On Tue, Sep 1, 2015 at 10:55 AM, David Singer <sin...@apple.com> wrote:
> > > On Sep 1, 2015, at 10:47 , Yay295 <yay...@gmail.com> wrote: > > > > On Tue, Sep 1, 2015 at 11:30 AM, David Singer <sin...@apple.com> wrote: > > > On Sep 1, 2015, at 4:03 , Robert O'Callahan <rob...@ocallahan.org> > wrote: > > >> On Tue, Sep 1, 2015 at 8:02 PM, Kevin Marks <kevinma...@gmail.com> > wrote: > > >> QuickTime supports full variable speed playback and has done for well > over > > >> a decade. With bidirectionally predicted frames you need a fair few > buffers > > >> anyway, so generalising to full variable wait is easier than posters > above > > >> claim - you need to work a GOP at a time, but memory buffering isn't > the > > >> big issue these days. > > > > > > "GOP”? > > > > Group of Pictures. Video-speak for the run between random access points. > > > > > How about a hard but realistic (IMHO) case: 4K video (4096 x 2160), 25 > fps, > > > keyframe every 10s. Storing all those frames takes 250 x 4096 x 2160 x > 2 > > > bytes = 4.32 GiB. Reading back those frames would kill performance so > that > > > all has to stay in VRAM. I respectfully deny that in such a case, > memory > > > buffering "isn't a big issue”. > > > > well, 10s is a pretty long random access interval. > > > > There's no way to know the distance between keyframes though. The video > could technically have only one keyframe and still work as a video. > > yes, but that is rare. There are indeed videos that don’t play well > backward, or consume lots of memory and/or CPU, but most are fine. > > > > > >> What QuickTime got right was having a ToC approach to video so being > able > > >> to seek rapidly was possible without thrashing , whereas the stream > > >> oriented approaches we are stuck with no wean knowing which bit of > the file > > >> to read to get the previous GOP is the hard part. > > > > > > I don't understand. Can you explain this in more detail? > I explained the essential difference a while ago here: http://lists.xiph.org/pipermail/vorbis-dev/2001-October/004846.html The QuickTime file format defines movies that have tracks made of media; the tracks are en edit list on the media; the media have the frame layout information encoded. > > > > The movie file structure (and hence MP4) has a table-of-contents > approach to file structure; each frame has its timestamps, file location, > size, and keyframe-nature stored in compact tables in the head of the file. > This makes trick modes and so on easier; you’re not reading the actual > video to seek for a keyframe, and so on. > > > > I suppose the browser could generate this data the first time it reads > through the video. It would use a lot less memory. Though that sounds like > a problem for the browsers to solve, not the standard. > > There is no *generation* on the browser side; these tables are part of the > file format. Well, when it imports stream-oriented media it has to construct these in memory, but they can be saved out again. I know that in theory this made its way into the mp4 format, but I'm not sure how much of it is real.
