Skip to content

ROADMAP.md

Latest commit

 

History

History
49 lines (42 loc) · 2.75 KB

ROADMAP.md

File metadata and controls

49 lines (42 loc) · 2.75 KB
Raw

Roadmap

Phases are ordered so each one proves something measurable before the next starts. Frame-rate targets are ceilings from bench.sh until a real loop measures the full cost on real scenes.

Phase 0 — Proof (done)

  • Measure parse vs trace vs wall-clock → bench.sh
  • Tri-brain architecture review → ARCHITECTURE.md
  • Wire protocol draft → daemon/PROTOCOL.md

Phase 1 — Resident daemon PoC (DONE — see FINDINGS.md)

  • Check out POV-Ray 3.7 source; build stock to confirm toolchain (tools/build_engine.sh)
  • Resident frontend: link libvfe+libpovray, init the engine ONCE, render every frame against the live session into an in-memory framebuffer — no PNG, no respawn (daemon/resident.cpp)
  • Found + cut the hidden 50 ms message-queue poll → 0.1 ms (patches/)
  • SDL2 live window with orbit camera + live fps (daemon/live.cpp)
  • Gate result: real raytraced frames live in memory; 1.4 → ~9–12 fps (thread-count-dependent). Gate substantially met; >30 fps is Phase 2.

Phase 2 — Resident thread-pool + scene (the .kkrieger lesson)

The remaining wall WAS per-frame regeneration. Measured truth (FINDINGS §7): the "thread churn" was mostly every TraceTask rebuilding identical random-sequence tables from a default-seeded mt19937 — 33% of daemon CPU.

  • Persistent task worker pool (parser + trace tasks; FD_NO_TASK_POOL=1 reverts) — neutral post-cache on this box, right shape for the engine
  • Memoize + share random sequences (the 33% bug); driver polls → 200µs; output verified bit-identical via deterministic selftest hash
  • Gate MET: 320×180 sustained — spin 107 fps, bee_world 91 fps, fd-game arena 138 fps end-to-end (floor 64×36: 9.85 → 2.34 ms)
  • Keep the scene database + bounding hierarchy resident; per frame mutate only transforms + camera (two-level BVH per GAME_ENGINE.md §2) — deferred until scenes are heavy enough for parse (~1-1.5ms) to matter

Phase 3 — DLSS-style temporal layer (RTX 5070 on .106)

  • Emit depth + object-ID buffers from the tracer (cheap)
  • Motion-vector reprojection from analytic transform deltas
  • Decide empirically: reproject-to-60 vs just render more real frames
  • Posterize + bayer-dither post pass (match the VGA look)

Phase 4 — Productize

  • Resolve AGPL posture before any binary ships (ARCHITECTURE.md → Licensing)
  • Health-check + auto-restart supervisor; separate daemon per consumer
  • Wire a playable demo into BoTTube / RustChain (feverdream addon sibling)

Non-goals

  • Replacing the batch pre-render pipeline (bottube-feverdream) — it stays.
  • Smooth full-res real-time. The aesthetic IS low-res + dither; that's the point.