utils/uhjdecoder.cpp

   1 /*
   2  * 2-channel UHJ Decoder
   3  *
   4  * Copyright (c) Chris Robinson <chris.kcat@gmail.com>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  */
  24
  25 #include "config.h"
  26
  27 #include <array>
  28 #include <complex>
  29 #include <cstring>
  30 #include <memory>
  31 #include <stddef.h>
  32 #include <string>
  33 #include <utility>
  34 #include <vector>
  35
  36 #include "albit.h"
  37 #include "albyte.h"
  38 #include "alcomplex.h"
  39 #include "almalloc.h"
  40 #include "alnumbers.h"
  41 #include "alspan.h"
  42 #include "vector.h"
  43 #include "opthelpers.h"
  44 #include "phase_shifter.h"
  45
  46 #include "sndfile.h"
  47
  48 #include "win_main_utf8.h"
  49
  50
  51 struct FileDeleter {
  52     void operator()(FILE *file) { fclose(file); }
  53 };
  54 using FilePtr = std::unique_ptr<FILE,FileDeleter>;
  55
  56 struct SndFileDeleter {
  57     void operator()(SNDFILE *sndfile) { sf_close(sndfile); }
  58 };
  59 using SndFilePtr = std::unique_ptr<SNDFILE,SndFileDeleter>;
  60
  61
  62 using ubyte = unsigned char;
  63 using ushort = unsigned short;
  64 using uint = unsigned int;
  65 using complex_d = std::complex<double>;
  66
  67 using byte4 = std::array<al::byte,4>;
  68
  69
  70 constexpr ubyte SUBTYPE_BFORMAT_FLOAT[]{
  71     0x03, 0x00, 0x00, 0x00, 0x21, 0x07, 0xd3, 0x11, 0x86, 0x44, 0xc8, 0xc1,
  72     0xca, 0x00, 0x00, 0x00
  73 };
  74
  75 void fwrite16le(ushort val, FILE *f)
  76 {
  77     ubyte data[2]{ static_cast<ubyte>(val&0xff), static_cast<ubyte>((val>>8)&0xff) };
  78     fwrite(data, 1, 2, f);
  79 }
  80
  81 void fwrite32le(uint val, FILE *f)
  82 {
  83     ubyte data[4]{ static_cast<ubyte>(val&0xff), static_cast<ubyte>((val>>8)&0xff),
  84         static_cast<ubyte>((val>>16)&0xff), static_cast<ubyte>((val>>24)&0xff) };
  85     fwrite(data, 1, 4, f);
  86 }
  87
  88 template<al::endian = al::endian::native>
  89 byte4 f32AsLEBytes(const float &value) = delete;
  90
  91 template<>
  92 byte4 f32AsLEBytes<al::endian::little>(const float &value)
  93 {
  94     byte4 ret{};
  95     std::memcpy(ret.data(), &value, 4);
  96     return ret;
  97 }
  98 template<>
  99 byte4 f32AsLEBytes<al::endian::big>(const float &value)
 100 {
 101     byte4 ret{};
 102     std::memcpy(ret.data(), &value, 4);
 103     std::swap(ret[0], ret[3]);
 104     std::swap(ret[1], ret[2]);
 105     return ret;
 106 }
 107
 108
 109 constexpr uint BufferLineSize{1024};
 110
 111 using FloatBufferLine = std::array<float,BufferLineSize>;
 112 using FloatBufferSpan = al::span<float,BufferLineSize>;
 113
 114
 115 struct UhjDecoder {
 116     constexpr static size_t sFilterDelay{1024};
 117
 118     alignas(16) std::array<float,BufferLineSize+sFilterDelay> mS{};
 119     alignas(16) std::array<float,BufferLineSize+sFilterDelay> mD{};
 120     alignas(16) std::array<float,BufferLineSize+sFilterDelay> mT{};
 121     alignas(16) std::array<float,BufferLineSize+sFilterDelay> mQ{};
 122
 123     /* History for the FIR filter. */
 124     alignas(16) std::array<float,sFilterDelay-1> mDTHistory{};
 125     alignas(16) std::array<float,sFilterDelay-1> mSHistory{};
 126
 127     alignas(16) std::array<float,BufferLineSize + sFilterDelay*2> mTemp{};
 128
 129     void decode(const float *RESTRICT InSamples, const size_t InChannels,
 130         const al::span<FloatBufferLine> OutSamples, const size_t SamplesToDo);
 131     void decode2(const float *RESTRICT InSamples, const al::span<FloatBufferLine,3> OutSamples,
 132         const size_t SamplesToDo);
 133
 134     DEF_NEWDEL(UhjDecoder)
 135 };
 136
 137 const PhaseShifterT<UhjDecoder::sFilterDelay*2> PShift{};
 138
 139
 140 /* Decoding UHJ is done as:
 141  *
 142  * S = Left + Right
 143  * D = Left - Right
 144  *
 145  * W = 0.981532*S + 0.197484*j(0.828331*D + 0.767820*T)
 146  * X = 0.418496*S - j(0.828331*D + 0.767820*T)
 147  * Y = 0.795968*D - 0.676392*T + j(0.186633*S)
 148  * Z = 1.023332*Q
 149  *
 150  * where j is a +90 degree phase shift. 3-channel UHJ excludes Q, while 2-
 151  * channel excludes Q and T. The B-Format signal reconstructed from 2-channel
 152  * UHJ should not be run through a normal B-Format decoder, as it needs
 153  * different shelf filters.
 154  *
 155  * NOTE: Some sources specify
 156  *
 157  * S = (Left + Right)/2
 158  * D = (Left - Right)/2
 159  *
 160  * However, this is incorrect. It's halving Left and Right even though they
 161  * were already halved during encoding, causing S and D to be half what they
 162  * initially were at the encoding stage. This division is not present in
 163  * Gerzon's original paper for deriving Sigma (S) or Delta (D) from the L and R
 164  * signals. As proof, taking Y for example:
 165  *
 166  * Y = 0.795968*D - 0.676392*T + j(0.186633*S)
 167  *
 168  * * Plug in the encoding parameters, using ? as a placeholder for whether S
 169  *   and D should receive an extra 0.5 factor
 170  * Y = 0.795968*(j(-0.3420201*W + 0.5098604*X) + 0.6554516*Y)*? -
 171  *     0.676392*(j(-0.1432*W + 0.6512*X) - 0.7071068*Y) +
 172  *     0.186633*j(0.9396926*W + 0.1855740*X)*?
 173  *
 174  * * Move common factors in
 175  * Y = (j(-0.3420201*0.795968*?*W + 0.5098604*0.795968*?*X) + 0.6554516*0.795968*?*Y) -
 176  *     (j(-0.1432*0.676392*W + 0.6512*0.676392*X) - 0.7071068*0.676392*Y) +
 177  *     j(0.9396926*0.186633*?*W + 0.1855740*0.186633*?*X)
 178  *
 179  * * Clean up extraneous groupings
 180  * Y = j(-0.3420201*0.795968*?*W + 0.5098604*0.795968*?*X) + 0.6554516*0.795968*?*Y -
 181  *     j(-0.1432*0.676392*W + 0.6512*0.676392*X) + 0.7071068*0.676392*Y +
 182  *     j*(0.9396926*0.186633*?*W + 0.1855740*0.186633*?*X)
 183  *
 184  * * Move phase shifts together and combine them
 185  * Y = j(-0.3420201*0.795968*?*W + 0.5098604*0.795968*?*X - -0.1432*0.676392*W -
 186  *        0.6512*0.676392*X + 0.9396926*0.186633*?*W + 0.1855740*0.186633*?*X) +
 187  *     0.6554516*0.795968*?*Y + 0.7071068*0.676392*Y
 188  *
 189  * * Reorder terms
 190  * Y = j(-0.3420201*0.795968*?*W +  0.1432*0.676392*W + 0.9396926*0.186633*?*W +
 191  *        0.5098604*0.795968*?*X + -0.6512*0.676392*X + 0.1855740*0.186633*?*X) +
 192  *     0.7071068*0.676392*Y + 0.6554516*0.795968*?*Y
 193  *
 194  * * Move common factors out
 195  * Y = j((-0.3420201*0.795968*? +  0.1432*0.676392 + 0.9396926*0.186633*?)*W +
 196  *       ( 0.5098604*0.795968*? + -0.6512*0.676392 + 0.1855740*0.186633*?)*X) +
 197  *     (0.7071068*0.676392 + 0.6554516*0.795968*?)*Y
 198  *
 199  * * Result w/ 0.5 factor:
 200  * -0.3420201*0.795968*0.5 +  0.1432*0.676392 + 0.9396926*0.186633*0.5 =  0.04843*W
 201  *  0.5098604*0.795968*0.5 + -0.6512*0.676392 + 0.1855740*0.186633*0.5 = -0.22023*X
 202  *  0.7071068*0.676392                        + 0.6554516*0.795968*0.5 =  0.73914*Y
 203  * -> Y = j(0.04843*W + -0.22023*X) + 0.73914*Y
 204  *
 205  * * Result w/o 0.5 factor:
 206  * -0.3420201*0.795968 +  0.1432*0.676392 + 0.9396926*0.186633 = 0.00000*W
 207  *  0.5098604*0.795968 + -0.6512*0.676392 + 0.1855740*0.186633 = 0.00000*X
 208  *  0.7071068*0.676392                    + 0.6554516*0.795968 = 1.00000*Y
 209  * -> Y = j(0.00000*W + 0.00000*X) + 1.00000*Y
 210  *
 211  * Not halving produces a result matching the original input.
 212  */
 213 void UhjDecoder::decode(const float *RESTRICT InSamples, const size_t InChannels,
 214     const al::span<FloatBufferLine> OutSamples, const size_t SamplesToDo)
 215 {
 216     ASSUME(SamplesToDo > 0);
 217
 218     float *woutput{OutSamples[0].data()};
 219     float *xoutput{OutSamples[1].data()};
 220     float *youtput{OutSamples[2].data()};
 221
 222     /* Add a delay to the input channels, to align it with the all-passed
 223      * signal.
 224      */
 225
 226     /* S = Left + Right */
 227     for(size_t i{0};i < SamplesToDo;++i)
 228         mS[sFilterDelay+i] = InSamples[i*InChannels + 0] + InSamples[i*InChannels + 1];
 229
 230     /* D = Left - Right */
 231     for(size_t i{0};i < SamplesToDo;++i)
 232         mD[sFilterDelay+i] = InSamples[i*InChannels + 0] - InSamples[i*InChannels + 1];
 233
 234     if(InChannels > 2)
 235     {
 236         /* T */
 237         for(size_t i{0};i < SamplesToDo;++i)
 238             mT[sFilterDelay+i] = InSamples[i*InChannels + 2];
 239     }
 240     if(InChannels > 3)
 241     {
 242         /* Q */
 243         for(size_t i{0};i < SamplesToDo;++i)
 244             mQ[sFilterDelay+i] = InSamples[i*InChannels + 3];
 245     }
 246
 247     /* Precompute j(0.828347*D + 0.767835*T) and store in xoutput. */
 248     auto tmpiter = std::copy(mDTHistory.cbegin(), mDTHistory.cend(), mTemp.begin());
 249     std::transform(mD.cbegin(), mD.cbegin()+SamplesToDo+sFilterDelay, mT.cbegin(), tmpiter,
 250         [](const float d, const float t) noexcept { return 0.828347f*d + 0.767835f*t; });
 251     std::copy_n(mTemp.cbegin()+SamplesToDo, mDTHistory.size(), mDTHistory.begin());
 252     PShift.process({xoutput, SamplesToDo}, mTemp.data());
 253
 254     for(size_t i{0};i < SamplesToDo;++i)
 255     {
 256         /* W = 0.981532*S + 0.197484*j(0.828347*D + 0.767835*T) */
 257         woutput[i] = 0.981532f*mS[i] + 0.197484f*xoutput[i];
 258         /* X = 0.418504*S - j(0.828347*D + 0.767835*T) */
 259         xoutput[i] = 0.418504f*mS[i] - xoutput[i];
 260     }
 261
 262     /* Precompute j*S and store in youtput. */
 263     tmpiter = std::copy(mSHistory.cbegin(), mSHistory.cend(), mTemp.begin());
 264     std::copy_n(mS.cbegin(), SamplesToDo+sFilterDelay, tmpiter);
 265     std::copy_n(mTemp.cbegin()+SamplesToDo, mSHistory.size(), mSHistory.begin());
 266     PShift.process({youtput, SamplesToDo}, mTemp.data());
 267
 268     for(size_t i{0};i < SamplesToDo;++i)
 269     {
 270         /* Y = 0.795968*D - 0.676392*T + j(0.186633*S) */
 271         youtput[i] = 0.795968f*mD[i] - 0.676392f*mT[i] + 0.186633f*youtput[i];
 272     }
 273
 274     if(OutSamples.size() > 3)
 275     {
 276         float *zoutput{OutSamples[3].data()};
 277         /* Z = 1.023332*Q */
 278         for(size_t i{0};i < SamplesToDo;++i)
 279             zoutput[i] = 1.023332f*mQ[i];
 280     }
 281
 282     std::copy(mS.begin()+SamplesToDo, mS.begin()+SamplesToDo+sFilterDelay, mS.begin());
 283     std::copy(mD.begin()+SamplesToDo, mD.begin()+SamplesToDo+sFilterDelay, mD.begin());
 284     std::copy(mT.begin()+SamplesToDo, mT.begin()+SamplesToDo+sFilterDelay, mT.begin());
 285     std::copy(mQ.begin()+SamplesToDo, mQ.begin()+SamplesToDo+sFilterDelay, mQ.begin());
 286 }
 287
 288 /* This is an alternative equation for decoding 2-channel UHJ. Not sure what
 289  * the intended benefit is over the above equation as this slightly reduces the
 290  * amount of the original left response and has more of the phase-shifted
 291  * forward response on the left response.
 292  *
 293  * This decoding is done as:
 294  *
 295  * S = Left + Right
 296  * D = Left - Right
 297  *
 298  * W = 0.981530*S + j*0.163585*D
 299  * X = 0.418504*S - j*0.828347*D
 300  * Y = 0.762956*D + j*0.384230*S
 301  *
 302  * where j is a +90 degree phase shift.
 303  *
 304  * NOTE: As above, S and D should not be halved. The only consequence of
 305  * halving here is merely a -6dB reduction in output, but it's still incorrect.
 306  */
 307 void UhjDecoder::decode2(const float *RESTRICT InSamples,
 308     const al::span<FloatBufferLine,3> OutSamples, const size_t SamplesToDo)
 309 {
 310     ASSUME(SamplesToDo > 0);
 311
 312     float *woutput{OutSamples[0].data()};
 313     float *xoutput{OutSamples[1].data()};
 314     float *youtput{OutSamples[2].data()};
 315
 316     /* S = Left + Right */
 317     for(size_t i{0};i < SamplesToDo;++i)
 318         mS[sFilterDelay+i] = InSamples[i*2 + 0] + InSamples[i*2 + 1];
 319
 320     /* D = Left - Right */
 321     for(size_t i{0};i < SamplesToDo;++i)
 322         mD[sFilterDelay+i] = InSamples[i*2 + 0] - InSamples[i*2 + 1];
 323
 324     /* Precompute j*D and store in xoutput. */
 325     auto tmpiter = std::copy(mDTHistory.cbegin(), mDTHistory.cend(), mTemp.begin());
 326     std::copy_n(mD.cbegin(), SamplesToDo+sFilterDelay, tmpiter);
 327     std::copy_n(mTemp.cbegin()+SamplesToDo, mDTHistory.size(), mDTHistory.begin());
 328     PShift.process({xoutput, SamplesToDo}, mTemp.data());
 329
 330     for(size_t i{0};i < SamplesToDo;++i)
 331     {
 332         /* W = 0.981530*S + j*0.163585*D */
 333         woutput[i] = 0.981530f*mS[i] + 0.163585f*xoutput[i];
 334         /* X = 0.418504*S - j*0.828347*D */
 335         xoutput[i] = 0.418504f*mS[i] - 0.828347f*xoutput[i];
 336     }
 337
 338     /* Precompute j*S and store in youtput. */
 339     tmpiter = std::copy(mSHistory.cbegin(), mSHistory.cend(), mTemp.begin());
 340     std::copy_n(mS.cbegin(), SamplesToDo+sFilterDelay, tmpiter);
 341     std::copy_n(mTemp.cbegin()+SamplesToDo, mSHistory.size(), mSHistory.begin());
 342     PShift.process({youtput, SamplesToDo}, mTemp.data());
 343
 344     for(size_t i{0};i < SamplesToDo;++i)
 345     {
 346         /* Y = 0.762956*D + j*0.384230*S */
 347         youtput[i] = 0.762956f*mD[i] + 0.384230f*youtput[i];
 348     }
 349
 350     std::copy(mS.begin()+SamplesToDo, mS.begin()+SamplesToDo+sFilterDelay, mS.begin());
 351     std::copy(mD.begin()+SamplesToDo, mD.begin()+SamplesToDo+sFilterDelay, mD.begin());
 352 }
 353
 354
 355 int main(int argc, char **argv)
 356 {
 357     if(argc < 2 || std::strcmp(argv[1], "-h") == 0 || std::strcmp(argv[1], "--help") == 0)
 358     {
 359         printf("Usage: %s <[options] filename.wav...>\n\n"
 360             "  Options:\n"
 361             "    --general      Use the general equations for 2-channel UHJ (default).\n"
 362             "    --alternative  Use the alternative equations for 2-channel UHJ.\n"
 363             "\n"
 364             "Note: When decoding 2-channel UHJ to an .amb file, the result should not use\n"
 365             "the normal B-Format shelf filters! Only 3- and 4-channel UHJ can accurately\n"
 366             "reconstruct the original B-Format signal.",
 367             argv[0]);
 368         return 1;
 369     }
 370
 371     size_t num_files{0}, num_decoded{0};
 372     bool use_general{true};
 373     for(int fidx{1};fidx < argc;++fidx)
 374     {
 375         if(std::strcmp(argv[fidx], "--general") == 0)
 376         {
 377             use_general = true;
 378             continue;
 379         }
 380         if(std::strcmp(argv[fidx], "--alternative") == 0)
 381         {
 382             use_general = false;
 383             continue;
 384         }
 385         ++num_files;
 386         SF_INFO ininfo{};
 387         SndFilePtr infile{sf_open(argv[fidx], SFM_READ, &ininfo)};
 388         if(!infile)
 389         {
 390             fprintf(stderr, "Failed to open %s\n", argv[fidx]);
 391             continue;
 392         }
 393         if(sf_command(infile.get(), SFC_WAVEX_GET_AMBISONIC, NULL, 0) == SF_AMBISONIC_B_FORMAT)
 394         {
 395             fprintf(stderr, "%s is already B-Format\n", argv[fidx]);
 396             continue;
 397         }
 398         uint outchans{};
 399         if(ininfo.channels == 2)
 400             outchans = 3;
 401         else if(ininfo.channels == 3 || ininfo.channels == 4)
 402             outchans = static_cast<uint>(ininfo.channels);
 403         else
 404         {
 405             fprintf(stderr, "%s is not a 2-, 3-, or 4-channel file\n", argv[fidx]);
 406             continue;
 407         }
 408         printf("Converting %s from %d-channel UHJ%s...\n", argv[fidx], ininfo.channels,
 409             (ininfo.channels == 2) ? use_general ? " (general)" : " (alternative)" : "");
 410
 411         std::string outname{argv[fidx]};
 412         auto lastslash = outname.find_last_of('/');
 413         if(lastslash != std::string::npos)
 414             outname.erase(0, lastslash+1);
 415         auto lastdot = outname.find_last_of('.');
 416         if(lastdot != std::string::npos)
 417             outname.resize(lastdot+1);
 418         outname += "amb";
 419
 420         FilePtr outfile{fopen(outname.c_str(), "wb")};
 421         if(!outfile)
 422         {
 423             fprintf(stderr, "Failed to create %s\n", outname.c_str());
 424             continue;
 425         }
 426
 427         fputs("RIFF", outfile.get());
 428         fwrite32le(0xFFFFFFFF, outfile.get()); // 'RIFF' header len; filled in at close
 429
 430         fputs("WAVE", outfile.get());
 431
 432         fputs("fmt ", outfile.get());
 433         fwrite32le(40, outfile.get()); // 'fmt ' header len; 40 bytes for EXTENSIBLE
 434
 435         // 16-bit val, format type id (extensible: 0xFFFE)
 436         fwrite16le(0xFFFE, outfile.get());
 437         // 16-bit val, channel count
 438         fwrite16le(static_cast<ushort>(outchans), outfile.get());
 439         // 32-bit val, frequency
 440         fwrite32le(static_cast<uint>(ininfo.samplerate), outfile.get());
 441         // 32-bit val, bytes per second
 442         fwrite32le(static_cast<uint>(ininfo.samplerate)*sizeof(float)*outchans, outfile.get());
 443         // 16-bit val, frame size
 444         fwrite16le(static_cast<ushort>(sizeof(float)*outchans), outfile.get());
 445         // 16-bit val, bits per sample
 446         fwrite16le(static_cast<ushort>(sizeof(float)*8), outfile.get());
 447         // 16-bit val, extra byte count
 448         fwrite16le(22, outfile.get());
 449         // 16-bit val, valid bits per sample
 450         fwrite16le(static_cast<ushort>(sizeof(float)*8), outfile.get());
 451         // 32-bit val, channel mask
 452         fwrite32le(0, outfile.get());
 453         // 16 byte GUID, sub-type format
 454         fwrite(SUBTYPE_BFORMAT_FLOAT, 1, 16, outfile.get());
 455
 456         fputs("data", outfile.get());
 457         fwrite32le(0xFFFFFFFF, outfile.get()); // 'data' header len; filled in at close
 458         if(ferror(outfile.get()))
 459         {
 460             fprintf(stderr, "Error writing wave file header: %s (%d)\n", strerror(errno), errno);
 461             continue;
 462         }
 463
 464         auto DataStart = ftell(outfile.get());
 465
 466         auto decoder = std::make_unique<UhjDecoder>();
 467         auto inmem = std::make_unique<float[]>(BufferLineSize*static_cast<uint>(ininfo.channels));
 468         auto decmem = al::vector<std::array<float,BufferLineSize>, 16>(outchans);
 469         auto outmem = std::make_unique<byte4[]>(BufferLineSize*outchans);
 470
 471         /* A number of initial samples need to be skipped to cut the lead-in
 472          * from the all-pass filter delay. The same number of samples need to
 473          * be fed through the decoder after reaching the end of the input file
 474          * to ensure none of the original input is lost.
 475          */
 476         size_t LeadIn{UhjDecoder::sFilterDelay};
 477         sf_count_t LeadOut{UhjDecoder::sFilterDelay};
 478         while(LeadOut > 0)
 479         {
 480             sf_count_t sgot{sf_readf_float(infile.get(), inmem.get(), BufferLineSize)};
 481             sgot = std::max<sf_count_t>(sgot, 0);
 482             if(sgot < BufferLineSize)
 483             {
 484                 const sf_count_t remaining{std::min(BufferLineSize - sgot, LeadOut)};
 485                 std::fill_n(inmem.get() + sgot*ininfo.channels, remaining*ininfo.channels, 0.0f);
 486                 sgot += remaining;
 487                 LeadOut -= remaining;
 488             }
 489
 490             auto got = static_cast<size_t>(sgot);
 491             if(ininfo.channels > 2 || use_general)
 492                 decoder->decode(inmem.get(), static_cast<uint>(ininfo.channels), decmem, got);
 493             else
 494                 decoder->decode2(inmem.get(), decmem, got);
 495             if(LeadIn >= got)
 496             {
 497                 LeadIn -= got;
 498                 continue;
 499             }
 500
 501             got -= LeadIn;
 502             for(size_t i{0};i < got;++i)
 503             {
 504                 /* Attenuate by -3dB for FuMa output levels. */
 505                 constexpr auto inv_sqrt2 = static_cast<float>(1.0/al::numbers::sqrt2);
 506                 for(size_t j{0};j < outchans;++j)
 507                     outmem[i*outchans + j] = f32AsLEBytes(decmem[j][LeadIn+i] * inv_sqrt2);
 508             }
 509             LeadIn = 0;
 510
 511             size_t wrote{fwrite(outmem.get(), sizeof(byte4)*outchans, got, outfile.get())};
 512             if(wrote < got)
 513             {
 514                 fprintf(stderr, "Error writing wave data: %s (%d)\n", strerror(errno), errno);
 515                 break;
 516             }
 517         }
 518
 519         auto DataEnd = ftell(outfile.get());
 520         if(DataEnd > DataStart)
 521         {
 522             long dataLen{DataEnd - DataStart};
 523             if(fseek(outfile.get(), 4, SEEK_SET) == 0)
 524                 fwrite32le(static_cast<uint>(DataEnd-8), outfile.get()); // 'WAVE' header len
 525             if(fseek(outfile.get(), DataStart-4, SEEK_SET) == 0)
 526                 fwrite32le(static_cast<uint>(dataLen), outfile.get()); // 'data' header len
 527         }
 528         fflush(outfile.get());
 529         ++num_decoded;
 530     }
 531     if(num_decoded == 0)
 532         fprintf(stderr, "Failed to decode any input files\n");
 533     else if(num_decoded < num_files)
 534         fprintf(stderr, "Decoded %zu of %zu files\n", num_decoded, num_files);
 535     else
 536         printf("Decoded %zu file%s\n", num_decoded, (num_decoded==1)?"":"s");
 537     return 0;
 538 }