Added some optimizations to the blur and drop shadow GL filters.

* Use ExpandToTransparentBorderPadMode since we can use GL_CLAMP_TO_EDGE
  to clamp to the texture.

* Shrink the bounding rects reported by the blur
  and drop shadow filters (expanding by 2 * radius isn't needed).

* Use a single-pass blur for radii <= 3 to avoid the overhead of
  rendering to an FBO.

* Made the fast blur setting generate filters for only a predefined set
  of radii, and then use the actual blur radius to spread the sample
  points outwards.

* Optimized the generated program to rely less on temporary variables,
  as those seemed to not be handled very well by certain GLSL compilers.

Reviewed-by: Gunnar Sletta

8 files changed, 307 insertions, 195 deletions

diff --git a/src/gui/effects/qgraphicseffect.cpp b/src/gui/effects/qgraphicseffect.cpp
index 9ed003c..3a6bab5 100644
--- a/src/gui/effects/qgraphicseffect.cpp
+++ b/src/gui/effects/qgraphicseffect.cpp
@@ -101,6 +101,7 @@
 
 #include <QtGui/qimage.h>
 #include <QtGui/qpainter.h>
+#include <QtGui/qpaintengine.h>
 #include <QtCore/qrect.h>
 #include <QtCore/qdebug.h>
 #include <private/qdrawhelper_p.h>
@@ -260,12 +261,13 @@ QPixmap QGraphicsEffectSource::pixmap(Qt::CoordinateSystem system, QPoint *offse
     }
 
     QPixmap pm;
-    if (d->m_cachedSystem == system)
+    if (d->m_cachedSystem == system && d->m_cachedMode == mode)
         QPixmapCache::find(d->m_cacheKey, &pm);
 
     if (pm.isNull()) {
         pm = d->pixmap(system, &d->m_cachedOffset, mode);
         d->m_cachedSystem = system;
+        d->m_cachedMode = mode;
 
         d->invalidateCache();
         d->m_cacheKey = QPixmapCache::insert(pm);
@@ -708,9 +710,13 @@ void QGraphicsBlurEffect::draw(QPainter *painter, QGraphicsEffectSource *source)
         return;
     }
 
+    QGraphicsEffectSource::PixmapPadMode mode = QGraphicsEffectSource::ExpandToEffectRectPadMode;
+    if (painter->paintEngine()->type() == QPaintEngine::OpenGL2)
+        mode = QGraphicsEffectSource::ExpandToTransparentBorderPadMode;
+
     // Draw pixmap in device coordinates to avoid pixmap scaling.
     QPoint offset;
-    const QPixmap pixmap = source->pixmap(Qt::DeviceCoordinates, &offset);
+    const QPixmap pixmap = source->pixmap(Qt::DeviceCoordinates, &offset, mode);
     QTransform restoreTransform = painter->worldTransform();
     painter->setWorldTransform(QTransform());
     d->filter->draw(painter, offset, pixmap);
@@ -893,9 +899,13 @@ void QGraphicsDropShadowEffect::draw(QPainter *painter, QGraphicsEffectSource *s
         return;
     }
 
+    QGraphicsEffectSource::PixmapPadMode mode = QGraphicsEffectSource::ExpandToEffectRectPadMode;
+    if (painter->paintEngine()->type() == QPaintEngine::OpenGL2)
+        mode = QGraphicsEffectSource::ExpandToTransparentBorderPadMode;
+
     // Draw pixmap in device coordinates to avoid pixmap scaling.
     QPoint offset;
-    const QPixmap pixmap = source->pixmap(Qt::DeviceCoordinates, &offset);
+    const QPixmap pixmap = source->pixmap(Qt::DeviceCoordinates, &offset, mode);
     QTransform restoreTransform = painter->worldTransform();
     painter->setWorldTransform(QTransform());
     d->filter->draw(painter, offset, pixmap);
diff --git a/src/gui/effects/qgraphicseffect_p.h b/src/gui/effects/qgraphicseffect_p.h
index dff84a1..1ed7103 100644
--- a/src/gui/effects/qgraphicseffect_p.h
+++ b/src/gui/effects/qgraphicseffect_p.h
@@ -66,7 +66,12 @@ class QGraphicsEffectSourcePrivate : public QObjectPrivate
 {
     Q_DECLARE_PUBLIC(QGraphicsEffectSource)
 public:
-    QGraphicsEffectSourcePrivate() : QObjectPrivate() {}
+    QGraphicsEffectSourcePrivate()
+        : QObjectPrivate()
+        , m_cachedSystem(Qt::DeviceCoordinates)
+        , m_cachedMode(QGraphicsEffectSource::ExpandToTransparentBorderPadMode)
+    {}
+
     virtual ~QGraphicsEffectSourcePrivate() { invalidateCache(); }
     virtual void detach() = 0;
     virtual QRectF boundingRect(Qt::CoordinateSystem system) const = 0;
@@ -88,6 +93,7 @@ public:
 
 private:
     mutable Qt::CoordinateSystem m_cachedSystem;
+    mutable QGraphicsEffectSource::PixmapPadMode m_cachedMode;
     mutable QPoint m_cachedOffset;
     mutable QPixmapCache::Key m_cacheKey;
 };
diff --git a/src/gui/graphicsview/qgraphicsitem.cpp b/src/gui/graphicsview/qgraphicsitem.cpp
index 7d9390c..738c6e3 100644
--- a/src/gui/graphicsview/qgraphicsitem.cpp
+++ b/src/gui/graphicsview/qgraphicsitem.cpp
@@ -10703,7 +10703,8 @@ QPixmap QGraphicsItemEffectSourcePrivate::pixmap(Qt::CoordinateSystem system, QP
     if (mode == QGraphicsEffectSource::ExpandToEffectRectPadMode) {
         effectRect = item->graphicsEffect()->boundingRectFor(sourceRect).toAlignedRect();
     } else if (mode == QGraphicsEffectSource::ExpandToTransparentBorderPadMode) {
-        effectRect = sourceRect.adjusted(-1, -1, 1, 1).toAlignedRect();
+        // adjust by 1.5 to account for cosmetic pens
+        effectRect = sourceRect.adjusted(-1.5, -1.5, 1.5, 1.5).toAlignedRect();
     } else {
         effectRect = sourceRect.toAlignedRect();
     }
diff --git a/src/gui/image/qpixmapfilter.cpp b/src/gui/image/qpixmapfilter.cpp
index f9ac79f..d0de03e 100644
--- a/src/gui/image/qpixmapfilter.cpp
+++ b/src/gui/image/qpixmapfilter.cpp
@@ -584,7 +584,7 @@ Qt::RenderHint QPixmapBlurFilter::blurHint() const
 QRectF QPixmapBlurFilter::boundingRectFor(const QRectF &rect) const
 {
     Q_D(const QPixmapBlurFilter);
-    const qreal delta = d->radius * 2;
+    const qreal delta = d->radius + 1;
     return rect.adjusted(-delta, -delta, delta, delta);
 }
 
@@ -1057,14 +1057,9 @@ void QPixmapDropShadowFilter::setOffset(const QPointF &offset)
 QRectF QPixmapDropShadowFilter::boundingRectFor(const QRectF &rect) const
 {
     Q_D(const QPixmapDropShadowFilter);
-
-    const qreal delta = qreal(d->radius * 2);
-    qreal x1 = qMin(rect.left(), rect.left() + d->offset.x() - delta);
-    qreal y1 = qMin(rect.top(), rect.top() + d->offset.y() - delta);
-    qreal x2 = qMax(rect.right(), rect.right() + d->offset.x() + delta);
-    qreal y2 = qMax(rect.bottom(), rect.bottom() + d->offset.y() + delta);
-
-    return QRectF(x1, y1, x2 - x1, y2 - y1);
+    qreal delta = d->radius + 1;
+    return rect.adjusted(-2, -2, 2, 2).united(
+            rect.translated(d->offset).adjusted(-delta, -delta, delta, delta));
 }
 
 /*!
diff --git a/src/opengl/gl2paintengineex/qglengineshadermanager.cpp b/src/opengl/gl2paintengineex/qglengineshadermanager.cpp
index e22303d..af9306f 100644
--- a/src/opengl/gl2paintengineex/qglengineshadermanager.cpp
+++ b/src/opengl/gl2paintengineex/qglengineshadermanager.cpp
@@ -476,7 +476,7 @@ bool QGLEngineShaderManager::useCorrectShaderProg()
         return false;
 
     bool useCustomSrc = customSrcStage != 0;
-    if (useCustomSrc && srcPixelType != QGLEngineShaderManager::ImageSrc) {
+    if (useCustomSrc && srcPixelType != QGLEngineShaderManager::ImageSrc && srcPixelType != Qt::TexturePattern) {
         useCustomSrc = false;
         qWarning("QGLEngineShaderManager - Ignoring custom shader stage for non image src");
     }
diff --git a/src/opengl/gl2paintengineex/qglengineshadersource_p.h b/src/opengl/gl2paintengineex/qglengineshadersource_p.h
index 3eef808..2407979 100644
--- a/src/opengl/gl2paintengineex/qglengineshadersource_p.h
+++ b/src/opengl/gl2paintengineex/qglengineshadersource_p.h
@@ -258,7 +258,7 @@ static const char* const qglslPositionWithTextureBrushVertexShader = "\
     uniform   mediump vec2  halfViewportSize; \
     uniform   highp   vec2  invertedTextureSize; \
     uniform   highp   mat3  brushTransform; \
-    varying   highp   vec2  brushTextureCoords; \
+    varying   highp   vec2  textureCoords; \
     void setPosition(void) { \
             gl_Position = pmvMatrix * vertexCoordsArray;\
             gl_Position.xy = gl_Position.xy / gl_Position.w; \
@@ -267,7 +267,7 @@ static const char* const qglslPositionWithTextureBrushVertexShader = "\
             mediump float invertedHTexCoordsZ = 1.0 / hTexCoords.z; \
             gl_Position.xy = gl_Position.xy * invertedHTexCoordsZ; \
             gl_Position.w = invertedHTexCoordsZ; \
-            brushTextureCoords.xy = (hTexCoords.xy * invertedTextureSize) * gl_Position.w; \
+            textureCoords.xy = (hTexCoords.xy * invertedTextureSize) * gl_Position.w; \
     }";
 
 static const char* const qglslAffinePositionWithTextureBrushVertexShader
@@ -278,26 +278,26 @@ static const char* const qglslAffinePositionWithTextureBrushVertexShader
 // we emulate GL_REPEAT by only taking the fractional part of the texture coords.
 // TODO: Special case POT textures which don't need this emulation
 static const char* const qglslTextureBrushSrcFragmentShader = "\
-    varying highp   vec2      brushTextureCoords; \
+    varying highp   vec2      textureCoords; \
     uniform lowp    sampler2D brushTexture; \
     lowp vec4 srcPixel() { \
-        return texture2D(brushTexture, fract(brushTextureCoords)); \
+        return texture2D(brushTexture, fract(textureCoords)); \
     }";
 #else
 static const char* const qglslTextureBrushSrcFragmentShader = "\
-    varying highp   vec2      brushTextureCoords; \
+    varying highp   vec2      textureCoords; \
     uniform lowp    sampler2D brushTexture; \
     lowp vec4 srcPixel() { \
-        return texture2D(brushTexture, brushTextureCoords); \
+        return texture2D(brushTexture, textureCoords); \
     }";
 #endif
 
 static const char* const qglslTextureBrushSrcWithPatternFragmentShader = "\
-    varying highp   vec2      brushTextureCoords; \
+    varying highp   vec2      textureCoords; \
     uniform lowp    vec4      patternColor; \
     uniform lowp    sampler2D brushTexture; \
     lowp vec4 srcPixel() { \
-        return patternColor * (1.0 - texture2D(brushTexture, brushTextureCoords).r); \
+        return patternColor * (1.0 - texture2D(brushTexture, textureCoords).r); \
     }";
 
 // Solid Fill Brush
diff --git a/src/opengl/gl2paintengineex/qpaintengineex_opengl2.cpp b/src/opengl/gl2paintengineex/qpaintengineex_opengl2.cpp
index b70810d..a9744b3 100644
--- a/src/opengl/gl2paintengineex/qpaintengineex_opengl2.cpp
+++ b/src/opengl/gl2paintengineex/qpaintengineex_opengl2.cpp
@@ -535,7 +535,7 @@ void QGL2PaintEngineExPrivate::updateBrushUniforms()
                 shaderManager->currentProgram()->setUniformValue(location(QGLEngineShaderManager::PatternColor), col);
             }
 
-            QSizeF invertedTextureSize(1.0 / texPixmap.width(), 1.0 * textureInvertedY / texPixmap.height());
+            QSizeF invertedTextureSize(1.0 / texPixmap.width(), 1.0 / texPixmap.height());
             shaderManager->currentProgram()->setUniformValue(location(QGLEngineShaderManager::InvertedTextureSize), invertedTextureSize);
 
             QVector2D halfViewportSize(width*0.5, height*0.5);
@@ -550,7 +550,11 @@ void QGL2PaintEngineExPrivate::updateBrushUniforms()
 
         QTransform translate(1, 0, 0, 1, -translationPoint.x(), -translationPoint.y());
         QTransform gl_to_qt(1, 0, 0, -1, 0, height);
-        QTransform inv_matrix = gl_to_qt * (brushQTransform * matrix).inverted() * translate;
+        QTransform inv_matrix;
+        if (style == Qt::TexturePattern && textureInvertedY == -1)
+            inv_matrix = gl_to_qt * (QTransform(1, 0, 0, -1, 0, currentBrush->texture().height()) * brushQTransform * matrix).inverted() * translate;
+        else
+            inv_matrix = gl_to_qt * (brushQTransform * matrix).inverted() * translate;
 
         shaderManager->currentProgram()->setUniformValue(location(QGLEngineShaderManager::BrushTransform), inv_matrix);
         shaderManager->currentProgram()->setUniformValue(location(QGLEngineShaderManager::BrushTexture), QT_BRUSH_TEXTURE_UNIT);
diff --git a/src/opengl/qglpixmapfilter.cpp b/src/opengl/qglpixmapfilter.cpp
index 656957d..e381a5d 100644
--- a/src/opengl/qglpixmapfilter.cpp
+++ b/src/opengl/qglpixmapfilter.cpp
@@ -104,7 +104,7 @@ public:
 
     void setUniforms(QGLShaderProgram *program);
 
-    static QByteArray generateGaussianShader(int radius, bool dropShadow = false);
+    static QByteArray generateGaussianShader(int radius, bool singlePass = false, bool dropShadow = false);
 
 protected:
     bool processGL(QPainter *painter, const QPointF &pos, const QPixmap &src, const QRectF &srcRect) const;
@@ -113,6 +113,7 @@ private:
 
     mutable QSize m_textureSize;
     mutable bool m_horizontalBlur;
+    mutable bool m_singlePass;
 
     mutable bool m_haveCached;
     mutable int m_cachedRadius;
@@ -132,6 +133,7 @@ protected:
 private:
     mutable QSize m_textureSize;
     mutable bool m_horizontalBlur;
+    mutable bool m_singlePass;
 
     mutable bool m_haveCached;
     mutable int m_cachedRadius;
@@ -298,123 +300,155 @@ bool QGLPixmapConvolutionFilter::processGL(QPainter *painter, const QPointF &pos
     return true;
 }
 
-static const char *qt_gl_blur_filter_fast =
-    "const int samples = 9;"
-    "uniform mediump vec2 delta;"
-    "lowp vec4 customShader(lowp sampler2D src, highp vec2 srcCoords) {"
-    "    mediump vec4 color = vec4(0.0, 0.0, 0.0, 0.0);"
-    "    mediump float offset = (float(samples) - 1.0) / 2.0;"
-    "    for (int i = 0; i < samples; i++) {"
-    "        mediump vec2 coord = srcCoords + delta * (offset - float(i)) / offset;"
-    "        color += texture2D(src, coord);"
-    "    }"
-    "    return color * (1.0 / float(samples));"
-    "}";
-
-static const char *qt_gl_drop_shadow_filter_fast =
-    "const int samples = 9;"
-    "uniform mediump vec2 delta;"
-    "uniform mediump vec4 shadowColor;"
-    "lowp vec4 customShader(lowp sampler2D src, highp vec2 srcCoords) {"
-    "    mediump vec4 color = vec4(0.0, 0.0, 0.0, 0.0);"
-    "    mediump float offset = (float(samples) - 1.0) / 2.0;"
-    "    for (int i = 0; i < samples; i++) {"
-    "        mediump vec2 coord = srcCoords + delta * (offset - float(i)) / offset;"
-    "        color += texture2D(src, coord).a * shadowColor;"
-    "    }"
-    "    return color * (1.0 / float(samples));"
-    "}";
+static const char *qt_gl_texture_sampling_helper =
+    "lowp float texture2DAlpha(lowp sampler2D src, highp vec2 srcCoords) {\n"
+    "   return texture2D(src, srcCoords).a;\n"
+    "}\n";
+
+static const char *qt_gl_clamped_texture_sampling_helper =
+    "highp vec4 texture_dimensions;\n" // x = width, y = height, z = 0.5/width, w = 0.5/height
+    "lowp float clampedTexture2DAlpha(lowp sampler2D src, highp vec2 srcCoords) {\n"
+    "   highp vec2 clampedCoords = clamp(srcCoords, texture_dimensions.zw, vec2(1.0) - texture_dimensions.zw);\n"
+    "   highp vec2 t = clamp(min(srcCoords, vec2(1.0) - srcCoords) * srcDim + 0.5, 0.0, 1.0);\n"
+    "   return texture2D(src, clampedCoords).a * t.x * t.y;\n"
+    "}\n"
+    "lowp vec4 clampedTexture2D(lowp sampler2D src, highp vec2 srcCoords) {\n"
+    "   highp vec2 clampedCoords = clamp(srcCoords, texture_dimensions.zw, vec2(1.0) - texture_dimensions.zw);\n"
+    "   highp vec2 t = clamp(min(srcCoords, vec2(1.0) - srcCoords) * srcDim + 0.5, 0.0, 1.0);\n"
+    "   return texture2D(src, clampedCoords) * t.x * t.y;\n"
+    "}\n";
+
+static QByteArray qt_gl_convertToClamped(const QByteArray &source)
+{
+    QByteArray result;
+    result.append(qt_gl_clamped_texture_sampling_helper);
+    result.append(QByteArray(source).replace("texture2DAlpha", "clampedTexture2DAlpha")
+                                    .replace("texture2D", "clampedTexture2D"));
+    return result;
+}
 
 QGLPixmapBlurFilter::QGLPixmapBlurFilter(Qt::RenderHint hint)
     : m_haveCached(false)
-    , m_cachedRadius(5)
+    , m_cachedRadius(0)
     , m_hint(hint)
 {
-    if (hint == Qt::PerformanceHint) {
-        QGLPixmapBlurFilter *filter = const_cast<QGLPixmapBlurFilter *>(this);
-        filter->setSource(qt_gl_blur_filter_fast);
-        m_haveCached = true;
-    }
 }
 
 bool QGLPixmapBlurFilter::processGL(QPainter *painter, const QPointF &pos, const QPixmap &src, const QRectF &) const
 {
     QGLPixmapBlurFilter *filter = const_cast<QGLPixmapBlurFilter *>(this);
 
-    int radius = qRound(this->radius());
-    if (!m_haveCached || (m_hint == Qt::QualityHint && radius != m_cachedRadius)) {
-        // Only regenerate the shader from source if parameters have changed.
-        m_haveCached = true;
-        m_cachedRadius = radius;
-        filter->setSource(generateGaussianShader(radius));
+    int actualRadius = qRound(radius());
+    int filterRadius = actualRadius;
+    int fastRadii[] = { 1, 2, 3, 5, 8, 15, 25 };
+    if (m_hint == Qt::PerformanceHint) {
+        uint i = 0;
+        for (; i < (sizeof(fastRadii)/sizeof(*fastRadii))-1; ++i) {
+            if (fastRadii[i+1] > filterRadius)
+                break;
+        }
+        filterRadius = fastRadii[i];
     }
 
-    QGLFramebufferObjectFormat format;
-    format.setInternalTextureFormat(GLenum(src.hasAlphaChannel() ? GL_RGBA : GL_RGB));
-    QGLFramebufferObject *fbo = qgl_fbo_pool()->acquire(src.size(), format);
+    m_singlePass = filterRadius <= 3;
 
-    if (!fbo)
-        return false;
+    if (!m_haveCached || filterRadius != m_cachedRadius) {
+        // Only regenerate the shader from source if parameters have changed.
+        m_haveCached = true;
+        m_cachedRadius = filterRadius;
+        QByteArray source = generateGaussianShader(filterRadius, m_singlePass);
+        filter->setSource(source);
+    }
 
-    glBindTexture(GL_TEXTURE_2D, fbo->texture());
+    QRect targetRect = QRectF(src.rect()).translated(pos).adjusted(-actualRadius, -actualRadius, actualRadius, actualRadius).toAlignedRect();
 
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
-    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
-    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
-    glBindTexture(GL_TEXTURE_2D, 0);
+    if (m_singlePass) {
+        // prepare for updateUniforms
+        m_textureSize = src.size();
 
-    // prepare for updateUniforms
-    m_textureSize = src.size();
+        // ensure GL_LINEAR filtering is used
+        painter->setRenderHint(QPainter::SmoothPixmapTransform);
+        filter->setOnPainter(painter);
+        QBrush pixmapBrush = src;
+        pixmapBrush.setTransform(QTransform::fromTranslate(pos.x(), pos.y()));
+        painter->fillRect(targetRect, pixmapBrush);
+        filter->removeFromPainter(painter);
+    } else {
+        QGLFramebufferObjectFormat format;
+        format.setInternalTextureFormat(GLenum(src.hasAlphaChannel() ? GL_RGBA : GL_RGB));
+        QGLFramebufferObject *fbo = qgl_fbo_pool()->acquire(targetRect.size(), format);
 
-    // horizontal pass, to pixmap
-    m_horizontalBlur = true;
+        if (!fbo)
+            return false;
 
-    QPainter fboPainter(fbo);
+        glBindTexture(GL_TEXTURE_2D, fbo->texture());
 
-    if (src.hasAlphaChannel()) {
-        glClearColor(0, 0, 0, 0);
-        glClear(GL_COLOR_BUFFER_BIT);
-    }
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+        glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+        glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+        glBindTexture(GL_TEXTURE_2D, 0);
 
-    // ensure GL_LINEAR filtering is used
-    fboPainter.setRenderHint(QPainter::SmoothPixmapTransform);
-    filter->setOnPainter(&fboPainter);
-    fboPainter.drawPixmap(0, 0, src);
-    filter->removeFromPainter(&fboPainter);
-    fboPainter.end();
+        // prepare for updateUniforms
+        m_textureSize = src.size();
 
-    QGL2PaintEngineEx *engine = static_cast<QGL2PaintEngineEx *>(painter->paintEngine());
+        // horizontal pass, to pixmap
+        m_horizontalBlur = true;
 
-    // vertical pass, to painter
-    m_horizontalBlur = false;
+        QPainter fboPainter(fbo);
 
-    painter->save();
-    // ensure GL_LINEAR filtering is used
-    painter->setRenderHint(QPainter::SmoothPixmapTransform);
-    filter->setOnPainter(painter);
-    engine->drawTexture(src.rect().translated(pos.x(), pos.y()), fbo->texture(), fbo->size(), src.rect().translated(0, fbo->height() - src.height()));
-    filter->removeFromPainter(painter);
-    painter->restore();
+        if (src.hasAlphaChannel()) {
+            glClearColor(0, 0, 0, 0);
+            glClear(GL_COLOR_BUFFER_BIT);
+        }
 
-    qgl_fbo_pool()->release(fbo);
+        // ensure GL_LINEAR filtering is used
+        fboPainter.setRenderHint(QPainter::SmoothPixmapTransform);
+        filter->setOnPainter(&fboPainter);
+        QBrush pixmapBrush = src;
+        pixmapBrush.setTransform(QTransform::fromTranslate(actualRadius, actualRadius));
+        fboPainter.fillRect(QRect(0, 0, targetRect.width(), targetRect.height()), pixmapBrush);
+        filter->removeFromPainter(&fboPainter);
+        fboPainter.end();
+
+        QGL2PaintEngineEx *engine = static_cast<QGL2PaintEngineEx *>(painter->paintEngine());
+
+        // vertical pass, to painter
+        m_horizontalBlur = false;
+        m_textureSize = fbo->size();
+
+        painter->save();
+        // ensure GL_LINEAR filtering is used
+        painter->setRenderHint(QPainter::SmoothPixmapTransform);
+        filter->setOnPainter(painter);
+        engine->drawTexture(targetRect, fbo->texture(), fbo->size(), QRect(QPoint(), targetRect.size()).translated(0, fbo->height() - targetRect.height()));
+        filter->removeFromPainter(painter);
+        painter->restore();
+
+        qgl_fbo_pool()->release(fbo);
+    }
 
     return true;
 }
 
 void QGLPixmapBlurFilter::setUniforms(QGLShaderProgram *program)
 {
+    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+
     if (m_hint == Qt::QualityHint) {
-        if (m_horizontalBlur)
+        if (m_singlePass)
+            program->setUniformValue("delta", 1.0 / m_textureSize.width(), 1.0 / m_textureSize.height());
+        else if (m_horizontalBlur)
             program->setUniformValue("delta", 1.0 / m_textureSize.width(), 0.0);
         else
             program->setUniformValue("delta", 0.0, 1.0 / m_textureSize.height());
     } else {
-        // 1.4 is chosen to most closely match the blurriness of the gaussian blur
-        // at low radii
-        qreal blur = radius() / 1.4f;
+        qreal blur = radius() / qreal(m_cachedRadius);
 
-        if (m_horizontalBlur)
+        if (m_singlePass)
+            program->setUniformValue("delta", blur / m_textureSize.width(), blur / m_textureSize.height());
+        else if (m_horizontalBlur)
             program->setUniformValue("delta", blur / m_textureSize.width(), 0.0);
         else
             program->setUniformValue("delta", 0.0, blur / m_textureSize.height());
@@ -426,12 +460,21 @@ static inline qreal gaussian(qreal dx, qreal sigma)
     return exp(-dx * dx / (2 * sigma * sigma)) / (Q_2PI * sigma * sigma);
 }
 
-QByteArray QGLPixmapBlurFilter::generateGaussianShader(int radius, bool dropShadow)
+QByteArray QGLPixmapBlurFilter::generateGaussianShader(int radius, bool singlePass, bool dropShadow)
 {
     Q_ASSERT(radius >= 1);
 
+    radius = qMin(127, radius);
+
+    static QCache<uint, QByteArray> shaderSourceCache;
+    uint key = radius | (int(singlePass) << 7) | (int(dropShadow) << 8);
+    QByteArray *cached = shaderSourceCache.object(key);
+    if (cached)
+        return *cached;
+
     QByteArray source;
     source.reserve(1000);
+    source.append(qt_gl_texture_sampling_helper);
 
     source.append("uniform highp vec2      delta;\n");
     if (dropShadow)
@@ -446,7 +489,7 @@ QByteArray QGLPixmapBlurFilter::generateGaussianShader(int radius, bool dropShad
     qreal sigma = radius / 1.65;
 
     qreal sum = 0;
-    for (int i = -radius; i <= radius; ++i) {
+    for (int i = -radius; i < radius; ++i) {
         float value = gaussian(i, sigma);
         gaussianComponents << value;
         sum += value;
@@ -464,43 +507,67 @@ QByteArray QGLPixmapBlurFilter::generateGaussianShader(int radius, bool dropShad
         weights << weight;
     }
 
-    // odd size ?
-    if (gaussianComponents.size() & 1) {
-        sampleOffsets << radius;
-        weights << gaussianComponents.last();
-    }
-
-    int currentVariable = 1;
-    source.append("    mediump vec4 sample = vec4(0.0);\n");
-    source.append("    mediump vec2 coord;\n");
-
-    qreal weightSum = 0;
-    source.append("    mediump float c;\n");
-    for (int i = 0; i < sampleOffsets.size(); ++i) {
-        qreal delta = sampleOffsets.at(i);
-
-        ++currentVariable;
+    int limit = sampleOffsets.size();
+    if (singlePass)
+        limit *= limit;
+
+    QByteArray baseCoordinate = "srcCoords";
+
+    for (int i = 0; i < limit; ++i) {
+        QByteArray coordinate = baseCoordinate;
+
+        qreal weight;
+        if (singlePass) {
+            const int xIndex = i % sampleOffsets.size();
+            const int yIndex = i / sampleOffsets.size();
+
+            const qreal deltaX = sampleOffsets.at(xIndex);
+            const qreal deltaY = sampleOffsets.at(yIndex);
+            weight = weights.at(xIndex) * weights.at(yIndex);
+
+            if (!qFuzzyCompare(deltaX, deltaY)) {
+                coordinate.append(" + vec2(delta.x * float(");
+                coordinate.append(QByteArray::number(deltaX));
+                coordinate.append("), delta.y * float(");
+                coordinate.append(QByteArray::number(deltaY));
+                coordinate.append("))");
+            } else if (!qFuzzyIsNull(deltaX)) {
+                coordinate.append(" + delta * float(");
+                coordinate.append(QByteArray::number(deltaX));
+                coordinate.append(")");
+            }
+        } else {
+            const qreal delta = sampleOffsets.at(i);
+            weight = weights.at(i);
+            if (!qFuzzyIsNull(delta)) {
+                coordinate.append(" + delta * float(");
+                coordinate.append(QByteArray::number(delta));
+                coordinate.append(")");
+            }
+        }
 
-        QByteArray coordinate = "srcCoords";
-        if (delta != qreal(0)) {
-            coordinate.append(" + delta * float(");
-            coordinate.append(QByteArray::number(delta));
-            coordinate.append(")");
+        if (i == 0) {
+            if (dropShadow)
+                source.append("    mediump float sample = ");
+            else
+                source.append("    mediump vec4 sample = ");
+        } else {
+            if (dropShadow)
+                source.append("    sample += ");
+            else
+                source.append("    sample += ");
         }
 
-        source.append("    coord = ");
+        source.append("texture2D(src, ");
         source.append(coordinate);
-        source.append(";\n");
+        source.append(")");
 
         if (dropShadow)
-            source.append("    sample += texture2D(src, coord).a * shadowColor");
-        else
-            source.append("    sample += texture2D(src, coord)");
+            source.append(".a");
 
-        weightSum += weights.at(i);
-        if (weights.at(i) != qreal(1)) {
+        if (!qFuzzyCompare(weight, qreal(1))) {
             source.append(" * float(");
-            source.append(QByteArray::number(weights.at(i)));
+            source.append(QByteArray::number(weight));
             source.append(");\n");
         } else {
             source.append(";\n");
@@ -508,86 +575,109 @@ QByteArray QGLPixmapBlurFilter::generateGaussianShader(int radius, bool dropShad
     }
 
     source.append("    return ");
+    if (dropShadow)
+        source.append("shadowColor * ");
     source.append("sample;\n");
     source.append("}\n");
 
+    cached = new QByteArray(source);
+    shaderSourceCache.insert(key, cached);
+
     return source;
 }
 
 QGLPixmapDropShadowFilter::QGLPixmapDropShadowFilter(Qt::RenderHint hint)
     : m_haveCached(false)
-    , m_cachedRadius(5)
+    , m_cachedRadius(0)
     , m_hint(hint)
 {
-    if (hint == Qt::PerformanceHint) {
-        QGLPixmapDropShadowFilter *filter = const_cast<QGLPixmapDropShadowFilter *>(this);
-        filter->setSource(qt_gl_drop_shadow_filter_fast);
-        m_haveCached = true;
-    }
 }
 
 bool QGLPixmapDropShadowFilter::processGL(QPainter *painter, const QPointF &pos, const QPixmap &src, const QRectF &srcRect) const
 {
     QGLPixmapDropShadowFilter *filter = const_cast<QGLPixmapDropShadowFilter *>(this);
 
-    int radius = qRound(this->blurRadius());
-    if (!m_haveCached || (m_hint == Qt::QualityHint && radius != m_cachedRadius)) {
+    int actualRadius = qRound(blurRadius());
+    int filterRadius = actualRadius;
+    m_singlePass = filterRadius <= 3;
+
+    if (!m_haveCached || filterRadius != m_cachedRadius) {
         // Only regenerate the shader from source if parameters have changed.
         m_haveCached = true;
-        m_cachedRadius = radius;
-        filter->setSource(QGLPixmapBlurFilter::generateGaussianShader(radius, true));
+        m_cachedRadius = filterRadius;
+        QByteArray source = QGLPixmapBlurFilter::generateGaussianShader(filterRadius, m_singlePass, true);
+        filter->setSource(source);
     }
 
-    QGLFramebufferObjectFormat format;
-    format.setInternalTextureFormat(GLenum(src.hasAlphaChannel() ? GL_RGBA : GL_RGB));
-    QGLFramebufferObject *fbo = qgl_fbo_pool()->acquire(src.size(), format);
-
-    if (!fbo)
-        return false;
-
-    glBindTexture(GL_TEXTURE_2D, fbo->texture());
-
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
-    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
-    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
-    glBindTexture(GL_TEXTURE_2D, 0);
+    QRect targetRect = QRectF(src.rect()).translated(pos + offset()).adjusted(-actualRadius, -actualRadius, actualRadius, actualRadius).toAlignedRect();
+
+    if (m_singlePass) {
+        // prepare for updateUniforms
+        m_textureSize = src.size();
+
+        painter->save();
+        // ensure GL_LINEAR filtering is used
+        painter->setRenderHint(QPainter::SmoothPixmapTransform);
+        filter->setOnPainter(painter);
+        QBrush pixmapBrush = src;
+        pixmapBrush.setTransform(QTransform::fromTranslate(pos.x() + offset().x(), pos.y() + offset().y()));
+        painter->fillRect(targetRect, pixmapBrush);
+        filter->removeFromPainter(painter);
+        painter->restore();
+    } else {
+        QGLFramebufferObjectFormat format;
+        format.setInternalTextureFormat(GLenum(src.hasAlphaChannel() ? GL_RGBA : GL_RGB));
+        QGLFramebufferObject *fbo = qgl_fbo_pool()->acquire(targetRect.size(), format);
 
-    // prepare for updateUniforms
-    m_textureSize = src.size();
+        if (!fbo)
+            return false;
 
-    // horizontal pass, to pixmap
-    m_horizontalBlur = true;
+        glBindTexture(GL_TEXTURE_2D, fbo->texture());
 
-    QPainter fboPainter(fbo);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+        glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+        glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+        glBindTexture(GL_TEXTURE_2D, 0);
 
-    if (src.hasAlphaChannel()) {
-        glClearColor(0, 0, 0, 0);
-        glClear(GL_COLOR_BUFFER_BIT);
-    }
+        // prepare for updateUniforms
+        m_textureSize = src.size();
 
-    // ensure GL_LINEAR filtering is used
-    fboPainter.setRenderHint(QPainter::SmoothPixmapTransform);
-    filter->setOnPainter(&fboPainter);
-    fboPainter.drawPixmap(0, 0, src);
-    filter->removeFromPainter(&fboPainter);
-    fboPainter.end();
+        // horizontal pass, to pixmap
+        m_horizontalBlur = true;
 
-    QGL2PaintEngineEx *engine = static_cast<QGL2PaintEngineEx *>(painter->paintEngine());
+        QPainter fboPainter(fbo);
 
-    // vertical pass, to painter
-    m_horizontalBlur = false;
-
-    painter->save();
-    // ensure GL_LINEAR filtering is used
-    painter->setRenderHint(QPainter::SmoothPixmapTransform);
-    filter->setOnPainter(painter);
-    QPointF ofs = offset();
-    engine->drawTexture(src.rect().translated(pos.x() + ofs.x(), pos.y() + ofs.y()), fbo->texture(), fbo->size(), src.rect().translated(0, fbo->height() - src.height()));
-    filter->removeFromPainter(painter);
-    painter->restore();
+        if (src.hasAlphaChannel()) {
+            glClearColor(0, 0, 0, 0);
+            glClear(GL_COLOR_BUFFER_BIT);
+        }
 
-    qgl_fbo_pool()->release(fbo);
+        // ensure GL_LINEAR filtering is used
+        fboPainter.setRenderHint(QPainter::SmoothPixmapTransform);
+        filter->setOnPainter(&fboPainter);
+        QBrush pixmapBrush = src;
+        pixmapBrush.setTransform(QTransform::fromTranslate(actualRadius, actualRadius));
+        fboPainter.fillRect(QRect(0, 0, targetRect.width(), targetRect.height()), pixmapBrush);
+        filter->removeFromPainter(&fboPainter);
+        fboPainter.end();
+
+        QGL2PaintEngineEx *engine = static_cast<QGL2PaintEngineEx *>(painter->paintEngine());
+
+        // vertical pass, to painter
+        m_horizontalBlur = false;
+        m_textureSize = fbo->size();
+
+        painter->save();
+        // ensure GL_LINEAR filtering is used
+        painter->setRenderHint(QPainter::SmoothPixmapTransform);
+        filter->setOnPainter(painter);
+        engine->drawTexture(targetRect, fbo->texture(), fbo->size(), src.rect().translated(0, fbo->height() - src.height()));
+        filter->removeFromPainter(painter);
+        painter->restore();
+
+        qgl_fbo_pool()->release(fbo);
+    }
 
     // Now draw the actual pixmap over the top.
     painter->drawPixmap(pos, src, srcRect);
@@ -597,8 +687,11 @@ bool QGLPixmapDropShadowFilter::processGL(QPainter *painter, const QPointF &pos,
 
 void QGLPixmapDropShadowFilter::setUniforms(QGLShaderProgram *program)
 {
+    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+
     QColor col = color();
-    if (m_horizontalBlur) {
+    if (m_horizontalBlur && !m_singlePass) {
         program->setUniformValue("shadowColor", 1.0f, 1.0f, 1.0f, 1.0f);
     } else {
         qreal alpha = col.alphaF();
@@ -607,17 +700,20 @@ void QGLPixmapDropShadowFilter::setUniforms(QGLShaderProgram *program)
                                                 col.blueF() * alpha,
                                                 alpha);
     }
+
     if (m_hint == Qt::QualityHint) {
-        if (m_horizontalBlur)
+        if (m_singlePass)
+            program->setUniformValue("delta", 1.0 / m_textureSize.width(), 1.0 / m_textureSize.height());
+        else if (m_horizontalBlur)
             program->setUniformValue("delta", 1.0 / m_textureSize.width(), 0.0);
         else
             program->setUniformValue("delta", 0.0, 1.0 / m_textureSize.height());
     } else {
-        // 1.4 is chosen to most closely match the blurriness of the gaussian blur
-        // at low radii
-        qreal blur = blurRadius() / 1.4f;
+        qreal blur = blurRadius() / qreal(m_cachedRadius);
 
-        if (m_horizontalBlur)
+        if (m_singlePass)
+            program->setUniformValue("delta", blur / m_textureSize.width(), blur / m_textureSize.height());
+        else if (m_horizontalBlur)
             program->setUniformValue("delta", blur / m_textureSize.width(), 0.0);
         else
             program->setUniformValue("delta", 0.0, blur / m_textureSize.height());