Add incomplete design doc.

[dcpomatic.git] / doc / design / timing.tex

diff --git a/doc/design/timing.tex b/doc/design/timing.tex
index 567ba024fbd7fae77aa2015747e76b71d076682b..0b6d73f3a259bc579cb4183f511c006116da8fca 100644 (file)
--- a/doc/design/timing.tex
+++ b/doc/design/timing.tex
@@ -8,7 +8,8 @@ We are trying to implement full-ish playlist based content specification.  The t
 Frame rates of things can vary a lot; content can be in pretty much
 anything, and DCP video and audio frame rates may change on a whim
 depending on what is best for a given set of content.  This suggests
-(albeit without strong justification) the need for a frame-rate-independent unit of time.
+(albeit without strong justification) the need for a
+frame-rate-independent unit of time.
 
 So far we've been using a time type called \texttt{Time} expressed in
 $\mathtt{TIME\_HZ}^{-1}$; e.g. \texttt{TIME\_HZ} units is 1 second.
@@ -37,6 +38,57 @@ scaling video, resampling audio).  The resampling is awkward, though,
 as you really need one resampler per source.  So it might make more sense
 to put stuff in the decoder.  But then, what's one map of resamplers between friends?
 
+On the other hand, having the resampler in the player is confusing.  Audio comes in
+at a frame `position', but then it gets resampled and not all of it may emerge from
+the resampler.  This means that the position is meaningless, and we want a count
+of samples out from the resampler (which can be done more elegantly by the decoder's
+\texttt{\_audio\_position}.
 
 
+\section{Options for what \texttt{Time} is a function of}
+
+I've been trying for a while with \texttt{Time} as a wall-clock
+`real-time' unit.  This means that the following is tricky:
+
+\begin{enumerate}
+\item Add content at 29.97 fps
+\item Length of this content is converted to \texttt{Time} using the
+  current DCP frame rate (which will be 29.97).
+\item Add more content at 25 fps.
+\item This causes the DCP frame rate to be changed to 25 fps, and so
+  the first piece of content is now being run slower and so its length
+  changes.
+\end{enumerate}
+
+I think this is the cause of content being overlapped in this case.
+
+It is tempting to solve this by making Time a subdivsion of DCP video
+frame rate.  This makes things nicer in many ways; you get a 1:1
+mapping of content video frames to Time in most cases, but not when
+video frames are skipped to halve the frame rate, say.  In this case
+you could have a piece of content at 50 fps which is some time $T$
+long at at DCP rate of 50 fps, but half as long at a DCP rate of 25 fps.
+
+I'm fairly sure that there is inherently not a nice representation which
+will obviate the need for things to be recalculated when DCP rate changes.
+
+On the plus side, lengths in \texttt{Time} are computed on-demand from
+lengths kept as source frames.
+
+
+\section{More musings}
+
+In version 2 things we changed, and a problem appeared.  We have / had
+\texttt{ContentTime} which is a metric time type, and it is used to
+describe video content length (amongst other things).  However if we
+load a set of TIFFs and then change the frame rate we don't have the
+length in frames i order to work out the new rate.
+
+This suggests that the content lengths, at least, should be described
+in frames.  Then to get metric lengths you would need to specify a
+timecode.
+
+I will probably have to try a frame-based ContentTime and see what
+problems arise.
+
 \end{document}