This reverts commit 489c5db079.
Treating EQUAL_MULTI_ROWS in the same way as the arbitrary-size cases is
just wrong. Consider 9 rows, 4 slices - we pick 4 slices with sizes
{ 3, 2, 2, 2 }, which EQUAL_MULTI_ROWS does not allow. It isn't possible
to split the frame into 4 slices at all with the EQUAL_MULTI_ROWS
structure - the closest options are 3 slices with sizes { 3, 3, 3 } or 5
slices with sizes { 2, 2, 2, 2, 1 }.
If the RPS we are predicting from has maximum size then at least one of
the pictures in it must be discarded before adding the current one.
Also revert 588114cea4, which added
now-redundant checks for the special case of a too-large RPS with all
pictures being in the same direction from the current one.
The previous code here did not handle passing a frames context when
ffmpeg itself did not know about the device it came from (for example,
because it was created by device derivation inside a filter graph), which
would break encoders requiring that input. Fix that by checking for HW
frames and device context methods independently, and prefer to use a
frames context method if possible. At the same time, revert the encoding
additions to the device matching function because the additional
complexity was not relevant to decoding.
Also fixes#8637, which is the same case but with the device creation
hidden in the ad-hoc libmfx setup code.
This can support encoders which want frames and/or device contexts. For
the device case, it currently picks the first initialised device of the
desired type to give to the encoder - a new option would be needed if it
were necessary to choose between multiple devices of the same type.
All of these encoders can accept libmfx surfaces directly in a hardware
frames context, or they can accept software frames if a suitable device
is supplied to use.
If we are given a non-render node, try to find the matching render node and
fail if that isn't possible.
libva will not accept a non-render device which is not DRM master, because
it requires legacy DRM authentication to succeed in that case:
<https://github.com/intel/libva/blob/master/va/drm/va_drm.c#L68-L75>. This
is annoying for kmsgrab because in most recording situations DRM master is
already held by something else (such as a windowing system), leading to
device derivation not working and forcing the user to create the target
VAAPI device separately.
Fix the quantisation offset - use the whole range, and don't change the
offset size based on bit depth.
Iterate the list in reverse order. The first region in the list is the one
that applies in the case of overlapping regions.
This removes the use of the nonstandard combined structures, which
generated some warnings with clang and will cause alignment problems
with some parameter buffer types.
Tries to find a device backed by the i915 kernel driver and loads the iHD
VAAPI driver to use with it. This reduces confusion on machines with
multiple DRM devices and removes the surprising requirement to set the
LIBVA_DRIVER_NAME environment variable to use libmfx at all.
Opening the device via X11 (DRI2/DRI3) rather than opening a DRM render
node directly is only useful if you intend to use the legacy X11 interop
functions. That's never true for the ffmpeg utility, and a library user
who does want this will likely provide their own display instance rather
than making a new one here.
For example: -init_hw_device vaapi:/dev/dri/renderD128,driver=foo
This may be more convenient that using the environment variable, and allows
loading different drivers for different devices in the same process.
Iterate over available render devices and pick the first one which looks
usable. Adds an option to specify the name of the kernel driver associated
with the desired device, so that it is possible to select a specific type
of device in a multiple-device system without knowing the card numbering.
For example: -init_hw_device vaapi:,kernel_driver=amdgpu will select only
devices using the "amdgpu" driver (as used with recent AMD graphics cards).
Kernel driver selection requires libdrm to work.
The implementation will use some default in this case. The empty string
is not a meaningful device for any existing hardware type, and indeed
OpenCL treats it identically to no device already to work around the lack
of this setting on the command line.
Attempts to pick the set of supported colour properties best matching the
input. Output is then set with the same values, except for the colour
matrix which may change when converting between RGB and YUV.
Doesn't change anything, but makes the behaviour better match that of the
other codecs (the CONSTANT_QUALITY_ONLY flag already ensures that CQP is
the only RC mode selectable for MJPEG).
Following b8c45bbcbc they contain allocated
unit arrays which will get leaked. These operations were inconsistently
applied and never actually needed (the old uninit left them in the correct
state), so just drop them entirely.
When making a new P-frame when B-frames are present the previous P-frame
is normally in the DPB because it will be referred to by subsequent
B-frames. However, this is not true if there are no B-frames, or in edge
cases where a GOP ends with two P-frames. Fix this by adding the direct
ref pics to the RPS explicitly.
Fixes#7699.
Tested-by: Ullysses A Eoff <ullysses.a.eoff@intel.com>
This attaches the logic of picking the mode of for the next picture to
the output, which simplifies some choices by removing the concept of
the picture for which input is not yet available. At the same time,
we allow more complex reference structures and track more reference
metadata (particularly the contents of the DPB) for use in the
codec-specific code.
It also adds flags to explicitly track the available features of the
different codecs. The new structure also allows open-GOP support, so
that is now available for codecs which can do it.
The specification allows 2^32-1 to be encoded as any number of zeroes
greater than 31, followed by a one. This previously failed because the
trace code would overflow the array containing the string representation
of the bits if there were more than 63 zeroes. Fix that by splitting the
trace output into batches, and at the same time move it out of the default
path.
(While this seems likely to be a specification error, libaom does support
it so we probably should as well.)
From a test case by keval shah <skeval65@gmail.com>.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Assertion avctx->codec_id != AV_CODEC_ID_NONE failed at src/libavcodec/parser.c:128
The setting on the internal AVCodecContext used for parsing only is
otherwise irrelevant, so just set it to avoid the assert.
The SAR of the input could have a numerator or denominator greater than
2^16 which would then be truncated to a 16-bit integer when written to
the VUI parameters, giving a random result. Instead, reduce the SAR to
the nearest representable fraction.
Fixes#7502.
These may be used by hwaccel decoders when the standard tables are not
otherwise available. At the same time, clean up that code into an array
so it's a little less repetitive.
Supports both prefix and suffix SEI, decoding all of the common SEI
types and some more obscure ones. Most of this is tested by the
existing tests in fate.
Adds an option to specify the number of tile rows and columns, then uses
a uniform tiling if possible and otherwise a fixed tiling with equal-sized
tiles to fill the frame.
Also adds -tile-columns and -tile-rows options to make tilings with
power-of-two numbers of tiles, matching the behaviour of the libvpx/VP9
encoder.
Mark Thompson
