agstack / field-carbon-model Goto Github PK

e.g., the SMAP L4SM User Guide, the L4C User Guide, the MERRA-2 documentation.

Create a REFERENCES file with relevant papers.

As discussed in recent commit...

The (P x N x T) data cube, and (N x T) sub-arrays, made sense during design phase because it allows for easy indexing of P variables. However, because of the way NumPy arrays are stored in memory and how they are broadcast, it makes more sense to work with (T x N) sub-arrays. They can be broadcast with (N,) arrays like vectorized parameters and summing the rows of (T, N) is more efficient than summing the columns of (N, T).

This isn't just an issue for performance (it is that, too) but is also necessary to fix for model code that involves vectorized parameters.

This leaves us with two options:

1. Change the drivers API so that (P x T x N) arrays are expected, instead.
2. Reshape driver arrays internally or use the transpose, a.T, where needed. Reshaping and transposing are both really fast because they just return views of the base array.