W3 Chunked terrain mesh, collision and LOD #10

Closed
opened 2026-11-09 05:26:05 +00:00 by claude · 1 comment
Owner

Depends on: W2

Build chunked render meshes with LOD and seam handling, biome splat materials, and server-side collision that needs no renderer.

Acceptance

  • Collision works in the headless server, and character tests pass on generated terrain.
  • No visible cracks between LOD levels in a capture.
  • Chunk streaming causes no frame spike over 8 ms in the profiler scenario.
**Depends on:** W2 Build chunked render meshes with LOD and seam handling, biome splat materials, and server-side collision that needs no renderer. **Acceptance** - Collision works in the headless server, and character tests pass on generated terrain. - No visible cracks between LOD levels in a capture. - Chunk streaming causes no frame spike over 8 ms in the profiler scenario.
claude added this to the v0.2 Movement and Island milestone 2026-11-09 05:26:05 +00:00
claude added the type/featurepriority/high labels 2026-11-09 05:26:05 +00:00
claude added this to the Tidewall v1.0 project 2026-11-09 15:00:04 +00:00
eric closed this issue 2026-11-10 19:28:36 +00:00
Author
Owner

Done in 8b74485.

What changed

  • src/client/terrain/terrain_view.gd: TerrainView draws the island as 16 × 16 chunks of 64 m. Each chunk is shown at one of four detail levels (2, 4, 8 and 16 m spacing), chosen by its distance to the camera. Geometry is built on WorkerThreadPool threads, nearest chunks first, with at most 6 in flight. The main thread only uploads finished meshes, within 2 ms per frame.
  • terrain_chunk_builder.gd: builds a chunk's arrays as a pure function. Seams: skirts hang from every edge, deeper than twice the most any level strays from the full-resolution edge. Normals come from the full-resolution heightfield at every level, so lighting matches across a boundary.
  • terrain.gdshader: a procedural splat with no textures. It blends sand, marsh, grass and rock using band weights averaged over 4 m. Steep ground turns to rock, and summits weather to pale stone.
  • Collision: the server and prediction use SimGround's heightfield only. Nothing from the renderer is involved.
  • Capture scenarios: terrain, terrain_lod (chunks tinted by level), terrain_cracks and terrain_cracks_control. tools/crack_check.gd counts sky pixels seen through the ground. make terrain-capture requires cracks in the skirtless control and none in the real terrain.

How it was tested

  • make check passes: 111 tests, self-check and smoke.
  • On seeds 11 and 12, 12 players move randomly for 900 ticks and never sink into the ground. A player sprints from 150 m out towards the tallest summit, climbs more than 15 m, and takes no damage.
  • Seams: on sampled shared edges, every pairing of the four levels is checked, and each gap is smaller than the skirt.
  • make terrain-capture: the control shows 437 crack pixels, and the real terrain shows 0. I also looked at the captures. The level rings are visible in terrain_lod, with no seams. Before a fix, the cracks view showed skirts lit as back faces, as thin dark lines. The shader now lights both faces from the terrain normal, and the lines are gone.
  • Profiler scenario (test_streaming_across_the_island_stays_under_the_frame_budget): the camera flies 1 km across the island at 30 m/s (0.5 m per frame at 60 fps), for 2,037 frames and 661 chunk rebuilds. The worst main-thread frame was 0.8 to 1.4 ms across 7 runs, against the 8 ms limit. This is headless CPU time only. GPU upload cost and real frame times need a GPU playtest.
  • One earlier run, under heavy host load (load average about 10), recorded a single 238 ms frame. I couldn't reproduce it in 7 later runs, so I opened a hardening follow-up to keep watching it.

The coastal views show bare sea floor, because water arrives with W6. Lowland is plain grass until W5 adds vegetation.

Done in 8b74485. **What changed** - `src/client/terrain/terrain_view.gd`: `TerrainView` draws the island as 16 × 16 chunks of 64 m. Each chunk is shown at one of four detail levels (2, 4, 8 and 16 m spacing), chosen by its distance to the camera. Geometry is built on `WorkerThreadPool` threads, nearest chunks first, with at most 6 in flight. The main thread only uploads finished meshes, within 2 ms per frame. - `terrain_chunk_builder.gd`: builds a chunk's arrays as a pure function. **Seams:** skirts hang from every edge, deeper than twice the most any level strays from the full-resolution edge. Normals come from the full-resolution heightfield at every level, so lighting matches across a boundary. - `terrain.gdshader`: a procedural splat with no textures. It blends sand, marsh, grass and rock using band weights averaged over 4 m. Steep ground turns to rock, and summits weather to pale stone. - **Collision:** the server and prediction use `SimGround`'s heightfield only. Nothing from the renderer is involved. - Capture scenarios: `terrain`, `terrain_lod` (chunks tinted by level), `terrain_cracks` and `terrain_cracks_control`. `tools/crack_check.gd` counts sky pixels seen through the ground. `make terrain-capture` requires cracks in the skirtless control and none in the real terrain. **How it was tested** - `make check` passes: 111 tests, self-check and smoke. - On seeds 11 and 12, 12 players move randomly for 900 ticks and never sink into the ground. A player sprints from 150 m out towards the tallest summit, climbs more than 15 m, and takes no damage. - Seams: on sampled shared edges, every pairing of the four levels is checked, and each gap is smaller than the skirt. - `make terrain-capture`: the control shows 437 crack pixels, and the real terrain shows **0**. I also looked at the captures. The level rings are visible in `terrain_lod`, with no seams. Before a fix, the cracks view showed skirts lit as back faces, as thin dark lines. The shader now lights both faces from the terrain normal, and the lines are gone. - **Profiler scenario** (`test_streaming_across_the_island_stays_under_the_frame_budget`): the camera flies 1 km across the island at 30 m/s (0.5 m per frame at 60 fps), for 2,037 frames and 661 chunk rebuilds. The worst main-thread frame was 0.8 to 1.4 ms across 7 runs, against the 8 ms limit. This is **headless CPU time only**. GPU upload cost and real frame times need a GPU playtest. - One earlier run, under heavy host load (load average about 10), recorded a single 238 ms frame. I couldn't reproduce it in 7 later runs, so I opened a hardening follow-up to keep watching it. The coastal views show bare sea floor, because water arrives with W6. Lowland is plain grass until W5 adds vegetation.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: claude/tidewall#10