cannam@56: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
cannam@56: 
cannam@56: /*
cannam@56:     Vamp
cannam@56: 
cannam@56:     An API for audio analysis and feature extraction plugins.
cannam@56: 
cannam@56:     Centre for Digital Music, Queen Mary, University of London.
cannam@56:     Copyright 2006 Chris Cannam.
cannam@56:   
cannam@56:     Permission is hereby granted, free of charge, to any person
cannam@56:     obtaining a copy of this software and associated documentation
cannam@56:     files (the "Software"), to deal in the Software without
cannam@56:     restriction, including without limitation the rights to use, copy,
cannam@56:     modify, merge, publish, distribute, sublicense, and/or sell copies
cannam@56:     of the Software, and to permit persons to whom the Software is
cannam@56:     furnished to do so, subject to the following conditions:
cannam@56: 
cannam@56:     The above copyright notice and this permission notice shall be
cannam@56:     included in all copies or substantial portions of the Software.
cannam@56: 
cannam@56:     THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
cannam@56:     EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
cannam@56:     MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
cannam@56:     NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR
cannam@56:     ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF
cannam@56:     CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
cannam@56:     WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
cannam@56: 
cannam@56:     Except as contained in this notice, the names of the Centre for
cannam@56:     Digital Music; Queen Mary, University of London; and Chris Cannam
cannam@56:     shall not be used in advertising or otherwise to promote the sale,
cannam@56:     use or other dealings in this Software without prior written
cannam@56:     authorization.
cannam@56: */
cannam@56: 
cannam@56: #include "PluginInputDomainAdapter.h"
cannam@56: 
cannam@56: #include <cmath>
cannam@56: 
cannam@56: namespace Vamp {
cannam@56: 
cannam@59: namespace HostExt {
cannam@59: 
cannam@56: PluginInputDomainAdapter::PluginInputDomainAdapter(Plugin *plugin) :
cannam@57:     PluginWrapper(plugin),
cannam@56:     m_channels(0),
cannam@56:     m_blockSize(0),
cannam@56:     m_freqbuf(0)
cannam@56: {
cannam@56: }
cannam@56: 
cannam@56: PluginInputDomainAdapter::~PluginInputDomainAdapter()
cannam@56: {
cannam@56: }
cannam@56:     
cannam@56: bool
cannam@56: PluginInputDomainAdapter::initialise(size_t channels, size_t stepSize, size_t blockSize)
cannam@56: {
cannam@61:     if (m_plugin->getInputDomain() == TimeDomain) {
cannam@56: 
cannam@61:         m_blockSize = blockSize;
cannam@61:         m_channels = channels;
cannam@61: 
cannam@61:         return m_plugin->initialise(channels, stepSize, blockSize);
cannam@56:     }
cannam@56: 
cannam@61:     if (blockSize < 2) {
cannam@61:         std::cerr << "ERROR: Vamp::HostExt::PluginInputDomainAdapter::initialise: blocksize < 2 not supported" << std::endl;
cannam@61:         return false;
cannam@61:     }                
cannam@61:         
cannam@61:     if (blockSize & (blockSize-1)) {
cannam@61:         std::cerr << "ERROR: Vamp::HostExt::PluginInputDomainAdapter::initialise: non-power-of-two\nblocksize " << blockSize << " not supported" << std::endl;
cannam@61:         return false;
cannam@61:     }
cannam@61: 
cannam@61:     if (m_channels > 0) {
cannam@61:         for (size_t c = 0; c < m_channels; ++c) {
cannam@61:             delete[] m_freqbuf[c];
cannam@61:         }
cannam@61:         delete[] m_freqbuf;
cannam@61:         delete[] m_ri;
cannam@61:         delete[] m_ro;
cannam@61:         delete[] m_io;
cannam@61:     }
cannam@61: 
cannam@61:     m_blockSize = blockSize;
cannam@56:     m_channels = channels;
cannam@56: 
cannam@61:     m_freqbuf = new float *[m_channels];
cannam@61:     for (size_t c = 0; c < m_channels; ++c) {
cannam@61:         m_freqbuf[c] = new float[m_blockSize + 2];
cannam@56:     }
cannam@61:     m_ri = new double[m_blockSize];
cannam@61:     m_ro = new double[m_blockSize];
cannam@61:     m_io = new double[m_blockSize];
cannam@56: 
cannam@56:     return m_plugin->initialise(channels, stepSize, blockSize);
cannam@56: }
cannam@56: 
cannam@57: Plugin::InputDomain
cannam@57: PluginInputDomainAdapter::getInputDomain() const
cannam@56: {
cannam@57:     return TimeDomain;
cannam@56: }
cannam@56: 
cannam@56: size_t
cannam@56: PluginInputDomainAdapter::getPreferredStepSize() const
cannam@56: {
cannam@56:     size_t step = m_plugin->getPreferredStepSize();
cannam@56: 
cannam@56:     if (step == 0 && (m_plugin->getInputDomain() == FrequencyDomain)) {
cannam@56:         step = getPreferredBlockSize() / 2;
cannam@56:     }
cannam@56: 
cannam@56:     return step;
cannam@56: }
cannam@56: 
cannam@56: size_t
cannam@56: PluginInputDomainAdapter::getPreferredBlockSize() const
cannam@56: {
cannam@56:     size_t block = m_plugin->getPreferredBlockSize();
cannam@56: 
cannam@61:     if (m_plugin->getInputDomain() == FrequencyDomain) {
cannam@61:         if (block == 0) {
cannam@61:             block = 1024;
cannam@61:         } else {
cannam@61:             block = makeBlockSizeAcceptable(block);
cannam@61:         }
cannam@56:     }
cannam@56: 
cannam@56:     return block;
cannam@56: }
cannam@56: 
cannam@61: size_t
cannam@61: PluginInputDomainAdapter::makeBlockSizeAcceptable(size_t blockSize) const
cannam@61: {
cannam@61:     if (blockSize < 2) {
cannam@61: 
cannam@61:         std::cerr << "WARNING: Vamp::HostExt::PluginInputDomainAdapter::initialise: blocksize < 2 not" << std::endl
cannam@61:                   << "supported, increasing from " << blockSize << " to 2" << std::endl;
cannam@61:         blockSize = 2;
cannam@61:         
cannam@61:     } else if (blockSize & (blockSize-1)) {
cannam@61:             
cannam@61:         // not a power of two, can't handle that with our current fft
cannam@61:         // implementation
cannam@61: 
cannam@61:         size_t nearest = blockSize;
cannam@61:         size_t power = 0;
cannam@61:         while (nearest > 1) {
cannam@61:             nearest >>= 1;
cannam@61:             ++power;
cannam@61:         }
cannam@61:         nearest = 1;
cannam@61:         while (power) {
cannam@61:             nearest <<= 1;
cannam@61:             --power;
cannam@61:         }
cannam@61:         
cannam@61:         if (blockSize - nearest > (nearest*2) - blockSize) {
cannam@61:             nearest = nearest*2;
cannam@61:         }
cannam@61:         
cannam@61:         std::cerr << "WARNING: Vamp::HostExt::PluginInputDomainAdapter::initialise: non-power-of-two\nblocksize " << blockSize << " not supported, using blocksize " << nearest << " instead" << std::endl;
cannam@61:         blockSize = nearest;
cannam@61:     }
cannam@61: 
cannam@61:     return blockSize;
cannam@61: }
cannam@61: 
cannam@56: Plugin::FeatureSet
cannam@56: PluginInputDomainAdapter::process(const float *const *inputBuffers, RealTime timestamp)
cannam@56: {
cannam@56:     if (m_plugin->getInputDomain() == TimeDomain) {
cannam@56:         return m_plugin->process(inputBuffers, timestamp);
cannam@56:     }
cannam@56: 
cannam@61:     // The timestamp supplied should be (according to the Vamp::Plugin
cannam@61:     // spec) the time of the start of the time-domain input block.
cannam@61:     // However, we want to pass to the plugin an FFT output calculated
cannam@61:     // from the block of samples _centred_ on that timestamp.
cannam@61:     // 
cannam@61:     // We have two options:
cannam@61:     // 
cannam@61:     // 1. Buffer the input, calculating the fft of the values at the
cannam@61:     // passed-in block minus blockSize/2 rather than starting at the
cannam@61:     // passed-in block.  So each time we call process on the plugin,
cannam@61:     // we are passing in the same timestamp as was passed to our own
cannam@61:     // process plugin, but not (the frequency domain representation
cannam@61:     // of) the same set of samples.  Advantages: avoids confusion in
cannam@61:     // the host by ensuring the returned values have timestamps
cannam@61:     // comparable with that passed in to this function (in fact this
cannam@61:     // is pretty much essential for one-value-per-block outputs);
cannam@61:     // consistent with hosts such as SV that deal with the
cannam@61:     // frequency-domain transform themselves.  Disadvantages: means
cannam@61:     // making the not necessarily correct assumption that the samples
cannam@61:     // preceding the first official block are all zero (or some other
cannam@61:     // known value).
cannam@61:     //
cannam@61:     // 2. Increase the passed-in timestamps by half the blocksize.  So
cannam@61:     // when we call process, we are passing in the frequency domain
cannam@61:     // representation of the same set of samples as passed to us, but
cannam@61:     // with a different timestamp.  Advantages: simplicity; avoids
cannam@61:     // iffy assumption mentioned above.  Disadvantages: inconsistency
cannam@61:     // with SV in cases where stepSize != blockSize/2; potential
cannam@61:     // confusion arising from returned timestamps being calculated
cannam@61:     // from the adjusted input timestamps rather than the original
cannam@61:     // ones (and inaccuracy where the returned timestamp is implied,
cannam@61:     // as in one-value-per-block).
cannam@61:     //
cannam@61:     // Neither way is ideal, but I don't think either is strictly
cannam@61:     // incorrect either.  I think this is just a case where the same
cannam@61:     // plugin can legitimately produce differing results from the same
cannam@61:     // input data, depending on how that data is packaged.
cannam@61:     // 
cannam@61:     // We'll go for option 2, adjusting the timestamps.  Note in
cannam@61:     // particular that this means some results can differ from those
cannam@61:     // produced by SV.
cannam@61: 
cannam@61:     std::cerr << "PluginInputDomainAdapter: sampleRate " << m_inputSampleRate << ", blocksize " << m_blockSize << ", adjusting time from " << timestamp;
cannam@61: 
cannam@61:     timestamp = timestamp + RealTime::frame2RealTime(m_blockSize/2,
cannam@61:                                                      m_inputSampleRate);
cannam@61: 
cannam@61:     std::cerr << " to " << timestamp << std::endl;
cannam@60: 
cannam@56:     for (size_t c = 0; c < m_channels; ++c) {
cannam@56: 
cannam@56:         for (size_t i = 0; i < m_blockSize; ++i) {
cannam@56:             // Hanning window
cannam@56:             m_ri[i] = double(inputBuffers[c][i])
cannam@56:                 * (0.50 - 0.50 * cos((2 * M_PI * i)
cannam@56:                                      / m_blockSize));
cannam@56:         }
cannam@56: 
cannam@56:         for (size_t i = 0; i < m_blockSize/2; ++i) {
cannam@56:             // FFT shift
cannam@56:             double value = m_ri[i];
cannam@56:             m_ri[i] = m_ri[i + m_blockSize/2];
cannam@56:             m_ri[i + m_blockSize/2] = value;
cannam@56:         }
cannam@56: 
cannam@56:         fft(m_blockSize, false, m_ri, 0, m_ro, m_io);
cannam@56: 
cannam@61:         for (size_t i = 0; i <= m_blockSize/2; ++i) {
cannam@56:             m_freqbuf[c][i * 2] = m_ro[i];
cannam@56:             m_freqbuf[c][i * 2 + 1] = m_io[i];
cannam@56:         }
cannam@56:     }
cannam@56: 
cannam@56:     return m_plugin->process(m_freqbuf, timestamp);
cannam@56: }
cannam@56: 
cannam@56: void
cannam@56: PluginInputDomainAdapter::fft(unsigned int n, bool inverse,
cannam@56:                               double *ri, double *ii, double *ro, double *io)
cannam@56: {
cannam@56:     if (!ri || !ro || !io) return;
cannam@56: 
cannam@56:     unsigned int bits;
cannam@56:     unsigned int i, j, k, m;
cannam@56:     unsigned int blockSize, blockEnd;
cannam@56: 
cannam@56:     double tr, ti;
cannam@56: 
cannam@56:     if (n < 2) return;
cannam@56:     if (n & (n-1)) return;
cannam@56: 
cannam@56:     double angle = 2.0 * M_PI;
cannam@56:     if (inverse) angle = -angle;
cannam@56: 
cannam@56:     for (i = 0; ; ++i) {
cannam@56: 	if (n & (1 << i)) {
cannam@56: 	    bits = i;
cannam@56: 	    break;
cannam@56: 	}
cannam@56:     }
cannam@56: 
cannam@56:     static unsigned int tableSize = 0;
cannam@56:     static int *table = 0;
cannam@56: 
cannam@56:     if (tableSize != n) {
cannam@56: 
cannam@56: 	delete[] table;
cannam@56: 
cannam@56: 	table = new int[n];
cannam@56: 
cannam@56: 	for (i = 0; i < n; ++i) {
cannam@56: 	
cannam@56: 	    m = i;
cannam@56: 
cannam@56: 	    for (j = k = 0; j < bits; ++j) {
cannam@56: 		k = (k << 1) | (m & 1);
cannam@56: 		m >>= 1;
cannam@56: 	    }
cannam@56: 
cannam@56: 	    table[i] = k;
cannam@56: 	}
cannam@56: 
cannam@56: 	tableSize = n;
cannam@56:     }
cannam@56: 
cannam@56:     if (ii) {
cannam@56: 	for (i = 0; i < n; ++i) {
cannam@56: 	    ro[table[i]] = ri[i];
cannam@56: 	    io[table[i]] = ii[i];
cannam@56: 	}
cannam@56:     } else {
cannam@56: 	for (i = 0; i < n; ++i) {
cannam@56: 	    ro[table[i]] = ri[i];
cannam@56: 	    io[table[i]] = 0.0;
cannam@56: 	}
cannam@56:     }
cannam@56: 
cannam@56:     blockEnd = 1;
cannam@56: 
cannam@56:     for (blockSize = 2; blockSize <= n; blockSize <<= 1) {
cannam@56: 
cannam@56: 	double delta = angle / (double)blockSize;
cannam@56: 	double sm2 = -sin(-2 * delta);
cannam@56: 	double sm1 = -sin(-delta);
cannam@56: 	double cm2 = cos(-2 * delta);
cannam@56: 	double cm1 = cos(-delta);
cannam@56: 	double w = 2 * cm1;
cannam@56: 	double ar[3], ai[3];
cannam@56: 
cannam@56: 	for (i = 0; i < n; i += blockSize) {
cannam@56: 
cannam@56: 	    ar[2] = cm2;
cannam@56: 	    ar[1] = cm1;
cannam@56: 
cannam@56: 	    ai[2] = sm2;
cannam@56: 	    ai[1] = sm1;
cannam@56: 
cannam@56: 	    for (j = i, m = 0; m < blockEnd; j++, m++) {
cannam@56: 
cannam@56: 		ar[0] = w * ar[1] - ar[2];
cannam@56: 		ar[2] = ar[1];
cannam@56: 		ar[1] = ar[0];
cannam@56: 
cannam@56: 		ai[0] = w * ai[1] - ai[2];
cannam@56: 		ai[2] = ai[1];
cannam@56: 		ai[1] = ai[0];
cannam@56: 
cannam@56: 		k = j + blockEnd;
cannam@56: 		tr = ar[0] * ro[k] - ai[0] * io[k];
cannam@56: 		ti = ar[0] * io[k] + ai[0] * ro[k];
cannam@56: 
cannam@56: 		ro[k] = ro[j] - tr;
cannam@56: 		io[k] = io[j] - ti;
cannam@56: 
cannam@56: 		ro[j] += tr;
cannam@56: 		io[j] += ti;
cannam@56: 	    }
cannam@56: 	}
cannam@56: 
cannam@56: 	blockEnd = blockSize;
cannam@56:     }
cannam@56: 
cannam@56:     if (inverse) {
cannam@56: 
cannam@56: 	double denom = (double)n;
cannam@56: 
cannam@56: 	for (i = 0; i < n; i++) {
cannam@56: 	    ro[i] /= denom;
cannam@56: 	    io[i] /= denom;
cannam@56: 	}
cannam@56:     }
cannam@56: }
cannam@56: 
cannam@59: }
cannam@56:         
cannam@56: }
cannam@56: