CUGL 1.3
Cornell University Game Library
|
#include <CUBiquadIIR.h>
Public Types | |
enum | Type : int { Type::UNDEFINED = 0, Type::LOWPASS = 1, Type::HIGHPASS = 2, Type::BANDPASS = 3, Type::ALLPASS = 4, Type::NOTCH = 5, Type::PEAK = 6, Type::LOWSHELF = 7, Type::HIGHSHELF = 8, Type::RESONANCE = 9 } |
Public Member Functions | |
BiquadIIR () | |
BiquadIIR (unsigned channels) | |
BiquadIIR (unsigned channels, Type type, float frequency, float gainDB, float qVal=INV_SQRT2) | |
BiquadIIR (const BiquadIIR ©) | |
BiquadIIR (BiquadIIR &&filter) | |
~BiquadIIR () | |
unsigned | getChannels () const |
void | setChannels (unsigned channels) |
void | setCoeff (const std::vector< float > &bvals, const std::vector< float > &avals) |
const std::vector< float > | getBCoeff () const |
const std::vector< float > | getACoeff () const |
void | setBCoeff (float b0, float b1, float b2) |
void | setACoeff (float a1, float a2) |
void | setType (Type type, float frequency, float gainDB, float qVal=INV_SQRT2) |
void | step (float gain, float *input, float *output) |
void | calculate (float gain, float *input, float *output, size_t size) |
void | clear () |
size_t | flush (float *output) |
Static Public Member Functions | |
static float | db2gain (float gainDB) |
static float | gain2db (float gain) |
static float | bandwidth2q (float width) |
static float | q2Bandwidth (float qVal) |
Static Public Attributes | |
static bool | VECTORIZE |
This class implements a biquad digital filter.
This is the most efficient filter acceptable for a parametric equalizer. As such, this filter has several types for quick creation of parameteric components. However, in most settings Butterworth filters are preferred because they have better roll off.
Frequencies are specified in "normalized" format. A normalized frequency is frequency/sample rate. For example, a 7 kHz frequency with a 44100 Hz sample rate has a normalized value 7000/44100 = 0.15873. However, filters are not intended to be model classes, and so it does not save the defining frequency.
This class supports vector optimizations for SSE and Neon 64. In timed simulations, these optimizations provide at least a 3-4x performance increase (and for 4 or 8 channel audio, much higher). These optimizations make use of the matrix precomputation outlined in "Implementation of Recursive Digital Filters into Vector SIMD DSP Architectures".
https://pdfs.semanticscholar.org/d150/a3f75dc033916f14029cd9101a8ea1d050bb.pdf
The algorithm in this paper performs extremely well in our tests, and even out-performs Apple's Acceleration library. However, our implementation is limited to 128-bit words as 256-bit (e.g. AVX) and higher show no significant increase in performance.
For performance reasons, this class does not have a (virtualized) subclass relationship with other IIR or FIR filters. However, the signature of the the calculation and coefficient methods has been standardized so that it can support templated polymorphism.
This class is not thread safe. External locking may be required when the filter is shared between multiple threads (such as between an audio thread and the main thread).
|
strong |
The underlying type of the biquad filter.
Most biquad filters are intended for a parameteric equalizer, and so will have one of the filters types below. If the coefficients of the biquad filter are set directly, it will have type UNDEFINED.
cugl::dsp::BiquadIIR::BiquadIIR | ( | ) |
Creates a second-order pass-through filter for a single channel.
cugl::dsp::BiquadIIR::BiquadIIR | ( | unsigned | channels | ) |
Creates a second-order pass-through filter for the given number of channels.
channels | The number of channels |
cugl::dsp::BiquadIIR::BiquadIIR | ( | unsigned | channels, |
Type | type, | ||
float | frequency, | ||
float | gainDB, | ||
float | qVal = INV_SQRT2 |
||
) |
Creates a special purpose filter of the given type
In addition to the type, the filter is defined by the target frequency and the gain for that frequency (which may be negative). This gain will be applied to the target frequency, but will roll-off or attenuate for other frequencies according to the type. The gain is specified in decibels, not as a multiplicative factor.
Frequencies are specified in "normalized" format. A normalized frequency is frequency/sample rate. For example, a 7 kHz frequency with a 44100 Hz sample rate has a normalized value 7000/44100 = 0.15873.
The Q factor is the inverse of the bandwidth, and is generally only relevant for the BANDPASS and BANDSTOP filter types. For the other types, the default value of 1/sqrt(2) is generally sufficient. For BANDPASS and BANDSTOP, the bandwidth2q() value produces the appropriate value for proper filter design.
Filters are not intended to be model classes, and so it does not save the defining frequency, type, gain, or other values.
If the type is undefined, the frequency and peakGain will be ignored, creating a pass-through filter.
channels | The number of channels to process |
type | The filter type |
frequency | The (normalized) target frequency |
gainDB | The gain at the target frequency in decibels |
qVal | The special Q factor |
cugl::dsp::BiquadIIR::BiquadIIR | ( | const BiquadIIR & | copy | ) |
Creates a copy of the biquad filter.
copy | The filter to copy |
cugl::dsp::BiquadIIR::BiquadIIR | ( | BiquadIIR && | filter | ) |
Creates a biquad filter with the resources of the original.
filter | The filter to acquire |
cugl::dsp::BiquadIIR::~BiquadIIR | ( | ) |
Destroys the filter, releasing all resources.
|
static |
Returns the q value for the given the filter bandwidth (in octaves)
The filter bandwidth is generally only relevant for the BANDPASS and BANDSTOP filter types. For the other types, the q value of 1/sqrt(2) is generally sufficient.
width | The filter bandwidth in octaves |
void cugl::dsp::BiquadIIR::calculate | ( | float | gain, |
float * | input, | ||
float * | output, | ||
size_t | size | ||
) |
Performs a filter of interleaved input data.
The output is written to the given output array, which should be the same size as the input array. The size is the number of frames, not samples. Hence the arrays must be size times the number of channels in size.
To provide real time processing, the output is delayed by the number of a-coefficients. Delayed results are buffered to be used the next time the filter is used (though they may be extracted with the flush method). The gain parameter is applied at the filter input, but does not affect the filter coefficients.
gain | The input gain factor |
input | The array of input samples |
output | The array to write the sample output |
size | The input size in frames |
void cugl::dsp::BiquadIIR::clear | ( | ) |
Clears the filter buffer of any delayed outputs or cached inputs
|
static |
Returns the gain factor for given value in decibels.
The factor is the amount to multiply the amplitude signal. The provided value is represented in decibels, so there may be some round-off error in conversion.
gainDB | The audio gain in decibels |
size_t cugl::dsp::BiquadIIR::flush | ( | float * | output | ) |
Flushes any delayed outputs to the provided array
The array size should be twice the number of channels. This method will also clear the buffer.
|
static |
Returns the decibel gain for given factor.
The factor is the amount to multiply the amplitude signal. The returned result is in decibels, so there may be some round-off error in conversion.
gain | The audio gain factor |
const std::vector<float> cugl::dsp::BiquadIIR::getACoeff | ( | ) | const |
Returns the lower coefficients for this IIR filter.
This filter implements the standard difference equation:
a[0]*y[n] = b[0]*x[n]+...+b[nb]*x[n-nb]-a[1]*y[n-1]-...-a[na]*y[n-na]
where y is the output and x in the input.
const std::vector<float> cugl::dsp::BiquadIIR::getBCoeff | ( | ) | const |
Returns the upper coefficients for this IIR filter.
This filter implements the standard difference equation:
a[0]*y[n] = b[0]*x[n]+...+b[nb]*x[n-nb]-a[1]*y[n-1]-...-a[na]*y[n-na]
where y is the output and x in the input.
|
inline |
Returns the number of channels for this filter
The data buffers depend on the number of channels. Changing this value will reset the data buffers to 0.
|
static |
Returns the filter bandwidth (in octaves) for the given q value
The filter bandwidth is generally only relevant for the BANDPASS and BANDSTOP filter types. For the other types, the q value of 1/sqrt(2) is generally sufficient.
qVal | The special Q factor |
void cugl::dsp::BiquadIIR::setACoeff | ( | float | a1, |
float | a2 | ||
) |
Sets the lower coefficients.
Setting this leaves the upper coefficients unchanged.
a1 | The lower first-order coefficient |
a2 | The lower second-order coefficient |
void cugl::dsp::BiquadIIR::setBCoeff | ( | float | b0, |
float | b1, | ||
float | b2 | ||
) |
Sets the upper coefficients.
Setting this leaves the lower coefficients unchanged.
b0 | The upper zero-order coefficient |
b1 | The upper first-order coefficient |
b2 | The upper second-order coefficient |
void cugl::dsp::BiquadIIR::setChannels | ( | unsigned | channels | ) |
Sets the number of channels for this filter
The data buffers depend on the number of channels. Changing this value will reset the data buffers to 0.
channels | The number of channels for this filter |
void cugl::dsp::BiquadIIR::setCoeff | ( | const std::vector< float > & | bvals, |
const std::vector< float > & | avals | ||
) |
Sets the coefficients for this IIR filter.
This filter implements the standard difference equation:
a[0]*y[n] = b[0]*x[n]+...+b[nb]*x[n-nb]-a[1]*y[n-1]-...-a[na]*y[n-na]
where y is the output and x in the input. If a[0] is not equal to 1, the filter coeffcients are normalized by a[0].
All b-coefficients and a-coefficients after the third are ignored. If any coefficients are missing, they are replaced with 1 for b[0] and a[0], and 0 otherwise.
bvals | The upper coefficients |
avals | The lower coefficients |
void cugl::dsp::BiquadIIR::setType | ( | Type | type, |
float | frequency, | ||
float | gainDB, | ||
float | qVal = INV_SQRT2 |
||
) |
Sets this filter to the special purpose one of the given type
In addition to the type, the filter is defined by the target frequency and the gain for that frequency (which may be negative). This gain will be applied to the target frequency, but will roll-off or attenuate for other frequencies according to the type. The gain is specified in decibels, not as a multiplicative factor,
Frequencies are specified in "normalized" format. A normalized frequency is frequency/sample rate. For example, a 7 kHz frequency with a 44100 Hz sample rate has a normalized value 7000/44100 = 0.15873.
The Q factor is the inverse of the bandwidth, and is generally only relevant for the BANDPASS and BANDSTOP filter types. For the other types, the default value of 1/sqrt(2) is generally sufficient. For BANDPASS and BANDSTOP, the bandwidth2q() value produces the appropriate value for proper filter design.
Filters are not intended to be model classes, and so it does not save the defining frequency, type, gain, or other values.
If the type is undefined, the frequency and peakGain will be ignored, creating a pass-through filter.
type | The filter type |
frequency | The (normalized) target frequency |
gainDB | The gain at the target frequency in decibels |
qVal | The special Q factor |
void cugl::dsp::BiquadIIR::step | ( | float | gain, |
float * | input, | ||
float * | output | ||
) |
Performs a filter of single frame of data.
The output is written to the given output array, which should be the same size as the input array. The size should be the number of channels.
To provide real time processing, the output is delayed by the number of a-coefficients. Delayed results are buffered to be used the next time the filter is used (though they may be extracted with the flush method). The gain parameter is applied at the filter input, but does not affect the filter coefficients.
gain | The input gain factor |
input | The input frame |
output | The frame to receive the output |
|
static |
Whether to use a vectorization algorithm (Access not thread safe)