Shader optimizations

[TokisanGames]

Performance

• 1. Precalculate this in C++ whenever terrain_size is changed and pass in the uniform: terrain_size_inv rather than in the shader. WIll save millions of divides per second.
  

  Terrain3D/src/terrain_material.cpp

  Line 98 in 0bb046e

  code += " vec2 ps = vec2(1.0) / terrain_size;\n"; // TODO can be precalculated

---

• 2. Replace a*b+c with fma(a,b,c), which is allegedly 10% faster. Not supported by webgl (but neither are texture arrays)
  https://docs.godotengine.org/en/latest/tutorials/shaders/shader_reference/shading_language.html

e.g.

• https://github.com/outobugi/GDExtensionTerrain/blob/0bb046eb2e0da0c137e91f74215f574f353f95fb/src/terrain_material.cpp#L73

• https://github.com/outobugi/GDExtensionTerrain/blob/0bb046eb2e0da0c137e91f74215f574f353f95fb/src/terrain_material.cpp#L80

• https://github.com/outobugi/GDExtensionTerrain/blob/0bb046eb2e0da0c137e91f74215f574f353f95fb/src/terrain_material.cpp#L141

• Maybe also make an FMA macro that uses fma or *+ depending on compatibility or not

---

• 3. Lookup region only once per vertex. Move into vertex() and save as a varrying for the colormap in fragment. - Looks bad on lower LODs when using this to calculate pixel data as the vertices are far apart.

---

• 4. Consider replacing texturegrad with texture, if we don't need it for 3D Projection. See Fragment & Texture Normals are wrong #169 update 2 and Fragment & Texture Normals are wrong #169 (comment)

---

• 5. As above remove all derivative functions and at least get them out of branches. Test optimizing branches

https://community.khronos.org/t/if-statements-in-shaders-confused-by-results/73518/15

---

• 6. Replace 16x16 Region map with an array. No need for a texture lookup

---

• 7. Replace 512x512 region blend map w/ a zoomed UV lookup of the region map, or a mathematical version from an array.

---

• 8. Since we're blending 4 adjacent vertices, move lookups to vertex() and weight in fragment. - Doesn't look good on lower LODs as the vertices are farther apart, but see redesign below.

---

• 9. Make a runtime virtual texture #245

---

• 10. Instead of sampling base/overlay texture lookups for each vertex, bilinear interpolate the UV for the texture lookup and sample albedo/normal once.

---

• 11. light() is wasteful for mobile according to lfxu. Rewrite it to do only what is needed.

Functional

• 1. May need to do separate transforms in vertex
  Firefly artifacts along seams between perfectly aligned meshes godotengine/godot#35067 (comment)
• 2. Rotating normal maps properly https://www.youtube.com/watch?v=iXqCszt3LbE
• 3. Single sample triplanar https://www.youtube.com/watch?v=dFApswlwsRc
• 4. Move uv scale to C++
• 5. Add vertex discard shader snippet to material options
• 6. Use antialiasing instead of distance fade

Sometimes I've had the need for antialiasing and have typically faded out (vertex grid, macrovariation) instead of aliasing. Perhaps anti-aliasing is better.
https://stackoverflow.com/questions/20204561/advanced-moir%C3%A9-a-pattern-reduction-in-hlsl-glsl-procedural-textures-shader-a
https://www.yaldex.com/open-gl/ch17lev1sec4.html#ch17fig04

• Triplanar mapping alternative methods.
  https://bgolus.medium.com/normal-mapping-for-a-triplanar-shader-10bf39dca05a

Redesign

• Consider looking up height, region, control, color all in vertex. Then with a large mesh, interpolate all in fragment. No lods, or very distant. No fragment lookups. Reducing memory usage (and query) is best optimization.
• Use triangle strips instead of triangles. And render terrain vertices without meshes.
  • https://www.youtube.com/watch?v=R-bjXOEQyX8
  • https://www.youtube.com/watch?v=5zlfJW2VGLM&t=174s
  • VERTEX_ID and mesh-less rendering godotengine/godot#19473
  • ArrayMesh with flag USES_EMPTY_VERTEX_ARRAY still needs vertex array for rendering work godotengine/godot#83446