OpenMP for faster implementation of pyramid schemes ;-)

This commit is contained in:
Emil Martinec
2010-09-22 22:43:39 -05:00
parent 539c39a92b
commit bde7141939
4 changed files with 102 additions and 59 deletions

View File

@@ -269,9 +269,7 @@ namespace rtengine {
int width = data_fine->W;
int height = data_fine->H;
float dirwt_l, dirwt_ab, norm_l, norm_ab;
//float lops,aops,bops;
float Lout, aout, bout;
//generate domain kernel
@@ -286,7 +284,11 @@ namespace rtengine {
}*/
//float domker[5][5] = {{1,1,1,1,1},{1,2,2,2,1},{1,2,4,2,1},{1,2,2,2,1},{1,1,1,1,1}};
for(int i = 0, i1=0; i < height; i+=pitch, i1++) {
#ifdef _OPENMP
#pragma omp parallel for
#endif
for(int i = 0; i < height; i+=pitch ) { int i1=i/pitch;
for(int j = 0, j1=0; j < width; j+=pitch, j1++)
{
//norm = DIRWT(i, j, i, j);
@@ -294,6 +296,9 @@ namespace rtengine {
//aout = -norm*data_fine->a[i][j];
//bout = -norm*data_fine->b[i][j];
//or
float dirwt_l, dirwt_ab, norm_l, norm_ab;
//float lops,aops,bops;
float Lout, aout, bout;
norm_l = norm_ab = 0;//if we do want to include the input pixel in the sum
Lout = 0;
aout = 0;
@@ -339,9 +344,6 @@ namespace rtengine {
int height = data_fine->H;
//float eps = 0.0;
double wtdsum[3], norm;
float hipass[3], hpffluct[3], tonefactor, nrfactor;
int i, j, ix, jx;
// c[0] noise_L
// c[1] noise_ab (relative to noise_L)
@@ -387,9 +389,14 @@ namespace rtengine {
if (pitch==1) {
// step (1-2-3-4)
for( i = 0; i < height; i++)
for( j = 0; j < width; j++) {
#ifdef _OPENMP
#pragma omp parallel for
#endif
for(int i = 0; i < height; i++)
for(int j = 0; j < width; j++) {
double wtdsum[3], norm;
float hipass[3], hpffluct[3], tonefactor, nrfactor;
tonefactor = ((NRWT_L(data_coarse->L[i][j])));
//Wiener filter
@@ -420,25 +427,37 @@ namespace rtengine {
LabImage* smooth;
smooth = new LabImage(width, height);
for( i = 0, ix=0; i < height; i+=pitch, ix++)
for( j = 0, jx=0; j < width; j+=pitch, jx++) {
#ifdef _OPENMP
#pragma omp parallel
#endif
{
#ifdef _OPENMP
#pragma omp for
#endif
for(int i = 0; i < height; i+=pitch)
{
int ix=i/pitch;
for(int j = 0, jx=0; j < width; j+=pitch, jx++) {
//copy common pixels
smooth->L[i][j] = data_coarse->L[ix][jx];
smooth->a[i][j] = data_coarse->a[ix][jx];
smooth->b[i][j] = data_coarse->b[ix][jx];
}
}
//if (pitch>1) {//pitch=2; step (1) expand coarse image, fill in missing data
for( i = 0; i < height-1; i+=2)
for( j = 0; j < width-1; j+=2) {
#ifdef _OPENMP
#pragma omp for
#endif
for(int i = 0; i < height-1; i+=2)
for(int j = 0; j < width-1; j+=2) {
//do midpoint first
norm=0;
wtdsum[0]=wtdsum[1]=wtdsum[2]=0.0;
for( ix=i; ix<MIN(height,i+3); ix+=2)
for ( jx=j; jx<MIN(width,j+3); jx+=2) {
double norm=0.0,wtdsum[3]={0.0,0.0,0.0};
//wtdsum[0]=wtdsum[1]=wtdsum[2]=0.0;
for(int ix=i; ix<MIN(height,i+3); ix+=2)
for (int jx=j; jx<MIN(width,j+3); jx+=2) {
wtdsum[0] += smooth->L[ix][jx];
wtdsum[1] += smooth->a[ix][jx];
wtdsum[2] += smooth->b[ix][jx];
@@ -449,20 +468,23 @@ namespace rtengine {
smooth->a[i+1][j+1]=wtdsum[1]*norm;
smooth->b[i+1][j+1]=wtdsum[2]*norm;
}
#ifdef _OPENMP
#pragma omp for
#endif
for( i = 0; i < height-1; i+=2)
for( j = 0; j < width-1; j+=2) {
for(int i = 0; i < height-1; i+=2)
for(int j = 0; j < width-1; j+=2) {
//now right neighbor
if (j+1==width) continue;
norm=0;
wtdsum[0]=wtdsum[1]=wtdsum[2]=0.0;
for (jx=j; jx<MIN(width,j+3); jx+=2) {
double norm=0.0,wtdsum[3]={0.0,0.0,0.0};
for (int jx=j; jx<MIN(width,j+3); jx+=2) {
wtdsum[0] += smooth->L[i][jx];
wtdsum[1] += smooth->a[i][jx];
wtdsum[2] += smooth->b[i][jx];
norm++;
}
for (ix=MAX(0,i-1); ix<MIN(height,i+2); ix+=2) {
for (int ix=MAX(0,i-1); ix<MIN(height,i+2); ix+=2) {
wtdsum[0] += smooth->L[ix][j+1];
wtdsum[1] += smooth->a[ix][j+1];
wtdsum[2] += smooth->b[ix][j+1];
@@ -475,15 +497,14 @@ namespace rtengine {
//now down neighbor
if (i+1==height) continue;
norm=0;
wtdsum[0]=wtdsum[1]=wtdsum[2]=0.0;
for (ix=i; ix<MIN(height,i+3); ix+=2) {
norm=0.0;wtdsum[0]=wtdsum[1]=wtdsum[2]=0.0;
for (int ix=i; ix<MIN(height,i+3); ix+=2) {
wtdsum[0] += smooth->L[ix][j];
wtdsum[1] += smooth->a[ix][j];
wtdsum[2] += smooth->b[ix][j];
norm++;
}
for (jx=MAX(0,j-1); jx<MIN(width,j+2); jx+=2) {
for (int jx=MAX(0,j-1); jx<MIN(width,j+2); jx+=2) {
wtdsum[0] += smooth->L[i+1][jx];
wtdsum[1] += smooth->a[i+1][jx];
wtdsum[2] += smooth->b[i+1][jx];
@@ -496,13 +517,17 @@ namespace rtengine {
}
#ifdef _OPENMP
#pragma omp for
#endif
// step (2-3-4)
for( i = 0; i < height; i++)
for( j = 0; j < width; j++) {
tonefactor = ((NRWT_L(smooth->L[i][j])));
for( int i = 0; i < height; i++)
for(int j = 0; j < width; j++) {
double tonefactor = ((NRWT_L(smooth->L[i][j])));
//double wtdsum[3], norm;
float hipass[3], hpffluct[3], nrfactor;
//Wiener filter
//luma
if (level<2) {
@@ -525,7 +550,7 @@ namespace rtengine {
data_fine->a[i][j] = hipass[1]+smooth->a[i][j];
data_fine->b[i][j] = hipass[2]+smooth->b[i][j];
}
} // end parallel
delete smooth;
}//end of pitch>1

View File

@@ -288,8 +288,7 @@ namespace rtengine {
int width = data_fine->W;
int height = data_fine->H;
float Lout, aout, bout;
float dirwt, norm;
//generate domain kernel
int halfwin = 1;//MIN(ceil(2*sig),3);
@@ -304,11 +303,14 @@ namespace rtengine {
//float domker[5][5] = {{1,1,1,1,1},{1,2,2,2,1},{1,2,4,2,1},{1,2,2,2,1},{1,1,1,1,1}};
//float domker[3][3] = {{1,1,1},{1,2,1},{1,1,1}};
for(int i = 0, i1=0; i < height; i+=pitch, i1++) {
#ifdef _OPENMP
#pragma omp parallel for
#endif
for(int i = 0; i < height; i+=pitch) { int i1=i/pitch;
for(int j = 0, j1=0; j < width; j+=pitch, j1++)
{
float Lout, aout, bout;
float norm;
norm = 0;//if we do want to include the input pixel in the sum
Lout = 0;
aout = 0;
@@ -316,7 +318,7 @@ namespace rtengine {
for(int inbr=MAX(0,i-scalewin); inbr<=MIN(height-1,i+scalewin); inbr+=scale) {
for (int jnbr=MAX(0,j-scalewin); jnbr<=MIN(width-1,j+scalewin); jnbr+=scale) {
dirwt = DIRWT(inbr, jnbr, i, j);
float dirwt = DIRWT(inbr, jnbr, i, j);
Lout += dirwt*data_fine->L[inbr][jnbr];
aout += dirwt*data_fine->a[inbr][jnbr];
bout += dirwt*data_fine->b[inbr][jnbr];
@@ -382,18 +384,21 @@ namespace rtengine {
if (pitch==1) {
// step (1-2-3-4)
#ifdef _OPENMP
#pragma omp parallel for
#endif
for(int i = 0; i < height; i++)
for(int j = 0; j < width; j++) {
//luma
hipass[0] = (float)data_fine->L[i][j]-data_coarse->L[i][j];
buffer[0][i*scale][j*scale] += hipass[0] * lumamult[level];//*luma;
float hipass0 = (float)data_fine->L[i][j]-data_coarse->L[i][j];
buffer[0][i*scale][j*scale] += hipass0 * lumamult[level];//*luma;
//chroma
hipass[1] = data_fine->a[i][j]-data_coarse->a[i][j];
hipass[2] = data_fine->b[i][j]-data_coarse->b[i][j];
buffer[1][i*scale][j*scale] += hipass[1] * chromamult[level]; //*chroma;
buffer[2][i*scale][j*scale] += hipass[2] * chromamult[level]; //*chroma;
float hipass1 = data_fine->a[i][j]-data_coarse->a[i][j];
float hipass2 = data_fine->b[i][j]-data_coarse->b[i][j];
buffer[1][i*scale][j*scale] += hipass1 * chromamult[level]; //*chroma;
buffer[2][i*scale][j*scale] += hipass2 * chromamult[level]; //*chroma;
}
} else {
@@ -402,10 +407,16 @@ namespace rtengine {
//if (pitch>1), pitch=2; expand coarse image, fill in missing data
LabImage* smooth;
smooth = new LabImage(width, height);
for(int i = 0, i2=0; i < height; i+=pitch, i2++)
#ifdef _OPENMP
#pragma omp parallel
#endif
{
#ifdef _OPENMP
#pragma omp for
#endif
for(int i = 0; i < height; i+=pitch){ int i2=i/pitch;
for(int j = 0, j2=0; j < width; j+=pitch, j2++) {
//copy common pixels
@@ -413,9 +424,11 @@ namespace rtengine {
smooth->a[i][j] = data_coarse->a[i2][j2];
smooth->b[i][j] = data_coarse->b[i2][j2];
}
//}
}
#ifdef _OPENMP
#pragma omp for
#endif
for(int i = 0; i < height-1; i+=2)
for(int j = 0; j < width-1; j+=2) {
//do midpoint first
@@ -440,7 +453,9 @@ namespace rtengine {
buffer[1][(i+1)*scale][(j+1)*scale]=wtdsum[4]*norm;
buffer[2][(i+1)*scale][(j+1)*scale]=wtdsum[5]*norm;
}
#ifdef _OPENMP
#pragma omp for
#endif
for(int i = 0; i < height-1; i+=2)
for(int j = 0; j < width-1; j+=2) {
//now right neighbor
@@ -511,6 +526,9 @@ namespace rtengine {
// step (2-3-4)
#ifdef _OPENMP
#pragma omp for
#endif
for(int i = 0; i < height; i++)
for(int j = 0; j < width; j++) {
@@ -524,7 +542,7 @@ namespace rtengine {
buffer[1][i*scale][j*scale] += hipass[1] * chromamult[level]; //*chroma;
buffer[2][i*scale][j*scale] += hipass[2] * chromamult[level]; //*chroma;
}
} // end parallel
delete smooth;
}

View File

@@ -350,7 +350,7 @@ int ProcParams::save (Glib::ustring fname) const {
{
std::stringstream ss;
ss << "Mult" << i;
keyFile.set_integer("Directional Pyramid Equalizer", ss.str(), dirpyrequalizer.mult[i]);
keyFile.set_double("Directional Pyramid Equalizer", ss.str(), dirpyrequalizer.mult[i]);
}
// save exif change list
@@ -611,7 +611,7 @@ if (keyFile.has_group ("Directional Pyramid Equalizer")) {
{
std::stringstream ss;
ss << "Mult" << i;
if(keyFile.has_key ("Directional Pyramid Equalizer", ss.str())) dirpyrequalizer.mult[i] = keyFile.get_integer ("Directional Pyramid Equalizer", ss.str());
if(keyFile.has_key ("Directional Pyramid Equalizer", ss.str())) dirpyrequalizer.mult[i] = keyFile.get_double ("Directional Pyramid Equalizer", ss.str());
}
}

View File

@@ -146,7 +146,7 @@ void DirPyrEqualizer::write (ProcParams* pp, ParamsEdited* pedited) {
pp->dirpyrequalizer.enabled = enabled->get_active ();
for (int i = 0; i < 8; i++) {
pp->dirpyrequalizer.mult[i] = (int) multiplier[i]->getValue();
pp->dirpyrequalizer.mult[i] = multiplier[i]->getValue();
}
if (pedited) {