AAC encoder: memoize quantize_band_cost
[ffmpeg.git] / libavcodec / aacenc.c
index 444ca0e..9e96cbc 100644 (file)
@@ -71,6 +71,16 @@ static void put_audio_specific_config(AVCodecContext *avctx)
     flush_put_bits(&pb);
 }
 
+void ff_quantize_band_cost_cache_init(struct AACEncContext *s)
+{
+    int sf, g;
+    for (sf = 0; sf < 256; sf++) {
+        for (g = 0; g < 128; g++) {
+            s->quantize_band_cost_cache[sf][g].bits = -1;
+        }
+    }
+}
+
 #define WINDOW_FUNC(type) \
 static void apply_ ##type ##_window(AVFloatDSPContext *fdsp, \
                                     SingleChannelElement *sce, \
@@ -196,37 +206,6 @@ static void adjust_frame_information(ChannelElement *cpe, int chans)
 {
     int i, w, w2, g, ch;
     int maxsfb, cmaxsfb;
-    IndividualChannelStream *ics;
-
-    if (cpe->common_window) {
-        ics = &cpe->ch[0].ics;
-        for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
-            for (w2 =  0; w2 < ics->group_len[w]; w2++) {
-                int start = (w+w2) * 128;
-                for (g = 0; g < ics->num_swb; g++) {
-                    //apply Intensity stereo coeffs transformation
-                    if (cpe->is_mask[w*16 + g]) {
-                        int p = -1 + 2 * (cpe->ch[1].band_type[w*16+g] - 14);
-                        float scale = cpe->ch[0].is_ener[w*16+g];
-                        for (i = 0; i < ics->swb_sizes[g]; i++) {
-                            cpe->ch[0].coeffs[start+i] = (cpe->ch[0].coeffs[start+i] + p*cpe->ch[1].coeffs[start+i]) * scale;
-                            cpe->ch[1].coeffs[start+i] = 0.0f;
-                        }
-                    } else if (cpe->ms_mask[w*16 + g] &&
-                               cpe->ch[0].band_type[w*16 + g] < NOISE_BT &&
-                               cpe->ch[1].band_type[w*16 + g] < NOISE_BT) {
-                        for (i = 0; i < ics->swb_sizes[g]; i++) {
-                            float L = (cpe->ch[0].coeffs[start+i] + cpe->ch[1].coeffs[start+i]) * 0.5f;
-                            float R = L - cpe->ch[1].coeffs[start+i];
-                            cpe->ch[0].coeffs[start+i] = L;
-                            cpe->ch[1].coeffs[start+i] = R;
-                        }
-                    }
-                    start += ics->swb_sizes[g];
-                }
-            }
-        }
-    }
 
     for (ch = 0; ch < chans; ch++) {
         IndividualChannelStream *ics = &cpe->ch[ch].ics;
@@ -273,6 +252,61 @@ static void adjust_frame_information(ChannelElement *cpe, int chans)
     }
 }
 
+static void apply_intensity_stereo(ChannelElement *cpe)
+{
+    int w, w2, g, i;
+    IndividualChannelStream *ics = &cpe->ch[0].ics;
+    if (!cpe->common_window)
+        return;
+    for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
+        for (w2 =  0; w2 < ics->group_len[w]; w2++) {
+            int start = (w+w2) * 128;
+            for (g = 0; g < ics->num_swb; g++) {
+                int p  = -1 + 2 * (cpe->ch[1].band_type[w*16+g] - 14);
+                float scale = cpe->ch[0].is_ener[w*16+g];
+                if (!cpe->is_mask[w*16 + g]) {
+                    start += ics->swb_sizes[g];
+                    continue;
+                }
+                if (cpe->ms_mask[w*16 + g])
+                    p *= -1;
+                for (i = 0; i < ics->swb_sizes[g]; i++) {
+                    float sum = (cpe->ch[0].coeffs[start+i] + p*cpe->ch[1].coeffs[start+i])*scale;
+                    cpe->ch[0].coeffs[start+i] = sum;
+                    cpe->ch[1].coeffs[start+i] = 0.0f;
+                }
+                start += ics->swb_sizes[g];
+            }
+        }
+    }
+}
+
+static void apply_mid_side_stereo(ChannelElement *cpe)
+{
+    int w, w2, g, i;
+    IndividualChannelStream *ics = &cpe->ch[0].ics;
+    if (!cpe->common_window)
+        return;
+    for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
+        for (w2 =  0; w2 < ics->group_len[w]; w2++) {
+            int start = (w+w2) * 128;
+            for (g = 0; g < ics->num_swb; g++) {
+                if (!cpe->ms_mask[w*16 + g] && !cpe->is_mask[w*16 + g]) {
+                    start += ics->swb_sizes[g];
+                    continue;
+                }
+                for (i = 0; i < ics->swb_sizes[g]; i++) {
+                    float L = (cpe->ch[0].coeffs[start+i] + cpe->ch[1].coeffs[start+i]) * 0.5f;
+                    float R = L - cpe->ch[1].coeffs[start+i];
+                    cpe->ch[0].coeffs[start+i] = L;
+                    cpe->ch[1].coeffs[start+i] = R;
+                }
+                start += ics->swb_sizes[g];
+            }
+        }
+    }
+}
+
 /**
  * Encode scalefactor band coding type.
  */
@@ -280,6 +314,9 @@ static void encode_band_info(AACEncContext *s, SingleChannelElement *sce)
 {
     int w;
 
+    if (s->coder->set_special_band_scalefactors)
+        s->coder->set_special_band_scalefactors(s, sce);
+
     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
         s->coder->encode_window_bands_info(s, sce, w, sce->ics.group_len[w], s->lambda);
 }
@@ -464,7 +501,8 @@ static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
     float **samples = s->planar_samples, *samples2, *la, *overlap;
     ChannelElement *cpe;
     SingleChannelElement *sce;
-    int i, ch, w, g, chans, tag, start_ch, ret;
+    int i, its, ch, w, chans, tag, start_ch, ret, frame_bits;
+    int target_bits, rate_bits, too_many_bits, too_few_bits;
     int ms_mode = 0, is_mode = 0, tns_mode = 0, pred_mode = 0;
     int chan_el_counter[4];
     FFPsyWindowInfo windows[AAC_MAX_CHANNELS];
@@ -556,14 +594,14 @@ static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
     }
     if ((ret = ff_alloc_packet2(avctx, avpkt, 8192 * s->channels, 0)) < 0)
         return ret;
+    frame_bits = its = 0;
     do {
-        int frame_bits;
-
         init_put_bits(&s->pb, avpkt->data, avpkt->size);
 
         if ((avctx->frame_number & 0xFF)==1 && !(avctx->flags & AV_CODEC_FLAG_BITEXACT))
             put_bitstream_info(s, LIBAVCODEC_IDENT);
         start_ch = 0;
+        target_bits = 0;
         memset(chan_el_counter, 0, sizeof(chan_el_counter));
         for (i = 0; i < s->chan_map[0]; i++) {
             FFPsyWindowInfo* wi = windows + start_ch;
@@ -586,9 +624,20 @@ static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
                     if (sce->band_type[w] > RESERVED_BT)
                         sce->band_type[w] = 0;
             }
+            s->psy.bitres.alloc = -1;
+            s->psy.bitres.bits = avctx->frame_bits / s->channels;
             s->psy.model->analyze(&s->psy, start_ch, coeffs, wi);
+            if (s->psy.bitres.alloc > 0) {
+                /* Lambda unused here on purpose, we need to take psy's unscaled allocation */
+                target_bits += s->psy.bitres.alloc
+                    * (s->lambda / (avctx->global_quality ? avctx->global_quality : 120));
+                s->psy.bitres.alloc /= chans;
+            }
+            s->cur_type = tag;
             for (ch = 0; ch < chans; ch++) {
                 s->cur_channel = start_ch + ch;
+                if (s->options.pns && s->coder->mark_pns)
+                    s->coder->mark_pns(s, avctx, &cpe->ch[ch]);
                 s->coder->search_for_quantizers(avctx, s, &cpe->ch[ch], s->lambda);
             }
             if (chans > 1
@@ -603,7 +652,7 @@ static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
                     }
                 }
             }
-            for (ch = 0; ch < chans; ch++) {
+            for (ch = 0; ch < chans; ch++) { /* TNS and PNS */
                 sce = &cpe->ch[ch];
                 s->cur_channel = start_ch + ch;
                 if (s->options.pns && s->coder->search_for_pns)
@@ -616,40 +665,38 @@ static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
                     tns_mode = 1;
             }
             s->cur_channel = start_ch;
-            if (s->options.stereo_mode && cpe->common_window) {
-                if (s->options.stereo_mode > 0) {
-                    IndividualChannelStream *ics = &cpe->ch[0].ics;
-                    for (w = 0; w < ics->num_windows; w += ics->group_len[w])
-                        for (g = 0;  g < ics->num_swb; g++)
-                            cpe->ms_mask[w*16+g] = 1;
-                } else if (s->coder->search_for_ms) {
-                    s->coder->search_for_ms(s, cpe);
-                }
-            }
-            if (s->options.intensity_stereo && s->coder->search_for_is) {
-                s->coder->search_for_is(s, avctx, cpe);
+            if (s->options.intensity_stereo) { /* Intensity Stereo */
+                if (s->coder->search_for_is)
+                    s->coder->search_for_is(s, avctx, cpe);
                 if (cpe->is_mode) is_mode = 1;
+                apply_intensity_stereo(cpe);
             }
-            if (s->coder->set_special_band_scalefactors)
-                for (ch = 0; ch < chans; ch++)
-                    s->coder->set_special_band_scalefactors(s, &cpe->ch[ch]);
-            adjust_frame_information(cpe, chans);
-            for (ch = 0; ch < chans; ch++) {
-                sce = &cpe->ch[ch];
-                s->cur_channel = start_ch + ch;
-                if (s->options.pred && s->coder->search_for_pred)
-                    s->coder->search_for_pred(s, sce);
-                if (cpe->ch[ch].ics.predictor_present) pred_mode = 1;
+            if (s->options.pred) { /* Prediction */
+                for (ch = 0; ch < chans; ch++) {
+                    sce = &cpe->ch[ch];
+                    s->cur_channel = start_ch + ch;
+                    if (s->options.pred && s->coder->search_for_pred)
+                        s->coder->search_for_pred(s, sce);
+                    if (cpe->ch[ch].ics.predictor_present) pred_mode = 1;
+                }
+                if (s->coder->adjust_common_prediction)
+                    s->coder->adjust_common_prediction(s, cpe);
+                for (ch = 0; ch < chans; ch++) {
+                    sce = &cpe->ch[ch];
+                    s->cur_channel = start_ch + ch;
+                    if (s->options.pred && s->coder->apply_main_pred)
+                        s->coder->apply_main_pred(s, sce);
+                }
+                s->cur_channel = start_ch;
             }
-            if (s->options.pred && s->coder->adjust_common_prediction)
-                s->coder->adjust_common_prediction(s, cpe);
-            for (ch = 0; ch < chans; ch++) {
-                sce = &cpe->ch[ch];
-                s->cur_channel = start_ch + ch;
-                if (s->options.pred && s->coder->apply_main_pred)
-                    s->coder->apply_main_pred(s, sce);
+            if (s->options.stereo_mode) { /* Mid/Side stereo */
+                if (s->options.stereo_mode == -1 && s->coder->search_for_ms)
+                    s->coder->search_for_ms(s, cpe);
+                else if (cpe->common_window)
+                    memset(cpe->ms_mask, 1, sizeof(cpe->ms_mask));
+                apply_mid_side_stereo(cpe);
             }
-            s->cur_channel = start_ch;
+            adjust_frame_information(cpe, chans);
             if (chans == 2) {
                 put_bits(&s->pb, 1, cpe->common_window);
                 if (cpe->common_window) {
@@ -667,35 +714,72 @@ static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
             start_ch += chans;
         }
 
-        frame_bits = put_bits_count(&s->pb);
-        if (frame_bits <= 6144 * s->channels - 3) {
-            s->psy.bitres.bits = frame_bits / s->channels;
+        if (avctx->flags & CODEC_FLAG_QSCALE) {
+            /* When using a constant Q-scale, don't mess with lambda */
             break;
         }
-        if (is_mode || ms_mode || tns_mode || pred_mode) {
-            for (i = 0; i < s->chan_map[0]; i++) {
-                // Must restore coeffs
-                chans = tag == TYPE_CPE ? 2 : 1;
-                cpe = &s->cpe[i];
-                for (ch = 0; ch < chans; ch++)
-                    memcpy(cpe->ch[ch].coeffs, cpe->ch[ch].pcoeffs, sizeof(cpe->ch[ch].coeffs));
-            }
-        }
 
-        s->lambda *= avctx->bit_rate * 1024.0f / avctx->sample_rate / frame_bits;
+        /* rate control stuff
+         * allow between the nominal bitrate, and what psy's bit reservoir says to target
+         * but drift towards the nominal bitrate always
+         */
+        frame_bits = put_bits_count(&s->pb);
+        rate_bits = avctx->bit_rate * 1024 / avctx->sample_rate;
+        rate_bits = FFMIN(rate_bits, 6144 * s->channels - 3);
+        too_many_bits = FFMAX(target_bits, rate_bits);
+        too_many_bits = FFMIN(too_many_bits, 6144 * s->channels - 3);
+        too_few_bits = FFMIN(FFMAX(rate_bits - rate_bits/4, target_bits), too_many_bits);
+
+        /* When using ABR, be strict (but only for increasing) */
+        too_few_bits = too_few_bits - too_few_bits/8;
+        too_many_bits = too_many_bits + too_many_bits/2;
+
+        if (   its == 0 /* for steady-state Q-scale tracking */
+            || (its < 5 && (frame_bits < too_few_bits || frame_bits > too_many_bits))
+            || frame_bits >= 6144 * s->channels - 3  )
+        {
+            float ratio = ((float)rate_bits) / frame_bits;
+
+            if (frame_bits >= too_few_bits && frame_bits <= too_many_bits) {
+                /*
+                 * This path is for steady-state Q-scale tracking
+                 * When frame bits fall within the stable range, we still need to adjust
+                 * lambda to maintain it like so in a stable fashion (large jumps in lambda
+                 * create artifacts and should be avoided), but slowly
+                 */
+                ratio = sqrtf(sqrtf(ratio));
+                ratio = av_clipf(ratio, 0.9f, 1.1f);
+            } else {
+                /* Not so fast though */
+                ratio = sqrtf(ratio);
+            }
+            s->lambda = FFMIN(s->lambda * ratio, 65536.f);
 
+            /* Keep iterating if we must reduce and lambda is in the sky */
+            if ((s->lambda < 300.f || ratio > 0.9f) && (s->lambda > 10.f || ratio < 1.1f)) {
+                break;
+            } else {
+                if (is_mode || ms_mode || tns_mode || pred_mode) {
+                    for (i = 0; i < s->chan_map[0]; i++) {
+                        // Must restore coeffs
+                        chans = tag == TYPE_CPE ? 2 : 1;
+                        cpe = &s->cpe[i];
+                        for (ch = 0; ch < chans; ch++)
+                            memcpy(cpe->ch[ch].coeffs, cpe->ch[ch].pcoeffs, sizeof(cpe->ch[ch].coeffs));
+                    }
+                }
+                its++;
+            }
+        } else {
+            break;
+        }
     } while (1);
 
     put_bits(&s->pb, 3, TYPE_END);
     flush_put_bits(&s->pb);
     avctx->frame_bits = put_bits_count(&s->pb);
-
-    // rate control stuff
-    if (!(avctx->flags & AV_CODEC_FLAG_QSCALE)) {
-        float ratio = avctx->bit_rate * 1024.0f / avctx->sample_rate / avctx->frame_bits;
-        s->lambda *= ratio;
-        s->lambda = FFMIN(s->lambda, 65536.f);
-    }
+    s->lambda_sum += s->lambda;
+    s->lambda_count++;
 
     if (!frame)
         s->last_frame++;
@@ -712,6 +796,8 @@ static av_cold int aac_encode_end(AVCodecContext *avctx)
 {
     AACEncContext *s = avctx->priv_data;
 
+    av_log(avctx, AV_LOG_INFO, "Qavg: %.3f\n", s->lambda_sum / s->lambda_count);
+
     ff_mdct_end(&s->mdct1024);
     ff_mdct_end(&s->mdct128);
     ff_psy_end(&s->psy);
@@ -836,6 +922,7 @@ static av_cold int aac_encode_init(AVCodecContext *avctx)
         ff_aac_coder_init_mips(s);
 
     s->lambda = avctx->global_quality > 0 ? avctx->global_quality : 120;
+    s->random_state = 0x1f2e3d4c;
 
     ff_aac_tableinit();
 
@@ -859,18 +946,10 @@ static const AVOption aacenc_options[] = {
         {"anmr",     "ANMR method",               0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_ANMR},    INT_MIN, INT_MAX, AACENC_FLAGS, "aac_coder"},
         {"twoloop",  "Two loop searching method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_TWOLOOP}, INT_MIN, INT_MAX, AACENC_FLAGS, "aac_coder"},
         {"fast",     "Constant quantizer",        0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_FAST},    INT_MIN, INT_MAX, AACENC_FLAGS, "aac_coder"},
-    {"aac_pns", "Perceptual Noise Substitution", offsetof(AACEncContext, options.pns), AV_OPT_TYPE_INT, {.i64 = 1}, 0, 1, AACENC_FLAGS, "aac_pns"},
-        {"disable",  "Disable perceptual noise substitution", 0, AV_OPT_TYPE_CONST, {.i64 =  0 }, INT_MIN, INT_MAX, AACENC_FLAGS, "aac_pns"},
-        {"enable",   "Enable perceptual noise substitution",  0, AV_OPT_TYPE_CONST, {.i64 =  1 }, INT_MIN, INT_MAX, AACENC_FLAGS, "aac_pns"},
-    {"aac_is", "Intensity stereo coding", offsetof(AACEncContext, options.intensity_stereo), AV_OPT_TYPE_INT, {.i64 = 1}, 0, 1, AACENC_FLAGS, "intensity_stereo"},
-        {"disable",  "Disable intensity stereo coding", 0, AV_OPT_TYPE_CONST, {.i64 = 0}, INT_MIN, INT_MAX, AACENC_FLAGS, "intensity_stereo"},
-        {"enable",   "Enable intensity stereo coding", 0, AV_OPT_TYPE_CONST, {.i64 = 1}, INT_MIN, INT_MAX, AACENC_FLAGS, "intensity_stereo"},
-    {"aac_tns", "Temporal noise shaping", offsetof(AACEncContext, options.tns), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, AACENC_FLAGS, "aac_tns"},
-        {"disable",  "Disable temporal noise shaping", 0, AV_OPT_TYPE_CONST, {.i64 = 0}, INT_MIN, INT_MAX, AACENC_FLAGS, "aac_tns"},
-        {"enable",   "Enable temporal noise shaping", 0, AV_OPT_TYPE_CONST, {.i64 = 1}, INT_MIN, INT_MAX, AACENC_FLAGS, "aac_tns"},
-    {"aac_pred", "AAC-Main prediction", offsetof(AACEncContext, options.pred), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, AACENC_FLAGS, "aac_pred"},
-        {"disable",  "Disable AAC-Main prediction", 0, AV_OPT_TYPE_CONST, {.i64 = 0}, INT_MIN, INT_MAX, AACENC_FLAGS, "aac_pred"},
-        {"enable",   "Enable AAC-Main prediction", 0, AV_OPT_TYPE_CONST, {.i64 = 1}, INT_MIN, INT_MAX, AACENC_FLAGS, "aac_pred"},
+    {"aac_pns", "Perceptual Noise Substitution", offsetof(AACEncContext, options.pns), AV_OPT_TYPE_BOOL, {.i64 = 1}, 0, 1, AACENC_FLAGS},
+    {"aac_is", "Intensity stereo coding", offsetof(AACEncContext, options.intensity_stereo), AV_OPT_TYPE_BOOL, {.i64 = 1}, 0, 1, AACENC_FLAGS},
+    {"aac_tns", "Temporal noise shaping", offsetof(AACEncContext, options.tns), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, AACENC_FLAGS},
+    {"aac_pred", "AAC-Main prediction", offsetof(AACEncContext, options.pred), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, AACENC_FLAGS},
     {NULL}
 };