Single-pixel optimizations are not yet implemented and can be useful for speeding up pathological cases like single-pixel GIFs used as padding. While at it, we can consider borrowing some of the techniques recently developed by these guys: https://bugzilla.mozilla.org/show_bug.cgi?id=371867 https://bugzilla.mozilla.org/show_bug.cgi?id=415854
Created attachment 28192 [details] single pixel optimization implementation of checkForSolidColor for Cairo after Adam Treat s patch: r41358.
Comment on attachment 28192 [details] single pixel optimization I don't really know enough about Cairo to know if your color extraction code is correct. I think I would have abstracted it into a nice little static function though: uint32_t* pixel = reinterpret_cast<uint32_t*>(data); 208 int alpha, red, green, blue; 209 210 if (alpha = (*pixel & 0xff000000) >> 24) { 211 red = ((*pixel & 0x00ff0000) >> 16) * 255 / alpha; 212 green = ((*pixel & 0x0000ff00) >> 8) * 255 / alpha; 213 blue = (*pixel & 0x000000ff) * 255 / alpha; 214 } else { 215 red = (*pixel & 0x00ff0000) >> 16; 216 green = (*pixel & 0x0000ff00) >> 8; 217 blue = (*pixel & 0x000000ff); 218 } Color color = colorFromCairoPixel(*pixel); I think someone who knows more about Cairo should comment before I can mark this r+. Also... is (*pixel & 0xff000000) >> 24) the cleanest way to get the alpha out of a cairo pixel? Talk about ugly, error-prone math... I think actually I'm going to r- for the way you grab colors out of the pixel. Even if that's the only way, we should make those static inlines so that they're not so error prone.
Created attachment 28221 [details] single pixel optimization with CairoColor I talked to eseidel in IRC and he mentioned to add a new typedef or somthing like that to avoid failured and make the code reading better. I added a new color schema to Cairo: CairoColor. CairoColor attends to premultiplying and every other calculation. I added some small callers to have access to the color values of a pixel. You can easily get the unpremutiplied colors of a pixel or set them by adding unpremultiplied colors. We just support image surfaces with 32bit for now. There are asserts before every direct manipulation of surfaces. But I don't see a problem to support other surfaces in the future if wanted.
Created attachment 28279 [details] single pixel optimization with CairoColor this patch doesn't add a new class but uses Color.cpp instead.
Created attachment 28302 [details] single pixel optimization with Color just added two calls to premultiply or unpremultiply the color of a pixel to Color.cpp.
Comment on attachment 28302 [details] single pixel optimization with Color halves, not halfes 148 Color colorFromPremultipliedARGB(RGBA32); should use unsigned, not RGBA32 This looks good. It would be nice if someone could check your pixel math before you land. Like Oliver?
landed in r41561.
Marking fixed so it drops out of the commit queue.