Makes the callback variable "removed" atomic, and on script unload,
first sets all callbacks to removed before actually unloading the script
out of a safety precaution. (See note at the bottom for further details)
This minimizes the possibility of a race condition where the script
callback could be called while those callbacks were being removed.
Big note for this change, this change should eventually be replaced with
a reference counting ownership method where script callbacks can hold a
reference and share ownership of the script if it's still alive while
the script callback is being called. That way the script callbacks can
safely execute. May require a fair amount of reworking of the script
object.
Deprecates:
obs_source_addref()
obs_output_addref()
obs_encoder_addref()
obs_service_addref()
obs_scene_addref()
These functions should be considered unsafe and not used. Instead, use:
obs_source_get_ref()
obs_output_get_ref()
obs_encoder_get_ref()
obs_service_get_ref()
obs_scene_get_ref()
These functions return a pointer to the incremented object only if the
object is still valid, otherwise they will return null, indicating that
the object is no longer valid or is unsafe to use.
The reason why this is being done is because certain third party plugins
seem to be using addref, and are somehow managing to call addref on
sources that have already been fully released. For the sake of safety,
almost all usage of these functions within OBS have also been replaced
as well.
We currently pass a 'buffer_size' option to avformat in bytes.
Unfortunately, avformat/librist.c admits an option with the same name but in milliseconds.
The default values set in the UI (2 MB) go beyond the range admitted in ms (0-30 000)
so avformat issues errors.
This fix disables the option when rist protocol is detected.
Signed-off-by: pkv <pkv@obsproject.com>
AVFormatContext::oformat was made const on April 27, 2021 [1]. If we
respect the constness of AVOutputFormat and do not cast results from
FFmpeg functions to non-const, we cannot modify the results after the
fact. Our choices are either to cast them to non-const (and presumably
have them implicitly casted back to const on later function calls), or
only try to modify the results in versions of FFmpeg where these are not
expected to be const.
Instead of relying on casts, we can set the encoder values in the
ffmpeg_cfg struct, which are later passed to new_stream.
Also modifies deps/media-playback. Removes compiler warnings.
Some avformat functions return const AV(In/Out)putFormat per [1], so
ifdef as needed.
[1]: 56450a0ee4
lavf 59.0.100 avformat.h
avformat: Constify the API wrt AV(In|Out)putFormat
Also constify AVProbeData.
If there is a PTS discontinuity in the playback stream that causes the PTS to
reset, it is likely that new audio frames will come in while there are still
video frames waiting to be dequeued. This will cause the audio frames to be
played while the video frames wait for the PTS to get back up to where they
were before the discontinuity.
Removes prior attempt to expose libcaption headers which really shouldnt
have public. This instead moves the obs-internal include out of the
public obs-scene.h and into it's implementation.
With certain audio encoders, gaps can be introduced into the audio
packets, causing the audio duration to be miscalculated because it
calculated audio duration based upon the PTS of the current packet to
the last packet. However, this audio encoder also did not store
timestamps for most of its audio packets, causing PTS values to be
calculated based upon duration values. So those two things combined
caused audio timestamps to go all out of whack when playing back certain
videos with the media source. This is particularly prevalent with WMV
files using Microsoft codecs.
So to fix it, when the duration needs to be calculated, just calculate
the duration based upon the sample count of the audio packet. This fixes
the issue with the video in question that caused problems, and likely
fixes issues in a lot of videos that may have been floating out there
for some time. Basically this is a fix for a potential long-standing
issue.
Closesobsproject/obs-studio#3683
Returning PyObject with borrowed reference will result in double free
and/or use after free issue. Issue seen as crash when running Python
script calling script_path().
Fixes a bug where playback would completely stop when manually seeking
to the end of the file. Playback should not officially stop until
playback has unpaused.
If seeking occurs too close to the previous time value, it could cause
the playback to delay an abnormal amount of time, or if the time value
is under the previous value, then it would cause an assert. So instead,
reset the next timestamp value to be instant in both cases if a seek
occurs to ensure timely playback.
(Note: This commit also modifies deps/media-playback)
Co-authored-by: Eric Lindvall <eric@5stops.com>
Co-authored-by: Ryan Foster <RytoEX@gmail.com>
Co-authored-by: Jim <obs.jim@gmail.com>
FFmpeg YUV to RGB conversions require extra destination padding.
The Mantis issue might have been fixed when some YUV format conversions
were moved to the GPU, but this may fix other YUV formats.
Fixes https://obsproject.com/mantis/view.php?id=1177
Instead of erroring out completely when it can't determine if the
certificate is valid, proceed anyway. This matches how web browsers
treat failed cert revocation checks. schannel just has somewhat
paranoid defaults.
Add PyDict_New and PyTuple_New to obs-scripting-python-import.[ch];
these functions are used by SWIG's generated code when I build OBS on
macOS with SWIG 4.0.1.
When using an IP camera on a local network, we wanted to minimize
delay. In order to achieve minimum delay, we allowed Media Source to
set BufferingMB to 0, and when it is 0, also enable AVFMT_FLAG_NOBUFFER
in the AVFormatContext flags.
SIMDE was introduced for aarch64 support, however, the library itself
supports non-SIMD fallback, which allows us provide support to other
platforms without code changes.
There is another world beyond x86. So we can simply enable SIMDE for
processors without SSE2 support.
Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
This fixes an issue where local files were using the avformat interrupt
callback, which is only supposed to be use to prevent blocking on
network calls.
The interrupt callback is designed to prevent the media source from
blocking; FFmpeg will internally call it periodically to prevent FFmpeg
function calls from blocking too long, and allow the caller to determine
whether blocking should stop.
The problem with this however is that AVERROR_EXIT causes the thread to
completely exit. This fixes it so that it treats it as an EOF rather
than as an abnormal error.
The current scripting directory path given to Python on OSX is
`../data/obs-scripting`. This works when bin/ and data/ are in the same
folder like rundir after a compile and in the old .app bundles but the
new normal .app bundle structure is different.
For bundles move obspython.py and _obspython.so to the binary folder,
just like the Lua one, and pass that bin path to the Python interpreter
so it can find it.
The current working dir isn't guaranteed to be set with .app bundles
so look up binary path at runtime.
_obspython.so currently links directly against python as @rpath/Python
and has 3 common python install paths set as rpaths so it would only
work if a user had python installed at one of those hardcoded paths.
Don't link _obspython against python at all but instead link it with
"-undefined dynamic_lookup" so when it is imported by python all it's
undefined python symbols get resolved at runtime against the user
supplied python instance loaded into the process earlier.
The rpaths aren't needed anymore this way.
The libpython .dylib files on MacOS are in the python_path/lib
subdir. The user supplied python_path still needs to be to the actual
python root and not lib/ as the root path get's used as python home dir
so just add lib/ when loading the dylib.
Cuda appears to be a bit buggy and can result in corruption where as
it's not being seen with other hardware decoders. Remove cuda for now
as a hardware decoder to prevent the possibility of corruption.
Fixes a bug where hardware frames would not transfer to RAM for the last
few frames of a video. The 'ret' variable can be 0 even though there
are still frames ready.
The hardware accelerated decoder context needs to be explicitly unrefed
when it's no longer in use, otherwise it and many resources associated
with it will leak.
When hardware accelerated decoding is enabled, sometimes it can't
initialize for whatever reason, so it will fall back to software on its
own. When this occurs, it will not use the hardware pixel format on the
frame; instead it will defer to a standard format on the frame. So if
the frame format does not match the expected format, assume software
decoding. (This was also what the hw-decode.c FFmpeg example did if the
format did not match the expected format)
I40A -> RGBA, 1080p, CPU
swscale (MMX), Intel i7-6700: 1910 us -> 0 us
GPU measurements used Intel GPA with SetStablePowerState.
I40A -> RGBA, 1080p, GPU
Intel HD Graphics 530: 0 us -> 677 us
NVIDIA RTX 2080 Ti: 0 us -> 43 us
PatTheMav
jpark37
jp9000
pkv
Ryan Foster
Scratch
Eric Lindvall
wangshaohui
Georges Basile Stavracas Neto
Kurt Kartaltepe
Colin Edwards
Exeldro
Per Heed
Richard Stanway
nkmerrill
yoho
Jimi Huotari
Clayton Groeneveld
Ed Maste
comex
LiamCoal
Jiaxun Yang
pkubaj
Rat
Peter Geis
Matt Gajownik
luz.paz
Ioan Loosley
VodBox
Jason Francis
Michael Fabian 'Xaymar' Dirks