void
Scaler::DownScaleBilinear(intType fromRow, int32 toRow)
{
BBitmap* src;
BBitmap* dest;
intType srcW, srcH;
intType destW, destH;
intType x, y;
const uchar* srcBits;
uchar* destBits;
intType srcBPR, destBPR;
const uchar* srcData;
uchar* destDataRow;
uchar* destData;
const int32 kBPP = 4;
DownScaleColumnData* columnData;
src = GetSrcImage();
dest = fScaledImage;
srcW = src->Bounds().IntegerWidth();
srcH = src->Bounds().IntegerHeight();
destW = dest->Bounds().IntegerWidth();
destH = dest->Bounds().IntegerHeight();
srcBits = (uchar*)src->Bits();
destBits = (uchar*)dest->Bits();
srcBPR = src->BytesPerRow();
destBPR = dest->BytesPerRow();
destDataRow = destBits + fromRow * destBPR;
const float deltaX = (srcW + 1.0) / (destW + 1.0);
const float deltaY = (srcH + 1.0) / (destH + 1.0);
const float deltaXY = deltaX * deltaY;
columnData = new DownScaleColumnData[destW+1];
DownScaleColumnData* cd = columnData;
for (x = 0; x <= destW; x ++, cd ++) {
const float fFromX = x * deltaX;
const float fToX = fFromX + deltaX;
cd->from = (intType)fFromX;
cd->to = (intType)fToX;
cd->alpha0 = 1.0 - (fFromX - cd->from);
cd->alpha1 = fToX - cd->to;
}
for (y = fromRow; IsRunning() && y <= toRow; y ++, destDataRow += destBPR) {
const float fFromY = y * deltaY;
const float fToY = fFromY + deltaY;
const intType fromY = (intType)fFromY;
const intType toY = (intType)fToY;
const float a0Y = 1.0 - (fFromY - fromY);
const float a1Y = fToY - toY;
const uchar* srcDataRow = srcBits + fromY * srcBPR;
destData = destDataRow;
cd = columnData;
for (x = 0; x <= destW; x ++, destData += kBPP, cd ++) {
const intType fromX = cd->from;
const intType toX = cd->to;
const float a0X = cd->alpha0;
const float a1X = cd->alpha1;
srcData = srcDataRow + fromX * kBPP;
float totalSum[3];
float sum[3];
RowValues(sum, srcData, srcW, fromX, toX, a0X, a1X, kBPP);
totalSum[0] = a0Y * sum[0];
totalSum[1] = a0Y * sum[1];
totalSum[2] = a0Y * sum[2];
srcData += srcBPR;
for (int32 r = fromY+1; r < toY; r ++, srcData += srcBPR) {
RowValues(sum, srcData, srcW, fromX, toX, a0X, a1X, kBPP);
totalSum[0] += sum[0];
totalSum[1] += sum[1];
totalSum[2] += sum[2];
}
if (toY <= srcH) {
RowValues(sum, srcData, srcW, fromX, toX, a0X, a1X, kBPP);
totalSum[0] += a1Y * sum[0];
totalSum[1] += a1Y * sum[1];
totalSum[2] += a1Y * sum[2];
}
destData[0] = static_cast<uchar>(totalSum[0] / deltaXY);
destData[1] = static_cast<uchar>(totalSum[1] / deltaXY);
destData[2] = static_cast<uchar>(totalSum[2] / deltaXY);
}
}
delete[] columnData;
}
//==========================================================================
int Ifpack_ICT::Compute()
{
if (!IsInitialized())
IFPACK_CHK_ERR(Initialize());
Time_.ResetStartTime();
IsComputed_ = false;
NumMyRows_ = A_.NumMyRows();
int Length = A_.MaxNumEntries();
vector<int> RowIndices(Length);
vector<double> RowValues(Length);
bool distributed = (Comm().NumProc() > 1)?true:false;
if (distributed)
{
SerialComm_ = Teuchos::rcp(new Epetra_SerialComm);
SerialMap_ = Teuchos::rcp(new Epetra_Map(NumMyRows_, 0, *SerialComm_));
assert (SerialComm_.get() != 0);
assert (SerialMap_.get() != 0);
}
else
SerialMap_ = Teuchos::rcp(const_cast<Epetra_Map*>(&A_.RowMatrixRowMap()), false);
int RowNnz;
#ifdef IFPACK_FLOPCOUNTERS
double flops = 0.0;
#endif
H_ = Teuchos::rcp(new Epetra_CrsMatrix(Copy,*SerialMap_,0));
if (H_.get() == 0)
IFPACK_CHK_ERR(-5); // memory allocation error
// get A(0,0) element and insert it (after sqrt)
IFPACK_CHK_ERR(A_.ExtractMyRowCopy(0,Length,RowNnz,
&RowValues[0],&RowIndices[0]));
// skip off-processor elements
if (distributed)
{
int count = 0;
for (int i = 0 ;i < RowNnz ; ++i)
{
if (RowIndices[i] < NumMyRows_){
RowIndices[count] = RowIndices[i];
RowValues[count] = RowValues[i];
++count;
}
else
continue;
}
RowNnz = count;
}
// modify diagonal
double diag_val = 0.0;
for (int i = 0 ;i < RowNnz ; ++i) {
if (RowIndices[i] == 0) {
double& v = RowValues[i];
diag_val = AbsoluteThreshold() * EPETRA_SGN(v) +
RelativeThreshold() * v;
break;
}
}
diag_val = sqrt(diag_val);
int diag_idx = 0;
EPETRA_CHK_ERR(H_->InsertGlobalValues(0,1,&diag_val, &diag_idx));
// The 10 is just a small constant to limit collisons as the actual keys
// we store are the indices and not integers
// [0..A_.MaxNumEntries()*LevelofFill()].
Ifpack_HashTable Hash( 10 * A_.MaxNumEntries() * LevelOfFill(), 1);
// start factorization for line 1
for (int row_i = 1 ; row_i < NumMyRows_ ; ++row_i) {
// get row `row_i' of the matrix
IFPACK_CHK_ERR(A_.ExtractMyRowCopy(row_i,Length,RowNnz,
&RowValues[0],&RowIndices[0]));
// skip off-processor elements
if (distributed)
{
int count = 0;
for (int i = 0 ;i < RowNnz ; ++i)
{
if (RowIndices[i] < NumMyRows_){
RowIndices[count] = RowIndices[i];
RowValues[count] = RowValues[i];
++count;
}
else
continue;
}
RowNnz = count;
}
// number of nonzeros in this row are defined as the nonzeros
// of the matrix, plus the level of fill
int LOF = (int)(LevelOfFill() * RowNnz);
if (LOF == 0) LOF = 1;
// convert line `row_i' into hash for fast access
Hash.reset();
double h_ii = 0.0;
for (int i = 0 ; i < RowNnz ; ++i) {
if (RowIndices[i] == row_i) {
double& v = RowValues[i];
h_ii = AbsoluteThreshold() * EPETRA_SGN(v) + RelativeThreshold() * v;
}
else if (RowIndices[i] < row_i)
{
Hash.set(RowIndices[i], RowValues[i], true);
}
}
// form element (row_i, col_j)
// I start from the first row that has a nonzero column
// index in row_i.
for (int col_j = RowIndices[0] ; col_j < row_i ; ++col_j) {
double h_ij = 0.0, h_jj = 0.0;
// note: get() returns 0.0 if col_j is not found
h_ij = Hash.get(col_j);
// get pointers to row `col_j'
int* ColIndices;
double* ColValues;
int ColNnz;
H_->ExtractGlobalRowView(col_j, ColNnz, ColValues, ColIndices);
for (int k = 0 ; k < ColNnz ; ++k) {
int col_k = ColIndices[k];
if (col_k == col_j)
h_jj = ColValues[k];
else {
double xxx = Hash.get(col_k);
if (xxx != 0.0)
{
h_ij -= ColValues[k] * xxx;
#ifdef IFPACK_FLOPCOUNTERS
flops += 2.0;
#endif
}
}
}
h_ij /= h_jj;
if (IFPACK_ABS(h_ij) > DropTolerance_)
{
Hash.set(col_j, h_ij);
}
#ifdef IFPACK_FLOPCOUNTERS
// only approx
ComputeFlops_ += 2.0 * flops + 1.0;
#endif
}
int size = Hash.getNumEntries();
vector<double> AbsRow(size);
int count = 0;
// +1 because I use the extra position for diagonal in insert
vector<int> keys(size + 1);
vector<double> values(size + 1);
Hash.arrayify(&keys[0], &values[0]);
for (int i = 0 ; i < size ; ++i)
{
AbsRow[i] = IFPACK_ABS(values[i]);
}
count = size;
double cutoff = 0.0;
if (count > LOF) {
nth_element(AbsRow.begin(), AbsRow.begin() + LOF, AbsRow.begin() + count,
std::greater<double>());
cutoff = AbsRow[LOF];
}
for (int i = 0 ; i < size ; ++i)
{
h_ii -= values[i] * values[i];
}
if (h_ii < 0.0) h_ii = 1e-12;;
h_ii = sqrt(h_ii);
#ifdef IFPACK_FLOPCOUNTERS
// only approx, + 1 == sqrt
ComputeFlops_ += 2 * size + 1;
#endif
double DiscardedElements = 0.0;
count = 0;
for (int i = 0 ; i < size ; ++i)
{
if (IFPACK_ABS(values[i]) > cutoff)
{
values[count] = values[i];
keys[count] = keys[i];
++count;
}
else
DiscardedElements += values[i];
}
if (RelaxValue() != 0.0) {
DiscardedElements *= RelaxValue();
h_ii += DiscardedElements;
}
values[count] = h_ii;
keys[count] = row_i;
++count;
H_->InsertGlobalValues(row_i, count, &(values[0]), (int*)&(keys[0]));
}
IFPACK_CHK_ERR(H_->FillComplete());
#if 0
// to check the complete factorization
Epetra_Vector LHS(Matrix().RowMatrixRowMap());
Epetra_Vector RHS1(Matrix().RowMatrixRowMap());
Epetra_Vector RHS2(Matrix().RowMatrixRowMap());
Epetra_Vector RHS3(Matrix().RowMatrixRowMap());
LHS.Random();
Matrix().Multiply(false,LHS,RHS1);
H_->Multiply(true,LHS,RHS2);
H_->Multiply(false,RHS2,RHS3);
RHS1.Update(-1.0, RHS3, 1.0);
cout << endl;
cout << RHS1;
#endif
int MyNonzeros = H_->NumGlobalNonzeros();
Comm().SumAll(&MyNonzeros, &GlobalNonzeros_, 1);
IsComputed_ = true;
#ifdef IFPACK_FLOPCOUNTERS
double TotalFlops; // sum across all the processors
A_.Comm().SumAll(&flops, &TotalFlops, 1);
ComputeFlops_ += TotalFlops;
#endif
++NumCompute_;
ComputeTime_ += Time_.ElapsedTime();
return(0);
}
