Skip to content

tech_debt.md

Latest commit

 

History

History
119 lines (84 loc) · 4.87 KB

tech_debt.md

File metadata and controls

119 lines (84 loc) · 4.87 KB
Raw

Tech Debt — Performance

Items identified during performance audit. Ordered by impact.

1. Replace IBL Importance Sampling with Split-Sum LUT

Impact: High — eliminates the biggest per-pixel cost
Effort: High

deferred_ibl.frag runs a GGX importance sampling loop up to 100 times per pixel per frame. At 1280×720 with 24 samples that's ~22M textureLod calls per frame, each with mip-level PDF computation (log2, sqrt).

Fix: Precompute two textures offline:

  • DFG LUT (2D, NdotV × roughness): stores ∫ F(cosθ) G(cosθ) / 4·NdotV dω
  • Pre-filtered env map (cubemap mip chain): stores irradiance at each roughness level

Runtime IBL becomes two texture lookups + multiply. Standard split-sum approximation (Karis 2013 / UE4 approach). Baking can reuse the existing Hammersley + GGX code, run once on HDRI load.

Files: deferred_ibl.frag, Renderer.cpp (add BakeSpecularLUT(), BakePrefilteredEnv())

2. Geometry Shader Cubemap for Shadow Passes

Impact: Medium — cuts 30 draw calls to 5
Effort: High

ShadowPass() and MSMShadowPass() call DrawSceneGeometry() 6× per light (one per cubemap face) × 5 lights = 30 full-scene draws per frame. Each iteration changes viewport, binds a different framebuffer attachment, and re-uploads model matrices.

Fix: Single draw call per light using a geometry shader that emits gl_Layer to route triangles to all 6 faces at once. Requires:

  • New shadow_depth_cube.geom geometry shader
  • Pack 6 view-projection matrices into a UBO
  • glFramebufferTexture (attach entire cubemap, not face-by-face)
  • Same fix applies to MSM moment pass

Files: Renderer.cpp (ShadowPass, MSMShadowPass), new shadow_depth_cube.geom

3. View-Space Depth in GBuffer (for Cel Outline)

Impact: Low–Medium — removes 8 matrix multiplies per pixel in outline pass
Effort: Medium

cel_outline.frag calls -(uView * vec4(worldPos, 1.0)).z once per neighbour sample (8 times per pixel). The view matrix is uniform but the per-sample multiply still executes in the fragment shader.

Fix: Store view-space depth directly in GBuffer during the GBuffer pass — pack it into the unused w component of the Normal attachment (currently RGBA16F, w is wasted). cel_outline.frag then reads a single float per sample instead of a mat4 multiply.

Alternatively: half-resolution outline pass + bilinear upscale, which also halves the texture fetch count.

Files: gbuffer.frag (write depth to gNormal.w), cel_outline.frag (read gNormal.w) GBuffer.h (document the new packing), ao.frag (already uses world pos, unaffected)

4. Cache Shadow Cubemaps for Static Lights

Impact: Medium — skips up to 30 scene redraws per frame when scene is static Effort: Low

ShadowPass() / MSMShadowPass() re-render every cubemap face every frame, even when neither the light nor any scene geometry has moved. In this project the Cornell box and all lights are static for the entire session.

Fix: Dirty-flag per light. Set on light position/range edit, on geometry transform change, or on m_showIBLProbes toggle. Skip the render pass for clean lights. Forcing a full rebuild on shader reload is acceptable.

Files: Renderer.cpp (ShadowPass, MSMShadowPass, light/geometry mutators), Renderer.h (m_shadowDirty[kMaxLights])

5. Instance or Batch Scene Submission

Impact: Low — reduces CPU-side draw overhead Effort: Low

DrawSceneGeometry() issues one Draw + one SetMat4("uModel", ...) upload per primitive. Cornell walls, ground, 3 cubes, sphere, up to 8 probe spheres = 14 draws per pass. Called once in GBuffer, once per shadow face per light in shadow passes. Probe spheres are all the same mesh at different transforms — textbook instancing case.

Fix: glDrawElementsInstanced for probe spheres with a per-instance model matrix in a VBO or SSBO. Secondary: collapse the three cubes into one instanced draw.

Files: Renderer.cpp (DrawSceneGeometry), Mesh.h/.cpp (add DrawInstanced)

6. Frustum Cull Before Shadow Face Submit

Impact: Low — skips submits that contribute nothing Effort: Low

Every primitive is re-submitted for all 6 cubemap faces of every light, even when the primitive sits entirely outside that face's 90° frustum. For Cornell the back wall never contributes to the +Z face from a light in front, etc.

Fix: Per-face frustum vs. per-primitive AABB test before the Draw call. Complements item #2 — if GS layered rendering lands, this moves into the geometry shader as per-triangle culling.

Files: Renderer.cpp (ShadowPass, MSMShadowPass), new AABB helper.

Already Done (not debt)

  • MSM on by default (m_useMSM = true) — PCF fallback is rarely hit
  • Hammersley dirty flag — only rebuilds when m_iblSamples changes
  • PCF trimmed from 20 → 12 samples
  • AO blur radius default reduced 8 → 5