Adding a dual stereo option.

Left and right are coded independently.
This commit is contained in:
Jean-Marc Valin 2010-12-02 13:46:48 -05:00
parent 100ae8ce34
commit e65978fea7
3 changed files with 69 additions and 10 deletions

View file

@ -902,13 +902,13 @@ static void quant_band(int encode, const CELTMode *m, int i, celt_norm *X, celt_
void quant_all_bands(int encode, const CELTMode *m, int start, int end,
celt_norm *_X, celt_norm *_Y, const celt_ener *bandE, int *pulses,
int shortBlocks, int fold, int intensity, int *tf_res, int resynth,
int shortBlocks, int fold, int dual_stereo, int intensity, int *tf_res, int resynth,
int total_bits, void *ec, int LM, int codedBands)
{
int i, balance;
celt_int32 remaining_bits;
const celt_int16 * restrict eBands = m->eBands;
celt_norm * restrict norm;
celt_norm * restrict norm, * restrict norm2;
VARDECL(celt_norm, _norm);
VARDECL(celt_norm, lowband_scratch);
int B;
@ -921,10 +921,10 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end,
M = 1<<LM;
B = shortBlocks ? M : 1;
ALLOC(_norm, M*eBands[m->nbEBands], celt_norm);
ALLOC(_norm, C*M*eBands[m->nbEBands], celt_norm);
ALLOC(lowband_scratch, M*(eBands[m->nbEBands]-eBands[m->nbEBands-1]), celt_norm);
norm = _norm;
norm2 = norm + M*eBands[m->nbEBands];
if (C==2)
{
int j;
@ -1015,10 +1015,28 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end,
if (effective_lowband < M*eBands[start])
effective_lowband = M*eBands[start];
}
quant_band(encode, m, i, X, Y, N, b, fold, B, intensity, tf_change,
effective_lowband != -1 ? norm+effective_lowband : NULL, resynth, ec, &remaining_bits, LM,
norm+M*eBands[i], bandE, 0, &seed, Q15ONE, lowband_scratch);
if (dual_stereo && i==intensity)
{
int j;
/* Switch off dual stereo to do intensity */
dual_stereo = 0;
for (j=0;j<M*eBands[i];j++)
norm[j] = HALF32(norm[j]+norm2[j]);
}
if (dual_stereo)
{
quant_band(encode, m, i, X, NULL, N, b/2, fold, B, intensity, tf_change,
effective_lowband != -1 ? norm+effective_lowband : NULL, resynth, ec, &remaining_bits, LM,
norm+M*eBands[i], bandE, 0, &seed, Q15ONE, lowband_scratch);
quant_band(encode, m, i, Y, NULL, N, b/2, fold, B, intensity, tf_change,
effective_lowband != -1 ? norm2+effective_lowband : NULL, resynth, ec, &remaining_bits, LM,
norm2+M*eBands[i], bandE, 0, &seed, Q15ONE, lowband_scratch);
} else {
quant_band(encode, m, i, X, Y, N, b, fold, B, intensity, tf_change,
effective_lowband != -1 ? norm+effective_lowband : NULL, resynth, ec, &remaining_bits, LM,
norm+M*eBands[i], bandE, 0, &seed, Q15ONE, lowband_scratch);
}
balance += pulses[i] + tell;
/* Update the folding position only as long as we have 1 bit/sample depth */

View file

@ -80,7 +80,7 @@ void haar1(celt_norm *X, int N0, int stride);
*/
void quant_all_bands(int encode, const CELTMode *m, int start, int end,
celt_norm * X, celt_norm * Y, const celt_ener *bandE, int *pulses,
int time_domain, int fold, int intensity, int *tf_res, int resynth,
int time_domain, int fold, int dual_stereo, int intensity, int *tf_res, int resynth,
int total_bits, void *enc, int M, int codedBands);

View file

@ -649,6 +649,37 @@ static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X,
return trim_index;
}
static int stereo_analysis(const CELTMode *m, const celt_norm *X,
int nbEBands, int LM, int C, int N0)
{
int i;
int thetas;
celt_word32 sumLR = EPSILON, sumMS = EPSILON;
/* Use the L1 norm to model the entropy of the L/R signal vs the M/S signal */
for (i=0;i<13;i++)
{
int j;
for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++)
{
celt_word16 L, R, M, S;
L = X[j];
R = X[N0+j];
M = L+R;
S = L-R;
sumLR += EXTEND32(ABS16(L)) + EXTEND32(ABS16(R));
sumMS += EXTEND32(ABS16(M)) + EXTEND32(ABS16(S));
}
}
sumMS = MULT16_32_Q15(QCONST16(0.707107f, 15), sumMS);
thetas = 13;
/* We don't need thetas for lower bands with LM<=1 */
if (LM<=1)
thetas -= 8;
return MULT16_32_Q15((m->eBands[13]<<(LM+1))+thetas, sumMS)
> MULT16_32_Q15(m->eBands[13]<<(LM+1), sumLR);
}
#ifdef FIXED_POINT
int celt_encode_with_ec(CELTEncoder * restrict st, const celt_int16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes, ec_enc *enc)
{
@ -689,6 +720,7 @@ int celt_encode_with_ec_float(CELTEncoder * restrict st, const celt_sig * pcm, i
int pitch_index=0;
celt_word16 gain1 = 0;
int intensity=0;
int dual_stereo=0;
SAVE_STACK;
if (nbCompressedBytes<0 || pcm==NULL)
@ -1001,6 +1033,11 @@ int celt_encode_with_ec_float(CELTEncoder * restrict st, const celt_sig * pcm, i
ec_byte_shrink(&buf, nbCompressedBytes);
}
if (C==2)
{
dual_stereo = stereo_analysis(st->mode, X, st->mode->nbEBands, LM, C, N);
ec_enc_bit_prob(enc, dual_stereo, 32768);
}
if (C==2)
{
int effectiveRate;
@ -1055,7 +1092,7 @@ int celt_encode_with_ec_float(CELTEncoder * restrict st, const celt_sig * pcm, i
/* Residual quantisation */
quant_all_bands(1, st->mode, st->start, st->end, X, C==2 ? X+N : NULL,
bandE, pulses, shortBlocks, has_fold, intensity, tf_res, resynth,
bandE, pulses, shortBlocks, has_fold, dual_stereo, intensity, tf_res, resynth,
nbCompressedBytes*8, enc, LM, codedBands);
quant_energy_finalise(st->mode, st->start, st->end, bandE, oldBandE, error, fine_quant, fine_priority, nbCompressedBytes*8-ec_enc_tell(enc, 0), enc, C);
@ -1619,6 +1656,7 @@ int celt_decode_with_ec_float(CELTDecoder * restrict st, const unsigned char *da
int postfilter_pitch;
celt_word16 postfilter_gain;
int intensity=0;
int dual_stereo=0;
SAVE_STACK;
if (pcm==NULL)
@ -1747,7 +1785,10 @@ int celt_decode_with_ec_float(CELTDecoder * restrict st, const unsigned char *da
}
if (C==2)
{
dual_stereo = ec_dec_bit_prob(dec, 32768);
intensity = ec_dec_uint(dec, 1+st->end-st->start);
}
bits = len*8 - ec_dec_tell(dec, 0) - 1;
codedBands = compute_allocation(st->mode, st->start, st->end, offsets, alloc_trim, bits, pulses, fine_quant, fine_priority, C, LM);
@ -1756,7 +1797,7 @@ int celt_decode_with_ec_float(CELTDecoder * restrict st, const unsigned char *da
/* Decode fixed codebook */
quant_all_bands(0, st->mode, st->start, st->end, X, C==2 ? X+N : NULL,
NULL, pulses, shortBlocks, has_fold, intensity, tf_res, 1,
NULL, pulses, shortBlocks, has_fold, dual_stereo, intensity, tf_res, 1,
len*8, dec, LM, codedBands);
unquant_energy_finalise(st->mode, st->start, st->end, bandE, oldBandE,