From a4833ffadaee1d293f230eb9e5fa5308128f0db8 Mon Sep 17 00:00:00 2001
From: Jean-Marc Valin <Jean-Marc.Valin@csiro.au>
Date: Thu, 10 Jan 2008 15:34:00 +1100
Subject: [PATCH] Stereo decoding working again (fixed a few issues in the
encoder at the same time)
---
libcelt/celt.c | 24 +++++-----
libcelt/modes.c | 4 +-
libcelt/quant_bands.c | 102 +++++++++++++++++++++++++++++-------------
libcelt/vq.c | 6 +--
4 files changed, 88 insertions(+), 48 deletions(-)
diff --git a/libcelt/celt.c b/libcelt/celt.c
index c165762ab..f01ac1e3c 100644
--- a/libcelt/celt.c
+++ b/libcelt/celt.c
@@ -107,7 +107,7 @@ CELTEncoder *celt_encoder_new(const CELTMode *mode)
= sin(.5*M_PI* sin(.5*M_PI*(i+.5)/st->overlap) * sin(.5*M_PI*(i+.5)/st->overlap));
for (i=0;i<2*N4;i++)
st->window[N-N4+i] = 1;
- st->oldBandE = celt_alloc(mode->nbEBands*sizeof(float));
+ st->oldBandE = celt_alloc(C*mode->nbEBands*sizeof(float));
st->preemph = 0.8;
st->preemph_memE = celt_alloc(C*sizeof(float));;
@@ -318,7 +318,6 @@ int celt_encode(CELTEncoder *st, short *pcm)
time_dct(X, N, B, C);
time_dct(P, N, B, C);
-
quant_energy(st->mode, bandE, st->oldBandE, &st->enc);
/* Pitch prediction */
@@ -447,7 +446,7 @@ CELTDecoder *celt_decoder_new(const CELTMode *mode)
for (i=0;i<2*N4;i++)
st->window[N-N4+i] = 1;
- st->oldBandE = celt_alloc(mode->nbEBands*sizeof(float));
+ st->oldBandE = celt_alloc(C*mode->nbEBands*sizeof(float));
st->preemph = 0.8;
st->preemph_memD = celt_alloc(C*sizeof(float));;
@@ -518,7 +517,7 @@ int celt_decode(CELTDecoder *st, char *data, int len, short *pcm)
float X[C*B*N]; /**< Interleaved signal MDCTs */
float P[C*B*N]; /**< Interleaved pitch MDCTs*/
- float bandE[st->mode->nbEBands];
+ float bandE[st->mode->nbEBands*C];
float gains[st->mode->nbPBands];
int pitch_index;
ec_dec dec;
@@ -543,16 +542,16 @@ int celt_decode(CELTDecoder *st, char *data, int len, short *pcm)
/* Pitch MDCT */
compute_mdcts(&st->mdct_lookup, st->window, st->out_mem+pitch_index*C, P, N, B, C);
- if (C==2)
- haar1(P, B*N*C, 1);
- time_dct(P, N, B, C);
-
{
float bandEp[st->mode->nbEBands];
compute_band_energies(st->mode, P, bandEp);
normalise_bands(st->mode, P, bandEp);
}
+ if (C==2)
+ haar1(P, B*N*C, 1);
+ time_dct(P, N, B, C);
+
/* Get the pitch gains */
unquant_pitch(gains, st->mode->nbPBands, &dec);
@@ -562,13 +561,16 @@ int celt_decode(CELTDecoder *st, char *data, int len, short *pcm)
/* Decode fixed codebook and merge with pitch */
unquant_bands(st->mode, X, P, &dec);
- /* Synthesis */
- denormalise_bands(st->mode, X, bandE);
-
time_idct(X, N, B, C);
if (C==2)
haar1(X, B*N*C, 1);
+ renormalise_bands(st->mode, X);
+
+ /* Synthesis */
+ denormalise_bands(st->mode, X, bandE);
+
+
CELT_MOVE(st->out_mem, st->out_mem+C*B*N, C*(MAX_PERIOD-B*N));
/* Compute inverse MDCTs */
compute_inv_mdcts(&st->mdct_lookup, st->window, X, st->out_mem, st->mdct_overlap, N, st->overlap, B, C);
diff --git a/libcelt/modes.c b/libcelt/modes.c
index 8d8ee6954..b354880f3 100644
--- a/libcelt/modes.c
+++ b/libcelt/modes.c
@@ -77,9 +77,9 @@ const int pbank1[PBANDS128+2] = {0, 2, 4, 6, 8, 12, 20, 28, PITCH_END128, 128};
#define NBANDS256 15
#define PBANDS256 8
#define PITCH_END256 88
-const int qbank3[NBANDS256+2] = {0, 4, 8, 12, 16, 24, 32, 40, 48, 56, 72, 88, 104, 126, 168, 232, 256};
+const int qbank3[NBANDS256+2] = {0, 4, 8, 12, 16, 24, 32, 40, 48, 56, 72, 88, 104, 136, 168, 232, 256};
//const int pbank3[PBANDS256+2] = {0, 8, 16, 24, 40, PITCH_END256, 256};
-const int pbank3[PBANDS256+2] = {0, 4, 8, 12, 16, 24, 40, 56, PITCH_END128, 128};
+const int pbank3[PBANDS256+2] = {0, 4, 8, 12, 16, 24, 40, 56, PITCH_END256, 256};
const CELTMode mode0 = {
128, /**< overlap */
diff --git a/libcelt/quant_bands.c b/libcelt/quant_bands.c
index c07915c60..d4dd2eab2 100644
--- a/libcelt/quant_bands.c
+++ b/libcelt/quant_bands.c
@@ -35,6 +35,38 @@
#include <math.h>
#include "os_support.h"
+static void quant_energy_mono(const CELTMode *m, float *eBands, float *oldEBands, ec_enc *enc)
+{
+ int i;
+ float prev = 0;
+ for (i=0;i<m->nbEBands;i++)
+ {
+ int qi;
+ float q;
+ float res;
+ float x;
+ float pred = m->ePredCoef*oldEBands[i]+m->eMeans[i];
+
+ x = 20*log10(.3+eBands[i]);
+ res = .25f*(i+3.f);
+ //res = 1;
+ qi = (int)floor(.5+(x-pred-prev)/res);
+ ec_laplace_encode(enc, qi, m->eDecay[i]);
+ q = qi*res;
+
+ //printf("%d ", qi);
+ //printf("%f %f ", pred+prev+q, x);
+ //printf("%f ", x-pred);
+
+ oldEBands[i] = pred+prev+q;
+ eBands[i] = pow(10, .05*oldEBands[i])-.3;
+ if (eBands[i] < 0)
+ eBands[i] = 0;
+ prev = (prev + .5*q);
+ }
+ //printf ("\n");
+}
+
void quant_energy(const CELTMode *m, float *eBands, float *oldEBands, ec_enc *enc)
{
int C;
@@ -43,7 +75,23 @@ void quant_energy(const CELTMode *m, float *eBands, float *oldEBands, ec_enc *en
if (C==1)
quant_energy_mono(m, eBands, oldEBands, enc);
- else if (C==2)
+ else
+#if 1
+ {
+ int c;
+ for (c=0;c<C;c++)
+ {
+ int i;
+ float E[m->nbEBands];
+ for (i=0;i<m->nbEBands;i++)
+ E[i] = eBands[C*i+c];
+ quant_energy_mono(m, E, oldEBands+c*m->nbEBands, enc);
+ for (i=0;i<m->nbEBands;i++)
+ eBands[C*i+c] = E[i];
+ }
+ }
+#else
+ if (C==2)
{
int i;
int NB = m->nbEBands;
@@ -75,9 +123,11 @@ void quant_energy(const CELTMode *m, float *eBands, float *oldEBands, ec_enc *en
} else {
celt_fatal("more than 2 channels not supported");
}
+#endif
}
-void quant_energy_mono(const CELTMode *m, float *eBands, float *oldEBands, ec_enc *enc)
+
+static void unquant_energy_mono(const CELTMode *m, float *eBands, float *oldEBands, ec_dec *dec)
{
int i;
float prev = 0;
@@ -86,19 +136,14 @@ void quant_energy_mono(const CELTMode *m, float *eBands, float *oldEBands, ec_en
int qi;
float q;
float res;
- float x;
float pred = m->ePredCoef*oldEBands[i]+m->eMeans[i];
- x = 20*log10(.3+eBands[i]);
res = .25f*(i+3.f);
- //res = 1;
- qi = (int)floor(.5+(x-pred-prev)/res);
- ec_laplace_encode(enc, qi, m->eDecay[i]);
+ qi = ec_laplace_decode(dec, m->eDecay[i]);
q = qi*res;
-
- //printf("%d ", qi);
//printf("%f %f ", pred+prev+q, x);
- //printf("%f ", x-pred);
+ //printf("%d ", qi);
+ //printf("%f ", x-pred-prev);
oldEBands[i] = pred+prev+q;
eBands[i] = pow(10, .05*oldEBands[i])-.3;
@@ -111,27 +156,20 @@ void quant_energy_mono(const CELTMode *m, float *eBands, float *oldEBands, ec_en
void unquant_energy(const CELTMode *m, float *eBands, float *oldEBands, ec_dec *dec)
{
- int i;
- float prev = 0;
- for (i=0;i<m->nbEBands;i++)
- {
- int qi;
- float q;
- float res;
- float pred = m->ePredCoef*oldEBands[i]+m->eMeans[i];
-
- res = .25f*(i+3.f);
- qi = ec_laplace_decode(dec, m->eDecay[i]);
- q = qi*res;
- //printf("%f %f ", pred+prev+q, x);
- //printf("%d ", qi);
- //printf("%f ", x-pred-prev);
-
- oldEBands[i] = pred+prev+q;
- eBands[i] = pow(10, .05*oldEBands[i])-.3;
- if (eBands[i] < 0)
- eBands[i] = 0;
- prev = (prev + .5*q);
+ int C;
+ C = m->nbChannels;
+
+ if (C==1)
+ unquant_energy_mono(m, eBands, oldEBands, dec);
+ else {
+ int c;
+ for (c=0;c<C;c++)
+ {
+ int i;
+ float E[m->nbEBands];
+ unquant_energy_mono(m, E, oldEBands+c*m->nbEBands, dec);
+ for (i=0;i<m->nbEBands;i++)
+ eBands[C*i+c] = E[i];
+ }
}
- //printf ("\n");
}
diff --git a/libcelt/vq.c b/libcelt/vq.c
index 9dbb108fc..e38090070 100644
--- a/libcelt/vq.c
+++ b/libcelt/vq.c
@@ -188,8 +188,9 @@ void alg_quant(float *x, float *W, int N, int K, float *p, float alpha, ec_enc *
pulse2comb(N, K, comb, signs, iy[0]);
ec_enc_uint64(enc,icwrs64(N, K, comb, signs),ncwrs64(N, K));
- /* Recompute the gain in one pass (to reduce errors) */
- if (0) {
+ /* Recompute the gain in one pass to reduce the encoder-decoder mismatch
+ due to the recursive computation used in quantisation */
+ if (1) {
float Ryp=0;
float Rpp=0;
float Ryy=0;
@@ -203,7 +204,6 @@ void alg_quant(float *x, float *W, int N, int K, float *p, float alpha, ec_enc *
for (i=0;i<N;i++)
y[0][i] = iy[0][i] - alpha*Ryp*p[i];
- /* Recompute after the projection (I think it's right) */
Ryp = 0;
for (i=0;i<N;i++)
Ryp += y[0][i]*p[i];
--
GitLab