soft3/mir.md

mirrender engineworld rendererR-1.0 implementation
cyber core entity cyber
π 0.01% 629 words · 4 min

mir

Russian мир: world, peace, community.

mir is the render engine for cyber. it reads two inputs and produces one output:

tru   →  positions (φ*, eigenvectors, spectral coords)  ─┐
                                                           ├→  mir  →  R-1.0 world
glia  →  inference outputs (neural features)            ─┘

every neuron running mir on the same graph state arrives at the same world. not approximately — topologically identical. the math island is at the same BVH path for everyone.

mir knows nothing about graphs, particles, or cyberlinks. it receives coordinates and features and makes them visible.

what mir does

epoch work (≤ 1 Hz) frame work (display refresh)
LOBPCG eigensolver on normalized Laplacian ℒ GPU BVH frustum cull → VisibleSet
Procrustes alignment to anchor-1024 tier dispatch (T0–T3 draw calls)
heat-kernel BVH at four τ scales focus luminosity animation
NRF training (Phase 2+) edge flow UV offset update
results double-buffered into EpochState IOSurface composite (zero copy)

between epochs, positions and cluster IDs are frozen. only focus luminosity and edge flow animate within a frame epoch.

rendering tiers

tier condition what renders
T0 content screen diameter > 200 px particle opens — content environment
T1 surface 40–200 px impostor + text label + edge preview
T2 shape 8–40 px analytic impostor (indirect draw, tile-shaded)
T3 splat 1–8 px 3D Gaussian splat (Kerbl 2023)
T∞ field < 1 px neural radiance field — cost per pixel, not per particle

T∞ is the tier that makes R-1.0 scale to unbounded graph size. every frame costs pixels × samples, independent of particle count.

hardware (honeycrisp backend)

component role
aruminium Metal GPU: BVH cull compute, impostor / splat / edge draw, composite
rane ANE: NRF head inference (Phase 2+)
acpu AMX: LOBPCG eigensolver, Procrustes alignment, BVH build, diffusion steps
unimem IOSurface: zero-copy frame handoff across CPU / GPU / ANE

target: M3 Pro at 4K 120 Hz (8.3 ms frame budget), 1 M-particle graph.

what mir is not

mir is not the graph engine. it does not compute φ*, eigenvectors, or cyberank — those come from tru. it does not run inference — that comes from glia. it receives results and renders them.

specs

  • specs/render.md — R-1.0 canonical protocol spec (Clifford primitives: compiled transformers spec §3)
  • specs/render-cyb.md — R-1.0-cyb: cyb implementation of R-1.0
  • docs/plan.md — implementation plan and architectural decisions

implementation order (Phase 1)

step module deliverable
1 graph/ snapshot loader, CSR adjacency, particle index
2 epoch/eigensolver LOBPCG + Procrustes aligned positions
3 epoch/bvh heat-kernel BVH, 4 τ scales
4 frame/cull GPU BVH traversal → VisibleSet + TierLevel
5 frame/tiers/t2 analytic impostor (indirect draw, tile-shaded)
6 frame/tiers/t3 Gaussian splat (Kerbl 2023)
7 bevy/ GraphWorldPlugin, WorldState::Graph, Cmd+5
8 frame/tiers/t1 labels via sugarloaf, world-space text
9 frame/edges bundled tubes + flow UV animation
10 frame/tiers/t0 content entry: camera transition + sandbox
11 conformance P-RENDER-TOPO, P-RENDER-POS, P-RENDER-FPS

Phase 2: T∞ hash-grid MLP (no CT-0.1). Phase 3: T∞ full NRF — CT-0.1 + Clifford block + volume ray-march.

in the stack

tru (compile model)  →  .model + φ* + eigenvectors
glia (run model)     →  neural features
mir (render)         →  R-1.0 world

see stack

Folder

Graph