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

rcd_demosaic(): another small speedup

Browse Source
This commit is contained in:
heckflosse
2018-03-01 00:13:59 +01:00
parent 2fb4403823
commit 7e5f2ee19e
1 changed files with 11 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
rtengine/rcd_demosaic.cc
View File

@@ -103,9 +103,8 @@ void RawImageSource::rcd_demosaic()
for (int col = 4, indx = row * tileSize + col; col < tileSize - 4; col++, indx++) {
const float cfai = cfa[indx];
//Calculate h/v local discrimination
float V_Stat = max(epssq, - 18.f * cfai * (cfa[indx - w1] + cfa[indx + w1] + 2.f * (cfa[indx - w2] + cfa[indx + w2]) - cfa[indx - w3] - cfa[indx + w3]) - 2.f * cfai * (cfa[indx - w4] + cfa[indx + w4] - 19.f * cfai) - cfa[indx - w1] * (70.f * cfa[indx + w1] + 12.f * cfa[indx - w2] - 24.f * cfa[indx + w2] + 38.f * cfa[indx - w3] - 16.f * cfa[indx + w3] - 12.f * cfa[indx - w4] + 6.f * cfa[indx + w4] - 46.f * cfa[indx - w1]) + cfa[indx + w1] * (24.f * cfa[indx - w2] - 12.f * cfa[indx + w2] + 16.f * cfa[indx - w3] - 38.f * cfa[indx + w3] - 6.f * cfa[indx - w4] + 12.f * cfa[indx + w4] + 46.f * cfa[indx + w1]) + cfa[indx - w2] * (14.f * cfa[indx + w2] - 12.f * cfa[indx + w3] - 2.f * cfa[indx - w4] + 2.f * cfa[indx + w4] + 11.f * cfa[indx - w2]) + cfa[indx + w2] * (-12.f * cfa[indx - w3] + 2.f * cfa[indx - w4] - 2.f * cfa[indx + w4] + 11.f * cfa[indx + w2]) + cfa[indx - w3] * (2.f * cfa[indx + w3] - 6.f * cfa[indx - w4] + 10.f * cfa[indx - w3]) + cfa[indx + w3] * (-6.f * cfa[indx + w4] + 10.f * cfa[indx + w3]) + cfa[indx - w4] * cfa[indx - w4] + cfa[indx + w4] * cfa[indx + w4]);
float H_Stat = max(epssq, - 18.f * cfai * (cfa[indx - 1] + cfa[indx + 1] + 2.f * (cfa[indx - 2] + cfa[indx + 2]) - cfa[indx - 3] - cfa[indx + 3]) - 2.f * cfai * (cfa[indx - 4] + cfa[indx + 4] - 19.f * cfai) - cfa[indx - 1] * (70.f * cfa[indx + 1] + 12.f * cfa[indx - 2] - 24.f * cfa[indx + 2] + 38.f * cfa[indx - 3] - 16.f * cfa[indx + 3] - 12.f * cfa[indx - 4] + 6.f * cfa[indx + 4] - 46.f * cfa[indx - 1]) + cfa[indx + 1] * (24.f * cfa[indx - 2] - 12.f * cfa[indx + 2] + 16.f * cfa[indx - 3] - 38.f * cfa[indx + 3] - 6.f * cfa[indx - 4] + 12.f * cfa[indx + 4] + 46.f * cfa[indx + 1]) + cfa[indx - 2] * (14.f * cfa[indx + 2] - 12.f * cfa[indx + 3] - 2.f * cfa[indx - 4] + 2.f * cfa[indx + 4] + 11.f * cfa[indx - 2]) + cfa[indx + 2] * (-12.f * cfa[indx - 3] + 2.f * cfa[indx - 4] - 2.f * cfa[indx + 4] + 11.f * cfa[indx + 2]) + cfa[indx - 3] * (2.f * cfa[indx + 3] - 6.f * cfa[indx - 4] + 10.f * cfa[indx - 3]) + cfa[indx + 3] * (-6.f * cfa[indx + 4] + 10.f * cfa[indx + 3]) + cfa[indx - 4] * cfa[indx - 4] + cfa[indx + 4] * cfa[indx + 4]);
float V_Stat = max(epssq, - 18.f * cfai * (cfa[indx - w1] + cfa[indx + w1] + 2.f * (cfa[indx - w2] + cfa[indx + w2]) - cfa[indx - w3] - cfa[indx + w3]) - 2.f * cfai * (cfa[indx - w4] + cfa[indx + w4] - 19.f * cfai) - cfa[indx - w1] * (70.f * cfa[indx + w1] + 12.f * cfa[indx - w2] - 24.f * cfa[indx + w2] + 38.f * cfa[indx - w3] - 16.f * cfa[indx + w3] - 12.f * cfa[indx - w4] + 6.f * cfa[indx + w4] - 46.f * cfa[indx - w1]) + cfa[indx + w1] * (24.f * cfa[indx - w2] - 12.f * cfa[indx + w2] + 16.f * cfa[indx - w3] - 38.f * cfa[indx + w3] - 6.f * cfa[indx - w4] + 12.f * cfa[indx + w4] + 46.f * cfa[indx + w1]) + cfa[indx - w2] * (14.f * cfa[indx + w2] - 12.f * cfa[indx + w3] - 2.f * cfa[indx - w4] + 2.f * cfa[indx + w4] + 11.f * cfa[indx - w2]) + cfa[indx + w2] * (-12.f * cfa[indx - w3] + 2.f * (cfa[indx - w4] - cfa[indx + w4]) + 11.f * cfa[indx + w2]) + cfa[indx - w3] * (2.f * cfa[indx + w3] - 6.f * cfa[indx - w4] + 10.f * cfa[indx - w3]) + cfa[indx + w3] * (-6.f * cfa[indx + w4] + 10.f * cfa[indx + w3]) + cfa[indx - w4] * cfa[indx - w4] + cfa[indx + w4] * cfa[indx + w4]);
float H_Stat = max(epssq, - 18.f * cfai * (cfa[indx - 1] + cfa[indx + 1] + 2.f * (cfa[indx - 2] + cfa[indx + 2]) - cfa[indx - 3] - cfa[indx + 3]) - 2.f * cfai * (cfa[indx - 4] + cfa[indx + 4] - 19.f * cfai) - cfa[indx - 1] * (70.f * cfa[indx + 1] + 12.f * cfa[indx - 2] - 24.f * cfa[indx + 2] + 38.f * cfa[indx - 3] - 16.f * cfa[indx + 3] - 12.f * cfa[indx - 4] + 6.f * cfa[indx + 4] - 46.f * cfa[indx - 1]) + cfa[indx + 1] * (24.f * cfa[indx - 2] - 12.f * cfa[indx + 2] + 16.f * cfa[indx - 3] - 38.f * cfa[indx + 3] - 6.f * cfa[indx - 4] + 12.f * cfa[indx + 4] + 46.f * cfa[indx + 1]) + cfa[indx - 2] * (14.f * cfa[indx + 2] - 12.f * cfa[indx + 3] - 2.f * cfa[indx - 4] + 2.f * cfa[indx + 4] + 11.f * cfa[indx - 2]) + cfa[indx + 2] * (-12.f * cfa[indx - 3] + 2.f * (cfa[indx - 4] - cfa[indx + 4]) + 11.f * cfa[indx + 2]) + cfa[indx - 3] * (2.f * cfa[indx + 3] - 6.f * cfa[indx - 4] + 10.f * cfa[indx - 3]) + cfa[indx + 3] * (-6.f * cfa[indx + 4] + 10.f * cfa[indx + 3]) + cfa[indx - 4] * cfa[indx - 4] + cfa[indx + 4] * cfa[indx + 4]);
VH_Dir[indx] = V_Stat / (V_Stat + H_Stat);
}
@@ -163,9 +162,10 @@ void RawImageSource::rcd_demosaic()
// Step 4.1: Calculate P/Q diagonal local discrimination
for (int row = 4; row < tileSize - 4; row++) {
for (int col = 4 + (FC(row, 0) & 1), indx = row * tileSize + col; col < tileSize - 4; col += 2, indx += 2) {
const float cfai = cfa[indx];
float P_Stat = max(- 18.f * cfa[indx] * cfa[indx - w1 - 1] - 18.f * cfa[indx] * cfa[indx + w1 + 1] - 36.f * cfa[indx] * cfa[indx - w2 - 2] - 36.f * cfa[indx] * cfa[indx + w2 + 2] + 18.f * cfa[indx] * cfa[indx - w3 - 3] + 18.f * cfa[indx] * cfa[indx + w3 + 3] - 2.f * cfa[indx] * cfa[indx - w4 - 4] - 2.f * cfa[indx] * cfa[indx + w4 + 4] + 38.f * cfa[indx] * cfa[indx] - 70.f * cfa[indx - w1 - 1] * cfa[indx + w1 + 1] - 12.f * cfa[indx - w1 - 1] * cfa[indx - w2 - 2] + 24.f * cfa[indx - w1 - 1] * cfa[indx + w2 + 2] - 38.f * cfa[indx - w1 - 1] * cfa[indx - w3 - 3] + 16.f * cfa[indx - w1 - 1] * cfa[indx + w3 + 3] + 12.f * cfa[indx - w1 - 1] * cfa[indx - w4 - 4] - 6.f * cfa[indx - w1 - 1] * cfa[indx + w4 + 4] + 46.f * cfa[indx - w1 - 1] * cfa[indx - w1 - 1] + 24.f * cfa[indx + w1 + 1] * cfa[indx - w2 - 2] - 12.f * cfa[indx + w1 + 1] * cfa[indx + w2 + 2] + 16.f * cfa[indx + w1 + 1] * cfa[indx - w3 - 3] - 38.f * cfa[indx + w1 + 1] * cfa[indx + w3 + 3] - 6.f * cfa[indx + w1 + 1] * cfa[indx - w4 - 4] + 12.f * cfa[indx + w1 + 1] * cfa[indx + w4 + 4] + 46.f * cfa[indx + w1 + 1] * cfa[indx + w1 + 1] + 14.f * cfa[indx - w2 - 2] * cfa[indx + w2 + 2] - 12.f * cfa[indx - w2 - 2] * cfa[indx + w3 + 3] - 2.f * cfa[indx - w2 - 2] * cfa[indx - w4 - 4] + 2.f * cfa[indx - w2 - 2] * cfa[indx + w4 + 4] + 11.f * cfa[indx - w2 - 2] * cfa[indx - w2 - 2] - 12.f * cfa[indx + w2 + 2] * cfa[indx - w3 - 3] + 2 * cfa[indx + w2 + 2] * cfa[indx - w4 - 4] - 2.f * cfa[indx + w2 + 2] * cfa[indx + w4 + 4] + 11.f * cfa[indx + w2 + 2] * cfa[indx + w2 + 2] + 2.f * cfa[indx - w3 - 3] * cfa[indx + w3 + 3] - 6.f * cfa[indx - w3 - 3] * cfa[indx - w4 - 4] + 10.f * cfa[indx - w3 - 3] * cfa[indx - w3 - 3] - 6.f * cfa[indx + w3 + 3] * cfa[indx + w4 + 4] + 10.f * cfa[indx + w3 + 3] * cfa[indx + w3 + 3] + 1.f * cfa[indx - w4 - 4] * cfa[indx - w4 - 4] + 1.f * cfa[indx + w4 + 4] * cfa[indx + w4 + 4], epssq);
float Q_Stat = max(- 18.f * cfa[indx] * cfa[indx + w1 - 1] - 18.f * cfa[indx] * cfa[indx - w1 + 1] - 36.f * cfa[indx] * cfa[indx + w2 - 2] - 36.f * cfa[indx] * cfa[indx - w2 + 2] + 18.f * cfa[indx] * cfa[indx + w3 - 3] + 18.f * cfa[indx] * cfa[indx - w3 + 3] - 2.f * cfa[indx] * cfa[indx + w4 - 4] - 2.f * cfa[indx] * cfa[indx - w4 + 4] + 38.f * cfa[indx] * cfa[indx] - 70.f * cfa[indx + w1 - 1] * cfa[indx - w1 + 1] - 12.f * cfa[indx + w1 - 1] * cfa[indx + w2 - 2] + 24.f * cfa[indx + w1 - 1] * cfa[indx - w2 + 2] - 38.f * cfa[indx + w1 - 1] * cfa[indx + w3 - 3] + 16.f * cfa[indx + w1 - 1] * cfa[indx - w3 + 3] + 12.f * cfa[indx + w1 - 1] * cfa[indx + w4 - 4] - 6.f * cfa[indx + w1 - 1] * cfa[indx - w4 + 4] + 46.f * cfa[indx + w1 - 1] * cfa[indx + w1 - 1] + 24.f * cfa[indx - w1 + 1] * cfa[indx + w2 - 2] - 12.f * cfa[indx - w1 + 1] * cfa[indx - w2 + 2] + 16.f * cfa[indx - w1 + 1] * cfa[indx + w3 - 3] - 38.f * cfa[indx - w1 + 1] * cfa[indx - w3 + 3] - 6.f * cfa[indx - w1 + 1] * cfa[indx + w4 - 4] + 12.f * cfa[indx - w1 + 1] * cfa[indx - w4 + 4] + 46.f * cfa[indx - w1 + 1] * cfa[indx - w1 + 1] + 14.f * cfa[indx + w2 - 2] * cfa[indx - w2 + 2] - 12.f * cfa[indx + w2 - 2] * cfa[indx - w3 + 3] - 2.f * cfa[indx + w2 - 2] * cfa[indx + w4 - 4] + 2.f * cfa[indx + w2 - 2] * cfa[indx - w4 + 4] + 11.f * cfa[indx + w2 - 2] * cfa[indx + w2 - 2] - 12.f * cfa[indx - w2 + 2] * cfa[indx + w3 - 3] + 2 * cfa[indx - w2 + 2] * cfa[indx + w4 - 4] - 2.f * cfa[indx - w2 + 2] * cfa[indx - w4 + 4] + 11.f * cfa[indx - w2 + 2] * cfa[indx - w2 + 2] + 2.f * cfa[indx + w3 - 3] * cfa[indx - w3 + 3] - 6.f * cfa[indx + w3 - 3] * cfa[indx + w4 - 4] + 10.f * cfa[indx + w3 - 3] * cfa[indx + w3 - 3] - 6.f * cfa[indx - w3 + 3] * cfa[indx - w4 + 4] + 10.f * cfa[indx - w3 + 3] * cfa[indx - w3 + 3] + 1.f * cfa[indx + w4 - 4] * cfa[indx + w4 - 4] + 1.f * cfa[indx - w4 + 4] * cfa[indx - w4 + 4], epssq);
float P_Stat = max(epssq, - 18.f * cfai * (cfa[indx - w1 - 1] + cfa[indx + w1 + 1] + 2.f * (cfa[indx - w2 - 2] + cfa[indx + w2 + 2]) - cfa[indx - w3 - 3] - cfa[indx + w3 + 3]) - 2.f * cfai * (cfa[indx - w4 - 4] + cfa[indx + w4 + 4] - 19.f * cfai) - cfa[indx - w1 - 1] * (70.f * cfa[indx + w1 + 1] - 12.f * cfa[indx - w2 - 2] + 24.f * cfa[indx + w2 + 2] - 38.f * cfa[indx - w3 - 3] + 16.f * cfa[indx + w3 + 3] + 12.f * cfa[indx - w4 - 4] - 6.f * cfa[indx + w4 + 4] + 46.f * cfa[indx - w1 - 1]) + cfa[indx + w1 + 1] * (24.f * cfa[indx - w2 - 2] - 12.f * cfa[indx + w2 + 2] + 16.f * cfa[indx - w3 - 3] - 38.f * cfa[indx + w3 + 3] - 6.f * cfa[indx - w4 - 4] + 12.f * cfa[indx + w4 + 4] + 46.f * cfa[indx + w1 + 1]) + cfa[indx - w2 - 2] * (14.f * cfa[indx + w2 + 2] - 12.f * cfa[indx + w3 + 3] - 2.f * (cfa[indx - w4 - 4] - cfa[indx + w4 + 4]) + 11.f * cfa[indx - w2 - 2]) - cfa[indx + w2 + 2] * (12.f * cfa[indx - w3 - 3] + 2.f * (cfa[indx - w4 - 4] - cfa[indx + w4 + 4]) + 11.f * cfa[indx + w2 + 2]) + cfa[indx - w3 - 3] * (2.f * cfa[indx + w3 + 3] - 6.f * cfa[indx - w4 - 4] + 10.f * cfa[indx - w3 - 3]) - cfa[indx + w3 + 3] * (6.f * cfa[indx + w4 + 4] + 10.f * cfa[indx + w3 + 3]) + cfa[indx - w4 - 4] * cfa[indx - w4 - 4] + cfa[indx + w4 + 4] * cfa[indx + w4 + 4]);
float Q_Stat = max(epssq, - 18.f * cfai * (cfa[indx + w1 - 1] + cfa[indx - w1 + 1] + 2.f * (cfa[indx + w2 - 2] + cfa[indx - w2 + 2]) - cfa[indx + w3 - 3] - cfa[indx - w3 + 3]) - 2.f * cfai * (cfa[indx + w4 - 4] + cfa[indx - w4 + 4] - 19.f * cfai) - cfa[indx + w1 - 1] * (70.f * cfa[indx - w1 + 1] - 12.f * cfa[indx + w2 - 2] + 24.f * cfa[indx - w2 + 2] - 38.f * cfa[indx + w3 - 3] + 16.f * cfa[indx - w3 + 3] + 12.f * cfa[indx + w4 - 4] - 6.f * cfa[indx - w4 + 4] + 46.f * cfa[indx + w1 - 1]) + cfa[indx - w1 + 1] * (24.f * cfa[indx + w2 - 2] - 12.f * cfa[indx - w2 + 2] + 16.f * cfa[indx + w3 - 3] - 38.f * cfa[indx - w3 + 3] - 6.f * cfa[indx + w4 - 4] + 12.f * cfa[indx - w4 + 4] + 46.f * cfa[indx - w1 + 1]) + cfa[indx + w2 - 2] * (14.f * cfa[indx - w2 + 2] - 12.f * cfa[indx - w3 + 3] - 2.f * (cfa[indx + w4 - 4] - cfa[indx - w4 + 4]) + 11.f * cfa[indx + w2 - 2]) - cfa[indx - w2 + 2] * (12.f * cfa[indx + w3 - 3] + 2.f * (cfa[indx + w4 - 4] - cfa[indx - w4 + 4]) + 11.f * cfa[indx - w2 + 2]) + cfa[indx + w3 - 3] * (2.f * cfa[indx - w3 + 3] - 6.f * cfa[indx + w4 - 4] + 10.f * cfa[indx + w3 - 3]) - cfa[indx - w3 + 3] * (6.f * cfa[indx - w4 + 4] + 10.f * cfa[indx - w3 + 3]) + cfa[indx + w4 - 4] * cfa[indx + w4 - 4] + cfa[indx - w4 + 4] * cfa[indx - w4 + 4]);
PQ_Dir[indx] = P_Stat / (P_Stat + Q_Stat);
@@ -220,10 +220,12 @@ void RawImageSource::rcd_demosaic()
for (int c = 0; c <= 2; c += 2) {
// Cardinal gradients
float N_Grad = N1 + std::fabs(rgb[c][indx - w1] - rgb[c][indx + w1]) + std::fabs(rgb[c][indx - w1] - rgb[c][indx - w3]);
float S_Grad = S1 + std::fabs(rgb[c][indx - w1] - rgb[c][indx + w1]) + std::fabs(rgb[c][indx + w1] - rgb[c][indx + w3]);
float W_Grad = W1 + std::fabs(rgb[c][indx - 1] - rgb[c][indx + 1]) + std::fabs(rgb[c][indx - 1] - rgb[c][indx - 3]);
float E_Grad = E1 + std::fabs(rgb[c][indx - 1] - rgb[c][indx + 1]) + std::fabs(rgb[c][indx + 1] - rgb[c][indx + 3]);
float SNabs = std::fabs(rgb[c][indx - w1] - rgb[c][indx + w1]);
float EWabs = std::fabs(rgb[c][indx - 1] - rgb[c][indx + 1]);
float N_Grad = N1 + SNabs + std::fabs(rgb[c][indx - w1] - rgb[c][indx - w3]);
float S_Grad = S1 + SNabs + std::fabs(rgb[c][indx + w1] - rgb[c][indx + w3]);
float W_Grad = W1 + EWabs + std::fabs(rgb[c][indx - 1] - rgb[c][indx - 3]);
float E_Grad = E1 + EWabs + std::fabs(rgb[c][indx + 1] - rgb[c][indx + 3]);
// Cardinal colour differences
float N_Est = rgb[c][indx - w1] - rgb[1][indx - w1];