// Copy constructor.
FFTFrame::FFTFrame(const FFTFrame& frame)
: m_FFTSize(frame.m_FFTSize),
m_log2FFTSize(frame.m_log2FFTSize),
m_realData(frame.m_FFTSize / 2),
m_imagData(frame.m_FFTSize / 2),
m_forwardContext(nullptr),
m_inverseContext(nullptr),
m_complexData(frame.m_FFTSize) {
m_forwardContext = contextForSize(m_FFTSize, DFT_R2C);
m_inverseContext = contextForSize(m_FFTSize, IDFT_C2R);
// Copy/setup frame data.
unsigned nbytes = sizeof(float) * (m_FFTSize / 2);
memcpy(realData(), frame.realData(), nbytes);
memcpy(imagData(), frame.imagData(), nbytes);
}
void Chroma::Consume(const FFTFrame &frame)
{
fill(m_features.begin(), m_features.end(), 0.0);
for (int i = m_min_index; i < m_max_index; i++) {
int note = m_notes[i];
double energy = frame.Energy(i);
if (m_interpolate) {
int note2 = note;
double a = 1.0;
if (m_notes_frac[i] < 0.5) {
note2 = (note + NUM_BANDS - 1) % NUM_BANDS;
a = 0.5 + m_notes_frac[i];
}
if (m_notes_frac[i] > 0.5) {
note2 = (note + 1) % NUM_BANDS;
a = 1.5 - m_notes_frac[i];
}
m_features[note] += energy * a;
m_features[note2] += energy * (1.0 - a);
}
else {
m_features[note] += energy;
}
}
m_consumer->Consume(m_features);
}
// Copy constructor.
FFTFrame::FFTFrame(const FFTFrame& frame)
: m_FFTSize(frame.m_FFTSize)
, m_log2FFTSize(frame.m_log2FFTSize)
, m_realData(unpackedFFTDataSize(frame.m_FFTSize))
, m_imagData(unpackedFFTDataSize(frame.m_FFTSize))
{
m_complexData = WTF::fastNewArray<GstFFTF32Complex>(unpackedFFTDataSize(m_FFTSize));
int fftLength = gst_fft_next_fast_length(m_FFTSize);
m_fft = gst_fft_f32_new(fftLength, FALSE);
m_inverseFft = gst_fft_f32_new(fftLength, TRUE);
// Copy/setup frame data.
memcpy(realData(), frame.realData(), sizeof(float) * unpackedFFTDataSize(m_FFTSize));
memcpy(imagData(), frame.imagData(), sizeof(float) * unpackedFFTDataSize(m_FFTSize));
}
void FFTLib::Compute(FFTFrame &frame) {
kiss_fftr(m_cfg, m_input, m_output);
auto input = m_output;
auto output = frame.data();
for (size_t i = 0; i <= m_frame_size / 2; ++i, ++input, ++output) {
*output = input->r * input->r + input->i * input->i;
}
}
// Copy constructor.
FFTFrame::FFTFrame(const FFTFrame& frame)
: m_FFTSize(frame.m_FFTSize)
, m_log2FFTSize(frame.m_log2FFTSize)
, m_forwardPlan(0)
, m_backwardPlan(0)
, m_data(2 * (3 + unpackedFFTWDataSize(fftSize()))) // enough space for real and imaginary data plus 16-byte alignment padding
{
// See the normal constructor for an explanation of the temporary pointer.
float temporary;
m_forwardPlan = fftwPlanForSize(m_FFTSize, Forward,
&temporary, realData(), imagData());
m_backwardPlan = fftwPlanForSize(m_FFTSize, Backward,
realData(), imagData(), &temporary);
// Copy/setup frame data.
size_t nbytes = sizeof(float) * unpackedFFTWDataSize(fftSize());
memcpy(realData(), frame.realData(), nbytes);
memcpy(imagData(), frame.imagData(), nbytes);
}
void FFTLib::Compute(FFTFrame &frame) {
av_rdft_calc(m_rdft_ctx, m_input);
auto input = m_input;
auto output = frame.begin();
output[0] = input[0] * input[0];
output[m_frame_size / 2] = input[1] * input[1];
output += 1;
input += 2;
for (size_t i = 1; i < m_frame_size / 2; i++) {
*output++ = input[0] * input[0] + input[1] * input[1];
input += 2;
}
}
void FFTLib::Compute(FFTFrame &frame) {
fftw_execute(m_plan);
auto output = frame.data();
auto in_ptr = m_output;
auto rev_in_ptr = m_output + m_frame_size - 1;
output[0] = in_ptr[0] * in_ptr[0];
output[m_frame_size / 2] = in_ptr[m_frame_size / 2] * in_ptr[m_frame_size / 2];
in_ptr += 1;
output += 1;
for (size_t i = 1; i < m_frame_size / 2; i++) {
*output++ = in_ptr[0] * in_ptr[0] + rev_in_ptr[0] * rev_in_ptr[0];
in_ptr++;
rev_in_ptr--;
}
}
std::unique_ptr<FFTFrame> FFTFrame::createInterpolatedFrame(const FFTFrame& frame1, const FFTFrame& frame2, double x)
{
std::unique_ptr<FFTFrame> newFrame(new FFTFrame(frame1.fftSize()));
newFrame->interpolateFrequencyComponents(frame1, frame2, x);
// In the time-domain, the 2nd half of the response must be zero, to avoid circular convolution aliasing...
int fftSize = newFrame->fftSize();
AudioFloatArray buffer(fftSize);
newFrame->doInverseFFT(buffer.data());
buffer.zeroRange(fftSize / 2, fftSize);
// Put back into frequency domain.
newFrame->doFFT(buffer.data());
return newFrame;
}
void FFTFrame::interpolateFrequencyComponents(const FFTFrame& frame1, const FFTFrame& frame2, double interp)
{
// FIXME : with some work, this method could be optimized
float* realP = realData();
float* imagP = imagData();
const float* realP1 = frame1.realData();
const float* imagP1 = frame1.imagData();
const float* realP2 = frame2.realData();
const float* imagP2 = frame2.imagData();
m_FFTSize = frame1.fftSize();
m_log2FFTSize = frame1.log2FFTSize();
double s1base = (1.0 - interp);
double s2base = interp;
double phaseAccum = 0.0;
double lastPhase1 = 0.0;
double lastPhase2 = 0.0;
realP[0] = static_cast<float>(s1base * realP1[0] + s2base * realP2[0]);
imagP[0] = static_cast<float>(s1base * imagP1[0] + s2base * imagP2[0]);
int n = m_FFTSize / 2;
for (int i = 1; i < n; ++i) {
Complex c1(realP1[i], imagP1[i]);
Complex c2(realP2[i], imagP2[i]);
double mag1 = abs(c1);
double mag2 = abs(c2);
// Interpolate magnitudes in decibels
double mag1db = 20.0 * log10(mag1);
double mag2db = 20.0 * log10(mag2);
double s1 = s1base;
double s2 = s2base;
double magdbdiff = mag1db - mag2db;
// Empirical tweak to retain higher-frequency zeroes
double threshold = (i > 16) ? 5.0 : 2.0;
if (magdbdiff < -threshold && mag1db < 0.0)
{
s1 = pow(s1, 0.75);
s2 = 1.0 - s1;
} else if (magdbdiff > threshold && mag2db < 0.0)
{
s2 = pow(s2, 0.75);
s1 = 1.0 - s2;
}
// Average magnitude by decibels instead of linearly
double magdb = s1 * mag1db + s2 * mag2db;
double mag = pow(10.0, 0.05 * magdb);
// Now, deal with phase
double phase1 = arg(c1);
double phase2 = arg(c2);
double deltaPhase1 = phase1 - lastPhase1;
double deltaPhase2 = phase2 - lastPhase2;
lastPhase1 = phase1;
lastPhase2 = phase2;
// Unwrap phase deltas
if (deltaPhase1 > piDouble)
deltaPhase1 -= 2.0 * piDouble;
if (deltaPhase1 < -piDouble)
deltaPhase1 += 2.0 * piDouble;
if (deltaPhase2 > piDouble)
deltaPhase2 -= 2.0 * piDouble;
if (deltaPhase2 < -piDouble)
deltaPhase2 += 2.0 * piDouble;
// Blend group-delays
double deltaPhaseBlend;
if (deltaPhase1 - deltaPhase2 > piDouble)
deltaPhaseBlend = s1 * deltaPhase1 + s2 * (2.0 * piDouble + deltaPhase2);
else if (deltaPhase2 - deltaPhase1 > piDouble)
deltaPhaseBlend = s1 * (2.0 * piDouble + deltaPhase1) + s2 * deltaPhase2;
else
deltaPhaseBlend = s1 * deltaPhase1 + s2 * deltaPhase2;
phaseAccum += deltaPhaseBlend;
// Unwrap
if (phaseAccum > piDouble)
phaseAccum -= 2.0 * piDouble;
if (phaseAccum < -piDouble)
phaseAccum += 2.0 * piDouble;
Complex c = std::polar(mag, phaseAccum);
realP[i] = static_cast<float>(c.real());
imagP[i] = static_cast<float>(c.imag());
}
}
