Current status of pressio-linalg API

[cwschilly]

This issue will document the current API of functions implemented in the repo and compare it to the API of other Python libraries (np.linalg, dask).

function	pressio-linalg	numpy	dask
max of a distributed array	max(a, axis=None, comm=None)	np.max(a, axis=None, out=None, keepdims=<no value>, initial=<no value>, where=<no value>)	dask.array.max(a, axis=None, keepdims=False, split_every=None, out=None)
min of a distributed array	min(a, axis=None, comm=None)	np.min(a, axis=None, out=None, keepdims=<no value>, initial=<no value>, where=<no value>)	dask.array.min(a, axis=None, keepdims=False, split_every=None, out=None)
mean of a distributed array	mean(a, axis=None, dtype=None, comm=None)	np.mean(a, axis=None, dtype=None, out=None, keepdims=, *, where=)	dask.array.mean(a, axis=None, dtype=None, keepdims=False, split_every=None, out=None)
standard deviation	std(a, axis=None, dtype=None, ddof=0, comm=None)	np.std(a, axis=None, dtype=None, out=None, ddof=0, keepdims=, *, where=)	dask.array.std(a, axis=None, dtype=None, keepdims=False, ddof=0, split_every=None, out=None)
matrix product 1	product(flag1, flag2, alpha, A, B, beta, C, comm=None)	np.matmul(x1, x2, /, out=None, *, casting='same_kind', order='K', dtype=None, subok=True[, signature, extobj, axes, axis])	dask.array.matmul(x1, x2, /, out=None, *, casting='same_kind', order='K', dtype=None, subok=True[, signature, extobj, axes, axis])
SVD (method of snapshots) 2	thin_svd(M, comm=None, method='auto')	np.linalg.svd(a, full_matrices=True, compute_uv=True, hermitian=False)	dask.array.linalg.svd(a, coerce_signs=True)

Footnotes

1. The public-facing name for the matrix product function has not been finalized ↩

2. The np and dask API are for standard SVD (not using the method of snapshots) ↩

[fnrizzi]

what is this:

...(vec, comm)

?

[cwschilly]

@fnrizzi The function names in the repo are somewhat long, so I left them out to keep things cleaner. I added a clarifying statement to the original comment. Would you prefer that I add them in? (for example, this would be _basic_min_via_python(vec, comm))

[fnrizzi]

@cwschilly those are the internal names of the functions, which we do not want to expose.
We need to write above the PUBLIC names of the functions

[cwschilly]

@fnrizzi Okay, I have added in the current public-facing function names

[fnrizzi]

some suggestions:

• dist_max -> max : and similar for min, we don't need to specify the dist_ since this implied by the fact that these are functions accepting a communicator
• A^T B: need to think about this
• svd_method_of_snapshots: need to think about this too, maybe svd_via_method_of_snapshots. There is no counterpart of this , because effectively this is internally just svd(A^T B) or similar. we callit svd_via_method_of_snapshots just as syntactic sugar

[fnrizzi]

At_dot_b(A, b, comm)

this we need to change into:

product(flag1, flag2, alpha, A, B, beta, A)

whre flag1 and flag2 indicate whether to transpose or not A and B, and alpoha and betaj are coefficients.
So we would have for example:

product('T', 'N', 1, A, A, 0, C)

to indicate we want C = A^T A

But wait, i need to think about this first to see how all this fits into the other repo

[cwschilly]

@fnrizzi This change has been made to main; I'll update the chart above to reflect the new API