shithub: opus

--- a/libcelt/bands.c

+++ b/libcelt/bands.c

@@ -36,7 +36,8 @@

 #include "cwrs.h"

 /* Applies a series of rotations so that pulses are spread like a two-sided

-exponential */

+exponential. The effect of this is to reduce the tonal noise created by the

+sparse spectrum resulting from the pulse codebook */

 static void exp_rotation(float *X, int len, float theta, int dir, int stride, int iter)

    int i, k;

@@ -87,7 +88,7 @@

-/* Compute the energy in each of the bands */

+/* Compute the amplitude (sqrt energy) in each of the bands */

 void compute_band_energies(const CELTMode *m, float *X, float *bank)

    int i, B;

@@ -197,6 +198,7 @@

       P[i] = 0;

+/* Quantisation of the residual */

 void quant_bands(const CELTMode *m, float *X, float *P, float *W, ec_enc *enc)

    int i, j, B;

@@ -234,6 +236,7 @@

       X[i] = 0;

+/* Decoding of the residual */

 void unquant_bands(const CELTMode *m, float *X, float *P, ec_dec *dec)

    int i, j, B;

--- a/libcelt/bands.h

+++ b/libcelt/bands.h

@@ -37,18 +37,54 @@

 #include "entenc.h"

 #include "entdec.h"

+/** Compute the amplitude (sqrt energy) in each of the bands

+ * @param m Mode data

+ * @param X Spectrum

+ * @param bands Square root of the energy for each band (returned)

+ */

 void compute_band_energies(const CELTMode *m, float *X, float *bands);

+/** Normalise each band of X such that the energy in each band is

+    equal to 1

+ * @param m Mode data

+ * @param X Spectrum (returned normalised)

+ * @param bands Square root of the energy for each band

+ */

 void normalise_bands(const CELTMode *m, float *X, float *bands);

+/** Denormalise each band of X to restore full amplitude

+ * @param m Mode data

+ * @param X Spectrum (returned de-normalised)

+ * @param bands Square root of the energy for each band

+ */

 void denormalise_bands(const CELTMode *m, float *X, float *bands);

+/** Compute the pitch predictor gain for each pitch band

+ * @param m Mode data

+ * @param X Spectrum to predict

+ * @param P Pitch vector (normalised)

+ * @param gains Gain computed for each pitch band (returned)

+ * @param bank Square root of the energy for each band

+ */

 void compute_pitch_gain(const CELTMode *m, float *X, float *P, float *gains, float *bank);

 void pitch_quant_bands(const CELTMode *m, float *X, float *P, float *gains);

+/** Quantisation/encoding of the residual spectrum

+ * @param m Mode data

+ * @param X Residual (normalised)

+ * @param P Pitch vector (normalised)

+ * @param W Perceptual weighting

+ * @param enc Entropy encoder

+ */

 void quant_bands(const CELTMode *m, float *X, float *P, float *W, ec_enc *enc);

+/** Decoding of the residual spectrum

+ * @param m Mode data

+ * @param X Residual (normalised)

+ * @param P Pitch vector (normalised)

+ * @param dec Entropy decoder

+*/

 void unquant_bands(const CELTMode *m, float *X, float *P, ec_dec *dec);

 #endif /* BANDS_H */

--- a/libcelt/celt.c

+++ b/libcelt/celt.c

@@ -282,6 +282,7 @@

       haar1(P, B*N*C, 1);

+   /* Get a tiny bit more frequency resolution and prevent unstable energy when quantising */

    time_dct(X, N, B, C);

    time_dct(P, N, B, C);

@@ -289,18 +290,16 @@

    compute_band_energies(st->mode, X, bandE);

    normalise_bands(st->mode, X, bandE);

    //for (i=0;i<st->mode->nbEBands;i++)printf("%f ", bandE[i]);printf("\n");

+   /* Normalise the pitch vector as well (discard the energies) */

       float bandEp[st->mode->nbEBands];

       compute_band_energies(st->mode, P, bandEp);

       normalise_bands(st->mode, P, bandEp);

-   //band_rotation(st->mode, X, -1);

-   //band_rotation(st->mode, P, -1);

    quant_energy(st->mode, bandE, st->oldBandE, &st->enc);

    /* Pitch prediction */

    compute_pitch_gain(st->mode, X, P, gains, bandE);

    quant_pitch(gains, st->mode->nbPBands, &st->enc);

@@ -307,7 +306,7 @@

    pitch_quant_bands(st->mode, X, P, gains);

    //for (i=0;i<B*N;i++) printf("%f ",P[i]);printf("\n");

-   /* Subtract the pitch prediction from the signal to encode */

+   /* Compute residual that we're going to encode */

    for (i=0;i<B*C*N;i++)

       X[i] -= P[i];

@@ -317,19 +316,7 @@

    printf ("%f\n", sum);*/

    /* Residual quantisation */

    quant_bands(st->mode, X, P, mask, &st->enc);

-   if (0) {//This is just for debugging

-      ec_enc_done(&st->enc);

-      ec_dec dec;

-      ec_byte_readinit(&st->buf,ec_byte_get_buffer(&st->buf),ec_byte_bytes(&st->buf));

-      ec_dec_init(&dec,&st->buf);

-      unquant_bands(st->mode, X, P, &dec);

-      //printf ("\n");

-   }

-   //band_rotation(st->mode, X, 1);

    /* Synthesis */

    denormalise_bands(st->mode, X, bandE);

@@ -336,11 +323,11 @@

    time_idct(X, N, B, C);

    if (C==2)

       haar1(X, B*N*C, 1);

    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, B, C);

+   compute_inv_mdcts(&st->mdct_lookup, st->window, X, st->out_mem, st->mdct_overlap, N, B, C);

+   /* De-emphasis and put everything back at the right place in the synthesis history */

    for (c=0;c<C;c++)

       for (i=0;i<B;i++)

@@ -532,7 +519,6 @@

       compute_band_energies(st->mode, P, bandEp);

       normalise_bands(st->mode, P, bandEp);

-   //band_rotation(st->mode, P, -1);

    /* Get the pitch gains */

    unquant_pitch(gains, st->mode->nbPBands, &dec);

@@ -543,8 +529,6 @@

    /* Decode fixed codebook and merge with pitch */

    unquant_bands(st->mode, X, P, &dec);

-   //band_rotation(st->mode, X, 1);

    /* Synthesis */

    denormalise_bands(st->mode, X, bandE);

--- a/libcelt/vq.h

+++ b/libcelt/vq.h

@@ -36,13 +36,41 @@

 #include "entdec.h"

-/* Algebraic pulse-based 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

-   the quantiser above is that the search is more complete. */

+/** Algebraic pulse-based 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 the quantiser above is that the search is more complete.

+ * @param x Residual signal to quantise/encode (returns quantised version)

+ * @param W Perceptual weight

+ * @param N Number of samples to encode

+ * @param K Number of pulses to use

+ * @param p Pitch vector (it is assumed that p+x is a unit vector)

+ * @param alpha compression factor in the pitch direction (magic!)

+ * @param enc Entropy encoder state

+*/

 void alg_quant(float *x, float *W, int N, int K, float *p, float alpha, ec_enc *enc);

+/** Algebraic pulse decoder

+ * @param x Decoded normalised spectrum (returned)

+ * @param N Number of samples to decode

+ * @param K Number of pulses to use

+ * @param p Pitch vector (automatically added to x)

+ * @param alpha compression factor in the pitch direction (magic!)

+ * @param dec Entropy decoder state

+ */

 void alg_unquant(float *x, int N, int K, float *p, float alpha, ec_dec *dec);

+/** Intra-frame predictor that matches a section of the current frame (at lower

+ * frequencies) to encode the current band.

+ * @param x Residual signal to quantise/encode (returns quantised version)

+ * @param W Perceptual weight

+ * @param N Number of samples to encode

+ * @param K Number of pulses to use

+ * @param Y Lower frequency spectrum to use, normalised to the same standard deviation

+ * @param p Pitch vector (it is assumed that p+x is a unit vector)

+ * @param B Stride (number of channels multiplied by the number of MDCTs per frame)

+ * @param N0 Number of valid offsets

+ * @param enc Entropy encoder state

+ */

 void intra_prediction(float *x, float *W, int N, int K, float *Y, float *P, int B, int N0, ec_enc *enc);

 void intra_unquant(float *x, int N, int K, float *Y, float *P, int B, int N0, ec_dec *dec);