rfcp/docs/devlog/gpu_supp/SESSION-2025-02-04-GPU-Acceleration-Complete.md

RFCP Session Summary — February 4, 2026

GPU Acceleration Complete: 195s → 11.2s (17.4x Speedup)

---

🎯 Session Goal

Complete GPU acceleration pipeline and optimize Full preset performance.

📊 Results

Performance Achievement

Metric	Before (3.7.0)	After (3.8.0)	Improvement
Full preset (6640 pts, 50m)	195s	11.2s	17.4x
Standard preset (1975 pts, 200m)	7.2s	2.3s (cached)	3.1x
Phase 2.5 (distances+path_loss)	0.33s	0.006s	55x
Phase 2.6 (terrain LOS)	7.29s	0.04s	182x
Per-point (workers)	1.1ms	0.1ms	11x

GPU Pipeline (Final Architecture)

Phase 1:   OSM data fetch (Overpass API)          ~6-10s (network)
Phase 2:   Terrain tile download + cache           ~4s first / 0s cached
Phase 2.5: GPU — distances + base path_loss        0.006s ⚡
Phase 2.6: GPU — terrain LOS + diffraction loss    0.04s  ⚡
Phase 2.7: GPU — antenna pattern loss              ~0s    ⚡
Phase 3:   CPU workers — buildings + vegetation     ~2s    
─────────────────────────────────────────────────
TOTAL (cached):                                    ~2.3s (Standard)
TOTAL (cached):                                    ~11.2s (Full)

---

🔧 Changes Made (Iterations 3.7.0 → 3.8.0)

Iteration 3.7.0 — GPU Precompute Foundation

• Added gpu_manager import to coverage_service.py
• Grid arrays created on GPU (CuPy)
• GPU precompute for distances + path_loss (vectorized)
• Fixed critical bug: CuPy worker process crashes (CUDA context sharing)
• Solution: GPU only in main process, workers use precomputed CPU values
• Fixed frontend duplicate calculation guard

Iteration 3.8.0 — Full Vectorization

• Phase 2.6: batch_terrain_los() in gpu_service.py
  • Vectorized terrain profile sampling for ALL points simultaneously
  • Earth curvature correction vectorized
  • Fresnel clearance + diffraction loss vectorized
• Phase 2.7: batch_antenna_pattern() in gpu_service.py
• Workers receive precomputed has_los, terrain_loss, antenna_loss
• Workers only compute buildings + reflections + vegetation

Critical Fix: _batch_elevation_lookup Vectorization

• Before: Python for loop over 59,250 coordinates (7.29s)
• After: Vectorized NumPy tile indexing, loop only over tiles (0.04s)
• Impact: 182x speedup on Phase 2.6 alone

Critical Fix: Vegetation Bbox Pre-filter

• Before: Each sample point checked ALL 683 vegetation polygons
• After: Bounding box pre-filter skips 95%+ of polygons
• Impact: Full preset 156s → 11.2s

---

📁 Files Modified

Backend

• app/services/coverage_service.py — precomputed values passthrough
• app/services/parallel_coverage_service.py — 5 worker functions updated
• app/services/gpu_service.py — batch_terrain_los, batch_antenna_pattern, batch_final_rsrp
• app/services/vegetation_service.py — bbox pre-filter on _point_in_vegetation

Build

• PyInstaller ONEDIR build: 1.6 GB dist → 1.2 GB NSIS installer
• CUDA DLLs bundled (cublas, cusparse, curand, etc.)
• Runtime hook for DLL directory setup

---

🏗️ Architecture (Final State)

Main Process (asyncio event loop)
├── Phase 2.5: GPU precompute
│   └── CuPy arrays: distances, path_loss (vectorized)
├── Phase 2.6: GPU terrain LOS  
│   └── Batch elevation lookup (vectorized NumPy)
│   └── Earth curvature + Fresnel (CuPy)
│   └── Diffraction loss (CuPy)
├── Phase 2.7: GPU antenna pattern
│   └── Bearing + pattern loss (CuPy)
│
└── Phase 3: CPU ProcessPool (3 workers)
    └── Receive precomputed dict per point
    └── Skip terrain/antenna (already computed)
    └── Only: buildings + reflections + vegetation
    └── Pure NumPy + CPU

Key Rule: GPU (CuPy) code ONLY in main process. Workers never import gpu_manager.

---

🎮 Side Activity: Dwarf Fortress Gamelog Analysis

Analyzed 102,669-line gamelog from fort "Lashderush (Prophethandle)":

• 8-9 years, 23 migrant waves, 1,943 masterpieces
• 51,599 combat actions, only 4 deaths (weredeer outbreak)
• Top crafter: Momuz Nëkorlibash (201 masterpieces)
• Sole survivor transforms between dwarf/weredeer

---

🔮 Next Steps

Immediate

• GPU acceleration ✅ COMPLETE
• SRTM terrain data integration (higher accuracy than current tiles)
• Session history persistence across app restarts

Short Term

• Multi-station dashboard
• Project export/import (JSON)
• Link budget analysis view

Medium Term

• LimeSDR hardware integration testing
• Real RF validation against field measurements
• 3D visualization mode

---

💡 Key Learnings

1. Python for-loops are the enemy — _batch_elevation_lookup went from 7.3s to 0.04s by replacing enumerate(zip()) with NumPy indexing
2. Spatial pre-filtering is massive — vegetation bbox check eliminated 95%+ of polygon tests
3. GPU context can't be shared across processes — spawn mode creates new CUDA contexts that OOM
4. Vectorize in main, distribute to workers — best pattern for GPU + multiprocessing
5. Profile before optimizing — Phase 2.6 bottleneck was invisible until measured

---

Session duration: ~4 hours Lines of code changed: ~300 Performance gain: 17.4x Feeling: 🚀