See vcpkg.json for a versioned list of our external dependencies, available via vcpkg.

Note: Use of vcpkg is not mandatory, as long as external dependencies have compatible versions and are discoverable by CMake's find_package mechanism.

CUDA

PyPI

pbatoolkit-gpu (downloaded from PyPI) requires dynamically linking to an instance of the

CUDA 12 Runtime library, and your
CUDA Driver.

Note: Recall that the CUDA Runtime is ABI compatible up to major version.

On 64-bit Windows, these are cudart64_12.dll and nvcuda.dll. Ensure that they are discoverable via Windows' DLL search order. We recommend adding <drive>:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.<minor>\bin (i.e. the binary folder of your CUDA Toolkit installation) to the PATH environment variable. The driver should already be on the search path by default after installation.

On Linux, they are libcudart.so.12 and libcuda.so.1. Ensure that they are discoverable via Linux's dynamic linker/loader. If they are not already in a default search path, we recommend simply updating the library search path, i.e. export LD_LIBRARY_PATH="path/to/driver/folder;path/to/runtime/folder;$LD_LIBRARY_PATH".

Note: MacOS does not support CUDA GPUs.

Our pbatoolkit-gpu prebuilt binaries include PTX, such that program load times will be delayed by JIT compilation on first use. Verify that your NVIDIA GPU supports compute capability at least 7.0. For example, only RTX 2060 up to 4090 chips are supported in the GeForce series. Runtime GPU performance may be constrained by the targeted compute capability.

Previous	Next
Build	Configuration