If h2645data.c were changed in a way that makes it rely on another
object file, all components that need h2645data.o would need to be
updated. And given that the VVC decoder->h2645data.o dependency is
currently hidden in a separate Makefile, it would likely be forgotten.
So move it to the libavcodec/Makefile.
Reviewed-by: Nuo Mi <nuomi2021@gmail.com>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This patch adds a fully-featured level 3 and 4 decoder for FFv1,
supporting Golomb and all Range coding variants, all pixel formats,
and all features, except for the newly added floating-point formats.
On a 6000 Ada, for 3840x2160 bgr0 content at 50Mbps (standard desktop
recording), it is able to do 400fps.
An Alder Lake with 24 threads can barely do 100fps.
This commit adds a reference to the buffer as an argument to
start_frame, and adapts all existing code.
This allows for asynchronous hardware accelerators to skip
copying packet data by referencing it.
This commit adds support for hardware accelerated decoding to
the decoder.
The previous commits already refactored the decoder, this commit
simply adds calls to hooks to decode.
Previously, the callback was only called on init. This makes it
get called on every frame.
We should switch to VK_KHR_video_maintenance2 and provide all params
upfront, but almost nothing supports it yet.
Nothing in this decoder would break if the generic code were to be
changed to allow slice "threading" with only one thread.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Fixes undefined integer overflows. The overflows could always
happen, yet before 4d8b706b1d
it was not undefined because the code implicitly used atomic
types, for which signed integer overflow is defined.
Reported-by: Kacper Michajlow <kasper93@gmail.com>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
If there were not enough data, checksum_size would be read
as zero (due to the implicit checks of the bytestream2 API)
and run into a "data block size invalid" error. Erroring out
earlier via "not enough extradata" is better.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It makes no sense to try to be exact with max depth
in get_vlc2(): It will mean that the compiler emits code
for all three stages of parsing and runtime checks for
whether max_depth is big enough to parse the next stage
when a not yet complete code has been encountered.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Up until now, decode_block() zeroed every block (of 128 bytes)
before decoding a block; yet this is suboptimal for all modes,
because all modes need to reset all the blocks they use anyway
and so it should be done in one go for all blocks.
For the alpha modes (where blocks need not be coded) all blocks
are zeroed initially anyway, because decode_block() might not
be doing it, so zeroing there again for the coded blocks is
a waste.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It allows to share them between frame threads.
dc11 can unfortunately not be made static without increasing
LOCALBUF_ELEMS in vlc.c.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This is trivial as well as beneficial because frame threads
now use the same table, saving cache/memory.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This avoids having to expose HQXContext in a header
and allows to make several symbols static.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
hqxvlc.c contains sort-of run-length VLCs in hardcoded form;
they amount to 26688 elements, taking 104KiB. These tables contain
many duplicated entries (they are partially created via a RPT_1024
macro). There are actually only 3039 different codes in all tables
combined, making this very wasteful.
This commit changes this by extracting the underlying entries
and creating a (static) RL-VLC. This only costs 3*3039 bytes
of .rodata. The resulting table needs only 15630 entries,
because our VLC init code uses smaller subtables when possible
(for an incomplete code, the negative of the length stored in
the VLC code is the number of bits the subtable uses; the hardcoded
tables uses a worst-case per table value).
Using GET_RL_VLC also gets rid of an unnecessary reload in case
a code is too long to be parsed in the first stage.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It will simplify creating RL-VLCs (which until now used
a stack-based VLC as temporary buffer).
Notice that there would be another option to merge them, namely:
struct VLCElem {
union {
VLCBaseType sym;
int16_t level;
};
int8_t len;
uint8_t run;
};
The main difference to the current patch is that this would change
the type of the length field of VLCElem to int8_t from int16_t.
It turns out that that this would entail additional sign extensions
on ppc, see https://godbolt.org/z/behWYEYE7. I have therefore refrained
from doing so, although it would make the patch simpler.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
These fields are meant to be for encoders; just use a local
table in pcm_decode_init() to convert from AVCodecID to AVSampleFormat
and also use it to provide sample_size.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Instead use a dedicated context for the few decoders that
use the LUT. Also use a dedicated init function for them.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
