Chris@0: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
Chris@0: 
Chris@0: /*
Chris@0:     Sonic Visualiser
Chris@0:     An audio file viewer and annotation editor.
Chris@0:     Centre for Digital Music, Queen Mary, University of London.
Chris@0:     This file copyright 2006 Chris Cannam.
Chris@0:     
Chris@0:     This program is free software; you can redistribute it and/or
Chris@0:     modify it under the terms of the GNU General Public License as
Chris@0:     published by the Free Software Foundation; either version 2 of the
Chris@0:     License, or (at your option) any later version.  See the file
Chris@0:     COPYING included with this distribution for more information.
Chris@0: */
Chris@0: 
Chris@14: #include "PhaseVocoderTimeStretcher.h"
Chris@0: 
Chris@0: #include <iostream>
Chris@0: #include <cassert>
Chris@0: 
Chris@14: //#define DEBUG_PHASE_VOCODER_TIME_STRETCHER 1
Chris@0: 
Chris@16: PhaseVocoderTimeStretcher::PhaseVocoderTimeStretcher(size_t channels,
Chris@16:                                                      float ratio,
Chris@16:                                                      bool sharpen,
Chris@15:                                                      size_t maxProcessInputBlockSize) :
Chris@16:     m_channels(channels),
Chris@16:     m_ratio(ratio),
Chris@16:     m_sharpen(sharpen)
Chris@0: {
Chris@16:     m_wlen = 1024;
Chris@16: 
Chris@15:     if (ratio < 1) {
Chris@16:         if (ratio < 0.4) {
Chris@16:             m_n1 = 1024;
Chris@16:             m_wlen = 2048;
Chris@16:         } else if (ratio < 0.8) {
Chris@16:             m_n1 = 512;
Chris@16:         } else {
Chris@16:             m_n1 = 256;
Chris@16:         }
Chris@16:         if (m_sharpen) {
Chris@17: //            m_n1 /= 2;
Chris@16:             m_wlen = 2048;
Chris@16:         }
Chris@15:         m_n2 = m_n1 * ratio;
Chris@15:     } else {
Chris@16:         if (ratio > 2) {
Chris@16:             m_n2 = 512;
Chris@16:             m_wlen = 4096; 
Chris@16:         } else if (ratio > 1.6) {
Chris@16:             m_n2 = 384;
Chris@16:             m_wlen = 2048;
Chris@16:         } else {
Chris@16:             m_n2 = 256;
Chris@16:         }
Chris@16:         if (m_sharpen) {
Chris@17: //            m_n2 /= 2;
Chris@16:             if (m_wlen < 2048) m_wlen = 2048;
Chris@16:         }
Chris@15:         m_n1 = m_n2 / ratio;
Chris@15:     }
Chris@16:         
Chris@20:     m_analysisWindow = new Window<float>(HanningWindow, m_wlen);
Chris@20:     m_synthesisWindow = new Window<float>(HanningWindow, m_wlen);
Chris@15: 
Chris@16:     m_prevPhase = new float *[m_channels];
Chris@16:     m_prevAdjustedPhase = new float *[m_channels];
Chris@15: 
Chris@20:     m_prevTransientMag = (float *)fftwf_malloc(sizeof(float) * (m_wlen / 2 + 1));
Chris@20:     m_prevTransientCount = 0;
Chris@20:     m_prevTransient = false;
Chris@20: 
Chris@20:     m_tempbuf = (float *)fftwf_malloc(sizeof(float) * m_wlen);
Chris@20: 
Chris@20:     m_time = new float *[m_channels];
Chris@20:     m_freq = new fftwf_complex *[m_channels];
Chris@20:     m_plan = new fftwf_plan[m_channels];
Chris@20:     m_iplan = new fftwf_plan[m_channels];
Chris@0: 
Chris@16:     m_inbuf = new RingBuffer<float> *[m_channels];
Chris@16:     m_outbuf = new RingBuffer<float> *[m_channels];
Chris@16:     m_mashbuf = new float *[m_channels];
Chris@16: 
Chris@16:     m_modulationbuf = (float *)fftwf_malloc(sizeof(float) * m_wlen);
Chris@16:         
Chris@16:     for (size_t c = 0; c < m_channels; ++c) {
Chris@16: 
Chris@20:         m_prevPhase[c] = (float *)fftwf_malloc(sizeof(float) * (m_wlen / 2 + 1));
Chris@20:         m_prevAdjustedPhase[c] = (float *)fftwf_malloc(sizeof(float) * (m_wlen / 2 + 1));
Chris@16: 
Chris@20:         m_time[c] = (float *)fftwf_malloc(sizeof(float) * m_wlen);
Chris@20:         m_freq[c] = (fftwf_complex *)fftwf_malloc(sizeof(fftwf_complex) *
Chris@20:                                                   (m_wlen / 2 + 1));
Chris@20:         
Chris@20:         m_plan[c] = fftwf_plan_dft_r2c_1d(m_wlen, m_time[c], m_freq[c], FFTW_ESTIMATE);
Chris@20:         m_iplan[c] = fftwf_plan_dft_c2r_1d(m_wlen, m_freq[c], m_time[c], FFTW_ESTIMATE);
Chris@16: 
Chris@16:         m_inbuf[c] = new RingBuffer<float>(m_wlen);
Chris@16:         m_outbuf[c] = new RingBuffer<float>
Chris@16:             (lrintf((maxProcessInputBlockSize + m_wlen) * ratio));
Chris@16:             
Chris@16:         m_mashbuf[c] = (float *)fftwf_malloc(sizeof(float) * m_wlen);
Chris@16:         
Chris@16:         for (int i = 0; i < m_wlen; ++i) {
Chris@16:             m_mashbuf[c][i] = 0.0;
Chris@20:         }
Chris@20: 
Chris@20:         for (int i = 0; i <= m_wlen/2; ++i) {
Chris@16:             m_prevPhase[c][i] = 0.0;
Chris@16:             m_prevAdjustedPhase[c][i] = 0.0;
Chris@16:         }
Chris@16:     }
Chris@16: 
Chris@0:     for (int i = 0; i < m_wlen; ++i) {
Chris@16:         m_modulationbuf[i] = 0.0;
Chris@0:     }
Chris@16: 
Chris@20:     for (int i = 0; i <= m_wlen/2; ++i) {
Chris@20:         m_prevTransientMag[i] = 0.0;
Chris@20:     }
Chris@20: 
Chris@16:     std::cerr << "PhaseVocoderTimeStretcher: channels = " << channels
Chris@16:               << ", ratio = " << ratio
Chris@16:               << ", n1 = " << m_n1 << ", n2 = " << m_n2 << ", wlen = "
Chris@16:               << m_wlen << ", max = " << maxProcessInputBlockSize
Chris@16:               << ", outbuflen = " << m_outbuf[0]->getSize() << std::endl;
Chris@0: }
Chris@0: 
Chris@14: PhaseVocoderTimeStretcher::~PhaseVocoderTimeStretcher()
Chris@0: {
Chris@14:     std::cerr << "PhaseVocoderTimeStretcher::~PhaseVocoderTimeStretcher" << std::endl;
Chris@0: 
Chris@20:     for (size_t c = 0; c < m_channels; ++c) {
Chris@0: 
Chris@20:         fftwf_destroy_plan(m_plan[c]);
Chris@20:         fftwf_destroy_plan(m_iplan[c]);
Chris@16: 
Chris@20:         fftwf_free(m_time[c]);
Chris@20:         fftwf_free(m_freq[c]);
Chris@16: 
Chris@16:         fftwf_free(m_mashbuf[c]);
Chris@16:         fftwf_free(m_prevPhase[c]);
Chris@16:         fftwf_free(m_prevAdjustedPhase[c]);
Chris@16: 
Chris@16:         delete m_inbuf[c];
Chris@16:         delete m_outbuf[c];
Chris@16:     }
Chris@16: 
Chris@20:     fftwf_free(m_tempbuf);
Chris@13:     fftwf_free(m_modulationbuf);
Chris@20:     fftwf_free(m_prevTransientMag);
Chris@0: 
Chris@16:     delete[] m_prevPhase;
Chris@16:     delete[] m_prevAdjustedPhase;
Chris@16:     delete[] m_inbuf;
Chris@16:     delete[] m_outbuf;
Chris@16:     delete[] m_mashbuf;
Chris@20:     delete[] m_time;
Chris@20:     delete[] m_freq;
Chris@20:     delete[] m_plan;
Chris@20:     delete[] m_iplan;
Chris@15: 
Chris@20:     delete m_analysisWindow;
Chris@20:     delete m_synthesisWindow;
Chris@0: }	
Chris@0: 
Chris@0: size_t
Chris@14: PhaseVocoderTimeStretcher::getProcessingLatency() const
Chris@0: {
Chris@0:     return getWindowSize() - getInputIncrement();
Chris@0: }
Chris@0: 
Chris@0: void
Chris@16: PhaseVocoderTimeStretcher::process(float **input, float **output, size_t samples)
Chris@16: {
Chris@16:     putInput(input, samples);
Chris@16:     getOutput(output, lrintf(samples * m_ratio));
Chris@16: }
Chris@16: 
Chris@16: size_t
Chris@16: PhaseVocoderTimeStretcher::getRequiredInputSamples() const
Chris@16: {
Chris@16:     if (m_inbuf[0]->getReadSpace() >= m_wlen) return 0;
Chris@16:     return m_wlen - m_inbuf[0]->getReadSpace();
Chris@16: }
Chris@16: 
Chris@16: void
Chris@16: PhaseVocoderTimeStretcher::putInput(float **input, size_t samples)
Chris@0: {
Chris@0:     // We need to add samples from input to our internal buffer.  When
Chris@0:     // we have m_windowSize samples in the buffer, we can process it,
Chris@0:     // move the samples back by m_n1 and write the output onto our
Chris@0:     // internal output buffer.  If we have (samples * ratio) samples
Chris@0:     // in that, we can write m_n2 of them back to output and return
Chris@0:     // (otherwise we have to write zeroes).
Chris@0: 
Chris@0:     // When we process, we write m_wlen to our fixed output buffer
Chris@0:     // (m_mashbuf).  We then pull out the first m_n2 samples from that
Chris@0:     // buffer, push them into the output ring buffer, and shift
Chris@0:     // m_mashbuf left by that amount.
Chris@0: 
Chris@0:     // The processing latency is then m_wlen - m_n2.
Chris@0: 
Chris@0:     size_t consumed = 0;
Chris@0: 
Chris@0:     while (consumed < samples) {
Chris@0: 
Chris@16: 	size_t writable = m_inbuf[0]->getWriteSpace();
Chris@0: 	writable = std::min(writable, samples - consumed);
Chris@0: 
Chris@0: 	if (writable == 0) {
Chris@0: 	    //!!! then what? I don't think this should happen, but
Chris@16: 	    std::cerr << "WARNING: PhaseVocoderTimeStretcher::putInput: writable == 0" << std::endl;
Chris@0: 	    break;
Chris@0: 	}
Chris@0: 
Chris@14: #ifdef DEBUG_PHASE_VOCODER_TIME_STRETCHER
Chris@0: 	std::cerr << "writing " << writable << " from index " << consumed << " to inbuf, consumed will be " << consumed + writable << std::endl;
Chris@0: #endif
Chris@16: 
Chris@16:         for (size_t c = 0; c < m_channels; ++c) {
Chris@16:             m_inbuf[c]->write(input[c] + consumed, writable);
Chris@16:         }
Chris@0: 	consumed += writable;
Chris@0: 
Chris@16: 	while (m_inbuf[0]->getReadSpace() >= m_wlen &&
Chris@16: 	       m_outbuf[0]->getWriteSpace() >= m_n2) {
Chris@0: 
Chris@0: 	    // We know we have at least m_wlen samples available
Chris@16: 	    // in m_inbuf.  We need to peek m_wlen of them for
Chris@0: 	    // processing, and then read m_n1 to advance the read
Chris@0: 	    // pointer.
Chris@16:             
Chris@20:             for (size_t c = 0; c < m_channels; ++c) {
Chris@20: 
Chris@20:                 size_t got = m_inbuf[c]->peek(m_tempbuf, m_wlen);
Chris@20:                 assert(got == m_wlen);
Chris@20: 
Chris@20:                 analyseBlock(c, m_tempbuf);
Chris@20:             }
Chris@20: 
Chris@20:             bool transient = false;
Chris@20:             if (m_sharpen) transient = isTransient();
Chris@20: 
Chris@16:             size_t n2 = m_n2;
Chris@20: 
Chris@20:             if (transient) {
Chris@20:                 n2 = m_n1;
Chris@20:             }
Chris@0: 
Chris@16:             for (size_t c = 0; c < m_channels; ++c) {
Chris@16: 
Chris@20:                 synthesiseBlock(c, m_mashbuf[c],
Chris@20:                                 c == 0 ? m_modulationbuf : 0,
Chris@20:                                 m_prevTransient ? m_n1 : m_n2);
Chris@16: 
Chris@0: 
Chris@14: #ifdef DEBUG_PHASE_VOCODER_TIME_STRETCHER
Chris@16:                 std::cerr << "writing first " << m_n2 << " from mashbuf, skipping " << m_n1 << " on inbuf " << std::endl;
Chris@0: #endif
Chris@16:                 m_inbuf[c]->skip(m_n1);
Chris@13: 
Chris@16:                 for (size_t i = 0; i < n2; ++i) {
Chris@16:                     if (m_modulationbuf[i] > 0.f) {
Chris@16:                         m_mashbuf[c][i] /= m_modulationbuf[i];
Chris@16:                     }
Chris@16:                 }
Chris@16: 
Chris@16:                 m_outbuf[c]->write(m_mashbuf[c], n2);
Chris@16: 
Chris@16:                 for (size_t i = 0; i < m_wlen - n2; ++i) {
Chris@16:                     m_mashbuf[c][i] = m_mashbuf[c][i + n2];
Chris@16:                 }
Chris@16: 
Chris@16:                 for (size_t i = m_wlen - n2; i < m_wlen; ++i) {
Chris@16:                     m_mashbuf[c][i] = 0.0f;
Chris@13:                 }
Chris@13:             }
Chris@13: 
Chris@20:             m_prevTransient = transient;
Chris@17: 
Chris@16:             for (size_t i = 0; i < m_wlen - n2; ++i) {
Chris@16:                 m_modulationbuf[i] = m_modulationbuf[i + n2];
Chris@0: 	    }
Chris@13: 
Chris@16: 	    for (size_t i = m_wlen - n2; i < m_wlen; ++i) {
Chris@13:                 m_modulationbuf[i] = 0.0f;
Chris@0: 	    }
Chris@0: 	}
Chris@0: 
Chris@0: 
Chris@14: #ifdef DEBUG_PHASE_VOCODER_TIME_STRETCHER
Chris@16: 	std::cerr << "loop ended: inbuf read space " << m_inbuf[0]->getReadSpace() << ", outbuf write space " << m_outbuf[0]->getWriteSpace() << std::endl;
Chris@0: #endif
Chris@0:     }
Chris@0: 
Chris@16: #ifdef DEBUG_PHASE_VOCODER_TIME_STRETCHER
Chris@16:     std::cerr << "PhaseVocoderTimeStretcher::putInput returning" << std::endl;
Chris@16: #endif
Chris@16: }
Chris@12: 
Chris@16: size_t
Chris@16: PhaseVocoderTimeStretcher::getAvailableOutputSamples() const
Chris@16: {
Chris@16:     return m_outbuf[0]->getReadSpace();
Chris@16: }
Chris@16: 
Chris@16: void
Chris@16: PhaseVocoderTimeStretcher::getOutput(float **output, size_t samples)
Chris@16: {
Chris@16:     if (m_outbuf[0]->getReadSpace() < samples) {
Chris@16: 	std::cerr << "WARNING: PhaseVocoderTimeStretcher::getOutput: not enough data (yet?) (" << m_outbuf[0]->getReadSpace() << " < " << samples << ")" << std::endl;
Chris@16: 	size_t fill = samples - m_outbuf[0]->getReadSpace();
Chris@16:         for (size_t c = 0; c < m_channels; ++c) {
Chris@16:             for (size_t i = 0; i < fill; ++i) {
Chris@16:                 output[c][i] = 0.0;
Chris@16:             }
Chris@16:             m_outbuf[c]->read(output[c] + fill, m_outbuf[c]->getReadSpace());
Chris@16:         }
Chris@0:     } else {
Chris@14: #ifdef DEBUG_PHASE_VOCODER_TIME_STRETCHER
Chris@16: 	std::cerr << "enough data - writing " << samples << " from outbuf" << std::endl;
Chris@0: #endif
Chris@16:         for (size_t c = 0; c < m_channels; ++c) {
Chris@16:             m_outbuf[c]->read(output[c], samples);
Chris@16:         }
Chris@0:     }
Chris@0: 
Chris@14: #ifdef DEBUG_PHASE_VOCODER_TIME_STRETCHER
Chris@16:     std::cerr << "PhaseVocoderTimeStretcher::getOutput returning" << std::endl;
Chris@0: #endif
Chris@0: }
Chris@0: 
Chris@20: void
Chris@20: PhaseVocoderTimeStretcher::analyseBlock(size_t c, float *buf)
Chris@0: {
Chris@0:     size_t i;
Chris@0: 
Chris@20:     // buf contains m_wlen samples
Chris@0: 
Chris@14: #ifdef DEBUG_PHASE_VOCODER_TIME_STRETCHER
Chris@20:     std::cerr << "PhaseVocoderTimeStretcher::analyseBlock (channel " << c << ")" << std::endl;
Chris@0: #endif
Chris@0: 
Chris@20:     m_analysisWindow->cut(buf);
Chris@0: 
Chris@0:     for (i = 0; i < m_wlen/2; ++i) {
Chris@0: 	float temp = buf[i];
Chris@0: 	buf[i] = buf[i + m_wlen/2];
Chris@0: 	buf[i + m_wlen/2] = temp;
Chris@0:     }
Chris@19: 
Chris@0:     for (i = 0; i < m_wlen; ++i) {
Chris@20: 	m_time[c][i] = buf[i];
Chris@0:     }
Chris@0: 
Chris@20:     fftwf_execute(m_plan[c]); // m_time -> m_freq
Chris@20: }
Chris@0: 
Chris@20: bool
Chris@20: PhaseVocoderTimeStretcher::isTransient()
Chris@20: {
Chris@20:     int count = 0;
Chris@16: 
Chris@20:     for (int i = 0; i <= m_wlen/2; ++i) {
Chris@16: 
Chris@20:         float real = 0.f, imag = 0.f;
Chris@20: 
Chris@20:         for (size_t c = 0; c < m_channels; ++c) {
Chris@20:             real += m_freq[c][i][0];
Chris@20:             imag += m_freq[c][i][1];
Chris@16:         }
Chris@16: 
Chris@20:         float sqrmag = (real * real + imag * imag);
Chris@20: 
Chris@20:         if (m_prevTransientMag[i] > 0.f) {
Chris@20:             float diff = 10.f * log10f(sqrmag / m_prevTransientMag[i]);
Chris@20:             if (diff > 3.f) ++count;
Chris@20:         }
Chris@20: 
Chris@20:         m_prevTransientMag[i] = sqrmag;
Chris@16:     }
Chris@16: 
Chris@20:     bool isTransient = false;
Chris@16: 
Chris@20:     if (count > m_wlen / 4.5 && //!!!
Chris@20:         count > m_prevTransientCount * 1.2) {
Chris@20:         isTransient = true;
Chris@20:         std::cerr << "isTransient (count = " << count << ", prev = " << m_prevTransientCount << ")" << std::endl;
Chris@20:     }
Chris@16: 
Chris@20:     m_prevTransientCount = count;
Chris@20: 
Chris@20:     return isTransient;
Chris@20: }
Chris@20: 
Chris@20: void
Chris@20: PhaseVocoderTimeStretcher::synthesiseBlock(size_t c,
Chris@20:                                            float *out,
Chris@20:                                            float *modulation,
Chris@20:                                            size_t lastStep)
Chris@20: {
Chris@20:     int i;
Chris@20: 
Chris@20:     bool unchanged = (lastStep == m_n1);
Chris@20: 
Chris@20:     for (i = 0; i <= m_wlen/2; ++i) {
Chris@0: 		
Chris@20:         float phase = princargf(atan2f(m_freq[c][i][1], m_freq[c][i][0]));
Chris@19:         float adjustedPhase = phase;
Chris@12: 
Chris@20:         if (!unchanged) {
Chris@16: 
Chris@20:             float mag = sqrtf(m_freq[c][i][0] * m_freq[c][i][0] +
Chris@20:                               m_freq[c][i][1] * m_freq[c][i][1]);
Chris@19: 
Chris@20:             float omega = (2 * M_PI * m_n1 * i) / m_wlen;
Chris@20: 	
Chris@20:             float expectedPhase = m_prevPhase[c][i] + omega;
Chris@20: 
Chris@20:             float phaseError = princargf(phase - expectedPhase);
Chris@20: 
Chris@20:             float phaseIncrement = (omega + phaseError) / m_n1;
Chris@20:             
Chris@20:             adjustedPhase = m_prevAdjustedPhase[c][i] +
Chris@20:                 lastStep * phaseIncrement;
Chris@20:             
Chris@20:             float real = mag * cosf(adjustedPhase);
Chris@20:             float imag = mag * sinf(adjustedPhase);
Chris@20:             m_freq[c][i][0] = real;
Chris@20:             m_freq[c][i][1] = imag;
Chris@19:         }
Chris@19: 
Chris@16:         m_prevPhase[c][i] = phase;
Chris@16:         m_prevAdjustedPhase[c][i] = adjustedPhase;
Chris@0:     }
Chris@20: 
Chris@20:     fftwf_execute(m_iplan[c]); // m_freq -> m_time, inverse fft
Chris@19: 
Chris@0:     for (i = 0; i < m_wlen/2; ++i) {
Chris@20:         float temp = m_time[c][i];
Chris@20:         m_time[c][i] = m_time[c][i + m_wlen/2];
Chris@20:         m_time[c][i + m_wlen/2] = temp;
Chris@20:     }
Chris@20:     
Chris@20:     for (i = 0; i < m_wlen; ++i) {
Chris@20:         m_time[c][i] = m_time[c][i] / m_wlen;
Chris@0:     }
Chris@15: 
Chris@20:     m_synthesisWindow->cut(m_time[c]);
Chris@19: 
Chris@19:     for (i = 0; i < m_wlen; ++i) {
Chris@20:         out[i] += m_time[c][i];
Chris@0:     }
Chris@16: 
Chris@16:     if (modulation) {
Chris@16: 
Chris@20:         float area = m_analysisWindow->getArea();
Chris@16: 
Chris@16:         for (i = 0; i < m_wlen; ++i) {
Chris@20:             float val = m_synthesisWindow->getValue(i);
Chris@16:             modulation[i] += val * area;
Chris@16:         }
Chris@16:     }
Chris@0: }
Chris@15: 
Chris@20: