Improving heif decoding performance

[roticv]

I have been looking into improving heic decoding performance. My use case is:

• Input: YUV420 HEVC, multiple tile
• Output: RGB interlaced

In my benchmark with hacky implementation with libyuv (see the question below about why libyuv) and fixing the separate RGB channel to interlaced RGB conversion, my numbers are roughly:

• 42ms with ffmpeg decoder plugin (and libheif yuv2rgb conversion)
• 30ms with ffmpeg decoder plugin + libyuv for yuv2rgh interlaced conversion

Questions:

• Does it make sense to use libyuv for yuv2rgb conversion? sharpyuv doesn't do yuv2rgb. The other alternative is libswscale (from ffmpeg), but following ffmpeg-devel mailing list, it has not been getting much love until recently and there is some suggestion that libyuv performs better than libswscale https://ffmpeg.org/pipermail/ffmpeg-devel/2024-June/330395.html
• Is the logic to default to using chroma_444 in decode_image_planar intentional? Is there any objection to change it to depend on out_chroma instead?

Thanks