gain code, gain pitch and pitch delay decoding for ACELP based codecs
authorVladimir Voroshilov <voroshil@gmail.com>
Mon, 30 Jun 2008 18:03:38 +0000 (18:03 +0000)
committerVladimir Voroshilov <voroshil@gmail.com>
Mon, 30 Jun 2008 18:03:38 +0000 (18:03 +0000)
Originally committed as revision 14037 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/acelp_pitch_delay.c [new file with mode: 0644]
libavcodec/acelp_pitch_delay.h [new file with mode: 0644]

diff --git a/libavcodec/acelp_pitch_delay.c b/libavcodec/acelp_pitch_delay.c
new file mode 100644 (file)
index 0000000..9b35fcc
--- /dev/null
@@ -0,0 +1,119 @@
+/*
+ * gain code, gain pitch and pitch delay decoding
+ *
+ * Copyright (c) 2008 Vladimir Voroshilov
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "avcodec.h"
+#include "acelp_pitch_delay.h"
+#include "acelp_math.h"
+
+int ff_acelp_decode_8bit_to_1st_delay3(int ac_index)
+{
+    ac_index += 58;
+    if(ac_index > 254)
+        ac_index = 3 * ac_index - 510;
+    return ac_index;
+}
+
+int ff_acelp_decode_4bit_to_2nd_delay3(
+        int ac_index,
+        int pitch_delay_min)
+{
+    if(ac_index < 4)
+        return 3 * (ac_index + pitch_delay_min);
+    else if(ac_index < 12)
+        return 3 * pitch_delay_min + ac_index + 6;
+    else
+        return 3 * (ac_index + pitch_delay_min) - 18;
+}
+
+int ff_acelp_decode_5_6_bit_to_2nd_delay3(
+        int ac_index,
+        int pitch_delay_min)
+{
+        return 3 * pitch_delay_min + ac_index - 2;
+}
+
+int ff_acelp_decode_9bit_to_1st_delay6(int ac_index)
+{
+    if(ac_index < 463)
+        return ac_index + 105;
+    else
+        return 6 * (ac_index - 368);
+}
+int ff_acelp_decode_6bit_to_2nd_delay6(
+        int ac_index,
+        int pitch_delay_min)
+{
+    return 6 * pitch_delay_min + ac_index - 3;
+}
+
+void ff_acelp_update_past_gain(
+    int16_t* quant_energy,
+    int gain_corr_factor,
+    int log2_ma_pred_order,
+    int erasure)
+{
+    int i;
+    int avg_gain=quant_energy[(1 << log2_ma_pred_order) - 1]; // (5.10)
+
+    for(i=(1 << log2_ma_pred_order) - 1; i>0; i--)
+    {
+        avg_gain       += quant_energy[i-1];
+        quant_energy[i] = quant_energy[i-1];
+    }
+
+    if(erasure)
+        quant_energy[0] = FFMAX(avg_gain >> log2_ma_pred_order, -10240) - 4096; // -10 and -4 in (5.10)
+    else
+        quant_energy[0] = (6165 * ((ff_log2(gain_corr_factor) >> 2) - (13 << 13))) >> 13;
+}
+
+int16_t ff_acelp_decode_gain_code(
+    int gain_corr_factor,
+    const int16_t* fc_v,
+    int mr_energy,
+    const int16_t* quant_energy,
+    const int16_t* ma_prediction_coeff,
+    int subframe_size,
+    int ma_pred_order)
+{
+    int i;
+
+    mr_energy <<= 10;
+
+    for(i=0; i<ma_pred_order; i++)
+        mr_energy += quant_energy[i] * ma_prediction_coeff[i];
+
+#ifdef G729_BITEXACT
+    mr_energy += (((-6165LL * ff_log2(sum_of_squares(fc_v, subframe_size, 0, 0))) >> 3) & ~0x3ff);
+
+    mr_energy = (5439 * (mr_energy >> 15)) >> 8;           // (0.15) = (0.15) * (7.23)
+
+    return bidir_sal(
+               ((ff_exp2(mr_energy & 0x7fff) + 16) >> 5) * (gain_corr_factor >> 1),
+               (mr_energy >> 15) - 25
+           );
+#else
+    mr_energy = gain_corr_factor * exp(M_LN10 / (20 << 23) * mr_energy) /
+                sqrt(sum_of_squares(fc_v, subframe_size, 0, 0));
+    return mr_energy >> 12;
+#endif
+}
diff --git a/libavcodec/acelp_pitch_delay.h b/libavcodec/acelp_pitch_delay.h
new file mode 100644 (file)
index 0000000..b5b9add
--- /dev/null
@@ -0,0 +1,220 @@
+/*
+ * gain code, gain pitch and pitch delay decoding
+ *
+ * Copyright (c) 2008 Vladimir Voroshilov
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef FFMPEG_ACELP_PITCH_DELAY_H
+#define FFMPEG_ACELP_PITCH_DELAY_H
+
+#include <stdint.h>
+
+#define PITCH_DELAY_MIN             20
+#define PITCH_DELAY_MAX             143
+
+/**
+ * \brief Decode pitch delay of the first subframe encoded by 8 bits with 1/3
+ *        resolution.
+ * \param ac_index adaptive codebook index (8 bits)
+ *
+ * \return pitch delay in 1/3 units
+ *
+ * Pitch delay is coded:
+ *    with 1/3 resolution, 19  < pitch_delay <  85
+ *    integers only,       85 <= pitch_delay <= 143
+ */
+int ff_acelp_decode_8bit_to_1st_delay3(int ac_index);
+
+/**
+ * \brief Decode pitch delay of the second subframe encoded by 5 or 6 bits
+ *        with 1/3 precision.
+ * \param ac_index adaptive codebook index (5 or 6 bits)
+ * \param pitch_delay_min lower bound (integer) of pitch delay interval
+ *                      for second subframe
+ *
+ * \return pitch delay in 1/3 units
+ *
+ * Pitch delay is coded:
+ *    with 1/3 resolution, -6 < pitch_delay - int(prev_pitch_delay) < 5
+ *
+ * \remark The routine is used in G.729 @8k, AMR @10.2k, AMR @7.95k,
+ *         AMR @7.4k for the second subframe.
+ */
+int ff_acelp_decode_5_6_bit_to_2nd_delay3(
+        int ac_index,
+        int pitch_delay_min);
+
+/**
+ * \brief Decode pitch delay with 1/3 precision.
+ * \param ac_index adaptive codebook index (4 bits)
+ * \param pitch_delay_min lower bound (integer) of pitch delay interval for
+ *                      second subframe
+ *
+ * \return pitch delay in 1/3 units
+ *
+ * Pitch delay is coded:
+ *    integers only,          -6  < pitch_delay - int(prev_pitch_delay) <= -2
+ *    with 1/3 resolution,    -2  < pitch_delay - int(prev_pitch_delay) <  1
+ *    integers only,           1 <= pitch_delay - int(prev_pitch_delay) <  5
+ *
+ * \remark The routine is used in G.729 @6.4k, AMR @6.7k, AMR @5.9k,
+ *         AMR @5.15k, AMR @4.75k for the second subframe.
+ */
+int ff_acelp_decode_4bit_to_2nd_delay3(
+        int ac_index,
+        int pitch_delay_min);
+
+/**
+ * \brief Decode pitch delay of the first subframe encoded by 9 bits
+ *        with 1/6 precision.
+ * \param ac_index adaptive codebook index (9 bits)
+ * \param pitch_delay_min lower bound (integer) of pitch delay interval for
+ *                      second subframe
+ *
+ * \return pitch delay in 1/6 units
+ *
+ * Pitch delay is coded:
+ *    with 1/6 resolution,  17  < pitch_delay <  95
+ *    integers only,        95 <= pitch_delay <= 143
+ *
+ * \remark The routine is used in AMR @12.2k for the first and third subframes.
+ */
+int ff_acelp_decode_9bit_to_1st_delay6(int ac_index);
+
+/**
+ * \brief Decode pitch delay of the second subframe encoded by 6 bits
+ *        with 1/6 precision.
+ * \param ac_index adaptive codebook index (6 bits)
+ * \param pitch_delay_min lower bound (integer) of pitch delay interval for
+ *                      second subframe
+ *
+ * \return pitch delay in 1/6 units
+ *
+ * Pitch delay is coded:
+ *    with 1/6 resolution, -6 < pitch_delay - int(prev_pitch_delay) < 5
+ *
+ * \remark The routine is used in AMR @12.2k for the second and fourth subframes.
+ */
+int ff_acelp_decode_6bit_to_2nd_delay6(
+        int ac_index,
+        int pitch_delay_min);
+
+/**
+ * \brief Update past quantized energies
+ * \param quant_energy [in/out] past quantized energies (5.10)
+ * \param gain_corr_factor gain correction factor
+ * \param log2_ma_pred_order log2() of MA prediction order
+ * \param erasure frame erasure flag
+ *
+ * If frame erasure flag is not equal to zero, memory is updated with
+ * averaged energy, attenuated by 4dB:
+ *     max(avg(quant_energy[i])-4, -14), i=0,ma_pred_order
+ *
+ * In normal mode memory is updated with
+ *     Er - Ep = 20 * log10(gain_corr_factor)
+ *
+ * \remark The routine is used in G.729 and AMR (all modes).
+ */
+void ff_acelp_update_past_gain(
+        int16_t* quant_energy,
+        int gain_corr_factor,
+        int log2_ma_pred_order,
+        int erasure);
+
+/**
+ * \brief Decode the adaptive codebook gain and add
+ *        correction (4.1.5 and 3.9.1 of G.729).
+ * \param gain_corr_factor gain correction factor (2.13)
+ * \param fc_v fixed-codebook vector (2.13)
+ * \param mr_energy mean innovation energy and fixed-point correction (7.13)
+ * \param quant_energy [in/out] past quantized energies (5.10)
+ * \param subframe_size length of subframe
+ * \param ma_pred_order MA prediction order
+ *
+ * \return quantized fixed-codebook gain (14.1)
+ *
+ * The routine implements equations 69, 66 and 71 of the G.729 specification (3.9.1)
+ *
+ *    Em   - mean innovation energy (dB, constant, depends on decoding algorithm)
+ *    Ep   - mean-removed predicted energy (dB)
+ *    Er   - mean-removed innovation energy (dB)
+ *    Ei   - mean energy of the fixed-codebook contribution (dB)
+ *    N    - subframe_size
+ *    M    - MA (Moving Average) prediction order
+ *    gc   - fixed-codebook gain
+ *    gc_p - predicted fixed-codebook gain
+ *
+ *    Fixed codebook gain is computed using predicted gain gc_p and
+ *    correction factor gain_corr_factor as shown below:
+ *
+ *        gc = gc_p * gain_corr_factor
+ *
+ *    The predicted fixed codebook gain gc_p is found by predicting
+ *    the energy of the fixed-codebook contribution from the energy
+ *    of previous fixed-codebook contributions.
+ *
+ *        mean = 1/N * sum(i,0,N){ fc_v[i] * fc_v[i] }
+ *
+ *        Ei = 10log(mean)
+ *
+ *        Er = 10log(1/N * gc^2 * mean) - Em = 20log(gc) + Ei - Em
+ *
+ *    Replacing Er with Ep and gc with gc_p we will receive:
+ *
+ *        Ep = 10log(1/N * gc_p^2 * mean) - Em = 20log(gc_p) + Ei - Em
+ *
+ *    and from above:
+ *
+ *        gc_p = 10^((Ep - Ei + Em) / 20)
+ *
+ *    Ep is predicted using past energies and prediction coefficients:
+ *
+ *        Ep = sum(i,0,M){ ma_prediction_coeff[i] * quant_energy[i] }
+ *
+ *    gc_p in fixed-point arithmetic is calculated as following:
+ *
+ *        mean = 1/N * sum(i,0,N){ (fc_v[i] / 2^13) * (fc_v[i] / 2^13) } =
+ *        = 1/N * sum(i,0,N) { fc_v[i] * fc_v[i] } / 2^26
+ *
+ *        Ei = 10log(mean) = -10log(N) - 10log(2^26) +
+ *        + 10log(sum(i,0,N) { fc_v[i] * fc_v[i] })
+ *
+ *        Ep - Ei + Em = Ep + Em + 10log(N) + 10log(2^26) -
+ *        - 10log(sum(i,0,N) { fc_v[i] * fc_v[i] }) =
+ *        = Ep + mr_energy - 10log(sum(i,0,N) { fc_v[i] * fc_v[i] })
+ *
+ *        gc_p = 10 ^ ((Ep - Ei + Em) / 20) =
+ *        = 2 ^ (3.3219 * (Ep - Ei + Em) / 20) = 2 ^ (0.166 * (Ep - Ei + Em))
+ *
+ *    where
+ *
+ *        mr_energy = Em + 10log(N) + 10log(2^26)
+ *
+ * \remark The routine is used in G.729 and AMR (all modes).
+ */
+int16_t ff_acelp_decode_gain_code(
+    int gain_corr_factor,
+    const int16_t* fc_v,
+    int mr_energy,
+    const int16_t* quant_energy,
+    const int16_t* ma_prediction_coeff,
+    int subframe_size,
+    int max_pred_order);
+
+#endif /* FFMPEG_ACELP_PITCH_DELAY_H */