diff options
author | Rostislav Pehlivanov <atomnuker@gmail.com> | 2017-12-30 17:02:54 +0000 |
---|---|---|
committer | Rostislav Pehlivanov <atomnuker@gmail.com> | 2017-12-30 17:05:23 +0000 |
commit | 51027d0b8b2835d4c70c9cb7b2ab5e28d5e3f22f (patch) | |
tree | 602207df544f63dac1fb907c6ddc59bf4ed06b79 | |
parent | 0c78b6a416c661afed745d44495b5194f714fb96 (diff) | |
download | ffmpeg-51027d0b8b2835d4c70c9cb7b2ab5e28d5e3f22f.tar.gz |
opus: merge encoder and decoder bitallocation functions into one
There's no difference apart from which entropy coding functions get called.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
-rw-r--r-- | libavcodec/opus.c | 348 | ||||
-rw-r--r-- | libavcodec/opus.h | 3 | ||||
-rw-r--r-- | libavcodec/opus_celt.c | 334 | ||||
-rw-r--r-- | libavcodec/opusenc.c | 337 | ||||
-rw-r--r-- | libavcodec/opusenc.h | 2 | ||||
-rw-r--r-- | libavcodec/opusenc_psy.c | 2 |
6 files changed, 358 insertions, 668 deletions
diff --git a/libavcodec/opus.c b/libavcodec/opus.c index 46b749cae6..9cbf4aed92 100644 --- a/libavcodec/opus.c +++ b/libavcodec/opus.c @@ -546,3 +546,351 @@ void ff_celt_quant_bands(CeltFrame *f, OpusRangeCoder *rc) update_lowband = (b > band_size << 3); } } + +#define NORMC(bits) ((bits) << (f->channels - 1) << f->size >> 2) + +void ff_celt_bitalloc(CeltFrame *f, OpusRangeCoder *rc, int encode) +{ + int i, j, low, high, total, done, bandbits, remaining, tbits_8ths; + int skip_startband = f->start_band; + int skip_bit = 0; + int intensitystereo_bit = 0; + int dualstereo_bit = 0; + int dynalloc = 6; + int extrabits = 0; + + int boost[CELT_MAX_BANDS] = { 0 }; + int trim_offset[CELT_MAX_BANDS]; + int threshold[CELT_MAX_BANDS]; + int bits1[CELT_MAX_BANDS]; + int bits2[CELT_MAX_BANDS]; + + /* Spread */ + if (opus_rc_tell(rc) + 4 <= f->framebits) + if (encode) + ff_opus_rc_enc_cdf(rc, f->spread, ff_celt_model_spread); + else + f->spread = ff_opus_rc_dec_cdf(rc, ff_celt_model_spread); + else + f->spread = CELT_SPREAD_NORMAL; + + /* Initialize static allocation caps */ + for (i = 0; i < CELT_MAX_BANDS; i++) + f->caps[i] = NORMC((ff_celt_static_caps[f->size][f->channels - 1][i] + 64) * ff_celt_freq_range[i]); + + /* Band boosts */ + tbits_8ths = f->framebits << 3; + for (i = f->start_band; i < f->end_band; i++) { + int quanta = ff_celt_freq_range[i] << (f->channels - 1) << f->size; + int b_dynalloc = dynalloc; + int boost_amount = f->alloc_boost[i]; + quanta = FFMIN(quanta << 3, FFMAX(6 << 3, quanta)); + + while (opus_rc_tell_frac(rc) + (b_dynalloc << 3) < tbits_8ths && boost[i] < f->caps[i]) { + int is_boost; + if (encode) { + is_boost = boost_amount--; + ff_opus_rc_enc_log(rc, is_boost, b_dynalloc); + } else { + is_boost = ff_opus_rc_dec_log(rc, b_dynalloc); + } + + if (!is_boost) + break; + + boost[i] += quanta; + tbits_8ths -= quanta; + + b_dynalloc = 1; + } + + if (boost[i]) + dynalloc = FFMAX(dynalloc - 1, 2); + } + + /* Allocation trim */ + if (opus_rc_tell_frac(rc) + (6 << 3) <= tbits_8ths) + if (encode) + ff_opus_rc_enc_cdf(rc, f->alloc_trim, ff_celt_model_alloc_trim); + else + f->alloc_trim = ff_opus_rc_dec_cdf(rc, ff_celt_model_alloc_trim); + + /* Anti-collapse bit reservation */ + tbits_8ths = (f->framebits << 3) - opus_rc_tell_frac(rc) - 1; + f->anticollapse_needed = 0; + if (f->transient && f->size >= 2 && tbits_8ths >= ((f->size + 2) << 3)) + f->anticollapse_needed = 1 << 3; + tbits_8ths -= f->anticollapse_needed; + + /* Band skip bit reservation */ + if (tbits_8ths >= 1 << 3) + skip_bit = 1 << 3; + tbits_8ths -= skip_bit; + + /* Intensity/dual stereo bit reservation */ + if (f->channels == 2) { + intensitystereo_bit = ff_celt_log2_frac[f->end_band - f->start_band]; + if (intensitystereo_bit <= tbits_8ths) { + tbits_8ths -= intensitystereo_bit; + if (tbits_8ths >= 1 << 3) { + dualstereo_bit = 1 << 3; + tbits_8ths -= 1 << 3; + } + } else { + intensitystereo_bit = 0; + } + } + + /* Trim offsets */ + for (i = f->start_band; i < f->end_band; i++) { + int trim = f->alloc_trim - 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 = NORMC(ff_celt_freq_range[i] * ff_celt_static_alloc[center][i]); + + if (bandbits) + bandbits = FFMAX(bandbits + trim_offset[i], 0); + bandbits += boost[i]; + + if (bandbits >= threshold[i] || done) { + done = 1; + total += FFMIN(bandbits, f->caps[i]); + } else if (bandbits >= f->channels << 3) { + total += f->channels << 3; + } + } + + if (total > tbits_8ths) + high = center - 1; + else + low = center + 1; + } + high = low--; + + /* Bisection */ + for (i = f->start_band; i < f->end_band; i++) { + bits1[i] = NORMC(ff_celt_freq_range[i] * ff_celt_static_alloc[low][i]); + bits2[i] = high >= CELT_VECTORS ? f->caps[i] : + NORMC(ff_celt_freq_range[i] * ff_celt_static_alloc[high][i]); + + if (bits1[i]) + bits1[i] = FFMAX(bits1[i] + trim_offset[i], 0); + if (bits2[i]) + bits2[i] = FFMAX(bits2[i] + trim_offset[i], 0); + + if (low) + bits1[i] += boost[i]; + bits2[i] += boost[i]; + + if (boost[i]) + skip_startband = i; + bits2[i] = FFMAX(bits2[i] - bits1[i], 0); + } + + /* 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, f->caps[j]); + } else if (bandbits >= f->channels << 3) + total += f->channels << 3; + } + if (total > tbits_8ths) + high = center; + else + low = center; + } + + /* Bisection */ + 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, f->caps[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_startband) { + /* all remaining bands are not skipped */ + tbits_8ths += skip_bit; + break; + } + + /* determine the number of bits available for coding "do not skip" markers */ + remaining = tbits_8ths - 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]; + allocation += FFMAX(remaining - (ff_celt_freq_bands[j] - ff_celt_freq_bands[f->start_band]), 0); + + /* 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)) { + int do_not_skip; + if (encode) { + do_not_skip = f->coded_bands <= f->skip_band_floor; + ff_opus_rc_enc_log(rc, do_not_skip, 1); + } else { + do_not_skip = ff_opus_rc_dec_log(rc, 1); + } + + if (do_not_skip) + break; + + total += 1 << 3; + allocation -= 1 << 3; + } + + /* the band is skipped, so reclaim its bits */ + total -= f->pulses[j]; + if (intensitystereo_bit) { + total -= intensitystereo_bit; + intensitystereo_bit = ff_celt_log2_frac[j - f->start_band]; + total += intensitystereo_bit; + } + + total += f->pulses[j] = (allocation >= f->channels << 3) ? f->channels << 3 : 0; + } + + /* IS start band */ + if (encode) { + if (intensitystereo_bit) { + f->intensity_stereo = FFMIN(f->intensity_stereo, f->coded_bands); + ff_opus_rc_enc_uint(rc, f->intensity_stereo, f->coded_bands + 1 - f->start_band); + } + } else { + f->intensity_stereo = f->dual_stereo = 0; + if (intensitystereo_bit) + f->intensity_stereo = f->start_band + ff_opus_rc_dec_uint(rc, f->coded_bands + 1 - f->start_band); + } + + /* DS flag */ + if (f->intensity_stereo <= f->start_band) + tbits_8ths += dualstereo_bit; /* no intensity stereo means no dual stereo */ + else if (dualstereo_bit) + if (encode) + ff_opus_rc_enc_log(rc, f->dual_stereo, 1); + else + f->dual_stereo = ff_opus_rc_dec_log(rc, 1); + + /* Supply the remaining bits in this frame to lower bands */ + remaining = tbits_8ths - 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++) { + const int bits = FFMIN(remaining, ff_celt_freq_range[i]); + f->pulses[i] += bits + bandbits * ff_celt_freq_range[i]; + remaining -= bits; + } + + /* Finally determine the allocation */ + 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 fine_bits; + int max_bits; + int offset; /* fine energy quantization offset, i.e. + * extra bits assigned over the standard + * totalbits/dof */ + + extrabits = FFMAX(f->pulses[i] - f->caps[i], 0); + 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(f->pulses[i] - (f->channels << 3), 0); + 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; + } +} diff --git a/libavcodec/opus.h b/libavcodec/opus.h index a10db7f0e8..edbaab5ce7 100644 --- a/libavcodec/opus.h +++ b/libavcodec/opus.h @@ -194,4 +194,7 @@ int ff_silk_decode_superframe(SilkContext *s, OpusRangeCoder *rc, /* Encode or decode CELT bands */ void ff_celt_quant_bands(CeltFrame *f, OpusRangeCoder *rc); +/* Encode or decode CELT bitallocation */ +void ff_celt_bitalloc(CeltFrame *f, OpusRangeCoder *rc, int encode); + #endif /* AVCODEC_OPUS_H */ diff --git a/libavcodec/opus_celt.c b/libavcodec/opus_celt.c index ff74e2f067..115dd8c63e 100644 --- a/libavcodec/opus_celt.c +++ b/libavcodec/opus_celt.c @@ -143,338 +143,6 @@ 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; @@ -753,7 +421,7 @@ 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); ff_celt_quant_bands (f, rc); diff --git a/libavcodec/opusenc.c b/libavcodec/opusenc.c index 93b2116e7b..4068c43594 100644 --- a/libavcodec/opusenc.c +++ b/libavcodec/opusenc.c @@ -282,333 +282,6 @@ static void celt_enc_tf(CeltFrame *f, OpusRangeCoder *rc) f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]]; } -void ff_celt_enc_bitalloc(CeltFrame *f, OpusRangeCoder *rc) -{ - int i, j, low, high, total, done, bandbits, remaining, tbits_8ths; - int skip_startband = f->start_band; - int skip_bit = 0; - int intensitystereo_bit = 0; - int dualstereo_bit = 0; - int dynalloc = 6; - int extrabits = 0; - - int *cap = f->caps; - int boost[CELT_MAX_BANDS]; - int trim_offset[CELT_MAX_BANDS]; - int threshold[CELT_MAX_BANDS]; - int bits1[CELT_MAX_BANDS]; - int bits2[CELT_MAX_BANDS]; - - /* Tell the spread to the decoder */ - if (opus_rc_tell(rc) + 4 <= f->framebits) - ff_opus_rc_enc_cdf(rc, f->spread, ff_celt_model_spread); - else - f->spread = CELT_SPREAD_NORMAL; - - /* 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; - } - - /* Band boosts */ - tbits_8ths = f->framebits << 3; - for (i = f->start_band; i < f->end_band; i++) { - int quanta, b_dynalloc, boost_amount = f->alloc_boost[i]; - - boost[i] = 0; - - quanta = ff_celt_freq_range[i] << (f->channels - 1) << f->size; - quanta = FFMIN(quanta << 3, FFMAX(6 << 3, quanta)); - b_dynalloc = dynalloc; - - while (opus_rc_tell_frac(rc) + (b_dynalloc << 3) < tbits_8ths && boost[i] < cap[i]) { - int is_boost = boost_amount--; - - ff_opus_rc_enc_log(rc, is_boost, b_dynalloc); - if (!is_boost) - break; - - boost[i] += quanta; - tbits_8ths -= quanta; - - b_dynalloc = 1; - } - - if (boost[i]) - dynalloc = FFMAX(2, dynalloc - 1); - } - - /* Put allocation trim */ - if (opus_rc_tell_frac(rc) + (6 << 3) <= tbits_8ths) - ff_opus_rc_enc_cdf(rc, f->alloc_trim, ff_celt_model_alloc_trim); - - /* Anti-collapse bit reservation */ - tbits_8ths = (f->framebits << 3) - opus_rc_tell_frac(rc) - 1; - f->anticollapse_needed = 0; - if (f->transient && f->size >= 2 && tbits_8ths >= ((f->size + 2) << 3)) - f->anticollapse_needed = 1 << 3; - tbits_8ths -= f->anticollapse_needed; - - /* Band skip bit reservation */ - if (tbits_8ths >= 1 << 3) - skip_bit = 1 << 3; - tbits_8ths -= skip_bit; - - /* Intensity/dual stereo bit reservation */ - if (f->channels == 2) { - intensitystereo_bit = ff_celt_log2_frac[f->end_band - f->start_band]; - if (intensitystereo_bit <= tbits_8ths) { - tbits_8ths -= intensitystereo_bit; - if (tbits_8ths >= 1 << 3) { - dualstereo_bit = 1 << 3; - tbits_8ths -= 1 << 3; - } - } else { - intensitystereo_bit = 0; - } - } - - /* Trim offsets */ - for (i = f->start_band; i < f->end_band; i++) { - int trim = f->alloc_trim - 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 > tbits_8ths) - high = center - 1; - else - low = center + 1; - } - high = low--; - - /* Bisection */ - 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_startband = 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 > tbits_8ths) - high = center; - else - low = center; - } - - /* Bisection */ - 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_startband) { - /* all remaining bands are not skipped */ - tbits_8ths += skip_bit; - break; - } - - /* determine the number of bits available for coding "do not skip" markers */ - remaining = tbits_8ths - 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)) { - const int do_not_skip = f->coded_bands <= f->skip_band_floor; - ff_opus_rc_enc_log(rc, do_not_skip, 1); - if (do_not_skip) - break; - - total += 1 << 3; - allocation -= 1 << 3; - } - - /* the band is skipped, so reclaim its bits */ - total -= f->pulses[j]; - if (intensitystereo_bit) { - total -= intensitystereo_bit; - intensitystereo_bit = ff_celt_log2_frac[j - f->start_band]; - total += intensitystereo_bit; - } - - total += f->pulses[j] = (allocation >= f->channels << 3) ? f->channels << 3 : 0; - } - - /* Encode stereo flags */ - if (intensitystereo_bit) { - f->intensity_stereo = FFMIN(f->intensity_stereo, f->coded_bands); - ff_opus_rc_enc_uint(rc, f->intensity_stereo, f->coded_bands + 1 - f->start_band); - } - if (f->intensity_stereo <= f->start_band) - tbits_8ths += dualstereo_bit; /* no intensity stereo means no dual stereo */ - else if (dualstereo_bit) - ff_opus_rc_enc_log(rc, f->dual_stereo, 1); - - /* Supply the remaining bits in this frame to lower bands */ - remaining = tbits_8ths - 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; - } - - /* Finally determine the allocation */ - 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_enc_quant_pfilter(OpusRangeCoder *rc, CeltFrame *f) { float gain = f->pf_gain; @@ -794,11 +467,11 @@ static void celt_encode_frame(OpusEncContext *s, OpusRangeCoder *rc, ff_opus_rc_enc_log(rc, f->transient, 3); /* Main encoding */ - celt_quant_coarse (f, rc, s->last_quantized_energy); - celt_enc_tf (f, rc); - ff_celt_enc_bitalloc(f, rc); - celt_quant_fine (f, rc); - ff_celt_quant_bands (f, rc); + celt_quant_coarse (f, rc, s->last_quantized_energy); + celt_enc_tf (f, rc); + ff_celt_bitalloc (f, rc, 1); + celt_quant_fine (f, rc); + ff_celt_quant_bands(f, rc); /* Anticollapse bit */ if (f->anticollapse_needed) diff --git a/libavcodec/opusenc.h b/libavcodec/opusenc.h index 56e4af67e3..b9162ebec6 100644 --- a/libavcodec/opusenc.h +++ b/libavcodec/opusenc.h @@ -51,6 +51,4 @@ typedef struct OpusPacketInfo { int frames; } OpusPacketInfo; -void ff_celt_enc_bitalloc(CeltFrame *f, OpusRangeCoder *rc); - #endif /* AVCODEC_OPUSENC_H */ diff --git a/libavcodec/opusenc_psy.c b/libavcodec/opusenc_psy.c index 5393c9d7de..50890c65d6 100644 --- a/libavcodec/opusenc_psy.c +++ b/libavcodec/opusenc_psy.c @@ -366,7 +366,7 @@ static int bands_dist(OpusPsyContext *s, CeltFrame *f, float *total_dist) OpusRangeCoder dump; ff_opus_rc_enc_init(&dump); - ff_celt_enc_bitalloc(f, &dump); + ff_celt_bitalloc(f, &dump, 1); for (i = 0; i < CELT_MAX_BANDS; i++) { float bits = 0.0f; |