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

Speedup for exp-loop

Browse Source
This commit is contained in:
heckflosse
2017-11-05 16:47:22 +01:00
parent 1e0c5bfb5e
commit facb37be91
1 changed files with 30 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
rtengine/tmo_fattal02.cc
View File

@@ -215,6 +215,7 @@ void gaussianBlur(const Array2Df& I, Array2Df& L)
void createGaussianPyramids( Array2Df* H, Array2Df** pyramids, int nlevels) void createGaussianPyramids( Array2Df* H, Array2Df** pyramids, int nlevels)
{ {
BENCHFUN
int width = H->getCols(); int width = H->getCols();
int height = H->getRows(); int height = H->getRows();
const int size = width*height; const int size = width*height;
@@ -256,6 +257,7 @@ void createGaussianPyramids( Array2Df* H, Array2Df** pyramids, int nlevels)
float calculateGradients(Array2Df* H, Array2Df* G, int k) float calculateGradients(Array2Df* H, Array2Df* G, int k)
{ {
BENCHFUN
const int width = H->getCols(); const int width = H->getCols();
const int height = H->getRows(); const int height = H->getRows();
const float divider = pow( 2.0f, k+1 ); const float divider = pow( 2.0f, k+1 );
@@ -331,6 +333,7 @@ void calculateFiMatrix(Array2Df* FI, Array2Df* gradients[],
float avgGrad[], int nlevels, int detail_level, float avgGrad[], int nlevels, int detail_level,
float alfa, float beta, float noise) float alfa, float beta, float noise)
{ {
BENCHFUN
const bool newfattal = true; const bool newfattal = true;
int width = gradients[nlevels-1]->getCols(); int width = gradients[nlevels-1]->getCols();
int height = gradients[nlevels-1]->getRows(); int height = gradients[nlevels-1]->getRows();
@@ -346,6 +349,7 @@ void calculateFiMatrix(Array2Df* FI, Array2Df* gradients[],
} }
} }
StopWatch Stop1("test");
for ( int k = nlevels-1; k >= 0 ; k-- ) for ( int k = nlevels-1; k >= 0 ; k-- )
{ {
width = gradients[k]->getCols(); width = gradients[k]->getCols();
@@ -375,7 +379,6 @@ void calculateFiMatrix(Array2Df* FI, Array2Df* gradients[],
} }
} }
// create next level // create next level
if ( k>1 ) if ( k>1 )
{ {
@@ -392,6 +395,7 @@ void calculateFiMatrix(Array2Df* FI, Array2Df* gradients[],
gaussianBlur(*fi[k-1], *fi[k-1]); gaussianBlur(*fi[k-1], *fi[k-1]);
} }
} }
Stop1.stop();
for ( int k=1 ; k<nlevels ; k++ ) for ( int k=1 ; k<nlevels ; k++ )
{ {
@@ -686,10 +690,33 @@ void tmo_fattal02(size_t width,
// return; // return;
// } // }
for ( size_t idx = 0 ; idx < height*width; ++idx ) StopWatch Stope("expf");
#pragma omp parallel
{ {
L(idx) = expf( gamma * U(idx) ); #ifdef __SSE2__
vfloat gammav = F2V(gamma);
#endif
#pragma omp for schedule(dynamic,16)
for ( size_t i=0 ; i<height ; i++ ) {
size_t j = 0;
#ifdef __SSE2__
for(; j < width - 3; j+=4)
{
STVFU(L[i][j], xexpf(gammav * LVFU(U[i][j])));
} }
#endif
for(; j < width; j++)
{
L[i][j] = xexpf( gamma * U[i][j]);
}
}
}
// for ( size_t idx = 0 ; idx < height*width; ++idx )
// {
// L(idx) = xexpf( gamma * U(idx) );
// }
Stope.stop();
} }
// ph.setValue(95); // ph.setValue(95);