> Have you done performance/quality comparisons of your interpolation > algorithms > with the reference implementation's? That's probably a good metric for what > is > acceptable performance.
When filtering rasters, this implementation is between 75 - 200 times slower (75 for NN and bicubic, 200 for bilinear) than the reference. I suspect they have some kind of optimization here, their bilinear is not any slower than nearest neighbour. Our code is completely unoptimized at the moment, but I'm wondering about using a completely different formula/algorithm vs optimizing the one I dug up. Quality-wise, this implementation is actually much, much better than Sun's. Filtering images is significantly faster since we get Cairo to handle it (2x slower than the reference), but I just realised that Cairo is not respecting our interpolation hint and always using bilinear. Anyways, committed for now, and I'll follow up with additional fixes. Francis
