You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
masque/masque/utils/pack2d.py

246 lines
10 KiB
Python

"""
2D bin-packing
"""
from typing import Sequence, Callable, Mapping
import numpy
from numpy.typing import NDArray, ArrayLike
from ..error import MasqueError
from ..pattern import Pattern
def maxrects_bssf(
rects: ArrayLike,
containers: ArrayLike,
presort: bool = True,
allow_rejects: bool = True,
) -> tuple[NDArray[numpy.float64], set[int]]:
"""
Pack rectangles `rects` into regions `containers` using the "maximal rectangles best short side fit"
algorithm (maxrects_bssf) from "A thousand ways to pack the bin", Jukka Jylanki, 2010.
This algorithm gives the best results, but is asymptotically slower than `guillotine_bssf_sas`.
Args:
rects: Nx2 array of rectangle sizes `[[x_size0, y_size0], ...]`.
containers: Mx4 array of regions into which `rects` will be placed, specified using their
corner coordinates ` [[x_min0, y_min0, x_max0, y_max0], ...]`.
presort: If `True` (default), largest-shortest-side rectangles will be placed
first. Otherwise, they will be placed in the order provided.
allow_rejects: If `False`, `MasqueError` will be raised if any rectangle cannot be placed.
Returns:
`[[x_min0, y_min0], ...]` placement locations for `rects`, with the same ordering.
The second argument is a set of indicies of `rects` entries which were rejected; their
corresponding placement locations should be ignored.
Raises:
MasqueError if `allow_rejects` is `True` but some `rects` could not be placed.
"""
regions = numpy.array(containers, copy=False, dtype=float)
rect_sizes = numpy.array(rects, copy=False, dtype=float)
rect_locs = numpy.zeros_like(rect_sizes)
rejected_inds = set()
if presort:
rotated_sizes = numpy.sort(rect_sizes, axis=1) # shortest side first
rect_order = numpy.lexsort(rotated_sizes.T)[::-1] # Descending shortest side
rect_sizes = rect_sizes[rect_order]
for rect_ind, rect_size in enumerate(rect_sizes):
''' Remove degenerate regions '''
# First remove duplicate regions (but keep one; code below would drop both)
regions = numpy.unique(regions, axis=0)
# Now remove regions enclosed in another
min_more = (regions[None, :, :2] >= regions[:, None, :2]).all(axis=2) # first axis > second axis
max_less = (regions[None, :, 2:] <= regions[:, None, 2:]).all(axis=2) # first axis < second axis
max_less &= ~numpy.eye(regions.shape[0], dtype=bool) # exclude self
degenerate = (min_more & max_less).any(axis=0)
regions = regions[~degenerate]
''' Place the rect '''
# Best short-side fit (bssf) to pick a region
region_sizes = regions[:, 2:] - regions[:, :2]
bssf_scores = (region_sizes - rect_size).min(axis=1).astype(float)
bssf_scores[bssf_scores < 0] = numpy.inf # doesn't fit!
rr = bssf_scores.argmin()
if numpy.isinf(bssf_scores[rr]):
if allow_rejects:
rejected_inds.add(rect_ind)
continue
else:
raise MasqueError(f'Failed to find a suitable location for rectangle {rect_ind}')
# Read out location
loc = regions[rr, :2]
rect_locs[rect_ind] = loc
''' Shatter regions '''
# Which regions does this rectangle intersect?
min_over = regions[:, :2] >= loc + rect_size
max_undr = regions[:, 2:] <= loc
intersects = ~(min_over | max_undr).any(axis=1)
# Which sides is there excess on?
region_past_botleft = intersects[:, None] & (regions[:, :2] < loc)
region_past_topright = intersects[:, None] & (regions[:, 2:] > loc + rect_size)
# Create new regions
r_lft = regions[region_past_botleft[:, 0]].copy()
r_bot = regions[region_past_botleft[:, 1]].copy()
r_rgt = regions[region_past_topright[:, 0]].copy()
r_top = regions[region_past_topright[:, 1]].copy()
r_lft[:, 2] = loc[0]
r_bot[:, 3] = loc[1]
r_rgt[:, 0] = loc[0] + rect_size[0]
r_top[:, 1] = loc[1] + rect_size[1]
regions = numpy.vstack((regions[~intersects], r_lft, r_bot, r_rgt, r_top))
if presort:
unsort_order = rect_order.argsort()
rect_locs = rect_locs[unsort_order]
rejected_inds = set(unsort_order[list(rejected_inds)])
return rect_locs, rejected_inds
def guillotine_bssf_sas(
rects: ArrayLike,
containers: ArrayLike,
presort: bool = True,
allow_rejects: bool = True,
) -> tuple[NDArray[numpy.float64], set[int]]:
"""
Pack rectangles `rects` into regions `containers` using the "guillotine best short side fit with
shorter axis split rule" algorithm (guillotine-BSSF-SAS) from "A thousand ways to pack the bin",
Jukka Jylanki, 2010.
This algorithm gives the worse results than `maxrects_bssf`, but is asymptotically faster.
# TODO consider adding rectangle-merge?
# TODO guillotine could use some additional testing
Args:
rects: Nx2 array of rectangle sizes `[[x_size0, y_size0], ...]`.
containers: Mx4 array of regions into which `rects` will be placed, specified using their
corner coordinates ` [[x_min0, y_min0, x_max0, y_max0], ...]`.
presort: If `True` (default), largest-shortest-side rectangles will be placed
first. Otherwise, they will be placed in the order provided.
allow_rejects: If `False`, `MasqueError` will be raised if any rectangle cannot be placed.
Returns:
`[[x_min0, y_min0], ...]` placement locations for `rects`, with the same ordering.
The second argument is a set of indicies of `rects` entries which were rejected; their
corresponding placement locations should be ignored.
Raises:
MasqueError if `allow_rejects` is `True` but some `rects` could not be placed.
"""
regions = numpy.array(containers, copy=False, dtype=float)
rect_sizes = numpy.array(rects, copy=False, dtype=float)
rect_locs = numpy.zeros_like(rect_sizes)
rejected_inds = set()
if presort:
rotated_sizes = numpy.sort(rect_sizes, axis=1) # shortest side first
rect_order = numpy.lexsort(rotated_sizes.T)[::-1] # Descending shortest side
rect_sizes = rect_sizes[rect_order]
for rect_ind, rect_size in enumerate(rect_sizes):
''' Place the rect '''
# Best short-side fit (bssf) to pick a region
region_sizes = regions[:, 2:] - regions[:, :2]
bssf_scores = (region_sizes - rect_size).min(axis=1).astype(float)
bssf_scores[bssf_scores < 0] = numpy.inf # doesn't fit!
rr = bssf_scores.argmin()
if numpy.isinf(bssf_scores[rr]):
if allow_rejects:
rejected_inds.add(rect_ind)
continue
else:
raise MasqueError(f'Failed to find a suitable location for rectangle {rect_ind}')
# Read out location
loc = regions[rr, :2]
rect_locs[rect_ind] = loc
region_size = region_sizes[rr]
split_horiz = region_size[0] < region_size[1]
new_region0 = regions[rr].copy()
new_region1 = new_region0.copy()
split_vertex = loc + rect_size
if split_horiz:
new_region0[2] = split_vertex[0]
new_region0[1] = split_vertex[1]
new_region1[0] = split_vertex[0]
else:
new_region0[3] = split_vertex[1]
new_region0[0] = split_vertex[0]
new_region1[1] = split_vertex[1]
regions = numpy.vstack((regions[:rr], regions[rr + 1:],
new_region0, new_region1))
if presort:
unsort_order = rect_order.argsort()
rect_locs = rect_locs[unsort_order]
rejected_inds = set(unsort_order[list(rejected_inds)])
return rect_locs, rejected_inds
def pack_patterns(
library: Mapping[str, Pattern],
patterns: Sequence[str],
containers: ArrayLike,
spacing: tuple[float, float],
presort: bool = True,
allow_rejects: bool = True,
packer: Callable = maxrects_bssf,
) -> tuple[Pattern, list[str]]:
"""
Pick placement locations for `patterns` inside the regions specified by `containers`.
No rotations are performed.
Args:
library: Library from which `Pattern` objects will be drawn.
patterns: Sequence of pattern names which are to be placed.
containers: Mx4 array of regions into which `patterns` will be placed, specified using their
corner coordinates ` [[x_min0, y_min0, x_max0, y_max0], ...]`.
spacing: (x, y) spacing between adjacent patterns. Patterns are effectively expanded outwards
by `spacing / 2` prior to placement, so this also affects pattern position relative to
container edges.
presort: If `True` (default), largest-shortest-side rectangles will be placed
first. Otherwise, they will be placed in the order provided.
allow_rejects: If `False`, `MasqueError` will be raised if any rectangle cannot be placed.
packer: Bin-packing method; see the other functions in this module (namely `maxrects_bssf`
and `guillotine_bssf_sas`).
Returns:
A `Pattern` containing one `Ref` for each entry in `patterns`.
A list of "rejected" pattern names, for which a valid placement location could not be found.
Raises:
MasqueError if `allow_rejects` is `True` but some `rects` could not be placed.
"""
half_spacing = numpy.array(spacing, copy=False, dtype=float) / 2
bounds = [library[pp].get_bounds() for pp in patterns]
sizes = [bb[1] - bb[0] + spacing if bb is not None else spacing for bb in bounds]
offsets = [half_spacing - bb[0] if bb is not None else (0, 0) for bb in bounds]
locations, reject_inds = packer(sizes, containers, presort=presort, allow_rejects=allow_rejects)
pat = Pattern()
for pp, oo, loc in zip(patterns, offsets, locations):
pat.ref(pp, offset=oo + loc)
rejects = [patterns[ii] for ii in reject_inds]
return pat, rejects