Raytracing Roadmap

[devshgraphicsprogramming]

Bugs and Hacky Workaround removal

• [x] Sporadic UIAutomationCore crash (OpenCL timeout)
• [x] Cache the sample sequence properly (record the params) and change its locality
• [x] Solve the weird coloured fireflies issue
• [x] Instancing transform bugs (stuff gets thrown all over the scene)

Minimum Usability

• [x] Read Max Samples from the XML file
• [x] Implement Russian Roulette and read it from the XML as well
• [ ] Robust Intersections
• [ ] Implement Rejection Sampled NEE & Multiple Importance Sampling
• [x] Scramble key advancement
• [x] Denoiser Integration (optionally enabled)
• [x] Autoexposure Integration
• [x] Tonemapping Integration
• [x] Saving intermediate outputs when screenshotting
• [x] Envmap light source and importance sampling
• [ ] Fast and FFT bloom choice

Tests

• [ ] test divergence between Mitsuba, ex 42 and ex 22 in daily_pt.xml
• [x] yank up the max bounces and see if it causes a problem

Ease of Use

• [x] Unify raygen.comp and closestHit.comp
• [ ] Extend nbl::system to be able to download textures and meshes over HTTP/URL
• [ ] Integrate NVENC/libvpx for producing video captures while rendering (CPU solution preferred)
• [ ] Cloud XR

Quality and Convergence with no bias

• [ ] OBB computation for light source meshes
• [ ] Blue Noise Sampling via Heitz's techniques of scramble key ordering
• [x] Schussler et al. BxDF fixes for bumpmaps and smooth shading normals
• [ ] CLEAR/LEADR/Toksvig (produce roughness map while mipmap filtering bumpmap)

Easy Quality Wins

• [ ] Covariance Rendering (possible memory bandwidth win of >15% by using mipmaps)
• [x] Light Path Expressions || Reweighting Firefly Samples
• [ ] GSAA (increase roughness of curvy surfaces)
• [ ] A-SVGF

Uber Quality

• [ ] ReSTIR
• [ ] Adaptive Sampling
• [ ] Spatial and Temporal Denoising (Nvidia SDK, ProRender, Intel, Renderman-like 7 frame AI?)
• [ ] Local TMO

Load Time Optimizations

• [ ] Converted Asset Cache
• [x] Polyphase Mip Mapping
• [ ] commit on GPU Virtual Texture and Virtual Geometry

Memory Optimizations

• [ ] Basis-U compressed prefiltered textures in the cache
• [ ] Compress geometry position attribute into 64bits

Rendering Optimizations

• [x] Pretransform (de-instance) if memory allows
• [x] Single BLAS for static scene elements
• [ ] Selectively disable normals and reonstruct from position derivatives
• [x] dont fetch UVs or compute some derivatives if no texture precomp stream
• [ ] Dont jitter and re-rasterize every frame if camera is still and we're accumulating
• [x] Occlusion and Triangle Culling when rasterizing
• [ ] Vulkan Raytracing port
• [ ] Path Guiding
• [ ] Descriptor Indexing
• [ ] Bidirectional or Global integrators such as VCM or UBP
• [ ] Manifold NEE for caustics or lightsources embedded in bulbs

Risky Optimizations

• [x] Optimize rejection sampling to do spherical convex hulls of light source or clipping to the hemisphere
• [ ] Material Compiler made out of RTX callables
• [ ] Last Krivanek paper MIS weights out of the [0,1] range
• [ ] Replace all scene BxDFs with LTC (texturing is problematic)
• [ ] Combined product sampling of LTC and light (requires the above) probably via composed warps
• [ ] LTC product path guiding

Endgame

• [ ] Baking (DDGI or Spherical Gaussian lightmaps)
• [ ] Animations & Temporal Tracing
• [ ] Subsurface Scattering (level up the BxDFs)
• [ ] Subdivision NEE for rendering of skin lit up by office lights
• [ ] PN-Triangle intersection
• [ ] Hair Rendering (geometry and BxDF)
• [ ] Geometry LoDs
• [ ] Out of Core Rendering (streaming Moana, here we come!)
• [ ] Spectral Rendering with Herowavelengths and resolving to multiple colourspaces simultaneously

[devshgraphicsprogramming]

albedo needs 8-10 bits of ray payload normal needs between 16 and 32

would a good idea to subsample (lower the spp) on the AoV calculations (they're noiseless anyway)

[devshgraphicsprogramming]

New stuff

• [ ] Register compression in Material Compiler (lots of benchmarks needed)
• [ ] Use tweaked GGX as a tweaked spherica sampler for diffuse ;)
• [ ] Adapative Sampling (cast more rays in pixels where variance is higher)