Chris@152: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
Chris@152: 
Chris@152: /*
Chris@152:     Sonic Visualiser
Chris@152:     An audio file viewer and annotation editor.
Chris@152:     Centre for Digital Music, Queen Mary, University of London.
Chris@152:     This file copyright 2006 Chris Cannam.
Chris@152:     
Chris@152:     This program is free software; you can redistribute it and/or
Chris@152:     modify it under the terms of the GNU General Public License as
Chris@152:     published by the Free Software Foundation; either version 2 of the
Chris@152:     License, or (at your option) any later version.  See the file
Chris@152:     COPYING included with this distribution for more information.
Chris@152: */
Chris@152: 
Chris@152: #include "FFTModel.h"
Chris@152: #include "DenseTimeValueModel.h"
Chris@152: 
Chris@183: #include "base/Profiler.h"
Chris@275: #include "base/Pitch.h"
Chris@183: 
Chris@402: #include <algorithm>
Chris@402: 
Chris@152: #include <cassert>
Chris@1090: #include <deque>
Chris@152: 
Chris@608: #ifndef __GNUC__
Chris@608: #include <alloca.h>
Chris@608: #endif
Chris@608: 
Chris@1090: using namespace std;
Chris@1090: 
Chris@152: FFTModel::FFTModel(const DenseTimeValueModel *model,
Chris@152:                    int channel,
Chris@152:                    WindowType windowType,
Chris@929:                    int windowSize,
Chris@929:                    int windowIncrement,
Chris@1090:                    int fftSize) :
Chris@1090:     m_model(model),
Chris@1090:     m_channel(channel),
Chris@1090:     m_windowType(windowType),
Chris@1090:     m_windowSize(windowSize),
Chris@1090:     m_windowIncrement(windowIncrement),
Chris@1090:     m_fftSize(fftSize),
Chris@1091:     m_windower(windowType, windowSize),
Chris@1093:     m_fft(fftSize),
Chris@1093:     m_cacheSize(3)
Chris@152: {
Chris@1091:     if (m_windowSize > m_fftSize) {
Chris@1091:         cerr << "ERROR: FFTModel::FFTModel: window size (" << m_windowSize
Chris@1091:              << ") must be at least FFT size (" << m_fftSize << ")" << endl;
Chris@1091:         throw invalid_argument("FFTModel window size must be at least FFT size");
Chris@1091:     }
Chris@1133: 
Chris@1133:     connect(model, SIGNAL(modelChanged()), this, SIGNAL(modelChanged()));
Chris@1133:     connect(model, SIGNAL(modelChangedWithin(sv_frame_t, sv_frame_t)),
Chris@1133:             this, SIGNAL(modelChangedWithin(sv_frame_t, sv_frame_t)));
Chris@152: }
Chris@152: 
Chris@152: FFTModel::~FFTModel()
Chris@152: {
Chris@152: }
Chris@152: 
Chris@360: void
Chris@360: FFTModel::sourceModelAboutToBeDeleted()
Chris@360: {
Chris@1090:     if (m_model) {
Chris@1090:         cerr << "FFTModel[" << this << "]::sourceModelAboutToBeDeleted(" << m_model << ")" << endl;
Chris@1090:         m_model = 0;
Chris@360:     }
Chris@360: }
Chris@360: 
Chris@1091: int
Chris@1091: FFTModel::getWidth() const
Chris@1091: {
Chris@1091:     if (!m_model) return 0;
Chris@1091:     return int((m_model->getEndFrame() - m_model->getStartFrame())
Chris@1091:                / m_windowIncrement) + 1;
Chris@1091: }
Chris@1091: 
Chris@1091: int
Chris@1091: FFTModel::getHeight() const
Chris@1091: {
Chris@1091:     return m_fftSize / 2 + 1;
Chris@1091: }
Chris@1091: 
Chris@152: QString
Chris@929: FFTModel::getBinName(int n) const
Chris@152: {
Chris@1040:     sv_samplerate_t sr = getSampleRate();
Chris@152:     if (!sr) return "";
Chris@204:     QString name = tr("%1 Hz").arg((n * sr) / ((getHeight()-1) * 2));
Chris@152:     return name;
Chris@152: }
Chris@152: 
Chris@1091: FFTModel::Column
Chris@1091: FFTModel::getColumn(int x) const
Chris@1091: {
Chris@1091:     auto cplx = getFFTColumn(x);
Chris@1091:     Column col;
Chris@1154:     col.reserve(cplx.size());
Chris@1091:     for (auto c: cplx) col.push_back(abs(c));
Chris@1154:     return move(col);
Chris@1091: }
Chris@1091: 
Chris@1091: float
Chris@1091: FFTModel::getMagnitudeAt(int x, int y) const
Chris@1091: {
Chris@1093:     if (x < 0 || x >= getWidth() || y < 0 || y >= getHeight()) return 0.f;
Chris@1093:     auto col = getFFTColumn(x);
Chris@1093:     return abs(col[y]);
Chris@1091: }
Chris@1091: 
Chris@1091: float
Chris@1091: FFTModel::getMaximumMagnitudeAt(int x) const
Chris@1091: {
Chris@1091:     Column col(getColumn(x));
Chris@1092:     float max = 0.f;
Chris@1154:     int n = int(col.size());
Chris@1154:     for (int i = 0; i < n; ++i) {
Chris@1092:         if (col[i] > max) max = col[i];
Chris@1092:     }
Chris@1092:     return max;
Chris@1091: }
Chris@1091: 
Chris@1091: float
Chris@1091: FFTModel::getPhaseAt(int x, int y) const
Chris@1091: {
Chris@1093:     if (x < 0 || x >= getWidth() || y < 0 || y >= getHeight()) return 0.f;
Chris@1091:     return arg(getFFTColumn(x)[y]);
Chris@1091: }
Chris@1091: 
Chris@1091: void
Chris@1091: FFTModel::getValuesAt(int x, int y, float &re, float &im) const
Chris@1091: {
Chris@1091:     auto col = getFFTColumn(x);
Chris@1091:     re = col[y].real();
Chris@1091:     im = col[y].imag();
Chris@1091: }
Chris@1091: 
Chris@1091: bool
Chris@1091: FFTModel::getMagnitudesAt(int x, float *values, int minbin, int count) const
Chris@1091: {
Chris@1091:     if (count == 0) count = getHeight();
Chris@1091:     auto col = getFFTColumn(x);
Chris@1091:     for (int i = 0; i < count; ++i) {
Chris@1091:         values[i] = abs(col[minbin + i]);
Chris@1091:     }
Chris@1091:     return true;
Chris@1091: }
Chris@1091: 
Chris@1091: bool
Chris@1091: FFTModel::getPhasesAt(int x, float *values, int minbin, int count) const
Chris@1091: {
Chris@1091:     if (count == 0) count = getHeight();
Chris@1091:     auto col = getFFTColumn(x);
Chris@1091:     for (int i = 0; i < count; ++i) {
Chris@1091:         values[i] = arg(col[minbin + i]);
Chris@1091:     }
Chris@1091:     return true;
Chris@1091: }
Chris@1091: 
Chris@1091: bool
Chris@1091: FFTModel::getValuesAt(int x, float *reals, float *imags, int minbin, int count) const
Chris@1091: {
Chris@1091:     if (count == 0) count = getHeight();
Chris@1091:     auto col = getFFTColumn(x);
Chris@1091:     for (int i = 0; i < count; ++i) {
Chris@1091:         reals[i] = col[minbin + i].real();
Chris@1091:     }
Chris@1091:     for (int i = 0; i < count; ++i) {
Chris@1091:         imags[i] = col[minbin + i].imag();
Chris@1091:     }
Chris@1091:     return true;
Chris@1091: }
Chris@1091: 
Chris@1091: vector<float>
Chris@1091: FFTModel::getSourceSamples(int column) const
Chris@1091: {
Chris@1094:     // m_fftSize may be greater than m_windowSize, but not the reverse
Chris@1094: 
Chris@1094: //    cerr << "getSourceSamples(" << column << ")" << endl;
Chris@1094:     
Chris@1091:     auto range = getSourceSampleRange(column);
Chris@1094:     auto data = getSourceData(range);
Chris@1094: 
Chris@1091:     int off = (m_fftSize - m_windowSize) / 2;
Chris@1094: 
Chris@1094:     if (off == 0) {
Chris@1094:         return data;
Chris@1094:     } else {
Chris@1094:         vector<float> pad(off, 0.f);
Chris@1094:         vector<float> padded;
Chris@1094:         padded.reserve(m_fftSize);
Chris@1094:         padded.insert(padded.end(), pad.begin(), pad.end());
Chris@1094:         padded.insert(padded.end(), data.begin(), data.end());
Chris@1094:         padded.insert(padded.end(), pad.begin(), pad.end());
Chris@1094:         return padded;
Chris@1094:     }
Chris@1094: }
Chris@1094: 
Chris@1094: vector<float>
Chris@1094: FFTModel::getSourceData(pair<sv_frame_t, sv_frame_t> range) const
Chris@1094: {
Chris@1094: //    cerr << "getSourceData(" << range.first << "," << range.second
Chris@1094: //         << "): saved range is (" << m_savedData.range.first
Chris@1094: //         << "," << m_savedData.range.second << ")" << endl;
Chris@1094: 
Chris@1100:     if (m_savedData.range == range) {
Chris@1100:         return m_savedData.data;
Chris@1100:     }
Chris@1094: 
Chris@1094:     if (range.first < m_savedData.range.second &&
Chris@1094:         range.first >= m_savedData.range.first &&
Chris@1094:         range.second > m_savedData.range.second) {
Chris@1094: 
Chris@1100:         sv_frame_t discard = range.first - m_savedData.range.first;
Chris@1100: 
Chris@1100:         vector<float> acc(m_savedData.data.begin() + discard,
Chris@1100:                           m_savedData.data.end());
Chris@1094: 
Chris@1095:         vector<float> rest =
Chris@1095:             getSourceDataUncached({ m_savedData.range.second, range.second });
Chris@1100: 
Chris@1100:         acc.insert(acc.end(), rest.begin(), rest.end());
Chris@1094:         
Chris@1095:         m_savedData = { range, acc };
Chris@1095:         return acc;
Chris@1095: 
Chris@1095:     } else {
Chris@1095: 
Chris@1095:         auto data = getSourceDataUncached(range);
Chris@1095:         m_savedData = { range, data };
Chris@1095:         return data;
Chris@1094:     }
Chris@1095: }
Chris@1094: 
Chris@1095: vector<float>
Chris@1095: FFTModel::getSourceDataUncached(pair<sv_frame_t, sv_frame_t> range) const
Chris@1095: {
Chris@1091:     decltype(range.first) pfx = 0;
Chris@1091:     if (range.first < 0) {
Chris@1091:         pfx = -range.first;
Chris@1091:         range = { 0, range.second };
Chris@1091:     }
Chris@1096: 
Chris@1096:     auto data = m_model->getData(m_channel,
Chris@1096:                                  range.first,
Chris@1096:                                  range.second - range.first);
Chris@1096: 
Chris@1096:     // don't return a partial frame
Chris@1096:     data.resize(range.second - range.first, 0.f);
Chris@1096: 
Chris@1096:     if (pfx > 0) {
Chris@1096:         vector<float> pad(pfx, 0.f);
Chris@1096:         data.insert(data.begin(), pad.begin(), pad.end());
Chris@1096:     }
Chris@1096:     
Chris@1091:     if (m_channel == -1) {
Chris@1091: 	int channels = m_model->getChannelCount();
Chris@1091: 	if (channels > 1) {
Chris@1096:             int n = int(data.size());
Chris@1096:             float factor = 1.f / float(channels);
Chris@1100:             // use mean instead of sum for fft model input
Chris@1096: 	    for (int i = 0; i < n; ++i) {
Chris@1096: 		data[i] *= factor;
Chris@1091: 	    }
Chris@1091: 	}
Chris@1091:     }
Chris@1094:     
Chris@1094:     return data;
Chris@1091: }
Chris@1091: 
Chris@1091: vector<complex<float>>
Chris@1093: FFTModel::getFFTColumn(int n) const
Chris@1091: {
Chris@1093:     for (auto &incache : m_cached) {
Chris@1093:         if (incache.n == n) {
Chris@1093:             return incache.col;
Chris@1093:         }
Chris@1093:     }
Chris@1093:     
Chris@1093:     auto samples = getSourceSamples(n);
Chris@1100:     m_windower.cut(samples.data());
Chris@1093:     auto col = m_fft.process(samples);
Chris@1093: 
Chris@1093:     SavedColumn sc { n, col };
Chris@1093:     if (m_cached.size() >= m_cacheSize) {
Chris@1093:         m_cached.pop_front();
Chris@1093:     }
Chris@1093:     m_cached.push_back(sc);
Chris@1093: 
Chris@1154:     return move(col);
Chris@1091: }
Chris@1091: 
Chris@275: bool
Chris@1045: FFTModel::estimateStableFrequency(int x, int y, double &frequency)
Chris@275: {
Chris@275:     if (!isOK()) return false;
Chris@275: 
Chris@1090:     frequency = double(y * getSampleRate()) / m_fftSize;
Chris@275: 
Chris@275:     if (x+1 >= getWidth()) return false;
Chris@275: 
Chris@275:     // At frequency f, a phase shift of 2pi (one cycle) happens in 1/f sec.
Chris@275:     // At hopsize h and sample rate sr, one hop happens in h/sr sec.
Chris@275:     // At window size w, for bin b, f is b*sr/w.
Chris@275:     // thus 2pi phase shift happens in w/(b*sr) sec.
Chris@275:     // We need to know what phase shift we expect from h/sr sec.
Chris@275:     // -> 2pi * ((h/sr) / (w/(b*sr)))
Chris@275:     //  = 2pi * ((h * b * sr) / (w * sr))
Chris@275:     //  = 2pi * (h * b) / w.
Chris@275: 
Chris@1038:     double oldPhase = getPhaseAt(x, y);
Chris@1038:     double newPhase = getPhaseAt(x+1, y);
Chris@275: 
Chris@929:     int incr = getResolution();
Chris@275: 
Chris@1090:     double expectedPhase = oldPhase + (2.0 * M_PI * y * incr) / m_fftSize;
Chris@275: 
Chris@1038:     double phaseError = princarg(newPhase - expectedPhase);
Chris@275: 
Chris@275:     // The new frequency estimate based on the phase error resulting
Chris@275:     // from assuming the "native" frequency of this bin
Chris@275: 
Chris@275:     frequency =
Chris@1090:         (getSampleRate() * (expectedPhase + phaseError - oldPhase)) /
Chris@1045:         (2.0 * M_PI * incr);
Chris@275: 
Chris@275:     return true;
Chris@275: }
Chris@275: 
Chris@275: FFTModel::PeakLocationSet
Chris@929: FFTModel::getPeaks(PeakPickType type, int x, int ymin, int ymax)
Chris@275: {
Chris@551:     Profiler profiler("FFTModel::getPeaks");
Chris@551: 
Chris@275:     FFTModel::PeakLocationSet peaks;
Chris@275:     if (!isOK()) return peaks;
Chris@275: 
Chris@275:     if (ymax == 0 || ymax > getHeight() - 1) {
Chris@275:         ymax = getHeight() - 1;
Chris@275:     }
Chris@275: 
Chris@275:     if (type == AllPeaks) {
Chris@551:         int minbin = ymin;
Chris@551:         if (minbin > 0) minbin = minbin - 1;
Chris@551:         int maxbin = ymax;
Chris@551:         if (maxbin < getHeight() - 1) maxbin = maxbin + 1;
Chris@551:         const int n = maxbin - minbin + 1;
Chris@608: #ifdef __GNUC__
Chris@551:         float values[n];
Chris@608: #else
Chris@608:         float *values = (float *)alloca(n * sizeof(float));
Chris@608: #endif
Chris@551:         getMagnitudesAt(x, values, minbin, maxbin - minbin + 1);
Chris@929:         for (int bin = ymin; bin <= ymax; ++bin) {
Chris@551:             if (bin == minbin || bin == maxbin) continue;
Chris@551:             if (values[bin - minbin] > values[bin - minbin - 1] &&
Chris@551:                 values[bin - minbin] > values[bin - minbin + 1]) {
Chris@275:                 peaks.insert(bin);
Chris@275:             }
Chris@275:         }
Chris@275:         return peaks;
Chris@275:     }
Chris@275: 
Chris@551:     Column values = getColumn(x);
Chris@1154:     int nv = int(values.size());
Chris@275: 
Chris@500:     float mean = 0.f;
Chris@1154:     for (int i = 0; i < nv; ++i) mean += values[i];
Chris@1154:     if (nv > 0) mean = mean / float(values.size());
Chris@1038:     
Chris@275:     // For peak picking we use a moving median window, picking the
Chris@275:     // highest value within each continuous region of values that
Chris@275:     // exceed the median.  For pitch adaptivity, we adjust the window
Chris@275:     // size to a roughly constant pitch range (about four tones).
Chris@275: 
Chris@1040:     sv_samplerate_t sampleRate = getSampleRate();
Chris@275: 
Chris@1090:     deque<float> window;
Chris@1090:     vector<int> inrange;
Chris@280:     float dist = 0.5;
Chris@500: 
Chris@929:     int medianWinSize = getPeakPickWindowSize(type, sampleRate, ymin, dist);
Chris@929:     int halfWin = medianWinSize/2;
Chris@275: 
Chris@929:     int binmin;
Chris@275:     if (ymin > halfWin) binmin = ymin - halfWin;
Chris@275:     else binmin = 0;
Chris@275: 
Chris@929:     int binmax;
Chris@1154:     if (ymax + halfWin < nv) binmax = ymax + halfWin;
Chris@1154:     else binmax = nv - 1;
Chris@275: 
Chris@929:     int prevcentre = 0;
Chris@500: 
Chris@929:     for (int bin = binmin; bin <= binmax; ++bin) {
Chris@275: 
Chris@275:         float value = values[bin];
Chris@275: 
Chris@275:         window.push_back(value);
Chris@275: 
Chris@280:         // so-called median will actually be the dist*100'th percentile
Chris@280:         medianWinSize = getPeakPickWindowSize(type, sampleRate, bin, dist);
Chris@275:         halfWin = medianWinSize/2;
Chris@275: 
Chris@929:         while ((int)window.size() > medianWinSize) {
Chris@500:             window.pop_front();
Chris@500:         }
Chris@500: 
Chris@1038:         int actualSize = int(window.size());
Chris@275: 
Chris@275:         if (type == MajorPitchAdaptivePeaks) {
Chris@1154:             if (ymax + halfWin < nv) binmax = ymax + halfWin;
Chris@1154:             else binmax = nv - 1;
Chris@275:         }
Chris@275: 
Chris@1090:         deque<float> sorted(window);
Chris@1090:         sort(sorted.begin(), sorted.end());
Chris@1038:         float median = sorted[int(float(sorted.size()) * dist)];
Chris@275: 
Chris@929:         int centrebin = 0;
Chris@500:         if (bin > actualSize/2) centrebin = bin - actualSize/2;
Chris@500:         
Chris@500:         while (centrebin > prevcentre || bin == binmin) {
Chris@275: 
Chris@500:             if (centrebin > prevcentre) ++prevcentre;
Chris@500: 
Chris@500:             float centre = values[prevcentre];
Chris@500: 
Chris@500:             if (centre > median) {
Chris@500:                 inrange.push_back(centrebin);
Chris@500:             }
Chris@500: 
Chris@1154:             if (centre <= median || centrebin+1 == nv) {
Chris@500:                 if (!inrange.empty()) {
Chris@929:                     int peakbin = 0;
Chris@500:                     float peakval = 0.f;
Chris@929:                     for (int i = 0; i < (int)inrange.size(); ++i) {
Chris@500:                         if (i == 0 || values[inrange[i]] > peakval) {
Chris@500:                             peakval = values[inrange[i]];
Chris@500:                             peakbin = inrange[i];
Chris@500:                         }
Chris@500:                     }
Chris@500:                     inrange.clear();
Chris@500:                     if (peakbin >= ymin && peakbin <= ymax) {
Chris@500:                         peaks.insert(peakbin);
Chris@275:                     }
Chris@275:                 }
Chris@275:             }
Chris@500: 
Chris@500:             if (bin == binmin) break;
Chris@275:         }
Chris@275:     }
Chris@275: 
Chris@275:     return peaks;
Chris@275: }
Chris@275: 
Chris@929: int
Chris@1040: FFTModel::getPeakPickWindowSize(PeakPickType type, sv_samplerate_t sampleRate,
Chris@929:                                 int bin, float &percentile) const
Chris@275: {
Chris@280:     percentile = 0.5;
Chris@275:     if (type == MajorPeaks) return 10;
Chris@275:     if (bin == 0) return 3;
Chris@280: 
Chris@1091:     double binfreq = (sampleRate * bin) / m_fftSize;
Chris@1038:     double hifreq = Pitch::getFrequencyForPitch(73, 0, binfreq);
Chris@280: 
Chris@1091:     int hibin = int(lrint((hifreq * m_fftSize) / sampleRate));
Chris@275:     int medianWinSize = hibin - bin;
Chris@275:     if (medianWinSize < 3) medianWinSize = 3;
Chris@280: 
Chris@1091:     percentile = 0.5f + float(binfreq / sampleRate);
Chris@280: 
Chris@275:     return medianWinSize;
Chris@275: }
Chris@275: 
Chris@275: FFTModel::PeakSet
Chris@929: FFTModel::getPeakFrequencies(PeakPickType type, int x,
Chris@929:                              int ymin, int ymax)
Chris@275: {
Chris@551:     Profiler profiler("FFTModel::getPeakFrequencies");
Chris@551: 
Chris@275:     PeakSet peaks;
Chris@275:     if (!isOK()) return peaks;
Chris@275:     PeakLocationSet locations = getPeaks(type, x, ymin, ymax);
Chris@275: 
Chris@1040:     sv_samplerate_t sampleRate = getSampleRate();
Chris@929:     int incr = getResolution();
Chris@275: 
Chris@275:     // This duplicates some of the work of estimateStableFrequency to
Chris@275:     // allow us to retrieve the phases in two separate vertical
Chris@275:     // columns, instead of jumping back and forth between columns x and
Chris@275:     // x+1, which may be significantly slower if re-seeking is needed
Chris@275: 
Chris@1090:     vector<float> phases;
Chris@275:     for (PeakLocationSet::iterator i = locations.begin();
Chris@275:          i != locations.end(); ++i) {
Chris@275:         phases.push_back(getPhaseAt(x, *i));
Chris@275:     }
Chris@275: 
Chris@929:     int phaseIndex = 0;
Chris@275:     for (PeakLocationSet::iterator i = locations.begin();
Chris@275:          i != locations.end(); ++i) {
Chris@1038:         double oldPhase = phases[phaseIndex];
Chris@1038:         double newPhase = getPhaseAt(x+1, *i);
Chris@1090:         double expectedPhase = oldPhase + (2.0 * M_PI * *i * incr) / m_fftSize;
Chris@1038:         double phaseError = princarg(newPhase - expectedPhase);
Chris@1038:         double frequency =
Chris@275:             (sampleRate * (expectedPhase + phaseError - oldPhase))
Chris@275:             / (2 * M_PI * incr);
Chris@1045:         peaks[*i] = frequency;
Chris@275:         ++phaseIndex;
Chris@275:     }
Chris@275: 
Chris@275:     return peaks;
Chris@275: }
Chris@275: