johnh2o2 / cuvarbase

Python library for fast time-series analysis on CUDA GPUs

cuvarbase

John Hoffman (c) 2017

cuvarbase is a Python library that uses PyCUDA to implement several time series tools used in astronomy on GPUs.

See the documentation.

This project is under active development, and currently includes implementations of

• Generalized Lomb Scargle periodogram
• Box-least squares (BLS )
• Non-equispaced fast Fourier transform (adjoint operation) (NFFT paper)
• Conditional entropy period finder (CE)

• Phase dispersion minimization (PDM2)

  • Currently operational but minimal unit testing or documentation (yet)

Hopefully future developments will have

• (Weighted) wavelet transforms
• Spectrograms (for PDM and GLS)
• Multiharmonic extensions for GLS

Dependencies

• PyCUDA <-essential

• scikit cuda <-also essential

  • used for access to the CUDA FFT runtime library

• matplotlib (for plotting utilities)
• nfft (for unit testing)
• astropy (for unit testing)

Using multiple GPUs

If you have more than one GPU, you can choose which one to use in a given script by setting the CUDA_DEVICE environment variable:

CUDA_DEVICE=1 python script.py

If anyone is interested in implementing multi-device load-balancing solution, they are encouraged to do so! At some point this may become important, but for the time being manually splitting up the jobs to different GPU's will have to suffice.