Several performance improvements with OpenMP.

- Faster image load.
- Faster image alignment.
- Faster boxblur.
- Faster result render.
- Faster writing DNG tiles.

DngFloatWriter.cpp

@@ -494,12 +494,12 @@ void DngFloatWriter::writeRawData() {
     int bytesps = bps >> 3;
     uLongf dstLen = tileWidth * tileLength * bytesps;
 
-//     #pragma omp parallel
+    #pragma omp parallel
     {
         Bytef cBuffer[dstLen];
         Bytef uBuffer[dstLen];
 
-//         #pragma omp for collapse(2)
+        #pragma omp for collapse(2) schedule(dynamic)
         for (size_t y = 0; y < height; y += tileLength) {
             for (size_t x = 0; x < width; x += tileWidth) {
                 size_t t = (y / tileLength) * tilesAcross + (x / tileWidth);
@@ -520,7 +520,7 @@ void DngFloatWriter::writeRawData() {
                 if (err != Z_OK) {
                     std::cerr << "DNG Deflate: Failed compressing tile " << t << ", with error " << err << std::endl;
                 } else {
-//                     #pragma omp critical
+                    #pragma omp critical
                     {
                         tileOffsets[t] = file.tellp();
                         file.write((const char *)cBuffer, tileBytes[t]);

EditableMask.cpp

@@ -174,6 +174,7 @@ void EditableMask::BoxBlur::boxBlur_4(size_t radius) {
 
 void EditableMask::BoxBlur::boxBlurH_4(size_t r) {
     float iarr = 1.0 / (r+r+1);
+    #pragma omp parallel for schedule(dynamic)
     for (size_t i = 0; i < m.height; ++i) {
         size_t ti = i * m.width, li = ti, ri = ti + r;
         float val = map[li] * (r + 1);
@@ -198,6 +199,7 @@ void EditableMask::BoxBlur::boxBlurH_4(size_t r) {
 
 void EditableMask::BoxBlur::boxBlurT_4(size_t r) {
     float iarr = 1.0 / (r+r+1);
+    #pragma omp parallel for schedule(dynamic)
     for (size_t i = 0; i < m.width; ++i) {
         size_t ti = i, li = ti, ri = ti + r*m.width;
         float val = map[li] * (r + 1);

Image.cpp

@@ -175,10 +175,6 @@ void Image::alignWith(const Image & r) {
         dx <<= 1;
         dy <<= 1;
     }
-    dx += r.dx;
-    dy += r.dy;
-    alignedPixels = &rawPixels[-dy*width - dx];
-    Log::msg(Log::DEBUG, "Image ", metaData->fileName, " displaced to (", dx, ", ", dy, ")");
 }
 
 

ImageStack.cpp

@@ -52,18 +52,35 @@ bool ImageStack::addImage(std::unique_ptr<Image> & i) {
 
 
 int ImageStack::load(const LoadOptions & options, ProgressIndicator & progress) {
-    int step = 100 / (options.fileNames.size() + 1);
+    int numImages = options.fileNames.size();
+    int step = 100 / (numImages + 1);
     int p = -step;
-    for (auto & name : options.fileNames) {
-        progress.advance(p += step, "Loading %1", name.c_str());
-        std::unique_ptr<Image> image(measureTime("Load raw", [&] () { return new Image(name.c_str()); }));
-        if (image.get() == nullptr || !image->good()) {
-            return 1;
-        }
-        if (!addImage(image)) {
-            return 2;
+    int error = 0, failedImage = 0;
+    {
+        Timer t("Load files");
+        #pragma omp parallel for schedule(dynamic)
+        for (int i = 0; i < numImages; ++i) {
+            if (!error) { // We cannot break from the for loop if we are using OpenMP
+                string name = options.fileNames[i];
+                #pragma omp critical
+                progress.advance(p += step, "Loading %1", name.c_str());
+                std::unique_ptr<Image> image(new Image(name.c_str()));
+                #pragma omp critical
+                if (!error) { // Report on the first image that fails, ignore the rest
+                    if (image.get() == nullptr || !image->good()) {
+                        error = 1;
+                        failedImage = i;
+                    } else if (!addImage(image)) {
+                        error = 2;
+                        failedImage = i;
+                    }
+                }
+            }
         }
     }
+    if (error) {
+        return (failedImage << 1) + error - 1;
+    }
     if (options.align) {
         Timer t("Align");
         progress.advance(p += step, "Aligning");
@@ -75,7 +92,7 @@ int ImageStack::load(const LoadOptions & options, ProgressIndicator & progress)
     computeRelExposures();
     mask.generateFrom(*this);
     progress.advance(100, "Done loading!");
-    return 0;
+    return numImages << 1;
 }
 
 
@@ -93,8 +110,12 @@ int ImageStack::save(const SaveOptions & options, ProgressIndicator & progress)
 
 void ImageStack::align() {
     if (images.size() > 1) {
-        for (auto cur = images.rbegin(), next = cur++; cur != images.rend(); next = cur++) {
-            (*cur)->alignWith(**next);
+        for (int i = images.size() - 2; i >= 0; --i) {
+            images[i]->alignWith(*images[i + 1]);
+        }
+        for (int i = images.size() - 2; i >= 0; --i) {
+            images[i]->displace(images[i + 1]->getDeltaX(), images[i + 1]->getDeltaY());
+            Log::msg(Log::DEBUG, "Image ", i, " displaced to (", images[i]->getDeltaX(), ", ", images[i]->getDeltaY(), ")");
         }
         for (auto & i : images) {
             i->releaseAlignData();
@@ -134,13 +155,14 @@ double ImageStack::value(size_t x, size_t y) const {
 
 
 void ImageStack::compose(float * dst) const {
-    Timer t("Compose");
     unique_ptr<float[]> map = measureTime("Blur", [&] () { return mask.blur(); });
+    Timer t("Compose");
     const MetaData & md = images.front()->getMetaData();
     int imageMax = images.size() - 1;
     float max = 0.0;
-    for (size_t y = 0, pos = 0; y < height; ++y) {
-        for (size_t x = 0; x < width; ++x, ++pos) {
+    #pragma omp parallel for schedule(dynamic)
+    for (size_t y = 0; y < height; ++y) {
+        for (size_t x = 0, pos = y*width; x < width; ++x, ++pos) {
             int j = map[pos] > imageMax ? imageMax : ceil(map[pos]);
             double v = 0.0, vv = 0.0, p;
             if (images[j]->contains(x, y)) {

Launcher.cpp

@@ -73,12 +73,17 @@ int Launcher::automaticMerge() {
     auto tr = [&] (const char * text) { return QCoreApplication::translate("LoadSave", text); };
     CoutProgressIndicator progress;
     ImageStack stack;
-    if (stack.load(options, progress)) {
-        int i = progress.getPercent() * (options.fileNames.size() + 1) / 100;
-        if (i) {
-            cerr << tr("Error loading %1").arg(options.fileNames[i].c_str()) << endl;
-            return 1;
+    int numImages = options.fileNames.size();
+    int result = stack.load(options, progress);
+    if (result < numImages * 2) {
+        int format = result & 1;
+        int i = result >> 1;
+        if (format) {
+            cerr << tr("Error loading %1, it has a different format.").arg(options.fileNames[i].c_str()) << endl;
+        } else {
+            cerr << tr("Error loading %1, file not found.").arg(options.fileNames[i].c_str()) << endl;
         }
+        return 1;
     }
     if (!wOptions.fileName.empty()) {
         size_t extPos = wOptions.fileName.find_last_of('.');

MainWindow.cpp

@@ -273,11 +273,12 @@ void MainWindow::loadImages(const LoadOptions & options) {
         QFuture<int> error = QtConcurrent::run([&] () { return newImages->load(options, progress); });
         while (error.isRunning())
             QApplication::instance()->processEvents(QEventLoop::ExcludeUserInputEvents);
-        if (error.result()) {
-            int i = progress.getPercent() * (numImages + 1) / 100;
-            QString message = error.result() == 1 ?
-                tr("Unable to open file %1.").arg(options.fileNames[i].c_str()) :
-                tr("File %1 has not the same format as the previous ones.").arg(options.fileNames[i].c_str());
+        int result = error.result();
+        if (result < numImages * 2) {
+            int i = result >> 1;
+            QString message = result & 1 ?
+            tr("File %1 has not the same format as the previous ones.").arg(options.fileNames[i].c_str()) :
+            tr("Unable to open file %1.").arg(options.fileNames[i].c_str());
             QMessageBox::warning(this, tr("Error opening file"), message);
             delete newImages;
             return;

PreviewWidget.cpp

@@ -121,6 +121,7 @@ void PreviewWidget::render(int minx, int miny, int maxx, int maxy) {
     if (area.isEmpty()) return;
     area.getCoords(&minx, &miny, &maxx, &maxy);
     QImage image(area.width() - 1, area.height() - 1, QImage::Format_RGB32);
+    #pragma omp parallel for schedule(dynamic)
     for (int row = miny; row < maxy; row++) {
         QRgb * scanLine = reinterpret_cast<QRgb *>(image.scanLine(row - miny));
         for (int col = minx; col < maxx; col++) {
@@ -188,6 +189,8 @@ void PreviewWidget::mousePressEvent(QMouseEvent * event) {
     } else
         event->ignore();
 }
+
+
 void PreviewWidget::mouseMoveEvent(QMouseEvent * event) {
     if (addPixels || rmPixels) {
         event->accept();

hdrmerge_es.ts

@@ -75,8 +75,12 @@
 <context>
     <name>LoadSave</name>
     <message>
-        <source>Error loading %1</source>
-        <translation type="finished">Error cargando %1</translation>
+        <source>Error loading %1, it has a different format.</source>
+        <translation type="finished">Error cargando %1, tiene un formato diferente.</translation>
+    </message>
+    <message>
+        <source>Error loading %1, file not found.</source>
+        <translation type="finished">Error cargando %1, no se encontró el fichero.</translation>
     </message>
     <message>
         <source>Writing result to %1</source>