Design Document

In which, the decisions for why things are the way they are are explained...

Computing the RIME involves calculating a 4D space of values, baseline x channel x timestep x source, and then summing (reducing) over the source dimension to produce a baseline x channel x timestep space of visibilities.

At present, the ordering is baseline x channel x timestep x source. The main reason for placing source as the last dimension is so that a parallel reduction can be computed on it, to produce the visibilities.

Problem dimensions

• na - Number of Antenna
• nbl - Number of Baselines
• nchan - Number of Channels
• ntime - Number of Timesteps
• npsrc - Number of Point Sources
• ngsrc - Number of Gaussian Sources
• nsrc - Number of Sources (Point and Gaussian Combined)

Subdividing the problem

In general, time and channel are easy to parallelise over.

Input

• UVW coordinates, 3 x nbl x ntime, real.
• LM coordinates, 2 x nsrc, real.
• brightness, 5 x ntime x nsrc, real.
• wavelength, 1 x nchan, real.
• pointing errors, 2 x na x ntime, real.
• gaussian shape, 3 x ngsrc, real.
• noise vector 4 x nbl x nchan x ntime, real.
• bayesian data, 4 x nbl x nchan x ntime, complex.

Intermediate Arrays

• scalar jones terms, na x nchan x ntime x nsrc , complex. Holds the result of the phase terms multiplied by the beam terms.
• chi squared terms, nbl x nchan x ntime , real. Holds the result of the sum of
  • the square of the real and imaginary components of the the four polarisations, obtained from the
  • difference of the visibilities (model) and the bayesian data

Output

• visibilities 4 x nbl x nchan x ntime