This should allow OpenMW to work better with git versions of openscenegraph. OSG dev version 3.5.5 added the setting of thread affinity for the main thread. The problem is that in the boost/standard threading libraries, the affinity of a thread is inherited by any further threads launched from that thread, leading to these threads always running on the same core as the main thread unless you tell them not to.
With OpenThreads, the default affinity of a thread is none, no matter what parent thread it was launched from.
So, when using custom threading with OSG 3.6+, we have these options:
1. explicitely tell OSG to *not* set the thread affinity
or 2. explicitely set the thread affinity of additional threads created (possible with boost, but not possible with std::thread)
or 3. use OpenThreads
or 4. accept the suboptimal performance of non-OSG threads (in OpenMW's case the sound streaming & video threads) running on the same core as the main thread
This patch opts for 3.)
Reference: http://forum.openscenegraph.org/viewtopic.php?t=16158
Instead of getting the loudness data for the whole file in advance, we now get it piece by piece as the sound is streamed.
The benefit is that we need to decode the audio just once instead of twice.
We no longer need to rewind() the stream when the first decoding is done, this should hopefully fix bug #3453 .
In between the startUpdate/finishUpdate calls, changes are deferred so that
they can happen all at once. This includes starting sounds, so when the
underwater sound is started it will be immediately checked to see if it's
playing. Since it's not yet playing, it'll be seen as stopped and get cleaned
up before ever playing.
Can't instantiate a container (at least with MSVC) without knowing the exact size of the object being stored, forward-declares only work with pointers.
I couldn't see a simple way to remove the forward declare, so pointers and memory management it is.
This helps avoid a lock during the movie player's read method, since it needs
to sync with the current playback position which would otherwise need to get
the movie decoder's current position.
The cache size is specified with a min/max range, intended to avoid constant
unloading once the limit is reached. This way, buffers can be unloaded down to
a reasonable mimimum, allowing some more buffers to be subsequently loaded
without causing more unloading.
A deque allows existing Sound_Buffer references to remain valid as long as new
ones are back-inserted. These references can be used instead of indices.
Certain OpenAL implementations, including Rapture3D, Creative's hardware
drivers, and more recent versions of OpenAL Soft, can batch together changes so
that they all occur at once, avoiding potential discontinuities with one sound
being changed before another, or the listeenr being changed before sounds are.
On other implementaitons, this is a no-op and maintains existing behavior.
This simply sets up the Sound record data to be used by the sound output. The
actual audio buffers, stored in the Sound_Handle, are still loaded on-demand.
Voices tend to be a bit long, and not individually replayed often. So it's
better to stream them instead of loading theminto a sound buffer. The loudness
data is very small, though, so that can be kept buffered indefinitely.
Now it returns true if *any* sounds matching the given Ptr and id are playing. The previous behaviour was causing problems with "zombie" sounds (sounds that have finished playing, but weren't removed from the map yet) making the isPlaying method return false even though there's another legitimately playing sound in the map.
Note, I suspect Rng::rollClosedProbability() is not needed. The only difference between it and rollProbability() is that one time in 37k (on Windows), it will give an output of 1.0.
On some versions of Linux, the value of 1.0 will occur about 1 time in 4 billion.
conversion from 'const float' to 'int', possible loss of data
conversion from 'double' to 'int', possible loss of data
conversion from 'float' to 'int', possible loss of data
- Don't try to use float audio or extended channel layouts if the hardware does not support them
- Add channel layout resampling support to ffmpeg_decoder
- Fix rindex overflow
- Fix audio sample size bugs (was using sample_fmt and channel count of the decoder, instead of the resampled settings). We didn't notice this bug before, because the OpenAL MovieAudioFactory tries to resample to a format of the same byte size.
- Add support for play/pause and seeking controls (not used by cutscenes in OpenMW)
- Closing the video when arriving at the stream end is now handled by the user (we may also want to keep the video open and seek back)
The video player now has a standalone demo, at https://github.com/scrawl/ogre-ffmpeg-videoplayer
- Split video player to separate source files.
- Move video player engine sources to extern/ (repository will be set up on github soon).
- Audio is handled in a MovieAudioFactory, implemented by the user (here in MWSound subsystem).
- Handle conversion of unsupported channel layouts via ffmpeg's swresample.
AV_NOPTS_VALUE is casted to int64_t when compared with pts,
so with libavc >= 56.1 it should be no-op because in that version
it's already signed int.
Signed-off-by: Lukasz Gromanowski <lgromanowski@gmail.com>
Added ifdefs because changes in API were introduced in libavc 56.1
and this code doesn't compile with older versions (ie. on Ubuntu, or Debian)
Signed-off-by: Lukasz Gromanowski <lgromanowski@gmail.com>
- Added REQUIRED to find_package(FFmpeg)
- Removed USE_FFMPEG option from CMakeLists.txt
- Always use FFmpeg for sound input
- Removed SOUND_DEFINE from CMakeLists.txt
- Removed #else branch from videoplayer.cpp with dummy VideoState code
(FFmpeg is now guaranteed to exist and the code was incomplete)
- Remove #ifdef OPENMW_USE_FFMPEG in ffmpeg_decoder.cpp, it is guaranteed to be used
- Remove #ifdef OPENMW_USE_FFMPEG from soundmanagerimp.cpp, it is guaranteed to be used
