annotate dsp/chromagram/ConstantQ.cpp @ 298:255e431ae3d4

* Key detector: when returning key strengths, use the peak value of the
  three underlying chromagram correlations (from 36-bin chromagram)
  corresponding to each key, instead of the mean.
  Rationale: This is the same method as used when returning the key value,
  and it's nice to have the same results in both returned value and plot.
  The peak performed better than the sum with a simple test set of triads,
  so it seems reasonable to change the plot to match the key output rather
  than the other way around.

* FFT: kiss_fftr returns only the non-conjugate bins, synthesise the rest
  rather than leaving them (perhaps dangerously) undefined.  Fixes an
  uninitialised data error in chromagram that could cause garbage results
  from key detector.

* Constant Q: remove precalculated values again, I reckon they're not
  proving such a good tradeoff.

/* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
/*
    QM DSP Library

    Centre for Digital Music, Queen Mary, University of London.
    This file copyright 2005-2006 Christian Landone.
    All rights reserved.
*/

#include "ConstantQ.h"
#include "dsp/transforms/FFT.h"

#include <iostream>

#ifdef NOT_DEFINED
// see note in CQprecalc

#include "CQprecalc.cpp"

static bool push_precalculated(int uk, int fftlength,
                               std::vector<unsigned> &is,
                               std::vector<unsigned> &js,
                               std::vector<double> &real,
                               std::vector<double> &imag)
{
    if (uk == 76 && fftlength == 16384) {
        push_76_16384(is, js, real, imag);
        return true;
    }
    if (uk == 144 && fftlength == 4096) {
        push_144_4096(is, js, real, imag);
        return true;
    }
    if (uk == 65 && fftlength == 2048) {
        push_65_2048(is, js, real, imag);
        return true;
    }
    if (uk == 84 && fftlength == 65536) {
        push_84_65536(is, js, real, imag);
        return true;
    }
    return false;
}
#endif

//---------------------------------------------------------------------------
// nextpow2 returns the smallest integer n such that 2^n >= x.
static double nextpow2(double x) {
    double y = ceil(log(x)/log(2.0));
    return(y);
}

static double squaredModule(const double & xx, const double & yy) {
    return xx*xx + yy*yy;
}

//----------------------------------------------------------------------------

ConstantQ::ConstantQ( CQConfig Config ) :
    m_sparseKernel(0)
{
    initialise( Config );
}

ConstantQ::~ConstantQ()
{
    deInitialise();
}

//----------------------------------------------------------------------------
void ConstantQ::sparsekernel()
{
//    std::cerr << "ConstantQ: initialising sparse kernel, uK = " << m_uK << ", FFTLength = " << m_FFTLength << "...";

    SparseKernel *sk = new SparseKernel();

#ifdef NOT_DEFINED
    if (push_precalculated(m_uK, m_FFTLength,
                           sk->is, sk->js, sk->real, sk->imag)) {
//        std::cerr << "using precalculated kernel" << std::endl;
        m_sparseKernel = sk;
        return;
    }
#endif

    //generates spectral kernel matrix (upside down?)
    // initialise temporal kernel with zeros, twice length to deal w. complex numbers

    double* hammingWindowRe = new double [ m_FFTLength ];
    double* hammingWindowIm = new double [ m_FFTLength ];
    double* transfHammingWindowRe = new double [ m_FFTLength ];
    double* transfHammingWindowIm = new double [ m_FFTLength ];

    for (unsigned u=0; u < m_FFTLength; u++) 
    {
	hammingWindowRe[u] = 0;
	hammingWindowIm[u] = 0;
    }

    // Here, fftleng*2 is a guess of the number of sparse cells in the matrix
    // The matrix K x fftlength but the non-zero cells are an antialiased
    // square root function. So mostly is a line, with some grey point.
    sk->is.reserve( m_FFTLength*2 );
    sk->js.reserve( m_FFTLength*2 );
    sk->real.reserve( m_FFTLength*2 );
    sk->imag.reserve( m_FFTLength*2 );
	
    // for each bin value K, calculate temporal kernel, take its fft to
    //calculate the spectral kernel then threshold it to make it sparse and 
    //add it to the sparse kernels matrix
    double squareThreshold = m_CQThresh * m_CQThresh;

    FFT m_FFT(m_FFTLength);
	
    for (unsigned k = m_uK; k--; ) 
    {
        for (unsigned u=0; u < m_FFTLength; u++) 
        {
            hammingWindowRe[u] = 0;
            hammingWindowIm[u] = 0;
        }
        
	// Computing a hamming window
	const unsigned hammingLength = (int) ceil( m_dQ * m_FS / ( m_FMin * pow(2,((double)(k))/(double)m_BPO)));

        unsigned origin = m_FFTLength/2 - hammingLength/2;

	for (unsigned i=0; i<hammingLength; i++) 
	{
	    const double angle = 2*PI*m_dQ*i/hammingLength;
	    const double real = cos(angle);
	    const double imag = sin(angle);
	    const double absol = hamming(hammingLength, i)/hammingLength;
	    hammingWindowRe[ origin + i ] = absol*real;
	    hammingWindowIm[ origin + i ] = absol*imag;
	}

        for (unsigned i = 0; i < m_FFTLength/2; ++i) {
            double temp = hammingWindowRe[i];
            hammingWindowRe[i] = hammingWindowRe[i + m_FFTLength/2];
            hammingWindowRe[i + m_FFTLength/2] = temp;
            temp = hammingWindowIm[i];
            hammingWindowIm[i] = hammingWindowIm[i + m_FFTLength/2];
            hammingWindowIm[i + m_FFTLength/2] = temp;
        }
    
	//do fft of hammingWindow
	m_FFT.process( 0, hammingWindowRe, hammingWindowIm, transfHammingWindowRe, transfHammingWindowIm );

		
	for (unsigned j=0; j<( m_FFTLength ); j++) 
	{
	    // perform thresholding
	    const double squaredBin = squaredModule( transfHammingWindowRe[ j ], transfHammingWindowIm[ j ]);
	    if (squaredBin <= squareThreshold) continue;
		
	    // Insert non-zero position indexes, doubled because they are floats
	    sk->is.push_back(j);
	    sk->js.push_back(k);

	    // take conjugate, normalise and add to array sparkernel
	    sk->real.push_back( transfHammingWindowRe[ j ]/m_FFTLength);
	    sk->imag.push_back(-transfHammingWindowIm[ j ]/m_FFTLength);
	}

    }

    delete [] hammingWindowRe;
    delete [] hammingWindowIm;
    delete [] transfHammingWindowRe;
    delete [] transfHammingWindowIm;

/*
    using std::cout;
    using std::endl;

    cout.precision(28);

    int n = sk->is.size();
    int w = 8;
    cout << "static unsigned int sk_i_" << m_uK << "_" << m_FFTLength << "[" << n << "] = {" << endl;
    for (int i = 0; i < n; ++i) {
        if (i % w == 0) cout << "    ";
        cout << sk->is[i];
        if (i + 1 < n) cout << ", ";
        if (i % w == w-1) cout << endl;
    };
    if (n % w != 0) cout << endl;
    cout << "};" << endl;

    n = sk->js.size();
    cout << "static unsigned int sk_j_" << m_uK << "_" << m_FFTLength << "[" << n << "] = {" << endl;
    for (int i = 0; i < n; ++i) {
        if (i % w == 0) cout << "    ";
        cout << sk->js[i];
        if (i + 1 < n) cout << ", ";
        if (i % w == w-1) cout << endl;
    };
    if (n % w != 0) cout << endl;
    cout << "};" << endl;

    w = 2;
    n = sk->real.size();
    cout << "static double sk_real_" << m_uK << "_" << m_FFTLength << "[" << n << "] = {" << endl;
    for (int i = 0; i < n; ++i) {
        if (i % w == 0) cout << "    ";
        cout << sk->real[i];
        if (i + 1 < n) cout << ", ";
        if (i % w == w-1) cout << endl;
    };
    if (n % w != 0) cout << endl;
    cout << "};" << endl;

    n = sk->imag.size();
    cout << "static double sk_imag_" << m_uK << "_" << m_FFTLength << "[" << n << "] = {" << endl;
    for (int i = 0; i < n; ++i) {
        if (i % w == 0) cout << "    ";
        cout << sk->imag[i];
        if (i + 1 < n) cout << ", ";
        if (i % w == w-1) cout << endl;
    };
    if (n % w != 0) cout << endl;
    cout << "};" << endl;

    cout << "static void push_" << m_uK << "_" << m_FFTLength << "(vector<unsigned int> &is, vector<unsigned int> &js, vector<double> &real, vector<double> &imag)" << endl;
    cout << "{\n    is.reserve(" << n << ");\n";
    cout << "    js.reserve(" << n << ");\n";
    cout << "    real.reserve(" << n << ");\n";
    cout << "    imag.reserve(" << n << ");\n";
    cout << "    for (int i = 0; i < " << n << "; ++i) {" << endl;
    cout << "        is.push_back(sk_i_" << m_uK << "_" << m_FFTLength << "[i]);" << endl;
    cout << "        js.push_back(sk_j_" << m_uK << "_" << m_FFTLength << "[i]);" << endl;
    cout << "        real.push_back(sk_real_" << m_uK << "_" << m_FFTLength << "[i]);" << endl;
    cout << "        imag.push_back(sk_imag_" << m_uK << "_" << m_FFTLength << "[i]);" << endl;
    cout << "    }" << endl;
    cout << "}" << endl;
*/
//    std::cerr << "done\n -> is: " << sk->is.size() << ", js: " << sk->js.size() << ", reals: " << sk->real.size() << ", imags: " << sk->imag.size() << std::endl;
    
    m_sparseKernel = sk;
    return;
}

//-----------------------------------------------------------------------------
double* ConstantQ::process( const double* fftdata )
{
    if (!m_sparseKernel) {
        std::cerr << "ERROR: ConstantQ::process: Sparse kernel has not been initialised" << std::endl;
        return m_CQdata;
    }

    SparseKernel *sk = m_sparseKernel;

    for (unsigned row=0; row<2*m_uK; row++) 
    {
	m_CQdata[ row ] = 0;
	m_CQdata[ row+1 ] = 0;
    }
    const unsigned *fftbin = &(sk->is[0]);
    const unsigned *cqbin  = &(sk->js[0]);
    const double   *real   = &(sk->real[0]);
    const double   *imag   = &(sk->imag[0]);
    const unsigned int sparseCells = sk->real.size();
	
    for (unsigned i = 0; i<sparseCells; i++)
    {
	const unsigned row = cqbin[i];
	const unsigned col = fftbin[i];
	const double & r1  = real[i];
	const double & i1  = imag[i];
	const double & r2  = fftdata[ (2*m_FFTLength) - 2*col - 2 ];
	const double & i2  = fftdata[ (2*m_FFTLength) - 2*col - 2 + 1 ];
	// add the multiplication
	m_CQdata[ 2*row  ] += (r1*r2 - i1*i2);
	m_CQdata[ 2*row+1] += (r1*i2 + i1*r2);
    }

    return m_CQdata;
}


void ConstantQ::initialise( CQConfig Config )
{
    m_FS = Config.FS;
    m_FMin = Config.min;		// min freq
    m_FMax = Config.max;		// max freq
    m_BPO = Config.BPO;		// bins per octave
    m_CQThresh = Config.CQThresh;// ConstantQ threshold for kernel generation

    m_dQ = 1/(pow(2,(1/(double)m_BPO))-1);	// Work out Q value for Filter bank
    m_uK = (unsigned int) ceil(m_BPO * log(m_FMax/m_FMin)/log(2.0));	// No. of constant Q bins

//    std::cerr << "ConstantQ::initialise: rate = " << m_FS << ", fmin = " << m_FMin << ", fmax = " << m_FMax << ", bpo = " << m_BPO << ", K = " << m_uK << ", Q = " << m_dQ << std::endl;

    // work out length of fft required for this constant Q Filter bank
    m_FFTLength = (int) pow(2, nextpow2(ceil( m_dQ*m_FS/m_FMin )));

    m_hop = m_FFTLength/8; // <------ hop size is window length divided by 32

//    std::cerr << "ConstantQ::initialise: -> fft length = " << m_FFTLength << ", hop = " << m_hop << std::endl;

    // allocate memory for cqdata
    m_CQdata = new double [2*m_uK];
}

void ConstantQ::deInitialise()
{
    delete [] m_CQdata;
    delete m_sparseKernel;
}

void ConstantQ::process(const double *FFTRe, const double* FFTIm,
                        double *CQRe, double *CQIm)
{
    if (!m_sparseKernel) {
        std::cerr << "ERROR: ConstantQ::process: Sparse kernel has not been initialised" << std::endl;
        return;
    }

    SparseKernel *sk = m_sparseKernel;

    for (unsigned row=0; row<m_uK; row++) 
    {
	CQRe[ row ] = 0;
	CQIm[ row ] = 0;
    }

    const unsigned *fftbin = &(sk->is[0]);
    const unsigned *cqbin  = &(sk->js[0]);
    const double   *real   = &(sk->real[0]);
    const double   *imag   = &(sk->imag[0]);
    const unsigned int sparseCells = sk->real.size();
	
    for (unsigned i = 0; i<sparseCells; i++)
    {
	const unsigned row = cqbin[i];
	const unsigned col = fftbin[i];
	const double & r1  = real[i];
	const double & i1  = imag[i];
	const double & r2  = FFTRe[ m_FFTLength - col - 1 ];
	const double & i2  = FFTIm[ m_FFTLength - col - 1 ];
	// add the multiplication
	CQRe[ row ] += (r1*r2 - i1*i2);
	CQIm[ row ] += (r1*i2 + i1*r2);
    }
}