Comment.

[dcpomatic.git] / src / lib / writer.cc

diff --git a/src/lib/writer.cc b/src/lib/writer.cc
index 93f5ed6d666d4395f16cf8ac09ce49dd84d50f0b..47c6b838142619a355cab1f0fbc298f84586df9b 100644 (file)
--- a/src/lib/writer.cc
+++ b/src/lib/writer.cc
@@ -570,7 +570,16 @@ operator== (QueueItem const & a, QueueItem const & b)
 void
 Writer::set_encoder_threads (int threads)
 {
-       _maximum_frames_in_memory = lrint (threads * 1.1);
+       /* I think the scaling factor here should be the ratio of the longest frame
+          encode time to the shortest; if the thread count is T, longest time is L
+          and the shortest time S we could encode L/S frames per thread whilst waiting
+          for the L frame to encode so we might have to store LT/S frames.
+
+          However we don't want to use too much memory, so keep it a bit lower than we'd
+          perhaps like.  A J2K frame is typically about 1Mb so 3 here will mean we could
+          use about 240Mb with 72 encoding threads.
+       */
+       _maximum_frames_in_memory = lrint (threads * 3);
 }
 
 void