RTLD_LAZY means that symbols will only be resolved when first used,
while RTLD_NOW tries to resolve them immediately. This means that if
there are missing symbols (e.g, because a function got removed from
libobs), dlopen with RTLD_LAZY will happily open that module but we get
a runtime crash when the module tries to use that symbol, while with
RTLD_NOW we instead get a (nicer) error on dlopen.
These were likely originally used for the obs_*_get_module functions,
but no longer are. Additionally, they weren't even implemented for
anything but obs_source_info. Let's remove this before it hits stable.
Phase 1 of the plugin manager provides the the ability to toggle off/on
plugins loading when OBS starts. Additionally, it implements loading of
a manifest file for plugins that allows plugin authors to provide more
detailed information about the plugin including authors, support site,
name, description.
In order to accomplish this, this change updates libobs to provide
more detailed tracking of modules- specifically tracking both enabled
and disabled modules, alone with a module load state which indicates
why a module is not loaded. Additionally, changes were made to establish
a links between a module and any features (inputs, outputs, encoders,
and services) it provides (and thus the ability to determine why a
feature might not be enabled). Along with these changes to modules,
this commit also provides an indicator and lookup for core modules which
can not be disabled by the plugin manager.
Finally, this change provides functions to properly load and retrieve
a standardized plugin metadata file.
The autoselect functionality was meant to enable user feedback for when
data was selected, but something else actually used, for example in case
the value "Auto" was chosen, or an incompatible one.
However, it has a few problems:
1. It was not really used, the only places in OBS were the legacy Video
Capture Device on macOS and the Video Capture Device on Windows.
2. It was never even fully implemented in the UI. In fact, the
properties view *only* supports it for obs_property_list, despite the
API theoretically allowing any data/property type.
3. The core obs_data seems like the wrong place for this kind of
interaction between plugins and UI. obs_data should carry data from one
place to another, some built-in feedback mechanism for the other
direction doesn't make too much sense. If the functionality is desired,
I think this should be rearchitectured as some place of explicit data
transfer.
Let's mark it as deprecated for future removal.
On MSVC, PRAGMA_WARN_DEPRECATION already does that, but on other
compilers it still throws a warning. PRAGMA_DISABLE_DEPRECATION is for
(rare) situations where it's intentional and correct that the deprecated
API is still used, such as to make sure some other API still works
internally.
This fixes the following issue:
- when an Audio Capture source (like 'Desktop Audio',
'wasapi_output_capture' or 'pulse_output_capture') and monitoring use
the same device, one gets duplicated audio for any monitored source,
since their audio is added through both the current scene and the output
capture.
This creates a shift in audio level (about 3 dBFS for white noise, and
oftentimes for music; up to 6 dBFS for very low frequency sine waves).
We solve the bug by:
- detecting when an Audio Capture source and monitoring share the same
device;
- silencing monitored sources whenever the Audio Capture source is
unmuted.
Signed-off-by: pkv <pkv@obsproject.com>
This fixes the following bug:
- a source might be copied into the same scene or through a nested scene.
The audio level will then increase by +6 dBFS.
An earlier fix [1] dealt with this bug at the scene audio rendering
level, which leaves some edge cases since the fix is not located
directly in the core audio callback.
The current fix consists in:
- tagging individual sources which appear several times in the audio
tree at each tick;
- promote them to root_nodes sources;
- bypass their mixing in scenes and transitions.
Due to being mixed as root_nodes, the audio of duplicated sources
appears only once in the final audio mix.
[1] https://github.com/obsproject/obs-studio/pull/10537
Signed-off-by: pkv <pkv@obsproject.com>
This reverts commit: 8a38e33 [1].
This also reverts commit: 72924ac [2].
(in favour of a fix in core audio pipeline, which allows to address
edge cases).
[1] libobs: Mix audio of each source in a scene only once
8a38e3375b
[2] libobs: Deduplicate audio for nested scenes/groups if not
transitioning
72924ac1f3
Signed-off-by: pkv <pkv@obsproject.com>
SIMD Everywhere finder needs to be installed alongside libobs CMake
package since its headers depends on it.
C++ cmath header is included on Windows on ARM to ensure that all math
functions are correctly defined when included in C++ code.
https://github.com/simd-everywhere/simde/issues/1304
macOS intrinsics is included to make sure that no redefinition error
happens.
Create a specific "out of memory" variant of os_breakpoint to facilitate
troubleshooting and to differentiate an "out of memory" crash from a
more generic crash.
The interleaver requires packets with sufficiently high timestamps from
all encoders to be available to send out before it outputs a packet.
This means that if one encoder stalls and does not submit packets for
a while the interleaver can end up with a permanent increase to its
delay due to multiple packets being received from the other encoders
without being able to send any.
This change introduces a recovery mechanism that relies on an upper
bound for the amount of packets that may become "streamable" at a time
being calculated, and flushing additional packets if that bound is
violated.
The upper bound is based on the largest expected interval between
packets and how many packets for each encoder could arrive within
double that interval (to give us a bit of wiggle room since not
all encoders will deliver frames 100% consistently).
This ensures that we still send data at a steady rate but allow the
buffer to drain if it becomes larger than it needs to be.
The OpenGL frontend is implemented in Metal for Apple Silicon Macs.
Metal does not provide a direct "clear" command (neither do D3D12 or
Vulkan) and instead a clear is a load command for a render target of a
render pass.
All writes (and color calculations) in Metal happen in linear space,
with appropriate gamma encoding and decoding happening automatically
depending on the pixel format of the the textures attached to the
pipeline.
Thus setting a clear color with linear gamma is the correct approach to
avoid the SRGB gamma curve being applied twice for both scenarios.
This fixes a change in PR 11605 that breaks Decklink video outputs.
get_mix_for_video() returns NULL for those outputs, causing
start_raw_video() to return before connecting the output.
A few recently introduced graphics API functions needed for Enhanced
Broadcasting were marked as `GRAPHICS_IMPORT_OPTIONAL`. Add the
proper checks in the wrapper to ensure they are safe on platforms
that do not implement the underlying function.
FreeBSD already implemented os_get_free_size. Move the new Linux
implementation into the existing not-__FreeBSD__ block.
Fixes: 935613816f ("libobs/util: Update `os_get_free_size()`")
This fixes a regression on Linux, introduced in commit 9bc3082.
According to the POSIX specification for the exec family of commands,
"The first argument is the filename or pathname of the executable to
be executed". This was done correctly before, but the above commit
removed "sh" from the arguments, breaking the pipe function on Linux.
With the removal of all legacy code paths, obsconfig.h always exists and
the compile definition always gets set. As such, it's no longer
necessary to check for it.
As removing the definition itself could be seen as a breaking change,
this simply moves the definition to pc.in and cmake.in files for now to
preserve the value for plugins that might expect this to be set. We may
remove the definition entirely in a later release.
In `parse_decl_string`, it jumps to `fail` label if an error occurs.
However, if the lexer encountered an error for example, `cfp.error_list`
may be empty.
The frontend has set a log handler which locks the mutex in `LogString`.
If `os_dlopen` fails in the child process, it calls `blog` and attempted
to lock the mutex that may have already been locked by another thread
before `fork()`. This change prevents the child process from locking the
mutex, resolving the potential deadlock.
