Digital Rendering Engineering

The equations don't run themselves.

The PathTrace() algorithm is derived in Volume 1. This volume builds the hardware infrastructure that executes it: the acceleration structure it traverses, the shader binding table that routes hits to materials, and the render graph that schedules the entire frame in 16.6 milliseconds.

Digital Rendering Engineering: The Hardware of Light is not a DirectX 12 tutorial. It is a production reference for Rendering Engineers, Graphics Programmers, and Engine TDs who already know what a BLAS is and need to know why their TLAS build is costing 3ms instead of 0.3ms.

• GPU Architecture - the execution model: Warp divergence measured with WaveActiveBallot. The occupancy cliff at 42 registers on Ampere. RDNA 3 comparison. Async compute: conditions for real overlap vs. accidental serialization. All with numbers from NSight, not estimates.
• DX12 Pipeline - production patterns: N=3 frames-in-flight fence protocol. Descriptor heap for bindless RT (SM 6.6, DESCRIPTORS_VOLATILE required). Render graph with complete barrier table for the DRE RT frame. PSO pipeline library: 200-500ms cold to under 10ms cached. All C++ targets Agility SDK 1.614.
• Hardware Ray Tracing - the complete DXR infrastructure: TLAS/BLAS: build flags, refit vs. rebuild, compaction. RTPSO: subobjects, hit groups, payload size vs. occupancy tradeoff. Shader Binding Table: indexing formula, the canonical stride bug, shadow ray optimization. All HLSL SM6.6, DXC-compiled.
• ReSTIR DI - 3-pass implementation: Reservoir struct (28 bytes/pixel). Initial candidates (N=32). Temporal reuse with M-cap. Spatial reuse with Jacobian correction and mandatory visibility ray. Measured: 1.81ms total on RTX 4090 at 1440p. Bitterli et al. 2020 derivation included.
• ReSTIR GI - one-bounce indirect illumination: GI reservoir with secondary hit validation. Temporal reuse: why the 0.2ms visibility ray is not optional (ghosting lasts 20 frames without it). Spatial: quality vs. cost. WRC for bounce 2+: 8MB hash grid, 0.6ms update, replaces multi-bounce at under 1 percent cost.
• Denoising hierarchy - SVGF, NRD, DLSS RR: SVGF complete (temporal + a-trous). DLSS RR: NGX integration, diffuse/specular separation requirement, resource state checklist (each missing item maps to a specific artifact). Glass mask fix: prevents indefinite boiling on refractive surfaces. Fallback hierarchy for non-RTX hardware.
• Neural Rendering - production implementation: 3DGS: covariance projection, GPU radix sort, tile rasterizer. NRC: multiresolution hash grid encoding + MLP, complete backpropagation, Adam with atomic hash updates, 33MB VRAM. SER (Ada): NvHitObject + NvReorderThread, 24 percent speedup at 50 materials. Work Graphs: adaptive dispatch with zero CPU round-trip.
• Engine Integration - 3 targets: UE5: Lumen override points, custom RT pass plugin. Proprietary DX12: 4 interface contracts (GBuffer, Scene, Material, Output). O3DE Atom: AZSL SRG, pass template, 3-8s hot reload cycle. CI/CD with CMake + GitHub Actions: zero-warning policy, automatic dependency tracking on Vol.1 functions.

All HLSL compiles on Shader Model 6.6. All C++ targets D3D12 on Windows 11. Vol. 1 companion code included as git submodule.

github.com/OpticsOptimizationLab/dre-hardware-of-light

Prerequisite: Volume 1 (The Physics of Light) - the BRDF, VNDF sampling, and path tracing mathematics are not re-derived here. They are called from the DXR pipeline built in this volume.

The chapter is not ready until the code compiles on production hardware and the frame budget closes at 16.6ms.