svenfoo

For GIMP 2.4 we dramatically improved the quality of the zoomed-out display. Much of this change was based on the introduction of a tile pyramid for the image projection. From a performance point of view this change helped for very large images as it reduces the need to access the image data. But overall it caused GIMP 2.4 to be slightly slower than GIMP 2.2 when it comes to updating the display.

The tile pyramid consists of a series of scaled down copies of the image. On level 0, we have the original projection created by combining all layers. Level 1 is a quarter of the original size, level 2 a quarter of the size of level 1 and so forth. So when the projection changes, in order to recompute one pixel on the upper level, four pixels from the pyramid level below are averaged. To do this correctly, the pixels have to be weighted by their alpha values. And this is where a lot of CPU cycles are burnt.

I have committed some changes to trunk last night that make the tile pyramid store the pixels in the upper levels with pre-multiplied alpha. Now averaging four pixels is as simple as adding their values and shifting the result two bits to the right (thus effectively dividing by four). Of course this means that the display rendering code had to be changed as well. But it turned out that this code also benefits from pre-multiplied alpha. It became simpler and we don’t any longer need the huge lookup tables that we used to use to blend the image pixels over the checkerboard pattern.

The graphics shows the time GIMP needs to recompute the image projection and to render it to the screen. I’ve measured this for a rather large image (10800 x 4856 pixels) and the time went down from 1.83 seconds to 1.45 seconds. This is an improvement that you actually notice as a user.

This entry was posted on Friday, December 14th, 2007 at 4:19 pm and is filed under GIMP. You can follow any responses to this entry through the RSS 2.0 feed. Both comments and pings are currently closed.