Feature Description

The frequency domain beamformers load the frequency-specific CSM from the respective freq_data object in a loop through all frequencies and allocate a new array each time using the numpy.array() function.

csm = array(self.freq_data.csm[i], dtype='complex128') (1)

If caching is used, it is necessary for the array to be loaded into memory from the hdf file. Doing this via

csm = self.freq_data.csm[i] (2)

would be slightly faster (about +10% speed for loading even though a new array is allocated).

If caching is deactivated, loading is 5 times faster with the current implementation anyway, but option (2) is then about 8 times as fast as option (1), as no allocation needs to take place.

The casting to complex128 is only needed for the numba implementations downstream if the CSM was intentionally calculated with only complex64 precision (not sure anybody ever uses this). Default for the CSM is complex128 (despite the docstring stating otherwise!)

It may be worth considering dropping the forced new allocation+casting in the loop. If complex64 support is necessary, a different numba function could be set to be called on a more general level.

Use Case

Every time frequency domain bf is done with a lot of frequencies, this would make calculations a little quicker.

Importance

Low: Nice to have, but not critical.