#include "ra288.h"
typedef struct {
- float history[8];
- float output[40];
- float pr1[36];
- float pr2[10];
- int phase;
-
- float st1a[111], st1b[37], st1[37];
- float st2a[38], st2b[11], st2[11];
- float sb[41];
- float lhist[10];
-} Real288_internal;
-
-static inline float scalar_product_float(float * v1, float * v2, int size)
+ float sp_lpc[36]; ///< LPC coefficients for speech data (spec: A)
+ float gain_lpc[10]; ///< LPC coefficients for gain (spec: GB)
+
+ float sp_hist[111]; ///< Speech data history (spec: SB)
+
+ /** Speech part of the gain autocorrelation (spec: REXP) */
+ float sp_rec[37];
+
+ float gain_hist[38]; ///< Log-gain history (spec: SBLG)
+
+ /** Recursive part of the gain autocorrelation (spec: REXPLG) */
+ float gain_rec[11];
+
+ float sp_block[41]; ///< Speech data of four blocks (spec: STTMP)
+ float gain_block[10]; ///< Gain data of four blocks (spec: GSTATE)
+} RA288Context;
+
+static av_cold int ra288_decode_init(AVCodecContext *avctx)
+{
+ avctx->sample_fmt = SAMPLE_FMT_S16;
+ return 0;
+}
+
+static inline float scalar_product_float(const float * v1, const float * v2,
+ int size)
{
float res = 0.;
return res;
}
-/* Decode and produce output */
-static void decode(Real288_internal *glob, float gain, int cb_coef)
+static void colmult(float *tgt, const float *m1, const float *m2, int n)
{
- int x, y;
+ while (n--)
+ *tgt++ = *m1++ * *m2++;
+}
+
+static void decode(RA288Context *ractx, float gain, int cb_coef)
+{
+ int i, j;
double sumsum;
float sum, buffer[5];
- memmove(glob->sb + 5, glob->sb, 36 * sizeof(*glob->sb));
+ memmove(ractx->sp_block + 5, ractx->sp_block, 36*sizeof(*ractx->sp_block));
- for (x=4; x >= 0; x--)
- glob->sb[x] = -scalar_product_float(glob->sb + x + 1, glob->pr1, 36);
+ for (i=4; i >= 0; i--)
+ ractx->sp_block[i] = -scalar_product_float(ractx->sp_block + i + 1,
+ ractx->sp_lpc, 36);
- /* convert log and do rms */
- sum = 32. - scalar_product_float(glob->pr2, glob->lhist, 10);
+ /* block 46 of G.728 spec */
+ sum = 32. - scalar_product_float(ractx->gain_lpc, ractx->gain_block, 10);
+ /* block 47 of G.728 spec */
sum = av_clipf(sum, 0, 60);
- sumsum = exp(sum * 0.1151292546497) * gain; /* pow(10.0,sum/20)*f */
+ /* block 48 of G.728 spec */
+ sumsum = exp(sum * 0.1151292546497) * gain; /* pow(10.0,sum/20)*gain */
- for (x=0; x < 5; x++)
- buffer[x] = codetable[cb_coef][x] * sumsum;
+ for (i=0; i < 5; i++)
+ buffer[i] = codetable[cb_coef][i] * sumsum;
sum = scalar_product_float(buffer, buffer, 5) / 5;
sum = FFMAX(sum, 1);
/* shift and store */
- memmove(glob->lhist, glob->lhist - 1, 10 * sizeof(*glob->lhist));
+ memmove(ractx->gain_block, ractx->gain_block - 1,
+ 10 * sizeof(*ractx->gain_block));
- *glob->lhist = glob->history[glob->phase] = 10 * log10(sum) - 32;
+ *ractx->gain_block = 10 * log10(sum) - 32;
- for (x=1; x < 5; x++)
- for (y=x-1; y >= 0; y--)
- buffer[x] -= glob->pr1[x-y-1] * buffer[y];
+ for (i=1; i < 5; i++)
+ for (j=i-1; j >= 0; j--)
+ buffer[i] -= ractx->sp_lpc[i-j-1] * buffer[j];
/* output */
- for (x=0; x < 5; x++) {
- glob->output[glob->phase*5+x] = glob->sb[4-x] =
- av_clipf(glob->sb[4-x] + buffer[x], -4095, 4095);
- }
+ for (i=0; i < 5; i++)
+ ractx->sp_block[4-i] =
+ av_clipf(ractx->sp_block[4-i] + buffer[i], -4095, 4095);
}
-/* column multiply */
-static void colmult(float *tgt, float *m1, const float *m2, int n)
-{
- while (n--)
- *(tgt++) = (*(m1++)) * (*(m2++));
-}
-
-static int pred(float *in, float *tgt, int n)
+/**
+ * Converts autocorrelation coefficients to LPC coefficients using the
+ * Levinson-Durbin algorithm. See blocks 37 and 50 of the G.728 specification.
+ *
+ * @return 0 if success, -1 if fail
+ */
+static int eval_lpc_coeffs(const float *in, float *tgt, int n)
{
- int x, y;
+ int i, j;
double f0, f1, f2;
if (in[n] == 0)
- return 0;
+ return -1;
if ((f0 = *in) <= 0)
- return 0;
+ return -1;
- for (x=1 ; ; x++) {
- float *p1 = in + x;
- float *p2 = tgt;
+ in--; // To avoid a -1 subtraction in the inner loop
- if (n < x)
- return 1;
+ for (i=1; i <= n; i++) {
+ f1 = in[i+1];
- f1 = *(p1--);
+ for (j=0; j < i - 1; j++)
+ f1 += in[i-j]*tgt[j];
- for (y=0; y < x - 1; y++)
- f1 += (*(p1--))*(*(p2++));
-
- p1 = tgt + x - 1;
- p2 = tgt;
- *(p1--) = f2 = -f1/f0;
- for (y=x >> 1; y--;) {
- float temp = *p2 + *p1 * f2;
- *(p1--) += *p2 * f2;
- *(p2++) = temp;
+ tgt[i-1] = f2 = -f1/f0;
+ for (j=0; j < i >> 1; j++) {
+ float temp = tgt[j] + tgt[i-j-2]*f2;
+ tgt[i-j-2] += tgt[j]*f2;
+ tgt[j] = temp;
}
if ((f0 += f1*f2) < 0)
- return 0;
+ return -1;
}
+
+ return 0;
}
-/* product sum (lsf) */
-static void prodsum(float *tgt, float *src, int len, int n)
+static void prodsum(float *tgt, const float *src, int len, int n)
{
for (; n >= 0; n--)
tgt[n] = scalar_product_float(src, src - n, len);
}
-static void co(int n, int i, int j, float *in, float *out, float *st1,
- float *st2, const float *table)
+/**
+ * Hybrid window filtering. See blocks 36 and 49 of the G.728 specification.
+ *
+ * @note This function is slightly different from that described in the spec.
+ * It expects in[0] to be the newest sample and in[n-1] to be the oldest
+ * one stored. The spec has in the more ordinary way (in[0] the oldest
+ * and in[n-1] the newest).
+ *
+ * @param order the order of the filter
+ * @param n the length of the input
+ * @param non_rec the number of non-recursive samples
+ * @param out the filter output
+ * @param in pointer to the input of the filter
+ * @param hist pointer to the input history of the filter. It is updated by
+ * this function.
+ * @param out pointer to the non-recursive part of the output
+ * @param out2 pointer to the recursive part of the output
+ * @param window pointer to the windowing function table
+ */
+static void do_hybrid_window(int order, int n, int non_rec, const float *in,
+ float *out, float *hist, float *out2,
+ const float *window)
{
- int a, b, c;
- unsigned int x;
- float *fp;
- float buffer1[37];
- float buffer2[37];
- float work[111];
-
- /* rotate and multiply */
- c = (b = (a = n + i) + j) - i;
- fp = st1 + i;
- for (x=0; x < b; x++) {
- if (x == c)
- fp=in;
- work[x] = *(table++) * (*(st1++) = *(fp++));
- }
+ int i;
+ float buffer1[order + 1];
+ float buffer2[order + 1];
+ float work[order + n + non_rec];
- prodsum(buffer1, work + n, i, n);
- prodsum(buffer2, work + a, j, n);
+ /* update history */
+ memmove(hist, hist + n, (order + non_rec)*sizeof(*hist));
- for (x=0;x<=n;x++) {
- *st2 = *st2 * (0.5625) + buffer1[x];
- out[x] = *(st2++) + buffer2[x];
+ for (i=0; i < n; i++)
+ hist[order + non_rec + i] = in[n-i-1];
+
+ colmult(work, window, hist, order + n + non_rec);
+
+ prodsum(buffer1, work + order , n , order);
+ prodsum(buffer2, work + order + n, non_rec, order);
+
+ for (i=0; i <= order; i++) {
+ out2[i] = out2[i] * 0.5625 + buffer1[i];
+ out [i] = out2[i] + buffer2[i];
}
- *out *= 1.00390625; /* to prevent clipping */
+
+ /* Multiply by the white noise correcting factor (WNCF) */
+ *out *= 257./256.;
}
-static void update(Real288_internal *glob)
+/**
+ * Backward synthesis filter. Find the LPC coefficients from past speech data.
+ */
+static void backward_filter(RA288Context *ractx)
{
- float buffer1[40], temp1[37];
- float buffer2[8], temp2[11];
+ float temp1[37]; // RTMP in the spec
+ float temp2[11]; // GPTPMP in the spec
- memcpy(buffer1 , glob->output + 20, 20*sizeof(*buffer1));
- memcpy(buffer1 + 20, glob->output , 20*sizeof(*buffer1));
+ do_hybrid_window(36, 40, 35, ractx->sp_block, temp1, ractx->sp_hist,
+ ractx->sp_rec, syn_window);
- co(36, 40, 35, buffer1, temp1, glob->st1a, glob->st1b, table1);
+ if (!eval_lpc_coeffs(temp1, ractx->sp_lpc, 36))
+ colmult(ractx->sp_lpc, ractx->sp_lpc, syn_bw_tab, 36);
- if (pred(temp1, glob->st1, 36))
- colmult(glob->pr1, glob->st1, table1a, 36);
+ do_hybrid_window(10, 8, 20, ractx->gain_block, temp2, ractx->gain_hist,
+ ractx->gain_rec, gain_window);
- memcpy(buffer2 , glob->history + 4, 4*sizeof(*buffer2));
- memcpy(buffer2 + 4, glob->history , 4*sizeof(*buffer2));
-
- co(10, 8, 20, buffer2, temp2, glob->st2a, glob->st2b, table2);
-
- if (pred(temp2, glob->st2, 10))
- colmult(glob->pr2, glob->st2, table2a, 10);
+ if (!eval_lpc_coeffs(temp2, ractx->gain_lpc, 10))
+ colmult(ractx->gain_lpc, ractx->gain_lpc, gain_bw_tab, 10);
}
-/* Decode a block (celp) */
static int ra288_decode_frame(AVCodecContext * avctx, void *data,
int *data_size, const uint8_t * buf,
int buf_size)
{
int16_t *out = data;
- int x, y;
- Real288_internal *glob = avctx->priv_data;
+ int i, j;
+ RA288Context *ractx = avctx->priv_data;
GetBitContext gb;
if (buf_size < avctx->block_align) {
init_get_bits(&gb, buf, avctx->block_align * 8);
- for (x=0; x < 32; x++) {
+ for (i=0; i < 32; i++) {
float gain = amptable[get_bits(&gb, 3)];
- int cb_coef = get_bits(&gb, 6 + (x&1));
- glob->phase = x & 7;
- decode(glob, gain, cb_coef);
+ int cb_coef = get_bits(&gb, 6 + (i&1));
- for (y=0; y < 5; y++)
- *(out++) = 8 * glob->output[glob->phase*5 + y];
+ decode(ractx, gain, cb_coef);
- if (glob->phase == 3)
- update(glob);
+ for (j=0; j < 5; j++)
+ *(out++) = 8 * ractx->sp_block[4 - j];
+
+ if ((i & 7) == 3)
+ backward_filter(ractx);
}
*data_size = (char *)out - (char *)data;
"real_288",
CODEC_TYPE_AUDIO,
CODEC_ID_RA_288,
- sizeof(Real288_internal),
- NULL,
+ sizeof(RA288Context),
+ ra288_decode_init,
NULL,
NULL,
ra288_decode_frame,