Why: under DLSS Frame Generation the Gossip Stones in Hyrule Town broke apart in made frames, each one drawn part way
to its neighbour (a player's report, RTX 5080, 4x; reproduced on an RTX 5090, 3x). Their actor motion came from
another stone.

- The pose (bones and origin) of each skinned draw is read when the draw is recorded (ub_read), not copied from the
  game's buffers after the frame, when an inline uniform update may have changed them. The status line counts the
  poses read at the draw and the poses copied later.
- Copies of one model paired with the frame before's copies by draw order (the n-th draw of the model). The game
  draws the copies in another order from frame to frame, the stones with two models, and a copy that comes into
  view has no frame before. Copies now pair by place: the origin plus the bones' mean translation (the copies share
  one origin on a grid of 10 units and are placed by their bones), all copies of a model at once, the nearest pairs
  first, never further apart than RX_VK_DLSS_ACTOR_REACH world units in one frame (default 1; a running Link moves
  about 0.1 a frame, the stones stand 2.3 apart). On a tie at one place the same n-th pairs first. A draw without a
  pose read at the draw pairs by its n-th, as before.

dlss_actor::place_of and dlss_actor::pair_by_place are checked in tests/test_dlss_actor.cpp.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

@@ -14,6 +14,7 @@ otherwise.
 
 | Date | Build | OS / desktop | GPU / driver | CPU / RAM | Settings | Where | Result |
 |---|---|---|---|---|---|---|---|
+| 2026-11-12 | v2.1.0-rc.3-281-g06ad86fe (source 8d4d275a): main ba3d3f73 + this change (before the review's small fixes), #2787's layouts and the close-up depth test; against v2.1.0-rc.3-271-gba3d3f73, main | Windows 11 Pro, 3840x2160 output | RTX 5090 32 GB, 621.24 | Ryzen 7 9800X3D | DLAA, frame generation 3x by DLSS (NGX), Reflex On + Boost, V-Sync Auto (FIFO) | Castle Town, the four Gossip Stones by the Temple of Time: running past them while the camera turns, about 30 s | Main: the stones break apart in made frames (a stone drawn part way to its neighbour); with `RX_VK_DLSS_ACTORS=0` they stay whole. This change: whole. Every pose read at the draw ("poses N read at the draw, 0 copied later"), 51-74 skinned draws a frame, actor motion 0.15-0.28 ms of CPU a frame, DLSS-G made 590 of 600 presents in play. No device lost |

@@ -24,0 +59,4 @@
+    std::vector<int> pairs(now.size(), -1);
+    std::vector<char> taken(then.size(), 0);
+    for (const Candidate& c : candidates)
+        if (pairs[c.i] < 0 && !taken[c.j]) pairs[c.i] = c.j, taken[c.j] = 1;

@@ -522,0 +544,4 @@
+    // first, never across more than RX_VK_DLSS_ACTOR_REACH world units between two engine frames (default 1: a running Link
+    // moves about 0.1 a frame, the Gossip Stones stand 2.3 apart; after a long frame a fast actor can go further and keeps the
+    // camera's motion for that frame): dlss_actor::pair_by_place. Else the same n-th draw of the model as before)
+    static const float reach = [] {

@@ -541,0 +620,4 @@
+        // draws too, else its n-th, as before)
+        auto it = old.end();
+        const auto m = placed ? models.find(k.key >> 8) : models.end();
+        if (m == models.end() || m->second.then.empty()) it = old.find(k.key);

@@ -815,0 +907,4 @@
+    // frame's or another draw's (the place pair_by_place needs, too). RX_VK_DLSS_ACTOR_LIVE=1: as before, for an A/B)
+    static const bool live = env_switch("RX_VK_DLSS_ACTOR_LIVE") == 1;
+    if (!live) {
+        d.pose.resize(16 + (size_t)d.nbones * 48);

@@ -14,6 +14,7 @@ otherwise.
 
 | Date | Build | OS / desktop | GPU / driver | CPU / RAM | Settings | Where | Result |
 |---|---|---|---|---|---|---|---|
+| 2026-11-12 | v2.1.0-rc.3-281-g06ad86fe (source 8d4d275a): main ba3d3f73 + this change (before the review's small fixes), [NotRexed/Hookshot PR 2787](https://github.com/NotRexed/Hookshot/pull/2787)'s layouts and the close-up depth test; against v2.1.0-rc.3-271-gba3d3f73, main | Windows 11 Pro, 3840x2160 output | RTX 5090 32 GB, 621.24 | Ryzen 7 9800X3D | DLAA, frame generation 3x by DLSS (NGX), Reflex On + Boost, V-Sync Auto (FIFO) | Castle Town, the four Gossip Stones by the Temple of Time: running past them while the camera turns, about 30 s | Main: the stones break apart in made frames (a stone drawn part way to its neighbour); with `RX_VK_DLSS_ACTORS=0` they stay whole. This change: whole. Every pose read at the draw ("poses N read at the draw, 0 copied later"), 51-74 skinned draws a frame, actor motion 0.15-0.28 ms of CPU a frame, DLSS-G made 590 of 600 presents in play. No device lost |

@@ -24,0 +88,4 @@
+        if (used[j] || !(d <= (double)reach)) return pairs;   // (not within the reach: nearest first; NaN fails here too)
+        by_order[i] = j, used[j] = 1, order_n++, order_total += d;
+    }
+    if (order_n >= nearest_n && order_total <= nearest_total + 1e-6) return by_order;

@@ -828,28 +945,52 @@ void temporal_note_skinned_draw(const SkinnedDraw& k) {
     static const bool off = env_switch("RX_VK_DLSS_ACTORS") == 0;
     if (off || !temporal_wanted() || !k.depth || !k.bones || k.bones_size < 48 || !k.object || k.object_size < 0x4c) return;
     State& s = st();
+    const std::chrono::steady_clock::time_point t0 = std::chrono::steady_clock::now();

@@ -831,0 +965,4 @@
+    // frame's or another draw's (the place pair_by_place needs, too). RX_VK_DLSS_ACTOR_LIVE=1: as before, for an A/B.
+    // Read without the lock, into this thread's own buffer (its room kept from draw to draw); the lock is taken only to add it
+    // to the frame's poses. ub_read is this recording thread's view of the stream, as the draw's own reads of its uniforms.
+    // (An assumption, open with the author: OoT writes its bone palettes and object blocks on the CPU only. ub_read sees the

@@ -14,6 +14,7 @@ otherwise.
 
 | Date | Build | OS / desktop | GPU / driver | CPU / RAM | Settings | Where | Result |
 |---|---|---|---|---|---|---|---|
+| 2026-11-12 | v2.1.0-rc.3-281-g06ad86fe (source 8d4d275a): main ba3d3f73 + this change (before the review's small fixes), [NotRexed/Hookshot PR 2787](https://github.com/NotRexed/Hookshot/pull/2787)'s layouts and the close-up depth test; against v2.1.0-rc.3-271-gba3d3f73, main | Windows 11 Pro, 3840x2160 output | RTX 5090 32 GB, 621.24 | Ryzen 7 9800X3D | DLAA, frame generation 3x by DLSS (NGX), Reflex On + Boost, V-Sync Auto (FIFO) | Castle Town, the four Gossip Stones by the Temple of Time: running past them while the camera turns, about 30 s | Main: the stones break apart in made frames (a stone drawn part way to its neighbour); with `RX_VK_DLSS_ACTORS=0` they stay whole. This change: whole. Every pose read at the draw ("poses N read at the draw, 0 copied later"), 51-74 skinned draws a frame, actor motion 0.15-0.28 ms of CPU a frame, DLSS-G made 590 of 600 presents in play. No device lost |

@@ -24,0 +90,4 @@
+        if (used[j] || !(d <= (double)reach)) return pairs;   // (not within the reach: nearest first; NaN fails here too)
+        by_order[i] = j, used[j] = 1, order_n++, order_total += d;
+    }
+    if (order_n > nearest_n || (order_n == nearest_n && order_total <= nearest_total + 1e-6)) return by_order;

@@ -543,0 +578,4 @@
+    // copy then is the next stone, 2.3 away), so it must not refuse an actor that really moved: as a distance a frame, a
+    // hitch or a low base frame rate (frame generation from 30 fps) took actors past it and they lost their motion for that
+    // frame, ghosting the ones that move most. 60 a second is 1 a frame at 60 fps, and some ten times a running Link (about
+    // 6 a second); it stays under the stones' 2.3 down to a base of 26 fps, and below that a lone stone can take its

@@ -832,0 +974,4 @@
+    // frame's or another draw's (the place pair_by_place needs, too). RX_VK_DLSS_ACTOR_LIVE=1: as before, for an A/B.
+    // Read without the lock, into this thread's own buffer (its room kept from draw to draw); the lock is taken only to add it
+    // to the frame's poses. ub_read is this recording thread's view of the stream, as the draw's own reads of its uniforms.
+    // (An assumption, open with the author: OoT writes its bone palettes and object blocks on the CPU only. ub_read sees the