You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

348 lines
12 KiB

/*
* AAC encoder main-type prediction
* Copyright (C) 2015 Rostislav Pehlivanov
*
* 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
* AAC encoder main-type prediction
* @author Rostislav Pehlivanov ( atomnuker gmail com )
*/
#include "aactab.h"
#include "aacenc_pred.h"
#include "aacenc_utils.h"
#include "aacenc_is.h" /* <- Needed for common window distortions */
#include "aacenc_quantization.h"
#define RESTORE_PRED(sce, sfb) \
if (sce->ics.prediction_used[sfb]) {\
sce->ics.prediction_used[sfb] = 0;\
sce->band_type[sfb] = sce->band_alt[sfb];\
}
static inline float flt16_round(float pf)
{
union av_intfloat32 tmp;
tmp.f = pf;
tmp.i = (tmp.i + 0x00008000U) & 0xFFFF0000U;
return tmp.f;
}
static inline float flt16_even(float pf)
{
union av_intfloat32 tmp;
tmp.f = pf;
tmp.i = (tmp.i + 0x00007FFFU + (tmp.i & 0x00010000U >> 16)) & 0xFFFF0000U;
return tmp.f;
}
static inline float flt16_trunc(float pf)
{
union av_intfloat32 pun;
pun.f = pf;
pun.i &= 0xFFFF0000U;
return pun.f;
}
static inline void predict(PredictorState *ps, float *coef, float *rcoef, int set)
{
float k2;
const float a = 0.953125; // 61.0 / 64
const float alpha = 0.90625; // 29.0 / 32
const float k1 = ps->k1;
const float r0 = ps->r0, r1 = ps->r1;
const float cor0 = ps->cor0, cor1 = ps->cor1;
const float var0 = ps->var0, var1 = ps->var1;
const float e0 = *coef - ps->x_est;
const float e1 = e0 - k1 * r0;
if (set)
*coef = e0;
ps->cor1 = flt16_trunc(alpha * cor1 + r1 * e1);
ps->var1 = flt16_trunc(alpha * var1 + 0.5f * (r1 * r1 + e1 * e1));
ps->cor0 = flt16_trunc(alpha * cor0 + r0 * e0);
ps->var0 = flt16_trunc(alpha * var0 + 0.5f * (r0 * r0 + e0 * e0));
ps->r1 = flt16_trunc(a * (r0 - k1 * e0));
ps->r0 = flt16_trunc(a * e0);
/* Prediction for next frame */
ps->k1 = ps->var0 > 1 ? ps->cor0 * flt16_even(a / ps->var0) : 0;
k2 = ps->var1 > 1 ? ps->cor1 * flt16_even(a / ps->var1) : 0;
*rcoef = ps->x_est = flt16_round(ps->k1*ps->r0 + k2*ps->r1);
}
static inline void reset_predict_state(PredictorState *ps)
{
ps->r0 = 0.0f;
ps->r1 = 0.0f;
ps->k1 = 0.0f;
ps->cor0 = 0.0f;
ps->cor1 = 0.0f;
ps->var0 = 1.0f;
ps->var1 = 1.0f;
ps->x_est = 0.0f;
}
static inline void reset_all_predictors(PredictorState *ps)
{
int i;
for (i = 0; i < MAX_PREDICTORS; i++)
reset_predict_state(&ps[i]);
}
static inline void reset_predictor_group(SingleChannelElement *sce, int group_num)
{
int i;
PredictorState *ps = sce->predictor_state;
for (i = group_num - 1; i < MAX_PREDICTORS; i += 30)
reset_predict_state(&ps[i]);
}
void ff_aac_apply_main_pred(AACEncContext *s, SingleChannelElement *sce)
{
int sfb, k;
const int pmax = FFMIN(sce->ics.max_sfb, ff_aac_pred_sfb_max[s->samplerate_index]);
if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
for (sfb = 0; sfb < pmax; sfb++) {
for (k = sce->ics.swb_offset[sfb]; k < sce->ics.swb_offset[sfb + 1]; k++) {
predict(&sce->predictor_state[k], &sce->coeffs[k], &sce->prcoeffs[k],
sce->ics.predictor_present && sce->ics.prediction_used[sfb]);
}
}
if (sce->ics.predictor_reset_group) {
reset_predictor_group(sce, sce->ics.predictor_reset_group);
}
} else {
reset_all_predictors(sce->predictor_state);
}
}
/* If inc = 0 you can check if this returns 0 to see if you can reset freely */
static inline int update_counters(IndividualChannelStream *ics, int inc)
{
int i;
for (i = 1; i < 31; i++) {
ics->predictor_reset_count[i] += inc;
if (ics->predictor_reset_count[i] > PRED_RESET_FRAME_MIN)
return i; /* Reset this immediately */
}
return 0;
}
void ff_aac_adjust_common_pred(AACEncContext *s, ChannelElement *cpe)
{
int start, w, w2, g, i, count = 0;
SingleChannelElement *sce0 = &cpe->ch[0];
SingleChannelElement *sce1 = &cpe->ch[1];
const int pmax0 = FFMIN(sce0->ics.max_sfb, ff_aac_pred_sfb_max[s->samplerate_index]);
const int pmax1 = FFMIN(sce1->ics.max_sfb, ff_aac_pred_sfb_max[s->samplerate_index]);
const int pmax = FFMIN(pmax0, pmax1);
if (!cpe->common_window ||
sce0->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE ||
sce1->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE)
return;
for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
start = 0;
for (g = 0; g < sce0->ics.num_swb; g++) {
int sfb = w*16+g;
int sum = sce0->ics.prediction_used[sfb] + sce1->ics.prediction_used[sfb];
float ener0 = 0.0f, ener1 = 0.0f, ener01 = 0.0f;
struct AACISError ph_err1, ph_err2, *erf;
if (sfb < PRED_SFB_START || sfb > pmax || sum != 2) {
RESTORE_PRED(sce0, sfb);
RESTORE_PRED(sce1, sfb);
start += sce0->ics.swb_sizes[g];
continue;
}
for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
float coef0 = sce0->pcoeffs[start+(w+w2)*128+i];
float coef1 = sce1->pcoeffs[start+(w+w2)*128+i];
ener0 += coef0*coef0;
ener1 += coef1*coef1;
ener01 += (coef0 + coef1)*(coef0 + coef1);
}
}
ph_err1 = ff_aac_is_encoding_err(s, cpe, start, w, g,
ener0, ener1, ener01, 1, -1);
ph_err2 = ff_aac_is_encoding_err(s, cpe, start, w, g,
ener0, ener1, ener01, 1, +1);
erf = ph_err1.error < ph_err2.error ? &ph_err1 : &ph_err2;
if (erf->pass) {
sce0->ics.prediction_used[sfb] = 1;
sce1->ics.prediction_used[sfb] = 1;
count++;
} else {
RESTORE_PRED(sce0, sfb);
RESTORE_PRED(sce1, sfb);
}
start += sce0->ics.swb_sizes[g];
}
}
sce1->ics.predictor_present = sce0->ics.predictor_present = !!count;
}
static void update_pred_resets(SingleChannelElement *sce)
{
int i, max_group_id_c, max_frame = 0;
float avg_frame = 0.0f;
IndividualChannelStream *ics = &sce->ics;
/* Update the counters and immediately update any frame behind schedule */
if ((ics->predictor_reset_group = update_counters(&sce->ics, 1)))
return;
for (i = 1; i < 31; i++) {
/* Count-based */
if (ics->predictor_reset_count[i] > max_frame) {
max_group_id_c = i;
max_frame = ics->predictor_reset_count[i];
}
avg_frame = (ics->predictor_reset_count[i] + avg_frame)/2;
}
if (max_frame > PRED_RESET_MIN) {
ics->predictor_reset_group = max_group_id_c;
} else {
ics->predictor_reset_group = 0;
}
}
void ff_aac_search_for_pred(AACEncContext *s, SingleChannelElement *sce)
{
int sfb, i, count = 0, cost_coeffs = 0, cost_pred = 0;
const int pmax = FFMIN(sce->ics.max_sfb, ff_aac_pred_sfb_max[s->samplerate_index]);
float *O34 = &s->scoefs[128*0], *P34 = &s->scoefs[128*1];
float *SENT = &s->scoefs[128*2], *S34 = &s->scoefs[128*3];
float *QERR = &s->scoefs[128*4];
if (sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE) {
sce->ics.predictor_present = 0;
return;
}
if (!sce->ics.predictor_initialized) {
reset_all_predictors(sce->predictor_state);
sce->ics.predictor_initialized = 1;
memcpy(sce->prcoeffs, sce->coeffs, 1024*sizeof(float));
for (i = 1; i < 31; i++)
sce->ics.predictor_reset_count[i] = i;
}
update_pred_resets(sce);
memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
for (sfb = PRED_SFB_START; sfb < pmax; sfb++) {
int cost1, cost2, cb_p;
float dist1, dist2, dist_spec_err = 0.0f;
const int cb_n = sce->zeroes[sfb] ? 0 : sce->band_type[sfb];
const int cb_min = sce->zeroes[sfb] ? 0 : 1;
const int cb_max = sce->zeroes[sfb] ? 0 : RESERVED_BT;
const int start_coef = sce->ics.swb_offset[sfb];
const int num_coeffs = sce->ics.swb_offset[sfb + 1] - start_coef;
const FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[sfb];
if (start_coef + num_coeffs > MAX_PREDICTORS ||
(s->cur_channel && sce->band_type[sfb] >= INTENSITY_BT2) ||
sce->band_type[sfb] == NOISE_BT)
continue;
/* Normal coefficients */
abs_pow34_v(O34, &sce->coeffs[start_coef], num_coeffs);
dist1 = quantize_and_encode_band_cost(s, NULL, &sce->coeffs[start_coef], NULL,
O34, num_coeffs, sce->sf_idx[sfb],
AAC encoder: Extensive improvements This finalizes merging of the work in the patches in ticket #2686. Improvements to twoloop and RC logic are extensive. The non-exhaustive list of twoloop improvments includes: - Tweaks to distortion limits on the RD optimization phase of twoloop - Deeper search in twoloop - PNS information marking to let twoloop decide when to use it (turned out having the decision made separately wasn't working) - Tonal band detection and priorization - Better band energy conservation rules - Strict hole avoidance For rate control: - Use psymodel's bit allocation to allow proper use of the bit reservoir. Don't work against the bit reservoir by moving lambda in the opposite direction when psymodel decides to allocate more/less bits to a frame. - Retry the encode if the effective rate lies outside a reasonable margin of psymodel's allocation or the selected ABR. - Log average lambda at the end. Useful info for everyone, but especially for tuning of the various encoder constants that relate to lambda feedback. Psy: - Do not apply lowpass with a FIR filter, instead just let the coder zero bands above the cutoff. The FIR filter induces group delay, and while zeroing bands causes ripple, it's lost in the quantization noise. - Experimental VBR bit allocation code - Tweak automatic lowpass filter threshold to maximize audio bandwidth at all bitrates while still providing acceptable, stable quality. I/S: - Phase decision fixes. Unrelated to #2686, but the bugs only surfaced when the merge was finalized. Measure I/S band energy accounting for phase, and prevent I/S and M/S from being applied both. PNS: - Avoid marking short bands with PNS when they're part of a window group in which there's a large variation of energy from one window to the next. PNS can't preserve those and the effect is extremely noticeable. M/S: - Implement BMLD protection similar to the specified in ISO-IEC/13818:7-2003, Appendix C Section 6.1. Since M/S decision doesn't conform to section 6.1, a different method had to be implemented, but should provide equivalent protection. - Move the decision logic closer to the method specified in ISO-IEC/13818:7-2003, Appendix C Section 6.1. Specifically, make sure M/S needs less bits than dual stereo. - Don't apply M/S in bands that are using I/S Now, this of course needed adjustments in the compare targets and fuzz factors of the AAC encoder's fate tests, but if wondering why the targets go up (more distortion), consider the previous coder was using too many bits on LF content (far more than required by psy), and thus those signals will now be more distorted, not less. The extra distortion isn't audible though, I carried extensive ABX testing to make sure. A very similar patch was also extensively tested by Kamendo2 in the context of #2686.
9 years ago
cb_n, s->lambda / band->threshold, INFINITY, &cost1, NULL, 0);
cost_coeffs += cost1;
/* Encoded coefficients - needed for #bits, band type and quant. error */
for (i = 0; i < num_coeffs; i++)
SENT[i] = sce->coeffs[start_coef + i] - sce->prcoeffs[start_coef + i];
abs_pow34_v(S34, SENT, num_coeffs);
if (cb_n < RESERVED_BT)
cb_p = av_clip(find_min_book(find_max_val(1, num_coeffs, S34), sce->sf_idx[sfb]), cb_min, cb_max);
else
cb_p = cb_n;
quantize_and_encode_band_cost(s, NULL, SENT, QERR, S34, num_coeffs,
sce->sf_idx[sfb], cb_p, s->lambda / band->threshold, INFINITY,
AAC encoder: Extensive improvements This finalizes merging of the work in the patches in ticket #2686. Improvements to twoloop and RC logic are extensive. The non-exhaustive list of twoloop improvments includes: - Tweaks to distortion limits on the RD optimization phase of twoloop - Deeper search in twoloop - PNS information marking to let twoloop decide when to use it (turned out having the decision made separately wasn't working) - Tonal band detection and priorization - Better band energy conservation rules - Strict hole avoidance For rate control: - Use psymodel's bit allocation to allow proper use of the bit reservoir. Don't work against the bit reservoir by moving lambda in the opposite direction when psymodel decides to allocate more/less bits to a frame. - Retry the encode if the effective rate lies outside a reasonable margin of psymodel's allocation or the selected ABR. - Log average lambda at the end. Useful info for everyone, but especially for tuning of the various encoder constants that relate to lambda feedback. Psy: - Do not apply lowpass with a FIR filter, instead just let the coder zero bands above the cutoff. The FIR filter induces group delay, and while zeroing bands causes ripple, it's lost in the quantization noise. - Experimental VBR bit allocation code - Tweak automatic lowpass filter threshold to maximize audio bandwidth at all bitrates while still providing acceptable, stable quality. I/S: - Phase decision fixes. Unrelated to #2686, but the bugs only surfaced when the merge was finalized. Measure I/S band energy accounting for phase, and prevent I/S and M/S from being applied both. PNS: - Avoid marking short bands with PNS when they're part of a window group in which there's a large variation of energy from one window to the next. PNS can't preserve those and the effect is extremely noticeable. M/S: - Implement BMLD protection similar to the specified in ISO-IEC/13818:7-2003, Appendix C Section 6.1. Since M/S decision doesn't conform to section 6.1, a different method had to be implemented, but should provide equivalent protection. - Move the decision logic closer to the method specified in ISO-IEC/13818:7-2003, Appendix C Section 6.1. Specifically, make sure M/S needs less bits than dual stereo. - Don't apply M/S in bands that are using I/S Now, this of course needed adjustments in the compare targets and fuzz factors of the AAC encoder's fate tests, but if wondering why the targets go up (more distortion), consider the previous coder was using too many bits on LF content (far more than required by psy), and thus those signals will now be more distorted, not less. The extra distortion isn't audible though, I carried extensive ABX testing to make sure. A very similar patch was also extensively tested by Kamendo2 in the context of #2686.
9 years ago
&cost2, NULL, 0);
/* Reconstructed coefficients - needed for distortion measurements */
for (i = 0; i < num_coeffs; i++)
sce->prcoeffs[start_coef + i] += QERR[i] != 0.0f ? (sce->prcoeffs[start_coef + i] - QERR[i]) : 0.0f;
abs_pow34_v(P34, &sce->prcoeffs[start_coef], num_coeffs);
if (cb_n < RESERVED_BT)
cb_p = av_clip(find_min_book(find_max_val(1, num_coeffs, P34), sce->sf_idx[sfb]), cb_min, cb_max);
else
cb_p = cb_n;
dist2 = quantize_and_encode_band_cost(s, NULL, &sce->prcoeffs[start_coef], NULL,
P34, num_coeffs, sce->sf_idx[sfb],
AAC encoder: Extensive improvements This finalizes merging of the work in the patches in ticket #2686. Improvements to twoloop and RC logic are extensive. The non-exhaustive list of twoloop improvments includes: - Tweaks to distortion limits on the RD optimization phase of twoloop - Deeper search in twoloop - PNS information marking to let twoloop decide when to use it (turned out having the decision made separately wasn't working) - Tonal band detection and priorization - Better band energy conservation rules - Strict hole avoidance For rate control: - Use psymodel's bit allocation to allow proper use of the bit reservoir. Don't work against the bit reservoir by moving lambda in the opposite direction when psymodel decides to allocate more/less bits to a frame. - Retry the encode if the effective rate lies outside a reasonable margin of psymodel's allocation or the selected ABR. - Log average lambda at the end. Useful info for everyone, but especially for tuning of the various encoder constants that relate to lambda feedback. Psy: - Do not apply lowpass with a FIR filter, instead just let the coder zero bands above the cutoff. The FIR filter induces group delay, and while zeroing bands causes ripple, it's lost in the quantization noise. - Experimental VBR bit allocation code - Tweak automatic lowpass filter threshold to maximize audio bandwidth at all bitrates while still providing acceptable, stable quality. I/S: - Phase decision fixes. Unrelated to #2686, but the bugs only surfaced when the merge was finalized. Measure I/S band energy accounting for phase, and prevent I/S and M/S from being applied both. PNS: - Avoid marking short bands with PNS when they're part of a window group in which there's a large variation of energy from one window to the next. PNS can't preserve those and the effect is extremely noticeable. M/S: - Implement BMLD protection similar to the specified in ISO-IEC/13818:7-2003, Appendix C Section 6.1. Since M/S decision doesn't conform to section 6.1, a different method had to be implemented, but should provide equivalent protection. - Move the decision logic closer to the method specified in ISO-IEC/13818:7-2003, Appendix C Section 6.1. Specifically, make sure M/S needs less bits than dual stereo. - Don't apply M/S in bands that are using I/S Now, this of course needed adjustments in the compare targets and fuzz factors of the AAC encoder's fate tests, but if wondering why the targets go up (more distortion), consider the previous coder was using too many bits on LF content (far more than required by psy), and thus those signals will now be more distorted, not less. The extra distortion isn't audible though, I carried extensive ABX testing to make sure. A very similar patch was also extensively tested by Kamendo2 in the context of #2686.
9 years ago
cb_p, s->lambda / band->threshold, INFINITY, NULL, NULL, 0);
for (i = 0; i < num_coeffs; i++)
dist_spec_err += (O34[i] - P34[i])*(O34[i] - P34[i]);
dist_spec_err *= s->lambda / band->threshold;
dist2 += dist_spec_err;
if (dist2 <= dist1 && cb_p <= cb_n) {
cost_pred += cost2;
sce->ics.prediction_used[sfb] = 1;
sce->band_alt[sfb] = cb_n;
sce->band_type[sfb] = cb_p;
count++;
} else {
cost_pred += cost1;
sce->band_alt[sfb] = cb_p;
}
}
if (count && cost_coeffs < cost_pred) {
count = 0;
for (sfb = PRED_SFB_START; sfb < pmax; sfb++)
RESTORE_PRED(sce, sfb);
memset(&sce->ics.prediction_used, 0, sizeof(sce->ics.prediction_used));
}
sce->ics.predictor_present = !!count;
}
/**
* Encoder predictors data.
*/
void ff_aac_encode_main_pred(AACEncContext *s, SingleChannelElement *sce)
{
int sfb;
IndividualChannelStream *ics = &sce->ics;
const int pmax = FFMIN(ics->max_sfb, ff_aac_pred_sfb_max[s->samplerate_index]);
if (s->profile != FF_PROFILE_AAC_MAIN ||
!ics->predictor_present)
return;
put_bits(&s->pb, 1, !!ics->predictor_reset_group);
if (ics->predictor_reset_group)
put_bits(&s->pb, 5, ics->predictor_reset_group);
for (sfb = 0; sfb < pmax; sfb++)
put_bits(&s->pb, 1, ics->prediction_used[sfb]);
}