Question

My program has the feature to export a hi-res image of the working canvas to the disk. Users will frequently try to export images of about 20,000 x 10,000 pixels @ 32bpp which equals about 800MB. Add that to the serious memory consumption already going on in your average 3D CAD program and you’ll pretty much guarantee an out-of-memory crash on 32-bit platforms.

So now I’m exporting tiles of 1000×1000 pixels which the user has to stitch together afterwards in a pixel editor. Is there a way I can solve this problem without the user doing any work?

I figured I could probably write a small exe that gets command-lined into the process and performs the stitching automatically. It would be a separate process and it would thus have 2GB of ram all to itself. Or is there a better way still? I’d like to support jpg, png and bmp so writing the image as a bytestream to the disk is not really possible.

Answer 1

If you want to do it language-agonistic then you only have two options:

• Either do as you have done or
• do it as the professional graphics
  applications do it. They keep the
  images on disk in full resolution but
  when loading/saving they don’t
  load/save the complete image at once
  but at runtime create such a
  tile-based approach in memory (and only load/save the parts that are currently relevant).

The disadvantage of the second aproach is that you have to develop EVERYTHING yourself. You cannot use any OS image handling, image en-/decoders or anything else.

I had a similar problem some time ago and could solve it by using only 1/16bpp (especially as 32bit for images usually contains 8 completely unused bits which do not need to be held in memory – alpha channel). This means I mostly had layers of 1bpp images. However this obviously only works if you use your own application to consume the images.