GDI 中的快速投影算法+

Question

在 GDI 中向图像添加阴影的有效方法是什么？

现在我从我的图像开始：

我使用 ImageAttributes 和 ColorMatrix 将图像的 alpha 蒙版绘制到新图像：

colorMatrix = (
    ( 0,  0,  0, 0, 0),
    ( 0,  0,  0, 0, 0),
    ( 0,  0,  0, 0, 0),
    (-1, -1, -1, 1, 0),
    ( 1,  1,  1, 0, 1)
    );

然后我应用高斯模糊卷积核，并稍微偏移它：

然后我将原始图像绘制回顶部：

问题是它太慢了，生成带有 drop-shadow 的图像大约需要 170ms，而没有 drop-shadow 则需要 2ms（慢 70 倍）：

• 带 drop阴影：171,332 µs
• 无投影：2,457us

当用户（例如我）滚动浏览项目列表时，额外的 169 毫秒延迟非常 值得注意。

---

您可以忽略下面的代码，它不会向问题或答案添加任何内容：

class function TImageEffects.GenerateDropShadow(image: TGPImage;
        const radius: Single; const OffsetX, OffsetY: Single; const Opacity: Single): TGPBitmap;
var
    width, height: Integer;
    alphaMask: TGPBitmap;
    shadow: TGPBitmap;
    graphics: TGPGraphics;
    imageAttributes: TGPImageAttributes;
    cm: TColorMatrix;
begin
{
    We generate a drop shadow by first getting the alpha mask. This will be a black
    sillouette on a transparent background. We then blur the black "shadow" by the amounts
    given.
    We then draw the original image on top of it's own shadow.
}

{
    http://msdn.microsoft.com/en-us/library/aa511280.aspx
    Windows Vista User Experience -> Guidelines -> Aesthetics -> Icons

    Basic Flat Icon Shadow Ranges

    Flat icons
    Flat icons are generally used for file icons and flat real-world objects,
    such as a document or a piece of paper.

    Flat icon lighting comes from the upper-left at 130 degrees.

    Smaller icons (for example, 16x16 and 32x32) are simplified for readability.
    However, if they contain a reflection within the icon (often simplified),
    they may have a tight drop shadow. The drop shadow ranges in opacity from
    30-50 percent.
    Layer effects can be used for flat icons, but should be compared with other
    flat icons. The shadows for objects will vary somewhat, according to what
    looks best and is most consistent within the size set and with the other
    icons in Windows Vista. On some occasions, it may even be necessary to
    modify the shadows. This will especially be true when objects are laid over
    others.
    A subtle range of colors may be used to achieve desired outcome. Shadows help
    objects sit in space. Color impacts the perceived weight of the shadow, and
    may distort the image if it is too heavy.

    Blend mode: Multiply
    Opacity: 22% to 50% - depends on color of the item.
    Angle: 130 to 120, use global light
    Distance: 3 (256 thru 48x), Distance = 1 (32x, 24x)
    Spread: 0
    Size: 7 (256x thru 48x), Spread = 2 (32x, 24x)
}
    width := image.GetWidth;
    height := image.GetHeight;

    //Get bitmap to hold final composited image and shadow
    Result := TGPBitmap.Create(width, height, PixelFormat32bppARGB);

    //Use ColorMatrix methods to "draw" the alpha image.
    alphaMask := TImageEffects.GetAlphaMask(image);
    try
        //Blur the black and white shadow image
//      shadow := TImageEffects.BoxBlur(alphaMask, radius);
        shadow := TImageEffects.GaussianBlur(alphaMask, radius); //because Gaussian Blur is linearly-separable into two 1d kernels, it's actually faster than the box blur
    finally
        alphaMask.Free;
    end;

    //Draw
    graphics := TGPGraphics.Create(Result);
    try
        //Draw the "shadow", using the passed in opacity value.
        {
            Color transformations are of the form
        c =  (r, g, b, a)
        c' = (r, g, b, a)
        c' = c*M
            = (r, g, b, a, 1) * (0 0 0 0 0)  //r
                                      (0 0 0 0 0)  //g
                                      (0 0 0 0 0)  //b
                                      (1 1 1 1 0)  //a
                                      (0 0 0 0 1)  //1
        }

        imageAttributes := TGPImageAttributes.Create;
    {   cm := (
                ( 1, 0, 0, 0,   0),
                ( 0, 1, 0, 0,   0),
                ( 0, 0, 1, 0,   0),
                ( 0, 0, 0, 0.5, 0),
                ( 0, 0, 0, 0,   1)
            );}
        cm[0, 0] :=  1; cm[0, 1] :=  0; cm[0, 2] :=  0; cm[0, 3] := 0;       cm[0, 4] := 0;
        cm[1, 0] :=  0; cm[1, 1] :=  1; cm[1, 2] :=  0; cm[1, 3] := 0;       cm[1, 4] := 0;
        cm[2, 0] :=  0; cm[2, 1] :=  0; cm[2, 2] :=  1; cm[2, 3] := 0;       cm[2, 4] := 0;
        cm[3, 0] :=  0; cm[3, 1] :=  0; cm[3, 2] :=  0; cm[3, 3] := Opacity; cm[3, 4] := 0;
        cm[4, 0] :=  0; cm[4, 1] :=  0; cm[4, 2] :=  0; cm[4, 3] := 0;       cm[4, 4] := 1;


        imageAttributes.SetColorMatrix(
                cm,
                ColorMatrixFlagsDefault,
                ColorAdjustTypeBitmap);
        try
            graphics.DrawImage(shadow,
                        MakeRectF(OffsetX, OffsetY, width, height), //destination rectangle
                        0, 0, //source (x,y)
                        width, height, //source width, height
                        UnitPixel,
                        ImageAttributes);

            //Draw original image over-top of it's shadow
            graphics.DrawImage(image, 0, 0);
        finally
            imageAttributes.Free;
        end;
    finally
        graphics.Free;
    end;
end;

其中使用函数来获取灰度 alpha 蒙版：

class function TImageEffects.GetAlphaMask(image: TGPImage): TGPBitmap;
var
    imageAttributes: TGPImageAttributes;
    cm: TColorMatrix;
    graphics: TGPGraphics;
    Width, Height: UINT;
begin
    {
        Color transformations are of the form
    c =  (r, g, b, a)
    c' = (r, g, b, a)
    c' = c*M
        = (r, g, b, a, 1) * (0 0 0 0 0)
                            (0 0 0 0 0)
                            (0 0 0 0 0)
                            (1 1 1 1 0)
                            (0 0 0 0 1)
    }

    imageAttributes := TGPImageAttributes.Create;

{   cm := (
            ( 0,  0,  0, 0, 0),
            ( 0,  0,  0, 0, 0),
            ( 0,  0,  0, 0, 0),
            (-1, -1, -1, 1, 0),
            ( 1,  1,  1, 0, 1)
        );}
    cm[0, 0] :=  0; cm[0, 1] :=  0; cm[0, 2] :=  0; cm[0, 3] := 0; cm[0, 4] := 0;
    cm[1, 0] :=  0; cm[1, 1] :=  0; cm[1, 2] :=  0; cm[1, 3] := 0; cm[1, 4] := 0;
    cm[2, 0] :=  0; cm[2, 1] :=  0; cm[2, 2] :=  0; cm[2, 3] := 0; cm[2, 4] := 0;
    cm[3, 0] := -1; cm[3, 1] := -1; cm[3, 2] := -1; cm[3, 3] := 1; cm[3, 4] := 0;
    cm[4, 0] :=  1; cm[4, 1] :=  1; cm[4, 2] :=  1; cm[4, 3] := 0; cm[4, 4] := 1;


    imageAttributes.SetColorMatrix(
            cm,
            ColorMatrixFlagsDefault,
            ColorAdjustTypeBitmap);

    width := image.GetWidth;
    height := image.GetHeight;

    Result := TGPBitmap.Create(Integer(width), Integer(height));
    graphics := TGPGraphics.Create(Result);
   try
        graphics.DrawImage(
                image,
                MakeRect(0, 0, width, height), //destination rectangle
             0, 0, //source (x,y)
             width, height,
             UnitPixel,
                ImageAttributes);
   finally
        graphics.Free;
    end;
end;

核心是高斯模糊：

class function TImageEffects.GaussianBlur(const bitmap: TGPBitmap;
  radius: Single): TGPBitmap;
var
    width, height: Integer;
    tempBitmap: TGPBitmap;
    bdSource: TBitmapData;
    bdTemp: TBitmapData;
    bdDest: TBitmapData;
    pSrc: PARGBArray;
    pTemp: PARGBArray;
    pDest: PARGBArray;
    stride: Integer;
    kernel: TKernel;
begin
//  kernel := MakeGaussianKernel2d(radius);
    kernel := MakeGaussianKernel1d(radius);
    try
//      Result := ConvolveBitmap(bitmap, kernel); brute 2d kernel

        width := bitmap.GetWidth;
        height := bitmap.GetHeight;

        // GDI+ still lies to us - the return format is BGR, NOT RGB.
        bitmap.LockBits(MakeRect(0, 0, width, height),
                ImageLockModeRead,
                PixelFormat32bppPARGB, bdSource);

        //intermediate bitmap
        tempBitmap := TGPBitmap.Create(width, height, PixelFormat32bppPARGB);
        tempBitmap.LockBits(MakeRect(0, 0, width, height),
                    ImageLockModeWrite,
                    PixelFormat32bppPARGB, bdTemp);

        //target bitmap
        Result := TGPBitmap.Create(width, height, PixelFormat32bppARGB);
        Result.LockBits(MakeRect(0, 0, width, height),
                    ImageLockModeWrite,
                    PixelFormat32bppPARGB, bdDest);

        pSrc := PARGBArray(bdSource.Scan0);
        pTemp := PARGBArray(bdTemp.Scan0);
        pDest := PARGBArray(bdDest.Scan0);
        stride := bdSource.Stride;

        ConvolveAndTranspose(kernel, pSrc^, pTemp^, width, height, stride, True, EdgeActionClampEdges);
        ConvolveAndTranspose(kernel, pTemp^, pDest^, height, width, stride, True, EdgeActionClampEdges);

        //Unlock source
       bitmap.UnlockBits(bdSource);
        tempBitmap.UnlockBits(bdTemp);
        Result.UnlockBits(bdDest);

        //get rid of temp
        tempBitmap.Free;
    finally
        kernel.Free;
    end;
end;

它需要一维内核：

class function TImageEffects.MakeGaussianKernel1d(radius: Single): TKernel;
var
    r: Integer;
    rows: Integer;
    matrix: TSingleDynArray;
    sigma: Single;
    sigma22: Single;
    sigmaPi2: Single;
    sqrtSigmaPi2: Single;
    radius2: Single;
    total: Single;
    index: Integer;
    row: Integer;
    distance: Single;
    i: Integer;
begin
    r := Ceil(radius);
    rows := r*2+1;

    SetLength(matrix, rows);
    sigma := radius/3.0;
    sigma22 := 2*sigma*sigma;
    sigmaPi2 := 2*pi*sigma;
    sqrtSigmaPi2 := Sqrt(sigmaPi2);
    radius2 := radius*radius;
    total := 0;

    Index := 0;
    for row := -r to r do
    begin
        distance := row*row;
        if (distance > radius2) then
            matrix[index] := 0
        else
        begin
            matrix[index] := Exp((-distance)/sigma22) / sqrtSigmaPi2;
            total := total + matrix[index];
            Inc(index);
        end;
    end;

    //Normalize the values
    for i := 0 to rows-1 do
        matrix[i] := matrix[i] / total;


    Result := TKernel.Create(rows, 1, matrix);
end;

然后高斯函数的神奇之处在于它可以分为两个一维卷积：

class procedure TImageEffects.convolveAndTranspose(kernel: TKernel;
  const inPixels: array of ARGB; var outPixels: array of ARGB; width,
  height, stride: Integer; alpha: Boolean; edgeAction: TEdgeAction);
var
    index: Integer;
    matrix: TSingleDynArray;
    rows: Integer; //number of rows in the kernel
    cols: Integer; //number of columns in the kernel
    rows2: Integer; //half row count
    cols2: Integer; //half column count

    x, y: Integer; //
    r, g, b, a: Single; //summed red, green, blue, alpha values
    row, col: Integer;
    ix, iy, ioffset: Integer;
    moffset: Integer;
    f: Single;
    rgb: ARGB;
    ir, ig, ib, ia: Integer;

   function ClampPixel(value: Single): Integer;
    begin
        Result := Trunc(value+0.5);
        if Result < 0 then
            Result := 0
        else if Result > 255 then
            Result := 255;
    end;
begin
    matrix := kernel.KernelData;
    cols := kernel.Width;
    cols2 := cols div 2;

    for y := 0 to height-1 do
    begin
        index := y;
        ioffset := y*width;
        for x := 0 to width-1 do
        begin
            r := 0;
            g := 0;
            b := 0;
            a := 0;

            moffset := cols2;
            for col := -cols2 to cols2 do
            begin
                f := matrix[moffset+col];

                if (f <> 0) then
                begin
                    ix := x+col;
                    if ( ix < 0 ) then
                    begin
                        if ( edgeAction = EdgeActionClampEdges ) then
                            ix := 0
                        else if ( edgeAction = EdgeActionWrapEdges ) then
                            ix := (x+width) mod width;
                    end
                    else if ( ix >= width) then
                    begin
                        if ( edgeAction = EdgeActionClampEdges ) then
                            ix := width-1
                        else if ( edgeAction = EdgeActionWrapEdges ) then
                            ix := (x+width) mod width;
                    end;
                    rgb := inPixels[ioffset+ix];
                    a := a + f * ((rgb shr 24) and $FF);
                    r := r + f * ((rgb shr 16) and $FF);
                    g := g + f * ((rgb shr  8) and $FF);
                    b := b + f * ((rgb       ) and $FF);
                end;
            end;
            if alpha then
                ia := ClampPixel(a)
         else
                ia := $FF;
            ir := ClampPixel(r);
            ig := ClampPixel(g);
            ib := ClampPixel(b);
            outPixels[index] := MakeARGB(ia, ir, ig, ib);

            Inc(index, height);
        end;
    end;
end;

通过示例使用，在我的 256x256 源图像上：

image := TImageEffects.GenerateDropShadow(localImage, 14, 2.12132, 2.12132, 1.0);

分析显示 88.62% 的时间花费在lines：

a := a + f * ((rgb shr 24) and $FF);
r := r + f * ((rgb shr 16) and $FF);
g := g + f * ((rgb shr  8) and $FF);
b := b + f * ((rgb       ) and $FF);

这是每像素 alpha 混合。

这让我认为，在 Windows 和 OSX 实时向窗口应用投影之后，有一种更好的方法来应用模糊效果来应用软投影。

Answer 1

该算法来自此博客条目： http://blog.ivank.net/fastest- gaussian-blur.html 。当然，它正在实现最后也是最快的版本。 :-)

它是直接从我的工作代码复制的，因此外部假设可能反映了这一点。该函数返回一个更大的位图以适应大小的增加。当然，在您的代码中，您需要相应地处理这个问题。它假定 32 位 alpha 图片，但可以轻松修改为仅处理 24 位（CHANNELS 常量和 PixelFormat 值）。

public static class DropShadow {
  const int CHANNELS = 4;

  public static Bitmap CreateShadow(Bitmap bitmap, int radius, float opacity) {
    // Alpha mask with opacity
    var matrix = new ColorMatrix(new float[][] {
            new float[] {  0F,  0F,  0F, 0F,      0F }, 
            new float[] {  0F,  0F,  0F, 0F,      0F }, 
            new float[] {  0F,  0F,  0F, 0F,      0F }, 
            new float[] { -1F, -1F, -1F, opacity, 0F },
            new float[] {  1F,  1F,  1F, 0F,      1F }
        });

    var imageAttributes = new ImageAttributes();
    imageAttributes.SetColorMatrix(matrix, ColorMatrixFlag.Default, ColorAdjustType.Bitmap);
    var shadow = new Bitmap(bitmap.Width + 4 * radius, bitmap.Height + 4 * radius);
    using (var graphics = Graphics.FromImage(shadow))
      graphics.DrawImage(bitmap, new Rectangle(2 * radius, 2 * radius, bitmap.Width, bitmap.Height), 0, 0, bitmap.Width, bitmap.Height, GraphicsUnit.Pixel, imageAttributes);

    // Gaussian blur
    var clone = shadow.Clone() as Bitmap;
    var shadowData = shadow.LockBits(new Rectangle(0, 0, shadow.Width, shadow.Height), ImageLockMode.ReadWrite, PixelFormat.Format32bppArgb);
    var cloneData = clone.LockBits(new Rectangle(0, 0, clone.Width, clone.Height), ImageLockMode.ReadWrite, PixelFormat.Format32bppArgb);

    var boxes = DetermineBoxes(radius, 3);
    BoxBlur(shadowData, cloneData, shadow.Width, shadow.Height, (boxes[0] - 1) / 2);
    BoxBlur(shadowData, cloneData, shadow.Width, shadow.Height, (boxes[1] - 1) / 2);
    BoxBlur(shadowData, cloneData, shadow.Width, shadow.Height, (boxes[2] - 1) / 2);

    shadow.UnlockBits(shadowData);
    clone.UnlockBits(cloneData);
    return shadow;
  }

  private static unsafe void BoxBlur(BitmapData data1, BitmapData data2, int width, int height, int radius) {
    byte* p1 = (byte*)(void*)data1.Scan0;
    byte* p2 = (byte*)(void*)data2.Scan0;

    int radius2 = 2 * radius + 1;
    int[] sum = new int[CHANNELS];
    int[] FirstValue = new int[CHANNELS];
    int[] LastValue = new int[CHANNELS];

    // Horizontal
    int stride = data1.Stride;
    for (var row = 0; row < height; row++) {
      int start = row * stride;
      int left = start;
      int right = start + radius * CHANNELS;

      for (int channel = 0; channel < CHANNELS; channel++) {
        FirstValue[channel] = p1[start + channel];
        LastValue[channel] = p1[start + (width - 1) * CHANNELS + channel];
        sum[channel] = (radius + 1) * FirstValue[channel];
      }
      for (var column = 0; column < radius; column++)
        for (int channel = 0; channel < CHANNELS; channel++)
          sum[channel] += p1[start + column * CHANNELS + channel];
      for (var column = 0; column <= radius; column++, right += CHANNELS, start += CHANNELS)
        for (int channel = 0; channel < CHANNELS; channel++) {
          sum[channel] += p1[right + channel] - FirstValue[channel];
          p2[start + channel] = (byte)(sum[channel] / radius2);
        }
      for (var column = radius + 1; column < width - radius; column++, left += CHANNELS, right += CHANNELS, start += CHANNELS)
        for (int channel = 0; channel < CHANNELS; channel++) {
          sum[channel] += p1[right + channel] - p1[left + channel];
          p2[start + channel] = (byte)(sum[channel] / radius2);
        }
      for (var column = width - radius; column < width; column++, left += CHANNELS, start += CHANNELS)
        for (int channel = 0; channel < CHANNELS; channel++) {
          sum[channel] += LastValue[channel] - p1[left + channel];
          p2[start + channel] = (byte)(sum[channel] / radius2);
        }
    }

    // Vertical
    stride = data2.Stride;
    for (int column = 0; column < width; column++) {
      int start = column * CHANNELS;
      int top = start;
      int bottom = start + radius * stride;

      for (int channel = 0; channel < CHANNELS; channel++) {
        FirstValue[channel] = p2[start + channel];
        LastValue[channel] = p2[start + (height - 1) * stride + channel];
        sum[channel] = (radius + 1) * FirstValue[channel];
      }
      for (int row = 0; row < radius; row++)
        for (int channel = 0; channel < CHANNELS; channel++)
          sum[channel] += p2[start + row * stride + channel];
      for (int row = 0; row <= radius; row++, bottom += stride, start += stride)
        for (int channel = 0; channel < CHANNELS; channel++) {
          sum[channel] += p2[bottom + channel] - FirstValue[channel];
          p1[start + channel] = (byte)(sum[channel] / radius2);
        }
      for (int row = radius + 1; row < height - radius; row++, top += stride, bottom += stride, start += stride)
        for (int channel = 0; channel < CHANNELS; channel++) {
          sum[channel] += p2[bottom + channel] - p2[top + channel];
          p1[start + channel] = (byte)(sum[channel] / radius2);
        }
      for (int row = height - radius; row < height; row++, top += stride, start += stride)
        for (int channel = 0; channel < CHANNELS; channel++) {
          sum[channel] += LastValue[channel] - p2[top + channel];
          p1[start + channel] = (byte)(sum[channel] / radius2);
        }
    }
  }

  private static int[] DetermineBoxes(double Sigma, int BoxCount) {
    double IdealWidth = Math.Sqrt((12 * Sigma * Sigma / BoxCount) + 1);
    int Lower = (int)Math.Floor(IdealWidth);
    if (Lower % 2 == 0)
      Lower--;
    int Upper = Lower + 2;

    double MedianWidth = (12 * Sigma * Sigma - BoxCount * Lower * Lower - 4 * BoxCount * Lower - 3 * BoxCount) / (-4 * Lower - 4);
    int Median = (int)Math.Round(MedianWidth);

    int[] BoxSizes = new int[BoxCount];
    for (int i = 0; i < BoxCount; i++)
      BoxSizes[i] = (i < Median) ? Lower : Upper;
    return BoxSizes;
  }

}

我认为将其转换为 Delphi 一定很简单。

附录：根据该博客的评论，如果您有一个整数半径和三个盒子，您实际上可以忘记 DetermineBoxes() 并使用：

BoxBlur(shadowData, cloneData, shadow.Width, shadow.Height, radius - 1);
BoxBlur(shadowData, cloneData, shadow.Width, shadow.Height, radius - 1);
BoxBlur(shadowData, cloneData, shadow.Width, shadow.Height, radius);

与位图本身相比，它的执行时间可以忽略不计，但仍然如此。 ..

Answer 2

我之所以要代码是为了看看您是否使用了“快速位图”方法或 GetPixel(), SetPixel() 方法。

既然您已经涵盖了这一点，我怀疑您是否能够在性能优化方面做更多的事情。 GDI+ 并不是为这种逐像素操作场景而设计的。实际上，您应该考虑实现一个更简单的阴影生成器，它看起来不会那么花哨，但也不会占用处理器资源。

这在很大程度上取决于您的使用场景（您还没有真正描述过）：

• 图像是否都相似（所有票证或者您是否仅将票证用作示例）？如果是，那么您可以生成一次阴影并重用该位图。
• 当用户正在做其他事情时，您可以生成并缓存图像的阴影版本（或只是阴影缩略图）作为后台进程。

您还可以在 Paint.NET（大多数内容使用 GDI+）中尝试高斯模糊并在那里测量它的速度。我怀疑你能否让它比 Paint.NET 更快，所以它是一个很好的基准。

Answer 3

如果您追求的是纯粹的性能，您也可以考虑仅对源图像的薄矩形边缘条进行卷积。这样，您就不会花时间对图像的中心（隐藏）部分进行卷积，而只对有机会在屏幕上绘制的部分进行卷积。

Answer 4

我测试了一些算法，最好的是 Gábor 实现的高斯模糊。在我的测试中，该算法的延迟约为 20 毫秒。

下面是它在 Delphi 中的实现，做了一些更改（它使用免费软件 Bilsen GDI+ 库）：

function CreateBlurShadow(ABitmap: IGPBitmap; ARadius: Integer; AOpacity: Double; AColor: TColor = clNone): IGPBitmap;

  procedure BoxBlur(const AData1, AData2: TGPBitmapData; AWidth, AHeight, ARadius: Integer);
  const
    CHANNELS = 4;
  var
    LScan1, LScan2: PByte;
    LSum, LFirstValue, LLastValue: array [0..CHANNELS-1] of Integer;
    LRadius2, LStride, LStart, LChannel, LLeft, LRight, LBottom, LTop, LRow, LColumn: Integer;
  begin
    LScan1 := AData1.Scan0;
    LScan2 := AData2.Scan0;
    LRadius2 := (2 * ARadius) + 1;
    LStride := AData1.Stride;
    for LRow := 0 to AHeight-1 do
    begin
      LStart := LRow * LStride;
      LLeft := LStart;
      LRight := LStart + ARadius * CHANNELS;
      for LChannel := 0 to CHANNELS-1 do
      begin
        LFirstValue[LChannel] := LScan1[LStart + LChannel];
        LLastValue[LChannel] := LScan1[LStart + ((AWidth - 1) * CHANNELS) + LChannel];
        LSum[LChannel] := (ARadius + 1) * LFirstValue[LChannel];
      end;
      for LColumn := 0 to ARadius-1 do
        for LChannel := 0 to CHANNELS-1 do
          LSum[LChannel] := LSum[LChannel] + LScan1[LStart + (LColumn * CHANNELS) + LChannel];
      for LColumn := 0 to ARadius do
      begin
        for LChannel := 0 to CHANNELS-1 do
        begin
          LSum[LChannel] := LSum[LChannel] + LScan1[LRight + LChannel] - LFirstValue[LChannel];
          LScan2[LStart + LChannel] := Byte(LSum[LChannel] div LRadius2);
        end;
        Inc(LRight, CHANNELS);
        Inc(LStart, CHANNELS);
      end;
      for LColumn := ARadius + 1 to AWidth-ARadius-1 do
      begin
        for LChannel := 0 to CHANNELS-1 do
        begin
          LSum[LChannel] := LSum[LChannel] + LScan1[LRight + LChannel] - LScan1[LLeft + LChannel];
          LScan2[LStart + LChannel] := Byte(LSum[LChannel] div LRadius2);
        end;
        Inc(LLeft, CHANNELS);
        Inc(LRight, CHANNELS);
        Inc(LStart, CHANNELS);
      end;
      for LColumn := AWidth-ARadius to AWidth-1 do
      begin
        for LChannel := 0 to CHANNELS-1 do
        begin
          LSum[LChannel] := LSum[LChannel] + LLastValue[LChannel] - LScan1[LLeft + LChannel];
          LScan2[LStart + LChannel] := Byte(LSum[LChannel] div LRadius2);
        end;
        Inc(LLeft, CHANNELS);
        Inc(LStart, CHANNELS);
      end;
    end;
    LStride := AData2.Stride;
    for LColumn := 0 to AWidth-1 do
    begin
      LStart := LColumn * CHANNELS;
      LTop := LStart;
      LBottom := LStart + (ARadius * LStride);
      for LChannel := 0 to CHANNELS-1 do
      begin
        LFirstValue[LChannel] := LScan2[LStart + LChannel];
        LLastValue[LChannel] := LScan2[LStart + ((AHeight - 1) * LStride) + LChannel];
        LSum[LChannel] := (ARadius + 1) * LFirstValue[LChannel];
      end;
      for LRow := 0 to ARadius-1 do
        for LChannel := 0 to CHANNELS-1 do
          LSum[LChannel] := LSum[LChannel] + LScan2[LStart + (LRow * LStride) + LChannel];
      for LRow := 0 to ARadius do
      begin
        for LChannel := 0 to CHANNELS-1 do
        begin
          LSum[LChannel] := LSum[LChannel] + LScan2[LBottom + LChannel] - LFirstValue[LChannel];
          LScan1[LStart + LChannel] := Byte(LSum[LChannel] div LRadius2);
        end;
        Inc(LBottom, LStride);
        Inc(LStart, LStride);
      end;
      for LRow := ARadius + 1 to AHeight - ARadius - 1 do
      begin
        for LChannel := 0 to CHANNELS-1 do
        begin
          LSum[LChannel] := LSum[LChannel] + LScan2[LBottom + LChannel] - LScan2[LTop + LChannel];
          LScan1[LStart + LChannel] := Byte(LSum[LChannel] div LRadius2);
        end;
        Inc(LTop, LStride);
        Inc(LBottom, LStride);
        Inc(LStart, LStride);
      end;
      for LRow := AHeight - ARadius to AHeight-1 do
      begin
        for LChannel := 0 to CHANNELS-1 do
        begin
          LSum[LChannel] := LSum[LChannel] + LLastValue[LChannel] - LScan2[LTop + LChannel];
          LScan1[LStart + LChannel] := Byte(LSum[LChannel] div LRadius2);
        end;
        Inc(LTop, LStride);
        Inc(LStart, LStride);
      end;
    end;
  end;

const
  INITIAL_MATRIX: array [0..4, 0..4] of Single =
   ((0.5,   0,   0, 0, 0),
    (0,   0.5,   0, 0, 0),
    (0,     0, 0.5, 0, 0),
    (0,     0,   0, 1, 0),
    (0,     0,   0, 0, 1));
var
  LMatrix: TGPColorMatrix;
  LImageAttributes: IGPImageAttributes;
  LShadow, LClone: IGPBitmap;
  LGraphics: IGPGraphics;
  LShadowData, LCloneData: TGPBitmapData;
  LColor: TGPColor;
begin
  ARadius := Max(ARadius, 0);
  LShadow := TGPBitmap.Create(ABitmap.Width + (4 * Cardinal(ARadius)),
    ABitmap.Height + (4 * Cardinal(ARadius)), PixelFormat32bppARGB);
  LGraphics := TGPGraphics.FromImage(LShadow);
  LGraphics.DrawImage(ABitmap, TGPRect.Create(2 * ARadius, 2 * ARadius,
    ABitmap.Width, ABitmap.Height), 0, 0, ABitmap.Width, ABitmap.Height,
    TGPUnit.UnitPixel);
  LClone := LShadow.Clone;
  LShadowData := LShadow.LockBits(TGPRect.Create(0, 0, LShadow.Width, LShadow.Height),
    [ImageLockModeRead, ImageLockModeWrite], PixelFormat32bppARGB);
  LCloneData := LClone.LockBits(TGPRect.Create(0, 0, LClone.Width, LClone.Height),
    [ImageLockModeRead, ImageLockModeWrite], PixelFormat32bppARGB);
  try
    BoxBlur(LShadowData, LCloneData, LShadow.Width, LShadow.Height, ARadius - 1);
    BoxBlur(LShadowData, LCloneData, LShadow.Width, LShadow.Height, ARadius - 1);
    BoxBlur(LShadowData, LCloneData, LShadow.Width, LShadow.Height, ARadius);
  finally
    LShadow.UnlockBits(LShadowData);
    LClone.UnlockBits(LCloneData);
  end;
  if (AColor = clNone) and (AOpacity = 1.0) then
    Result := LShadow
  else
  begin
    LColor := TGPColor.CreateFromColorRef(ColorToRGB(AColor));
    Move(INITIAL_MATRIX[0, 0], LMatrix.M[0, 0], SizeOf(INITIAL_MATRIX));
    LMatrix.M[4, 0] := Min((Integer(LColor.R) - 127) / 127, 1.0);
    LMatrix.M[4, 1] := Min((Integer(LColor.G) - 127) / 127, 1.0);
    LMatrix.M[4, 2] := Min((Integer(LColor.B) - 127) / 127, 1.0);
    LMatrix.M[4, 3] := AOpacity-1;
    LImageAttributes := TGPImageAttributes.Create;
    LImageAttributes.SetColorMatrix(LMatrix, TGPColorMatrixFlags.ColorMatrixFlagsDefault,
      TGPColorAdjustType.ColorAdjustTypeBitmap);
    Result := TGPBitmap.Create(LShadow.Width, LShadow.Height, PixelFormat32bppARGB);
    LGraphics := TGPGraphics.FromImage(Result);
    LGraphics.DrawImage(LShadow, TGPRect.Create(0, 0, LShadow.Width, LShadow.Height),
      0, 0, Result.Width, Result.Height, TGPUnit.UnitPixel, LImageAttributes);
  end;
end;

Answer 5

我知道逐像素操作相当慢，但从未做过基准测试； 70x 看起来很多，超出了我的预期。也许您使用托管语言这一事实​​会导致这种情况，因为这是虚拟机开销最大化的一种情况。您是否尝试过用本机代码制作该部分程序？此链接有一个本机实现，您可以使用它进行快速测试：

http://www .codeproject.com/KB/GDI/Glow_and_Shadow_effects.aspx

不幸的是，它们唯一的区别是使用可以生成本机代码的语言，但它们仍然使用双层循环来访问像素。例如，如果您可以使用 CUDA，如果您可以假设运行应用程序的计算机具有此类硬件，那就更好了。但在这种情况下，您将不再使用 GDI+。无论如何，也许另一个问题有帮助：

使用显卡而不是 GDI+ 进行图像操作