liz/rawTherapee
liz/rawTherapee
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Performance optimization for impulse_nr on multi-core-systems (Issue 1671)

Browse Source
This commit is contained in:
Ingo 2013-01-22 19:00:36 +01:00
parent ce43113575
commit 7b0cda5c61
1 changed files with 40 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

68
rtengine/impulse_denoise.h
View File

@ -5,7 +5,7 @@
* it under the terms of the GNU General Public License as published by * it under the terms of the GNU General Public License as published by
* the Free Software Foundation, either version 3 of the License, or * the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version. * (at your option) any later version.
* *
* RawTherapee is distributed in the hope that it will be useful, * RawTherapee is distributed in the hope that it will be useful,
* but widthITheightOUT ANY widthARRANTY; without even the implied warranty of * but widthITheightOUT ANY widthARRANTY; without even the implied warranty of
* MERCheightANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * MERCheightANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
@ -15,7 +15,7 @@
* along with RawTherapee. If not, see <http://www.gnu.org/licenses/>. * along with RawTherapee. If not, see <http://www.gnu.org/licenses/>.
* *
* 2010 Emil Martinec <ejmartin@uchicago.edu> * 2010 Emil Martinec <ejmartin@uchicago.edu>
* *
*/ */
#include <cstddef> #include <cstddef>
#include "rt_math.h" #include "rt_math.h"
@ -29,17 +29,17 @@ using namespace std;
namespace rtengine { namespace rtengine {
void ImProcFunctions::impulse_nr (LabImage* lab, double thresh) { void ImProcFunctions::impulse_nr (LabImage* lab, double thresh) {
// %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% // %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
// impulse noise removal // impulse noise removal
// local variables // local variables
int width = lab->W; int width = lab->W;
int height = lab->H; int height = lab->H;
float hpfabs, hfnbrave; float hpfabs, hfnbrave;
// buffer for the lowpass image // buffer for the lowpass image
float ** lpf = new float *[height]; float ** lpf = new float *[height];
// buffer for the highpass image // buffer for the highpass image
@ -51,45 +51,57 @@ void ImProcFunctions::impulse_nr (LabImage* lab, double thresh) {
//memset (impish[i], 0, width*sizeof(unsigned short)); //memset (impish[i], 0, width*sizeof(unsigned short));
} }
//The cleaning algorithm starts here //The cleaning algorithm starts here
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% //%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
// modified bilateral filter for lowpass image, omitting input pixel; or Gaussian blur // modified bilateral filter for lowpass image, omitting input pixel; or Gaussian blur
static float eps = 1.0; static float eps = 1.0;
float wtdsum[3], dirwt, norm; float wtdsum[3], dirwt, norm;
int i1, j1; int i1, j1;
//rangeblur<unsigned short, unsigned int> (lab->L, lpf, impish /*used as buffer here*/, width, height, thresh, false); //rangeblur<unsigned short, unsigned int> (lab->L, lpf, impish /*used as buffer here*/, width, height, thresh, false);
#ifdef _OPENMP #ifdef _OPENMP
#pragma omp parallel #pragma omp parallel
#endif #endif
{ {
AlignedBufferMP<double> buffer(max(width,height)); AlignedBufferMP<double> buffer(max(width,height));
gaussHorizontal<float> (lab->L, lpf, buffer, width, height, max(2.0,thresh-1.0)); gaussHorizontal<float> (lab->L, lpf, buffer, width, height, max(2.0,thresh-1.0));
gaussVertical<float> (lpf, lpf, buffer, width, height, max(2.0,thresh-1.0)); gaussVertical<float> (lpf, lpf, buffer, width, height, max(2.0,thresh-1.0));
} }
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% //%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
float impthr = max(1.0,5.5-thresh); float impthr = max(1.0,5.5-thresh);
float impthrDiv24 = impthr / 24.0f; //Issue 1671: moved the Division outside the loop, impthr can be optimized out too, but I let in the code at the moment
#ifdef _OPENMP
#pragma omp parallel for private(hpfabs, hfnbrave,i1,j1)
#endif
for (int i=0; i < height; i++) for (int i=0; i < height; i++)
for (int j=0; j < width; j++) { for (int j=0; j < width; j++) {
hpfabs = fabs(lab->L[i][j]-lpf[i][j]); hpfabs = fabs(lab->L[i][j]-lpf[i][j]);
//block average of high pass data //block average of high pass data
for (i1=max(0,i-2), hfnbrave=0; i1<=min(i+2,height-1); i1++ ) for (i1=max(0,i-2), hfnbrave=0; i1<=min(i+2,height-1); i1++ )
for (j1=max(0,j-2); j1<=min(j+2,width-1); j1++ ) { for (j1=max(0,j-2); j1<=min(j+2,width-1); j1++ ) {
hfnbrave += fabs(lab->L[i1][j1]-lpf[i1][j1]); hfnbrave += fabs(lab->L[i1][j1]-lpf[i1][j1]);
} }
hfnbrave = (hfnbrave-hpfabs)/24; impish[i][j] = (hpfabs>((hfnbrave-hpfabs)*impthrDiv24));
hpfabs>(hfnbrave*impthr) ? impish[i][j]=1 : impish[i][j]=0;
}//now impulsive values have been identified }//now impulsive values have been identified
// Issue 1671:
// often, noise isn't evenly distributed, e.g. only a few noisy pixels in the bright sky, but many in the dark foreground,
// so it's better to schedule dynamic and let every thread only process 16 rows, to avoid running big threads out of work
// Measured it and in fact gives better performance than without schedule(dynamic,16). Of course, there could be a better
// choice for the chunk_size than 16
// race conditions are avoided by the array impish
#ifdef _OPENMP
#pragma omp parallel for private(wtdsum,norm,dirwt,i1,j1) schedule(dynamic,16)
#endif
for (int i=0; i < height; i++) for (int i=0; i < height; i++)
for (int j=0; j < width; j++) { for (int j=0; j < width; j++) {
if (!impish[i][j]) continue; if (!impish[i][j]) continue;
@ -110,17 +122,17 @@ void ImProcFunctions::impulse_nr (LabImage* lab, double thresh) {
lab->L[i][j]=wtdsum[0]/norm;//low pass filter lab->L[i][j]=wtdsum[0]/norm;//low pass filter
lab->a[i][j]=wtdsum[1]/norm;//low pass filter lab->a[i][j]=wtdsum[1]/norm;//low pass filter
lab->b[i][j]=wtdsum[2]/norm;//low pass filter lab->b[i][j]=wtdsum[2]/norm;//low pass filter
} }
}//now impulsive values have been corrected }//now impulsive values have been corrected
for (int i=0; i<height; i++) { for (int i=0; i<height; i++) {
delete [] lpf[i]; delete [] lpf[i];
delete [] impish[i]; delete [] impish[i];
} }
delete [] lpf; delete [] lpf;
delete [] impish; delete [] impish;
} }
} }