Avoid copy of image data from decoder to avifImage. about libavif

[joedrago]

Yes, definitely.

I am planning to stubbornly keep my basic avifDecoderRead() as a braindead-simple path for people that aren't trying to use image sequences or squeak out extra performance, as it allows for avifImage objects to be decoupled almost entirely from encoders and decoders.

That said, I'm in the process now of putting in image sequence support, and a side effect of that medium-sized code reorg, I'm planning to make avifDecoderRead() simply call 3-4 other avifDecoder functions which ends up with a copy of the first image, with functions that perform something like:

• Begin the decode/parse
• "get" the first image, which will be owned by the decoder (decoder->image will be a not-copied avifImage)
• "copy" decoder->image into the passed in image ptr

I plan to reuse the decoder->image structure as you "get" the next image, so if you want YUV copies of multiple frames, you're going to have to copy it anyways, but I imagine you just want to take the current decoder->image and run it through your own YUV->RGB pipeline, in which decoder->image can be safely reused, right?

[dalecurtis]

That break up sounds good.

Yes in the ideal case we can output directly to YUV8 planes, but we can't reuse the buffers from dav1d/libavif for that. We do have to copy into them. In the less ideal case (where RBG8/RGB_F16) is required we have to convert into a new structure.

So I don't think we can ever get to a zero copy solution, but a one copy one would be fine too.

[joedrago]

This is done, see the Advanced Decoding section of the README.