nishi code wiki / research / voxel rendering on WebGL2

Real-time rendering on a constrained stack: voxels, WebGL2, no compute

SOTA census · compiled 2026-08-01 · 5 axes · ~30 sourced claims · 8 declared gaps

lineage: forked from research_rtrender parent domain: real-time rendering + GPU architecture corpus slug: reference-voxelrender-webgl2-sota-2026-08-01

Claims are labelled SHIPPING / RESEARCH / ANNOUNCED / DEPRECATED. Gaps are declared UNVERIFIED rather than guessed.

1. The constraint filter — what shipping-AAA SOTA cannot survive

Constraints, stated once: WebGL2 / GLSL ES 3.00, fragment + vertex shaders only, no compute shaders, integer/fixed-point core, no third-party engines, everything reimplementable from papers. Current renderer: per-pixel integer DDA voxel raycast in a fragment shader, skinned-mesh NPC pass, sun shadow march, bump mapping, procedural clouds, 60 fps at 480×300 upscaled, scoring 983/1000 on an internal groundtruth critic that saturates at 1000.

All parent figures above are quoted verbatim from research_rtrender (compiled 2026-07-31) and carry their sources there; the Cooperative Vectors verdict is sourced to devblogs.microsoft.com/directx/shader-model-6-9-retail-and-more/ (2026-02-26). What survives the filter is narrow: anything expressible as render-to-texture fragment passes, vertex-stage skinning, and CPU-side wasm. That narrowness is the subject of this brief.

2. The browser platform, honestly, in 2026

What WebGL2 / GLSL ES 3.00 actually gives this stack: real integer and bitwise ops (the integer DDA core is legal and exact), texelFetch, 3D textures, integer textures, multiple render targets, instancing, transform feedback, uniform buffers, float render targets via EXT_color_buffer_float. What it withholds: compute shaders, storage buffers, atomics, shared memory, bitCount() (that is GLSL ES 3.1+ — popcount must be emulated with SWAR arithmetic or an 8-bit lookup texture), and any 64-bit integer type (64-bit voxel masks split into two uints). GPGPU on this stack means the classic form: ping-pong render-to-texture fragment passes. That is not a toy — it is exactly the shape Radiance Cascades wants (axis 4).

The WebGPU verdict for 2026: at ~82–85% global support with Firefox default-on only for Windows and Apple-Silicon macOS, WebGPU is a legitimate second backend but not a floor. A renderer that requires it silently drops one user in six. WebGL2 remains the correctness baseline; WebGPU is an acceleration tier — revisit when caniuse crosses ~90% and Firefox ships default-on across desktop Linux.

3. Voxel-specific SOTA — the one axis where the constraint is not a handicap

Teardown is the existence proof for this whole stack's thesis: a shipped, beloved game whose every surface is voxel ray-marched with true dynamic lighting, built by a tiny team with no triangle-GI machinery at all (Software Engineering Daily interview with Dennis Gustafsson, 2025-01). It is the closest shipped analogue to the Nishi renderer's architecture — the difference is scale of acceleration structure, not kind.

The structure ladder is well mapped. SVO → SVDAG (2013) bought memory: identical subtrees merged, billions of voxels in GPU memory. 64-trees buy traversal: a 4³ branching factor packs 64 children's occupancy into one 64-bit mask, halving tree depth versus an octree (each level divides an axis by 4, not 2) and letting one fetch answer “which of my 64 cells are solid”. The 2024 guide by dubiousconst282 is the best practical treatment and comes with a working repo. NAADF (Eurographics 2026) is the new frontier: instead of a deep hierarchy it uses three layers augmented with in-cell axis-aligned distance fields acting as empty-space-skip caches, and the authors measure 3–5x faster ray tracing than state-of-the-art DAG variants — while remaining much cheaper to edit than SVOs/DAGs because the three layers live in separate buffers and edits need no restructuring. For a destructible world that is the pairing that matters: DAGs bought compression by making edits hard; NAADF buys speed and keeps edits cheap.

WebGL2 porting notes (analysis, not sourced claims): 64-bit masks become uvec2; child indexing needs popcount, emulated in ES 3.00 by SWAR (v = v - ((v>>1u)&0x55555555u)…) or an 8-bit LUT texture; node pools live in integer textures addressed with texelFetch. Nothing in 64-tree or NAADF-style traversal requires compute — both are per-ray algorithms, and this renderer already runs one ray per pixel in a fragment shader. The integer core is an asset here: traversal against integer coordinates can be gated bit-exactly against the flat DDA.

4. Noiseless GI — Radiance Cascades, and why noiseless is load-bearing here

The idea in one paragraph. Sannikov's penumbra hypothesis: resolving illumination needs high spatial resolution only near surfaces and high angular resolution only far from them — never both at once. Radiance Cascades exploits the inverse relationship directly: cascade level i+1 doubles ray count (angular resolution) while halving probe density (spatial resolution), each level tracing only a bounded ring of ray intervals; merging the cascades reconstructs a full radiance field. Total cost is constant and scene-independent; there is no stochastic sampling, hence no noise, no denoiser, no temporal accumulation, zero temporal latency.

Why this is the decision-changer for a no-compute stack. The parent brief's GI ladder — BVH → ReSTIR DI → A-trous → radiance cache — is a pipeline for laundering 1-spp noise into an image, and every stage past the first assumes machinery (compute traversal, denoiser, TAA) this stack cannot host. Radiance Cascades inverts the premise: spend the budget on a deterministic structure instead of stochastic samples. The cascades are just textures; building and merging them are just fragment passes over quads; the probe traces are the same ray-march this renderer already does per-pixel for its sun shadow. The 2D screen-space form is proven in GLSL-ES-class shaders today (tmpvar, GM Shaders); the 3D world-space form is where the open research frontier sits — and where a voxel grid is the friendliest possible host, since interval traces become the engine's native integer DDA.

5. What this means for the Nishi stack — the ranked adoptable ladder

Explicit non-adoptions, so nobody relitigates them: ReSTIR (1-spp noise with no denoiser to launder it), 3D Gaussian Splatting (the tile radix sort is the correctness bottleneck per the parent, and it wants compute; also wrong representation — this is a voxel engine), Nanite-class mesh virtualization (compute raster + 64-bit atomics), frame generation (the parent already rules it out on latency grounds even with compute). None of these are “later” — they are wrong for this boundary, and the ladder above is not a consolation prize: rungs 1–3 are the current voxel research frontier, and the frontier happens to fit in a fragment shader.

Declared UNVERIFIED — do not treat as measured