[FFmpeg-devel] [PATCH 8/9] avcodec/opus: Rename opus.c->opus_celt.c, opus_celt.c->opusdec_celt.c
Andreas Rheinhardt
andreas.rheinhardt at outlook.com
Fri Oct 7 23:25:07 EEST 2022
Since commit 4fc2531fff112836026aad2bdaf128c9d15a72e3 opus.c
contains only the celt stuff shared between decoder and encoder.
meanwhile, opus_celt.c is decoder-only. So the new names
reflect the actual content better than the current ones.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt at outlook.com>
---
libavcodec/Makefile | 4 +-
libavcodec/opus.c | 522 ----------------------
libavcodec/opus_celt.c | 899 ++++++++++++++++++--------------------
libavcodec/opusdec_celt.c | 587 +++++++++++++++++++++++++
4 files changed, 1006 insertions(+), 1006 deletions(-)
delete mode 100644 libavcodec/opus.c
create mode 100644 libavcodec/opusdec_celt.c
diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index b7eb3b1e48..949c65a0e3 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -554,10 +554,10 @@ OBJS-$(CONFIG_NELLYMOSER_ENCODER) += nellymoserenc.o nellymoser.o
OBJS-$(CONFIG_NOTCHLC_DECODER) += notchlc.o
OBJS-$(CONFIG_NUV_DECODER) += nuv.o rtjpeg.o
OBJS-$(CONFIG_ON2AVC_DECODER) += on2avc.o on2avcdata.o
-OBJS-$(CONFIG_OPUS_DECODER) += opusdec.o opus.o opus_celt.o \
+OBJS-$(CONFIG_OPUS_DECODER) += opusdec.o opusdec_celt.o opus_celt.o \
opus_pvq.o opus_silk.o opustab.o vorbis_data.o \
opusdec_rc.o opusdsp.o opus_parse.o
-OBJS-$(CONFIG_OPUS_ENCODER) += opusenc.o opus.o opusenc_psy.o \
+OBJS-$(CONFIG_OPUS_ENCODER) += opusenc.o opus_celt.o opusenc_psy.o \
opusenc_rc.o opustab.o opus_pvq.o
OBJS-$(CONFIG_PAF_AUDIO_DECODER) += pafaudio.o
OBJS-$(CONFIG_PAF_VIDEO_DECODER) += pafvideo.o
diff --git a/libavcodec/opus.c b/libavcodec/opus.c
deleted file mode 100644
index 8def5e6e34..0000000000
--- a/libavcodec/opus.c
+++ /dev/null
@@ -1,522 +0,0 @@
-/*
- * Copyright (c) 2012 Andrew D'Addesio
- * Copyright (c) 2013-2014 Mozilla Corporation
- *
- * This file is part of FFmpeg.
- *
- * FFmpeg is free software; you can redistribute it and/or
- * modify it under the terms of the GNU Lesser General Public
- * License as published by the Free Software Foundation; either
- * version 2.1 of the License, or (at your option) any later version.
- *
- * FFmpeg is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * Lesser General Public License for more details.
- *
- * You should have received a copy of the GNU Lesser General Public
- * License along with FFmpeg; if not, write to the Free Software
- * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
- */
-
-#include <stdint.h>
-
-#include "config_components.h"
-#include "opus_celt.h"
-#include "opus_pvq.h"
-#include "opustab.h"
-#include "opus_rc.h"
-#include "opusdec_rc.h"
-#include "opusenc_rc.h"
-
-#if !CONFIG_OPUS_ENCODER
-#define ff_opus_rc_enc_log(...)
-#define ff_opus_rc_enc_cdf(...)
-#define ff_opus_rc_enc_uint(...)
-#endif
-
-#if !CONFIG_OPUS_DECODER
-#define ff_opus_rc_dec_log(...) 0
-#define ff_opus_rc_dec_cdf(...) 0
-#define ff_opus_rc_dec_uint(...) 0
-#endif
-
-static inline void opus_rc_enc_log(OpusRangeCoder *rc, int val, uint32_t bits)
-{
- ff_opus_rc_enc_log((OpusEncRangeCoder*)rc, val, bits);
-}
-
-static inline uint32_t opus_rc_dec_log(OpusRangeCoder *rc, uint32_t bits)
-{
- return ff_opus_rc_dec_log((OpusDecRangeCoder*)rc, bits);
-}
-
-static inline void opus_rc_enc_cdf(OpusRangeCoder *rc, int val, const uint16_t *cdf)
-{
- ff_opus_rc_enc_cdf((OpusEncRangeCoder*)rc, val, cdf);
-}
-
-static inline uint32_t opus_rc_dec_cdf(OpusRangeCoder *rc, const uint16_t *cdf)
-{
- return ff_opus_rc_dec_cdf((OpusDecRangeCoder*)rc, cdf);
-}
-
-void ff_celt_quant_bands(CeltFrame *f, OpusRangeCoder *rc)
-{
- float lowband_scratch[8 * 22];
- float norm1[2 * 8 * 100];
- float *norm2 = norm1 + 8 * 100;
-
- int totalbits = (f->framebits << 3) - f->anticollapse_needed;
-
- int update_lowband = 1;
- int lowband_offset = 0;
-
- int i, j;
-
- 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;
- float *norm_loc1, *norm_loc2;
-
- int consumed = opus_rc_tell_frac(rc);
- 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] ||
- i == f->start_band + 1) && (update_lowband || lowband_offset == 0))
- lowband_offset = i;
-
- if (i == f->start_band + 1) {
- /* Special Hybrid Folding (RFC 8251 section 9). Copy the first band into
- the second to ensure the second band never has to use the LCG. */
- int count = (ff_celt_freq_range[i] - ff_celt_freq_range[i-1]) << f->size;
-
- memcpy(&norm1[band_offset], &norm1[band_offset - count], count * sizeof(float));
-
- if (f->channels == 2)
- memcpy(&norm2[band_offset], &norm2[band_offset - count], count * sizeof(float));
- }
-
- /* 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 (++foldend < i && 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++)
- norm1[j] = (norm1[j] + norm2[j]) / 2;
- }
-
- norm_loc1 = effective_lowband != -1 ? norm1 + (effective_lowband << f->size) : NULL;
- norm_loc2 = effective_lowband != -1 ? norm2 + (effective_lowband << f->size) : NULL;
-
- if (f->dual_stereo) {
- cm[0] = f->pvq->quant_band(f->pvq, f, rc, i, X, NULL, band_size, b >> 1,
- f->blocks, norm_loc1, f->size,
- norm1 + band_offset, 0, 1.0f,
- lowband_scratch, cm[0]);
-
- cm[1] = f->pvq->quant_band(f->pvq, f, rc, i, Y, NULL, band_size, b >> 1,
- f->blocks, norm_loc2, f->size,
- norm2 + band_offset, 0, 1.0f,
- lowband_scratch, cm[1]);
- } else {
- cm[0] = f->pvq->quant_band(f->pvq, f, rc, i, X, Y, band_size, b >> 0,
- f->blocks, norm_loc1, f->size,
- norm1 + 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);
- }
-}
-
-#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];
-
- if (!CONFIG_OPUS_DECODER || !CONFIG_OPUS_ENCODER)
- encode = CONFIG_OPUS_ENCODER;
- /* Spread */
- if (opus_rc_tell(rc) + 4 <= f->framebits) {
- if (encode)
- opus_rc_enc_cdf(rc, f->spread, ff_celt_model_spread);
- else
- f->spread = 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--;
- opus_rc_enc_log(rc, is_boost, b_dynalloc);
- } else {
- is_boost = 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 (!encode)
- f->alloc_trim = 5;
- if (opus_rc_tell_frac(rc) + (6 << 3) <= tbits_8ths)
- if (encode)
- opus_rc_enc_cdf(rc, f->alloc_trim, ff_celt_model_alloc_trim);
- else
- f->alloc_trim = 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;
- opus_rc_enc_log(rc, do_not_skip, 1);
- } else {
- do_not_skip = 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((OpusEncRangeCoder*)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((OpusDecRangeCoder*)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)
- opus_rc_enc_log(rc, f->dual_stereo, 1);
- else
- f->dual_stereo = 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_celt.c b/libavcodec/opus_celt.c
index a14764ec18..8def5e6e34 100644
--- a/libavcodec/opus_celt.c
+++ b/libavcodec/opus_celt.c
@@ -1,7 +1,6 @@
/*
* Copyright (c) 2012 Andrew D'Addesio
* Copyright (c) 2013-2014 Mozilla Corporation
- * Copyright (c) 2016 Rostislav Pehlivanov <atomnuker at gmail.com>
*
* This file is part of FFmpeg.
*
@@ -20,568 +19,504 @@
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
-/**
- * @file
- * Opus CELT decoder
- */
-
-#include <float.h>
+#include <stdint.h>
+#include "config_components.h"
#include "opus_celt.h"
-#include "opusdec_rc.h"
-#include "opustab.h"
#include "opus_pvq.h"
+#include "opustab.h"
+#include "opus_rc.h"
+#include "opusdec_rc.h"
+#include "opusenc_rc.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, OpusDecRangeCoder *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->c) + 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;
- }
+#if !CONFIG_OPUS_ENCODER
+#define ff_opus_rc_enc_log(...)
+#define ff_opus_rc_enc_cdf(...)
+#define ff_opus_rc_enc_uint(...)
+#endif
- available = f->framebits - opus_rc_tell(&rc->c);
- 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;
- }
- }
-}
+#if !CONFIG_OPUS_DECODER
+#define ff_opus_rc_dec_log(...) 0
+#define ff_opus_rc_dec_cdf(...) 0
+#define ff_opus_rc_dec_uint(...) 0
+#endif
-static void celt_decode_fine_energy(CeltFrame *f, OpusDecRangeCoder *rc)
+static inline void opus_rc_enc_log(OpusRangeCoder *rc, int val, uint32_t bits)
{
- 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;
- }
- }
+ ff_opus_rc_enc_log((OpusEncRangeCoder*)rc, val, bits);
}
-static void celt_decode_final_energy(CeltFrame *f, OpusDecRangeCoder *rc)
+static inline uint32_t opus_rc_dec_log(OpusRangeCoder *rc, uint32_t bits)
{
- int priority, i, j;
- int bits_left = f->framebits - opus_rc_tell(&rc->c);
-
- 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--;
- }
- }
- }
+ return ff_opus_rc_dec_log((OpusDecRangeCoder*)rc, bits);
}
-static void celt_decode_tf_changes(CeltFrame *f, OpusDecRangeCoder *rc)
+static inline void opus_rc_enc_cdf(OpusRangeCoder *rc, int val, const uint16_t *cdf)
{
- 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->c);
- 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->c);
- 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]];
- }
+ ff_opus_rc_enc_cdf((OpusEncRangeCoder*)rc, val, cdf);
}
-static void celt_denormalize(CeltFrame *f, CeltBlock *block, float *data)
+static inline uint32_t opus_rc_dec_cdf(OpusRangeCoder *rc, const uint16_t *cdf)
{
- 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;
- }
+ return ff_opus_rc_dec_cdf((OpusDecRangeCoder*)rc, cdf);
}
-static void celt_postfilter_apply_transition(CeltBlock *block, float *data)
+void ff_celt_quant_bands(CeltFrame *f, OpusRangeCoder *rc)
{
- 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;
- }
-}
+ float lowband_scratch[8 * 22];
+ float norm1[2 * 8 * 100];
+ float *norm2 = norm1 + 8 * 100;
-static void celt_postfilter(CeltFrame *f, CeltBlock *block)
-{
- int len = f->blocksize * f->blocks;
- const int filter_len = len - 2 * CELT_OVERLAP;
+ int totalbits = (f->framebits << 3) - f->anticollapse_needed;
- celt_postfilter_apply_transition(block, block->buf + 1024);
+ int update_lowband = 1;
+ int lowband_offset = 0;
- block->pf_period_old = block->pf_period;
- memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
+ int i, j;
- block->pf_period = block->pf_period_new;
- memcpy(block->pf_gains, block->pf_gains_new, sizeof(block->pf_gains));
+ 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;
+ float *norm_loc1, *norm_loc2;
+
+ int consumed = opus_rc_tell_frac(rc);
+ 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 (len > CELT_OVERLAP) {
- celt_postfilter_apply_transition(block, block->buf + 1024 + CELT_OVERLAP);
+ if ((ff_celt_freq_bands[i] - ff_celt_freq_range[i] >= ff_celt_freq_bands[f->start_band] ||
+ i == f->start_band + 1) && (update_lowband || lowband_offset == 0))
+ lowband_offset = i;
- 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);
+ if (i == f->start_band + 1) {
+ /* Special Hybrid Folding (RFC 8251 section 9). Copy the first band into
+ the second to ensure the second band never has to use the LCG. */
+ int count = (ff_celt_freq_range[i] - ff_celt_freq_range[i-1]) << f->size;
- block->pf_period_old = block->pf_period;
- memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
- }
+ memcpy(&norm1[band_offset], &norm1[band_offset - count], count * sizeof(float));
- memmove(block->buf, block->buf + len, (1024 + CELT_OVERLAP / 2) * sizeof(float));
-}
+ if (f->channels == 2)
+ memcpy(&norm2[band_offset], &norm2[band_offset - count], count * sizeof(float));
+ }
-static int parse_postfilter(CeltFrame *f, OpusDecRangeCoder *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->c) + 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];
+ /* 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 (++foldend < i && 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];
}
}
- consumed = opus_rc_tell(&rc->c);
- }
+ 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++)
+ norm1[j] = (norm1[j] + norm2[j]) / 2;
+ }
- return consumed;
-}
+ norm_loc1 = effective_lowband != -1 ? norm1 + (effective_lowband << f->size) : NULL;
+ norm_loc2 = effective_lowband != -1 ? norm2 + (effective_lowband << f->size) : NULL;
-static void process_anticollapse(CeltFrame *f, CeltBlock *block, float *X)
-{
- int i, j, k;
+ if (f->dual_stereo) {
+ cm[0] = f->pvq->quant_band(f->pvq, f, rc, i, X, NULL, band_size, b >> 1,
+ f->blocks, norm_loc1, f->size,
+ norm1 + band_offset, 0, 1.0f,
+ lowband_scratch, cm[0]);
- 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;
- }
+ cm[1] = f->pvq->quant_band(f->pvq, f, rc, i, Y, NULL, band_size, b >> 1,
+ f->blocks, norm_loc2, f->size,
+ norm2 + band_offset, 0, 1.0f,
+ lowband_scratch, cm[1]);
+ } else {
+ cm[0] = f->pvq->quant_band(f->pvq, f, rc, i, X, Y, band_size, b >> 0,
+ f->blocks, norm_loc1, f->size,
+ norm1 + band_offset, 0, 1.0f,
+ lowband_scratch, cm[0] | cm[1]);
+ cm[1] = cm[0];
}
- /* 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);
+ 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, OpusDecRangeCoder *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
- AVTXContext *imdct;
- av_tx_fn imdct_fn;
-
- 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;
- }
+#define NORMC(bits) ((bits) << (f->channels - 1) << f->size >> 2)
- 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->c.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;
+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];
+
+ if (!CONFIG_OPUS_DECODER || !CONFIG_OPUS_ENCODER)
+ encode = CONFIG_OPUS_ENCODER;
+ /* Spread */
+ if (opus_rc_tell(rc) + 4 <= f->framebits) {
+ if (encode)
+ opus_rc_enc_cdf(rc, f->spread, ff_celt_model_spread);
+ else
+ f->spread = opus_rc_dec_cdf(rc, ff_celt_model_spread);
+ } else {
+ f->spread = CELT_SPREAD_NORMAL;
}
- if (!f->output_channels)
- f->output_channels = channels;
+ /* 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]);
- 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));
- }
+ /* 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--;
+ opus_rc_enc_log(rc, is_boost, b_dynalloc);
+ } else {
+ is_boost = opus_rc_dec_log(rc, b_dynalloc);
+ }
- consumed = opus_rc_tell(&rc->c);
+ if (!is_boost)
+ break;
- /* obtain silence flag */
- if (consumed >= f->framebits)
- f->silence = 1;
- else if (consumed == 1)
- f->silence = ff_opus_rc_dec_log(rc, 15);
+ boost[i] += quanta;
+ tbits_8ths -= quanta;
+ b_dynalloc = 1;
+ }
- if (f->silence) {
- consumed = f->framebits;
- rc->c.total_bits += f->framebits - opus_rc_tell(&rc->c);
+ if (boost[i])
+ dynalloc = FFMAX(dynalloc - 1, 2);
}
- /* 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->tx[f->transient ? 0 : f->size];
- imdct_fn = f->tx_fn[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]);
+ /* Allocation trim */
+ if (!encode)
+ f->alloc_trim = 5;
+ if (opus_rc_tell_frac(rc) + (6 << 3) <= tbits_8ths)
+ if (encode)
+ opus_rc_enc_cdf(rc, f->alloc_trim, ff_celt_model_alloc_trim);
+ else
+ f->alloc_trim = 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;
+ }
}
- celt_decode_coarse_energy(f, rc);
- celt_decode_tf_changes (f, rc);
- ff_celt_bitalloc (f, &rc->c, 0);
- celt_decode_fine_energy (f, rc);
- ff_celt_quant_bands (f, &rc->c);
-
- if (f->anticollapse_needed)
- f->anticollapse = ff_opus_rc_get_raw(rc, 1);
-
- celt_decode_final_energy(f, rc);
+ /* 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;
- /* apply anti-collapse processing and denormalization to
- * each coded channel */
- for (i = 0; i < f->channels; i++) {
- CeltBlock *block = &f->block[i];
+ /* 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);
- if (f->anticollapse)
- process_anticollapse(f, block, f->block[i].coeffs);
+ trim_offset[i] = trim * (band << scale) >> 6;
- celt_denormalize(f, block, f->block[i].coeffs);
+ if (ff_celt_freq_range[i] << f->size == 1)
+ trim_offset[i] -= f->channels << 3;
}
- /* 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;
+ /* 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;
+ }
}
- 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];
+ if (total > tbits_8ths)
+ high = center - 1;
+ else
+ low = center + 1;
+ }
+ high = low--;
- /* iMDCT and overlap-add */
- for (j = 0; j < f->blocks; j++) {
- float *dst = block->buf + 1024 + j * f->blocksize;
+ /* 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);
+ }
- imdct_fn(imdct, dst + CELT_OVERLAP / 2, f->block[i].coeffs + j,
- sizeof(float)*f->blocks);
- f->dsp->vector_fmul_window(dst, dst, dst + CELT_OVERLAP / 2,
- ff_celt_window, CELT_OVERLAP / 2);
+ /* 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 (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 (total > tbits_8ths)
+ high = center;
+ else
+ low = center;
}
- if (channels == 1)
- memcpy(f->block[1].energy, f->block[0].energy, sizeof(f->block[0].energy));
+ /* Bisection */
+ done = total = 0;
+ for (i = f->end_band - 1; i >= f->start_band; i--) {
+ bandbits = bits1[i] + (low * bits2[i] >> CELT_ALLOC_STEPS);
- for (i = 0; i < 2; i++ ) {
- CeltBlock *block = &f->block[i];
+ if (bandbits >= threshold[i] || done)
+ done = 1;
+ else
+ bandbits = (bandbits >= f->channels << 3) ?
+ f->channels << 3 : 0;
- 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;
- }
+ bandbits = FFMIN(bandbits, f->caps[i]);
+ f->pulses[i] = bandbits;
+ total += bandbits;
}
- f->seed = rc->c.range;
+ /* Band skipping */
+ for (f->coded_bands = f->end_band; ; f->coded_bands--) {
+ int allocation;
+ j = f->coded_bands - 1;
- return 0;
-}
-
-void ff_celt_flush(CeltFrame *f)
-{
- int i, j;
-
- if (f->flushed)
- return;
+ if (j == skip_startband) {
+ /* all remaining bands are not skipped */
+ tbits_8ths += skip_bit;
+ break;
+ }
- for (i = 0; i < 2; i++) {
- CeltBlock *block = &f->block[i];
+ /* 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;
+ opus_rc_enc_log(rc, do_not_skip, 1);
+ } else {
+ do_not_skip = opus_rc_dec_log(rc, 1);
+ }
- for (j = 0; j < CELT_MAX_BANDS; j++)
- block->prev_energy[0][j] = block->prev_energy[1][j] = CELT_ENERGY_SILENCE;
+ if (do_not_skip)
+ break;
- memset(block->energy, 0, sizeof(block->energy));
- memset(block->buf, 0, sizeof(block->buf));
+ total += 1 << 3;
+ allocation -= 1 << 3;
+ }
- 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));
+ /* 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;
+ }
- /* 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;
+ total += f->pulses[j] = (allocation >= f->channels << 3) ? f->channels << 3 : 0;
}
- 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->tx); i++)
- av_tx_uninit(&frm->tx[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;
+ /* IS start band */
+ if (encode) {
+ if (intensitystereo_bit) {
+ f->intensity_stereo = FFMIN(f->intensity_stereo, f->coded_bands);
+ ff_opus_rc_enc_uint((OpusEncRangeCoder*)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((OpusDecRangeCoder*)rc, f->coded_bands + 1 - f->start_band);
+ }
- 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);
+ /* 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)
+ opus_rc_enc_log(rc, f->dual_stereo, 1);
+ else
+ f->dual_stereo = 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;
}
- frm = av_mallocz(sizeof(*frm));
- if (!frm)
- return AVERROR(ENOMEM);
+ /* 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;
+ }
- frm->avctx = avctx;
- frm->output_channels = output_channels;
- frm->apply_phase_inv = apply_phase_inv;
+ /* 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;
- for (i = 0; i < FF_ARRAY_ELEMS(frm->tx); i++) {
- const float scale = -1.0f/32768;
- if ((ret = av_tx_init(&frm->tx[i], &frm->tx_fn[i], AV_TX_FLOAT_MDCT, 1, 15 << (i + 3), &scale, 0)) < 0)
- goto fail;
+ fineextra <<= f->channels + 2;
+ f->fine_priority[i] = (fineextra >= extrabits - prev_extra);
+ extrabits -= fineextra;
+ }
}
+ f->remaining = extrabits;
- 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;
+ /* 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;
}
-
- ff_opus_dsp_init(&frm->opusdsp);
- ff_celt_flush(frm);
-
- *f = frm;
-
- return 0;
-fail:
- ff_celt_free(&frm);
- return ret;
}
diff --git a/libavcodec/opusdec_celt.c b/libavcodec/opusdec_celt.c
new file mode 100644
index 0000000000..a14764ec18
--- /dev/null
+++ b/libavcodec/opusdec_celt.c
@@ -0,0 +1,587 @@
+/*
+ * Copyright (c) 2012 Andrew D'Addesio
+ * Copyright (c) 2013-2014 Mozilla Corporation
+ * Copyright (c) 2016 Rostislav Pehlivanov <atomnuker at 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 <float.h>
+
+#include "opus_celt.h"
+#include "opusdec_rc.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, OpusDecRangeCoder *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->c) + 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->c);
+ 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, OpusDecRangeCoder *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, OpusDecRangeCoder *rc)
+{
+ int priority, i, j;
+ int bits_left = f->framebits - opus_rc_tell(&rc->c);
+
+ 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, OpusDecRangeCoder *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->c);
+ 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->c);
+ 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, OpusDecRangeCoder *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->c) + 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->c);
+ }
+
+ 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, OpusDecRangeCoder *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
+ AVTXContext *imdct;
+ av_tx_fn imdct_fn;
+
+ 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->c.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->c);
+
+ /* 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->c.total_bits += f->framebits - opus_rc_tell(&rc->c);
+ }
+
+ /* 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->tx[f->transient ? 0 : f->size];
+ imdct_fn = f->tx_fn[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->c, 0);
+ celt_decode_fine_energy (f, rc);
+ ff_celt_quant_bands (f, &rc->c);
+
+ 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_fn(imdct, dst + CELT_OVERLAP / 2, f->block[i].coeffs + j,
+ sizeof(float)*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->c.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->tx); i++)
+ av_tx_uninit(&frm->tx[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->tx); i++) {
+ const float scale = -1.0f/32768;
+ if ((ret = av_tx_init(&frm->tx[i], &frm->tx_fn[i], AV_TX_FLOAT_MDCT, 1, 15 << (i + 3), &scale, 0)) < 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;
+}
--
2.34.1
More information about the ffmpeg-devel
mailing list