Implemented faster image transformation in the raster engine.

Task-number: 245650
Reviewed-by: Gunnar

3 files changed, 664 insertions, 13 deletions

diff --git a/src/gui/painting/qblendfunctions.cpp b/src/gui/painting/qblendfunctions.cpp
index 1121c0e..e447301 100644
--- a/src/gui/painting/qblendfunctions.cpp
+++ b/src/gui/painting/qblendfunctions.cpp
@@ -793,8 +793,351 @@ void qt_scale_image_argb32_on_argb32(uchar *destPixels, int dbpl,
     }
 }
 
+struct QTransformImageVertex
+{
+    qreal x, y, u, v; // destination coordinates (x, y) and source coordinates (u, v)
+};
+
+template <class SrcT, class DestT, class Blender>
+void qt_transform_image_rasterize(DestT *destPixels, int dbpl,
+                                  const SrcT *srcPixels, int sbpl,
+                                  const QTransformImageVertex &topLeft, const QTransformImageVertex &bottomLeft,
+                                  const QTransformImageVertex &topRight, const QTransformImageVertex &bottomRight,
+                                  const QRect &sourceRect,
+                                  const QRect &clip,
+                                  qreal topY, qreal bottomY,
+                                  int dudx, int dvdx, int dudy, int dvdy, int u0, int v0,
+                                  Blender blender)
+{
+    int fromY = qMax(qRound(topY), clip.top());
+    int toY = qMin(qRound(bottomY), clip.top() + clip.height());
+    if (fromY >= toY)
+        return;
+
+    qreal leftSlope = (bottomLeft.x - topLeft.x) / (bottomLeft.y - topLeft.y);
+    qreal rightSlope = (bottomRight.x - topRight.x) / (bottomRight.y - topRight.y);
+    int dx_l = int(leftSlope * 0x10000);
+    int dx_r = int(rightSlope * 0x10000);
+    int x_l = int((topLeft.x + (0.5 + fromY - topLeft.y) * leftSlope + 0.5) * 0x10000);
+    int x_r = int((topRight.x + (0.5 + fromY - topRight.y) * rightSlope + 0.5) * 0x10000);
+
+    int fromX, toX, x1, x2, u, v, i, ii;
+    DestT *line;
+    for (int y = fromY; y < toY; ++y) {
+        line = reinterpret_cast<DestT *>(reinterpret_cast<uchar *>(destPixels) + y * dbpl);
+
+        fromX = qMax(x_l >> 16, clip.left());
+        toX = qMin(x_r >> 16, clip.left() + clip.width());
+        if (fromX < toX) {
+            // Because of rounding, we can get source coordinates outside the source image.
+            // Clamp these coordinates to the source rect to avoid segmentation fault and
+            // garbage on the screen.
+
+            // Find the first pixel on the current scan line where the source coordinates are within the source rect.
+            x1 = fromX;
+            u = x1 * dudx + y * dudy + u0;
+            v = x1 * dvdx + y * dvdy + v0;
+            for (; x1 < toX; ++x1) {
+                int uu = u >> 16;
+                int vv = v >> 16;
+                if (uu >= sourceRect.left() && uu < sourceRect.left() + sourceRect.width()
+                    && vv >= sourceRect.top() && vv < sourceRect.top() + sourceRect.height()) {
+                    break;
+                }
+                u += dudx;
+                v += dvdx;
+            }
+
+            // Find the last pixel on the current scan line where the source coordinates are within the source rect.
+            x2 = toX;
+            u = (x2 - 1) * dudx + y * dudy + u0;
+            v = (x2 - 1) * dvdx + y * dvdy + v0;
+            for (; x2 > x1; --x2) {
+                int uu = u >> 16;
+                int vv = v >> 16;
+                if (uu >= sourceRect.left() && uu < sourceRect.left() + sourceRect.width()
+                    && vv >= sourceRect.top() && vv < sourceRect.top() + sourceRect.height()) {
+                    break;
+                }
+                u -= dudx;
+                v -= dvdx;
+            }
+
+            // Set up values at the beginning of the scan line.
+            u = fromX * dudx + y * dudy + u0;
+            v = fromX * dvdx + y * dvdy + v0;
+            line += fromX;
+
+            // Beginning of the scan line, with per-pixel checks.
+            i = x1 - fromX;
+            while (i) {
+                int uu = qBound(sourceRect.left(), u >> 16, sourceRect.left() + sourceRect.width() - 1);
+                int vv = qBound(sourceRect.top(), v >> 16, sourceRect.top() + sourceRect.height() - 1);
+                blender.write(line, reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + vv * sbpl)[uu]);
+                u += dudx;
+                v += dvdx;
+                ++line;
+                --i;
+            }
+
+            // Middle of the scan line, without checks.
+            // Manual loop unrolling.
+            i = x2 - x1;
+            ii = i >> 3;
+            while (ii) {
+                blender.write(&line[0], reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx;
+                blender.write(&line[1], reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx;
+                blender.write(&line[2], reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx;
+                blender.write(&line[3], reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx;
+                blender.write(&line[4], reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx;
+                blender.write(&line[5], reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx;
+                blender.write(&line[6], reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx;
+                blender.write(&line[7], reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx;
+                line += 8;
+                --ii;
+            }
+            switch (i & 7) {
+                case 7: blender.write(line, reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx; ++line;
+                case 6: blender.write(line, reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx; ++line;
+                case 5: blender.write(line, reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx; ++line;
+                case 4: blender.write(line, reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx; ++line;
+                case 3: blender.write(line, reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx; ++line;
+                case 2: blender.write(line, reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx; ++line;
+                case 1: blender.write(line, reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + (v >> 16) * sbpl)[u >> 16]); u += dudx; v += dvdx; ++line;
+            }
+
+            // End of the scan line, with per-pixel checks.
+            i = toX - x2;
+            while (i) {
+                int uu = qBound(sourceRect.left(), u >> 16, sourceRect.left() + sourceRect.width() - 1);
+                int vv = qBound(sourceRect.top(), v >> 16, sourceRect.top() + sourceRect.height() - 1);
+                blender.write(line, reinterpret_cast<const SrcT *>(reinterpret_cast<const uchar *>(srcPixels) + vv * sbpl)[uu]);
+                u += dudx;
+                v += dvdx;
+                ++line;
+                --i;
+            }
+        }
+        x_l += dx_l;
+        x_r += dx_r;
+    }
+}
+
+template <class SrcT, class DestT, class Blender>
+void qt_transform_image(DestT *destPixels, int dbpl,
+                        const SrcT *srcPixels, int sbpl,
+                        const QRectF &targetRect,
+                        const QRectF &sourceRect,
+                        const QRect &clip,
+                        const QTransform &targetRectTransform,
+                        Blender blender)
+{
+    enum Corner
+    {
+        TopLeft,
+        TopRight,
+        BottomRight,
+        BottomLeft
+    };
+
+    // map source rectangle to destination.
+    QTransformImageVertex v[4];
+    v[TopLeft].u = v[BottomLeft].u = sourceRect.left();
+    v[TopLeft].v = v[TopRight].v = sourceRect.top();
+    v[TopRight].u = v[BottomRight].u = sourceRect.right();
+    v[BottomLeft].v = v[BottomRight].v = sourceRect.bottom();
+    targetRectTransform.map(targetRect.left(), targetRect.top(), &v[TopLeft].x, &v[TopLeft].y);
+    targetRectTransform.map(targetRect.right(), targetRect.top(), &v[TopRight].x, &v[TopRight].y);
+    targetRectTransform.map(targetRect.left(), targetRect.bottom(), &v[BottomLeft].x, &v[BottomLeft].y);
+    targetRectTransform.map(targetRect.right(), targetRect.bottom(), &v[BottomRight].x, &v[BottomRight].y);
+
+    // find topmost vertex.
+    int topmost = 0;
+    for (int i = 1; i < 4; ++i) {
+        if (v[i].y < v[topmost].y)
+            topmost = i;
+    }
+    // rearrange array such that topmost vertex is at index 0.
+    switch (topmost) {
+    case 1:
+        {
+            QTransformImageVertex t = v[0];
+            for (int i = 0; i < 3; ++i)
+                v[i] = v[i+1];
+            v[3] = t;
+        }
+        break;
+    case 2:
+        qSwap(v[0], v[2]);
+        qSwap(v[1], v[3]);
+        break;
+    case 3:
+        {
+            QTransformImageVertex t = v[3];
+            for (int i = 3; i > 0; --i)
+                v[i] = v[i-1];
+            v[0] = t;
+        }
+        break;
+    }
+
+    // if necessary, swap vertex 1 and 3 such that 1 is to the left of 3.
+    qreal dx1 = v[1].x - v[0].x;
+    qreal dy1 = v[1].y - v[0].y;
+    qreal dx2 = v[3].x - v[0].x;
+    qreal dy2 = v[3].y - v[0].y;
+    if (dx1 * dy2 - dx2 * dy1 > 0)
+        qSwap(v[1], v[3]);
+
+    QTransformImageVertex u = {v[1].x - v[0].x, v[1].y - v[0].y, v[1].u - v[0].u, v[1].v - v[0].v};
+    QTransformImageVertex w = {v[2].x - v[0].x, v[2].y - v[0].y, v[2].u - v[0].u, v[2].v - v[0].v};
+
+    qreal det = u.x * w.y - u.y * w.x;
+    if (det == 0)
+        return;
+
+    qreal invDet = 1.0 / det;
+    qreal m11, m12, m21, m22, mdx, mdy;
+
+    m11 = (u.u * w.y - u.y * w.u) * invDet;
+    m12 = (u.x * w.u - u.u * w.x) * invDet;
+    m21 = (u.v * w.y - u.y * w.v) * invDet;
+    m22 = (u.x * w.v - u.v * w.x) * invDet;
+    mdx = v[0].u - m11 * v[0].x - m12 * v[0].y;
+    mdy = v[0].v - m21 * v[0].x - m22 * v[0].y;
+
+    int dudx = int(m11 * 0x10000);
+    int dvdx = int(m21 * 0x10000);
+    int dudy = int(m12 * 0x10000);
+    int dvdy = int(m22 * 0x10000);
+    int u0 = qCeil((0.5 * m11 + 0.5 * m12 + mdx) * 0x10000) - 1;
+    int v0 = qCeil((0.5 * m21 + 0.5 * m22 + mdy) * 0x10000) - 1;
+
+    int x1 = qFloor(sourceRect.left());
+    int y1 = qFloor(sourceRect.top());
+    int x2 = qCeil(sourceRect.right());
+    int y2 = qCeil(sourceRect.bottom());
+    QRect sourceRectI(x1, y1, x2 - x1, y2 - y1);
+
+    // rasterize trapezoids.
+    if (v[1].y < v[3].y) {
+        qt_transform_image_rasterize(destPixels, dbpl, srcPixels, sbpl, v[0], v[1], v[0], v[3], sourceRectI, clip, v[0].y, v[1].y, dudx, dvdx, dudy, dvdy, u0, v0, blender);
+        qt_transform_image_rasterize(destPixels, dbpl, srcPixels, sbpl, v[1], v[2], v[0], v[3], sourceRectI, clip, v[1].y, v[3].y, dudx, dvdx, dudy, dvdy, u0, v0, blender);
+        qt_transform_image_rasterize(destPixels, dbpl, srcPixels, sbpl, v[1], v[2], v[3], v[2], sourceRectI, clip, v[3].y, v[2].y, dudx, dvdx, dudy, dvdy, u0, v0, blender);
+    } else {
+        qt_transform_image_rasterize(destPixels, dbpl, srcPixels, sbpl, v[0], v[1], v[0], v[3], sourceRectI, clip, v[0].y, v[3].y, dudx, dvdx, dudy, dvdy, u0, v0, blender);
+        qt_transform_image_rasterize(destPixels, dbpl, srcPixels, sbpl, v[0], v[1], v[3], v[2], sourceRectI, clip, v[3].y, v[1].y, dudx, dvdx, dudy, dvdy, u0, v0, blender);
+        qt_transform_image_rasterize(destPixels, dbpl, srcPixels, sbpl, v[1], v[2], v[3], v[2], sourceRectI, clip, v[1].y, v[2].y, dudx, dvdx, dudy, dvdy, u0, v0, blender);
+    }
+}
+
+void qt_transform_image_rgb16_on_rgb16(uchar *destPixels, int dbpl,
+                                       const uchar *srcPixels, int sbpl,
+                                       const QRectF &targetRect,
+                                       const QRectF &sourceRect,
+                                       const QRect &clip,
+                                       const QTransform &targetRectTransform,
+                                       int const_alpha)
+{
+    if (const_alpha == 256) {
+        Blend_RGB16_on_RGB16_NoAlpha noAlpha;
+        qt_transform_image(reinterpret_cast<quint16 *>(destPixels), dbpl,
+                           reinterpret_cast<const quint16 *>(srcPixels), sbpl,
+                           targetRect, sourceRect, clip, targetRectTransform, noAlpha);
+    } else {
+        Blend_RGB16_on_RGB16_ConstAlpha constAlpha(const_alpha);
+        qt_transform_image(reinterpret_cast<quint16 *>(destPixels), dbpl,
+                           reinterpret_cast<const quint16 *>(srcPixels), sbpl,
+                           targetRect, sourceRect, clip, targetRectTransform, constAlpha);
+    }
+}
+
+void qt_transform_image_argb24_on_rgb16(uchar *destPixels, int dbpl,
+                                        const uchar *srcPixels, int sbpl,
+                                        const QRectF &targetRect,
+                                        const QRectF &sourceRect,
+                                        const QRect &clip,
+                                        const QTransform &targetRectTransform,
+                                        int const_alpha)
+{
+    if (const_alpha == 256) {
+        Blend_ARGB24_on_RGB16_SourceAlpha noAlpha;
+        qt_transform_image(reinterpret_cast<quint16 *>(destPixels), dbpl,
+                           reinterpret_cast<const qargb8565 *>(srcPixels), sbpl,
+                           targetRect, sourceRect, clip, targetRectTransform, noAlpha);
+    } else {
+        Blend_ARGB24_on_RGB16_SourceAndConstAlpha constAlpha(const_alpha);
+        qt_transform_image(reinterpret_cast<quint16 *>(destPixels), dbpl,
+                           reinterpret_cast<const qargb8565 *>(srcPixels), sbpl,
+                           targetRect, sourceRect, clip, targetRectTransform, constAlpha);
+    }
+}
 
 
+void qt_transform_image_argb32_on_rgb16(uchar *destPixels, int dbpl,
+                                        const uchar *srcPixels, int sbpl,
+                                        const QRectF &targetRect,
+                                        const QRectF &sourceRect,
+                                        const QRect &clip,
+                                        const QTransform &targetRectTransform,
+                                        int const_alpha)
+{
+    if (const_alpha == 256) {
+        Blend_ARGB32_on_RGB16_SourceAlpha noAlpha;
+        qt_transform_image(reinterpret_cast<quint16 *>(destPixels), dbpl,
+                           reinterpret_cast<const quint32 *>(srcPixels), sbpl,
+                           targetRect, sourceRect, clip, targetRectTransform, noAlpha);
+    } else {
+        Blend_ARGB32_on_RGB16_SourceAndConstAlpha constAlpha(const_alpha);
+        qt_transform_image(reinterpret_cast<quint16 *>(destPixels), dbpl,
+                           reinterpret_cast<const quint32 *>(srcPixels), sbpl,
+                           targetRect, sourceRect, clip, targetRectTransform, constAlpha);
+    }
+}
+
+
+void qt_transform_image_rgb32_on_rgb32(uchar *destPixels, int dbpl,
+                                       const uchar *srcPixels, int sbpl,
+                                       const QRectF &targetRect,
+                                       const QRectF &sourceRect,
+                                       const QRect &clip,
+                                       const QTransform &targetRectTransform,
+                                       int const_alpha)
+{
+    if (const_alpha == 256) {
+        Blend_RGB32_on_RGB32_NoAlpha noAlpha;
+        qt_transform_image(reinterpret_cast<quint32 *>(destPixels), dbpl,
+                           reinterpret_cast<const quint32 *>(srcPixels), sbpl,
+                           targetRect, sourceRect, clip, targetRectTransform, noAlpha);
+    } else {
+        Blend_RGB32_on_RGB32_ConstAlpha constAlpha(const_alpha);
+        qt_transform_image(reinterpret_cast<quint32 *>(destPixels), dbpl,
+                           reinterpret_cast<const quint32 *>(srcPixels), sbpl,
+                           targetRect, sourceRect, clip, targetRectTransform, constAlpha);
+    }
+}
+
+void qt_transform_image_argb32_on_argb32(uchar *destPixels, int dbpl,
+                                         const uchar *srcPixels, int sbpl,
+                                         const QRectF &targetRect,
+                                         const QRectF &sourceRect,
+                                         const QRect &clip,
+                                         const QTransform &targetRectTransform,
+                                         int const_alpha)
+{
+    if (const_alpha == 256) {
+        Blend_ARGB32_on_ARGB32_SourceAlpha sourceAlpha;
+        qt_transform_image(reinterpret_cast<quint32 *>(destPixels), dbpl,
+                           reinterpret_cast<const quint32 *>(srcPixels), sbpl,
+                           targetRect, sourceRect, clip, targetRectTransform, sourceAlpha);
+    } else {
+        Blend_ARGB32_on_ARGB32_SourceAndConstAlpha constAlpha(const_alpha);
+        qt_transform_image(reinterpret_cast<quint32 *>(destPixels), dbpl,
+                           reinterpret_cast<const quint32 *>(srcPixels), sbpl,
+                           targetRect, sourceRect, clip, targetRectTransform, constAlpha);
+    }
+}
+
 SrcOverScaleFunc qScaleFunctions[QImage::NImageFormats][QImage::NImageFormats] = {
     {   // Format_Invalid
         0,      // Format_Invalid,
@@ -1378,5 +1721,295 @@ SrcOverBlendFunc qBlendFunctions[QImage::NImageFormats][QImage::NImageFormats] =
     }
 };
 
+SrcOverTransformFunc qTransformFunctions[QImage::NImageFormats][QImage::NImageFormats] = {
+    {   // Format_Invalid
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_Mono
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_MonoLSB
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_Indexed8
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_RGB32
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        qt_transform_image_rgb32_on_rgb32,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        qt_transform_image_argb32_on_argb32,    // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_ARGB32
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_ARGB32_Premultiplied
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        qt_transform_image_rgb32_on_rgb32,          // Format_RGB32,
+        0,      // Format_ARGB32,
+        qt_transform_image_argb32_on_argb32,        // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_RGB16
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        qt_transform_image_argb32_on_rgb16,       // Format_ARGB32_Premultiplied,
+        qt_transform_image_rgb16_on_rgb16,        // Format_RGB16,
+        qt_transform_image_argb24_on_rgb16,       // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_ARGB8565_Premultiplied
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_RGB666
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_ARGB6666_Premultiplied
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_RGB555
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_ARGB8555_Premultiplied
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_RGB888
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_RGB444
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    },
+    {   // Format_ARGB4444_Premultiplied
+        0,      // Format_Invalid,
+        0,      // Format_Mono,
+        0,      // Format_MonoLSB,
+        0,      // Format_Indexed8,
+        0,      // Format_RGB32,
+        0,      // Format_ARGB32,
+        0,      // Format_ARGB32_Premultiplied,
+        0,      // Format_RGB16,
+        0,      // Format_ARGB8565_Premultiplied,
+        0,      // Format_RGB666,
+        0,      // Format_ARGB6666_Premultiplied,
+        0,      // Format_RGB555,
+        0,      // Format_ARGB8555_Premultiplied,
+        0,      // Format_RGB888,
+        0,      // Format_RGB444,
+        0       // Format_ARGB4444_Premultiplied,
+    }
+};
 
 QT_END_NAMESPACE
diff --git a/src/gui/painting/qdrawhelper_p.h b/src/gui/painting/qdrawhelper_p.h
index 18c3358..83d2671 100644
--- a/src/gui/painting/qdrawhelper_p.h
+++ b/src/gui/painting/qdrawhelper_p.h
@@ -146,6 +146,14 @@ typedef void (*SrcOverScaleFunc)(uchar *destPixels, int dbpl,
                                  const QRect &clipRect,
                                  int const_alpha);
 
+typedef void (*SrcOverTransformFunc)(uchar *destPixels, int dbpl,
+                                     const uchar *src, int spbl,
+                                     const QRectF &targetRect,
+                                     const QRectF &sourceRect,
+                                     const QRect &clipRect,
+                                     const QTransform &targetRectTransform,
+                                     int const_alpha);
+
 
 struct DrawHelper {
     ProcessSpans blendColor;
@@ -158,6 +166,7 @@ struct DrawHelper {
 
 extern SrcOverBlendFunc qBlendFunctions[QImage::NImageFormats][QImage::NImageFormats];
 extern SrcOverScaleFunc qScaleFunctions[QImage::NImageFormats][QImage::NImageFormats];
+extern SrcOverTransformFunc qTransformFunctions[QImage::NImageFormats][QImage::NImageFormats];
 
 extern DrawHelper qDrawHelper[QImage::NImageFormats];
 
diff --git a/src/gui/painting/qpaintengine_raster.cpp b/src/gui/painting/qpaintengine_raster.cpp
index d00329b..8b83f02 100644
--- a/src/gui/painting/qpaintengine_raster.cpp
+++ b/src/gui/painting/qpaintengine_raster.cpp
@@ -2495,10 +2495,7 @@ void QRasterPaintEngine::drawImage(const QPointF &p, const QImage &img)
         const QClipData *clip = d->clip();
         QPointF pt(p.x() + s->matrix.dx(), p.y() + s->matrix.dy());
 
-        // ### TODO: remove this eventually...
-        static bool NO_BLEND_FUNC = !qgetenv("QT_NO_BLEND_FUNCTIONS").isNull();
-
-        if (s->flags.fast_images && !NO_BLEND_FUNC) {
+        if (s->flags.fast_images) {
             SrcOverBlendFunc func = qBlendFunctions[d->rasterBuffer->format][img.format()];
             if (func) {
                 if (!clip) {
@@ -2511,6 +2508,8 @@ void QRasterPaintEngine::drawImage(const QPointF &p, const QImage &img)
             }
         }
 
+
+
         d->image_filler.clip = clip;
         d->image_filler.initTexture(&img, s->intOpacity, QTextureData::Plain, img.rect());
         if (!d->image_filler.blend)
@@ -2562,14 +2561,24 @@ void QRasterPaintEngine::drawImage(const QRectF &r, const QImage &img, const QRe
     if (s->matrix.type() > QTransform::TxTranslate || stretch_sr) {
 
         if (s->flags.fast_images) {
-            SrcOverScaleFunc func = qScaleFunctions[d->rasterBuffer->format][img.format()];
-            if (func && (!clip || clip->hasRectClip)) {
-                func(d->rasterBuffer->buffer(), d->rasterBuffer->bytesPerLine(),
-                     img.bits(), img.bytesPerLine(),
-                     qt_mapRect_non_normalizing(r, s->matrix), sr,
-                     !clip ? d->deviceRect : clip->clipRect,
-                     s->intOpacity);
-                return;
+            if (s->matrix.type() > QTransform::TxScale) {
+                SrcOverTransformFunc func = qTransformFunctions[d->rasterBuffer->format][img.format()];
+                if (func && (!clip || clip->hasRectClip)) {
+                    func(d->rasterBuffer->buffer(), d->rasterBuffer->bytesPerLine(), img.bits(),
+                         img.bytesPerLine(), r, sr, !clip ? d->deviceRect : clip->clipRect,
+                         s->matrix, s->intOpacity);
+                    return;
+                }
+            } else {
+                SrcOverScaleFunc func = qScaleFunctions[d->rasterBuffer->format][img.format()];
+                if (func && (!clip || clip->hasRectClip)) {
+                    func(d->rasterBuffer->buffer(), d->rasterBuffer->bytesPerLine(),
+                         img.bits(), img.bytesPerLine(),
+                         qt_mapRect_non_normalizing(r, s->matrix), sr,
+                         !clip ? d->deviceRect : clip->clipRect,
+                         s->intOpacity);
+                    return;
+                }
             }
         }
 
@@ -4056,7 +4065,7 @@ void QRasterPaintEnginePrivate::recalculateFastImages()
 
     s->flags.fast_images = !(s->renderHints & QPainter::SmoothPixmapTransform)
                            && rasterBuffer->compositionMode == QPainter::CompositionMode_SourceOver
-                           && s->matrix.type() <= QTransform::TxScale;
+                           && s->matrix.type() <= QTransform::TxShear;
 }