#include "modes.h"
#include "vq.h"
#include "cwrs.h"
+#include "stack_alloc.h"
#include "os_support.h"
#include "mathops.h"
/* Compute the amplitude (sqrt energy) in each of the bands */
void compute_band_energies(const CELTMode *m, const celt_sig_t *X, celt_ener_t *bank)
{
- int i, c, B, C;
+ int i, c;
const celt_int16_t *eBands = m->eBands;
- B = m->nbMdctBlocks;
- C = m->nbChannels;
+ const int C = CHANNELS(m);
for (c=0;c<C;c++)
{
for (i=0;i<m->nbEBands;i++)
int j;
celt_word32_t maxval=0;
celt_word32_t sum = 0;
- for (j=B*eBands[i];j<B*eBands[i+1];j++)
+ for (j=eBands[i];j<eBands[i+1];j++)
maxval = MAX32(maxval, ABS32(X[j*C+c]));
if (maxval > 0)
{
int shift = celt_ilog2(maxval)-10;
- for (j=B*eBands[i];j<B*eBands[i+1];j++)
- sum += VSHR32(X[j*C+c],shift)*VSHR32(X[j*C+c],shift);
+ for (j=eBands[i];j<eBands[i+1];j++)
+ sum += MULT16_16(EXTRACT16(VSHR32(X[j*C+c],shift)),EXTRACT16(VSHR32(X[j*C+c],shift)));
/* We're adding one here to make damn sure we never end up with a pitch vector that's
larger than unity norm */
- bank[i*C+c] = 1+VSHR32(EXTEND32(celt_sqrt(sum)),-shift);
+ bank[i*C+c] = EPSILON+VSHR32(EXTEND32(celt_sqrt(sum)),-shift);
} else {
- bank[i*C+c] = 0;
+ bank[i*C+c] = EPSILON;
}
/*printf ("%f ", bank[i*C+c]);*/
}
}
/* Normalise each band such that the energy is one. */
-void normalise_bands(const CELTMode *m, const celt_sig_t *freq, celt_norm_t *X, const celt_ener_t *bank)
+void normalise_bands(const CELTMode *m, const celt_sig_t * restrict freq, celt_norm_t * restrict X, const celt_ener_t *bank)
{
- int i, c, B, C;
+ int i, c;
const celt_int16_t *eBands = m->eBands;
- B = m->nbMdctBlocks;
- C = m->nbChannels;
+ const int C = CHANNELS(m);
for (c=0;c<C;c++)
{
for (i=0;i<m->nbEBands;i++)
celt_word16_t g;
int j,shift;
celt_word16_t E;
- shift = celt_ilog2(bank[i*C+c])-13;
+ shift = celt_zlog2(bank[i*C+c])-13;
E = VSHR32(bank[i*C+c], shift);
- if (E>0)
- g = DIV32_16(SHL32(Q15ONE,13),MULT16_16_Q14(E,sqrtC_1[C-1]));
- else
- g = 0;
- for (j=B*eBands[i];j<B*eBands[i+1];j++)
+ g = EXTRACT16(celt_rcp(SHR32(MULT16_16(E,sqrtC_1[C-1]),11)));
+ for (j=eBands[i];j<eBands[i+1];j++)
X[j*C+c] = MULT16_16_Q14(VSHR32(freq[j*C+c],shift),g);
}
}
- for (i=B*C*eBands[m->nbEBands];i<B*C*eBands[m->nbEBands+1];i++)
+ for (i=C*eBands[m->nbEBands];i<C*eBands[m->nbEBands+1];i++)
X[i] = 0;
}
-void renormalise_bands(const CELTMode *m, celt_norm_t *X)
+#ifndef DISABLE_STEREO
+void renormalise_bands(const CELTMode *m, celt_norm_t * restrict X)
{
int i;
VARDECL(celt_ener_t, tmpE);
VARDECL(celt_sig_t, freq);
SAVE_STACK;
ALLOC(tmpE, m->nbEBands*m->nbChannels, celt_ener_t);
- ALLOC(freq, m->nbMdctBlocks*m->nbChannels*m->eBands[m->nbEBands+1], celt_sig_t);
- for (i=0;i<m->nbMdctBlocks*m->nbChannels*m->eBands[m->nbEBands+1];i++)
+ ALLOC(freq, m->nbChannels*m->eBands[m->nbEBands+1], celt_sig_t);
+ for (i=0;i<m->nbChannels*m->eBands[m->nbEBands+1];i++)
freq[i] = SHL32(EXTEND32(X[i]), 10);
compute_band_energies(m, freq, tmpE);
normalise_bands(m, freq, X, tmpE);
RESTORE_STACK;
}
-#else
+#endif /* DISABLE_STEREO */
+#else /* FIXED_POINT */
/* Compute the amplitude (sqrt energy) in each of the bands */
void compute_band_energies(const CELTMode *m, const celt_sig_t *X, celt_ener_t *bank)
{
- int i, c, B, C;
+ int i, c;
const celt_int16_t *eBands = m->eBands;
- B = m->nbMdctBlocks;
- C = m->nbChannels;
+ const int C = CHANNELS(m);
for (c=0;c<C;c++)
{
for (i=0;i<m->nbEBands;i++)
{
int j;
celt_word32_t sum = 1e-10;
- for (j=B*eBands[i];j<B*eBands[i+1];j++)
+ for (j=eBands[i];j<eBands[i+1];j++)
sum += X[j*C+c]*X[j*C+c];
bank[i*C+c] = sqrt(sum);
/*printf ("%f ", bank[i*C+c]);*/
}
/* Normalise each band such that the energy is one. */
-void normalise_bands(const CELTMode *m, const celt_sig_t *freq, celt_norm_t *X, const celt_ener_t *bank)
+void normalise_bands(const CELTMode *m, const celt_sig_t * restrict freq, celt_norm_t * restrict X, const celt_ener_t *bank)
{
- int i, c, B, C;
+ int i, c;
const celt_int16_t *eBands = m->eBands;
- B = m->nbMdctBlocks;
- C = m->nbChannels;
+ const int C = CHANNELS(m);
for (c=0;c<C;c++)
{
for (i=0;i<m->nbEBands;i++)
{
int j;
celt_word16_t g = 1.f/(1e-10+bank[i*C+c]*sqrt(C));
- for (j=B*eBands[i];j<B*eBands[i+1];j++)
+ for (j=eBands[i];j<eBands[i+1];j++)
X[j*C+c] = freq[j*C+c]*g;
}
}
- for (i=B*C*eBands[m->nbEBands];i<B*C*eBands[m->nbEBands+1];i++)
+ for (i=C*eBands[m->nbEBands];i<C*eBands[m->nbEBands+1];i++)
X[i] = 0;
}
-void renormalise_bands(const CELTMode *m, celt_norm_t *X)
+#ifndef DISABLE_STEREO
+void renormalise_bands(const CELTMode *m, celt_norm_t * restrict X)
{
VARDECL(celt_ener_t, tmpE);
SAVE_STACK;
normalise_bands(m, X, X, tmpE);
RESTORE_STACK;
}
-#endif
+#endif /* DISABLE_STEREO */
+#endif /* FIXED_POINT */
/* De-normalise the energy to produce the synthesis from the unit-energy bands */
-void denormalise_bands(const CELTMode *m, const celt_norm_t *X, celt_sig_t *freq, const celt_ener_t *bank)
+void denormalise_bands(const CELTMode *m, const celt_norm_t * restrict X, celt_sig_t * restrict freq, const celt_ener_t *bank)
{
- int i, c, B, C;
+ int i, c;
const celt_int16_t *eBands = m->eBands;
- B = m->nbMdctBlocks;
- C = m->nbChannels;
+ const int C = CHANNELS(m);
if (C>2)
celt_fatal("denormalise_bands() not implemented for >2 channels");
for (c=0;c<C;c++)
{
int j;
celt_word32_t g = MULT16_32_Q14(sqrtC_1[C-1],bank[i*C+c]);
- for (j=B*eBands[i];j<B*eBands[i+1];j++)
+ for (j=eBands[i];j<eBands[i+1];j++)
freq[j*C+c] = MULT16_32_Q14(X[j*C+c], g);
}
}
- for (i=B*C*eBands[m->nbEBands];i<B*C*eBands[m->nbEBands+1];i++)
+ for (i=C*eBands[m->nbEBands];i<C*eBands[m->nbEBands+1];i++)
freq[i] = 0;
}
/* Compute the best gain for each "pitch band" */
void compute_pitch_gain(const CELTMode *m, const celt_norm_t *X, const celt_norm_t *P, celt_pgain_t *gains)
{
- int i, B;
+ int i;
const celt_int16_t *pBands = m->pBands;
- B = m->nbMdctBlocks*m->nbChannels;
-
+ const int C = CHANNELS(m);
+
for (i=0;i<m->nbPBands;i++)
{
celt_word32_t Sxy=0, Sxx=0;
int j;
/* We know we're not going to overflow because Sxx can't be more than 1 (Q28) */
- for (j=B*pBands[i];j<B*pBands[i+1];j++)
+ for (j=C*pBands[i];j<C*pBands[i+1];j++)
{
Sxy = MAC16_16(Sxy, X[j], P[j]);
Sxx = MAC16_16(Sxx, X[j], X[j]);
residual doesn't quantise well */
Sxy = MULT16_32_Q15(QCONST16(.9f, 15), Sxy);
/* gain = Sxy/Sxx */
- gains[i] = DIV32_16(Sxy,ADD32(SHR32(Sxx, PGAIN_SHIFT),EPSILON));
+ gains[i] = EXTRACT16(celt_div(Sxy,ADD32(SHR32(Sxx, PGAIN_SHIFT),EPSILON)));
/*printf ("%f ", 1-sqrt(1-gain*gain));*/
}
/*if(rand()%10==0)
}
/* Apply the (quantised) gain to each "pitch band" */
-void pitch_quant_bands(const CELTMode *m, celt_norm_t *P, const celt_pgain_t *gains)
+void pitch_quant_bands(const CELTMode *m, celt_norm_t * restrict P, const celt_pgain_t * restrict gains)
{
- int i, B;
+ int i;
const celt_int16_t *pBands = m->pBands;
- B = m->nbMdctBlocks*m->nbChannels;
+ const int C = CHANNELS(m);
for (i=0;i<m->nbPBands;i++)
{
int j;
- for (j=B*pBands[i];j<B*pBands[i+1];j++)
+ for (j=C*pBands[i];j<C*pBands[i+1];j++)
P[j] = MULT16_16_Q15(gains[i], P[j]);
/*printf ("%f ", gain);*/
}
- for (i=B*pBands[m->nbPBands];i<B*pBands[m->nbPBands+1];i++)
+ for (i=C*pBands[m->nbPBands];i<C*pBands[m->nbPBands+1];i++)
P[i] = 0;
}
/* Quantisation of the residual */
-void quant_bands(const CELTMode *m, celt_norm_t *X, celt_norm_t *P, celt_mask_t *W, int total_bits, ec_enc *enc)
+void quant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, celt_mask_t *W, int total_bits, ec_enc *enc)
{
- int i, j, B, bits;
+ int i, j, bits;
const celt_int16_t *eBands = m->eBands;
- celt_word16_t alpha;
- VARDECL(celt_norm_t, norm);
+ celt_norm_t * restrict norm;
+ VARDECL(celt_norm_t, _norm);
VARDECL(int, pulses);
VARDECL(int, offsets);
+ const int C = CHANNELS(m);
SAVE_STACK;
- B = m->nbMdctBlocks*m->nbChannels;
-
- ALLOC(norm, B*eBands[m->nbEBands+1], celt_norm_t);
+ ALLOC(_norm, C*eBands[m->nbEBands+1], celt_norm_t);
ALLOC(pulses, m->nbEBands, int);
ALLOC(offsets, m->nbEBands, int);
+ norm = _norm;
for (i=0;i<m->nbEBands;i++)
offsets[i] = 0;
int q;
celt_word16_t n;
q = pulses[i];
- n = SHL16(celt_sqrt(B*(eBands[i+1]-eBands[i])),11);
+ n = SHL16(celt_sqrt(C*(eBands[i+1]-eBands[i])),11);
/* If pitch isn't available, use intra-frame prediction */
if (eBands[i] >= m->pitchEnd || q<=0)
{
q -= 1;
- alpha = 0;
if (q<0)
- intra_fold(X+B*eBands[i], B*(eBands[i+1]-eBands[i]), norm, P+B*eBands[i], B, eBands[i], eBands[m->nbEBands+1]);
+ intra_fold(X+C*eBands[i], eBands[i+1]-eBands[i], norm, P+C*eBands[i], C, eBands[i], eBands[m->nbEBands+1]);
else
- intra_prediction(X+B*eBands[i], W+B*eBands[i], B*(eBands[i+1]-eBands[i]), q, norm, P+B*eBands[i], B, eBands[i], enc);
- } else {
- alpha = QCONST16(.7f,15);
+ intra_prediction(X+C*eBands[i], W+C*eBands[i], eBands[i+1]-eBands[i], q, norm, P+C*eBands[i], C, eBands[i], enc);
}
if (q > 0)
{
- int nb_rotations = (B*(eBands[i+1]-eBands[i])+4*q)/(8*q);
- exp_rotation(P+B*eBands[i], B*(eBands[i+1]-eBands[i]), -1, B, nb_rotations);
- exp_rotation(X+B*eBands[i], B*(eBands[i+1]-eBands[i]), -1, B, nb_rotations);
- alg_quant(X+B*eBands[i], W+B*eBands[i], B*(eBands[i+1]-eBands[i]), q, P+B*eBands[i], alpha, enc);
- exp_rotation(X+B*eBands[i], B*(eBands[i+1]-eBands[i]), 1, B, nb_rotations);
+ int nb_rotations = q <= 2*C ? 2*C/q : 0;
+ if (nb_rotations != 0)
+ {
+ exp_rotation(P+C*eBands[i], C*(eBands[i+1]-eBands[i]), -1, C, nb_rotations);
+ exp_rotation(X+C*eBands[i], C*(eBands[i+1]-eBands[i]), -1, C, nb_rotations);
+ }
+ alg_quant(X+C*eBands[i], W+C*eBands[i], C*(eBands[i+1]-eBands[i]), q, P+C*eBands[i], enc);
+ if (nb_rotations != 0)
+ exp_rotation(X+C*eBands[i], C*(eBands[i+1]-eBands[i]), 1, C, nb_rotations);
}
- for (j=B*eBands[i];j<B*eBands[i+1];j++)
+ for (j=C*eBands[i];j<C*eBands[i+1];j++)
norm[j] = MULT16_16_Q15(n,X[j]);
}
- for (i=B*eBands[m->nbEBands];i<B*eBands[m->nbEBands+1];i++)
+ for (i=C*eBands[m->nbEBands];i<C*eBands[m->nbEBands+1];i++)
X[i] = 0;
RESTORE_STACK;
}
/* Decoding of the residual */
-void unquant_bands(const CELTMode *m, celt_norm_t *X, celt_norm_t *P, int total_bits, ec_dec *dec)
+void unquant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, int total_bits, ec_dec *dec)
{
- int i, j, B, bits;
+ int i, j, bits;
const celt_int16_t *eBands = m->eBands;
- celt_word16_t alpha;
- VARDECL(celt_norm_t, norm);
+ celt_norm_t * restrict norm;
+ VARDECL(celt_norm_t, _norm);
VARDECL(int, pulses);
VARDECL(int, offsets);
+ const int C = CHANNELS(m);
SAVE_STACK;
- B = m->nbMdctBlocks*m->nbChannels;
-
- ALLOC(norm, B*eBands[m->nbEBands+1], celt_norm_t);
+ ALLOC(_norm, C*eBands[m->nbEBands+1], celt_norm_t);
ALLOC(pulses, m->nbEBands, int);
ALLOC(offsets, m->nbEBands, int);
+ norm = _norm;
for (i=0;i<m->nbEBands;i++)
offsets[i] = 0;
int q;
celt_word16_t n;
q = pulses[i];
- n = SHL16(celt_sqrt(B*(eBands[i+1]-eBands[i])),11);
+ n = SHL16(celt_sqrt(C*(eBands[i+1]-eBands[i])),11);
/* If pitch isn't available, use intra-frame prediction */
if (eBands[i] >= m->pitchEnd || q<=0)
{
q -= 1;
- alpha = 0;
if (q<0)
- intra_fold(X+B*eBands[i], B*(eBands[i+1]-eBands[i]), norm, P+B*eBands[i], B, eBands[i], eBands[m->nbEBands+1]);
+ intra_fold(X+C*eBands[i], eBands[i+1]-eBands[i], norm, P+C*eBands[i], C, eBands[i], eBands[m->nbEBands+1]);
else
- intra_unquant(X+B*eBands[i], B*(eBands[i+1]-eBands[i]), q, norm, P+B*eBands[i], B, eBands[i], dec);
- } else {
- alpha = QCONST16(.7f,15);
+ intra_unquant(X+C*eBands[i], eBands[i+1]-eBands[i], q, norm, P+C*eBands[i], C, eBands[i], dec);
}
if (q > 0)
{
- int nb_rotations = (B*(eBands[i+1]-eBands[i])+4*q)/(8*q);
- exp_rotation(P+B*eBands[i], B*(eBands[i+1]-eBands[i]), -1, B, nb_rotations);
- alg_unquant(X+B*eBands[i], B*(eBands[i+1]-eBands[i]), q, P+B*eBands[i], alpha, dec);
- exp_rotation(X+B*eBands[i], B*(eBands[i+1]-eBands[i]), 1, B, nb_rotations);
+ int nb_rotations = q <= 2*C ? 2*C/q : 0;
+ if (nb_rotations != 0)
+ exp_rotation(P+C*eBands[i], C*(eBands[i+1]-eBands[i]), -1, C, nb_rotations);
+ alg_unquant(X+C*eBands[i], C*(eBands[i+1]-eBands[i]), q, P+C*eBands[i], dec);
+ if (nb_rotations != 0)
+ exp_rotation(X+C*eBands[i], C*(eBands[i+1]-eBands[i]), 1, C, nb_rotations);
}
- for (j=B*eBands[i];j<B*eBands[i+1];j++)
+ for (j=C*eBands[i];j<C*eBands[i+1];j++)
norm[j] = MULT16_16_Q15(n,X[j]);
}
- for (i=B*eBands[m->nbEBands];i<B*eBands[m->nbEBands+1];i++)
+ for (i=C*eBands[m->nbEBands];i<C*eBands[m->nbEBands+1];i++)
X[i] = 0;
RESTORE_STACK;
}
+#ifndef DISABLE_STEREO
void stereo_mix(const CELTMode *m, celt_norm_t *X, const celt_ener_t *bank, int dir)
{
- int i, B, C;
+ int i;
const celt_int16_t *eBands = m->eBands;
- B = m->nbMdctBlocks;
- C = m->nbChannels;
+ const int C = CHANNELS(m);
for (i=0;i<m->nbEBands;i++)
{
int j;
celt_word16_t a1, a2;
celt_word16_t norm;
#ifdef FIXED_POINT
- int shift = celt_ilog2(MAX32(bank[i*C], bank[i*C+1]))-13;
+ int shift = celt_zlog2(MAX32(bank[i*C], bank[i*C+1]))-13;
#endif
left = VSHR32(bank[i*C],shift);
right = VSHR32(bank[i*C+1],shift);
norm = EPSILON + celt_sqrt(EPSILON+MULT16_16(left,left)+MULT16_16(right,right));
a1 = DIV32_16(SHL32(EXTEND32(left),14),norm);
a2 = dir*DIV32_16(SHL32(EXTEND32(right),14),norm);
- for (j=B*eBands[i];j<B*eBands[i+1];j++)
+ for (j=eBands[i];j<eBands[i+1];j++)
{
celt_norm_t r, l;
l = X[j*C];
X[j*C+1] = MULT16_16_Q14(a1,r) - MULT16_16_Q14(a2,l);
}
}
- for (i=B*C*eBands[m->nbEBands];i<B*C*eBands[m->nbEBands+1];i++)
+ for (i=C*eBands[m->nbEBands];i<C*eBands[m->nbEBands+1];i++)
X[i] = 0;
-
}
+#endif
\ No newline at end of file
xiph / opus.git / blobdiff