On Mon, Jan 25, 2016 at 10:37 PM, Michael Niedermayer <mich...@niedermayer.cc> wrote: > On Fri, Jan 22, 2016 at 02:54:21PM -0800, Neil Birkbeck wrote: >> Hmm. I don't have a good idea of how likely it is for this conversion to >> float (by dividing a constant) to not be bit-exact on different >> architectures (compilers?) when there should not be any other math >> transforming the metadata (other than the conversion back to the integer >> coding for cases like hevc, which for a given architecture is possible >> without loss). The fact that this could happen at all does make it annoying >> in terms of bit-exact test expectations across arch, and this is the main >> concern, right? (for this type of metadata, it is really a hint to >> TVs/algorithms, and some will ignore it altogether) > > bitexactness is one concern, also theres the issue with what is ideally > correct. > that is what are the ideal values dictated by various standards > that hardware (cammeras, ...) aim at ? > are these rational or float or what can represent them better ? >
Both HEVC and the HDMI Info Frames use fixed-point integers (the same scales too, apparently), I do not know of the formats anything else uses. Maybe we should be using AVRationals? I would argue that MKV storing floats is a terrible idea, and someone should bonk them over the head and store fixed point as well. - Hendrik _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel