Chris@69: /* Copyright (c) 2007-2008 CSIRO
Chris@69:    Copyright (c) 2007-2010 Xiph.Org Foundation
Chris@69:    Copyright (c) 2008 Gregory Maxwell
Chris@69:    Written by Jean-Marc Valin and Gregory Maxwell */
Chris@69: /*
Chris@69:    Redistribution and use in source and binary forms, with or without
Chris@69:    modification, are permitted provided that the following conditions
Chris@69:    are met:
Chris@69: 
Chris@69:    - Redistributions of source code must retain the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer.
Chris@69: 
Chris@69:    - Redistributions in binary form must reproduce the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer in the
Chris@69:    documentation and/or other materials provided with the distribution.
Chris@69: 
Chris@69:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
Chris@69:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
Chris@69:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
Chris@69:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
Chris@69:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
Chris@69:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
Chris@69:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
Chris@69:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
Chris@69:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
Chris@69:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
Chris@69:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
Chris@69: */
Chris@69: 
Chris@69: #ifdef HAVE_CONFIG_H
Chris@69: #include "config.h"
Chris@69: #endif
Chris@69: 
Chris@69: #define CELT_DECODER_C
Chris@69: 
Chris@69: #include "cpu_support.h"
Chris@69: #include "os_support.h"
Chris@69: #include "mdct.h"
Chris@69: #include <math.h>
Chris@69: #include "celt.h"
Chris@69: #include "pitch.h"
Chris@69: #include "bands.h"
Chris@69: #include "modes.h"
Chris@69: #include "entcode.h"
Chris@69: #include "quant_bands.h"
Chris@69: #include "rate.h"
Chris@69: #include "stack_alloc.h"
Chris@69: #include "mathops.h"
Chris@69: #include "float_cast.h"
Chris@69: #include <stdarg.h>
Chris@69: #include "celt_lpc.h"
Chris@69: #include "vq.h"
Chris@69: 
Chris@69: /* The maximum pitch lag to allow in the pitch-based PLC. It's possible to save
Chris@69:    CPU time in the PLC pitch search by making this smaller than MAX_PERIOD. The
Chris@69:    current value corresponds to a pitch of 66.67 Hz. */
Chris@69: #define PLC_PITCH_LAG_MAX (720)
Chris@69: /* The minimum pitch lag to allow in the pitch-based PLC. This corresponds to a
Chris@69:    pitch of 480 Hz. */
Chris@69: #define PLC_PITCH_LAG_MIN (100)
Chris@69: 
Chris@69: #if defined(SMALL_FOOTPRINT) && defined(FIXED_POINT)
Chris@69: #define NORM_ALIASING_HACK
Chris@69: #endif
Chris@69: /**********************************************************************/
Chris@69: /*                                                                    */
Chris@69: /*                             DECODER                                */
Chris@69: /*                                                                    */
Chris@69: /**********************************************************************/
Chris@69: #define DECODE_BUFFER_SIZE 2048
Chris@69: 
Chris@69: /** Decoder state
Chris@69:  @brief Decoder state
Chris@69:  */
Chris@69: struct OpusCustomDecoder {
Chris@69:    const OpusCustomMode *mode;
Chris@69:    int overlap;
Chris@69:    int channels;
Chris@69:    int stream_channels;
Chris@69: 
Chris@69:    int downsample;
Chris@69:    int start, end;
Chris@69:    int signalling;
Chris@69:    int disable_inv;
Chris@69:    int arch;
Chris@69: 
Chris@69:    /* Everything beyond this point gets cleared on a reset */
Chris@69: #define DECODER_RESET_START rng
Chris@69: 
Chris@69:    opus_uint32 rng;
Chris@69:    int error;
Chris@69:    int last_pitch_index;
Chris@69:    int loss_count;
Chris@69:    int skip_plc;
Chris@69:    int postfilter_period;
Chris@69:    int postfilter_period_old;
Chris@69:    opus_val16 postfilter_gain;
Chris@69:    opus_val16 postfilter_gain_old;
Chris@69:    int postfilter_tapset;
Chris@69:    int postfilter_tapset_old;
Chris@69: 
Chris@69:    celt_sig preemph_memD[2];
Chris@69: 
Chris@69:    celt_sig _decode_mem[1]; /* Size = channels*(DECODE_BUFFER_SIZE+mode->overlap) */
Chris@69:    /* opus_val16 lpc[],  Size = channels*LPC_ORDER */
Chris@69:    /* opus_val16 oldEBands[], Size = 2*mode->nbEBands */
Chris@69:    /* opus_val16 oldLogE[], Size = 2*mode->nbEBands */
Chris@69:    /* opus_val16 oldLogE2[], Size = 2*mode->nbEBands */
Chris@69:    /* opus_val16 backgroundLogE[], Size = 2*mode->nbEBands */
Chris@69: };
Chris@69: 
Chris@69: #if defined(ENABLE_HARDENING) || defined(ENABLE_ASSERTIONS)
Chris@69: /* Make basic checks on the CELT state to ensure we don't end
Chris@69:    up writing all over memory. */
Chris@69: void validate_celt_decoder(CELTDecoder *st)
Chris@69: {
Chris@69: #ifndef CUSTOM_MODES
Chris@69:    celt_assert(st->mode == opus_custom_mode_create(48000, 960, NULL));
Chris@69:    celt_assert(st->overlap == 120);
Chris@69: #endif
Chris@69:    celt_assert(st->channels == 1 || st->channels == 2);
Chris@69:    celt_assert(st->stream_channels == 1 || st->stream_channels == 2);
Chris@69:    celt_assert(st->downsample > 0);
Chris@69:    celt_assert(st->start == 0 || st->start == 17);
Chris@69:    celt_assert(st->start < st->end);
Chris@69:    celt_assert(st->end <= 21);
Chris@69: #ifdef OPUS_ARCHMASK
Chris@69:    celt_assert(st->arch >= 0);
Chris@69:    celt_assert(st->arch <= OPUS_ARCHMASK);
Chris@69: #endif
Chris@69:    celt_assert(st->last_pitch_index <= PLC_PITCH_LAG_MAX);
Chris@69:    celt_assert(st->last_pitch_index >= PLC_PITCH_LAG_MIN || st->last_pitch_index == 0);
Chris@69:    celt_assert(st->postfilter_period < MAX_PERIOD);
Chris@69:    celt_assert(st->postfilter_period >= COMBFILTER_MINPERIOD || st->postfilter_period == 0);
Chris@69:    celt_assert(st->postfilter_period_old < MAX_PERIOD);
Chris@69:    celt_assert(st->postfilter_period_old >= COMBFILTER_MINPERIOD || st->postfilter_period_old == 0);
Chris@69:    celt_assert(st->postfilter_tapset <= 2);
Chris@69:    celt_assert(st->postfilter_tapset >= 0);
Chris@69:    celt_assert(st->postfilter_tapset_old <= 2);
Chris@69:    celt_assert(st->postfilter_tapset_old >= 0);
Chris@69: }
Chris@69: #endif
Chris@69: 
Chris@69: int celt_decoder_get_size(int channels)
Chris@69: {
Chris@69:    const CELTMode *mode = opus_custom_mode_create(48000, 960, NULL);
Chris@69:    return opus_custom_decoder_get_size(mode, channels);
Chris@69: }
Chris@69: 
Chris@69: OPUS_CUSTOM_NOSTATIC int opus_custom_decoder_get_size(const CELTMode *mode, int channels)
Chris@69: {
Chris@69:    int size = sizeof(struct CELTDecoder)
Chris@69:             + (channels*(DECODE_BUFFER_SIZE+mode->overlap)-1)*sizeof(celt_sig)
Chris@69:             + channels*LPC_ORDER*sizeof(opus_val16)
Chris@69:             + 4*2*mode->nbEBands*sizeof(opus_val16);
Chris@69:    return size;
Chris@69: }
Chris@69: 
Chris@69: #ifdef CUSTOM_MODES
Chris@69: CELTDecoder *opus_custom_decoder_create(const CELTMode *mode, int channels, int *error)
Chris@69: {
Chris@69:    int ret;
Chris@69:    CELTDecoder *st = (CELTDecoder *)opus_alloc(opus_custom_decoder_get_size(mode, channels));
Chris@69:    ret = opus_custom_decoder_init(st, mode, channels);
Chris@69:    if (ret != OPUS_OK)
Chris@69:    {
Chris@69:       opus_custom_decoder_destroy(st);
Chris@69:       st = NULL;
Chris@69:    }
Chris@69:    if (error)
Chris@69:       *error = ret;
Chris@69:    return st;
Chris@69: }
Chris@69: #endif /* CUSTOM_MODES */
Chris@69: 
Chris@69: int celt_decoder_init(CELTDecoder *st, opus_int32 sampling_rate, int channels)
Chris@69: {
Chris@69:    int ret;
Chris@69:    ret = opus_custom_decoder_init(st, opus_custom_mode_create(48000, 960, NULL), channels);
Chris@69:    if (ret != OPUS_OK)
Chris@69:       return ret;
Chris@69:    st->downsample = resampling_factor(sampling_rate);
Chris@69:    if (st->downsample==0)
Chris@69:       return OPUS_BAD_ARG;
Chris@69:    else
Chris@69:       return OPUS_OK;
Chris@69: }
Chris@69: 
Chris@69: OPUS_CUSTOM_NOSTATIC int opus_custom_decoder_init(CELTDecoder *st, const CELTMode *mode, int channels)
Chris@69: {
Chris@69:    if (channels < 0 || channels > 2)
Chris@69:       return OPUS_BAD_ARG;
Chris@69: 
Chris@69:    if (st==NULL)
Chris@69:       return OPUS_ALLOC_FAIL;
Chris@69: 
Chris@69:    OPUS_CLEAR((char*)st, opus_custom_decoder_get_size(mode, channels));
Chris@69: 
Chris@69:    st->mode = mode;
Chris@69:    st->overlap = mode->overlap;
Chris@69:    st->stream_channels = st->channels = channels;
Chris@69: 
Chris@69:    st->downsample = 1;
Chris@69:    st->start = 0;
Chris@69:    st->end = st->mode->effEBands;
Chris@69:    st->signalling = 1;
Chris@69: #ifndef DISABLE_UPDATE_DRAFT
Chris@69:    st->disable_inv = channels == 1;
Chris@69: #else
Chris@69:    st->disable_inv = 0;
Chris@69: #endif
Chris@69:    st->arch = opus_select_arch();
Chris@69: 
Chris@69:    opus_custom_decoder_ctl(st, OPUS_RESET_STATE);
Chris@69: 
Chris@69:    return OPUS_OK;
Chris@69: }
Chris@69: 
Chris@69: #ifdef CUSTOM_MODES
Chris@69: void opus_custom_decoder_destroy(CELTDecoder *st)
Chris@69: {
Chris@69:    opus_free(st);
Chris@69: }
Chris@69: #endif /* CUSTOM_MODES */
Chris@69: 
Chris@69: #ifndef CUSTOM_MODES
Chris@69: /* Special case for stereo with no downsampling and no accumulation. This is
Chris@69:    quite common and we can make it faster by processing both channels in the
Chris@69:    same loop, reducing overhead due to the dependency loop in the IIR filter. */
Chris@69: static void deemphasis_stereo_simple(celt_sig *in[], opus_val16 *pcm, int N, const opus_val16 coef0,
Chris@69:       celt_sig *mem)
Chris@69: {
Chris@69:    celt_sig * OPUS_RESTRICT x0;
Chris@69:    celt_sig * OPUS_RESTRICT x1;
Chris@69:    celt_sig m0, m1;
Chris@69:    int j;
Chris@69:    x0=in[0];
Chris@69:    x1=in[1];
Chris@69:    m0 = mem[0];
Chris@69:    m1 = mem[1];
Chris@69:    for (j=0;j<N;j++)
Chris@69:    {
Chris@69:       celt_sig tmp0, tmp1;
Chris@69:       /* Add VERY_SMALL to x[] first to reduce dependency chain. */
Chris@69:       tmp0 = x0[j] + VERY_SMALL + m0;
Chris@69:       tmp1 = x1[j] + VERY_SMALL + m1;
Chris@69:       m0 = MULT16_32_Q15(coef0, tmp0);
Chris@69:       m1 = MULT16_32_Q15(coef0, tmp1);
Chris@69:       pcm[2*j  ] = SCALEOUT(SIG2WORD16(tmp0));
Chris@69:       pcm[2*j+1] = SCALEOUT(SIG2WORD16(tmp1));
Chris@69:    }
Chris@69:    mem[0] = m0;
Chris@69:    mem[1] = m1;
Chris@69: }
Chris@69: #endif
Chris@69: 
Chris@69: #ifndef RESYNTH
Chris@69: static
Chris@69: #endif
Chris@69: void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, const opus_val16 *coef,
Chris@69:       celt_sig *mem, int accum)
Chris@69: {
Chris@69:    int c;
Chris@69:    int Nd;
Chris@69:    int apply_downsampling=0;
Chris@69:    opus_val16 coef0;
Chris@69:    VARDECL(celt_sig, scratch);
Chris@69:    SAVE_STACK;
Chris@69: #ifndef CUSTOM_MODES
Chris@69:    /* Short version for common case. */
Chris@69:    if (downsample == 1 && C == 2 && !accum)
Chris@69:    {
Chris@69:       deemphasis_stereo_simple(in, pcm, N, coef[0], mem);
Chris@69:       return;
Chris@69:    }
Chris@69: #endif
Chris@69: #ifndef FIXED_POINT
Chris@69:    (void)accum;
Chris@69:    celt_assert(accum==0);
Chris@69: #endif
Chris@69:    ALLOC(scratch, N, celt_sig);
Chris@69:    coef0 = coef[0];
Chris@69:    Nd = N/downsample;
Chris@69:    c=0; do {
Chris@69:       int j;
Chris@69:       celt_sig * OPUS_RESTRICT x;
Chris@69:       opus_val16  * OPUS_RESTRICT y;
Chris@69:       celt_sig m = mem[c];
Chris@69:       x =in[c];
Chris@69:       y = pcm+c;
Chris@69: #ifdef CUSTOM_MODES
Chris@69:       if (coef[1] != 0)
Chris@69:       {
Chris@69:          opus_val16 coef1 = coef[1];
Chris@69:          opus_val16 coef3 = coef[3];
Chris@69:          for (j=0;j<N;j++)
Chris@69:          {
Chris@69:             celt_sig tmp = x[j] + m + VERY_SMALL;
Chris@69:             m = MULT16_32_Q15(coef0, tmp)
Chris@69:                           - MULT16_32_Q15(coef1, x[j]);
Chris@69:             tmp = SHL32(MULT16_32_Q15(coef3, tmp), 2);
Chris@69:             scratch[j] = tmp;
Chris@69:          }
Chris@69:          apply_downsampling=1;
Chris@69:       } else
Chris@69: #endif
Chris@69:       if (downsample>1)
Chris@69:       {
Chris@69:          /* Shortcut for the standard (non-custom modes) case */
Chris@69:          for (j=0;j<N;j++)
Chris@69:          {
Chris@69:             celt_sig tmp = x[j] + VERY_SMALL + m;
Chris@69:             m = MULT16_32_Q15(coef0, tmp);
Chris@69:             scratch[j] = tmp;
Chris@69:          }
Chris@69:          apply_downsampling=1;
Chris@69:       } else {
Chris@69:          /* Shortcut for the standard (non-custom modes) case */
Chris@69: #ifdef FIXED_POINT
Chris@69:          if (accum)
Chris@69:          {
Chris@69:             for (j=0;j<N;j++)
Chris@69:             {
Chris@69:                celt_sig tmp = x[j] + m + VERY_SMALL;
Chris@69:                m = MULT16_32_Q15(coef0, tmp);
Chris@69:                y[j*C] = SAT16(ADD32(y[j*C], SCALEOUT(SIG2WORD16(tmp))));
Chris@69:             }
Chris@69:          } else
Chris@69: #endif
Chris@69:          {
Chris@69:             for (j=0;j<N;j++)
Chris@69:             {
Chris@69:                celt_sig tmp = x[j] + VERY_SMALL + m;
Chris@69:                m = MULT16_32_Q15(coef0, tmp);
Chris@69:                y[j*C] = SCALEOUT(SIG2WORD16(tmp));
Chris@69:             }
Chris@69:          }
Chris@69:       }
Chris@69:       mem[c] = m;
Chris@69: 
Chris@69:       if (apply_downsampling)
Chris@69:       {
Chris@69:          /* Perform down-sampling */
Chris@69: #ifdef FIXED_POINT
Chris@69:          if (accum)
Chris@69:          {
Chris@69:             for (j=0;j<Nd;j++)
Chris@69:                y[j*C] = SAT16(ADD32(y[j*C], SCALEOUT(SIG2WORD16(scratch[j*downsample]))));
Chris@69:          } else
Chris@69: #endif
Chris@69:          {
Chris@69:             for (j=0;j<Nd;j++)
Chris@69:                y[j*C] = SCALEOUT(SIG2WORD16(scratch[j*downsample]));
Chris@69:          }
Chris@69:       }
Chris@69:    } while (++c<C);
Chris@69:    RESTORE_STACK;
Chris@69: }
Chris@69: 
Chris@69: #ifndef RESYNTH
Chris@69: static
Chris@69: #endif
Chris@69: void celt_synthesis(const CELTMode *mode, celt_norm *X, celt_sig * out_syn[],
Chris@69:                     opus_val16 *oldBandE, int start, int effEnd, int C, int CC,
Chris@69:                     int isTransient, int LM, int downsample,
Chris@69:                     int silence, int arch)
Chris@69: {
Chris@69:    int c, i;
Chris@69:    int M;
Chris@69:    int b;
Chris@69:    int B;
Chris@69:    int N, NB;
Chris@69:    int shift;
Chris@69:    int nbEBands;
Chris@69:    int overlap;
Chris@69:    VARDECL(celt_sig, freq);
Chris@69:    SAVE_STACK;
Chris@69: 
Chris@69:    overlap = mode->overlap;
Chris@69:    nbEBands = mode->nbEBands;
Chris@69:    N = mode->shortMdctSize<<LM;
Chris@69:    ALLOC(freq, N, celt_sig); /**< Interleaved signal MDCTs */
Chris@69:    M = 1<<LM;
Chris@69: 
Chris@69:    if (isTransient)
Chris@69:    {
Chris@69:       B = M;
Chris@69:       NB = mode->shortMdctSize;
Chris@69:       shift = mode->maxLM;
Chris@69:    } else {
Chris@69:       B = 1;
Chris@69:       NB = mode->shortMdctSize<<LM;
Chris@69:       shift = mode->maxLM-LM;
Chris@69:    }
Chris@69: 
Chris@69:    if (CC==2&&C==1)
Chris@69:    {
Chris@69:       /* Copying a mono streams to two channels */
Chris@69:       celt_sig *freq2;
Chris@69:       denormalise_bands(mode, X, freq, oldBandE, start, effEnd, M,
Chris@69:             downsample, silence);
Chris@69:       /* Store a temporary copy in the output buffer because the IMDCT destroys its input. */
Chris@69:       freq2 = out_syn[1]+overlap/2;
Chris@69:       OPUS_COPY(freq2, freq, N);
Chris@69:       for (b=0;b<B;b++)
Chris@69:          clt_mdct_backward(&mode->mdct, &freq2[b], out_syn[0]+NB*b, mode->window, overlap, shift, B, arch);
Chris@69:       for (b=0;b<B;b++)
Chris@69:          clt_mdct_backward(&mode->mdct, &freq[b], out_syn[1]+NB*b, mode->window, overlap, shift, B, arch);
Chris@69:    } else if (CC==1&&C==2)
Chris@69:    {
Chris@69:       /* Downmixing a stereo stream to mono */
Chris@69:       celt_sig *freq2;
Chris@69:       freq2 = out_syn[0]+overlap/2;
Chris@69:       denormalise_bands(mode, X, freq, oldBandE, start, effEnd, M,
Chris@69:             downsample, silence);
Chris@69:       /* Use the output buffer as temp array before downmixing. */
Chris@69:       denormalise_bands(mode, X+N, freq2, oldBandE+nbEBands, start, effEnd, M,
Chris@69:             downsample, silence);
Chris@69:       for (i=0;i<N;i++)
Chris@69:          freq[i] = ADD32(HALF32(freq[i]), HALF32(freq2[i]));
Chris@69:       for (b=0;b<B;b++)
Chris@69:          clt_mdct_backward(&mode->mdct, &freq[b], out_syn[0]+NB*b, mode->window, overlap, shift, B, arch);
Chris@69:    } else {
Chris@69:       /* Normal case (mono or stereo) */
Chris@69:       c=0; do {
Chris@69:          denormalise_bands(mode, X+c*N, freq, oldBandE+c*nbEBands, start, effEnd, M,
Chris@69:                downsample, silence);
Chris@69:          for (b=0;b<B;b++)
Chris@69:             clt_mdct_backward(&mode->mdct, &freq[b], out_syn[c]+NB*b, mode->window, overlap, shift, B, arch);
Chris@69:       } while (++c<CC);
Chris@69:    }
Chris@69:    /* Saturate IMDCT output so that we can't overflow in the pitch postfilter
Chris@69:       or in the */
Chris@69:    c=0; do {
Chris@69:       for (i=0;i<N;i++)
Chris@69:          out_syn[c][i] = SATURATE(out_syn[c][i], SIG_SAT);
Chris@69:    } while (++c<CC);
Chris@69:    RESTORE_STACK;
Chris@69: }
Chris@69: 
Chris@69: static void tf_decode(int start, int end, int isTransient, int *tf_res, int LM, ec_dec *dec)
Chris@69: {
Chris@69:    int i, curr, tf_select;
Chris@69:    int tf_select_rsv;
Chris@69:    int tf_changed;
Chris@69:    int logp;
Chris@69:    opus_uint32 budget;
Chris@69:    opus_uint32 tell;
Chris@69: 
Chris@69:    budget = dec->storage*8;
Chris@69:    tell = ec_tell(dec);
Chris@69:    logp = isTransient ? 2 : 4;
Chris@69:    tf_select_rsv = LM>0 && tell+logp+1<=budget;
Chris@69:    budget -= tf_select_rsv;
Chris@69:    tf_changed = curr = 0;
Chris@69:    for (i=start;i<end;i++)
Chris@69:    {
Chris@69:       if (tell+logp<=budget)
Chris@69:       {
Chris@69:          curr ^= ec_dec_bit_logp(dec, logp);
Chris@69:          tell = ec_tell(dec);
Chris@69:          tf_changed |= curr;
Chris@69:       }
Chris@69:       tf_res[i] = curr;
Chris@69:       logp = isTransient ? 4 : 5;
Chris@69:    }
Chris@69:    tf_select = 0;
Chris@69:    if (tf_select_rsv &&
Chris@69:      tf_select_table[LM][4*isTransient+0+tf_changed] !=
Chris@69:      tf_select_table[LM][4*isTransient+2+tf_changed])
Chris@69:    {
Chris@69:       tf_select = ec_dec_bit_logp(dec, 1);
Chris@69:    }
Chris@69:    for (i=start;i<end;i++)
Chris@69:    {
Chris@69:       tf_res[i] = tf_select_table[LM][4*isTransient+2*tf_select+tf_res[i]];
Chris@69:    }
Chris@69: }
Chris@69: 
Chris@69: static int celt_plc_pitch_search(celt_sig *decode_mem[2], int C, int arch)
Chris@69: {
Chris@69:    int pitch_index;
Chris@69:    VARDECL( opus_val16, lp_pitch_buf );
Chris@69:    SAVE_STACK;
Chris@69:    ALLOC( lp_pitch_buf, DECODE_BUFFER_SIZE>>1, opus_val16 );
Chris@69:    pitch_downsample(decode_mem, lp_pitch_buf,
Chris@69:          DECODE_BUFFER_SIZE, C, arch);
Chris@69:    pitch_search(lp_pitch_buf+(PLC_PITCH_LAG_MAX>>1), lp_pitch_buf,
Chris@69:          DECODE_BUFFER_SIZE-PLC_PITCH_LAG_MAX,
Chris@69:          PLC_PITCH_LAG_MAX-PLC_PITCH_LAG_MIN, &pitch_index, arch);
Chris@69:    pitch_index = PLC_PITCH_LAG_MAX-pitch_index;
Chris@69:    RESTORE_STACK;
Chris@69:    return pitch_index;
Chris@69: }
Chris@69: 
Chris@69: static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
Chris@69: {
Chris@69:    int c;
Chris@69:    int i;
Chris@69:    const int C = st->channels;
Chris@69:    celt_sig *decode_mem[2];
Chris@69:    celt_sig *out_syn[2];
Chris@69:    opus_val16 *lpc;
Chris@69:    opus_val16 *oldBandE, *oldLogE, *oldLogE2, *backgroundLogE;
Chris@69:    const OpusCustomMode *mode;
Chris@69:    int nbEBands;
Chris@69:    int overlap;
Chris@69:    int start;
Chris@69:    int loss_count;
Chris@69:    int noise_based;
Chris@69:    const opus_int16 *eBands;
Chris@69:    SAVE_STACK;
Chris@69: 
Chris@69:    mode = st->mode;
Chris@69:    nbEBands = mode->nbEBands;
Chris@69:    overlap = mode->overlap;
Chris@69:    eBands = mode->eBands;
Chris@69: 
Chris@69:    c=0; do {
Chris@69:       decode_mem[c] = st->_decode_mem + c*(DECODE_BUFFER_SIZE+overlap);
Chris@69:       out_syn[c] = decode_mem[c]+DECODE_BUFFER_SIZE-N;
Chris@69:    } while (++c<C);
Chris@69:    lpc = (opus_val16*)(st->_decode_mem+(DECODE_BUFFER_SIZE+overlap)*C);
Chris@69:    oldBandE = lpc+C*LPC_ORDER;
Chris@69:    oldLogE = oldBandE + 2*nbEBands;
Chris@69:    oldLogE2 = oldLogE + 2*nbEBands;
Chris@69:    backgroundLogE = oldLogE2  + 2*nbEBands;
Chris@69: 
Chris@69:    loss_count = st->loss_count;
Chris@69:    start = st->start;
Chris@69:    noise_based = loss_count >= 5 || start != 0 || st->skip_plc;
Chris@69:    if (noise_based)
Chris@69:    {
Chris@69:       /* Noise-based PLC/CNG */
Chris@69: #ifdef NORM_ALIASING_HACK
Chris@69:       celt_norm *X;
Chris@69: #else
Chris@69:       VARDECL(celt_norm, X);
Chris@69: #endif
Chris@69:       opus_uint32 seed;
Chris@69:       int end;
Chris@69:       int effEnd;
Chris@69:       opus_val16 decay;
Chris@69:       end = st->end;
Chris@69:       effEnd = IMAX(start, IMIN(end, mode->effEBands));
Chris@69: 
Chris@69: #ifdef NORM_ALIASING_HACK
Chris@69:       /* This is an ugly hack that breaks aliasing rules and would be easily broken,
Chris@69:          but it saves almost 4kB of stack. */
Chris@69:       X = (celt_norm*)(out_syn[C-1]+overlap/2);
Chris@69: #else
Chris@69:       ALLOC(X, C*N, celt_norm);   /**< Interleaved normalised MDCTs */
Chris@69: #endif
Chris@69: 
Chris@69:       /* Energy decay */
Chris@69:       decay = loss_count==0 ? QCONST16(1.5f, DB_SHIFT) : QCONST16(.5f, DB_SHIFT);
Chris@69:       c=0; do
Chris@69:       {
Chris@69:          for (i=start;i<end;i++)
Chris@69:             oldBandE[c*nbEBands+i] = MAX16(backgroundLogE[c*nbEBands+i], oldBandE[c*nbEBands+i] - decay);
Chris@69:       } while (++c<C);
Chris@69:       seed = st->rng;
Chris@69:       for (c=0;c<C;c++)
Chris@69:       {
Chris@69:          for (i=start;i<effEnd;i++)
Chris@69:          {
Chris@69:             int j;
Chris@69:             int boffs;
Chris@69:             int blen;
Chris@69:             boffs = N*c+(eBands[i]<<LM);
Chris@69:             blen = (eBands[i+1]-eBands[i])<<LM;
Chris@69:             for (j=0;j<blen;j++)
Chris@69:             {
Chris@69:                seed = celt_lcg_rand(seed);
Chris@69:                X[boffs+j] = (celt_norm)((opus_int32)seed>>20);
Chris@69:             }
Chris@69:             renormalise_vector(X+boffs, blen, Q15ONE, st->arch);
Chris@69:          }
Chris@69:       }
Chris@69:       st->rng = seed;
Chris@69: 
Chris@69:       c=0; do {
Chris@69:          OPUS_MOVE(decode_mem[c], decode_mem[c]+N,
Chris@69:                DECODE_BUFFER_SIZE-N+(overlap>>1));
Chris@69:       } while (++c<C);
Chris@69: 
Chris@69:       celt_synthesis(mode, X, out_syn, oldBandE, start, effEnd, C, C, 0, LM, st->downsample, 0, st->arch);
Chris@69:    } else {
Chris@69:       int exc_length;
Chris@69:       /* Pitch-based PLC */
Chris@69:       const opus_val16 *window;
Chris@69:       opus_val16 *exc;
Chris@69:       opus_val16 fade = Q15ONE;
Chris@69:       int pitch_index;
Chris@69:       VARDECL(opus_val32, etmp);
Chris@69:       VARDECL(opus_val16, _exc);
Chris@69:       VARDECL(opus_val16, fir_tmp);
Chris@69: 
Chris@69:       if (loss_count == 0)
Chris@69:       {
Chris@69:          st->last_pitch_index = pitch_index = celt_plc_pitch_search(decode_mem, C, st->arch);
Chris@69:       } else {
Chris@69:          pitch_index = st->last_pitch_index;
Chris@69:          fade = QCONST16(.8f,15);
Chris@69:       }
Chris@69: 
Chris@69:       /* We want the excitation for 2 pitch periods in order to look for a
Chris@69:          decaying signal, but we can't get more than MAX_PERIOD. */
Chris@69:       exc_length = IMIN(2*pitch_index, MAX_PERIOD);
Chris@69: 
Chris@69:       ALLOC(etmp, overlap, opus_val32);
Chris@69:       ALLOC(_exc, MAX_PERIOD+LPC_ORDER, opus_val16);
Chris@69:       ALLOC(fir_tmp, exc_length, opus_val16);
Chris@69:       exc = _exc+LPC_ORDER;
Chris@69:       window = mode->window;
Chris@69:       c=0; do {
Chris@69:          opus_val16 decay;
Chris@69:          opus_val16 attenuation;
Chris@69:          opus_val32 S1=0;
Chris@69:          celt_sig *buf;
Chris@69:          int extrapolation_offset;
Chris@69:          int extrapolation_len;
Chris@69:          int j;
Chris@69: 
Chris@69:          buf = decode_mem[c];
Chris@69:          for (i=0;i<MAX_PERIOD+LPC_ORDER;i++)
Chris@69:             exc[i-LPC_ORDER] = ROUND16(buf[DECODE_BUFFER_SIZE-MAX_PERIOD-LPC_ORDER+i], SIG_SHIFT);
Chris@69: 
Chris@69:          if (loss_count == 0)
Chris@69:          {
Chris@69:             opus_val32 ac[LPC_ORDER+1];
Chris@69:             /* Compute LPC coefficients for the last MAX_PERIOD samples before
Chris@69:                the first loss so we can work in the excitation-filter domain. */
Chris@69:             _celt_autocorr(exc, ac, window, overlap,
Chris@69:                    LPC_ORDER, MAX_PERIOD, st->arch);
Chris@69:             /* Add a noise floor of -40 dB. */
Chris@69: #ifdef FIXED_POINT
Chris@69:             ac[0] += SHR32(ac[0],13);
Chris@69: #else
Chris@69:             ac[0] *= 1.0001f;
Chris@69: #endif
Chris@69:             /* Use lag windowing to stabilize the Levinson-Durbin recursion. */
Chris@69:             for (i=1;i<=LPC_ORDER;i++)
Chris@69:             {
Chris@69:                /*ac[i] *= exp(-.5*(2*M_PI*.002*i)*(2*M_PI*.002*i));*/
Chris@69: #ifdef FIXED_POINT
Chris@69:                ac[i] -= MULT16_32_Q15(2*i*i, ac[i]);
Chris@69: #else
Chris@69:                ac[i] -= ac[i]*(0.008f*0.008f)*i*i;
Chris@69: #endif
Chris@69:             }
Chris@69:             _celt_lpc(lpc+c*LPC_ORDER, ac, LPC_ORDER);
Chris@69: #ifdef FIXED_POINT
Chris@69:          /* For fixed-point, apply bandwidth expansion until we can guarantee that
Chris@69:             no overflow can happen in the IIR filter. This means:
Chris@69:             32768*sum(abs(filter)) < 2^31 */
Chris@69:          while (1) {
Chris@69:             opus_val16 tmp=Q15ONE;
Chris@69:             opus_val32 sum=QCONST16(1., SIG_SHIFT);
Chris@69:             for (i=0;i<LPC_ORDER;i++)
Chris@69:                sum += ABS16(lpc[c*LPC_ORDER+i]);
Chris@69:             if (sum < 65535) break;
Chris@69:             for (i=0;i<LPC_ORDER;i++)
Chris@69:             {
Chris@69:                tmp = MULT16_16_Q15(QCONST16(.99f,15), tmp);
Chris@69:                lpc[c*LPC_ORDER+i] = MULT16_16_Q15(lpc[c*LPC_ORDER+i], tmp);
Chris@69:             }
Chris@69:          }
Chris@69: #endif
Chris@69:          }
Chris@69:          /* Initialize the LPC history with the samples just before the start
Chris@69:             of the region for which we're computing the excitation. */
Chris@69:          {
Chris@69:             /* Compute the excitation for exc_length samples before the loss. We need the copy
Chris@69:                because celt_fir() cannot filter in-place. */
Chris@69:             celt_fir(exc+MAX_PERIOD-exc_length, lpc+c*LPC_ORDER,
Chris@69:                   fir_tmp, exc_length, LPC_ORDER, st->arch);
Chris@69:             OPUS_COPY(exc+MAX_PERIOD-exc_length, fir_tmp, exc_length);
Chris@69:          }
Chris@69: 
Chris@69:          /* Check if the waveform is decaying, and if so how fast.
Chris@69:             We do this to avoid adding energy when concealing in a segment
Chris@69:             with decaying energy. */
Chris@69:          {
Chris@69:             opus_val32 E1=1, E2=1;
Chris@69:             int decay_length;
Chris@69: #ifdef FIXED_POINT
Chris@69:             int shift = IMAX(0,2*celt_zlog2(celt_maxabs16(&exc[MAX_PERIOD-exc_length], exc_length))-20);
Chris@69: #endif
Chris@69:             decay_length = exc_length>>1;
Chris@69:             for (i=0;i<decay_length;i++)
Chris@69:             {
Chris@69:                opus_val16 e;
Chris@69:                e = exc[MAX_PERIOD-decay_length+i];
Chris@69:                E1 += SHR32(MULT16_16(e, e), shift);
Chris@69:                e = exc[MAX_PERIOD-2*decay_length+i];
Chris@69:                E2 += SHR32(MULT16_16(e, e), shift);
Chris@69:             }
Chris@69:             E1 = MIN32(E1, E2);
Chris@69:             decay = celt_sqrt(frac_div32(SHR32(E1, 1), E2));
Chris@69:          }
Chris@69: 
Chris@69:          /* Move the decoder memory one frame to the left to give us room to
Chris@69:             add the data for the new frame. We ignore the overlap that extends
Chris@69:             past the end of the buffer, because we aren't going to use it. */
Chris@69:          OPUS_MOVE(buf, buf+N, DECODE_BUFFER_SIZE-N);
Chris@69: 
Chris@69:          /* Extrapolate from the end of the excitation with a period of
Chris@69:             "pitch_index", scaling down each period by an additional factor of
Chris@69:             "decay". */
Chris@69:          extrapolation_offset = MAX_PERIOD-pitch_index;
Chris@69:          /* We need to extrapolate enough samples to cover a complete MDCT
Chris@69:             window (including overlap/2 samples on both sides). */
Chris@69:          extrapolation_len = N+overlap;
Chris@69:          /* We also apply fading if this is not the first loss. */
Chris@69:          attenuation = MULT16_16_Q15(fade, decay);
Chris@69:          for (i=j=0;i<extrapolation_len;i++,j++)
Chris@69:          {
Chris@69:             opus_val16 tmp;
Chris@69:             if (j >= pitch_index) {
Chris@69:                j -= pitch_index;
Chris@69:                attenuation = MULT16_16_Q15(attenuation, decay);
Chris@69:             }
Chris@69:             buf[DECODE_BUFFER_SIZE-N+i] =
Chris@69:                   SHL32(EXTEND32(MULT16_16_Q15(attenuation,
Chris@69:                         exc[extrapolation_offset+j])), SIG_SHIFT);
Chris@69:             /* Compute the energy of the previously decoded signal whose
Chris@69:                excitation we're copying. */
Chris@69:             tmp = ROUND16(
Chris@69:                   buf[DECODE_BUFFER_SIZE-MAX_PERIOD-N+extrapolation_offset+j],
Chris@69:                   SIG_SHIFT);
Chris@69:             S1 += SHR32(MULT16_16(tmp, tmp), 10);
Chris@69:          }
Chris@69:          {
Chris@69:             opus_val16 lpc_mem[LPC_ORDER];
Chris@69:             /* Copy the last decoded samples (prior to the overlap region) to
Chris@69:                synthesis filter memory so we can have a continuous signal. */
Chris@69:             for (i=0;i<LPC_ORDER;i++)
Chris@69:                lpc_mem[i] = ROUND16(buf[DECODE_BUFFER_SIZE-N-1-i], SIG_SHIFT);
Chris@69:             /* Apply the synthesis filter to convert the excitation back into
Chris@69:                the signal domain. */
Chris@69:             celt_iir(buf+DECODE_BUFFER_SIZE-N, lpc+c*LPC_ORDER,
Chris@69:                   buf+DECODE_BUFFER_SIZE-N, extrapolation_len, LPC_ORDER,
Chris@69:                   lpc_mem, st->arch);
Chris@69: #ifdef FIXED_POINT
Chris@69:             for (i=0; i < extrapolation_len; i++)
Chris@69:                buf[DECODE_BUFFER_SIZE-N+i] = SATURATE(buf[DECODE_BUFFER_SIZE-N+i], SIG_SAT);
Chris@69: #endif
Chris@69:          }
Chris@69: 
Chris@69:          /* Check if the synthesis energy is higher than expected, which can
Chris@69:             happen with the signal changes during our window. If so,
Chris@69:             attenuate. */
Chris@69:          {
Chris@69:             opus_val32 S2=0;
Chris@69:             for (i=0;i<extrapolation_len;i++)
Chris@69:             {
Chris@69:                opus_val16 tmp = ROUND16(buf[DECODE_BUFFER_SIZE-N+i], SIG_SHIFT);
Chris@69:                S2 += SHR32(MULT16_16(tmp, tmp), 10);
Chris@69:             }
Chris@69:             /* This checks for an "explosion" in the synthesis. */
Chris@69: #ifdef FIXED_POINT
Chris@69:             if (!(S1 > SHR32(S2,2)))
Chris@69: #else
Chris@69:             /* The float test is written this way to catch NaNs in the output
Chris@69:                of the IIR filter at the same time. */
Chris@69:             if (!(S1 > 0.2f*S2))
Chris@69: #endif
Chris@69:             {
Chris@69:                for (i=0;i<extrapolation_len;i++)
Chris@69:                   buf[DECODE_BUFFER_SIZE-N+i] = 0;
Chris@69:             } else if (S1 < S2)
Chris@69:             {
Chris@69:                opus_val16 ratio = celt_sqrt(frac_div32(SHR32(S1,1)+1,S2+1));
Chris@69:                for (i=0;i<overlap;i++)
Chris@69:                {
Chris@69:                   opus_val16 tmp_g = Q15ONE
Chris@69:                         - MULT16_16_Q15(window[i], Q15ONE-ratio);
Chris@69:                   buf[DECODE_BUFFER_SIZE-N+i] =
Chris@69:                         MULT16_32_Q15(tmp_g, buf[DECODE_BUFFER_SIZE-N+i]);
Chris@69:                }
Chris@69:                for (i=overlap;i<extrapolation_len;i++)
Chris@69:                {
Chris@69:                   buf[DECODE_BUFFER_SIZE-N+i] =
Chris@69:                         MULT16_32_Q15(ratio, buf[DECODE_BUFFER_SIZE-N+i]);
Chris@69:                }
Chris@69:             }
Chris@69:          }
Chris@69: 
Chris@69:          /* Apply the pre-filter to the MDCT overlap for the next frame because
Chris@69:             the post-filter will be re-applied in the decoder after the MDCT
Chris@69:             overlap. */
Chris@69:          comb_filter(etmp, buf+DECODE_BUFFER_SIZE,
Chris@69:               st->postfilter_period, st->postfilter_period, overlap,
Chris@69:               -st->postfilter_gain, -st->postfilter_gain,
Chris@69:               st->postfilter_tapset, st->postfilter_tapset, NULL, 0, st->arch);
Chris@69: 
Chris@69:          /* Simulate TDAC on the concealed audio so that it blends with the
Chris@69:             MDCT of the next frame. */
Chris@69:          for (i=0;i<overlap/2;i++)
Chris@69:          {
Chris@69:             buf[DECODE_BUFFER_SIZE+i] =
Chris@69:                MULT16_32_Q15(window[i], etmp[overlap-1-i])
Chris@69:                + MULT16_32_Q15(window[overlap-i-1], etmp[i]);
Chris@69:          }
Chris@69:       } while (++c<C);
Chris@69:    }
Chris@69: 
Chris@69:    st->loss_count = loss_count+1;
Chris@69: 
Chris@69:    RESTORE_STACK;
Chris@69: }
Chris@69: 
Chris@69: int celt_decode_with_ec(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data,
Chris@69:       int len, opus_val16 * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum)
Chris@69: {
Chris@69:    int c, i, N;
Chris@69:    int spread_decision;
Chris@69:    opus_int32 bits;
Chris@69:    ec_dec _dec;
Chris@69: #ifdef NORM_ALIASING_HACK
Chris@69:    celt_norm *X;
Chris@69: #else
Chris@69:    VARDECL(celt_norm, X);
Chris@69: #endif
Chris@69:    VARDECL(int, fine_quant);
Chris@69:    VARDECL(int, pulses);
Chris@69:    VARDECL(int, cap);
Chris@69:    VARDECL(int, offsets);
Chris@69:    VARDECL(int, fine_priority);
Chris@69:    VARDECL(int, tf_res);
Chris@69:    VARDECL(unsigned char, collapse_masks);
Chris@69:    celt_sig *decode_mem[2];
Chris@69:    celt_sig *out_syn[2];
Chris@69:    opus_val16 *lpc;
Chris@69:    opus_val16 *oldBandE, *oldLogE, *oldLogE2, *backgroundLogE;
Chris@69: 
Chris@69:    int shortBlocks;
Chris@69:    int isTransient;
Chris@69:    int intra_ener;
Chris@69:    const int CC = st->channels;
Chris@69:    int LM, M;
Chris@69:    int start;
Chris@69:    int end;
Chris@69:    int effEnd;
Chris@69:    int codedBands;
Chris@69:    int alloc_trim;
Chris@69:    int postfilter_pitch;
Chris@69:    opus_val16 postfilter_gain;
Chris@69:    int intensity=0;
Chris@69:    int dual_stereo=0;
Chris@69:    opus_int32 total_bits;
Chris@69:    opus_int32 balance;
Chris@69:    opus_int32 tell;
Chris@69:    int dynalloc_logp;
Chris@69:    int postfilter_tapset;
Chris@69:    int anti_collapse_rsv;
Chris@69:    int anti_collapse_on=0;
Chris@69:    int silence;
Chris@69:    int C = st->stream_channels;
Chris@69:    const OpusCustomMode *mode;
Chris@69:    int nbEBands;
Chris@69:    int overlap;
Chris@69:    const opus_int16 *eBands;
Chris@69:    ALLOC_STACK;
Chris@69: 
Chris@69:    VALIDATE_CELT_DECODER(st);
Chris@69:    mode = st->mode;
Chris@69:    nbEBands = mode->nbEBands;
Chris@69:    overlap = mode->overlap;
Chris@69:    eBands = mode->eBands;
Chris@69:    start = st->start;
Chris@69:    end = st->end;
Chris@69:    frame_size *= st->downsample;
Chris@69: 
Chris@69:    lpc = (opus_val16*)(st->_decode_mem+(DECODE_BUFFER_SIZE+overlap)*CC);
Chris@69:    oldBandE = lpc+CC*LPC_ORDER;
Chris@69:    oldLogE = oldBandE + 2*nbEBands;
Chris@69:    oldLogE2 = oldLogE + 2*nbEBands;
Chris@69:    backgroundLogE = oldLogE2  + 2*nbEBands;
Chris@69: 
Chris@69: #ifdef CUSTOM_MODES
Chris@69:    if (st->signalling && data!=NULL)
Chris@69:    {
Chris@69:       int data0=data[0];
Chris@69:       /* Convert "standard mode" to Opus header */
Chris@69:       if (mode->Fs==48000 && mode->shortMdctSize==120)
Chris@69:       {
Chris@69:          data0 = fromOpus(data0);
Chris@69:          if (data0<0)
Chris@69:             return OPUS_INVALID_PACKET;
Chris@69:       }
Chris@69:       st->end = end = IMAX(1, mode->effEBands-2*(data0>>5));
Chris@69:       LM = (data0>>3)&0x3;
Chris@69:       C = 1 + ((data0>>2)&0x1);
Chris@69:       data++;
Chris@69:       len--;
Chris@69:       if (LM>mode->maxLM)
Chris@69:          return OPUS_INVALID_PACKET;
Chris@69:       if (frame_size < mode->shortMdctSize<<LM)
Chris@69:          return OPUS_BUFFER_TOO_SMALL;
Chris@69:       else
Chris@69:          frame_size = mode->shortMdctSize<<LM;
Chris@69:    } else {
Chris@69: #else
Chris@69:    {
Chris@69: #endif
Chris@69:       for (LM=0;LM<=mode->maxLM;LM++)
Chris@69:          if (mode->shortMdctSize<<LM==frame_size)
Chris@69:             break;
Chris@69:       if (LM>mode->maxLM)
Chris@69:          return OPUS_BAD_ARG;
Chris@69:    }
Chris@69:    M=1<<LM;
Chris@69: 
Chris@69:    if (len<0 || len>1275 || pcm==NULL)
Chris@69:       return OPUS_BAD_ARG;
Chris@69: 
Chris@69:    N = M*mode->shortMdctSize;
Chris@69:    c=0; do {
Chris@69:       decode_mem[c] = st->_decode_mem + c*(DECODE_BUFFER_SIZE+overlap);
Chris@69:       out_syn[c] = decode_mem[c]+DECODE_BUFFER_SIZE-N;
Chris@69:    } while (++c<CC);
Chris@69: 
Chris@69:    effEnd = end;
Chris@69:    if (effEnd > mode->effEBands)
Chris@69:       effEnd = mode->effEBands;
Chris@69: 
Chris@69:    if (data == NULL || len<=1)
Chris@69:    {
Chris@69:       celt_decode_lost(st, N, LM);
Chris@69:       deemphasis(out_syn, pcm, N, CC, st->downsample, mode->preemph, st->preemph_memD, accum);
Chris@69:       RESTORE_STACK;
Chris@69:       return frame_size/st->downsample;
Chris@69:    }
Chris@69: 
Chris@69:    /* Check if there are at least two packets received consecutively before
Chris@69:     * turning on the pitch-based PLC */
Chris@69:    st->skip_plc = st->loss_count != 0;
Chris@69: 
Chris@69:    if (dec == NULL)
Chris@69:    {
Chris@69:       ec_dec_init(&_dec,(unsigned char*)data,len);
Chris@69:       dec = &_dec;
Chris@69:    }
Chris@69: 
Chris@69:    if (C==1)
Chris@69:    {
Chris@69:       for (i=0;i<nbEBands;i++)
Chris@69:          oldBandE[i]=MAX16(oldBandE[i],oldBandE[nbEBands+i]);
Chris@69:    }
Chris@69: 
Chris@69:    total_bits = len*8;
Chris@69:    tell = ec_tell(dec);
Chris@69: 
Chris@69:    if (tell >= total_bits)
Chris@69:       silence = 1;
Chris@69:    else if (tell==1)
Chris@69:       silence = ec_dec_bit_logp(dec, 15);
Chris@69:    else
Chris@69:       silence = 0;
Chris@69:    if (silence)
Chris@69:    {
Chris@69:       /* Pretend we've read all the remaining bits */
Chris@69:       tell = len*8;
Chris@69:       dec->nbits_total+=tell-ec_tell(dec);
Chris@69:    }
Chris@69: 
Chris@69:    postfilter_gain = 0;
Chris@69:    postfilter_pitch = 0;
Chris@69:    postfilter_tapset = 0;
Chris@69:    if (start==0 && tell+16 <= total_bits)
Chris@69:    {
Chris@69:       if(ec_dec_bit_logp(dec, 1))
Chris@69:       {
Chris@69:          int qg, octave;
Chris@69:          octave = ec_dec_uint(dec, 6);
Chris@69:          postfilter_pitch = (16<<octave)+ec_dec_bits(dec, 4+octave)-1;
Chris@69:          qg = ec_dec_bits(dec, 3);
Chris@69:          if (ec_tell(dec)+2<=total_bits)
Chris@69:             postfilter_tapset = ec_dec_icdf(dec, tapset_icdf, 2);
Chris@69:          postfilter_gain = QCONST16(.09375f,15)*(qg+1);
Chris@69:       }
Chris@69:       tell = ec_tell(dec);
Chris@69:    }
Chris@69: 
Chris@69:    if (LM > 0 && tell+3 <= total_bits)
Chris@69:    {
Chris@69:       isTransient = ec_dec_bit_logp(dec, 3);
Chris@69:       tell = ec_tell(dec);
Chris@69:    }
Chris@69:    else
Chris@69:       isTransient = 0;
Chris@69: 
Chris@69:    if (isTransient)
Chris@69:       shortBlocks = M;
Chris@69:    else
Chris@69:       shortBlocks = 0;
Chris@69: 
Chris@69:    /* Decode the global flags (first symbols in the stream) */
Chris@69:    intra_ener = tell+3<=total_bits ? ec_dec_bit_logp(dec, 3) : 0;
Chris@69:    /* Get band energies */
Chris@69:    unquant_coarse_energy(mode, start, end, oldBandE,
Chris@69:          intra_ener, dec, C, LM);
Chris@69: 
Chris@69:    ALLOC(tf_res, nbEBands, int);
Chris@69:    tf_decode(start, end, isTransient, tf_res, LM, dec);
Chris@69: 
Chris@69:    tell = ec_tell(dec);
Chris@69:    spread_decision = SPREAD_NORMAL;
Chris@69:    if (tell+4 <= total_bits)
Chris@69:       spread_decision = ec_dec_icdf(dec, spread_icdf, 5);
Chris@69: 
Chris@69:    ALLOC(cap, nbEBands, int);
Chris@69: 
Chris@69:    init_caps(mode,cap,LM,C);
Chris@69: 
Chris@69:    ALLOC(offsets, nbEBands, int);
Chris@69: 
Chris@69:    dynalloc_logp = 6;
Chris@69:    total_bits<<=BITRES;
Chris@69:    tell = ec_tell_frac(dec);
Chris@69:    for (i=start;i<end;i++)
Chris@69:    {
Chris@69:       int width, quanta;
Chris@69:       int dynalloc_loop_logp;
Chris@69:       int boost;
Chris@69:       width = C*(eBands[i+1]-eBands[i])<<LM;
Chris@69:       /* quanta is 6 bits, but no more than 1 bit/sample
Chris@69:          and no less than 1/8 bit/sample */
Chris@69:       quanta = IMIN(width<<BITRES, IMAX(6<<BITRES, width));
Chris@69:       dynalloc_loop_logp = dynalloc_logp;
Chris@69:       boost = 0;
Chris@69:       while (tell+(dynalloc_loop_logp<<BITRES) < total_bits && boost < cap[i])
Chris@69:       {
Chris@69:          int flag;
Chris@69:          flag = ec_dec_bit_logp(dec, dynalloc_loop_logp);
Chris@69:          tell = ec_tell_frac(dec);
Chris@69:          if (!flag)
Chris@69:             break;
Chris@69:          boost += quanta;
Chris@69:          total_bits -= quanta;
Chris@69:          dynalloc_loop_logp = 1;
Chris@69:       }
Chris@69:       offsets[i] = boost;
Chris@69:       /* Making dynalloc more likely */
Chris@69:       if (boost>0)
Chris@69:          dynalloc_logp = IMAX(2, dynalloc_logp-1);
Chris@69:    }
Chris@69: 
Chris@69:    ALLOC(fine_quant, nbEBands, int);
Chris@69:    alloc_trim = tell+(6<<BITRES) <= total_bits ?
Chris@69:          ec_dec_icdf(dec, trim_icdf, 7) : 5;
Chris@69: 
Chris@69:    bits = (((opus_int32)len*8)<<BITRES) - ec_tell_frac(dec) - 1;
Chris@69:    anti_collapse_rsv = isTransient&&LM>=2&&bits>=((LM+2)<<BITRES) ? (1<<BITRES) : 0;
Chris@69:    bits -= anti_collapse_rsv;
Chris@69: 
Chris@69:    ALLOC(pulses, nbEBands, int);
Chris@69:    ALLOC(fine_priority, nbEBands, int);
Chris@69: 
Chris@69:    codedBands = clt_compute_allocation(mode, start, end, offsets, cap,
Chris@69:          alloc_trim, &intensity, &dual_stereo, bits, &balance, pulses,
Chris@69:          fine_quant, fine_priority, C, LM, dec, 0, 0, 0);
Chris@69: 
Chris@69:    unquant_fine_energy(mode, start, end, oldBandE, fine_quant, dec, C);
Chris@69: 
Chris@69:    c=0; do {
Chris@69:       OPUS_MOVE(decode_mem[c], decode_mem[c]+N, DECODE_BUFFER_SIZE-N+overlap/2);
Chris@69:    } while (++c<CC);
Chris@69: 
Chris@69:    /* Decode fixed codebook */
Chris@69:    ALLOC(collapse_masks, C*nbEBands, unsigned char);
Chris@69: 
Chris@69: #ifdef NORM_ALIASING_HACK
Chris@69:    /* This is an ugly hack that breaks aliasing rules and would be easily broken,
Chris@69:       but it saves almost 4kB of stack. */
Chris@69:    X = (celt_norm*)(out_syn[CC-1]+overlap/2);
Chris@69: #else
Chris@69:    ALLOC(X, C*N, celt_norm);   /**< Interleaved normalised MDCTs */
Chris@69: #endif
Chris@69: 
Chris@69:    quant_all_bands(0, mode, start, end, X, C==2 ? X+N : NULL, collapse_masks,
Chris@69:          NULL, pulses, shortBlocks, spread_decision, dual_stereo, intensity, tf_res,
Chris@69:          len*(8<<BITRES)-anti_collapse_rsv, balance, dec, LM, codedBands, &st->rng, 0,
Chris@69:          st->arch, st->disable_inv);
Chris@69: 
Chris@69:    if (anti_collapse_rsv > 0)
Chris@69:    {
Chris@69:       anti_collapse_on = ec_dec_bits(dec, 1);
Chris@69:    }
Chris@69: 
Chris@69:    unquant_energy_finalise(mode, start, end, oldBandE,
Chris@69:          fine_quant, fine_priority, len*8-ec_tell(dec), dec, C);
Chris@69: 
Chris@69:    if (anti_collapse_on)
Chris@69:       anti_collapse(mode, X, collapse_masks, LM, C, N,
Chris@69:             start, end, oldBandE, oldLogE, oldLogE2, pulses, st->rng, st->arch);
Chris@69: 
Chris@69:    if (silence)
Chris@69:    {
Chris@69:       for (i=0;i<C*nbEBands;i++)
Chris@69:          oldBandE[i] = -QCONST16(28.f,DB_SHIFT);
Chris@69:    }
Chris@69: 
Chris@69:    celt_synthesis(mode, X, out_syn, oldBandE, start, effEnd,
Chris@69:                   C, CC, isTransient, LM, st->downsample, silence, st->arch);
Chris@69: 
Chris@69:    c=0; do {
Chris@69:       st->postfilter_period=IMAX(st->postfilter_period, COMBFILTER_MINPERIOD);
Chris@69:       st->postfilter_period_old=IMAX(st->postfilter_period_old, COMBFILTER_MINPERIOD);
Chris@69:       comb_filter(out_syn[c], out_syn[c], st->postfilter_period_old, st->postfilter_period, mode->shortMdctSize,
Chris@69:             st->postfilter_gain_old, st->postfilter_gain, st->postfilter_tapset_old, st->postfilter_tapset,
Chris@69:             mode->window, overlap, st->arch);
Chris@69:       if (LM!=0)
Chris@69:          comb_filter(out_syn[c]+mode->shortMdctSize, out_syn[c]+mode->shortMdctSize, st->postfilter_period, postfilter_pitch, N-mode->shortMdctSize,
Chris@69:                st->postfilter_gain, postfilter_gain, st->postfilter_tapset, postfilter_tapset,
Chris@69:                mode->window, overlap, st->arch);
Chris@69: 
Chris@69:    } while (++c<CC);
Chris@69:    st->postfilter_period_old = st->postfilter_period;
Chris@69:    st->postfilter_gain_old = st->postfilter_gain;
Chris@69:    st->postfilter_tapset_old = st->postfilter_tapset;
Chris@69:    st->postfilter_period = postfilter_pitch;
Chris@69:    st->postfilter_gain = postfilter_gain;
Chris@69:    st->postfilter_tapset = postfilter_tapset;
Chris@69:    if (LM!=0)
Chris@69:    {
Chris@69:       st->postfilter_period_old = st->postfilter_period;
Chris@69:       st->postfilter_gain_old = st->postfilter_gain;
Chris@69:       st->postfilter_tapset_old = st->postfilter_tapset;
Chris@69:    }
Chris@69: 
Chris@69:    if (C==1)
Chris@69:       OPUS_COPY(&oldBandE[nbEBands], oldBandE, nbEBands);
Chris@69: 
Chris@69:    /* In case start or end were to change */
Chris@69:    if (!isTransient)
Chris@69:    {
Chris@69:       opus_val16 max_background_increase;
Chris@69:       OPUS_COPY(oldLogE2, oldLogE, 2*nbEBands);
Chris@69:       OPUS_COPY(oldLogE, oldBandE, 2*nbEBands);
Chris@69:       /* In normal circumstances, we only allow the noise floor to increase by
Chris@69:          up to 2.4 dB/second, but when we're in DTX, we allow up to 6 dB
Chris@69:          increase for each update.*/
Chris@69:       if (st->loss_count < 10)
Chris@69:          max_background_increase = M*QCONST16(0.001f,DB_SHIFT);
Chris@69:       else
Chris@69:          max_background_increase = QCONST16(1.f,DB_SHIFT);
Chris@69:       for (i=0;i<2*nbEBands;i++)
Chris@69:          backgroundLogE[i] = MIN16(backgroundLogE[i] + max_background_increase, oldBandE[i]);
Chris@69:    } else {
Chris@69:       for (i=0;i<2*nbEBands;i++)
Chris@69:          oldLogE[i] = MIN16(oldLogE[i], oldBandE[i]);
Chris@69:    }
Chris@69:    c=0; do
Chris@69:    {
Chris@69:       for (i=0;i<start;i++)
Chris@69:       {
Chris@69:          oldBandE[c*nbEBands+i]=0;
Chris@69:          oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT);
Chris@69:       }
Chris@69:       for (i=end;i<nbEBands;i++)
Chris@69:       {
Chris@69:          oldBandE[c*nbEBands+i]=0;
Chris@69:          oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT);
Chris@69:       }
Chris@69:    } while (++c<2);
Chris@69:    st->rng = dec->rng;
Chris@69: 
Chris@69:    deemphasis(out_syn, pcm, N, CC, st->downsample, mode->preemph, st->preemph_memD, accum);
Chris@69:    st->loss_count = 0;
Chris@69:    RESTORE_STACK;
Chris@69:    if (ec_tell(dec) > 8*len)
Chris@69:       return OPUS_INTERNAL_ERROR;
Chris@69:    if(ec_get_error(dec))
Chris@69:       st->error = 1;
Chris@69:    return frame_size/st->downsample;
Chris@69: }
Chris@69: 
Chris@69: 
Chris@69: #ifdef CUSTOM_MODES
Chris@69: 
Chris@69: #ifdef FIXED_POINT
Chris@69: int opus_custom_decode(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data, int len, opus_int16 * OPUS_RESTRICT pcm, int frame_size)
Chris@69: {
Chris@69:    return celt_decode_with_ec(st, data, len, pcm, frame_size, NULL, 0);
Chris@69: }
Chris@69: 
Chris@69: #ifndef DISABLE_FLOAT_API
Chris@69: int opus_custom_decode_float(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data, int len, float * OPUS_RESTRICT pcm, int frame_size)
Chris@69: {
Chris@69:    int j, ret, C, N;
Chris@69:    VARDECL(opus_int16, out);
Chris@69:    ALLOC_STACK;
Chris@69: 
Chris@69:    if (pcm==NULL)
Chris@69:       return OPUS_BAD_ARG;
Chris@69: 
Chris@69:    C = st->channels;
Chris@69:    N = frame_size;
Chris@69: 
Chris@69:    ALLOC(out, C*N, opus_int16);
Chris@69:    ret=celt_decode_with_ec(st, data, len, out, frame_size, NULL, 0);
Chris@69:    if (ret>0)
Chris@69:       for (j=0;j<C*ret;j++)
Chris@69:          pcm[j]=out[j]*(1.f/32768.f);
Chris@69: 
Chris@69:    RESTORE_STACK;
Chris@69:    return ret;
Chris@69: }
Chris@69: #endif /* DISABLE_FLOAT_API */
Chris@69: 
Chris@69: #else
Chris@69: 
Chris@69: int opus_custom_decode_float(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data, int len, float * OPUS_RESTRICT pcm, int frame_size)
Chris@69: {
Chris@69:    return celt_decode_with_ec(st, data, len, pcm, frame_size, NULL, 0);
Chris@69: }
Chris@69: 
Chris@69: int opus_custom_decode(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data, int len, opus_int16 * OPUS_RESTRICT pcm, int frame_size)
Chris@69: {
Chris@69:    int j, ret, C, N;
Chris@69:    VARDECL(celt_sig, out);
Chris@69:    ALLOC_STACK;
Chris@69: 
Chris@69:    if (pcm==NULL)
Chris@69:       return OPUS_BAD_ARG;
Chris@69: 
Chris@69:    C = st->channels;
Chris@69:    N = frame_size;
Chris@69:    ALLOC(out, C*N, celt_sig);
Chris@69: 
Chris@69:    ret=celt_decode_with_ec(st, data, len, out, frame_size, NULL, 0);
Chris@69: 
Chris@69:    if (ret>0)
Chris@69:       for (j=0;j<C*ret;j++)
Chris@69:          pcm[j] = FLOAT2INT16 (out[j]);
Chris@69: 
Chris@69:    RESTORE_STACK;
Chris@69:    return ret;
Chris@69: }
Chris@69: 
Chris@69: #endif
Chris@69: #endif /* CUSTOM_MODES */
Chris@69: 
Chris@69: int opus_custom_decoder_ctl(CELTDecoder * OPUS_RESTRICT st, int request, ...)
Chris@69: {
Chris@69:    va_list ap;
Chris@69: 
Chris@69:    va_start(ap, request);
Chris@69:    switch (request)
Chris@69:    {
Chris@69:       case CELT_SET_START_BAND_REQUEST:
Chris@69:       {
Chris@69:          opus_int32 value = va_arg(ap, opus_int32);
Chris@69:          if (value<0 || value>=st->mode->nbEBands)
Chris@69:             goto bad_arg;
Chris@69:          st->start = value;
Chris@69:       }
Chris@69:       break;
Chris@69:       case CELT_SET_END_BAND_REQUEST:
Chris@69:       {
Chris@69:          opus_int32 value = va_arg(ap, opus_int32);
Chris@69:          if (value<1 || value>st->mode->nbEBands)
Chris@69:             goto bad_arg;
Chris@69:          st->end = value;
Chris@69:       }
Chris@69:       break;
Chris@69:       case CELT_SET_CHANNELS_REQUEST:
Chris@69:       {
Chris@69:          opus_int32 value = va_arg(ap, opus_int32);
Chris@69:          if (value<1 || value>2)
Chris@69:             goto bad_arg;
Chris@69:          st->stream_channels = value;
Chris@69:       }
Chris@69:       break;
Chris@69:       case CELT_GET_AND_CLEAR_ERROR_REQUEST:
Chris@69:       {
Chris@69:          opus_int32 *value = va_arg(ap, opus_int32*);
Chris@69:          if (value==NULL)
Chris@69:             goto bad_arg;
Chris@69:          *value=st->error;
Chris@69:          st->error = 0;
Chris@69:       }
Chris@69:       break;
Chris@69:       case OPUS_GET_LOOKAHEAD_REQUEST:
Chris@69:       {
Chris@69:          opus_int32 *value = va_arg(ap, opus_int32*);
Chris@69:          if (value==NULL)
Chris@69:             goto bad_arg;
Chris@69:          *value = st->overlap/st->downsample;
Chris@69:       }
Chris@69:       break;
Chris@69:       case OPUS_RESET_STATE:
Chris@69:       {
Chris@69:          int i;
Chris@69:          opus_val16 *lpc, *oldBandE, *oldLogE, *oldLogE2;
Chris@69:          lpc = (opus_val16*)(st->_decode_mem+(DECODE_BUFFER_SIZE+st->overlap)*st->channels);
Chris@69:          oldBandE = lpc+st->channels*LPC_ORDER;
Chris@69:          oldLogE = oldBandE + 2*st->mode->nbEBands;
Chris@69:          oldLogE2 = oldLogE + 2*st->mode->nbEBands;
Chris@69:          OPUS_CLEAR((char*)&st->DECODER_RESET_START,
Chris@69:                opus_custom_decoder_get_size(st->mode, st->channels)-
Chris@69:                ((char*)&st->DECODER_RESET_START - (char*)st));
Chris@69:          for (i=0;i<2*st->mode->nbEBands;i++)
Chris@69:             oldLogE[i]=oldLogE2[i]=-QCONST16(28.f,DB_SHIFT);
Chris@69:          st->skip_plc = 1;
Chris@69:       }
Chris@69:       break;
Chris@69:       case OPUS_GET_PITCH_REQUEST:
Chris@69:       {
Chris@69:          opus_int32 *value = va_arg(ap, opus_int32*);
Chris@69:          if (value==NULL)
Chris@69:             goto bad_arg;
Chris@69:          *value = st->postfilter_period;
Chris@69:       }
Chris@69:       break;
Chris@69:       case CELT_GET_MODE_REQUEST:
Chris@69:       {
Chris@69:          const CELTMode ** value = va_arg(ap, const CELTMode**);
Chris@69:          if (value==0)
Chris@69:             goto bad_arg;
Chris@69:          *value=st->mode;
Chris@69:       }
Chris@69:       break;
Chris@69:       case CELT_SET_SIGNALLING_REQUEST:
Chris@69:       {
Chris@69:          opus_int32 value = va_arg(ap, opus_int32);
Chris@69:          st->signalling = value;
Chris@69:       }
Chris@69:       break;
Chris@69:       case OPUS_GET_FINAL_RANGE_REQUEST:
Chris@69:       {
Chris@69:          opus_uint32 * value = va_arg(ap, opus_uint32 *);
Chris@69:          if (value==0)
Chris@69:             goto bad_arg;
Chris@69:          *value=st->rng;
Chris@69:       }
Chris@69:       break;
Chris@69:       case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST:
Chris@69:       {
Chris@69:           opus_int32 value = va_arg(ap, opus_int32);
Chris@69:           if(value<0 || value>1)
Chris@69:           {
Chris@69:              goto bad_arg;
Chris@69:           }
Chris@69:           st->disable_inv = value;
Chris@69:       }
Chris@69:       break;
Chris@69:       case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST:
Chris@69:       {
Chris@69:           opus_int32 *value = va_arg(ap, opus_int32*);
Chris@69:           if (!value)
Chris@69:           {
Chris@69:              goto bad_arg;
Chris@69:           }
Chris@69:           *value = st->disable_inv;
Chris@69:       }
Chris@69:       break;
Chris@69:       default:
Chris@69:          goto bad_request;
Chris@69:    }
Chris@69:    va_end(ap);
Chris@69:    return OPUS_OK;
Chris@69: bad_arg:
Chris@69:    va_end(ap);
Chris@69:    return OPUS_BAD_ARG;
Chris@69: bad_request:
Chris@69:       va_end(ap);
Chris@69:   return OPUS_UNIMPLEMENTED;
Chris@69: }