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

Speedup for green equilibration

Browse Source
This commit is contained in:
heckflosse 2017-08-22 01:21:09 +02:00
parent ad20c39907
commit f409f1333e
2 changed files with 136 additions and 98 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

183
rtengine/green_equil_RT.cc
View File

@ -30,6 +30,8 @@
#include "rt_math.h"
#include "rawimagesource.h"
#include "opthelper.h"
namespace rtengine
{
@ -42,15 +44,25 @@ void RawImageSource::green_equilibrate(float thresh, array2D<float> &rawData)
int height = H, width = W;
// local variables
float** rawptr = rawData;
array2D<float> cfa (width, height, rawptr);
//array2D<int> checker (width,height,ARRAY2D_CLEAR_DATA);
array2D<float> cfa(width / 2 + (width & 1), height);
#ifdef _OPENMP
#pragma omp parallel for schedule(dynamic,16)
#endif
for(int i = 0; i < height; ++i) {
int j = (FC(i,0) & 1) ^ 1;
#ifdef __SSE2__
for(; j < width - 7; j += 8) {
STVFU(cfa[i][j>>1], LC2VFU(rawData[i][j]));
}
#endif
for(; j < width; j += 2) {
cfa[i][j>>1] = rawData[i][j];
}
}
//int verbose=1;
static const float eps = 1.0; //tolerance to avoid dividing by zero
constexpr float eps = 1.f; //tolerance to avoid dividing by zero
const float thresh6 = 6 * thresh;
// %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
// Fill G interpolated values with border interpolation and input values
@ -59,93 +71,116 @@ void RawImageSource::green_equilibrate(float thresh, array2D<float> &rawData)
//int counter, vtest;
//The green equilibration algorithm starts here
/*
#ifdef _OPENMP
#pragma omp parallel for
#endif
for (int rr=1; rr < height-1; rr++)
for (int cc=3-(FC(rr,2)&1); cc < width-2; cc+=2) {
float pcorr = (cfa[rr+1][cc+1]-cfa[rr][cc])*(cfa[rr-1][cc-1]-cfa[rr][cc]);
float mcorr = (cfa[rr-1][cc+1]-cfa[rr][cc])*(cfa[rr+1][cc-1]-cfa[rr][cc]);
if (pcorr>0 && mcorr>0) {checker[rr][cc]=1;} else {checker[rr][cc]=0;}
checker[rr][cc]=1;//test what happens if we always interpolate
}
counter=vtest=0;
*/
//now smooth the cfa data
#ifdef _OPENMP
#pragma omp parallel for schedule(dynamic,16)
#pragma omp parallel
#endif
{
#ifdef __SSE2__
vfloat zd5v = F2V(0.5f);
vfloat onev = F2V(1.f);
vfloat threshv = F2V(thresh);
vfloat thresh6v = F2V(thresh6);
vfloat epsv = F2V(eps);
#endif
#ifdef _OPENMP
#pragma omp for schedule(dynamic,16)
#endif
for (int rr = 4; rr < height - 4; rr++)
for (int cc = 5 - (FC(rr, 2) & 1); cc < width - 6; cc += 2) {
//if (checker[rr][cc]) {
//%%%%%%%%%%%%%%%%%%%%%%
//neighbor checking code from Manuel Llorens Garcia
float o1_1 = cfa[(rr - 1)][cc - 1];
float o1_2 = cfa[(rr - 1)][cc + 1];
float o1_3 = cfa[(rr + 1)][cc - 1];
float o1_4 = cfa[(rr + 1)][cc + 1];
float o2_1 = cfa[(rr - 2)][cc];
float o2_2 = cfa[(rr + 2)][cc];
float o2_3 = cfa[(rr)][cc - 2];
float o2_4 = cfa[(rr)][cc + 2];
for (int rr = 4; rr < height - 4; rr++) {
int cc = 5 - (FC(rr, 2) & 1);
#ifdef __SSE2__
for (; cc < width - 12; cc += 8) {
//neighbour checking code from Manuel Llorens Garcia
vfloat o1_1 = LVFU(cfa[rr - 1][(cc - 1)>>1]);
vfloat o1_2 = LVFU(cfa[rr - 1][(cc + 1)>>1]);
vfloat o1_3 = LVFU(cfa[rr + 1][(cc - 1)>>1]);
vfloat o1_4 = LVFU(cfa[rr + 1][(cc + 1)>>1]);
vfloat o2_1 = LVFU(cfa[rr - 2][cc>>1]);
vfloat o2_2 = LVFU(cfa[rr + 2][cc>>1]);
vfloat o2_3 = LVFU(cfa[rr][(cc - 2)>>1]);
vfloat o2_4 = LVFU(cfa[rr][(cc + 2)>>1]);
float d1 = (o1_1 + o1_2 + o1_3 + o1_4) * 0.25f;
float d2 = (o2_1 + o2_2 + o2_3 + o2_4) * 0.25f;
vfloat d1 = (o1_1 + o1_2 + o1_3 + o1_4);
vfloat d2 = (o2_1 + o2_2 + o2_3 + o2_4);
float c1 = (fabs(o1_1 - o1_2) + fabs(o1_1 - o1_3) + fabs(o1_1 - o1_4) + fabs(o1_2 - o1_3) + fabs(o1_3 - o1_4) + fabs(o1_2 - o1_4)) / 6.f;
float c2 = (fabs(o2_1 - o2_2) + fabs(o2_1 - o2_3) + fabs(o2_1 - o2_4) + fabs(o2_2 - o2_3) + fabs(o2_3 - o2_4) + fabs(o2_2 - o2_4)) / 6.f;
//%%%%%%%%%%%%%%%%%%%%%%
vfloat c1 = (vabsf(o1_1 - o1_2) + vabsf(o1_1 - o1_3) + vabsf(o1_1 - o1_4) + vabsf(o1_2 - o1_3) + vabsf(o1_3 - o1_4) + vabsf(o1_2 - o1_4));
vfloat c2 = (vabsf(o2_1 - o2_2) + vabsf(o2_1 - o2_3) + vabsf(o2_1 - o2_4) + vabsf(o2_2 - o2_3) + vabsf(o2_3 - o2_4) + vabsf(o2_2 - o2_4));
//vote1=(checker[rr-2][cc]+checker[rr][cc-2]+checker[rr][cc+2]+checker[rr+2][cc]);
//vote2=(checker[rr+1][cc-1]+checker[rr+1][cc+1]+checker[rr-1][cc-1]+checker[rr-1][cc+1]);
//if ((vote1==0 || vote2==0) && (c1+c2)<2*thresh*fabs(d1-d2)) vtest++;
//if (vote1>0 && vote2>0 && (c1+c2)<4*thresh*fabs(d1-d2)) {
if ((c1 + c2) < 4 * thresh * fabs(d1 - d2)) {
vmask mask1 = vmaskf_lt(c1 + c2, thresh6v * vabsf(d1 - d2));
if(_mm_movemask_ps((vfloat)mask1)) { // if for any of the 4 pixels the condition is true, do the maths for all 4 pixels and mask the unused out at the end
//pixel interpolation
float gin = cfa[rr][cc];
vfloat gin = LVFU(cfa[rr][cc>>1]);
float gse = (cfa[rr + 1][cc + 1]) + 0.5f * (cfa[rr][cc] - cfa[rr + 2][cc + 2]);
float gnw = (cfa[rr - 1][cc - 1]) + 0.5f * (cfa[rr][cc] - cfa[rr - 2][cc - 2]);
float gne = (cfa[rr - 1][cc + 1]) + 0.5f * (cfa[rr][cc] - cfa[rr - 2][cc + 2]);
float gsw = (cfa[rr + 1][cc - 1]) + 0.5f * (cfa[rr][cc] - cfa[rr + 2][cc - 2]);
vfloat gmp2p2 = gin - LVFU(cfa[rr + 2][(cc + 2)>>1]);
vfloat gmm2m2 = gin - LVFU(cfa[rr - 2][(cc - 2)>>1]);
vfloat gmm2p2 = gin - LVFU(cfa[rr - 2][(cc + 2)>>1]);
vfloat gmp2m2 = gin - LVFU(cfa[rr + 2][(cc - 2)>>1]);
vfloat gse = o1_4 + zd5v * gmp2p2;
vfloat gnw = o1_1 + zd5v * gmm2m2;
vfloat gne = o1_2 + zd5v * gmm2p2;
vfloat gsw = o1_3 + zd5v * gmp2m2;
vfloat wtse = onev / (epsv + SQRV(gmp2p2) + SQRV(LVFU(cfa[rr + 3][(cc + 3)>>1]) - o1_4));
vfloat wtnw = onev / (epsv + SQRV(gmm2m2) + SQRV(LVFU(cfa[rr - 3][(cc - 3)>>1]) - o1_1));
vfloat wtne = onev / (epsv + SQRV(gmm2p2) + SQRV(LVFU(cfa[rr - 3][(cc + 3)>>1]) - o1_2));
vfloat wtsw = onev / (epsv + SQRV(gmp2m2) + SQRV(LVFU(cfa[rr + 3][(cc - 3)>>1]) - o1_3));
float wtse = 1.0f / (eps + SQR(cfa[rr + 2][cc + 2] - cfa[rr][cc]) + SQR(cfa[rr + 3][cc + 3] - cfa[rr + 1][cc + 1]));
float wtnw = 1.0f / (eps + SQR(cfa[rr - 2][cc - 2] - cfa[rr][cc]) + SQR(cfa[rr - 3][cc - 3] - cfa[rr - 1][cc - 1]));
float wtne = 1.0f / (eps + SQR(cfa[rr - 2][cc + 2] - cfa[rr][cc]) + SQR(cfa[rr - 3][cc + 3] - cfa[rr - 1][cc + 1]));
float wtsw = 1.0f / (eps + SQR(cfa[rr + 2][cc - 2] - cfa[rr][cc]) + SQR(cfa[rr + 3][cc - 3] - cfa[rr + 1][cc - 1]));
vfloat ginterp = (gse * wtse + gnw * wtnw + gne * wtne + gsw * wtsw) / (wtse + wtnw + wtne + wtsw);
vfloat val = vself(vmaskf_lt(ginterp - gin, threshv * (ginterp + gin)), zd5v * (ginterp + gin), gin);
val = vself(mask1, val, gin);
STC2VFU(rawData[rr][cc], val);
}
}
#endif
for (; cc < width - 6; cc += 2) {
//neighbour checking code from Manuel Llorens Garcia
float o1_1 = cfa[rr - 1][(cc - 1)>>1];
float o1_2 = cfa[rr - 1][(cc + 1)>>1];
float o1_3 = cfa[rr + 1][(cc - 1)>>1];
float o1_4 = cfa[rr + 1][(cc + 1)>>1];
float o2_1 = cfa[rr - 2][cc>>1];
float o2_2 = cfa[rr + 2][cc>>1];
float o2_3 = cfa[rr][(cc - 2)>>1];
float o2_4 = cfa[rr][(cc + 2)>>1];
float d1 = (o1_1 + o1_2) + (o1_3 + o1_4);
float d2 = (o2_1 + o2_2) + (o2_3 + o2_4);
float c1 = (fabs(o1_1 - o1_2) + fabs(o1_1 - o1_3) + fabs(o1_1 - o1_4) + fabs(o1_2 - o1_3) + fabs(o1_3 - o1_4) + fabs(o1_2 - o1_4));
float c2 = (fabs(o2_1 - o2_2) + fabs(o2_1 - o2_3) + fabs(o2_1 - o2_4) + fabs(o2_2 - o2_3) + fabs(o2_3 - o2_4) + fabs(o2_2 - o2_4));
if (c1 + c2 < thresh6 * fabs(d1 - d2)) {
//pixel interpolation
float gin = cfa[rr][cc>>1];
float gmp2p2 = gin - cfa[rr + 2][(cc + 2)>>1];
float gmm2m2 = gin - cfa[rr - 2][(cc - 2)>>1];
float gmm2p2 = gin - cfa[rr - 2][(cc + 2)>>1];
float gmp2m2 = gin - cfa[rr + 2][(cc - 2)>>1];
float gse = o1_4 + 0.5f * gmp2p2;
float gnw = o1_1 + 0.5f * gmm2m2;
float gne = o1_2 + 0.5f * gmm2p2;
float gsw = o1_3 + 0.5f * gmp2m2;
float wtse = 1.f / (eps + SQR(gmp2p2) + SQR(cfa[rr + 3][(cc + 3)>>1] - o1_4));
float wtnw = 1.f / (eps + SQR(gmm2m2) + SQR(cfa[rr - 3][(cc - 3)>>1] - o1_1));
float wtne = 1.f / (eps + SQR(gmm2p2) + SQR(cfa[rr - 3][(cc + 3)>>1] - o1_2));
float wtsw = 1.f / (eps + SQR(gmp2m2) + SQR(cfa[rr + 3][(cc - 3)>>1] - o1_3));
float ginterp = (gse * wtse + gnw * wtnw + gne * wtne + gsw * wtsw) / (wtse + wtnw + wtne + wtsw);
if ( ((ginterp - gin) < thresh * (ginterp + gin)) ) {
if (ginterp - gin < thresh * (ginterp + gin)) {
rawData[rr][cc] = 0.5f * (ginterp + gin);
//counter++;
}
}
// }
}
//printf("pixfix count= %d; vtest= %d \n",counter,vtest);
// %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
// done
/*t2 = clock();
dt = ((double)(t2-t1)) / CLOCKS_PER_SEC;
if (verbose) {
fprintf(stderr,_("elapsed time = %5.3fs\n"),dt);
}*/
}
}
}
}

41
rtengine/rawimagesource.cc
View File

@ -1915,43 +1915,46 @@ void RawImageSource::preprocess (const RAWParams &raw, const LensProfParams &le
// check if it is an olympus E camera, if yes, compute G channel pre-compensation factors
if ( ri->getSensorType() == ST_BAYER && (raw.bayersensor.greenthresh || (((idata->getMake().size() >= 7 && idata->getMake().substr(0, 7) == "OLYMPUS" && idata->getModel()[0] == 'E') || (idata->getMake().size() >= 9 && idata->getMake().substr(0, 9) == "Panasonic")) && raw.bayersensor.method != RAWParams::BayerSensor::methodstring[ RAWParams::BayerSensor::vng4])) ) {
StopWatch stop1("gel part one");
// global correction
int ng1 = 0, ng2 = 0, i = 0;
int ng1 = 0, ng2 = 0;
double avgg1 = 0., avgg2 = 0.;
#ifdef _OPENMP
#pragma omp parallel for default(shared) private(i) reduction(+: ng1, ng2, avgg1, avgg2)
#pragma omp parallel for reduction(+: ng1, ng2, avgg1, avgg2) schedule(dynamic,16)
#endif
for (i = border; i < H - border; i++)
for (int j = border; j < W - border; j++)
if (ri->ISGREEN(i, j)) {
for (int i = border; i < H - border; i++) {
double avgg = 0.;
for (int j = border + ((FC(i, border) & 1) ^ 1); j < W - border; j += 2) {
avgg += rawData[i][j];
}
int ng = (W - 2 * border + (FC(i, border) & 1)) / 2;
if (i & 1) {
avgg2 += rawData[i][j];
ng2++;
avgg2 += avgg;
ng2 += ng;
} else {
avgg1 += rawData[i][j];
ng1++;
avgg1 += avgg;
ng1 += ng;
}
}
double corrg1 = ((double)avgg1 / ng1 + (double)avgg2 / ng2) / 2.0 / ((double)avgg1 / ng1);
double corrg2 = ((double)avgg1 / ng1 + (double)avgg2 / ng2) / 2.0 / ((double)avgg2 / ng2);
double corrg1 = (avgg1 / ng1 + avgg2 / ng2) / 2.0 / (avgg1 / ng1);
double corrg2 = (avgg1 / ng1 + avgg2 / ng2) / 2.0 / (avgg2 / ng2);
#ifdef _OPENMP
#pragma omp parallel for default(shared)
#pragma omp parallel for schedule(dynamic,16)
#endif
for (int i = border; i < H - border; i++)
for (int j = border; j < W - border; j++)
if (ri->ISGREEN(i, j)) {
float currData;
currData = (float)(rawData[i][j] * ((i & 1) ? corrg2 : corrg1));
rawData[i][j] = (currData);
for (int i = border; i < H - border; i++) {
double corrg = (i & 1) ? corrg2 : corrg1;
for (int j = border + ((FC(i, border) & 1) ^ 1); j < W - border; j += 2) {
rawData[i][j] *= corrg;
}
}
}
if ( ri->getSensorType() == ST_BAYER && raw.bayersensor.greenthresh > 0) {
StopWatch Stop1("gel part two");
if (plistener) {
plistener->setProgressStr ("Green equilibrate...");
plistener->setProgress (0.0);