path: root/utils/makemhr/makemhr.cpp

/*
 * HRTF utility for producing and demonstrating the process of creating an
 * OpenAL Soft compatible HRIR data set.
 *
 * Copyright (C) 2011-2019  Christopher Fitzgerald
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License along
 * with this program; if not, write to the Free Software Foundation, Inc.,
 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
 *
 * Or visit:  http://www.gnu.org/licenses/old-licenses/gpl-2.0.html
 *
 * --------------------------------------------------------------------------
 *
 * A big thanks goes out to all those whose work done in the field of
 * binaural sound synthesis using measured HRTFs makes this utility and the
 * OpenAL Soft implementation possible.
 *
 * The algorithm for diffuse-field equalization was adapted from the work
 * done by Rio Emmanuel and Larcher Veronique of IRCAM and Bill Gardner of
 * MIT Media Laboratory.  It operates as follows:
 *
 *  1.  Take the FFT of each HRIR and only keep the magnitude responses.
 *  2.  Calculate the diffuse-field power-average of all HRIRs weighted by
 *      their contribution to the total surface area covered by their
 *      measurement. This has since been modified to use coverage volume for
 *      multi-field HRIR data sets.
 *  3.  Take the diffuse-field average and limit its magnitude range.
 *  4.  Equalize the responses by using the inverse of the diffuse-field
 *      average.
 *  5.  Reconstruct the minimum-phase responses.
 *  5.  Zero the DC component.
 *  6.  IFFT the result and truncate to the desired-length minimum-phase FIR.
 *
 * The spherical head algorithm for calculating propagation delay was adapted
 * from the paper:
 *
 *  Modeling Interaural Time Difference Assuming a Spherical Head
 *  Joel David Miller
 *  Music 150, Musical Acoustics, Stanford University
 *  December 2, 2001
 *
 * The formulae for calculating the Kaiser window metrics are from the
 * the textbook:
 *
 *  Discrete-Time Signal Processing
 *  Alan V. Oppenheim and Ronald W. Schafer
 *  Prentice-Hall Signal Processing Series
 *  1999
 */

#define _UNICODE
#include "config.h"

#include "makemhr.h"

#include <algorithm>
#include <atomic>
#include <chrono>
#include <cmath>
#include <complex>
#include <cstdint>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <functional>
#include <limits>
#include <memory>
#include <numeric>
#include <thread>
#include <utility>
#include <vector>

#ifdef HAVE_GETOPT
#include <unistd.h>
#else
#include "../getopt.h"
#endif

#include "loaddef.h"
#include "loadsofa.h"

#include "win_main_utf8.h"


namespace {

using namespace std::placeholders;

} // namespace

#ifndef M_PI
#define M_PI                         (3.14159265358979323846)
#endif


// Head model used for calculating the impulse delays.
enum HeadModelT {
    HM_NONE,
    HM_DATASET, // Measure the onset from the dataset.
    HM_SPHERE   // Calculate the onset using a spherical head model.
};


// The epsilon used to maintain signal stability.
#define EPSILON                      (1e-9)

// The limits to the FFT window size override on the command line.
#define MIN_FFTSIZE                  (65536)
#define MAX_FFTSIZE                  (131072)

// The limits to the equalization range limit on the command line.
#define MIN_LIMIT                    (2.0)
#define MAX_LIMIT                    (120.0)

// The limits to the truncation window size on the command line.
#define MIN_TRUNCSIZE                (16)
#define MAX_TRUNCSIZE                (512)

// The limits to the custom head radius on the command line.
#define MIN_CUSTOM_RADIUS            (0.05)
#define MAX_CUSTOM_RADIUS            (0.15)

// The truncation window size must be a multiple of the below value to allow
// for vectorized convolution.
#define MOD_TRUNCSIZE                (8)

// The defaults for the command line options.
#define DEFAULT_FFTSIZE              (65536)
#define DEFAULT_EQUALIZE             (1)
#define DEFAULT_SURFACE              (1)
#define DEFAULT_LIMIT                (24.0)
#define DEFAULT_TRUNCSIZE            (32)
#define DEFAULT_HEAD_MODEL           (HM_DATASET)
#define DEFAULT_CUSTOM_RADIUS        (0.0)

// The maximum propagation delay value supported by OpenAL Soft.
#define MAX_HRTD                     (63.0)

// The OpenAL Soft HRTF format marker.  It stands for minimum-phase head
// response protocol 02.
#define MHR_FORMAT                   ("MinPHR02")

/* Channel index enums. Mono uses LeftChannel only. */
enum ChannelIndex : uint {
    LeftChannel = 0u,
    RightChannel = 1u
};


/* Performs a string substitution.  Any case-insensitive occurrences of the
 * pattern string are replaced with the replacement string.  The result is
 * truncated if necessary.
 */
static int StrSubst(const char *in, const char *pat, const char *rep, const size_t maxLen, char *out)
{
    size_t inLen, patLen, repLen;
    size_t si, di;
    int truncated;

    inLen = strlen(in);
    patLen = strlen(pat);
    repLen = strlen(rep);
    si = 0;
    di = 0;
    truncated = 0;
    while(si < inLen && di < maxLen)
    {
        if(patLen <= inLen-si)
        {
            if(strncasecmp(&in[si], pat, patLen) == 0)
            {
                if(repLen > maxLen-di)
                {
                    repLen = maxLen - di;
                    truncated = 1;
                }
                strncpy(&out[di], rep, repLen);
                si += patLen;
                di += repLen;
            }
        }
        out[di] = in[si];
        si++;
        di++;
    }
    if(si < inLen)
        truncated = 1;
    out[di] = '\0';
    return !truncated;
}


/*********************
 *** Math routines ***
 *********************/

// Simple clamp routine.
static double Clamp(const double val, const double lower, const double upper)
{
    return std::min(std::max(val, lower), upper);
}

static inline uint dither_rng(uint *seed)
{
    *seed = *seed * 96314165 + 907633515;
    return *seed;
}

// Performs a triangular probability density function dither. The input samples
// should be normalized (-1 to +1).
static void TpdfDither(double *RESTRICT out, const double *RESTRICT in, const double scale,
                       const int count, const int step, uint *seed)
{
    static constexpr double PRNG_SCALE = 1.0 / std::numeric_limits<uint>::max();

    for(int i{0};i < count;i++)
    {
        uint prn0{dither_rng(seed)};
        uint prn1{dither_rng(seed)};
        out[i*step] = std::round(in[i]*scale + (prn0*PRNG_SCALE - prn1*PRNG_SCALE));
    }
}

/* Fast Fourier transform routines. The number of points must be a power of
 * two.
 */

// Performs bit-reversal ordering.
static void FftArrange(const uint n, complex_d *inout)
{
    // Handle in-place arrangement.
    uint rk{0u};
    for(uint k{0u};k < n;k++)
    {
        if(rk > k)
            std::swap(inout[rk], inout[k]);

        uint m{n};
        while(rk&(m >>= 1))
            rk &= ~m;
        rk |= m;
    }
}

// Performs the summation.
static void FftSummation(const int n, const double s, complex_d *cplx)
{
    double pi;
    int m, m2;
    int i, k, mk;

    pi = s * M_PI;
    for(m = 1, m2 = 2;m < n; m <<= 1, m2 <<= 1)
    {
        // v = Complex (-2.0 * sin (0.5 * pi / m) * sin (0.5 * pi / m), -sin (pi / m))
        double sm = sin(0.5 * pi / m);
        auto v = complex_d{-2.0*sm*sm, -sin(pi / m)};
        auto w = complex_d{1.0, 0.0};
        for(i = 0;i < m;i++)
        {
            for(k = i;k < n;k += m2)
            {
                mk = k + m;
                auto t = w * cplx[mk];
                cplx[mk] = cplx[k] - t;
                cplx[k] = cplx[k] + t;
            }
            w += v*w;
        }
    }
}

// Performs a forward FFT.
void FftForward(const uint n, complex_d *inout)
{
    FftArrange(n, inout);
    FftSummation(n, 1.0, inout);
}

// Performs an inverse FFT.
void FftInverse(const uint n, complex_d *inout)
{
    FftArrange(n, inout);
    FftSummation(n, -1.0, inout);
    double f{1.0 / n};
    for(uint i{0};i < n;i++)
        inout[i] *= f;
}

/* Calculate the complex helical sequence (or discrete-time analytical signal)
 * of the given input using the Hilbert transform. Given the natural logarithm
 * of a signal's magnitude response, the imaginary components can be used as
 * the angles for minimum-phase reconstruction.
 */
static void Hilbert(const uint n, complex_d *inout)
{
    uint i;

    // Handle in-place operation.
    for(i = 0;i < n;i++)
        inout[i].imag(0.0);

    FftInverse(n, inout);
    for(i = 1;i < (n+1)/2;i++)
        inout[i] *= 2.0;
    /* Increment i if n is even. */
    i += (n&1)^1;
    for(;i < n;i++)
        inout[i] = complex_d{0.0, 0.0};
    FftForward(n, inout);
}

/* Calculate the magnitude response of the given input.  This is used in
 * place of phase decomposition, since the phase residuals are discarded for
 * minimum phase reconstruction.  The mirrored half of the response is also
 * discarded.
 */
void MagnitudeResponse(const uint n, const complex_d *in, double *out)
{
    const uint m = 1 + (n / 2);
    uint i;
    for(i = 0;i < m;i++)
        out[i] = std::max(std::abs(in[i]), EPSILON);
}

/* Apply a range limit (in dB) to the given magnitude response.  This is used
 * to adjust the effects of the diffuse-field average on the equalization
 * process.
 */
static void LimitMagnitudeResponse(const uint n, const uint m, const double limit, const double *in, double *out)
{
    double halfLim;
    uint i, lower, upper;
    double ave;

    halfLim = limit / 2.0;
    // Convert the response to dB.
    for(i = 0;i < m;i++)
        out[i] = 20.0 * std::log10(in[i]);
    // Use six octaves to calculate the average magnitude of the signal.
    lower = (static_cast<uint>(std::ceil(n / std::pow(2.0, 8.0)))) - 1;
    upper = (static_cast<uint>(std::floor(n / std::pow(2.0, 2.0)))) - 1;
    ave = 0.0;
    for(i = lower;i <= upper;i++)
        ave += out[i];
    ave /= upper - lower + 1;
    // Keep the response within range of the average magnitude.
    for(i = 0;i < m;i++)
        out[i] = Clamp(out[i], ave - halfLim, ave + halfLim);
    // Convert the response back to linear magnitude.
    for(i = 0;i < m;i++)
        out[i] = std::pow(10.0, out[i] / 20.0);
}

/* Reconstructs the minimum-phase component for the given magnitude response
 * of a signal.  This is equivalent to phase recomposition, sans the missing
 * residuals (which were discarded).  The mirrored half of the response is
 * reconstructed.
 */
static void MinimumPhase(const uint n, const double *in, complex_d *out)
{
    const uint m = 1 + (n / 2);
    std::vector<double> mags(n);

    uint i;
    for(i = 0;i < m;i++)
    {
        mags[i] = std::max(EPSILON, in[i]);
        out[i] = complex_d{std::log(mags[i]), 0.0};
    }
    for(;i < n;i++)
    {
        mags[i] = mags[n - i];
        out[i] = out[n - i];
    }
    Hilbert(n, out);
    // Remove any DC offset the filter has.
    mags[0] = EPSILON;
    for(i = 0;i < n;i++)
    {
        auto a = std::exp(complex_d{0.0, out[i].imag()});
        out[i] = complex_d{mags[i], 0.0} * a;
    }
}


/***************************
 *** Resampler functions ***
 ***************************/

/* This is the normalized cardinal sine (sinc) function.
 *
 *   sinc(x) = { 1,                   x = 0
 *             { sin(pi x) / (pi x),  otherwise.
 */
static double Sinc(const double x)
{
    if(std::abs(x) < EPSILON)
        return 1.0;
    return std::sin(M_PI * x) / (M_PI * x);
}

/* The zero-order modified Bessel function of the first kind, used for the
 * Kaiser window.
 *
 *   I_0(x) = sum_{k=0}^inf (1 / k!)^2 (x / 2)^(2 k)
 *          = sum_{k=0}^inf ((x / 2)^k / k!)^2
 */
static double BesselI_0(const double x)
{
    double term, sum, x2, y, last_sum;
    int k;

    // Start at k=1 since k=0 is trivial.
    term = 1.0;
    sum = 1.0;
    x2 = x/2.0;
    k = 1;

    // Let the integration converge until the term of the sum is no longer
    // significant.
    do {
        y = x2 / k;
        k++;
        last_sum = sum;
        term *= y * y;
        sum += term;
    } while(sum != last_sum);
    return sum;
}

/* Calculate a Kaiser window from the given beta value and a normalized k
 * [-1, 1].
 *
 *   w(k) = { I_0(B sqrt(1 - k^2)) / I_0(B),  -1 <= k <= 1
 *          { 0,                              elsewhere.
 *
 * Where k can be calculated as:
 *
 *   k = i / l,         where -l <= i <= l.
 *
 * or:
 *
 *   k = 2 i / M - 1,   where 0 <= i <= M.
 */
static double Kaiser(const double b, const double k)
{
    if(!(k >= -1.0 && k <= 1.0))
        return 0.0;
    return BesselI_0(b * std::sqrt(1.0 - k*k)) / BesselI_0(b);
}

// Calculates the greatest common divisor of a and b.
static uint Gcd(uint x, uint y)
{
    while(y > 0)
    {
        uint z{y};
        y = x % y;
        x = z;
    }
    return x;
}

/* Calculates the size (order) of the Kaiser window.  Rejection is in dB and
 * the transition width is normalized frequency (0.5 is nyquist).
 *
 *   M = { ceil((r - 7.95) / (2.285 2 pi f_t)),  r > 21
 *       { ceil(5.79 / 2 pi f_t),                r <= 21.
 *
 */
static uint CalcKaiserOrder(const double rejection, const double transition)
{
    double w_t = 2.0 * M_PI * transition;
    if(rejection > 21.0)
        return static_cast<uint>(std::ceil((rejection - 7.95) / (2.285 * w_t)));
    return static_cast<uint>(std::ceil(5.79 / w_t));
}

// Calculates the beta value of the Kaiser window.  Rejection is in dB.
static double CalcKaiserBeta(const double rejection)
{
    if(rejection > 50.0)
        return 0.1102 * (rejection - 8.7);
    if(rejection >= 21.0)
        return (0.5842 * std::pow(rejection - 21.0, 0.4)) +
               (0.07886 * (rejection - 21.0));
    return 0.0;
}

/* Calculates a point on the Kaiser-windowed sinc filter for the given half-
 * width, beta, gain, and cutoff.  The point is specified in non-normalized
 * samples, from 0 to M, where M = (2 l + 1).
 *
 *   w(k) 2 p f_t sinc(2 f_t x)
 *
 *   x    -- centered sample index (i - l)
 *   k    -- normalized and centered window index (x / l)
 *   w(k) -- window function (Kaiser)
 *   p    -- gain compensation factor when sampling
 *   f_t  -- normalized center frequency (or cutoff; 0.5 is nyquist)
 */
static double SincFilter(const int l, const double b, const double gain, const double cutoff, const int i)
{
    return Kaiser(b, static_cast<double>(i - l) / l) * 2.0 * gain * cutoff * Sinc(2.0 * cutoff * (i - l));
}

/* This is a polyphase sinc-filtered resampler.
 *
 *              Upsample                      Downsample
 *
 *              p/q = 3/2                     p/q = 3/5
 *
 *          M-+-+-+->                     M-+-+-+->
 *         -------------------+          ---------------------+
 *   p  s * f f f f|f|        |    p  s * f f f f f           |
 *   |  0 *   0 0 0|0|0       |    |  0 *   0 0 0 0|0|        |
 *   v  0 *     0 0|0|0 0     |    v  0 *     0 0 0|0|0       |
 *      s *       f|f|f f f   |       s *       f f|f|f f     |
 *      0 *        |0|0 0 0 0 |       0 *         0|0|0 0 0   |
 *         --------+=+--------+       0 *          |0|0 0 0 0 |
 *          d . d .|d|. d . d            ----------+=+--------+
 *                                        d . . . .|d|. . . .
 *          q->
 *                                        q-+-+-+->
 *
 *   P_f(i,j) = q i mod p + pj
 *   P_s(i,j) = floor(q i / p) - j
 *   d[i=0..N-1] = sum_{j=0}^{floor((M - 1) / p)} {
 *                   { f[P_f(i,j)] s[P_s(i,j)],  P_f(i,j) < M
 *                   { 0,                        P_f(i,j) >= M. }
 */

// Calculate the resampling metrics and build the Kaiser-windowed sinc filter
// that's used to cut frequencies above the destination nyquist.
void ResamplerSetup(ResamplerT *rs, const uint srcRate, const uint dstRate)
{
    double cutoff, width, beta;
    uint gcd, l;
    int i;

    gcd = Gcd(srcRate, dstRate);
    rs->mP = dstRate / gcd;
    rs->mQ = srcRate / gcd;
    /* The cutoff is adjusted by half the transition width, so the transition
     * ends before the nyquist (0.5).  Both are scaled by the downsampling
     * factor.
     */
    if(rs->mP > rs->mQ)
    {
        cutoff = 0.475 / rs->mP;
        width = 0.05 / rs->mP;
    }
    else
    {
        cutoff = 0.475 / rs->mQ;
        width = 0.05 / rs->mQ;
    }
    // A rejection of -180 dB is used for the stop band. Round up when
    // calculating the left offset to avoid increasing the transition width.
    l = (CalcKaiserOrder(180.0, width)+1) / 2;
    beta = CalcKaiserBeta(180.0);
    rs->mM = l*2 + 1;
    rs->mL = l;
    rs->mF.resize(rs->mM);
    for(i = 0;i < (static_cast<int>(rs->mM));i++)
        rs->mF[i] = SincFilter(static_cast<int>(l), beta, rs->mP, cutoff, i);
}

// Perform the upsample-filter-downsample resampling operation using a
// polyphase filter implementation.
void ResamplerRun(ResamplerT *rs, const uint inN, const double *in, const uint outN, double *out)
{
    const uint p = rs->mP, q = rs->mQ, m = rs->mM, l = rs->mL;
    std::vector<double> workspace;
    const double *f = rs->mF.data();
    uint j_f, j_s;
    double *work;
    uint i;

    if(outN == 0)
        return;

    // Handle in-place operation.
    if(in == out)
    {
        workspace.resize(outN);
        work = workspace.data();
    }
    else
        work = out;
    // Resample the input.
    for(i = 0;i < outN;i++)
    {
        double r = 0.0;
        // Input starts at l to compensate for the filter delay.  This will
        // drop any build-up from the first half of the filter.
        j_f = (l + (q * i)) % p;
        j_s = (l + (q * i)) / p;
        while(j_f < m)
        {
            // Only take input when 0 <= j_s < inN.  This single unsigned
            // comparison catches both cases.
            if(j_s < inN)
                r += f[j_f] * in[j_s];
            j_f += p;
            j_s--;
        }
        work[i] = r;
    }
    // Clean up after in-place operation.
    if(work != out)
    {
        for(i = 0;i < outN;i++)
            out[i] = work[i];
    }
}


/***************************
 *** File storage output ***
 ***************************/

// Write an ASCII string to a file.
static int WriteAscii(const char *out, FILE *fp, const char *filename)
{
    size_t len;

    len = strlen(out);
    if(fwrite(out, 1, len, fp) != len)
    {
        fclose(fp);
        fprintf(stderr, "\nError: Bad write to file '%s'.\n", filename);
        return 0;
    }
    return 1;
}

// Write a binary value of the given byte order and byte size to a file,
// loading it from a 32-bit unsigned integer.
static int WriteBin4(const uint bytes, const uint32_t in, FILE *fp, const char *filename)
{
    uint8_t out[4];
    uint i;

    for(i = 0;i < bytes;i++)
        out[i] = (in>>(i*8)) & 0x000000FF;

    if(fwrite(out, 1, bytes, fp) != bytes)
    {
        fprintf(stderr, "\nError: Bad write to file '%s'.\n", filename);
        return 0;
    }
    return 1;
}

// Store the OpenAL Soft HRTF data set.
static int StoreMhr(const HrirDataT *hData, const char *filename)
{
    uint channels = (hData->mChannelType == CT_STEREO) ? 2 : 1;
    uint n = hData->mIrPoints;
    FILE *fp;
    uint fi, ei, ai, i;
    uint dither_seed = 22222;

    if((fp=fopen(filename, "wb")) == nullptr)
    {
        fprintf(stderr, "\nError: Could not open MHR file '%s'.\n", filename);
        return 0;
    }
    if(!WriteAscii(MHR_FORMAT, fp, filename))
        return 0;
    if(!WriteBin4(4, hData->mIrRate, fp, filename))
        return 0;
    if(!WriteBin4(1, static_cast<uint32_t>(hData->mSampleType), fp, filename))
        return 0;
    if(!WriteBin4(1, static_cast<uint32_t>(hData->mChannelType), fp, filename))
        return 0;
    if(!WriteBin4(1, hData->mIrPoints, fp, filename))
        return 0;
    if(!WriteBin4(1, hData->mFdCount, fp, filename))
        return 0;
    for(fi = 0;fi < hData->mFdCount;fi++)
    {
        auto fdist = static_cast<uint32_t>(std::round(1000.0 * hData->mFds[fi].mDistance));
        if(!WriteBin4(2, fdist, fp, filename))
            return 0;
        if(!WriteBin4(1, hData->mFds[fi].mEvCount, fp, filename))
            return 0;
        for(ei = 0;ei < hData->mFds[fi].mEvCount;ei++)
        {
            if(!WriteBin4(1, hData->mFds[fi].mEvs[ei].mAzCount, fp, filename))
                return 0;
        }
    }

    for(fi = 0;fi < hData->mFdCount;fi++)
    {
        const double scale = (hData->mSampleType == ST_S16) ? 32767.0 :
                             ((hData->mSampleType == ST_S24) ? 8388607.0 : 0.0);
        const int bps = (hData->mSampleType == ST_S16) ? 2 :
                        ((hData->mSampleType == ST_S24) ? 3 : 0);

        for(ei = 0;ei < hData->mFds[fi].mEvCount;ei++)
        {
            for(ai = 0;ai < hData->mFds[fi].mEvs[ei].mAzCount;ai++)
            {
                HrirAzT *azd = &hData->mFds[fi].mEvs[ei].mAzs[ai];
                double out[2 * MAX_TRUNCSIZE];

                TpdfDither(out, azd->mIrs[0], scale, n, channels, &dither_seed);
                if(hData->mChannelType == CT_STEREO)
                    TpdfDither(out+1, azd->mIrs[1], scale, n, channels, &dither_seed);
                for(i = 0;i < (channels * n);i++)
                {
                    int v = static_cast<int>(Clamp(out[i], -scale-1.0, scale));
                    if(!WriteBin4(bps, static_cast<uint32_t>(v), fp, filename))
                        return 0;
                }
            }
        }
    }
    for(fi = 0;fi < hData->mFdCount;fi++)
    {
        for(ei = 0;ei < hData->mFds[fi].mEvCount;ei++)
        {
            for(ai = 0;ai < hData->mFds[fi].mEvs[ei].mAzCount;ai++)
            {
                const HrirAzT &azd = hData->mFds[fi].mEvs[ei].mAzs[ai];
                int v = static_cast<int>(std::min(std::round(hData->mIrRate * azd.mDelays[0]), MAX_HRTD));

                if(!WriteBin4(1, static_cast<uint32_t>(v), fp, filename))
                    return 0;
                if(hData->mChannelType == CT_STEREO)
                {
                    v = static_cast<int>(std::min(std::round(hData->mIrRate * azd.mDelays[1]), MAX_HRTD));

                    if(!WriteBin4(1, static_cast<uint32_t>(v), fp, filename))
                        return 0;
                }
            }
        }
    }
    fclose(fp);
    return 1;
}


/***********************
 *** HRTF processing ***
 ***********************/

/* Balances the maximum HRIR magnitudes of multi-field data sets by
 * independently normalizing each field in relation to the overall maximum.
 * This is done to ignore distance attenuation.
 */
static void BalanceFieldMagnitudes(const HrirDataT *hData, const uint channels, const uint m)
{
    double maxMags[MAX_FD_COUNT];
    uint fi, ei, ai, ti, i;

    double maxMag{0.0};
    for(fi = 0;fi < hData->mFdCount;fi++)
    {
        maxMags[fi] = 0.0;

        for(ei = hData->mFds[fi].mEvStart;ei < hData->mFds[fi].mEvCount;ei++)
        {
            for(ai = 0;ai < hData->mFds[fi].mEvs[ei].mAzCount;ai++)
            {
                HrirAzT *azd = &hData->mFds[fi].mEvs[ei].mAzs[ai];
                for(ti = 0;ti < channels;ti++)
                {
                    for(i = 0;i < m;i++)
                        maxMags[fi] = std::max(azd->mIrs[ti][i], maxMags[fi]);
                }
            }
        }

        maxMag = std::max(maxMags[fi], maxMag);
    }

    for(fi = 0;fi < hData->mFdCount;fi++)
    {
        const double magFactor{maxMag / maxMags[fi]};

        for(ei = hData->mFds[fi].mEvStart;ei < hData->mFds[fi].mEvCount;ei++)
        {
            for(ai = 0;ai < hData->mFds[fi].mEvs[ei].mAzCount;ai++)
            {
                HrirAzT *azd = &hData->mFds[fi].mEvs[ei].mAzs[ai];
                for(ti = 0;ti < channels;ti++)
                {
                    for(i = 0;i < m;i++)
                        azd->mIrs[ti][i] *= magFactor;
                }
            }
        }
    }
}

/* Calculate the contribution of each HRIR to the diffuse-field average based
 * on its coverage volume.  All volumes are centered at the spherical HRIR
 * coordinates and measured by extruded solid angle.
 */
static void CalculateDfWeights(const HrirDataT *hData, double *weights)
{
    double sum, innerRa, outerRa, evs, ev, upperEv, lowerEv;
    double solidAngle, solidVolume;
    uint fi, ei;

    sum = 0.0;
    // The head radius acts as the limit for the inner radius.
    innerRa = hData->mRadius;
    for(fi = 0;fi < hData->mFdCount;fi++)
    {
        // Each volume ends half way between progressive field measurements.
        if((fi + 1) < hData->mFdCount)
            outerRa = 0.5f * (hData->mFds[fi].mDistance + hData->mFds[fi + 1].mDistance);
        // The final volume has its limit extended to some practical value.
        // This is done to emphasize the far-field responses in the average.
        else
            outerRa = 10.0f;

        evs = M_PI / 2.0 / (hData->mFds[fi].mEvCount - 1);
        for(ei = hData->mFds[fi].mEvStart;ei < hData->mFds[fi].mEvCount;ei++)
        {
            // For each elevation, calculate the upper and lower limits of
            // the patch band.
            ev = hData->mFds[fi].mEvs[ei].mElevation;
            lowerEv = std::max(-M_PI / 2.0, ev - evs);
            upperEv = std::min(M_PI / 2.0, ev + evs);
            // Calculate the surface area of the patch band.
            solidAngle = 2.0 * M_PI * (std::sin(upperEv) - std::sin(lowerEv));
            // Then the volume of the extruded patch band.
            solidVolume = solidAngle * (std::pow(outerRa, 3.0) - std::pow(innerRa, 3.0)) / 3.0;
            // Each weight is the volume of one extruded patch.
            weights[(fi * MAX_EV_COUNT) + ei] = solidVolume / hData->mFds[fi].mEvs[ei].mAzCount;
            // Sum the total coverage volume of the HRIRs for all fields.
            sum += solidAngle;
        }

        innerRa = outerRa;
    }

    for(fi = 0;fi < hData->mFdCount;fi++)
    {
        // Normalize the weights given the total surface coverage for all
        // fields.
        for(ei = hData->mFds[fi].mEvStart;ei < hData->mFds[fi].mEvCount;ei++)
            weights[(fi * MAX_EV_COUNT) + ei] /= sum;
    }
}

/* Calculate the diffuse-field average from the given magnitude responses of
 * the HRIR set.  Weighting can be applied to compensate for the varying
 * coverage of each HRIR.  The final average can then be limited by the
 * specified magnitude range (in positive dB; 0.0 to skip).
 */
static void CalculateDiffuseFieldAverage(const HrirDataT *hData, const uint channels, const uint m, const int weighted, const double limit, double *dfa)
{
    std::vector<double> weights(hData->mFdCount * MAX_EV_COUNT);
    uint count, ti, fi, ei, i, ai;

    if(weighted)
    {
        // Use coverage weighting to calculate the average.
        CalculateDfWeights(hData, weights.data());
    }
    else
    {
        double weight;

        // If coverage weighting is not used, the weights still need to be
        // averaged by the number of existing HRIRs.
        count = hData->mIrCount;
        for(fi = 0;fi < hData->mFdCount;fi++)
        {
            for(ei = 0;ei < hData->mFds[fi].mEvStart;ei++)
                count -= hData->mFds[fi].mEvs[ei].mAzCount;
        }
        weight = 1.0 / count;

        for(fi = 0;fi < hData->mFdCount;fi++)
        {
            for(ei = hData->mFds[fi].mEvStart;ei < hData->mFds[fi].mEvCount;ei++)
                weights[(fi * MAX_EV_COUNT) + ei] = weight;
        }
    }
    for(ti = 0;ti < channels;ti++)
    {
        for(i = 0;i < m;i++)
            dfa[(ti * m) + i] = 0.0;
        for(fi = 0;fi < hData->mFdCount;fi++)
        {
            for(ei = hData->mFds[fi].mEvStart;ei < hData->mFds[fi].mEvCount;ei++)
            {
                for(ai = 0;ai < hData->mFds[fi].mEvs[ei].mAzCount;ai++)
                {
                    HrirAzT *azd = &hData->mFds[fi].mEvs[ei].mAzs[ai];
                    // Get the weight for this HRIR's contribution.
                    double weight = weights[(fi * MAX_EV_COUNT) + ei];

                    // Add this HRIR's weighted power average to the total.
                    for(i = 0;i < m;i++)
                        dfa[(ti * m) + i] += weight * azd->mIrs[ti][i] * azd->mIrs[ti][i];
                }
            }
        }
        // Finish the average calculation and keep it from being too small.
        for(i = 0;i < m;i++)
            dfa[(ti * m) + i] = std::max(sqrt(dfa[(ti * m) + i]), EPSILON);
        // Apply a limit to the magnitude range of the diffuse-field average
        // if desired.
        if(limit > 0.0)
            LimitMagnitudeResponse(hData->mFftSize, m, limit, &dfa[ti * m], &dfa[ti * m]);
    }
}

// Perform diffuse-field equalization on the magnitude responses of the HRIR
// set using the given average response.
static void DiffuseFieldEqualize(const uint channels, const uint m, const double *dfa, const HrirDataT *hData)
{
    uint ti, fi, ei, ai, i;

    for(fi = 0;fi < hData->mFdCount;fi++)
    {
        for(ei = hData->mFds[fi].mEvStart;ei < hData->mFds[fi].mEvCount;ei++)
        {
            for(ai = 0;ai < hData->mFds[fi].mEvs[ei].mAzCount;ai++)
            {
                HrirAzT *azd = &hData->mFds[fi].mEvs[ei].mAzs[ai];

                for(ti = 0;ti < channels;ti++)
                {
                    for(i = 0;i < m;i++)
                        azd->mIrs[ti][i] /= dfa[(ti * m) + i];
                }
            }
        }
    }
}

/* Perform minimum-phase reconstruction using the magnitude responses of the
 * HRIR set. Work is delegated to this struct, which runs asynchronously on one
 * or more threads (sharing the same reconstructor object).
 */
struct HrirReconstructor {
    std::vector<double*> mIrs;
    std::atomic<size_t> mCurrent;
    std::atomic<size_t> mDone;
    size_t mFftSize;
    size_t mIrPoints;

    void Worker()
    {
        auto h = std::vector<complex_d>(mFftSize);

        while(1)
        {
            /* Load the current index to process. */
            size_t idx{mCurrent.load()};
            do {
                /* If the index is at the end, we're done. */
                if(idx >= mIrs.size())
                    return;
                /* Otherwise, increment the current index atomically so other
                 * threads know to go to the next one. If this call fails, the
                 * current index was just changed by another thread and the new
                 * value is loaded into idx, which we'll recheck.
                 */
            } while(!mCurrent.compare_exchange_weak(idx, idx+1, std::memory_order_relaxed));

            /* Now do the reconstruction, and apply the inverse FFT to get the
             * time-domain response.
             */
            MinimumPhase(mFftSize, mIrs[idx], h.data());
            FftInverse(mFftSize, h.data());
            for(size_t i{0u};i < mIrPoints;++i)
                mIrs[idx][i] = h[i].real();

            /* Increment the number of IRs done. */
            mDone.fetch_add(1);
        }
    }
};

static void ReconstructHrirs(const HrirDataT *hData)
{
    const uint channels{(hData->mChannelType == CT_STEREO) ? 2u : 1u};

    /* Count the number of IRs to process (excluding elevations that will be
     * synthesized later).
     */
    size_t total{hData->mIrCount};
    for(uint fi{0u};fi < hData->mFdCount;fi++)
    {
        for(uint ei{0u};ei < hData->mFds[fi].mEvStart;ei++)
            total -= hData->mFds[fi].mEvs[ei].mAzCount;
    }
    total *= channels;

    /* Set up the reconstructor with the needed size info and pointers to the
     * IRs to process.
     */
    HrirReconstructor reconstructor;
    reconstructor.mIrs.reserve(total);
    reconstructor.mCurrent.store(0, std::memory_order_relaxed);
    reconstructor.mDone.store(0, std::memory_order_relaxed);
    reconstructor.mFftSize = hData->mFftSize;
    reconstructor.mIrPoints = hData->mIrPoints;
    for(uint fi{0u};fi < hData->mFdCount;fi++)
    {
        const HrirFdT &field = hData->mFds[fi];
        for(uint ei{field.mEvStart};ei < field.mEvCount;ei++)
        {
            const HrirEvT &elev = field.mEvs[ei];
            for(uint ai{0u};ai < elev.mAzCount;ai++)
            {
                const HrirAzT &azd = elev.mAzs[ai];
                for(uint ti{0u};ti < channels;ti++)
                    reconstructor.mIrs.push_back(azd.mIrs[ti]);
            }
        }
    }

    /* Launch two threads to work on reconstruction. */
    std::thread thrd1{std::mem_fn(&HrirReconstructor::Worker), &reconstructor};
    std::thread thrd2{std::mem_fn(&HrirReconstructor::Worker), &reconstructor};

    /* Keep track of the number of IRs done, periodically reporting it. */
    size_t count;
    while((count=reconstructor.mDone.load()) != total)
    {
        size_t pcdone{count * 100 / total};

        printf("\r%3zu%% done (%zu of %zu)", pcdone, count, total);
        fflush(stdout);

        std::this_thread::sleep_for(std::chrono::milliseconds{50});
    }
    size_t pcdone{count * 100 / total};
    printf("\r%3zu%% done (%zu of %zu)\n", pcdone, count, total);

    if(thrd2.joinable()) thrd2.join();
    if(thrd1.joinable()) thrd1.join();
}

// Resamples the HRIRs for use at the given sampling rate.
static void ResampleHrirs(const uint rate, HrirDataT *hData)
{
    uint channels = (hData->mChannelType == CT_STEREO) ? 2 : 1;
    uint n = hData->mIrPoints;
    uint ti, fi, ei, ai;
    ResamplerT rs;

    ResamplerSetup(&rs, hData->mIrRate, rate);
    for(fi = 0;fi < hData->mFdCount;fi++)
    {
        for(ei = hData->mFds[fi].mEvStart;ei < hData->mFds[fi].mEvCount;ei++)
        {
            for(ai = 0;ai < hData->mFds[fi].mEvs[ei].mAzCount;ai++)
            {
                HrirAzT *azd = &hData->mFds[fi].mEvs[ei].mAzs[ai];
                for(ti = 0;ti < channels;ti++)
                    ResamplerRun(&rs, n, azd->mIrs[ti], n, azd->mIrs[ti]);
            }
        }
    }
    hData->mIrRate = rate;
}

/* Given field and elevation indices and an azimuth, calculate the indices of
 * the two HRIRs that bound the coordinate along with a factor for
 * calculating the continuous HRIR using interpolation.
 */
static void CalcAzIndices(const HrirFdT &field, const uint ei, const double az, uint *a0, uint *a1, double *af)
{
    double f{(2.0*M_PI + az) * field.mEvs[ei].mAzCount / (2.0*M_PI)};
    uint i{static_cast<uint>(f) % field.mEvs[ei].mAzCount};

    f -= std::floor(f);
    *a0 = i;
    *a1 = (i + 1) % field.mEvs[ei].mAzCount;
    *af = f;
}

/* Synthesize any missing onset timings at the bottom elevations of each field.
 * This just mirrors some top elevations for the bottom, and blends the
 * remaining elevations (not an accurate model).
 */
static void SynthesizeOnsets(HrirDataT *hData)
{
    const uint channels{(hData->mChannelType == CT_STEREO) ? 2u : 1u};

    auto proc_field = [channels](HrirFdT &field) -> void
    {
        /* Get the starting elevation from the measurements, and use it as the
         * upper elevation limit for what needs to be calculated.
         */
        const uint upperElevReal{field.mEvStart};
        if(upperElevReal <= 0) return;

        /* Get the lowest half of the missing elevations' delays by mirroring
         * the top elevation delays. The responses are on a spherical grid
         * centered between the ears, so these should align.
         */
        uint ei{};
        if(channels > 1)
        {
            /* Take the polar opposite position of the desired measurement and
             * swap the ears.
             */
            field.mEvs[0].mAzs[0].mDelays[0] = field.mEvs[field.mEvCount-1].mAzs[0].mDelays[1];
            field.mEvs[0].mAzs[0].mDelays[1] = field.mEvs[field.mEvCount-1].mAzs[0].mDelays[0];
            for(ei = 1u;ei < (upperElevReal+1)/2;++ei)
            {
                const uint topElev{field.mEvCount-ei-1};

                for(uint ai{0u};ai < field.mEvs[ei].mAzCount;ai++)
                {
                    uint a0, a1;
                    double af;

                    /* Rotate this current azimuth by a half-circle, and lookup
                     * the mirrored elevation to find the indices for the polar
                     * opposite position (may need blending).
                     */
                    const double az{field.mEvs[ei].mAzs[ai].mAzimuth + M_PI};
                    CalcAzIndices(field, topElev, az, &a0, &a1, &af);

                    /* Blend the delays, and again, swap the ears. */
                    field.mEvs[ei].mAzs[ai].mDelays[0] = Lerp(
                        field.mEvs[topElev].mAzs[a0].mDelays[1],
                        field.mEvs[topElev].mAzs[a1].mDelays[1], af);
                    field.mEvs[ei].mAzs[ai].mDelays[1] = Lerp(
                        field.mEvs[topElev].mAzs[a0].mDelays[0],
                        field.mEvs[topElev].mAzs[a1].mDelays[0], af);
                }
            }
        }
        else
        {
            field.mEvs[0].mAzs[0].mDelays[0] = field.mEvs[field.mEvCount-1].mAzs[0].mDelays[0];
            for(ei = 1u;ei < (upperElevReal+1)/2;++ei)
            {
                const uint topElev{field.mEvCount-ei-1};

                for(uint ai{0u};ai < field.mEvs[ei].mAzCount;ai++)
                {
                    uint a0, a1;
                    double af;

                    /* For mono data sets, mirror the azimuth front<->back
                     * since the other ear is a mirror of what we have (e.g.
                     * the left ear's back-left is simulated with the right
                     * ear's front-right, which uses the left ear's front-left
                     * measurement).
                     */
                    double az{field.mEvs[ei].mAzs[ai].mAzimuth};
                    if(az <= M_PI) az = M_PI - az;
                    else az = (M_PI*2.0)-az + M_PI;
                    CalcAzIndices(field, topElev, az, &a0, &a1, &af);

                    field.mEvs[ei].mAzs[ai].mDelays[0] = Lerp(
                        field.mEvs[topElev].mAzs[a0].mDelays[0],
                        field.mEvs[topElev].mAzs[a1].mDelays[0], af);
                }
            }
        }
        /* Record the lowest elevation filled in with the mirrored top. */
        const uint lowerElevFake{ei-1u};

        /* Fill in the remaining delays using bilinear interpolation. This
         * helps smooth the transition back to the real delays.
         */
        for(;ei < upperElevReal;++ei)
        {
            const double ef{(field.mEvs[upperElevReal].mElevation - field.mEvs[ei].mElevation) /
                (field.mEvs[upperElevReal].mElevation - field.mEvs[lowerElevFake].mElevation)};

            for(uint ai{0u};ai < field.mEvs[ei].mAzCount;ai++)
            {
                uint a0, a1, a2, a3;
                double af0, af1;

                double az{field.mEvs[ei].mAzs[ai].mAzimuth};
                CalcAzIndices(field, upperElevReal, az, &a0, &a1, &af0);
                CalcAzIndices(field, lowerElevFake, az, &a2, &a3, &af1);
                double blend[4]{
                    (1.0-ef) * (1.0-af0),
                    (1.0-ef) * (    af0),
                    (    ef) * (1.0-af1),
                    (    ef) * (    af1)
                };

                for(uint ti{0u};ti < channels;ti++)
                {
                    field.mEvs[ei].mAzs[ai].mDelays[ti] =
                        field.mEvs[upperElevReal].mAzs[a0].mDelays[ti]*blend[0] +
                        field.mEvs[upperElevReal].mAzs[a1].mDelays[ti]*blend[1] +
                        field.mEvs[lowerElevFake].mAzs[a2].mDelays[ti]*blend[2] +
                        field.mEvs[lowerElevFake].mAzs[a3].mDelays[ti]*blend[3];
                }
            }
        }
    };
    std::for_each(hData->mFds.begin(), hData->mFds.begin()+hData->mFdCount, proc_field);
}

/* Attempt to synthesize any missing HRIRs at the bottom elevations of each
 * field.  Right now this just blends the lowest elevation HRIRs together and
 * applies some attenuation and high frequency damping.  It is a simple, if
 * inaccurate model.
 */
static void SynthesizeHrirs(HrirDataT *hData)
{
    const uint channels{(hData->mChannelType == CT_STEREO) ? 2u : 1u};
    const uint irSize{hData->mIrPoints};
    const double beta{3.5e-6 * hData->mIrRate};

    auto proc_field = [channels,irSize,beta](HrirFdT &field) -> void
    {
        const uint oi{field.mEvStart};
        if(oi <= 0) return;

        for(uint ti{0u};ti < channels;ti++)
        {
            for(uint i{0u};i < irSize;i++)
                field.mEvs[0].mAzs[0].mIrs[ti][i] = 0.0;
            /* Blend the lowest defined elevation's responses for an average
             * -90 degree elevation response.
             */
            double blend_count{0.0};
            for(uint ai{0u};ai < field.mEvs[oi].mAzCount;ai++)
            {
                /* Only include the left responses for the left ear, and the
                 * right responses for the right ear. This removes the cross-
                 * talk that shouldn't exist for the -90 degree elevation
                 * response (and would be mistimed anyway). NOTE: Azimuth goes
                 * from 0...2pi rather than -pi...+pi (0 in front, clockwise).
                 */
                if(std::abs(field.mEvs[oi].mAzs[ai].mAzimuth) < EPSILON ||
                   (ti == LeftChannel && field.mEvs[oi].mAzs[ai].mAzimuth > M_PI-EPSILON) ||
                   (ti == RightChannel && field.mEvs[oi].mAzs[ai].mAzimuth < M_PI+EPSILON))
                {
                    for(uint i{0u};i < irSize;i++)
                        field.mEvs[0].mAzs[0].mIrs[ti][i] += field.mEvs[oi].mAzs[ai].mIrs[ti][i];
                    blend_count += 1.0;
                }
            }
            if(blend_count > 0.0)
            {
                for(uint i{0u};i < irSize;i++)
                    field.mEvs[0].mAzs[0].mIrs[ti][i] /= blend_count;
            }

            for(uint ei{1u};ei < field.mEvStart;ei++)
            {
                const double of{static_cast<double>(ei) / field.mEvStart};
                const double b{(1.0 - of) * beta};
                for(uint ai{0u};ai < field.mEvs[ei].mAzCount;ai++)
                {
                    uint a0, a1;
                    double af;

                    CalcAzIndices(field, oi, field.mEvs[ei].mAzs[ai].mAzimuth, &a0, &a1, &af);
                    double lp[4]{};
                    for(uint i{0u};i < irSize;i++)
                    {
                        /* Blend the two defined HRIRs closest to this azimuth,
                         * then blend that with the synthesized -90 elevation.
                         */
                        const double s1{Lerp(field.mEvs[oi].mAzs[a0].mIrs[ti][i],
                            field.mEvs[oi].mAzs[a1].mIrs[ti][i], af)};
                        const double s0{Lerp(field.mEvs[0].mAzs[0].mIrs[ti][i], s1, of)};
                        /* Apply a low-pass to simulate body occlusion. */
                        lp[0] = Lerp(s0, lp[0], b);
                        lp[1] = Lerp(lp[0], lp[1], b);
                        lp[2] = Lerp(lp[1], lp[2], b);
                        lp[3] = Lerp(lp[2], lp[3], b);
                        field.mEvs[ei].mAzs[ai].mIrs[ti][i] = lp[3];
                    }
                }
            }
            const double b{beta};
            double lp[4]{};
            for(uint i{0u};i < irSize;i++)
            {
                const double s0{field.mEvs[0].mAzs[0].mIrs[ti][i]};
                lp[0] = Lerp(s0, lp[0], b);
                lp[1] = Lerp(lp[0], lp[1], b);
                lp[2] = Lerp(lp[1], lp[2], b);
                lp[3] = Lerp(lp[2], lp[3], b);
                field.mEvs[0].mAzs[0].mIrs[ti][i] = lp[3];
            }
        }
        field.mEvStart = 0;
    };
    std::for_each(hData->mFds.begin(), hData->mFds.begin()+hData->mFdCount, proc_field);
}

// The following routines assume a full set of HRIRs for all elevations.

// Normalize the HRIR set and slightly attenuate the result.
static void NormalizeHrirs(HrirDataT *hData)
{
    const uint channels{(hData->mChannelType == CT_STEREO) ? 2u : 1u};
    const uint irSize{hData->mIrPoints};

    /* Find the maximum amplitude and RMS out of all the IRs. */
    struct LevelPair { double amp, rms; };
    auto proc0_field = [channels,irSize](const LevelPair levels, const HrirFdT &field) -> LevelPair
    {
        auto proc_elev = [channels,irSize](const LevelPair levels, const HrirEvT &elev) -> LevelPair
        {
            auto proc_azi = [channels,irSize](const LevelPair levels, const HrirAzT &azi) -> LevelPair
            {
                auto proc_channel = [irSize](const LevelPair levels, const double *ir) -> LevelPair
                {
                    /* Calculate the peak amplitude and RMS of this IR. */
                    auto current = std::accumulate(ir, ir+irSize, LevelPair{0.0, 0.0},
                        [](const LevelPair current, const double impulse) -> LevelPair
                        {
                            return LevelPair{std::max(std::abs(impulse), current.amp),
                                current.rms + impulse*impulse};
                        });
                    current.rms = std::sqrt(current.rms / irSize);

                    /* Accumulate levels by taking the maximum amplitude and RMS. */
                    return LevelPair{std::max(current.amp, levels.amp),
                        std::max(current.rms, levels.rms)};
                };
                return std::accumulate(azi.mIrs, azi.mIrs+channels, levels, proc_channel);
            };
            return std::accumulate(elev.mAzs, elev.mAzs+elev.mAzCount, levels, proc_azi);
        };
        return std::accumulate(field.mEvs, field.mEvs+field.mEvCount, levels, proc_elev);
    };
    const auto maxlev = std::accumulate(hData->mFds.begin(), hData->mFds.begin()+hData->mFdCount,
        LevelPair{0.0, 0.0}, proc0_field);

    /* Normalize using the maximum RMS of the HRIRs. The RMS measure for the
     * non-filtered signal is of an impulse with equal length (to the filter):
     *
     * rms_impulse = sqrt(sum([ 1^2, 0^2, 0^2, ... ]) / n)
     *             = sqrt(1 / n)
     *
     * This helps keep a more consistent volume between the non-filtered signal
     * and various data sets.
     */
    double factor{std::sqrt(1.0 / irSize) / maxlev.rms};

    /* Also ensure the samples themselves won't clip. */
    factor = std::min(factor, 0.99/maxlev.amp);

    /* Now scale all IRs by the given factor. */
    auto proc1_field = [channels,irSize,factor](HrirFdT &field) -> void
    {
        auto proc_elev = [channels,irSize,factor](HrirEvT &elev) -> void
        {
            auto proc_azi = [channels,irSize,factor](HrirAzT &azi) -> void
            {
                auto proc_channel = [irSize,factor](double *ir) -> void
                {
                    std::transform(ir, ir+irSize, ir,
                        std::bind(std::multiplies<double>{}, _1, factor));
                };
                std::for_each(azi.mIrs, azi.mIrs+channels, proc_channel);
            };
            std::for_each(elev.mAzs, elev.mAzs+elev.mAzCount, proc_azi);
        };
        std::for_each(field.mEvs, field.mEvs+field.mEvCount, proc_elev);
    };
    std::for_each(hData->mFds.begin(), hData->mFds.begin()+hData->mFdCount, proc1_field);
}

// Calculate the left-ear time delay using a spherical head model.
static double CalcLTD(const double ev, const double az, const double rad, const double dist)
{
    double azp, dlp, l, al;

    azp = std::asin(std::cos(ev) * std::sin(az));
    dlp = std::sqrt((dist*dist) + (rad*rad) + (2.0*dist*rad*sin(azp)));
    l = std::sqrt((dist*dist) - (rad*rad));
    al = (0.5 * M_PI) + azp;
    if(dlp > l)
        dlp = l + (rad * (al - std::acos(rad / dist)));
    return dlp / 343.3;
}

// Calculate the effective head-related time delays for each minimum-phase
// HRIR. This is done per-field since distance delay is ignored.
static void CalculateHrtds(const HeadModelT model, const double radius, HrirDataT *hData)
{
    uint channels = (hData->mChannelType == CT_STEREO) ? 2 : 1;
    double customRatio{radius / hData->mRadius};
    uint ti, fi, ei, ai;

    if(model == HM_SPHERE)
    {
        for(fi = 0;fi < hData->mFdCount;fi++)
        {
            for(ei = 0;ei < hData->mFds[fi].mEvCount;ei++)
            {
                HrirEvT *evd = &hData->mFds[fi].mEvs[ei];

                for(ai = 0;ai < evd->mAzCount;ai++)
                {
                    HrirAzT *azd = &evd->mAzs[ai];

                    for(ti = 0;ti < channels;ti++)
                        azd->mDelays[ti] = CalcLTD(evd->mElevation, azd->mAzimuth, radius, hData->mFds[fi].mDistance);
                }
            }
        }
    }
    else if(customRatio != 1.0)
    {
        for(fi = 0;fi < hData->mFdCount;fi++)
        {
            for(ei = 0;ei < hData->mFds[fi].mEvCount;ei++)
            {
                HrirEvT *evd = &hData->mFds[fi].mEvs[ei];

                for(ai = 0;ai < evd->mAzCount;ai++)
                {
                    HrirAzT *azd = &evd->mAzs[ai];
                    for(ti = 0;ti < channels;ti++)
                        azd->mDelays[ti] *= customRatio;
                }
            }
        }
    }

    for(fi = 0;fi < hData->mFdCount;fi++)
    {
        double minHrtd{std::numeric_limits<double>::infinity()};
        for(ei = 0;ei < hData->mFds[fi].mEvCount;ei++)
        {
            for(ai = 0;ai < hData->mFds[fi].mEvs[ei].mAzCount;ai++)
            {
                HrirAzT *azd = &hData->mFds[fi].mEvs[ei].mAzs[ai];

                for(ti = 0;ti < channels;ti++)
                    minHrtd = std::min(azd->mDelays[ti], minHrtd);
            }
        }

        for(ei = 0;ei < hData->mFds[fi].mEvCount;ei++)
        {
            for(ti = 0;ti < channels;ti++)
            {
                for(ai = 0;ai < hData->mFds[fi].mEvs[ei].mAzCount;ai++)
                    hData->mFds[fi].mEvs[ei].mAzs[ai].mDelays[ti] -= minHrtd;
            }
        }
    }
}

// Allocate and configure dynamic HRIR structures.
int PrepareHrirData(const uint fdCount, const double (&distances)[MAX_FD_COUNT],
    const uint (&evCounts)[MAX_FD_COUNT], const uint azCounts[MAX_FD_COUNT * MAX_EV_COUNT],
    HrirDataT *hData)
{
    uint evTotal = 0, azTotal = 0, fi, ei, ai;

    for(fi = 0;fi < fdCount;fi++)
    {
        evTotal += evCounts[fi];
        for(ei = 0;ei < evCounts[fi];ei++)
            azTotal += azCounts[(fi * MAX_EV_COUNT) + ei];
    }
    if(!fdCount || !evTotal || !azTotal)
        return 0;

    hData->mEvsBase.resize(evTotal);
    hData->mAzsBase.resize(azTotal);
    hData->mFds.resize(fdCount);
    hData->mIrCount = azTotal;
    hData->mFdCount = fdCount;
    evTotal = 0;
    azTotal = 0;
    for(fi = 0;fi < fdCount;fi++)
    {
        hData->mFds[fi].mDistance = distances[fi];
        hData->mFds[fi].mEvCount = evCounts[fi];
        hData->mFds[fi].mEvStart = 0;
        hData->mFds[fi].mEvs = &hData->mEvsBase[evTotal];
        evTotal += evCounts[fi];
        for(ei = 0;ei < evCounts[fi];ei++)
        {
            uint azCount = azCounts[(fi * MAX_EV_COUNT) + ei];

            hData->mFds[fi].mIrCount += azCount;
            hData->mFds[fi].mEvs[ei].mElevation = -M_PI / 2.0 + M_PI * ei / (evCounts[fi] - 1);
            hData->mFds[fi].mEvs[ei].mIrCount += azCount;
            hData->mFds[fi].mEvs[ei].mAzCount = azCount;
            hData->mFds[fi].mEvs[ei].mAzs = &hData->mAzsBase[azTotal];
            for(ai = 0;ai < azCount;ai++)
            {
                hData->mFds[fi].mEvs[ei].mAzs[ai].mAzimuth = 2.0 * M_PI * ai / azCount;
                hData->mFds[fi].mEvs[ei].mAzs[ai].mIndex = azTotal + ai;
                hData->mFds[fi].mEvs[ei].mAzs[ai].mDelays[0] = 0.0;
                hData->mFds[fi].mEvs[ei].mAzs[ai].mDelays[1] = 0.0;
                hData->mFds[fi].mEvs[ei].mAzs[ai].mIrs[0] = nullptr;
                hData->mFds[fi].mEvs[ei].mAzs[ai].mIrs[1] = nullptr;
            }
            azTotal += azCount;
        }
    }
    return 1;
}


/* Parse the data set definition and process the source data, storing the
 * resulting data set as desired.  If the input name is NULL it will read
 * from standard input.
 */
static int ProcessDefinition(const char *inName, const uint outRate, const ChannelModeT chanMode, const uint fftSize, const int equalize, const int surface, const double limit, const uint truncSize, const HeadModelT model, const double radius, const char *outName)
{
    char rateStr[8+1], expName[MAX_PATH_LEN];
    char startbytes[4]{};
    size_t startbytecount{0u};
    HrirDataT hData;
    FILE *fp;
    int ret;

    if(!inName)
    {
        inName = "stdin";
        fp = stdin;
    }
    else
    {
        fp = fopen(inName, "r");
        if(fp == nullptr)
        {
            fprintf(stderr, "Error: Could not open input file '%s'\n", inName);
            return 0;
        }

        startbytecount = fread(startbytes, 1, sizeof(startbytes), fp);
        if(startbytecount != sizeof(startbytes))
        {
            fclose(fp);
            fprintf(stderr, "Error: Could not read input file '%s'\n", inName);
            return 0;
        }

        if(startbytes[0] == '\x89' && startbytes[1] == 'H' && startbytes[2] == 'D' &&
           startbytes[3] == 'F')
        {
            fclose(fp);
            fp = nullptr;

            fprintf(stdout, "Reading HRTF data from %s...\n", inName);
            if(!LoadSofaFile(inName, fftSize, truncSize, chanMode, &hData))
                return 0;
        }
    }
    if(fp != nullptr)
    {
        fprintf(stdout, "Reading HRIR definition from %s...\n", inName);
        const bool success{LoadDefInput(fp, startbytes, startbytecount, inName, fftSize, truncSize,
            chanMode, &hData)};
        if(fp != stdin)
            fclose(fp);
        if(!success)
            return 0;
    }

    if(equalize)
    {
        uint c = (hData.mChannelType == CT_STEREO) ? 2 : 1;
        uint m = 1 + hData.mFftSize / 2;
        std::vector<double> dfa(c * m);

        if(hData.mFdCount > 1)
        {
            fprintf(stdout, "Balancing field magnitudes...\n");
            BalanceFieldMagnitudes(&hData, c, m);
        }
        fprintf(stdout, "Calculating diffuse-field average...\n");
        CalculateDiffuseFieldAverage(&hData, c, m, surface, limit, dfa.data());
        fprintf(stdout, "Performing diffuse-field equalization...\n");
        DiffuseFieldEqualize(c, m, dfa.data(), &hData);
    }
    fprintf(stdout, "Performing minimum phase reconstruction...\n");
    ReconstructHrirs(&hData);
    if(outRate != 0 && outRate != hData.mIrRate)
    {
        fprintf(stdout, "Resampling HRIRs...\n");
        ResampleHrirs(outRate, &hData);
    }
    fprintf(stdout, "Truncating minimum-phase HRIRs...\n");
    hData.mIrPoints = truncSize;
    fprintf(stdout, "Synthesizing missing elevations...\n");
    if(model == HM_DATASET)
        SynthesizeOnsets(&hData);
    SynthesizeHrirs(&hData);
    fprintf(stdout, "Normalizing final HRIRs...\n");
    NormalizeHrirs(&hData);
    fprintf(stdout, "Calculating impulse delays...\n");
    CalculateHrtds(model, (radius > DEFAULT_CUSTOM_RADIUS) ? radius : hData.mRadius, &hData);
    snprintf(rateStr, 8, "%u", hData.mIrRate);
    StrSubst(outName, "%r", rateStr, MAX_PATH_LEN, expName);
    fprintf(stdout, "Creating MHR data set %s...\n", expName);
    ret = StoreMhr(&hData, expName);

    return ret;
}

static void PrintHelp(const char *argv0, FILE *ofile)
{
    fprintf(ofile, "Usage:  %s [<option>...]\n\n", argv0);
    fprintf(ofile, "Options:\n");
    fprintf(ofile, " -r <rate>       Change the data set sample rate to the specified value and\n");
    fprintf(ofile, "                 resample the HRIRs accordingly.\n");
    fprintf(ofile, " -m              Change the data set to mono, mirroring the left ear for the\n");
    fprintf(ofile, "                 right ear.\n");
    fprintf(ofile, " -f <points>     Override the FFT window size (default: %u).\n", DEFAULT_FFTSIZE);
    fprintf(ofile, " -e {on|off}     Toggle diffuse-field equalization (default: %s).\n", (DEFAULT_EQUALIZE ? "on" : "off"));
    fprintf(ofile, " -s {on|off}     Toggle surface-weighted diffuse-field average (default: %s).\n", (DEFAULT_SURFACE ? "on" : "off"));
    fprintf(ofile, " -l {<dB>|none}  Specify a limit to the magnitude range of the diffuse-field\n");
    fprintf(ofile, "                 average (default: %.2f).\n", DEFAULT_LIMIT);
    fprintf(ofile, " -w <points>     Specify the size of the truncation window that's applied\n");
    fprintf(ofile, "                 after minimum-phase reconstruction (default: %u).\n", DEFAULT_TRUNCSIZE);
    fprintf(ofile, " -d {dataset|    Specify the model used for calculating the head-delay timing\n");
    fprintf(ofile, "     sphere}     values (default: %s).\n", ((DEFAULT_HEAD_MODEL == HM_DATASET) ? "dataset" : "sphere"));
    fprintf(ofile, " -c <radius>     Use a customized head radius measured to-ear in meters.\n");
    fprintf(ofile, " -i <filename>   Specify an HRIR definition file to use (defaults to stdin).\n");
    fprintf(ofile, " -o <filename>   Specify an output file. Use of '%%r' will be substituted with\n");
    fprintf(ofile, "                 the data set sample rate.\n");
}

// Standard command line dispatch.
int main(int argc, char *argv[])
{
    const char *inName = nullptr, *outName = nullptr;
    uint outRate, fftSize;
    int equalize, surface;
    char *end = nullptr;
    ChannelModeT chanMode;
    HeadModelT model;
    uint truncSize;
    double radius;
    double limit;
    int opt;

    GET_UNICODE_ARGS(&argc, &argv);

    if(argc < 2)
    {
        fprintf(stdout, "HRTF Processing and Composition Utility\n\n");
        PrintHelp(argv[0], stdout);
        exit(EXIT_SUCCESS);
    }

    outName = "./oalsoft_hrtf_%r.mhr";
    outRate = 0;
    chanMode = CM_AllowStereo;
    fftSize = DEFAULT_FFTSIZE;
    equalize = DEFAULT_EQUALIZE;
    surface = DEFAULT_SURFACE;
    limit = DEFAULT_LIMIT;
    truncSize = DEFAULT_TRUNCSIZE;
    model = DEFAULT_HEAD_MODEL;
    radius = DEFAULT_CUSTOM_RADIUS;

    while((opt=getopt(argc, argv, "r:mf:e:s:l:w:d:c:e:i:o:h")) != -1)
    {
        switch(opt)
        {
        case 'r':
            outRate = strtoul(optarg, &end, 10);
            if(end[0] != '\0' || outRate < MIN_RATE || outRate > MAX_RATE)
            {
                fprintf(stderr, "\nError: Got unexpected value \"%s\" for option -%c, expected between %u to %u.\n", optarg, opt, MIN_RATE, MAX_RATE);
                exit(EXIT_FAILURE);
            }
            break;

        case 'm':
            chanMode = CM_ForceMono;
            break;

        case 'f':
            fftSize = strtoul(optarg, &end, 10);
            if(end[0] != '\0' || (fftSize&(fftSize-1)) || fftSize < MIN_FFTSIZE || fftSize > MAX_FFTSIZE)
            {
                fprintf(stderr, "\nError: Got unexpected value \"%s\" for option -%c, expected a power-of-two between %u to %u.\n", optarg, opt, MIN_FFTSIZE, MAX_FFTSIZE);
                exit(EXIT_FAILURE);
            }
            break;

        case 'e':
            if(strcmp(optarg, "on") == 0)
                equalize = 1;
            else if(strcmp(optarg, "off") == 0)
                equalize = 0;
            else
            {
                fprintf(stderr, "\nError: Got unexpected value \"%s\" for option -%c, expected on or off.\n", optarg, opt);
                exit(EXIT_FAILURE);
            }
            break;

        case 's':
            if(strcmp(optarg, "on") == 0)
                surface = 1;
            else if(strcmp(optarg, "off") == 0)
                surface = 0;
            else
            {
                fprintf(stderr, "\nError: Got unexpected value \"%s\" for option -%c, expected on or off.\n", optarg, opt);
                exit(EXIT_FAILURE);
            }
            break;

        case 'l':
            if(strcmp(optarg, "none") == 0)
                limit = 0.0;
            else
            {
                limit = strtod(optarg, &end);
                if(end[0] != '\0' || limit < MIN_LIMIT || limit > MAX_LIMIT)
                {
                    fprintf(stderr, "\nError: Got unexpected value \"%s\" for option -%c, expected between %.0f to %.0f.\n", optarg, opt, MIN_LIMIT, MAX_LIMIT);
                    exit(EXIT_FAILURE);
                }
            }
            break;

        case 'w':
            truncSize = strtoul(optarg, &end, 10);
            if(end[0] != '\0' || truncSize < MIN_TRUNCSIZE || truncSize > MAX_TRUNCSIZE || (truncSize%MOD_TRUNCSIZE))
            {
                fprintf(stderr, "\nError: Got unexpected value \"%s\" for option -%c, expected multiple of %u between %u to %u.\n", optarg, opt, MOD_TRUNCSIZE, MIN_TRUNCSIZE, MAX_TRUNCSIZE);
                exit(EXIT_FAILURE);
            }
            break;

        case 'd':
            if(strcmp(optarg, "dataset") == 0)
                model = HM_DATASET;
            else if(strcmp(optarg, "sphere") == 0)
                model = HM_SPHERE;
            else
            {
                fprintf(stderr, "\nError: Got unexpected value \"%s\" for option -%c, expected dataset or sphere.\n", optarg, opt);
                exit(EXIT_FAILURE);
            }
            break;

        case 'c':
            radius = strtod(optarg, &end);
            if(end[0] != '\0' || radius < MIN_CUSTOM_RADIUS || radius > MAX_CUSTOM_RADIUS)
            {
                fprintf(stderr, "\nError: Got unexpected value \"%s\" for option -%c, expected between %.2f to %.2f.\n", optarg, opt, MIN_CUSTOM_RADIUS, MAX_CUSTOM_RADIUS);
                exit(EXIT_FAILURE);
            }
            break;

        case 'i':
            inName = optarg;
            break;

        case 'o':
            outName = optarg;
            break;

        case 'h':
            PrintHelp(argv[0], stdout);
            exit(EXIT_SUCCESS);

        default: /* '?' */
            PrintHelp(argv[0], stderr);
            exit(EXIT_FAILURE);
        }
    }

    int ret = ProcessDefinition(inName, outRate, chanMode, fftSize, equalize, surface, limit,
        truncSize, model, radius, outName);
    if(!ret) return -1;
    fprintf(stdout, "Operation completed.\n");

    return EXIT_SUCCESS;
}