Skip to content

Build from Source

Build Methods

NovaPhy can be built in two ways:

  1. Python package (recommended) — via pip install -e . with scikit-build-core
  2. Standalone CMake — for C++ projects or custom integration

CMake Presets

NovaPhy provides two CMake presets:

Preset Description
default Standard build without IPC
ipc Full build with IPC + CUDA support 
# Using presets with pip
CMAKE_ARGS="--preset=default" pip install -e .
CMAKE_ARGS="--preset=ipc" pip install -e .

Build Flags

Flag Default Description
NOVAPHY_WITH_IPC OFF Enable IPC contact via libuipc (requires CUDA)
NOVAPHY_WITH_VBD_CUDA OFF Enable VBD CUDA backend
NOVAPHY_WITH_SPH_CUDA OFF Enable SPH CUDA backend; CPU SPH is always built
CMAKE_CUDA_COMPILER auto Path to nvcc
CMAKE_TOOLCHAIN_FILE — Path to vcpkg toolchain

Standalone CMake Build

For using NovaPhy in a C++ project:

mkdir build && cd build
cmake -S .. -B . --preset=default --install-prefix=/path/to/install
cmake --build .
cmake --install . --component core

Or use add_subdirectory:

add_subdirectory(/path/to/NovaPhy)
target_link_libraries(your_target PRIVATE novaphy::core)

Compiler Compatibility

Without IPC

Any C++20-capable compiler works:

  • MSVC 2022+
  • GCC 11+
  • Clang 14+

With IPC (libuipc)

IPC has stricter compiler requirements:

Compiler Status Notes
MSVC 2019+ Works —
GCC 11 Works —
GCC 12 Works —
GCC 13 Works —
GCC 14-15 Fails Incompatible with nvcc-12
GCC 9 Fails libuipc needs <span>, introduced in GCC 10
GCC 10 Fails PSTL in libstdc++ 10 incompatible with oneTBB
GCC 15 Fails urdfdom expects non-standard uint32_t import removed from libstdc++1
Clang 19+ Works With libstdc++ 13
Clang <= 18 Fails libuipc uses alias template deduction unsupported by older Clang2
CUDA 12.4 Works Minimum for the NVIDIA IPC path
CUDA 12.8 Works Required for RTX 50-series / compute_120 targets such as RTX 5070 Ti
CUDA 13 TODO muda expects non-strict dependent name resolution
CoreX toolchain Works Use NOVAPHY_GPU_PLATFORM=COREX

Packaging

Build a wheel package for distribution:

pip install build
CMAKE_ARGS="--preset=default" python -m build

Troubleshooting

pybind11 version conflict

Always use pybind11 from pip, NOT from vcpkg. The vcpkg version conflicts with conda's Python.

CUDA 12 + glibc 2.43

CUDA 12 has conflicting math function implementations with glibc 2.43's noexcept declarations. This breaks strict noexcept checking in Clang. Use GCC as the CUDA host compiler if affected.

Clang with libstdc++

Clang uses libstdc++ by default. Any crash caused by libstdc++ also affects Clang.

  1. Fixed upstream via a patch, but not yet in the current vcpkg baseline (2025.07.25). Add -include cstdint to CXXFLAGS as a workaround. ↩

  2. However, alias template deduction is only required for CUDA targets. If you have another compiler as the CUDA host compiler (e.g. GCC 11), Clang 14+ can still compile non-CUDA C/C++ targets. ↩