Skip to content

Latest commit

 

History

History
44 lines (27 loc) · 1.53 KB

refactor_202510.md

File metadata and controls

44 lines (27 loc) · 1.53 KB

🔄 Refactor Summary (October 2025)

This note summarizes the refactor made in October 2025.

  • 🔧 Reliability: Custom CUDA kernels replace problematic Thrust dependency
  • 🎯 Per-Object Parameters: Parameters can now be assigned per object instead of globally
  • 📐 Code Quality: Better code organization and centralized parameter management

🎮 Custom CUDA Kernels

NVIDIA Thrust library reports runtime errors with driver version 560.94+. A set of custom CUDA kernels are implemented as workaround in /src/cpp/kernels/:

  • reduce.cu / reduce.hpp - Reduction operations ⚡
  • exclusive_scan.cu / exclusive_scan.hpp - Prefix sum operations 🔢
  • vec_ops.cu / vec_ops.hpp - Vector operations 📊

Now the use-thrust option has been removed. Our internal benchmarks show that the new kernels perform as efficiently as thrust for our specific use cases. ⚡

🧩 Parameter System Refactor

  • Parameters can now be defined per object (e.g., Young's modulus, contact gaps) instead of globally 🎯
  • ⚠️ This introduces breaking changes to our Python APIs

🐳 Docker Improvements

  • Python packages now install locally to ~/.local/share/ppf-cts/venv instead of globally 🐍

📊 Code Metrics

Component Before After Change
Rust source 3,426 lines 3,891 lines +13.6% 📈
Python frontend 6,161 lines 7,908 lines +28.4% 🚀
CUDA source ~15KB ~39KB +160% 🎮

Last Updated: October 3, 2025 📅