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

1 files changed, 93 insertions, 43 deletions

diff --git a/libavcodec/aaccoder_twoloop.h b/libavcodec/aaccoder_twoloop.h
index d4290b4023..397a4db5e5 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;