Installation

Quick install

Conda (recommended)

conda install -c conda-forge -c set3mah simcoon

pip

pip install simcoon

Pre-built wheels are available for Linux (x86_64, aarch64), macOS (arm64), and Windows (x64) on Python 3.10–3.14.

BLAS/LAPACK and OpenMP

Simcoon uses Armadillo for linear algebra, which in turn relies on a BLAS/LAPACK implementation. The threading model of the BLAS library matters because it can conflict with OpenMP if both are loaded in the same process.

OpenMP is enabled on Linux and macOS for parallel batch operations. On Windows, OpenMP is disabled in the Python bindings to avoid conflicts between different OpenMP runtimes (e.g. vcomp140.dll from MSVC and runtimes from other packages). Batch operations on Windows run sequentially while BLAS handles internal threading.

BLAS/OpenMP summary
Platform	BLAS	OpenMP	Conflict risk
macOS	Apple Accelerate	ON (libomp, bundled in wheel)	Duplicate libomp when mixed with conda packages (see below)
Linux	System OpenBLAS	ON (libgomp)	None
Windows	vcpkg OpenBLAS (pip) / netlib+MKL (conda)	OFF	None

Duplicate OpenMP runtimes on macOS

The rule: one environment = one OpenMP runtime. In a conda environment that runtime is conda-forge’s llvm-openmp (libomp.dylib), and every native package must share it. Two setups break the rule:

a PyPI wheel that bundles its own libomp (PyTorch wheels do; the simcoon wheel does, via delocate) next to conda numpy/scipy/simcoon;
a source build of simcoon linked against an Armadillo that does not come from the environment (a system or Homebrew copy): its OpenBLAS drags a second libomp into the process.

Symptoms range from the explicit abort message (OMP: Error #15: Initializing libomp.dylib, but found libomp.dylib already initialized) to silent crashes (SIGSEGV inside libomp worker threads under threaded runs).

What to do:

conda-forge for everything native: simcoon, numpy, scipy, armadillo and PyTorch (conda install -c conda-forge pytorch, not pip install torch). Do not mix in Homebrew or system libraries.
Source builds in a conda environment pick the environment’s Armadillo automatically (CMake detects CONDA_PREFIX) and warn when they do not. Install it first and purge any stale build cache when switching:
```
conda install -c conda-forge armadillo
rm -rf build/           # a cached non-conda Armadillo path would be kept otherwise
pip install -e . --no-build-isolation
```
Check which runtimes a process really loads:
```
python -m simcoon.doctor
```
It imports numpy, scipy, simcoon and torch (if present), lists the OpenMP runtimes each one brings and exits with 1 when there are several.

KMP_DUPLICATE_LIB_OK=TRUE is not a remedy: it silences the guard and leaves two runtimes fighting (crashes or silently wrong results remain possible, this is Intel’s own warning).

Using MKL with conda on Linux

If you prefer Intel MKL for performance, switch the BLAS backend and set the threading layer to avoid conflicts between libiomp5 (Intel) and libgomp (GCC):

conda install libblas=*=*mkl mkl
export MKL_THREADING_LAYER=GNU

Developer installation

Prerequisites (conda)

conda create --name simcoon_dev
conda activate simcoon_dev

Linux:

conda env update -f environment.yml

macOS (Apple Silicon):

conda env update -f environment_arm64.yml

Windows:

conda env update -f environment_win.yml

Prerequisites (system packages)

Debian/Ubuntu:

sudo apt-get install libarmadillo-dev libopenblas-dev liblapack-dev \
    libgtest-dev ninja-build cmake

macOS: use the conda environment (environment_arm64.yml above: conda-forge Armadillo, Accelerate BLAS, cmake, ninja). Do not install the dependencies with Homebrew: a Homebrew Armadillo links Homebrew’s OpenBLAS and libomp, i.e. a second OpenMP runtime next to the environment’s (see Duplicate OpenMP runtimes on macOS above).

Windows (vcpkg):

vcpkg install armadillo:x64-windows openblas:x64-windows

Building from source

git clone https://github.com/3MAH/simcoon.git
cd simcoon
pip install -e . --no-build-isolation

This builds the C++ library and Python bindings in one step using scikit-build-core.

Enabling OpenMP (optional, for conda environments):

pip install -e . --no-build-isolation \
    --config-settings=cmake.define.SIMCOON_USE_OPENMP=ON

Running tests

Python tests:

pytest

C++ tests:

mkdir build && cd build
cmake .. -DSIMCOON_BUILD_TESTS=ON -DCMAKE_BUILD_TYPE=Release
cmake --build .
ctest --output-on-failure

Using simcoon with fedoo

Simcoon is designed to work with fedoo for finite-element simulations. Both packages can be installed together:

# conda
conda install -c conda-forge -c set3mah simcoon fedoo

# pip
pip install simcoon fedoo

Keep both packages (and numpy/scipy) on the same channel — all-conda or all-PyPI — and the BLAS/OpenMP setup described above ensures they coexist without runtime conflicts (on macOS, mixing channels can load a second OpenMP runtime; see Duplicate OpenMP runtimes on macOS).