]> git.sesse.net Git - ffmpeg/blobdiff - libavcodec/aacenc_ltp.c
aacenc: add support for encoding files using Long Term Prediction
[ffmpeg] / libavcodec / aacenc_ltp.c
diff --git a/libavcodec/aacenc_ltp.c b/libavcodec/aacenc_ltp.c
new file mode 100644 (file)
index 0000000..066e195
--- /dev/null
@@ -0,0 +1,227 @@
+/*
+ * AAC encoder TNS
+ * 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 long term prediction
+ * @author Rostislav Pehlivanov ( atomnuker gmail com )
+ */
+
+#include "aacenc_ltp.h"
+#include "aacenc_quantization.h"
+#include "aacenc_utils.h"
+
+/**
+ * Encode LTP data.
+ */
+void ff_aac_encode_ltp_info(AACEncContext *s, SingleChannelElement *sce,
+                            int common_window)
+{
+    int i;
+    IndividualChannelStream *ics = &sce->ics;
+    if (s->profile != FF_PROFILE_AAC_LTP || !ics->predictor_present)
+        return;
+    if (common_window)
+        put_bits(&s->pb, 1, 0);
+    put_bits(&s->pb, 1, ics->ltp.present);
+    if (!ics->ltp.present)
+        return;
+    put_bits(&s->pb, 11, ics->ltp.lag);
+    put_bits(&s->pb, 3,  ics->ltp.coef_idx);
+    for (i = 0; i < FFMIN(ics->max_sfb, MAX_LTP_LONG_SFB); i++)
+        put_bits(&s->pb, 1, ics->ltp.used[i]);
+}
+
+void ff_aac_ltp_insert_new_frame(AACEncContext *s)
+{
+    int i, ch, tag, chans, cur_channel, start_ch = 0;
+    ChannelElement *cpe;
+    SingleChannelElement *sce;
+    for (i = 0; i < s->chan_map[0]; i++) {
+        cpe = &s->cpe[i];
+        tag      = s->chan_map[i+1];
+        chans    = tag == TYPE_CPE ? 2 : 1;
+        for (ch = 0; ch < chans; ch++) {
+            sce = &cpe->ch[ch];
+            cur_channel = start_ch + ch;
+            /* New sample + overlap */
+            memcpy(&sce->ltp_state[0],    &sce->ltp_state[1024], 1024*sizeof(sce->ltp_state[0]));
+            memcpy(&sce->ltp_state[1024], &s->planar_samples[cur_channel][2048], 1024*sizeof(sce->ltp_state[0]));
+            memcpy(&sce->ltp_state[2048], &sce->ret_buf[0], 1024*sizeof(sce->ltp_state[0]));
+        }
+        start_ch += chans;
+    }
+}
+
+/**
+ * Process LTP parameters
+ * @see Patent WO2006070265A1
+ */
+void ff_aac_update_ltp(AACEncContext *s, SingleChannelElement *sce)
+{
+    int i, j, lag;
+    float corr, s0, s1, max_corr = 0.0f;
+    float *samples = &s->planar_samples[s->cur_channel][1024];
+    float *pred_signal = &sce->ltp_state[0];
+    int samples_num = 2048;
+
+    if (s->profile != FF_PROFILE_AAC_LTP)
+        return;
+
+    /* Calculate lag */
+    for (i = 0; i < samples_num; i++) {
+        s0 = s1 = 0.0f;
+        for (j = 0; j < samples_num; j++) {
+            if (j + 1024 < i)
+                continue;
+            s0 += samples[j]*pred_signal[j-i+1024];
+            s1 += pred_signal[j-i+1024]*pred_signal[j-i+1024];
+        }
+        corr = s1 > 0.0f ? s0/sqrt(s1) : 0.0f;
+        if (corr > max_corr) {
+            max_corr = corr;
+            lag = i;
+        }
+    }
+    lag = av_clip(lag, 0, 2048); /* 11 bits => 2^11 = 2048 */
+
+    if (!lag) {
+        sce->ics.ltp.lag = lag;
+        return;
+    }
+
+    s0 = s1 = 0.0f;
+    for (i = 0; i < lag; i++) {
+        s0 += samples[i];
+        s1 += pred_signal[i-lag+1024];
+    }
+
+    sce->ics.ltp.coef_idx = quant_array_idx(s0/s1, ltp_coef, 8);
+    sce->ics.ltp.coef     = ltp_coef[sce->ics.ltp.coef_idx];
+
+    /* Predict the new samples */
+    if (lag < 1024)
+        samples_num = lag + 1024;
+    for (i = 0; i < samples_num; i++)
+        pred_signal[i+1024] = sce->ics.ltp.coef*pred_signal[i-lag+1024];
+    memset(&pred_signal[samples_num], 0, (2048 - samples_num)*sizeof(float));
+
+    sce->ics.ltp.lag = lag;
+}
+
+void ff_aac_adjust_common_ltp(AACEncContext *s, ChannelElement *cpe)
+{
+    int sfb, count = 0;
+    SingleChannelElement *sce0 = &cpe->ch[0];
+    SingleChannelElement *sce1 = &cpe->ch[1];
+
+    if (!cpe->common_window ||
+        sce0->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE ||
+        sce1->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE)
+        return;
+
+    for (sfb = 0; sfb < FFMIN(sce0->ics.max_sfb, MAX_LTP_LONG_SFB); sfb++) {
+        int sum = sce0->ics.ltp.used[sfb] + sce1->ics.ltp.used[sfb];
+        if (sum != 2) {
+            sce0->ics.ltp.used[sfb] = 0;
+        } else if (sum == 2) {
+            count++;
+        }
+    }
+
+    sce0->ics.ltp.present = !!count;
+    sce0->ics.predictor_present = !!count;
+}
+
+/**
+ * Mark LTP sfb's
+ */
+void ff_aac_search_for_ltp(AACEncContext *s, SingleChannelElement *sce,
+                           int common_window)
+{
+    int w, g, w2, i, start = 0, count = 0;
+    int saved_bits = -(15 + FFMIN(sce->ics.max_sfb, MAX_LTP_LONG_SFB));
+    float *C34 = &s->scoefs[128*0], *PCD = &s->scoefs[128*1];
+    float *PCD34 = &s->scoefs[128*2];
+    const int max_ltp = FFMIN(sce->ics.max_sfb, MAX_LTP_LONG_SFB);
+
+    if (sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE ||
+        !sce->ics.ltp.lag)
+        return;
+
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+        start = 0;
+        for (g = 0;  g < sce->ics.num_swb; g++) {
+            int bits1 = 0, bits2 = 0;
+            float dist1 = 0.0f, dist2 = 0.0f;
+            if (w*16+g > max_ltp) {
+                start += sce->ics.swb_sizes[g];
+                continue;
+            }
+            for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
+                int bits_tmp1, bits_tmp2;
+                FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
+                for (i = 0; i < sce->ics.swb_sizes[g]; i++)
+                    PCD[i] = sce->coeffs[start+(w+w2)*128+i] - sce->lcoeffs[start+(w+w2)*128+i];
+                abs_pow34_v(C34,  &sce->coeffs[start+(w+w2)*128],  sce->ics.swb_sizes[g]);
+                abs_pow34_v(PCD34, PCD, sce->ics.swb_sizes[g]);
+                dist1 += quantize_band_cost(s, &sce->coeffs[start+(w+w2)*128], C34, sce->ics.swb_sizes[g],
+                                            sce->sf_idx[(w+w2)*16+g], sce->band_type[(w+w2)*16+g],
+                                            s->lambda/band->threshold, INFINITY, &bits_tmp1, NULL, 0);
+                dist2 += quantize_band_cost(s, PCD, PCD34, sce->ics.swb_sizes[g],
+                                            sce->sf_idx[(w+w2)*16+g],
+                                            sce->band_type[(w+w2)*16+g],
+                                            s->lambda/band->threshold, INFINITY, &bits_tmp2, NULL, 0);
+                bits1 += bits_tmp1;
+                bits2 += bits_tmp2;
+            }
+            if (dist2 < dist1 && bits2 < bits1) {
+                for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
+                    for (i = 0; i < sce->ics.swb_sizes[g]; i++)
+                        sce->coeffs[start+(w+w2)*128+i] -= sce->lcoeffs[start+(w+w2)*128+i];
+                sce->ics.ltp.used[w*16+g] = 1;
+                saved_bits += bits1 - bits2;
+                count++;
+            }
+            start += sce->ics.swb_sizes[g];
+        }
+    }
+
+    sce->ics.ltp.present = !!count && (saved_bits >= 0);
+    sce->ics.predictor_present = !!sce->ics.ltp.present;
+
+    /* Reset any marked sfbs */
+    if (!sce->ics.ltp.present && !!count) {
+        for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+            start = 0;
+            for (g = 0;  g < sce->ics.num_swb; g++) {
+                if (sce->ics.ltp.used[w*16+g]) {
+                    for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
+                        for (i = 0; i < sce->ics.swb_sizes[g]; i++) {
+                            sce->coeffs[start+(w+w2)*128+i] += sce->lcoeffs[start+(w+w2)*128+i];
+                        }
+                    }
+                }
+                start += sce->ics.swb_sizes[g];
+            }
+        }
+    }
+}