Propagating perceptual weighting around (not used yet).

This commit is contained in:
Jean-Marc Valin 2007-12-14 13:47:04 +11:00
parent 36d78e6f74
commit 46014ca49a
6 changed files with 21 additions and 12 deletions

View file

@ -214,7 +214,7 @@ void pitch_quant_bands(const CELTMode *m, float *X, float *P, float *gains)
P[i] = 0; P[i] = 0;
} }
void quant_bands(const CELTMode *m, float *X, float *P, ec_enc *enc) void quant_bands(const CELTMode *m, float *X, float *P, float *W, ec_enc *enc)
{ {
int i, j, B; int i, j, B;
const int *eBands = m->eBands; const int *eBands = m->eBands;
@ -227,14 +227,14 @@ void quant_bands(const CELTMode *m, float *X, float *P, ec_enc *enc)
q = m->nbPulses[i]; q = m->nbPulses[i];
if (q>0) { if (q>0) {
float n = sqrt(B*(eBands[i+1]-eBands[i])); float n = sqrt(B*(eBands[i+1]-eBands[i]));
alg_quant(X+B*eBands[i], B*(eBands[i+1]-eBands[i]), q, P+B*eBands[i], 0.7, enc); alg_quant(X+B*eBands[i], W+B*eBands[i], B*(eBands[i+1]-eBands[i]), q, P+B*eBands[i], 0.7, enc);
for (j=B*eBands[i];j<B*eBands[i+1];j++) for (j=B*eBands[i];j<B*eBands[i+1];j++)
norm[j] = X[j] * n; norm[j] = X[j] * n;
//printf ("%f ", log2(ncwrs64(B*(eBands[i+1]-eBands[i]), q))/(B*(eBands[i+1]-eBands[i]))); //printf ("%f ", log2(ncwrs64(B*(eBands[i+1]-eBands[i]), q))/(B*(eBands[i+1]-eBands[i])));
//printf ("%f ", log2(ncwrs64(B*(eBands[i+1]-eBands[i]), q))); //printf ("%f ", log2(ncwrs64(B*(eBands[i+1]-eBands[i]), q)));
} else { } else {
float n = sqrt(B*(eBands[i+1]-eBands[i])); float n = sqrt(B*(eBands[i+1]-eBands[i]));
copy_quant(X+B*eBands[i], B*(eBands[i+1]-eBands[i]), -q, norm, B, eBands[i], enc); copy_quant(X+B*eBands[i], W+B*eBands[i], B*(eBands[i+1]-eBands[i]), -q, norm, B, eBands[i], enc);
for (j=B*eBands[i];j<B*eBands[i+1];j++) for (j=B*eBands[i];j<B*eBands[i+1];j++)
norm[j] = X[j] * n; norm[j] = X[j] * n;
//printf ("%f ", (1+log2(eBands[i]-(eBands[i+1]-eBands[i]))+log2(ncwrs64(B*(eBands[i+1]-eBands[i]), -q)))/(B*(eBands[i+1]-eBands[i]))); //printf ("%f ", (1+log2(eBands[i]-(eBands[i+1]-eBands[i]))+log2(ncwrs64(B*(eBands[i+1]-eBands[i]), -q)))/(B*(eBands[i+1]-eBands[i])));

View file

@ -47,7 +47,7 @@ void compute_pitch_gain(const CELTMode *m, float *X, float *P, float *gains, flo
void pitch_quant_bands(const CELTMode *m, float *X, float *P, float *gains); void pitch_quant_bands(const CELTMode *m, float *X, float *P, float *gains);
void quant_bands(const CELTMode *m, float *X, float *P, ec_enc *enc); void quant_bands(const CELTMode *m, float *X, float *P, float *W, ec_enc *enc);
void unquant_bands(const CELTMode *m, float *X, float *P, ec_dec *dec); void unquant_bands(const CELTMode *m, float *X, float *P, ec_dec *dec);

View file

@ -40,6 +40,7 @@
#include "probenc.h" #include "probenc.h"
#include "quant_pitch.h" #include "quant_pitch.h"
#include "quant_bands.h" #include "quant_bands.h"
#include "psy.h"
#define MAX_PERIOD 1024 #define MAX_PERIOD 1024
@ -50,6 +51,7 @@ struct CELTEncoder {
int block_size; int block_size;
int nb_blocks; int nb_blocks;
int channels; int channels;
int Fs;
ec_byte_buffer buf; ec_byte_buffer buf;
ec_enc enc; ec_enc enc;
@ -83,6 +85,7 @@ CELTEncoder *celt_encoder_new(const CELTMode *mode)
st->frame_size = B*N; st->frame_size = B*N;
st->block_size = N; st->block_size = N;
st->nb_blocks = B; st->nb_blocks = B;
st->Fs = 44100;
ec_byte_writeinit(&st->buf); ec_byte_writeinit(&st->buf);
ec_enc_init(&st->enc,&st->buf); ec_enc_init(&st->enc,&st->buf);
@ -207,6 +210,7 @@ int celt_encode(CELTEncoder *st, short *pcm)
float X[B*C*N]; /**< Interleaved signal MDCTs */ float X[B*C*N]; /**< Interleaved signal MDCTs */
float P[B*C*N]; /**< Interleaved pitch MDCTs*/ float P[B*C*N]; /**< Interleaved pitch MDCTs*/
float mask[B*C*N]; /**< Masking curve */
float bandE[st->mode->nbEBands]; float bandE[st->mode->nbEBands];
float gains[st->mode->nbPBands]; float gains[st->mode->nbPBands];
int pitch_index; int pitch_index;
@ -228,6 +232,11 @@ int celt_encode(CELTEncoder *st, short *pcm)
/* Compute MDCTs */ /* Compute MDCTs */
compute_mdcts(&st->mdct_lookup, st->window, in, X, N, B, C); compute_mdcts(&st->mdct_lookup, st->window, in, X, N, B, C);
compute_masking(X, mask, B*C*N, st->Fs);
/* Invert and stretch the mask to length of X */
for (i=B*C*N-1;i>=0;i--)
mask[i] = 1/(1+mask[i>>1]);
/* Pitch analysis */ /* Pitch analysis */
for (c=0;c<C;c++) for (c=0;c<C;c++)
{ {
@ -286,7 +295,7 @@ int celt_encode(CELTEncoder *st, short *pcm)
sum += X[i]*X[i]; sum += X[i]*X[i];
printf ("%f\n", sum);*/ printf ("%f\n", sum);*/
/* Residual quantisation */ /* Residual quantisation */
quant_bands(st->mode, X, P, &st->enc); quant_bands(st->mode, X, P, mask, &st->enc);
if (0) {//This is just for debugging if (0) {//This is just for debugging
ec_enc_done(&st->enc); ec_enc_done(&st->enc);

View file

@ -52,8 +52,8 @@ int main(int argc, char *argv[])
outFile = argv[2]; outFile = argv[2];
fout = fopen(outFile, "wb+"); fout = fopen(outFile, "wb+");
enc = celt_encoder_new(celt_mode2); enc = celt_encoder_new(celt_mode1);
dec = celt_decoder_new(celt_mode2); dec = celt_decoder_new(celt_mode1);
while (!feof(fin)) while (!feof(fin))
{ {

View file

@ -38,7 +38,7 @@
/* Improved algebraic pulse-base quantiser. The signal x is replaced by the sum of the pitch /* Improved algebraic pulse-base quantiser. The signal x is replaced by the sum of the pitch
a combination of pulses such that its norm is still equal to 1. The only difference with a combination of pulses such that its norm is still equal to 1. The only difference with
the quantiser above is that the search is more complete. */ the quantiser above is that the search is more complete. */
void alg_quant(float *x, int N, int K, float *p, float alpha, ec_enc *enc) void alg_quant(float *x, float *W, int N, int K, float *p, float alpha, ec_enc *enc)
{ {
int L = 3; int L = 3;
//float tata[200]; //float tata[200];
@ -223,7 +223,7 @@ void alg_quant(float *x, int N, int K, float *p, float alpha, ec_enc *enc)
static const float pg[5] = {1.f, .6f, .45f, 0.35f, 0.25f}; static const float pg[5] = {1.f, .6f, .45f, 0.35f, 0.25f};
/* Finds the right offset into Y and copy it */ /* Finds the right offset into Y and copy it */
void copy_quant(float *x, int N, int K, float *Y, int B, int N0, ec_enc *enc) void copy_quant(float *x, float *W, int N, int K, float *Y, int B, int N0, ec_enc *enc)
{ {
int i,j; int i,j;
int best=0; int best=0;
@ -287,7 +287,7 @@ void copy_quant(float *x, int N, int K, float *Y, int B, int N0, ec_enc *enc)
E = .8/sqrt(E); E = .8/sqrt(E);
for (j=0;j<N;j++) for (j=0;j<N;j++)
P[j] *= E; P[j] *= E;
alg_quant(x, N, K, P, 0, enc); alg_quant(x, W, N, K, P, 0, enc);
} }
} }

View file

@ -39,12 +39,12 @@
/* Algebraic pulse-base quantiser. The signal x is replaced by the sum of the pitch /* Algebraic pulse-base quantiser. The signal x is replaced by the sum of the pitch
a combination of pulses such that its norm is still equal to 1. The only difference with a combination of pulses such that its norm is still equal to 1. The only difference with
the quantiser above is that the search is more complete. */ the quantiser above is that the search is more complete. */
void alg_quant(float *x, int N, int K, float *p, float alpha, ec_enc *enc); void alg_quant(float *x, float *W, int N, int K, float *p, float alpha, ec_enc *enc);
void alg_unquant(float *x, int N, int K, float *p, float alpha, ec_dec *dec); void alg_unquant(float *x, int N, int K, float *p, float alpha, ec_dec *dec);
/* Finds the right offset into Y and copy it */ /* Finds the right offset into Y and copy it */
void copy_quant(float *x, int N, int K, float *Y, int B, int N0, ec_enc *enc); void copy_quant(float *x, float *W, int N, int K, float *Y, int B, int N0, ec_enc *enc);
void copy_unquant(float *x, int N, int K, float *Y, int B, int N0, ec_dec *dec); void copy_unquant(float *x, int N, int K, float *Y, int B, int N0, ec_dec *dec);
#endif /* VQ_H */ #endif /* VQ_H */