PERF: fix SparseArray._simple_new object initialization

[jorisvandenbossche]

Apart from this being more idiomatic, it also avoids creating a SparseArray through the normal machinery (including validation of the input etc) for the empty list.

With this PR:

In [1]: data = np.array([1, 2, 3], dtype=float) In [2]: index = pd.core.arrays.sparse.IntIndex(5, np.array([0, 2, 4])) In [3]: dtype = pd.SparseDtype("float64", 0) In [4]: pd.arrays.SparseArray._simple_new(data, index, dtype) Out[4]: [1.0, 0, 2.0, 0, 3.0] Fill: 0 IntIndex Indices: array([0, 2, 4], dtype=int32) In [5]: %timeit pd.arrays.SparseArray._simple_new(data, index, dtype) 381 ns ± 4.83 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each) 

while on released version this gives around 50µs (100x slower)

Noticed while investigating #32196

[rth]

Thanks @jorisvandenbossche! Quick benchmark result below when running pd.DataFrame.sparse.from_spmatrix on a random sparse CSR array of given n_samples, n_features with a density=0.01,

label master (s) PR (s) n_samples n_features 100 100000 14.5374 10.1247 10000 10000 1.5599 1.1037 100000 100 0.0180 0.0134 

so overall this makes that method around 30% faster.

[simonjayhawkins]

Thanks @jorisvandenbossche lgtm

[jorisvandenbossche]

@rth thanks for the timings! Yes, it was indeed a large part of the original slow from_spmatrix, the snippet in the issue does most of the rest

[jreback]

so we have asvs for this? also add a whats new note

[rth]

Note added in #32825 that should work for both PRs I think.