void tonality_get_info(TonalityAnalysisState *tonal, AnalysisInfo *info_out, int len)
{
int pos;
int curr_lookahead;
float psum;
int i;
pos = tonal->read_pos;
curr_lookahead = tonal->write_pos-tonal->read_pos;
if (curr_lookahead<0)
curr_lookahead += DETECT_SIZE;
/* On long frames, look at the second analysis window rather than the first. */
if (len > tonal->Fs/50 && pos != tonal->write_pos)
{
pos++;
if (pos==DETECT_SIZE)
pos=0;
}
if (pos == tonal->write_pos)
pos--;
if (pos<0)
pos = DETECT_SIZE-1;
OPUS_COPY(info_out, &tonal->info[pos], 1);
/* If possible, look ahead for a tone to compensate for the delay in the tone detector. */
for (i=0;i<3;i++)
{
pos++;
if (pos==DETECT_SIZE)
pos = 0;
if (pos == tonal->write_pos)
break;
info_out->tonality = MAX32(0, -.03f + MAX32(info_out->tonality, tonal->info[pos].tonality-.05f));
}
tonal->read_subframe += len/(tonal->Fs/400);
while (tonal->read_subframe>=8)
{
tonal->read_subframe -= 8;
tonal->read_pos++;
}
if (tonal->read_pos>=DETECT_SIZE)
tonal->read_pos-=DETECT_SIZE;
/* The -1 is to compensate for the delay in the features themselves. */
curr_lookahead = IMAX(curr_lookahead-1, 0);
psum=0;
/* Summing the probability of transition patterns that involve music at
time (DETECT_SIZE-curr_lookahead-1) */
for (i=0;i<DETECT_SIZE-curr_lookahead;i++)
psum += tonal->pmusic[i];
for (;i<DETECT_SIZE;i++)
psum += tonal->pspeech[i];
psum = psum*tonal->music_confidence + (1-psum)*tonal->speech_confidence;
/*printf("%f %f %f %f %f\n", psum, info_out->music_prob, info_out->vad_prob, info_out->activity_probability, info_out->tonality);*/
info_out->music_prob = psum;
}
static opus_val32 downmix_and_resample(downmix_func downmix, const void *_x, opus_val32 *y, opus_val32 S[3], int subframe, int offset, int c1, int c2, int C, int Fs)
{
VARDECL(opus_val32, tmp);
opus_val32 scale;
int j;
opus_val32 ret = 0;
SAVE_STACK;
if (subframe==0) return 0;
if (Fs == 48000)
{
subframe *= 2;
offset *= 2;
} else if (Fs == 16000) {
subframe = subframe*2/3;
offset = offset*2/3;
}
ALLOC(tmp, subframe, opus_val32);
downmix(_x, tmp, subframe, offset, c1, c2, C);
#ifdef FIXED_POINT
scale = (1<<SIG_SHIFT);
#else
scale = 1.f/32768;
#endif
if (c2==-2)
scale /= C;
else if (c2>-1)
scale /= 2;
for (j=0;j<subframe;j++)
tmp[j] *= scale;
if (Fs == 48000)
{
ret = silk_resampler_down2_hp(S, y, tmp, subframe);
} else if (Fs == 24000) {
OPUS_COPY(y, tmp, subframe);
} else if (Fs == 16000) {
VARDECL(opus_val32, tmp3x);
ALLOC(tmp3x, 3*subframe, opus_val32);
/* Don't do this at home! This resampler is horrible and it's only (barely)
usable for the purpose of the analysis because we don't care about all
the aliasing between 8 kHz and 12 kHz. */
for (j=0;j<subframe;j++)
{
tmp3x[3*j] = tmp[j];
tmp3x[3*j+1] = tmp[j];
tmp3x[3*j+2] = tmp[j];
}
silk_resampler_down2_hp(S, y, tmp3x, 3*subframe);
}
RESTORE_STACK;
return ret;
}
void tonality_get_info(TonalityAnalysisState *tonal, AnalysisInfo *info_out, int len)
{
int pos;
int curr_lookahead;
float psum;
int i;
pos = tonal->read_pos;
curr_lookahead = tonal->write_pos-tonal->read_pos;
if (curr_lookahead<0)
curr_lookahead += DETECT_SIZE;
if (len > 480 && pos != tonal->write_pos)
{
pos++;
if (pos==DETECT_SIZE)
pos=0;
}
if (pos == tonal->write_pos)
pos--;
if (pos<0)
pos = DETECT_SIZE-1;
OPUS_COPY(info_out, &tonal->info[pos], 1);
tonal->read_subframe += len/120;
while (tonal->read_subframe>=4)
{
tonal->read_subframe -= 4;
tonal->read_pos++;
}
if (tonal->read_pos>=DETECT_SIZE)
tonal->read_pos-=DETECT_SIZE;
/* Compensate for the delay in the features themselves.
FIXME: Need a better estimate the 10 I just made up */
curr_lookahead = IMAX(curr_lookahead-10, 0);
psum=0;
/* Summing the probability of transition patterns that involve music at
time (DETECT_SIZE-curr_lookahead-1) */
for (i=0;i<DETECT_SIZE-curr_lookahead;i++)
psum += tonal->pmusic[i];
for (;i<DETECT_SIZE;i++)
psum += tonal->pspeech[i];
psum = psum*tonal->music_confidence + (1-psum)*tonal->speech_confidence;
/*printf("%f %f %f\n", psum, info_out->music_prob, info_out->tonality);*/
info_out->music_prob = psum;
}
void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info_out, const CELTMode *celt_mode, const void *x, int len, int offset, int c1, int c2, int C, int lsb_depth, downmix_func downmix)
{
int i, b;
const kiss_fft_state *kfft;
VARDECL(kiss_fft_cpx, in);
VARDECL(kiss_fft_cpx, out);
int N = 480, N2=240;
float * OPUS_RESTRICT A = tonal->angle;
float * OPUS_RESTRICT dA = tonal->d_angle;
float * OPUS_RESTRICT d2A = tonal->d2_angle;
VARDECL(float, tonality);
VARDECL(float, noisiness);
float band_tonality[NB_TBANDS];
float logE[NB_TBANDS];
float BFCC[8];
float features[25];
float frame_tonality;
float max_frame_tonality;
/*float tw_sum=0;*/
float frame_noisiness;
const float pi4 = (float)(M_PI*M_PI*M_PI*M_PI);
float slope=0;
float frame_stationarity;
float relativeE;
float frame_probs[2];
float alpha, alphaE, alphaE2;
float frame_loudness;
float bandwidth_mask;
int bandwidth=0;
float maxE = 0;
float noise_floor;
int remaining;
AnalysisInfo *info;
SAVE_STACK;
tonal->last_transition++;
alpha = 1.f/IMIN(20, 1+tonal->count);
alphaE = 1.f/IMIN(50, 1+tonal->count);
alphaE2 = 1.f/IMIN(1000, 1+tonal->count);
if (tonal->count<4)
tonal->music_prob = .5;
kfft = celt_mode->mdct.kfft[0];
if (tonal->count==0)
tonal->mem_fill = 240;
downmix(x, &tonal->inmem[tonal->mem_fill], IMIN(len, ANALYSIS_BUF_SIZE-tonal->mem_fill), offset, c1, c2, C);
if (tonal->mem_fill+len < ANALYSIS_BUF_SIZE)
{
tonal->mem_fill += len;
/* Don't have enough to update the analysis */
RESTORE_STACK;
return;
}
info = &tonal->info[tonal->write_pos++];
if (tonal->write_pos>=DETECT_SIZE)
tonal->write_pos-=DETECT_SIZE;
ALLOC(in, 480, kiss_fft_cpx);
ALLOC(out, 480, kiss_fft_cpx);
ALLOC(tonality, 240, float);
ALLOC(noisiness, 240, float);
for (i=0;i<N2;i++)
{
float w = analysis_window[i];
in[i].r = (kiss_fft_scalar)(w*tonal->inmem[i]);
in[i].i = (kiss_fft_scalar)(w*tonal->inmem[N2+i]);
in[N-i-1].r = (kiss_fft_scalar)(w*tonal->inmem[N-i-1]);
in[N-i-1].i = (kiss_fft_scalar)(w*tonal->inmem[N+N2-i-1]);
}
OPUS_MOVE(tonal->inmem, tonal->inmem+ANALYSIS_BUF_SIZE-240, 240);
remaining = len - (ANALYSIS_BUF_SIZE-tonal->mem_fill);
downmix(x, &tonal->inmem[240], remaining, offset+ANALYSIS_BUF_SIZE-tonal->mem_fill, c1, c2, C);
tonal->mem_fill = 240 + remaining;
opus_fft(kfft, in, out);
for (i=1;i<N2;i++)
{
float X1r, X2r, X1i, X2i;
float angle, d_angle, d2_angle;
float angle2, d_angle2, d2_angle2;
float mod1, mod2, avg_mod;
X1r = (float)out[i].r+out[N-i].r;
X1i = (float)out[i].i-out[N-i].i;
X2r = (float)out[i].i+out[N-i].i;
X2i = (float)out[N-i].r-out[i].r;
angle = (float)(.5f/M_PI)*fast_atan2f(X1i, X1r);
d_angle = angle - A[i];
d2_angle = d_angle - dA[i];
angle2 = (float)(.5f/M_PI)*fast_atan2f(X2i, X2r);
d_angle2 = angle2 - angle;
d2_angle2 = d_angle2 - d_angle;
mod1 = d2_angle - (float)floor(.5+d2_angle);
noisiness[i] = ABS16(mod1);
mod1 *= mod1;
mod1 *= mod1;
mod2 = d2_angle2 - (float)floor(.5+d2_angle2);
noisiness[i] += ABS16(mod2);
mod2 *= mod2;
mod2 *= mod2;
avg_mod = .25f*(d2A[i]+2.f*mod1+mod2);
tonality[i] = 1.f/(1.f+40.f*16.f*pi4*avg_mod)-.015f;
A[i] = angle2;
dA[i] = d_angle2;
d2A[i] = mod2;
}
frame_tonality = 0;
max_frame_tonality = 0;
/*tw_sum = 0;*/
info->activity = 0;
frame_noisiness = 0;
frame_stationarity = 0;
if (!tonal->count)
{
for (b=0;b<NB_TBANDS;b++)
{
tonal->lowE[b] = 1e10;
tonal->highE[b] = -1e10;
}
}
relativeE = 0;
frame_loudness = 0;
for (b=0;b<NB_TBANDS;b++)
{
float E=0, tE=0, nE=0;
float L1, L2;
float stationarity;
for (i=tbands[b];i<tbands[b+1];i++)
{
float binE = out[i].r*(float)out[i].r + out[N-i].r*(float)out[N-i].r
+ out[i].i*(float)out[i].i + out[N-i].i*(float)out[N-i].i;
#ifdef FIXED_POINT
/* FIXME: It's probably best to change the BFCC filter initial state instead */
binE *= 5.55e-17f;
#endif
E += binE;
tE += binE*tonality[i];
nE += binE*2.f*(.5f-noisiness[i]);
}
tonal->E[tonal->E_count][b] = E;
frame_noisiness += nE/(1e-15f+E);
frame_loudness += (float)sqrt(E+1e-10f);
logE[b] = (float)log(E+1e-10f);
tonal->lowE[b] = MIN32(logE[b], tonal->lowE[b]+.01f);
tonal->highE[b] = MAX32(logE[b], tonal->highE[b]-.1f);
if (tonal->highE[b] < tonal->lowE[b]+1.f)
{
tonal->highE[b]+=.5f;
tonal->lowE[b]-=.5f;
}
relativeE += (logE[b]-tonal->lowE[b])/(1e-15f+tonal->highE[b]-tonal->lowE[b]);
L1=L2=0;
for (i=0;i<NB_FRAMES;i++)
{
L1 += (float)sqrt(tonal->E[i][b]);
L2 += tonal->E[i][b];
}
stationarity = MIN16(0.99f,L1/(float)sqrt(1e-15+NB_FRAMES*L2));
stationarity *= stationarity;
stationarity *= stationarity;
frame_stationarity += stationarity;
/*band_tonality[b] = tE/(1e-15+E)*/;
band_tonality[b] = MAX16(tE/(1e-15f+E), stationarity*tonal->prev_band_tonality[b]);
#if 0
if (b>=NB_TONAL_SKIP_BANDS)
{
frame_tonality += tweight[b]*band_tonality[b];
tw_sum += tweight[b];
}
#else
frame_tonality += band_tonality[b];
if (b>=NB_TBANDS-NB_TONAL_SKIP_BANDS)
frame_tonality -= band_tonality[b-NB_TBANDS+NB_TONAL_SKIP_BANDS];
#endif
max_frame_tonality = MAX16(max_frame_tonality, (1.f+.03f*(b-NB_TBANDS))*frame_tonality);
slope += band_tonality[b]*(b-8);
/*printf("%f %f ", band_tonality[b], stationarity);*/
tonal->prev_band_tonality[b] = band_tonality[b];
}
bandwidth_mask = 0;
bandwidth = 0;
maxE = 0;
noise_floor = 5.7e-4f/(1<<(IMAX(0,lsb_depth-8)));
#ifdef FIXED_POINT
noise_floor *= 1<<(15+SIG_SHIFT);
#endif
noise_floor *= noise_floor;
for (b=0;b<NB_TOT_BANDS;b++)
{
float E=0;
int band_start, band_end;
/* Keep a margin of 300 Hz for aliasing */
band_start = extra_bands[b];
band_end = extra_bands[b+1];
for (i=band_start;i<band_end;i++)
{
float binE = out[i].r*(float)out[i].r + out[N-i].r*(float)out[N-i].r
+ out[i].i*(float)out[i].i + out[N-i].i*(float)out[N-i].i;
E += binE;
}
maxE = MAX32(maxE, E);
tonal->meanE[b] = MAX32((1-alphaE2)*tonal->meanE[b], E);
E = MAX32(E, tonal->meanE[b]);
/* Use a simple follower with 13 dB/Bark slope for spreading function */
bandwidth_mask = MAX32(.05f*bandwidth_mask, E);
/* Consider the band "active" only if all these conditions are met:
1) less than 10 dB below the simple follower
2) less than 90 dB below the peak band (maximal masking possible considering
both the ATH and the loudness-dependent slope of the spreading function)
3) above the PCM quantization noise floor
*/
if (E>.1*bandwidth_mask && E*1e9f > maxE && E > noise_floor*(band_end-band_start))
bandwidth = b;
}
if (tonal->count<=2)
bandwidth = 20;
frame_loudness = 20*(float)log10(frame_loudness);
tonal->Etracker = MAX32(tonal->Etracker-.03f, frame_loudness);
tonal->lowECount *= (1-alphaE);
if (frame_loudness < tonal->Etracker-30)
tonal->lowECount += alphaE;
for (i=0;i<8;i++)
{
float sum=0;
for (b=0;b<16;b++)
sum += dct_table[i*16+b]*logE[b];
BFCC[i] = sum;
}
frame_stationarity /= NB_TBANDS;
relativeE /= NB_TBANDS;
if (tonal->count<10)
relativeE = .5;
frame_noisiness /= NB_TBANDS;
#if 1
info->activity = frame_noisiness + (1-frame_noisiness)*relativeE;
#else
info->activity = .5*(1+frame_noisiness-frame_stationarity);
#endif
frame_tonality = (max_frame_tonality/(NB_TBANDS-NB_TONAL_SKIP_BANDS));
frame_tonality = MAX16(frame_tonality, tonal->prev_tonality*.8f);
tonal->prev_tonality = frame_tonality;
slope /= 8*8;
info->tonality_slope = slope;
tonal->E_count = (tonal->E_count+1)%NB_FRAMES;
tonal->count++;
info->tonality = frame_tonality;
for (i=0;i<4;i++)
features[i] = -0.12299f*(BFCC[i]+tonal->mem[i+24]) + 0.49195f*(tonal->mem[i]+tonal->mem[i+16]) + 0.69693f*tonal->mem[i+8] - 1.4349f*tonal->cmean[i];
for (i=0;i<4;i++)
tonal->cmean[i] = (1-alpha)*tonal->cmean[i] + alpha*BFCC[i];
for (i=0;i<4;i++)
features[4+i] = 0.63246f*(BFCC[i]-tonal->mem[i+24]) + 0.31623f*(tonal->mem[i]-tonal->mem[i+16]);
for (i=0;i<3;i++)
features[8+i] = 0.53452f*(BFCC[i]+tonal->mem[i+24]) - 0.26726f*(tonal->mem[i]+tonal->mem[i+16]) -0.53452f*tonal->mem[i+8];
if (tonal->count > 5)
{
for (i=0;i<9;i++)
tonal->std[i] = (1-alpha)*tonal->std[i] + alpha*features[i]*features[i];
}
for (i=0;i<8;i++)
{
tonal->mem[i+24] = tonal->mem[i+16];
tonal->mem[i+16] = tonal->mem[i+8];
tonal->mem[i+8] = tonal->mem[i];
tonal->mem[i] = BFCC[i];
}
for (i=0;i<9;i++)
features[11+i] = (float)sqrt(tonal->std[i]);
features[20] = info->tonality;
features[21] = info->activity;
features[22] = frame_stationarity;
features[23] = info->tonality_slope;
features[24] = tonal->lowECount;
#ifndef DISABLE_FLOAT_API
mlp_process(&net, features, frame_probs);
frame_probs[0] = .5f*(frame_probs[0]+1);
/* Curve fitting between the MLP probability and the actual probability */
frame_probs[0] = .01f + 1.21f*frame_probs[0]*frame_probs[0] - .23f*(float)pow(frame_probs[0], 10);
/* Probability of active audio (as opposed to silence) */
frame_probs[1] = .5f*frame_probs[1]+.5f;
/* Consider that silence has a 50-50 probability. */
frame_probs[0] = frame_probs[1]*frame_probs[0] + (1-frame_probs[1])*.5f;
/*printf("%f %f ", frame_probs[0], frame_probs[1]);*/
{
/* Probability of state transition */
float tau;
/* Represents independence of the MLP probabilities, where
beta=1 means fully independent. */
float beta;
/* Denormalized probability of speech (p0) and music (p1) after update */
float p0, p1;
/* Probabilities for "all speech" and "all music" */
float s0, m0;
/* Probability sum for renormalisation */
float psum;
/* Instantaneous probability of speech and music, with beta pre-applied. */
float speech0;
float music0;
/* One transition every 3 minutes of active audio */
tau = .00005f*frame_probs[1];
beta = .05f;
if (1) {
/* Adapt beta based on how "unexpected" the new prob is */
float p, q;
p = MAX16(.05f,MIN16(.95f,frame_probs[0]));
q = MAX16(.05f,MIN16(.95f,tonal->music_prob));
beta = .01f+.05f*ABS16(p-q)/(p*(1-q)+q*(1-p));
}
/* p0 and p1 are the probabilities of speech and music at this frame
using only information from previous frame and applying the
state transition model */
p0 = (1-tonal->music_prob)*(1-tau) + tonal->music_prob *tau;
p1 = tonal->music_prob *(1-tau) + (1-tonal->music_prob)*tau;
/* We apply the current probability with exponent beta to work around
the fact that the probability estimates aren't independent. */
p0 *= (float)pow(1-frame_probs[0], beta);
p1 *= (float)pow(frame_probs[0], beta);
/* Normalise the probabilities to get the Marokv probability of music. */
tonal->music_prob = p1/(p0+p1);
info->music_prob = tonal->music_prob;
/* This chunk of code deals with delayed decision. */
psum=1e-20f;
/* Instantaneous probability of speech and music, with beta pre-applied. */
speech0 = (float)pow(1-frame_probs[0], beta);
music0 = (float)pow(frame_probs[0], beta);
if (tonal->count==1)
{
tonal->pspeech[0]=.5;
tonal->pmusic [0]=.5;
}
/* Updated probability of having only speech (s0) or only music (m0),
before considering the new observation. */
s0 = tonal->pspeech[0] + tonal->pspeech[1];
m0 = tonal->pmusic [0] + tonal->pmusic [1];
/* Updates s0 and m0 with instantaneous probability. */
tonal->pspeech[0] = s0*(1-tau)*speech0;
tonal->pmusic [0] = m0*(1-tau)*music0;
/* Propagate the transition probabilities */
for (i=1;i<DETECT_SIZE-1;i++)
{
tonal->pspeech[i] = tonal->pspeech[i+1]*speech0;
tonal->pmusic [i] = tonal->pmusic [i+1]*music0;
}
/* Probability that the latest frame is speech, when all the previous ones were music. */
tonal->pspeech[DETECT_SIZE-1] = m0*tau*speech0;
/* Probability that the latest frame is music, when all the previous ones were speech. */
tonal->pmusic [DETECT_SIZE-1] = s0*tau*music0;
/* Renormalise probabilities to 1 */
for (i=0;i<DETECT_SIZE;i++)
psum += tonal->pspeech[i] + tonal->pmusic[i];
psum = 1.f/psum;
for (i=0;i<DETECT_SIZE;i++)
{
tonal->pspeech[i] *= psum;
tonal->pmusic [i] *= psum;
}
psum = tonal->pmusic[0];
for (i=1;i<DETECT_SIZE;i++)
psum += tonal->pspeech[i];
/* Estimate our confidence in the speech/music decisions */
if (frame_probs[1]>.75)
{
if (tonal->music_prob>.9)
{
float adapt;
adapt = 1.f/(++tonal->music_confidence_count);
tonal->music_confidence_count = IMIN(tonal->music_confidence_count, 500);
tonal->music_confidence += adapt*MAX16(-.2f,frame_probs[0]-tonal->music_confidence);
}
if (tonal->music_prob<.1)
{
float adapt;
adapt = 1.f/(++tonal->speech_confidence_count);
tonal->speech_confidence_count = IMIN(tonal->speech_confidence_count, 500);
tonal->speech_confidence += adapt*MIN16(.2f,frame_probs[0]-tonal->speech_confidence);
}
} else {
if (tonal->music_confidence_count==0)
tonal->music_confidence = .9f;
if (tonal->speech_confidence_count==0)
tonal->speech_confidence = .1f;
}
}
if (tonal->last_music != (tonal->music_prob>.5f))
tonal->last_transition=0;
tonal->last_music = tonal->music_prob>.5f;
#else
info->music_prob = 0;
#endif
/*for (i=0;i<25;i++)
printf("%f ", features[i]);
printf("\n");*/
info->bandwidth = bandwidth;
/*printf("%d %d\n", info->bandwidth, info->opus_bandwidth);*/
info->noisiness = frame_noisiness;
info->valid = 1;
if (info_out!=NULL)
OPUS_COPY(info_out, info, 1);
RESTORE_STACK;
}
void quant_coarse_energy(const CELTMode *m, int start, int end, int effEnd,
const opus_val16 *eBands, opus_val16 *oldEBands, opus_uint32 budget,
opus_val16 *error, ec_enc *enc, int C, int LM, int nbAvailableBytes,
int force_intra, opus_val32 *delayedIntra, int two_pass, int loss_rate)
{
int intra;
opus_val16 max_decay;
VARDECL(opus_val16, oldEBands_intra);
VARDECL(opus_val16, error_intra);
ec_enc enc_start_state;
opus_uint32 tell;
int badness1=0;
opus_int32 intra_bias;
opus_val32 new_distortion;
SAVE_STACK;
intra = force_intra || (!two_pass && *delayedIntra>2*C*(end-start) && nbAvailableBytes > (end-start)*C);
intra_bias = (opus_int32)((budget**delayedIntra*loss_rate)/(C*512));
new_distortion = loss_distortion(eBands, oldEBands, start, effEnd, m->nbEBands, C);
tell = ec_tell(enc);
if (tell+3 > budget)
two_pass = intra = 0;
/* Encode the global flags using a simple probability model
(first symbols in the stream) */
max_decay = QCONST16(16.f,DB_SHIFT);
if (end-start>10)
{
#ifdef FIXED_POINT
max_decay = MIN32(max_decay, SHL32(EXTEND32(nbAvailableBytes),DB_SHIFT-3));
#else
max_decay = MIN32(max_decay, .125f*nbAvailableBytes);
#endif
}
enc_start_state = *enc;
ALLOC(oldEBands_intra, C*m->nbEBands, opus_val16);
ALLOC(error_intra, C*m->nbEBands, opus_val16);
OPUS_COPY(oldEBands_intra, oldEBands, C*m->nbEBands);
if (two_pass || intra)
{
badness1 = quant_coarse_energy_impl(m, start, end, eBands, oldEBands_intra, budget,
tell, e_prob_model[LM][1], error_intra, enc, C, LM, 1, max_decay);
}
if (!intra)
{
unsigned char *intra_buf;
ec_enc enc_intra_state;
opus_int32 tell_intra;
opus_uint32 nstart_bytes;
opus_uint32 nintra_bytes;
int badness2;
VARDECL(unsigned char, intra_bits);
tell_intra = ec_tell_frac(enc);
enc_intra_state = *enc;
nstart_bytes = ec_range_bytes(&enc_start_state);
nintra_bytes = ec_range_bytes(&enc_intra_state);
intra_buf = ec_get_buffer(&enc_intra_state) + nstart_bytes;
ALLOC(intra_bits, nintra_bytes-nstart_bytes, unsigned char);
/* Copy bits from intra bit-stream */
OPUS_COPY(intra_bits, intra_buf, nintra_bytes - nstart_bytes);
*enc = enc_start_state;
badness2 = quant_coarse_energy_impl(m, start, end, eBands, oldEBands, budget,
tell, e_prob_model[LM][intra], error, enc, C, LM, 0, max_decay);
if (two_pass && (badness1 < badness2 || (badness1 == badness2 && ((opus_int32)ec_tell_frac(enc))+intra_bias > tell_intra)))
{
*enc = enc_intra_state;
/* Copy intra bits to bit-stream */
OPUS_COPY(intra_buf, intra_bits, nintra_bytes - nstart_bytes);
OPUS_COPY(oldEBands, oldEBands_intra, C*m->nbEBands);
OPUS_COPY(error, error_intra, C*m->nbEBands);
intra = 1;
}
} else {
void surround_analysis(const CELTMode *celt_mode, const void *pcm, opus_val16 *bandLogE, opus_val32 *mem, opus_val32 *preemph_mem,
int len, int overlap, int channels, int rate, opus_copy_channel_in_func copy_channel_in, int arch
)
{
int c;
int i;
int LM;
int pos[8] = {0};
int upsample;
int frame_size;
int freq_size;
opus_val16 channel_offset;
opus_val32 bandE[21];
opus_val16 maskLogE[3][21];
VARDECL(opus_val32, in);
VARDECL(opus_val16, x);
VARDECL(opus_val32, freq);
SAVE_STACK;
upsample = resampling_factor(rate);
frame_size = len*upsample;
freq_size = IMIN(960, frame_size);
/* LM = log2(frame_size / 120) */
for (LM=0;LM<celt_mode->maxLM;LM++)
if (celt_mode->shortMdctSize<<LM==frame_size)
break;
ALLOC(in, frame_size+overlap, opus_val32);
ALLOC(x, len, opus_val16);
ALLOC(freq, freq_size, opus_val32);
channel_pos(channels, pos);
for (c=0;c<3;c++)
for (i=0;i<21;i++)
maskLogE[c][i] = -QCONST16(28.f, DB_SHIFT);
for (c=0;c<channels;c++)
{
int frame;
int nb_frames = frame_size/freq_size;
celt_assert(nb_frames*freq_size == frame_size);
OPUS_COPY(in, mem+c*overlap, overlap);
(*copy_channel_in)(x, 1, pcm, channels, c, len);
celt_preemphasis(x, in+overlap, frame_size, 1, upsample, celt_mode->preemph, preemph_mem+c, 0);
#ifndef FIXED_POINT
{
opus_val32 sum;
sum = celt_inner_prod(in, in, frame_size+overlap, 0);
/* This should filter out both NaNs and ridiculous signals that could
cause NaNs further down. */
if (!(sum < 1e9f) || celt_isnan(sum))
{
OPUS_CLEAR(in, frame_size+overlap);
preemph_mem[c] = 0;
}
}
#endif
OPUS_CLEAR(bandE, 21);
for (frame=0;frame<nb_frames;frame++)
{
opus_val32 tmpE[21];
clt_mdct_forward(&celt_mode->mdct, in+960*frame, freq, celt_mode->window,
overlap, celt_mode->maxLM-LM, 1, arch);
if (upsample != 1)
{
int bound = freq_size/upsample;
for (i=0;i<bound;i++)
freq[i] *= upsample;
for (;i<freq_size;i++)
freq[i] = 0;
}
compute_band_energies(celt_mode, freq, tmpE, 21, 1, LM);
/* If we have multiple frames, take the max energy. */
for (i=0;i<21;i++)
bandE[i] = MAX32(bandE[i], tmpE[i]);
}
amp2Log2(celt_mode, 21, 21, bandE, bandLogE+21*c, 1);
/* Apply spreading function with -6 dB/band going up and -12 dB/band going down. */
for (i=1;i<21;i++)
bandLogE[21*c+i] = MAX16(bandLogE[21*c+i], bandLogE[21*c+i-1]-QCONST16(1.f, DB_SHIFT));
for (i=19;i>=0;i--)
bandLogE[21*c+i] = MAX16(bandLogE[21*c+i], bandLogE[21*c+i+1]-QCONST16(2.f, DB_SHIFT));
if (pos[c]==1)
{
for (i=0;i<21;i++)
maskLogE[0][i] = logSum(maskLogE[0][i], bandLogE[21*c+i]);
} else if (pos[c]==3)
{
for (i=0;i<21;i++)
maskLogE[2][i] = logSum(maskLogE[2][i], bandLogE[21*c+i]);
} else if (pos[c]==2)
{
for (i=0;i<21;i++)
{
maskLogE[0][i] = logSum(maskLogE[0][i], bandLogE[21*c+i]-QCONST16(.5f, DB_SHIFT));
maskLogE[2][i] = logSum(maskLogE[2][i], bandLogE[21*c+i]-QCONST16(.5f, DB_SHIFT));
}
}
#if 0
for (i=0;i<21;i++)
printf("%f ", bandLogE[21*c+i]);
float sum=0;
for (i=0;i<21;i++)
sum += bandLogE[21*c+i];
printf("%f ", sum/21);
#endif
OPUS_COPY(mem+c*overlap, in+frame_size, overlap);
}
for (i=0;i<21;i++)
maskLogE[1][i] = MIN32(maskLogE[0][i],maskLogE[2][i]);
channel_offset = HALF16(celt_log2(QCONST32(2.f,14)/(channels-1)));
for (c=0;c<3;c++)
for (i=0;i<21;i++)
maskLogE[c][i] += channel_offset;
#if 0
for (c=0;c<3;c++)
{
for (i=0;i<21;i++)
printf("%f ", maskLogE[c][i]);
}
#endif
for (c=0;c<channels;c++)
{
opus_val16 *mask;
if (pos[c]!=0)
{
mask = &maskLogE[pos[c]-1][0];
for (i=0;i<21;i++)
bandLogE[21*c+i] = bandLogE[21*c+i] - mask[i];
} else {
for (i=0;i<21;i++)
bandLogE[21*c+i] = 0;
}
#if 0
for (i=0;i<21;i++)
printf("%f ", bandLogE[21*c+i]);
printf("\n");
#endif
#if 0
float sum=0;
for (i=0;i<21;i++)
sum += bandLogE[21*c+i];
printf("%f ", sum/(float)QCONST32(21.f, DB_SHIFT));
printf("\n");
#endif
}
RESTORE_STACK;
}
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);
} else if (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] = HALF32(ADD32(freq[i],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);
}
RESTORE_STACK;
}