Head

GCC Code Coverage Report

Directory:	../../../ffmpeg/		Exec	Total	Coverage
File:	src/libavcodec/opus_celt.c	Lines:	293	315	93.0 %
Date:	2020-01-27 20:17:04	Branches:	190	208	91.3 %


/*
 * Copyright (c) 2012 Andrew D'Addesio
 * Copyright (c) 2013-2014 Mozilla Corporation
 * Copyright (c) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
 *
 * 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
 */

/**
 * @file
 * Opus CELT decoder
 */

#include "opus_celt.h"
#include "opustab.h"
#include "opus_pvq.h"

/* Use the 2D z-transform to apply prediction in both the time domain (alpha)
 * and the frequency domain (beta) */
static void celt_decode_coarse_energy(CeltFrame *f, OpusRangeCoder *rc)
{
    int i, j;
    float prev[2] = { 0 };
    float alpha = ff_celt_alpha_coef[f->size];
    float beta  = ff_celt_beta_coef[f->size];
    const uint8_t *model = ff_celt_coarse_energy_dist[f->size][0];

    /* intra frame */
    if (opus_rc_tell(rc) + 3 <= f->framebits && ff_opus_rc_dec_log(rc, 3)) {
        alpha = 0.0f;
        beta  = 1.0f - (4915.0f/32768.0f);
        model = ff_celt_coarse_energy_dist[f->size][1];
    }

    for (i = 0; i < CELT_MAX_BANDS; i++) {
        for (j = 0; j < f->channels; j++) {
            CeltBlock *block = &f->block[j];
            float value;
            int available;

            if (i < f->start_band || i >= f->end_band) {
                block->energy[i] = 0.0;
                continue;
            }

            available = f->framebits - opus_rc_tell(rc);
            if (available >= 15) {
                /* decode using a Laplace distribution */
                int k = FFMIN(i, 20) << 1;
                value = ff_opus_rc_dec_laplace(rc, model[k] << 7, model[k+1] << 6);
            } else if (available >= 2) {
                int x = ff_opus_rc_dec_cdf(rc, ff_celt_model_energy_small);
                value = (x>>1) ^ -(x&1);
            } else if (available >= 1) {
                value = -(float)ff_opus_rc_dec_log(rc, 1);
            } else value = -1;

            block->energy[i] = FFMAX(-9.0f, block->energy[i]) * alpha + prev[j] + value;
            prev[j] += beta * value;
        }
    }
}

static void celt_decode_fine_energy(CeltFrame *f, OpusRangeCoder *rc)
{
    int i;
    for (i = f->start_band; i < f->end_band; i++) {
        int j;
        if (!f->fine_bits[i])
            continue;

        for (j = 0; j < f->channels; j++) {
            CeltBlock *block = &f->block[j];
            int q2;
            float offset;
            q2 = ff_opus_rc_get_raw(rc, f->fine_bits[i]);
            offset = (q2 + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f - 0.5f;
            block->energy[i] += offset;
        }
    }
}

static void celt_decode_final_energy(CeltFrame *f, OpusRangeCoder *rc)
{
    int priority, i, j;
    int bits_left = f->framebits - opus_rc_tell(rc);

    for (priority = 0; priority < 2; priority++) {
        for (i = f->start_band; i < f->end_band && bits_left >= f->channels; i++) {
            if (f->fine_priority[i] != priority || f->fine_bits[i] >= CELT_MAX_FINE_BITS)
                continue;

            for (j = 0; j < f->channels; j++) {
                int q2;
                float offset;
                q2 = ff_opus_rc_get_raw(rc, 1);
                offset = (q2 - 0.5f) * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
                f->block[j].energy[i] += offset;
                bits_left--;
            }
        }
    }
}

static void celt_decode_tf_changes(CeltFrame *f, OpusRangeCoder *rc)
{
    int i, diff = 0, tf_select = 0, tf_changed = 0, tf_select_bit;
    int consumed, bits = f->transient ? 2 : 4;

    consumed = opus_rc_tell(rc);
    tf_select_bit = (f->size != 0 && consumed+bits+1 <= f->framebits);

    for (i = f->start_band; i < f->end_band; i++) {
        if (consumed+bits+tf_select_bit <= f->framebits) {
            diff ^= ff_opus_rc_dec_log(rc, bits);
            consumed = opus_rc_tell(rc);
            tf_changed |= diff;
        }
        f->tf_change[i] = diff;
        bits = f->transient ? 4 : 5;
    }

    if (tf_select_bit && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
                         ff_celt_tf_select[f->size][f->transient][1][tf_changed])
        tf_select = ff_opus_rc_dec_log(rc, 1);

    for (i = f->start_band; i < f->end_band; i++) {
        f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
    }
}

static void celt_denormalize(CeltFrame *f, CeltBlock *block, float *data)
{
    int i, j;

    for (i = f->start_band; i < f->end_band; i++) {
        float *dst = data + (ff_celt_freq_bands[i] << f->size);
        float log_norm = block->energy[i] + ff_celt_mean_energy[i];
        float norm = exp2f(FFMIN(log_norm, 32.0f));

        for (j = 0; j < ff_celt_freq_range[i] << f->size; j++)
            dst[j] *= norm;
    }
}

static void celt_postfilter_apply_transition(CeltBlock *block, float *data)
{
    const int T0 = block->pf_period_old;
    const int T1 = block->pf_period;

    float g00, g01, g02;
    float g10, g11, g12;

    float x0, x1, x2, x3, x4;

    int i;

    if (block->pf_gains[0]     == 0.0 &&
        block->pf_gains_old[0] == 0.0)
        return;

    g00 = block->pf_gains_old[0];
    g01 = block->pf_gains_old[1];
    g02 = block->pf_gains_old[2];
    g10 = block->pf_gains[0];
    g11 = block->pf_gains[1];
    g12 = block->pf_gains[2];

    x1 = data[-T1 + 1];
    x2 = data[-T1];
    x3 = data[-T1 - 1];
    x4 = data[-T1 - 2];

    for (i = 0; i < CELT_OVERLAP; i++) {
        float w = ff_celt_window2[i];
        x0 = data[i - T1 + 2];

        data[i] +=  (1.0 - w) * g00 * data[i - T0]                          +
                    (1.0 - w) * g01 * (data[i - T0 - 1] + data[i - T0 + 1]) +
                    (1.0 - w) * g02 * (data[i - T0 - 2] + data[i - T0 + 2]) +
                    w         * g10 * x2                                    +
                    w         * g11 * (x1 + x3)                             +
                    w         * g12 * (x0 + x4);
        x4 = x3;
        x3 = x2;
        x2 = x1;
        x1 = x0;
    }
}

static void celt_postfilter(CeltFrame *f, CeltBlock *block)
{
    int len = f->blocksize * f->blocks;
    const int filter_len = len - 2 * CELT_OVERLAP;

    celt_postfilter_apply_transition(block, block->buf + 1024);

    block->pf_period_old = block->pf_period;
    memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));

    block->pf_period = block->pf_period_new;
    memcpy(block->pf_gains, block->pf_gains_new, sizeof(block->pf_gains));

    if (len > CELT_OVERLAP) {
        celt_postfilter_apply_transition(block, block->buf + 1024 + CELT_OVERLAP);

        if (block->pf_gains[0] > FLT_EPSILON && filter_len > 0)
            f->opusdsp.postfilter(block->buf + 1024 + 2 * CELT_OVERLAP,
                                  block->pf_period, block->pf_gains,
                                  filter_len);

        block->pf_period_old = block->pf_period;
        memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
    }

    memmove(block->buf, block->buf + len, (1024 + CELT_OVERLAP / 2) * sizeof(float));
}

static int parse_postfilter(CeltFrame *f, OpusRangeCoder *rc, int consumed)
{
    int i;

    memset(f->block[0].pf_gains_new, 0, sizeof(f->block[0].pf_gains_new));
    memset(f->block[1].pf_gains_new, 0, sizeof(f->block[1].pf_gains_new));

    if (f->start_band == 0 && consumed + 16 <= f->framebits) {
        int has_postfilter = ff_opus_rc_dec_log(rc, 1);
        if (has_postfilter) {
            float gain;
            int tapset, octave, period;

            octave = ff_opus_rc_dec_uint(rc, 6);
            period = (16 << octave) + ff_opus_rc_get_raw(rc, 4 + octave) - 1;
            gain   = 0.09375f * (ff_opus_rc_get_raw(rc, 3) + 1);
            tapset = (opus_rc_tell(rc) + 2 <= f->framebits) ?
                     ff_opus_rc_dec_cdf(rc, ff_celt_model_tapset) : 0;

            for (i = 0; i < 2; i++) {
                CeltBlock *block = &f->block[i];

                block->pf_period_new = FFMAX(period, CELT_POSTFILTER_MINPERIOD);
                block->pf_gains_new[0] = gain * ff_celt_postfilter_taps[tapset][0];
                block->pf_gains_new[1] = gain * ff_celt_postfilter_taps[tapset][1];
                block->pf_gains_new[2] = gain * ff_celt_postfilter_taps[tapset][2];
            }
        }

        consumed = opus_rc_tell(rc);
    }

    return consumed;
}

static void process_anticollapse(CeltFrame *f, CeltBlock *block, float *X)
{
    int i, j, k;

    for (i = f->start_band; i < f->end_band; i++) {
        int renormalize = 0;
        float *xptr;
        float prev[2];
        float Ediff, r;
        float thresh, sqrt_1;
        int depth;

        /* depth in 1/8 bits */
        depth = (1 + f->pulses[i]) / (ff_celt_freq_range[i] << f->size);
        thresh = exp2f(-1.0 - 0.125f * depth);
        sqrt_1 = 1.0f / sqrtf(ff_celt_freq_range[i] << f->size);

        xptr = X + (ff_celt_freq_bands[i] << f->size);

        prev[0] = block->prev_energy[0][i];
        prev[1] = block->prev_energy[1][i];
        if (f->channels == 1) {
            CeltBlock *block1 = &f->block[1];

            prev[0] = FFMAX(prev[0], block1->prev_energy[0][i]);
            prev[1] = FFMAX(prev[1], block1->prev_energy[1][i]);
        }
        Ediff = block->energy[i] - FFMIN(prev[0], prev[1]);
        Ediff = FFMAX(0, Ediff);

        /* r needs to be multiplied by 2 or 2*sqrt(2) depending on LM because
        short blocks don't have the same energy as long */
        r = exp2f(1 - Ediff);
        if (f->size == 3)
            r *= M_SQRT2;
        r = FFMIN(thresh, r) * sqrt_1;
        for (k = 0; k < 1 << f->size; k++) {
            /* Detect collapse */
            if (!(block->collapse_masks[i] & 1 << k)) {
                /* Fill with noise */
                for (j = 0; j < ff_celt_freq_range[i]; j++)
                    xptr[(j << f->size) + k] = (celt_rng(f) & 0x8000) ? r : -r;
                renormalize = 1;
            }
        }

        /* We just added some energy, so we need to renormalize */
        if (renormalize)
            celt_renormalize_vector(xptr, ff_celt_freq_range[i] << f->size, 1.0f);
    }
}

int ff_celt_decode_frame(CeltFrame *f, OpusRangeCoder *rc,
                         float **output, int channels, int frame_size,
                         int start_band,  int end_band)
{
    int i, j, downmix = 0;
    int consumed;           // bits of entropy consumed thus far for this frame
    MDCT15Context *imdct;

    if (channels != 1 && channels != 2) {
        av_log(f->avctx, AV_LOG_ERROR, "Invalid number of coded channels: %d\n",
               channels);
        return AVERROR_INVALIDDATA;
    }
    if (start_band < 0 || start_band > end_band || end_band > CELT_MAX_BANDS) {
        av_log(f->avctx, AV_LOG_ERROR, "Invalid start/end band: %d %d\n",
               start_band, end_band);
        return AVERROR_INVALIDDATA;
    }

    f->silence        = 0;
    f->transient      = 0;
    f->anticollapse   = 0;
    f->flushed        = 0;
    f->channels       = channels;
    f->start_band     = start_band;
    f->end_band       = end_band;
    f->framebits      = rc->rb.bytes * 8;

    f->size = av_log2(frame_size / CELT_SHORT_BLOCKSIZE);
    if (f->size > CELT_MAX_LOG_BLOCKS ||
        frame_size != CELT_SHORT_BLOCKSIZE * (1 << f->size)) {
        av_log(f->avctx, AV_LOG_ERROR, "Invalid CELT frame size: %d\n",
               frame_size);
        return AVERROR_INVALIDDATA;
    }

    if (!f->output_channels)
        f->output_channels = channels;

    for (i = 0; i < f->channels; i++) {
        memset(f->block[i].coeffs,         0, sizeof(f->block[i].coeffs));
        memset(f->block[i].collapse_masks, 0, sizeof(f->block[i].collapse_masks));
    }

    consumed = opus_rc_tell(rc);

    /* obtain silence flag */
    if (consumed >= f->framebits)
        f->silence = 1;
    else if (consumed == 1)
        f->silence = ff_opus_rc_dec_log(rc, 15);


    if (f->silence) {
        consumed = f->framebits;
        rc->total_bits += f->framebits - opus_rc_tell(rc);
    }

    /* obtain post-filter options */
    consumed = parse_postfilter(f, rc, consumed);

    /* obtain transient flag */
    if (f->size != 0 && consumed+3 <= f->framebits)
        f->transient = ff_opus_rc_dec_log(rc, 3);

    f->blocks    = f->transient ? 1 << f->size : 1;
    f->blocksize = frame_size / f->blocks;

    imdct = f->imdct[f->transient ? 0 : f->size];

    if (channels == 1) {
        for (i = 0; i < CELT_MAX_BANDS; i++)
            f->block[0].energy[i] = FFMAX(f->block[0].energy[i], f->block[1].energy[i]);
    }

    celt_decode_coarse_energy(f, rc);
    celt_decode_tf_changes   (f, rc);
    ff_celt_bitalloc         (f, rc, 0);
    celt_decode_fine_energy  (f, rc);
    ff_celt_quant_bands      (f, rc);

    if (f->anticollapse_needed)
        f->anticollapse = ff_opus_rc_get_raw(rc, 1);

    celt_decode_final_energy(f, rc);

    /* apply anti-collapse processing and denormalization to
     * each coded channel */
    for (i = 0; i < f->channels; i++) {
        CeltBlock *block = &f->block[i];

        if (f->anticollapse)
            process_anticollapse(f, block, f->block[i].coeffs);

        celt_denormalize(f, block, f->block[i].coeffs);
    }

    /* stereo -> mono downmix */
    if (f->output_channels < f->channels) {
        f->dsp->vector_fmac_scalar(f->block[0].coeffs, f->block[1].coeffs, 1.0, FFALIGN(frame_size, 16));
        downmix = 1;
    } else if (f->output_channels > f->channels)
        memcpy(f->block[1].coeffs, f->block[0].coeffs, frame_size * sizeof(float));

    if (f->silence) {
        for (i = 0; i < 2; i++) {
            CeltBlock *block = &f->block[i];

            for (j = 0; j < FF_ARRAY_ELEMS(block->energy); j++)
                block->energy[j] = CELT_ENERGY_SILENCE;
        }
        memset(f->block[0].coeffs, 0, sizeof(f->block[0].coeffs));
        memset(f->block[1].coeffs, 0, sizeof(f->block[1].coeffs));
    }

    /* transform and output for each output channel */
    for (i = 0; i < f->output_channels; i++) {
        CeltBlock *block = &f->block[i];

        /* iMDCT and overlap-add */
        for (j = 0; j < f->blocks; j++) {
            float *dst  = block->buf + 1024 + j * f->blocksize;

            imdct->imdct_half(imdct, dst + CELT_OVERLAP / 2, f->block[i].coeffs + j,
                              f->blocks);
            f->dsp->vector_fmul_window(dst, dst, dst + CELT_OVERLAP / 2,
                                       ff_celt_window, CELT_OVERLAP / 2);
        }

        if (downmix)
            f->dsp->vector_fmul_scalar(&block->buf[1024], &block->buf[1024], 0.5f, frame_size);

        /* postfilter */
        celt_postfilter(f, block);

        /* deemphasis */
        block->emph_coeff = f->opusdsp.deemphasis(output[i],
                                                  &block->buf[1024 - frame_size],
                                                  block->emph_coeff, frame_size);
    }

    if (channels == 1)
        memcpy(f->block[1].energy, f->block[0].energy, sizeof(f->block[0].energy));

    for (i = 0; i < 2; i++ ) {
        CeltBlock *block = &f->block[i];

        if (!f->transient) {
            memcpy(block->prev_energy[1], block->prev_energy[0], sizeof(block->prev_energy[0]));
            memcpy(block->prev_energy[0], block->energy,         sizeof(block->prev_energy[0]));
        } else {
            for (j = 0; j < CELT_MAX_BANDS; j++)
                block->prev_energy[0][j] = FFMIN(block->prev_energy[0][j], block->energy[j]);
        }

        for (j = 0; j < f->start_band; j++) {
            block->prev_energy[0][j] = CELT_ENERGY_SILENCE;
            block->energy[j]         = 0.0;
        }
        for (j = f->end_band; j < CELT_MAX_BANDS; j++) {
            block->prev_energy[0][j] = CELT_ENERGY_SILENCE;
            block->energy[j]         = 0.0;
        }
    }

    f->seed = rc->range;

    return 0;
}

void ff_celt_flush(CeltFrame *f)
{
    int i, j;

    if (f->flushed)
        return;

    for (i = 0; i < 2; i++) {
        CeltBlock *block = &f->block[i];

        for (j = 0; j < CELT_MAX_BANDS; j++)
            block->prev_energy[0][j] = block->prev_energy[1][j] = CELT_ENERGY_SILENCE;

        memset(block->energy, 0, sizeof(block->energy));
        memset(block->buf,    0, sizeof(block->buf));

        memset(block->pf_gains,     0, sizeof(block->pf_gains));
        memset(block->pf_gains_old, 0, sizeof(block->pf_gains_old));
        memset(block->pf_gains_new, 0, sizeof(block->pf_gains_new));

        /* libopus uses CELT_EMPH_COEFF on init, but 0 is better since there's
         * a lesser discontinuity when seeking.
         * The deemphasis functions differ from libopus in that they require
         * an initial state divided by the coefficient. */
        block->emph_coeff = 0.0f / CELT_EMPH_COEFF;
    }
    f->seed = 0;

    f->flushed = 1;
}

void ff_celt_free(CeltFrame **f)
{
    CeltFrame *frm = *f;
    int i;

    if (!frm)
        return;

    for (i = 0; i < FF_ARRAY_ELEMS(frm->imdct); i++)
        ff_mdct15_uninit(&frm->imdct[i]);

    ff_celt_pvq_uninit(&frm->pvq);

    av_freep(&frm->dsp);
    av_freep(f);
}

int ff_celt_init(AVCodecContext *avctx, CeltFrame **f, int output_channels,
                 int apply_phase_inv)
{
    CeltFrame *frm;
    int i, ret;

    if (output_channels != 1 && output_channels != 2) {
        av_log(avctx, AV_LOG_ERROR, "Invalid number of output channels: %d\n",
               output_channels);
        return AVERROR(EINVAL);
    }

    frm = av_mallocz(sizeof(*frm));
    if (!frm)
        return AVERROR(ENOMEM);

    frm->avctx           = avctx;
    frm->output_channels = output_channels;
    frm->apply_phase_inv = apply_phase_inv;

    for (i = 0; i < FF_ARRAY_ELEMS(frm->imdct); i++)
        if ((ret = ff_mdct15_init(&frm->imdct[i], 1, i + 3, -1.0f/32768)) < 0)
            goto fail;

    if ((ret = ff_celt_pvq_init(&frm->pvq, 0)) < 0)
        goto fail;

    frm->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
    if (!frm->dsp) {
        ret = AVERROR(ENOMEM);
        goto fail;
    }

    ff_opus_dsp_init(&frm->opusdsp);
    ff_celt_flush(frm);

    *f = frm;

    return 0;
fail:
    ff_celt_free(&frm);
    return ret;
}


Generated by: GCOVR (Version 4.1)

Line	Branch	Exec	Source
1			/*
2			* Copyright (c) 2012 Andrew D'Addesio
3			* Copyright (c) 2013-2014 Mozilla Corporation
4			* Copyright (c) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
5			*
6			* This file is part of FFmpeg.
7			*
8			* FFmpeg is free software; you can redistribute it and/or
9			* modify it under the terms of the GNU Lesser General Public
10			* License as published by the Free Software Foundation; either
11			* version 2.1 of the License, or (at your option) any later version.
12			*
13			* FFmpeg is distributed in the hope that it will be useful,
14			* but WITHOUT ANY WARRANTY; without even the implied warranty of
15			* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16			* Lesser General Public License for more details.
17			*
18			* You should have received a copy of the GNU Lesser General Public
19			* License along with FFmpeg; if not, write to the Free Software
20			* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21			*/
22
23			/**
24			* @file
25			* Opus CELT decoder
26			*/
27
28			#include "opus_celt.h"
29			#include "opustab.h"
30			#include "opus_pvq.h"
31
32			/* Use the 2D z-transform to apply prediction in both the time domain (alpha)
33			* and the frequency domain (beta) */
34		29869	static void celt_decode_coarse_energy(CeltFrame f, OpusRangeCoder rc)
35			{
36			int i, j;
37		29869	float prev[2] = { 0 };
38		29869	float alpha = ff_celt_alpha_coef[f->size];
39		29869	float beta = ff_celt_beta_coef[f->size];
40		29869	const uint8_t *model = ff_celt_coarse_energy_dist[f->size][0];
41
42			/* intra frame */
43	✓✓✓✓	29869	if (opus_rc_tell(rc) + 3 <= f->framebits && ff_opus_rc_dec_log(rc, 3)) {
44		5873	alpha = 0.0f;
45		5873	beta = 1.0f - (4915.0f/32768.0f);
46		5873	model = ff_celt_coarse_energy_dist[f->size][1];
47			}
48
49	✓✓	657118	for (i = 0; i < CELT_MAX_BANDS; i++) {
50	✓✓	1647849	for (j = 0; j < f->channels; j++) {
51		1020600	CeltBlock *block = &f->block[j];
52			float value;
53			int available;
54
55	✓✓✓✓	1020600	if (i < f->start_band \|\| i >= f->end_band) {
56		170908	block->energy[i] = 0.0;
57		170908	continue;
58			}
59
60		849692	available = f->framebits - opus_rc_tell(rc);
61	✓✓	849692	if (available >= 15) {
62			/* decode using a Laplace distribution */
63		845898	int k = FFMIN(i, 20) << 1;
64		845898	value = ff_opus_rc_dec_laplace(rc, model[k] << 7, model[k+1] << 6);
65	✓✓	3794	} else if (available >= 2) {
66		113	int x = ff_opus_rc_dec_cdf(rc, ff_celt_model_energy_small);
67		113	value = (x>>1) ^ -(x&1);
68	✓✓	3681	} else if (available >= 1) {
69		3	value = -(float)ff_opus_rc_dec_log(rc, 1);
70		3678	} else value = -1;
71
72	✓✓	849692	block->energy[i] = FFMAX(-9.0f, block->energy[i]) * alpha + prev[j] + value;
73		849692	prev[j] += beta * value;
74			}
75			}
76		29869	}
77
78		29869	static void celt_decode_fine_energy(CeltFrame f, OpusRangeCoder rc)
79			{
80			int i;
81	✓✓	537903	for (i = f->start_band; i < f->end_band; i++) {
82			int j;
83	✓✓	508034	if (!f->fine_bits[i])
84		95330	continue;
85
86	✓✓	1094007	for (j = 0; j < f->channels; j++) {
87		681303	CeltBlock *block = &f->block[j];
88			int q2;
89			float offset;
90		681303	q2 = ff_opus_rc_get_raw(rc, f->fine_bits[i]);
91		681303	offset = (q2 + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f - 0.5f;
92		681303	block->energy[i] += offset;
93			}
94			}
95		29869	}
96
97		29869	static void celt_decode_final_energy(CeltFrame f, OpusRangeCoder rc)
98			{
99			int priority, i, j;
100		29869	int bits_left = f->framebits - opus_rc_tell(rc);
101
102	✓✓	89607	for (priority = 0; priority < 2; priority++) {
103	✓✓✓✓	162460	for (i = f->start_band; i < f->end_band && bits_left >= f->channels; i++) {
104	✓✓✓✓	102722	if (f->fine_priority[i] != priority \|\| f->fine_bits[i] >= CELT_MAX_FINE_BITS)
105		65402	continue;
106
107	✓✓	96124	for (j = 0; j < f->channels; j++) {
108			int q2;
109			float offset;
110		58804	q2 = ff_opus_rc_get_raw(rc, 1);
111		58804	offset = (q2 - 0.5f) * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
112		58804	f->block[j].energy[i] += offset;
113		58804	bits_left--;
114			}
115			}
116			}
117		29869	}
118
119		29869	static void celt_decode_tf_changes(CeltFrame f, OpusRangeCoder rc)
120			{
121		29869	int i, diff = 0, tf_select = 0, tf_changed = 0, tf_select_bit;
122	✓✓	29869	int consumed, bits = f->transient ? 2 : 4;
123
124		29869	consumed = opus_rc_tell(rc);
125	✓✓✓✓	29869	tf_select_bit = (f->size != 0 && consumed+bits+1 <= f->framebits);
126
127	✓✓	537903	for (i = f->start_band; i < f->end_band; i++) {
128	✓✓	508034	if (consumed+bits+tf_select_bit <= f->framebits) {
129		505558	diff ^= ff_opus_rc_dec_log(rc, bits);
130		505558	consumed = opus_rc_tell(rc);
131		505558	tf_changed \|= diff;
132			}
133		508034	f->tf_change[i] = diff;
134	✓✓	508034	bits = f->transient ? 4 : 5;
135			}
136
137	✓✓	29869	if (tf_select_bit && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
138	✓✓	16071	ff_celt_tf_select[f->size][f->transient][1][tf_changed])
139		8088	tf_select = ff_opus_rc_dec_log(rc, 1);
140
141	✓✓	537903	for (i = f->start_band; i < f->end_band; i++) {
142		508034	f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
143			}
144		29869	}
145
146		48600	static void celt_denormalize(CeltFrame f, CeltBlock block, float *data)
147			{
148			int i, j;
149
150	✓✓	898292	for (i = f->start_band; i < f->end_band; i++) {
151		849692	float *dst = data + (ff_celt_freq_bands[i] << f->size);
152		849692	float log_norm = block->energy[i] + ff_celt_mean_energy[i];
153	✗✓	849692	float norm = exp2f(FFMIN(log_norm, 32.0f));
154
155	✓✓	13154552	for (j = 0; j < ff_celt_freq_range[i] << f->size; j++)
156		12304860	dst[j] *= norm;
157			}
158		48600	}
159
160		89012	static void celt_postfilter_apply_transition(CeltBlock block, float data)
161			{
162		89012	const int T0 = block->pf_period_old;
163		89012	const int T1 = block->pf_period;
164
165			float g00, g01, g02;
166			float g10, g11, g12;
167
168			float x0, x1, x2, x3, x4;
169
170			int i;
171
172	✓✓	89012	if (block->pf_gains[0] == 0.0 &&
173	✓✓	41570	block->pf_gains_old[0] == 0.0)
174		37390	return;
175
176		51622	g00 = block->pf_gains_old[0];
177		51622	g01 = block->pf_gains_old[1];
178		51622	g02 = block->pf_gains_old[2];
179		51622	g10 = block->pf_gains[0];
180		51622	g11 = block->pf_gains[1];
181		51622	g12 = block->pf_gains[2];
182
183		51622	x1 = data[-T1 + 1];
184		51622	x2 = data[-T1];
185		51622	x3 = data[-T1 - 1];
186		51622	x4 = data[-T1 - 2];
187
188	✓✓	6246262	for (i = 0; i < CELT_OVERLAP; i++) {
189		6194640	float w = ff_celt_window2[i];
190		6194640	x0 = data[i - T1 + 2];
191
192		6194640	data[i] += (1.0 - w) * g00 * data[i - T0] +
193		6194640	(1.0 - w) * g01 * (data[i - T0 - 1] + data[i - T0 + 1]) +
194		6194640	(1.0 - w) * g02 * (data[i - T0 - 2] + data[i - T0 + 2]) +
195		6194640	w * g10 * x2 +
196		6194640	w * g11 * (x1 + x3) +
197		6194640	w * g12 * (x0 + x4);
198		6194640	x4 = x3;
199		6194640	x3 = x2;
200		6194640	x2 = x1;
201		6194640	x1 = x0;
202			}
203			}
204
205		56678	static void celt_postfilter(CeltFrame f, CeltBlock block)
206			{
207		56678	int len = f->blocksize * f->blocks;
208		56678	const int filter_len = len - 2 * CELT_OVERLAP;
209
210		56678	celt_postfilter_apply_transition(block, block->buf + 1024);
211
212		56678	block->pf_period_old = block->pf_period;
213		56678	memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
214
215		56678	block->pf_period = block->pf_period_new;
216		56678	memcpy(block->pf_gains, block->pf_gains_new, sizeof(block->pf_gains));
217
218	✓✓	56678	if (len > CELT_OVERLAP) {
219		32334	celt_postfilter_apply_transition(block, block->buf + 1024 + CELT_OVERLAP);
220
221	✓✓✓✓	32334	if (block->pf_gains[0] > FLT_EPSILON && filter_len > 0)
222		10124	f->opusdsp.postfilter(block->buf + 1024 + 2 * CELT_OVERLAP,
223		10124	block->pf_period, block->pf_gains,
224			filter_len);
225
226		32334	block->pf_period_old = block->pf_period;
227		32334	memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
228			}
229
230		56678	memmove(block->buf, block->buf + len, (1024 + CELT_OVERLAP / 2) * sizeof(float));
231		56678	}
232
233		29869	static int parse_postfilter(CeltFrame f, OpusRangeCoder rc, int consumed)
234			{
235			int i;
236
237		29869	memset(f->block[0].pf_gains_new, 0, sizeof(f->block[0].pf_gains_new));
238		29869	memset(f->block[1].pf_gains_new, 0, sizeof(f->block[1].pf_gains_new));
239
240	✓✓✓✓	29869	if (f->start_band == 0 && consumed + 16 <= f->framebits) {
241		24809	int has_postfilter = ff_opus_rc_dec_log(rc, 1);
242	✓✓	24809	if (has_postfilter) {
243			float gain;
244			int tapset, octave, period;
245
246		15873	octave = ff_opus_rc_dec_uint(rc, 6);
247		15873	period = (16 << octave) + ff_opus_rc_get_raw(rc, 4 + octave) - 1;
248		15873	gain = 0.09375f * (ff_opus_rc_get_raw(rc, 3) + 1);
249		15873	tapset = (opus_rc_tell(rc) + 2 <= f->framebits) ?
250	✓✗	15873	ff_opus_rc_dec_cdf(rc, ff_celt_model_tapset) : 0;
251
252	✓✓	47619	for (i = 0; i < 2; i++) {
253		31746	CeltBlock *block = &f->block[i];
254
255		31746	block->pf_period_new = FFMAX(period, CELT_POSTFILTER_MINPERIOD);
256		31746	block->pf_gains_new[0] = gain * ff_celt_postfilter_taps[tapset][0];
257		31746	block->pf_gains_new[1] = gain * ff_celt_postfilter_taps[tapset][1];
258		31746	block->pf_gains_new[2] = gain * ff_celt_postfilter_taps[tapset][2];
259			}
260			}
261
262		24809	consumed = opus_rc_tell(rc);
263			}
264
265		29869	return consumed;
266			}
267
268		2750	static void process_anticollapse(CeltFrame f, CeltBlock block, float *X)
269			{
270			int i, j, k;
271
272	✓✓	37875	for (i = f->start_band; i < f->end_band; i++) {
273		35125	int renormalize = 0;
274			float *xptr;
275			float prev[2];
276			float Ediff, r;
277			float thresh, sqrt_1;
278			int depth;
279
280			/* depth in 1/8 bits */
281		35125	depth = (1 + f->pulses[i]) / (ff_celt_freq_range[i] << f->size);
282		35125	thresh = exp2f(-1.0 - 0.125f * depth);
283		35125	sqrt_1 = 1.0f / sqrtf(ff_celt_freq_range[i] << f->size);
284
285		35125	xptr = X + (ff_celt_freq_bands[i] << f->size);
286
287		35125	prev[0] = block->prev_energy[0][i];
288		35125	prev[1] = block->prev_energy[1][i];
289	✓✓	35125	if (f->channels == 1) {
290		9975	CeltBlock *block1 = &f->block[1];
291
292	✓✓	9975	prev[0] = FFMAX(prev[0], block1->prev_energy[0][i]);
293	✓✓	9975	prev[1] = FFMAX(prev[1], block1->prev_energy[1][i]);
294			}
295	✓✓	35125	Ediff = block->energy[i] - FFMIN(prev[0], prev[1]);
296	✓✓	35125	Ediff = FFMAX(0, Ediff);
297
298			/* r needs to be multiplied by 2 or 2*sqrt(2) depending on LM because
299			short blocks don't have the same energy as long */
300		35125	r = exp2f(1 - Ediff);
301	✓✓	35125	if (f->size == 3)
302		23564	r *= M_SQRT2;
303	✓✓	35125	r = FFMIN(thresh, r) * sqrt_1;
304	✓✓	269881	for (k = 0; k < 1 << f->size; k++) {
305			/* Detect collapse */
306	✓✓	234756	if (!(block->collapse_masks[i] & 1 << k)) {
307			/* Fill with noise */
308	✓✓	115623	for (j = 0; j < ff_celt_freq_range[i]; j++)
309	✓✓	95928	xptr[(j << f->size) + k] = (celt_rng(f) & 0x8000) ? r : -r;
310		19695	renormalize = 1;
311			}
312			}
313
314			/* We just added some energy, so we need to renormalize */
315	✓✓	35125	if (renormalize)
316		8458	celt_renormalize_vector(xptr, ff_celt_freq_range[i] << f->size, 1.0f);
317			}
318		2750	}
319
320		29869	int ff_celt_decode_frame(CeltFrame f, OpusRangeCoder rc,
321			float **output, int channels, int frame_size,
322			int start_band, int end_band)
323			{
324		29869	int i, j, downmix = 0;
325			int consumed; // bits of entropy consumed thus far for this frame
326			MDCT15Context *imdct;
327
328	✓✓✗✓	29869	if (channels != 1 && channels != 2) {
329			av_log(f->avctx, AV_LOG_ERROR, "Invalid number of coded channels: %d\n",
330			channels);
331			return AVERROR_INVALIDDATA;
332			}
333	✓✗✓✗ ✗✓	29869	if (start_band < 0 \|\| start_band > end_band \|\| end_band > CELT_MAX_BANDS) {
334			av_log(f->avctx, AV_LOG_ERROR, "Invalid start/end band: %d %d\n",
335			start_band, end_band);
336			return AVERROR_INVALIDDATA;
337			}
338
339		29869	f->silence = 0;
340		29869	f->transient = 0;
341		29869	f->anticollapse = 0;
342		29869	f->flushed = 0;
343		29869	f->channels = channels;
344		29869	f->start_band = start_band;
345		29869	f->end_band = end_band;
346		29869	f->framebits = rc->rb.bytes * 8;
347
348		29869	f->size = av_log2(frame_size / CELT_SHORT_BLOCKSIZE);
349	✓✗	29869	if (f->size > CELT_MAX_LOG_BLOCKS \|\|
350	✗✓	29869	frame_size != CELT_SHORT_BLOCKSIZE * (1 << f->size)) {
351			av_log(f->avctx, AV_LOG_ERROR, "Invalid CELT frame size: %d\n",
352			frame_size);
353			return AVERROR_INVALIDDATA;
354			}
355
356	✗✓	29869	if (!f->output_channels)
357			f->output_channels = channels;
358
359	✓✓	78469	for (i = 0; i < f->channels; i++) {
360		48600	memset(f->block[i].coeffs, 0, sizeof(f->block[i].coeffs));
361		48600	memset(f->block[i].collapse_masks, 0, sizeof(f->block[i].collapse_masks));
362			}
363
364		29869	consumed = opus_rc_tell(rc);
365
366			/* obtain silence flag */
367	✗✓	29869	if (consumed >= f->framebits)
368			f->silence = 1;
369	✓✓	29869	else if (consumed == 1)
370		24934	f->silence = ff_opus_rc_dec_log(rc, 15);
371
372
373	✓✓	29869	if (f->silence) {
374		125	consumed = f->framebits;
375		125	rc->total_bits += f->framebits - opus_rc_tell(rc);
376			}
377
378			/* obtain post-filter options */
379		29869	consumed = parse_postfilter(f, rc, consumed);
380
381			/* obtain transient flag */
382	✓✓✓✓	29869	if (f->size != 0 && consumed+3 <= f->framebits)
383		16079	f->transient = ff_opus_rc_dec_log(rc, 3);
384
385	✓✓	29869	f->blocks = f->transient ? 1 << f->size : 1;
386		29869	f->blocksize = frame_size / f->blocks;
387
388	✓✓	29869	imdct = f->imdct[f->transient ? 0 : f->size];
389
390	✓✓	29869	if (channels == 1) {
391	✓✓	245036	for (i = 0; i < CELT_MAX_BANDS; i++)
392	✓✓	233898	f->block[0].energy[i] = FFMAX(f->block[0].energy[i], f->block[1].energy[i]);
393			}
394
395		29869	celt_decode_coarse_energy(f, rc);
396		29869	celt_decode_tf_changes (f, rc);
397		29869	ff_celt_bitalloc (f, rc, 0);
398		29869	celt_decode_fine_energy (f, rc);
399		29869	ff_celt_quant_bands (f, rc);
400
401	✓✓	29869	if (f->anticollapse_needed)
402		3123	f->anticollapse = ff_opus_rc_get_raw(rc, 1);
403
404		29869	celt_decode_final_energy(f, rc);
405
406			/* apply anti-collapse processing and denormalization to
407			* each coded channel */
408	✓✓	78469	for (i = 0; i < f->channels; i++) {
409		48600	CeltBlock *block = &f->block[i];
410
411	✓✓	48600	if (f->anticollapse)
412		2750	process_anticollapse(f, block, f->block[i].coeffs);
413
414		48600	celt_denormalize(f, block, f->block[i].coeffs);
415			}
416
417			/* stereo -> mono downmix */
418	✗✓	29869	if (f->output_channels < f->channels) {
419			f->dsp->vector_fmac_scalar(f->block[0].coeffs, f->block[1].coeffs, 1.0, FFALIGN(frame_size, 16));
420			downmix = 1;
421	✓✓	29869	} else if (f->output_channels > f->channels)
422		8078	memcpy(f->block[1].coeffs, f->block[0].coeffs, frame_size * sizeof(float));
423
424	✓✓	29869	if (f->silence) {
425	✓✓	375	for (i = 0; i < 2; i++) {
426		250	CeltBlock *block = &f->block[i];
427
428	✓✓	5500	for (j = 0; j < FF_ARRAY_ELEMS(block->energy); j++)
429		5250	block->energy[j] = CELT_ENERGY_SILENCE;
430			}
431		125	memset(f->block[0].coeffs, 0, sizeof(f->block[0].coeffs));
432		125	memset(f->block[1].coeffs, 0, sizeof(f->block[1].coeffs));
433			}
434
435			/* transform and output for each output channel */
436	✓✓	86547	for (i = 0; i < f->output_channels; i++) {
437		56678	CeltBlock *block = &f->block[i];
438
439			/* iMDCT and overlap-add */
440	✓✓	149134	for (j = 0; j < f->blocks; j++) {
441		92456	float dst = block->buf + 1024 + j f->blocksize;
442
443		92456	imdct->imdct_half(imdct, dst + CELT_OVERLAP / 2, f->block[i].coeffs + j,
444		92456	f->blocks);
445		92456	f->dsp->vector_fmul_window(dst, dst, dst + CELT_OVERLAP / 2,
446			ff_celt_window, CELT_OVERLAP / 2);
447			}
448
449	✗✓	56678	if (downmix)
450			f->dsp->vector_fmul_scalar(&block->buf[1024], &block->buf[1024], 0.5f, frame_size);
451
452			/* postfilter */
453		56678	celt_postfilter(f, block);
454
455			/* deemphasis */
456		56678	block->emph_coeff = f->opusdsp.deemphasis(output[i],
457		56678	&block->buf[1024 - frame_size],
458			block->emph_coeff, frame_size);
459			}
460
461	✓✓	29869	if (channels == 1)
462		11138	memcpy(f->block[1].energy, f->block[0].energy, sizeof(f->block[0].energy));
463
464	✓✓	89607	for (i = 0; i < 2; i++ ) {
465		59738	CeltBlock *block = &f->block[i];
466
467	✓✓	59738	if (!f->transient) {
468		51720	memcpy(block->prev_energy[1], block->prev_energy[0], sizeof(block->prev_energy[0]));
469		51720	memcpy(block->prev_energy[0], block->energy, sizeof(block->prev_energy[0]));
470			} else {
471	✓✓	176396	for (j = 0; j < CELT_MAX_BANDS; j++)
472	✓✓	168378	block->prev_energy[0][j] = FFMIN(block->prev_energy[0][j], block->energy[j]);
473			}
474
475	✓✓	227528	for (j = 0; j < f->start_band; j++) {
476		167790	block->prev_energy[0][j] = CELT_ENERGY_SILENCE;
477		167790	block->energy[j] = 0.0;
478			}
479	✓✓	130378	for (j = f->end_band; j < CELT_MAX_BANDS; j++) {
480		70640	block->prev_energy[0][j] = CELT_ENERGY_SILENCE;
481		70640	block->energy[j] = 0.0;
482			}
483			}
484
485		29869	f->seed = rc->range;
486
487		29869	return 0;
488			}
489
490		4597	void ff_celt_flush(CeltFrame *f)
491			{
492			int i, j;
493
494	✓✓	4597	if (f->flushed)
495		4533	return;
496
497	✓✓	192	for (i = 0; i < 2; i++) {
498		128	CeltBlock *block = &f->block[i];
499
500	✓✓	2816	for (j = 0; j < CELT_MAX_BANDS; j++)
501		2688	block->prev_energy[0][j] = block->prev_energy[1][j] = CELT_ENERGY_SILENCE;
502
503		128	memset(block->energy, 0, sizeof(block->energy));
504		128	memset(block->buf, 0, sizeof(block->buf));
505
506		128	memset(block->pf_gains, 0, sizeof(block->pf_gains));
507		128	memset(block->pf_gains_old, 0, sizeof(block->pf_gains_old));
508		128	memset(block->pf_gains_new, 0, sizeof(block->pf_gains_new));
509
510			/* libopus uses CELT_EMPH_COEFF on init, but 0 is better since there's
511			* a lesser discontinuity when seeking.
512			* The deemphasis functions differ from libopus in that they require
513			* an initial state divided by the coefficient. */
514		128	block->emph_coeff = 0.0f / CELT_EMPH_COEFF;
515			}
516		64	f->seed = 0;
517
518		64	f->flushed = 1;
519			}
520
521		41	void ff_celt_free(CeltFrame **f)
522			{
523		41	CeltFrame frm = f;
524			int i;
525
526	✗✓	41	if (!frm)
527			return;
528
529	✓✓	205	for (i = 0; i < FF_ARRAY_ELEMS(frm->imdct); i++)
530		164	ff_mdct15_uninit(&frm->imdct[i]);
531
532		41	ff_celt_pvq_uninit(&frm->pvq);
533
534		41	av_freep(&frm->dsp);
535		41	av_freep(f);
536			}
537
538		41	int ff_celt_init(AVCodecContext avctx, CeltFrame *f, int output_channels,
539			int apply_phase_inv)
540			{
541			CeltFrame *frm;
542			int i, ret;
543
544	✓✓✗✓	41	if (output_channels != 1 && output_channels != 2) {
545			av_log(avctx, AV_LOG_ERROR, "Invalid number of output channels: %d\n",
546			output_channels);
547			return AVERROR(EINVAL);
548			}
549
550		41	frm = av_mallocz(sizeof(*frm));
551	✗✓	41	if (!frm)
552			return AVERROR(ENOMEM);
553
554		41	frm->avctx = avctx;
555		41	frm->output_channels = output_channels;
556		41	frm->apply_phase_inv = apply_phase_inv;
557
558	✓✓	205	for (i = 0; i < FF_ARRAY_ELEMS(frm->imdct); i++)
559	✗✓	164	if ((ret = ff_mdct15_init(&frm->imdct[i], 1, i + 3, -1.0f/32768)) < 0)
560			goto fail;
561
562	✗✓	41	if ((ret = ff_celt_pvq_init(&frm->pvq, 0)) < 0)
563			goto fail;
564
565		41	frm->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
566	✗✓	41	if (!frm->dsp) {
567			ret = AVERROR(ENOMEM);
568			goto fail;
569			}
570
571		41	ff_opus_dsp_init(&frm->opusdsp);
572		41	ff_celt_flush(frm);
573
574		41	*f = frm;
575
576		41	return 0;
577			fail:
578			ff_celt_free(&frm);
579			return ret;
580			}