aptx: implement the aptX bluetooth codec
authorAurelien Jacobs <aurel@gnuage.org>
Fri, 10 Nov 2017 21:09:23 +0000 (22:09 +0100)
committerRostislav Pehlivanov <atomnuker@gmail.com>
Fri, 10 Nov 2017 21:32:06 +0000 (21:32 +0000)
The encoder was reverse engineered from binary library and from
EP0398973B1 patent (long expired).
The decoder was simply deduced from the encoder.

doc/general.texi
libavcodec/Makefile
libavcodec/allcodecs.c
libavcodec/aptx.c [new file with mode: 0644]
libavcodec/avcodec.h
libavcodec/codec_desc.c

index e6ae277..de4efee 100644 (file)
@@ -993,6 +993,8 @@ following image formats are supported:
 @item Amazing Studio PAF Audio @tab     @tab  X
 @item Apple lossless audio   @tab  X  @tab  X
     @tab QuickTime fourcc 'alac'
+@item aptX                   @tab  X  @tab  X
+    @tab Used in Bluetooth A2DP
 @item ATRAC1                 @tab     @tab  X
 @item ATRAC3                 @tab     @tab  X
 @item ATRAC3+                @tab     @tab  X
index 45f4db5..95c843d 100644 (file)
@@ -188,6 +188,8 @@ OBJS-$(CONFIG_AMV_ENCODER)             += mjpegenc.o mjpegenc_common.o \
 OBJS-$(CONFIG_ANM_DECODER)             += anm.o
 OBJS-$(CONFIG_ANSI_DECODER)            += ansi.o cga_data.o
 OBJS-$(CONFIG_APE_DECODER)             += apedec.o
+OBJS-$(CONFIG_APTX_DECODER)            += aptx.o
+OBJS-$(CONFIG_APTX_ENCODER)            += aptx.o
 OBJS-$(CONFIG_APNG_DECODER)            += png.o pngdec.o pngdsp.o
 OBJS-$(CONFIG_APNG_ENCODER)            += png.o pngenc.o
 OBJS-$(CONFIG_SSA_DECODER)             += assdec.o ass.o
index d96e499..463f7ed 100644 (file)
@@ -406,6 +406,7 @@ static void register_all(void)
     REGISTER_DECODER(AMRNB,             amrnb);
     REGISTER_DECODER(AMRWB,             amrwb);
     REGISTER_DECODER(APE,               ape);
+    REGISTER_ENCDEC (APTX,              aptx);
     REGISTER_DECODER(ATRAC1,            atrac1);
     REGISTER_DECODER(ATRAC3,            atrac3);
     REGISTER_DECODER(ATRAC3AL,          atrac3al);
diff --git a/libavcodec/aptx.c b/libavcodec/aptx.c
new file mode 100644 (file)
index 0000000..d09ce8f
--- /dev/null
@@ -0,0 +1,860 @@
+/*
+ * Audio Processing Technology codec for Bluetooth (aptX)
+ *
+ * Copyright (C) 2017  Aurelien Jacobs <aurel@gnuage.org>
+ *
+ * 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 "libavutil/intreadwrite.h"
+#include "avcodec.h"
+#include "internal.h"
+#include "mathops.h"
+#include "audio_frame_queue.h"
+
+
+enum channels {
+    LEFT,
+    RIGHT,
+    NB_CHANNELS
+};
+
+enum subbands {
+    LF,  // Low Frequency (0-5.5 kHz)
+    MLF, // Medium-Low Frequency (5.5-11kHz)
+    MHF, // Medium-High Frequency (11-16.5kHz)
+    HF,  // High Frequency (16.5-22kHz)
+    NB_SUBBANDS
+};
+
+#define NB_FILTERS 2
+#define FILTER_TAPS 16
+
+typedef struct {
+    int pos;
+    int32_t buffer[2*FILTER_TAPS];
+} FilterSignal;
+
+typedef struct {
+    FilterSignal outer_filter_signal[NB_FILTERS];
+    FilterSignal inner_filter_signal[NB_FILTERS][NB_FILTERS];
+} QMFAnalysis;
+
+typedef struct {
+    int32_t quantized_sample;
+    int32_t quantized_sample_parity_change;
+    int32_t error;
+} Quantize;
+
+typedef struct {
+    int32_t quantization_factor;
+    int32_t factor_select;
+    int32_t reconstructed_difference;
+} InvertQuantize;
+
+typedef struct {
+    int32_t prev_sign[2];
+    int32_t s_weight[2];
+    int32_t d_weight[24];
+    int32_t pos;
+    int32_t reconstructed_differences[48];
+    int32_t previous_reconstructed_sample;
+    int32_t predicted_difference;
+    int32_t predicted_sample;
+} Prediction;
+
+typedef struct {
+    int32_t codeword_history;
+    int32_t dither_parity;
+    int32_t dither[NB_SUBBANDS];
+
+    QMFAnalysis qmf;
+    Quantize quantize[NB_SUBBANDS];
+    InvertQuantize invert_quantize[NB_SUBBANDS];
+    Prediction prediction[NB_SUBBANDS];
+} Channel;
+
+typedef struct {
+    int32_t sync_idx;
+    Channel channels[NB_CHANNELS];
+    AudioFrameQueue afq;
+} AptXContext;
+
+
+static const int32_t quantize_intervals_LF[65] = {
+      -9948,    9948,   29860,   49808,   69822,   89926,  110144,  130502,
+     151026,  171738,  192666,  213832,  235264,  256982,  279014,  301384,
+     324118,  347244,  370790,  394782,  419250,  444226,  469742,  495832,
+     522536,  549890,  577936,  606720,  636290,  666700,  698006,  730270,
+     763562,  797958,  833538,  870398,  908640,  948376,  989740, 1032874,
+    1077948, 1125150, 1174700, 1226850, 1281900, 1340196, 1402156, 1468282,
+    1539182, 1615610, 1698514, 1789098, 1888944, 2000168, 2125700, 2269750,
+    2438670, 2642660, 2899462, 3243240, 3746078, 4535138, 5664098, 7102424,
+    8897462,
+};
+static const int32_t invert_quantize_dither_factors_LF[65] = {
+       9948,   9948,   9962,   9988,  10026,  10078,  10142,  10218,
+      10306,  10408,  10520,  10646,  10784,  10934,  11098,  11274,
+      11462,  11664,  11880,  12112,  12358,  12618,  12898,  13194,
+      13510,  13844,  14202,  14582,  14988,  15422,  15884,  16380,
+      16912,  17484,  18098,  18762,  19480,  20258,  21106,  22030,
+      23044,  24158,  25390,  26760,  28290,  30008,  31954,  34172,
+      36728,  39700,  43202,  47382,  52462,  58762,  66770,  77280,
+      91642, 112348, 144452, 199326, 303512, 485546, 643414, 794914,
+    1000124,
+};
+static const int32_t quantize_dither_factors_LF[65] = {
+        0,     4,     7,    10,    13,    16,    19,    22,
+       26,    28,    32,    35,    38,    41,    44,    47,
+       51,    54,    58,    62,    65,    70,    74,    79,
+       84,    90,    95,   102,   109,   116,   124,   133,
+      143,   154,   166,   180,   195,   212,   231,   254,
+      279,   308,   343,   383,   430,   487,   555,   639,
+      743,   876,  1045,  1270,  1575,  2002,  2628,  3591,
+     5177,  8026, 13719, 26047, 45509, 39467, 37875, 51303,
+        0,
+};
+static const int16_t quantize_factor_select_offset_LF[65] = {
+      0, -21, -19, -17, -15, -12, -10,  -8,
+     -6,  -4,  -1,   1,   3,   6,   8,  10,
+     13,  15,  18,  20,  23,  26,  29,  31,
+     34,  37,  40,  43,  47,  50,  53,  57,
+     60,  64,  68,  72,  76,  80,  85,  89,
+     94,  99, 105, 110, 116, 123, 129, 136,
+    144, 152, 161, 171, 182, 194, 207, 223,
+    241, 263, 291, 328, 382, 467, 522, 522,
+    522,
+};
+
+
+static const int32_t quantize_intervals_MLF[9] = {
+    -89806, 89806, 278502, 494338, 759442, 1113112, 1652322, 2720256, 5190186,
+};
+static const int32_t invert_quantize_dither_factors_MLF[9] = {
+    89806, 89806, 98890, 116946, 148158, 205512, 333698, 734236, 1735696,
+};
+static const int32_t quantize_dither_factors_MLF[9] = {
+    0, 2271, 4514, 7803, 14339, 32047, 100135, 250365, 0,
+};
+static const int16_t quantize_factor_select_offset_MLF[9] = {
+    0, -14, 6, 29, 58, 96, 154, 270, 521,
+};
+
+
+static const int32_t quantize_intervals_MHF[3] = {
+    -194080, 194080, 890562,
+};
+static const int32_t invert_quantize_dither_factors_MHF[3] = {
+    194080, 194080, 502402,
+};
+static const int32_t quantize_dither_factors_MHF[3] = {
+    0, 77081, 0,
+};
+static const int16_t quantize_factor_select_offset_MHF[3] = {
+    0, -33, 136,
+};
+
+
+static const int32_t quantize_intervals_HF[5] = {
+    -163006, 163006, 542708, 1120554, 2669238,
+};
+static const int32_t invert_quantize_dither_factors_HF[5] = {
+    163006, 163006, 216698, 361148, 1187538,
+};
+static const int32_t quantize_dither_factors_HF[5] = {
+    0, 13423, 36113, 206598, 0,
+};
+static const int16_t quantize_factor_select_offset_HF[5] = {
+    0, -8, 33, 95, 262,
+};
+
+typedef const struct {
+    const int32_t *quantize_intervals;
+    const int32_t *invert_quantize_dither_factors;
+    const int32_t *quantize_dither_factors;
+    const int16_t *quantize_factor_select_offset;
+    int tables_size;
+    int32_t quantized_bits;
+    int32_t prediction_order;
+} ConstTables;
+
+static ConstTables tables[NB_SUBBANDS] = {
+    [LF]  = { quantize_intervals_LF,
+              invert_quantize_dither_factors_LF,
+              quantize_dither_factors_LF,
+              quantize_factor_select_offset_LF,
+              FF_ARRAY_ELEMS(quantize_intervals_LF),
+              7, 24 },
+    [MLF] = { quantize_intervals_MLF,
+              invert_quantize_dither_factors_MLF,
+              quantize_dither_factors_MLF,
+              quantize_factor_select_offset_MLF,
+              FF_ARRAY_ELEMS(quantize_intervals_MLF),
+              4, 12 },
+    [MHF] = { quantize_intervals_MHF,
+              invert_quantize_dither_factors_MHF,
+              quantize_dither_factors_MHF,
+              quantize_factor_select_offset_MHF,
+              FF_ARRAY_ELEMS(quantize_intervals_MHF),
+              2, 6 },
+    [HF]  = { quantize_intervals_HF,
+              invert_quantize_dither_factors_HF,
+              quantize_dither_factors_HF,
+              quantize_factor_select_offset_HF,
+              FF_ARRAY_ELEMS(quantize_intervals_HF),
+              3, 12 },
+};
+
+static const int16_t quantization_factors[32] = {
+    2048, 2093, 2139, 2186, 2233, 2282, 2332, 2383,
+    2435, 2489, 2543, 2599, 2656, 2714, 2774, 2834,
+    2896, 2960, 3025, 3091, 3158, 3228, 3298, 3371,
+    3444, 3520, 3597, 3676, 3756, 3838, 3922, 4008,
+};
+
+
+/* Rounded right shift with optionnal clipping */
+#define RSHIFT_SIZE(size)                                                     \
+av_always_inline                                                              \
+static int##size##_t rshift##size(int##size##_t value, int shift)             \
+{                                                                             \
+    int##size##_t rounding = (int##size##_t)1 << (shift - 1);                 \
+    int##size##_t mask = ((int##size##_t)1 << (shift + 1)) - 1;               \
+    return ((value + rounding) >> shift) - ((value & mask) == rounding);      \
+}                                                                             \
+av_always_inline                                                              \
+static int##size##_t rshift##size##_clip24(int##size##_t value, int shift)    \
+{                                                                             \
+    return av_clip_intp2(rshift##size(value, shift), 23);                     \
+}
+RSHIFT_SIZE(32)
+RSHIFT_SIZE(64)
+
+
+av_always_inline
+static void aptx_update_codeword_history(Channel *channel)
+{
+    int32_t cw = ((channel->quantize[0].quantized_sample & 3) << 0) +
+                 ((channel->quantize[1].quantized_sample & 2) << 1) +
+                 ((channel->quantize[2].quantized_sample & 1) << 3);
+    channel->codeword_history = (cw << 8) + (channel->codeword_history << 4);
+}
+
+static void aptx_generate_dither(Channel *channel)
+{
+    int subband;
+    int64_t m;
+    int32_t d;
+
+    aptx_update_codeword_history(channel);
+
+    m = (int64_t)5184443 * (channel->codeword_history >> 7);
+    d = (m << 2) + (m >> 22);
+    for (subband = 0; subband < NB_SUBBANDS; subband++)
+        channel->dither[subband] = d << (23 - 5*subband);
+    channel->dither_parity = (d >> 25) & 1;
+}
+
+/*
+ * Convolution filter coefficients for the outer QMF of the QMF tree.
+ * The 2 sets are a mirror of each other.
+ */
+static const int32_t aptx_qmf_outer_coeffs[NB_FILTERS][FILTER_TAPS] = {
+    {
+        730, -413, -9611, 43626, -121026, 269973, -585547, 2801966,
+        697128, -160481, 27611, 8478, -10043, 3511, 688, -897,
+    },
+    {
+        -897, 688, 3511, -10043, 8478, 27611, -160481, 697128,
+        2801966, -585547, 269973, -121026, 43626, -9611, -413, 730,
+    },
+};
+
+/*
+ * Convolution filter coefficients for the inner QMF of the QMF tree.
+ * The 2 sets are a mirror of each other.
+ */
+static const int32_t aptx_qmf_inner_coeffs[NB_FILTERS][FILTER_TAPS] = {
+    {
+       1033, -584, -13592, 61697, -171156, 381799, -828088, 3962579,
+       985888, -226954, 39048, 11990, -14203, 4966, 973, -1268,
+    },
+    {
+      -1268, 973, 4966, -14203, 11990, 39048, -226954, 985888,
+      3962579, -828088, 381799, -171156, 61697, -13592, -584, 1033,
+    },
+};
+
+/*
+ * Push one sample into a circular signal buffer.
+ */
+av_always_inline
+static void aptx_qmf_filter_signal_push(FilterSignal *signal, int32_t sample)
+{
+    signal->buffer[signal->pos            ] = sample;
+    signal->buffer[signal->pos+FILTER_TAPS] = sample;
+    signal->pos = (signal->pos + 1) & (FILTER_TAPS - 1);
+}
+
+/*
+ * Compute the convolution of the signal with the coefficients, and reduce
+ * to 24 bits by applying the specified right shifting.
+ */
+av_always_inline
+static int32_t aptx_qmf_convolution(FilterSignal *signal,
+                                    const int32_t coeffs[FILTER_TAPS],
+                                    int shift)
+{
+    int32_t *sig = &signal->buffer[signal->pos];
+    int64_t e = 0;
+    int i;
+
+    for (i = 0; i < FILTER_TAPS; i++)
+        e += MUL64(sig[i], coeffs[i]);
+
+    return rshift64_clip24(e, shift);
+}
+
+/*
+ * Half-band QMF analysis filter realized with a polyphase FIR filter.
+ * Split into 2 subbands and downsample by 2.
+ * So for each pair of samples that goes in, one sample goes out,
+ * split into 2 separate subbands.
+ */
+av_always_inline
+static void aptx_qmf_polyphase_analysis(FilterSignal signal[NB_FILTERS],
+                                        const int32_t coeffs[NB_FILTERS][FILTER_TAPS],
+                                        int shift,
+                                        int32_t samples[NB_FILTERS],
+                                        int32_t *low_subband_output,
+                                        int32_t *high_subband_output)
+{
+    int32_t subbands[NB_FILTERS];
+    int i;
+
+    for (i = 0; i < NB_FILTERS; i++) {
+        aptx_qmf_filter_signal_push(&signal[i], samples[NB_FILTERS-1-i]);
+        subbands[i] = aptx_qmf_convolution(&signal[i], coeffs[i], shift);
+    }
+
+    *low_subband_output  = av_clip_intp2(subbands[0] + subbands[1], 23);
+    *high_subband_output = av_clip_intp2(subbands[0] - subbands[1], 23);
+}
+
+/*
+ * Two stage QMF analysis tree.
+ * Split 4 input samples into 4 subbands and downsample by 4.
+ * So for each group of 4 samples that goes in, one sample goes out,
+ * split into 4 separate subbands.
+ */
+static void aptx_qmf_tree_analysis(QMFAnalysis *qmf,
+                                   int32_t samples[4],
+                                   int32_t subband_samples[4])
+{
+    int32_t intermediate_samples[4];
+    int i;
+
+    /* Split 4 input samples into 2 intermediate subbands downsampled to 2 samples */
+    for (i = 0; i < 2; i++)
+        aptx_qmf_polyphase_analysis(qmf->outer_filter_signal,
+                                    aptx_qmf_outer_coeffs, 23,
+                                    &samples[2*i],
+                                    &intermediate_samples[0+i],
+                                    &intermediate_samples[2+i]);
+
+    /* Split 2 intermediate subband samples into 4 final subbands downsampled to 1 sample */
+    for (i = 0; i < 2; i++)
+        aptx_qmf_polyphase_analysis(qmf->inner_filter_signal[i],
+                                    aptx_qmf_inner_coeffs, 23,
+                                    &intermediate_samples[2*i],
+                                    &subband_samples[2*i+0],
+                                    &subband_samples[2*i+1]);
+}
+
+/*
+ * Half-band QMF synthesis filter realized with a polyphase FIR filter.
+ * Join 2 subbands and upsample by 2.
+ * So for each 2 subbands sample that goes in, a pair of samples goes out.
+ */
+av_always_inline
+static void aptx_qmf_polyphase_synthesis(FilterSignal signal[NB_FILTERS],
+                                         const int32_t coeffs[NB_FILTERS][FILTER_TAPS],
+                                         int shift,
+                                         int32_t low_subband_input,
+                                         int32_t high_subband_input,
+                                         int32_t samples[NB_FILTERS])
+{
+    int32_t subbands[NB_FILTERS];
+    int i;
+
+    subbands[0] = low_subband_input + high_subband_input;
+    subbands[1] = low_subband_input - high_subband_input;
+
+    for (i = 0; i < NB_FILTERS; i++) {
+        aptx_qmf_filter_signal_push(&signal[i], subbands[1-i]);
+        samples[i] = aptx_qmf_convolution(&signal[i], coeffs[i], shift);
+    }
+}
+
+/*
+ * Two stage QMF synthesis tree.
+ * Join 4 subbands and upsample by 4.
+ * So for each 4 subbands sample that goes in, a group of 4 samples goes out.
+ */
+static void aptx_qmf_tree_synthesis(QMFAnalysis *qmf,
+                                    int32_t subband_samples[4],
+                                    int32_t samples[4])
+{
+    int32_t intermediate_samples[4];
+    int i;
+
+    /* Join 4 subbands into 2 intermediate subbands upsampled to 2 samples. */
+    for (i = 0; i < 2; i++)
+        aptx_qmf_polyphase_synthesis(qmf->inner_filter_signal[i],
+                                     aptx_qmf_inner_coeffs, 22,
+                                     subband_samples[2*i+0],
+                                     subband_samples[2*i+1],
+                                     &intermediate_samples[2*i]);
+
+    /* Join 2 samples from intermediate subbands upsampled to 4 samples. */
+    for (i = 0; i < 2; i++)
+        aptx_qmf_polyphase_synthesis(qmf->outer_filter_signal,
+                                     aptx_qmf_outer_coeffs, 21,
+                                     intermediate_samples[0+i],
+                                     intermediate_samples[2+i],
+                                     &samples[2*i]);
+}
+
+
+av_always_inline
+static int32_t aptx_bin_search(int32_t value, int32_t factor,
+                               const int32_t *intervals, int32_t nb_intervals)
+{
+    int32_t idx = 0;
+    int i;
+
+    for (i = nb_intervals >> 1; i > 0; i >>= 1)
+        if (MUL64(factor, intervals[idx + i]) <= ((int64_t)value << 24))
+            idx += i;
+
+    return idx;
+}
+
+static void aptx_quantize_difference(Quantize *quantize,
+                                     int32_t sample_difference,
+                                     int32_t dither,
+                                     int32_t quantization_factor,
+                                     ConstTables *tables)
+{
+    const int32_t *intervals = tables->quantize_intervals;
+    int32_t quantized_sample, dithered_sample, parity_change;
+    int32_t d, mean, interval, inv;
+    int64_t error;
+
+    quantized_sample = aptx_bin_search(FFABS(sample_difference) >> 4,
+                                       quantization_factor,
+                                       intervals, tables->tables_size);
+
+    d = rshift32_clip24(MULH(dither, dither), 7) - (1 << 23);
+    d = rshift64(MUL64(d, tables->quantize_dither_factors[quantized_sample]), 23);
+
+    intervals += quantized_sample;
+    mean = (intervals[1] + intervals[0]) / 2;
+    interval = (intervals[1] - intervals[0]) * (-(sample_difference < 0) | 1);
+
+    dithered_sample = rshift64_clip24(MUL64(dither, interval) + ((int64_t)(mean + d) << 32), 32);
+    error = ((int64_t)FFABS(sample_difference) << 20) - MUL64(dithered_sample, quantization_factor);
+    quantize->error = FFABS(rshift64(error, 23));
+
+    parity_change = quantized_sample;
+    if (error < 0)
+        quantized_sample--;
+    else
+        parity_change--;
+
+    inv = -(sample_difference < 0);
+    quantize->quantized_sample               = quantized_sample ^ inv;
+    quantize->quantized_sample_parity_change = parity_change    ^ inv;
+}
+
+static void aptx_encode_channel(Channel *channel, int32_t samples[4])
+{
+    int32_t subband_samples[4];
+    int subband;
+    aptx_qmf_tree_analysis(&channel->qmf, samples, subband_samples);
+    aptx_generate_dither(channel);
+    for (subband = 0; subband < NB_SUBBANDS; subband++) {
+        int32_t diff = av_clip_intp2(subband_samples[subband] - channel->prediction[subband].predicted_sample, 23);
+        aptx_quantize_difference(&channel->quantize[subband], diff,
+                                 channel->dither[subband],
+                                 channel->invert_quantize[subband].quantization_factor,
+                                 &tables[subband]);
+    }
+}
+
+static void aptx_decode_channel(Channel *channel, int32_t samples[4])
+{
+    int32_t subband_samples[4];
+    int subband;
+    for (subband = 0; subband < NB_SUBBANDS; subband++)
+        subband_samples[subband] = channel->prediction[subband].previous_reconstructed_sample;
+    aptx_qmf_tree_synthesis(&channel->qmf, subband_samples, samples);
+}
+
+
+static void aptx_invert_quantization(InvertQuantize *invert_quantize,
+                                     int32_t quantized_sample, int32_t dither,
+                                     ConstTables *tables)
+{
+    int32_t qr, idx, shift, factor_select;
+
+    idx = (quantized_sample ^ -(quantized_sample < 0)) + 1;
+    qr = tables->quantize_intervals[idx] / 2;
+    if (quantized_sample < 0)
+        qr = -qr;
+
+    qr = rshift64_clip24(((int64_t)qr<<32) + MUL64(dither, tables->invert_quantize_dither_factors[idx]), 32);
+    invert_quantize->reconstructed_difference = MUL64(invert_quantize->quantization_factor, qr) >> 19;
+
+    shift = 24 - tables->quantized_bits;
+
+    /* update factor_select */
+    factor_select = 32620 * invert_quantize->factor_select;
+    factor_select = rshift32(factor_select + (tables->quantize_factor_select_offset[idx] << 15), 15);
+    invert_quantize->factor_select = av_clip(factor_select, 0, (shift << 8) | 0xFF);
+
+    /* update quantization factor */
+    idx = (invert_quantize->factor_select & 0xFF) >> 3;
+    shift -= invert_quantize->factor_select >> 8;
+    invert_quantize->quantization_factor = (quantization_factors[idx] << 11) >> shift;
+}
+
+static int32_t *aptx_reconstructed_differences_update(Prediction *prediction,
+                                                      int32_t reconstructed_difference,
+                                                      int order)
+{
+    int32_t *rd1 = prediction->reconstructed_differences, *rd2 = rd1 + order;
+    int p = prediction->pos;
+
+    rd1[p] = rd2[p];
+    prediction->pos = p = (p + 1) % order;
+    rd2[p] = reconstructed_difference;
+    return &rd2[p];
+}
+
+static void aptx_prediction_filtering(Prediction *prediction,
+                                      int32_t reconstructed_difference,
+                                      int order)
+{
+    int32_t reconstructed_sample, predictor, srd0;
+    int32_t *reconstructed_differences;
+    int64_t predicted_difference = 0;
+    int i;
+
+    reconstructed_sample = av_clip_intp2(reconstructed_difference + prediction->predicted_sample, 23);
+    predictor = av_clip_intp2((MUL64(prediction->s_weight[0], prediction->previous_reconstructed_sample)
+                             + MUL64(prediction->s_weight[1], reconstructed_sample)) >> 22, 23);
+    prediction->previous_reconstructed_sample = reconstructed_sample;
+
+    reconstructed_differences = aptx_reconstructed_differences_update(prediction, reconstructed_difference, order);
+    srd0 = FFDIFFSIGN(reconstructed_difference, 0) << 23;
+    for (i = 0; i < order; i++) {
+        int32_t srd = FF_SIGNBIT(reconstructed_differences[-i-1]) | 1;
+        prediction->d_weight[i] -= rshift32(prediction->d_weight[i] - srd*srd0, 8);
+        predicted_difference += MUL64(reconstructed_differences[-i], prediction->d_weight[i]);
+    }
+
+    prediction->predicted_difference = av_clip_intp2(predicted_difference >> 22, 23);
+    prediction->predicted_sample = av_clip_intp2(predictor + prediction->predicted_difference, 23);
+}
+
+static void aptx_process_subband(InvertQuantize *invert_quantize,
+                                 Prediction *prediction,
+                                 int32_t quantized_sample, int32_t dither,
+                                 ConstTables *tables)
+{
+    int32_t sign, same_sign[2], weight[2], sw1, range;
+
+    aptx_invert_quantization(invert_quantize, quantized_sample, dither, tables);
+
+    sign = FFDIFFSIGN(invert_quantize->reconstructed_difference,
+                      -prediction->predicted_difference);
+    same_sign[0] = sign * prediction->prev_sign[0];
+    same_sign[1] = sign * prediction->prev_sign[1];
+    prediction->prev_sign[0] = prediction->prev_sign[1];
+    prediction->prev_sign[1] = sign | 1;
+
+    range = 0x100000;
+    sw1 = rshift32(-same_sign[1] * prediction->s_weight[1], 1);
+    sw1 = (av_clip(sw1, -range, range) & ~0xF) << 4;
+
+    range = 0x300000;
+    weight[0] = 254 * prediction->s_weight[0] + 0x800000*same_sign[0] + sw1;
+    prediction->s_weight[0] = av_clip(rshift32(weight[0], 8), -range, range);
+
+    range = 0x3C0000 - prediction->s_weight[0];
+    weight[1] = 255 * prediction->s_weight[1] + 0xC00000*same_sign[1];
+    prediction->s_weight[1] = av_clip(rshift32(weight[1], 8), -range, range);
+
+    aptx_prediction_filtering(prediction,
+                              invert_quantize->reconstructed_difference,
+                              tables->prediction_order);
+}
+
+static void aptx_invert_quantize_and_prediction(Channel *channel)
+{
+    int subband;
+    for (subband = 0; subband < NB_SUBBANDS; subband++)
+        aptx_process_subband(&channel->invert_quantize[subband],
+                             &channel->prediction[subband],
+                             channel->quantize[subband].quantized_sample,
+                             channel->dither[subband],
+                             &tables[subband]);
+}
+
+static int32_t aptx_quantized_parity(Channel *channel)
+{
+    int32_t parity = channel->dither_parity;
+    int subband;
+
+    for (subband = 0; subband < NB_SUBBANDS; subband++)
+        parity ^= channel->quantize[subband].quantized_sample;
+
+    return parity & 1;
+}
+
+/* For each sample, ensure that the parity of all subbands of all channels
+ * is 0 except once every 8 samples where the parity is forced to 1. */
+static int aptx_check_parity(Channel channels[NB_CHANNELS], int32_t *idx)
+{
+    int32_t parity = aptx_quantized_parity(&channels[LEFT])
+                   ^ aptx_quantized_parity(&channels[RIGHT]);
+
+    int eighth = *idx == 7;
+    *idx = (*idx + 1) & 7;
+
+    return parity ^ eighth;
+}
+
+static void aptx_insert_sync(Channel channels[NB_CHANNELS], int32_t *idx)
+{
+    if (aptx_check_parity(channels, idx)) {
+        int i;
+        Channel *c;
+        static const int map[] = { 1, 2, 0, 3 };
+        Quantize *min = &channels[NB_CHANNELS-1].quantize[map[0]];
+        for (c = &channels[NB_CHANNELS-1]; c >= channels; c--)
+            for (i = 0; i < NB_SUBBANDS; i++)
+                if (c->quantize[map[i]].error < min->error)
+                    min = &c->quantize[map[i]];
+
+        /* Forcing the desired parity is done by offsetting by 1 the quantized
+         * sample from the subband featuring the smallest quantization error. */
+        min->quantized_sample = min->quantized_sample_parity_change;
+    }
+}
+
+static uint16_t aptx_pack_codeword(Channel *channel)
+{
+    int32_t parity = aptx_quantized_parity(channel);
+    return (((channel->quantize[3].quantized_sample & 0x06) | parity) << 13)
+         | (((channel->quantize[2].quantized_sample & 0x03)         ) << 11)
+         | (((channel->quantize[1].quantized_sample & 0x0F)         ) <<  7)
+         | (((channel->quantize[0].quantized_sample & 0x7F)         ) <<  0);
+}
+
+static void aptx_unpack_codeword(Channel *channel, uint16_t codeword)
+{
+    channel->quantize[0].quantized_sample = sign_extend(codeword >>  0, 7);
+    channel->quantize[1].quantized_sample = sign_extend(codeword >>  7, 4);
+    channel->quantize[2].quantized_sample = sign_extend(codeword >> 11, 2);
+    channel->quantize[3].quantized_sample = sign_extend(codeword >> 13, 3);
+    channel->quantize[3].quantized_sample = (channel->quantize[3].quantized_sample & ~1)
+                                          | aptx_quantized_parity(channel);
+}
+
+static void aptx_encode_samples(AptXContext *ctx,
+                                int32_t samples[NB_CHANNELS][4],
+                                uint8_t output[2*NB_CHANNELS])
+{
+    int channel;
+    for (channel = 0; channel < NB_CHANNELS; channel++)
+        aptx_encode_channel(&ctx->channels[channel], samples[channel]);
+
+    aptx_insert_sync(ctx->channels, &ctx->sync_idx);
+
+    for (channel = 0; channel < NB_CHANNELS; channel++) {
+        aptx_invert_quantize_and_prediction(&ctx->channels[channel]);
+        AV_WB16(output + 2*channel, aptx_pack_codeword(&ctx->channels[channel]));
+    }
+}
+
+static int aptx_decode_samples(AptXContext *ctx,
+                                const uint8_t input[2*NB_CHANNELS],
+                                int32_t samples[NB_CHANNELS][4])
+{
+    int channel, ret;
+
+    for (channel = 0; channel < NB_CHANNELS; channel++) {
+        uint16_t codeword;
+        aptx_generate_dither(&ctx->channels[channel]);
+
+        codeword = AV_RB16(input + 2*channel);
+        aptx_unpack_codeword(&ctx->channels[channel], codeword);
+        aptx_invert_quantize_and_prediction(&ctx->channels[channel]);
+    }
+
+    ret = aptx_check_parity(ctx->channels, &ctx->sync_idx);
+
+    for (channel = 0; channel < NB_CHANNELS; channel++)
+        aptx_decode_channel(&ctx->channels[channel], samples[channel]);
+
+    return ret;
+}
+
+
+static av_cold int aptx_init(AVCodecContext *avctx)
+{
+    AptXContext *s = avctx->priv_data;
+    int chan, subband;
+
+    if (avctx->frame_size == 0)
+        avctx->frame_size = 1024;
+
+    if (avctx->frame_size & 3) {
+        av_log(avctx, AV_LOG_ERROR, "Frame size must be a multiple of 4 samples\n");
+        return AVERROR(EINVAL);
+    }
+
+    for (chan = 0; chan < NB_CHANNELS; chan++) {
+        Channel *channel = &s->channels[chan];
+        for (subband = 0; subband < NB_SUBBANDS; subband++) {
+            Prediction *prediction = &channel->prediction[subband];
+            prediction->prev_sign[0] = 1;
+            prediction->prev_sign[1] = 1;
+        }
+    }
+
+    ff_af_queue_init(avctx, &s->afq);
+    return 0;
+}
+
+static int aptx_decode_frame(AVCodecContext *avctx, void *data,
+                             int *got_frame_ptr, AVPacket *avpkt)
+{
+    AptXContext *s = avctx->priv_data;
+    AVFrame *frame = data;
+    int pos, channel, sample, ret;
+
+    if (avpkt->size < 4) {
+        av_log(avctx, AV_LOG_ERROR, "Packet is too small\n");
+        return AVERROR_INVALIDDATA;
+    }
+
+    /* get output buffer */
+    frame->channels = NB_CHANNELS;
+    frame->format = AV_SAMPLE_FMT_S32P;
+    frame->nb_samples = avpkt->size & ~3;
+    if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
+        return ret;
+
+    for (pos = 0; pos < frame->nb_samples; pos += 4) {
+        int32_t samples[NB_CHANNELS][4];
+
+        if (aptx_decode_samples(s, &avpkt->data[pos], samples)) {
+            av_log(avctx, AV_LOG_ERROR, "Synchronization error\n");
+            return AVERROR_INVALIDDATA;
+        }
+
+        for (channel = 0; channel < NB_CHANNELS; channel++)
+            for (sample = 0; sample < 4; sample++)
+                AV_WN32A(&frame->data[channel][4*(sample+pos)],
+                         samples[channel][sample] << 8);
+    }
+
+    *got_frame_ptr = 1;
+    return frame->nb_samples;
+}
+
+static int aptx_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
+                             const AVFrame *frame, int *got_packet_ptr)
+{
+    AptXContext *s = avctx->priv_data;
+    int pos, channel, sample, ret;
+
+    if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
+        return ret;
+
+    if ((ret = ff_alloc_packet2(avctx, avpkt, frame->nb_samples, 0)) < 0)
+        return ret;
+
+    for (pos = 0; pos < frame->nb_samples; pos += 4) {
+        int32_t samples[NB_CHANNELS][4];
+
+        for (channel = 0; channel < NB_CHANNELS; channel++)
+            for (sample = 0; sample < 4; sample++)
+                samples[channel][sample] = (int32_t)AV_RN32A(&frame->data[channel][4*(sample+pos)]) >> 8;
+
+        aptx_encode_samples(s, samples, avpkt->data + pos);
+    }
+
+    ff_af_queue_remove(&s->afq, frame->nb_samples, &avpkt->pts, &avpkt->duration);
+    *got_packet_ptr = 1;
+    return 0;
+}
+
+static av_cold int aptx_close(AVCodecContext *avctx)
+{
+    AptXContext *s = avctx->priv_data;
+    ff_af_queue_close(&s->afq);
+    return 0;
+}
+
+
+#if CONFIG_APTX_DECODER
+AVCodec ff_aptx_decoder = {
+    .name                  = "aptx",
+    .long_name             = NULL_IF_CONFIG_SMALL("aptX (Audio Processing Technology for Bluetooth)"),
+    .type                  = AVMEDIA_TYPE_AUDIO,
+    .id                    = AV_CODEC_ID_APTX,
+    .priv_data_size        = sizeof(AptXContext),
+    .init                  = aptx_init,
+    .decode                = aptx_decode_frame,
+    .close                 = aptx_close,
+    .capabilities          = AV_CODEC_CAP_DR1,
+    .channel_layouts       = (const uint64_t[]) { AV_CH_LAYOUT_STEREO, 0},
+    .sample_fmts           = (const enum AVSampleFormat[]) { AV_SAMPLE_FMT_S32P,
+                                                             AV_SAMPLE_FMT_NONE },
+};
+#endif
+
+#if CONFIG_APTX_ENCODER
+AVCodec ff_aptx_encoder = {
+    .name                  = "aptx",
+    .long_name             = NULL_IF_CONFIG_SMALL("aptX (Audio Processing Technology for Bluetooth)"),
+    .type                  = AVMEDIA_TYPE_AUDIO,
+    .id                    = AV_CODEC_ID_APTX,
+    .priv_data_size        = sizeof(AptXContext),
+    .init                  = aptx_init,
+    .encode2               = aptx_encode_frame,
+    .close                 = aptx_close,
+    .channel_layouts       = (const uint64_t[]) { AV_CH_LAYOUT_STEREO, 0},
+    .sample_fmts           = (const enum AVSampleFormat[]) { AV_SAMPLE_FMT_S32P,
+                                                             AV_SAMPLE_FMT_NONE },
+    .supported_samplerates = (const int[]) {8000, 16000, 24000, 32000, 44100, 48000, 0},
+};
+#endif
index c413442..429d62a 100644 (file)
@@ -632,6 +632,7 @@ enum AVCodecID {
     AV_CODEC_ID_ATRAC3AL,
     AV_CODEC_ID_ATRAC3PAL,
     AV_CODEC_ID_DOLBY_E,
+    AV_CODEC_ID_APTX,
 
     /* subtitle codecs */
     AV_CODEC_ID_FIRST_SUBTITLE = 0x17000,          ///< A dummy ID pointing at the start of subtitle codecs.
index 92bf1d2..c3688de 100644 (file)
@@ -2859,6 +2859,13 @@ static const AVCodecDescriptor codec_descriptors[] = {
         .long_name = NULL_IF_CONFIG_SMALL("ADPCM MTAF"),
         .props     = AV_CODEC_PROP_LOSSY,
     },
+    {
+        .id        = AV_CODEC_ID_APTX,
+        .type      = AVMEDIA_TYPE_AUDIO,
+        .name      = "aptx",
+        .long_name = NULL_IF_CONFIG_SMALL("aptX (Audio Processing Technology for Bluetooth)"),
+        .props     = AV_CODEC_PROP_LOSSY,
+    },
 
     /* subtitle codecs */
     {