comparison ac3dec.c @ 6430:de7502093922 libavcodec

Use 24-bit fixed-point transform coefficients until just before MDCT. This gives 7% faster decoding on average.

6429:3c67363d6fa0

 * Table of bin locations for rematrixing bands
 * reference: Section 7.5.2 Rematrixing : Frequency Band Definitions
 */
static const uint8_t rematrix_band_tab[5] = { 13, 25, 37, 61, 253 };

/**
 * table for exponent to scale_factor mapping
 * scale_factors[i] = 2 ^ -i
 */
static float scale_factors[25];

/** table for grouping exponents */
static uint8_t exp_ungroup_tab[128][3];


/** tables for ungrouping mantissas */
static float b1_mantissas[32][3];
static float b2_mantissas[128][3];
static float b3_mantissas[8];
static float b4_mantissas[128][2];
static float b5_mantissas[16];

/**
 * Quantization table: levels for symmetric. bits for asymmetric.
 * reference: Table 7.18 Mapping of bap to Quantizer
 */

    int center_mix_level;                   ///< Center mix level index
    int surround_mix_level;                 ///< Surround mix level index
    float downmix_coeffs[AC3_MAX_CHANNELS][2];  ///< stereo downmix coefficients
    float dynamic_range[2];                 ///< dynamic range
    float cpl_coords[AC3_MAX_CHANNELS][18]; ///< coupling coordinates
    int   num_cpl_bands;                    ///< number of coupling bands
    int   num_cpl_subbands;                 ///< number of coupling sub bands
    int   start_freq[AC3_MAX_CHANNELS];     ///< start frequency bin
    int   end_freq[AC3_MAX_CHANNELS];       ///< end frequency bin
    AC3BitAllocParameters bit_alloc_params; ///< bit allocation parameters

    uint8_t bap[AC3_MAX_CHANNELS][256];     ///< bit allocation pointers
    int16_t psd[AC3_MAX_CHANNELS][256];     ///< scaled exponents
    int16_t band_psd[AC3_MAX_CHANNELS][50]; ///< interpolated exponents
    int16_t mask[AC3_MAX_CHANNELS][50];     ///< masking curve values

    DECLARE_ALIGNED_16(float, transform_coeffs[AC3_MAX_CHANNELS][256]);  ///< transform coefficients

    /* For IMDCT. */
    MDCTContext imdct_512;                  ///< for 512 sample IMDCT
    MDCTContext imdct_256;                  ///< for 256 sample IMDCT

/**
 * Symmetrical Dequantization
 * reference: Section 7.3.3 Expansion of Mantissas for Symmetrical Quantization
 *            Tables 7.19 to 7.23
 */
static inline float
symmetric_dequant(int code, int levels)
{
    return (code - (levels >> 1)) * (2.0f / levels);
}

/*
 * Initialize tables at runtime.
 */

       reference: Section 7.7.1 Dynamic Range Control */
    for(i=0; i<256; i++) {
        int v = (i >> 5) - ((i >> 7) << 3) - 5;
        dynamic_range_tab[i] = powf(2.0f, v) * ((i & 0x1F) | 0x20);
    }

    /* generate scale factors for exponents and asymmetrical dequantization
       reference: Section 7.3.2 Expansion of Mantissas for Asymmetric Quantization */
    for (i = 0; i < 25; i++)
        scale_factors[i] = pow(2.0, -i);

    /* generate exponent tables
       reference: Section 7.1.3 Exponent Decoding */
    for(i=0; i<128; i++) {
        exp_ungroup_tab[i][0] =  i / 25;

        do {
            subbnd++;
            for(j=0; j<12; j++) {
                for(ch=1; ch<=s->fbw_channels; ch++) {
                    if(s->channel_in_cpl[ch]) {
                        s->transform_coeffs[ch][i] = s->transform_coeffs[CPL_CH][i] * s->cpl_coords[ch][bnd] * 8.0f;
                        if (ch == 2 && s->phase_flags[bnd])
                            s->transform_coeffs[ch][i] = -s->transform_coeffs[ch][i];
                    }
                }
                i++;
            }
        } while(s->cpl_band_struct[subbnd]);

/**
 * Grouped mantissas for 3-level 5-level and 11-level quantization
 */
typedef struct {
    float b1_mant[3];
    float b2_mant[3];
    float b4_mant[2];
    int b1ptr;
    int b2ptr;
    int b4ptr;
} mant_groups;

{
    GetBitContext *gbc = &s->gbc;
    int i, gcode, tbap, start, end;
    uint8_t *exps;
    uint8_t *bap;
    float *coeffs;

    exps = s->dexps[ch_index];
    bap = s->bap[ch_index];
    coeffs = s->transform_coeffs[ch_index];
    start = s->start_freq[ch_index];
    end = s->end_freq[ch_index];

    for (i = start; i < end; i++) {
        tbap = bap[i];
        switch (tbap) {
            case 0:
                coeffs[i] = ((av_random(&s->dith_state) & 0xFFFF) / 65535.0f) - 0.5f;
                break;

            case 1:
                if(m->b1ptr > 2) {
                    gcode = get_bits(gbc, 5);

            case 5:
                coeffs[i] = b5_mantissas[get_bits(gbc, 4)];
                break;

            default:
                /* asymmetric dequantization */
                coeffs[i] = get_sbits(gbc, quantization_tab[tbap]) * scale_factors[quantization_tab[tbap]-1];
                break;
        }
        coeffs[i] *= scale_factors[exps[i]];
    }

    return 0;
}

 * reference: Section 7.3.4 Dither for Zero Bit Mantissas (bap=0)
 */
static void remove_dithering(AC3DecodeContext *s) {
    int ch, i;
    int end=0;
    float *coeffs;
    uint8_t *bap;

    for(ch=1; ch<=s->fbw_channels; ch++) {
        if(!s->dither_flag[ch]) {
            coeffs = s->transform_coeffs[ch];
            bap = s->bap[ch];
            if(s->channel_in_cpl[ch])
                end = s->start_freq[CPL_CH];
            else
                end = s->end_freq[ch];
            for(i=0; i<end; i++) {
                if(!bap[i])
                    coeffs[i] = 0.0f;
            }
            if(s->channel_in_cpl[ch]) {
                bap = s->bap[CPL_CH];
                for(; i<s->end_freq[CPL_CH]; i++) {
                    if(!bap[i])
                        coeffs[i] = 0.0f;
                }
            }
        }
    }
}

 */
static void do_rematrixing(AC3DecodeContext *s)
{
    int bnd, i;
    int end, bndend;
    float tmp0, tmp1;

    end = FFMIN(s->end_freq[1], s->end_freq[2]);

    for(bnd=0; bnd<s->num_rematrixing_bands; bnd++) {
        if(s->rematrixing_flags[bnd]) {
            bndend = FFMIN(end, rematrix_band_tab[bnd+1]);
            for(i=rematrix_band_tab[bnd]; i<bndend; i++) {
                tmp0 = s->transform_coeffs[1][i];
                tmp1 = s->transform_coeffs[2][i];
                s->transform_coeffs[1][i] = tmp0 + tmp1;
                s->transform_coeffs[2][i] = tmp0 - tmp1;
            }
        }
    }
}

                    master_cpl_coord = 3 * get_bits(gbc, 2);
                    for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
                        cpl_coord_exp = get_bits(gbc, 4);
                        cpl_coord_mant = get_bits(gbc, 4);
                        if (cpl_coord_exp == 15)
                            s->cpl_coords[ch][bnd] = cpl_coord_mant / 16.0f;
                        else
                            s->cpl_coords[ch][bnd] = (cpl_coord_mant + 16.0f) / 32.0f;
                        s->cpl_coords[ch][bnd] *= scale_factors[cpl_coord_exp + master_cpl_coord];
                    }
                }
            }
        }
        /* phase flags */

    if(s->channel_mode == AC3_CHMODE_STEREO)
        do_rematrixing(s);

    /* apply scaling to coefficients (headroom, dynrng) */
    for(ch=1; ch<=s->channels; ch++) {
        float gain = 2.0f * s->mul_bias;
        if(s->channel_mode == AC3_CHMODE_DUALMONO) {
            gain *= s->dynamic_range[ch-1];
        } else {
            gain *= s->dynamic_range[0];
        }
        for(i=0; i<s->end_freq[ch]; i++) {
            s->transform_coeffs[ch][i] *= gain;
        }
    }

    do_imdct(s);

6430:de7502093922

 * Table of bin locations for rematrixing bands
 * reference: Section 7.5.2 Rematrixing : Frequency Band Definitions
 */
static const uint8_t rematrix_band_tab[5] = { 13, 25, 37, 61, 253 };

/** table for grouping exponents */
static uint8_t exp_ungroup_tab[128][3];


/** tables for ungrouping mantissas */
static int b1_mantissas[32][3];
static int b2_mantissas[128][3];
static int b3_mantissas[8];
static int b4_mantissas[128][2];
static int b5_mantissas[16];

/**
 * Quantization table: levels for symmetric. bits for asymmetric.
 * reference: Table 7.18 Mapping of bap to Quantizer
 */

    int center_mix_level;                   ///< Center mix level index
    int surround_mix_level;                 ///< Surround mix level index
    float downmix_coeffs[AC3_MAX_CHANNELS][2];  ///< stereo downmix coefficients
    float dynamic_range[2];                 ///< dynamic range
    int   cpl_coords[AC3_MAX_CHANNELS][18]; ///< coupling coordinates
    int   num_cpl_bands;                    ///< number of coupling bands
    int   num_cpl_subbands;                 ///< number of coupling sub bands
    int   start_freq[AC3_MAX_CHANNELS];     ///< start frequency bin
    int   end_freq[AC3_MAX_CHANNELS];       ///< end frequency bin
    AC3BitAllocParameters bit_alloc_params; ///< bit allocation parameters

    uint8_t bap[AC3_MAX_CHANNELS][256];     ///< bit allocation pointers
    int16_t psd[AC3_MAX_CHANNELS][256];     ///< scaled exponents
    int16_t band_psd[AC3_MAX_CHANNELS][50]; ///< interpolated exponents
    int16_t mask[AC3_MAX_CHANNELS][50];     ///< masking curve values

    int fixed_coeffs[AC3_MAX_CHANNELS][256];    ///> fixed-point transform coefficients
    DECLARE_ALIGNED_16(float, transform_coeffs[AC3_MAX_CHANNELS][256]);  ///< transform coefficients

    /* For IMDCT. */
    MDCTContext imdct_512;                  ///< for 512 sample IMDCT
    MDCTContext imdct_256;                  ///< for 256 sample IMDCT

/**
 * Symmetrical Dequantization
 * reference: Section 7.3.3 Expansion of Mantissas for Symmetrical Quantization
 *            Tables 7.19 to 7.23
 */
static inline int
symmetric_dequant(int code, int levels)
{
    return ((code - (levels >> 1)) << 24) / levels;
}

/*
 * Initialize tables at runtime.
 */

       reference: Section 7.7.1 Dynamic Range Control */
    for(i=0; i<256; i++) {
        int v = (i >> 5) - ((i >> 7) << 3) - 5;
        dynamic_range_tab[i] = powf(2.0f, v) * ((i & 0x1F) | 0x20);
    }

    /* generate exponent tables
       reference: Section 7.1.3 Exponent Decoding */
    for(i=0; i<128; i++) {
        exp_ungroup_tab[i][0] =  i / 25;

        do {
            subbnd++;
            for(j=0; j<12; j++) {
                for(ch=1; ch<=s->fbw_channels; ch++) {
                    if(s->channel_in_cpl[ch]) {
                        s->fixed_coeffs[ch][i] = ((int64_t)s->fixed_coeffs[CPL_CH][i] * (int64_t)s->cpl_coords[ch][bnd]) >> 23;
                        if (ch == 2 && s->phase_flags[bnd])
                            s->fixed_coeffs[ch][i] = -s->fixed_coeffs[ch][i];
                    }
                }
                i++;
            }
        } while(s->cpl_band_struct[subbnd]);

/**
 * Grouped mantissas for 3-level 5-level and 11-level quantization
 */
typedef struct {
    int b1_mant[3];
    int b2_mant[3];
    int b4_mant[2];
    int b1ptr;
    int b2ptr;
    int b4ptr;
} mant_groups;

{
    GetBitContext *gbc = &s->gbc;
    int i, gcode, tbap, start, end;
    uint8_t *exps;
    uint8_t *bap;
    int *coeffs;

    exps = s->dexps[ch_index];
    bap = s->bap[ch_index];
    coeffs = s->fixed_coeffs[ch_index];
    start = s->start_freq[ch_index];
    end = s->end_freq[ch_index];

    for (i = start; i < end; i++) {
        tbap = bap[i];
        switch (tbap) {
            case 0:
                coeffs[i] = (av_random(&s->dith_state) & 0x7FFFFF) - 4194304;
                break;

            case 1:
                if(m->b1ptr > 2) {
                    gcode = get_bits(gbc, 5);

            case 5:
                coeffs[i] = b5_mantissas[get_bits(gbc, 4)];
                break;

            default: {
                /* asymmetric dequantization */
                int qlevel = quantization_tab[tbap];
                coeffs[i] = get_sbits(gbc, qlevel) << (24 - qlevel);
                break;
            }
        }
        coeffs[i] >>= exps[i];
    }

    return 0;
}

 * reference: Section 7.3.4 Dither for Zero Bit Mantissas (bap=0)
 */
static void remove_dithering(AC3DecodeContext *s) {
    int ch, i;
    int end=0;
    int *coeffs;
    uint8_t *bap;

    for(ch=1; ch<=s->fbw_channels; ch++) {
        if(!s->dither_flag[ch]) {
            coeffs = s->fixed_coeffs[ch];
            bap = s->bap[ch];
            if(s->channel_in_cpl[ch])
                end = s->start_freq[CPL_CH];
            else
                end = s->end_freq[ch];
            for(i=0; i<end; i++) {
                if(!bap[i])
                    coeffs[i] = 0;
            }
            if(s->channel_in_cpl[ch]) {
                bap = s->bap[CPL_CH];
                for(; i<s->end_freq[CPL_CH]; i++) {
                    if(!bap[i])
                        coeffs[i] = 0;
                }
            }
        }
    }
}

 */
static void do_rematrixing(AC3DecodeContext *s)
{
    int bnd, i;
    int end, bndend;
    int tmp0, tmp1;

    end = FFMIN(s->end_freq[1], s->end_freq[2]);

    for(bnd=0; bnd<s->num_rematrixing_bands; bnd++) {
        if(s->rematrixing_flags[bnd]) {
            bndend = FFMIN(end, rematrix_band_tab[bnd+1]);
            for(i=rematrix_band_tab[bnd]; i<bndend; i++) {
                tmp0 = s->fixed_coeffs[1][i];
                tmp1 = s->fixed_coeffs[2][i];
                s->fixed_coeffs[1][i] = tmp0 + tmp1;
                s->fixed_coeffs[2][i] = tmp0 - tmp1;
            }
        }
    }
}

                    master_cpl_coord = 3 * get_bits(gbc, 2);
                    for (bnd = 0; bnd < s->num_cpl_bands; bnd++) {
                        cpl_coord_exp = get_bits(gbc, 4);
                        cpl_coord_mant = get_bits(gbc, 4);
                        if (cpl_coord_exp == 15)
                            s->cpl_coords[ch][bnd] = cpl_coord_mant << 22;
                        else
                            s->cpl_coords[ch][bnd] = (cpl_coord_mant + 16) << 21;
                        s->cpl_coords[ch][bnd] >>= (cpl_coord_exp + master_cpl_coord);
                    }
                }
            }
        }
        /* phase flags */

    if(s->channel_mode == AC3_CHMODE_STEREO)
        do_rematrixing(s);

    /* apply scaling to coefficients (headroom, dynrng) */
    for(ch=1; ch<=s->channels; ch++) {
        float gain = s->mul_bias / 4194304.0f;
        if(s->channel_mode == AC3_CHMODE_DUALMONO) {
            gain *= s->dynamic_range[ch-1];
        } else {
            gain *= s->dynamic_range[0];
        }
        for(i=0; i<256; i++) {
            s->transform_coeffs[ch][i] = s->fixed_coeffs[ch][i] * gain;
        }
    }

    do_imdct(s);