xref: /aosp_15_r20/external/aac/libFDK/src/mdct.cpp (revision e54365361535b070c2db7374cec45c159c7d0e7a)

/* -----------------------------------------------------------------------------
Software License for The Fraunhofer FDK AAC Codec Library for Android

© Copyright  1995 - 2019 Fraunhofer-Gesellschaft zur Förderung der angewandten
Forschung e.V. All rights reserved.

 1.    INTRODUCTION
The Fraunhofer FDK AAC Codec Library for Android ("FDK AAC Codec") is software
that implements the MPEG Advanced Audio Coding ("AAC") encoding and decoding
scheme for digital audio. This FDK AAC Codec software is intended to be used on
a wide variety of Android devices.

AAC's HE-AAC and HE-AAC v2 versions are regarded as today's most efficient
general perceptual audio codecs. AAC-ELD is considered the best-performing
full-bandwidth communications codec by independent studies and is widely
deployed. AAC has been standardized by ISO and IEC as part of the MPEG
specifications.

Patent licenses for necessary patent claims for the FDK AAC Codec (including
those of Fraunhofer) may be obtained through Via Licensing
(www.vialicensing.com) or through the respective patent owners individually for
the purpose of encoding or decoding bit streams in products that are compliant
with the ISO/IEC MPEG audio standards. Please note that most manufacturers of
Android devices already license these patent claims through Via Licensing or
directly from the patent owners, and therefore FDK AAC Codec software may
already be covered under those patent licenses when it is used for those
licensed purposes only.

Commercially-licensed AAC software libraries, including floating-point versions
with enhanced sound quality, are also available from Fraunhofer. Users are
encouraged to check the Fraunhofer website for additional applications
information and documentation.

2.    COPYRIGHT LICENSE

Redistribution and use in source and binary forms, with or without modification,
are permitted without payment of copyright license fees provided that you
satisfy the following conditions:

You must retain the complete text of this software license in redistributions of
the FDK AAC Codec or your modifications thereto in source code form.

You must retain the complete text of this software license in the documentation
and/or other materials provided with redistributions of the FDK AAC Codec or
your modifications thereto in binary form. You must make available free of
charge copies of the complete source code of the FDK AAC Codec and your
modifications thereto to recipients of copies in binary form.

The name of Fraunhofer may not be used to endorse or promote products derived
from this library without prior written permission.

You may not charge copyright license fees for anyone to use, copy or distribute
the FDK AAC Codec software or your modifications thereto.

Your modified versions of the FDK AAC Codec must carry prominent notices stating
that you changed the software and the date of any change. For modified versions
of the FDK AAC Codec, the term "Fraunhofer FDK AAC Codec Library for Android"
must be replaced by the term "Third-Party Modified Version of the Fraunhofer FDK
AAC Codec Library for Android."

3.    NO PATENT LICENSE

NO EXPRESS OR IMPLIED LICENSES TO ANY PATENT CLAIMS, including without
limitation the patents of Fraunhofer, ARE GRANTED BY THIS SOFTWARE LICENSE.
Fraunhofer provides no warranty of patent non-infringement with respect to this
software.

You may use this FDK AAC Codec software or modifications thereto only for
purposes that are authorized by appropriate patent licenses.

4.    DISCLAIMER

This FDK AAC Codec software is provided by Fraunhofer on behalf of the copyright
holders and contributors "AS IS" and WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES,
including but not limited to the implied warranties of merchantability and
fitness for a particular purpose. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR
CONTRIBUTORS BE LIABLE for any direct, indirect, incidental, special, exemplary,
or consequential damages, including but not limited to procurement of substitute
goods or services; loss of use, data, or profits, or business interruption,
however caused and on any theory of liability, whether in contract, strict
liability, or tort (including negligence), arising in any way out of the use of
this software, even if advised of the possibility of such damage.

5.    CONTACT INFORMATION

Fraunhofer Institute for Integrated Circuits IIS
Attention: Audio and Multimedia Departments - FDK AAC LL
Am Wolfsmantel 33
91058 Erlangen, Germany

www.iis.fraunhofer.de/amm
[email protected]
----------------------------------------------------------------------------- */

/******************* Library for basic calculation routines ********************

   Author(s):   Josef Hoepfl, Manuel Jander, Youliy Ninov, Daniel Hagel

   Description: MDCT/MDST routines

*******************************************************************************/

#include "mdct.h"

#include "FDK_tools_rom.h"
#include "dct.h"
#include "fixpoint_math.h"

void mdct_init(H_MDCT hMdct, FIXP_DBL *overlap, INT overlapBufferSize) {
  hMdct->overlap.freq = overlap;
  // FDKmemclear(overlap, overlapBufferSize*sizeof(FIXP_DBL));
  hMdct->prev_fr = 0;
  hMdct->prev_nr = 0;
  hMdct->prev_tl = 0;
  hMdct->ov_size = overlapBufferSize;
  hMdct->prevAliasSymmetry = 0;
  hMdct->prevPrevAliasSymmetry = 0;
  hMdct->pFacZir = NULL;
  hMdct->pAsymOvlp = NULL;
}

/*
This program implements the forward MDCT transform on an input block of data.
The input block is in a form (A,B,C,D) where A,B,C and D are the respective
1/4th segments of the block. The program takes the input block and folds it in
the form:
(-D-Cr,A-Br). This block is twice shorter and here the 'r' suffix denotes
flipping of the sequence (reversing the order of the samples). While folding the
input block in the above mentioned shorter block the program windows the data.
Because the two operations (windowing and folding) are not implemented
sequentially, but together the program's structure is not easy to understand.
Once the output (already windowed) block (-D-Cr,A-Br) is ready it is passed to
the DCT IV for processing.
*/
INT mdct_block(H_MDCT hMdct, const INT_PCM *RESTRICT timeData,
               const INT noInSamples, FIXP_DBL *RESTRICT mdctData,
               const INT nSpec, const INT tl, const FIXP_WTP *pRightWindowPart,
               const INT fr, SHORT *pMdctData_e) {
  int i, n;
  /* tl: transform length
     fl: left window slope length
     nl: left window slope offset
     fr: right window slope length
     nr: right window slope offset
     See FDK_tools/doc/intern/mdct.tex for more detail. */
  int fl, nl, nr;
  const FIXP_WTP *wls, *wrs;

  wrs = pRightWindowPart;

  /* Detect FRprevious / FL mismatches and override parameters accordingly */
  if (hMdct->prev_fr ==
      0) { /* At start just initialize and pass parameters as they are */
    hMdct->prev_fr = fr;
    hMdct->prev_wrs = wrs;
    hMdct->prev_tl = tl;
  }

  /* Derive NR */
  nr = (tl - fr) >> 1;

  /* Skip input samples if tl is smaller than block size */
  timeData += (noInSamples - tl) >> 1;

  /* windowing */
  for (n = 0; n < nSpec; n++) {
    /*
     * MDCT scale:
     * + 1: fMultDiv2() in windowing.
     * + 1: Because of factor 1/2 in Princen-Bradley compliant windowed TDAC.
     */
    INT mdctData_e = 1 + 1;

    /* Derive left parameters */
    wls = hMdct->prev_wrs;
    fl = hMdct->prev_fr;
    nl = (tl - fl) >> 1;

    /* Here we implement a simplified version of what happens after the this
    piece of code (see the comments below). We implement the folding of A and B
    segments to (A-Br) but A is zero, because in this part of the MDCT sequence
    the window coefficients with which A must be multiplied are zero.    */
    for (i = 0; i < nl; i++) {
#if SAMPLE_BITS == DFRACT_BITS /* SPC_BITS and DFRACT_BITS should be equal. */
      mdctData[(tl / 2) + i] = -((FIXP_DBL)timeData[tl - i - 1] >> (1));
#else
      mdctData[(tl / 2) + i] = -(FIXP_DBL)timeData[tl - i - 1]
                               << (DFRACT_BITS - SAMPLE_BITS - 1); /* 0(A)-Br */
#endif
    }

    /* Implements the folding and windowing of the left part of the sequence,
    that is segments A and B. The A segment is multiplied by the respective left
    window coefficient and placed in a temporary variable.

    tmp0 = fMultDiv2((FIXP_PCM)timeData[i+nl], pLeftWindowPart[i].v.im);

    After this the B segment taken in reverse order is multiplied by the left
    window and subtracted from the previously derived temporary variable, so
    that finally we implement the A-Br operation. This output is written to the
    right part of the MDCT output : (-D-Cr,A-Br).

    mdctData[(tl/2)+i+nl] = fMultSubDiv2(tmp0, (FIXP_PCM)timeData[tl-nl-i-1],
    pLeftWindowPart[i].v.re);//A*window-Br*window

    The (A-Br) data is written to the output buffer (mdctData) without being
    flipped.     */
    for (i = 0; i < fl / 2; i++) {
      FIXP_DBL tmp0;
      tmp0 = fMultDiv2((FIXP_PCM)timeData[i + nl], wls[i].v.im); /* a*window */
      mdctData[(tl / 2) + i + nl] =
          fMultSubDiv2(tmp0, (FIXP_PCM)timeData[tl - nl - i - 1],
                       wls[i].v.re); /* A*window-Br*window */
    }

    /* Right window slope offset */
    /* Here we implement a simplified version of what happens after the this
    piece of code (see the comments below). We implement the folding of C and D
    segments to (-D-Cr) but D is zero, because in this part of the MDCT sequence
    the window coefficients with which D must be multiplied are zero.    */
    for (i = 0; i < nr; i++) {
#if SAMPLE_BITS == \
    DFRACT_BITS /* This should be SPC_BITS instead of DFRACT_BITS. */
      mdctData[(tl / 2) - 1 - i] = -((FIXP_DBL)timeData[tl + i] >> (1));
#else
      mdctData[(tl / 2) - 1 - i] =
          -(FIXP_DBL)timeData[tl + i]
          << (DFRACT_BITS - SAMPLE_BITS - 1); /* -C flipped at placing */
#endif
    }

    /* Implements the folding and windowing of the right part of the sequence,
    that is, segments C and D. The C segment is multiplied by the respective
    right window coefficient and placed in a temporary variable.

    tmp1 = fMultDiv2((FIXP_PCM)timeData[tl+nr+i], pRightWindowPart[i].v.re);

    After this the D segment taken in reverse order is multiplied by the right
    window and added from the previously derived temporary variable, so that we
    get (C+Dr) operation. This output is negated to get (-C-Dr) and written to
    the left part of the MDCT output while being reversed (flipped) at the same
    time, so that from (-C-Dr) we get (-D-Cr)=> (-D-Cr,A-Br).

    mdctData[(tl/2)-nr-i-1] = -fMultAddDiv2(tmp1,
    (FIXP_PCM)timeData[(tl*2)-nr-i-1], pRightWindowPart[i].v.im);*/
    for (i = 0; i < fr / 2; i++) {
      FIXP_DBL tmp1;
      tmp1 = fMultDiv2((FIXP_PCM)timeData[tl + nr + i],
                       wrs[i].v.re); /* C*window */
      mdctData[(tl / 2) - nr - i - 1] =
          -fMultAddDiv2(tmp1, (FIXP_PCM)timeData[(tl * 2) - nr - i - 1],
                        wrs[i].v.im); /* -(C*window+Dr*window) and flip before
                                         placing -> -Cr - D */
    }

    /* We pass the shortened folded data (-D-Cr,A-Br) to the MDCT function */
    dct_IV(mdctData, tl, &mdctData_e);

    pMdctData_e[n] = (SHORT)mdctData_e;

    timeData += tl;
    mdctData += tl;

    hMdct->prev_wrs = wrs;
    hMdct->prev_fr = fr;
    hMdct->prev_tl = tl;
  }

  return nSpec * tl;
}

void imdct_gain(FIXP_DBL *pGain_m, int *pGain_e, int tl) {
  FIXP_DBL gain_m = *pGain_m;
  int gain_e = *pGain_e;
  int log2_tl;

  gain_e += -MDCT_OUTPUT_GAIN - MDCT_OUT_HEADROOM + 1;
  if (tl == 0) {
    /* Dont regard the 2/N factor from the IDCT. It is compensated for somewhere
     * else. */
    *pGain_e = gain_e;
    return;
  }

  log2_tl = DFRACT_BITS - 1 - fNormz((FIXP_DBL)tl);
  gain_e += -log2_tl;

  /* Detect non-radix 2 transform length and add amplitude compensation factor
     which cannot be included into the exponent above */
  switch ((tl) >> (log2_tl - 2)) {
    case 0x7: /* 10 ms, 1/tl = 1.0/(FDKpow(2.0, -log2_tl) *
                 0.53333333333333333333) */
      if (gain_m == (FIXP_DBL)0) {
        gain_m = FL2FXCONST_DBL(0.53333333333333333333f);
      } else {
        gain_m = fMult(gain_m, FL2FXCONST_DBL(0.53333333333333333333f));
      }
      break;
    case 0x6: /* 3/4 of radix 2, 1/tl = 1.0/(FDKpow(2.0, -log2_tl) * 2.0/3.0) */
      if (gain_m == (FIXP_DBL)0) {
        gain_m = FL2FXCONST_DBL(2.0 / 3.0f);
      } else {
        gain_m = fMult(gain_m, FL2FXCONST_DBL(2.0 / 3.0f));
      }
      break;
    case 0x5: /* 0.8 of radix 2 (e.g. tl 160), 1/tl = 1.0/(FDKpow(2.0, -log2_tl)
               * 0.8/1.5) */
      if (gain_m == (FIXP_DBL)0) {
        gain_m = FL2FXCONST_DBL(0.53333333333333333333f);
      } else {
        gain_m = fMult(gain_m, FL2FXCONST_DBL(0.53333333333333333333f));
      }
      break;
    case 0x4:
      /* radix 2, nothing to do. */
      break;
    default:
      /* unsupported */
      FDK_ASSERT(0);
      break;
  }

  *pGain_m = gain_m;
  *pGain_e = gain_e;
}

INT imdct_drain(H_MDCT hMdct, FIXP_DBL *output, INT nrSamplesRoom) {
  int buffered_samples = 0;

  if (nrSamplesRoom > 0) {
    buffered_samples = hMdct->ov_offset;

    FDK_ASSERT(buffered_samples <= nrSamplesRoom);

    if (buffered_samples > 0) {
      FDKmemcpy(output, hMdct->overlap.time,
                buffered_samples * sizeof(FIXP_DBL));
      hMdct->ov_offset = 0;
    }
  }
  return buffered_samples;
}

INT imdct_copy_ov_and_nr(H_MDCT hMdct, FIXP_DBL *pTimeData, INT nrSamples) {
  FIXP_DBL *pOvl;
  int nt, nf, i;

  nt = fMin(hMdct->ov_offset, nrSamples);
  nrSamples -= nt;
  nf = fMin(hMdct->prev_nr, nrSamples);
  FDKmemcpy(pTimeData, hMdct->overlap.time, nt * sizeof(FIXP_DBL));
  pTimeData += nt;

  pOvl = hMdct->overlap.freq + hMdct->ov_size - 1;
  if (hMdct->prevPrevAliasSymmetry == 0) {
    for (i = 0; i < nf; i++) {
      FIXP_DBL x = -(*pOvl--);
      *pTimeData = IMDCT_SCALE_DBL(x);
      pTimeData++;
    }
  } else {
    for (i = 0; i < nf; i++) {
      FIXP_DBL x = (*pOvl--);
      *pTimeData = IMDCT_SCALE_DBL(x);
      pTimeData++;
    }
  }

  return (nt + nf);
}

void imdct_adapt_parameters(H_MDCT hMdct, int *pfl, int *pnl, int tl,
                            const FIXP_WTP *wls, int noOutSamples) {
  int fl = *pfl, nl = *pnl;
  int window_diff, use_current = 0, use_previous = 0;
  if (hMdct->prev_tl == 0) {
    hMdct->prev_wrs = wls;
    hMdct->prev_fr = fl;
    hMdct->prev_nr = (noOutSamples - fl) >> 1;
    hMdct->prev_tl = noOutSamples;
    hMdct->ov_offset = 0;
    use_current = 1;
  }

  window_diff = (hMdct->prev_fr - fl) >> 1;

  /* check if the previous window slope can be adjusted to match the current
   * window slope */
  if (hMdct->prev_nr + window_diff > 0) {
    use_current = 1;
  }
  /* check if the current window slope can be adjusted to match the previous
   * window slope */
  if (nl - window_diff > 0) {
    use_previous = 1;
  }

  /* if both is possible choose the larger of both window slope lengths */
  if (use_current && use_previous) {
    if (fl < hMdct->prev_fr) {
      use_current = 0;
    }
  }
  /*
   * If the previous transform block is big enough, enlarge previous window
   * overlap, if not, then shrink current window overlap.
   */
  if (use_current) {
    hMdct->prev_nr += window_diff;
    hMdct->prev_fr = fl;
    hMdct->prev_wrs = wls;
  } else {
    nl -= window_diff;
    fl = hMdct->prev_fr;
  }

  *pfl = fl;
  *pnl = nl;
}

/*
This program implements the inverse modulated lapped transform, a generalized
version of the inverse MDCT transform. Setting none of the MLT_*_ALIAS_FLAG
flags computes the IMDCT, setting all of them computes the IMDST. Other
combinations of these flags compute type III transforms used by the RSVD60
multichannel tool for transitions between MDCT/MDST. The following description
relates to the IMDCT only.

If we pass the data block (A,B,C,D,E,F) to the FORWARD MDCT it will produce two
outputs. The first one will be over the (A,B,C,D) part =>(-D-Cr,A-Br) and the
second one will be over the (C,D,E,F) part => (-F-Er,C-Dr), since there is a
overlap between consequtive passes of the algorithm. This overlap is over the
(C,D) segments. The two outputs will be given sequentially to the DCT IV
algorithm. At the INVERSE MDCT side we get two consecutive outputs from the IDCT
IV algorithm, namely the same blocks: (-D-Cr,A-Br) and (-F-Er,C-Dr). The first
of them lands in the Overlap buffer and the second is in the working one, which,
one algorithm pass later will substitute the one residing in the overlap
register. The IMDCT algorithm has to produce the C and D segments from the two
buffers. In order to do this we take the left part of the overlap
buffer(-D-Cr,A-Br), namely (-D-Cr) and add it appropriately to the right part of
the working buffer (-F-Er,C-Dr), namely (C-Dr), so that we get first the C
segment and later the D segment. We do this in the following way: From the right
part of the working buffer(C-Dr) we subtract the flipped left part of the
overlap buffer(-D-Cr):

Result = (C-Dr) - flipped(-D-Cr) = C -Dr + Dr + C = 2C
We divide by two and get the C segment. What we did is adding the right part of
the first frame to the left part of the second one.   While applying these
operation we multiply the respective segments with the appropriate window
functions.

In order to get the D segment we do the following:
From the negated second part of the working buffer(C-Dr) we subtract the flipped
first part of the overlap buffer (-D-Cr):

Result= - (C -Dr) - flipped(-D-Cr)= -C +Dr +Dr +C = 2Dr.
After dividing by two and flipping we get the D segment.What we did is adding
the right part of the first frame to the left part of the second one.   While
applying these operation we multiply the respective segments with the
appropriate window functions.

Once we have obtained the C and D segments the overlap buffer is emptied and the
current buffer is sent in it, so that the E and F segments are available for
decoding in the next algorithm pass.*/
INT imlt_block(H_MDCT hMdct, FIXP_DBL *output, FIXP_DBL *spectrum,
               const SHORT scalefactor[], const INT nSpec,
               const INT noOutSamples, const INT tl, const FIXP_WTP *wls,
               INT fl, const FIXP_WTP *wrs, const INT fr, FIXP_DBL gain,
               int flags) {
  FIXP_DBL *pOvl;
  FIXP_DBL *pOut0 = output, *pOut1;
  INT nl, nr;
  int w, i, nrSamples = 0, specShiftScale, transform_gain_e = 0;
  int currAliasSymmetry = (flags & MLT_FLAG_CURR_ALIAS_SYMMETRY);

  /* Derive NR and NL */
  nr = (tl - fr) >> 1;
  nl = (tl - fl) >> 1;

  /* Include 2/N IMDCT gain into gain factor and exponent. */
  imdct_gain(&gain, &transform_gain_e, tl);

  /* Detect FRprevious / FL mismatches and override parameters accordingly */
  if (hMdct->prev_fr != fl) {
    imdct_adapt_parameters(hMdct, &fl, &nl, tl, wls, noOutSamples);
  }

  pOvl = hMdct->overlap.freq + hMdct->ov_size - 1;

  if (noOutSamples > nrSamples) {
    /* Purge buffered output. */
    for (i = 0; i < hMdct->ov_offset; i++) {
      *pOut0 = hMdct->overlap.time[i];
      pOut0++;
    }
    nrSamples = hMdct->ov_offset;
    hMdct->ov_offset = 0;
  }

  for (w = 0; w < nSpec; w++) {
    FIXP_DBL *pSpec, *pCurr;
    const FIXP_WTP *pWindow;

    /* Detect FRprevious / FL mismatches and override parameters accordingly */
    if (hMdct->prev_fr != fl) {
      imdct_adapt_parameters(hMdct, &fl, &nl, tl, wls, noOutSamples);
    }

    specShiftScale = transform_gain_e;

    /* Setup window pointers */
    pWindow = hMdct->prev_wrs;

    /* Current spectrum */
    pSpec = spectrum + w * tl;

    /* DCT IV of current spectrum. */
    if (currAliasSymmetry == 0) {
      if (hMdct->prevAliasSymmetry == 0) {
        dct_IV(pSpec, tl, &specShiftScale);
      } else {
        FIXP_DBL _tmp[1024 + ALIGNMENT_DEFAULT / sizeof(FIXP_DBL)];
        FIXP_DBL *tmp = (FIXP_DBL *)ALIGN_PTR(_tmp);
        C_ALLOC_ALIGNED_REGISTER(tmp, sizeof(_tmp));
        dct_III(pSpec, tmp, tl, &specShiftScale);
        C_ALLOC_ALIGNED_UNREGISTER(tmp);
      }
    } else {
      if (hMdct->prevAliasSymmetry == 0) {
        FIXP_DBL _tmp[1024 + ALIGNMENT_DEFAULT / sizeof(FIXP_DBL)];
        FIXP_DBL *tmp = (FIXP_DBL *)ALIGN_PTR(_tmp);
        C_ALLOC_ALIGNED_REGISTER(tmp, sizeof(_tmp));
        dst_III(pSpec, tmp, tl, &specShiftScale);
        C_ALLOC_ALIGNED_UNREGISTER(tmp);
      } else {
        dst_IV(pSpec, tl, &specShiftScale);
      }
    }

    /* Optional scaling of time domain - no yet windowed - of current spectrum
     */
    /* and de-scale current spectrum signal (time domain, no yet windowed) */
    if (gain != (FIXP_DBL)0) {
      for (i = 0; i < tl; i++) {
        pSpec[i] = fMult(pSpec[i], gain);
      }
    }

    {
      int loc_scale =
          fixmin_I(scalefactor[w] + specShiftScale, (INT)DFRACT_BITS - 1);
      DWORD_ALIGNED(pSpec);
      scaleValuesSaturate(pSpec, tl, loc_scale);
    }

    if (noOutSamples <= nrSamples) {
      /* Divert output first half to overlap buffer if we already got enough
       * output samples. */
      pOut0 = hMdct->overlap.time + hMdct->ov_offset;
      hMdct->ov_offset += hMdct->prev_nr + fl / 2;
    } else {
      /* Account output samples */
      nrSamples += hMdct->prev_nr + fl / 2;
    }

    /* NR output samples 0 .. NR. -overlap[TL/2..TL/2-NR] */
    if ((hMdct->pFacZir != 0) && (hMdct->prev_nr == fl / 2)) {
      /* In the case of ACELP -> TCX20 -> FD short add FAC ZIR on nr signal part
       */
      for (i = 0; i < hMdct->prev_nr; i++) {
        FIXP_DBL x = -(*pOvl--);
        *pOut0 = fAddSaturate(x, IMDCT_SCALE_DBL(hMdct->pFacZir[i]));
        pOut0++;
      }
      hMdct->pFacZir = NULL;
    } else {
      /* Here we implement a simplified version of what happens after the this
      piece of code (see the comments below). We implement the folding of C and
      D segments from (-D-Cr) but D is zero, because in this part of the MDCT
      sequence the window coefficients with which D must be multiplied are zero.
      "pOut0" writes sequentially the C block from left to right.   */
      if (hMdct->prevPrevAliasSymmetry == 0) {
        for (i = 0; i < hMdct->prev_nr; i++) {
          FIXP_DBL x = -(*pOvl--);
          *pOut0 = IMDCT_SCALE_DBL(x);
          pOut0++;
        }
      } else {
        for (i = 0; i < hMdct->prev_nr; i++) {
          FIXP_DBL x = *pOvl--;
          *pOut0 = IMDCT_SCALE_DBL(x);
          pOut0++;
        }
      }
    }

    if (noOutSamples <= nrSamples) {
      /* Divert output second half to overlap buffer if we already got enough
       * output samples. */
      pOut1 = hMdct->overlap.time + hMdct->ov_offset + fl / 2 - 1;
      hMdct->ov_offset += fl / 2 + nl;
    } else {
      pOut1 = pOut0 + (fl - 1);
      nrSamples += fl / 2 + nl;
    }

    /* output samples before window crossing point NR .. TL/2.
     * -overlap[TL/2-NR..TL/2-NR-FL/2] + current[NR..TL/2] */
    /* output samples after window crossing point TL/2 .. TL/2+FL/2.
     * -overlap[0..FL/2] - current[TL/2..FL/2] */
    pCurr = pSpec + tl - fl / 2;
    DWORD_ALIGNED(pCurr);
    C_ALLOC_ALIGNED_REGISTER(pWindow, fl);
    DWORD_ALIGNED(pWindow);
    C_ALLOC_ALIGNED_UNREGISTER(pWindow);

    if (hMdct->prevPrevAliasSymmetry == 0) {
      if (hMdct->prevAliasSymmetry == 0) {
        if (!hMdct->pAsymOvlp) {
          for (i = 0; i < fl / 2; i++) {
            FIXP_DBL x0, x1;
            cplxMultDiv2(&x1, &x0, *pCurr++, -*pOvl--, pWindow[i]);
            *pOut0 = IMDCT_SCALE_DBL_LSH1(x0);
            *pOut1 = IMDCT_SCALE_DBL_LSH1(-x1);
            pOut0++;
            pOut1--;
          }
        } else {
          FIXP_DBL *pAsymOvl = hMdct->pAsymOvlp + fl / 2 - 1;
          for (i = 0; i < fl / 2; i++) {
            FIXP_DBL x0, x1;
            x1 = -fMultDiv2(*pCurr, pWindow[i].v.re) +
                 fMultDiv2(*pAsymOvl, pWindow[i].v.im);
            x0 = fMultDiv2(*pCurr, pWindow[i].v.im) -
                 fMultDiv2(*pOvl, pWindow[i].v.re);
            pCurr++;
            pOvl--;
            pAsymOvl--;
            *pOut0++ = IMDCT_SCALE_DBL_LSH1(x0);
            *pOut1-- = IMDCT_SCALE_DBL_LSH1(x1);
          }
          hMdct->pAsymOvlp = NULL;
        }
      } else { /* prevAliasingSymmetry == 1 */
        for (i = 0; i < fl / 2; i++) {
          FIXP_DBL x0, x1;
          cplxMultDiv2(&x1, &x0, *pCurr++, -*pOvl--, pWindow[i]);
          *pOut0 = IMDCT_SCALE_DBL_LSH1(x0);
          *pOut1 = IMDCT_SCALE_DBL_LSH1(x1);
          pOut0++;
          pOut1--;
        }
      }
    } else { /* prevPrevAliasingSymmetry == 1 */
      if (hMdct->prevAliasSymmetry == 0) {
        for (i = 0; i < fl / 2; i++) {
          FIXP_DBL x0, x1;
          cplxMultDiv2(&x1, &x0, *pCurr++, *pOvl--, pWindow[i]);
          *pOut0 = IMDCT_SCALE_DBL_LSH1(x0);
          *pOut1 = IMDCT_SCALE_DBL_LSH1(-x1);
          pOut0++;
          pOut1--;
        }
      } else { /* prevAliasingSymmetry == 1 */
        for (i = 0; i < fl / 2; i++) {
          FIXP_DBL x0, x1;
          cplxMultDiv2(&x1, &x0, *pCurr++, *pOvl--, pWindow[i]);
          *pOut0 = IMDCT_SCALE_DBL_LSH1(x0);
          *pOut1 = IMDCT_SCALE_DBL_LSH1(x1);
          pOut0++;
          pOut1--;
        }
      }
    }

    if (hMdct->pFacZir != 0) {
      /* add FAC ZIR of previous ACELP -> mdct transition */
      FIXP_DBL *pOut = pOut0 - fl / 2;
      FDK_ASSERT(fl / 2 <= 128);
      for (i = 0; i < fl / 2; i++) {
        pOut[i] = fAddSaturate(pOut[i], IMDCT_SCALE_DBL(hMdct->pFacZir[i]));
      }
      hMdct->pFacZir = NULL;
    }
    pOut0 += (fl / 2) + nl;

    /* NL output samples TL/2+FL/2..TL. - current[FL/2..0] */
    pOut1 += (fl / 2) + 1;
    pCurr = pSpec + tl - fl / 2 - 1;
    /* Here we implement a simplified version of what happens above the this
    piece of code (see the comments above). We implement the folding of C and D
    segments from (C-Dr) but C is zero, because in this part of the MDCT
    sequence the window coefficients with which C must be multiplied are zero.
    "pOut1" writes sequentially the D block from left to right.   */
    if (hMdct->prevAliasSymmetry == 0) {
      for (i = 0; i < nl; i++) {
        FIXP_DBL x = -(*pCurr--);
        *pOut1++ = IMDCT_SCALE_DBL(x);
      }
    } else {
      for (i = 0; i < nl; i++) {
        FIXP_DBL x = *pCurr--;
        *pOut1++ = IMDCT_SCALE_DBL(x);
      }
    }

    /* Set overlap source pointer for next window pOvl = pSpec + tl/2 - 1; */
    pOvl = pSpec + tl / 2 - 1;

    /* Previous window values. */
    hMdct->prev_nr = nr;
    hMdct->prev_fr = fr;
    hMdct->prev_tl = tl;
    hMdct->prev_wrs = wrs;

    /* Previous aliasing symmetry */
    hMdct->prevPrevAliasSymmetry = hMdct->prevAliasSymmetry;
    hMdct->prevAliasSymmetry = currAliasSymmetry;
  }

  /* Save overlap */

  pOvl = hMdct->overlap.freq + hMdct->ov_size - tl / 2;
  FDKmemcpy(pOvl, &spectrum[(nSpec - 1) * tl], (tl / 2) * sizeof(FIXP_DBL));

  return nrSamples;
}