Added info.c to libtheoraenc source and dropped library dependency on libtheoradec.

Addresses issue reported in https://bugs.debian.org/923940 about some symbols being unresolved.

GStreamer's "theoradec" element always calls th_decode_ctl() for all telemetry
options, passing 0 if they are not enabled. This results in our creating a
Cairo buffer and performing YUV/RGB conversion, even if all telemetry is
disabled, which is very slow and can cause noticeable lag during playback.

Jan Niklas Hasse authored 5 years ago and Ralph Giles committed 5 years ago

Signed-off-by: Ralph Giles <giles@thaumas.net>

Sebastian Buchwald authored 6 years ago and Ralph Giles committed 5 years ago

From GCC documentation (for GCC 5 and above):
- ‘+’ means that this operand is both read and written by the instruction.
- Only read-only operands can use ‘%’.

Thus, ‘+’ and ‘%’ cannot be used at the same time.

Signed-off-by: Timothy B. Terriberry <tterribe@xiph.org>
Signed-off-by: Ralph Giles <giles@thaumas.net>

This will bail on files exceeding 16384x16384 (useful for fuzzing).

This prefixes generated object file names with their relative
path. The idea is to avoid conflicts. We don't have any, but
automake warns about the issue, so set the option to address that.

Cameron Cawley authored 7 years ago and Ralph Giles committed 7 years ago

Signed-off-by: Ralph Giles <giles@thaumas.net>

Seems to only effect OS X according to Travis CI

We made sure to emit emms before potentially calling/returning to
 applicate code, but not before calling libc code.
It turns out musl libc's malloc/free implementation use floating
 point, so they don't work correctly unless we do.
Add some extra calls to make sure the FPU is in the proper state
 any place we might potentially call malloc or free.

Thanks to Rune (ztsdztsd on Trac) for the report.

Fixes #2287

If we have some dupe frames right near the end of the buffer, our
 metric window can be slightly larger than the buffer.
However, the clamping we use to determine how many frames' worth of
 metrics to add to the buffer had a mix of signed and unsigned
 types, so if that number when negative, instead of clamping
 against zero, it would ask for all remaining frames in the file.
That would cause an infinite loop when we tried to find the last
 keyframe in the circular buffer in oc_enc_select_qi() (which was
 not big enough to actually hold that many frames).

This patch changes the clamp to a form that works with all unsigned
 values.

Thanks to Brion Vibber for the report.

Fixes #2229

Forward-port of r19507 from svn.

In the case where we got a data packet before receiving all three
 headers, the old code would check to see if the second through
 seventh bytes matched the magic string "theora" (extremely
 unlikely, but possible), and if so, return TH_EBADHEADER, otherwise
 return TH_ENOTFORMAT.
That this was not consistent was a bit non-sensical.

5a5f5bb2 changed to returning TH_EBADHEADER if we got a data
 packet after receiving the first header, but left the old behavior
 for the first packet.
Change instead to explicitly return TH_ENOTFORMAT if the first
 header was missing (since we only check one bit of the packet to
 determine whether or not it's a data packet, odds are it's a packet
 for some other kind of data, like a Vorbis header).
We continue to return TH_EBADHEADER if we see a data packet after
 encountering a valid header, but before reading all three.

Also re-arrange the NULL checks to follow continue to allow the
 undocumented ability to pass in NULL for parameters which are not
 needed by the next header in the sequence.
E.g., it's perfectly all right to pass NULL for _setup when
 expecting to read the comment header next.
In this case we'll now return TH_EBADHEADER instead of TH_EFAULT if
 the packet was actually a data packet.

Our documentation says we will, so we should.

Thanks to Chris Peterson for the report.

Fixes #2279.

oc_mb_fill_cmapping11() was defined without the last two parameters
 of its fellow mapping functions, meaning it got called with extra
 undeclared parameters.

This could confuse the emscripten JavaScript cross-compiler's asm.js
 optimizations, as it's very picky about function signatures.

Fixes #2068.
Patch by Brion Vibber <brion@pobox.com>.

svn path=/trunk/theora/; revision=19261

Patch by valtri.
https://trac.xiph.org/ticket/2032

svn path=/trunk/theora/; revision=19181

svn path=/trunk/theora/; revision=19178

svn path=/trunk/theora/; revision=19086

svn path=/trunk/theora/; revision=19019

I accidentally broke it in r18218.

svn path=/trunk/theora/; revision=18794

Newer versions of png.h don't include it for us.
This is only needed when compiling with OC_DUMP_IMAGES (i.e., for
 debugging).

svn path=/trunk/theora/; revision=18760

There was never any reason for these to not be const, and it
 shouldn't break backwards API or ABI compatibility.

svn path=/trunk/theora/; revision=18348

This was broken when the dequant_tables indexing changed in commit
 r16102, but it only affected post-processing quality, so we never
 noticed.
With gcc 4.8.0, this can now trigger a segfault during decoder
 initialization.

svn path=/trunk/theora/; revision=18268

Save activity from first pass. Gives a small SSIM improvement (e.g. 0.04dB for 360p elephants dream at 750kbit/s)

svn path=/trunk/theora/; revision=18223

Thanks to Holden for the report.

svn path=/trunk/theora/; revision=18221

ref_frame_data was being allocated with the aligned allocator, but
 freed with the normal _ogg_free() function on failure, which
 doesn't work.
This would only cause a problem if there was just enough memory to
 satisfy the reference frame allocation (just over 4.5 or 9 bytes
 per pixel) but not enough for the fragment buffer offets (1/16 or
 1/8th byte per pixel).

svn path=/trunk/theora/; revision=18219

This allows it to be used from dump_video without having to disable
 striped decoding there.

svn path=/trunk/theora/; revision=18218

1) Allow specification of the section alignment and make the
    default alignment 4 bytes (like the RVCT manual says it should
    be).
   This ensures that linking our ARM asm into a Thumb executable
    won't get the section placed on a 2-byte boundary.
2) Add a linker hint to indicate we don't need an executable stack.
   Otherwise it assumes we do.
3) Stop building and linking to an (empty) armopts object file.
   This is needed to make 2) work, since we include this file, so
    we can't add the linker hint at the end of it, and linking in
    just one object without that hint makes the whole library ask
    for an executable stack.

svn path=/trunk/theora/; revision=18031

Fixes #1751.

svn path=/trunk/theora/; revision=17992

Patch by mnemo.
Fixes #1804.

svn path=/trunk/theora/; revision=17989

This extends the fix from r17276 dealing with the removal of the DC
 skip-prevention flag in r17174.
If we allow SKIPing even when the cost is right near the maximum
 representable SSD, then the cost of other blocks can push it over
 the edge, causing us to randomly skip a block halfway through the
 macroblock.

svn path=/trunk/theora/; revision=17959

This gives CIF encoding a 37.5% speed up on a Cortex A8.

svn path=/trunk/theora/; revision=17874

svn path=/trunk/theora/; revision=17825

svn path=/trunk/theora/; revision=17822

This change inserts a new speed level OC_SP_LEVEL_NOSATD (3)
between OC_SP_LEVEL_FAST_ANALYSIS (still 2) and OC_SP_LEVEL_NOMC
(now 4). The new speed level further accelerates analysis by
using the cheaper SAD (sum of absolute differences) metric
for measuring distortion instead of the more expensive SATD
(sum of absolute transformed differences).

The new level is intermediate in both quality and encoding
speed between the adjacent speed levels, allowing smoother
adaptation to realtime encoding contraints. In particular
it's much closer to realtime 720p30 on current Core2 cpus.

The cheaper SAD metric is now also used for the fastest NOMC
speed level. This is a reduction in quality for this speed level
as well, but it also about 10% faster. Since the NOMC mode is
already very poor quality, used as a last resort when performance
is critical, we deemed this a better option than adding a further
NOMC_SATD mode or keeping the old NOMC behaviour.

To support the new speed level, parallel copies of the OC_MODE_RD
table are kept, one for each metric. The encoder context keeps
track of which on to use based ont eh current speed level. Analysis
code chooses which metric to calculate and use on the fly. The extra
conditionals did not measurably affect encoding time in benchmarks.

The metrics collection and fitting code used to generate modedec.h
is also updated to track both SAD and SATD measurements.

Thanks to Tim Terriberry and Greg Maxwell for their comments and
and contributions in reviewing this patch. The implementation
of oc_enc_frag_intra_sad is based on one from earlier development
work by Greg Maxwell.

svn path=/trunk/theora/; revision=17821

This was only used when a stream started without a keyframe, and was
 initializing the wrong region of memory.
This was harmless (the region it initialized was still safely
 allocated, it just didn't cover all of the frame we wanted to use
 as a reference frame), but lead to unpredictable results and
 valgrind errors.
See https://bugzilla.mozilla.org/show_bug.cgi?id=625773

svn path=/trunk/theora/; revision=17780

The specification says that the frame rate numerator and
denominator MUST both be greater than zero. This wasn't
checked as part of the structure validation code, and in
fact passing a zero value here caused a segfault.

The code now checks for this in oc_state_init() and returns
TH_EINVAL, which is causes th_encode_alloc() and oc_decode_alloc()
to return NULL.

A check for that NULL return is also added to encoder_example.c
to show that it may fail when an invalid th_info parameter is
passed.

svn path=/trunk/theora/; revision=17760

svn path=/trunk/theora/; revision=17756

Right now we still immediately compute abs(dc), but we will need the signed
 value to do real DC costing.
This commit should not change the output of the encoder.

svn path=/trunk/theora/; revision=17755

The changes in r17745 made PROC and ENDP directives emit multiple
 lines, leading to mis-matched line numbers between the original RVCT
 source and the GNU translation.
gas allows multiple statements per line, so use them to keep the line
 numbers in sync.
Also, contrary to the commit message in r17745, function names actually
 do show up correctly in valgrind; I was testing the wrong binary.

svn path=/trunk/theora/; revision=17752

Also knock another instruction off the PMOVMSKB emulation.

svn path=/trunk/theora/; revision=17751

The wrong offset was being used for the shift factor in
 oc_enc_enquant_table_fixup_neon(), and the wrong value was being returned from
 oc_enc_quantize_neon() when the entire block was quantized to zero.

svn path=/trunk/theora/; revision=17750