Mtp1 » History » Version 8

« Previous - Version 8/12 (diff) - Next » - Current version
Chris Cannam, 2012-02-23 05:28 PM


From Method to Plugin: Building a new plugin on OS/X with make

Note: This tutorial is specific to OS/X. Go here for a version that uses Windows and Visual C++.

We're going to walk through the process of making, and compiling, a new Vamp plugin based on the skeleton files included with the Vamp plugin SDK.

We will start by setting up a project in which we just get the skeleton plugin to compile without it doing any actual work, and then we'll add some substance to it afterwards.

All work will be done using the terminal window and make (rather than using, say, the Xcode IDE), so a passing familiarity with the command line will really help. Some familiarity with C++ will be necessary too in the later steps.

The focus here is on the practical details of what you need to put in a plugin and how to get it to build and run -- not on the real mathematical or signal-processing aspect. We will pick a very simple method (time-domain signal power, block by block) for this example. Please refer to the Vamp plugin API programmer's guide for further reading, with information about returning more sophisticated features.

Before you begin: Make sure you have the Xcode tools (the OS/X developer SDK) installed! You can't compile anything without it.

Note on build architectures: Before OS/X 10.6, the default for the Xcode tools was to build 32-bit Intel binaries (known as "i386") when running on an Intel Mac, and 32-bit PowerPC ("ppc") when running on a PowerPC. This was changed in 10.6 so as to build 64-bit Intel binaries ("x86_64") by default. Unfortunately, plugins that are 64-bit only cannot be loaded into 32-bit hosts, such as the commonly distributed versions of all current Vamp hosts. OS/X does support building for more than one architecture at once (storing the results in a fat file or "universal binary"), and that is the approach we take in this tutorial. If we were to build for only a single architecture, i386 would currently be the more useful choice.

1. Download and build the SDK

Download the Vamp plugin SDK version 2.2 from the "development headers and source code" link on the developer page at http://vamp-plugins.org/develop.html -- the file you want is vamp-plugin-sdk-2.2.tar.gz. Save it into your home directory, open a terminal window, and unpack it. We'll also rename its directory from vamp-plugin-sdk-2.2 to vamp-plugin-sdk for easier reference later on.

mac:~ chris$ ls vamp*
vamp-plugin-sdk-2.2.tar.gz
mac:~ chris$ tar xvzf vamp-plugin-sdk-2.2.tar.gz
 ... lots of output ...
mac:~ chris$ mv vamp-plugin-sdk-2.2 vamp-plugin-sdk
mac:~ chris$

At this point you really ought to read the README file in the SDK directory, and the README.osx file in the SDK's build subdirectory. But for the tutorial we'll skip that and plunge in and build the SDK directly.

We'll only build the SDK libraries and example plugins. We won't build the test host, because it requires an additional library (libsndfile). We'll download a pre-compiled binary of the test host later instead. (There are pre-compiled libraries too, but since we still need the SDK for the header files, we might as well compile it in place.)

mac:~ chris$ cd vamp-plugin-sdk
mac:~/vamp-plugin-sdk chris$ make -f build/Makefile.osx sdk
 ... lots of output ...
mac:~/vamp-plugin-sdk chris$ make -f build/Makefile.osx plugins
 ... lots of output ...
mac:~/vamp-plugin-sdk chris$

2. Copy the skeleton files to our new project home

We're going to build our plugin in a new directory called tutorial in our home directory.

mac:~/vamp-plugin-sdk chris$ cd 
mac:~ chris$ mkdir tutorial
mac:~ chris$ cd tutorial
mac:~/tutorial chris$

The starting point will be the set of skeleton source files provided with the SDK. These compile into a valid, "working" Vamp plugin that happens to do nothing at all.

mac:~/tutorial chris$ cp ../vamp-plugin-sdk/skeleton/* .
mac:~/tutorial chris$ ls
Makefile.skeleton       MyPlugin.h              vamp-plugin.list
MyPlugin.cpp            plugins.cpp             vamp-plugin.map
mac:~/tutorial chris$

The bulk of the skeleton plugin code is contained in the files MyPlugin.cpp and MyPlugin.h. These two files implement a single C++ class, called MyPlugin. For the sake of brevity in the tutorial we'll leave these names unchanged, but you might prefer to change them! To do so, rename the two files as you wish, and replace every occurrence of the text MyPlugin in both of them, and in plugins.cpp, with your preferred plugin class name.

The file plugins.cpp contains the entry point for the plugin library. A library can hold more than one plugin, and the job of plugins.cpp is to provide a single known public function (vampGetPluginDescriptor) which the host can use to find out what plugins are available in the library. The skeleton version of plugins.cpp just returns the single MyPlugin plugin class.

Note that it makes absolutely no difference to the operation of the plugin what its class is called, or what any of these files is called; MyPlugin is (in purely technical terms) as good a name as any. It also shouldn't matter if two different libraries happen to use the same class name. But if you have more than one plugin in the same library, they'll need to have different class names then!

3. Get the skeleton build working

The first thing we'll do with this skeleton project is build it into a "working" (although pointless) plugin.

To build it we're going to use the tool make, which takes a set of production rules described in a Makefile and uses them to turn source files into targets, in this case with the help of the C++ compiler.

The skeleton project contains a file Makefile.skeleton which will be the basis of our Makefile.

mac:~/tutorial chris$ cp Makefile.skeleton Makefile

Now, open the Makefile in the text editor; we need to edit it to suit our new project. We haven't changed the names of any of the skeleton source files, so we don't need to edit those, but we do need to uncomment the lines that are specific to compiling on OS/X. We want to make a universal binary (32- and 64-bit) rather than a native build, so we look for:
##  Uncomment these for an OS/X universal binary (PPC, 32- and 64-bit Intel) using command-line tools:

# CXXFLAGS = -isysroot /Developer/SDKs/MacOSX10.4u.sdk -arch i386 -arch x86_64 -arch ppc -I$(VAMP_SDK_DIR) -Wall -fPIC
# PLUGIN_EXT = .dylib
# PLUGIN = $(PLUGIN_LIBRARY_NAME)$(PLUGIN_EXT)
# LDFLAGS = -dynamiclib -install_name $(PLUGIN) $(VAMP_SDK_DIR)/libvamp-sdk.a -exported_symbols_list vamp-plugin.list

Remove the # characters from the starts of the four lines in that block:
##  Uncomment these for an OS/X universal binary (PPC, 32- and 64-bit Intel) using command-line tools:

CXXFLAGS = -isysroot /Developer/SDKs/MacOSX10.4u.sdk -arch i386 -arch x86_64 -arch ppc -I$(VAMP_SDK_DIR) -Wall -fPIC
PLUGIN_EXT = .dylib
PLUGIN = $(PLUGIN_LIBRARY_NAME)$(PLUGIN_EXT)
LDFLAGS = -dynamiclib -install_name $(PLUGIN) $(VAMP_SDK_DIR)/libvamp-sdk.a -exported_symbols_list vamp-plugin.list

Then, without changing anything else, save the file and run make.

mac:~/tutorial chris$ make
g++ -isysroot /Developer/SDKs/MacOSX10.4u.sdk -arch i386 -arch x86_64 -arch ppc -I../vamp-plugin-sdk -Wall -fPIC -c -o MyPlugin.o MyPlugin.cpp
g++ -isysroot /Developer/SDKs/MacOSX10.4u.sdk -arch i386 -arch x86_64 -arch ppc -I../vamp-plugin-sdk -Wall -fPIC -c -o plugins.o plugins.cpp
g++ -o myplugins.dylib MyPlugin.o plugins.o -dynamiclib -install_name myplugins.dylib ../vamp-plugin-sdk/libvamp-sdk.a -exported_symbols_list vamp-plugin.list
mac:~/tutorial chris$ 

You should now have a plugin library file called myplugins.dylib, as well as some .o files created during the build process.

mac:~/tutorial chris$ ls
Makefile                MyPlugin.o              vamp-plugin.list
Makefile.skeleton       myplugins.dylib         vamp-plugin.map
MyPlugin.cpp            plugins.cpp
MyPlugin.h              plugins.o
mac:~/tutorial chris$

This myplugins.dylib file is a valid and complete Vamp plugin library. It doesn't do anything worthwhile, but it can be loaded and "used" in any host. It defines a single Vamp plugin, whose identifier is "myplugin" (this is coded into the MyPlugin.cpp file, we'll be changing it later).

4. Check that the plugin works with some test programs

The next thing to do is gather some programs we can use to test our plugin, so that we can check it built correctly, and so that we'll be well placed to test it properly when it actually does something.

vamp-simple-host

The first one is the vamp-simple-host that is part of the Vamp SDK. This is the part of the SDK that we didn't build in step 1 (because of its dependency on libsndfile). Download it from the "pre-compiled library and host binaries" link at http://vamp-plugins.org/develop.html; the file you're downloading will be vamp-plugin-sdk-2.1-binaries-osx-universal.tar.gz. For the time being, just open this archive in the OS/X Finder and unpack the single file vamp-simple-host into our project directory (the tutorial one).

Like all Vamp hosts, vamp-simple-host understands the VAMP_PATH environment variable to tell it where to look for plugins. We can set this variable for a single run of the program by prefixing the program name with the variable assignment on the command line. And, with the -l option, we can ask it to list the plugins it finds there.

mac:~/tutorial chris$ ls vamp-simple-host
vamp-simple-host
mac:~/tutorial chris$ VAMP_PATH=. ./vamp-simple-host -l
Vamp plugin search path: [.]

Vamp plugin libraries found in search path:

  ./myplugins.dylib:
    [A] [v2] My Plugin, "myplugin" []

mac:~/tutorial chris$

Huzzah. We can use this host to run the plugin on some test audio files, not just list it, but there isn't much point yet.

vamp-plugin-tester

The other test host worth setting up at the start is vamp-plugin-tester, a program that tests your plugin for a number of possible problems and pitfalls. You can download this from http://vamp-plugins.org/develop.html as well; copy it into the same directory as well, for the time being.

mac:~/tutorial chris$ ls vamp-plugin-tester
vamp-plugin-tester
mac:~/tutorial chris$ VAMP_PATH=. ./vamp-plugin-tester -a
vamp-plugin-tester: Running...
Testing plugin: myplugins:myplugin
 -- Performing test: A1 Invalid identifiers
 -- Performing test: A2 Empty metadata fields
 ** WARNING: Plugin description is empty
 ** WARNING: Plugin maker is empty
 ** WARNING: Plugin copyright is empty
 ** WARNING: Plugin parameter "parameter" description is empty
 ** WARNING: Plugin output "output" description is empty
 -- Performing test: A3 Inappropriate value extents
 -- Performing test: B1 Output number mismatching
 ** NOTE: No results returned for output "output" 
 -- Performing test: B2 Invalid or dubious timestamp usage
 -- Performing test: C1 Normal input
 -- Performing test: C2 Empty input
 -- Performing test: C3 Short input
 -- Performing test: C4 Absolutely silent input
 -- Performing test: C5 Input beyond traditional +/-1 range
 -- Performing test: C6 Random input
 -- Performing test: D1 Consecutive runs with separate instances
 -- Performing test: D2 Consecutive runs with a single instance using reset
 -- Performing test: D3 Simultaneous interleaved runs in a single thread
 -- Performing test: D4 Consecutive runs with different start times
 ** WARNING: Consecutive runs with different starting timestamps produce the same result
 -- Performing test: E1 Inconsistent default program
 -- Performing test: E2 Inconsistent default parameters
 -- Performing test: F1 Different sample rates
 -- Performing test: F2 Lengthy constructor
vamp-plugin-tester: All tests succeeded for this plugin

vamp-plugin-tester: All tests succeeded, with 6 warning(s) and 1 other note(s)
mac:~/tutorial chris$

As you see, vamp-plugin-tester runs quite a number of tests -- see its README file for more details about the error and warning messages it might give. It's a good idea to use the tester right from the start of plugin development.

5. Now, the code!

Right, let's make the plugin do something. We're going to calculate the mean power for each processing block. The work we do in this section will involve editing the MyPlugin.cpp file (and at one point also MyPlugins.h) in the text editor.

The calculation we want is sum(x[i]^2) / N, where x[i] is audio sample number i, for i in the range 0 to N-1, with N the number of samples in the processing block.

Describing the input and output formats

Our calculation is a time-domain one (working directly from the PCM audio data), which means we don't need to change this function (found at line 63):

MyPlugin::InputDomain
MyPlugin::getInputDomain() const
{
    return TimeDomain;
}

We are going to write code to handle a single audio channel only, and leave it to the host to decide what to do if more than one channel is provided (most hosts will mix-down the input for us). So that means we don't need to change these functions either:

size_t
MyPlugin::getMinChannelCount() const
{
    return 1;
}

size_t
MyPlugin::getMaxChannelCount() const
{
    return 1;
}

Nothing about our calculation requires us to constrain the processing block size -- we can handle any block size. So we can leave this function unchanged as well:

size_t
MyPlugin::getPreferredBlockSize() const
{
    return 0; // 0 means "I can handle any block size" 
}

The function getOutputDescriptors describes what sort of features we intend to return. As it happens, the skeleton already contains pretty much the description we are going to need: a single feature, with a single value, returned for each processing block. We should probably change the name of the output, at least:

MyPlugin::OutputList
MyPlugin::getOutputDescriptors() const
{
    OutputList list;

    OutputDescriptor d;
    d.identifier = "power";
    d.name = "Power";
    d.description = "";
    d.unit = "";
    d.hasFixedBinCount = true;
    d.binCount = 1;
    d.hasKnownExtents = false;
    d.isQuantized = false;
    d.sampleType = OutputDescriptor::OneSamplePerStep;
    d.hasDuration = false;
    list.push_back(d);

    return list;
}

Initialisation

We said that we can accept any block size -- but we do need to know what the block size is.

This is told to us in the initialise function. Looking at that function, we can see the argument is size_t blockSize. It's our job to remember the value of this.

We need to add a class data member for this. In MyPlugin.h, look for this line at line 54 (near the bottom of the file):

    // plugin-specific data and methods go here

and add a line after it:

    // plugin-specific data and methods go here
    size_t m_blockSize;

Then, back in MyPlugin.cpp, find this line at line 187 in the initialise function:

    // Real initialisation work goes here!

and add a line to set the data member:

    // Real initialisation work goes here!
    m_blockSize = blockSize;

Also it's very good practice to make sure the data member is initialised to zero in the class constructor. That, at line 10 of MyPlugin.cpp, initially reads:

MyPlugin::MyPlugin(float inputSampleRate) :
    Plugin(inputSampleRate)
{
}

and we want it to read:

MyPlugin::MyPlugin(float inputSampleRate) :
    Plugin(inputSampleRate),
    m_blockSize(0)
{
}

At this point, after saving both the header and .cpp file, it's probably worth going back to the terminal window and making sure it still compiles. (Just run make again.)

Processing

The core of our calculation happens in the process method:

MyPlugin::FeatureSet
MyPlugin::process(const float *const *inputBuffers, Vamp::RealTime timestamp)
{
    // Do actual work!
    return FeatureSet();
}

Here inputBuffers is effectively an array of arrays -- to retrieve a single audio sample, we index it first by audio channel number (we know that we only have one channel, so the only valid index is 0) and then by audio sample number (from 0 to the processing block size less 1).

What we want to do is add up the squares of the audio sample values, and divide by the number of samples.

MyPlugin::FeatureSet
MyPlugin::process(const float *const *inputBuffers, Vamp::RealTime timestamp)
{
    float sumOfSquares = 0.0f;

    size_t i = 0; // note: same type as m_blockSize

    while (i < m_blockSize) {
        float sample = inputBuffers[0][i];
        sumOfSquares += sample * sample;
        ++i;
    }

    float meanPower = sumOfSquares / m_blockSize;

    // now what?

    return FeatureSet();
}

So we've calculated the mean power value -- now how to return it?

In Vamp plugin terms, what we have is a plugin that has a single output, on which is returned a single audio feature for each process block, with one value. We need to construct a Feature object, give it a single value, and then push it as the only feature in output 0 (the first) of a new FeatureSet object. See the Vamp plugin API programmer's guide for more information about feature representation.

Here's the code:

MyPlugin::FeatureSet
MyPlugin::process(const float *const *inputBuffers, Vamp::RealTime timestamp)
{
    float sumOfSquares = 0.0f;

    size_t i = 0;

    while (i < m_blockSize) {
        float sample = inputBuffers[0][i];
        sumOfSquares += sample * sample;
        ++i;
    }

    float meanPower = sumOfSquares / m_blockSize;

    Feature f;
    f.hasTimestamp = false;
    f.values.push_back(meanPower);

    FeatureSet fs;
    fs[0].push_back(f);
    return fs;
}

After making this change and returning to the terminal window to run make again, we now have a plugin that actually does something.

With a suitable input file (.wav or .aiff or a similar uncompressed format that vamp-simple-host understands), we can now run it:

mac:~/tutorial chris$ VAMP_PATH=. ./vamp-simple-host myplugins:myplugin ~/my-song.wav
vamp-simple-host: Running...
Reading file: "/Users/chris/my-song.wav", writing to standard output
Running plugin: "myplugin"...                                        
Using block size = 1024, step size = 1024                            
Plugin accepts 1 -> 1 channel(s)                                     
Sound file has 2 (will mix/augment if necessary)                     
Output is: "output"                                                  
 0.000000000: 0
 0.023219954: 0
 0.046439909: 0
 0.069659863: 0
 0.092879818: 0
 0.116099773: 0
 0.139319727: 0
 0.162539682: 1.56888e-11
 0.185759637: 4.90218e-09
 0.208979591: 2.135e-07
 0.232199546: 0.00666197
 ... and lots and lots and lots and lots more output ...
mac:~/tutorial chris$

Try using the vamp-plugin-tester again as well.

6. Fill in descriptions and other metadata

Now we have a working plugin, but it still has the rather awkward name of MyPlugin. There are several functions at the top of MyPlugin.cpp which we can use to give it a more sensible name and description.

For example:

string
MyPlugin::getIdentifier() const
{
    return "myplugin";
}

The identifier is a string that is not normally used by people (for example, it never appears when plugins are listed in a menu of a graphical application), but that uniquely identifies the plugin within its library. Something like "power" is perfectly appropriate here.

You should fill in all of getIdentifier, getName, getDescription, getMaker, getPluginVersion, and getCopyright for every plugin you write.

In my case, I would need something like:

string
MyPlugin::getIdentifier() const
{
    return "power";
}

string
MyPlugin::getName() const
{
    return "Signal power level";
}

string
MyPlugin::getDescription() const
{
    return "Calculate the mean signal power for each processing block";
}

string
MyPlugin::getMaker() const
{
    return "Chris Cannam";
}

int
MyPlugin::getPluginVersion() const
{
    return 1;
}

string
MyPlugin::getCopyright() const
{
    return "Freely redistributable (tutorial example code)";
}