Optimization for the referenced video case.

With referenced video from a DCP decoder, no video will ever
be fetched from the decoder.  Hence the code to discard given video
will be activated after _decoded builds up to the magic size.

Before this commit the code would attempt to fill with black up to
given frame N (with N very large) from the last frame in _decoded when
_decoded had been trimmed.  This would result in exponential growth
in execution time for the VideoDecoder::give() path.

diff --git a/src/lib/video_decoder.cc b/src/lib/video_decoder.cc
index edc746010f082a0558617763798544b750a940d1..f856d93aa7e81cd81bcb0a472903d03f2944aee7 100644 (file)
--- a/src/lib/video_decoder.cc
+++ b/src/lib/video_decoder.cc
@@ -294,39 +294,45 @@ VideoDecoder::give (shared_ptr<const ImageProxy> image, Frame frame)
        /* If we've pre-rolled on a seek we may now receive out-of-order frames
           (frames before the last seek time) which we can just ignore.
        */
-
        if (from && (*from) > to_push.front().frame) {
                return;
        }
 
-       if (from) {
-               switch (_content->video->frame_type ()) {
-               case VIDEO_FRAME_TYPE_2D:
-                       fill_one_eye (from->index(), to_push.front().frame.index(), EYES_BOTH);
-                       break;
-               case VIDEO_FRAME_TYPE_3D:
-               case VIDEO_FRAME_TYPE_3D_LEFT_RIGHT:
-               case VIDEO_FRAME_TYPE_3D_TOP_BOTTOM:
-               case VIDEO_FRAME_TYPE_3D_ALTERNATE:
-                       fill_both_eyes (from.get(), to_push.front().frame);
-                       break;
-               case VIDEO_FRAME_TYPE_3D_LEFT:
-                       fill_one_eye (from->index(), to_push.front().frame.index(), EYES_LEFT);
-                       break;
-               case VIDEO_FRAME_TYPE_3D_RIGHT:
-                       fill_one_eye (from->index(), to_push.front().frame.index(), EYES_RIGHT);
-                       break;
+       int const max_decoded_size = 96;
+
+       /* If _decoded is already `full' there is no point in adding anything more to it,
+          as the new stuff will just be removed again.
+       */
+       if (_decoded.size() < max_decoded_size) {
+               if (from) {
+                       switch (_content->video->frame_type ()) {
+                       case VIDEO_FRAME_TYPE_2D:
+                               fill_one_eye (from->index(), to_push.front().frame.index(), EYES_BOTH);
+                               break;
+                       case VIDEO_FRAME_TYPE_3D:
+                       case VIDEO_FRAME_TYPE_3D_LEFT_RIGHT:
+                       case VIDEO_FRAME_TYPE_3D_TOP_BOTTOM:
+                       case VIDEO_FRAME_TYPE_3D_ALTERNATE:
+                               fill_both_eyes (from.get(), to_push.front().frame);
+                               break;
+                       case VIDEO_FRAME_TYPE_3D_LEFT:
+                               fill_one_eye (from->index(), to_push.front().frame.index(), EYES_LEFT);
+                               break;
+                       case VIDEO_FRAME_TYPE_3D_RIGHT:
+                               fill_one_eye (from->index(), to_push.front().frame.index(), EYES_RIGHT);
+                               break;
+                       }
                }
-       }
 
-       copy (to_push.begin(), to_push.end(), back_inserter (_decoded));
+               copy (to_push.begin(), to_push.end(), back_inserter (_decoded));
+       }
 
        /* We can't let this build up too much or we will run out of memory.  There is a
           `best' value for the allowed size of _decoded which balances memory use
           with decoding efficiency (lack of seeks).  Throwing away video frames here
           is not a problem for correctness, so do it.
        */
-       while (_decoded.size() > 96) {
+       while (_decoded.size() > max_decoded_size) {
                _decoded.pop_back ();
        }
 }