Commit 1003be3c6d for ffmpeg
commit 1003be3c6dc0de5e6703aed968943ec08a1fae0a
Author: Lynne <dev@lynne.ee>
Date: Sun Nov 8 02:19:58 2026 +0800
avcodec/aaccoder_nmr: keep the scalefactor chain legal
A coded band's candidates only reach up to the scalefactor at which it
quantizes to zero, so two neighbours in the coding sequence more than
SCALE_MAX_DIFF apart leave the trellis without a legal path: every later
state stays at FLT_MAX, the backtrack returns the finest candidates, and
the commit clamps loud bands below their finest legal scalefactor into
an oversized frame.
Walk the coding sequence tracking the scalefactors still reachable from
its start, and drop the quieter band of a pair that cannot connect;
within reach it would quantize to zero anyway. Run it on the coarse grid
after setup (the fine pass refines around the coarse solve, so it stays
legal), and again whenever bands leave the sequence: in the decoder
buffer shed loop, where a dropped band may have bridged its neighbours,
and when bands turn into noise.
The psy model no longer codes such bands in the first place; this is a
no-op on every legal chain and bit-exact on normal content.
diff --git a/libavcodec/aaccoder_nmr.h b/libavcodec/aaccoder_nmr.h
index fd7d6e1cbc..c9fc6646e0 100644
--- a/libavcodec/aaccoder_nmr.h
+++ b/libavcodec/aaccoder_nmr.h
@@ -291,6 +291,63 @@ static void nmr_bail_channel(SingleChannelElement *sce)
}
}
+/* Zero coded band b, subwindow flags included (the group flag is re-derived
+ * by ANDing). */
+static void nmr_drop_band(NMRSlot *t, int b)
+{
+ for (int w2 = 0; w2 < t->sce->ics.group_len[t->bw[b]]; w2++)
+ t->sce->zeroes[(t->bw[b] + w2)*16 + t->bg[b]] = 1;
+}
+
+/* Keep the coding sequence act[] scalefactor-chain legal on grid `step`.
+ * A band's candidates only reach up to where it quantizes to zero, so two
+ * neighbours more than SCALE_MAX_DIFF apart leave the trellis without any
+ * legal path: every later state stays FLT_MAX and the backtrack returns the
+ * finest candidates, an oversized frame. Walk the chain tracking the
+ * scalefactor range still reachable from its start, and drop the quieter
+ * band of a pair that cannot connect: within reach it would quantize to
+ * zero anyway. A no-op on a legal chain. */
+static void nmr_chain_repair(NMRSlot *t, int step)
+{
+ int rlo[128], rhi[128], n = 0;
+
+ for (int k = 0; k < t->nact; k++) {
+ const int b = t->act[k];
+ const int lo = t->blo[b], hi = lo + (t->bnc[b] - 1)*step;
+ int clo = lo, chi = hi, keep = 1;
+ while (n) {
+ const int p = t->act[n - 1];
+ clo = FFMAX(lo, rlo[n - 1] - SCALE_MAX_DIFF);
+ chi = FFMIN(hi, rhi[n - 1] + SCALE_MAX_DIFF);
+ clo = lo + (clo - lo + step - 1) / step * step;
+ if (clo <= chi) {
+ chi = lo + (chi - lo) / step * step;
+ break;
+ }
+ if (t->maxvals[t->bidx[p]] > t->maxvals[t->bidx[b]]) {
+ keep = 0;
+ break;
+ }
+ nmr_drop_band(t, p);
+ n--;
+ }
+ if (!keep) {
+ nmr_drop_band(t, b);
+ continue;
+ }
+ if (!n) {
+ clo = lo;
+ chi = hi;
+ }
+ /* n <= k: compacting in place never overwrites an unread entry */
+ t->act[n] = b;
+ rlo[n] = clo;
+ rhi[n] = chi;
+ n++;
+ }
+ t->nact = n;
+}
+
/* Per-channel setup into slot t: short-block threshold shaping, the
* allocation law, zero decisions, and the PASS 1 coarse candidate curves.
* Returns the coded-band count; 0 = nothing codeable (caller bails). */
@@ -546,6 +603,9 @@ static int nmr_setup_channel(AVCodecContext *avctx, AACEncContext *s,
t->is_pns[b] = 0;
}
t->nact = nbnd;
+ /* the fine pass refines around the coarse solve, so a chain legal on
+ * the coarse grid stays legal there */
+ nmr_chain_repair(t, cstep);
return nbnd;
}
@@ -1100,19 +1160,17 @@ static void nmr_solve_group(AVCodecContext *avctx, AACEncContext *s,
int dropped = 0;
for (int k = 0; k < nsl; k++) {
NMRSlot *t = sl[k];
- SingleChannelElement *sce = t->sce;
- int hi = 1, b, w0, g;
+ int hi = 1;
if (t->nact <= 1)
continue;
for (int b_ = 2; b_ < t->nact; b_++)
if (t->bg[t->act[b_]] >= t->bg[t->act[hi]])
hi = b_;
- b = t->act[hi];
- w0 = t->bw[b]; g = t->bg[b];
- for (int w2 = 0; w2 < sce->ics.group_len[w0]; w2++)
- sce->zeroes[(w0+w2)*16+g] = 1;
+ nmr_drop_band(t, t->act[hi]);
memmove(&t->act[hi], &t->act[hi + 1], (t->nact - hi - 1) * sizeof(t->act[0]));
t->nact--;
+ /* the dropped band may have bridged its neighbours */
+ nmr_chain_repair(t, NMR_STEP);
dropped = 1;
}
if (!dropped)
@@ -1316,6 +1374,8 @@ static void nmr_solve_group(AVCodecContext *avctx, AACEncContext *s,
if (!t->is_pns[t->act[b_]])
t->act[n++] = t->act[b_];
t->nact = n;
+ /* noise bands leave the scalefactor chain */
+ nmr_chain_repair(t, NMR_STEP);
}
pns_total += pns_count;
}