13 months agoFix typo in a comment in opus_decoder.c
Matt Brubeck [Thu, 24 Aug 2017 19:03:11 +0000 (12:03 -0700)]
Fix typo in a comment in opus_decoder.c

Signed-off-by: Ralph Giles <giles@thaumas.net>
13 months agofix alignment exceptions
Ray Essick [Thu, 24 Aug 2017 12:32:42 +0000 (05:32 -0700)]
fix alignment exceptions

some SSE optimizations were using an instruction sequence that required
128-bit alignment, even though this is not always guaranteed. The error
was in handling a 64-bit entity; made the same macro changes that had
been done previously for 32-bit entities.

Signed-off-by: Felicia Lim <flim@google.com>
13 months agoMore fixes for IESG
Jean-Marc Valin [Thu, 17 Aug 2017 23:02:59 +0000 (19:02 -0400)]
More fixes for IESG

13 months agoupdate draft: addressing IETF LC comments
Jean-Marc Valin [Thu, 17 Aug 2017 18:23:51 +0000 (14:23 -0400)]
update draft: addressing IETF LC comments

14 months agoForce crlf line endings for VS2015 project files.
Pavel P [Thu, 10 Aug 2017 08:22:33 +0000 (01:22 -0700)]
Force crlf line endings for VS2015 project files.

Signed-off-by: Ricardo Constantino <wiiaboo@gmail.com>
14 months agoUpdate draft: addressing AD comments
Jean-Marc Valin [Wed, 26 Jul 2017 03:09:14 +0000 (23:09 -0400)]
Update draft: addressing AD comments

14 months agoFix celt_pitch_xcorr() ARM jump table compiling error
Linfeng Zhang [Thu, 20 Jul 2017 20:00:07 +0000 (13:00 -0700)]
Fix celt_pitch_xcorr() ARM jump table compiling error

This bug is introduced in commit 85fb8a1.

Change-Id: I3822cdfc1308ca62ce03da008260de8f4b04596c

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
14 months agoUpdate links, fix bits, add security considerations
Jean-Marc Valin [Wed, 12 Jul 2017 20:16:45 +0000 (16:16 -0400)]
Update links, fix bits, add security considerations

15 months agoAdd missing #include for gcc
Felicia Lim [Thu, 13 Jul 2017 17:33:57 +0000 (10:33 -0700)]
Add missing #include for gcc

15 months agoAvoid -Wsometimes-uninitialized error for valid test code
Felicia Lim [Wed, 12 Jul 2017 00:22:32 +0000 (17:22 -0700)]
Avoid -Wsometimes-uninitialized error for valid test code

|frame_size_enum| in tests/test_opus_encode.cl:117 is flagged as potentially
uninitialized but get_frame_size_enum() will fail anyway if a valid value is
not found.

15 months agoUpdate testvector download command.
Ralph Giles [Wed, 12 Jul 2017 15:45:33 +0000 (08:45 -0700)]
Update testvector download command.

Migrating opus-codec.org to a static site generator inserted
a redirect into the download path, which the actual testvectors
at https://opus-codec.org/static/testvectors/opus_testvectors.tar.gz.

Add the -L option to the suggested curl invocation so the redirect
is followed to download the actual resource. Without this one
gets the html redirection fallback page.

15 months agoFix uninitialized variables in decide_dtx_mode()
Andrew Larkin [Thu, 6 Jul 2017 18:00:36 +0000 (11:00 -0700)]
Fix uninitialized variables in decide_dtx_mode()

Signed-off-by: Felicia Lim <flim@google.com>
15 months agoAdd checksum for opus 1.2.1 release
Tristan Matthews [Wed, 28 Jun 2017 20:29:02 +0000 (16:29 -0400)]
Add checksum for opus 1.2.1 release

15 months agobump LT revision v1.2.1
Jean-Marc Valin [Mon, 26 Jun 2017 18:04:09 +0000 (14:04 -0400)]
bump LT revision

15 months agoWorking around misdetected audio bandwidth
Jean-Marc Valin [Mon, 26 Jun 2017 17:02:23 +0000 (13:02 -0400)]
Working around misdetected audio bandwidth

The new detector for SWB vs FB is unreliable due to the hack
that was required to get 24 kHz analysis working. We're now err
on the side of FB just to make sure.

15 months agoPrevent update draft from expiring
Jean-Marc Valin [Tue, 20 Jun 2017 17:07:43 +0000 (13:07 -0400)]
Prevent update draft from expiring

15 months agoAdd checksum for opus 1.2 release.
Ralph Giles [Tue, 20 Jun 2017 20:18:32 +0000 (13:18 -0700)]
Add checksum for opus 1.2 release.

15 months agoAdd "f" suffix to float constants v1.2
Jean-Marc Valin [Mon, 19 Jun 2017 21:55:56 +0000 (17:55 -0400)]
Add "f" suffix to float constants

15 months agobump LT version numbers
Jean-Marc Valin [Mon, 19 Jun 2017 20:30:38 +0000 (16:30 -0400)]
bump LT version numbers

15 months agoUpdate release checklist for website move to git.
Ralph Giles [Fri, 16 Jun 2017 16:17:15 +0000 (09:17 -0700)]
Update release checklist for website move to git.

16 months agoFix spacing in celt_sources.mk.
Ralph Giles [Tue, 13 Jun 2017 21:13:52 +0000 (14:13 -0700)]
Fix spacing in celt_sources.mk.

Uniform layout is easier for external tools to parse,
especially maintaining one source file per line.

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
16 months agotests/run_vectors.sh: Fix argument quoting
Mark Harris [Wed, 14 Jun 2017 04:35:28 +0000 (21:35 -0700)]
tests/run_vectors.sh: Fix argument quoting

Allows spaces in path arguments

16 months agoFree all allocated memory in unit tests
Mark Harris [Wed, 14 Jun 2017 04:22:51 +0000 (21:22 -0700)]
Free all allocated memory in unit tests

16 months agoFix a duplicate symbol name.
Ralph Giles [Tue, 13 Jun 2017 17:23:06 +0000 (10:23 -0700)]
Fix a duplicate symbol name.

Both versions of silk/tables_NLSF_CB had symbols
named silk_NLSF_CB1_Wght_Q9 which causes problems
if both files are build as part of the same
compilation unit. Rename the wideband version
to include a 'WB' string to avoid the conflict.

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
16 months agoImproving tonality estimation with delayed decision
Jean-Marc Valin [Sat, 10 Jun 2017 02:16:53 +0000 (22:16 -0400)]
Improving tonality estimation with delayed decision

16 months agoUpdate releases.sha2
Jean-Marc Valin [Fri, 9 Jun 2017 17:42:55 +0000 (13:42 -0400)]
Update releases.sha2

16 months agoClean celt_pitch_xcorr_float_neon() v1.2-rc1
Linfeng Zhang [Thu, 8 Sep 2016 00:14:05 +0000 (17:14 -0700)]
Clean celt_pitch_xcorr_float_neon()

Call celt_inner_prod_neon() and remove redundant code.

Change-Id: I980e94330ae75c10297b9035fac221515aee144f

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
16 months agoOptimize floating-point celt_inner_prod() and dual_inner_prod() for ARM NEON
Linfeng Zhang [Wed, 25 Jan 2017 17:45:48 +0000 (09:45 -0800)]
Optimize floating-point celt_inner_prod() and dual_inner_prod() for ARM NEON

The floating-point optimizations are not bit exact with C functions,
because of the different orders of floating-point operations.
But they are bit exact with the simulation C functions which simulate
the floating operations in the optimizations.

Change-Id: I149fda5b602fd5712b16fc8983df3c6c0c9e76ad

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
16 months agoOptimize fixed-point celt_inner_prod() and dual_inner_prod() for ARM NEON
Linfeng Zhang [Wed, 7 Sep 2016 23:43:54 +0000 (16:43 -0700)]
Optimize fixed-point celt_inner_prod() and dual_inner_prod() for ARM NEON

This optimization is bit exact with C functions.

Change-Id: Ia9ce6dd3c20d2f56dbd43ddc02d1a6fd6554608d

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
16 months agoReplace call of celt_inner_prod_c() (step 2)
Linfeng Zhang [Wed, 7 Sep 2016 22:44:23 +0000 (15:44 -0700)]
Replace call of celt_inner_prod_c() (step 2)

Should call celt_inner_prod().
This requires the API change of celt_pitch_xcorr() by passing in

We tested on x86 and arm, and got bit exact results as original.

Change-Id: I606915da6a196f327ce81f4a5ae32811f4c1fabb

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
16 months agoReplace call of celt_inner_prod_c() (step 1)
Linfeng Zhang [Wed, 7 Sep 2016 22:29:03 +0000 (15:29 -0700)]
Replace call of celt_inner_prod_c() (step 1)

Should call celt_inner_prod().

This change is bit exact as original, except for x86 floating-point.
In x86 floating-point, it calls celt_inner_prod_sse() which may have
different output with the change of floating-point operations' orders.

Change-Id: Ia2381e2e198a84296ac28305183f15be842b3454

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
16 months agoInitialize speech/music prob based on application
Jean-Marc Valin [Sun, 4 Jun 2017 21:45:06 +0000 (17:45 -0400)]
Initialize speech/music prob based on application

16 months agoProperly compute redundancy_bytes
Jean-Marc Valin [Sat, 3 Jun 2017 03:01:20 +0000 (23:01 -0400)]
Properly compute redundancy_bytes

16 months agoProperly init speech/music confidence values
Jean-Marc Valin [Sun, 4 Jun 2017 06:59:33 +0000 (02:59 -0400)]
Properly init speech/music confidence values

This was sometimes causing the music probability to be stuck at
zero at the beginning of files

16 months agoDon't use MAY_HAVE_NEON in arm_silk_map.c.
Jonathan Lennox [Fri, 2 Jun 2017 19:53:11 +0000 (15:53 -0400)]
Don't use MAY_HAVE_NEON in arm_silk_map.c.

It's unnecessary, and isn't defined correctly on floating-point.

This makes us correctly use Neon functions (in floating-point mode) on
platforms where Neon is detected by RTCD.

Signed-off-by: Timothy B. Terriberry <tterribe@xiph.org>
16 months agoRecalibrate VBR to hit target bitrate on average
Jean-Marc Valin [Fri, 2 Jun 2017 19:16:24 +0000 (15:16 -0400)]
Recalibrate VBR to hit target bitrate on average

16 months agoUpdate dynalloc VBR calibration
Jean-Marc Valin [Fri, 2 Jun 2017 18:01:36 +0000 (14:01 -0400)]
Update dynalloc VBR calibration

16 months agofix comments
Jean-Marc Valin [Thu, 1 Jun 2017 21:56:31 +0000 (17:56 -0400)]
fix comments

16 months agoAvoid using log2() since it's not in C90
Jean-Marc Valin [Thu, 1 Jun 2017 17:21:59 +0000 (13:21 -0400)]
Avoid using log2() since it's not in C90

16 months agoAdding leakage modelling to boost bands exp_leakage7
Jean-Marc Valin [Thu, 1 Jun 2017 07:14:13 +0000 (03:14 -0400)]
Adding leakage modelling to boost bands

We boost bands that either cause leakage or are filled with leakage

16 months agoAnalysis refactoring for fixed-point scaling
Jean-Marc Valin [Thu, 1 Jun 2017 07:13:01 +0000 (03:13 -0400)]
Analysis refactoring for fixed-point scaling

Now scaling the energy to the same value is would be in floating-point

16 months agoLet CBR use more bits for dynalloc
Jean-Marc Valin [Thu, 1 Jun 2017 07:08:58 +0000 (03:08 -0400)]
Let CBR use more bits for dynalloc

It seems like letting CBR use up to 2/3 of the bit is still a win

16 months agosilk_LIMIT_32() should return an opus_int32 (not opus_int)
Jean-Marc Valin [Tue, 30 May 2017 18:26:38 +0000 (14:26 -0400)]
silk_LIMIT_32() should return an opus_int32 (not opus_int)

Thanks to petrufm for pointing that out:

16 months agoFix memory leaks in opus_compare.c
Jean-Marc Valin [Fri, 26 May 2017 20:51:48 +0000 (16:51 -0400)]
Fix memory leaks in opus_compare.c

This makes it possible to run the testvectors with asan enabled.

16 months agoFixes unit tests that need arch-specific code
Jean-Marc Valin [Thu, 25 May 2017 20:28:53 +0000 (16:28 -0400)]
Fixes unit tests that need arch-specific code

We now include the object files for those rather than attempt to

16 months agoappveyor: fix typoed appveyor command
Ricardo Constantino [Thu, 25 May 2017 13:09:30 +0000 (14:09 +0100)]
appveyor: fix typoed appveyor command

And run the build request only at the last configuration.
Also add a check if api_key is unset to not try to request a new build.

Signed-off-by: Ralph Giles <giles@thaumas.net>
16 months agoappveyor: Add the api key for the linked ci builds.
Ralph Giles [Wed, 24 May 2017 21:15:42 +0000 (14:15 -0700)]
appveyor:  Add the api key for the linked ci builds.

The appveyor script needs an encrypted api key to trigger
downstream builds, and there doesn't seem to be an easy
way to make the builds we run for the main repo something
other contributors can control.

Signed-off-by: Ricardo Constantino <wiiaboo@gmail.com>
16 months agoappveyor: request rebuild of opus-tools on successful build
Ricardo Constantino [Wed, 22 Feb 2017 23:39:48 +0000 (23:39 +0000)]
appveyor: request rebuild of opus-tools on successful build

Using my API key for now, but needs rillian's to work.

Signed-off-by: Ralph Giles <giles@thaumas.net>
16 months agoappveyor: simplify and preserve paths
Ricardo Constantino [Fri, 10 Feb 2017 00:58:55 +0000 (00:58 +0000)]
appveyor: simplify and preserve paths

Artifact with preserved paths is more useful for CI.

Signed-off-by: Ralph Giles <giles@thaumas.net>
16 months agoappveyor: move to dotfile like travis
Ricardo Constantino [Fri, 10 Feb 2017 00:53:03 +0000 (00:53 +0000)]
appveyor: move to dotfile like travis

Less clutter on systems that hide dotfiles.

Signed-off-by: Ralph Giles <giles@thaumas.net>
16 months agoUpdated SHA256 of releases
Jean-Marc Valin [Wed, 24 May 2017 19:28:50 +0000 (15:28 -0400)]
Updated SHA256 of releases

16 months agoFixes a double->float conversion warning
Jean-Marc Valin [Wed, 24 May 2017 18:21:08 +0000 (14:21 -0400)]
Fixes a double->float conversion warning

16 months agoFixes some fixed-point 16-bit int overflows v1.2-beta
Jean-Marc Valin [Wed, 24 May 2017 05:21:51 +0000 (01:21 -0400)]
Fixes some fixed-point 16-bit int overflows

The code would have run fine on 32-bit archs, but would have overflowed
on a 16-bit arch

16 months agoReducing trim at low bitrate
Jean-Marc Valin [Wed, 24 May 2017 05:07:20 +0000 (01:07 -0400)]
Reducing trim at low bitrate

Some informal tests seem to confirm that reducing the trim at 32-64 kbps
improves quality (better HF). It's not clear whether it's also the case
at 96 kb/s and above, so we're leaving it as is for those rates.

This corresponds to buildC in this thread:

Also see:

16 months agocosmetics,silk: correct input/output arg comments
Linfeng Zhang [Sat, 21 Jan 2017 21:35:03 +0000 (13:35 -0800)]
cosmetics,silk: correct input/output arg comments

Change-Id: I607a8b75b0711a485384d6f854cf6e2ec18b38b7

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
16 months agoFixes a surround bug on loud signal
Jean-Marc Valin [Tue, 23 May 2017 07:32:18 +0000 (03:32 -0400)]
Fixes a surround bug on loud signal

The constant was copied from the analysis code which uses +/-1 as float
scaling, but surround_analysis() uses +/-32767.

16 months agoOptimize silk_biquad_alt_stride2() for ARM NEON
Linfeng Zhang [Thu, 1 Sep 2016 21:20:57 +0000 (14:20 -0700)]
Optimize silk_biquad_alt_stride2() for ARM NEON

The optimization is bit exact with C function.

Change-Id: Ifb8f04b19f2d576e79ce5dcfa7e0fc374d71d6c8

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
16 months agoUpdate silk_biquad_alt()
Linfeng Zhang [Thu, 1 Sep 2016 20:44:11 +0000 (13:44 -0700)]
Update silk_biquad_alt()

Split to silk_biquad_alt_stride1() and silk_biquad_alt_stride2(),
so that it can be optimized more efficiently when stride is 2.

This change in C code is bit exact with the origin.

Change-Id: Idaefe670397016ace2a489e3435ac61b7dbe79d5

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
17 months agoIndent commands in the README.
Ralph Giles [Wed, 10 May 2017 17:46:22 +0000 (10:46 -0700)]
Indent commands in the README.

Makes things a little easier to read.

Signed-off-by: Gian-Carlo Pascutto <gcp@sjeng.org>
17 months agoSuggest basic build dependencies in the README.
Ralph Giles [Wed, 10 May 2017 17:37:46 +0000 (10:37 -0700)]
Suggest basic build dependencies in the README.

Make it easier for users unfamiliar with C applications
to installed the necessary build dependencies.

Signed-off-by: Gian-Carlo Pascutto <gcp@sjeng.org>
17 months agoAvoid warning when __STDC_VERSION__ is not defined
Carlos Alberto Lopez Perez [Mon, 10 Apr 2017 23:07:35 +0000 (01:07 +0200)]
Avoid warning when __STDC_VERSION__ is not defined

This fixes an issue with gcc 4.9 on Debian 8, at least,
which defines __STDC__ but not __STDC_VERSION__, unlike
more recent gcc.

Signed-off-by: Ralph Giles <giles@thaumas.net>
17 months agooops, fix build broken by previous patch
Jean-Marc Valin [Fri, 28 Apr 2017 21:41:54 +0000 (17:41 -0400)]
oops, fix build broken by previous patch

17 months agoNon-diegetic support for Ambisonics Mapping 254.
Drew Allen [Fri, 28 Apr 2017 16:58:04 +0000 (09:58 -0700)]
Non-diegetic support for Ambisonics Mapping 254.

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
18 months agoOptimize silk_warped_autocorrelation_FIX() for ARM NEON
Linfeng Zhang [Wed, 13 Jul 2016 23:40:48 +0000 (16:40 -0700)]
Optimize silk_warped_autocorrelation_FIX() for ARM NEON

The optimization is bit exact with C function.

This optimization speeds up fixed-point SILK encoder on NEON about
5% to 8%. (Tested on Acer Chromebook, ARMv7 Processor rev 3 (v7l).)

Change-Id: I582f6f3585b7946149e16a2ad3084ebc0ae79a4f

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
18 months agoFix OPUS_ARG_NONNULL indices in opus_multistream.h
Felicia Lim [Mon, 27 Mar 2017 17:36:46 +0000 (10:36 -0700)]
Fix OPUS_ARG_NONNULL indices in opus_multistream.h

19 months agoVS2015: remove previously ignored C4146 in kiss_fft.c
Ricardo Constantino [Thu, 2 Mar 2017 14:48:11 +0000 (14:48 +0000)]
VS2015: remove previously ignored C4146 in kiss_fft.c

C4146: unary minus operator applied to unsigned type, result still unsigned

Fixed in a previous commit.

Signed-off-by: Mark Harris <mark.hsj@gmail.com>
19 months agoVS2015: Ignore LNK4221 and properly force SSE off for Win32
Ricardo Constantino [Mon, 27 Feb 2017 19:04:06 +0000 (19:04 +0000)]
VS2015: Ignore LNK4221 and properly force SSE off for Win32

This warning complains of files that don't define any new symbols
not being included in linking.
GCC doesn't seem to complain about those, so neither should VS.

Signed-off-by: Mark Harris <mark.hsj@gmail.com>
Closes https://github.com/xiph/opus/pull/34

19 months agoVS2015: ignore C4244 warning in opus_compare.c
Ricardo Constantino [Mon, 27 Feb 2017 14:56:06 +0000 (14:56 +0000)]
VS2015: ignore C4244 warning in opus_compare.c

opus_compare.c defines the bitstream so we avoid changing it by
ignoring the non-breaking warning instead of changing the file.

Signed-off-by: Mark Harris <mark.hsj@gmail.com>
Resolves https://github.com/xiph/opus/issues/21

19 months agoFix compiler warnings
Mark Harris [Tue, 21 Feb 2017 03:51:40 +0000 (19:51 -0800)]
Fix compiler warnings

- celt/modes.c:430:14: warning: cast from 'const unsigned char *' to
  'opus_int16 *' increases required alignment from 1 to 2 [-Wcast-align]
- 'C[0][1]' may be used uninitialized [-Wmaybe-uninitialized]
- Unused variable/parameter
- Value stored is never read
- MSVC warnings about "possible loss of data" due to type conversions
- MSVC warning C4146: unary minus operator applied to unsigned type
- silk/NLSF_del_dec_quant.c:137:20: warning: array subscript is above
  array bounds [-Warray-bounds] (gcc -O3 false positive)
- src/mlp_train.h:39:20: warning: function declaration isn't a prototype
- Remove SMALL_FOOTPRINT code from SSE 4.1 FIR implementation, matching
  the C implementation.

The clang -Wcast-align warnings with SSE intrinsics are a known
clang issue: https://llvm.org/bugs/show_bug.cgi?id=20670

19 months agoAdd common.props to Makefile.am
Ricardo Constantino [Tue, 21 Feb 2017 02:24:27 +0000 (02:24 +0000)]
Add common.props to Makefile.am

Signed-off-by: Mark Harris <mark.hsj@gmail.com>
19 months agoVS2015: only compile fixed silk for appropriate configs
Ricardo Constantino [Mon, 20 Feb 2017 22:53:31 +0000 (22:53 +0000)]
VS2015: only compile fixed silk for appropriate configs

Signed-off-by: Mark Harris <mark.hsj@gmail.com>
Closes https://github.com/xiph/opus/pull/26

19 months agoVS2015: Move back PreprocessorDefinitions to opus
Ricardo Constantino [Mon, 20 Feb 2017 22:52:06 +0000 (22:52 +0000)]
VS2015: Move back PreprocessorDefinitions to opus

They're only needed in that project, no need to be in common.props.

Signed-off-by: Mark Harris <mark.hsj@gmail.com>
19 months agoVS2015: join common options in a properties file
Ricardo Constantino [Tue, 10 Jan 2017 00:48:49 +0000 (00:48 +0000)]
VS2015: join common options in a properties file

Should make the projects a bit easier to digest and the differences
between the different configurations and platforms be more obvious.

A (happy?) side effect of this is the static libraries are now about
the same size as the ones produced by GCC and seem to build fine with

Needs testing with other third party applications to make sure nothing
is missing that shouldn't.

Signed-off-by: Mark Harris <mark.hsj@gmail.com>
19 months agoVS2015: remove dead options for linker in static configurations
Ricardo Constantino [Tue, 10 Jan 2017 00:45:10 +0000 (00:45 +0000)]
VS2015: remove dead options for linker in static configurations

Signed-off-by: Mark Harris <mark.hsj@gmail.com>
19 months agotest_opus_decode: Fix build on GCC 3.4 to 4.5.x
Mark Harris [Sun, 19 Feb 2017 04:54:19 +0000 (20:54 -0800)]
test_opus_decode: Fix build on GCC 3.4 to 4.5.x

Broken by b0949f11, reported by ko-zu on GitHub.

19 months agoFix sum initializer in celt_fir()
Jean-Marc Valin [Sat, 18 Feb 2017 22:04:24 +0000 (17:04 -0500)]
Fix sum initializer in celt_fir()

19 months agoFix stability test
Jean-Marc Valin [Sat, 18 Feb 2017 22:04:14 +0000 (17:04 -0500)]
Fix stability test

19 months agoSpeed up test_unit_LPC_inv_pred_gain by returning early on stable filters
Jean-Marc Valin [Sat, 18 Feb 2017 21:43:48 +0000 (16:43 -0500)]
Speed up test_unit_LPC_inv_pred_gain by returning early on stable filters

19 months agoRemove SMALL_FOOTPRINT code for celt_fir()
Jean-Marc Valin [Sat, 18 Feb 2017 00:28:28 +0000 (19:28 -0500)]
Remove SMALL_FOOTPRINT code for celt_fir()

Wasn't worth it given the small code size of the alternative, which also
got refactored a little (still bit exact).

19 months agoIncreasing GLOBAL_STACK_SIZE to 120000 to avoid failure
Jean-Marc Valin [Wed, 15 Feb 2017 22:34:24 +0000 (17:34 -0500)]
Increasing GLOBAL_STACK_SIZE to 120000 to avoid failure

Cannot prove it's the correct value, but it's better than the previous
values, which sometimes segfaults. The increase was made necessary due
to 120 ms frame size support.

19 months agoReduce the scope of Ne10 includes
Michael Bradshaw [Sat, 28 Jan 2017 00:07:26 +0000 (16:07 -0800)]
Reduce the scope of Ne10 includes

libopus only uses the DSP module of Ne10, and never uses the init functions.

Signed-off-by: Michael Bradshaw <mjbshaw@google.com>
Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
19 months agoRemove #include "config.h" from ARM Ne10 headers
Michael Bradshaw [Sat, 28 Jan 2017 01:07:57 +0000 (17:07 -0800)]
Remove #include "config.h" from ARM Ne10 headers

Signed-off-by: Michael Bradshaw <mjbshaw@google.com>
Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
19 months agoOptimize silk_LPC_inverse_pred_gain() for ARM NEON
Linfeng Zhang [Thu, 14 Jul 2016 00:25:49 +0000 (17:25 -0700)]
Optimize silk_LPC_inverse_pred_gain() for ARM NEON

The optimization is bit exact with C function.

Change-Id: Ib3bdc26a5a4ebe02e7f24be85104e8e9a2a9a738

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
19 months agoAdding a unit test for LPC_inv_pred_gain()
Jean-Marc Valin [Wed, 15 Feb 2017 00:06:10 +0000 (19:06 -0500)]
Adding a unit test for LPC_inv_pred_gain()

It checks that no clearly unstable filter passes the LPC_inv_pred_gain()
test. Also, this will make it possible to check assembly for correctness.

Modified from an original patch from Linfeng Zhang <linfengz@google.com>.

19 months agoAdd more coverage to travis tests.
Stephen [Sat, 21 Jan 2017 16:20:56 +0000 (08:20 -0800)]
Add more coverage to travis tests.

Signed-off-by: Ralph Giles <giles@mozilla.com>
19 months agoRename y0/y1 to out0/out1 to avoid name clash with the Bessel functions
Jean-Marc Valin [Tue, 14 Feb 2017 06:32:19 +0000 (01:32 -0500)]
Rename y0/y1 to out0/out1 to avoid name clash with the Bessel functions

20 months agoAdd decoder fuzz target.
Felicia Lim [Fri, 10 Feb 2017 19:42:08 +0000 (11:42 -0800)]
Add decoder fuzz target.

20 months agoRemove silk_LPC_inverse_pred_gain_Q24() which is no longer used anywhere
Jean-Marc Valin [Fri, 10 Feb 2017 04:01:26 +0000 (23:01 -0500)]
Remove silk_LPC_inverse_pred_gain_Q24() which is no longer used anywhere

20 months agoarm2gnu.pl: Fix GNU ARM .type directive
Mark Harris [Sun, 5 Feb 2017 00:57:41 +0000 (16:57 -0800)]
arm2gnu.pl: Fix GNU ARM .type directive

Broken by 76e831d.  Without the .type directive, SIGILL may be produced
if the C code is compiled in Thumb mode, because the compiler may assume
that the asm symbol is also Thumb and call it using a BL instruction.

20 months agofix ARM build w/--disable-intrinsics --enable-asm
James Zern [Sat, 4 Feb 2017 02:55:44 +0000 (18:55 -0800)]
fix ARM build w/--disable-intrinsics --enable-asm

and rtcd disabled (CFLAGS=-mfpu=neon)

broken since:
cfdaf365 Optimize silk_NSQ_del_dec() for ARM NEON

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
20 months agoOptimize silk_NSQ_del_dec() for ARM NEON
Linfeng Zhang [Fri, 26 Aug 2016 01:12:54 +0000 (18:12 -0700)]
Optimize silk_NSQ_del_dec() for ARM NEON

The optimization is bit exact with C function.

This optimization speeds up SILK encoder on NEON as following.

Complexity 0-5:  0%
Complexity 6-7:  6%
Complexity 8-9: 10%
Complexity  10:  8%

Got similar results on floating-point.

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
20 months agoAvoids undefined behaviour in ARM-optimized code
Jean-Marc Valin [Thu, 26 Jan 2017 02:56:48 +0000 (21:56 -0500)]
Avoids undefined behaviour in ARM-optimized code

Casting to unsigned to avoid shifting negative values left.

20 months agoAdd --enable-check-asm
Felicia Lim [Wed, 18 Jan 2017 22:47:26 +0000 (14:47 -0800)]
Add --enable-check-asm

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
20 months agoUpdate NSQ_LPC_BUF_LENGTH macro
Linfeng Zhang [Thu, 11 Aug 2016 20:21:58 +0000 (13:21 -0700)]


Signed-off-by: Felicia Lim <flim@google.com>
20 months agoconfigure: Improve error message v1.2-alpha2
Mark Harris [Wed, 18 Jan 2017 08:13:37 +0000 (00:13 -0800)]
configure: Improve error message

20 months agoAdd FIXED_POINT check if using celt_fir in SILK's LPC analysis filter
Felicia Lim [Wed, 18 Jan 2017 00:23:27 +0000 (16:23 -0800)]
Add FIXED_POINT check if using celt_fir in SILK's LPC analysis filter

Use of celt_fir remains off by default

20 months agoRevise celt_fir_c() to not pass in argument "mem"
Linfeng Zhang [Thu, 16 Jun 2016 23:21:02 +0000 (16:21 -0700)]
Revise celt_fir_c() to not pass in argument "mem"

The "mem" in celt_fir_c() either is contained in the head of input "x"
in reverse order already, or can be easily attached to the head of "x"
before calling the function. Removing argument "mem" can eliminate the
redundant buffer copies inside.
Update celt_fir_sse4_1() accordingly.

21 months agofixed_debug.h: delete some dead code
James Zern [Fri, 13 Jan 2017 04:29:54 +0000 (20:29 -0800)]
fixed_debug.h: delete some dead code

this has the side-effect of removing some C++ style comments

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>
21 months agoFix error message
Chris Rudmin [Tue, 3 Jan 2017 22:03:32 +0000 (17:03 -0500)]
Fix error message

Signed-off-by: Timothy B. Terriberry <tterribe@xiph.org>
21 months agoAvoids pre-echo in hybrid mode caused by noise being injected in the first band
Jean-Marc Valin [Fri, 6 Jan 2017 04:39:44 +0000 (23:39 -0500)]
Avoids pre-echo in hybrid mode caused by noise being injected in the first band

This could happen when we had more than 32 bits on the first hybrid band with
a transient just in the middle of the frame. The band would be split and the
first half of the frame could end up with non-zero energy, but not enough
bits for a pulse. Because it's the first band, no folding would be possible.
This would cause noise to be injected for the entire duration of the first
half and that noise should then get folded to higher bands.