David Sinden

I tried this but had conflicts between the version of cuda I have and the version it wanted to install

Can you reproduce this in matlab? I can reproduce this in your notebook. I get sensible results when I lower the cfl number to 0.03 when looking at the `p_max`....

The two-points-per-wavelength rule is to be able to resolve the wave. The simulations seem to be unstable above the a cfl of 0.045, given the huge ´p_max´. In general I...

May investigate additional options such as - `shuffle` [details](https://docs.h5py.org/en/latest/high/dataset.html#shuffle-filter) for speed-up in available compression options. - checksums [details](https://docs.h5py.org/en/latest/high/dataset.html#fletcher32-filter) - scaleoffset for saving large datasets of indices such as p_source_index: [details](https://docs.h5py.org/en/latest/high/dataset.html#scale-offset-filter)

> When large running 3D simulations repeatedly with the exact same source geometry I can save the precomputed mask, weights and indices to a h5 file and load them up...

In `simulation_options` the argument `data_cast` is either either `off`, `single` or `double`. I think setting a default as `np.float64` with another accepted option being `np.float32`.

Seems to replicate the matlab behaviour in [writeMatrix](https://github.com/ucl-bug/k-wave/blob/main/k-Wave/writeMatrix.m#L129)

It may be useful to see how this has been done for with [Feltor](https://github.com/feltor-dev/xFELTOR) and look at this presentation [here](https://nbviewer.org/github/TomNicholas/CPSFR_xarray_talk/blob/master/CPSFR_xarray.ipynb) This is implemented in [open_pyfus](https://github.com/OpenwaterHealth/open_pyfus) Other possible options, more focused...

What would be the best approach? There are a few things which can be done to accelerate the code: refactoring with list comprehension to remove loops; joblib; JIT with numba;...

Closed as will be covered by another PR.