not a bugfix as this path was only used for float16 but it makes sense
to allow using this for integers too
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Fixes failures in some systems since 171060d5dc.
This can be further improved by only allocating the arrays when needed.
Signed-off-by: James Almer <jamrial@gmail.com>
ATM this performs as well or better as any other algorithm tried.
Its simple for the decoder.
On the encoder side complexity depends on how parameters are
chosen. But with a fixed mul_count of 1 and basic heuristic
it performs as well as any more complex choice i tried so far.
The encoder code here is flexible and allows mul_count > 1
and also can easily be used to exactly test various parameters.
With mul_count=512 we can gain another 6% in remap table size
for fixed point in float data. (this is not implemented in this
patch though)
Sponsored-by: Sovereign Tech Fund
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This is a replacement in ffmpeg for the deprecated avcodec flag AV_CODEC_FLAG_DROPCHANGED.
This option is meant to be used when the filtergraph should not be
reinited upon input parameter changes as that leads to loss of state
in the filtergraph potentially leading to broken or aborted output,
e.g. inserting of silence with first_pts specified in aresample.
Generally useful to avoid corrupted yet decodable packets in live
streaming inputs.
This option when enabled takes precedence over reinit_filters
The existing default value is the same as the primary readrate.
This maintains the flow rate as best as possible but does not
reduce the lag. This new value allows catchup to gradually happen
without overwhelming the receiver.
Addresses a concern brought up in #11469
Otherwise the decoder may enter an loop of repeated state resets.
This can happen for streams with CRA but no IDR.
Signed-off-by: Zhao Zhili <zhilizhao@tencent.com>
It allows the callee to clobber the MMX state,
yet since 1e3dc705df this is no longer
done. So use the stricter declare_func instead.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This commit adds option for enabling SmartAccess Video (SAV)
in AMF encoders. SAV is an AMD hardware-specific feature which
enables the parallelization of encode and decode streams across
multiple Video Codec Engine (VCN) hardware instances.
SDL2's headers #define HAVE_PTHREAD_SETNAME_NP (to one here);
we also have a config define which leads the compiler to
warn about this when the values are distinct (as they are here).
Suppress this warning by including config.h before SDL.h.
The latter will then redefine our define which doesn't matter
and which does not lead to warnings because compilers don't warn
about system headers by default (depending upon -Wsystem-headers).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The field width in a %s directive gives the amount of characters
to read from the input and not the size of the receiving buffer;
the latter must be of course also have space for the trailing \0
which has been forgotten here. The commit adds it (and fixes a
-Wfortify-source warning from Clang).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The underlying integer type of an enumeration is
implementation-defined (see C11, 6.7.2.2 (4)); GCC defaults
to unsigned if there are no negative values like for all enums
from pixfmt.h except enum AVPixelFormat.
This means that tests like "if (csp >= AVCOL_SPC_NB)" for
invalid colorspaces need not work as expected (namely if
enum AVColorSpace is signed). It also means that testing
for such an enum variable to be >= 0 may be tautologically
true. Clang emits a -Wtautological-unsigned-enum-zero-compare
warning for this.
Fix both of these issues by casting to unsigned.
Also do the same in libswscale/format.c.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It is undefined behavior. Pass a dummy buffer instead.
Fixes "runtime error: applying non-zero offset 1024 to null pointer".
affected the SVQ1 vsynth FATE tests.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It is undefined behavior.
Fixes "runtime error: applying non-zero offset 8 to null pointer".
Fixes the Snow vsynth FATE-tests.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Fixes "runtime error: addition of unsigned offset to 0x765a09523a90
overflowed to 0x765a09523a8e". This fixes all H.261 tests when run
under UBsan.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
codec20 is raw uncompressed image data.
It exists in Rebel Assault 1 as a special format for STOR/FTCH
and is used again in the Full Throttle Remaster from 2017.
Signed-off-by: Manuel Lauss <manuel.lauss@gmail.com>
Codec45 is used in some RA2 videos on top of codec37.
It consists of 2 tables (768 and 32768 bytes), and datapackets
like codec2 (delta-x, delta-y) with a pixel counter instead of a color value.
It then reads the 4 surrounding pixels, looks up 3 separate values
for each in table1, adds them together to form an index into
table2 for a new pixel value, in a row.
The data coming in gets the x/y coordinates out of the visible
area very quickly (2-3 iterations); I don't see any visual
difference between the ffmpeg-produced frames and ones captured
with dosbox from rebel assault 2, which leads me to believe this
codec never worked as intended.
Signed-off-by: Manuel Lauss <manuel.lauss@gmail.com>
Handle STOR/FTCH the same way the RA1 game engine does:
On STOR, save the next following FOBJ (not the decoded image)
in a buffer; decode it on FTCH.
The RA1 codecs and the fobj handler now take an explicit
GetByteContext in order to be able to replay stored data.
Used extensively by Rebel Assault 1 for e.g. backgrounds and
the cockpit overlay.
For ANIMv2 keep the current system to store the decoded image, as
replaying a FOBJ would not work with codecs37/47/48 due to sequence
violations.
Signed-off-by: Manuel Lauss <manuel.lauss@gmail.com>
Compression 4 code 0 means copy from delta buffer without mv,
AND start of a skip run. This gets rid of the extra case and column
index manipulation and implements this as it is implemented in the
original game exe, i.e. as a special case for after mv copy.
Signed-off-by: Manuel Lauss <manuel.lauss@gmail.com>
This codec works on 4x4 pixel tiles, which can be generated and read
from the datastream. They're both identical, codec5 misses a tile
index to skip the current tile.
Codecs33/34 are the same as 4/5 but with a different tileset generator.
Used only in Rebel Assault 1.
Signed-off-by: Manuel Lauss <manuel.lauss@gmail.com>
This codec alternatingly skips and changes existing pixels.
A second 16bit parameter in the FOBJ header indicates how to do
the pixel changes: either by specifying a LUT in the codec datastream
or by adding a constant value to the pixel.
Signed-off-by: Manuel Lauss <manuel.lauss@gmail.com>
this codec consists of 4 byte packets: 2bytes delta-x, 1 byte delta-y
and 1 byte color to put at that spot.
Used in Rebel Assault 1 only.
Signed-off-by: Manuel Lauss <manuel.lauss@gmail.com>
codec3 is codec1 which writes zero values instead of skipping them.
This fixes a lot of RA1 videos.
Signed-off-by: Manuel Lauss <manuel.lauss@gmail.com>
- clear the front buffer with color 0 on the first FOBJ.
Fixes a lot of Rebel Assault 1 videos and Rebel Assault 2 space
scenes (e.g. 08PLAY.SAN which consists only of codec1/2/21 objects
which only ever touch parts of the buffer).
- for ANIMv1 (Rebel Assault 1): set palette index 0 to all zeroes.
This fixes a lot of stray colors in e.g L1HANGAR.ANM, L2INTRO.ANM,
space scenes.
- Esp in RA1, there are a lot of FRME objects which don't contain
any video data (prebuffering some audio only). Account for that.
- In RA1 L2PLAY.ANM there are a few unaligned FOBJs, handle this
in a generic way.
Signed-off-by: Manuel Lauss <manuel.lauss@gmail.com>
The size of the video frame (FOBJ) of the old codecs (ANIMv0/1/2) can
very reliably be determined:
- ANIMv0/1 (=Rebel Assault 1) uses a 384x242 internal buffer for
everything. The codec parameters only describe the size and offset
of the specific FOBJ on that buffer.
- ANIMv2 titles usually use one of the fullscreen codecs (37/47/48)
as first FOBJ, and their dimensions can generally be trusted.
- RA2 uses 424x260 as internal buffer, use that if encountered:
08PLAY.SAN does not use codec37 so we need to guess using the
codec coordinates.
- ignore sizes smaller than 2x2 or larger than 800x600.
- some game videos have an initial fobj with either 1x1 or -1x-1
pixels in size, ignore them with a warning (Full Throttle
and the Rebel Assault 2 xxRETRY.SAN videos).
Once a known/valid dimension set has been discovered, use it and
don't change it for subsequent FOBJs, rather clamp the large frame
to the determined dimensions.
Tested with RA1, RA2, Full Throttle, Dig, Outlaws, SotE and MotS
videos.
Signed-off-by: Manuel Lauss <manuel.lauss@gmail.com>
