Roadmap

Roadmap

Phase 1: Core Engine Foundations — ✅ COMPLETED

• ECS + kernel Ring Buffer proof of concept
• Basic CUDA kernel (gravity + semi-implicit Euler)
• Pinned Memory (zero-copy CPU↔GPU)
• Race conditions resolved (atomics, double buffering, sparse set)
• Kernel module (LKM) with direct packet injection
• Dynamic packets ([EntityID][CompID][Data]...)
• Generic component dispatcher
• Validation: 62.55µs latency, 0% loss, 495 pkt/s

---

Phase 2: Architecture Refactor & Modularization — ✅ COMPLETED

• Refactor from monolithic engine/+shared/+plugins/ to 20 flat modules
• Migrate build system from Make to xmake (C++23, -fno-rtti, -fno-exceptions)
• Apply SOLID principles, Doxygen documentation, .inl inline files, guard clauses
• Merge plugins/bci/ into root bci/ module with full DSP pipeline
• Integrate a Vulkan backend into the render/ module (alongside a software rasterizer)
• Create engine/ facade aggregating all module dependencies
• Port CUDA physics from legacy PhysicsGPU.cu to gpu/ module (PhysicsKernel.cu)
• Port kernel module install/uninstall to xmake custom targets (kmod-*)
• Implement automatic socket fallback when kernel module unavailable
• Decouple GPU compute (--cuda) from the renderer (--renderer) for headless server

---

Phase 3: Simulation & Network Research — 🔄 IN PROGRESS

• Complete authoritative server (apps/server/main.cpp)
• Client with client-side prediction
• EntityRegistry (sparse set + generational IDs)
• SystemScheduler (DAG, component conflict analysis, PreSwap/PostSwap)
• Global double buffering (WorldPartition + Partition)
• Octree memory optimization (std::allocator + reserve)
• Empty chunk garbage collection (inline GC, FlatAtomicsHashMap::remove())
• ThreadPool for SystemScheduler (Fork-Join pattern, parallel stages)
• Unified Network class (driver + socket fallback + dispatch → PacketQueue)
• Bidirectional kernel module (TX ring buffer + kernel thread + ioctl wake)
• Sleeping entities (active/dormant) to save computation
• Real Octree integration (broadphase collision)
• SIMD Physics (AVX/SSE vectorization)
• Dirty List (migration optimization: check only moving entities)
• Advanced client prediction (Hermite splines)
• Tests with simulated network latency (50-200ms)
• Anti-tunneling: raycast across chunks to detect fast entity tunneling
• Chunk activation/deactivation (_active flag): selective spatial update

Success criteria: 1000+ stable entities at 60Hz with CPU physics.

---

Phase 4: BCI Plugin & Neurofeedback Research — 🔄 IN PROGRESS

This phase develops the bci/ module as an experimental research vector, aligned with the SEAMLESS team's paradigms at Inria.

• OpenBCI Cyton driver — 8-channel, 250 Hz, USB serial, lock-free ring buffer
• Per-channel FFT with Hann window (256 pts) — PSD per channel
• SignalMetrics — Schumacher $R(t)$, sliding RMS, baseline calibration (12 tests ✅)
• RiemannianGeometry — $\delta_R$ affine-invariant, $D_M$ Mahalanobis, Jacobi (12 tests ✅)
• NeuralMetrics — normalised struct from_state(), muscle_alert
• BCI module architecture (sources, DSP pipeline, metrics, streams)
• Enable BrainFlow, LSL, Eigen external dependencies in xmake (merged into the official xmake-repo)
• LSL Outlet — Broadcast corrected EEG timestamps
• [~] Auto-calibration phase (30s rest → compute baseline_R) — in progress (bci/calibration/)
• OpenViBE Box Algorithms — CBoxAlgorithmStabilityMonitor, CBoxAlgorithmMuscleRelaxation
• Visual feedback loop conditioned on NeuralMetrics
• Neural signal mapping (motor imagery decoding) → ECS components

Success criteria: Closed-loop BCI pipeline with enriched feedback running < 20ms e2e.

---

Phase 5: Massive Simulation & Infrastructure — 📋 PLANNED

• Broadphase collision via octree + adaptive brute-force (GPU BVH if possible)
• Narrow-phase AABB (GJK, simplified SAT)
• Collision resolution: impulse-based + positional correction
• Benchmark: 100k+ entities @ 60 FPS stable
• GPU streaming with CUDA streams
• State compression (delta, bitpacking)
• Session management (auth, per-client stats, RCU)
• Server Meshing (spatial sharding, chunk migration)
• Inter-server ghosting: migrated entities remain readable on the old server during transition
• TTL for unloaded chunks: time in RAM before serialization (5-30s?)
• Fast binary serialization format for offloadChunk()
• Dynamic load balancing
• NUMA-aware allocators (if multi-socket server)

Success criteria: 100k+ entities maintained @ 60 FPS stable.

---

Phase 6: Total Immersion — 🔮 FUTURE

• Photorealistic rendering (NeRF or PBR pipeline)
• Haptic feedback (via haptic/ module)
• Spatial audio (via audio/ module)
• Custom RTOS for strict determinism
• GPUDirect RDMA (NIC → VRAM direct)

Success criteria: Functional FullDive prototype with multi-modal sensory feedback.

---

← Executables | Next: Future Ideas →