Game: DLSS/DLAA ghosting with the camera inside Link, a second case: skin_check() misses his largest mesh (follow-up to #1370, #1447) #1529

Open
opened 2026-11-18 02:15:37 +00:00 by magicbeanenjoyerguyperson · 1 comment

Longshot version

v3.0.0-rc.7-own (source f3ef4d06): my own build of the v3.0.0-rc.7 tag with the first patch from #1370

Channel

Built from source

What kind of problem

Graphics (missing, black or wrong-coloured things, lighting, flicker)

What happened

Follow-up to #1370 (closed) and pull request #1447 (merged). There is one more case, and I think #1447 has it too.

What I saw. Crawling through a tunnel with DLAA, a patch in the middle of the picture smears into streaks while the camera moves. Everything around it stays sharp. It shows with frame generation on and off, only in motion, and most when crawling into the small room. First clip below.

What I expected. No smear, as in the other crawl tunnel and at the vine walls since the fix for #1370.

Steps. DLAA or DLSS. Go to the spot under "Where in the game" and crawl through the tunnel towards the room with the grate.

The cause, as far as I could track it. One of Link's meshes is never treated as a skinned draw, so the hidden-mesh rule never follows it, and its depth stays in the scene while the camera is inside him.

  • skin_check() in gpu/vk/vk_draw.cpp only takes a draw when the bone palette is in vertex-stage binding 3: const nvn_p_bind_buf& bones = s.ubo[0][3]; if (!bones.addr || bones.size < 48 || bones.size % 48) return;
  • Link's largest mesh (31,608 indices; depth program e56b2a466be24f0f, colour program d4474e51baeb2570) has a block of 16,960 bytes in binding 3 and the palette in binding 4. It is the same 280-bone buffer his other meshes have in binding 3.
  • With the camera inside Link the game still makes 32 draws of him with his depth-pass programs, late in the frame. 31 pass skin_check(). This one does not.
  • One mesh of another character has the same two bindings (programs a6d37b703547222a and bd7da7a670ae8f68, a 141-bone palette in binding 4).
  • Because of the same test, these meshes get no actor motion either.
  • Related: #1449 (motion vectors for content that moves without bones). That issue counts skinned actors as covered. These are skinned draws that skin_check() does not take, so they have no actor motion. It may belong in its coverage inventory.

How I found it. With RX_VK_OCC_QUERIES=0 the rule from #1370 left out all 30 meshes it follows in every frame, and the patch was still on screen. Then RX_VK_DLSS_TRACE=300 listed the frame's draws. The attached longshot_log-second-tunnel-important-lines.txt has those lines and says how to read them.

What I have not tested. This is on my own build of the v3.0.0-rc.7 tag with my original patch from #1370, not on main or a nightly. The description of #1447 says its check sits right after skin_check(), so I expect main to miss the same mesh, but I have not run it. Someone with a nightly can check it in a minute at the spot.

What fixed it on my build. For the hidden-mesh rule only, a draw counts as skinned by its vertex layout (the joint and weight attributes skin_check() already looks at), wherever the palette is bound. The actor motion gets exactly the draws it got before. With that, 31 depth draws a frame are left out where it was 30, and the patch is gone (second clip). Attached:

  • dlss-hidden-body-depth-v3.0.0-rc.7-v2.patch: the whole change as a patch against v3.0.0-rc.7 (the second form of the patch on #1370).
  • skin_check-the-three-places.txt: the three changed places in skin_check() by themselves, for reading.

You are welcome to use it under the project's licence, as with the first patch. No Co-Authored-By line needed.

There could be other cases like this one. I don't expect many, if any.

Found and patched with an AI assistant (Claude). I built and ran everything myself, on one PC.

Where in the game

Hyrule Castle: the crawl tunnel at the upper end of the moat on the keep's east side (by the ladder), crawling into the small room with the grate

How often

Every time

Graphics card and driver version

RTX 5090, NVIDIA 621.14

CPU, RAM and operating system

Ryzen 9 9950X3D, 64 GB, Windows 11

Changes from the defaults

Upscaler: DLAA at 3840x2160. Frame generation on in one run and off in the others: it shows both ways.

  • This is my own build from source (the v3.0.0-rc.7 tag) with a few changes of my own: button mapping and prompt options, hold to quit, a hold at 60 fps for the game's scenes, and the patch from #1370. Only the last one touches rendering.
  • longshot_env.txt: HOS_DISCORD=0 and HOS_CRASH_WATCH=0. For the two test runs one more line each: RX_VK_OCC_QUERIES=0, then RX_VK_DLSS_TRACE=300.
  • no ReShade, shader-cache tool or injector, clips recorded with obs, xbox game bar enabled

longshot_log.txt

In the two whole logs my Windows account name is replaced by "account" in four paths (C:\Users\account...). Nothing else in them is changed.

  • longshot_log-second-tunnel-important-lines.txt: start here. The lines that matter out of four runs, with a short note on each.
  • longshot_log-rc7-second-tunnel-draw-list.txt: the whole log of the run with the problem, with RX_VK_DLSS_TRACE=300.
  • longshot_log-rc7-second-tunnel-fixed.txt: the whole log of the run with the second form of the patch.
  • dlss-hidden-body-depth-v3.0.0-rc.7-v2.patch and skin_check-the-three-places.txt: the change.

Screenshots or video

First clip: the problem, on the build with the first patch from #1370. Second clip: the same crawl with the second form of the patch.

Before you send it

  • The problem is still there on the newest release or nightly.
  • I searched the open issues for it.
  • I'm not attaching game files, keys, the game's executable, its .bin caches or saves.
### Longshot version v3.0.0-rc.7-own (source f3ef4d06): my own build of the v3.0.0-rc.7 tag with the first patch from #1370 ### Channel Built from source ### What kind of problem Graphics (missing, black or wrong-coloured things, lighting, flicker) ### What happened Follow-up to #1370 (closed) and pull request #1447 (merged). There is one more case, and I think #1447 has it too. **What I saw.** Crawling through a tunnel with DLAA, a patch in the middle of the picture smears into streaks while the camera moves. Everything around it stays sharp. It shows with frame generation on and off, only in motion, and most when crawling into the small room. First clip below. **What I expected.** No smear, as in the other crawl tunnel and at the vine walls since the fix for #1370. **Steps.** DLAA or DLSS. Go to the spot under "Where in the game" and crawl through the tunnel towards the room with the grate. **The cause, as far as I could track it.** One of Link's meshes is never treated as a skinned draw, so the hidden-mesh rule never follows it, and its depth stays in the scene while the camera is inside him. - `skin_check()` in `gpu/vk/vk_draw.cpp` only takes a draw when the bone palette is in vertex-stage binding 3: `const nvn_p_bind_buf& bones = s.ubo[0][3]; if (!bones.addr || bones.size < 48 || bones.size % 48) return;` - Link's largest mesh (31,608 indices; depth program `e56b2a466be24f0f`, colour program `d4474e51baeb2570`) has a block of 16,960 bytes in binding 3 and the palette in binding 4. It is the same 280-bone buffer his other meshes have in binding 3. - With the camera inside Link the game still makes 32 draws of him with his depth-pass programs, late in the frame. 31 pass `skin_check()`. This one does not. - One mesh of another character has the same two bindings (programs `a6d37b703547222a` and `bd7da7a670ae8f68`, a 141-bone palette in binding 4). - Because of the same test, these meshes get no actor motion either. - Related: #1449 (motion vectors for content that moves without bones). That issue counts skinned actors as covered. These are skinned draws that skin_check() does not take, so they have no actor motion. It may belong in its coverage inventory. **How I found it.** With `RX_VK_OCC_QUERIES=0` the rule from #1370 left out all 30 meshes it follows in every frame, and the patch was still on screen. Then `RX_VK_DLSS_TRACE=300` listed the frame's draws. The attached `longshot_log-second-tunnel-important-lines.txt` has those lines and says how to read them. **What I have not tested.** This is on my own build of the v3.0.0-rc.7 tag with my original patch from #1370, not on main or a nightly. The description of #1447 says its check sits right after `skin_check()`, so I expect main to miss the same mesh, but I have not run it. Someone with a nightly can check it in a minute at the spot. **What fixed it on my build.** For the hidden-mesh rule only, a draw counts as skinned by its vertex layout (the joint and weight attributes `skin_check()` already looks at), wherever the palette is bound. The actor motion gets exactly the draws it got before. With that, 31 depth draws a frame are left out where it was 30, and the patch is gone (second clip). Attached: - `dlss-hidden-body-depth-v3.0.0-rc.7-v2.patch`: the whole change as a patch against v3.0.0-rc.7 (the second form of the patch on #1370). - `skin_check-the-three-places.txt`: the three changed places in `skin_check()` by themselves, for reading. You are welcome to use it under the project's licence, as with the first patch. No Co-Authored-By line needed. There could be other cases like this one. I don't expect many, if any. Found and patched with an AI assistant (Claude). I built and ran everything myself, on one PC. ### Where in the game Hyrule Castle: the crawl tunnel at the upper end of the moat on the keep's east side (by the ladder), crawling into the small room with the grate ### How often Every time ### Graphics card and driver version RTX 5090, NVIDIA 621.14 ### CPU, RAM and operating system Ryzen 9 9950X3D, 64 GB, Windows 11 ### Changes from the defaults Upscaler: DLAA at 3840x2160. Frame generation on in one run and off in the others: it shows both ways. - This is my own build from source (the v3.0.0-rc.7 tag) with a few changes of my own: button mapping and prompt options, hold to quit, a hold at 60 fps for the game's scenes, and the patch from #1370. Only the last one touches rendering. - `longshot_env.txt`: `HOS_DISCORD=0` and `HOS_CRASH_WATCH=0`. For the two test runs one more line each: `RX_VK_OCC_QUERIES=0`, then `RX_VK_DLSS_TRACE=300`. - no ReShade, shader-cache tool or injector, clips recorded with obs, xbox game bar enabled ### longshot_log.txt In the two whole logs my Windows account name is replaced by "account" in four paths (C:\Users\account\...). Nothing else in them is changed. - `longshot_log-second-tunnel-important-lines.txt`: start here. The lines that matter out of four runs, with a short note on each. - `longshot_log-rc7-second-tunnel-draw-list.txt`: the whole log of the run with the problem, with `RX_VK_DLSS_TRACE=300`. - `longshot_log-rc7-second-tunnel-fixed.txt`: the whole log of the run with the second form of the patch. - `dlss-hidden-body-depth-v3.0.0-rc.7-v2.patch` and `skin_check-the-three-places.txt`: the change. ### Screenshots or video First clip: the problem, on the build with the first patch from #1370. Second clip: the same crawl with the second form of the patch. ### Before you send it - [x] The problem is still there on the newest release or nightly. - [x] I searched the open issues for it. - [x] I'm not attaching game files, keys, the game's executable, its .bin caches or saves.
Member

Possibly related, for context:

  • issue #1370 Game: DLSS/DLAA ghosting when camera clips inside Link — same camera-inside-Link depth ghosting; this reports a missed mesh case
  • PR #1447 fix(vk): leave out the depth draws of a mesh the game no longer shows: no ghosting with the camera inside Link (#1370) — merged fix misses skinned draws whose palette is in another binding
Possibly related, for context: - issue #1370 Game: DLSS/DLAA ghosting when camera clips inside Link — same camera-inside-Link depth ghosting; this reports a missed mesh case - PR #1447 fix(vk): leave out the depth draws of a mesh the game no longer shows: no ghosting with the camera inside Link (#1370) — merged fix misses skinned draws whose palette is in another binding
Sign in to join this conversation.
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot#1529
No description provided.