/* Copyright (c) 2007-2008 CSIRO Copyright (c) 2007-2010 Xiph.Org Foundation Copyright (c) 2008 Gregory Maxwell
Written by Jean-Marc Valin and Gregory Maxwell */ /* Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
- Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER 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 OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
/* The maximum pitch lag to allow in the pitch-based PLC. It's possible to save CPU time in the PLC pitch search by making this smaller than MAX_PERIOD. The
current value corresponds to a pitch of 66.67 Hz. */ #define PLC_PITCH_LAG_MAX (720) /* The minimum pitch lag to allow in the pitch-based PLC. This corresponds to a
pitch of 480 Hz. */ #define PLC_PITCH_LAG_MIN (100)
/** Decoder state @brief Decoder state
*/ struct OpusCustomDecoder { const OpusCustomMode *mode; int overlap; int channels; int stream_channels;
int downsample; int start, end; int signalling; int disable_inv; int complexity; int arch;
/* Everything beyond this point gets cleared on a reset */ #define DECODER_RESET_START rng
opus_uint32 rng; int error; int last_pitch_index; int loss_duration; int skip_plc; int postfilter_period; int postfilter_period_old;
opus_val16 postfilter_gain;
opus_val16 postfilter_gain_old; int postfilter_tapset; int postfilter_tapset_old; int prefilter_and_fold;
celt_sig preemph_memD[2];
#ifdef ENABLE_DEEP_PLC
opus_int16 plc_pcm[PLC_UPDATE_SAMPLES]; int plc_fill; float plc_preemphasis_mem; #endif
#ifdefined(ENABLE_HARDENING) || defined(ENABLE_ASSERTIONS) /* Make basic checks on the CELT state to ensure we don't end
up writing all over memory. */ void validate_celt_decoder(CELTDecoder *st)
{ #ifndef CUSTOM_MODES
celt_assert(st->mode == opus_custom_mode_create(48000, 960, NULL));
celt_assert(st->overlap == 120);
celt_assert(st->end <= 21); #else /* From Section 4.3 in the spec: "The normal CELT layer uses 21 of those bands, though Opus Custom (see Section 6.2) may use a different number of bands"
#ifndef CUSTOM_MODES /* Special case for stereo with no downsampling and no accumulation. This is quite common and we can make it faster by processing both channels in the
same loop, reducing overhead due to the dependency loop in the IIR filter. */ staticvoid deemphasis_stereo_simple(celt_sig *in[], opus_val16 *pcm, int N, const opus_val16 coef0,
celt_sig *mem)
{
celt_sig * OPUS_RESTRICT x0;
celt_sig * OPUS_RESTRICT x1;
celt_sig m0, m1; int j;
x0=in[0];
x1=in[1];
m0 = mem[0];
m1 = mem[1]; for (j=0;j<N;j++)
{
celt_sig tmp0, tmp1; /* Add VERY_SMALL to x[] first to reduce dependency chain. */
tmp0 = x0[j] + VERY_SMALL + m0;
tmp1 = x1[j] + VERY_SMALL + m1;
m0 = MULT16_32_Q15(coef0, tmp0);
m1 = MULT16_32_Q15(coef0, tmp1);
pcm[2*j ] = SCALEOUT(SIG2WORD16(tmp0));
pcm[2*j+1] = SCALEOUT(SIG2WORD16(tmp1));
}
mem[0] = m0;
mem[1] = m1;
} #endif
#ifndef RESYNTH static #endif void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, const opus_val16 *coef,
celt_sig *mem, int accum)
{ int c; int Nd; int apply_downsampling=0;
opus_val16 coef0;
VARDECL(celt_sig, scratch);
SAVE_STACK; #ifndef CUSTOM_MODES /* Short version for common case. */ if (downsample == 1 && C == 2 && !accum)
{
deemphasis_stereo_simple(in, pcm, N, coef[0], mem); return;
} #endif #ifndef FIXED_POINT
(void)accum;
celt_assert(accum==0); #endif
ALLOC(scratch, N, celt_sig);
coef0 = coef[0];
Nd = N/downsample;
c=0; do { int j;
celt_sig * OPUS_RESTRICT x;
opus_val16 * OPUS_RESTRICT y;
celt_sig m = mem[c];
x =in[c];
y = pcm+c; #ifdef CUSTOM_MODES if (coef[1] != 0)
{
opus_val16 coef1 = coef[1];
opus_val16 coef3 = coef[3]; for (j=0;j<N;j++)
{
celt_sig tmp = x[j] + m + VERY_SMALL;
m = MULT16_32_Q15(coef0, tmp)
- MULT16_32_Q15(coef1, x[j]);
tmp = SHL32(MULT16_32_Q15(coef3, tmp), 2);
scratch[j] = tmp;
}
apply_downsampling=1;
} else #endif if (downsample>1)
{ /* Shortcut for the standard (non-custom modes) case */ for (j=0;j<N;j++)
{
celt_sig tmp = x[j] + VERY_SMALL + m;
m = MULT16_32_Q15(coef0, tmp);
scratch[j] = tmp;
}
apply_downsampling=1;
} else { /* Shortcut for the standard (non-custom modes) case */ #ifdef FIXED_POINT if (accum)
{ for (j=0;j<N;j++)
{
celt_sig tmp = x[j] + m + VERY_SMALL;
m = MULT16_32_Q15(coef0, tmp);
y[j*C] = SAT16(ADD32(y[j*C], SCALEOUT(SIG2WORD16(tmp))));
}
} else #endif
{ for (j=0;j<N;j++)
{
celt_sig tmp = x[j] + VERY_SMALL + m;
m = MULT16_32_Q15(coef0, tmp);
y[j*C] = SCALEOUT(SIG2WORD16(tmp));
}
}
}
mem[c] = m;
if (apply_downsampling)
{ /* Perform down-sampling */ #ifdef FIXED_POINT if (accum)
{ for (j=0;j<Nd;j++)
y[j*C] = SAT16(ADD32(y[j*C], SCALEOUT(SIG2WORD16(scratch[j*downsample]))));
} else #endif
{ for (j=0;j<Nd;j++)
y[j*C] = SCALEOUT(SIG2WORD16(scratch[j*downsample]));
}
}
} while (++c<C);
RESTORE_STACK;
}
#ifndef RESYNTH static #endif void celt_synthesis(const CELTMode *mode, celt_norm *X, celt_sig * out_syn[],
opus_val16 *oldBandE, int start, int effEnd, int C, int CC, int isTransient, int LM, int downsample, int silence, int arch)
{ int c, i; int M; int b; int B; int N, NB; int shift; int nbEBands; int overlap;
VARDECL(celt_sig, freq);
SAVE_STACK;
overlap = mode->overlap;
nbEBands = mode->nbEBands;
N = mode->shortMdctSize<<LM;
ALLOC(freq, N, celt_sig); /**< Interleaved signal MDCTs */
M = 1<<LM;
if (isTransient)
{
B = M;
NB = mode->shortMdctSize;
shift = mode->maxLM;
} else {
B = 1;
NB = mode->shortMdctSize<<LM;
shift = mode->maxLM-LM;
}
if (CC==2&&C==1)
{ /* Copying a mono streams to two channels */
celt_sig *freq2;
denormalise_bands(mode, X, freq, oldBandE, start, effEnd, M,
downsample, silence); /* Store a temporary copy in the output buffer because the IMDCT destroys its input. */
freq2 = out_syn[1]+overlap/2;
OPUS_COPY(freq2, freq, N); for (b=0;b<B;b++)
clt_mdct_backward(&mode->mdct, &freq2[b], out_syn[0]+NB*b, mode->window, overlap, shift, B, arch); for (b=0;b<B;b++)
clt_mdct_backward(&mode->mdct, &freq[b], out_syn[1]+NB*b, mode->window, overlap, shift, B, arch);
} elseif (CC==1&&C==2)
{ /* Downmixing a stereo stream to mono */
celt_sig *freq2;
freq2 = out_syn[0]+overlap/2;
denormalise_bands(mode, X, freq, oldBandE, start, effEnd, M,
downsample, silence); /* Use the output buffer as temp array before downmixing. */
denormalise_bands(mode, X+N, freq2, oldBandE+nbEBands, start, effEnd, M,
downsample, silence); for (i=0;i<N;i++)
freq[i] = ADD32(HALF32(freq[i]), HALF32(freq2[i])); for (b=0;b<B;b++)
clt_mdct_backward(&mode->mdct, &freq[b], out_syn[0]+NB*b, mode->window, overlap, shift, B, arch);
} else { /* Normal case (mono or stereo) */
c=0; do {
denormalise_bands(mode, X+c*N, freq, oldBandE+c*nbEBands, start, effEnd, M,
downsample, silence); for (b=0;b<B;b++)
clt_mdct_backward(&mode->mdct, &freq[b], out_syn[c]+NB*b, mode->window, overlap, shift, B, arch);
} while (++c<CC);
} /* Saturate IMDCT output so that we can't overflow in the pitch postfilter
or in the */
c=0; do { for (i=0;i<N;i++)
out_syn[c][i] = SATURATE(out_syn[c][i], SIG_SAT);
} while (++c<CC);
RESTORE_STACK;
}
staticvoid tf_decode(int start, int end, int isTransient, int *tf_res, int LM, ec_dec *dec)
{ int i, curr, tf_select; int tf_select_rsv; int tf_changed; int logp;
opus_uint32 budget;
opus_uint32 tell;
staticvoid prefilter_and_fold(CELTDecoder * OPUS_RESTRICT st, int N)
{ int c; int CC; int i; int overlap;
celt_sig *decode_mem[2]; const OpusCustomMode *mode;
VARDECL(opus_val32, etmp);
mode = st->mode;
overlap = st->overlap;
CC = st->channels;
ALLOC(etmp, overlap, opus_val32);
c=0; do {
decode_mem[c] = st->_decode_mem + c*(DECODE_BUFFER_SIZE+overlap);
} while (++c<CC);
c=0; do { /* Apply the pre-filter to the MDCT overlap for the next frame because the post-filter will be re-applied in the decoder after the MDCT
overlap. */
comb_filter(etmp, decode_mem[c]+DECODE_BUFFER_SIZE-N,
st->postfilter_period_old, st->postfilter_period, overlap,
-st->postfilter_gain_old, -st->postfilter_gain,
st->postfilter_tapset_old, st->postfilter_tapset, NULL, 0, st->arch);
/* Simulate TDAC on the concealed audio so that it blends with the
MDCT of the next frame. */ for (i=0;i<overlap/2;i++)
{
decode_mem[c][DECODE_BUFFER_SIZE-N+i] =
MULT16_32_Q15(mode->window[i], etmp[overlap-1-i])
+ MULT16_32_Q15(mode->window[overlap-i-1], etmp[i]);
}
} while (++c<CC);
}
void update_plc_state(LPCNetPLCState *lpcnet, celt_sig *decode_mem[2], float *plc_preemphasis_mem, int CC)
{ int i; int tmp_read_post, tmp_fec_skip; int offset;
celt_sig buf48k[DECODE_BUFFER_SIZE];
opus_int16 buf16k[PLC_UPDATE_SAMPLES]; if (CC == 1) OPUS_COPY(buf48k, decode_mem[0], DECODE_BUFFER_SIZE); else { for (i=0;i<DECODE_BUFFER_SIZE;i++) {
buf48k[i] = .5*(decode_mem[0][i] + decode_mem[1][i]);
}
} /* Down-sample the last 40 ms. */ for (i=1;i<DECODE_BUFFER_SIZE;i++) buf48k[i] += PREEMPHASIS*buf48k[i-1];
*plc_preemphasis_mem = buf48k[DECODE_BUFFER_SIZE-1];
offset = DECODE_BUFFER_SIZE-SINC_ORDER-1 - 3*(PLC_UPDATE_SAMPLES-1);
celt_assert(3*(PLC_UPDATE_SAMPLES-1) + SINC_ORDER + offset == DECODE_BUFFER_SIZE-1); for (i=0;i<PLC_UPDATE_SAMPLES;i++) { int j; float sum = 0; for (j=0;j<SINC_ORDER+1;j++) {
sum += buf48k[3*i + j + offset]*sinc_filter[j];
}
buf16k[i] = float2int(MIN32(32767.f, MAX32(-32767.f, sum)));
}
tmp_read_post = lpcnet->fec_read_pos;
tmp_fec_skip = lpcnet->fec_skip; for (i=0;i<PLC_UPDATE_FRAMES;i++) {
lpcnet_plc_update(lpcnet, &buf16k[FRAME_SIZE*i]);
}
lpcnet->fec_read_pos = tmp_read_post;
lpcnet->fec_skip = tmp_fec_skip;
} #endif
staticvoid celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM #ifdef ENABLE_DEEP_PLC
,LPCNetPLCState *lpcnet #endif
)
{ int c; int i; constint C = st->channels;
celt_sig *decode_mem[2];
celt_sig *out_syn[2];
opus_val16 *lpc;
opus_val16 *oldBandE, *oldLogE, *oldLogE2, *backgroundLogE; const OpusCustomMode *mode; int nbEBands; int overlap; int start; int loss_duration; int noise_based; const opus_int16 *eBands;
SAVE_STACK;
/* We want the excitation for 2 pitch periods in order to look for a
decaying signal, but we can't get more than MAX_PERIOD. */
exc_length = IMIN(2*pitch_index, MAX_PERIOD);
ALLOC(_exc, MAX_PERIOD+CELT_LPC_ORDER, opus_val16);
ALLOC(fir_tmp, exc_length, opus_val16);
exc = _exc+CELT_LPC_ORDER;
window = mode->window;
c=0; do {
opus_val16 decay;
opus_val16 attenuation;
opus_val32 S1=0;
celt_sig *buf; int extrapolation_offset; int extrapolation_len; int j;
buf = decode_mem[c]; for (i=0;i<MAX_PERIOD+CELT_LPC_ORDER;i++)
exc[i-CELT_LPC_ORDER] = SROUND16(buf[DECODE_BUFFER_SIZE-MAX_PERIOD-CELT_LPC_ORDER+i], SIG_SHIFT);
if (loss_duration == 0)
{
opus_val32 ac[CELT_LPC_ORDER+1]; /* Compute LPC coefficients for the last MAX_PERIOD samples before
the first loss so we can work in the excitation-filter domain. */
_celt_autocorr(exc, ac, window, overlap,
CELT_LPC_ORDER, MAX_PERIOD, st->arch); /* Add a noise floor of -40 dB. */ #ifdef FIXED_POINT
ac[0] += SHR32(ac[0],13); #else
ac[0] *= 1.0001f; #endif /* Use lag windowing to stabilize the Levinson-Durbin recursion. */ for (i=1;i<=CELT_LPC_ORDER;i++)
{ /*ac[i] *= exp(-.5*(2*M_PI*.002*i)*(2*M_PI*.002*i));*/ #ifdef FIXED_POINT
ac[i] -= MULT16_32_Q15(2*i*i, ac[i]); #else
ac[i] -= ac[i]*(0.008f*0.008f)*i*i; #endif
}
_celt_lpc(lpc+c*CELT_LPC_ORDER, ac, CELT_LPC_ORDER); #ifdef FIXED_POINT /* For fixed-point, apply bandwidth expansion until we can guarantee that no overflow can happen in the IIR filter. This means:
32768*sum(abs(filter)) < 2^31 */ while (1) {
opus_val16 tmp=Q15ONE;
opus_val32 sum=QCONST16(1., SIG_SHIFT); for (i=0;i<CELT_LPC_ORDER;i++)
sum += ABS16(lpc[c*CELT_LPC_ORDER+i]); if (sum < 65535) break; for (i=0;i<CELT_LPC_ORDER;i++)
{
tmp = MULT16_16_Q15(QCONST16(.99f,15), tmp);
lpc[c*CELT_LPC_ORDER+i] = MULT16_16_Q15(lpc[c*CELT_LPC_ORDER+i], tmp);
}
} #endif
} /* Initialize the LPC history with the samples just before the start
of the region for which we're computing the excitation. */
{ /* Compute the excitation for exc_length samples before the loss. We need the copy
because celt_fir() cannot filter in-place. */
celt_fir(exc+MAX_PERIOD-exc_length, lpc+c*CELT_LPC_ORDER,
fir_tmp, exc_length, CELT_LPC_ORDER, st->arch);
OPUS_COPY(exc+MAX_PERIOD-exc_length, fir_tmp, exc_length);
}
/* Check if the waveform is decaying, and if so how fast. We do this to avoid adding energy when concealing in a segment
with decaying energy. */
{
opus_val32 E1=1, E2=1; int decay_length; #ifdef FIXED_POINT int shift = IMAX(0,2*celt_zlog2(celt_maxabs16(&exc[MAX_PERIOD-exc_length], exc_length))-20); #endif
decay_length = exc_length>>1; for (i=0;i<decay_length;i++)
{
opus_val16 e;
e = exc[MAX_PERIOD-decay_length+i];
E1 += SHR32(MULT16_16(e, e), shift);
e = exc[MAX_PERIOD-2*decay_length+i];
E2 += SHR32(MULT16_16(e, e), shift);
}
E1 = MIN32(E1, E2);
decay = celt_sqrt(frac_div32(SHR32(E1, 1), E2));
}
/* Move the decoder memory one frame to the left to give us room to add the data for the new frame. We ignore the overlap that extends
past the end of the buffer, because we aren't going to use it. */
OPUS_MOVE(buf, buf+N, DECODE_BUFFER_SIZE-N);
/* Extrapolate from the end of the excitation with a period of "pitch_index", scaling down each period by an additional factor of
"decay". */
extrapolation_offset = MAX_PERIOD-pitch_index; /* We need to extrapolate enough samples to cover a complete MDCT
window (including overlap/2 samples on both sides). */
extrapolation_len = N+overlap; /* We also apply fading if this is not the first loss. */
attenuation = MULT16_16_Q15(fade, decay); for (i=j=0;i<extrapolation_len;i++,j++)
{
opus_val16 tmp; if (j >= pitch_index) {
j -= pitch_index;
attenuation = MULT16_16_Q15(attenuation, decay);
}
buf[DECODE_BUFFER_SIZE-N+i] =
SHL32(EXTEND32(MULT16_16_Q15(attenuation,
exc[extrapolation_offset+j])), SIG_SHIFT); /* Compute the energy of the previously decoded signal whose
excitation we're copying. */
tmp = SROUND16(
buf[DECODE_BUFFER_SIZE-MAX_PERIOD-N+extrapolation_offset+j],
SIG_SHIFT);
S1 += SHR32(MULT16_16(tmp, tmp), 10);
}
{
opus_val16 lpc_mem[CELT_LPC_ORDER]; /* Copy the last decoded samples (prior to the overlap region) to
synthesis filter memory so we can have a continuous signal. */ for (i=0;i<CELT_LPC_ORDER;i++)
lpc_mem[i] = SROUND16(buf[DECODE_BUFFER_SIZE-N-1-i], SIG_SHIFT); /* Apply the synthesis filter to convert the excitation back into
the signal domain. */
celt_iir(buf+DECODE_BUFFER_SIZE-N, lpc+c*CELT_LPC_ORDER,
buf+DECODE_BUFFER_SIZE-N, extrapolation_len, CELT_LPC_ORDER,
lpc_mem, st->arch); #ifdef FIXED_POINT for (i=0; i < extrapolation_len; i++)
buf[DECODE_BUFFER_SIZE-N+i] = SATURATE(buf[DECODE_BUFFER_SIZE-N+i], SIG_SAT); #endif
}
/* Check if the synthesis energy is higher than expected, which can happen with the signal changes during our window. If so,
attenuate. */
{
opus_val32 S2=0; for (i=0;i<extrapolation_len;i++)
{
opus_val16 tmp = SROUND16(buf[DECODE_BUFFER_SIZE-N+i], SIG_SHIFT);
S2 += SHR32(MULT16_16(tmp, tmp), 10);
} /* This checks for an "explosion" in the synthesis. */ #ifdef FIXED_POINT if (!(S1 > SHR32(S2,2))) #else /* The float test is written this way to catch NaNs in the output
of the IIR filter at the same time. */ if (!(S1 > 0.2f*S2)) #endif
{ for (i=0;i<extrapolation_len;i++)
buf[DECODE_BUFFER_SIZE-N+i] = 0;
} elseif (S1 < S2)
{
opus_val16 ratio = celt_sqrt(frac_div32(SHR32(S1,1)+1,S2+1)); for (i=0;i<overlap;i++)
{
opus_val16 tmp_g = Q15ONE
- MULT16_16_Q15(window[i], Q15ONE-ratio);
buf[DECODE_BUFFER_SIZE-N+i] =
MULT16_32_Q15(tmp_g, buf[DECODE_BUFFER_SIZE-N+i]);
} for (i=overlap;i<extrapolation_len;i++)
{
buf[DECODE_BUFFER_SIZE-N+i] =
MULT16_32_Q15(ratio, buf[DECODE_BUFFER_SIZE-N+i]);
}
}
}
} while (++c<C);
#ifdef ENABLE_DEEP_PLC if (lpcnet->loaded && (st->complexity >= 5 || lpcnet->fec_fill_pos > 0)) { float overlap_mem; int samples_needed16k;
celt_sig *buf;
VARDECL(float, buf_copy);
buf = decode_mem[0];
ALLOC(buf_copy, C*overlap, float);
c=0; do {
OPUS_COPY(buf_copy+c*overlap, &decode_mem[c][DECODE_BUFFER_SIZE-N], overlap);
} while (++c<C);
/* Need enough samples from the PLC to cover the frame size, resampling delay,
and the overlap at the end. */
samples_needed16k = (N+SINC_ORDER+overlap)/3; if (loss_duration == 0) {
st->plc_fill = 0;
} while (st->plc_fill < samples_needed16k) {
lpcnet_plc_conceal(lpcnet, &st->plc_pcm[st->plc_fill]);
st->plc_fill += FRAME_SIZE;
} /* Resample to 48 kHz. */ for (i=0;i<(N+overlap)/3;i++) { int j; float sum; for (sum=0, j=0;j<17;j++) sum += 3*st->plc_pcm[i+j]*sinc_filter[3*j];
buf[DECODE_BUFFER_SIZE-N+3*i] = sum; for (sum=0, j=0;j<16;j++) sum += 3*st->plc_pcm[i+j+1]*sinc_filter[3*j+2];
buf[DECODE_BUFFER_SIZE-N+3*i+1] = sum; for (sum=0, j=0;j<16;j++) sum += 3*st->plc_pcm[i+j+1]*sinc_filter[3*j+1];
buf[DECODE_BUFFER_SIZE-N+3*i+2] = sum;
}
OPUS_MOVE(st->plc_pcm, &st->plc_pcm[N/3], st->plc_fill-N/3);
st->plc_fill -= N/3; for (i=0;i<N;i++) { float tmp = buf[DECODE_BUFFER_SIZE-N+i];
buf[DECODE_BUFFER_SIZE-N+i] -= PREEMPHASIS*st->plc_preemphasis_mem;
st->plc_preemphasis_mem = tmp;
}
overlap_mem = st->plc_preemphasis_mem; for (i=0;i<overlap;i++) { float tmp = buf[DECODE_BUFFER_SIZE+i];
buf[DECODE_BUFFER_SIZE+i] -= PREEMPHASIS*overlap_mem;
overlap_mem = tmp;
} /* For now, we just do mono PLC. */ if (C==2) OPUS_COPY(decode_mem[1], decode_mem[0], DECODE_BUFFER_SIZE+overlap);
c=0; do { /* Cross-fade with 48-kHz non-neural PLC for the first 2.5 ms to avoid a discontinuity. */ if (loss_duration == 0) { for (i=0;i<overlap;i++) decode_mem[c][DECODE_BUFFER_SIZE-N+i] = (1-window[i])*buf_copy[c*overlap+i] + (window[i])*decode_mem[c][DECODE_BUFFER_SIZE-N+i];
}
} while (++c<C);
} #endif
st->prefilter_and_fold = 1;
}
/* Saturate to soemthing large to avoid wrap-around. */
st->loss_duration = IMIN(10000, loss_duration+(1<<LM));
RESTORE_STACK;
}
int celt_decode_with_ec_dred(CELTDecoder * OPUS_RESTRICT st, constunsignedchar *data, int len, opus_val16 * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum #ifdef ENABLE_DEEP_PLC
,LPCNetPLCState *lpcnet #endif
)
{ int c, i, N; int spread_decision;
opus_int32 bits;
ec_dec _dec;
VARDECL(celt_norm, X);
VARDECL(int, fine_quant);
VARDECL(int, pulses);
VARDECL(int, cap);
VARDECL(int, offsets);
VARDECL(int, fine_priority);
VARDECL(int, tf_res);
VARDECL(unsignedchar, collapse_masks);
celt_sig *decode_mem[2];
celt_sig *out_syn[2];
opus_val16 *lpc;
opus_val16 *oldBandE, *oldLogE, *oldLogE2, *backgroundLogE;
int shortBlocks; int isTransient; int intra_ener; constint CC = st->channels; int LM, M; int start; int end; int effEnd; int codedBands; int alloc_trim; int postfilter_pitch;
opus_val16 postfilter_gain; int intensity=0; int dual_stereo=0;
opus_int32 total_bits;
opus_int32 balance;
opus_int32 tell; int dynalloc_logp; int postfilter_tapset; int anti_collapse_rsv; int anti_collapse_on=0; int silence; int C = st->stream_channels; const OpusCustomMode *mode; int nbEBands; int overlap; const opus_int16 *eBands;
opus_val16 max_background_increase;
ALLOC_STACK;
#ifdef CUSTOM_MODES if (st->signalling && data!=NULL)
{ int data0=data[0]; /* Convert "standard mode" to Opus header */ if (mode->Fs==48000 && mode->shortMdctSize==120)
{
data0 = fromOpus(data0); if (data0<0) return OPUS_INVALID_PACKET;
}
st->end = end = IMAX(1, mode->effEBands-2*(data0>>5));
LM = (data0>>3)&0x3;
C = 1 + ((data0>>2)&0x1);
data++;
len--; if (LM>mode->maxLM) return OPUS_INVALID_PACKET; if (frame_size < mode->shortMdctSize<<LM) return OPUS_BUFFER_TOO_SMALL; else
frame_size = mode->shortMdctSize<<LM;
} else { #else
{ #endif for (LM=0;LM<=mode->maxLM;LM++) if (mode->shortMdctSize<<LM==frame_size) break; if (LM>mode->maxLM) return OPUS_BAD_ARG;
}
M=1<<LM;
if (len<0 || len>1275 || pcm==NULL) return OPUS_BAD_ARG;
N = M*mode->shortMdctSize;
c=0; do {
decode_mem[c] = st->_decode_mem + c*(DECODE_BUFFER_SIZE+overlap);
out_syn[c] = decode_mem[c]+DECODE_BUFFER_SIZE-N;
} while (++c<CC);
effEnd = end; if (effEnd > mode->effEBands)
effEnd = mode->effEBands;
if (data == NULL || len<=1)
{
celt_decode_lost(st, N, LM #ifdef ENABLE_DEEP_PLC
, lpcnet #endif
);
deemphasis(out_syn, pcm, N, CC, st->downsample, mode->preemph, st->preemph_memD, accum);
RESTORE_STACK; return frame_size/st->downsample;
} #ifdef ENABLE_DEEP_PLC else { /* FIXME: This is a bit of a hack just to make sure opus_decode_native() knows we're no longer in PLC. */ if (lpcnet) lpcnet->blend = 0;
} #endif
/* Check if there are at least two packets received consecutively before
* turning on the pitch-based PLC */ if (st->loss_duration == 0) st->skip_plc = 0;
if (dec == NULL)
{
ec_dec_init(&_dec,(unsignedchar*)data,len);
dec = &_dec;
}
if (C==1)
{ for (i=0;i<nbEBands;i++)
oldBandE[i]=MAX16(oldBandE[i],oldBandE[nbEBands+i]);
}
total_bits = len*8;
tell = ec_tell(dec);
if (tell >= total_bits)
silence = 1; elseif (tell==1)
silence = ec_dec_bit_logp(dec, 15); else
silence = 0; if (silence)
{ /* Pretend we've read all the remaining bits */
tell = len*8;
dec->nbits_total+=tell-ec_tell(dec);
}
if (isTransient)
shortBlocks = M; else
shortBlocks = 0;
/* Decode the global flags (first symbols in the stream) */
intra_ener = tell+3<=total_bits ? ec_dec_bit_logp(dec, 3) : 0; /* If recovering from packet loss, make sure we make the energy prediction safe to reduce the
risk of getting loud artifacts. */ if (!intra_ener && st->loss_duration != 0) {
c=0; do
{
opus_val16 safety = 0; int missing = IMIN(10, st->loss_duration>>LM); if (LM==0) safety = QCONST16(1.5f,DB_SHIFT); elseif (LM==1) safety = QCONST16(.5f,DB_SHIFT); for (i=start;i<end;i++)
{ if (oldBandE[c*nbEBands+i] < MAX16(oldLogE[c*nbEBands+i], oldLogE2[c*nbEBands+i])) { /* If energy is going down already, continue the trend. */
opus_val32 slope;
opus_val32 E0, E1, E2;
E0 = oldBandE[c*nbEBands+i];
E1 = oldLogE[c*nbEBands+i];
E2 = oldLogE2[c*nbEBands+i];
slope = MAX32(E1 - E0, HALF32(E2 - E0));
E0 -= MAX32(0, (1+missing)*slope);
oldBandE[c*nbEBands+i] = MAX32(-QCONST16(20.f,DB_SHIFT), E0);
} else { /* Otherwise take the min of the last frames. */
oldBandE[c*nbEBands+i] = MIN16(MIN16(oldBandE[c*nbEBands+i], oldLogE[c*nbEBands+i]), oldLogE2[c*nbEBands+i]);
} /* Shorter frames have more natural fluctuations -- play it safe. */
oldBandE[c*nbEBands+i] -= safety;
}
} while (++c<2);
} /* Get band energies */
unquant_coarse_energy(mode, start, end, oldBandE,
intra_ener, dec, C, LM);
dynalloc_logp = 6;
total_bits<<=BITRES;
tell = ec_tell_frac(dec); for (i=start;i<end;i++)
{ int width, quanta; int dynalloc_loop_logp; int boost;
width = C*(eBands[i+1]-eBands[i])<<LM; /* quanta is 6 bits, but no more than 1 bit/sample
and no less than 1/8 bit/sample */
quanta = IMIN(width<<BITRES, IMAX(6<<BITRES, width));
dynalloc_loop_logp = dynalloc_logp;
boost = 0; while (tell+(dynalloc_loop_logp<<BITRES) < total_bits && boost < cap[i])
{ int flag;
flag = ec_dec_bit_logp(dec, dynalloc_loop_logp);
tell = ec_tell_frac(dec); if (!flag) break;
boost += quanta;
total_bits -= quanta;
dynalloc_loop_logp = 1;
}
offsets[i] = boost; /* Making dynalloc more likely */ if (boost>0)
dynalloc_logp = IMAX(2, dynalloc_logp-1);
}
if (C==1)
OPUS_COPY(&oldBandE[nbEBands], oldBandE, nbEBands);
if (!isTransient)
{
OPUS_COPY(oldLogE2, oldLogE, 2*nbEBands);
OPUS_COPY(oldLogE, oldBandE, 2*nbEBands);
} else { for (i=0;i<2*nbEBands;i++)
oldLogE[i] = MIN16(oldLogE[i], oldBandE[i]);
} /* In normal circumstances, we only allow the noise floor to increase by up to 2.4 dB/second, but when we're in DTX we give the weight of
all missing packets to the update packet. */
max_background_increase = IMIN(160, st->loss_duration+M)*QCONST16(0.001f,DB_SHIFT); for (i=0;i<2*nbEBands;i++)
backgroundLogE[i] = MIN16(backgroundLogE[i] + max_background_increase, oldBandE[i]); /* In case start or end were to change */
c=0; do
{ for (i=0;i<start;i++)
{
oldBandE[c*nbEBands+i]=0;
oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT);
} for (i=end;i<nbEBands;i++)
{
oldBandE[c*nbEBands+i]=0;
oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT);
}
} while (++c<2);
st->rng = dec->rng;
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.