X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fopus_celt.c;h=115dd8c63e5b2b5eade765c15208b05fe390b4a3;hb=2ca65fc7b74444edd51d5803a2c1e05a801a6023;hp=84d484753b361fe9292d3941a2510a0f771ace51;hpb=24204eae1c7dba402d88f0cd3077814dbb7f9376;p=ffmpeg diff --git a/libavcodec/opus_celt.c b/libavcodec/opus_celt.c index 84d484753b3..115dd8c63e5 100644 --- a/libavcodec/opus_celt.c +++ b/libavcodec/opus_celt.c @@ -143,345 +143,14 @@ static void celt_decode_tf_changes(CeltFrame *f, OpusRangeCoder *rc) } } -static void celt_decode_allocation(CeltFrame *f, OpusRangeCoder *rc) -{ - // approx. maximum bit allocation for each band before boost/trim - int cap[CELT_MAX_BANDS]; - int boost[CELT_MAX_BANDS]; - int threshold[CELT_MAX_BANDS]; - int bits1[CELT_MAX_BANDS]; - int bits2[CELT_MAX_BANDS]; - int trim_offset[CELT_MAX_BANDS]; - - int skip_start_band = f->start_band; - int dynalloc = 6; - int alloctrim = 5; - int extrabits = 0; - - int skip_bit = 0; - int intensity_stereo_bit = 0; - int dual_stereo_bit = 0; - - int remaining, bandbits; - int low, high, total, done; - int totalbits; - int consumed; - int i, j; - - consumed = opus_rc_tell(rc); - - /* obtain spread flag */ - f->spread = CELT_SPREAD_NORMAL; - if (consumed + 4 <= f->framebits) - f->spread = ff_opus_rc_dec_cdf(rc, ff_celt_model_spread); - - /* generate static allocation caps */ - for (i = 0; i < CELT_MAX_BANDS; i++) { - cap[i] = (ff_celt_static_caps[f->size][f->channels - 1][i] + 64) - * ff_celt_freq_range[i] << (f->channels - 1) << f->size >> 2; - } - - /* obtain band boost */ - totalbits = f->framebits << 3; // convert to 1/8 bits - consumed = opus_rc_tell_frac(rc); - for (i = f->start_band; i < f->end_band; i++) { - int quanta, band_dynalloc; - - boost[i] = 0; - - quanta = ff_celt_freq_range[i] << (f->channels - 1) << f->size; - quanta = FFMIN(quanta << 3, FFMAX(6 << 3, quanta)); - band_dynalloc = dynalloc; - while (consumed + (band_dynalloc<<3) < totalbits && boost[i] < cap[i]) { - int add = ff_opus_rc_dec_log(rc, band_dynalloc); - consumed = opus_rc_tell_frac(rc); - if (!add) - break; - - boost[i] += quanta; - totalbits -= quanta; - band_dynalloc = 1; - } - /* dynalloc is more likely to occur if it's already been used for earlier bands */ - if (boost[i]) - dynalloc = FFMAX(2, dynalloc - 1); - } - - /* obtain allocation trim */ - if (consumed + (6 << 3) <= totalbits) - alloctrim = ff_opus_rc_dec_cdf(rc, ff_celt_model_alloc_trim); - - /* anti-collapse bit reservation */ - totalbits = (f->framebits << 3) - opus_rc_tell_frac(rc) - 1; - f->anticollapse_needed = 0; - if (f->blocks > 1 && f->size >= 2 && - totalbits >= ((f->size + 2) << 3)) - f->anticollapse_needed = 1 << 3; - totalbits -= f->anticollapse_needed; - - /* band skip bit reservation */ - if (totalbits >= 1 << 3) - skip_bit = 1 << 3; - totalbits -= skip_bit; - - /* intensity/dual stereo bit reservation */ - if (f->channels == 2) { - intensity_stereo_bit = ff_celt_log2_frac[f->end_band - f->start_band]; - if (intensity_stereo_bit <= totalbits) { - totalbits -= intensity_stereo_bit; - if (totalbits >= 1 << 3) { - dual_stereo_bit = 1 << 3; - totalbits -= 1 << 3; - } - } else - intensity_stereo_bit = 0; - } - - for (i = f->start_band; i < f->end_band; i++) { - int trim = alloctrim - 5 - f->size; - int band = ff_celt_freq_range[i] * (f->end_band - i - 1); - int duration = f->size + 3; - int scale = duration + f->channels - 1; - - /* PVQ minimum allocation threshold, below this value the band is - * skipped */ - threshold[i] = FFMAX(3 * ff_celt_freq_range[i] << duration >> 4, - f->channels << 3); - - trim_offset[i] = trim * (band << scale) >> 6; - - if (ff_celt_freq_range[i] << f->size == 1) - trim_offset[i] -= f->channels << 3; - } - - /* bisection */ - low = 1; - high = CELT_VECTORS - 1; - while (low <= high) { - int center = (low + high) >> 1; - done = total = 0; - - for (i = f->end_band - 1; i >= f->start_band; i--) { - bandbits = ff_celt_freq_range[i] * ff_celt_static_alloc[center][i] - << (f->channels - 1) << f->size >> 2; - - if (bandbits) - bandbits = FFMAX(0, bandbits + trim_offset[i]); - bandbits += boost[i]; - - if (bandbits >= threshold[i] || done) { - done = 1; - total += FFMIN(bandbits, cap[i]); - } else if (bandbits >= f->channels << 3) - total += f->channels << 3; - } - - if (total > totalbits) - high = center - 1; - else - low = center + 1; - } - high = low--; - - for (i = f->start_band; i < f->end_band; i++) { - bits1[i] = ff_celt_freq_range[i] * ff_celt_static_alloc[low][i] - << (f->channels - 1) << f->size >> 2; - bits2[i] = high >= CELT_VECTORS ? cap[i] : - ff_celt_freq_range[i] * ff_celt_static_alloc[high][i] - << (f->channels - 1) << f->size >> 2; - - if (bits1[i]) - bits1[i] = FFMAX(0, bits1[i] + trim_offset[i]); - if (bits2[i]) - bits2[i] = FFMAX(0, bits2[i] + trim_offset[i]); - if (low) - bits1[i] += boost[i]; - bits2[i] += boost[i]; - - if (boost[i]) - skip_start_band = i; - bits2[i] = FFMAX(0, bits2[i] - bits1[i]); - } - - /* bisection */ - low = 0; - high = 1 << CELT_ALLOC_STEPS; - for (i = 0; i < CELT_ALLOC_STEPS; i++) { - int center = (low + high) >> 1; - done = total = 0; - - for (j = f->end_band - 1; j >= f->start_band; j--) { - bandbits = bits1[j] + (center * bits2[j] >> CELT_ALLOC_STEPS); - - if (bandbits >= threshold[j] || done) { - done = 1; - total += FFMIN(bandbits, cap[j]); - } else if (bandbits >= f->channels << 3) - total += f->channels << 3; - } - if (total > totalbits) - high = center; - else - low = center; - } - - done = total = 0; - for (i = f->end_band - 1; i >= f->start_band; i--) { - bandbits = bits1[i] + (low * bits2[i] >> CELT_ALLOC_STEPS); - - if (bandbits >= threshold[i] || done) - done = 1; - else - bandbits = (bandbits >= f->channels << 3) ? - f->channels << 3 : 0; - - bandbits = FFMIN(bandbits, cap[i]); - f->pulses[i] = bandbits; - total += bandbits; - } - - /* band skipping */ - for (f->coded_bands = f->end_band; ; f->coded_bands--) { - int allocation; - j = f->coded_bands - 1; - - if (j == skip_start_band) { - /* all remaining bands are not skipped */ - totalbits += skip_bit; - break; - } - - /* determine the number of bits available for coding "do not skip" markers */ - remaining = totalbits - total; - bandbits = remaining / (ff_celt_freq_bands[j+1] - ff_celt_freq_bands[f->start_band]); - remaining -= bandbits * (ff_celt_freq_bands[j+1] - ff_celt_freq_bands[f->start_band]); - allocation = f->pulses[j] + bandbits * ff_celt_freq_range[j] - + FFMAX(0, remaining - (ff_celt_freq_bands[j] - ff_celt_freq_bands[f->start_band])); - - /* a "do not skip" marker is only coded if the allocation is - above the chosen threshold */ - if (allocation >= FFMAX(threshold[j], (f->channels + 1) <<3 )) { - if (ff_opus_rc_dec_log(rc, 1)) - break; - - total += 1 << 3; - allocation -= 1 << 3; - } - - /* the band is skipped, so reclaim its bits */ - total -= f->pulses[j]; - if (intensity_stereo_bit) { - total -= intensity_stereo_bit; - intensity_stereo_bit = ff_celt_log2_frac[j - f->start_band]; - total += intensity_stereo_bit; - } - - total += f->pulses[j] = (allocation >= f->channels << 3) ? - f->channels << 3 : 0; - } - - /* obtain stereo flags */ - f->intensity_stereo = 0; - f->dual_stereo = 0; - if (intensity_stereo_bit) - f->intensity_stereo = f->start_band + - ff_opus_rc_dec_uint(rc, f->coded_bands + 1 - f->start_band); - if (f->intensity_stereo <= f->start_band) - totalbits += dual_stereo_bit; /* no intensity stereo means no dual stereo */ - else if (dual_stereo_bit) - f->dual_stereo = ff_opus_rc_dec_log(rc, 1); - - /* supply the remaining bits in this frame to lower bands */ - remaining = totalbits - total; - bandbits = remaining / (ff_celt_freq_bands[f->coded_bands] - ff_celt_freq_bands[f->start_band]); - remaining -= bandbits * (ff_celt_freq_bands[f->coded_bands] - ff_celt_freq_bands[f->start_band]); - for (i = f->start_band; i < f->coded_bands; i++) { - int bits = FFMIN(remaining, ff_celt_freq_range[i]); - - f->pulses[i] += bits + bandbits * ff_celt_freq_range[i]; - remaining -= bits; - } - - for (i = f->start_band; i < f->coded_bands; i++) { - int N = ff_celt_freq_range[i] << f->size; - int prev_extra = extrabits; - f->pulses[i] += extrabits; - - if (N > 1) { - int dof; // degrees of freedom - int temp; // dof * channels * log(dof) - int offset; // fine energy quantization offset, i.e. - // extra bits assigned over the standard - // totalbits/dof - int fine_bits, max_bits; - - extrabits = FFMAX(0, f->pulses[i] - cap[i]); - f->pulses[i] -= extrabits; - - /* intensity stereo makes use of an extra degree of freedom */ - dof = N * f->channels - + (f->channels == 2 && N > 2 && !f->dual_stereo && i < f->intensity_stereo); - temp = dof * (ff_celt_log_freq_range[i] + (f->size<<3)); - offset = (temp >> 1) - dof * CELT_FINE_OFFSET; - if (N == 2) /* dof=2 is the only case that doesn't fit the model */ - offset += dof<<1; - - /* grant an additional bias for the first and second pulses */ - if (f->pulses[i] + offset < 2 * (dof << 3)) - offset += temp >> 2; - else if (f->pulses[i] + offset < 3 * (dof << 3)) - offset += temp >> 3; - - fine_bits = (f->pulses[i] + offset + (dof << 2)) / (dof << 3); - max_bits = FFMIN((f->pulses[i]>>3) >> (f->channels - 1), - CELT_MAX_FINE_BITS); - - max_bits = FFMAX(max_bits, 0); - - f->fine_bits[i] = av_clip(fine_bits, 0, max_bits); - - /* if fine_bits was rounded down or capped, - give priority for the final fine energy pass */ - f->fine_priority[i] = (f->fine_bits[i] * (dof<<3) >= f->pulses[i] + offset); - - /* the remaining bits are assigned to PVQ */ - f->pulses[i] -= f->fine_bits[i] << (f->channels - 1) << 3; - } else { - /* all bits go to fine energy except for the sign bit */ - extrabits = FFMAX(0, f->pulses[i] - (f->channels << 3)); - f->pulses[i] -= extrabits; - f->fine_bits[i] = 0; - f->fine_priority[i] = 1; - } - - /* hand back a limited number of extra fine energy bits to this band */ - if (extrabits > 0) { - int fineextra = FFMIN(extrabits >> (f->channels + 2), - CELT_MAX_FINE_BITS - f->fine_bits[i]); - f->fine_bits[i] += fineextra; - - fineextra <<= f->channels + 2; - f->fine_priority[i] = (fineextra >= extrabits - prev_extra); - extrabits -= fineextra; - } - } - f->remaining = extrabits; - - /* skipped bands dedicate all of their bits for fine energy */ - for (; i < f->end_band; i++) { - f->fine_bits[i] = f->pulses[i] >> (f->channels - 1) >> 3; - f->pulses[i] = 0; - f->fine_priority[i] = f->fine_bits[i] < 1; - } -} - 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 norm = exp2f(block->energy[i] + ff_celt_mean_energy[i]); + 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; @@ -675,98 +344,6 @@ static void process_anticollapse(CeltFrame *f, CeltBlock *block, float *X) } } -static void celt_decode_bands(CeltFrame *f, OpusRangeCoder *rc) -{ - float lowband_scratch[8 * 22]; - float norm[2 * 8 * 100]; - - int totalbits = (f->framebits << 3) - f->anticollapse_needed; - - int update_lowband = 1; - int lowband_offset = 0; - - int i, j; - - memset(f->block[0].coeffs, 0, sizeof(f->block[0].coeffs)); - memset(f->block[1].coeffs, 0, sizeof(f->block[0].coeffs)); - - for (i = f->start_band; i < f->end_band; i++) { - uint32_t cm[2] = { (1 << f->blocks) - 1, (1 << f->blocks) - 1 }; - int band_offset = ff_celt_freq_bands[i] << f->size; - int band_size = ff_celt_freq_range[i] << f->size; - float *X = f->block[0].coeffs + band_offset; - float *Y = (f->channels == 2) ? f->block[1].coeffs + band_offset : NULL; - - int consumed = opus_rc_tell_frac(rc); - float *norm2 = norm + 8 * 100; - int effective_lowband = -1; - int b = 0; - - /* Compute how many bits we want to allocate to this band */ - if (i != f->start_band) - f->remaining -= consumed; - f->remaining2 = totalbits - consumed - 1; - if (i <= f->coded_bands - 1) { - int curr_balance = f->remaining / FFMIN(3, f->coded_bands-i); - b = av_clip_uintp2(FFMIN(f->remaining2 + 1, f->pulses[i] + curr_balance), 14); - } - - if (ff_celt_freq_bands[i] - ff_celt_freq_range[i] >= ff_celt_freq_bands[f->start_band] && - (update_lowband || lowband_offset == 0)) - lowband_offset = i; - - /* Get a conservative estimate of the collapse_mask's for the bands we're - going to be folding from. */ - if (lowband_offset != 0 && (f->spread != CELT_SPREAD_AGGRESSIVE || - f->blocks > 1 || f->tf_change[i] < 0)) { - int foldstart, foldend; - - /* This ensures we never repeat spectral content within one band */ - effective_lowband = FFMAX(ff_celt_freq_bands[f->start_band], - ff_celt_freq_bands[lowband_offset] - ff_celt_freq_range[i]); - foldstart = lowband_offset; - while (ff_celt_freq_bands[--foldstart] > effective_lowband); - foldend = lowband_offset - 1; - while (ff_celt_freq_bands[++foldend] < effective_lowband + ff_celt_freq_range[i]); - - cm[0] = cm[1] = 0; - for (j = foldstart; j < foldend; j++) { - cm[0] |= f->block[0].collapse_masks[j]; - cm[1] |= f->block[f->channels - 1].collapse_masks[j]; - } - } - - if (f->dual_stereo && i == f->intensity_stereo) { - /* Switch off dual stereo to do intensity */ - f->dual_stereo = 0; - for (j = ff_celt_freq_bands[f->start_band] << f->size; j < band_offset; j++) - norm[j] = (norm[j] + norm2[j]) / 2; - } - - if (f->dual_stereo) { - cm[0] = f->pvq->decode_band(f->pvq, f, rc, i, X, NULL, band_size, b / 2, f->blocks, - effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size, - norm + band_offset, 0, 1.0f, lowband_scratch, cm[0]); - - cm[1] = f->pvq->decode_band(f->pvq, f, rc, i, Y, NULL, band_size, b/2, f->blocks, - effective_lowband != -1 ? norm2 + (effective_lowband << f->size) : NULL, f->size, - norm2 + band_offset, 0, 1.0f, lowband_scratch, cm[1]); - } else { - cm[0] = f->pvq->decode_band(f->pvq, f, rc, i, X, Y, band_size, b, f->blocks, - effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size, - norm + band_offset, 0, 1.0f, lowband_scratch, cm[0]|cm[1]); - cm[1] = cm[0]; - } - - f->block[0].collapse_masks[i] = (uint8_t)cm[0]; - f->block[f->channels - 1].collapse_masks[i] = (uint8_t)cm[1]; - f->remaining += f->pulses[i] + consumed; - - /* Update the folding position only as long as we have 1 bit/sample depth */ - update_lowband = (b > band_size << 3); - } -} - int ff_celt_decode_frame(CeltFrame *f, OpusRangeCoder *rc, float **output, int channels, int frame_size, int start_band, int end_band) @@ -806,8 +383,10 @@ int ff_celt_decode_frame(CeltFrame *f, OpusRangeCoder *rc, if (!f->output_channels) f->output_channels = channels; - memset(f->block[0].collapse_masks, 0, sizeof(f->block[0].collapse_masks)); - memset(f->block[1].collapse_masks, 0, sizeof(f->block[1].collapse_masks)); + 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); @@ -842,9 +421,9 @@ int ff_celt_decode_frame(CeltFrame *f, OpusRangeCoder *rc, celt_decode_coarse_energy(f, rc); celt_decode_tf_changes (f, rc); - celt_decode_allocation (f, rc); + ff_celt_bitalloc (f, rc, 0); celt_decode_fine_energy (f, rc); - celt_decode_bands (f, rc); + ff_celt_quant_bands (f, rc); if (f->anticollapse_needed) f->anticollapse = ff_opus_rc_get_raw(rc, 1); @@ -984,7 +563,8 @@ void ff_celt_free(CeltFrame **f) av_freep(f); } -int ff_celt_init(AVCodecContext *avctx, CeltFrame **f, int output_channels) +int ff_celt_init(AVCodecContext *avctx, CeltFrame **f, int output_channels, + int apply_phase_inv) { CeltFrame *frm; int i, ret; @@ -1001,12 +581,13 @@ int ff_celt_init(AVCodecContext *avctx, CeltFrame **f, int output_channels) 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) + if ((ret = ff_celt_pvq_init(&frm->pvq, 0)) < 0) goto fail; frm->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);