[4ian]

Some testers report stuttering in starters using the 3D terrain (GDevelopApp/GDevelop-examples#1106). Their Chrome traces show the graphics card is the bottleneck: the terrain's shader costs too much for each pixel of ground. JavaScript time per frame is unchanged.

Changes

• Noise: the noise making layer borders irregular is read from a 256×256 texture of random values: 1 texture read instead of 16 sin() hashes per pixel. With texture variation, it is 3 reads instead of 12 hashes.
• Height blending: the average color of each layer texture is computed once on the CPU and passed as a uniform, instead of 4 extra texture reads per pixel. If the image can't be read by the CPU (from another domain without CORS, or not loaded yet), that layer still reads the smallest mipmap.
• Shader variants: the shader is compiled only with what the terrain uses, through #defines for layer textures, normal maps, height blending and texture variation.
• Vertex shader: the world normal comes from the camera-space normal three.js already computes, instead of an inverse() for each vertex.
• New advanced property "Anisotropic filtering", 4x by default as before.

Compatibility

• Projects keep the same properties and values.
• The shader is compiled again only when the set of #defines changes. Properties change only in the editor (property actions are private), and holes already worked this way.
• The irregular shapes of layer borders and texture-variation patches use new random values, so they look the same overall but sit in different places.

Testing

The gameplay tests of 5 examples using the terrain, grass and scattered models were run with GDevelop 5.6.286: 3D island terrain, 3D driving, 3D tank, first-person horror and 3D flight sim. Each game was run twice with the old extension and twice with the new one.

• 25 of 25 tests pass with both versions.
• Screenshots: with the same version, two runs give pixel-identical screenshots, except the tank game's gameplay screenshots, whose physics vary from run to run. Between old and new, differences are small and limited to the fine texture of the ground and the position of the noise patterns.
• Rendering time: measured with a temporary benchmark test (not committed): a 640×360 frame, waiting for the graphics work to finish, median of 20 frames, averaged over the two rounds. Times come from Chromium's software renderer, so they show the direction of the change rather than the gain on a real graphics card. "Terrain only" hides the grass and scattered models, whose code didn't change.

Game	Full scene	Terrain only
3D island terrain	2077 → 1888 ms (−9.1%)	1865 → 1697 ms (−9.0%)
3D driving	913 → 817 ms (−10.6%)	747 → 709 ms (−5.1%)
3D flight sim	972 → 842 ms (−13.5%)	808 → 704 ms (−12.9%)
3D tank	1030 → 906 ms (−12.0%)	791 → 671 ms (−15.2%)
First-person horror	997 → 900 ms (−9.7%)	449 → 388 ms (−13.7%)

🤖 Generated with Claude Code

https://claude.ai/code/session_01GFZnZwjYz8dSQxu3B2QVUJ