Memory, caching and large data

Size grows quickly

An order-\(k\) square has \(4^k\) cells. Coordinates alone take 16 bytes per cell as two 64-bit integers. These figures exclude the data image, transforms, plotting objects, caches, and temporary arrays.

Order	Side	Cells	Coordinate array alone
8	256	65,536	1 MiB
10	1,024	1,048,576	16 MiB
12	4,096	16,777,216	256 MiB
14	16,384	268,435,456	4 GiB

plot_curve() automatically changes from line to filled cells at effective order 8. Forcing millions of line segments can be expensive. Direct single-cell indexing avoids allocating a complete path, but plotting and array conversion necessarily materialize data.

The cell limit

A guard rejects accidental oversized materialization. It is a ceiling per request, not a memory reservation or a guarantee of safe total peak memory. Its default is 2**28 cells.

import hilbertplot as hp

previous = hp.config.cell_limit()
try:
    hp.config.set_cell_limit(4**10)
    points = hp.curves.points(0, order=5)
finally:
    hp.config.set_cell_limit(previous)

Only raise the limit when you have budgeted the intermediates and outputs. None restores the default. It does not disable the guard. The setting is process-global, so coordinate changes across threads in applications that share it.

Inspect and release cached construction data

stats = hp.config.cache_info()
print(stats.hits, stats.misses, stats.currsize)
hp.config.clear_cache()

The generation cache memoizes recursive stroke arrays shared across curves. cache_info() returns the standard functools cache statistics; its count is entries, not bytes. clear_cache() clears those cached arrays. It does not close figures, free arrays that your own code still references, or guarantee a decrease in the operating system’s RSS metric.

For loops creating figures, save and close each one:

import matplotlib.pyplot as plt

for curve in [0, 1, 12]:
    fig = hp.plot_curve(curve, order=3)
    fig.savefig(f"curve-{curve}.svg")
    plt.close(fig)

Exact conversion versus display

to_grid() / from_grid() preserve dtype and masks. plot() converts real samples to floating point; color rendering is not an archival encoding of the original values. Averaging, truncation, Fourier preprocessing and color normalization each have a different meaning. Conversion guarantees · Transform details