(feat): use np.zeros for buffer creation with fill_value=0

[ilan-gold]

I was profiling code and noticed a non-trivial amount of time spent on np.full with a fill_value=0 so I did a little digging and at least on my machine:

In [1]: import numpy as np

In [2]: %timeit np.full((10_000, 10_000), dtype=np.float32, fill_value=0, order="C")
33.2 ms ± 784 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [3]: %timeit np.zeros((10_000, 10_000), dtype=np.float32, order="C")
1.04 μs ± 1.48 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

For example:

zarr-python/src/zarr/core/array.py

Lines 1285 to 1290 in 481550a

	out_buffer = prototype.nd_buffer.create(
	shape=indexer.shape,
	dtype=out_dtype,
	order=self.order,
	fill_value=self.metadata.fill_value,
	)

and

zarr-python/src/zarr/codecs/sharding.py

Lines 441 to 443 in 481550a

	out = chunk_spec.prototype.nd_buffer.create(
	shape=shard_shape, dtype=shard_spec.dtype, order=shard_spec.order, fill_value=0
	)

are both points where this happens

TODO:

• Add unit tests and/or doctests in docstrings
• Add docstrings and API docs for any new/modified user-facing classes and functions
• New/modified features documented in docs/user-guide/*.rst
• Changes documented as a new file in changes/
• GitHub Actions have all passed
• Test coverage is 100% (Codecov passes)

[d-v-b]

out of curiosity, is there any performance difference between np.zeros(...)[:] = fill_value and np.full(..., fill_value)? I would expect not but I have no idea.

[ilan-gold]

Not much :/

In [1]: import numpy as np
In [2]: %timeit d = np.zeros((10_000, 10_000), dtype=np.float32, order="C"); d[:] = 1
35.1 ms ± 934 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [3]: %timeit np.full((10_000, 10_000), dtype=np.float32, fill_value=1, order="C")
37.7 ms ± 2.15 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

[ilan-gold]

@d-v-b I think the failing test is unrelated to this PR. I'm also seeing it locally on main