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.

1 files changed, 241 insertions, 56 deletions

diff --git a/libavcodec/aaccoder.c b/libavcodec/aaccoder.c
index 10ea14b..dafdc9f 100644
--- a/libavcodec/aaccoder.c
+++ b/libavcodec/aaccoder.c
@@ -33,7 +33,9 @@
 #include "libavutil/libm.h" // brought forward to work around cygwin header breakage
 
 #include <float.h>
+
 #include "libavutil/mathematics.h"
+#include "mathops.h"
 #include "avcodec.h"
 #include "put_bits.h"
 #include "aac.h"
@@ -50,9 +52,6 @@
 
 #include "libavcodec/aaccoder_twoloop.h"
 
-/** Frequency in Hz for lower limit of noise substitution **/
-#define NOISE_LOW_LIMIT 4000
-
 /* Parameter of f(x) = a*(lambda/100), defines the maximum fourier spread
  * beyond which no PNS is used (since the SFBs contain tone rather than noise) */
 #define NOISE_SPREAD_THRESHOLD 0.5073f
@@ -124,7 +123,7 @@ static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce
                     rd += quantize_band_cost(s, &sce->coeffs[start + w*128],
                                              &s->scoefs[start + w*128], size,
                                              sce->sf_idx[(win+w)*16+swb], aac_cb_out_map[cb],
-                                             lambda / band->threshold, INFINITY, NULL, 0);
+                                             lambda / band->threshold, INFINITY, NULL, NULL, 0);
                 }
                 cost_stay_here = path[swb][cb].cost + rd;
                 cost_get_here  = minrd              + rd + run_bits + 4;
@@ -335,7 +334,7 @@ static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                         FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
                         dist += quantize_band_cost(s, coefs + w2*128, s->scoefs + start + w2*128, sce->ics.swb_sizes[g],
-                                                   q + q0, cb, lambda / band->threshold, INFINITY, NULL, 0);
+                                                   q + q0, cb, lambda / band->threshold, INFINITY, NULL, NULL, 0);
                     }
                     minrd = FFMIN(minrd, dist);
 
@@ -499,7 +498,7 @@ static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
                                                ESC_BT,
                                                lambda,
                                                INFINITY,
-                                               &b,
+                                               &b, NULL,
                                                0);
                     dist -= b;
                 }
@@ -588,12 +587,36 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne
 {
     FFPsyBand *band;
     int w, g, w2, i;
+    int wlen = 1024 / sce->ics.num_windows;
+    int bandwidth, cutoff;
     float *PNS = &s->scoefs[0*128], *PNS34 = &s->scoefs[1*128];
     float *NOR34 = &s->scoefs[3*128];
     const float lambda = s->lambda;
-    const float freq_mult = avctx->sample_rate/(1024.0f/sce->ics.num_windows)/2.0f;
+    const float freq_mult = avctx->sample_rate*0.5f/wlen;
     const float thr_mult = NOISE_LAMBDA_REPLACE*(100.0f/lambda);
-    const float spread_threshold = NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f);
+    const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
+    const float dist_bias = av_clipf(4.f * 120 / lambda, 0.25f, 4.0f);
+    const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
+
+    int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
+        / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
+        * (lambda / 120.f);
+
+    /** Keep this in sync with twoloop's cutoff selection */
+    float rate_bandwidth_multiplier = 1.5f;
+    int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
+        ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
+        : (avctx->bit_rate / avctx->channels);
+
+    frame_bit_rate *= 1.15f;
+
+    if (avctx->cutoff > 0) {
+        bandwidth = avctx->cutoff;
+    } else {
+        bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
+    }
+
+    cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
 
     memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
@@ -602,32 +625,44 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne
             int noise_sfi;
             float dist1 = 0.0f, dist2 = 0.0f, noise_amp;
             float pns_energy = 0.0f, pns_tgt_energy, energy_ratio, dist_thresh;
-            float sfb_energy = 0.0f, threshold = 0.0f, spread = 0.0f;
+            float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
+            float min_energy = -1.0f, max_energy = 0.0f;
             const int start = wstart+sce->ics.swb_offset[g];
             const float freq = (start-wstart)*freq_mult;
             const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
-            if (freq < NOISE_LOW_LIMIT || avctx->cutoff && freq >= avctx->cutoff)
+            if (freq < NOISE_LOW_LIMIT || (start-wstart) >= cutoff)
                 continue;
             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                 band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
                 sfb_energy += band->energy;
-                spread     += band->spread;
+                spread     = FFMIN(spread, band->spread);
                 threshold  += band->threshold;
+                if (!w2) {
+                    min_energy = max_energy = band->energy;
+                } else {
+                    min_energy = FFMIN(min_energy, band->energy);
+                    max_energy = FFMAX(max_energy, band->energy);
+                }
             }
 
             /* Ramps down at ~8000Hz and loosens the dist threshold */
-            dist_thresh = FFMIN(2.5f*NOISE_LOW_LIMIT/freq, 2.5f);
-
-            /* zero and energy close to threshold usually means hole avoidance,
-             * we do want to remain avoiding holes with PNS
+            dist_thresh = av_clipf(2.5f*NOISE_LOW_LIMIT/freq, 0.5f, 2.5f) * dist_bias;
+
+            /* PNS is acceptable when all of these are true:
+             * 1. high spread energy (noise-like band)
+             * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
+             * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
+             *
+             * At this stage, point 2 is relaxed for zeroed bands near the noise threshold (hole avoidance is more important)
              */
             if (((sce->zeroes[w*16+g] || !sce->band_alt[w*16+g]) && sfb_energy < threshold*sqrtf(1.5f/freq_boost)) || spread < spread_threshold ||
-                (sce->band_alt[w*16+g] && sfb_energy > threshold*thr_mult*freq_boost)) {
+                (!sce->zeroes[w*16+g] && sce->band_alt[w*16+g] && sfb_energy > threshold*thr_mult*freq_boost) ||
+                min_energy < pns_transient_energy_r * max_energy ) {
                 sce->pns_ener[w*16+g] = sfb_energy;
                 continue;
             }
 
-            pns_tgt_energy = sfb_energy*spread*spread/sce->ics.group_len[w];
+            pns_tgt_energy = sfb_energy*FFMIN(1.0f, spread*spread);
             noise_sfi = av_clip(roundf(log2f(pns_tgt_energy)*2), -100, 155); /* Quantize */
             noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO];    /* Dequantize */
             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
@@ -648,13 +683,18 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne
                                             sce->ics.swb_sizes[g],
                                             sce->sf_idx[(w+w2)*16+g],
                                             sce->band_alt[(w+w2)*16+g],
-                                            lambda/band->threshold, INFINITY, NULL, 0);
-                /* Estimate rd on average as 9 bits for CB and sf + spread energy * lambda/thr */
-                dist2 += 9+band->energy/(band->spread*band->spread)*lambda/band->threshold;
+                                            lambda/band->threshold, INFINITY, NULL, NULL, 0);
+                /* Estimate rd on average as 5 bits for SF, 4 for the CB, plus spread energy * lambda/thr */
+                dist2 += band->energy/(band->spread*band->spread)*lambda*dist_thresh/band->threshold;
+            }
+            if (g && sce->sf_idx[(w+w2)*16+g-1] == NOISE_BT) {
+                dist2 += 5;
+            } else {
+                dist2 += 9;
             }
             energy_ratio = pns_tgt_energy/pns_energy; /* Compensates for quantization error */
             sce->pns_ener[w*16+g] = energy_ratio*pns_tgt_energy;
-            if (energy_ratio > 0.85f && energy_ratio < 1.25f && (sce->zeroes[w*16+g] || !sce->band_alt[w*16+g] || dist2*dist_thresh < dist1)) {
+            if (sce->zeroes[w*16+g] || !sce->band_alt[w*16+g] || (energy_ratio > 0.85f && energy_ratio < 1.25f && dist2 < dist1)) {
                 sce->band_type[w*16+g] = NOISE_BT;
                 sce->zeroes[w*16+g] = 0;
             }
@@ -662,62 +702,203 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne
     }
 }
 
+static void mark_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelElement *sce)
+{
+    FFPsyBand *band;
+    int w, g, w2;
+    int wlen = 1024 / sce->ics.num_windows;
+    int bandwidth, cutoff;
+    const float lambda = s->lambda;
+    const float freq_mult = avctx->sample_rate*0.5f/wlen;
+    const float spread_threshold = FFMIN(0.75f, NOISE_SPREAD_THRESHOLD*FFMAX(0.5f, lambda/100.f));
+    const float pns_transient_energy_r = FFMIN(0.7f, lambda / 140.f);
+
+    int refbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
+        / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
+        * (lambda / 120.f);
+
+    /** Keep this in sync with twoloop's cutoff selection */
+    float rate_bandwidth_multiplier = 1.5f;
+    int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
+        ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
+        : (avctx->bit_rate / avctx->channels);
+
+    frame_bit_rate *= 1.15f;
+
+    if (avctx->cutoff > 0) {
+        bandwidth = avctx->cutoff;
+    } else {
+        bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
+    }
+
+    cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
+
+    memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+        for (g = 0;  g < sce->ics.num_swb; g++) {
+            float sfb_energy = 0.0f, threshold = 0.0f, spread = 2.0f;
+            float min_energy = -1.0f, max_energy = 0.0f;
+            const int start = sce->ics.swb_offset[g];
+            const float freq = start*freq_mult;
+            const float freq_boost = FFMAX(0.88f*freq/NOISE_LOW_LIMIT, 1.0f);
+            if (freq < NOISE_LOW_LIMIT || start >= cutoff) {
+                sce->can_pns[w*16+g] = 0;
+                continue;
+            }
+            for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
+                band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
+                sfb_energy += band->energy;
+                spread     = FFMIN(spread, band->spread);
+                threshold  += band->threshold;
+                if (!w2) {
+                    min_energy = max_energy = band->energy;
+                } else {
+                    min_energy = FFMIN(min_energy, band->energy);
+                    max_energy = FFMAX(max_energy, band->energy);
+                }
+            }
+
+            /* PNS is acceptable when all of these are true:
+             * 1. high spread energy (noise-like band)
+             * 2. near-threshold energy (high PE means the random nature of PNS content will be noticed)
+             * 3. on short window groups, all windows have similar energy (variations in energy would be destroyed by PNS)
+             */
+            sce->pns_ener[w*16+g] = sfb_energy;
+            if (sfb_energy < threshold*sqrtf(1.5f/freq_boost) || spread < spread_threshold || min_energy < pns_transient_energy_r * max_energy) {
+                sce->can_pns[w*16+g] = 0;
+            } else {
+                sce->can_pns[w*16+g] = 1;
+            }
+        }
+    }
+}
+
 static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
 {
-    int start = 0, i, w, w2, g;
+    int start = 0, i, w, w2, g, sid_sf_boost;
     float M[128], S[128];
     float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
     const float lambda = s->lambda;
+    const float mslambda = FFMIN(1.0f, lambda / 120.f);
     SingleChannelElement *sce0 = &cpe->ch[0];
     SingleChannelElement *sce1 = &cpe->ch[1];
     if (!cpe->common_window)
         return;
     for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) {
+        int min_sf_idx_mid = SCALE_MAX_POS;
+        int min_sf_idx_side = SCALE_MAX_POS;
+        for (g = 0; g < sce0->ics.num_swb; g++) {
+            if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT)
+                min_sf_idx_mid = FFMIN(min_sf_idx_mid, sce0->sf_idx[w*16+g]);
+            if (!sce1->zeroes[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT)
+                min_sf_idx_side = FFMIN(min_sf_idx_side, sce1->sf_idx[w*16+g]);
+        }
+
         start = 0;
         for (g = 0;  g < sce0->ics.num_swb; g++) {
+            float bmax = bval2bmax(g * 17.0f / sce0->ics.num_swb) / 0.0045f;
+            cpe->ms_mask[w*16+g] = 0;
             if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
-                float dist1 = 0.0f, dist2 = 0.0f;
+                float Mmax = 0.0f, Smax = 0.0f;
+
+                /* Must compute mid/side SF and book for the whole window group */
                 for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
-                    FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
-                    FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
-                    float minthr = FFMIN(band0->threshold, band1->threshold);
-                    float maxthr = FFMAX(band0->threshold, band1->threshold);
                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
                         M[i] = (sce0->coeffs[start+(w+w2)*128+i]
                               + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
                         S[i] =  M[i]
                               - sce1->coeffs[start+(w+w2)*128+i];
                     }
-                    abs_pow34_v(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
-                    abs_pow34_v(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
-                    abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
-                    abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
-                    dist1 += quantize_band_cost(s, &sce0->coeffs[start + (w+w2)*128],
-                                                L34,
-                                                sce0->ics.swb_sizes[g],
-                                                sce0->sf_idx[(w+w2)*16+g],
-                                                sce0->band_type[(w+w2)*16+g],
-                                                lambda / band0->threshold, INFINITY, NULL, 0);
-                    dist1 += quantize_band_cost(s, &sce1->coeffs[start + (w+w2)*128],
-                                                R34,
-                                                sce1->ics.swb_sizes[g],
-                                                sce1->sf_idx[(w+w2)*16+g],
-                                                sce1->band_type[(w+w2)*16+g],
-                                                lambda / band1->threshold, INFINITY, NULL, 0);
-                    dist2 += quantize_band_cost(s, M,
-                                                M34,
-                                                sce0->ics.swb_sizes[g],
-                                                sce0->sf_idx[(w+w2)*16+g],
-                                                sce0->band_type[(w+w2)*16+g],
-                                                lambda / maxthr, INFINITY, NULL, 0);
-                    dist2 += quantize_band_cost(s, S,
-                                                S34,
-                                                sce1->ics.swb_sizes[g],
-                                                sce1->sf_idx[(w+w2)*16+g],
-                                                sce1->band_type[(w+w2)*16+g],
-                                                lambda / minthr, INFINITY, NULL, 0);
+                    abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
+                    abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
+                    for (i = 0; i < sce0->ics.swb_sizes[g]; i++ ) {
+                        Mmax = FFMAX(Mmax, M34[i]);
+                        Smax = FFMAX(Smax, S34[i]);
+                    }
+                }
+
+                for (sid_sf_boost = 0; sid_sf_boost < 4; sid_sf_boost++) {
+                    float dist1 = 0.0f, dist2 = 0.0f;
+                    int B0 = 0, B1 = 0;
+                    int minidx;
+                    int mididx, sididx;
+                    int midcb, sidcb;
+
+                    minidx = FFMIN(sce0->sf_idx[w*16+g], sce1->sf_idx[w*16+g]);
+                    mididx = av_clip(minidx, min_sf_idx_mid, min_sf_idx_mid + SCALE_MAX_DIFF);
+                    sididx = av_clip(minidx - sid_sf_boost * 3, min_sf_idx_side, min_sf_idx_side + SCALE_MAX_DIFF);
+                    midcb = find_min_book(Mmax, mididx);
+                    sidcb = find_min_book(Smax, sididx);
+
+                    if ((mididx > minidx) || (sididx > minidx)) {
+                        /* scalefactor range violation, bad stuff, will decrease quality unacceptably */
+                        continue;
+                    }
+
+                    /* No CB can be zero */
+                    midcb = FFMAX(1,midcb);
+                    sidcb = FFMAX(1,sidcb);
+
+                    for (w2 = 0; w2 < sce0->ics.group_len[w]; w2++) {
+                        FFPsyBand *band0 = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g];
+                        FFPsyBand *band1 = &s->psy.ch[s->cur_channel+1].psy_bands[(w+w2)*16+g];
+                        float minthr = FFMIN(band0->threshold, band1->threshold);
+                        int b1,b2,b3,b4;
+                        for (i = 0; i < sce0->ics.swb_sizes[g]; i++) {
+                            M[i] = (sce0->coeffs[start+(w+w2)*128+i]
+                                  + sce1->coeffs[start+(w+w2)*128+i]) * 0.5;
+                            S[i] =  M[i]
+                                  - sce1->coeffs[start+(w+w2)*128+i];
+                        }
+
+                        abs_pow34_v(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
+                        abs_pow34_v(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
+                        abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
+                        abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
+                        dist1 += quantize_band_cost(s, &sce0->coeffs[start + (w+w2)*128],
+                                                    L34,
+                                                    sce0->ics.swb_sizes[g],
+                                                    sce0->sf_idx[(w+w2)*16+g],
+                                                    sce0->band_type[(w+w2)*16+g],
+                                                    lambda / band0->threshold, INFINITY, &b1, NULL, 0);
+                        dist1 += quantize_band_cost(s, &sce1->coeffs[start + (w+w2)*128],
+                                                    R34,
+                                                    sce1->ics.swb_sizes[g],
+                                                    sce1->sf_idx[(w+w2)*16+g],
+                                                    sce1->band_type[(w+w2)*16+g],
+                                                    lambda / band1->threshold, INFINITY, &b2, NULL, 0);
+                        dist2 += quantize_band_cost(s, M,
+                                                    M34,
+                                                    sce0->ics.swb_sizes[g],
+                                                    sce0->sf_idx[(w+w2)*16+g],
+                                                    sce0->band_type[(w+w2)*16+g],
+                                                    lambda / minthr, INFINITY, &b3, NULL, 0);
+                        dist2 += quantize_band_cost(s, S,
+                                                    S34,
+                                                    sce1->ics.swb_sizes[g],
+                                                    sce1->sf_idx[(w+w2)*16+g],
+                                                    sce1->band_type[(w+w2)*16+g],
+                                                    mslambda / (minthr * bmax), INFINITY, &b4, NULL, 0);
+                        B0 += b1+b2;
+                        B1 += b3+b4;
+                        dist1 -= B0;
+                        dist2 -= B1;
+                    }
+                    cpe->ms_mask[w*16+g] = dist2 <= dist1 && B1 < B0;
+                    if (cpe->ms_mask[w*16+g]) {
+                        /* Setting the M/S mask is useful with I/S, but only the flag */
+                        if (!cpe->is_mask[w*16+g]) {
+                            sce0->sf_idx[w*16+g] = mididx;
+                            sce1->sf_idx[w*16+g] = sididx;
+                            sce0->band_type[w*16+g] = midcb;
+                            sce1->band_type[w*16+g] = sidcb;
+                        }
+                        break;
+                    } else if (B1 > B0) {
+                        /* More boost won't fix this */
+                        break;
+                    }
                 }
-                cpe->ms_mask[w*16+g] = dist2 < dist1;
             }
             start += sce0->ics.swb_sizes[g];
         }
@@ -736,6 +917,7 @@ AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
         ff_aac_apply_tns,
         set_special_band_scalefactors,
         search_for_pns,
+        mark_pns,
         ff_aac_search_for_tns,
         search_for_ms,
         ff_aac_search_for_is,
@@ -752,6 +934,7 @@ AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
         ff_aac_apply_tns,
         set_special_band_scalefactors,
         search_for_pns,
+        mark_pns,
         ff_aac_search_for_tns,
         search_for_ms,
         ff_aac_search_for_is,
@@ -768,6 +951,7 @@ AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
         ff_aac_apply_tns,
         set_special_band_scalefactors,
         search_for_pns,
+        mark_pns,
         ff_aac_search_for_tns,
         search_for_ms,
         ff_aac_search_for_is,
@@ -784,6 +968,7 @@ AACCoefficientsEncoder ff_aac_coders[AAC_CODER_NB] = {
         ff_aac_apply_tns,
         set_special_band_scalefactors,
         search_for_pns,
+        mark_pns,
         ff_aac_search_for_tns,
         search_for_ms,
         ff_aac_search_for_is,