annotate twinvq.c @ 11032:01bd040f8607 libavcodec

Unroll main loop so the edge==0 case is seperate. This allows many things to be simplified away. h264 decoder is overall 1% faster with a mbaff sample and 0.1% slower with the cathedral sample, probably because the slow loop filter code must be loaded into the code cache for each first MB of each row but isnt used for the following MBs.

/*
 * TwinVQ decoder
 * Copyright (c) 2009 Vitor Sessak
 *
 * 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 "get_bits.h"
#include "dsputil.h"
#include "lsp.h"

#include <math.h>
#include <stdint.h>

#include "twinvq_data.h"

enum FrameType {
    FT_SHORT = 0,  ///< Short frame  (divided in n   sub-blocks)
    FT_MEDIUM,     ///< Medium frame (divided in m<n sub-blocks)
    FT_LONG,       ///< Long frame   (single sub-block + PPC)
    FT_PPC,        ///< Periodic Peak Component (part of the long frame)
};

/**
 * Parameters and tables that are different for each frame type
 */
struct FrameMode {
    uint8_t         sub;      ///< Number subblocks in each frame
    const uint16_t *bark_tab;

    /** number of distinct bark scale envelope values */
    uint8_t         bark_env_size;

    const int16_t  *bark_cb;    ///< codebook for the bark scale envelope (BSE)
    uint8_t         bark_n_coef;///< number of BSE CB coefficients to read
    uint8_t         bark_n_bit; ///< number of bits of the BSE coefs

    //@{
    /** main codebooks for spectrum data */
    const int16_t    *cb0;
    const int16_t    *cb1;
    //@}

    uint8_t         cb_len_read; ///< number of spectrum coefficients to read
};

/**
 * Parameters and tables that are different for every combination of
 * bitrate/sample rate
 */
typedef struct {
    struct FrameMode fmode[3]; ///< frame type-dependant parameters

    uint16_t     size;        ///< frame size in samples
    uint8_t      n_lsp;       ///< number of lsp coefficients
    const float *lspcodebook;

    /* number of bits of the different LSP CB coefficients */
    uint8_t      lsp_bit0;
    uint8_t      lsp_bit1;
    uint8_t      lsp_bit2;

    uint8_t      lsp_split;      ///< number of CB entries for the LSP decoding
    const int16_t *ppc_shape_cb; ///< PPC shape CB

    /** number of the bits for the PPC period value */
    uint8_t      ppc_period_bit;

    uint8_t      ppc_shape_bit;  ///< number of bits of the PPC shape CB coeffs
    uint8_t      ppc_shape_len;  ///< size of PPC shape CB
    uint8_t      pgain_bit;      ///< bits for PPC gain

    /** constant for peak period to peak width conversion */
    uint16_t     peak_per2wid;
} ModeTab;

static const ModeTab mode_08_08 = {
    {
        { 8, bark_tab_s08_64,  10, tab.fcb08s  , 1, 5, tab.cb0808s0, tab.cb0808s1, 18},
        { 2, bark_tab_m08_256, 20, tab.fcb08m  , 2, 5, tab.cb0808m0, tab.cb0808m1, 16},
        { 1, bark_tab_l08_512, 30, tab.fcb08l  , 3, 6, tab.cb0808l0, tab.cb0808l1, 17}
    },
    512 , 12, tab.lsp08,   1, 5, 3, 3, tab.shape08  , 8, 28, 20, 6, 40
};

static const ModeTab mode_11_08 = {
    {
        { 8, bark_tab_s11_64,  10, tab.fcb11s  , 1, 5, tab.cb1108s0, tab.cb1108s1, 29},
        { 2, bark_tab_m11_256, 20, tab.fcb11m  , 2, 5, tab.cb1108m0, tab.cb1108m1, 24},
        { 1, bark_tab_l11_512, 30, tab.fcb11l  , 3, 6, tab.cb1108l0, tab.cb1108l1, 27}
    },
    512 , 16, tab.lsp11,   1, 6, 4, 3, tab.shape11  , 9, 36, 30, 7, 90
};

static const ModeTab mode_11_10 = {
    {
        { 8, bark_tab_s11_64,  10, tab.fcb11s  , 1, 5, tab.cb1110s0, tab.cb1110s1, 21},
        { 2, bark_tab_m11_256, 20, tab.fcb11m  , 2, 5, tab.cb1110m0, tab.cb1110m1, 18},
        { 1, bark_tab_l11_512, 30, tab.fcb11l  , 3, 6, tab.cb1110l0, tab.cb1110l1, 20}
    },
    512 , 16, tab.lsp11,   1, 6, 4, 3, tab.shape11  , 9, 36, 30, 7, 90
};

static const ModeTab mode_16_16 = {
    {
        { 8, bark_tab_s16_128, 10, tab.fcb16s  , 1, 5, tab.cb1616s0, tab.cb1616s1, 16},
        { 2, bark_tab_m16_512, 20, tab.fcb16m  , 2, 5, tab.cb1616m0, tab.cb1616m1, 15},
        { 1, bark_tab_l16_1024,30, tab.fcb16l  , 3, 6, tab.cb1616l0, tab.cb1616l1, 16}
    },
    1024, 16, tab.lsp16,   1, 6, 4, 3, tab.shape16  , 9, 56, 60, 7, 180
};

static const ModeTab mode_22_20 = {
    {
        { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2220s0, tab.cb2220s1, 18},
        { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2220m0, tab.cb2220m1, 17},
        { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2220l0, tab.cb2220l1, 18}
    },
    1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
};

static const ModeTab mode_22_24 = {
    {
        { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2224s0, tab.cb2224s1, 15},
        { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2224m0, tab.cb2224m1, 14},
        { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2224l0, tab.cb2224l1, 15}
    },
    1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
};

static const ModeTab mode_22_32 = {
    {
        { 4, bark_tab_s22_128, 10, tab.fcb22s_2, 1, 6, tab.cb2232s0, tab.cb2232s1, 11},
        { 2, bark_tab_m22_256, 20, tab.fcb22m_2, 2, 6, tab.cb2232m0, tab.cb2232m1, 11},
        { 1, bark_tab_l22_512, 32, tab.fcb22l_2, 4, 6, tab.cb2232l0, tab.cb2232l1, 12}
    },
    512 , 16, tab.lsp22_2, 1, 6, 4, 4, tab.shape22_2, 9, 56, 36, 7, 72
};

static const ModeTab mode_44_40 = {
    {
        {16, bark_tab_s44_128, 10, tab.fcb44s  , 1, 6, tab.cb4440s0, tab.cb4440s1, 18},
        { 4, bark_tab_m44_512, 20, tab.fcb44m  , 2, 6, tab.cb4440m0, tab.cb4440m1, 17},
        { 1, bark_tab_l44_2048,40, tab.fcb44l  , 4, 6, tab.cb4440l0, tab.cb4440l1, 17}
    },
    2048, 20, tab.lsp44,   1, 6, 4, 4, tab.shape44  , 9, 84, 54, 7, 432
};

static const ModeTab mode_44_48 = {
    {
        {16, bark_tab_s44_128, 10, tab.fcb44s  , 1, 6, tab.cb4448s0, tab.cb4448s1, 15},
        { 4, bark_tab_m44_512, 20, tab.fcb44m  , 2, 6, tab.cb4448m0, tab.cb4448m1, 14},
        { 1, bark_tab_l44_2048,40, tab.fcb44l  , 4, 6, tab.cb4448l0, tab.cb4448l1, 14}
    },
    2048, 20, tab.lsp44,   1, 6, 4, 4, tab.shape44  , 9, 84, 54, 7, 432
};

typedef struct TwinContext {
    AVCodecContext *avctx;
    DSPContext      dsp;
    FFTContext mdct_ctx[3];

    const ModeTab *mtab;

    // history
    float lsp_hist[2][20];           ///< LSP coefficients of the last frame
    float bark_hist[3][2][40];       ///< BSE coefficients of last frame

    // bitstream parameters
    int16_t permut[4][4096];
    uint8_t length[4][2];            ///< main codebook stride
    uint8_t length_change[4];
    uint8_t bits_main_spec[2][4][2]; ///< bits for the main codebook
    int bits_main_spec_change[4];
    int n_div[4];

    float *spectrum;
    float *curr_frame;               ///< non-interleaved output
    float *prev_frame;               ///< non-interleaved previous frame
    int last_block_pos[2];

    float *cos_tabs[3];

    // scratch buffers
    float *tmp_buf;
} TwinContext;

#define PPC_SHAPE_CB_SIZE 64
#define SUB_AMP_MAX       4500.0
#define MULAW_MU          100.0
#define GAIN_BITS         8
#define AMP_MAX           13000.0
#define SUB_GAIN_BITS     5
#define WINDOW_TYPE_BITS  4
#define PGAIN_MU          200

/** @note not speed critical, hence not optimized */
static void memset_float(float *buf, float val, int size)
{
    while (size--)
        *buf++ = val;
}

/**
 * Evaluate a single LPC amplitude spectrum envelope coefficient from the line
 * spectrum pairs.
 *
 * @param lsp a vector of the cosinus of the LSP values
 * @param cos_val cos(PI*i/N) where i is the index of the LPC amplitude
 * @param order the order of the LSP (and the size of the *lsp buffer). Must
 *        be a multiple of four.
 * @return the LPC value
 *
 * @todo reuse code from vorbis_dec.c: vorbis_floor0_decode
 */
static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
{
    int j;
    float p = 0.5f;
    float q = 0.5f;
    float two_cos_w = 2.0f*cos_val;

    for (j = 0; j + 1 < order; j += 2*2) {
        // Unroll the loop once since order is a multiple of four
        q *= lsp[j  ] - two_cos_w;
        p *= lsp[j+1] - two_cos_w;

        q *= lsp[j+2] - two_cos_w;
        p *= lsp[j+3] - two_cos_w;
    }

    p *= p * (2.0f - two_cos_w);
    q *= q * (2.0f + two_cos_w);

    return 0.5 / (p + q);
}

/**
 * Evaluates the LPC amplitude spectrum envelope from the line spectrum pairs.
 */
static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc)
{
    int i;
    const ModeTab *mtab = tctx->mtab;
    int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;

    for (i = 0; i < size_s/2; i++) {
        float cos_i = tctx->cos_tabs[0][i];
        lpc[i]          = eval_lpc_spectrum(cos_vals,  cos_i, mtab->n_lsp);
        lpc[size_s-i-1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
    }
}

static void interpolate(float *out, float v1, float v2, int size)
{
    int i;
    float step = (v1 - v2)/(size + 1);

    for (i = 0; i < size; i++) {
        v2 += step;
        out[i] = v2;
    }
}

static inline float get_cos(int idx, int part, const float *cos_tab, int size)
{
    return part ? -cos_tab[size - idx - 1] :
                   cos_tab[       idx    ];
}

/**
 * Evaluates the LPC amplitude spectrum envelope from the line spectrum pairs.
 * Probably for speed reasons, the coefficients are evaluated as
 * siiiibiiiisiiiibiiiisiiiibiiiisiiiibiiiis ...
 * where s is an evaluated value, i is a value interpolated from the others
 * and b might be either calculated or interpolated, depending on an
 * unexplained condition.
 *
 * @param step the size of a block "siiiibiiii"
 * @param in the cosinus of the LSP data
 * @param part is 0 for 0...PI (positive cossinus values) and 1 for PI...2PI
          (negative cossinus values)
 * @param size the size of the whole output
 */
static inline void eval_lpcenv_or_interp(TwinContext *tctx,
                                         enum FrameType ftype,
                                         float *out, const float *in,
                                         int size, int step, int part)
{
    int i;
    const ModeTab *mtab = tctx->mtab;
    const float *cos_tab = tctx->cos_tabs[ftype];

    // Fill the 's'
    for (i = 0; i < size; i += step)
        out[i] =
            eval_lpc_spectrum(in,
                              get_cos(i, part, cos_tab, size),
                              mtab->n_lsp);

    // Fill the 'iiiibiiii'
    for (i = step; i <= size - 2*step; i += step) {
        if (out[i + step] + out[i - step] >  1.95*out[i] ||
            out[i + step]                 >=  out[i - step]) {
            interpolate(out + i - step + 1, out[i], out[i-step], step - 1);
        } else {
            out[i - step/2] =
                eval_lpc_spectrum(in,
                                  get_cos(i-step/2, part, cos_tab, size),
                                  mtab->n_lsp);
            interpolate(out + i - step   + 1, out[i-step/2], out[i-step  ], step/2 - 1);
            interpolate(out + i - step/2 + 1, out[i       ], out[i-step/2], step/2 - 1);
        }
    }

    interpolate(out + size - 2*step + 1, out[size-step], out[size - 2*step], step - 1);
}

static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype,
                               const float *buf, float *lpc,
                               int size, int step)
{
    eval_lpcenv_or_interp(tctx, ftype, lpc         , buf, size/2,   step, 0);
    eval_lpcenv_or_interp(tctx, ftype, lpc + size/2, buf, size/2, 2*step, 1);

    interpolate(lpc+size/2-step+1, lpc[size/2], lpc[size/2-step], step);

    memset_float(lpc + size - 2*step + 1, lpc[size - 2*step], 2*step - 1);
}

/**
 * Inverse quantization. Read CB coefficients for cb1 and cb2 from the
 * bitstream, sum the corresponding vectors and write the result to *out
 * after permutation.
 */
static void dequant(TwinContext *tctx, GetBitContext *gb, float *out,
                    enum FrameType ftype,
                    const int16_t *cb0, const int16_t *cb1, int cb_len)
{
    int pos = 0;
    int i, j;

    for (i = 0; i < tctx->n_div[ftype]; i++) {
        int tmp0, tmp1;
        int sign0 = 1;
        int sign1 = 1;
        const int16_t *tab0, *tab1;
        int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
        int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);

        int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
        if (bits == 7) {
            if (get_bits1(gb))
                sign0 = -1;
            bits = 6;
        }
        tmp0 = get_bits(gb, bits);

        bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];

        if (bits == 7) {
            if (get_bits1(gb))
                sign1 = -1;

            bits = 6;
        }
        tmp1 = get_bits(gb, bits);

        tab0 = cb0 + tmp0*cb_len;
        tab1 = cb1 + tmp1*cb_len;

        for (j = 0; j < length; j++)
            out[tctx->permut[ftype][pos+j]] = sign0*tab0[j] + sign1*tab1[j];

        pos += length;
    }

}

static inline float mulawinv(float y, float clip, float mu)
{
    y = av_clipf(y/clip, -1, 1);
    return clip * FFSIGN(y) * (exp(log(1+mu) * fabs(y)) - 1) / mu;
}

/**
 * Evaluate a*b/400 rounded to the nearest integer. When, for example,
 * a*b == 200 and the nearest integer is ill-defined, use a table to emulate
 * the following broken float-based implementation used by the binary decoder:
 *
 * \code
 * static int very_broken_op(int a, int b)
 * {
 *    static float test; // Ugh, force gcc to do the division first...
 *
 *    test = a/400.;
 *    return b * test +  0.5;
 * }
 * \endcode
 *
 * @note if this function is replaced by just ROUNDED_DIV(a*b,400.), the stddev
 * between the original file (before encoding with Yamaha encoder) and the
 * decoded output increases, which leads one to believe that the encoder expects
 * exactly this broken calculation.
 */
static int very_broken_op(int a, int b)
{
    int x = a*b + 200;
    int size;
    const uint8_t *rtab;

    if (x%400 || b%5)
        return x/400;

    x /= 400;

    size = tabs[b/5].size;
    rtab = tabs[b/5].tab;
    return x - rtab[size*av_log2(2*(x - 1)/size)+(x - 1)%size];
}

/**
 * Sum to data a periodic peak of a given period, width and shape.
 *
 * @param period the period of the peak divised by 400.0
 */
static void add_peak(int period, int width, const float *shape,
                     float ppc_gain, float *speech, int len)
{
    int i, j;

    const float *shape_end = shape + len;
    int center;

    // First peak centered around zero
    for (i = 0; i < width/2; i++)
        speech[i] += ppc_gain * *shape++;

    for (i = 1; i < ROUNDED_DIV(len,width) ; i++) {
        center = very_broken_op(period, i);
        for (j = -width/2; j < (width+1)/2; j++)
            speech[j+center] += ppc_gain * *shape++;
    }

    // For the last block, be careful not to go beyond the end of the buffer
    center = very_broken_op(period, i);
    for (j = -width/2; j < (width + 1)/2 && shape < shape_end; j++)
        speech[j+center] += ppc_gain * *shape++;
}

static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape,
                       float ppc_gain, float *speech)
{
    const ModeTab *mtab = tctx->mtab;
    int isampf = tctx->avctx->sample_rate/1000;
    int ibps = tctx->avctx->bit_rate/(1000 * tctx->avctx->channels);
    int min_period = ROUNDED_DIV(  40*2*mtab->size, isampf);
    int max_period = ROUNDED_DIV(6*40*2*mtab->size, isampf);
    int period_range = max_period - min_period;

    // This is actually the period multiplied by 400. It is just linearly coded
    // between its maximum and minimum value.
    int period = min_period +
        ROUNDED_DIV(period_coef*period_range, (1 << mtab->ppc_period_bit) - 1);
    int width;

    if (isampf == 22 && ibps == 32) {
        // For some unknown reason, NTT decided to code this case differently...
        width = ROUNDED_DIV((period + 800)* mtab->peak_per2wid, 400*mtab->size);
    } else
        width =             (period      )* mtab->peak_per2wid/(400*mtab->size);

    add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len);
}

static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype,
                     float *out)
{
    const ModeTab *mtab = tctx->mtab;
    int i, j;
    int sub = mtab->fmode[ftype].sub;
    float step     = AMP_MAX     / ((1 <<     GAIN_BITS) - 1);
    float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1);

    if (ftype == FT_LONG) {
        for (i = 0; i < tctx->avctx->channels; i++)
            out[i] = (1./(1<<13)) *
                mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
                         AMP_MAX, MULAW_MU);
    } else {
        for (i = 0; i < tctx->avctx->channels; i++) {
            float val = (1./(1<<23)) *
                mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
                         AMP_MAX, MULAW_MU);

            for (j = 0; j < sub; j++) {
                out[i*sub + j] =
                    val*mulawinv(sub_step* 0.5 +
                                 sub_step* get_bits(gb, SUB_GAIN_BITS),
                                 SUB_AMP_MAX, MULAW_MU);
            }
        }
    }
}

/**
 * Rearrange the LSP coefficients so that they have a minimum distance of
 * min_dist. This function does it exactly as described in section of 3.2.4
 * of the G.729 specification (but interestingly is different from what the
 * reference decoder actually does).
 */
static void rearrange_lsp(int order, float *lsp, float min_dist)
{
    int i;
    float min_dist2 = min_dist * 0.5;
    for (i = 1; i < order; i++)
        if (lsp[i] - lsp[i-1] < min_dist) {
            float avg = (lsp[i] + lsp[i-1]) * 0.5;

            lsp[i-1] = avg - min_dist2;
            lsp[i  ] = avg + min_dist2;
        }
}

static void decode_lsp(TwinContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
                       int lpc_hist_idx, float *lsp, float *hist)
{
    const ModeTab *mtab = tctx->mtab;
    int i, j;

    const float *cb  =  mtab->lspcodebook;
    const float *cb2 =  cb  + (1 << mtab->lsp_bit1)*mtab->n_lsp;
    const float *cb3 =  cb2 + (1 << mtab->lsp_bit2)*mtab->n_lsp;

    const int8_t funny_rounding[4] = {
        -2,
        mtab->lsp_split == 4 ? -2 : 1,
        mtab->lsp_split == 4 ? -2 : 1,
        0
    };

    j = 0;
    for (i = 0; i < mtab->lsp_split; i++) {
        int chunk_end = ((i + 1)*mtab->n_lsp + funny_rounding[i])/mtab->lsp_split;
        for (; j < chunk_end; j++)
            lsp[j] = cb [lpc_idx1    * mtab->n_lsp + j] +
                     cb2[lpc_idx2[i] * mtab->n_lsp + j];
    }

    rearrange_lsp(mtab->n_lsp, lsp, 0.0001);

    for (i = 0; i < mtab->n_lsp; i++) {
        float tmp1 = 1. -          cb3[lpc_hist_idx*mtab->n_lsp + i];
        float tmp2 =     hist[i] * cb3[lpc_hist_idx*mtab->n_lsp + i];
        hist[i] = lsp[i];
        lsp[i]  = lsp[i] * tmp1 + tmp2;
    }

    rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
    rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
    ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
}

static void dec_lpc_spectrum_inv(TwinContext *tctx, float *lsp,
                                 enum FrameType ftype, float *lpc)
{
    int i;
    int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;

    for (i = 0; i < tctx->mtab->n_lsp; i++)
        lsp[i] =  2*cos(lsp[i]);

    switch (ftype) {
    case FT_LONG:
        eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
        break;
    case FT_MEDIUM:
        eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
        break;
    case FT_SHORT:
        eval_lpcenv(tctx, lsp, lpc);
        break;
    }
}

static void imdct_and_window(TwinContext *tctx, enum FrameType ftype, int wtype,
                            float *in, float *prev, int ch)
{
    const ModeTab *mtab = tctx->mtab;
    int bsize = mtab->size / mtab->fmode[ftype].sub;
    int size  = mtab->size;
    float *buf1 = tctx->tmp_buf;
    int j;
    int wsize; // Window size
    float *out = tctx->curr_frame + 2*ch*mtab->size;
    float *out2 = out;
    float *prev_buf;
    int first_wsize;

    static const uint8_t wtype_to_wsize[]      = {0, 0, 2, 2, 2, 1, 0, 1, 1};
    int types_sizes[] = {
        mtab->size /    mtab->fmode[FT_LONG  ].sub,
        mtab->size /    mtab->fmode[FT_MEDIUM].sub,
        mtab->size / (2*mtab->fmode[FT_SHORT ].sub),
    };

    wsize = types_sizes[wtype_to_wsize[wtype]];
    first_wsize = wsize;
    prev_buf = prev + (size - bsize)/2;

    for (j = 0; j < mtab->fmode[ftype].sub; j++) {
        int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype;

        if (!j && wtype == 4)
            sub_wtype = 4;
        else if (j == mtab->fmode[ftype].sub-1 && wtype == 7)
            sub_wtype = 7;

        wsize = types_sizes[wtype_to_wsize[sub_wtype]];

        ff_imdct_half(&tctx->mdct_ctx[ftype], buf1 + bsize*j, in + bsize*j);

        tctx->dsp.vector_fmul_window(out2,
                                     prev_buf + (bsize-wsize)/2,
                                     buf1 + bsize*j,
                                     ff_sine_windows[av_log2(wsize)],
                                     0.0,
                                     wsize/2);
        out2 += wsize;

        memcpy(out2, buf1 + bsize*j + wsize/2, (bsize - wsize/2)*sizeof(float));

        out2 += ftype == FT_MEDIUM ? (bsize-wsize)/2 : bsize - wsize;

        prev_buf = buf1 + bsize*j + bsize/2;
    }

    tctx->last_block_pos[ch] = (size + first_wsize)/2;
}

static void imdct_output(TwinContext *tctx, enum FrameType ftype, int wtype,
                         float *out)
{
    const ModeTab *mtab = tctx->mtab;
    float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0];
    int i, j;

    for (i = 0; i < tctx->avctx->channels; i++) {
        imdct_and_window(tctx, ftype, wtype,
                         tctx->spectrum + i*mtab->size,
                         prev_buf + 2*i*mtab->size,
                         i);
    }

    if (tctx->avctx->channels == 2) {
        for (i = 0; i < mtab->size - tctx->last_block_pos[0]; i++) {
            float f1 = prev_buf[               i];
            float f2 = prev_buf[2*mtab->size + i];
            out[2*i    ] = f1 + f2;
            out[2*i + 1] = f1 - f2;
        }
        for (j = 0; i < mtab->size; j++,i++) {
            float f1 = tctx->curr_frame[               j];
            float f2 = tctx->curr_frame[2*mtab->size + j];
            out[2*i    ] = f1 + f2;
            out[2*i + 1] = f1 - f2;
        }
    } else {
        memcpy(out, prev_buf,
               (mtab->size - tctx->last_block_pos[0]) * sizeof(*out));

        out +=  mtab->size - tctx->last_block_pos[0];

        memcpy(out, tctx->curr_frame,
               (tctx->last_block_pos[0]) * sizeof(*out));
    }

}

static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist,
                         int ch, float *out, float gain, enum FrameType ftype)
{
    const ModeTab *mtab = tctx->mtab;
    int i,j;
    float *hist = tctx->bark_hist[ftype][ch];
    float val = ((const float []) {0.4, 0.35, 0.28})[ftype];
    int bark_n_coef  = mtab->fmode[ftype].bark_n_coef;
    int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef;
    int idx = 0;

    for (i = 0; i < fw_cb_len; i++)
        for (j = 0; j < bark_n_coef; j++, idx++) {
            float tmp2 =
                mtab->fmode[ftype].bark_cb[fw_cb_len*in[j] + i] * (1./4096);
            float st = use_hist ?
                (1. - val) * tmp2 + val*hist[idx] + 1. : tmp2 + 1.;

            hist[idx] = tmp2;
            if (st < -1.) st = 1.;

            memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]);
            out += mtab->fmode[ftype].bark_tab[idx];
        }

}

static void read_and_decode_spectrum(TwinContext *tctx, GetBitContext *gb,
                                     float *out, enum FrameType ftype)
{
    const ModeTab *mtab = tctx->mtab;
    int channels = tctx->avctx->channels;
    int sub = mtab->fmode[ftype].sub;
    int block_size = mtab->size / sub;
    float gain[channels*sub];
    float ppc_shape[mtab->ppc_shape_len * channels * 4];
    uint8_t bark1[channels][sub][mtab->fmode[ftype].bark_n_coef];
    uint8_t bark_use_hist[channels][sub];

    uint8_t lpc_idx1[channels];
    uint8_t lpc_idx2[channels][tctx->mtab->lsp_split];
    uint8_t lpc_hist_idx[channels];

    int i, j, k;

    dequant(tctx, gb, out, ftype,
            mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
            mtab->fmode[ftype].cb_len_read);

    for (i = 0; i < channels; i++)
        for (j = 0; j < sub; j++)
            for (k = 0; k < mtab->fmode[ftype].bark_n_coef; k++)
                bark1[i][j][k] =
                    get_bits(gb, mtab->fmode[ftype].bark_n_bit);

    for (i = 0; i < channels; i++)
        for (j = 0; j < sub; j++)
            bark_use_hist[i][j] = get_bits1(gb);

    dec_gain(tctx, gb, ftype, gain);

    for (i = 0; i < channels; i++) {
        lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0);
        lpc_idx1    [i] = get_bits(gb, tctx->mtab->lsp_bit1);

        for (j = 0; j < tctx->mtab->lsp_split; j++)
            lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2);
    }

    if (ftype == FT_LONG) {
        int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len*channels - 1)/
            tctx->n_div[3];
        dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb,
                mtab->ppc_shape_cb + cb_len_p*PPC_SHAPE_CB_SIZE, cb_len_p);
    }

    for (i = 0; i < channels; i++) {
        float *chunk = out + mtab->size * i;
        float lsp[tctx->mtab->n_lsp];

        for (j = 0; j < sub; j++) {
            dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i,
                         tctx->tmp_buf, gain[sub*i+j], ftype);

            tctx->dsp.vector_fmul(chunk + block_size*j, tctx->tmp_buf,
                                  block_size);

        }

        if (ftype == FT_LONG) {
            float pgain_step = 25000. / ((1 << mtab->pgain_bit) - 1);
            int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit);
            int g_coef = get_bits(gb, tctx->mtab->pgain_bit);
            float v = 1./8192*
                mulawinv(pgain_step*g_coef+ pgain_step/2, 25000., PGAIN_MU);

            decode_ppc(tctx, p_coef, ppc_shape + i*mtab->ppc_shape_len, v,
                       chunk);
        }

        decode_lsp(tctx, lpc_idx1[i], lpc_idx2[i], lpc_hist_idx[i], lsp,
                   tctx->lsp_hist[i]);

        dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);

        for (j = 0; j < mtab->fmode[ftype].sub; j++) {
            tctx->dsp.vector_fmul(chunk, tctx->tmp_buf, block_size);
            chunk += block_size;
        }
    }
}

static int twin_decode_frame(AVCodecContext * avctx, void *data,
                             int *data_size, AVPacket *avpkt)
{
    const uint8_t *buf = avpkt->data;
    int buf_size = avpkt->size;
    TwinContext *tctx = avctx->priv_data;
    GetBitContext gb;
    const ModeTab *mtab = tctx->mtab;
    float *out = data;
    enum FrameType ftype;
    int window_type;
    static const enum FrameType wtype_to_ftype_table[] = {
        FT_LONG,   FT_LONG, FT_SHORT, FT_LONG,
        FT_MEDIUM, FT_LONG, FT_LONG,  FT_MEDIUM, FT_MEDIUM
    };

    if (buf_size*8 < avctx->bit_rate*mtab->size/avctx->sample_rate + 8) {
        av_log(avctx, AV_LOG_ERROR,
               "Frame too small (%d bytes). Truncated file?\n", buf_size);
        *data_size = 0;
        return buf_size;
    }

    init_get_bits(&gb, buf, buf_size * 8);
    skip_bits(&gb, get_bits(&gb, 8));
    window_type = get_bits(&gb, WINDOW_TYPE_BITS);

    if (window_type > 8) {
        av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n");
        return -1;
    }

    ftype = wtype_to_ftype_table[window_type];

    read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype);

    imdct_output(tctx, ftype, window_type, out);

    FFSWAP(float*, tctx->curr_frame, tctx->prev_frame);

    if (tctx->avctx->frame_number < 2) {
        *data_size=0;
        return buf_size;
    }

    tctx->dsp.vector_clipf(out, out, -32700./(1<<15), 32700./(1<<15),
                           avctx->channels * mtab->size);

    *data_size = mtab->size*avctx->channels*4;

    return buf_size;
}

/**
 * Init IMDCT and windowing tables
 */
static av_cold void init_mdct_win(TwinContext *tctx)
{
    int i,j;
    const ModeTab *mtab = tctx->mtab;
    int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
    int size_m = mtab->size / mtab->fmode[FT_MEDIUM].sub;
    int channels = tctx->avctx->channels;
    float norm = channels == 1 ? 2. : 1.;

    for (i = 0; i < 3; i++) {
        int bsize = tctx->mtab->size/tctx->mtab->fmode[i].sub;
        ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
                     -sqrt(norm/bsize) / (1<<15));
    }

    tctx->tmp_buf  = av_malloc(mtab->size            * sizeof(*tctx->tmp_buf));

    tctx->spectrum  = av_malloc(2*mtab->size*channels*sizeof(float));
    tctx->curr_frame = av_malloc(2*mtab->size*channels*sizeof(float));
    tctx->prev_frame  = av_malloc(2*mtab->size*channels*sizeof(float));

    for (i = 0; i < 3; i++) {
        int m = 4*mtab->size/mtab->fmode[i].sub;
        double freq = 2*M_PI/m;
        tctx->cos_tabs[i] = av_malloc((m/4)*sizeof(*tctx->cos_tabs));

        for (j = 0; j <= m/8; j++)
            tctx->cos_tabs[i][j] = cos((2*j + 1)*freq);
        for (j = 1; j <  m/8; j++)
            tctx->cos_tabs[i][m/4-j] = tctx->cos_tabs[i][j];
    }


    ff_init_ff_sine_windows(av_log2(size_m));
    ff_init_ff_sine_windows(av_log2(size_s/2));
    ff_init_ff_sine_windows(av_log2(mtab->size));
}

/**
 * Interpret the data as if it were a num_blocks x line_len[0] matrix and for
 * each line do a cyclic permutation, i.e.
 * abcdefghijklm -> defghijklmabc
 * where the amount to be shifted is evaluated depending on the column.
 */
static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
                              int block_size,
                              const uint8_t line_len[2], int length_div,
                              enum FrameType ftype)

{
    int i,j;

    for (i = 0; i < line_len[0]; i++) {
        int shift;

        if (num_blocks == 1 ||
            (ftype == FT_LONG && num_vect % num_blocks) ||
            (ftype != FT_LONG && num_vect & 1         ) ||
            i == line_len[1]) {
            shift = 0;
        } else if (ftype == FT_LONG) {
            shift = i;
        } else
            shift = i*i;

        for (j = 0; j < num_vect && (j+num_vect*i < block_size*num_blocks); j++)
            tab[i*num_vect+j] = i*num_vect + (j + shift) % num_vect;
    }
}

/**
 * Interpret the input data as in the following table:
 *
 * \verbatim
 *
 * abcdefgh
 * ijklmnop
 * qrstuvw
 * x123456
 *
 * \endverbatim
 *
 * and transpose it, giving the output
 * aiqxbjr1cks2dlt3emu4fvn5gow6hp
 */
static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
                           const uint8_t line_len[2], int length_div)
{
    int i,j;
    int cont= 0;
    for (i = 0; i < num_vect; i++)
        for (j = 0; j < line_len[i >= length_div]; j++)
            out[cont++] = in[j*num_vect + i];
}

static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
{
    int block_size = size/n_blocks;
    int i;

    for (i = 0; i < size; i++)
        out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
}

static av_cold void construct_perm_table(TwinContext *tctx,enum FrameType ftype)
{
    int block_size;
    const ModeTab *mtab = tctx->mtab;
    int size = tctx->avctx->channels*mtab->fmode[ftype].sub;
    int16_t *tmp_perm = (int16_t *) tctx->tmp_buf;

    if (ftype == FT_PPC) {
        size  = tctx->avctx->channels;
        block_size = mtab->ppc_shape_len;
    } else
        block_size = mtab->size / mtab->fmode[ftype].sub;

    permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
                      block_size, tctx->length[ftype],
                      tctx->length_change[ftype], ftype);

    transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
                   tctx->length[ftype], tctx->length_change[ftype]);

    linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
                size*block_size);
}

static av_cold void init_bitstream_params(TwinContext *tctx)
{
    const ModeTab *mtab = tctx->mtab;
    int n_ch = tctx->avctx->channels;
    int total_fr_bits = tctx->avctx->bit_rate*mtab->size/
                             tctx->avctx->sample_rate;

    int lsp_bits_per_block = n_ch*(mtab->lsp_bit0 + mtab->lsp_bit1 +
                                   mtab->lsp_split*mtab->lsp_bit2);

    int ppc_bits = n_ch*(mtab->pgain_bit + mtab->ppc_shape_bit +
                         mtab->ppc_period_bit);

    int bsize_no_main_cb[3];
    int bse_bits[3];
    int i;
    enum FrameType frametype;

    for (i = 0; i < 3; i++)
        // +1 for history usage switch
        bse_bits[i] = n_ch *
            (mtab->fmode[i].bark_n_coef * mtab->fmode[i].bark_n_bit + 1);

    bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
                          WINDOW_TYPE_BITS + n_ch*GAIN_BITS;

    for (i = 0; i < 2; i++)
        bsize_no_main_cb[i] =
            lsp_bits_per_block + n_ch*GAIN_BITS + WINDOW_TYPE_BITS +
            mtab->fmode[i].sub*(bse_bits[i] + n_ch*SUB_GAIN_BITS);

    // The remaining bits are all used for the main spectrum coefficients
    for (i = 0; i < 4; i++) {
        int bit_size;
        int vect_size;
        int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
        if (i == 3) {
            bit_size  = n_ch * mtab->ppc_shape_bit;
            vect_size = n_ch * mtab->ppc_shape_len;
        } else {
            bit_size = total_fr_bits - bsize_no_main_cb[i];
            vect_size = n_ch * mtab->size;
        }

        tctx->n_div[i] = (bit_size + 13) / 14;

        rounded_up   = (bit_size + tctx->n_div[i] - 1)/tctx->n_div[i];
        rounded_down = (bit_size           )/tctx->n_div[i];
        num_rounded_down = rounded_up * tctx->n_div[i] - bit_size;
        num_rounded_up = tctx->n_div[i] - num_rounded_down;
        tctx->bits_main_spec[0][i][0] = (rounded_up   + 1)/2;
        tctx->bits_main_spec[1][i][0] = (rounded_up      )/2;
        tctx->bits_main_spec[0][i][1] = (rounded_down + 1)/2;
        tctx->bits_main_spec[1][i][1] = (rounded_down    )/2;
        tctx->bits_main_spec_change[i] = num_rounded_up;

        rounded_up   = (vect_size + tctx->n_div[i] - 1)/tctx->n_div[i];
        rounded_down = (vect_size                     )/tctx->n_div[i];
        num_rounded_down = rounded_up * tctx->n_div[i] - vect_size;
        num_rounded_up = tctx->n_div[i] - num_rounded_down;
        tctx->length[i][0] = rounded_up;
        tctx->length[i][1] = rounded_down;
        tctx->length_change[i] = num_rounded_up;
    }

    for (frametype = FT_SHORT; frametype <= FT_PPC; frametype++)
        construct_perm_table(tctx, frametype);
}

static av_cold int twin_decode_init(AVCodecContext *avctx)
{
    TwinContext *tctx = avctx->priv_data;
    int isampf = avctx->sample_rate/1000;
    int ibps = avctx->bit_rate/(1000 * avctx->channels);

    tctx->avctx       = avctx;
    avctx->sample_fmt = SAMPLE_FMT_FLT;

    if (avctx->channels > 2) {
        av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n",
               avctx->channels);
        return -1;
    }

    switch ((isampf << 8) +  ibps) {
    case (8 <<8) +  8: tctx->mtab = &mode_08_08; break;
    case (11<<8) +  8: tctx->mtab = &mode_11_08; break;
    case (11<<8) + 10: tctx->mtab = &mode_11_10; break;
    case (16<<8) + 16: tctx->mtab = &mode_16_16; break;
    case (22<<8) + 20: tctx->mtab = &mode_22_20; break;
    case (22<<8) + 24: tctx->mtab = &mode_22_24; break;
    case (22<<8) + 32: tctx->mtab = &mode_22_32; break;
    case (44<<8) + 40: tctx->mtab = &mode_44_40; break;
    case (44<<8) + 48: tctx->mtab = &mode_44_48; break;
    default:
        av_log(avctx, AV_LOG_ERROR, "This version does not support %d kHz - %d kbit/s/ch mode.\n", isampf, isampf);
        return -1;
    }

    dsputil_init(&tctx->dsp, avctx);
    init_mdct_win(tctx);
    init_bitstream_params(tctx);

    memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist));

    return 0;
}

static av_cold int twin_decode_close(AVCodecContext *avctx)
{
    TwinContext *tctx = avctx->priv_data;
    int i;

    for (i = 0; i < 3; i++) {
        ff_mdct_end(&tctx->mdct_ctx[i]);
        av_free(tctx->cos_tabs[i]);
    }


    av_free(tctx->curr_frame);
    av_free(tctx->spectrum);
    av_free(tctx->prev_frame);
    av_free(tctx->tmp_buf);

    return 0;
}

AVCodec twinvq_decoder =
{
    "twinvq",
    CODEC_TYPE_AUDIO,
    CODEC_ID_TWINVQ,
    sizeof(TwinContext),
    twin_decode_init,
    NULL,
    twin_decode_close,
    twin_decode_frame,
    .long_name = NULL_IF_CONFIG_SMALL("VQF TwinVQ"),
};