AAC encoder: improve SF range utilization

This patch does 4 things, all of which interact and thus it
woudln't be possible to commit them separately without causing
either quality regressions or assertion failures.

Fate comparison targets don't all reflect improvements in
quality, yet listening tests show substantially improved quality
and stability.

1. Increase SF range utilization.

The spec requires SF delta values to be constrained within the
range -60..60. The previous code was applying that range to
the whole SF array and not only the deltas of consecutive values,
because doing so requires smarter code: zeroing or otherwise
skipping a band may invalidate lots of SF choices.

This patch implements that logic to allow the coders to utilize
the full dynamic range of scalefactors, increasing quality quite
considerably, and fixing delta-SF-related assertion failures,
since now the limitation is enforced rather than asserted.

2. PNS tweaks

The previous modification makes big improvements in twoloop's
efficiency, and every time that happens PNS logic needs to be
tweaked accordingly to avoid it from stepping all over twoloop's
decisions. This patch includes modifications of the sort.

3. Account for lowpass cutoff during PSY analysis

The closer PSY's allocation is to final allocation the better
the quality is, and given these modifications, twoloop is now
very efficient at avoiding holes. Thus, to compute accurate
thresholds, PSY needs to account for the lowpass applied
implicitly during twoloop (by zeroing high bands).

This patch makes twoloop set the cutoff in psymodel's context
the first time it runs, and makes PSY account for it during
threshold computation, making PE and threshold computations
closer to the final allocation and thus achieving better
subjective quality.

4. Tweaks to RC lambda tracking loop in relation to PNS

Without this tweak some corner cases cause quality regressions.
Basically, lambda needs to react faster to overall bitrate
efficiency changes since now PNS can be quite successful in
enforcing maximum bitrates, when PSY allocates too many bits
to the lower bands, suppressing the signals RC logic uses to
lower lambda in those cases and causing aggressive PNS.

This tweak makes PNS much less aggressive, though it can still
use some further tweaks.

Also update MIPS specializations and adjust fuzz

Also in lavc/mips/aacpsy_mips.h: remove trailing whitespace

11 files changed, 393 insertions, 177 deletions

diff --git a/libavcodec/aaccoder.c b/libavcodec/aaccoder.c
index 2a66045..2a0cb1f 100644
--- a/libavcodec/aaccoder.c
+++ b/libavcodec/aaccoder.c
@@ -54,7 +54,7 @@
 
 /* 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
+#define NOISE_SPREAD_THRESHOLD 0.9f
 
 /* Parameter of f(x) = a*(100/lambda), defines how much PNS is allowed to
  * replace low energy non zero bands */
@@ -591,6 +591,7 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne
     int bandwidth, cutoff;
     float *PNS = &s->scoefs[0*128], *PNS34 = &s->scoefs[1*128];
     float *NOR34 = &s->scoefs[3*128];
+    uint8_t nextband[128];
     const float lambda = s->lambda;
     const float freq_mult = avctx->sample_rate*0.5f/wlen;
     const float thr_mult = NOISE_LAMBDA_REPLACE*(100.0f/lambda);
@@ -604,6 +605,7 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne
 
     /** Keep this in sync with twoloop's cutoff selection */
     float rate_bandwidth_multiplier = 1.5f;
+    int prev = -1000, prev_sf = -1;
     int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
         ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
         : (avctx->bit_rate / avctx->channels);
@@ -619,6 +621,7 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne
     cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
 
     memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type));
+    ff_init_nextband_map(sce, nextband);
     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
         int wstart = w*128;
         for (g = 0;  g < sce->ics.num_swb; g++) {
@@ -655,16 +658,27 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne
              *
              * 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 ||
+            if ((!sce->zeroes[w*16+g] && !ff_sfdelta_can_remove_band(sce, nextband, prev_sf, w*16+g)) ||
+                ((sce->zeroes[w*16+g] || !sce->band_alt[w*16+g]) && sfb_energy < threshold*sqrtf(1.0f/freq_boost)) || spread < spread_threshold ||
                 (!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;
+                if (!sce->zeroes[w*16+g])
+                    prev_sf = sce->sf_idx[w*16+g];
                 continue;
             }
 
             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 */
+            if (prev != -1000) {
+                int noise_sfdiff = noise_sfi - prev + SCALE_DIFF_ZERO;
+                if (noise_sfdiff < 0 || noise_sfdiff > 2*SCALE_MAX_DIFF) {
+                    if (!sce->zeroes[w*16+g])
+                        prev_sf = sce->sf_idx[w*16+g];
+                    continue;
+                }
+            }
             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                 float band_energy, scale, pns_senergy;
                 const int start_c = (w+w2)*128+sce->ics.swb_offset[g];
@@ -697,7 +711,10 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne
             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;
+                prev = noise_sfi;
             }
+            if (!sce->zeroes[w*16+g])
+                prev_sf = sce->sf_idx[w*16+g];
         }
     }
 }
@@ -775,7 +792,8 @@ static void mark_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelEleme
 
 static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
 {
-    int start = 0, i, w, w2, g, sid_sf_boost;
+    int start = 0, i, w, w2, g, sid_sf_boost, prev_mid, prev_side;
+    uint8_t nextband0[128], nextband1[128];
     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;
@@ -784,21 +802,19 @@ static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
     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]);
-        }
 
+    /** Scout out next nonzero bands */
+    ff_init_nextband_map(sce0, nextband0);
+    ff_init_nextband_map(sce1, nextband1);
+
+    prev_mid = sce0->sf_idx[0];
+    prev_side = sce1->sf_idx[0];
+    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++) {
             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]) {
+            if (!sce0->zeroes[w*16+g] && !sce1->zeroes[w*16+g]) {
                 float Mmax = 0.0f, Smax = 0.0f;
 
                 /* Must compute mid/side SF and book for the whole window group */
@@ -825,16 +841,18 @@ static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
                     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)) {
+                    mididx = av_clip(minidx, 0, SCALE_MAX_POS - SCALE_DIV_512);
+                    sididx = av_clip(minidx - sid_sf_boost * 3, 0, SCALE_MAX_POS - SCALE_DIV_512);
+                    if (!cpe->is_mask[w*16+g] && sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT
+                        && (   !ff_sfdelta_can_replace(sce0, nextband0, prev_mid, mididx, w*16+g)
+                            || !ff_sfdelta_can_replace(sce1, nextband1, prev_side, sididx, w*16+g))) {
                         /* scalefactor range violation, bad stuff, will decrease quality unacceptably */
                         continue;
                     }
 
+                    midcb = find_min_book(Mmax, mididx);
+                    sidcb = find_min_book(Smax, sididx);
+
                     /* No CB can be zero */
                     midcb = FFMAX(1,midcb);
                     sidcb = FFMAX(1,sidcb);
@@ -900,6 +918,10 @@ static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
                     }
                 }
             }
+            if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT)
+                prev_mid = sce0->sf_idx[w*16+g];
+            if (!sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT)
+                prev_side = sce1->sf_idx[w*16+g];
             start += sce0->ics.swb_sizes[g];
         }
     }
diff --git a/libavcodec/aaccoder_twoloop.h b/libavcodec/aaccoder_twoloop.h
index d4290b4..397a4db 100644
--- a/libavcodec/aaccoder_twoloop.h
+++ b/libavcodec/aaccoder_twoloop.h
@@ -76,6 +76,7 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
     int refbits = destbits;
     int toomanybits, toofewbits;
     char nzs[128];
+    uint8_t nextband[128];
     int maxsf[128];
     float dists[128] = { 0 }, qenergies[128] = { 0 }, uplims[128], euplims[128], energies[128];
     float maxvals[128], spread_thr_r[128];
@@ -102,7 +103,7 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
      */
     float sfoffs = av_clipf(log2f(120.0f / lambda) * 4.0f, -5, 10);
 
-    int fflag, minscaler, maxscaler, nminscaler, minrdsf;
+    int fflag, minscaler, maxscaler, nminscaler;
     int its  = 0;
     int maxits = 30;
     int allz = 0;
@@ -158,9 +159,13 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
         /** search further */
         maxits *= 2;
     } else {
-        /** When using ABR, be strict */
-        toomanybits = destbits + destbits/16;
-        toofewbits = destbits - destbits/4;
+        /* When using ABR, be strict, but a reasonable leeway is
+         * critical to allow RC to smoothly track desired bitrate
+         * without sudden quality drops that cause audible artifacts.
+         * Symmetry is also desirable, to avoid systematic bias.
+         */
+        toomanybits = destbits + destbits/8;
+        toofewbits = destbits - destbits/8;
 
         sfoffs = 0;
         rdlambda = sqrtf(rdlambda);
@@ -191,6 +196,7 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
             bandwidth = avctx->cutoff;
         } else {
             bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
+            s->psy.cutoff = bandwidth;
         }
 
         cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
@@ -241,7 +247,7 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
             nzs[w*16+g] = nz;
             sce->zeroes[w*16+g] = !nz;
             allz |= nz;
-            if (nz) {
+            if (nz && sce->can_pns[w*16+g]) {
                 spread_thr_r[w*16+g] = energy * nz / (uplim * spread);
                 if (min_spread_thr_r < 0) {
                     min_spread_thr_r = max_spread_thr_r = spread_thr_r[w*16+g];
@@ -433,6 +439,7 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
         } while (qstep);
 
         overdist = 1;
+        fflag = tbits < toofewbits;
         for (i = 0; i < 2 && (overdist || recomprd); ++i) {
             if (recomprd) {
                 /** Must recompute distortion */
@@ -484,13 +491,13 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
                     }
                 }
             }
-            if (!i && s->options.pns && its > maxits/2) {
+            if (!i && s->options.pns && its > maxits/2 && tbits > toofewbits) {
                 float maxoverdist = 0.0f;
+                float ovrfactor = 1.f+(maxits-its)*16.f/maxits;
                 overdist = recomprd = 0;
                 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
-                    float ovrfactor = 2.f+(maxits-its)*16.f/maxits;
                     for (g = start = 0;  g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
-                        if (!sce->zeroes[w*16+g] && dists[w*16+g] > uplims[w*16+g]*ovrfactor) {
+                        if (!sce->zeroes[w*16+g] && sce->sf_idx[w*16+g] > SCALE_ONE_POS && dists[w*16+g] > uplims[w*16+g]*ovrfactor) {
                             float ovrdist = dists[w*16+g] / FFMAX(uplims[w*16+g],euplims[w*16+g]);
                             maxoverdist = FFMAX(maxoverdist, ovrdist);
                             overdist++;
@@ -506,7 +513,7 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
                     float zspread;
                     int zeroable = 0;
                     int zeroed = 0;
-                    int maxzeroed;
+                    int maxzeroed, zloop;
                     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
                         for (g = start = 0;  g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
                             if (start >= pns_start_pos && !sce->zeroes[w*16+g] && sce->can_pns[w*16+g]) {
@@ -517,21 +524,41 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
                         }
                     }
                     zspread = (maxspread-minspread) * 0.0125f + minspread;
-                    zspread = FFMIN(maxoverdist, zspread);
-                    maxzeroed = zeroable * its / (2 * maxits);
-                    for (g = sce->ics.num_swb-1; g > 0 && zeroed < maxzeroed; g--) {
-                        if (sce->ics.swb_offset[g] < pns_start_pos)
-                            continue;
-                        for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
-                            if (!sce->zeroes[w*16+g] && sce->can_pns[w*16+g] && spread_thr_r[w*16+g] <= zspread) {
-                                sce->zeroes[w*16+g] = 1;
-                                sce->band_type[w*16+g] = 0;
-                                zeroed++;
+                    /* Don't PNS everything even if allowed. It suppresses bit starvation signals from RC,
+                     * and forced the hand of the later search_for_pns step.
+                     * Instead, PNS a fraction of the spread_thr_r range depending on how starved for bits we are,
+                     * and leave further PNSing to search_for_pns if worthwhile.
+                     */
+                    zspread = FFMIN3(min_spread_thr_r * 8.f, zspread,
+                        ((toomanybits - tbits) * min_spread_thr_r + (tbits - toofewbits) * max_spread_thr_r) / (toomanybits - toofewbits + 1));
+                    maxzeroed = FFMIN(zeroable, FFMAX(1, (zeroable * its + maxits - 1) / (2 * maxits)));
+                    for (zloop = 0; zloop < 2; zloop++) {
+                        /* Two passes: first distorted stuff - two birds in one shot and all that,
+                         * then anything viable. Viable means not zero, but either CB=zero-able
+                         * (too high SF), not SF <= 1 (that means we'd be operating at very high
+                         * quality, we don't want PNS when doing VHQ), PNS allowed, and within
+                         * the lowest ranking percentile.
+                         */
+                        float loopovrfactor = (zloop) ? 1.0f : ovrfactor;
+                        int loopminsf = (zloop) ? (SCALE_ONE_POS - SCALE_DIV_512) : SCALE_ONE_POS;
+                        int mcb;
+                        for (g = sce->ics.num_swb-1; g > 0 && zeroed < maxzeroed; g--) {
+                            if (sce->ics.swb_offset[g] < pns_start_pos)
+                                continue;
+                            for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+                                if (!sce->zeroes[w*16+g] && sce->can_pns[w*16+g] && spread_thr_r[w*16+g] <= zspread
+                                    && sce->sf_idx[w*16+g] > loopminsf
+                                    && (dists[w*16+g] > loopovrfactor*uplims[w*16+g] || !(mcb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]))
+                                        || (mcb <= 1 && dists[w*16+g] > FFMIN(uplims[w*16+g], euplims[w*16+g]))) ) {
+                                    sce->zeroes[w*16+g] = 1;
+                                    sce->band_type[w*16+g] = 0;
+                                    zeroed++;
+                                }
                             }
                         }
                     }
                     if (zeroed)
-                        recomprd = 1;
+                        recomprd = fflag = 1;
                 } else {
                     overdist = 0;
                 }
@@ -549,9 +576,8 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
             }
         }
 
-        fflag = 0;
         minscaler = nminscaler = av_clip(minscaler, SCALE_ONE_POS - SCALE_DIV_512, SCALE_MAX_POS - SCALE_DIV_512);
-        minrdsf = FFMAX3(60, minscaler - 1, maxscaler - SCALE_MAX_DIFF - 1);
+        prev = -1;
         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
             /** Start with big steps, end up fine-tunning */
             int depth = (its > maxits/2) ? ((its > maxits*2/3) ? 1 : 3) : 10;
@@ -561,19 +587,22 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
             start = w * 128;
             for (g = 0; g < sce->ics.num_swb; g++) {
                 int prevsc = sce->sf_idx[w*16+g];
-                int minrdsfboost = (sce->ics.num_windows > 1) ? av_clip(g-4, -2, 0) : av_clip(g-16, -4, 0);
+                if (prev < 0 && !sce->zeroes[w*16+g])
+                    prev = sce->sf_idx[0];
                 if (!sce->zeroes[w*16+g]) {
                     const float *coefs = sce->coeffs + start;
                     const float *scaled = s->scoefs + start;
                     int cmb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
-                    if ((!cmb || dists[w*16+g] > uplims[w*16+g]) && sce->sf_idx[w*16+g] > minrdsf) {
+                    int mindeltasf = FFMAX(0, prev - SCALE_MAX_DIFF);
+                    int maxdeltasf = FFMIN(SCALE_MAX_POS - SCALE_DIV_512, prev + SCALE_MAX_DIFF);
+                    if ((!cmb || dists[w*16+g] > uplims[w*16+g]) && sce->sf_idx[w*16+g] > mindeltasf) {
                         /* Try to make sure there is some energy in every nonzero band
                          * NOTE: This algorithm must be forcibly imbalanced, pushing harder
                          *  on holes or more distorted bands at first, otherwise there's
                          *  no net gain (since the next iteration will offset all bands
                          *  on the opposite direction to compensate for extra bits)
                          */
-                        for (i = 0; i < edepth; ++i) {
+                        for (i = 0; i < edepth && sce->sf_idx[w*16+g] > mindeltasf; ++i) {
                             int cb, bits;
                             float dist, qenergy;
                             int mb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1);
@@ -585,6 +614,12 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
                             } else if (i >= depth && dists[w*16+g] < euplims[w*16+g]) {
                                 break;
                             }
+                            /* !g is the DC band, it's important, since quantization error here
+                             * applies to less than a cycle, it creates horrible intermodulation
+                             * distortion if it doesn't stick to what psy requests
+                             */
+                            if (!g && sce->ics.num_windows > 1 && dists[w*16+g] >= euplims[w*16+g])
+                                maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g], maxsf[w*16+g]);
                             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                                 int b;
                                 float sqenergy;
@@ -603,19 +638,19 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
                             sce->sf_idx[w*16+g]--;
                             dists[w*16+g] = dist - bits;
                             qenergies[w*16+g] = qenergy;
-                            if (mb && (sce->sf_idx[w*16+g] < (minrdsf+minrdsfboost) || (
+                            if (mb && (sce->sf_idx[w*16+g] < mindeltasf || (
                                     (dists[w*16+g] < FFMIN(uplmax*uplims[w*16+g], euplims[w*16+g]))
                                     && (fabsf(qenergies[w*16+g]-energies[w*16+g]) < euplims[w*16+g])
                                 ) )) {
                                 break;
                             }
                         }
-                    } else if (tbits > toofewbits && sce->sf_idx[w*16+g] < maxscaler
+                    } else if (tbits > toofewbits && sce->sf_idx[w*16+g] < FFMIN(maxdeltasf, maxsf[w*16+g])
                             && (dists[w*16+g] < FFMIN(euplims[w*16+g], uplims[w*16+g]))
                             && (fabsf(qenergies[w*16+g]-energies[w*16+g]) < euplims[w*16+g])
                         ) {
                         /** Um... over target. Save bits for more important stuff. */
-                        for (i = 0; i < depth; ++i) {
+                        for (i = 0; i < depth && sce->sf_idx[w*16+g] < maxdeltasf; ++i) {
                             int cb, bits;
                             float dist, qenergy;
                             cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]+1);
@@ -651,38 +686,53 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
                             }
                         }
                     }
+                    prev = sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], mindeltasf, maxdeltasf);
+                    if (sce->sf_idx[w*16+g] != prevsc)
+                        fflag = 1;
+                    nminscaler = FFMIN(nminscaler, sce->sf_idx[w*16+g]);
+                    sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
                 }
-                sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minrdsf, minscaler + SCALE_MAX_DIFF);
-                sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], SCALE_MAX_POS - SCALE_DIV_512);
-                if (sce->sf_idx[w*16+g] != prevsc)
-                    fflag = 1;
-                nminscaler = FFMIN(nminscaler, sce->sf_idx[w*16+g]);
-                sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
                 start += sce->ics.swb_sizes[g];
             }
         }
-        if (nminscaler < minscaler || sce->ics.num_windows > 1) {
-            /** SF difference limit violation risk. Must re-clamp. */
-            minscaler = nminscaler;
-            for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
-                for (g = 0; g < sce->ics.num_swb; g++) {
-                    sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
+
+        /** SF difference limit violation risk. Must re-clamp. */
+        prev = -1;
+        for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+            for (g = 0; g < sce->ics.num_swb; g++) {
+                if (!sce->zeroes[w*16+g]) {
+                    int prevsf = sce->sf_idx[w*16+g];
+                    if (prev < 0)
+                        prev = prevsf;
+                    sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], prev - SCALE_MAX_DIFF, prev + SCALE_MAX_DIFF);
                     sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
+                    prev = sce->sf_idx[w*16+g];
+                    if (!fflag && prevsf != sce->sf_idx[w*16+g])
+                        fflag = 1;
                 }
             }
         }
+
         its++;
     } while (fflag && its < maxits);
 
+    /** Scout out next nonzero bands */
+    ff_init_nextband_map(sce, nextband);
+
     prev = -1;
     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
         /** Make sure proper codebooks are set */
-        for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
+        for (g = 0; g < sce->ics.num_swb; g++) {
             if (!sce->zeroes[w*16+g]) {
                 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
                 if (sce->band_type[w*16+g] <= 0) {
-                    sce->zeroes[w*16+g] = 1;
-                    sce->band_type[w*16+g] = 0;
+                    if (!ff_sfdelta_can_remove_band(sce, nextband, prev, w*16+g)) {
+                        /** Cannot zero out, make sure it's not attempted */
+                        sce->band_type[w*16+g] = 1;
+                    } else {
+                        sce->zeroes[w*16+g] = 1;
+                        sce->band_type[w*16+g] = 0;
+                    }
                 }
             } else {
                 sce->band_type[w*16+g] = 0;
diff --git a/libavcodec/aacenc.c b/libavcodec/aacenc.c
index 0bd8891..971f8ab 100644
--- a/libavcodec/aacenc.c
+++ b/libavcodec/aacenc.c
@@ -793,7 +793,7 @@ static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
             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)) {
+            if (ratio > 0.9f && ratio < 1.1f) {
                 break;
             } else {
                 if (is_mode || ms_mode || tns_mode || pred_mode) {
diff --git a/libavcodec/aacenc_is.c b/libavcodec/aacenc_is.c
index 97be9b3..a4b8916 100644
--- a/libavcodec/aacenc_is.c
+++ b/libavcodec/aacenc_is.c
@@ -99,18 +99,23 @@ void ff_aac_search_for_is(AACEncContext *s, AVCodecContext *avctx, ChannelElemen
 {
     SingleChannelElement *sce0 = &cpe->ch[0];
     SingleChannelElement *sce1 = &cpe->ch[1];
-    int start = 0, count = 0, w, w2, g, i;
+    int start = 0, count = 0, w, w2, g, i, prev_sf1 = -1;
     const float freq_mult = avctx->sample_rate/(1024.0f/sce0->ics.num_windows)/2.0f;
+    uint8_t nextband1[128];
 
     if (!cpe->common_window)
         return;
 
+    /** Scout out next nonzero bands */
+    ff_init_nextband_map(sce1, nextband1);
+
     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++) {
             if (start*freq_mult > INT_STEREO_LOW_LIMIT*(s->lambda/170.0f) &&
                 cpe->ch[0].band_type[w*16+g] != NOISE_BT && !cpe->ch[0].zeroes[w*16+g] &&
-                cpe->ch[1].band_type[w*16+g] != NOISE_BT && !cpe->ch[1].zeroes[w*16+g]) {
+                cpe->ch[1].band_type[w*16+g] != NOISE_BT && !cpe->ch[1].zeroes[w*16+g] &&
+                ff_sfdelta_can_remove_band(sce1, nextband1, prev_sf1, w*16+g)) {
                 float ener0 = 0.0f, ener1 = 0.0f, ener01 = 0.0f, ener01p = 0.0f;
                 struct AACISError ph_err1, ph_err2, *erf;
                 if (sce0->band_type[w*16+g] == NOISE_BT ||
@@ -142,6 +147,8 @@ void ff_aac_search_for_is(AACEncContext *s, AVCodecContext *avctx, ChannelElemen
                     count++;
                 }
             }
+            if (!sce1->zeroes[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT)
+                prev_sf1 = sce1->sf_idx[w*16+g];
             start += sce0->ics.swb_sizes[g];
         }
     }
diff --git a/libavcodec/aacenc_utils.h b/libavcodec/aacenc_utils.h
index 40e1746..5b308f2 100644
--- a/libavcodec/aacenc_utils.h
+++ b/libavcodec/aacenc_utils.h
@@ -191,6 +191,69 @@ static inline int lcg_random(unsigned previous_val)
     return v.s;
 }
 
+
+/*
+ * Compute a nextband map to be used with SF delta constraint utilities.
+ * The nextband array should contain 128 elements, and positions that don't
+ * map to valid, nonzero bands of the form w*16+g (with w being the initial
+ * window of the window group, only) are left indetermined.
+ */
+static inline void ff_init_nextband_map(const SingleChannelElement *sce, uint8_t *nextband)
+{
+    unsigned char prevband = 0;
+    int w, g;
+    /** Just a safe default */
+    for (g = 0; g < 128; g++)
+        nextband[g] = g;
+
+    /** Now really navigate the nonzero band chain */
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+        for (g = 0; g < sce->ics.num_swb; g++) {
+            if (!sce->zeroes[w*16+g] && sce->band_type[w*16+g] < RESERVED_BT)
+                prevband = nextband[prevband] = w*16+g;
+        }
+    }
+    nextband[prevband] = prevband; /* terminate */
+}
+
+/*
+ * Updates nextband to reflect a removed band (equivalent to
+ * calling ff_init_nextband_map after marking a band as zero)
+ */
+static inline void ff_nextband_remove(uint8_t *nextband, int prevband, int band)
+{
+    nextband[prevband] = nextband[band];
+}
+
+/*
+ * Checks whether the specified band could be removed without inducing
+ * scalefactor delta that violates SF delta encoding constraints.
+ * prev_sf has to be the scalefactor of the previous nonzero, nonspecial
+ * band, in encoding order, or negative if there was no such band.
+ */
+static inline int ff_sfdelta_can_remove_band(const SingleChannelElement *sce,
+    const uint8_t *nextband, int prev_sf, int band)
+{
+    return prev_sf >= 0
+        && sce->sf_idx[nextband[band]] >= (prev_sf - SCALE_MAX_DIFF)
+        && sce->sf_idx[nextband[band]] <= (prev_sf + SCALE_MAX_DIFF);
+}
+
+/*
+ * Checks whether the specified band's scalefactor could be replaced
+ * with another one without violating SF delta encoding constraints.
+ * prev_sf has to be the scalefactor of the previous nonzero, nonsepcial
+ * band, in encoding order, or negative if there was no such band.
+ */
+static inline int ff_sfdelta_can_replace(const SingleChannelElement *sce,
+    const uint8_t *nextband, int prev_sf, int new_sf, int band)
+{
+    return new_sf >= (prev_sf - SCALE_MAX_DIFF)
+        && new_sf <= (prev_sf + SCALE_MAX_DIFF)
+        && sce->sf_idx[nextband[band]] >= (new_sf - SCALE_MAX_DIFF)
+        && sce->sf_idx[nextband[band]] <= (new_sf + SCALE_MAX_DIFF);
+}
+
 #define ERROR_IF(cond, ...) \
     if (cond) { \
         av_log(avctx, AV_LOG_ERROR, __VA_ARGS__); \
diff --git a/libavcodec/aacpsy.c b/libavcodec/aacpsy.c
index 40b3b41..71eeb3e 100644
--- a/libavcodec/aacpsy.c
+++ b/libavcodec/aacpsy.c
@@ -305,7 +305,7 @@ static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
     float prev, minscale, minath, minsnr, pe_min;
     int chan_bitrate = ctx->avctx->bit_rate / ((ctx->avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : ctx->avctx->channels);
 
-    const int bandwidth    = ctx->avctx->cutoff ? ctx->avctx->cutoff : AAC_CUTOFF(ctx->avctx);
+    const int bandwidth    = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx);
     const float num_bark   = calc_bark((float)bandwidth);
 
     ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
@@ -595,26 +595,30 @@ static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
 
 #ifndef calc_thr_3gpp
 static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch,
-                          const uint8_t *band_sizes, const float *coefs)
+                          const uint8_t *band_sizes, const float *coefs, const int cutoff)
 {
     int i, w, g;
-    int start = 0;
+    int start = 0, wstart = 0;
     for (w = 0; w < wi->num_windows*16; w += 16) {
+        wstart = 0;
         for (g = 0; g < num_bands; g++) {
             AacPsyBand *band = &pch->band[w+g];
 
             float form_factor = 0.0f;
             float Temp;
             band->energy = 0.0f;
-            for (i = 0; i < band_sizes[g]; i++) {
-                band->energy += coefs[start+i] * coefs[start+i];
-                form_factor  += sqrtf(fabs(coefs[start+i]));
+            if (wstart < cutoff) {
+                for (i = 0; i < band_sizes[g]; i++) {
+                    band->energy += coefs[start+i] * coefs[start+i];
+                    form_factor  += sqrtf(fabs(coefs[start+i]));
+                }
             }
             Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0;
             band->thr      = band->energy * 0.001258925f;
             band->nz_lines = form_factor * sqrtf(Temp);
 
             start += band_sizes[g];
+            wstart += band_sizes[g];
         }
     }
 }
@@ -655,9 +659,11 @@ static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
     const uint8_t *band_sizes  = ctx->bands[wi->num_windows == 8];
     AacPsyCoeffs  *coeffs      = pctx->psy_coef[wi->num_windows == 8];
     const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
+    const int bandwidth        = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx);
+    const int cutoff           = bandwidth * 2048 / wi->num_windows / ctx->avctx->sample_rate;
 
     //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
-    calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs);
+    calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs, cutoff);
 
     //modify thresholds and energies - spread, threshold in quiet, pre-echo control
     for (w = 0; w < wi->num_windows*16; w += 16) {
diff --git a/libavcodec/mips/aaccoder_mips.c b/libavcodec/mips/aaccoder_mips.c
index e8e1e62..04b3dca 100644
--- a/libavcodec/mips/aaccoder_mips.c
+++ b/libavcodec/mips/aaccoder_mips.c
@@ -2336,74 +2336,136 @@ static float quantize_band_cost(struct AACEncContext *s, const float *in,
 
 static void search_for_ms_mips(AACEncContext *s, ChannelElement *cpe)
 {
-    int start = 0, i, w, w2, g;
+    int start = 0, i, w, w2, g, sid_sf_boost, prev_mid, prev_side;
+    uint8_t nextband0[128], nextband1[128];
     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;
+
+    /** Scout out next nonzero bands */
+    ff_init_nextband_map(sce0, nextband0);
+    ff_init_nextband_map(sce1, nextband1);
+
+    prev_mid = sce0->sf_idx[0];
+    prev_side = sce1->sf_idx[0];
     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++) {
-            if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) {
-                float dist1 = 0.0f, dist2 = 0.0f;
+            float bmax = bval2bmax(g * 17.0f / sce0->ics.num_swb) / 0.0045f;
+            cpe->ms_mask[w*16+g] = 0;
+            if (!sce0->zeroes[w*16+g] && !sce1->zeroes[w*16+g]) {
+                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+=4) {
-                        M[i  ] = (sce0->coeffs[start+w2*128+i  ]
-                                + sce1->coeffs[start+w2*128+i  ]) * 0.5;
-                        M[i+1] = (sce0->coeffs[start+w2*128+i+1]
-                                + sce1->coeffs[start+w2*128+i+1]) * 0.5;
-                        M[i+2] = (sce0->coeffs[start+w2*128+i+2]
-                                + sce1->coeffs[start+w2*128+i+2]) * 0.5;
-                        M[i+3] = (sce0->coeffs[start+w2*128+i+3]
-                                + sce1->coeffs[start+w2*128+i+3]) * 0.5;
-
-                        S[i  ] =  M[i  ]
-                                - sce1->coeffs[start+w2*128+i  ];
-                        S[i+1] =  M[i+1]
-                                - sce1->coeffs[start+w2*128+i+1];
-                        S[i+2] =  M[i+2]
-                                - sce1->coeffs[start+w2*128+i+2];
-                        S[i+3] =  M[i+3]
-                                - sce1->coeffs[start+w2*128+i+3];
-                   }
-                    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, 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, 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, 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, NULL, 0);
+                    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(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, 0, SCALE_MAX_POS - SCALE_DIV_512);
+                    sididx = av_clip(minidx - sid_sf_boost * 3, 0, SCALE_MAX_POS - SCALE_DIV_512);
+                    if (!cpe->is_mask[w*16+g] && sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT
+                        && (   !ff_sfdelta_can_replace(sce0, nextband0, prev_mid, mididx, w*16+g)
+                            || !ff_sfdelta_can_replace(sce1, nextband1, prev_side, sididx, w*16+g))) {
+                        /* scalefactor range violation, bad stuff, will decrease quality unacceptably */
+                        continue;
+                    }
+
+                    midcb = find_min_book(Mmax, mididx);
+                    sidcb = find_min_book(Smax, sididx);
+
+                    /* 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 or PNS, but only the flag */
+                        if (!cpe->is_mask[w*16+g] && sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT) {
+                            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;
             }
+            if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT)
+                prev_mid = sce0->sf_idx[w*16+g];
+            if (!sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT)
+                prev_side = sce1->sf_idx[w*16+g];
             start += sce0->ics.swb_sizes[g];
         }
     }
diff --git a/libavcodec/mips/aacpsy_mips.h b/libavcodec/mips/aacpsy_mips.h
index 596dcad..42ff442 100644
--- a/libavcodec/mips/aacpsy_mips.h
+++ b/libavcodec/mips/aacpsy_mips.h
@@ -61,58 +61,62 @@
 #if HAVE_INLINE_ASM && HAVE_MIPSFPU && ( PSY_LAME_FIR_LEN == 21 )
 static void calc_thr_3gpp_mips(const FFPsyWindowInfo *wi, const int num_bands,
                                AacPsyChannel *pch, const uint8_t *band_sizes,
-                               const float *coefs)
+                               const float *coefs, const int cutoff)
 {
     int i, w, g;
-    int start = 0;
+    int start = 0, wstart = 0;
     for (w = 0; w < wi->num_windows*16; w += 16) {
+        wstart = 0;
         for (g = 0; g < num_bands; g++) {
             AacPsyBand *band = &pch->band[w+g];
 
             float form_factor = 0.0f;
             float Temp;
             band->energy = 0.0f;
-            for (i = 0; i < band_sizes[g]; i+=4) {
-                float a, b, c, d;
-                float ax, bx, cx, dx;
-                float *cf = (float *)&coefs[start+i];
-
-                __asm__ volatile (
-                    "lwc1   %[a],   0(%[cf])                \n\t"
-                    "lwc1   %[b],   4(%[cf])                \n\t"
-                    "lwc1   %[c],   8(%[cf])                \n\t"
-                    "lwc1   %[d],   12(%[cf])               \n\t"
-                    "abs.s  %[a],   %[a]                    \n\t"
-                    "abs.s  %[b],   %[b]                    \n\t"
-                    "abs.s  %[c],   %[c]                    \n\t"
-                    "abs.s  %[d],   %[d]                    \n\t"
-                    "sqrt.s %[ax],  %[a]                    \n\t"
-                    "sqrt.s %[bx],  %[b]                    \n\t"
-                    "sqrt.s %[cx],  %[c]                    \n\t"
-                    "sqrt.s %[dx],  %[d]                    \n\t"
-                    "madd.s %[e],   %[e],   %[a],   %[a]    \n\t"
-                    "madd.s %[e],   %[e],   %[b],   %[b]    \n\t"
-                    "madd.s %[e],   %[e],   %[c],   %[c]    \n\t"
-                    "madd.s %[e],   %[e],   %[d],   %[d]    \n\t"
-                    "add.s  %[f],   %[f],   %[ax]           \n\t"
-                    "add.s  %[f],   %[f],   %[bx]           \n\t"
-                    "add.s  %[f],   %[f],   %[cx]           \n\t"
-                    "add.s  %[f],   %[f],   %[dx]           \n\t"
-
-                    : [a]"=&f"(a), [b]"=&f"(b),
-                      [c]"=&f"(c), [d]"=&f"(d),
-                      [e]"+f"(band->energy), [f]"+f"(form_factor),
-                      [ax]"=&f"(ax), [bx]"=&f"(bx),
-                      [cx]"=&f"(cx), [dx]"=&f"(dx)
-                    : [cf]"r"(cf)
-                    : "memory"
-                );
+            if (wstart < cutoff) {
+                for (i = 0; i < band_sizes[g]; i+=4) {
+                    float a, b, c, d;
+                    float ax, bx, cx, dx;
+                    float *cf = (float *)&coefs[start+i];
+
+                    __asm__ volatile (
+                        "lwc1   %[a],   0(%[cf])                \n\t"
+                        "lwc1   %[b],   4(%[cf])                \n\t"
+                        "lwc1   %[c],   8(%[cf])                \n\t"
+                        "lwc1   %[d],   12(%[cf])               \n\t"
+                        "abs.s  %[a],   %[a]                    \n\t"
+                        "abs.s  %[b],   %[b]                    \n\t"
+                        "abs.s  %[c],   %[c]                    \n\t"
+                        "abs.s  %[d],   %[d]                    \n\t"
+                        "sqrt.s %[ax],  %[a]                    \n\t"
+                        "sqrt.s %[bx],  %[b]                    \n\t"
+                        "sqrt.s %[cx],  %[c]                    \n\t"
+                        "sqrt.s %[dx],  %[d]                    \n\t"
+                        "madd.s %[e],   %[e],   %[a],   %[a]    \n\t"
+                        "madd.s %[e],   %[e],   %[b],   %[b]    \n\t"
+                        "madd.s %[e],   %[e],   %[c],   %[c]    \n\t"
+                        "madd.s %[e],   %[e],   %[d],   %[d]    \n\t"
+                        "add.s  %[f],   %[f],   %[ax]           \n\t"
+                        "add.s  %[f],   %[f],   %[bx]           \n\t"
+                        "add.s  %[f],   %[f],   %[cx]           \n\t"
+                        "add.s  %[f],   %[f],   %[dx]           \n\t"
+
+                        : [a]"=&f"(a), [b]"=&f"(b),
+                          [c]"=&f"(c), [d]"=&f"(d),
+                          [e]"+f"(band->energy), [f]"+f"(form_factor),
+                          [ax]"=&f"(ax), [bx]"=&f"(bx),
+                          [cx]"=&f"(cx), [dx]"=&f"(dx)
+                        : [cf]"r"(cf)
+                        : "memory"
+                    );
+                }
             }
 
             Temp = sqrtf((float)band_sizes[g] / band->energy);
             band->thr      = band->energy * 0.001258925f;
             band->nz_lines = form_factor * sqrtf(Temp);
             start += band_sizes[g];
+            wstart += band_sizes[g];
         }
     }
 }
diff --git a/libavcodec/psymodel.c b/libavcodec/psymodel.c
index 7622124..6274a49 100644
--- a/libavcodec/psymodel.c
+++ b/libavcodec/psymodel.c
@@ -39,6 +39,7 @@ av_cold int ff_psy_init(FFPsyContext *ctx, AVCodecContext *avctx, int num_lens,
     ctx->group     = av_mallocz_array(sizeof(ctx->group[0]), num_groups);
     ctx->bands     = av_malloc_array (sizeof(ctx->bands[0]),      num_lens);
     ctx->num_bands = av_malloc_array (sizeof(ctx->num_bands[0]),  num_lens);
+    ctx->cutoff    = avctx->cutoff;
 
     if (!ctx->ch || !ctx->group || !ctx->bands || !ctx->num_bands) {
         ff_psy_end(ctx);
diff --git a/libavcodec/psymodel.h b/libavcodec/psymodel.h
index 565117d..35d184c 100644
--- a/libavcodec/psymodel.h
+++ b/libavcodec/psymodel.h
@@ -93,6 +93,7 @@ typedef struct FFPsyContext {
     FFPsyChannel      *ch;            ///< single channel information
     FFPsyChannelGroup *group;         ///< channel group information
     int num_groups;                   ///< number of channel groups
+    int cutoff;                       ///< lowpass frequency cutoff for analysis
 
     uint8_t **bands;                  ///< scalefactor band sizes for possible frame sizes
     int     *num_bands;               ///< number of scalefactor bands for possible frame sizes
diff --git a/tests/fate/aac.mak b/tests/fate/aac.mak
index 2d41888..7a6c734 100644
--- a/tests/fate/aac.mak
+++ b/tests/fate/aac.mak
@@ -146,16 +146,16 @@ fate-aac-aref-encode: CMD = enc_dec_pcm adts wav s16le $(REF) -strict -2 -c:a aa
 fate-aac-aref-encode: CMP = stddev
 fate-aac-aref-encode: REF = ./tests/data/asynth-44100-2.wav
 fate-aac-aref-encode: CMP_SHIFT = -4096
-fate-aac-aref-encode: CMP_TARGET = 1139
+fate-aac-aref-encode: CMP_TARGET = 670
 fate-aac-aref-encode: SIZE_TOLERANCE = 2464
-fate-aac-aref-encode: FUZZ = 6
+fate-aac-aref-encode: FUZZ = 89
 
 FATE_AAC_ENCODE += fate-aac-ln-encode
 fate-aac-ln-encode: CMD = enc_dec_pcm adts wav s16le $(TARGET_SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav -strict -2 -c:a aac -aac_is 0 -aac_pns 0 -aac_ms 0 -aac_tns 0 -b:a 512k
 fate-aac-ln-encode: CMP = stddev
 fate-aac-ln-encode: REF = $(SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav
 fate-aac-ln-encode: CMP_SHIFT = -4096
-fate-aac-ln-encode: CMP_TARGET = 80
+fate-aac-ln-encode: CMP_TARGET = 50
 fate-aac-ln-encode: SIZE_TOLERANCE = 3560
 fate-aac-ln-encode: FUZZ = 30
 
@@ -164,7 +164,7 @@ fate-aac-ln-encode-128k: CMD = enc_dec_pcm adts wav s16le $(TARGET_SAMPLES)/audi
 fate-aac-ln-encode-128k: CMP = stddev
 fate-aac-ln-encode-128k: REF = $(SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav
 fate-aac-ln-encode-128k: CMP_SHIFT = -4096
-fate-aac-ln-encode-128k: CMP_TARGET = 745
+fate-aac-ln-encode-128k: CMP_TARGET = 798
 fate-aac-ln-encode-128k: SIZE_TOLERANCE = 3560
 fate-aac-ln-encode-128k: FUZZ = 5
 
@@ -173,16 +173,16 @@ fate-aac-pns-encode: CMD = enc_dec_pcm adts wav s16le $(TARGET_SAMPLES)/audio-re
 fate-aac-pns-encode: CMP = stddev
 fate-aac-pns-encode: REF = $(SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav
 fate-aac-pns-encode: CMP_SHIFT = -4096
-fate-aac-pns-encode: CMP_TARGET = 695
+fate-aac-pns-encode: CMP_TARGET = 663
 fate-aac-pns-encode: SIZE_TOLERANCE = 3560
-fate-aac-pns-encode: FUZZ = 25
+fate-aac-pns-encode: FUZZ = 72
 
 FATE_AAC_ENCODE += fate-aac-tns-encode
 fate-aac-tns-encode: CMD = enc_dec_pcm adts wav s16le $(TARGET_SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav -strict -2 -c:a aac -aac_tns 1 -aac_is 0 -aac_pns 0 -aac_ms 0 -b:a 128k -cutoff 22050
 fate-aac-tns-encode: CMP = stddev
 fate-aac-tns-encode: REF = $(SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav
 fate-aac-tns-encode: CMP_SHIFT = -4096
-fate-aac-tns-encode: CMP_TARGET = 766
+fate-aac-tns-encode: CMP_TARGET = 857
 fate-aac-tns-encode: FUZZ = 6
 fate-aac-tns-encode: SIZE_TOLERANCE = 3560
 
@@ -191,25 +191,25 @@ fate-aac-is-encode: CMD = enc_dec_pcm adts wav s16le $(TARGET_SAMPLES)/audio-ref
 fate-aac-is-encode: CMP = stddev
 fate-aac-is-encode: REF = $(SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav
 fate-aac-is-encode: CMP_SHIFT = -4096
-fate-aac-is-encode: CMP_TARGET = 584
+fate-aac-is-encode: CMP_TARGET = 725
 fate-aac-is-encode: SIZE_TOLERANCE = 3560
-fate-aac-is-encode: FUZZ = 1
+fate-aac-is-encode: FUZZ = 5
 
 FATE_AAC_ENCODE += fate-aac-ms-encode
 fate-aac-ms-encode: CMD = enc_dec_pcm adts wav s16le $(TARGET_SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav -strict -2 -c:a aac -aac_pns 0 -aac_is 0 -aac_ms 1 -aac_tns 0 -b:a 128k -cutoff 22050
 fate-aac-ms-encode: CMP = stddev
 fate-aac-ms-encode: REF = $(SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav
 fate-aac-ms-encode: CMP_SHIFT = -4096
-fate-aac-ms-encode: CMP_TARGET = 615
+fate-aac-ms-encode: CMP_TARGET = 682
 fate-aac-ms-encode: SIZE_TOLERANCE = 3560
-fate-aac-ms-encode: FUZZ = 10
+fate-aac-ms-encode: FUZZ = 15
 
 FATE_AAC_ENCODE += fate-aac-ltp-encode
 fate-aac-ltp-encode: CMD = enc_dec_pcm adts wav s16le $(TARGET_SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav -strict -2 -c:a aac -profile:a aac_ltp -aac_pns 0 -aac_is 0 -aac_ms 0 -aac_tns 0 -b:a 36k -fflags +bitexact -flags +bitexact
 fate-aac-ltp-encode: CMP = stddev
 fate-aac-ltp-encode: REF = $(SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav
 fate-aac-ltp-encode: CMP_SHIFT = -4096
-fate-aac-ltp-encode: CMP_TARGET = 1120
+fate-aac-ltp-encode: CMP_TARGET = 1284
 fate-aac-ltp-encode: SIZE_TOLERANCE = 3560
 fate-aac-ltp-encode: FUZZ = 17
 
@@ -218,7 +218,7 @@ fate-aac-pred-encode: CMD = enc_dec_pcm adts wav s16le $(TARGET_SAMPLES)/audio-r
 fate-aac-pred-encode: CMP = stddev
 fate-aac-pred-encode: REF = $(SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav
 fate-aac-pred-encode: CMP_SHIFT = -4096
-fate-aac-pred-encode: CMP_TARGET = 790
+fate-aac-pred-encode: CMP_TARGET = 835
 fate-aac-pred-encode: FUZZ = 12
 fate-aac-pred-encode: SIZE_TOLERANCE = 3560