From ca6533cd88e7de8a6275a7a576a07f1db1459814 Mon Sep 17 00:00:00 2001
From: Jean-Marc Valin <jean-marc.valin@octasic.com>
Date: Fri, 23 Jul 2010 16:10:36 -0400
Subject: [PATCH] Energy quantization tuning.
Prediction coefficient now depends on the frame size. Also, the pdfs and means
have been adjusted and better quantized. This breaks compatibility of course.
---
libcelt/celt.c | 4 ++--
libcelt/dump_modes.c | 1 -
libcelt/modes.c | 1 -
libcelt/modes.h | 12 +-----------
libcelt/quant_bands.c | 36 +++++++++++++++++++++++-------------
libcelt/quant_bands.h | 4 ++--
6 files changed, 28 insertions(+), 30 deletions(-)
diff --git a/libcelt/celt.c b/libcelt/celt.c
index 763aea4e3..8e9e4f61d 100644
--- a/libcelt/celt.c
+++ b/libcelt/celt.c
@@ -986,7 +986,7 @@ int celt_encode_with_ec_float(CELTEncoder * restrict st, const celt_sig * pcm, c
#else
max_decay = .125*nbAvailableBytes;
#endif
- quant_coarse_energy(st->mode, st->start, st->end, bandLogE, st->oldBandE, nbCompressedBytes*8, intra_ener, st->mode->prob, error, enc, C, max_decay);
+ quant_coarse_energy(st->mode, st->start, st->end, bandLogE, st->oldBandE, nbCompressedBytes*8, intra_ener, st->mode->prob, error, enc, C, LM, max_decay);
/* Variable bitrate */
if (vbr_rate>0)
{
@@ -1800,7 +1800,7 @@ int celt_decode_with_ec_float(CELTDecoder * restrict st, const unsigned char *da
ALLOC(fine_quant, st->mode->nbEBands, int);
/* Get band energies */
- unquant_coarse_energy(st->mode, st->start, st->end, bandE, st->oldBandE, intra_ener, st->mode->prob, dec, C);
+ unquant_coarse_energy(st->mode, st->start, st->end, bandE, st->oldBandE, intra_ener, st->mode->prob, dec, C, LM);
ALLOC(tf_res, st->mode->nbEBands, int);
tf_decode(st->start, st->end, C, isTransient, tf_res, nbAvailableBytes, LM, dec);
diff --git a/libcelt/dump_modes.c b/libcelt/dump_modes.c
index 86aa796a0..d980764c5 100644
--- a/libcelt/dump_modes.c
+++ b/libcelt/dump_modes.c
@@ -152,7 +152,6 @@ void dump_modes(FILE *file, CELTMode **modes, int nb_modes)
fprintf(file, "%d,\t/* nbEBands */\n", mode->nbEBands);
fprintf(file, "%d,\t/* pitchEnd */\n", mode->pitchEnd);
fprintf(file, "eBands%d_%d,\t/* eBands */\n", mode->Fs, mdctSize);
- fprintf(file, WORD16 ",\t/* ePredCoef */\n", mode->ePredCoef);
fprintf(file, "%d,\t/* nbAllocVectors */\n", mode->nbAllocVectors);
fprintf(file, "allocVectors%d_%d,\t/* allocVectors */\n", mode->Fs, mdctSize);
fprintf(file, "NULL,\t/* bits */\n");
diff --git a/libcelt/modes.c b/libcelt/modes.c
index d1512d03d..5d990b60a 100644
--- a/libcelt/modes.c
+++ b/libcelt/modes.c
@@ -334,7 +334,6 @@ CELTMode *celt_mode_create(celt_int32 Fs, int frame_size, int *error)
goto failure;
mode->marker_start = MODEPARTIAL;
mode->Fs = Fs;
- mode->ePredCoef = QCONST16(.8f,15);
/* Pre/de-emphasis depends on sampling rate. The "standard" pre-emphasis
is defined as A(z) = 1 - 0.85*z^-1 at 48 kHz. Other rates should
diff --git a/libcelt/modes.h b/libcelt/modes.h
index b24034695..44be1d818 100644
--- a/libcelt/modes.h
+++ b/libcelt/modes.h
@@ -44,7 +44,7 @@
#define MAX_CONFIG_SIZES 5
-#define CELT_BITSTREAM_VERSION 0x8000000c
+#define CELT_BITSTREAM_VERSION 0x8000000d
#ifdef STATIC_MODES
#include "static_modes.h"
@@ -52,14 +52,6 @@
#define MAX_PERIOD 1024
-#ifndef MCHANNELS
-# ifdef DISABLE_STEREO
-# define MCHANNELS(mode) (1)
-# else
-# define MCHANNELS(mode) ((mode)->nbChannels)
-# endif
-#endif
-
#ifndef CHANNELS
# ifdef DISABLE_STEREO
# define CHANNELS(_C) (1)
@@ -90,8 +82,6 @@ struct CELTMode {
celt_word16 preemph[4];
const celt_int16 *eBands; /**< Definition for each "pseudo-critical band" */
- celt_word16 ePredCoef;/**< Prediction coefficient for the energy encoding */
-
int nbAllocVectors; /**< Number of lines in the matrix below */
const unsigned char *allocVectors; /**< Number of bits in each band for several rates */
diff --git a/libcelt/quant_bands.c b/libcelt/quant_bands.c
index 4a7abf974..9e4529217 100644
--- a/libcelt/quant_bands.c
+++ b/libcelt/quant_bands.c
@@ -42,9 +42,16 @@
#include "mathops.h"
#include "stack_alloc.h"
-#define E_MEANS_SIZE (5)
+#define E_MEANS_SIZE (3)
-const celt_word16 eMeans[E_MEANS_SIZE] = {QCONST16(7.5f,DB_SHIFT), -QCONST16(1.33f,DB_SHIFT), -QCONST16(2.f,DB_SHIFT), -QCONST16(0.42f,DB_SHIFT), QCONST16(0.17f,DB_SHIFT)};
+static const celt_word16 eMeans[E_MEANS_SIZE] = {QCONST16(7.5f,DB_SHIFT), -QCONST16(1.f,DB_SHIFT), -QCONST16(.5f,DB_SHIFT)};
+
+/* prediction coefficients: 0.9, 0.8, 0.65, 0.5 */
+#ifdef FIXED_POINT
+static const celt_word16 pred_coef[4] = {29440, 26112, 21248, 16384};
+#else
+static const celt_word16 pred_coef[4] = {29440/32768., 26112/32768., 21248/32768., 16384/32768.};
+#endif
/* FIXME: Implement for stereo */
int intra_decision(celt_word16 *eBands, celt_word16 *oldEBands, int len)
@@ -68,12 +75,12 @@ int *quant_prob_alloc(const CELTMode *m)
return NULL;
for (i=0;i<m->nbEBands;i++)
{
- prob[2*i] = 6000-i*200;
+ prob[2*i] = 7000-i*200;
prob[2*i+1] = ec_laplace_get_start_freq(prob[2*i]);
}
for (i=0;i<m->nbEBands;i++)
{
- prob[2*m->nbEBands+2*i] = 9000-i*240;
+ prob[2*m->nbEBands+2*i] = 9000-i*220;
prob[2*m->nbEBands+2*i+1] = ec_laplace_get_start_freq(prob[2*m->nbEBands+2*i]);
}
return prob;
@@ -84,22 +91,23 @@ void quant_prob_free(int *freq)
celt_free(freq);
}
-void quant_coarse_energy(const CELTMode *m, int start, int end, const celt_word16 *eBands, celt_word16 *oldEBands, int budget, int intra, int *prob, celt_word16 *error, ec_enc *enc, int _C, celt_word16 max_decay)
+void quant_coarse_energy(const CELTMode *m, int start, int end, const celt_word16 *eBands, celt_word16 *oldEBands, int budget, int intra, int *prob, celt_word16 *error, ec_enc *enc, int _C, int LM, celt_word16 max_decay)
{
int i, c;
celt_word32 prev[2] = {0,0};
- celt_word16 coef = m->ePredCoef;
+ celt_word16 coef;
celt_word16 beta;
const int C = CHANNELS(_C);
+ coef = pred_coef[LM];
+
if (intra)
{
coef = 0;
prob += 2*m->nbEBands;
}
- /* The .8 is a heuristic */
- beta = MULT16_16_P15(QCONST16(.8f,15),coef);
-
+ /* No theoretical justification for this, it just works */
+ beta = MULT16_16_P15(coef,coef);
/* Encode at a fixed coarse resolution */
for (i=start;i<end;i++)
{
@@ -228,21 +236,23 @@ void quant_energy_finalise(const CELTMode *m, int start, int end, celt_ener *eBa
} while (++c < C);
}
-void unquant_coarse_energy(const CELTMode *m, int start, int end, celt_ener *eBands, celt_word16 *oldEBands, int intra, int *prob, ec_dec *dec, int _C)
+void unquant_coarse_energy(const CELTMode *m, int start, int end, celt_ener *eBands, celt_word16 *oldEBands, int intra, int *prob, ec_dec *dec, int _C, int LM)
{
int i, c;
celt_word32 prev[2] = {0, 0};
- celt_word16 coef = m->ePredCoef;
+ celt_word16 coef;
celt_word16 beta;
const int C = CHANNELS(_C);
+ coef = pred_coef[LM];
+
if (intra)
{
coef = 0;
prob += 2*m->nbEBands;
}
- /* The .8 is a heuristic */
- beta = MULT16_16_P15(QCONST16(.8f,15),coef);
+ /* No theoretical justification for this, it just works */
+ beta = MULT16_16_P15(coef,coef);
/* Decode at a fixed coarse resolution */
for (i=start;i<end;i++)
diff --git a/libcelt/quant_bands.h b/libcelt/quant_bands.h
index ab61f7ace..722a5343a 100644
--- a/libcelt/quant_bands.h
+++ b/libcelt/quant_bands.h
@@ -56,13 +56,13 @@ void compute_fine_allocation(const CELTMode *m, int *bits, int budget);
int intra_decision(celt_word16 *eBands, celt_word16 *oldEBands, int len);
-void quant_coarse_energy(const CELTMode *m, int start, int end, const celt_word16 *eBands, celt_word16 *oldEBands, int budget, int intra, int *prob, celt_word16 *error, ec_enc *enc, int _C, celt_word16 max_decay);
+void quant_coarse_energy(const CELTMode *m, int start, int end, const celt_word16 *eBands, celt_word16 *oldEBands, int budget, int intra, int *prob, celt_word16 *error, ec_enc *enc, int _C, int LM, celt_word16 max_decay);
void quant_fine_energy(const CELTMode *m, int start, int end, celt_ener *eBands, celt_word16 *oldEBands, celt_word16 *error, int *fine_quant, ec_enc *enc, int _C);
void quant_energy_finalise(const CELTMode *m, int start, int end, celt_ener *eBands, celt_word16 *oldEBands, celt_word16 *error, int *fine_quant, int *fine_priority, int bits_left, ec_enc *enc, int _C);
-void unquant_coarse_energy(const CELTMode *m, int start, int end, celt_ener *eBands, celt_word16 *oldEBands, int intra, int *prob, ec_dec *dec, int _C);
+void unquant_coarse_energy(const CELTMode *m, int start, int end, celt_ener *eBands, celt_word16 *oldEBands, int intra, int *prob, ec_dec *dec, int _C, int LM);
void unquant_fine_energy(const CELTMode *m, int start, int end, celt_ener *eBands, celt_word16 *oldEBands, int *fine_quant, ec_dec *dec, int _C);
--
GitLab