & update
// x = x + W^{-1} (b - A x)
for(int j=0; j<lclNumRows; j++){
double diag=d_ptr[j];
for (int m=0 ; m<NumVectors; m++) {
double dtmp=0.0;
// Note: Since the diagonal is in the matrix, we need to zero that entry of T2 here to make sure it doesn't contribute.
t2_ptr[m][j]=0.0;
for(int k=rowptr[j];k<rowptr[j+1];k++){
dtmp+= values[k]*t2_ptr[m][colind[k]];
}
// Yes, we need to update both of these.
t2_ptr[m][j] = (t_ptr[m][j]- dtmp)/diag;
y_ptr[m][j] += omega*t2_ptr[m][j];
}
}
}
// Counter update
NumApplyInverse_++;
ApplyInverseTime_ += Time_.ElapsedTime();
return 0;
}
//==============================================================================
int Ifpack_Chebyshev::
ApplyInverse(const Epetra_MultiVector& X, Epetra_MultiVector& Y) const
{
if (!IsComputed())
IFPACK_CHK_ERR(-3);
if (PolyDegree_ == 0)
return 0;
int nVec = X.NumVectors();
int len = X.MyLength();
if (nVec != Y.NumVectors())
IFPACK_CHK_ERR(-2);
Time_->ResetStartTime();
// AztecOO gives X and Y pointing to the same memory location,
// need to create an auxiliary vector, Xcopy
Teuchos::RefCountPtr<const Epetra_MultiVector> Xcopy;
if (X.Pointers()[0] == Y.Pointers()[0])
Xcopy = Teuchos::rcp( new Epetra_MultiVector(X) );
else
Xcopy = Teuchos::rcp( &X, false );
double **xPtr = 0, **yPtr = 0;
Xcopy->ExtractView(&xPtr);
Y.ExtractView(&yPtr);
#ifdef HAVE_IFPACK_EPETRAEXT
EpetraExt_PointToBlockDiagPermute* IBD=0;
if (UseBlockMode_) IBD=&*InvBlockDiagonal_;
#endif
//--- Do a quick solve when the matrix is identity
double *invDiag=0;
if(!UseBlockMode_) invDiag=InvDiagonal_->Values();
if ((LambdaMin_ == 1.0) && (LambdaMax_ == LambdaMin_)) {
#ifdef HAVE_IFPACK_EPETRAEXT
if(UseBlockMode_) IBD->ApplyInverse(*Xcopy,Y);
else
#endif
if (nVec == 1) {
double *yPointer = yPtr[0], *xPointer = xPtr[0];
for (int i = 0; i < len; ++i)
yPointer[i] = xPointer[i]*invDiag[i];
}
else {
int i, k;
for (i = 0; i < len; ++i) {
double coeff = invDiag[i];
for (k = 0; k < nVec; ++k)
yPtr[k][i] = xPtr[k][i] * coeff;
}
} // if (nVec == 1)
return 0;
} // if ((LambdaMin_ == 1.0) && (LambdaMax_ == LambdaMin_))
//--- Initialize coefficients
// Note that delta stores the inverse of ML_Cheby::delta
double alpha = LambdaMax_ / EigRatio_;
double beta = 1.1 * LambdaMax_;
double delta = 2.0 / (beta - alpha);
double theta = 0.5 * (beta + alpha);
double s1 = theta * delta;
//--- Define vectors
// In ML_Cheby, V corresponds to pAux and W to dk
Epetra_MultiVector V(X);
Epetra_MultiVector W(X);
#ifdef HAVE_IFPACK_EPETRAEXT
Epetra_MultiVector Temp(X);
#endif
double *vPointer = V.Values(), *wPointer = W.Values();
double oneOverTheta = 1.0/theta;
int i, j, k;
//--- If solving normal equations, multiply RHS by A^T
if(SolveNormalEquations_){
Apply_Transpose(Operator_,Y,V);
Y=V;
}
// Do the smoothing when block scaling is turned OFF
// --- Treat the initial guess
if (ZeroStartingSolution_ == false) {
Operator_->Apply(Y, V);
// Compute W = invDiag * ( X - V )/ Theta
#ifdef HAVE_IFPACK_EPETRAEXT
if(UseBlockMode_) {
Temp.Update(oneOverTheta,X,-oneOverTheta,V,0.0);
IBD->ApplyInverse(Temp,W);
// Perform additional matvecs for normal equations
// CMS: Testing this only in block mode FOR NOW
if(SolveNormalEquations_){
IBD->ApplyInverse(W,Temp);
Apply_Transpose(Operator_,Temp,W);
}
}
else
#endif
if (nVec == 1) {
double *xPointer = xPtr[0];
for (i = 0; i < len; ++i)
wPointer[i] = invDiag[i] * (xPointer[i] - vPointer[i]) * oneOverTheta;
}
else {
for (i = 0; i < len; ++i) {
double coeff = invDiag[i]*oneOverTheta;
double *wi = wPointer + i, *vi = vPointer + i;
for (k = 0; k < nVec; ++k) {
*wi = (xPtr[k][i] - (*vi)) * coeff;
wi = wi + len; vi = vi + len;
}
}
} // if (nVec == 1)
// Update the vector Y
Y.Update(1.0, W, 1.0);
}
else {
// Compute W = invDiag * X / Theta
#ifdef HAVE_IFPACK_EPETRAEXT
if(UseBlockMode_) {
IBD->ApplyInverse(X,W);
// Perform additional matvecs for normal equations
// CMS: Testing this only in block mode FOR NOW
if(SolveNormalEquations_){
IBD->ApplyInverse(W,Temp);
Apply_Transpose(Operator_,Temp,W);
}
W.Scale(oneOverTheta);
Y.Update(1.0, W, 0.0);
}
else
#endif
if (nVec == 1) {
double *xPointer = xPtr[0];
for (i = 0; i < len; ++i){
wPointer[i] = invDiag[i] * xPointer[i] * oneOverTheta;
}
memcpy(yPtr[0], wPointer, len*sizeof(double));
}
else {
for (i = 0; i < len; ++i) {
double coeff = invDiag[i]*oneOverTheta;
double *wi = wPointer + i;
for (k = 0; k < nVec; ++k) {
*wi = xPtr[k][i] * coeff;
wi = wi + len;
}
}
for (k = 0; k < nVec; ++k)
memcpy(yPtr[k], wPointer + k*len, len*sizeof(double));
} // if (nVec == 1)
} // if (ZeroStartingSolution_ == false)
//--- Apply the polynomial
double rhok = 1.0/s1, rhokp1;
double dtemp1, dtemp2;
int degreeMinusOne = PolyDegree_ - 1;
if (nVec == 1) {
double *xPointer = xPtr[0];
for (k = 0; k < degreeMinusOne; ++k) {
Operator_->Apply(Y, V);
rhokp1 = 1.0 / (2.0*s1 - rhok);
dtemp1 = rhokp1 * rhok;
dtemp2 = 2.0 * rhokp1 * delta;
rhok = rhokp1;
// Compute W = dtemp1 * W
W.Scale(dtemp1);
// Compute W = W + dtemp2 * invDiag * ( X - V )
#ifdef HAVE_IFPACK_EPETRAEXT
if(UseBlockMode_) {
//NTS: We can clobber V since it will be reset in the Apply
V.Update(dtemp2,X,-dtemp2);
IBD->ApplyInverse(V,Temp);
// Perform additional matvecs for normal equations
// CMS: Testing this only in block mode FOR NOW
if(SolveNormalEquations_){
IBD->ApplyInverse(V,Temp);
Apply_Transpose(Operator_,Temp,V);
}
W.Update(1.0,Temp,1.0);
}
else{
#endif
for (i = 0; i < len; ++i)
wPointer[i] += dtemp2* invDiag[i] * (xPointer[i] - vPointer[i]);
#ifdef HAVE_IFPACK_EPETRAEXT
}
#endif
// Update the vector Y
Y.Update(1.0, W, 1.0);
} // for (k = 0; k < degreeMinusOne; ++k)
}
else {
for (k = 0; k < degreeMinusOne; ++k) {
Operator_->Apply(Y, V);
rhokp1 = 1.0 / (2.0*s1 - rhok);
dtemp1 = rhokp1 * rhok;
dtemp2 = 2.0 * rhokp1 * delta;
rhok = rhokp1;
// Compute W = dtemp1 * W
W.Scale(dtemp1);
// Compute W = W + dtemp2 * invDiag * ( X - V )
#ifdef HAVE_IFPACK_EPETRAEXT
if(UseBlockMode_) {
//We can clobber V since it will be reset in the Apply
V.Update(dtemp2,X,-dtemp2);
IBD->ApplyInverse(V,Temp);
// Perform additional matvecs for normal equations
// CMS: Testing this only in block mode FOR NOW
if(SolveNormalEquations_){
IBD->ApplyInverse(V,Temp);
Apply_Transpose(Operator_,Temp,V);
}
W.Update(1.0,Temp,1.0);
}
else{
#endif
for (i = 0; i < len; ++i) {
double coeff = invDiag[i]*dtemp2;
double *wi = wPointer + i, *vi = vPointer + i;
for (j = 0; j < nVec; ++j) {
*wi += (xPtr[j][i] - (*vi)) * coeff;
wi = wi + len; vi = vi + len;
}
}
#ifdef HAVE_IFPACK_EPETRAEXT
}
#endif
// Update the vector Y
Y.Update(1.0, W, 1.0);
} // for (k = 0; k < degreeMinusOne; ++k)
} // if (nVec == 1)
// Flops are updated in each of the following.
++NumApplyInverse_;
ApplyInverseTime_ += Time_->ElapsedTime();
return(0);
}
int Amesos_Scalapack::Solve() {
if( debug_ == 1 ) std::cout << "Entering `Solve()'" << std::endl;
NumSolve_++;
Epetra_MultiVector *vecX = Problem_->GetLHS() ;
Epetra_MultiVector *vecB = Problem_->GetRHS() ;
//
// Compute the number of right hands sides
// (and check that X and B have the same shape)
//
int nrhs;
if ( vecX == 0 ) {
nrhs = 0 ;
EPETRA_CHK_ERR( vecB != 0 ) ;
} else {
nrhs = vecX->NumVectors() ;
EPETRA_CHK_ERR( vecB->NumVectors() != nrhs ) ;
}
Epetra_MultiVector *ScalapackB =0;
Epetra_MultiVector *ScalapackX =0;
//
// Extract Scalapack versions of X and B
//
double *ScalapackXvalues ;
Epetra_RowMatrix *RowMatrixA = dynamic_cast<Epetra_RowMatrix *>(Problem_->GetOperator());
Time_->ResetStartTime(); // track time to broadcast vectors
//
// Copy B to the scalapack version of B
//
const Epetra_Map &OriginalMap = RowMatrixA->RowMatrixRowMap();
Epetra_MultiVector *ScalapackXextract = new Epetra_MultiVector( *VectorMap_, nrhs ) ;
Epetra_MultiVector *ScalapackBextract = new Epetra_MultiVector( *VectorMap_, nrhs ) ;
Epetra_Import ImportToScalapack( *VectorMap_, OriginalMap );
ScalapackBextract->Import( *vecB, ImportToScalapack, Insert ) ;
ScalapackB = ScalapackBextract ;
ScalapackX = ScalapackXextract ;
VecTime_ += Time_->ElapsedTime();
//
// Call SCALAPACKs PDGETRS to perform the solve
//
int DescX[10];
ScalapackX->Scale(1.0, *ScalapackB) ;
int ScalapackXlda ;
Time_->ResetStartTime(); // tract time to solve
//
// Setup DescX
//
if( nrhs > nb_ ) {
EPETRA_CHK_ERR( -2 );
}
int Ierr[1] ;
Ierr[0] = 0 ;
const int zero = 0 ;
const int one = 1 ;
if ( iam_ < nprow_ * npcol_ ) {
assert( ScalapackX->ExtractView( &ScalapackXvalues, &ScalapackXlda ) == 0 ) ;
if ( false ) std::cout << "Amesos_Scalapack.cpp: " << __LINE__ << " ScalapackXlda = " << ScalapackXlda
<< " lda_ = " << lda_
<< " nprow_ = " << nprow_
<< " npcol_ = " << npcol_
<< " myprow_ = " << myprow_
<< " mypcol_ = " << mypcol_
<< " iam_ = " << iam_ << std::endl ;
if ( TwoD_distribution_ ) assert( mypcol_ >0 || EPETRA_MAX(ScalapackXlda,1) == lda_ ) ;
DESCINIT_F77(DescX,
&NumGlobalElements_,
&nrhs,
&nb_,
&nb_,
&zero,
&zero,
&ictxt_,
&lda_,
Ierr ) ;
assert( Ierr[0] == 0 ) ;
//
// For the 1D data distribution, we factor the transposed
// matrix, hence we must invert the sense of the transposition
//
char trans = 'N';
if ( TwoD_distribution_ ) {
if ( UseTranspose() ) trans = 'T' ;
} else {
if ( ! UseTranspose() ) trans = 'T' ;
}
if ( nprow_ * npcol_ == 1 ) {
DGETRS_F77(&trans,
&NumGlobalElements_,
&nrhs,
&DenseA_[0],
&lda_,
&Ipiv_[0],
ScalapackXvalues,
&lda_,
Ierr ) ;
} else {
PDGETRS_F77(&trans,
&NumGlobalElements_,
&nrhs,
&DenseA_[0],
&one,
&one,
DescA_,
&Ipiv_[0],
ScalapackXvalues,
&one,
&one,
DescX,
Ierr ) ;
}
}
SolTime_ += Time_->ElapsedTime();
Time_->ResetStartTime(); // track time to broadcast vectors
//
// Copy X back to the original vector
//
Epetra_Import ImportFromScalapack( OriginalMap, *VectorMap_ );
vecX->Import( *ScalapackX, ImportFromScalapack, Insert ) ;
delete ScalapackBextract ;
delete ScalapackXextract ;
VecTime_ += Time_->ElapsedTime();
// All processes should return the same error code
if ( nprow_ * npcol_ < Comm().NumProc() )
Comm().Broadcast( Ierr, 1, 0 ) ;
// MS // compute vector norms
if( ComputeVectorNorms_ == true || verbose_ == 2 ) {
double NormLHS, NormRHS;
for( int i=0 ; i<nrhs ; ++i ) {
assert((*vecX)(i)->Norm2(&NormLHS)==0);
assert((*vecB)(i)->Norm2(&NormRHS)==0);
if( verbose_ && Comm().MyPID() == 0 ) {
std::cout << "Amesos_Scalapack : vector " << i << ", ||x|| = " << NormLHS
<< ", ||b|| = " << NormRHS << std::endl;
}
}
}
// MS // compute true residual
if( ComputeTrueResidual_ == true || verbose_ == 2 ) {
double Norm;
Epetra_MultiVector Ax(vecB->Map(),nrhs);
for( int i=0 ; i<nrhs ; ++i ) {
(Problem_->GetMatrix()->Multiply(UseTranspose(), *((*vecX)(i)), Ax));
(Ax.Update(1.0, *((*vecB)(i)), -1.0));
(Ax.Norm2(&Norm));
if( verbose_ && Comm().MyPID() == 0 ) {
std::cout << "Amesos_Scalapack : vector " << i << ", ||Ax - b|| = " << Norm << std::endl;
}
}
}
return Ierr[0];
}
int ShyLU_Probing_Operator::Apply(const Epetra_MultiVector &X,
Epetra_MultiVector &Y) const
{
#ifdef TIMING_OUTPUT
apply_time_->start();
#endif
int nvectors = X.NumVectors();
bool local = (C_->Comm().NumProc() == 1);
int err;
//cout << "No of colors after probing" << nvectors << endl;
#ifdef TIMING_OUTPUT
matvec_time_->start();
#endif
err = G_->Multiply(false, X, *temp2);
assert(err == 0);
if (!local)
err = C_->Multiply(false, X, *temp);
else
{
// localize X
double *values;
int mylda;
X.ExtractView(&values, &mylda);
Epetra_SerialComm LComm; // Use Serial Comm for the local blocks.
Epetra_Map SerialMap(X.Map().NumMyElements(), X.Map().NumMyElements(),
X.Map().MyGlobalElements(), 0, LComm);
Epetra_MultiVector Xl(View, SerialMap, values, mylda, X.NumVectors());
err = C_->Multiply(false, Xl, *temp);
}
assert(err == 0);
#ifdef TIMING_OUTPUT
matvec_time_->stop();
#endif
int nrows = C_->RowMap().NumMyElements();
#ifdef DEBUG
cout << "DEBUG MODE" << endl;
assert(nrows == localDRowMap_->NumGlobalElements());
int gids[nrows], gids1[nrows];
C_->RowMap().MyGlobalElements(gids);
localDRowMap_->MyGlobalElements(gids1);
for (int i = 0; i < nrows; i++)
{
assert(gids[i] == gids1[i]);
}
#endif
#ifdef TIMING_OUTPUT
localize_time_->start();
#endif
//int err;
int lda;
double *values;
if (!local)
{
err = temp->ExtractView(&values, &lda);
assert (err == 0);
// copy to local vector //TODO: OMP parallel
assert(lda == nrows);
//#pragma omp parallel for shared(nvectors, nrows, values)
for (int v = 0; v < nvectors; v++)
{
for (int i = 0; i < nrows; i++)
{
err = ltemp->ReplaceMyValue(i, v, values[i+v*lda]);
assert (err == 0);
}
}
}
#ifdef TIMING_OUTPUT
localize_time_->stop();
trisolve_time_->start();
#endif
if (!local)
{
LP_->SetRHS(ltemp.getRawPtr());
}
else
{
//LP_->SetRHS(temp.getRawPtr());
}
//LP_->SetLHS(localX.getRawPtr());
//TODO: Why not just in Reset(). Check the distr path.
ssym_->OrigLP->SetLHS(localX.getRawPtr());
ssym_->OrigLP->SetRHS(temp.getRawPtr());
ssym_->ReIdx_LP->fwd();
solver_->Solve();
#ifdef TIMING_OUTPUT
trisolve_time_->stop();
dist_time_->start();
#endif
if (!local)
{
err = localX->ExtractView(&values, &lda);
assert (err == 0);
//Copy back to dist vector //TODO: OMP parallel
//#pragma omp parallel for
for (int v = 0; v < nvectors; v++)
{
for (int i = 0; i < nrows; i++)
{
err = temp->ReplaceMyValue(i, v, values[i+v*lda]);
assert (err == 0);
}
}
}
#ifdef TIMING_OUTPUT
dist_time_->stop();
matvec2_time_->start();
#endif
if (!local)
{
R_->Multiply(false, *temp, Y);
}
else
{
// Should Y be localY in Multiply and then exported to Y ?? TODO:
// Use view mode ?
double *values;
int mylda;
Y.ExtractView(&values, &mylda);
Epetra_SerialComm LComm; // Use Serial Comm for the local blocks.
Epetra_Map SerialMap(Y.Map().NumMyElements(), Y.Map().NumMyElements(),
Y.Map().MyGlobalElements(), 0, LComm);
Epetra_MultiVector Yl(View, SerialMap, values, mylda, Y.NumVectors());
R_->Multiply(false, *localX, Yl);
}
#ifdef TIMING_OUTPUT
matvec2_time_->stop();
update_time_->start();
#endif
err = Y.Update(1.0, *temp2, -1.0);
//cout << Y.MyLength() << " " << temp2.MyLength() << endl;
assert(err == 0);
#ifdef TIMING_OUTPUT
update_time_->stop();
apply_time_->stop();
#endif
cntApply++;
return 0;
}
int BuildMultiVectorTests (Epetra_MultiVector & C, const double alpha,
Epetra_MultiVector& A,
Epetra_MultiVector& sqrtA,
Epetra_MultiVector& B,
Epetra_MultiVector& C_alphaA,
Epetra_MultiVector& C_alphaAplusB,
Epetra_MultiVector& C_plusB,
double* const dotvec_AB,
double* const norm1_A,
double* const norm2_sqrtA,
double* const norminf_A,
double* const normw_A,
Epetra_MultiVector& Weights,
double* const minval_A,
double* const maxval_A,
double* const meanval_A ) {
// For given values alpha and a (possibly zero) filled
// Epetra_MultiVector (the this object), allocated double * arguments dotvec_AB,
// norm1_A, and norm2_A, and allocated Epetra_MultiVectors A, sqrtA,
// B, C_alpha, C_alphaAplusB and C_plusB, this method will generate values for
// Epetra_MultiVectors A, B and all of the additional arguments on
// the list above such that, if A, B and (this) are used with the methods in
// this class, the results should match the results generated by this routine.
// Specifically, the results in dotvec_AB should match those from a call to
// A.dotProd (B,dotvec). Similarly for other routines.
int i,j;
double fi, fj; // Used for casting loop variables to floats
// Define some useful constants
int A_nrows = A.MyLength();
int A_ncols = A.NumVectors();
int sqrtA_nrows = sqrtA.MyLength();
int sqrtA_ncols = sqrtA.NumVectors();
int B_nrows = B.MyLength();
int B_ncols = B.NumVectors();
double **Ap = 0;
double **sqrtAp = 0;
double **Bp = 0;
double **Cp = 0;
double **C_alphaAp = 0;
double **C_alphaAplusBp = 0;
double **C_plusBp = 0;
double **Weightsp = 0;
A.ExtractView(&Ap);
sqrtA.ExtractView(&sqrtAp);
B.ExtractView(&Bp);
C.ExtractView(&Cp);
C_alphaA.ExtractView(&C_alphaAp);
C_alphaAplusB.ExtractView(&C_alphaAplusBp);
C_plusB.ExtractView(&C_plusBp);
Weights.ExtractView(&Weightsp);
bool A_is_local = (A.MyLength() == A.GlobalLength());
bool B_is_local = (B.MyLength() == B.GlobalLength());
bool C_is_local = (C.MyLength() == C.GlobalLength());
int A_IndexBase = A.Map().IndexBase();
int B_IndexBase = B.Map().IndexBase();
// Build two new maps that we can use for defining global equation indices below
Epetra_Map * A_Map = new Epetra_Map(-1, A_nrows, A_IndexBase, A.Map().Comm());
Epetra_Map * B_Map = new Epetra_Map(-1, B_nrows, B_IndexBase, B.Map().Comm());
int* A_MyGlobalElements = new int[A_nrows];
A_Map->MyGlobalElements(A_MyGlobalElements);
int* B_MyGlobalElements = new int[B_nrows];
B_Map->MyGlobalElements(B_MyGlobalElements);
// Check for compatible dimensions
if (C.MyLength() != A_nrows ||
A_nrows != B_nrows ||
C.NumVectors() != A_ncols ||
A_ncols != B_ncols ||
sqrtA_nrows != A_nrows ||
sqrtA_ncols != A_ncols ||
C.MyLength() != C_alphaA.MyLength() ||
C.NumVectors() != C_alphaA.NumVectors() ||
C.MyLength() != C_alphaAplusB.MyLength() ||
C.NumVectors() != C_alphaAplusB.NumVectors() ||
C.MyLength() != C_plusB.MyLength() ||
C.NumVectors() != C_plusB.NumVectors() ) return(-2); // Return error
bool Case1 = ( A_is_local && B_is_local && C_is_local); // Case 1
bool Case2 = (!A_is_local && !B_is_local && !C_is_local);// Case 2
// Test for meaningful cases
if (!(Case1 || Case2)) return(-3); // Meaningless case
/* Fill A and B with values as follows:
If A_is_local is false:
A(i,j) = A_MyGlobalElements[i]*j, i=1,...,numLocalEquations, j=1,...,NumVectors
else
A(i,j) = i*j, i=1,...,numLocalEquations, j=1,...,NumVectors
If B_is_local is false:
B(i,j) = 1/(A_MyGlobalElements[i]*j), i=1,...,numLocalEquations, j=1,...,NumVectors
else
B(i,j) = 1/(i*j), i=1,...,numLocalEquations, j=1,...,NumVectors
In addition, scale each entry by GlobalLength for A and
1/GlobalLength for B--keeps the magnitude of entries in check
*/
//Define scale factor
double sf = A.GlobalLength();
double sfinv = 1.0/sf;
// Define A
if (A_is_local)
{
for (j = 0; j <A_ncols ; j++)
{
for (i = 0; i<A_nrows; i++)
{
fi = i+1; // Get float version of i and j, offset by 1.
fj = j+1;
Ap[j][i] = (fi*sfinv)*fj;
sqrtAp[j][i] = std::sqrt(Ap[j][i]);
}
}
}
else
{
for (j = 0; j <A_ncols ; j++)
{
for (i = 0; i<A_nrows; i++)
{
fi = A_MyGlobalElements[i]+1; // Get float version of i and j, offset by 1.
fj = j+1;
Ap[j][i] = (fi*sfinv)*fj;
sqrtAp[j][i] = std::sqrt(Ap[j][i]);
}
}
}
// Define B depending on TransB and B_is_local
if (B_is_local)
{
for (j = 0; j <B_ncols ; j++)
{
for (i = 0; i<B_nrows; i++)
{
fi = i+1; // Get float version of i and j, offset by 1.
fj = j+1;
Bp[j][i] = 1.0/((fi*sfinv)*fj);
}
}
}
else
{
for (j = 0; j <B_ncols ; j++)
{
for (i = 0; i<B_nrows; i++)
{
fi = B_MyGlobalElements[i]+1; // Get float version of i and j, offset by 1.
fj = j+1;
Bp[j][i] = 1.0/((fi*sfinv)*fj);
}
}
}
// Generate C_alphaA = alpha * A
for (j = 0; j <A_ncols ; j++)
for (i = 0; i<A_nrows; i++)
C_alphaAp[j][i] = alpha * Ap[j][i];
// Generate C_alphaA = alpha * A + B
for (j = 0; j <A_ncols ; j++)
for (i = 0; i<A_nrows; i++)
C_alphaAplusBp[j][i] = alpha * Ap[j][i] + Bp[j][i];
// Generate C_plusB = this + B
for (j = 0; j <A_ncols ; j++)
for (i = 0; i<A_nrows; i++)
C_plusBp[j][i] = Cp[j][i] + Bp[j][i];
// Generate dotvec_AB. Because B(i,j) = 1/A(i,j), dotvec[i] = C.GlobalLength()
for (i=0; i< A.NumVectors(); i++) dotvec_AB[i] = C.GlobalLength();
// For the next two results we want to be careful how we do arithmetic
// to avoid very large numbers.
// We are computing sfinv*(C.GlobalLength()*(C.GlobalLength()+1)/2)
double result = C.GlobalLength();
result *= sfinv;
result /= 2.0;
result *= (double)(C.GlobalLength()+1);
// Generate norm1_A. Can use formula for sum of first n integers.
for (i=0; i< A.NumVectors(); i++)
// m1_A[i] = (i+1)*C.GlobalLength()*(C.GlobalLength()+1)/2;
norm1_A[i] = result * ((double) (i+1));
// Generate norm2_sqrtA. Can use formula for sum of first n integers.
for (i=0; i< A.NumVectors(); i++)
// norm2_sqrtA[i] = std::sqrt((double) ((i+1)*C.GlobalLength()*(C.GlobalLength()+1)/2));
norm2_sqrtA[i] = std::sqrt(result * ((double) (i+1)));
// Generate norminf_A, minval_A, maxval_A, meanval_A.
for (i=0; i< A.NumVectors(); i++)
{
norminf_A[i] = (double) (i+1);
minval_A[i] = (double) (i+1)/ (double) A.GlobalLength();
maxval_A[i] = (double) (i+1);
meanval_A[i] = norm1_A[i]/((double) (A.GlobalLength()));
}
// Define weights and expected weighted norm
for (i=0; i< A.NumVectors(); i++)
{
double ip1 = (double) i+1;
normw_A[i] = ip1;
for (j=0; j<A_nrows; j++) Weightsp[i][j] = Ap[i][j]/ip1;
}
delete A_Map;
delete B_Map;
delete [] A_MyGlobalElements;
delete [] B_MyGlobalElements;
return(0);
}
int
LOCA::Epetra::AugmentedOp::Apply(const Epetra_MultiVector& Input,
Epetra_MultiVector& Result) const
{
// Get number of vectors
int n = Input.NumVectors();
// Check num vectors is correct
if (importedInput == NULL || n != importedInput->NumVectors())
globalData->locaErrorCheck->throwError("LOCA::Epetra::AugmentedOp::Apply()"
"Must call init() before Apply()!");
// Import parameter components
importedInput->Import(Input, *extendedImporter, Insert);
// get views
double **input_view;
double **result_view;
importedInput->ExtractView(&input_view);
Result.ExtractView(&result_view);
// get views of paramter components
// this is done by setting the pointer for each column to start at
// underlyingLength
double **input_view_y = new double*[n];
for (int i=0; i<n; i++)
input_view_y[i] = input_view[i]+underlyingLength;
// break input, result into components
Epetra_MultiVector input_x(View, underlyingMap, input_view, n);
Epetra_MultiVector input_y(View, localMap, input_view_y, n);
Epetra_MultiVector result_x(View, underlyingMap, result_view, n);
// Compute J*input_x
jacOperator->Apply(input_x, result_x);
if (useTranspose) {
// Compute J^T*input_x + b*input_y
result_x.Multiply('N', 'N', 1.0, *b, input_y, 1.0);
// Compute a^T*input_x + c^T*input_y
result_y->Multiply('T', 'N', 1.0, *a, input_x, 0.0);
result_y->Multiply('T', 'N', 1.0, c, input_y, 1.0);
}
else {
// Compute J*input_x + a*input_y
result_x.Multiply('N', 'N', 1.0, *a, input_y, 1.0);
// Compute b^T*input_x + c*input_y
result_y->Multiply('T', 'N', 1.0, *b, input_x, 0.0);
result_y->Multiply('N', 'N', 1.0, c, input_y, 1.0);
}
// Set parameter component
if (haveParamComponent)
for (int j=0; j<n; j++)
for (int i=0; i<numConstraints; i++)
result_view[j][underlyingLength+i] = (*result_y)[j][i];
delete [] input_view_y;
return 0;
}
int main(int argc, char *argv[])
{
#ifdef HAVE_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm(MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
#define ML_SCALING
#ifdef ML_SCALING
const int ntimers=4;
enum {total, probBuild, precBuild, solve};
ml_DblLoc timeVec[ntimers], maxTime[ntimers], minTime[ntimers];
for (int i=0; i<ntimers; i++) timeVec[i].rank = Comm.MyPID();
timeVec[total].value = MPI_Wtime();
#endif
int nx;
if (argc > 1) nx = (int) strtol(argv[1],NULL,10);
else nx = 256;
if (nx < 1) nx = 256; // input a nonpositive integer if you want to specify
// the XML input file name.
nx = nx*(int)sqrt((double)Comm.NumProc());
int ny = nx;
printf("nx = %d\nny = %d\n",nx,ny);
fflush(stdout);
char xmlFile[80];
bool readXML=false;
if (argc > 2) {strcpy(xmlFile,argv[2]); readXML = true;}
else sprintf(xmlFile,"%s","params.xml");
ParameterList GaleriList;
GaleriList.set("nx", nx);
GaleriList.set("ny", ny);
#ifdef ML_SCALING
timeVec[probBuild].value = MPI_Wtime();
#endif
Epetra_Map* Map = CreateMap("Cartesian2D", Comm, GaleriList);
Epetra_CrsMatrix* A = CreateCrsMatrix("Laplace2D", Map, GaleriList);
if (!Comm.MyPID()) printf("nx = %d, ny = %d, mx = %d, my = %d\n",nx,ny,GaleriList.get("mx",-1),GaleriList.get("my",-1));
fflush(stdout);
//avoid potential overflow
double numMyRows = A->NumMyRows();
double numGlobalRows;
Comm.SumAll(&numMyRows,&numGlobalRows,1);
if (!Comm.MyPID()) printf("# global rows = %1.0f\n",numGlobalRows);
//printf("pid %d: #rows = %d\n",Comm.MyPID(),A->NumMyRows());
fflush(stdout);
Epetra_MultiVector *coords = CreateCartesianCoordinates("2D", Map,GaleriList);
double *x_coord=0,*y_coord=0,*z_coord=0;
double **ttt;
if (!coords->ExtractView(&ttt)) {
x_coord = ttt[0];
y_coord = ttt[1];
} else {
if (!Comm.MyPID()) printf("Error extracting coordinate vectors\n");
MPI_Finalize();
exit(EXIT_FAILURE);
}
Epetra_Vector LHS(*Map); LHS.Random();
Epetra_Vector RHS(*Map); RHS.PutScalar(0.0);
Epetra_LinearProblem Problem(A, &LHS, &RHS);
AztecOO solver(Problem);
#ifdef ML_SCALING
timeVec[probBuild].value = MPI_Wtime() - timeVec[probBuild].value;
#endif
// =========================== begin of ML part ===========================
#ifdef ML_SCALING
timeVec[precBuild].value = MPI_Wtime();
#endif
ParameterList MLList;
if (readXML) {
MLList.set("read XML",true);
MLList.set("XML input file",xmlFile);
}
else {
cout << "here" << endl;
ML_Epetra::SetDefaults("SA",MLList);
MLList.set("smoother: type","Chebyshev");
MLList.set("smoother: sweeps",3);
MLList.set("coarse: max size",1);
}
MLList.set("x-coordinates", x_coord);
MLList.set("y-coordinates", y_coord);
MLList.set("z-coordinates", z_coord);
/*
RCP<std::vector<int> >
m_smootherAztecOptions = rcp(new std::vector<int>(AZ_OPTIONS_SIZE));
RCP<std::vector<double> >
m_smootherAztecParams = rcp(new std::vector<double>(AZ_PARAMS_SIZE));
//int m_smootherAztecOptions[AZ_OPTIONS_SIZE];
//double m_smootherAztecParams[AZ_PARAMS_SIZE];
std::string smootherType("Aztec");
AZ_defaults(&(*m_smootherAztecOptions)[0],&(*m_smootherAztecParams)[0]);
(*m_smootherAztecOptions)[AZ_precond] = AZ_dom_decomp;
(*m_smootherAztecOptions)[AZ_subdomain_solve] = AZ_icc;
bool smootherAztecAsSolver = true;
*/
MLList.set("ML output",10);
MLList.set("repartition: enable",1);
MLList.set("repartition: max min ratio",1.3);
MLList.set("repartition: min per proc",200);
MLList.set("repartition: partitioner","Zoltan");
MLList.set("repartition: Zoltan dimensions",2);
MLList.set("repartition: put on single proc",1);
MLList.set("repartition: Zoltan type","hypergraph");
MLList.set("repartition: estimated iterations",13);
/*
MLList.set("smoother: Aztec options",m_smootherAztecOptions);
MLList.set("smoother: Aztec params",m_smootherAztecParams);
MLList.set("smoother: type",smootherType.c_str());
MLList.set("smoother: Aztec as solver", smootherAztecAsSolver);
MLList.set("ML print initial list", 0);
MLList.set("ML print final list", 0);
*/
ML_Epetra::MultiLevelPreconditioner* MLPrec =
new ML_Epetra::MultiLevelPreconditioner(*A, MLList);
// verify unused parameters on process 0 (put -1 to print on all
// processes)
MLPrec->PrintUnused(0);
#ifdef ML_SCALING
timeVec[precBuild].value = MPI_Wtime() - timeVec[precBuild].value;
#endif
// =========================== end of ML part =============================
#ifdef ML_SCALING
timeVec[solve].value = MPI_Wtime();
#endif
solver.SetPrecOperator(MLPrec);
solver.SetAztecOption(AZ_solver, AZ_cg);
solver.SetAztecOption(AZ_output, 1);
solver.Iterate(500, 1e-12);
#ifdef ML_SCALING
timeVec[solve].value = MPI_Wtime() - timeVec[solve].value;
#endif
// destroy the preconditioner
delete MLPrec;
// compute the real residual
double residual;
LHS.Norm2(&residual);
if( Comm.MyPID()==0 ) {
cout << "||b-Ax||_2 = " << residual << endl;
}
// for testing purposes
if (residual > 1e-5)
exit(EXIT_FAILURE);
delete A;
delete Map;
delete coords;
#ifdef ML_SCALING
timeVec[total].value = MPI_Wtime() - timeVec[total].value;
//avg
double dupTime[ntimers],avgTime[ntimers];
for (int i=0; i<ntimers; i++) dupTime[i] = timeVec[i].value;
MPI_Reduce(dupTime,avgTime,ntimers,MPI_DOUBLE,MPI_SUM,0,MPI_COMM_WORLD);
for (int i=0; i<ntimers; i++) avgTime[i] = avgTime[i]/Comm.NumProc();
//min
MPI_Reduce(timeVec,minTime,ntimers,MPI_DOUBLE_INT,MPI_MINLOC,0,MPI_COMM_WORLD);
//max
MPI_Reduce(timeVec,maxTime,ntimers,MPI_DOUBLE_INT,MPI_MAXLOC,0,MPI_COMM_WORLD);
if (Comm.MyPID() == 0) {
printf("timing : max (pid) min (pid) avg\n");
printf("Problem build : %2.3e (%d) %2.3e (%d) %2.3e \n",
maxTime[probBuild].value,maxTime[probBuild].rank,
minTime[probBuild].value,minTime[probBuild].rank,
avgTime[probBuild]);
printf("Preconditioner build : %2.3e (%d) %2.3e (%d) %2.3e \n",
maxTime[precBuild].value,maxTime[precBuild].rank,
minTime[precBuild].value,minTime[precBuild].rank,
avgTime[precBuild]);
printf("Solve : %2.3e (%d) %2.3e (%d) %2.3e \n",
maxTime[solve].value,maxTime[solve].rank,
minTime[solve].value,minTime[solve].rank,
avgTime[solve]);
printf("Total : %2.3e (%d) %2.3e (%d) %2.3e \n",
maxTime[total].value,maxTime[total].rank,
minTime[total].value,minTime[total].rank,
avgTime[total]);
}
#endif
#ifdef HAVE_MPI
MPI_Finalize();
#endif
return(EXIT_SUCCESS);
}
//==============================================================================
// this requires Epetra_CrsMatrix + OptimizeStorage()
int Ifpack_PointRelaxation::
ApplyInverseSGS_FastCrsMatrix(const Epetra_CrsMatrix* A, const Epetra_MultiVector& X, Epetra_MultiVector& Y) const
{
int* IndexOffset;
int* Indices;
double* Values;
IFPACK_CHK_ERR(A->ExtractCrsDataPointers(IndexOffset, Indices, Values));
int NumVectors = X.NumVectors();
Teuchos::RefCountPtr< Epetra_MultiVector > Y2;
if (IsParallel_) {
Y2 = Teuchos::rcp( new Epetra_MultiVector(Importer_->TargetMap(), NumVectors) );
}
else
Y2 = Teuchos::rcp( &Y, false );
double** y_ptr, ** y2_ptr, ** x_ptr, *d_ptr;
X.ExtractView(&x_ptr);
Y.ExtractView(&y_ptr);
Y2->ExtractView(&y2_ptr);
Diagonal_->ExtractView(&d_ptr);
for (int iter = 0 ; iter < NumSweeps_ ; ++iter) {
// only one data exchange per sweep
if (IsParallel_)
IFPACK_CHK_ERR(Y2->Import(Y,*Importer_,Insert));
for (int i = 0 ; i < NumMyRows_ ; ++i) {
int col;
double diag = d_ptr[i];
// no need to extract row i
//IFPACK_CHK_ERR(A->ExtractMyRowView(i, NumEntries, Values, Indices));
for (int m = 0 ; m < NumVectors ; ++m) {
double dtemp = 0.0;
for (int k = IndexOffset[i] ; k < IndexOffset[i + 1] ; ++k) {
col = Indices[k];
dtemp += Values[k] * y2_ptr[m][col];
}
y2_ptr[m][i] += DampingFactor_ * (x_ptr[m][i] - dtemp) * diag;
}
}
for (int i = NumMyRows_ - 1 ; i > -1 ; --i) {
int col;
double diag = d_ptr[i];
// no need to extract row i
//IFPACK_CHK_ERR(A->ExtractMyRowView(i, NumEntries, Values, Indices));
for (int m = 0 ; m < NumVectors ; ++m) {
double dtemp = 0.0;
for (int k = IndexOffset[i] ; k < IndexOffset[i + 1] ; ++k) {
col = Indices[k];
dtemp += Values[k] * y2_ptr[m][col];
}
y2_ptr[m][i] += DampingFactor_ * (x_ptr[m][i] - dtemp) * diag;
}
}
if (IsParallel_)
for (int m = 0 ; m < NumVectors ; ++m)
for (int i = 0 ; i < NumMyRows_ ; ++i)
y_ptr[m][i] = y2_ptr[m][i];
}
ApplyInverseFlops_ += NumVectors * (8 * NumGlobalRows_ + 4 * NumGlobalNonzeros_);
return(0);
}
int main(int argc, char *argv[])
{
#ifdef HAVE_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm(MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
int nx;
if (argc > 1)
nx = (int) strtol(argv[1],NULL,10);
else
nx = 256;
int ny = nx * Comm.NumProc(); // each subdomain is a square
ParameterList GaleriList;
GaleriList.set("nx", nx);
GaleriList.set("ny", ny);
GaleriList.set("mx", 1);
GaleriList.set("my", Comm.NumProc());
int NumNodes = nx*ny;
int NumPDEEqns = 2;
Epetra_Map* Map = CreateMap("Cartesian2D", Comm, GaleriList);
Epetra_CrsMatrix* CrsA = CreateCrsMatrix("Laplace2D", Map, GaleriList);
Epetra_VbrMatrix* A = CreateVbrMatrix(CrsA, NumPDEEqns);
Epetra_Vector LHS(A->DomainMap()); LHS.PutScalar(0);
Epetra_Vector RHS(A->DomainMap()); RHS.Random();
Epetra_LinearProblem Problem(A, &LHS, &RHS);
AztecOO solver(Problem);
double *x_coord = 0, *y_coord = 0, *z_coord = 0;
Epetra_MultiVector *coords = CreateCartesianCoordinates("2D", &(CrsA->Map()),
GaleriList);
double **ttt;
if (!coords->ExtractView(&ttt)) {
x_coord = ttt[0];
y_coord = ttt[1];
} else {
printf("Error extracting coordinate vectors\n");
# ifdef HAVE_MPI
MPI_Finalize() ;
# endif
exit(EXIT_FAILURE);
}
ParameterList MLList;
SetDefaults("SA",MLList);
MLList.set("ML output",10);
MLList.set("max levels",10);
MLList.set("increasing or decreasing","increasing");
MLList.set("smoother: type", "Chebyshev");
MLList.set("smoother: sweeps", 3);
// *) if a low number, it will use all the available processes
// *) if a big number, it will use only processor 0 on the next level
MLList.set("aggregation: next-level aggregates per process", 1);
MLList.set("aggregation: type (level 0)", "Zoltan");
MLList.set("aggregation: type (level 1)", "Uncoupled");
MLList.set("aggregation: type (level 2)", "Zoltan");
MLList.set("aggregation: smoothing sweeps", 2);
MLList.set("x-coordinates", x_coord);
MLList.set("y-coordinates", y_coord);
MLList.set("z-coordinates", z_coord);
// specify the reduction with respect to the previous level
// (very small values can break the code)
int ratio = 16;
MLList.set("aggregation: global aggregates (level 0)",
NumNodes / ratio);
MLList.set("aggregation: global aggregates (level 1)",
NumNodes / (ratio * ratio));
MLList.set("aggregation: global aggregates (level 2)",
NumNodes / (ratio * ratio * ratio));
MultiLevelPreconditioner* MLPrec =
new MultiLevelPreconditioner(*A, MLList, true);
solver.SetPrecOperator(MLPrec);
solver.SetAztecOption(AZ_solver, AZ_cg_condnum);
solver.SetAztecOption(AZ_output, 1);
solver.Iterate(100, 1e-12);
// compute the real residual
Epetra_Vector Residual(A->DomainMap());
//1.0 * RHS + 0.0 * RHS - 1.0 * (A * LHS)
A->Apply(LHS,Residual);
Residual.Update(1.0, RHS, 0.0, RHS, -1.0);
double rn;
Residual.Norm2(&rn);
if (Comm.MyPID() == 0 )
std::cout << "||b-Ax||_2 = " << rn << endl;
if (Comm.MyPID() == 0 && rn > 1e-5) {
std::cout << "TEST FAILED!!!!" << endl;
# ifdef HAVE_MPI
MPI_Finalize() ;
# endif
exit(EXIT_FAILURE);
}
delete MLPrec;
delete coords;
delete Map;
delete CrsA;
delete A;
if (Comm.MyPID() == 0)
std::cout << "TEST PASSED" << endl;
#ifdef HAVE_MPI
MPI_Finalize() ;
#endif
exit(EXIT_SUCCESS);
}
int BuildMatrixTests (Epetra_MultiVector & C,
const char TransA, const char TransB,
const double alpha,
Epetra_MultiVector& A,
Epetra_MultiVector& B,
const double beta,
Epetra_MultiVector& C_GEMM ) {
// For given values of TransA, TransB, alpha and beta, a (possibly
// zero) filled Epetra_MultiVector C, and allocated
// Epetra_MultiVectors A, B and C_GEMM this routine will generate values for
// Epetra_MultiVectors A, B and C_GEMM such that, if A, B and (this) are
// used with GEMM in this class, the results should match the results
// generated by this routine.
// Test for Strided multivectors (required for GEMM ops)
if (!A.ConstantStride() ||
!B.ConstantStride() ||
!C_GEMM.ConstantStride() ||
!C.ConstantStride()) return(-1); // Error
int i, j;
double fi, fj; // Used for casting loop variables to floats
// Get a view of the MultiVectors
double *Ap = 0;
double *Bp = 0;
double *Cp = 0;
double *C_GEMMp = 0;
int A_nrows = A.MyLength();
int A_ncols = A.NumVectors();
int B_nrows = B.MyLength();
int B_ncols = B.NumVectors();
int C_nrows = C.MyLength();
int C_ncols = C.NumVectors();
int A_Stride = 0;
int B_Stride = 0;
int C_Stride = 0;
int C_GEMM_Stride = 0;
A.ExtractView(&Ap, &A_Stride);
B.ExtractView(&Bp, &B_Stride);
C.ExtractView(&Cp, &C_Stride);
C_GEMM.ExtractView(&C_GEMMp, &C_GEMM_Stride);
// Define some useful constants
int opA_ncols = (TransA=='N') ? A.NumVectors() : A.MyLength();
int opB_nrows = (TransB=='N') ? B.MyLength() : B.NumVectors();
int C_global_inner_dim = (TransA=='N') ? A.NumVectors() : A.GlobalLength();
bool A_is_local = (!A.DistributedGlobal());
bool B_is_local = (!B.DistributedGlobal());
bool C_is_local = (!C.DistributedGlobal());
int A_IndexBase = A.Map().IndexBase();
int B_IndexBase = B.Map().IndexBase();
// Build two new maps that we can use for defining global equation indices below
Epetra_Map * A_Map = new Epetra_Map(-1, A_nrows, A_IndexBase, A.Map().Comm());
Epetra_Map * B_Map = new Epetra_Map(-1, B_nrows, B_IndexBase, B.Map().Comm());
int* A_MyGlobalElements = new int[A_nrows];
A_Map->MyGlobalElements(A_MyGlobalElements);
int* B_MyGlobalElements = new int[B_nrows];
B_Map->MyGlobalElements(B_MyGlobalElements);
// Check for compatible dimensions
if (C.MyLength() != C_nrows ||
opA_ncols != opB_nrows ||
C.NumVectors() != C_ncols ||
C.MyLength() != C_GEMM.MyLength() ||
C.NumVectors() != C_GEMM.NumVectors() ) {
delete A_Map;
delete B_Map;
delete [] A_MyGlobalElements;
delete [] B_MyGlobalElements;
return(-2); // Return error
}
bool Case1 = ( A_is_local && B_is_local && C_is_local); // Case 1 above
bool Case2 = (!A_is_local && !B_is_local && C_is_local && TransA=='T' );// Case 2
bool Case3 = (!A_is_local && B_is_local && !C_is_local && TransA=='N');// Case 3
// Test for meaningful cases
if (!(Case1 || Case2 || Case3)) {
delete A_Map;
delete B_Map;
delete [] A_MyGlobalElements;
delete [] B_MyGlobalElements;
return(-3); // Meaningless case
}
/* Fill A, B and C with values as follows:
If A_is_local is false:
A(i,j) = A_MyGlobalElements[i]*j, i=1,...,numLocalEquations, j=1,...,NumVectors
else
A(i,j) = i*j, i=1,...,numLocalEquations, j=1,...,NumVectors
If B_is_local is false:
B(i,j) = 1/(A_MyGlobalElements[i]*j), i=1,...,numLocalEquations, j=1,...,NumVectors
else
B(i,j) = 1/(i*j), i=1,...,numLocalEquations, j=1,...,NumVectors
In addition, scale each entry by GlobalLength for A and
1/GlobalLength for B--keeps the magnitude of entries in check
C_GEMM will depend on A_is_local and B_is_local. Three cases:
1) A_is_local true and B_is_local true:
C_GEMM will be local replicated and equal to A*B = i*NumVectors/j
2) A_is_local false and B_is_local false
C_GEMM will be local replicated = A(trans)*B(i,j) = i*numGlobalEquations/j
3) A_is_local false B_is_local true
C_GEMM will distributed global and equals A*B = A_MyGlobalElements[i]*NumVectors/j
*/
// Define a scalar to keep magnitude of entries reasonable
double sf = C_global_inner_dim;
double sfinv = 1.0/sf;
// Define A depending on A_is_local
if (A_is_local)
{
for (j = 0; j <A_ncols ; j++)
for (i = 0; i<A_nrows; i++)
{
fi = i+1; // Get float version of i and j, offset by 1.
fj = j+1;
Ap[i + A_Stride*j] = (fi*sfinv)*fj;
}
}
else
{
for (j = 0; j <A_ncols ; j++)
for (i = 0; i<A_nrows; i++)
{
fi = A_MyGlobalElements[i]+1; // Get float version of i and j, offset by 1.
fj = j+1;
Ap[i + A_Stride*j] = (fi*sfinv)*fj;
}
}
// Define B depending on TransB and B_is_local
if (B_is_local)
{
for (j = 0; j <B_ncols ; j++)
for (i = 0; i<B_nrows; i++)
{
fi = i+1; // Get float version of i and j, offset by 1.
fj = j+1;
Bp[i + B_Stride*j] = 1.0/((fi*sfinv)*fj);
}
}
else
{
for (j = 0; j <B_ncols ; j++)
for (i = 0; i<B_nrows; i++)
{
fi = B_MyGlobalElements[i]+1; // Get float version of i and j, offset by 1.
fj = j+1;
Bp[i + B_Stride*j] = 1.0/((fi*sfinv)*fj);
}
}
// Define C_GEMM depending on A_is_local and B_is_local. C_GEMM is also a
// function of alpha, beta, TransA, TransB:
// C_GEMM = alpha*A(TransA)*B(TransB) + beta*C_GEMM
if (Case1)
{
for (j = 0; j <C_ncols ; j++)
for (i = 0; i<C_nrows; i++)
{
// Get float version of i and j, offset by 1.
fi = (i+1)*C_global_inner_dim;
fj = j+1;
C_GEMMp[i + C_GEMM_Stride*j] = alpha * (fi/fj)
+ beta * Cp[i + C_Stride*j];
}
}
else if (Case2)
{
for (j = 0; j <C_ncols ; j++)
for (i = 0; i<C_nrows; i++)
{
// Get float version of i and j, offset by 1.
fi = (i+1)*C_global_inner_dim;
fj = j+1;
C_GEMMp[i + C_GEMM_Stride*j] = alpha * (fi/fj)
+ beta * Cp[i + C_Stride*j];
}
}
else
{
for (j = 0; j <C_ncols ; j++)
for (i = 0; i<C_nrows; i++)
{
// Get float version of i and j.
fi = (A_MyGlobalElements[i]+1)*C_global_inner_dim;
fj = j+1;
C_GEMMp[i + C_GEMM_Stride*j] = alpha * (fi/fj)
+ beta * Cp[i + C_Stride*j];
}
}
delete A_Map;
delete B_Map;
delete [] A_MyGlobalElements;
delete [] B_MyGlobalElements;
return(0);
}
int Ifpack_PointRelaxation::
ApplyInverseGS_LocalFastCrsMatrix(const Epetra_CrsMatrix* A, const Epetra_MultiVector& X, Epetra_MultiVector& Y) const
{
int* IndexOffset;
int* Indices;
double* Values;
IFPACK_CHK_ERR(A->ExtractCrsDataPointers(IndexOffset, Indices, Values));
int NumVectors = X.NumVectors();
Teuchos::RefCountPtr< Epetra_MultiVector > Y2;
if (IsParallel_) {
Y2 = Teuchos::rcp( new Epetra_MultiVector(Importer_->TargetMap(), NumVectors) );
}
else
Y2 = Teuchos::rcp( &Y, false );
double** y_ptr, ** y2_ptr, ** x_ptr, *d_ptr;
X.ExtractView(&x_ptr);
Y.ExtractView(&y_ptr);
Y2->ExtractView(&y2_ptr);
Diagonal_->ExtractView(&d_ptr);
for (int iter = 0 ; iter < NumSweeps_ ; ++iter) {
// only one data exchange per sweep
if (IsParallel_)
IFPACK_CHK_ERR(Y2->Import(Y,*Importer_,Insert));
if(!DoBackwardGS_){
/* Forward Mode */
for (int ii = 0 ; ii < NumLocalSmoothingIndices_ ; ++ii) {
int i=LocalSmoothingIndices_[ii];
int col;
double diag = d_ptr[i];
for (int m = 0 ; m < NumVectors ; ++m) {
double dtemp = 0.0;
for (int k = IndexOffset[i] ; k < IndexOffset[i + 1] ; ++k) {
col = Indices[k];
dtemp += Values[k] * y2_ptr[m][col];
}
y2_ptr[m][i] += DampingFactor_ * (x_ptr[m][i] - dtemp) * diag;
}
}
}
else {
/* Backward Mode */
for (int ii = NumLocalSmoothingIndices_ - 1 ; ii > -1 ; --ii) {
int i=LocalSmoothingIndices_[ii];
int col;
double diag = d_ptr[i];
for (int m = 0 ; m < NumVectors ; ++m) {
double dtemp = 0.0;
for (int k = IndexOffset[i] ; k < IndexOffset[i + 1] ; ++k) {
col = Indices[k];
dtemp += Values[k] * y2_ptr[m][col];
}
y2_ptr[m][i] += DampingFactor_ * (x_ptr[m][i] - dtemp) * diag;
}
}
}
if (IsParallel_)
for (int m = 0 ; m < NumVectors ; ++m)
for (int ii = 0 ; ii < NumLocalSmoothingIndices_ ; ++ii) {
int i=LocalSmoothingIndices_[ii];
y_ptr[m][i] = y2_ptr[m][i];
}
}
#ifdef IFPACK_FLOPCOUNTERS
ApplyInverseFlops_ += NumVectors * (8 * NumGlobalRows_ + 4 * NumGlobalNonzeros_);
#endif
return(0);
} //ApplyInverseGS_LocalFastCrsMatrix()
int
LOCA::Epetra::AugmentedOp::ApplyInverse(const Epetra_MultiVector& Input,
Epetra_MultiVector& Result) const
{
// Get number of vectors
int n = Input.NumVectors();
// Check num vectors is correct
if (importedInput == NULL || n != importedInput->NumVectors())
globalData->locaErrorCheck->throwError(
"LOCA::Epetra::AugmentedOp::ApplyInverse()"
"Must call init() before ApplyInverse()!");
// Import parameter components
importedInput->Import(Input, *extendedImporter, Insert);
// get views
double **input_view;
double **result_view;
importedInput->ExtractView(&input_view);
Result.ExtractView(&result_view);
// get views of paramter components
// this is done by setting the pointer for each column to start at
// underlyingLength
double **input_view_y = new double*[n];
for (int i=0; i<n; i++)
input_view_y[i] = input_view[i]+underlyingLength;
// break input, result into components
Epetra_MultiVector input_x(View, underlyingMap, input_view, n);
Epetra_MultiVector input_y(View, localMap, input_view_y, n);
Epetra_MultiVector result_x(View, underlyingMap, result_view, n);
// copy input_y into result_y
result_y->Scale(1.0, input_y);
// Temporary epetra vectors
Epetra_MultiVector tmp2(c);
tmp->PutScalar(0.0);
// Solve J*result_x = input_x
jacOperator->ApplyInverse(input_x, result_x);
if (useTranspose) {
// Solve J*tmp = b
jacOperator->ApplyInverse(*b, *tmp);
// Compute input_y - a^T*result_x
result_y->Multiply('T', 'N', -1.0, *a, result_x, 1.0);
// Compute c - a^T*tmp
tmp2.Multiply('T', 'N', -1.0, *a, *tmp, 1.0);
}
else {
// Solve J*tmp = a
jacOperator->ApplyInverse(*a, *tmp);
// Compute input_y - b^T*result_x
result_y->Multiply('T', 'N', -1.0, *b, result_x, 1.0);
// Compute c - b^T*tmp1
tmp2.Multiply('T', 'N', -1.0, *b, *tmp, 1.0);
}
// Solve (c - z^T*tmp1)^-1 * (input_y - z^T*result_x) where z = a or b
int *ipiv = new int[numConstraints];
int info;
double *result_y_view;
int result_y_lda;
result_y->ExtractView(&result_y_view, &result_y_lda);
double *tmp2_view;
int tmp2_lda;
tmp2.ExtractView(&tmp2_view, &tmp2_lda);
dlapack.GESV(numConstraints, n, tmp2_view, tmp2_lda, ipiv, result_y_view,
result_y_lda, &info);
delete [] ipiv;
if (info != 0) {
globalData->locaErrorCheck->throwError(
"LOCA::Epetra::AugmentedOp::ApplyInverse()"
"Solve of dense matrix failed!");
}
// Compute result_x = result_x - tmp*result_y
result_x.Multiply('N', 'N', -1.0, *tmp, *result_y, 1.0);
// Set parameter component
if (haveParamComponent)
for (int j=0; j<n; j++)
for (int i=0; i<numConstraints; i++)
result_view[j][underlyingLength+i] = (*result_y)[j][i];
delete [] input_view_y;
return 0;
}
//==============================================================================
int Ifpack_PointRelaxation::
ApplyInverseSGS_RowMatrix(const Epetra_MultiVector& X, Epetra_MultiVector& Y) const
{
int NumVectors = X.NumVectors();
int Length = Matrix().MaxNumEntries();
std::vector<int> Indices(Length);
std::vector<double> Values(Length);
Teuchos::RefCountPtr< Epetra_MultiVector > Y2;
if (IsParallel_) {
Y2 = Teuchos::rcp( new Epetra_MultiVector(Importer_->TargetMap(), NumVectors) );
}
else
Y2 = Teuchos::rcp( &Y, false );
double** y_ptr, ** y2_ptr, ** x_ptr, *d_ptr;
X.ExtractView(&x_ptr);
Y.ExtractView(&y_ptr);
Y2->ExtractView(&y2_ptr);
Diagonal_->ExtractView(&d_ptr);
for (int iter = 0 ; iter < NumSweeps_ ; ++iter) {
// only one data exchange per sweep
if (IsParallel_)
IFPACK_CHK_ERR(Y2->Import(Y,*Importer_,Insert));
for (int ii = 0 ; ii < NumLocalSmoothingIndices_ ; ++ii) {
int i = (!LocalSmoothingIndices_)? ii : LocalSmoothingIndices_[ii];
int NumEntries;
int col;
double diag = d_ptr[i];
IFPACK_CHK_ERR(Matrix_->ExtractMyRowCopy(i, Length,NumEntries,
&Values[0], &Indices[0]));
for (int m = 0 ; m < NumVectors ; ++m) {
double dtemp = 0.0;
for (int k = 0 ; k < NumEntries ; ++k) {
col = Indices[k];
dtemp += Values[k] * y2_ptr[m][col];
}
y2_ptr[m][i] += DampingFactor_ * (x_ptr[m][i] - dtemp) * diag;
}
}
for (int ii = NumLocalSmoothingIndices_ - 1 ; ii > -1 ; --ii) {
int i = (!LocalSmoothingIndices_)? ii : LocalSmoothingIndices_[ii];
int NumEntries;
int col;
double diag = d_ptr[i];
IFPACK_CHK_ERR(Matrix_->ExtractMyRowCopy(i, Length,NumEntries,
&Values[0], &Indices[0]));
for (int m = 0 ; m < NumVectors ; ++m) {
double dtemp = 0.0;
for (int k = 0 ; k < NumEntries ; ++k) {
col = Indices[k];
dtemp += Values[k] * y2_ptr[m][col];
}
y2_ptr[m][i] += DampingFactor_ * (x_ptr[m][i] - dtemp) * diag;
}
}
if (IsParallel_)
for (int m = 0 ; m < NumVectors ; ++m)
for (int ii = 0 ; ii < NumLocalSmoothingIndices_ ; ++ii) {
int i = (!LocalSmoothingIndices_)? ii : LocalSmoothingIndices_[ii];
y_ptr[m][i] = y2_ptr[m][i];
}
}
#ifdef IFPACK_FLOPCOUNTERS
ApplyInverseFlops_ += NumVectors * (8 * NumGlobalRows_ + 4 * NumGlobalNonzeros_);
#endif
return(0);
}
// ================================================ ====== ==== ==== == =
// Apply the preconditioner w/ RHS B and get result X
int ML_Epetra::LevelWrap::ApplyInverse(const Epetra_MultiVector& B, Epetra_MultiVector& X_) const{
#ifdef ML_TIMING
double t_time,t_diff;
StartTimer(&t_time);
#endif
// Sanity Checks
if (!B.Map().SameAs(OperatorDomainMap())) return -1;
if (!X_.Map().SameAs(OperatorRangeMap())) return -1;
if (!X_.Map().SameAs(B.Map())) return -1;
if (B.NumVectors() != X_.NumVectors()) return -1;
// Build new work vector X
Epetra_MultiVector X(X_.Map(),X_.NumVectors(),true);
Epetra_MultiVector tmp0(X_.Map(),X_.NumVectors(),true);
Epetra_MultiVector tmp1(P0_->DomainMap(),X_.NumVectors(),true);
Epetra_MultiVector tmp2(P0_->DomainMap(),X_.NumVectors(),true);
double *Bptr, *Xptr;
int BLDA, XLDA;
// Pre Smoother
if(pre_or_post==ML_BOTH || pre_or_post==ML_PRESMOOTHER){
if (use_mlsmoother_) { // ML smoothing is experimental
B.ExtractView(&Bptr,&BLDA);
X.ExtractView(&Xptr,&XLDA);
ML_Smoother_Apply(&(ml_subproblem_->pre_smoother[0]),
X.MyLength(), Xptr,
X.MyLength(), Bptr, ML_ZERO);
}
else Smoother_->ApplyInverse(B,X);
A0_->Apply(X,tmp0);
tmp0.Update(1.0,B,-1.0);
}
else tmp0 = B;
// Form coarse residual
if(use_pt_) P0_->Multiply(true,tmp0,tmp1);
else R0_->Multiply(false,tmp0,tmp1);
// Solve coarse problem
A1prec_->ApplyInverse(tmp1,tmp2);
// Update solution
P0_->Multiply(false,tmp2,tmp0);
X.Update(1.0,tmp0,1.0);
// Post Smoother
if(pre_or_post==ML_BOTH || pre_or_post==ML_POSTSMOOTHER){
if (use_mlsmoother_) { // ML smoothing is experimental
Epetra_MultiVector tmp3(X_.Map(),X_.NumVectors(),true);
A0_->Apply(X,tmp0);
tmp0.Update(1.0,B,-1.0);
tmp3.PutScalar(0.0);
tmp0.ExtractView(&Bptr,&BLDA);
tmp3.ExtractView(&Xptr,&XLDA);
ML_Smoother_Apply(&(ml_subproblem_->pre_smoother[0]),X.MyLength(),Xptr,
X.MyLength(), Bptr, ML_ZERO);
X.Update(1.0,tmp3,1.0);
}
else Smoother_->ApplyInverse(B,X);
}
// Copy to output
X_=X;
#ifdef ML_TIMING
StopTimer(&t_time,&t_diff);
/* Output */
ML_Comm *comm_;
ML_Comm_Create(&comm_);
ApplicationTime_+= t_diff;
if(FirstApplication_){
FirstApplication_=false;
FirstApplicationTime_=ApplicationTime_;
}/*end if*/
ML_Comm_Destroy(&comm_);
#endif
return 0;
}
