libavcodec/atrac3.c

Go to the documentation of this file.

/*
 * Atrac 3 compatible decoder
 * Copyright (c) 2006-2008 Maxim Poliakovski
 * Copyright (c) 2006-2008 Benjamin Larsson
 *
 * 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 <math.h>
#include <stddef.h>
#include <stdio.h>

#include "libavutil/float_dsp.h"
#include "libavutil/libm.h"
#include "avcodec.h"
#include "bytestream.h"
#include "fft.h"
#include "fmtconvert.h"
#include "get_bits.h"
#include "internal.h"

#include "atrac.h"
#include "atrac3data.h"

#define JOINT_STEREO    0x12
#define STEREO          0x2

#define SAMPLES_PER_FRAME 1024
#define MDCT_SIZE          512

typedef struct GainInfo {
    int num_gain_data;
    int lev_code[8];
    int loc_code[8];
} GainInfo;

typedef struct GainBlock {
    GainInfo g_block[4];
} GainBlock;

typedef struct TonalComponent {
    int pos;
    int num_coefs;
    float coef[8];
} TonalComponent;

typedef struct ChannelUnit {
    int            bands_coded;
    int            num_components;
    float          prev_frame[SAMPLES_PER_FRAME];
    int            gc_blk_switch;
    TonalComponent components[64];
    GainBlock      gain_block[2];

    DECLARE_ALIGNED(32, float, spectrum)[SAMPLES_PER_FRAME];
    DECLARE_ALIGNED(32, float, imdct_buf)[SAMPLES_PER_FRAME];

    float          delay_buf1[46]; 
    float          delay_buf2[46];
    float          delay_buf3[46];
} ChannelUnit;

typedef struct ATRAC3Context {
    AVFrame frame;
    GetBitContext gb;

    int coding_mode;

    ChannelUnit *units;


    int matrix_coeff_index_prev[4];
    int matrix_coeff_index_now[4];
    int matrix_coeff_index_next[4];
    int weighting_delay[6];


    uint8_t *decoded_bytes_buffer;
    float temp_buf[1070];


    int scrambled_stream;

    FFTContext mdct_ctx;
    FmtConvertContext fmt_conv;
    AVFloatDSPContext fdsp;
} ATRAC3Context;

static DECLARE_ALIGNED(32, float, mdct_window)[MDCT_SIZE];
static VLC_TYPE atrac3_vlc_table[4096][2];
static VLC   spectral_coeff_tab[7];
static float gain_tab1[16];
static float gain_tab2[31];


static void imlt(ATRAC3Context *q, float *input, float *output, int odd_band)
{
    int i;

    if (odd_band) {
        for (i = 0; i < 128; i++)
            FFSWAP(float, input[i], input[255 - i]);
    }

    q->mdct_ctx.imdct_calc(&q->mdct_ctx, output, input);

    /* Perform windowing on the output. */
    q->fdsp.vector_fmul(output, output, mdct_window, MDCT_SIZE);
}

/*
 * indata descrambling, only used for data coming from the rm container
 */
static int decode_bytes(const uint8_t *input, uint8_t *out, int bytes)
{
    int i, off;
    uint32_t c;
    const uint32_t *buf;
    uint32_t *output = (uint32_t *)out;

    off = (intptr_t)input & 3;
    buf = (const uint32_t *)(input - off);
    c   = av_be2ne32((0x537F6103 >> (off * 8)) | (0x537F6103 << (32 - (off * 8))));
    bytes += 3 + off;
    for (i = 0; i < bytes / 4; i++)
        output[i] = c ^ buf[i];

    if (off)
        av_log_ask_for_sample(NULL, "Offset of %d not handled.\n", off);

    return off;
}

static av_cold void init_atrac3_window(void)
{
    int i, j;

    /* generate the mdct window, for details see
     * http://wiki.multimedia.cx/index.php?title=RealAudio_atrc#Windows */
    for (i = 0, j = 255; i < 128; i++, j--) {
        float wi = sin(((i + 0.5) / 256.0 - 0.5) * M_PI) + 1.0;
        float wj = sin(((j + 0.5) / 256.0 - 0.5) * M_PI) + 1.0;
        float w  = 0.5 * (wi * wi + wj * wj);
        mdct_window[i] = mdct_window[511 - i] = wi / w;
        mdct_window[j] = mdct_window[511 - j] = wj / w;
    }
}

static av_cold int atrac3_decode_close(AVCodecContext *avctx)
{
    ATRAC3Context *q = avctx->priv_data;

    av_free(q->units);
    av_free(q->decoded_bytes_buffer);

    ff_mdct_end(&q->mdct_ctx);

    return 0;
}

static void read_quant_spectral_coeffs(GetBitContext *gb, int selector,
                                       int coding_flag, int *mantissas,
                                       int num_codes)
{
    int i, code, huff_symb;

    if (selector == 1)
        num_codes /= 2;

    if (coding_flag != 0) {
        /* constant length coding (CLC) */
        int num_bits = clc_length_tab[selector];

        if (selector > 1) {
            for (i = 0; i < num_codes; i++) {
                if (num_bits)
                    code = get_sbits(gb, num_bits);
                else
                    code = 0;
                mantissas[i] = code;
            }
        } else {
            for (i = 0; i < num_codes; i++) {
                if (num_bits)
                    code = get_bits(gb, num_bits); // num_bits is always 4 in this case
                else
                    code = 0;
                mantissas[i * 2    ] = mantissa_clc_tab[code >> 2];
                mantissas[i * 2 + 1] = mantissa_clc_tab[code &  3];
            }
        }
    } else {
        /* variable length coding (VLC) */
        if (selector != 1) {
            for (i = 0; i < num_codes; i++) {
                huff_symb = get_vlc2(gb, spectral_coeff_tab[selector-1].table,
                                     spectral_coeff_tab[selector-1].bits, 3);
                huff_symb += 1;
                code = huff_symb >> 1;
                if (huff_symb & 1)
                    code = -code;
                mantissas[i] = code;
            }
        } else {
            for (i = 0; i < num_codes; i++) {
                huff_symb = get_vlc2(gb, spectral_coeff_tab[selector - 1].table,
                                     spectral_coeff_tab[selector - 1].bits, 3);
                mantissas[i * 2    ] = mantissa_vlc_tab[huff_symb * 2    ];
                mantissas[i * 2 + 1] = mantissa_vlc_tab[huff_symb * 2 + 1];
            }
        }
    }
}

static int decode_spectrum(GetBitContext *gb, float *output)
{
    int num_subbands, coding_mode, i, j, first, last, subband_size;
    int subband_vlc_index[32], sf_index[32];
    int mantissas[128];
    float scale_factor;

    num_subbands = get_bits(gb, 5);  // number of coded subbands
    coding_mode  = get_bits1(gb);    // coding Mode: 0 - VLC/ 1-CLC

    /* get the VLC selector table for the subbands, 0 means not coded */
    for (i = 0; i <= num_subbands; i++)
        subband_vlc_index[i] = get_bits(gb, 3);

    /* read the scale factor indexes from the stream */
    for (i = 0; i <= num_subbands; i++) {
        if (subband_vlc_index[i] != 0)
            sf_index[i] = get_bits(gb, 6);
    }

    for (i = 0; i <= num_subbands; i++) {
        first = subband_tab[i    ];
        last  = subband_tab[i + 1];

        subband_size = last - first;

        if (subband_vlc_index[i] != 0) {
            /* decode spectral coefficients for this subband */
            /* TODO: This can be done faster is several blocks share the
             * same VLC selector (subband_vlc_index) */
            read_quant_spectral_coeffs(gb, subband_vlc_index[i], coding_mode,
                                       mantissas, subband_size);

            /* decode the scale factor for this subband */
            scale_factor = ff_atrac_sf_table[sf_index[i]] *
                           inv_max_quant[subband_vlc_index[i]];

            /* inverse quantize the coefficients */
            for (j = 0; first < last; first++, j++)
                output[first] = mantissas[j] * scale_factor;
        } else {
            /* this subband was not coded, so zero the entire subband */
            memset(output + first, 0, subband_size * sizeof(*output));
        }
    }

    /* clear the subbands that were not coded */
    first = subband_tab[i];
    memset(output + first, 0, (SAMPLES_PER_FRAME - first) * sizeof(*output));
    return num_subbands;
}

static int decode_tonal_components(GetBitContext *gb,
                                   TonalComponent *components, int num_bands)
{
    int i, b, c, m;
    int nb_components, coding_mode_selector, coding_mode;
    int band_flags[4], mantissa[8];
    int component_count = 0;

    nb_components = get_bits(gb, 5);

    /* no tonal components */
    if (nb_components == 0)
        return 0;

    coding_mode_selector = get_bits(gb, 2);
    if (coding_mode_selector == 2)
        return AVERROR_INVALIDDATA;

    coding_mode = coding_mode_selector & 1;

    for (i = 0; i < nb_components; i++) {
        int coded_values_per_component, quant_step_index;

        for (b = 0; b <= num_bands; b++)
            band_flags[b] = get_bits1(gb);

        coded_values_per_component = get_bits(gb, 3);

        quant_step_index = get_bits(gb, 3);
        if (quant_step_index <= 1)
            return AVERROR_INVALIDDATA;

        if (coding_mode_selector == 3)
            coding_mode = get_bits1(gb);

        for (b = 0; b < (num_bands + 1) * 4; b++) {
            int coded_components;

            if (band_flags[b >> 2] == 0)
                continue;

            coded_components = get_bits(gb, 3);

            for (c = 0; c < coded_components; c++) {
                TonalComponent *cmp = &components[component_count];
                int sf_index, coded_values, max_coded_values;
                float scale_factor;

                sf_index = get_bits(gb, 6);
                if (component_count >= 64)
                    return AVERROR_INVALIDDATA;

                cmp->pos = b * 64 + get_bits(gb, 6);

                max_coded_values = SAMPLES_PER_FRAME - cmp->pos;
                coded_values     = coded_values_per_component + 1;
                coded_values     = FFMIN(max_coded_values, coded_values);

                scale_factor = ff_atrac_sf_table[sf_index] *
                               inv_max_quant[quant_step_index];

                read_quant_spectral_coeffs(gb, quant_step_index, coding_mode,
                                           mantissa, coded_values);

                cmp->num_coefs = coded_values;

                /* inverse quant */
                for (m = 0; m < coded_values; m++)
                    cmp->coef[m] = mantissa[m] * scale_factor;

                component_count++;
            }
        }
    }

    return component_count;
}

static int decode_gain_control(GetBitContext *gb, GainBlock *block,
                               int num_bands)
{
    int i, cf, num_data;
    int *level, *loc;

    GainInfo *gain = block->g_block;

    for (i = 0; i <= num_bands; i++) {
        num_data              = get_bits(gb, 3);
        gain[i].num_gain_data = num_data;
        level                 = gain[i].lev_code;
        loc                   = gain[i].loc_code;

        for (cf = 0; cf < gain[i].num_gain_data; cf++) {
            level[cf] = get_bits(gb, 4);
            loc  [cf] = get_bits(gb, 5);
            if (cf && loc[cf] <= loc[cf - 1])
                return AVERROR_INVALIDDATA;
        }
    }

    /* Clear the unused blocks. */
    for (; i < 4 ; i++)
        gain[i].num_gain_data = 0;

    return 0;
}

static void gain_compensate_and_overlap(float *input, float *prev,
                                        float *output, GainInfo *gain1,
                                        GainInfo *gain2)
{
    float g1, g2, gain_inc;
    int i, j, num_data, start_loc, end_loc;


    if (gain2->num_gain_data == 0)
        g1 = 1.0;
    else
        g1 = gain_tab1[gain2->lev_code[0]];

    if (gain1->num_gain_data == 0) {
        for (i = 0; i < 256; i++)
            output[i] = input[i] * g1 + prev[i];
    } else {
        num_data = gain1->num_gain_data;
        gain1->loc_code[num_data] = 32;
        gain1->lev_code[num_data] = 4;

        for (i = 0, j = 0; i < num_data; i++) {
            start_loc = gain1->loc_code[i] * 8;
            end_loc   = start_loc + 8;

            g2       = gain_tab1[gain1->lev_code[i]];
            gain_inc = gain_tab2[gain1->lev_code[i + 1] -
                                 gain1->lev_code[i    ] + 15];

            /* interpolate */
            for (; j < start_loc; j++)
                output[j] = (input[j] * g1 + prev[j]) * g2;

            /* interpolation is done over eight samples */
            for (; j < end_loc; j++) {
                output[j] = (input[j] * g1 + prev[j]) * g2;
                g2 *= gain_inc;
            }
        }

        for (; j < 256; j++)
            output[j] = input[j] * g1 + prev[j];
    }

    /* Delay for the overlapping part. */
    memcpy(prev, &input[256], 256 * sizeof(*prev));
}

static int add_tonal_components(float *spectrum, int num_components,
                                TonalComponent *components)
{
    int i, j, last_pos = -1;
    float *input, *output;

    for (i = 0; i < num_components; i++) {
        last_pos = FFMAX(components[i].pos + components[i].num_coefs, last_pos);
        input    = components[i].coef;
        output   = &spectrum[components[i].pos];

        for (j = 0; j < components[i].num_coefs; j++)
            output[i] += input[i];
    }

    return last_pos;
}

#define INTERPOLATE(old, new, nsample) \
    ((old) + (nsample) * 0.125 * ((new) - (old)))

static void reverse_matrixing(float *su1, float *su2, int *prev_code,
                              int *curr_code)
{
    int i, nsample, band;
    float mc1_l, mc1_r, mc2_l, mc2_r;

    for (i = 0, band = 0; band < 4 * 256; band += 256, i++) {
        int s1 = prev_code[i];
        int s2 = curr_code[i];
        nsample = band;

        if (s1 != s2) {
            /* Selector value changed, interpolation needed. */
            mc1_l = matrix_coeffs[s1 * 2    ];
            mc1_r = matrix_coeffs[s1 * 2 + 1];
            mc2_l = matrix_coeffs[s2 * 2    ];
            mc2_r = matrix_coeffs[s2 * 2 + 1];

            /* Interpolation is done over the first eight samples. */
            for (; nsample < band + 8; nsample++) {
                float c1 = su1[nsample];
                float c2 = su2[nsample];
                c2 = c1 * INTERPOLATE(mc1_l, mc2_l, nsample - band) +
                     c2 * INTERPOLATE(mc1_r, mc2_r, nsample - band);
                su1[nsample] = c2;
                su2[nsample] = c1 * 2.0 - c2;
            }
        }

        /* Apply the matrix without interpolation. */
        switch (s2) {
        case 0:     /* M/S decoding */
            for (; nsample < band + 256; nsample++) {
                float c1 = su1[nsample];
                float c2 = su2[nsample];
                su1[nsample] =  c2       * 2.0;
                su2[nsample] = (c1 - c2) * 2.0;
            }
            break;
        case 1:
            for (; nsample < band + 256; nsample++) {
                float c1 = su1[nsample];
                float c2 = su2[nsample];
                su1[nsample] = (c1 + c2) *  2.0;
                su2[nsample] =  c2       * -2.0;
            }
            break;
        case 2:
        case 3:
            for (; nsample < band + 256; nsample++) {
                float c1 = su1[nsample];
                float c2 = su2[nsample];
                su1[nsample] = c1 + c2;
                su2[nsample] = c1 - c2;
            }
            break;
        default:
            av_assert1(0);
        }
    }
}

static void get_channel_weights(int index, int flag, float ch[2])
{
    if (index == 7) {
        ch[0] = 1.0;
        ch[1] = 1.0;
    } else {
        ch[0] = (index & 7) / 7.0;
        ch[1] = sqrt(2 - ch[0] * ch[0]);
        if (flag)
            FFSWAP(float, ch[0], ch[1]);
    }
}

static void channel_weighting(float *su1, float *su2, int *p3)
{
    int band, nsample;
    /* w[x][y] y=0 is left y=1 is right */
    float w[2][2];

    if (p3[1] != 7 || p3[3] != 7) {
        get_channel_weights(p3[1], p3[0], w[0]);
        get_channel_weights(p3[3], p3[2], w[1]);

        for (band = 256; band < 4 * 256; band += 256) {
            for (nsample = band; nsample < band + 8; nsample++) {
                su1[nsample] *= INTERPOLATE(w[0][0], w[0][1], nsample - band);
                su2[nsample] *= INTERPOLATE(w[1][0], w[1][1], nsample - band);
            }
            for(; nsample < band + 256; nsample++) {
                su1[nsample] *= w[1][0];
                su2[nsample] *= w[1][1];
            }
        }
    }
}

static int decode_channel_sound_unit(ATRAC3Context *q, GetBitContext *gb,
                                     ChannelUnit *snd, float *output,
                                     int channel_num, int coding_mode)
{
    int band, ret, num_subbands, last_tonal, num_bands;
    GainBlock *gain1 = &snd->gain_block[    snd->gc_blk_switch];
    GainBlock *gain2 = &snd->gain_block[1 - snd->gc_blk_switch];

    if (coding_mode == JOINT_STEREO && channel_num == 1) {
        if (get_bits(gb, 2) != 3) {
            av_log(NULL,AV_LOG_ERROR,"JS mono Sound Unit id != 3.\n");
            return AVERROR_INVALIDDATA;
        }
    } else {
        if (get_bits(gb, 6) != 0x28) {
            av_log(NULL,AV_LOG_ERROR,"Sound Unit id != 0x28.\n");
            return AVERROR_INVALIDDATA;
        }
    }

    /* number of coded QMF bands */
    snd->bands_coded = get_bits(gb, 2);

    ret = decode_gain_control(gb, gain2, snd->bands_coded);
    if (ret)
        return ret;

    snd->num_components = decode_tonal_components(gb, snd->components,
                                                  snd->bands_coded);
    if (snd->num_components == -1)
        return -1;

    num_subbands = decode_spectrum(gb, snd->spectrum);

    /* Merge the decoded spectrum and tonal components. */
    last_tonal = add_tonal_components(snd->spectrum, snd->num_components,
                                      snd->components);


    /* calculate number of used MLT/QMF bands according to the amount of coded
       spectral lines */
    num_bands = (subband_tab[num_subbands] - 1) >> 8;
    if (last_tonal >= 0)
        num_bands = FFMAX((last_tonal + 256) >> 8, num_bands);


    /* Reconstruct time domain samples. */
    for (band = 0; band < 4; band++) {
        /* Perform the IMDCT step without overlapping. */
        if (band <= num_bands)
            imlt(q, &snd->spectrum[band * 256], snd->imdct_buf, band & 1);
        else
            memset(snd->imdct_buf, 0, 512 * sizeof(*snd->imdct_buf));

        /* gain compensation and overlapping */
        gain_compensate_and_overlap(snd->imdct_buf,
                                    &snd->prev_frame[band * 256],
                                    &output[band * 256],
                                    &gain1->g_block[band],
                                    &gain2->g_block[band]);
    }

    /* Swap the gain control buffers for the next frame. */
    snd->gc_blk_switch ^= 1;

    return 0;
}

static int decode_frame(AVCodecContext *avctx, const uint8_t *databuf,
                        float **out_samples)
{
    ATRAC3Context *q = avctx->priv_data;
    int ret, i;
    uint8_t *ptr1;

    if (q->coding_mode == JOINT_STEREO) {
        /* channel coupling mode */
        /* decode Sound Unit 1 */
        init_get_bits(&q->gb, databuf, avctx->block_align * 8);

        ret = decode_channel_sound_unit(q, &q->gb, q->units, out_samples[0], 0,
                                        JOINT_STEREO);
        if (ret != 0)
            return ret;

        /* Framedata of the su2 in the joint-stereo mode is encoded in
         * reverse byte order so we need to swap it first. */
        if (databuf == q->decoded_bytes_buffer) {
            uint8_t *ptr2 = q->decoded_bytes_buffer + avctx->block_align - 1;
            ptr1          = q->decoded_bytes_buffer;
            for (i = 0; i < avctx->block_align / 2; i++, ptr1++, ptr2--)
                FFSWAP(uint8_t, *ptr1, *ptr2);
        } else {
            const uint8_t *ptr2 = databuf + avctx->block_align - 1;
            for (i = 0; i < avctx->block_align; i++)
                q->decoded_bytes_buffer[i] = *ptr2--;
        }

        /* Skip the sync codes (0xF8). */
        ptr1 = q->decoded_bytes_buffer;
        for (i = 4; *ptr1 == 0xF8; i++, ptr1++) {
            if (i >= avctx->block_align)
                return AVERROR_INVALIDDATA;
        }


        /* set the bitstream reader at the start of the second Sound Unit*/
        init_get_bits(&q->gb, ptr1, avctx->block_align * 8);

        /* Fill the Weighting coeffs delay buffer */
        memmove(q->weighting_delay, &q->weighting_delay[2],
                4 * sizeof(*q->weighting_delay));
        q->weighting_delay[4] = get_bits1(&q->gb);
        q->weighting_delay[5] = get_bits(&q->gb, 3);

        for (i = 0; i < 4; i++) {
            q->matrix_coeff_index_prev[i] = q->matrix_coeff_index_now[i];
            q->matrix_coeff_index_now[i]  = q->matrix_coeff_index_next[i];
            q->matrix_coeff_index_next[i] = get_bits(&q->gb, 2);
        }

        /* Decode Sound Unit 2. */
        ret = decode_channel_sound_unit(q, &q->gb, &q->units[1],
                                        out_samples[1], 1, JOINT_STEREO);
        if (ret != 0)
            return ret;

        /* Reconstruct the channel coefficients. */
        reverse_matrixing(out_samples[0], out_samples[1],
                          q->matrix_coeff_index_prev,
                          q->matrix_coeff_index_now);

        channel_weighting(out_samples[0], out_samples[1], q->weighting_delay);
    } else {
        /* normal stereo mode or mono */
        /* Decode the channel sound units. */
        for (i = 0; i < avctx->channels; i++) {
            /* Set the bitstream reader at the start of a channel sound unit. */
            init_get_bits(&q->gb,
                          databuf + i * avctx->block_align / avctx->channels,
                          avctx->block_align * 8 / avctx->channels);

            ret = decode_channel_sound_unit(q, &q->gb, &q->units[i],
                                            out_samples[i], i, q->coding_mode);
            if (ret != 0)
                return ret;
        }
    }

    /* Apply the iQMF synthesis filter. */
    for (i = 0; i < avctx->channels; i++) {
        float *p1 = out_samples[i];
        float *p2 = p1 + 256;
        float *p3 = p2 + 256;
        float *p4 = p3 + 256;
        ff_atrac_iqmf(p1, p2, 256, p1, q->units[i].delay_buf1, q->temp_buf);
        ff_atrac_iqmf(p4, p3, 256, p3, q->units[i].delay_buf2, q->temp_buf);
        ff_atrac_iqmf(p1, p3, 512, p1, q->units[i].delay_buf3, q->temp_buf);
    }

    return 0;
}

static int atrac3_decode_frame(AVCodecContext *avctx, void *data,
                               int *got_frame_ptr, AVPacket *avpkt)
{
    const uint8_t *buf = avpkt->data;
    int buf_size = avpkt->size;
    ATRAC3Context *q = avctx->priv_data;
    int ret;
    const uint8_t *databuf;

    if (buf_size < avctx->block_align) {
        av_log(avctx, AV_LOG_ERROR,
               "Frame too small (%d bytes). Truncated file?\n", buf_size);
        return AVERROR_INVALIDDATA;
    }

    /* get output buffer */
    q->frame.nb_samples = SAMPLES_PER_FRAME;
    if ((ret = ff_get_buffer(avctx, &q->frame)) < 0) {
        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
        return ret;
    }

    /* Check if we need to descramble and what buffer to pass on. */
    if (q->scrambled_stream) {
        decode_bytes(buf, q->decoded_bytes_buffer, avctx->block_align);
        databuf = q->decoded_bytes_buffer;
    } else {
        databuf = buf;
    }

    ret = decode_frame(avctx, databuf, (float **)q->frame.extended_data);
    if (ret) {
        av_log(NULL, AV_LOG_ERROR, "Frame decoding error!\n");
        return ret;
    }

    *got_frame_ptr   = 1;
    *(AVFrame *)data = q->frame;

    return avctx->block_align;
}

static void atrac3_init_static_data(void)
{
    int i;

    init_atrac3_window();
    ff_atrac_generate_tables();

    /* Initialize the VLC tables. */
    for (i = 0; i < 7; i++) {
        spectral_coeff_tab[i].table = &atrac3_vlc_table[atrac3_vlc_offs[i]];
        spectral_coeff_tab[i].table_allocated = atrac3_vlc_offs[i + 1] -
                                                atrac3_vlc_offs[i    ];
        init_vlc(&spectral_coeff_tab[i], 9, huff_tab_sizes[i],
                 huff_bits[i],  1, 1,
                 huff_codes[i], 1, 1, INIT_VLC_USE_NEW_STATIC);
    }

    /* Generate gain tables */
    for (i = 0; i < 16; i++)
        gain_tab1[i] = exp2f (4 - i);

    for (i = -15; i < 16; i++)
        gain_tab2[i + 15] = exp2f (i * -0.125);
}

static av_cold int atrac3_decode_init(AVCodecContext *avctx)
{
    static int static_init_done;
    int i, ret;
    int version, delay, samples_per_frame, frame_factor;
    const uint8_t *edata_ptr = avctx->extradata;
    ATRAC3Context *q = avctx->priv_data;

    if (avctx->channels <= 0 || avctx->channels > 2) {
        av_log(avctx, AV_LOG_ERROR, "Channel configuration error!\n");
        return AVERROR(EINVAL);
    }

    if (!static_init_done)
        atrac3_init_static_data();
    static_init_done = 1;

    /* Take care of the codec-specific extradata. */
    if (avctx->extradata_size == 14) {
        /* Parse the extradata, WAV format */
        av_log(avctx, AV_LOG_DEBUG, "[0-1] %d\n",
               bytestream_get_le16(&edata_ptr));  // Unknown value always 1
        edata_ptr += 4;                             // samples per channel
        q->coding_mode = bytestream_get_le16(&edata_ptr);
        av_log(avctx, AV_LOG_DEBUG,"[8-9] %d\n",
               bytestream_get_le16(&edata_ptr));  //Dupe of coding mode
        frame_factor = bytestream_get_le16(&edata_ptr);  // Unknown always 1
        av_log(avctx, AV_LOG_DEBUG,"[12-13] %d\n",
               bytestream_get_le16(&edata_ptr));  // Unknown always 0

        /* setup */
        samples_per_frame    = SAMPLES_PER_FRAME * avctx->channels;
        version              = 4;
        delay                = 0x88E;
        q->coding_mode       = q->coding_mode ? JOINT_STEREO : STEREO;
        q->scrambled_stream  = 0;

        if (avctx->block_align !=  96 * avctx->channels * frame_factor &&
            avctx->block_align != 152 * avctx->channels * frame_factor &&
            avctx->block_align != 192 * avctx->channels * frame_factor) {
            av_log(avctx, AV_LOG_ERROR, "Unknown frame/channel/frame_factor "
                   "configuration %d/%d/%d\n", avctx->block_align,
                   avctx->channels, frame_factor);
            return AVERROR_INVALIDDATA;
        }
    } else if (avctx->extradata_size == 10) {
        /* Parse the extradata, RM format. */
        version                = bytestream_get_be32(&edata_ptr);
        samples_per_frame      = bytestream_get_be16(&edata_ptr);
        delay                  = bytestream_get_be16(&edata_ptr);
        q->coding_mode         = bytestream_get_be16(&edata_ptr);
        q->scrambled_stream    = 1;

    } else {
        av_log(NULL, AV_LOG_ERROR, "Unknown extradata size %d.\n",
               avctx->extradata_size);
        return AVERROR(EINVAL);
    }

    if (q->coding_mode == JOINT_STEREO && avctx->channels < 2) {
        av_log(avctx, AV_LOG_ERROR, "Invalid coding mode\n");
        return AVERROR_INVALIDDATA;
    }

    /* Check the extradata */

    if (version != 4) {
        av_log(avctx, AV_LOG_ERROR, "Version %d != 4.\n", version);
        return AVERROR_INVALIDDATA;
    }

    if (samples_per_frame != SAMPLES_PER_FRAME &&
        samples_per_frame != SAMPLES_PER_FRAME * 2) {
        av_log(avctx, AV_LOG_ERROR, "Unknown amount of samples per frame %d.\n",
               samples_per_frame);
        return AVERROR_INVALIDDATA;
    }

    if (delay != 0x88E) {
        av_log(avctx, AV_LOG_ERROR, "Unknown amount of delay %x != 0x88E.\n",
               delay);
        return AVERROR_INVALIDDATA;
    }

    if (q->coding_mode == STEREO)
        av_log(avctx, AV_LOG_DEBUG, "Normal stereo detected.\n");
    else if (q->coding_mode == JOINT_STEREO)
        av_log(avctx, AV_LOG_DEBUG, "Joint stereo detected.\n");
    else {
        av_log(avctx, AV_LOG_ERROR, "Unknown channel coding mode %x!\n",
               q->coding_mode);
        return AVERROR_INVALIDDATA;
    }

    if (avctx->block_align >= UINT_MAX / 2)
        return AVERROR(EINVAL);

    q->decoded_bytes_buffer = av_mallocz(FFALIGN(avctx->block_align, 4) +
                                         FF_INPUT_BUFFER_PADDING_SIZE);
    if (q->decoded_bytes_buffer == NULL)
        return AVERROR(ENOMEM);

    avctx->sample_fmt = AV_SAMPLE_FMT_FLTP;

    /* initialize the MDCT transform */
    if ((ret = ff_mdct_init(&q->mdct_ctx, 9, 1, 1.0 / 32768)) < 0) {
        av_log(avctx, AV_LOG_ERROR, "Error initializing MDCT\n");
        av_freep(&q->decoded_bytes_buffer);
        return ret;
    }

    /* init the joint-stereo decoding data */
    q->weighting_delay[0] = 0;
    q->weighting_delay[1] = 7;
    q->weighting_delay[2] = 0;
    q->weighting_delay[3] = 7;
    q->weighting_delay[4] = 0;
    q->weighting_delay[5] = 7;

    for (i = 0; i < 4; i++) {
        q->matrix_coeff_index_prev[i] = 3;
        q->matrix_coeff_index_now[i]  = 3;
        q->matrix_coeff_index_next[i] = 3;
    }

    avpriv_float_dsp_init(&q->fdsp, avctx->flags & CODEC_FLAG_BITEXACT);
    ff_fmt_convert_init(&q->fmt_conv, avctx);

    q->units = av_mallocz(sizeof(*q->units) * avctx->channels);
    if (!q->units) {
        atrac3_decode_close(avctx);
        return AVERROR(ENOMEM);
    }

    avcodec_get_frame_defaults(&q->frame);
    avctx->coded_frame = &q->frame;

    return 0;
}

AVCodec ff_atrac3_decoder = {
    .name             = "atrac3",
    .type             = AVMEDIA_TYPE_AUDIO,
    .id               = AV_CODEC_ID_ATRAC3,
    .priv_data_size   = sizeof(ATRAC3Context),
    .init             = atrac3_decode_init,
    .close            = atrac3_decode_close,
    .decode           = atrac3_decode_frame,
    .capabilities     = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
    .long_name        = NULL_IF_CONFIG_SMALL("Atrac 3 (Adaptive TRansform Acoustic Coding 3)"),
    .sample_fmts      = (const enum AVSampleFormat[]) { AV_SAMPLE_FMT_FLTP,
                                                        AV_SAMPLE_FMT_NONE },
};

---