ShuriZma
/
pcsx2
mirror of https://github.com/PCSX2/pcsx2.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

3rdparty: Upgrade soundtouch lib to 2.3.1

This commit is contained in:
Christian Kenny 2021-11-21 21:36:08 -05:00 committed by refractionpcsx2
parent 791f2a63ac
commit e37afd6976
18 changed files with 1141 additions and 1020 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
3rdparty/soundtouch/COPYING.TXT vendored
View File

@ -2,7 +2,7 @@
Version 2.1, February 1999 Version 2.1, February 1999
Copyright (C) 1991, 1999 Free Software Foundation, Inc. Copyright (C) 1991, 1999 Free Software Foundation, Inc.
59 Temple Place, Suite 330, Boston, MA 02111-1307 USA 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Everyone is permitted to copy and distribute verbatim copies Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed. of this license document, but changing it is not allowed.
@ -117,7 +117,7 @@ be combined with the library in order to run.
0. This License Agreement applies to any software library or other 0. This License Agreement applies to any software library or other
program which contains a notice placed by the copyright holder or program which contains a notice placed by the copyright holder or
other authoried party saying it may be distributed under the terms of other authorized party saying it may be distributed under the terms of
this Lesser General Public License (also called "this License"). this Lesser General Public License (also called "this License").
Each licensee is addressed as "you". Each licensee is addressed as "you".

198
3rdparty/soundtouch/README.html vendored
View File

@ -1,21 +1,22 @@
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"> <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html> <html>
<head> <head>
<title>SoundTouch library README</title> <title>SoundTouch library README</title>
<meta http-equiv="Content-Type"
content="text/html; charset=windows-1252">
<meta http-equiv="Content-Language" content="en-us"> <meta http-equiv="Content-Language" content="en-us">
<meta name="author" content="Olli Parviainen"> <meta name="author" content="Olli Parviainen">
<meta name="description" <meta name="description" content="Readme file for SoundTouch audio processing library">
content="Readme file for SoundTouch audio processing library">
<style> <style>
body {font-family: Arial, Helvetica; } body {
font-family: Arial, Helvetica;
}
</style> </style>
</head> </head>
<body class="normal"> <body class="normal">
<hr> <hr>
<h1>SoundTouch audio processing library v2.1.2</h1> <h1>SoundTouch audio processing library v2.3.1</h1>
<p class="normal">SoundTouch library Copyright &copy; Olli Parviainen 2001-2018</p> <p class="normal">SoundTouch library Copyright &copy; Olli Parviainen 2001-2021</p>
<hr> <hr>
<h2>1. Introduction </h2> <h2>1. Introduction </h2>
<p>SoundTouch is an open-source audio processing library that allows <p>SoundTouch is an open-source audio processing library that allows
@ -33,11 +34,12 @@ same time</li>
<h3>1.1 Contact information </h3> <h3>1.1 Contact information </h3>
<p>Author email: oparviai 'at' iki.fi </p> <p>Author email: oparviai 'at' iki.fi </p>
<p>SoundTouch WWW page: <a href="http://soundtouch.surina.net">http://soundtouch.surina.net</a></p> <p>SoundTouch WWW page: <a href="http://soundtouch.surina.net">http://soundtouch.surina.net</a></p>
<p>SoundTouch git repository: <a href="https://gitlab.com/soundtouch/soundtouch.git">https://gitlab.com/soundtouch/soundtouch.git</a></p> <p>SoundTouch git repository: <a
href="https://gitlab.com/soundtouch/soundtouch.git">https://gitlab.com/soundtouch/soundtouch.git</a></p>
<hr> <hr>
<h2>2. Compiling SoundTouch</h2> <h2>2. Compiling SoundTouch</h2>
<p>Before compiling, notice that you can choose the sample data format if it's <p>Before compiling, notice that you can choose the sample data format if it's
desirable to use floating point sample data instead of 16bit integers. See desirable to use 16bit integer sample data instead of floating point samples. See
section &quot;sample data format&quot; for more information.</p> section &quot;sample data format&quot; for more information.</p>
<p>Also notice that SoundTouch can use OpenMP instructions for parallel <p>Also notice that SoundTouch can use OpenMP instructions for parallel
computation to accelerate the runtime processing speed in multi-core systems, computation to accelerate the runtime processing speed in multi-core systems,
@ -70,15 +72,17 @@ folders.</p>
<li>x64 64bit: C:\Program Files (x86)\Microsoft Visual Studio <li>x64 64bit: C:\Program Files (x86)\Microsoft Visual Studio
9.0\VC\redist\amd64\Microsoft.VC90.OPENMP\vcomp90.dll</li> 9.0\VC\redist\amd64\Microsoft.VC90.OPENMP\vcomp90.dll</li>
</ul> </ul>
<p>In Visual Studio 2008, a SP1 version may be required for these libraries. In <p>In other VC++ versions the required library will be expectedly found in similar
other VC++ versions the required library will be expectedly found in similar
&quot;redist&quot; location.</p> &quot;redist&quot; location.</p>
<p>Notice that as minor demonstration of a &quot;dll hell&quot; phenomenon both the 32-bit <p>Notice that as minor demonstration of a &quot;dll hell&quot; phenomenon both the 32-bit
and 64-bit version of vcomp90.dll have the same filename but different contents, and 64-bit version of vcomp90.dll have the same filename but different contents,
thus choose the proper version to allow the program start.</p> thus choose the proper version to allow the program to start.</p>
<h3>2.2. Building in Gnu platforms</h3> <h3>2.2. Building in Gnu platforms</h3>
<p>The SoundTouch library compiles in practically any platform <p>The SoundTouch library compiles in practically any platform
supporting GNU compiler (GCC) tools. SoundTouch requires GCC version 4.3 or later.</p> supporting GNU compiler (GCC) tools.
<h4>2.2.1 Compiling with autotools</h4>
<p>To install build prerequisites for 'autotools' tool chain:</p>
<pre> sudo apt-get install automake autoconf libtool build-essential</pre>
<p>To build and install the binaries, run the following commands in <p>To build and install the binaries, run the following commands in
/soundtouch directory:</p> /soundtouch directory:</p>
<table border="0" cellpadding="0" cellspacing="4"> <table border="0" cellpadding="0" cellspacing="4">
@ -125,41 +129,27 @@ destination locations.</p>
</tr> </tr>
</tbody> </tbody>
</table> </table>
<h4><b>2.2.1 Required GNU tools</b></h4>
<p> <span style="font-weight: bold;">Bash shell</span>, <span
style="font-weight: bold;">GNU C++ compiler</span>, <span
style="font-weight: bold;">libtool</span>, <span
style="font-weight: bold;">autoconf</span> and <span
style="font-weight: bold;">automake</span> tools
are required for compiling the SoundTouch library. These are usually
included with the GNU/Linux distribution, but if not, install these
packages first. For example, Ubuntu Linux can acquire and install
these with the following command:</p>
<pre><b>sudo apt-get install automake autoconf libtool build-essential</b></pre>
<h4><b>2.2.2 Problems with GCC compiler compatibility</b></h4>
<p>At the release time the SoundTouch package has been tested to
compile in GNU/Linux platform. However, If you have problems getting the
SoundTouch library compiled, try disabling optimizations that are specific for
x86 processors by running <b>./configure</b> script with switch
<blockquote>
<pre>--enable-x86-optimizations=no</pre>
</blockquote>
Alternatively, if you don't use GNU Configure system, edit file "include/STTypes.h" <b>Compiling portable Shared Library / DLL version</b>
directly and remove the following definition:<blockquote> <p> The GNU autotools compilation does not automatically create a shared-library version of
<pre>#define SOUNDTOUCH_ALLOW_X86_OPTIMIZATIONS 1</pre> SoundTouch (.so or .dll) that features position-independent code and C-language
</blockquote> api that are more suitable for cross-language development than C++ libraries.</p>
<p> Use script "make-gnu-dll-sh" to build a portable dynamic library version if such is desired.</p>
<h4><b>2.2.3 Compiling Shared Library / DLL version in Cygwin</b></h4> <h4><b>2.2.2 Compiling with cmake</b></h4>
<p> <p>'cmake' build scripts are provided as an alternative to the autotools toolchain.</p>
The GNU compilation does not automatically create a shared-library version of <p>To install cmake build prerequisites:</p>
SoundTouch (.so or .dll). If such is desired, then you can create it as follows <pre> sudo apt-get install libtool build-essential cmake</pre>
after running the usual compilation:</p> <p>To build:</p>
<blockquote> <pre>
<pre>g++ -shared -static -DDLL_EXPORTS -I../../include -o SoundTouch.dll \ cmake .
SoundTouchDLL.cpp ../SoundTouch/.libs/libSoundTouch.a make -j
sstrip SoundTouch.dll</pre> make install</pre>
</blockquote> <p>To compile the additional portable Shared Library / DLL version with the native C-language API:</p>
<pre>
cmake . -DSOUNDTOUCH_DLL=ON
make -j
make install</pre>
<h3>2.3. Building in Android</h3> <h3>2.3. Building in Android</h3>
<p>Android compilation instructions are within the <p>Android compilation instructions are within the
@ -174,18 +164,26 @@ library binary version to use.</p>
example application that processes WAV audio files using SoundTouch library in example application that processes WAV audio files using SoundTouch library in
Android devices.</p> Android devices.</p>
<h3>2.4. Building in Mac</h3>
<p>Install autoconf tool as instructed in <a
href="http://macappstore.org/autoconf/">http://macappstore.org/autoconf/</a>, or alternatively the 'cmake' toolchain.</p>
<p>Then, build as described above in section "Building in Gnu platforms".</p>
<hr> <hr>
<h2>3. About implementation &amp; Usage tips <h3>3.1. Supported sample data formats</h3> <h2>3. About implementation &amp; Usage tips <h3>3.1. Supported sample data formats</h3>
<p>The sample data format can be chosen between 16bit signed integer <p>The sample data format can be chosen between 16bit signed integer
and 32bit floating point values. The default is 32bit floating point format, and 32bit floating point values.</p>
which will also provide slightly better sound quality over the integer format. </p> </p> The default sample type is 32bit floating point format,
which also provides better sound quality than integer format because
integer algorithms need to scale already intermediate calculation results to
avoid integer overflows. These early integer scalings can slightly degrade
output quality.</p>
<p> In Windows environment, the sample data format is chosen in file <p> In Windows environment, the sample data format is chosen in file
"STTypes.h" by choosing one of the following defines:</p> "STTypes.h" by choosing one of the following defines:</p>
<ul> <ul>
<li> <span style="font-weight: bold;">#define <li> <span style="font-weight: bold;">#define
SOUNDTOUCH_INTEGER_SAMPLES</span> for 16bit signed integer</li> SOUNDTOUCH_INTEGER_SAMPLES</span> for 16bit signed integer</li>
<li> <span style="font-weight: bold;">#define </span><span <li> <span style="font-weight: bold;">#define </span><span style="font-weight: bold;">SOUNDTOUCH_</span><span
style="font-weight: bold;">SOUNDTOUCH_</span><span
style="font-weight: bold;">FLOAT_SAMPLES</span> for 32bit floating style="font-weight: bold;">FLOAT_SAMPLES</span> for 32bit floating
point</li> point</li>
</ul> </ul>
@ -301,7 +299,8 @@ ends of the consecutive sequences will overlap with each other.<br>
<br> <br>
This shouldn't be that critical parameter. If you reduce the This shouldn't be that critical parameter. If you reduce the
DEFAULT_SEQUENCE_MS setting by a large amount, you might wish to try a DEFAULT_SEQUENCE_MS setting by a large amount, you might wish to try a
smaller value on this.</li> smaller value on this.
</li>
</ul> </ul>
<p>Notice that these parameters can also be set during execution time <p>Notice that these parameters can also be set during execution time
with functions "<strong>TDStretch::setParameters()</strong>" and "<strong>SoundTouch::setSetting()</strong>".</p> with functions "<strong>TDStretch::setParameters()</strong>" and "<strong>SoundTouch::setSetting()</strong>".</p>
@ -359,6 +358,29 @@ computation burden</td>
</tbody> </tbody>
</table> </table>
<h3>3.5 Performance Optimizations </h3> <h3>3.5 Performance Optimizations </h3>
<p><strong>Integer vs floating point:</strong></p>
<p>Floating point sample type is generally recommended because it provides
better sound quality.</p>
<p>However, execution speed difference between integer and floating point processing
depends on the CPU architecture. As rule of thumb,
<ul>
<li>in 32-bit x86 floating point and integer are roughly equally fast</li>
<li>in 64-bit x86/x64 floating point can be significantly faster than integer
version, because MMX integer optimizations are not available in the x64 architecture.
That depends on the compiler however, so that gcc can autovectorize integer routines
to work equally fast as floating point, where as Visual C++ (2017) does not
perform equally well and produces integer code that runs some 3x slower than
SSE-optimized floating poing code.
</li>
<li>in ARMv7 integer routines are twice as fast as floating point. Their
relative difference is roughly the same both with and without NEON; NEON
vfpu can however bring 2.4x speed improvement.
</li>
<li>in other platforms: try out if the execution time performance makes a
big difference</li>
</ul>
</p>
<p><strong>General optimizations:</strong></p> <p><strong>General optimizations:</strong></p>
<p>The time-stretch routine has a 'quick' mode that substantially <p>The time-stretch routine has a 'quick' mode that substantially
speeds up the algorithm but may slightly compromise the sound quality. speeds up the algorithm but may slightly compromise the sound quality.
@ -373,11 +395,15 @@ function with parameter id of SETTING_USE_QUICKSEEK and value
intrinsics, providing about a 3x processing speedup for x86 compatible intrinsics, providing about a 3x processing speedup for x86 compatible
processors vs. non-SIMD implementation:</p> processors vs. non-SIMD implementation:</p>
<ul> <ul>
<li> Intel MMX optimized routines are used with x86 CPUs when 16bit integer <li> MMX optimized routines are used in 32-bit x86 build when 16bit integer
sample type is used</li> sample type is used</li>
<li> Intel SSE optimized routines are used with x86 CPUs when 32bit floating <li> SSE optimized routines are used in 32- and 64-bit x86 CPUs when 32bit
point sample type is used</li> floating point sample type is used</li>
</ul> </ul>
<p>The algorithms are tuned to utilize autovectorization efficiently
also in other CPU architectures, for example ARM cpus see approx 2.4x processing
speedup when NEON SIMD support is present.
</p>
<h3>3.5 OpenMP parallel computation</h3> <h3>3.5 OpenMP parallel computation</h3>
<p>SoundTouch 1.9 onwards support running the algorithms parallel in several CPU <p>SoundTouch 1.9 onwards support running the algorithms parallel in several CPU
cores. Based on benchmark the experienced multi-core processing speed-up gain cores. Based on benchmark the experienced multi-core processing speed-up gain
@ -385,7 +411,8 @@ ranges between +30% (on a high-spec dual-core x86 Windows PC) to 215% (on a mode
quad-core ARM of Raspberry Pi2). </p> quad-core ARM of Raspberry Pi2). </p>
<p>See an external blog article with more detailed discussion about the <p>See an external blog article with more detailed discussion about the
<a href="http://www.softwarecoven.com/parallel-computing-in-embedded-mobile-devices/"> <a href="http://www.softwarecoven.com/parallel-computing-in-embedded-mobile-devices/">
SoundTouch OpenMP optimization</a>.</p> SoundTouch OpenMP optimization</a>.
</p>
<p>The parallel computing support is implemented using OpenMP spec 3.0 <p>The parallel computing support is implemented using OpenMP spec 3.0
instructions. These instructions are supported by Visual C++ 2008 and later, and instructions. These instructions are supported by Visual C++ 2008 and later, and
GCC v4.2 and later. Compilers that do not supporting OpenMP will ignore these GCC v4.2 and later. Compilers that do not supporting OpenMP will ignore these
@ -409,7 +436,8 @@ library as follows:</p>
</strong>settings. Set </strong>settings. Set
there &quot;<strong>OpenMP support</strong>&quot; to &quot;<strong>Yes</strong>&quot;. Alternatively add there &quot;<strong>OpenMP support</strong>&quot; to &quot;<strong>Yes</strong>&quot;. Alternatively add
<strong>/openmp</strong> switch to command-line <strong>/openmp</strong> switch to command-line
parameters</li> parameters
</li>
<li><strong>GNU</strong>: Run the configure script with &quot;<strong>./configure <li><strong>GNU</strong>: Run the configure script with &quot;<strong>./configure
--enable-openmp</strong>&quot; switch, then run make as usually</li> --enable-openmp</strong>&quot; switch, then run make as usually</li>
<li><strong>Android</strong>: Add &quot;<strong>-fopenmp</strong>&quot; switches to compiler &amp; linker <li><strong>Android</strong>: Add &quot;<strong>-fopenmp</strong>&quot; switches to compiler &amp; linker
@ -575,13 +603,49 @@ this corresponds to lowering the pitch by -0.318 semitones:</p>
<hr> <hr>
<h2>5. Change History</h2> <h2>5. Change History</h2>
<h3>5.1. SoundTouch library Change History </h3> <h3>5.1. SoundTouch library Change History </h3>
<p><b>2.3.1:</b></p>
<ul>
<li>Adjusted cmake build settings and header files that cmake installs</li>
</ul>
<p><b>2.3.0:</b></p>
<ul>
<li>Disable setting "SOUNDTOUCH_ALLOW_NONEXACT_SIMD_OPTIMIZATION" by default. The original
purpose of this setting was to avoid performance penalty due to unaligned SIMD memory
accesses in old CPUs, but that is not any more issue in concurrent CPU SIMD implementations
and having this setting enabled can cause slight compromise in result quality.
</li>
<li>Bugfix: soundtouch.clear() to really clear whole processing pipeline state. Earlier
individual variables were left uncleared, which caused slightly different result if
the same audio stream were processed again after calling clear().
</li>
<li>Bugfix: TDstretch to align initial offset position to be in middle of correlation search
window. This ensures that with zero tempo change the output will be same as input.
</li>
<li>Bugfix: Fix a bug in TDstrectch with too small initial skipFract value that occurred
with certain processing parameter settings: Replace assert with assignment that
corrects the situation.
</li>
<li>Remove OpenMP "_init_threading" workaround from Android build as it's not needed with concurrent
Android SDKs any more.</li>
</ul>
<p><b>2.2:</b></p>
<ul>
<li>Improved source codes so that compiler can autovectorize them more effectively.
This brings remarkable improvement e.g. ARM cpus equipped with NEON vfpu: Bencmarked
2.4x improvement in execution speed in ARMv7l vs the previous SoundTouch version
for both integer and floating point sample types.
</li>
<li>Bugfix: Resolved bad sound quality when using integer sample types in non-x86 CPU</li>
<li>Bugfix: Fixed possible reading past end of array in BPM peak detection algorithm</li>
</ul>
<p><b>2.1.2:</b></p> <p><b>2.1.2:</b></p>
<ul> <ul>
<li>Bump version to 2.1.2 also in configure.ac. The earlier release had old version info for GNU autotools.</li> <li>Bump version to 2.1.2 also in configure.ac. The earlier release had old version info for GNU autotools.</li>
</ul> </ul>
<p><b>2.1.1:</b></p> <p><b>2.1.1:</b></p>
<ul> <ul>
<li>Bugfixes: Fixed potential buffer overwrite bugs in WavFile routines. Replaced asserts with runtime exceptions.</li> <li>Bugfixes: Fixed potential buffer overwrite bugs in WavFile routines. Replaced asserts with runtime exceptions.
</li>
<li>Android: Migrated the SoundTouch Android example to new Android Studio</li> <li>Android: Migrated the SoundTouch Android example to new Android Studio</li>
<li>Automake: unset ACLOCAL in bootstrap script in case earlier build script has set it</li> <li>Automake: unset ACLOCAL in bootstrap script in case earlier build script has set it</li>
@ -602,11 +666,13 @@ this corresponds to lowering the pitch by -0.318 semitones:</p>
</ul> </ul>
<p><b>2.0:</b></p> <p><b>2.0:</b></p>
<ul> <ul>
<li>Added functions to get initial processing latency, duration ratio between the original input and processed output tracks, and clarified reporting of input/output batch sizes</li> <li>Added functions to get initial processing latency, duration ratio between the original input and processed
output tracks, and clarified reporting of input/output batch sizes</li>
<li>Fixed issue that added brief sequence of silence to beginning of output audio</li> <li>Fixed issue that added brief sequence of silence to beginning of output audio</li>
<li>Adjusted algorithm parameters to reduce reverberating effect at tempo slowdown</li> <li>Adjusted algorithm parameters to reduce reverberating effect at tempo slowdown</li>
<li>Bugfix: Fixed a glitch that could cause negative array indexing in quick seek algorithm</li> <li>Bugfix: Fixed a glitch that could cause negative array indexing in quick seek algorithm</li>
<li>Bugfix: flush() didn't properly flush final samples from the pipeline on 2nd time in case that soundtouch object instance was recycled and used for processing a second audio stream.</li> <li>Bugfix: flush() didn't properly flush final samples from the pipeline on 2nd time in case that soundtouch
object instance was recycled and used for processing a second audio stream.</li>
<li>Bugfix: Pi value had incorrect 9th/10th decimals</li> <li>Bugfix: Pi value had incorrect 9th/10th decimals</li>
<li>Added C# example application that uses SoundTouch dll library for processing MP3 files</li> <li>Added C# example application that uses SoundTouch dll library for processing MP3 files</li>
</ul> </ul>
@ -616,8 +682,10 @@ this corresponds to lowering the pitch by -0.318 semitones:</p>
</ul> </ul>
<p><b>1.9.1:</b></p> <p><b>1.9.1:</b></p>
<ul> <ul>
<li>Improved SoundTouch::flush() function so that it returns precisely the desired amount of samples for exact output duration control</li> <li>Improved SoundTouch::flush() function so that it returns precisely the desired amount of samples for exact
<li>Redesigned quickseek algorithm for improved sound quality when using the quickseek mode. The new quickseek algorithm can find 99% as good results as the output duration control</li>
<li>Redesigned quickseek algorithm for improved sound quality when using the quickseek mode. The new quickseek
algorithm can find 99% as good results as the
default full-scan mode, while the quickseek algorithm is remarkable less default full-scan mode, while the quickseek algorithm is remarkable less
CPU intensive.</li> CPU intensive.</li>
<li>Added adaptive integer divider scaling for improved sound quality when using integer processing algorithm <li>Added adaptive integer divider scaling for improved sound quality when using integer processing algorithm
@ -625,7 +693,8 @@ this corresponds to lowering the pitch by -0.318 semitones:</p>
</ul> </ul>
<p><b>1.9:</b></p> <p><b>1.9:</b></p>
<ul> <ul>
<li>Added support for parallel computation support via OpenMP primitives for better performance in multicore systems. <li>Added support for parallel computation support via OpenMP primitives for better performance in multicore
systems.
Benchmarks show that achieved parallel processing speedup improvement Benchmarks show that achieved parallel processing speedup improvement
typically range from +30% (x86 dual-core) to +180% (ARM quad-core). The typically range from +30% (x86 dual-core) to +180% (ARM quad-core). The
OpenMP optimizations are disabled by default, see OpenMP notes above in this OpenMP optimizations are disabled by default, see OpenMP notes above in this
@ -869,8 +938,10 @@ submitted bugfixes:</p>
<li> Jamie Bullock</li> <li> Jamie Bullock</li>
<li> Chris Bryan</li> <li> Chris Bryan</li>
<li> Jacek Caban</li> <li> Jacek Caban</li>
<li> Marketa Calabkova</li>
<li> Brian Cameron</li> <li> Brian Cameron</li>
<li> Jason Champion</li> <li> Jason Champion</li>
<li> Giuseppe Cigala</li>
<li> David Clark</li> <li> David Clark</li>
<li> Patrick Colis</li> <li> Patrick Colis</li>
<li> Miquel Colon</li> <li> Miquel Colon</li>
@ -901,6 +972,7 @@ submitted bugfixes:</p>
<li> Tyson Smith</li> <li> Tyson Smith</li>
<li> John Stumpo</li> <li> John Stumpo</li>
<li> Mario di Vece</li> <li> Mario di Vece</li>
<li> Rémi Verschelde</li>
<li> Katja Vetter</li> <li> Katja Vetter</li>
<li> Wu Q.</li> <li> Wu Q.</li>
</ul> </ul>
@ -922,6 +994,6 @@ Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA</p>
<p>---</p> <p>---</p>
<p>commercial license alternative also available, contact author for details.</p> <p>commercial license alternative also available, contact author for details.</p>
<hr> <hr>
<p><i>README.html file updated in November-2018</i></p>
</body> </body>
</html> </html>

3
3rdparty/soundtouch/soundtouch/FIFOSampleBuffer.h vendored
View File

@ -170,6 +170,9 @@ public:
/// allow trimming (downwards) amount of samples in pipeline. /// allow trimming (downwards) amount of samples in pipeline.
/// Returns adjusted amount of samples /// Returns adjusted amount of samples
uint adjustAmountOfSamples(uint numSamples); uint adjustAmountOfSamples(uint numSamples);
/// Add silence to end of buffer
void addSilent(uint nSamples);
}; };
} }

21
3rdparty/soundtouch/soundtouch/STTypes.h vendored
View File

@ -121,10 +121,10 @@ namespace soundtouch
#endif #endif
// If defined, allows the SIMD-optimized routines to take minor shortcuts // If defined, allows the SIMD-optimized routines to skip unevenly aligned
// for improved performance. Undefine to require faithfully similar SIMD // memory offsets that can cause performance penalty in some SIMD implementations.
// calculations as in normal C implementation. // Causes slight compromise in sound quality.
#define SOUNDTOUCH_ALLOW_NONEXACT_SIMD_OPTIMIZATION 1 // #define SOUNDTOUCH_ALLOW_NONEXACT_SIMD_OPTIMIZATION 1
#ifdef SOUNDTOUCH_INTEGER_SAMPLES #ifdef SOUNDTOUCH_INTEGER_SAMPLES
@ -149,8 +149,9 @@ namespace soundtouch
// floating point samples // floating point samples
typedef float SAMPLETYPE; typedef float SAMPLETYPE;
// data type for sample accumulation: Use double to utilize full precision. // data type for sample accumulation: Use float also here to enable
typedef double LONG_SAMPLETYPE; // efficient autovectorization
typedef float LONG_SAMPLETYPE;
#ifdef SOUNDTOUCH_ALLOW_X86_OPTIMIZATIONS #ifdef SOUNDTOUCH_ALLOW_X86_OPTIMIZATIONS
// Allow SSE optimizations // Allow SSE optimizations
@ -159,7 +160,13 @@ namespace soundtouch
#endif // SOUNDTOUCH_INTEGER_SAMPLES #endif // SOUNDTOUCH_INTEGER_SAMPLES
}; #if ((SOUNDTOUCH_ALLOW_SSE) || (__SSE__) || (SOUNDTOUCH_USE_NEON))
#if SOUNDTOUCH_ALLOW_NONEXACT_SIMD_OPTIMIZATION
#define ST_SIMD_AVOID_UNALIGNED
#endif
#endif
}
// define ST_NO_EXCEPTION_HANDLING switch to disable throwing std exceptions: // define ST_NO_EXCEPTION_HANDLING switch to disable throwing std exceptions:
// #define ST_NO_EXCEPTION_HANDLING 1 // #define ST_NO_EXCEPTION_HANDLING 1

4
3rdparty/soundtouch/soundtouch/SoundTouch.h vendored
View File

@ -72,10 +72,10 @@ namespace soundtouch
{ {
/// Soundtouch library version string /// Soundtouch library version string
#define SOUNDTOUCH_VERSION "2.1.2" #define SOUNDTOUCH_VERSION "2.3.1"
/// SoundTouch library version id /// SoundTouch library version id
#define SOUNDTOUCH_VERSION_ID (20102) #define SOUNDTOUCH_VERSION_ID (20301)
// //
// Available setting IDs for the 'setSetting' & 'get_setting' functions: // Available setting IDs for the 'setSetting' & 'get_setting' functions:

7
3rdparty/soundtouch/source/SoundTouch/BPMDetect.cpp vendored
View File

@ -313,7 +313,7 @@ void BPMDetect::updateXCorr(int process_samples)
#pragma omp parallel for #pragma omp parallel for
for (offs = windowStart; offs < windowLen; offs ++) for (offs = windowStart; offs < windowLen; offs ++)
{ {
double sum; float sum;
int i; int i;
sum = 0; sum = 0;
@ -341,7 +341,6 @@ void BPMDetect::updateBeatPos(int process_samples)
// static double thr = 0.0003; // static double thr = 0.0003;
double posScale = (double)this->decimateBy / (double)this->sampleRate; double posScale = (double)this->decimateBy / (double)this->sampleRate;
int resetDur = (int)(0.12 / posScale + 0.5); int resetDur = (int)(0.12 / posScale + 0.5);
double corrScale = 1.0 / (double)(windowLen - windowStart);
// prescale pbuffer // prescale pbuffer
float tmp[XCORR_UPDATE_SEQUENCE / 2]; float tmp[XCORR_UPDATE_SEQUENCE / 2];
@ -353,7 +352,7 @@ void BPMDetect::updateBeatPos(int process_samples)
#pragma omp parallel for #pragma omp parallel for
for (int offs = windowStart; offs < windowLen; offs++) for (int offs = windowStart; offs < windowLen; offs++)
{ {
double sum = 0; float sum = 0;
for (int i = 0; i < process_samples; i++) for (int i = 0; i < process_samples; i++)
{ {
sum += tmp[i] * pBuffer[offs + i]; sum += tmp[i] * pBuffer[offs + i];
@ -562,7 +561,7 @@ float BPMDetect::getBpm()
/// \return number of beats in the arrays. /// \return number of beats in the arrays.
int BPMDetect::getBeats(float *pos, float *values, int max_num) int BPMDetect::getBeats(float *pos, float *values, int max_num)
{ {
int num = beats.size(); int num = (int)beats.size();
if ((!pos) || (!values)) return num; // pos or values NULL, return just size if ((!pos) || (!values)) return num; // pos or values NULL, return just size
for (int i = 0; (i < num) && (i < max_num); i++) for (int i = 0; (i < num) && (i < max_num); i++)

8
3rdparty/soundtouch/source/SoundTouch/FIFOSampleBuffer.cpp vendored
View File

@ -265,3 +265,11 @@ uint FIFOSampleBuffer::adjustAmountOfSamples(uint numSamples)
} }
return samplesInBuffer; return samplesInBuffer;
} }
/// Add silence to end of buffer
void FIFOSampleBuffer::addSilent(uint nSamples)
{
memset(ptrEnd(nSamples), 0, sizeof(SAMPLETYPE) * nSamples * channels);
samplesInBuffer += nSamples;
}

81
3rdparty/soundtouch/source/SoundTouch/FIRFilter.cpp vendored
View File

@ -60,12 +60,14 @@ FIRFilter::FIRFilter()
length = 0; length = 0;
lengthDiv8 = 0; lengthDiv8 = 0;
filterCoeffs = NULL; filterCoeffs = NULL;
filterCoeffsStereo = NULL;
} }
FIRFilter::~FIRFilter() FIRFilter::~FIRFilter()
{ {
delete[] filterCoeffs; delete[] filterCoeffs;
delete[] filterCoeffsStereo;
} }
@ -78,35 +80,26 @@ uint FIRFilter::evaluateFilterStereo(SAMPLETYPE *dest, const SAMPLETYPE *src, ui
// because division is much slower operation than multiplying. // because division is much slower operation than multiplying.
double dScaler = 1.0 / (double)resultDivider; double dScaler = 1.0 / (double)resultDivider;
#endif #endif
// hint compiler autovectorization that loop length is divisible by 8
int ilength = length & -8;
assert(length != 0); assert((length != 0) && (length == ilength) && (src != NULL) && (dest != NULL) && (filterCoeffs != NULL));
assert(src != NULL);
assert(dest != NULL);
assert(filterCoeffs != NULL);
end = 2 * (numSamples - length); end = 2 * (numSamples - ilength);
#pragma omp parallel for #pragma omp parallel for
for (j = 0; j < end; j += 2) for (j = 0; j < end; j += 2)
{ {
const SAMPLETYPE *ptr; const SAMPLETYPE *ptr;
LONG_SAMPLETYPE suml, sumr; LONG_SAMPLETYPE suml, sumr;
uint i;
suml = sumr = 0; suml = sumr = 0;
ptr = src + j; ptr = src + j;
for (i = 0; i < length; i += 4) for (int i = 0; i < ilength; i ++)
{ {
// loop is unrolled by factor of 4 here for efficiency suml += ptr[2 * i] * filterCoeffsStereo[2 * i];
suml += ptr[2 * i + 0] * filterCoeffs[i + 0] + sumr += ptr[2 * i + 1] * filterCoeffsStereo[2 * i + 1];
ptr[2 * i + 2] * filterCoeffs[i + 1] +
ptr[2 * i + 4] * filterCoeffs[i + 2] +
ptr[2 * i + 6] * filterCoeffs[i + 3];
sumr += ptr[2 * i + 1] * filterCoeffs[i + 0] +
ptr[2 * i + 3] * filterCoeffs[i + 1] +
ptr[2 * i + 5] * filterCoeffs[i + 2] +
ptr[2 * i + 7] * filterCoeffs[i + 3];
} }
#ifdef SOUNDTOUCH_INTEGER_SAMPLES #ifdef SOUNDTOUCH_INTEGER_SAMPLES
@ -116,14 +109,11 @@ uint FIRFilter::evaluateFilterStereo(SAMPLETYPE *dest, const SAMPLETYPE *src, ui
suml = (suml < -32768) ? -32768 : (suml > 32767) ? 32767 : suml; suml = (suml < -32768) ? -32768 : (suml > 32767) ? 32767 : suml;
// saturate to 16 bit integer limits // saturate to 16 bit integer limits
sumr = (sumr < -32768) ? -32768 : (sumr > 32767) ? 32767 : sumr; sumr = (sumr < -32768) ? -32768 : (sumr > 32767) ? 32767 : sumr;
#else
suml *= dScaler;
sumr *= dScaler;
#endif // SOUNDTOUCH_INTEGER_SAMPLES #endif // SOUNDTOUCH_INTEGER_SAMPLES
dest[j] = (SAMPLETYPE)suml; dest[j] = (SAMPLETYPE)suml;
dest[j + 1] = (SAMPLETYPE)sumr; dest[j + 1] = (SAMPLETYPE)sumr;
} }
return numSamples - length; return numSamples - ilength;
} }
@ -137,31 +127,28 @@ uint FIRFilter::evaluateFilterMono(SAMPLETYPE *dest, const SAMPLETYPE *src, uint
double dScaler = 1.0 / (double)resultDivider; double dScaler = 1.0 / (double)resultDivider;
#endif #endif
assert(length != 0); // hint compiler autovectorization that loop length is divisible by 8
int ilength = length & -8;
end = numSamples - length; assert(ilength != 0);
end = numSamples - ilength;
#pragma omp parallel for #pragma omp parallel for
for (j = 0; j < end; j ++) for (j = 0; j < end; j ++)
{ {
const SAMPLETYPE *pSrc = src + j; const SAMPLETYPE *pSrc = src + j;
LONG_SAMPLETYPE sum; LONG_SAMPLETYPE sum;
uint i; int i;
sum = 0; sum = 0;
for (i = 0; i < length; i += 4) for (i = 0; i < ilength; i ++)
{ {
// loop is unrolled by factor of 4 here for efficiency sum += pSrc[i] * filterCoeffs[i];
sum += pSrc[i + 0] * filterCoeffs[i + 0] +
pSrc[i + 1] * filterCoeffs[i + 1] +
pSrc[i + 2] * filterCoeffs[i + 2] +
pSrc[i + 3] * filterCoeffs[i + 3];
} }
#ifdef SOUNDTOUCH_INTEGER_SAMPLES #ifdef SOUNDTOUCH_INTEGER_SAMPLES
sum >>= resultDivFactor; sum >>= resultDivFactor;
// saturate to 16 bit integer limits // saturate to 16 bit integer limits
sum = (sum < -32768) ? -32768 : (sum > 32767) ? 32767 : sum; sum = (sum < -32768) ? -32768 : (sum > 32767) ? 32767 : sum;
#else
sum *= dScaler;
#endif // SOUNDTOUCH_INTEGER_SAMPLES #endif // SOUNDTOUCH_INTEGER_SAMPLES
dest[j] = (SAMPLETYPE)sum; dest[j] = (SAMPLETYPE)sum;
} }
@ -185,14 +172,18 @@ uint FIRFilter::evaluateFilterMulti(SAMPLETYPE *dest, const SAMPLETYPE *src, uin
assert(filterCoeffs != NULL); assert(filterCoeffs != NULL);
assert(numChannels < 16); assert(numChannels < 16);
end = numChannels * (numSamples - length); // hint compiler autovectorization that loop length is divisible by 8
int ilength = length & -8;
end = numChannels * (numSamples - ilength);
#pragma omp parallel for #pragma omp parallel for
for (j = 0; j < end; j += numChannels) for (j = 0; j < end; j += numChannels)
{ {
const SAMPLETYPE *ptr; const SAMPLETYPE *ptr;
LONG_SAMPLETYPE sums[16]; LONG_SAMPLETYPE sums[16];
uint c, i; uint c;
int i;
for (c = 0; c < numChannels; c ++) for (c = 0; c < numChannels; c ++)
{ {
@ -201,7 +192,7 @@ uint FIRFilter::evaluateFilterMulti(SAMPLETYPE *dest, const SAMPLETYPE *src, uin
ptr = src + j; ptr = src + j;
for (i = 0; i < length; i ++) for (i = 0; i < ilength; i ++)
{ {
SAMPLETYPE coef=filterCoeffs[i]; SAMPLETYPE coef=filterCoeffs[i];
for (c = 0; c < numChannels; c ++) for (c = 0; c < numChannels; c ++)
@ -215,13 +206,11 @@ uint FIRFilter::evaluateFilterMulti(SAMPLETYPE *dest, const SAMPLETYPE *src, uin
{ {
#ifdef SOUNDTOUCH_INTEGER_SAMPLES #ifdef SOUNDTOUCH_INTEGER_SAMPLES
sums[c] >>= resultDivFactor; sums[c] >>= resultDivFactor;
#else
sums[c] *= dScaler;
#endif // SOUNDTOUCH_INTEGER_SAMPLES #endif // SOUNDTOUCH_INTEGER_SAMPLES
dest[j+c] = (SAMPLETYPE)sums[c]; dest[j+c] = (SAMPLETYPE)sums[c];
} }
} }
return numSamples - length; return numSamples - ilength;
} }
@ -233,6 +222,13 @@ void FIRFilter::setCoefficients(const SAMPLETYPE *coeffs, uint newLength, uint u
assert(newLength > 0); assert(newLength > 0);
if (newLength % 8) ST_THROW_RT_ERROR("FIR filter length not divisible by 8"); if (newLength % 8) ST_THROW_RT_ERROR("FIR filter length not divisible by 8");
#ifdef SOUNDTOUCH_FLOAT_SAMPLES
// scale coefficients already here if using floating samples
double scale = 1.0 / resultDivider;
#else
short scale = 1;
#endif
lengthDiv8 = newLength / 8; lengthDiv8 = newLength / 8;
length = lengthDiv8 * 8; length = lengthDiv8 * 8;
assert(length == newLength); assert(length == newLength);
@ -242,7 +238,16 @@ void FIRFilter::setCoefficients(const SAMPLETYPE *coeffs, uint newLength, uint u
delete[] filterCoeffs; delete[] filterCoeffs;
filterCoeffs = new SAMPLETYPE[length]; filterCoeffs = new SAMPLETYPE[length];
memcpy(filterCoeffs, coeffs, length * sizeof(SAMPLETYPE)); delete[] filterCoeffsStereo;
filterCoeffsStereo = new SAMPLETYPE[length*2];
for (uint i = 0; i < length; i ++)
{
filterCoeffs[i] = (SAMPLETYPE)(coeffs[i] * scale);
// create also stereo set of filter coefficients: this allows compiler
// to autovectorize filter evaluation much more efficiently
filterCoeffsStereo[2 * i] = (SAMPLETYPE)(coeffs[i] * scale);
filterCoeffsStereo[2 * i + 1] = (SAMPLETYPE)(coeffs[i] * scale);
}
} }

1
3rdparty/soundtouch/source/SoundTouch/FIRFilter.h vendored
View File

@ -57,6 +57,7 @@ protected:
// Memory for filter coefficients // Memory for filter coefficients
SAMPLETYPE *filterCoeffs; SAMPLETYPE *filterCoeffs;
SAMPLETYPE *filterCoeffsStereo;
virtual uint evaluateFilterStereo(SAMPLETYPE *dest, virtual uint evaluateFilterStereo(SAMPLETYPE *dest,
const SAMPLETYPE *src, const SAMPLETYPE *src,

8
3rdparty/soundtouch/source/SoundTouch/InterpolateCubic.h vendored
View File

@ -41,7 +41,6 @@ namespace soundtouch
class InterpolateCubic : public TransposerBase class InterpolateCubic : public TransposerBase
{ {
protected: protected:
virtual void resetRegisters();
virtual int transposeMono(SAMPLETYPE *dest, virtual int transposeMono(SAMPLETYPE *dest,
const SAMPLETYPE *src, const SAMPLETYPE *src,
int &srcSamples); int &srcSamples);
@ -56,6 +55,13 @@ protected:
public: public:
InterpolateCubic(); InterpolateCubic();
virtual void resetRegisters();
int getLatency() const
{
return 1;
}
}; };
} }

2
3rdparty/soundtouch/source/SoundTouch/InterpolateLinear.cpp vendored
View File

@ -142,7 +142,7 @@ int InterpolateLinearInteger::transposeMulti(SAMPLETYPE *dest, const SAMPLETYPE
LONG_SAMPLETYPE temp, vol1; LONG_SAMPLETYPE temp, vol1;
assert(iFract < SCALE); assert(iFract < SCALE);
vol1 = (SCALE - iFract); vol1 = (LONG_SAMPLETYPE)(SCALE - iFract);
for (int c = 0; c < numChannels; c ++) for (int c = 0; c < numChannels; c ++)
{ {
temp = vol1 * src[c] + iFract * src[c + numChannels]; temp = vol1 * src[c] + iFract * src[c + numChannels];

18
3rdparty/soundtouch/source/SoundTouch/InterpolateLinear.h vendored
View File

@ -45,8 +45,6 @@ protected:
int iFract; int iFract;
int iRate; int iRate;
virtual void resetRegisters();
virtual int transposeMono(SAMPLETYPE *dest, virtual int transposeMono(SAMPLETYPE *dest,
const SAMPLETYPE *src, const SAMPLETYPE *src,
int &srcSamples); int &srcSamples);
@ -60,6 +58,13 @@ public:
/// Sets new target rate. Normal rate = 1.0, smaller values represent slower /// Sets new target rate. Normal rate = 1.0, smaller values represent slower
/// rate, larger faster rates. /// rate, larger faster rates.
virtual void setRate(double newRate); virtual void setRate(double newRate);
virtual void resetRegisters();
int getLatency() const
{
return 0;
}
}; };
@ -69,8 +74,6 @@ class InterpolateLinearFloat : public TransposerBase
protected: protected:
double fract; double fract;
virtual void resetRegisters();
virtual int transposeMono(SAMPLETYPE *dest, virtual int transposeMono(SAMPLETYPE *dest,
const SAMPLETYPE *src, const SAMPLETYPE *src,
int &srcSamples); int &srcSamples);
@ -81,6 +84,13 @@ protected:
public: public:
InterpolateLinearFloat(); InterpolateLinearFloat();
virtual void resetRegisters();
int getLatency() const
{
return 0;
}
}; };
} }

8
3rdparty/soundtouch/source/SoundTouch/InterpolateShannon.h vendored
View File

@ -46,7 +46,6 @@ namespace soundtouch
class InterpolateShannon : public TransposerBase class InterpolateShannon : public TransposerBase
{ {
protected: protected:
void resetRegisters();
int transposeMono(SAMPLETYPE *dest, int transposeMono(SAMPLETYPE *dest,
const SAMPLETYPE *src, const SAMPLETYPE *src,
int &srcSamples); int &srcSamples);
@ -61,6 +60,13 @@ protected:
public: public:
InterpolateShannon(); InterpolateShannon();
void resetRegisters();
int getLatency() const
{
return 3;
}
}; };
} }

6
3rdparty/soundtouch/source/SoundTouch/PeakFinder.cpp vendored
View File

@ -57,7 +57,7 @@ int PeakFinder::findTop(const float *data, int peakpos) const
refvalue = data[peakpos]; refvalue = data[peakpos];
// seek within <EFBFBD>10 points // seek within ±10 points
start = peakpos - 10; start = peakpos - 10;
if (start < minPos) start = minPos; if (start < minPos) start = minPos;
end = peakpos + 10; end = peakpos + 10;
@ -142,7 +142,7 @@ int PeakFinder::findCrossingLevel(const float *data, float level, int peakpos, i
peaklevel = data[peakpos]; peaklevel = data[peakpos];
assert(peaklevel >= level); assert(peaklevel >= level);
pos = peakpos; pos = peakpos;
while ((pos >= minPos) && (pos < maxPos)) while ((pos >= minPos) && (pos + direction < maxPos))
{ {
if (data[pos + direction] < level) return pos; // crossing found if (data[pos + direction] < level) return pos; // crossing found
pos += direction; pos += direction;
@ -256,7 +256,7 @@ double PeakFinder::detectPeak(const float *data, int aminPos, int amaxPos)
// accept harmonic peak if // accept harmonic peak if
// (a) it is found // (a) it is found
// (b) is within <EFBFBD>4% of the expected harmonic interval // (b) is within ±4% of the expected harmonic interval
// (c) has at least half x-corr value of the max. peak // (c) has at least half x-corr value of the max. peak
double diff = harmonic * peaktmp / highPeak; double diff = harmonic * peaktmp / highPeak;

10
3rdparty/soundtouch/source/SoundTouch/RateTransposer.cpp vendored
View File

@ -61,6 +61,7 @@ RateTransposer::RateTransposer() : FIFOProcessor(&outputBuffer)
// Instantiates the anti-alias filter // Instantiates the anti-alias filter
pAAFilter = new AAFilter(64); pAAFilter = new AAFilter(64);
pTransposer = TransposerBase::newInstance(); pTransposer = TransposerBase::newInstance();
clear();
} }
@ -77,6 +78,7 @@ void RateTransposer::enableAAFilter(bool newMode)
#ifndef SOUNDTOUCH_PREVENT_CLICK_AT_RATE_CROSSOVER #ifndef SOUNDTOUCH_PREVENT_CLICK_AT_RATE_CROSSOVER
// Disable Anti-alias filter if desirable to avoid click at rate change zero value crossover // Disable Anti-alias filter if desirable to avoid click at rate change zero value crossover
bUseAAFilter = newMode; bUseAAFilter = newMode;
clear();
#endif #endif
} }
@ -192,6 +194,11 @@ void RateTransposer::clear()
outputBuffer.clear(); outputBuffer.clear();
midBuffer.clear(); midBuffer.clear();
inputBuffer.clear(); inputBuffer.clear();
pTransposer->resetRegisters();
// prefill buffer to avoid losing first samples at beginning of stream
int prefill = getLatency();
inputBuffer.addSilent(prefill);
} }
@ -209,7 +216,8 @@ int RateTransposer::isEmpty() const
/// Return approximate initial input-output latency /// Return approximate initial input-output latency
int RateTransposer::getLatency() const int RateTransposer::getLatency() const
{ {
return (bUseAAFilter) ? pAAFilter->getLength() : 0; return pTransposer->getLatency() +
((bUseAAFilter) ? (pAAFilter->getLength() / 2) : 0);
} }

5
3rdparty/soundtouch/source/SoundTouch/RateTransposer.h vendored
View File

@ -59,8 +59,6 @@ public:
}; };
protected: protected:
virtual void resetRegisters() = 0;
virtual int transposeMono(SAMPLETYPE *dest, virtual int transposeMono(SAMPLETYPE *dest,
const SAMPLETYPE *src, const SAMPLETYPE *src,
int &srcSamples) = 0; int &srcSamples) = 0;
@ -83,6 +81,9 @@ public:
virtual int transpose(FIFOSampleBuffer &dest, FIFOSampleBuffer &src); virtual int transpose(FIFOSampleBuffer &dest, FIFOSampleBuffer &src);
virtual void setRate(double newRate); virtual void setRate(double newRate);
virtual void setChannels(int channels); virtual void setChannels(int channels);
virtual int getLatency() const = 0;
virtual void resetRegisters() = 0;
// static factory function // static factory function
static TransposerBase *newInstance(); static TransposerBase *newInstance();

121
3rdparty/soundtouch/source/SoundTouch/TDStretch.cpp vendored
View File

@ -1,4 +1,4 @@
//////////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////////
/// ///
/// Sampled sound tempo changer/time stretch algorithm. Changes the sound tempo /// Sampled sound tempo changer/time stretch algorithm. Changes the sound tempo
/// while maintaining the original pitch by using a time domain WSOLA-like /// while maintaining the original pitch by using a time domain WSOLA-like
@ -54,7 +54,6 @@ using namespace soundtouch;
#define max(x, y) (((x) > (y)) ? (x) : (y)) #define max(x, y) (((x) > (y)) ? (x) : (y))
/***************************************************************************** /*****************************************************************************
* *
* Constant definitions * Constant definitions
@ -93,11 +92,6 @@ TDStretch::TDStretch() : FIFOProcessor(&outputBuffer)
bAutoSeqSetting = true; bAutoSeqSetting = true;
bAutoSeekSetting = true; bAutoSeekSetting = true;
maxnorm = 0;
maxnormf = 1e8;
skipFract = 0;
tempo = 1.0f; tempo = 1.0f;
setParameters(44100, DEFAULT_SEQUENCE_MS, DEFAULT_SEEKWINDOW_MS, DEFAULT_OVERLAP_MS); setParameters(44100, DEFAULT_SEQUENCE_MS, DEFAULT_SEEKWINDOW_MS, DEFAULT_OVERLAP_MS);
setTempo(1.0f); setTempo(1.0f);
@ -224,6 +218,9 @@ void TDStretch::clearInput()
inputBuffer.clear(); inputBuffer.clear();
clearMidBuffer(); clearMidBuffer();
isBeginning = true; isBeginning = true;
maxnorm = 0;
maxnormf = 1e8;
skipFract = 0;
} }
@ -315,9 +312,10 @@ int TDStretch::seekBestOverlapPositionFull(const SAMPLETYPE *refPos)
{ {
double corr; double corr;
// Calculates correlation value for the mixing position corresponding to 'i' // Calculates correlation value for the mixing position corresponding to 'i'
#ifdef _OPENMP #if defined(_OPENMP) || defined(ST_SIMD_AVOID_UNALIGNED)
// in parallel OpenMP mode, can't use norm accumulator version as parallel executor won't // in parallel OpenMP mode, can't use norm accumulator version as parallel executor won't
// iterate the loop in sequential order // iterate the loop in sequential order
// in SIMD mode, avoid accumulator version to allow avoiding unaligned positions
corr = calcCrossCorr(refPos + channels * i, pMidBuffer, norm); corr = calcCrossCorr(refPos + channels * i, pMidBuffer, norm);
#else #else
// In non-parallel version call "calcCrossCorrAccumulate" that is otherwise same // In non-parallel version call "calcCrossCorrAccumulate" that is otherwise same
@ -675,11 +673,10 @@ void TDStretch::processSamples()
// Adjust processing offset at beginning of track by not perform initial overlapping // Adjust processing offset at beginning of track by not perform initial overlapping
// and compensating that in the 'input buffer skip' calculation // and compensating that in the 'input buffer skip' calculation
isBeginning = false; isBeginning = false;
int skip = (int)(tempo * overlapLength + 0.5); int skip = (int)(tempo * overlapLength + 0.5 * seekLength + 0.5);
#ifdef SOUNDTOUCH_ALLOW_NONEXACT_SIMD_OPTIMIZATION #ifdef ST_SIMD_AVOID_UNALIGNED
#ifdef SOUNDTOUCH_ALLOW_SSE // in SIMD mode, round the skip amount to value corresponding to aligned memory address
// if SSE mode, round the skip amount to value corresponding to aligned memory address
if (channels == 1) if (channels == 1)
{ {
skip &= -4; skip &= -4;
@ -689,9 +686,11 @@ void TDStretch::processSamples()
skip &= -2; skip &= -2;
} }
#endif #endif
#endif
skipFract -= skip; skipFract -= skip;
assert(nominalSkip >= -skipFract); if (skipFract <= -nominalSkip)
{
skipFract = -nominalSkip;
}
} }
// ... then copy sequence samples from 'inputBuffer' to output: // ... then copy sequence samples from 'inputBuffer' to output:
@ -830,21 +829,19 @@ void TDStretch::overlapStereo(short *poutput, const short *input) const
// Overlaps samples in 'midBuffer' with the samples in 'input'. The 'Multi' // Overlaps samples in 'midBuffer' with the samples in 'input'. The 'Multi'
// version of the routine. // version of the routine.
void TDStretch::overlapMulti(SAMPLETYPE *poutput, const SAMPLETYPE *input) const void TDStretch::overlapMulti(short *poutput, const short *input) const
{ {
SAMPLETYPE m1=(SAMPLETYPE)0; short m1;
SAMPLETYPE m2;
int i = 0; int i = 0;
for (m2 = (SAMPLETYPE)overlapLength; m2; m2 --) for (m1 = 0; m1 < overlapLength; m1 ++)
{ {
short m2 = (short)(overlapLength - m1);
for (int c = 0; c < channels; c ++) for (int c = 0; c < channels; c ++)
{ {
poutput[i] = (input[i] * m1 + pMidBuffer[i] * m2) / overlapLength; poutput[i] = (input[i] * m1 + pMidBuffer[i] * m2) / overlapLength;
i++; i++;
} }
m1++;
} }
} }
@ -889,20 +886,23 @@ double TDStretch::calcCrossCorr(const short *mixingPos, const short *compare, do
unsigned long lnorm; unsigned long lnorm;
int i; int i;
#ifdef ST_SIMD_AVOID_UNALIGNED
// in SIMD mode skip 'mixingPos' positions that aren't aligned to 16-byte boundary
if (((ulongptr)mixingPos) & 15) return -1e50;
#endif
// hint compiler autovectorization that loop length is divisible by 8
int ilength = (channels * overlapLength) & -8;
corr = lnorm = 0; corr = lnorm = 0;
// Same routine for stereo and mono. For stereo, unroll loop for better // Same routine for stereo and mono
// efficiency and gives slightly better resolution against rounding. for (i = 0; i < ilength; i += 2)
// For mono it same routine, just unrolls loop by factor of 4
for (i = 0; i < channels * overlapLength; i += 4)
{ {
corr += (mixingPos[i] * compare[i] + corr += (mixingPos[i] * compare[i] +
mixingPos[i + 1] * compare[i + 1]) >> overlapDividerBitsNorm; // notice: do intermediate division here to avoid integer overflow mixingPos[i + 1] * compare[i + 1]) >> overlapDividerBitsNorm;
corr += (mixingPos[i + 2] * compare[i + 2] +
mixingPos[i + 3] * compare[i + 3]) >> overlapDividerBitsNorm;
lnorm += (mixingPos[i] * mixingPos[i] + lnorm += (mixingPos[i] * mixingPos[i] +
mixingPos[i + 1] * mixingPos[i + 1]) >> overlapDividerBitsNorm; // notice: do intermediate division here to avoid integer overflow mixingPos[i + 1] * mixingPos[i + 1]) >> overlapDividerBitsNorm;
lnorm += (mixingPos[i + 2] * mixingPos[i + 2] + // do intermediate scalings to avoid integer overflow
mixingPos[i + 3] * mixingPos[i + 3]) >> overlapDividerBitsNorm;
} }
if (lnorm > maxnorm) if (lnorm > maxnorm)
@ -925,9 +925,12 @@ double TDStretch::calcCrossCorr(const short *mixingPos, const short *compare, do
double TDStretch::calcCrossCorrAccumulate(const short *mixingPos, const short *compare, double &norm) double TDStretch::calcCrossCorrAccumulate(const short *mixingPos, const short *compare, double &norm)
{ {
long corr; long corr;
unsigned long lnorm; long lnorm;
int i; int i;
// hint compiler autovectorization that loop length is divisible by 8
int ilength = (channels * overlapLength) & -8;
// cancel first normalizer tap from previous round // cancel first normalizer tap from previous round
lnorm = 0; lnorm = 0;
for (i = 1; i <= channels; i ++) for (i = 1; i <= channels; i ++)
@ -936,15 +939,11 @@ double TDStretch::calcCrossCorrAccumulate(const short *mixingPos, const short *c
} }
corr = 0; corr = 0;
// Same routine for stereo and mono. For stereo, unroll loop for better // Same routine for stereo and mono.
// efficiency and gives slightly better resolution against rounding. for (i = 0; i < ilength; i += 2)
// For mono it same routine, just unrolls loop by factor of 4
for (i = 0; i < channels * overlapLength; i += 4)
{ {
corr += (mixingPos[i] * compare[i] + corr += (mixingPos[i] * compare[i] +
mixingPos[i + 1] * compare[i + 1]) >> overlapDividerBitsNorm; // notice: do intermediate division here to avoid integer overflow mixingPos[i + 1] * compare[i + 1]) >> overlapDividerBitsNorm;
corr += (mixingPos[i + 2] * compare[i + 2] +
mixingPos[i + 3] * compare[i + 3]) >> overlapDividerBitsNorm;
} }
// update normalizer with last samples of this round // update normalizer with last samples of this round
@ -1045,27 +1044,24 @@ void TDStretch::calculateOverlapLength(int overlapInMsec)
/// Calculate cross-correlation /// Calculate cross-correlation
double TDStretch::calcCrossCorr(const float *mixingPos, const float *compare, double &anorm) double TDStretch::calcCrossCorr(const float *mixingPos, const float *compare, double &anorm)
{ {
double corr; float corr;
double norm; float norm;
int i; int i;
#ifdef ST_SIMD_AVOID_UNALIGNED
// in SIMD mode skip 'mixingPos' positions that aren't aligned to 16-byte boundary
if (((ulongptr)mixingPos) & 15) return -1e50;
#endif
// hint compiler autovectorization that loop length is divisible by 8
int ilength = (channels * overlapLength) & -8;
corr = norm = 0; corr = norm = 0;
// Same routine for stereo and mono. For Stereo, unroll by factor of 2. // Same routine for stereo and mono
// For mono it's same routine yet unrollsd by factor of 4. for (i = 0; i < ilength; i ++)
for (i = 0; i < channels * overlapLength; i += 4)
{ {
corr += mixingPos[i] * compare[i] + corr += mixingPos[i] * compare[i];
mixingPos[i + 1] * compare[i + 1]; norm += mixingPos[i] * mixingPos[i];
norm += mixingPos[i] * mixingPos[i] +
mixingPos[i + 1] * mixingPos[i + 1];
// unroll the loop for better CPU efficiency:
corr += mixingPos[i + 2] * compare[i + 2] +
mixingPos[i + 3] * compare[i + 3];
norm += mixingPos[i + 2] * mixingPos[i + 2] +
mixingPos[i + 3] * mixingPos[i + 3];
} }
anorm = norm; anorm = norm;
@ -1076,7 +1072,7 @@ double TDStretch::calcCrossCorr(const float *mixingPos, const float *compare, do
/// Update cross-correlation by accumulating "norm" coefficient by previously calculated value /// Update cross-correlation by accumulating "norm" coefficient by previously calculated value
double TDStretch::calcCrossCorrAccumulate(const float *mixingPos, const float *compare, double &norm) double TDStretch::calcCrossCorrAccumulate(const float *mixingPos, const float *compare, double &norm)
{ {
double corr; float corr;
int i; int i;
corr = 0; corr = 0;
@ -1087,14 +1083,13 @@ double TDStretch::calcCrossCorrAccumulate(const float *mixingPos, const float *c
norm -= mixingPos[-i] * mixingPos[-i]; norm -= mixingPos[-i] * mixingPos[-i];
} }
// Same routine for stereo and mono. For Stereo, unroll by factor of 2. // hint compiler autovectorization that loop length is divisible by 8
// For mono it's same routine yet unrollsd by factor of 4. int ilength = (channels * overlapLength) & -8;
for (i = 0; i < channels * overlapLength; i += 4)
// Same routine for stereo and mono
for (i = 0; i < ilength; i ++)
{ {
corr += mixingPos[i] * compare[i] + corr += mixingPos[i] * compare[i];
mixingPos[i + 1] * compare[i + 1] +
mixingPos[i + 2] * compare[i + 2] +
mixingPos[i + 3] * compare[i + 3];
} }
// update normalizer with last samples of this round // update normalizer with last samples of this round

2
3rdparty/soundtouch/source/SoundTouch/sse_optimized.cpp vendored
View File

@ -80,7 +80,7 @@ double TDStretchSSE::calcCrossCorr(const float *pV1, const float *pV2, double &a
// Compile-time define SOUNDTOUCH_ALLOW_NONEXACT_SIMD_OPTIMIZATION is provided // Compile-time define SOUNDTOUCH_ALLOW_NONEXACT_SIMD_OPTIMIZATION is provided
// for choosing if this little cheating is allowed. // for choosing if this little cheating is allowed.
#ifdef SOUNDTOUCH_ALLOW_NONEXACT_SIMD_OPTIMIZATION #ifdef ST_SIMD_AVOID_UNALIGNED
// Little cheating allowed, return valid correlation only for // Little cheating allowed, return valid correlation only for
// aligned locations, meaning every second round for stereo sound. // aligned locations, meaning every second round for stereo sound.