CUDArray is a CUDA-accelerated subset of the NumPy library. The goal of CUDArray is to combine the easy of development from the NumPy with the computational power of Nvidia GPUs in a lightweight and extensible framework.

CUDArray currently imposes many limitations in order to span a manageable subset of the NumPy library. Nonetheless, it supports a neural network pipeline as demonstrated in the project deeppy.

• Drop-in replacement for NumPy (limitations apply).
• Fast array operations based on cuBLAS, cuRAND and cuDNN.
• (somewhat) Simple C++/CUDA wrapper based on Cython.
• Extends NumPy with specialized functions for neural networks.
• CPU fall-back when CUDA is not available.

First, you should consider specifying the following environment variables.

• INSTALL_PREFIX (default: /usr/local). Path where to install libcudarray. For the Anaconda Python distribution this should be /path/to/anaconda.
• CUDA_PREFIX (default: /usr/local/cuda). Path to the CUDA SDK organized in bin/, lib/, include/ folders.
• CUDNN_ENABLED. Set CUDNN_ENABLED to 1 to include cuDNN operations in libcudarray.

Then build and install libcudarray with

make
make install

Finally, install the cudarray Python package:

python setup.py install

Install the cudarray Python package:

python setup.py --without-cuda install

Please consult the technical report for now. Proper documentation is on the TODO list.

Feel free to report an issue for feature requests and bug reports.

For a more informal chat, visit #cudarray on the freenode IRC network.

If you use CUDArray for research, please cite the technical report:

@techreport{larsen2014cudarray,
  author = "Larsen, Anders Boesen Lindbo",
  title = "{CUDArray}: {CUDA}-based {NumPy}",
  institution = "Department of Applied Mathematics and Computer Science, Technical University of Denmark",
  year = "2014",
  number = "DTU Compute 2014-21",
}

• Proper transpose support,
• Add functionality for copying from NumPy array to existing CUDArray array.
• FFT module based on cuFFT.
• Unit tests!
• Add documentation to wiki.
• Windows/OS X support.

Thanks to the following projects for inspiration.

• cudamat
• PyCUDA
• mshadow
• Caffe
• CUDPP