H.264: bound decoder work from overlapping slices, gaps and queued output

- A macroblock that an earlier slice of the same picture already decoded is
  refused, so overlapping slices cannot redo the picture's reconstruction
  once per slice.
- The memory budget check counts pictures waiting in the output queue that
  have left the DPB (one decode call may carry many access units) and the
  DPB's actual size, counting shared buffers once.
- frame_num gap concealment advances the POC state across the skipped run in
  constant time instead of once per missing frame_num.
- Storing a reference picture evicts until the DPB is under capacity, rather
  than evicting one picture and inserting regardless.

crates/h264/src/cabac_mb.rs

@@ -48,7 +48,7 @@ impl SliceDecoder<'_> {
         let intra_slice = self.sh.slice_type.is_intra();
         loop {
             ensure!(addr < total, "macroblock address {addr} out of range");
-            self.start_mb(addr);
+            self.start_mb(addr)?;
             let skip = !intra_slice && self.decode_skip_flag(&mut c);
             if skip {
                 self.decode_skip()?;

crates/h264/src/decoder.rs

@@ -451,7 +451,11 @@ impl Decoder {
             chroma_planes: if chroma == 0 { 0 } else { 2 },
             bytes_per_sample: 1,
         };
-        let pictures = (sps.max_dpb_frames() + 1 + self.max_in_flight) as u64;
+        // Pictures waiting for output that already left the DPB are held too (one decode call may
+        // carry many access units). A buffer shared with a DPB entry is counted once.
+        let queued = self.out_queue.iter().filter(|o| !self.dpb.entries.iter().any(|e| Arc::ptr_eq(&e.frame, &o.frame))).count();
+        let dpb = sps.max_dpb_frames().max(self.dpb.entries.len());
+        let pictures = (dpb + 1 + self.max_in_flight.max(queued)) as u64;
         ensure!(
             shape.fits(pictures, self.memory_budget),
             "{}x{} pictures with {pictures} in the decoder exceed the memory budget",

crates/h264/src/dpb.rs

@@ -139,8 +139,9 @@ impl Dpb {
         let mut unused = (prev_ref_frame_num + 1) % max_frame_num;
         // Only the last `max_num_ref_frames` "non-existing" frames can survive the sliding window, and
         // inserting more than that evicts every earlier short-term reference. Skip straight to that tail
-        // (keeping the POC state walk for every missing frame_num) so a hostile gap of up to 65535
-        // frames costs at most `max_num_ref_frames` frame allocations instead of one per missing frame.
+        // (advancing the POC state across the skipped run in constant time) so a hostile gap of up to
+        // 65535 frames costs at most `max_num_ref_frames` frame allocations and POC updates instead of
+        // one per missing frame.
         let gap = frame_num.wrapping_add(max_frame_num).wrapping_sub(unused) % max_frame_num;
         let keep = u32::try_from(max_num_ref_frames.max(1)).unwrap_or(u32::MAX);
         let skip = gap.saturating_sub(keep);
@@ -149,10 +150,17 @@ impl Dpb {
                 e.mark = RefMark::Unused;
             }
             self.prune();
-            for _ in 0..skip {
-                poc_state(unused);
-                unused = (unused + 1) % max_frame_num;
+            // The POC state only advances FrameNumOffset where frame_num decreases, so walking the
+            // skipped run is the same as visiting its first value, the wrap (if the run crosses
+            // it) and its last value: constant work however long the gap.
+            let last = (unused + (skip - 1) % max_frame_num) % max_frame_num;
+            poc_state(unused);
+            if last < unused {
+                poc_state(max_frame_num - 1);
+                poc_state(0);
             }
+            poc_state(last);
+            unused = (last + 1) % max_frame_num;
         }
         let mut guard = 0;
         while unused != frame_num && guard < max_frame_num {
@@ -438,14 +446,15 @@ impl Dpb {
                 }
             }
         } else {
-            while self.entries.len() >= self.capacity {
+            // Evict until there is room (frame_num gap concealment can leave the DPB over-full), so
+            // the store never holds more pictures than the memory budget was checked for.
+            while self.entries.len() >= self.capacity.max(1) {
                 if !self.bump(out) {
                     if let Some(pos) = self.entries.iter().position(|e| !e.needed_for_output) {
                         self.entries.remove(pos);
                     } else {
                         self.entries.remove(0);
                     }
-                    break;
                 }
             }
         }
@@ -546,6 +555,37 @@ mod tests {
         (made, walk, left)
     }
 
+    /// (FrameNumOffset, prev frame_num) after `walk`, by the rule of `PocState::update_gap_frame`.
+    fn poc_after(walk: &[u32], prev: u32, max: u32) -> (u64, u32) {
+        walk.iter().fold((0, prev), |(off, prev), &f| (if prev > f { off + u64::from(max) } else { off }, f))
+    }
+
+    #[test]
+    fn gap_walk_matches_every_frame_num_for_any_skip() {
+        for (prev, cur) in [(1, 9), (10, 3), (14, 0), (15, 14), (0, 15), (7, 6)] {
+            for refs in 1..4 {
+                let mut dpb = Dpb::new();
+                let (_, walk, _) = fill(&mut dpb, prev, cur, 16, refs);
+                let every: Vec<u32> = (1..16).map(|i| (prev + i) % 16).take_while(|&f| f != cur).collect();
+                assert_eq!(poc_after(&walk, prev, 16), poc_after(&every, prev, 16), "{prev} -> {cur}, {refs} refs");
+                assert_eq!(walk.last(), every.last(), "{prev} -> {cur}, {refs} refs");
+            }
+        }
+    }
+
+    #[test]
+    fn reference_store_evicts_down_to_capacity() {
+        let mut dpb = Dpb::new();
+        dpb.capacity = 2;
+        // Gap concealment can leave more non-output references than the capacity.
+        for i in 0..5 {
+            short(&mut dpb, i, i);
+        }
+        let mut out = Vec::new();
+        assert!(!dpb.bump_for_space(RefMark::Short, 0, &mut out));
+        assert!(dpb.entries.len() < 2, "{} pictures left before storing into a 2-picture DPB", dpb.entries.len());
+    }
+
     #[test]
     fn huge_gap_makes_only_the_frames_that_stay_referenced() {
         let mut dpb = Dpb::new();
@@ -553,8 +593,9 @@ mod tests {
         short(&mut dpb, 2, 1);
         let (made, walk, left) = fill(&mut dpb, 1, 65535, 65536, 4);
         assert_eq!(made, 4);
-        // The POC state still walks every missing frame_num, in order.
-        assert_eq!(walk.len(), 65533);
+        // The POC state ends where walking every missing frame_num would leave it, in a few steps.
+        assert_eq!(poc_after(&walk, 1, 65536), poc_after(&(2..65535).collect::<Vec<_>>(), 1, 65536));
+        assert!(walk.len() <= 8, "{} POC updates", walk.len());
         assert_eq!(walk.first(), Some(&2));
         assert_eq!(walk.last(), Some(&65534));
         assert_eq!(left, vec![(65531, true), (65532, true), (65533, true), (65534, true)]);
@@ -566,7 +607,9 @@ mod tests {
         short(&mut dpb, 1, 65000);
         let (made, walk, left) = fill(&mut dpb, 65000, 3, 65536, 2);
         assert_eq!(made, 2);
-        assert_eq!(walk.len(), 538);
+        let every: Vec<u32> = (65001..65536).chain(0..3).collect();
+        assert_eq!(poc_after(&walk, 65000, 65536), poc_after(&every, 65000, 65536));
+        assert!(walk.len() <= 6, "{} POC updates", walk.len());
         assert_eq!(walk.first(), Some(&65001));
         assert_eq!(walk.last(), Some(&2));
         assert_eq!(left, vec![(1, true), (2, true)]);

crates/h264/src/slicedec.rs

@@ -346,8 +346,11 @@ impl<'a> SliceDecoder<'a> {
     // Neighbour helpers
     // ------------------------------------------------------------------------------------------
 
-    /// Set up the current macroblock address and neighbour availability.
-    pub fn start_mb(&mut self, addr: usize) {
+    /// Set up the current macroblock address and neighbour availability. A macroblock an earlier
+    /// slice of this picture already decoded is refused: overlapping slices would otherwise redo
+    /// the picture's reconstruction once per slice.
+    pub fn start_mb(&mut self, addr: usize) -> Result<()> {
+        ensure!(self.pic.mbs[addr].slice_num == u32::MAX, "macroblock {addr} overlaps an earlier slice");
         self.mb_addr = addr;
         self.mb_x = addr % self.mb_w;
         self.mb_y = addr / self.mb_w;
@@ -364,6 +367,7 @@ impl<'a> SliceDecoder<'a> {
         ];
         self.mv_done = 0;
         self.pic.mbs[addr] = MbState { slice_num: sn, ..MbState::default() };
+        Ok(())
     }
 
     #[inline]
@@ -1467,7 +1471,7 @@ impl<'a> SliceDecoder<'a> {
                 let run = r.read_ue()? as usize;
                 ensure!(addr + run <= total, "mb_skip_run out of range");
                 for _ in 0..run {
-                    self.start_mb(addr);
+                    self.start_mb(addr)?;
                     self.decode_skip()?;
                     self.pic.mb_done(addr);
                     addr += 1;
@@ -1477,7 +1481,7 @@ impl<'a> SliceDecoder<'a> {
                 }
                 ensure!(addr < total, "macroblock address {addr} out of range");
             }
-            self.start_mb(addr);
+            self.start_mb(addr)?;
             self.decode_mb_cavlc(r)?;
             self.pic.mb_done(addr);
             addr += 1;

crates/h264/src/synth_tests.rs

@@ -113,6 +113,7 @@ fn avcc_rejects_a_valid_picture_slice_without_starting_a_picture() {
             poc_lsb: Some(0),
             poc_delta: None,
             wt: None,
+            first_mb: 0,
         },
         false,
     );
@@ -153,6 +154,8 @@ struct Pic {
     /// Explicit weights (luma, cb, cr) as (weight, offset) for refIdx 0 of list 0 and list 1;
     /// log2 denominators are 5 (luma) and 3 (chroma).
     wt: Option<[[(i32, i32); 3]; 2]>,
+    /// first_mb_in_slice (`mbs` then lists the slice's macroblocks from there).
+    first_mb: u32,
 }
 
 impl Pic {
@@ -165,7 +168,7 @@ impl Pic {
 }
 
 fn slice_header(w: &mut BitWriter, p: &Pic, cabac: bool) {
-    w.write_ue(0); // first_mb
+    w.write_ue(p.first_mb);
     w.write_ue(if p.b {
         6
     } else if p.is_p() {
@@ -660,3 +663,55 @@ fn frame_num_gap_is_filled_when_the_sps_allows_gaps() {
     assert_eq!(out.len(), 2);
     assert_eq!(dec.stats().frame_num_gaps, 1);
 }
+
+/// An IDR picture followed by `pictures` reference P pictures of P_Skip macroblocks.
+fn skip_stream(pictures: u32) -> (Vec<u8>, Vec<Vec<u8>>) {
+    let mut seed = 11u32;
+    let a: Vec<Option<Vec<u8>>> = (0..NMB).map(|_| Some(random_mb(&mut seed))).collect();
+    let mut head = sps(66, 2, false, PocType::T2);
+    head.extend(pps(false, false));
+    head.extend(slice_nal(&Pic { idr: true, frame_num: 0, ref_idc: 3, mbs: a, ..Default::default() }, false));
+    let ps = (1..=pictures).map(|k| slice_nal(&Pic { frame_num: k % 16, ref_idc: 2, mbs: vec![None; NMB], ..Default::default() }, false)).collect();
+    (head, ps)
+}
+
+#[test]
+fn overlapping_slices_are_not_reconstructed_twice() {
+    let (head, ps) = skip_stream(1);
+    let mut once = Decoder::with_threads(1);
+    let mut out = once.decode(&[head.clone(), ps[0].clone()].concat(), 0).unwrap();
+    out.extend(once.flush());
+    assert!(once.take_error().is_none());
+    assert_eq!(out.len(), 2);
+    // A second slice of the same picture starting at macroblock 1 covers those macroblocks again.
+    let again = slice_nal(&Pic { frame_num: 1, ref_idc: 2, mbs: vec![None; NMB - 1], first_mb: 1, ..Default::default() }, false);
+    let mut twice = Decoder::with_threads(1);
+    let result = twice.decode(&[head, ps[0].clone(), again].concat(), 0);
+    twice.flush();
+    let err = result.err().or_else(|| twice.take_error()).expect("overlapping slice refused");
+    assert!(err.to_string().contains("overlaps"), "{err}");
+}
+
+#[test]
+fn output_queue_counts_against_the_memory_budget() {
+    let (head, ps) = skip_stream(40);
+    let sps_nal = sps(66, 2, false, PocType::T2);
+    let parsed = crate::params::Sps::parse(&deckcraft_bitstream::unescape_rbsp(&sps_nal[5..])).unwrap();
+    let one = (NMB * 256 + 2 * NMB * 64) as u64;
+    let budget = (parsed.max_dpb_frames() + 1 + 3) as u64 * one;
+    // One access unit per call: output drains as it goes, and the stream fits.
+    let mut dec = Decoder::with_threads(1);
+    dec.set_memory_budget(budget);
+    let mut out = dec.decode(&head, 0).unwrap();
+    for (i, p) in ps.iter().enumerate() {
+        out.extend(dec.decode(p, i as i64 + 1).unwrap());
+    }
+    out.extend(dec.flush());
+    assert!(dec.take_error().is_none());
+    assert_eq!(out.len(), 41);
+    // Every access unit in one call: the pictures held for output stop fitting the budget.
+    let mut dec = Decoder::with_threads(1);
+    dec.set_memory_budget(budget);
+    let err = dec.decode(&[head, ps.concat()].concat(), 0).expect_err("output queue is over budget");
+    assert!(err.to_string().contains("memory budget"), "{err}");
+}