CUDA-based KPMΒΆ

CUDA enables the execution of general purpose code on Nvidia GPUs. It can be used to accelerate computational algorithms which feature natural parallelism. Pybinding features experimental support for CUDA. It’s used for kernel polynomial method (KPM) calculations – see the Green’s function tutorial page and API reference.

The CUDA-base KPM implementation is available via the greens.kpm_cuda() function. It mirrors the API of the regular CPU-based greens.kpm(). The only difference between them is where the calculation will take place. Note that the CUDA implementation is still experimental and that only diagonal Green’s function elements will be computed on the GPU, while off-diagonal falls back to regular CPU code. This will be addressed in a future version.

By default, CUDA support is disabled. You will need to turn it on manually by recompiling the package. First, ensure that you have CUDA Toolkit 7.5 or newer installed. Next, remove any existing Pybinding installation by executing the following command in terminal:

pip3 uninstall pybinding

Finally, reinstall it with CUDA turned on:

PB_CUDA=ON pip3 install pybinding --no-binary pybinding

Note that pybinding is written twice. This is not a mistake. The --no-binary pybinding flag tells pip to compile from source. Since this is all experimental: expect errors and no support.