dsp::Convolution Class Reference

Performs stereo partitioned convolution of an input signal with an impulse response in the frequency domain, using the JUCE FFT class. More...

Classes
struct	Latency
	Contains configuration information for a convolution with a fixed latency. More...
struct	NonUniform
	Contains configuration information for a non-uniform convolution. More...

Public Types
enum class	Stereo { no , yes }
enum class	Trim { no , yes }
enum class	Normalise { no , yes }

Public Member Functions
	Convolution ()
	Initialises an object for performing convolution in the frequency domain.
	Convolution (ConvolutionMessageQueue &queue)
	Initialises a convolution engine using a shared background message queue.
	Convolution (const Latency &requiredLatency)
	Initialises an object for performing convolution with a fixed latency.
	Convolution (const NonUniform &requiredHeadSize)
	Initialises an object for performing convolution in the frequency domain using a non-uniform partitioned algorithm.
	Convolution (const Latency &, ConvolutionMessageQueue &)
	Behaves the same as the constructor taking a single Latency argument, but with a shared background message queue.
	Convolution (const NonUniform &, ConvolutionMessageQueue &)
	Behaves the same as the constructor taking a single NonUniform argument, but with a shared background message queue.
	~Convolution () noexcept
void	prepare (const ProcessSpec &)
	Must be called before first calling process.
void	reset () noexcept
	Resets the processing pipeline ready to start a new stream of data.
template<typename ProcessContext, std::enable_if_t< std::is_same_v< typename ProcessContext::SampleType, float >, int > = 0>
void	process (const ProcessContext &context) noexcept
	Performs the filter operation on the given set of samples with optional stereo processing.
void	loadImpulseResponse (const void *sourceData, size_t sourceDataSize, Stereo isStereo, Trim requiresTrimming, size_t size, Normalise requiresNormalisation=Normalise::yes)
	This function loads an impulse response audio file from memory, added in a JUCE project with the Projucer as binary data.
void	loadImpulseResponse (const File &fileImpulseResponse, Stereo isStereo, Trim requiresTrimming, size_t size, Normalise requiresNormalisation=Normalise::yes)
	This function loads an impulse response from an audio file.
void	loadImpulseResponse (AudioBuffer< float > &&buffer, double bufferSampleRate, Stereo isStereo, Trim requiresTrimming, Normalise requiresNormalisation)
	This function loads an impulse response from an audio buffer.
int	getCurrentIRSize () const
	This function returns the size of the current IR in samples.
int	getLatency () const
	This function returns the current latency of the process in samples.

Detailed Description

Performs stereo partitioned convolution of an input signal with an impulse response in the frequency domain, using the JUCE FFT class.

This class provides some thread-safe functions to load impulse responses from audio files or memory on-the-fly without noticeable artefacts, performing resampling and trimming if necessary.

The processing performed by this class is equivalent to the time domain convolution done in the FIRFilter class, with a FIRFilter::Coefficients object having the samples of the impulse response as its coefficients. However, in general it is more efficient to do frequency domain convolution when the size of the impulse response is 64 samples or greater.

Note: The default operation of this class uses zero latency and a uniform partitioned algorithm. If the impulse response size is large, or if the algorithm is too CPU intensive, it is possible to use either a fixed latency version of the algorithm, or a simple non-uniform partitioned convolution algorithm.

Threading: It is not safe to interleave calls to the methods of this class. If you need to load new impulse responses during processing the load() calls must be synchronised with process() calls, which in practice means making the load() call from the audio thread. The loadImpulseResponse() functions are wait-free and are therefore suitable for use in a realtime context.

See also

FIRFilter, FIRFilter::Coefficients, FFT

Member Enumeration Documentation

Stereo

enum class dsp::Convolution::Stereo

strong

Enumerator
no
yes

Trim

enum class dsp::Convolution::Trim

strong

Enumerator
no
yes

Normalise

enum class dsp::Convolution::Normalise

strong

Enumerator
no
yes

Constructor & Destructor Documentation

Convolution() [1/6]

dsp::Convolution::Convolution

(

)

Initialises an object for performing convolution in the frequency domain.

Referenced by getLatency().

Convolution() [2/6]

dsp::Convolution::Convolution ( ConvolutionMessageQueue & queue )

explicit

Initialises a convolution engine using a shared background message queue.

IMPORTANT: the queue must remain alive throughout the lifetime of the Convolution.

Convolution() [3/6]

dsp::Convolution::Convolution ( const Latency & requiredLatency )

explicit

Initialises an object for performing convolution with a fixed latency.

If the requested latency is zero, the actual latency will also be zero. For requested latencies greater than zero, the actual latency will always at least as large as the requested latency. Using a fixed non-zero latency can reduce the CPU consumption of the convolution algorithm.

Parameters

requiredLatency

the minimum latency

Convolution() [4/6]

dsp::Convolution::Convolution ( const NonUniform & requiredHeadSize )

explicit

Initialises an object for performing convolution in the frequency domain using a non-uniform partitioned algorithm.

A requiredHeadSize of 256 samples or greater will improve the efficiency of the processing for IR sizes of 4096 samples or greater (recommended for reverberation IRs).

Parameters

requiredHeadSize

the head IR size for two stage non-uniform partitioned convolution

Convolution() [5/6]

dsp::Convolution::Convolution	(	const Latency &	,
		ConvolutionMessageQueue &	)

Behaves the same as the constructor taking a single Latency argument, but with a shared background message queue.

IMPORTANT: the queue must remain alive throughout the lifetime of the Convolution.

Convolution() [6/6]

dsp::Convolution::Convolution	(	const NonUniform &	,
		ConvolutionMessageQueue &	)

Behaves the same as the constructor taking a single NonUniform argument, but with a shared background message queue.

IMPORTANT: the queue must remain alive throughout the lifetime of the Convolution.

~Convolution()

dsp::Convolution::~Convolution ( )

noexcept

Member Function Documentation

prepare()

void dsp::Convolution::prepare

(

const ProcessSpec &

)

Must be called before first calling process.

In general, calls to loadImpulseResponse() load the impulse response (IR) asynchronously. The IR will become active once it has been completely loaded and processed, which may take some time.

Calling prepare() will ensure that the IR supplied to the most recent call to loadImpulseResponse() is fully initialised. This IR will then be active during the next call to process(). It is recommended to call loadImpulseResponse() before prepare() if a specific IR must be active during the first process() call.

References prepare().

Referenced by getLatency(), and prepare().

reset()

void dsp::Convolution::reset ( )

noexcept

Resets the processing pipeline ready to start a new stream of data.

References reset().

Referenced by getLatency(), and reset().

process()

template<typename ProcessContext, std::enable_if_t< std::is_same_v< typename ProcessContext::SampleType, float >, int > = 0>

void dsp::Convolution::process ( const ProcessContext & context )

noexcept

Performs the filter operation on the given set of samples with optional stereo processing.

References process().

Referenced by process().

loadImpulseResponse() [1/3]

void dsp::Convolution::loadImpulseResponse	(	const void *	sourceData,
		size_t	sourceDataSize,
		Stereo	isStereo,
		Trim	requiresTrimming,
		size_t	size,
		Normalise	requiresNormalisation = Normalise::yes )

This function loads an impulse response audio file from memory, added in a JUCE project with the Projucer as binary data.

It can load any of the audio formats registered in JUCE, and performs some resampling and pre-processing as well if needed.

Note: Don't try to use this function on float samples, since the data is expected to be an audio file in its binary format. Be sure that the original data remains constant throughout the lifetime of the Convolution object, as the loading process will happen on a background thread once this function has returned.

Parameters

sourceData	the block of data to use as the stream's source
sourceDataSize	the number of bytes in the source data block
isStereo	selects either stereo or mono
requiresTrimming	optionally trim the start and the end of the impulse response
size	the expected size for the impulse response after loading, can be set to 0 to requesting the original impulse response size
requiresNormalisation	optionally normalise the impulse response amplitude

References yes.

loadImpulseResponse() [2/3]

void dsp::Convolution::loadImpulseResponse	(	const File &	fileImpulseResponse,
		Stereo	isStereo,
		Trim	requiresTrimming,
		size_t	size,
		Normalise	requiresNormalisation = Normalise::yes )

This function loads an impulse response from an audio file.

It can load any of the audio formats registered in JUCE, and performs some resampling and pre-processing as well if needed.

Parameters

fileImpulseResponse	the location of the audio file
isStereo	selects either stereo or mono
requiresTrimming	optionally trim the start and the end of the impulse response
size	the expected size for the impulse response after loading, can be set to 0 to requesting the original impulse response size
requiresNormalisation	optionally normalise the impulse response amplitude

References yes.

loadImpulseResponse() [3/3]

void dsp::Convolution::loadImpulseResponse	(	AudioBuffer< float > &&	buffer,
		double	bufferSampleRate,
		Stereo	isStereo,
		Trim	requiresTrimming,
		Normalise	requiresNormalisation )

This function loads an impulse response from an audio buffer.

To avoid memory allocation on the audio thread, this function takes ownership of the buffer passed in.

If calling this function during processing, make sure that the buffer is not allocated on the audio thread (be careful of accidental copies!). If you need to pass arbitrary/generated buffers it's recommended to create these buffers on a separate thread and to use some wait-free construct (a lock-free queue or a SpinLock/GenericScopedTryLock combination) to transfer ownership to the audio thread without allocating.

Parameters

buffer	the AudioBuffer to use
bufferSampleRate	the sampleRate of the data in the AudioBuffer
isStereo	selects either stereo or mono
requiresTrimming	optionally trim the start and the end of the impulse response
requiresNormalisation	optionally normalise the impulse response amplitude

getCurrentIRSize()

int dsp::Convolution::getCurrentIRSize

(

)

const

This function returns the size of the current IR in samples.

getLatency()

int dsp::Convolution::getLatency

(

)

const

This function returns the current latency of the process in samples.

Note: This is the latency of the convolution engine, not the latency associated with the current impulse response choice that has to be considered separately (linear phase filters, for example).

References Convolution(), dsp::isBypassed(), JUCE_DECLARE_NON_COPYABLE_WITH_LEAK_DETECTOR, prepare(), and reset().