void Lanczose_Scale(IplImage *in, IplImage *out, int owidth, int oheight)
{
int channels = in->nChannels;
int iwidth = in->width;
int iheight = in->height;
int i,j,k,x,y;
for(x=0;x<owidth;x++)
for(y=0;y<oheight;y++)
{
double x0 = floor(x*1.0*iwidth/owidth);
double y0 = floor(y*1.0*iheight/oheight);
CvScalar temp, temp2;
for(k=0;k<channels;k++)
temp2.val[k] = 0;
for(i=x0-a+1;i<=x0+a;i++)
for(j=y0-a+1;j<=y0+a;j++)
if(i>=0 && i<iwidth && j>=0 && j<iheight)
{
temp = cvGet2D(in,j,i);
for(k=0;k<channels;k++)
{
temp2.val[k] += temp.val[k]*Lanczos(x0-i)*Lanczos(y0-j);
}
}
cvSet2D(out,y,x,temp2);
}
}
void BaseArray::ResampleLanczos(int sizex, int sizey, int n){
double xratio, yratio, sum, weight, x, y, Lx, Ly;
std::pair<int, int> size = Size();
std::vector< std::vector<double> > tmp (sizex, std::vector<double>(sizey));
xratio = sizex/(double)size.first;
yratio = sizey/(double)size.second;
for (int i = 0; i < sizex; i++){
for (int j = 0; j < sizey; j++){
x = i/xratio;
y = j/yratio;
sum = 0;
weight = 0;
for (int r = -n+1; r < n; r++){
for (int s = -n+1; s < n; s++){
int cx = floor(x)+r, cy = floor(y)+s;
Lx = Lanczos(cx-x, n);
Ly = Lanczos(cy-y, n);
weight += Lx * Ly;
sum += At(cx, cy, true) * Lx * Ly;
}
}
sum = sum/weight;
sum = (sum < 0) ? 0 : sum;
sum = (sum > 255) ? 255 : sum;
tmp[i][j] = sum;
}
}
data = tmp;
}
pair<Base<F>,Base<F>>
HermitianHelper( const DistSparseMatrix<F>& A, Int basisSize )
{
typedef Base<F> Real;
Grid grid( A.Comm() );
DistMatrix<Real,STAR,STAR> T(grid);
Lanczos( A, T, basisSize );
const Int k = T.Height();
if( k == 0 )
return pair<Real,Real>(0,0);
auto d = GetDiagonal( T.Matrix() );
auto dSub = GetDiagonal( T.Matrix(), -1 );
Matrix<Real> w;
HermitianTridiagEig( d, dSub, w );
pair<Real,Real> extremal;
extremal.second = MaxNorm(w);
extremal.first = extremal.second;
for( Int i=0; i<k; ++i )
extremal.first = Min(extremal.first,Abs(w.Get(i,0)));
return extremal;
}
pair<Base<F>,Base<F>>
HermitianExtremalSingValEst( const SparseMatrix<F>& A, Int basisSize )
{
EL_DEBUG_CSE
typedef Base<F> Real;
Matrix<Real> T;
Lanczos( A, T, basisSize );
const Int k = T.Height();
if( k == 0 )
return pair<Real,Real>(0,0);
Matrix<Real> d, dSub;
d = GetDiagonal( T );
dSub = GetDiagonal( T, -1 );
Matrix<Real> w;
HermitianTridiagEig( d, dSub, w );
pair<Real,Real> extremal;
extremal.second = MaxNorm(w);
extremal.first = extremal.second;
for( Int i=0; i<k; ++i )
extremal.first = Min(extremal.first,Abs(w(i)));
return extremal;
}
/*------------------------------------------------------------------------------
* Constructor is used as the main driver
*----------------------------------------------------------------------------*/
Green::Green(const int ntm, const int sdim, const int niter, const double min, const double max,
const int ndos, const double eps, double **Hessian, const int itm, double **lpdos)
{
const double tpi = 8.*atan(1.);
natom = ntm; sysdim = sdim; nit = niter; epson = eps;
wmin = min*tpi; wmax = max*tpi; nw = ndos + (ndos+1)%2;
H = Hessian; iatom = itm;
ldos = lpdos;
memory = new Memory();
if (natom < 1 || iatom < 0 || iatom >= natom){
printf("\nError: Wrong number of total atoms or wrong index of interested atom!\n");
return;
}
ndim = natom * sysdim;
if (nit < 1){printf("\nError: Wrong input of maximum iterations!\n"); return;}
if (nit > ndim){printf("\nError: # Lanczos iterations is not expected to exceed the degree of freedom!\n"); return;}
if (nw < 1){printf("\nError: Wrong input of points in LDOS!\n"); return;}
// initialize variables and allocate local memories
dw = (wmax - wmin)/double(nw-1);
memory->create(alpha, sysdim,nit, "Green_Green:alpha");
memory->create(beta, sysdim,nit+1,"Green_Green:beta");
//memory->create(ldos, nw,sysdim, "Green_Green:ldos");
// use Lanczos algorithm to diagonalize the Hessian
Lanczos();
// Get the inverser of the treated hessian by continued fractional method
Recursion();
return;
}
void Img::ResampleLanczos(int sizex, int sizey, int n){
double xratio, yratio, sum, weight, x, y, Lx, Ly;
std::pair<int, int> size = SizeOriginal();
std::vector< std::vector<double> > tmp (sizex, std::vector<double>(sizey));
xratio = sizex/(double)size.first;
yratio = sizey/(double)size.second;
for (int i = 0; i < sizex; i++){
for (int j = 0; j < sizey; j++){
x = i/xratio;
y = j/yratio;
sum = 0;
weight = 0;
for (int r = -n+1; r < n; r++){
for (int s = -n+1; s < n; s++){
int cx = floor(x)+r;
int cy = floor(y)+s;
Lx = Lanczos(cx-x, n);
Ly = Lanczos(cy-y, n);
weight += (Lx * Ly);
sum += (AtOriginal(cx, cy, true) * Lx * Ly);
}
}
sum = sum/weight;
/*if (sum < 0){
x = floor(x);
y = floor(y);
sum = AtOriginal(x, y, true);
//sum = 0.25 * (AtOriginal(x-1, y-1, true) + AtOriginal(x-1, y+1, true) + AtOriginal(x+1, y-1, true) + AtOriginal(x+1, y+1, true));
} else if (sum > 255) {
x = floor(x);
y = floor(y);
sum = AtOriginal(x, y, true);
//sum = 0.25 * (AtOriginal(x-1, y-1, true) + AtOriginal(x-1, y+1, true) + AtOriginal(x+1, y-1, true) + AtOriginal(x+1, y+1, true));
}
*/
sum = (sum < 0) ? 0 : sum;
sum = (sum > 255) ? 255 : sum;
tmp[i][j] = sum;
}
}
data = tmp;
}
// Thread-safe routine
void PCL_FUNC PCL_InitializeLanczosLUT( float*& LUT, int n )
{
static Mutex mutex;
volatile AutoLock lock( mutex );
if ( LUT == 0 )
{
LUT = new float[ n*__PCL_LANCZOS_LUT_RESOLUTION + 1 ];
for ( int i = 0, k = 0; i < n; ++i )
for ( int j = 0; j < __PCL_LANCZOS_LUT_RESOLUTION; ++j, ++k )
LUT[k] = Lanczos( i + double( j )/__PCL_LANCZOS_LUT_RESOLUTION, n );
LUT[n*__PCL_LANCZOS_LUT_RESOLUTION] = 0;
}
}
const float* PCL_FUNC PCL_InitializeLanczosIntLUT( int n )
{
ValidateLanczosLUTOrder( n );
{
static Mutex mutex;
volatile AutoLock lock( mutex );
float*& LUT = PCL_Lanczos_Int_LUT[n-1];
if ( LUT == nullptr )
{
LUT = new float[ n*__PCL_LANCZOS_LUT_INT_RESOLUTION + 1 ];
for ( int i = 0, k = 0; i < n; ++i )
for ( int j = 0; j < __PCL_LANCZOS_LUT_INT_RESOLUTION; ++j, ++k )
LUT[k] = Lanczos( i + double( j )/__PCL_LANCZOS_LUT_INT_RESOLUTION, n );
LUT[n*__PCL_LANCZOS_LUT_INT_RESOLUTION] = 0;
}
return reinterpret_cast<const float*>( LUT );
}
}
const double** PCL_FUNC PCL_InitializeLanczosRealLUT( int n )
{
ValidateLanczosLUTOrder( n );
{
static Mutex mutex;
volatile AutoLock lock( mutex );
double**& LUT = PCL_Lanczos_LUT[n-1];
if ( LUT == nullptr )
{
LUT = new double*[ 2*n ];
for ( int j = -n + 1, k = 0; j <= n; ++j, ++k )
{
LUT[k] = new double[ __PCL_LANCZOS_LUT_REAL_RESOLUTION ];
for ( int i = 0; i < __PCL_LANCZOS_LUT_REAL_RESOLUTION; ++i )
LUT[k][i] = Lanczos( j - double( i )/__PCL_LANCZOS_LUT_REAL_RESOLUTION, n );
}
}
return const_cast<const double**>( reinterpret_cast<double**>( LUT ) );
}
}
static inline void Resample(SDL_Surface * src, SDL_Surface * dst, int filter)
{
const double blur = 1.0;
double factor = dst->w / (double)src->w;
double scale = std::min(factor, 1.0) / blur;
int FilterRadius = filter;
if (filter < 1 )
FilterRadius = 1;
if (filter > 3) //automatically determine fastest filter setting
{
FilterRadius = 3;
if (scale < 0.67) FilterRadius = 2;
if (scale <= 0.5) FilterRadius = 1;
}
double support = FilterRadius / scale;
std::vector<double> contribution_x(std::min((size_t)src->w, 5+(size_t)(2*support)));
/* 5 = room for rounding up in calculations of start, stop and support */
Uint32 ** temp = new Uint32 * [src->h]; //array of source->height * dest->width
for (int i = 0 ; i < src->h; i++)
temp[i] = new Uint32 [dst->w];
if (support <= 0.5) { support = 0.5 + 1E-12; scale = 1.0; }
for (int x = 0; x < dst->w; ++x)
{
double center = (x + 0.5) / factor;
size_t start = (size_t)std::max(center - support + 0.5, (double)0);
size_t stop = (size_t)std::min(center + support + 0.5, (double)src->w);
double density = 0.0;
size_t nmax = stop - start;
double s = start - center + 0.5;
double point[4] = {0,0,0,0};
Uint8 v;
double diff;
for (int y = 0; y < src->h; y++)
{
for (size_t n = 0; n < nmax; ++n)
{
if (y == 0)
{ //only come up with the contribution list once per column.
contribution_x[n] = Lanczos (s * scale, FilterRadius);
density += contribution_x[n];
s++;
}
//it MUST be a 32-bit surface for following code to work correctly
Uint8 * p = (Uint8 *)src->pixels + y * src->pitch + (start+n) * 4;
for (int c = 0; c < 4; c++)
point[c] += p[c] * contribution_x[n];
}
/* Normalize. Truncate to Uint8 values. Place in temp array*/
Uint8 * p = (Uint8 *)&temp[y][x];
for (size_t c = 0; c < 4; c++)
{
if (density != 0.0 && density != 1.0)
point[c] /= density;
if (point[c] < 0)
point[c] = 0;
if (point[c] > 255)
point[c] = 255;
v = (Uint8) point[c];
diff = point[c] - (double)v;
if (diff < 0)
diff = -diff;
if (diff >= 0.5)
v++;
p[c] = v;
point[c] = 0; //reset value for next loop
}
}
}
factor = dst->h / (double)src->h;
scale = std::min(factor, 1.0) / blur;
if (filter > 3) //automatically determine fastest filter setting
{
FilterRadius = 3;
if (scale < 0.67) FilterRadius = 2;
if (scale <= 0.5) FilterRadius = 1;
}
support = FilterRadius / scale;
std::vector<double> contribution_y(std::min((size_t)src->h, 5+(size_t)(2*support)));
if (support <= 0.5) { support = 0.5 + 1E-12; scale = 1.0; }
for (int y = 0; y<dst->h; ++y)
{
double center = (y + 0.5) / factor;
size_t start = (size_t)std::max(center - support + 0.5, (double)0);
size_t stop = (size_t)std::min(center + support + 0.5, (double)src->h);
double density = 0.0;
size_t nmax = stop-start;
double s = start - center+0.5;
double point[4] = {0,0,0,0};
Uint8 v;
double diff;
for (int x=0; x<dst->w; x++)
{
for (size_t n=0; n<nmax; ++n)
{
if (x == 0)
{
contribution_y[n] = Lanczos(s * scale, FilterRadius);
density += contribution_y[n];
s++;
}
Uint8 * p = (Uint8 *)&temp[start+n][x];
for (int c = 0; c < 4; c++)
point[c] += p[c] * contribution_y[n];
}
//destination must also be a 32 bit surface for this to work!
Uint8 * p = (Uint8 *)dst->pixels + y * dst->pitch + x * 4;
for (size_t c = 0; c < 4; c++)
{
if (density != 0.0 && density != 1.0)
point[c] /= density;
if (point[c] < 0)
point[c] = 0;
if (point[c] > 255)
point[c] = 255;
v = (Uint8) point[c];
diff = point[c] - (double)v;
if (diff < 0)
diff = -diff;
if (diff >= 0.5)
v++;
p[c] = v;
point[c] = 0;
}
}
}
//free the temp array, so we don't leak any memory
for (int i = 0 ; i < src->h; i++)
delete [] temp[i];
delete [] temp;
}