RealVideo 4.0 has a VLC that encodes two intra types per code; each
intra type is in the range 0..8 (inclusive) and up until now the VLC
used symbols in the range 0..80; one type was encoded as the remainder
when dividing the symbol by 9 whereas the other type was encoded as
symbol / 9. This is suboptimal; a better way would be to use the high
and low nibble to encode each symbol. But an even better way is to use
16bit symbols so that the two intra types can be directly written as
a 16bit value.
This commit implements this; in order to avoid huge tables the symbols
are stored as uint8_t with high and low nibbles encoding one type each;
they are only unpacked to uint16_t during initialization.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
After permuting the codes, symbols and lengths tables used to initialize
the VLC so that the codes are ordered from left to right in the Huffman
tree, the codes become redundant as they can be easily computed from the
lengths at runtime; in this case one has to use explicit symbol tables,
but all the symbols used here fit into an uint8_t, whereas some codes
needed uint16_t. This saves about 1.6KB.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
After permuting the codes, symbols and lengths tables used to initialize
the VLCs so that the codes are ordered from left to right in the Huffman
tree, the codes become redundant as they can be easily computed from the
lengths at runtime (or at compile time with --enable-hardcoded-tables);
in this case one has to use explicit symbol tables, but all the symbols
used here fit into an uint8_t, whereas some codes needed uint16_t.
Furthermore, the codes had holes because the range of the symbols was not
contiguous; these have also been removed.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
If both codes, lengths and symbols tables are ordered so that the codes
are sorted from left to right in the tree, the codes can be easily
derived from the lengths and therefore become redundant. This is
exploited in this commit to remove the codes tables for the mobiclip
decoder; notice that tables for the run-length VLC were already ordered
correctly.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Up until now, VLCs that were part of an array of VLCs were often not
initialized in a loop, but separately. The probable reason for this
was that these VLCs differed slightly in the parameters to be used for
them (i.e. the number of codes or the number of bits to be used
differs), so that one would have to provide these parameters e.g. via
arrays.
Yet these problems have actually largely been solved by now: The length
information is contained in a run-length encoded form that is the same
for all VLCs and both the number of codes as well as the number of bits
to use for each VLC can be easily derived from them.
There is just one problem to be solved: When the underlying tables have
a different number of elements, putting them into an array of arrays
would be wasteful; using an array of pointers to the arrays would
also be wasteful. Therefore this commit combines the tables into bigger
tables. (Given that all the length tables have the same layout this
applies only to the symbols tables.)
Finally, the array containing the offset of the VLC's buffer in the big
buffer has also been removed.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Several of the quantisation VLCs come in pairs and up until now the
number of bits used for each VLC was set to the same value for both VLCs
in such a pair even when one of the two required only a lower number.
This is a waste given that the get_vlc2() call is compatible with these
two VLCs using a different number of bits (it uses vlc->bits).
Given that the code lengths are descending it is easily possible to know
the length of the longest code for a given VLC: It is the length of the
first one. With this information one can easily use the least amount of
bits.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
After permuting both length, code as well as symbol tables so that
the codes are ordered from left to right in the tree, it became apparent
that the length of the codes decreases from left to right. Therefore one
can run-length encode the lengths to save space. This commit implements
this.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
By switching to ff_init_vlc_from_lengths() one can make a table of
codes of type uint8_t superfluous, saving space.
Other VLCs (those without dedicated symbols table and with codes of
type uint8_t) have been made to use ff_init_vlc_from_lengths(), too,
because it reduces codesize (ff_init_vlc_from_lengths() has two
parameters less than ff_init_vlc_sparse()) and because it allows to
use the offset parameter in future commits.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
By switching to ff_init_vlc_from_lengths() one can replace tables of
codes of type uint16_t with tables of symbols of type uint8_t, saving
space.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
By using ff_init_vlc_from_lengths(), we do not have to keep track of the
codes themselves, but can offload this to ff_init_vlc_from_lengths().
Furthermore, the old code presumed sizeof(int) == 4; this is no longer
so.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
By switching to ff_init_vlc_from_lengths() one can replace a table of
codes of type uint32_t with a table of symbols of type uint8_t saving
space. The old tables also had holes in it (because of the symbols) which
are now superfluous, saving ever more space.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
The VLC tables to be used for parsing RealVideo 1.0 DC coefficients are
weird: The luma table contains a block of 2^11 codes beginning with the
same prefix and length that all have the same symbol (i.e. value only
depends upon the prefix); the same goes for the chroma block (except
it's only 2^9 codes). Up until now, these entries (which generally could
be parsed like ordinary entries with subtables) have been treated
specially: They have been treated like open ends of the tree, so that
get_vlc2() returned a value < 0 upon encountering them; afterwards it
was checked whether the right prefix was used and if so, the appropriate
number of bytes was skipped.
But there is actually an easy albeit slightly hacky way to support them
directly without pointless subtables: Just modify the VLC table so that
all the entries sharing the right prefix have a length that equals the
length of the whole entry.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
These tables were huge (14 bits) because one needed 14 bits in order to
find out whether a code is valid and in the VLC table or a valid code that
required hacky workarounds due to RealVideo 1.0 using multiple codes
for the same symbol and the code predating the introduction of symbols
tables for VLCs.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
The RealVideo 1.0 decoder uses VLCs to parse DC coefficients. But the
values returned from get_vlc2() are not directly used; instead
-(val - 128) (which is in the range -127..128) is. This transformation
is unnecessary as it can effectively be done when initializing the VLC
by modifying the symbols table used. There is just one minor
complication: The chroma table is incomplete and in order to distinguish
an error from get_vlc2() (due to an invalid code) the ordinary return
range is modified to 0..255. This is possible because the only caller of
this function is (on success) only interested in the return value modulo
256.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
RealVideo 1.0 uses an insane way to encode DC coefficients: There are
several symbols that (for no good reason whatsoever) have multiple
encodings, leading to longer codes than necessary.
More specifically, the tree for the 256 luma symbols contains 255 codes
belonging to 255 different symbols on the left; going further right,
the tree consists of two blocks of 128 codes each of length 14 encoding
consecutive numbers (including two encodings for the symbol missing among
the 255 codes on the left); this is followed by two blocks of codes of
length 16 each containing 256 elements with consecutive symbols (i.e.
each of the blocks allows to encode all symbols). The rest of the tree
consists of 2^11 codes that all encode the same symbol.
The tree for the 256 chroma symbols is similar, but is missing the
blocks of length 256 and there are only 2^9 consecutive codes that
encode the same symbol; furthermore, the chroma tree is incomplete:
The right-most node has no right child.
All of this caused problems when parsing these codes; the reason is that
the code for this predates commit b613bacca9
which added support for explicit symbol tables and thereby removed the
requirement that different codes have different symbols. In order to
address this, the trees used for parsing were incomplete: They contained
the 255 codes on the left and one code for the remaining symbol. Whenever
a code not in these trees was encountered, it was dealt with in
special cases (one for each of the blocks mentioned above).
This commit reduces the number of special cases: Using a symbols table
allows to treat the blocks of consecutive symbols like ordinary codes;
only the blocks encoding a single symbol are still treated specially
(in order not to waste memory on tables for them).
In order to not increment the size of the tables used to initialize the
VLCs both the symbols as well as the lengths are now run-length encoded.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
This can be achieved by switching to ff_init_vlc_from_lengths() which
allows to replace two uint16_t tables for codes with uint8_t tables for
the symbols by permuting the tables so that the codes are ordered from
left to right in the tree in which case they can be easily computed from
the lengths at runtime.
And after doing so, it became apparent that the tables for the symbols
are actually the same for luma and chroma, so that one can even omit one
of them.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Permuting the tables used to initialize the Cook VLCs so that the code
tables are ordered from left to right in the tree revealed that the
length of the codes are ascending from left to right. Therefore one can
run-length encode them to avoid the big length tables; this saves a bit
more than 1KB.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Up until now, the Cook decoder used tables for the lengths of codes and
tables of the codes itself to initialize VLCs; the tables for the codes
were of type uint16_t because the codes were so long. It did not use
explicit symbol tables. This commit instead reorders the tables so that
the code tables are sorted from left to right in the tree. Then the
codes can be easily derived from the lengths and therefore be omitted.
This comes at the price of explicitly coding the symbols, but this is
nevertheless a net win because most of the symbols tables can be coded
on one byte. Furthermore, Cook actually does not use a contiguous range
of symbols for its main VLC tables and the old code compensated for that
by adding holes (codes of length zero) to the tables (that are skipped by
ff_init_vlc_sparse()). This is no longer necessary with the new
approach. All in all, this saves about 1.7KB.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
This is possible by switching to ff_init_vlc_from_lengths() which allows
to replace the table for the codes (which need an uint16_t) by a table
of symbols which fit into an uint8_t. Also switch to an ordinary
INIT_VLC macro while just at it.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Both the motion vector as well as the bias VLCs have an escape code;
for the motion vectors, this value depended on the specific VLC table,
whereas all the bias VLCs used the same value; the escape value has not
been inlined in the latter case.
But for both kinds of VLCs there are lots of values that are unused for
all the VLCs of each kind and each of these can be used as common escape
value, thus allowing to inline the escape value. This commit implements
this.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
After the motion vector and bias values tables have been reordered so
that the codes are ordered from left to right, it emerged that the
length of these entries are actually ascending for every table.
Therefore it is possible to encode them in a run-length style and create
the actual length tables during runtime. This commit implements this.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
The ClearVideo decoder uses VLC tables that are initialized at runtime
from static length, symbol and codes tables. Yet the code tables can be
omitted by subjecting all of these tables to the permutation that orders
the codes from left to right in the tree. After this is done, the codes
can be easily computed at runtime from the lengths and therefore
omitted. This saves about 10KB.
Only one minor complication is encountered when doing so: The tree
corresponding to the AC VLC codes is incomplete; but this can be
handled by adding an entry with negative length.
Furthermore, there are also VLCs that are only initialized with lengths
and codes tables with codes of type uint16_t. These have also been
switched to ff_init_vlc_from_lengths() as this means that one can
replace the uint16_t codes tables with uint8_t symbols tables.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
The IMC decoder uses Huffman tables which are created at runtime from
length tables of type uint8_t and code tables of type uint16_t together
with an implicit symbols table (i.e. symbol[i] == i). This commit
changes this: All three tables are subjected to the same permutation to
order the codes from left to right in the tree; afterwards the codes can
be omitted because they are easily computable at runtime from the
lengths, whereas the symbols need to be explicitly coded now. But said
symbols fit into an uint8_t, so one saves space.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Using one big table for the codebook symbols and lengths makes it
possible to remove the pointers to the individual tables.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
The On2 audio decoder uses huge tables to initialize VLC tables. These
tables (mostly) use symbols tables in addition to the codes tables and
the lengths tables. This commit makes the codes tables redundant and
removes them: If all tables are permuted so that the codes are ordered
from left to right in the Huffman tree, the codes become redundant and
can be easily calculated at runtime from the lengths
(via ff_init_vlc_from_lengths()); this also avoids sorting the codes in
ff_init_vlc_sparse()*.
The symbols tables are always 16bit, the codes tables are 32bit, 16bit
or (rarely) 8bit, the lengths tables are always 8bit. Even though some
symbols tables have been used twice (which is no longer possible now
because different permutations need to be performed on the code tables
sharing the same symbol table in order to order them from left to right),
this nevertheless saves about 28KB.
*: If the initializations of the VLCs are repeated 2048 times
(interleaved with calls to free the VLCs which have not been timed), the
number of decicycles spent on each round of initializations improves
from 27669656 to 7356159.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Andreas Rheinhardt