On some systems (eg. mine), VAAPI fails on vaInitialize. Valgrind was
able to spot that the device was not being closed, and it appears to
have been correct. This fixes a memory leak.
Previously, we just checked for H264 regardless of which codec was
selected. This mostly worked, but on Fedora they ship AV1 without H264
or HEVC. In that case users wont see AV1 as an option.
During high graphics thread pressure it can take a significant time to
acquire the graphics lock. This change releases the OpenGL texture after
rendering to avoid the 2nd lock after sending the frame to FFmpeg. This
improves 99%-tile/100%-tile and median encode in a near encoder overload
scenario, and modestly raises the ceiling before encoder overload in my
test scene.
Master:
min=0 ms, median=4.29 ms, max=33.072 ms, 99th percentile=8.877 ms
min=0 ms, median=4.438 ms, max=77.157 ms, 99th percentile=9.853 ms
min=0 ms, median=4.527 ms, max=57.292 ms, 99th percentile=9.282 ms
This commit:
min=0.97 ms, median=3.009 ms, max=13.215 ms, 99th percentile=5.899 ms
min=1.181 ms, median=2.91 ms, max=9.854 ms, 99th percentile=5.56 ms
min=0.461 ms, median=3.013 ms, max=10.693 ms, 99th percentile=5.871 ms
Not strictly necessary, and does not fix any bug. This just corrects a
nitpick that technically audio encoder packets are start points, but
that they aren't being labeled accordingly as keyframes.
The frame rate used to initialize an AMF encoder should be aligned
with the derived frame rate in video_output_info instead of the global
obs_video_info structure. With this change, IDRs can be aligned when
multiple renditions are being encoded.
Using video_output_info members for the format, colorspace, and range
parameters in addition to the frame rate provides a single source for
this information and obs_video_info is no longer needed.
FFmpeg is already very effective at detecting the correct input format
for socket-style protocols (SRT, RIST, TCP, UDP, etc.). By overriding
the format of SRT and RIST manually to mpegts, the user is being
prevented from using other container formats via these protocols.
In the rare case that libavformat is unable to detect the correct
container format for an SRT or RIST stream, the user may manually
specify that format in the existing format field.
With the FFmpeg options field which was recently added, probe options
may be specified by the user to further tune format detection.
Add a COMPAT version of NV_ENC_INITIALIZE_PARAMS_VER, which changed
between SDK 12.0 and 12.1.
If using drivers older than the configured SDK/API requires, fall back
to the compatibility version of the API (11.1).
An example scenario would be encoding in H.264 on driver version 512.15.
SDK 12.1 requires driver version 531.61 on Windows. Currently, in this
scenario, OBS will fall back to the FFmpeg encoder path, which will do
its own driver check, which will fail on this driver version.
With this change, this scenario will instead leverage the existing
compatibility hack to fall back to the compatibility version of the API
without falling back to FFmpeg.
It should be noted that while neat, the compatibility hack that enables
Kepler GPUs to continue to be able to use NVENC in this way is expected
to be removed at some point.
Because 4a8e8bb1bd changed the
NVENCAPI_STRUCT_VERSION define to rely on enc->needs_compat_ver, we must
do the conditional check and define needs_compat_ver before any structs
are initialized because they will check the version when doing so by
calling NVENCAPI_STRUCT_VERSION internally.
In this case, make sure that we define enc->needs_compat_ver before
NV_ENCODE_API_FUNCTION_LIST_VER is called for
NV_ENCODE_API_FUNCTION_LIST.
Previously, we always assumed we could push more frames before an encoded
packet would be available. Actually, we can get EAGAIN when the
frame queue is full and before any encoded packets have been completed.
Instead, attempt to read packets more often as the worst that can
happen is we get EAGAIN again.
Frame rate should be defined before AMF encoder initialization,
because this information is used for setting vui_time_scale in SPS.
If frame rate isn't defined before initialization,
then AMF encoder set default frame rate into VUI header (30 fps).
To avoid a compiler error, this commit also bumps the configured NVENC
version to match the updated version in the deps package.
Notable changes:
* deps.ffmpeg: Update zlib to 1.3
* deps.ffmpeg: Update SVT-AV1 to 1.8.0
* deps.ffmpeg: Update aom to 3.8.0
* deps.ffmpeg: Update libsrt to 1.5.3
* deps.ffmpeg: Update nv-codec-headers to 12.1.14.0
* deps.ffmpeg: Update AMF to 1.4.32
* deps.ffmpeg: Update FFmpeg to 6.1
* deps.macos: Update LuaJIT to 2.1 c525bcb902
* deps.macos: Update FreeType to 2.13.2
* deps.macos: Update Asio to 1.29.0
* deps.macos: Update nlohmann JSON to 3.11.3
* deps.macos: Update Sparkle to 2.5.2
* deps.macos: Update Syphon Framework to 5.0 39e31383ff
* deps.windows: Update FreeType to 2.13.2
* deps.windows: Update curl to 8.5.0
* deps.windows: Update LuaJIT to 2.1 c525bcb902
* deps.windows: Update Asio to 1.29.0
* deps.windows: Update nlohmann JSON to 3.11.3
* deps.windows: Update VPL to v2.10.1
* deps.qt: Update Qt6 to 6.6.1 for Windows
* deps.qt: Update Qt6 to 6.6.1 for macOS
If a user sets both AdaptiveMiniGOP=true and EnablePreAnalysis=true
in the AMF/FFmpeg options field, AMF will adaptively insert
B-pictures, and no longer uses the fixed B pattern.
For a fixed B-frames pattern, it is expected that increasing B-frames
can cause a quality drop for certain content such as with high motion.
AdaptiveMiniGOP is recommended when using B-frames to improve the
quality in such cases. AdaptiveMiniGOP is dependent on PreAnalysis
which means that trying to enable it without having PreAnalysis turned
ON will have no negative effect (AdaptiveMiniGOP won't be enabled).
