Silverlight: Super-fast Dymanic Image Generation Code (Revisited)

  • 아티클

I decided to give a third spin of Joe Stegman's dynamic image generation code. This time, it's many times faster (about 10x) than the original implementation and the png is generated in-place (no recoding necessary).

I used some hacky optimizations :) For example, replacing the big CRC loop with a constant (0), definitely improves speed a lot! :)

There are no more intermediate memory streams or buffers: whenever SetPixel() is called, the value is written directly into PNG. Also all the header and size information is created just once. That also helped performance quite a bit.

With the new code my raindrops sample runs 20% faster (90% of the time used in calculating the drops): www.nokola.com/raindrops

Download the source code containing the new PngEncoder class here: www.nokola.com/sources/water.zip

Here's how to use the new image generator class:

PngEncoder surface = new PngEncoder(640, 480); // image dimension

surface.SetPixelSlow(40, 30, 200, 135, 32, 255); // set pixel at (40,30) with color RGBA=(200,135,32,255)

           

// draw a white horizontal line fast

int rowStart = surface.GetRowStart(30);

for (int i = 0; i < 10; i++) {

// SetPixelAtRowStart() is good for blitting/copying existing images onto this one

surface.SetPixelAtRowStart(i + 40, rowStart, 255, 255, 255, 255);

}

// display the image

BitmapImage img = new BitmapImage();

img.SetSource(surface.GetImageStream());

imgWater.Source = img; // this is just a normal Silverlight Image

And, here's the GetImageStream() function for comparison with the previous implementation:

public Stream GetImageStream()

{

MemoryStream ms = new MemoryStream();

ms.Write(_buffer, 0, _buffer.Length);

ms.Seek(0, SeekOrigin.Begin);

return ms;

}

Compare this to the previous function, that had numerous for()-s, encoding logic, etc :)

hoho! I'm very happy for doing this!

Edit: here's the source code for the SetPixelXX functions - to prove they are actually not slow :)

public void SetPixelSlow(int col, int row, byte red, byte green, byte blue, byte alpha)

{

int start = _rowLength * row + col * 4 + 1;

      int blockNum = start / _blockSize;

      start += ((blockNum + 1) * 5);

      start += _dataStart;

      _buffer[start] = red;

      _buffer[start + 1] = green;

      _buffer[start + 2] = blue;

      _buffer[start + 3] = alpha;

}

public void SetPixelAtRowStart(int col, int rowStart, byte red, byte green, byte blue, byte alpha)

{

int start = rowStart + (col << 2);

      _buffer[start] = red;

      _buffer[start + 1] = green;

      _buffer[start + 2] = blue;

      _buffer[start + 3] = alpha;

}

public int GetRowStart(int row)

{

int start = _rowLength * row + 1;

      int blockNum = start / _blockSize;

      start += ((blockNum + 1) * 5);

      start += _dataStart;

      return start;

}

Comments

  • Anonymous
    March 05, 2009
    The comment has been removed

  • Anonymous
    May 26, 2009
    Thanks!

  • Anonymous
    June 22, 2009
    When I try to use the code above, I cannot read the generated png stream, all imaging programs say that it's not a valid png format (including .NET's libraries under WPF). Did anybody else have issues?

  • Anonymous
    June 27, 2009
    Hi, Unfortunately the code strangely doesn't work with WPF, the produced image seems to be not a valid PNG image. Any ideas?

  • Anonymous
    July 01, 2009
    Yes, the produced image is not a valid PNG (the CRC is set to 0). This is a trick that allows to generate the PNG much faster and it happens to work in SL 2. For WPF/other programs, check out my previous posts about PNG image generation. The generators there make valid PNGs according to the spec, but slower.

  • Anonymous
    July 01, 2009
    Ah, okay, perfect. And sorry for double posting, I though the blog ate my comment. Thanks, Christoph

  • Anonymous
    July 01, 2009
    no problem :) I just came back from 3 weeks holiday and saw your comment that was left for review by the blog software.

  • Anonymous
    August 17, 2009
    Humm... interesting, So it is truely faster!! Thanks for bringing this up

  • Anonymous
    August 18, 2009
    yw :) Silverlight 3 has a new WritableBitmap class that makes the png code above (partially) obsolete. http://www.wintellect.com/CS/blogs/jprosise/archive/2009/03/23/silverlight-3-s-new-writeablebitmap.aspx I haven't perf-ed it but I bet it's faster than the previous method.

  • Anonymous
    September 24, 2009
    Hi, I have a problem with this approach when my canvas size is more than 3000... If i write the resultant image and try to open it, i am not able to view anything.. the image is corrupted... But the approach works fine for smaller images(Canvas). any idea on this?? Thanks, Ravi

  • Anonymous
    September 24, 2009
    I'm not sure why the image gets corrupted over 3000 size...might be an issue with the algorithm I'm using. Anyway, a better approach is to use WritableBitmap, which is now available in SL3. The approach above is for SL2, which does not have WritableBitmap. Check this blog for more info: http://www.wintellect.com/CS/blogs/jprosise/archive/2009/03/23/silverlight-3-s-new-writeablebitmap.aspx hope this helps - please let me know!