Features

What are GPUVMEM's current features?

1. Object oriented Framework: Object orientation makes easier the creation and the building and the plugging of a feature that you are interested in.
2. Multi-GPU: Radio datasets (frequency channels) in GPU memory are distributed between different GPU using the P2P approach.
3. Multi-Field/Mosaic support: Mosaic images in interferometry allow the study of large scale objects in the sky. In this Bayesian approach we fit model images to the ensemble of all pointings.
4. Parameterized reconstruction: Although GPUVMEM requires a few parameters, it does not require user assistance during algorithm iteration as CLEAN. In this sense, we say GPUVMEM is an unsupervised image synthesis algorithm.
5. Regular grid-to-irregular grid approach: An interpolation step is done to compute model visibilities from the 2D Fast Fourier transform of the image estimate to compare them directly with observed visibilities.
6. Gridding to speed up reconstructions: Gridding is the process of resampling visibility data into a regular grid. Usually this is done convolving the data with a kernel to avoid aliasing. Although convolutional gridding kernels are not yet implemented in GPUVMEM, the gridding process reduce the amount of data and speeds up the algorithm reducing computation time.
7. Briggs weighting scheme for gridding: When gridding is used, users can add the robust parameter to vary the visibility weighting between uniform weighting (R=-2.0) and natural weighting (R=2.0).
8. Degridding to study residual visibilities with CASA: After convergence, if gridding is being used the algorithm will use a degridding by bilinear interpolation such that residual visibilities can be studied using NRAO CASA just like in CLEAN.
9. Multi-frequency support: Spectral dependency can introduce strong effects into image synthesis. In this implementation we use the following image representation for multi-frequency synthesis:

where is the image at reference frequency and alpha is the spectral index image.

10. Use of multiple priors/regularization terms: Since this problem does not have a unique solution. Image/s can be regularized with different terms as:

• Entropy

Where is the minimum value for your pixels and the term enables the image to have non-positive values such that pixels could be bounded by:

By default, the eta value is:

• Laplacian term

where and are the image pixel sizes.

• Total Variation

• Quadratic Penalization

• Total Squared Variation

• L1 Norm

11. Different gradient-based optimization algorithms to solve the inverse problem:

• Conjugate Gradient (Polak-Ribiere).
• Limited Memory Broyden–Fletcher–Goldfarb–Shanno (LBFGS).