GLSLTestbed

A Windows only OpenGL 4.6 renderer for testing different rendering techniques & solutions.

Development of this project has been discontinued in favor of PKRenderer.

Implemented Features

• Clustered forward rendering.
• Volumetric fog & lighting.
• Realtime global illumination (Voxel cone tracing).
• Temporal reprojection for volumetrics & screen space GI.
• Variance shadow mapping.
• Light cookies.
• Point, spot & directional lights.
• Cacaded shadow maps for directional lights.
• Physically based shading (Cook-Torrance brdf).
• HDR bloom.
• Tone mapping & color grading.
• Film grain.
• Auto exposure.
• Depth of field with auto focus.
• Screen space ambient occlusion.
• Octahedron hdr environment maps.
• Multi compile shader variants.
• Asset hot reloading.
• Shader material/property block system.
• Instanced dynamic batching.

Planned Features

• Rectangular area light support.
• Exponential variance shadow maps.
• Motion vectors.
• Render queues (currently only supports opaque renderers).
  • Transparent queue with back to front sorting.
• Documentation.

Render Pipeline Execution Order

A rough overview of the steps taken to render a frame. (Some steps are omitted to avoid repetition).

• Cull geometry & lights.
• Update lights system.
  • Sort lights by type & shadowmap usage.
  • Update light data buffers.
  • Render shadowmaps for shadow casting lights.
    • Render in batches of 4 (or 1 in the case of directional lights).
    • Gather shadow casting geometry for the batch.
    • Render intermediate shadow map.
    • Perform blur.
    • Blit into shadow map atlas.
• Render scene depth, normals & roughness.
• Compute light clusters.
  • compute max depth per 2d tile.
  • assign lights to clusters & cull clusters outside of max depth range.
• Render screen space ambient occlusion from scene depth & normals.
• Render visible geometry into gi volume.
• Render screen space gi.
• Forward render opaque objects.
  • Update instancing buffers.
    • Gather material properties to per shader property buffers.
    • Gather matrices to matrix buffers.
  • Draw instanced (PBR fragment shader overview).
    • Sample scene OEM for ambient specular & diffuse.
    • Sample & apply SSAO (Only affects ambient lighting).
    • Access contributing lights list from light clusters.
    • Process each light in the list through the material's BRDF.
• Render Volumetrics.
  • Compute Lighting & density per volume cell.
    • Process visible lights
    • Inject light from gi volume.
  • Compute integrated scattering per volume cell.
  • Composite with forward output.
• (TODO) Render transparent objects.
• Render depth of field
  • Compute auto focus distance.
  • Downsample forward output.
  • Render blurred foreground & background into two layers.
  • Upsample & composite layers with high res forward output.
• Bloom & Tonemapping
  • Compute luminance histogram from forward output.
  • Compute & interpolate auto exposure from luminance histogram.
  • Render bloom layers (35 passes of separable blur outputted into 6 different bloom layers).
  • Composite bloom layers.
  • Apply auto exposure & tonemapping.
  • Apply gamma correction.

Performance Metrics

• Average frame timings profiled on a NVIDIA RTX 2080 TI at 1080p resolution (These were recorded before global illumination was implemented).
• The test scene has only 512 objects with 3 different materials & 2 different meshes.
• Light types are distributed so that there exists one directional light & an equal amount point & spot lights.
• Light radii are set at 40m to achieve good saturation in light clusters.
• The test scene has all features enabled (a test without volumetrics would probably run a lot faster).

Results with shadow casting lights:

• 0 lights : 1.4ms
• 5 lights : 2.1ms
• 17 lights : 3.0ms
• 33 lights : 4.3ms
• 54 lights : 5.8ms

Results with non shadow casting lights:

• 32 lights : 1.9ms
• 64 lights : 2.4ms
• 128 lights : 3.5ms
• 256 lights : 5.8ms
• 512 lights : 7.2ms

Known Issues & Performance Drawbacks

• Matrix buffers are currently mapped per pass. Which is causing a lot of data duplication & some driver stalls.
  • A more optimal solution would be to gather matrices that contribute to the current frame into a single buffer & only build index buffers per pass.
• OpenGL doesn't support multiple command queues or async dispatches. Which leads to otherwise easily multithreadable passes having to be executed consecutively.
• Using sparse textures would probably be more performant in the following scenarios:
  • Shadow map atlas allocation. Currently the entire atlas is allocated in full. As opposed to only commiting sparse tiles for active areas.
  • Shader property instancing currently uses bindless texture handles for material properties. Creating a sparse texture atlas for them would probably be more cache friendly.
• Culling of for each pass is currently done on the cpu. This could be moved to the gpu instead.
• Vertex shader input attribute layouts are not ensured to match with vao attribute layouts in anyway.
• Vertex array objects are not shared among meshes but instead created per mesh.
  • A better approach could be to have attribute layout based vao cache & switch vertex buffers between drawcalls instead.
• Volumetric sample dithering causes artifacts on far away high frequency lighting effects.
  • This could be fixed by breaking up the low resolution sampling pattern in the composite pass with high frequency noise & then using temporal AA to hide the high frequency noise.

Asset Sources

• HDRI Haven
• CC0 Textures

Libraries & Other Dependencies

• C++ 17 support required
• OpenGL ARB Bindless texture & viewport array extension support required.
• OpenGL 4.6 support required.
• KTX
• yaml-cpp
• Glad
• GLFW
• GLM
• mikktspace
• tinyobjloader