[arrow] improve test coverage and error handling

examples/generate_gds_perf.py

@@ -0,0 +1,5 @@
+from masque.file.gdsii_perf import main
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())

examples/profile_gdsii_readers.py

@@ -0,0 +1,131 @@
+from __future__ import annotations
+
+import argparse
+import importlib
+import json
+import time
+from pathlib import Path
+from typing import Any
+
+from masque import LibraryError
+
+
+READERS: dict[str, tuple[str, tuple[str, ...]]] = {
+    'gdsii': ('masque.file.gdsii', ('readfile',)),
+    'gdsii_arrow': ('masque.file.gdsii_arrow', ('readfile', 'arrow_import', 'arrow_convert')),
+    }
+
+
+def _summarize_library(path: Path, elapsed_s: float, info: dict[str, object], lib: object) -> dict[str, object]:
+    assert hasattr(lib, '__len__')
+    assert hasattr(lib, 'tops')
+    tops = lib.tops()      # type: ignore[no-any-return, attr-defined]
+    try:
+        unique_top = lib.top()      # type: ignore[no-any-return, attr-defined]
+    except LibraryError:
+        unique_top = None
+
+    return {
+        'path': str(path),
+        'elapsed_s': elapsed_s,
+        'library_name': info['name'],
+        'cell_count': len(lib),     # type: ignore[arg-type]
+        'topcells': tops,
+        'topcell': unique_top,
+        }
+
+
+def _summarize_arrow_import(path: Path, elapsed_s: float, arrow_arr: Any) -> dict[str, object]:
+    libarr = arrow_arr[0]
+    return {
+        'path': str(path),
+        'elapsed_s': elapsed_s,
+        'arrow_rows': len(arrow_arr),
+        'library_name': libarr['lib_name'].as_py(),
+        'cell_count': len(libarr['cells']),
+        'layer_count': len(libarr['layers']),
+        }
+
+
+def _profile_stage(module: Any, stage: str, path: Path) -> dict[str, object]:
+    start = time.perf_counter()
+
+    if stage == 'readfile':
+        lib, info = module.readfile(path)
+        elapsed_s = time.perf_counter() - start
+        return _summarize_library(path, elapsed_s, info, lib)
+
+    if stage == 'arrow_import':
+        if hasattr(module, 'readfile_arrow'):
+            libarr, _info = module.readfile_arrow(path)
+            elapsed_s = time.perf_counter() - start
+            return {
+                'path': str(path),
+                'elapsed_s': elapsed_s,
+                'arrow_rows': 1,
+                'library_name': libarr['lib_name'].as_py(),
+                'cell_count': len(libarr['cells']),
+                'layer_count': len(libarr['layers']),
+                }
+
+        arrow_arr = module._read_to_arrow(path)
+        elapsed_s = time.perf_counter() - start
+        return _summarize_arrow_import(path, elapsed_s, arrow_arr)
+
+    if stage == 'arrow_convert':
+        arrow_arr = module._read_to_arrow(path)
+        libarr = arrow_arr[0]
+        start = time.perf_counter()
+        lib, info = module.read_arrow(libarr)
+        elapsed_s = time.perf_counter() - start
+        return _summarize_library(path, elapsed_s, info, lib)
+
+    raise ValueError(f'Unsupported stage {stage!r}')
+
+
+def build_arg_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(description='Profile GDS readers with a stable end-to-end workload.')
+    parser.add_argument('--reader', choices=sorted(READERS), required=True)
+    parser.add_argument('--stage', default='readfile')
+    parser.add_argument('--path', type=Path, required=True)
+    parser.add_argument('--warmup', type=int, default=1)
+    parser.add_argument('--repeat', type=int, default=1)
+    parser.add_argument('--output-json', type=Path)
+    return parser
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = build_arg_parser()
+    args = parser.parse_args(argv)
+
+    module_name, stages = READERS[args.reader]
+    if args.stage not in stages:
+        parser.error(f'reader {args.reader!r} only supports stages: {", ".join(stages)}')
+
+    module = importlib.import_module(module_name)
+    path = args.path.expanduser().resolve()
+
+    for _ in range(args.warmup):
+        _profile_stage(module, args.stage, path)
+
+    runs = []
+    for _ in range(args.repeat):
+        runs.append(_profile_stage(module, args.stage, path))
+
+    payload = {
+        'reader': args.reader,
+        'stage': args.stage,
+        'warmup': args.warmup,
+        'repeat': args.repeat,
+        'runs': runs,
+        }
+    rendered = json.dumps(payload, indent=2, sort_keys=True)
+    if args.output_json is not None:
+        args.output_json.parent.mkdir(parents=True, exist_ok=True)
+        args.output_json.write_text(rendered + '\n')
+    print(rendered)
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())

masque/__init__.py

@@ -42,6 +42,7 @@ from .error import (
 from .shapes import (
     Shape as Shape,
     Polygon as Polygon,
+    RectCollection as RectCollection,
     Path as Path,
     Circle as Circle,
     Arc as Arc,

masque/file/gdsii.py

@@ -37,7 +37,7 @@ from klamath import records
 
 from .utils import is_gzipped, tmpfile
 from .. import Pattern, Ref, PatternError, LibraryError, Label, Shape
-from ..shapes import Polygon, Path
+from ..shapes import Polygon, Path, RectCollection
 from ..repetition import Grid
 from ..utils import layer_t, annotations_t
 from ..library import LazyLibrary, Library, ILibrary, ILibraryView
@@ -323,26 +323,40 @@ def _gpath_to_mpath(gpath: klamath.library.Path, raw_mode: bool) -> tuple[layer_
     else:
         raise PatternError(f'Unrecognized path type: {gpath.path_type}')
 
-    mpath = Path(
+    vertices = gpath.xy.astype(float)
-        vertices=gpath.xy.astype(float),
+    annotations = _properties_to_annotations(gpath.properties)
+    cap_extensions = None
+    if cap == Path.Cap.SquareCustom:
+        cap_extensions = numpy.asarray(gpath.extension, dtype=float)
+
+    if raw_mode:
+        mpath = Path._from_raw(
+            vertices=vertices,
             width=gpath.width,
             cap=cap,
-        offset=numpy.zeros(2),
+            cap_extensions=cap_extensions,
-        annotations=_properties_to_annotations(gpath.properties),
+            annotations=annotations,
-        raw=raw_mode,
+            )
+    else:
+        mpath = Path(
+            vertices=vertices,
+            width=gpath.width,
+            cap=cap,
+            cap_extensions=cap_extensions,
+            offset=numpy.zeros(2),
+            annotations=annotations,
             )
-    if cap == Path.Cap.SquareCustom:
-        mpath.cap_extensions = gpath.extension
     return gpath.layer, mpath
 
 
 def _boundary_to_polygon(boundary: klamath.library.Boundary, raw_mode: bool) -> tuple[layer_t, Polygon]:
-    return boundary.layer, Polygon(
+    vertices = boundary.xy[:-1].astype(float)
-        vertices=boundary.xy[:-1].astype(float),
+    annotations = _properties_to_annotations(boundary.properties)
-        offset=numpy.zeros(2),
+    if raw_mode:
-        annotations=_properties_to_annotations(boundary.properties),
+        poly = Polygon._from_raw(vertices=vertices, annotations=annotations)
-        raw=raw_mode,
+    else:
-        )
+        poly = Polygon(vertices=vertices, offset=numpy.zeros(2), annotations=annotations)
+    return boundary.layer, poly
 
 
 def _mrefs_to_grefs(refs: dict[str | None, list[Ref]]) -> list[klamath.library.Reference]:
@@ -466,6 +480,20 @@ def _shapes_to_elements(
                     properties=properties,
                     )
                 elements.append(path)
+            elif isinstance(shape, RectCollection):
+                for rect in shape.rects:
+                    xy_closed = numpy.empty((5, 2), dtype=numpy.int32)
+                    xy_closed[0] = rint_cast((rect[0], rect[1]))
+                    xy_closed[1] = rint_cast((rect[0], rect[3]))
+                    xy_closed[2] = rint_cast((rect[2], rect[3]))
+                    xy_closed[3] = rint_cast((rect[2], rect[1]))
+                    xy_closed[4] = xy_closed[0]
+                    boundary = klamath.elements.Boundary(
+                        layer=(layer, data_type),
+                        xy=xy_closed,
+                        properties=properties,
+                        )
+                    elements.append(boundary)
             elif isinstance(shape, Polygon):
                 polygon = shape
                 xy_closed = numpy.empty((polygon.vertices.shape[0] + 1, 2), dtype=numpy.int32)

masque/file/gdsii_arrow.py

@@ -0,0 +1,882 @@
+# ruff: noqa: ARG001, F401
+"""
+GDSII file format readers and writers using the `TODO` library.
+
+Note that GDSII references follow the same convention as `masque`,
+  with this order of operations:
+   1. Mirroring
+   2. Rotation
+   3. Scaling
+   4. Offset and array expansion (no mirroring/rotation/scaling applied to offsets)
+
+  Scaling, rotation, and mirroring apply to individual instances, not grid
+   vectors or offsets.
+
+Notes:
+ * absolute positioning is not supported
+ * PLEX is not supported
+ * ELFLAGS are not supported
+ * GDS does not support library- or structure-level annotations
+ * GDS creation/modification/access times are set to 1900-01-01 for reproducibility.
+ * Gzip modification time is set to 0 (start of current epoch, usually 1970-01-01)
+
+ TODO writing
+ TODO warn on boxes, nodes
+"""
+from typing import IO, cast, Any
+from collections.abc import Iterable, Mapping, Callable
+from importlib.machinery import EXTENSION_SUFFIXES
+import importlib.util
+import mmap
+import logging
+import os
+import pathlib
+import gzip
+import string
+import sys
+import tempfile
+from pprint import pformat
+
+from klamath.basic import KlamathError
+import numpy
+from numpy.typing import ArrayLike, NDArray
+import pyarrow
+from pyarrow.cffi import ffi
+
+from .utils import is_gzipped, tmpfile
+from .. import Pattern, Ref, PatternError, LibraryError, Label, Shape
+from ..shapes import Polygon, Path, PolyCollection, RectCollection
+from ..repetition import Grid
+from ..utils import layer_t, annotations_t
+from ..library import LazyLibrary, Library, ILibrary, ILibraryView
+
+
+logger = logging.getLogger(__name__)
+
+ffi.cdef(
+    """
+    const char* last_error_message(void);
+    int read_path(const char* path, struct ArrowArray* array, struct ArrowSchema* schema);
+    int scan_bytes(uint8_t* data, size_t size, struct ArrowArray* array, struct ArrowSchema* schema);
+    int read_cells_bytes(
+        uint8_t* data,
+        size_t size,
+        uint64_t* ranges,
+        size_t range_count,
+        struct ArrowArray* array,
+        struct ArrowSchema* schema
+    );
+    """
+)
+
+clib: Any | None = None
+
+
+path_cap_map = {
+    0: Path.Cap.Flush,
+    1: Path.Cap.Circle,
+    2: Path.Cap.Square,
+    4: Path.Cap.SquareCustom,
+    }
+
+
+def rint_cast(val: ArrayLike) -> NDArray[numpy.int32]:
+    return numpy.rint(val).astype(numpy.int32)
+
+
+def _packed_layer_u32_to_pairs(values: NDArray[numpy.unsignedinteger[Any]]) -> NDArray[numpy.int16]:
+    layer = (values >> numpy.uint32(16)).astype(numpy.uint16).view(numpy.int16)
+    dtype = (values & numpy.uint32(0xffff)).astype(numpy.uint16).view(numpy.int16)
+    return numpy.stack((layer, dtype), axis=-1)
+
+
+def _packed_counts_u32_to_pairs(values: NDArray[numpy.unsignedinteger[Any]]) -> NDArray[numpy.int64]:
+    a_count = (values >> numpy.uint32(16)).astype(numpy.uint16).astype(numpy.int64)
+    b_count = (values & numpy.uint32(0xffff)).astype(numpy.uint16).astype(numpy.int64)
+    return numpy.stack((a_count, b_count), axis=-1)
+
+
+def _packed_xy_u64_to_pairs(values: NDArray[numpy.unsignedinteger[Any]]) -> NDArray[numpy.int32]:
+    xx = (values >> numpy.uint64(32)).astype(numpy.uint32).view(numpy.int32)
+    yy = (values & numpy.uint64(0xffff_ffff)).astype(numpy.uint32).view(numpy.int32)
+    return numpy.stack((xx, yy), axis=-1)
+
+
+def _local_library_filename() -> str:
+    if sys.platform.startswith('linux'):
+        return 'libklamath_rs_ext.so'
+    if sys.platform == 'darwin':
+        return 'libklamath_rs_ext.dylib'
+    if sys.platform == 'win32':
+        return 'klamath_rs_ext.dll'
+    raise OSError(f'Unsupported platform for klamath_rs_ext: {sys.platform!r}')
+
+
+def _installed_library_candidates() -> list[pathlib.Path]:
+    candidates: list[pathlib.Path] = []
+
+    try:
+        spec = importlib.util.find_spec('klamath_rs_ext.klamath_rs_ext')
+    except ModuleNotFoundError:
+        spec = None
+    if spec is not None and spec.origin is not None:
+        candidates.append(pathlib.Path(spec.origin))
+
+    try:
+        pkg_spec = importlib.util.find_spec('klamath_rs_ext')
+    except ModuleNotFoundError:
+        pkg_spec = None
+    if pkg_spec is not None and pkg_spec.submodule_search_locations is not None:
+        for location in pkg_spec.submodule_search_locations:
+            pkg_dir = pathlib.Path(location)
+            for suffix in EXTENSION_SUFFIXES:
+                candidates.extend(sorted(pkg_dir.glob(f'klamath_rs_ext*{suffix}')))
+
+    return candidates
+
+
+def _repo_library_candidates() -> list[pathlib.Path]:
+    repo_root = pathlib.Path(__file__).resolve().parents[2]
+    library_name = _local_library_filename()
+    return [
+        repo_root / 'klamath-rs' / 'target' / 'release' / library_name,
+        repo_root / 'klamath-rs' / 'target' / 'debug' / library_name,
+        ]
+
+
+def find_klamath_rs_library() -> pathlib.Path | None:
+    env_path = os.environ.get('KLAMATH_RS_EXT_LIB')
+    if env_path:
+        candidate = pathlib.Path(env_path).expanduser()
+        if candidate.exists():
+            return candidate.resolve()
+
+    seen: set[pathlib.Path] = set()
+    for candidate in _installed_library_candidates() + _repo_library_candidates():
+        resolved = candidate.expanduser()
+        if resolved in seen:
+            continue
+        seen.add(resolved)
+        if resolved.exists():
+            return resolved.resolve()
+    return None
+
+
+def is_available() -> bool:
+    return find_klamath_rs_library() is not None
+
+
+def _get_clib() -> Any:
+    global clib
+    if clib is None:
+        lib_path = find_klamath_rs_library()
+        if lib_path is None:
+            raise ImportError(
+                'Could not locate klamath_rs_ext shared library. '
+                'Build klamath-rs with `cargo build --release --manifest-path klamath-rs/Cargo.toml` '
+                'or set KLAMATH_RS_EXT_LIB to the built library path.'
+                )
+        clib = ffi.dlopen(str(lib_path))
+    return clib
+
+
+def _read_annotations(
+        prop_offs: NDArray[numpy.integer[Any]],
+        prop_key: NDArray[numpy.integer[Any]],
+        prop_val: list[str],
+        ee: int,
+        ) -> annotations_t:
+    prop_ii, prop_ff = prop_offs[ee], prop_offs[ee + 1]
+    if prop_ii >= prop_ff:
+        return None
+    return {str(prop_key[off]): [prop_val[off]] for off in range(prop_ii, prop_ff)}
+
+
+def _read_to_arrow(
+        filename: str | pathlib.Path,
+        ) -> pyarrow.Array:
+    path = pathlib.Path(filename).expanduser().resolve()
+    ptr_array = ffi.new('struct ArrowArray[]', 1)
+    ptr_schema = ffi.new('struct ArrowSchema[]', 1)
+    if is_gzipped(path):
+        with gzip.open(path, mode='rb') as src:
+            data = src.read()
+        with tempfile.NamedTemporaryFile(suffix='.gds', delete=False) as tmp_stream:
+            tmp_stream.write(data)
+            tmp_name = tmp_stream.name
+        try:
+            _call_native(_get_clib().read_path(tmp_name.encode(), ptr_array, ptr_schema), 'read_path')
+        finally:
+            pathlib.Path(tmp_name).unlink(missing_ok=True)
+    else:
+        _call_native(_get_clib().read_path(str(path).encode(), ptr_array, ptr_schema), 'read_path')
+    return _import_arrow_array(ptr_array, ptr_schema)
+
+
+def _import_arrow_array(ptr_array: Any, ptr_schema: Any) -> pyarrow.Array:
+    iptr_schema = int(ffi.cast('uintptr_t', ptr_schema))
+    iptr_array = int(ffi.cast('uintptr_t', ptr_array))
+    return pyarrow.Array._import_from_c(iptr_array, iptr_schema)
+
+
+def _call_native(status: int, action: str) -> None:
+    if status == 0:
+        return
+
+    err_ptr = _get_clib().last_error_message()
+    if err_ptr == ffi.NULL:
+        raise KlamathError(f'{action} failed')
+
+    message = ffi.string(err_ptr).decode(errors='replace')
+    raise KlamathError(message)
+
+
+def _scan_buffer_to_arrow(buffer: bytes | mmap.mmap | memoryview) -> pyarrow.Array:
+    ptr_array = ffi.new('struct ArrowArray[]', 1)
+    ptr_schema = ffi.new('struct ArrowSchema[]', 1)
+    buf_view = memoryview(buffer)
+    cbuf = ffi.from_buffer('uint8_t[]', buf_view)
+    _call_native(_get_clib().scan_bytes(cbuf, len(buf_view), ptr_array, ptr_schema), 'scan_bytes')
+    return _import_arrow_array(ptr_array, ptr_schema)
+
+
+def _read_selected_cells_to_arrow(
+        buffer: bytes | mmap.mmap | memoryview,
+        ranges: NDArray[numpy.uint64],
+        ) -> pyarrow.Array:
+    ptr_array = ffi.new('struct ArrowArray[]', 1)
+    ptr_schema = ffi.new('struct ArrowSchema[]', 1)
+    buf_view = memoryview(buffer)
+    cbuf = ffi.from_buffer('uint8_t[]', buf_view)
+    flat_ranges = numpy.require(ranges, dtype=numpy.uint64, requirements=('C_CONTIGUOUS', 'ALIGNED'))
+    cranges = ffi.from_buffer('uint64_t[]', flat_ranges)
+    _call_native(
+        _get_clib().read_cells_bytes(cbuf, len(buf_view), cranges, int(flat_ranges.shape[0]), ptr_array, ptr_schema),
+        'read_cells_bytes',
+    )
+    return _import_arrow_array(ptr_array, ptr_schema)
+
+
+def readfile(
+        filename: str | pathlib.Path,
+        ) -> tuple[Library, dict[str, Any]]:
+    """
+    Read a GDSII file from a path into `masque.Library` / `Pattern` objects.
+
+    Will automatically decompress gzipped files.
+
+    Args:
+        filename: Filename to read.
+
+    For callers that can consume Arrow directly, prefer `readfile_arrow()`
+    to skip Python `Pattern` construction entirely.
+    """
+    arrow_arr = _read_to_arrow(filename)
+    assert len(arrow_arr) == 1
+
+    results = read_arrow(arrow_arr[0])
+
+    return results
+
+
+def readfile_arrow(
+        filename: str | pathlib.Path,
+        ) -> tuple[pyarrow.StructScalar, dict[str, Any]]:
+    """
+    Read a GDSII file into the native Arrow representation without converting
+    it into `masque.Library` / `Pattern` objects.
+
+    This is the lowest-overhead public read path exposed by this module.
+
+    Args:
+        filename: Filename to read.
+
+    Returns:
+        - Arrow struct scalar for the library payload
+        - dict of GDSII library info
+    """
+    arrow_arr = _read_to_arrow(filename)
+    assert len(arrow_arr) == 1
+    libarr = arrow_arr[0]
+    return libarr, _read_header(libarr)
+
+
+def read_arrow(
+        libarr: pyarrow.Array,
+        ) -> tuple[Library, dict[str, Any]]:
+    """
+    # TODO check GDSII file for cycles!
+    Read a gdsii file and translate it into a dict of Pattern objects. GDSII structures are
+     translated into Pattern objects; boundaries are translated into polygons, and srefs and arefs
+     are translated into Ref objects.
+
+    Additional library info is returned in a dict, containing:
+      'name': name of the library
+      'meters_per_unit': number of meters per database unit (all values are in database units)
+      'logical_units_per_unit': number of "logical" units displayed by layout tools (typically microns)
+                                per database unit
+
+    Args:
+        libarr: Arrow library payload as returned by `readfile_arrow()`.
+
+    Returns:
+        - dict of pattern_name:Patterns generated from GDSII structures
+        - dict of GDSII library info
+    """
+    library_info = _read_header(libarr)
+
+    layer_names_np = _packed_layer_u32_to_pairs(libarr['layers'].values.to_numpy())
+    layer_tups = [(int(pair[0]), int(pair[1])) for pair in layer_names_np]
+
+    cell_ids = libarr['cells'].values.field('id').to_numpy()
+    cell_names = libarr['cell_names'].as_py()
+
+    def get_geom(libarr: pyarrow.Array, geom_type: str) -> dict[str, Any]:
+        el = libarr['cells'].values.field(geom_type)
+        elem = dict(
+            offsets = el.offsets.to_numpy(),
+            xy_arr = el.values.field('xy').values.to_numpy().reshape((-1, 2)),
+            xy_off = el.values.field('xy').offsets.to_numpy() // 2,
+            layer_inds = el.values.field('layer').to_numpy(),
+            prop_off = el.values.field('properties').offsets.to_numpy(),
+            prop_key = el.values.field('properties').values.field('key').to_numpy(),
+            prop_val = el.values.field('properties').values.field('value').to_pylist(),
+            )
+        return elem
+
+    def get_boundary_batches(libarr: pyarrow.Array) -> dict[str, Any]:
+        batches = libarr['cells'].values.field('boundary_batches')
+        return dict(
+            offsets = batches.offsets.to_numpy(),
+            layer_inds = batches.values.field('layer').to_numpy(),
+            vert_arr = batches.values.field('vertices').values.to_numpy().reshape((-1, 2)),
+            vert_off = batches.values.field('vertices').offsets.to_numpy() // 2,
+            poly_off = batches.values.field('vertex_offsets').offsets.to_numpy(),
+            poly_offsets = batches.values.field('vertex_offsets').values.to_numpy(),
+            )
+
+    def get_rect_batches(libarr: pyarrow.Array) -> dict[str, Any]:
+        batches = libarr['cells'].values.field('rect_batches')
+        return dict(
+            offsets = batches.offsets.to_numpy(),
+            layer_inds = batches.values.field('layer').to_numpy(),
+            rect_arr = batches.values.field('rects').values.to_numpy().reshape((-1, 4)),
+            rect_off = batches.values.field('rects').offsets.to_numpy() // 4,
+            )
+
+    def get_boundary_props(libarr: pyarrow.Array) -> dict[str, Any]:
+        boundaries = libarr['cells'].values.field('boundary_props')
+        return dict(
+            offsets = boundaries.offsets.to_numpy(),
+            layer_inds = boundaries.values.field('layer').to_numpy(),
+            vert_arr = boundaries.values.field('vertices').values.to_numpy().reshape((-1, 2)),
+            vert_off = boundaries.values.field('vertices').offsets.to_numpy() // 2,
+            prop_off = boundaries.values.field('properties').offsets.to_numpy(),
+            prop_key = boundaries.values.field('properties').values.field('key').to_numpy(),
+            prop_val = boundaries.values.field('properties').values.field('value').to_pylist(),
+            )
+
+    def get_refs(libarr: pyarrow.Array, geom_type: str, has_repetition: bool) -> dict[str, Any]:
+        refs = libarr['cells'].values.field(geom_type)
+        values = refs.values
+        elem = dict(
+            offsets = refs.offsets.to_numpy(),
+            targets = values.field('target').to_numpy(),
+            xy = _packed_xy_u64_to_pairs(values.field('xy').to_numpy()),
+            invert_y = values.field('invert_y').to_numpy(zero_copy_only=False),
+            angle_rad = values.field('angle_rad').to_numpy(),
+            scale = values.field('scale').to_numpy(),
+            )
+        if has_repetition:
+            elem.update(dict(
+                xy0 = _packed_xy_u64_to_pairs(values.field('xy0').to_numpy()),
+                xy1 = _packed_xy_u64_to_pairs(values.field('xy1').to_numpy()),
+                counts = _packed_counts_u32_to_pairs(values.field('counts').to_numpy()),
+                ))
+        return elem
+
+    def get_ref_props(libarr: pyarrow.Array, geom_type: str, has_repetition: bool) -> dict[str, Any]:
+        refs = libarr['cells'].values.field(geom_type)
+        values = refs.values
+        elem = dict(
+            offsets = refs.offsets.to_numpy(),
+            targets = values.field('target').to_numpy(),
+            xy = _packed_xy_u64_to_pairs(values.field('xy').to_numpy()),
+            invert_y = values.field('invert_y').to_numpy(zero_copy_only=False),
+            angle_rad = values.field('angle_rad').to_numpy(),
+            scale = values.field('scale').to_numpy(),
+            prop_off = values.field('properties').offsets.to_numpy(),
+            prop_key = values.field('properties').values.field('key').to_numpy(),
+            prop_val = values.field('properties').values.field('value').to_pylist(),
+            )
+        if has_repetition:
+            elem.update(dict(
+                xy0 = _packed_xy_u64_to_pairs(values.field('xy0').to_numpy()),
+                xy1 = _packed_xy_u64_to_pairs(values.field('xy1').to_numpy()),
+                counts = _packed_counts_u32_to_pairs(values.field('counts').to_numpy()),
+                ))
+        return elem
+
+    txt = libarr['cells'].values.field('texts')
+    texts = dict(
+        offsets = txt.offsets.to_numpy(),
+        layer_inds = txt.values.field('layer').to_numpy(),
+        xy = _packed_xy_u64_to_pairs(txt.values.field('xy').to_numpy()),
+        string = txt.values.field('string').to_pylist(),
+        prop_off = txt.values.field('properties').offsets.to_numpy(),
+        prop_key = txt.values.field('properties').values.field('key').to_numpy(),
+        prop_val = txt.values.field('properties').values.field('value').to_pylist(),
+        )
+
+    elements = dict(
+        srefs = get_refs(libarr, 'srefs', has_repetition=False),
+        arefs = get_refs(libarr, 'arefs', has_repetition=True),
+        sref_props = get_ref_props(libarr, 'sref_props', has_repetition=False),
+        aref_props = get_ref_props(libarr, 'aref_props', has_repetition=True),
+        rect_batches = get_rect_batches(libarr),
+        boundary_batches = get_boundary_batches(libarr),
+        boundary_props = get_boundary_props(libarr),
+        paths = get_geom(libarr, 'paths'),
+        texts = texts,
+        )
+
+    paths = libarr['cells'].values.field('paths')
+    elements['paths'].update(dict(
+        width = paths.values.field('width').fill_null(0).to_numpy(),
+        path_type = paths.values.field('path_type').fill_null(0).to_numpy(),
+        extensions = numpy.stack((
+            paths.values.field('extension_start').fill_null(0).to_numpy(),
+            paths.values.field('extension_end').fill_null(0).to_numpy(),
+            ), axis=-1),
+        ))
+
+    global_args = dict(
+        cell_names = cell_names,
+        layer_tups = layer_tups,
+        )
+
+    mlib = Library()
+    for cc in range(len(libarr['cells'])):
+        name = cell_names[int(cell_ids[cc])]
+        pat = Pattern()
+        _rect_batches_to_rectcollections(pat, global_args, elements['rect_batches'], cc)
+        _boundary_batches_to_polygons(pat, global_args, elements['boundary_batches'], cc)
+        _boundary_props_to_polygons(pat, global_args, elements['boundary_props'], cc)
+        _gpaths_to_mpaths(pat, global_args, elements['paths'], cc)
+        _srefs_to_mrefs(pat, global_args, elements['srefs'], cc)
+        _arefs_to_mrefs(pat, global_args, elements['arefs'], cc)
+        _sref_props_to_mrefs(pat, global_args, elements['sref_props'], cc)
+        _aref_props_to_mrefs(pat, global_args, elements['aref_props'], cc)
+        _texts_to_labels(pat, global_args, elements['texts'], cc)
+        mlib[name] = pat
+
+    return mlib, library_info
+
+
+def _read_header(libarr: pyarrow.Array) -> dict[str, Any]:
+    """
+    Read the file header and create the library_info dict.
+    """
+    library_info = dict(
+        name = libarr['lib_name'].as_py(),
+        meters_per_unit = libarr['meters_per_db_unit'].as_py(),
+        logical_units_per_unit = libarr['user_units_per_db_unit'].as_py(),
+        )
+    return library_info
+
+
+def _srefs_to_mrefs(
+        pat: Pattern,
+        global_args: dict[str, Any],
+        elem: dict[str, Any],
+        cc: int,
+        ) -> None:
+    cell_names = global_args['cell_names']
+    elem_off = elem['offsets']
+    elem_count = elem_off[cc + 1] - elem_off[cc]
+    if elem_count == 0:
+        return
+
+    start = elem_off[cc]
+    stop = elem_off[cc + 1]
+    elem_targets = elem['targets'][start:stop]
+    elem_xy = elem['xy'][start:stop]
+    elem_invert_y = elem['invert_y'][start:stop]
+    elem_angle_rad = elem['angle_rad'][start:stop]
+    elem_scale = elem['scale'][start:stop]
+
+    _append_plain_refs_sorted(
+        pat=pat,
+        cell_names=cell_names,
+        elem_targets=elem_targets,
+        elem_xy=elem_xy,
+        elem_invert_y=elem_invert_y,
+        elem_angle_rad=elem_angle_rad,
+        elem_scale=elem_scale,
+        )
+
+
+def _append_plain_refs_sorted(
+        *,
+        pat: Pattern,
+        cell_names: list[str],
+        elem_targets: NDArray[numpy.integer[Any]],
+        elem_xy: NDArray[numpy.integer[Any]],
+        elem_invert_y: NDArray[numpy.bool_ | numpy.bool],
+        elem_angle_rad: NDArray[numpy.floating[Any]],
+        elem_scale: NDArray[numpy.floating[Any]],
+        ) -> None:
+    elem_count = len(elem_targets)
+    if elem_count == 0:
+        return
+
+    target_start = 0
+    while target_start < elem_count:
+        target_id = int(elem_targets[target_start])
+        target_stop = target_start + 1
+        while target_stop < elem_count and elem_targets[target_stop] == target_id:
+            target_stop += 1
+
+        append_refs = pat.refs[cell_names[target_id]].extend
+        append_refs(
+            Ref._from_raw(
+                offset=elem_xy[ee],
+                mirrored=elem_invert_y[ee],
+                rotation=elem_angle_rad[ee],
+                scale=elem_scale[ee],
+                repetition=None,
+                annotations=None,
+                )
+            for ee in range(target_start, target_stop)
+            )
+
+        target_start = target_stop
+
+
+def _arefs_to_mrefs(
+        pat: Pattern,
+        global_args: dict[str, Any],
+        elem: dict[str, Any],
+        cc: int,
+        ) -> None:
+    cell_names = global_args['cell_names']
+    elem_off = elem['offsets']
+    elem_count = elem_off[cc + 1] - elem_off[cc]
+    if elem_count == 0:
+        return
+
+    start = elem_off[cc]
+    stop = elem_off[cc + 1]
+    elem_targets = elem['targets'][start:stop]
+    elem_xy = elem['xy'][start:stop]
+    elem_invert_y = elem['invert_y'][start:stop]
+    elem_angle_rad = elem['angle_rad'][start:stop]
+    elem_scale = elem['scale'][start:stop]
+    elem_xy0 = elem['xy0'][start:stop]
+    elem_xy1 = elem['xy1'][start:stop]
+    elem_counts = elem['counts'][start:stop]
+
+    if len(elem_targets) == 0:
+        return
+
+    target = None
+    append_ref: Callable[[Ref], Any] | None = None
+    for ee in range(len(elem_targets)):
+        target_id = int(elem_targets[ee])
+        if target != target_id:
+            target = target_id
+            append_ref = pat.refs[cell_names[target_id]].append
+        assert append_ref is not None
+        a_count, b_count = elem_counts[ee]
+        append_ref(Ref._from_raw(
+            offset=elem_xy[ee],
+            mirrored=elem_invert_y[ee],
+            rotation=elem_angle_rad[ee],
+            scale=elem_scale[ee],
+            repetition=Grid._from_raw(a_vector=elem_xy0[ee], b_vector=elem_xy1[ee], a_count=a_count, b_count=b_count),
+            annotations=None,
+            ))
+
+
+def _sref_props_to_mrefs(
+        pat: Pattern,
+        global_args: dict[str, Any],
+        elem: dict[str, Any],
+        cc: int,
+        ) -> None:
+    cell_names = global_args['cell_names']
+    elem_off = elem['offsets']
+    prop_key = elem['prop_key']
+    prop_val = elem['prop_val']
+
+    elem_count = elem_off[cc + 1] - elem_off[cc]
+    if elem_count == 0:
+        return
+
+    elem_slc = slice(elem_off[cc], elem_off[cc] + elem_count + 1)
+    prop_offs = elem['prop_off'][elem_slc]
+    elem_targets = elem['targets'][elem_off[cc]:elem_off[cc + 1]]
+    elem_xy = elem['xy'][elem_off[cc]:elem_off[cc + 1]]
+    elem_invert_y = elem['invert_y'][elem_off[cc]:elem_off[cc + 1]]
+    elem_angle_rad = elem['angle_rad'][elem_off[cc]:elem_off[cc + 1]]
+    elem_scale = elem['scale'][elem_off[cc]:elem_off[cc + 1]]
+
+    for ee in range(elem_count):
+        annotations = _read_annotations(prop_offs, prop_key, prop_val, ee)
+        ref = Ref._from_raw(
+            offset=elem_xy[ee],
+            mirrored=elem_invert_y[ee],
+            rotation=elem_angle_rad[ee],
+            scale=elem_scale[ee],
+            repetition=None,
+            annotations=annotations,
+            )
+        pat.refs[cell_names[int(elem_targets[ee])]].append(ref)
+
+
+def _aref_props_to_mrefs(
+        pat: Pattern,
+        global_args: dict[str, Any],
+        elem: dict[str, Any],
+        cc: int,
+        ) -> None:
+    cell_names = global_args['cell_names']
+    elem_off = elem['offsets']
+    prop_key = elem['prop_key']
+    prop_val = elem['prop_val']
+
+    elem_count = elem_off[cc + 1] - elem_off[cc]
+    if elem_count == 0:
+        return
+
+    elem_slc = slice(elem_off[cc], elem_off[cc] + elem_count + 1)
+    prop_offs = elem['prop_off'][elem_slc]
+    elem_targets = elem['targets'][elem_off[cc]:elem_off[cc + 1]]
+    elem_xy = elem['xy'][elem_off[cc]:elem_off[cc + 1]]
+    elem_invert_y = elem['invert_y'][elem_off[cc]:elem_off[cc + 1]]
+    elem_angle_rad = elem['angle_rad'][elem_off[cc]:elem_off[cc + 1]]
+    elem_scale = elem['scale'][elem_off[cc]:elem_off[cc + 1]]
+    elem_xy0 = elem['xy0'][elem_off[cc]:elem_off[cc + 1]]
+    elem_xy1 = elem['xy1'][elem_off[cc]:elem_off[cc + 1]]
+    elem_counts = elem['counts'][elem_off[cc]:elem_off[cc + 1]]
+
+    for ee in range(elem_count):
+        a_count, b_count = elem_counts[ee]
+        annotations = _read_annotations(prop_offs, prop_key, prop_val, ee)
+        ref = Ref._from_raw(
+            offset=elem_xy[ee],
+            mirrored=elem_invert_y[ee],
+            rotation=elem_angle_rad[ee],
+            scale=elem_scale[ee],
+            repetition=Grid._from_raw(a_vector=elem_xy0[ee], b_vector=elem_xy1[ee], a_count=a_count, b_count=b_count),
+            annotations=annotations,
+            )
+        pat.refs[cell_names[int(elem_targets[ee])]].append(ref)
+
+
+def _texts_to_labels(
+        pat: Pattern,
+        global_args: dict[str, Any],
+        elem: dict[str, Any],
+        cc: int,
+        ) -> None:
+    elem_off = elem['offsets']      # which elements belong to each cell
+    xy = elem['xy']
+    layer_tups = global_args['layer_tups']
+    layer_inds = elem['layer_inds']
+    prop_key = elem['prop_key']
+    prop_val = elem['prop_val']
+
+    elem_count = elem_off[cc + 1] - elem_off[cc]
+    elem_slc = slice(elem_off[cc], elem_off[cc] + elem_count + 1)   # +1 to capture ending location for last elem
+    prop_offs = elem['prop_off'][elem_slc]  # which props belong to each element
+    elem_xy = xy[elem_slc][:elem_count]
+    elem_layer_inds = layer_inds[elem_slc][:elem_count]
+    elem_strings = elem['string'][elem_slc][:elem_count]
+
+    for ee in range(elem_count):
+        layer = layer_tups[int(elem_layer_inds[ee])]
+        offset = elem_xy[ee]
+        string = elem_strings[ee]
+
+        annotations = _read_annotations(prop_offs, prop_key, prop_val, ee)
+        mlabel = Label._from_raw(string=string, offset=offset, annotations=annotations)
+        pat.labels[layer].append(mlabel)
+
+
+def _gpaths_to_mpaths(
+        pat: Pattern,
+        global_args: dict[str, Any],
+        elem: dict[str, Any],
+        cc: int,
+        ) -> None:
+    elem_off = elem['offsets']      # which elements belong to each cell
+    xy_val = elem['xy_arr']
+    layer_tups = global_args['layer_tups']
+    layer_inds = elem['layer_inds']
+    prop_key = elem['prop_key']
+    prop_val = elem['prop_val']
+
+    elem_count = elem_off[cc + 1] - elem_off[cc]
+    elem_slc = slice(elem_off[cc], elem_off[cc] + elem_count + 1)   # +1 to capture ending location for last elem
+    xy_offs = elem['xy_off'][elem_slc]      # which xy coords belong to each element
+    prop_offs = elem['prop_off'][elem_slc]  # which props belong to each element
+    elem_layer_inds = layer_inds[elem_slc][:elem_count]
+    elem_widths = elem['width'][elem_slc][:elem_count]
+    elem_path_types = elem['path_type'][elem_slc][:elem_count]
+    elem_extensions = elem['extensions'][elem_slc][:elem_count]
+
+    for ee in range(elem_count):
+        layer = layer_tups[int(elem_layer_inds[ee])]
+        vertices = xy_val[xy_offs[ee]:xy_offs[ee + 1]]
+        width = elem_widths[ee]
+        cap_int = int(elem_path_types[ee])
+        if cap_int not in path_cap_map:
+            raise PatternError(f'Unrecognized path type: {cap_int}')
+        cap = path_cap_map[cap_int]
+        if cap_int == 4:
+            cap_extensions = elem_extensions[ee]
+        else:
+            cap_extensions = None
+
+        annotations = _read_annotations(prop_offs, prop_key, prop_val, ee)
+        path = Path._from_raw(
+            vertices=vertices,
+            width=width,
+            cap=cap,
+            cap_extensions=cap_extensions,
+            annotations=annotations,
+            )
+        pat.shapes[layer].append(path)
+
+
+def _boundary_batches_to_polygons(
+        pat: Pattern,
+        global_args: dict[str, Any],
+        elem: dict[str, Any],
+        cc: int,
+        ) -> None:
+    elem_off = elem['offsets']      # which elements belong to each cell
+    vert_arr = elem['vert_arr']
+    vert_off = elem['vert_off']
+    layer_inds = elem['layer_inds']
+    layer_tups = global_args['layer_tups']
+    poly_off = elem['poly_off']
+    poly_offsets = elem['poly_offsets']
+
+    batch_count = elem_off[cc + 1] - elem_off[cc]
+    if batch_count == 0:
+        return
+
+    elem_slc = slice(elem_off[cc], elem_off[cc] + batch_count + 1)   # +1 to capture ending location for last elem
+    elem_vert_off = vert_off[elem_slc]
+    elem_poly_off = poly_off[elem_slc]
+    elem_layer_inds = layer_inds[elem_slc][:batch_count]
+
+    for bb in range(batch_count):
+        layer = layer_tups[int(elem_layer_inds[bb])]
+        vertices = vert_arr[elem_vert_off[bb]:elem_vert_off[bb + 1]]
+        vertex_offsets = poly_offsets[elem_poly_off[bb]:elem_poly_off[bb + 1]]
+
+        if vertex_offsets.size == 1:
+            poly = Polygon._from_raw(vertices=vertices, annotations=None)
+            pat.shapes[layer].append(poly)
+        else:
+            polys = PolyCollection._from_raw(vertex_lists=vertices, vertex_offsets=vertex_offsets, annotations=None)
+            pat.shapes[layer].append(polys)
+
+
+def _rect_batches_to_rectcollections(
+        pat: Pattern,
+        global_args: dict[str, Any],
+        elem: dict[str, Any],
+        cc: int,
+        ) -> None:
+    elem_off = elem['offsets']
+    rect_arr = elem['rect_arr']
+    rect_off = elem['rect_off']
+    layer_inds = elem['layer_inds']
+    layer_tups = global_args['layer_tups']
+
+    batch_count = elem_off[cc + 1] - elem_off[cc]
+    if batch_count == 0:
+        return
+
+    elem_slc = slice(elem_off[cc], elem_off[cc] + batch_count + 1)
+    elem_rect_off = rect_off[elem_slc]
+    elem_layer_inds = layer_inds[elem_slc][:batch_count]
+
+    for bb in range(batch_count):
+        layer = layer_tups[int(elem_layer_inds[bb])]
+        rects = rect_arr[elem_rect_off[bb]:elem_rect_off[bb + 1]]
+        rect_collection = RectCollection._from_raw(rects=rects, annotations=None)
+        pat.shapes[layer].append(rect_collection)
+
+
+def _boundary_props_to_polygons(
+        pat: Pattern,
+        global_args: dict[str, Any],
+        elem: dict[str, Any],
+        cc: int,
+        ) -> None:
+    elem_off = elem['offsets']
+    vert_arr = elem['vert_arr']
+    vert_off = elem['vert_off']
+    layer_inds = elem['layer_inds']
+    layer_tups = global_args['layer_tups']
+    prop_key = elem['prop_key']
+    prop_val = elem['prop_val']
+
+    elem_count = elem_off[cc + 1] - elem_off[cc]
+    if elem_count == 0:
+        return
+
+    elem_slc = slice(elem_off[cc], elem_off[cc] + elem_count + 1)
+    elem_vert_off = vert_off[elem_slc]
+    prop_offs = elem['prop_off'][elem_slc]
+    elem_layer_inds = layer_inds[elem_slc][:elem_count]
+
+    for ee in range(elem_count):
+        layer = layer_tups[int(elem_layer_inds[ee])]
+        vertices = vert_arr[elem_vert_off[ee]:elem_vert_off[ee + 1]]
+        annotations = _read_annotations(prop_offs, prop_key, prop_val, ee)
+        poly = Polygon._from_raw(vertices=vertices, annotations=annotations)
+        pat.shapes[layer].append(poly)
+
+
+#def _properties_to_annotations(properties: pyarrow.Array) -> annotations_t:
+#    return {prop['key'].as_py(): prop['value'].as_py() for prop in properties}
+
+
+def check_valid_names(
+        names: Iterable[str],
+        max_length: int = 32,
+        ) -> None:
+    """
+    Check all provided names to see if they're valid GDSII cell names.
+
+    Args:
+        names: Collection of names to check
+        max_length: Max allowed length
+
+    """
+    allowed_chars = set(string.ascii_letters + string.digits + '_?$')
+
+    bad_chars = [
+        name for name in names
+        if not set(name).issubset(allowed_chars)
+        ]
+
+    bad_lengths = [
+        name for name in names
+        if len(name) > max_length
+        ]
+
+    if bad_chars:
+        logger.error('Names contain invalid characters:\n' + pformat(bad_chars))
+
+    if bad_lengths:
+        logger.error(f'Names too long (>{max_length}:\n' + pformat(bad_chars))
+
+    if bad_chars or bad_lengths:
+        raise LibraryError('Library contains invalid names, see log above')

masque/file/gdsii_lazy_arrow.py

@@ -0,0 +1,960 @@
+"""
+Lazy GDSII readers and writers backed by native Arrow scan/materialize paths.
+
+This module is intentionally separate from `gdsii_arrow` so the eager read path
+keeps its current behavior and performance profile.
+"""
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import IO, Any, cast
+from collections import defaultdict
+from collections.abc import Callable, Iterator, Mapping, Sequence
+import copy
+import gzip
+import logging
+import mmap
+import pathlib
+
+import numpy
+from numpy.typing import NDArray
+import pyarrow
+import klamath
+
+from . import gdsii, gdsii_arrow
+from .utils import is_gzipped, tmpfile
+from ..error import LibraryError
+from ..library import ILibrary, ILibraryView, Library, LibraryView, dangling_mode_t
+from ..pattern import Pattern, map_targets
+from ..utils import apply_transforms
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class _StructRange:
+    start: int
+    end: int
+
+
+@dataclass
+class _SourceBuffer:
+    path: pathlib.Path
+    data: bytes | mmap.mmap
+    handle: IO[bytes] | None = None
+
+    def raw_slice(self, start: int, end: int) -> bytes:
+        return self.data[start:end]
+
+
+@dataclass
+class _ScanRefs:
+    offsets: NDArray[numpy.integer[Any]]
+    targets: NDArray[numpy.integer[Any]]
+    xy: NDArray[numpy.int32]
+    xy0: NDArray[numpy.int32]
+    xy1: NDArray[numpy.int32]
+    counts: NDArray[numpy.int64]
+    invert_y: NDArray[numpy.bool_ | numpy.bool]
+    angle_rad: NDArray[numpy.floating[Any]]
+    scale: NDArray[numpy.floating[Any]]
+
+
+@dataclass(frozen=True)
+class _CellScan:
+    cell_id: int
+    struct_range: _StructRange
+    ref_start: int
+    ref_stop: int
+    children: set[str]
+
+
+@dataclass
+class _ScanPayload:
+    libarr: pyarrow.StructScalar
+    library_info: dict[str, Any]
+    cell_names: list[str]
+    cell_order: list[str]
+    cells: dict[str, _CellScan]
+    refs: _ScanRefs
+
+
+@dataclass
+class _SourceLayer:
+    library: ILibraryView
+    source_to_visible: dict[str, str]
+    visible_to_source: dict[str, str]
+    child_graph: dict[str, set[str]]
+    order: list[str]
+
+
+@dataclass(frozen=True)
+class _SourceEntry:
+    layer_index: int
+    source_name: str
+
+
+def is_available() -> bool:
+    return gdsii_arrow.is_available()
+
+
+def _read_header(libarr: pyarrow.StructScalar) -> dict[str, Any]:
+    return gdsii_arrow._read_header(libarr)
+
+
+def _open_source_buffer(path: pathlib.Path) -> _SourceBuffer:
+    if is_gzipped(path):
+        with gzip.open(path, mode='rb') as stream:
+            data = stream.read()
+        return _SourceBuffer(path=path, data=data)
+
+    handle = path.open(mode='rb', buffering=0)
+    mapped = mmap.mmap(handle.fileno(), 0, access=mmap.ACCESS_READ)
+    return _SourceBuffer(path=path, data=mapped, handle=handle)
+
+
+def _extract_scan_payload(libarr: pyarrow.StructScalar) -> _ScanPayload:
+    library_info = _read_header(libarr)
+    cell_names = libarr['cell_names'].as_py()
+
+    cells = libarr['cells']
+    cell_values = cells.values
+    cell_ids = cell_values.field('id').to_numpy()
+    struct_starts = cell_values.field('struct_start_offset').to_numpy()
+    struct_ends = cell_values.field('struct_end_offset').to_numpy()
+
+    refs = cell_values.field('refs')
+    ref_values = refs.values
+    ref_offsets = refs.offsets.to_numpy()
+    targets = ref_values.field('target').to_numpy()
+    xy = gdsii_arrow._packed_xy_u64_to_pairs(ref_values.field('xy').to_numpy())
+    xy0 = gdsii_arrow._packed_xy_u64_to_pairs(ref_values.field('xy0').to_numpy())
+    xy1 = gdsii_arrow._packed_xy_u64_to_pairs(ref_values.field('xy1').to_numpy())
+    counts = gdsii_arrow._packed_counts_u32_to_pairs(ref_values.field('counts').to_numpy())
+    invert_y = ref_values.field('invert_y').to_numpy(zero_copy_only=False)
+    angle_rad = ref_values.field('angle_rad').to_numpy()
+    scale = ref_values.field('scale').to_numpy()
+
+    ref_payload = _ScanRefs(
+        offsets=ref_offsets,
+        targets=targets,
+        xy=xy,
+        xy0=xy0,
+        xy1=xy1,
+        counts=counts,
+        invert_y=invert_y,
+        angle_rad=angle_rad,
+        scale=scale,
+    )
+
+    cell_order = [cell_names[int(cell_id)] for cell_id in cell_ids]
+    cell_scan: dict[str, _CellScan] = {}
+    for cc, name in enumerate(cell_order):
+        ref_start = int(ref_offsets[cc])
+        ref_stop = int(ref_offsets[cc + 1])
+        children = {
+            cell_names[int(target)]
+            for target in targets[ref_start:ref_stop]
+        }
+        cell_scan[name] = _CellScan(
+            cell_id=int(cell_ids[cc]),
+            struct_range=_StructRange(int(struct_starts[cc]), int(struct_ends[cc])),
+            ref_start=ref_start,
+            ref_stop=ref_stop,
+            children=children,
+        )
+
+    return _ScanPayload(
+        libarr=libarr,
+        library_info=library_info,
+        cell_names=cell_names,
+        cell_order=cell_order,
+        cells=cell_scan,
+        refs=ref_payload,
+    )
+
+
+def _pattern_children(pat: Pattern) -> set[str]:
+    return {child for child, refs in pat.refs.items() if child is not None and refs}
+
+
+def _remap_pattern_targets(pat: Pattern, remap: Callable[[str | None], str | None]) -> Pattern:
+    if not pat.refs:
+        return pat
+    pat.refs = map_targets(pat.refs, remap)
+    return pat
+
+
+def _coerce_library_view(source: Mapping[str, Pattern] | ILibraryView) -> ILibraryView:
+    if isinstance(source, ILibraryView):
+        return source
+    return LibraryView(source)
+
+
+def _source_order(source: ILibraryView) -> list[str]:
+    if isinstance(source, ArrowLibrary):
+        return list(source.source_order())
+    return list(source.keys())
+
+
+def _make_ref_rows(
+        xy: NDArray[numpy.integer[Any]],
+        angle_rad: NDArray[numpy.floating[Any]],
+        invert_y: NDArray[numpy.bool_ | numpy.bool],
+        scale: NDArray[numpy.floating[Any]],
+        ) -> NDArray[numpy.float64]:
+    rows = numpy.empty((len(xy), 5), dtype=float)
+    rows[:, :2] = xy
+    rows[:, 2] = angle_rad
+    rows[:, 3] = invert_y.astype(float)
+    rows[:, 4] = scale
+    return rows
+
+
+def _expand_aref_row(
+        xy: NDArray[numpy.integer[Any]],
+        xy0: NDArray[numpy.integer[Any]],
+        xy1: NDArray[numpy.integer[Any]],
+        counts: NDArray[numpy.integer[Any]],
+        angle_rad: float,
+        invert_y: bool,
+        scale: float,
+        ) -> NDArray[numpy.float64]:
+    a_count = int(counts[0])
+    b_count = int(counts[1])
+    aa, bb = numpy.meshgrid(numpy.arange(a_count), numpy.arange(b_count), indexing='ij')
+    displacements = aa.reshape(-1, 1) * xy0[None, :] + bb.reshape(-1, 1) * xy1[None, :]
+    rows = numpy.empty((displacements.shape[0], 5), dtype=float)
+    rows[:, :2] = xy + displacements
+    rows[:, 2] = angle_rad
+    rows[:, 3] = float(invert_y)
+    rows[:, 4] = scale
+    return rows
+
+
+class ArrowLibrary(ILibraryView):
+    """
+    Read-only library backed by the native lazy Arrow scan schema.
+
+    Materializing a cell via `__getitem__` caches a real `Pattern` for that cell.
+    Cached cells are treated as edited for future writes from this module.
+    """
+
+    path: pathlib.Path
+    library_info: dict[str, Any]
+
+    def __init__(
+            self,
+            *,
+            path: pathlib.Path,
+            payload: _ScanPayload,
+            source: _SourceBuffer,
+            ) -> None:
+        self.path = path
+        self.library_info = payload.library_info
+        self._payload = payload
+        self._source = source
+        self._cache: dict[str, Pattern] = {}
+
+    @classmethod
+    def from_file(cls, filename: str | pathlib.Path) -> ArrowLibrary:
+        path = pathlib.Path(filename).expanduser().resolve()
+        source = _open_source_buffer(path)
+        scan_arr = gdsii_arrow._scan_buffer_to_arrow(source.data)
+        assert len(scan_arr) == 1
+        payload = _extract_scan_payload(scan_arr[0])
+        return cls(path=path, payload=payload, source=source)
+
+    def __getitem__(self, key: str) -> Pattern:
+        return self._materialize_pattern(key, persist=True)
+
+    def __iter__(self) -> Iterator[str]:
+        return iter(self._payload.cell_order)
+
+    def __len__(self) -> int:
+        return len(self._payload.cell_order)
+
+    def __contains__(self, key: object) -> bool:
+        return key in self._payload.cells
+
+    def source_order(self) -> tuple[str, ...]:
+        return tuple(self._payload.cell_order)
+
+    def raw_struct_bytes(self, name: str) -> bytes:
+        struct_range = self._payload.cells[name].struct_range
+        return self._source.raw_slice(struct_range.start, struct_range.end)
+
+    def materialize_many(
+            self,
+            names: Sequence[str],
+            *,
+            persist: bool = True,
+            ) -> LibraryView:
+        mats = self._materialize_patterns(names, persist=persist)
+        return LibraryView(mats)
+
+    def _materialize_patterns(
+            self,
+            names: Sequence[str],
+            *,
+            persist: bool,
+            ) -> dict[str, Pattern]:
+        ordered_names = list(dict.fromkeys(names))
+        missing = [name for name in ordered_names if name not in self._payload.cells]
+        if missing:
+            raise KeyError(missing[0])
+
+        materialized: dict[str, Pattern] = {}
+        uncached = [name for name in ordered_names if name not in self._cache]
+        if uncached:
+            ranges = numpy.asarray(
+                [
+                    [
+                        self._payload.cells[name].struct_range.start,
+                        self._payload.cells[name].struct_range.end,
+                    ]
+                    for name in uncached
+                ],
+                dtype=numpy.uint64,
+            )
+            arrow_arr = gdsii_arrow._read_selected_cells_to_arrow(self._source.data, ranges)
+            assert len(arrow_arr) == 1
+            selected_lib, _info = gdsii_arrow.read_arrow(arrow_arr[0])
+            for name in uncached:
+                pat = selected_lib[name]
+                materialized[name] = pat
+                if persist:
+                    self._cache[name] = pat
+
+        for name in ordered_names:
+            if name in self._cache:
+                materialized[name] = self._cache[name]
+        return materialized
+
+    def _materialize_pattern(self, name: str, *, persist: bool) -> Pattern:
+        return self._materialize_patterns((name,), persist=persist)[name]
+
+    def _raw_children(self, name: str) -> set[str]:
+        return set(self._payload.cells[name].children)
+
+    def _collect_raw_transforms(self, cell: _CellScan, target_id: int) -> list[NDArray[numpy.float64]]:
+        refs = self._payload.refs
+        start = cell.ref_start
+        stop = cell.ref_stop
+        if stop <= start:
+            return []
+
+        targets = refs.targets[start:stop]
+        mask = targets == target_id
+        if not mask.any():
+            return []
+
+        rows: list[NDArray[numpy.float64]] = []
+        counts = refs.counts[start:stop]
+        unit_mask = mask & (counts[:, 0] == 1) & (counts[:, 1] == 1)
+        if unit_mask.any():
+            rows.append(_make_ref_rows(
+                refs.xy[start:stop][unit_mask],
+                refs.angle_rad[start:stop][unit_mask],
+                refs.invert_y[start:stop][unit_mask],
+                refs.scale[start:stop][unit_mask],
+            ))
+
+        aref_indices = numpy.nonzero(mask & ~unit_mask)[0]
+        for idx in aref_indices:
+            abs_idx = start + int(idx)
+            rows.append(_expand_aref_row(
+                xy=refs.xy[abs_idx],
+                xy0=refs.xy0[abs_idx],
+                xy1=refs.xy1[abs_idx],
+                counts=refs.counts[abs_idx],
+                angle_rad=float(refs.angle_rad[abs_idx]),
+                invert_y=bool(refs.invert_y[abs_idx]),
+                scale=float(refs.scale[abs_idx]),
+            ))
+        return rows
+
+    def child_graph(
+            self,
+            dangling: dangling_mode_t = 'error',
+            ) -> dict[str, set[str]]:
+        graph: dict[str, set[str]] = {}
+        for name in self._payload.cell_order:
+            if name in self._cache:
+                graph[name] = _pattern_children(self._cache[name])
+            else:
+                graph[name] = self._raw_children(name)
+
+        existing = set(graph)
+        dangling_refs = set().union(*(children - existing for children in graph.values()))
+        if dangling == 'error':
+            if dangling_refs:
+                raise self._dangling_refs_error(cast('set[str]', dangling_refs), 'building child graph')
+            return graph
+        if dangling == 'ignore':
+            return {name: {child for child in children if child in existing} for name, children in graph.items()}
+
+        for child in dangling_refs:
+            graph.setdefault(cast('str', child), set())
+        return graph
+
+    def parent_graph(
+            self,
+            dangling: dangling_mode_t = 'error',
+            ) -> dict[str, set[str]]:
+        child_graph = self.child_graph(dangling='include' if dangling == 'include' else 'ignore')
+        existing = set(self.keys())
+        igraph: dict[str, set[str]] = {name: set() for name in child_graph}
+        for parent, children in child_graph.items():
+            for child in children:
+                if child in existing or dangling == 'include':
+                    igraph.setdefault(child, set()).add(parent)
+        if dangling == 'error':
+            raw = self.child_graph(dangling='include')
+            dangling_refs = set().union(*(children - existing for children in raw.values()))
+            if dangling_refs:
+                raise self._dangling_refs_error(cast('set[str]', dangling_refs), 'building parent graph')
+        return igraph
+
+    def subtree(
+            self,
+            tops: str | Sequence[str],
+            ) -> ILibraryView:
+        if isinstance(tops, str):
+            tops = (tops,)
+        keep = cast('set[str]', self.referenced_patterns(tops) - {None})
+        keep |= set(tops)
+        return self.materialize_many(tuple(keep), persist=True)
+
+    def tops(self) -> list[str]:
+        graph = self.child_graph(dangling='ignore')
+        names = set(graph)
+        not_toplevel: set[str] = set()
+        for children in graph.values():
+            not_toplevel |= children
+        return list(names - not_toplevel)
+
+    def find_refs_local(
+            self,
+            name: str,
+            parent_graph: dict[str, set[str]] | None = None,
+            dangling: dangling_mode_t = 'error',
+            ) -> dict[str, list[NDArray[numpy.float64]]]:
+        instances: dict[str, list[NDArray[numpy.float64]]] = defaultdict(list)
+        if parent_graph is None:
+            graph_mode = 'ignore' if dangling == 'ignore' else 'include'
+            parent_graph = self.parent_graph(dangling=graph_mode)
+
+        if name not in self:
+            if name not in parent_graph:
+                return instances
+            if dangling == 'error':
+                raise self._dangling_refs_error({name}, f'finding local refs for {name!r}')
+            if dangling == 'ignore':
+                return instances
+
+        target_id = self._payload.cells.get(name)
+        for parent in parent_graph.get(name, set()):
+            if parent in self._cache:
+                for ref in self._cache[parent].refs.get(name, []):
+                    instances[parent].append(ref.as_transforms())
+                continue
+
+            if target_id is None or parent not in self._payload.cells:
+                continue
+            rows = self._collect_raw_transforms(self._payload.cells[parent], target_id.cell_id)
+            if rows:
+                instances[parent].extend(rows)
+        return instances
+
+    def find_refs_global(
+            self,
+            name: str,
+            order: list[str] | None = None,
+            parent_graph: dict[str, set[str]] | None = None,
+            dangling: dangling_mode_t = 'error',
+            ) -> dict[tuple[str, ...], NDArray[numpy.float64]]:
+        graph_mode = 'ignore' if dangling == 'ignore' else 'include'
+        if order is None:
+            order = self.child_order(dangling=graph_mode)
+        if parent_graph is None:
+            parent_graph = self.parent_graph(dangling=graph_mode)
+
+        if name not in self:
+            if name not in parent_graph:
+                return {}
+            if dangling == 'error':
+                raise self._dangling_refs_error({name}, f'finding global refs for {name!r}')
+            if dangling == 'ignore':
+                return {}
+
+        self_keys = set(self.keys())
+        transforms: dict[str, list[tuple[tuple[str, ...], NDArray[numpy.float64]]]]
+        transforms = defaultdict(list)
+        for parent, vals in self.find_refs_local(name, parent_graph=parent_graph, dangling=dangling).items():
+            transforms[parent] = [((name,), numpy.concatenate(vals))]
+
+        for next_name in order:
+            if next_name not in transforms:
+                continue
+            if not parent_graph.get(next_name, set()) & self_keys:
+                continue
+
+            outers = self.find_refs_local(next_name, parent_graph=parent_graph, dangling=dangling)
+            inners = transforms.pop(next_name)
+            for parent, outer in outers.items():
+                outer_tf = numpy.concatenate(outer)
+                for path, inner in inners:
+                    combined = apply_transforms(outer_tf, inner)
+                    transforms[parent].append(((next_name,) + path, combined))
+
+        result = {}
+        for parent, targets in transforms.items():
+            for path, instances in targets:
+                full_path = (parent,) + path
+                result[full_path] = instances
+        return result
+
+
+class OverlayLibrary(ILibrary):
+    """
+    Mutable overlay over one or more source libraries.
+
+    Source-backed cells remain lazy until accessed through `__getitem__`, at
+    which point that visible cell is promoted into an overlay-owned materialized
+    `Pattern`.
+    """
+
+    def __init__(self) -> None:
+        self._layers: list[_SourceLayer] = []
+        self._entries: dict[str, Pattern | _SourceEntry] = {}
+        self._order: list[str] = []
+        self._target_remap: dict[str, str] = {}
+
+    def __iter__(self) -> Iterator[str]:
+        return (name for name in self._order if name in self._entries)
+
+    def __len__(self) -> int:
+        return len(self._entries)
+
+    def __contains__(self, key: object) -> bool:
+        return key in self._entries
+
+    def __getitem__(self, key: str) -> Pattern:
+        return self._materialize_pattern(key, persist=True)
+
+    def __setitem__(
+            self,
+            key: str,
+            value: Pattern | Callable[[], Pattern],
+            ) -> None:
+        if key in self._entries:
+            raise LibraryError(f'"{key}" already exists in the library. Overwriting is not allowed!')
+        pattern = value() if callable(value) else value
+        self._entries[key] = pattern
+        if key not in self._order:
+            self._order.append(key)
+
+    def __delitem__(self, key: str) -> None:
+        if key not in self._entries:
+            raise KeyError(key)
+        del self._entries[key]
+
+    def _merge(self, key_self: str, other: Mapping[str, Pattern], key_other: str) -> None:
+        self[key_self] = copy.deepcopy(other[key_other])
+
+    def add_source(
+            self,
+            source: Mapping[str, Pattern] | ILibraryView,
+            *,
+            rename_theirs: Callable[[ILibraryView, str], str] | None = None,
+            ) -> dict[str, str]:
+        view = _coerce_library_view(source)
+        source_order = _source_order(view)
+        child_graph = view.child_graph(dangling='include')
+
+        source_to_visible: dict[str, str] = {}
+        visible_to_source: dict[str, str] = {}
+        rename_map: dict[str, str] = {}
+
+        for name in source_order:
+            visible = name
+            if visible in self._entries or visible in visible_to_source:
+                if rename_theirs is None:
+                    raise LibraryError(f'Conflicting name while adding source: {name!r}')
+                visible = rename_theirs(self, name)
+                if visible in self._entries or visible in visible_to_source:
+                    raise LibraryError(f'Unresolved duplicate key encountered while adding source: {name!r} -> {visible!r}')
+                rename_map[name] = visible
+            source_to_visible[name] = visible
+            visible_to_source[visible] = name
+
+        layer = _SourceLayer(
+            library=view,
+            source_to_visible=source_to_visible,
+            visible_to_source=visible_to_source,
+            child_graph=child_graph,
+            order=[source_to_visible[name] for name in source_order],
+        )
+        layer_index = len(self._layers)
+        self._layers.append(layer)
+
+        for source_name, visible_name in source_to_visible.items():
+            self._entries[visible_name] = _SourceEntry(layer_index=layer_index, source_name=source_name)
+            if visible_name not in self._order:
+                self._order.append(visible_name)
+
+        return rename_map
+
+    def rename(
+            self,
+            old_name: str,
+            new_name: str,
+            move_references: bool = False,
+            ) -> OverlayLibrary:
+        if old_name not in self._entries:
+            raise LibraryError(f'"{old_name}" does not exist in the library.')
+        if old_name == new_name:
+            return self
+        if new_name in self._entries:
+            raise LibraryError(f'"{new_name}" already exists in the library.')
+
+        entry = self._entries.pop(old_name)
+        self._entries[new_name] = entry
+        if isinstance(entry, _SourceEntry):
+            layer = self._layers[entry.layer_index]
+            layer.source_to_visible[entry.source_name] = new_name
+            del layer.visible_to_source[old_name]
+            layer.visible_to_source[new_name] = entry.source_name
+
+        idx = self._order.index(old_name)
+        self._order[idx] = new_name
+
+        if move_references:
+            self.move_references(old_name, new_name)
+        return self
+
+    def _resolve_target(self, target: str) -> str:
+        seen: set[str] = set()
+        current = target
+        while current in self._target_remap:
+            if current in seen:
+                raise LibraryError(f'Cycle encountered while resolving target remap for {target!r}')
+            seen.add(current)
+            current = self._target_remap[current]
+        return current
+
+    def _set_target_remap(self, old_target: str, new_target: str) -> None:
+        resolved_new = self._resolve_target(new_target)
+        if resolved_new == old_target:
+            raise LibraryError(f'Ref target remap would create a cycle: {old_target!r} -> {new_target!r}')
+        self._target_remap[old_target] = resolved_new
+        for key in list(self._target_remap):
+            self._target_remap[key] = self._resolve_target(self._target_remap[key])
+
+    def move_references(self, old_target: str, new_target: str) -> OverlayLibrary:
+        if old_target == new_target:
+            return self
+        self._set_target_remap(old_target, new_target)
+        for entry in list(self._entries.values()):
+            if isinstance(entry, Pattern) and old_target in entry.refs:
+                entry.refs[new_target].extend(entry.refs[old_target])
+                del entry.refs[old_target]
+        return self
+
+    def _effective_target(self, layer: _SourceLayer, target: str) -> str:
+        visible = layer.source_to_visible.get(target, target)
+        return self._resolve_target(visible)
+
+    def _materialize_pattern(self, name: str, *, persist: bool) -> Pattern:
+        if name not in self._entries:
+            raise KeyError(name)
+        entry = self._entries[name]
+        if isinstance(entry, Pattern):
+            return entry
+
+        layer = self._layers[entry.layer_index]
+        source_pat = layer.library[entry.source_name].deepcopy()
+        remap = lambda target: None if target is None else self._effective_target(layer, target)
+        pat = _remap_pattern_targets(source_pat, remap)
+        if persist:
+            self._entries[name] = pat
+        return pat
+
+    def child_graph(
+            self,
+            dangling: dangling_mode_t = 'error',
+            ) -> dict[str, set[str]]:
+        graph: dict[str, set[str]] = {}
+        for name in self._order:
+            if name not in self._entries:
+                continue
+            entry = self._entries[name]
+            if isinstance(entry, Pattern):
+                graph[name] = _pattern_children(entry)
+                continue
+            layer = self._layers[entry.layer_index]
+            children = {self._effective_target(layer, child) for child in layer.child_graph.get(entry.source_name, set())}
+            graph[name] = children
+
+        existing = set(graph)
+        dangling_refs = set().union(*(children - existing for children in graph.values()))
+        if dangling == 'error':
+            if dangling_refs:
+                raise self._dangling_refs_error(cast('set[str]', dangling_refs), 'building child graph')
+            return graph
+        if dangling == 'ignore':
+            return {name: {child for child in children if child in existing} for name, children in graph.items()}
+
+        for child in dangling_refs:
+            graph.setdefault(cast('str', child), set())
+        return graph
+
+    def parent_graph(
+            self,
+            dangling: dangling_mode_t = 'error',
+            ) -> dict[str, set[str]]:
+        child_graph = self.child_graph(dangling='include' if dangling == 'include' else 'ignore')
+        existing = set(self.keys())
+        igraph: dict[str, set[str]] = {name: set() for name in child_graph}
+        for parent, children in child_graph.items():
+            for child in children:
+                if child in existing or dangling == 'include':
+                    igraph.setdefault(child, set()).add(parent)
+        if dangling == 'error':
+            raw = self.child_graph(dangling='include')
+            dangling_refs = set().union(*(children - existing for children in raw.values()))
+            if dangling_refs:
+                raise self._dangling_refs_error(cast('set[str]', dangling_refs), 'building parent graph')
+        return igraph
+
+    def subtree(
+            self,
+            tops: str | Sequence[str],
+            ) -> ILibraryView:
+        if isinstance(tops, str):
+            tops = (tops,)
+        keep = cast('set[str]', self.referenced_patterns(tops) - {None})
+        keep |= set(tops)
+        return LibraryView({name: self[name] for name in keep})
+
+    def find_refs_local(
+            self,
+            name: str,
+            parent_graph: dict[str, set[str]] | None = None,
+            dangling: dangling_mode_t = 'error',
+            ) -> dict[str, list[NDArray[numpy.float64]]]:
+        instances: dict[str, list[NDArray[numpy.float64]]] = defaultdict(list)
+        if parent_graph is None:
+            graph_mode = 'ignore' if dangling == 'ignore' else 'include'
+            parent_graph = self.parent_graph(dangling=graph_mode)
+
+        if name not in self:
+            if name not in parent_graph:
+                return instances
+            if dangling == 'error':
+                raise self._dangling_refs_error({name}, f'finding local refs for {name!r}')
+            if dangling == 'ignore':
+                return instances
+
+        for parent in parent_graph.get(name, set()):
+            pat = self._materialize_pattern(parent, persist=False)
+            for ref in pat.refs.get(name, []):
+                instances[parent].append(ref.as_transforms())
+        return instances
+
+    def find_refs_global(
+            self,
+            name: str,
+            order: list[str] | None = None,
+            parent_graph: dict[str, set[str]] | None = None,
+            dangling: dangling_mode_t = 'error',
+            ) -> dict[tuple[str, ...], NDArray[numpy.float64]]:
+        graph_mode = 'ignore' if dangling == 'ignore' else 'include'
+        if order is None:
+            order = self.child_order(dangling=graph_mode)
+        if parent_graph is None:
+            parent_graph = self.parent_graph(dangling=graph_mode)
+
+        if name not in self:
+            if name not in parent_graph:
+                return {}
+            if dangling == 'error':
+                raise self._dangling_refs_error({name}, f'finding global refs for {name!r}')
+            if dangling == 'ignore':
+                return {}
+
+        self_keys = set(self.keys())
+        transforms: dict[str, list[tuple[tuple[str, ...], NDArray[numpy.float64]]]]
+        transforms = defaultdict(list)
+        for parent, vals in self.find_refs_local(name, parent_graph=parent_graph, dangling=dangling).items():
+            transforms[parent] = [((name,), numpy.concatenate(vals))]
+
+        for next_name in order:
+            if next_name not in transforms:
+                continue
+            if not parent_graph.get(next_name, set()) & self_keys:
+                continue
+
+            outers = self.find_refs_local(next_name, parent_graph=parent_graph, dangling=dangling)
+            inners = transforms.pop(next_name)
+            for parent, outer in outers.items():
+                outer_tf = numpy.concatenate(outer)
+                for path, inner in inners:
+                    combined = apply_transforms(outer_tf, inner)
+                    transforms[parent].append(((next_name,) + path, combined))
+
+        result = {}
+        for parent, targets in transforms.items():
+            for path, instances in targets:
+                result[(parent,) + path] = instances
+        return result
+
+    def source_order(self) -> tuple[str, ...]:
+        return tuple(name for name in self._order if name in self._entries)
+
+
+def readfile(
+        filename: str | pathlib.Path,
+        ) -> tuple[ArrowLibrary, dict[str, Any]]:
+    lib = ArrowLibrary.from_file(filename)
+    return lib, lib.library_info
+
+
+def load_libraryfile(
+        filename: str | pathlib.Path,
+        ) -> tuple[ArrowLibrary, dict[str, Any]]:
+    return readfile(filename)
+
+
+def _get_write_info(
+        library: Mapping[str, Pattern] | ILibraryView,
+        *,
+        meters_per_unit: float | None,
+        logical_units_per_unit: float | None,
+        library_name: str | None,
+        ) -> tuple[float, float, str]:
+    if meters_per_unit is not None and logical_units_per_unit is not None and library_name is not None:
+        return meters_per_unit, logical_units_per_unit, library_name
+
+    infos: list[dict[str, Any]] = []
+    if isinstance(library, ArrowLibrary):
+        infos.append(library.library_info)
+    elif isinstance(library, OverlayLibrary):
+        for layer in library._layers:
+            if isinstance(layer.library, ArrowLibrary):
+                infos.append(layer.library.library_info)
+
+    if infos:
+        unit_pairs = {(info['meters_per_unit'], info['logical_units_per_unit']) for info in infos}
+        if len(unit_pairs) > 1:
+            raise LibraryError('Merged lazy GDS sources must have identical units before writing')
+        info = infos[0]
+        meters = info['meters_per_unit'] if meters_per_unit is None else meters_per_unit
+        logical = info['logical_units_per_unit'] if logical_units_per_unit is None else logical_units_per_unit
+        name = info['name'] if library_name is None else library_name
+        return meters, logical, name
+
+    if meters_per_unit is None or logical_units_per_unit is None or library_name is None:
+        raise LibraryError('meters_per_unit, logical_units_per_unit, and library_name are required for non-GDS-backed lazy writes')
+    return meters_per_unit, logical_units_per_unit, library_name
+
+
+def _can_copy_arrow_cell(library: ArrowLibrary, name: str) -> bool:
+    return name not in library._cache
+
+
+def _can_copy_overlay_cell(library: OverlayLibrary, name: str, entry: _SourceEntry) -> bool:
+    layer = library._layers[entry.layer_index]
+    if not isinstance(layer.library, ArrowLibrary):
+        return False
+    if name != entry.source_name:
+        return False
+    children = layer.child_graph.get(entry.source_name, set())
+    return all(library._effective_target(layer, child) == child for child in children)
+
+
+def _write_pattern_struct(stream: IO[bytes], name: str, pat: Pattern) -> None:
+    elements: list[klamath.elements.Element] = []
+    elements += gdsii._shapes_to_elements(pat.shapes)
+    elements += gdsii._labels_to_texts(pat.labels)
+    elements += gdsii._mrefs_to_grefs(pat.refs)
+    klamath.library.write_struct(stream, name=name.encode('ASCII'), elements=elements)
+
+
+def write(
+        library: Mapping[str, Pattern] | ILibraryView,
+        stream: IO[bytes],
+        *,
+        meters_per_unit: float | None = None,
+        logical_units_per_unit: float | None = None,
+        library_name: str | None = None,
+        ) -> None:
+    meters_per_unit, logical_units_per_unit, library_name = _get_write_info(
+        library,
+        meters_per_unit=meters_per_unit,
+        logical_units_per_unit=logical_units_per_unit,
+        library_name=library_name,
+    )
+
+    header = klamath.library.FileHeader(
+        name=library_name.encode('ASCII'),
+        user_units_per_db_unit=logical_units_per_unit,
+        meters_per_db_unit=meters_per_unit,
+    )
+    header.write(stream)
+
+    if isinstance(library, ArrowLibrary):
+        for name in library.source_order():
+            if _can_copy_arrow_cell(library, name):
+                stream.write(library.raw_struct_bytes(name))
+            else:
+                _write_pattern_struct(stream, name, library._materialize_pattern(name, persist=False))
+        klamath.records.ENDLIB.write(stream, None)
+        return
+
+    if isinstance(library, OverlayLibrary):
+        for name in library.source_order():
+            entry = library._entries[name]
+            if isinstance(entry, _SourceEntry) and _can_copy_overlay_cell(library, name, entry):
+                layer = library._layers[entry.layer_index]
+                assert isinstance(layer.library, ArrowLibrary)
+                stream.write(layer.library.raw_struct_bytes(entry.source_name))
+            else:
+                _write_pattern_struct(stream, name, library._materialize_pattern(name, persist=False))
+        klamath.records.ENDLIB.write(stream, None)
+        return
+
+    gdsii.write(cast('Mapping[str, Pattern]', library), stream, meters_per_unit, logical_units_per_unit, library_name)
+
+
+def writefile(
+        library: Mapping[str, Pattern] | ILibraryView,
+        filename: str | pathlib.Path,
+        *,
+        meters_per_unit: float | None = None,
+        logical_units_per_unit: float | None = None,
+        library_name: str | None = None,
+        ) -> None:
+    path = pathlib.Path(filename)
+
+    with tmpfile(path) as base_stream:
+        streams: tuple[Any, ...] = (base_stream,)
+        if path.suffix == '.gz':
+            stream = cast('IO[bytes]', gzip.GzipFile(filename='', mtime=0, fileobj=base_stream, mode='wb', compresslevel=6))
+            streams = (stream,) + streams
+        else:
+            stream = base_stream
+
+        try:
+            write(
+                library,
+                stream,
+                meters_per_unit=meters_per_unit,
+                logical_units_per_unit=logical_units_per_unit,
+                library_name=library_name,
+            )
+        finally:
+            for ss in streams:
+                ss.close()

masque/file/gdsii_perf.py

@@ -0,0 +1,633 @@
+"""
+Synthetic GDS fixture generation for reader/writer performance testing.
+
+The presets here are intentionally hierarchical and deterministic. They aim to
+approximate a pair of real-world layout families discussed during GDS reader and
+writer work:
+
+* `many_cells`: tens of thousands of cells, moderate reference count, very heavy
+  box usage after flattening, and moderate polygon density.
+* `many_instances`: a much smaller cell library with very high reference count,
+  similar box density, and far fewer polygons.
+
+Fixtures are written by streaming structures through `klamath` directly so large
+benchmark files can be produced without first materializing an equally large
+`masque.Library` in Python.
+"""
+from __future__ import annotations
+
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Any
+import argparse
+import json
+import math
+
+import numpy
+import klamath
+from klamath import elements
+
+
+EMPTY_PROPERTIES: dict[int, bytes] = {}
+METERS_PER_DB_UNIT = 1e-9
+USER_UNITS_PER_DB_UNIT = 1e-3
+TOTAL_LAYERS = 200
+
+
+@dataclass(frozen=True)
+class FixturePreset:
+    name: str
+    total_layers: int
+    box_layers: int
+    heavy_box_layers: int
+    polygon_layers: int
+    box_cells: int
+    poly_cells: int
+    box_wrappers: int
+    poly_wrappers: int
+    box_clusters: int
+    poly_clusters: int
+    box_cluster_refs: int
+    poly_cluster_refs: int
+    top_direct_box_refs: int
+    top_direct_poly_refs: int
+    heavy_boxes_per_cell: int
+    regular_boxes_per_cell: int
+    polygons_per_cell: int
+    path_stride: int
+    text_stride: int
+    box_cluster_array: tuple[int, int]
+    top_box_array: tuple[int, int]
+    poly_cluster_array: tuple[int, int]
+    top_poly_array: tuple[int, int]
+    rare_annotation_stride: int
+
+
+PRESETS: dict[str, FixturePreset] = {
+    'many_cells': FixturePreset(
+        name='many_cells',
+        total_layers=TOTAL_LAYERS,
+        box_layers=20,
+        heavy_box_layers=3,
+        polygon_layers=20,
+        box_cells=17_000,
+        poly_cells=6_000,
+        box_wrappers=18_000,
+        poly_wrappers=6_000,
+        box_clusters=2_000,
+        poly_clusters=999,
+        box_cluster_refs=24,
+        poly_cluster_refs=16,
+        top_direct_box_refs=21_000,
+        top_direct_poly_refs=7_000,
+        heavy_boxes_per_cell=6,
+        regular_boxes_per_cell=2,
+        polygons_per_cell=50,
+        path_stride=2,
+        text_stride=3,
+        box_cluster_array=(24, 16),
+        top_box_array=(8, 8),
+        poly_cluster_array=(4, 2),
+        top_poly_array=(3, 2),
+        rare_annotation_stride=1_250,
+        ),
+    'many_instances': FixturePreset(
+        name='many_instances',
+        total_layers=TOTAL_LAYERS,
+        box_layers=25,
+        heavy_box_layers=3,
+        polygon_layers=10,
+        box_cells=2_500,
+        poly_cells=500,
+        box_wrappers=1_000,
+        poly_wrappers=500,
+        box_clusters=1_000,
+        poly_clusters=499,
+        box_cluster_refs=1_200,
+        poly_cluster_refs=400,
+        top_direct_box_refs=102_001,
+        top_direct_poly_refs=0,
+        heavy_boxes_per_cell=40,
+        regular_boxes_per_cell=16,
+        polygons_per_cell=60,
+        path_stride=1,
+        text_stride=2,
+        box_cluster_array=(1, 1),
+        top_box_array=(1, 1),
+        poly_cluster_array=(1, 1),
+        top_poly_array=(1, 1),
+        rare_annotation_stride=250,
+        ),
+    }
+
+
+@dataclass(frozen=True)
+class FixtureManifest:
+    preset: str
+    scale: float
+    gds_path: str
+    library_name: str
+    cells: int
+    refs: int
+    layers: int
+    box_layers: int
+    heavy_box_layers: list[list[int]]
+    polygon_layers: list[list[int]]
+    hierarchical_boxes_per_heavy_layer: int
+    hierarchical_boxes_per_regular_layer: int
+    hierarchical_polygons_total: int
+    hierarchical_paths_total: int
+    hierarchical_texts_total: int
+    flattened_box_placements: int
+    flattened_poly_placements: int
+    estimated_flat_boxes_per_heavy_layer: int
+    estimated_flat_polygons_per_active_polygon_layer: int
+
+
+def _scaled_count(value: int, scale: float, minimum: int = 0) -> int:
+    if value == 0:
+        return 0
+    scaled = int(math.ceil(value * scale))
+    return max(minimum, scaled)
+
+
+def _scaled_preset(preset: FixturePreset, scale: float) -> FixturePreset:
+    if scale <= 0:
+        raise ValueError(f'scale must be positive, got {scale!r}')
+
+    return FixturePreset(
+        name=preset.name,
+        total_layers=preset.total_layers,
+        box_layers=min(preset.box_layers, preset.total_layers),
+        heavy_box_layers=min(preset.heavy_box_layers, preset.box_layers),
+        polygon_layers=min(preset.polygon_layers, preset.total_layers),
+        box_cells=_scaled_count(preset.box_cells, scale, minimum=1),
+        poly_cells=_scaled_count(preset.poly_cells, scale, minimum=1),
+        box_wrappers=_scaled_count(preset.box_wrappers, scale),
+        poly_wrappers=_scaled_count(preset.poly_wrappers, scale),
+        box_clusters=_scaled_count(preset.box_clusters, scale, minimum=1),
+        poly_clusters=_scaled_count(preset.poly_clusters, scale, minimum=1),
+        box_cluster_refs=_scaled_count(preset.box_cluster_refs, scale, minimum=1),
+        poly_cluster_refs=_scaled_count(preset.poly_cluster_refs, scale, minimum=1),
+        top_direct_box_refs=_scaled_count(preset.top_direct_box_refs, scale),
+        top_direct_poly_refs=_scaled_count(preset.top_direct_poly_refs, scale),
+        heavy_boxes_per_cell=max(1, preset.heavy_boxes_per_cell),
+        regular_boxes_per_cell=max(1, preset.regular_boxes_per_cell),
+        polygons_per_cell=max(1, preset.polygons_per_cell),
+        path_stride=max(1, preset.path_stride),
+        text_stride=max(1, preset.text_stride),
+        box_cluster_array=preset.box_cluster_array,
+        top_box_array=preset.top_box_array,
+        poly_cluster_array=preset.poly_cluster_array,
+        top_poly_array=preset.top_poly_array,
+        rare_annotation_stride=max(1, _scaled_count(preset.rare_annotation_stride, scale, minimum=1)),
+        )
+
+
+def _rect_xy(xmin: int, ymin: int, xmax: int, ymax: int) -> numpy.ndarray[Any, numpy.dtype[numpy.int32]]:
+    return numpy.array(
+        [[xmin, ymin], [xmin, ymax], [xmax, ymax], [xmax, ymin], [xmin, ymin]],
+        dtype=numpy.int32,
+        )
+
+
+def _poly_xy(points: list[tuple[int, int]]) -> numpy.ndarray[Any, numpy.dtype[numpy.int32]]:
+    closed = points + [points[0]]
+    return numpy.array(closed, dtype=numpy.int32)
+
+
+def _sref(
+        target: str,
+        xy: tuple[int, int],
+        properties: dict[int, bytes] | None = None,
+        ) -> elements.Reference:
+    return klamath.library.Reference(
+        struct_name=target.encode('ASCII'),
+        invert_y=False,
+        mag=1.0,
+        angle_deg=0.0,
+        xy=numpy.array([xy], dtype=numpy.int32),
+        colrow=None,
+        properties=EMPTY_PROPERTIES if properties is None else properties,
+        )
+
+
+def _aref(
+        target: str,
+        origin: tuple[int, int],
+        counts: tuple[int, int],
+        step: tuple[int, int],
+        properties: dict[int, bytes] | None = None,
+        ) -> elements.Reference:
+    cols, rows = counts
+    dx, dy = step
+    xy = numpy.array(
+        [
+            origin,
+            (origin[0] + cols * dx, origin[1]),
+            (origin[0], origin[1] + rows * dy),
+            ],
+        dtype=numpy.int32,
+        )
+    return klamath.library.Reference(
+        struct_name=target.encode('ASCII'),
+        invert_y=False,
+        mag=1.0,
+        angle_deg=0.0,
+        xy=xy,
+        colrow=(cols, rows),
+        properties=EMPTY_PROPERTIES if properties is None else properties,
+        )
+
+
+def _annotation(index: int) -> dict[int, bytes]:
+    return {1: f'perf-{index}'.encode('ASCII')}
+
+
+def _make_box_cell(name: str, index: int, cfg: FixturePreset) -> list[elements.Element]:
+    cell_elements: list[elements.Element] = []
+    xbase = (index % 17) * 600
+    ybase = (index // 17) * 180
+
+    for layer in range(cfg.heavy_box_layers):
+        for box_idx in range(cfg.heavy_boxes_per_cell):
+            x0 = xbase + box_idx * 22
+            y0 = ybase + layer * 40
+            width = 10 + ((index + box_idx + layer) % 7) * 6
+            height = 10 + ((index * 3 + box_idx + layer) % 5) * 8
+            properties = _annotation(index) if index % cfg.rare_annotation_stride == 0 and box_idx == 0 and layer == 0 else EMPTY_PROPERTIES
+            cell_elements.append(elements.Boundary(
+                layer=(layer, 0),
+                xy=_rect_xy(x0, y0, x0 + width, y0 + height),
+                properties=properties,
+                ))
+
+    for layer in range(cfg.heavy_box_layers, cfg.box_layers):
+        for box_idx in range(cfg.regular_boxes_per_cell):
+            x0 = xbase + box_idx * 38
+            y0 = ybase + (layer - cfg.heavy_box_layers) * 28 + 400
+            width = 18 + ((index + layer + box_idx) % 9) * 4
+            height = 12 + ((index + 2 * layer + box_idx) % 6) * 5
+            cell_elements.append(elements.Boundary(
+                layer=(layer, 0),
+                xy=_rect_xy(x0, y0, x0 + width, y0 + height),
+                properties=EMPTY_PROPERTIES,
+                ))
+
+    return cell_elements
+
+
+def _make_poly_cell(name: str, index: int, cfg: FixturePreset) -> list[elements.Element]:
+    cell_elements: list[elements.Element] = []
+    xbase = (index % 19) * 900
+    ybase = (index // 19) * 260
+
+    for poly_idx in range(cfg.polygons_per_cell):
+        layer = poly_idx % cfg.polygon_layers
+        dx = xbase + (poly_idx % 5) * 120
+        dy = ybase + (poly_idx // 5) * 80
+        size = 18 + ((index + poly_idx + layer) % 11) * 7
+        points = [
+            (dx, dy),
+            (dx + size, dy + size // 5),
+            (dx + size + size // 3, dy + size),
+            (dx + size // 2, dy + size + size // 2),
+            (dx - size // 4, dy + size // 2),
+            ]
+        properties = _annotation(index) if poly_idx == 0 and index % cfg.rare_annotation_stride == 0 else EMPTY_PROPERTIES
+        cell_elements.append(elements.Boundary(
+            layer=(layer, 0),
+            xy=_poly_xy(points),
+            properties=properties,
+            ))
+
+    if index % cfg.path_stride == 0:
+        layer = index % cfg.polygon_layers
+        cell_elements.append(elements.Path(
+            layer=(layer, 1),
+            path_type=2,
+            width=12 + (index % 5) * 4,
+            extension=(0, 0),
+            xy=numpy.array(
+                [
+                    [xbase, ybase + 900],
+                    [xbase + 240, ybase + 930],
+                    [xbase + 420, ybase + 960],
+                    ],
+                dtype=numpy.int32,
+                ),
+            properties=EMPTY_PROPERTIES,
+            ))
+
+    if index % cfg.text_stride == 0:
+        layer = index % cfg.polygon_layers
+        properties = _annotation(index) if index % cfg.rare_annotation_stride == 0 else EMPTY_PROPERTIES
+        cell_elements.append(elements.Text(
+            layer=(layer, 2),
+            presentation=0,
+            path_type=0,
+            width=0,
+            invert_y=False,
+            mag=1.0,
+            angle_deg=0.0,
+            xy=numpy.array([[xbase + 64, ybase + 1536]], dtype=numpy.int32),
+            string=f'T{index:05d}'.encode('ASCII'),
+            properties=properties,
+            ))
+
+    return cell_elements
+
+
+def _write_struct(stream: Any, name: str, cell_elements: list[elements.Element]) -> None:
+    klamath.library.write_struct(stream, name=name.encode('ASCII'), elements=cell_elements)
+
+
+def _box_name(index: int) -> str:
+    return f'box_{index:05d}'
+
+
+def _poly_name(index: int) -> str:
+    return f'poly_{index:05d}'
+
+
+def _box_wrapper_name(index: int) -> str:
+    return f'box_wrap_{index:05d}'
+
+
+def _poly_wrapper_name(index: int) -> str:
+    return f'poly_wrap_{index:05d}'
+
+
+def _box_cluster_name(index: int) -> str:
+    return f'box_cluster_{index:05d}'
+
+
+def _poly_cluster_name(index: int) -> str:
+    return f'poly_cluster_{index:05d}'
+
+
+def _write_box_cells(stream: Any, cfg: FixturePreset) -> None:
+    for idx in range(cfg.box_cells):
+        _write_struct(stream, _box_name(idx), _make_box_cell(_box_name(idx), idx, cfg))
+
+
+def _write_poly_cells(stream: Any, cfg: FixturePreset) -> None:
+    for idx in range(cfg.poly_cells):
+        _write_struct(stream, _poly_name(idx), _make_poly_cell(_poly_name(idx), idx, cfg))
+
+
+def _write_wrappers(stream: Any, cfg: FixturePreset) -> None:
+    for idx in range(cfg.box_wrappers):
+        target = _box_name(idx % cfg.box_cells)
+        origin = ((idx % 97) * 2_000, (idx // 97) * 2_000)
+        _write_struct(stream, _box_wrapper_name(idx), [_sref(target, origin)])
+
+    for idx in range(cfg.poly_wrappers):
+        target = _poly_name(idx % cfg.poly_cells)
+        origin = ((idx % 61) * 3_200, (idx // 61) * 3_200)
+        _write_struct(stream, _poly_wrapper_name(idx), [_sref(target, origin)])
+
+
+def _write_box_clusters(stream: Any, cfg: FixturePreset) -> None:
+    array_refs = min(cfg.box_cluster_refs, max(1, (3 * cfg.box_cluster_refs) // 4))
+    for idx in range(cfg.box_clusters):
+        cell_elements: list[elements.Element] = []
+        for ref_idx in range(cfg.box_cluster_refs):
+            target = _box_name((idx * cfg.box_cluster_refs + ref_idx) % cfg.box_cells)
+            origin = (
+                (ref_idx % 6) * 48_000,
+                (ref_idx // 6) * 48_000,
+                )
+            if ref_idx < array_refs:
+                cell_elements.append(_aref(target, origin, cfg.box_cluster_array, (720, 900)))
+            else:
+                cell_elements.append(_sref(target, origin))
+        _write_struct(stream, _box_cluster_name(idx), cell_elements)
+
+
+def _write_poly_clusters(stream: Any, cfg: FixturePreset) -> None:
+    array_refs = min(cfg.poly_cluster_refs, cfg.poly_cluster_refs // 2)
+    for idx in range(cfg.poly_clusters):
+        cell_elements: list[elements.Element] = []
+        for ref_idx in range(cfg.poly_cluster_refs):
+            target = _poly_name((idx * cfg.poly_cluster_refs + ref_idx) % cfg.poly_cells)
+            origin = (
+                (ref_idx % 10) * 96_000,
+                (ref_idx // 10) * 96_000,
+                )
+            if ref_idx < array_refs:
+                cell_elements.append(_aref(target, origin, cfg.poly_cluster_array, (12_000, 8_500)))
+            else:
+                cell_elements.append(_sref(target, origin))
+        _write_struct(stream, _poly_cluster_name(idx), cell_elements)
+
+
+def _top_box_refs(cfg: FixturePreset) -> list[elements.Reference]:
+    refs: list[elements.Reference] = []
+
+    for idx in range(cfg.box_wrappers):
+        refs.append(_sref(
+            _box_wrapper_name(idx),
+            ((idx % 240) * 240_000, (idx // 240) * 240_000),
+            ))
+
+    for idx in range(cfg.box_clusters):
+        refs.append(_sref(
+            _box_cluster_name(idx),
+            ((idx % 100) * 800_000, (idx // 100) * 800_000 + 14_000_000),
+            ))
+
+    for idx in range(cfg.top_direct_box_refs):
+        target = _box_name(idx % cfg.box_cells)
+        origin = (
+            (idx % 150) * 160_000,
+            (idx // 150) * 160_000 + 26_000_000,
+            )
+        if cfg.top_box_array == (1, 1):
+            refs.append(_sref(target, origin))
+        else:
+            refs.append(_aref(target, origin, cfg.top_box_array, (1_100, 1_350)))
+
+    return refs
+
+
+def _top_poly_refs(cfg: FixturePreset) -> list[elements.Reference]:
+    refs: list[elements.Reference] = []
+
+    for idx in range(cfg.poly_wrappers):
+        refs.append(_sref(
+            _poly_wrapper_name(idx),
+            ((idx % 180) * 360_000, (idx // 180) * 360_000 + 44_000_000),
+            ))
+
+    for idx in range(cfg.poly_clusters):
+        refs.append(_sref(
+            _poly_cluster_name(idx),
+            ((idx % 70) * 1_100_000, (idx // 70) * 1_100_000 + 58_000_000),
+            ))
+
+    for idx in range(cfg.top_direct_poly_refs):
+        target = _poly_name(idx % cfg.poly_cells)
+        origin = (
+            (idx % 110) * 420_000,
+            (idx // 110) * 420_000 + 72_000_000,
+            )
+        if cfg.top_poly_array == (1, 1):
+            refs.append(_sref(target, origin))
+        else:
+            refs.append(_aref(target, origin, cfg.top_poly_array, (16_000, 14_000)))
+
+    return refs
+
+
+def _write_top(stream: Any, cfg: FixturePreset) -> None:
+    cell_elements: list[elements.Element] = []
+    cell_elements.extend(_top_box_refs(cfg))
+    cell_elements.extend(_top_poly_refs(cfg))
+    _write_struct(stream, 'TOP', cell_elements)
+
+
+def _poly_paths_total(cfg: FixturePreset) -> int:
+    return (cfg.poly_cells - 1) // cfg.path_stride + 1
+
+
+def _poly_texts_total(cfg: FixturePreset) -> int:
+    return (cfg.poly_cells - 1) // cfg.text_stride + 1
+
+
+def _ref_instances_per_box_cluster(cfg: FixturePreset) -> int:
+    array_refs = min(cfg.box_cluster_refs, max(1, (3 * cfg.box_cluster_refs) // 4))
+    array_mult = cfg.box_cluster_array[0] * cfg.box_cluster_array[1]
+    return array_refs * array_mult + (cfg.box_cluster_refs - array_refs)
+
+
+def _ref_instances_per_poly_cluster(cfg: FixturePreset) -> int:
+    array_refs = min(cfg.poly_cluster_refs, cfg.poly_cluster_refs // 2)
+    array_mult = cfg.poly_cluster_array[0] * cfg.poly_cluster_array[1]
+    return array_refs * array_mult + (cfg.poly_cluster_refs - array_refs)
+
+
+def fixture_manifest(path: str | Path, preset: str, scale: float = 1.0) -> FixtureManifest:
+    base = PRESETS[preset]
+    cfg = _scaled_preset(base, scale)
+
+    flattened_box_placements = (
+        cfg.box_wrappers
+        + cfg.box_clusters * _ref_instances_per_box_cluster(cfg)
+        + cfg.top_direct_box_refs * cfg.top_box_array[0] * cfg.top_box_array[1]
+        )
+    flattened_poly_placements = (
+        cfg.poly_wrappers
+        + cfg.poly_clusters * _ref_instances_per_poly_cluster(cfg)
+        + cfg.top_direct_poly_refs * cfg.top_poly_array[0] * cfg.top_poly_array[1]
+        )
+    polygon_layers = max(1, cfg.polygon_layers)
+    polys_per_layer = (cfg.poly_cells * cfg.polygons_per_cell) // polygon_layers
+
+    return FixtureManifest(
+        preset=cfg.name,
+        scale=scale,
+        gds_path=str(Path(path)),
+        library_name=f'masque-perf-{cfg.name}',
+        cells=cfg.box_cells + cfg.poly_cells + cfg.box_wrappers + cfg.poly_wrappers + cfg.box_clusters + cfg.poly_clusters + 1,
+        refs=(
+            cfg.box_wrappers
+            + cfg.poly_wrappers
+            + cfg.box_clusters * cfg.box_cluster_refs
+            + cfg.poly_clusters * cfg.poly_cluster_refs
+            + cfg.box_wrappers + cfg.poly_wrappers + cfg.box_clusters + cfg.poly_clusters
+            + cfg.top_direct_box_refs + cfg.top_direct_poly_refs
+            ),
+        layers=cfg.total_layers,
+        box_layers=cfg.box_layers,
+        heavy_box_layers=[[layer, 0] for layer in range(cfg.heavy_box_layers)],
+        polygon_layers=[[layer, 0] for layer in range(cfg.polygon_layers)],
+        hierarchical_boxes_per_heavy_layer=cfg.box_cells * cfg.heavy_boxes_per_cell,
+        hierarchical_boxes_per_regular_layer=cfg.box_cells * cfg.regular_boxes_per_cell,
+        hierarchical_polygons_total=cfg.poly_cells * cfg.polygons_per_cell,
+        hierarchical_paths_total=_poly_paths_total(cfg),
+        hierarchical_texts_total=_poly_texts_total(cfg),
+        flattened_box_placements=flattened_box_placements,
+        flattened_poly_placements=flattened_poly_placements,
+        estimated_flat_boxes_per_heavy_layer=flattened_box_placements * cfg.heavy_boxes_per_cell,
+        estimated_flat_polygons_per_active_polygon_layer=flattened_poly_placements * polys_per_layer // cfg.poly_cells if cfg.poly_cells else 0,
+        )
+
+
+def write_fixture(
+        path: str | Path,
+        *,
+        preset: str,
+        scale: float = 1.0,
+        write_manifest: bool = True,
+        ) -> FixtureManifest:
+    if preset not in PRESETS:
+        known = ', '.join(sorted(PRESETS))
+        raise KeyError(f'unknown preset {preset!r}; expected one of: {known}')
+
+    manifest = fixture_manifest(path, preset, scale)
+    cfg = _scaled_preset(PRESETS[preset], scale)
+    output = Path(path)
+    output.parent.mkdir(parents=True, exist_ok=True)
+
+    with output.open('wb') as stream:
+        header = klamath.library.FileHeader(
+            name=manifest.library_name.encode('ASCII'),
+            user_units_per_db_unit=USER_UNITS_PER_DB_UNIT,
+            meters_per_db_unit=METERS_PER_DB_UNIT,
+            )
+        header.write(stream)
+        _write_box_cells(stream, cfg)
+        _write_poly_cells(stream, cfg)
+        _write_wrappers(stream, cfg)
+        _write_box_clusters(stream, cfg)
+        _write_poly_clusters(stream, cfg)
+        _write_top(stream, cfg)
+        klamath.records.ENDLIB.write(stream, None)
+
+    if write_manifest:
+        manifest_path = output.with_suffix(output.suffix + '.json')
+        manifest_path.write_text(json.dumps(asdict(manifest), indent=2, sort_keys=True) + '\n')
+
+    return manifest
+
+
+def build_arg_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(description='Generate synthetic GDS fixtures for GDS reader/writer performance work.')
+    parser.add_argument(
+        'preset',
+        nargs='?',
+        default='many_cells',
+        choices=sorted(PRESETS),
+        help='Fixture family to generate.',
+        )
+    parser.add_argument(
+        'output',
+        nargs='?',
+        help='Output .gds path. Defaults to build/gds_perf/<preset>.gds',
+        )
+    parser.add_argument(
+        '--scale',
+        type=float,
+        default=1.0,
+        help='Scale the preset counts down or up while keeping the same shape mix. Default: 1.0',
+        )
+    parser.add_argument(
+        '--no-manifest',
+        action='store_true',
+        help='Do not write the sidecar JSON manifest.',
+        )
+    return parser
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = build_arg_parser()
+    args = parser.parse_args(argv)
+    output = Path(args.output) if args.output is not None else Path('build/gds_perf') / f'{args.preset}.gds'
+    manifest = write_fixture(output, preset=args.preset, scale=args.scale, write_manifest=not args.no_manifest)
+    print(json.dumps(asdict(manifest), indent=2, sort_keys=True))
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())

masque/file/svg.py

@@ -45,6 +45,10 @@ def _make_svg_ids(names: Mapping[str, Pattern]) -> dict[str, str]:
     return svg_ids
 
 
+def _detached_library(library: Mapping[str, Pattern]) -> dict[str, Pattern]:
+    return {name: pat.deepcopy() for name, pat in library.items()}
+
+
 def writefile(
         library: Mapping[str, Pattern],
         top: str,
@@ -53,13 +57,12 @@ def writefile(
         annotate_ports: bool = False,
         ) -> None:
     """
-    Write a Pattern to an SVG file, by first calling .polygonize() on it
+    Write a Pattern to an SVG file, by first calling .polygonize() on a detached
+     materialized copy
      to change the shapes into polygons, and then writing patterns as SVG
      groups (<g>, inside <defs>), polygons as paths (<path>), and refs
      as <use> elements.
 
-    Note that this function modifies the Pattern.
-
     If `custom_attributes` is `True`, a non-standard `pattern_layer` attribute
      is written to the relevant elements.
 
@@ -71,19 +74,21 @@ def writefile(
      prior to calling this function.
 
     Args:
-        pattern: Pattern to write to file. Modified by this function.
+        library: Mapping of pattern names to patterns.
+        top: Name of the top-level pattern to render.
         filename: Filename to write to.
         custom_attributes: Whether to write non-standard `pattern_layer` attribute to the
             SVG elements.
         annotate_ports: If True, draw an arrow for each port (similar to
             `Pattern.visualize(..., ports=True)`).
     """
-    pattern = library[top]
+    detached = _detached_library(library)
+    pattern = detached[top]
 
     # Polygonize pattern
     pattern.polygonize()
 
-    bounds = pattern.get_bounds(library=library)
+    bounds = pattern.get_bounds(library=detached)
     if bounds is None:
         bounds_min, bounds_max = numpy.array([[-1, -1], [1, 1]])
         logger.warning('Pattern had no bounds (empty?); setting arbitrary viewbox', stacklevel=1)
@@ -96,10 +101,10 @@ def writefile(
     # Create file
     svg = svgwrite.Drawing(filename, profile='full', viewBox=viewbox_string,
                            debug=(not custom_attributes))
-    svg_ids = _make_svg_ids(library)
+    svg_ids = _make_svg_ids(detached)
 
     # Now create a group for each pattern and add in any Boundary and Use elements
-    for name, pat in library.items():
+    for name, pat in detached.items():
         svg_group = svg.g(id=svg_ids[name], fill='blue', stroke='red')
 
         for layer, shapes in pat.shapes.items():
@@ -158,21 +163,21 @@ def writefile_inverted(
      box and drawing the polygons with reverse vertex order inside it, all within
      one `<path>` element.
 
-    Note that this function modifies the Pattern.
-
     If you want pattern polygonized with non-default arguments, just call `pattern.polygonize()`
      prior to calling this function.
 
     Args:
-        pattern: Pattern to write to file. Modified by this function.
+        library: Mapping of pattern names to patterns.
+        top: Name of the top-level pattern to render.
         filename: Filename to write to.
     """
-    pattern = library[top]
+    detached = _detached_library(library)
+    pattern = detached[top]
 
     # Polygonize and flatten pattern
-    pattern.polygonize().flatten(library)
+    pattern.polygonize().flatten(detached)
 
-    bounds = pattern.get_bounds(library=library)
+    bounds = pattern.get_bounds(library=detached)
     if bounds is None:
         bounds_min, bounds_max = numpy.array([[-1, -1], [1, 1]])
         logger.warning('Pattern had no bounds (empty?); setting arbitrary viewbox', stacklevel=1)

masque/label.py

@@ -53,6 +53,22 @@ class Label(PositionableImpl, RepeatableImpl, AnnotatableImpl, Bounded, Pivotabl
         self.repetition = repetition
         self.annotations = annotations if annotations is not None else {}
 
+    @classmethod
+    def _from_raw(
+            cls,
+            string: str,
+            *,
+            offset: NDArray[numpy.float64],
+            repetition: Repetition | None = None,
+            annotations: annotations_t | None = None,
+            ) -> Self:
+        new = cls.__new__(cls)
+        new._string = string
+        new._offset = offset
+        new._repetition = repetition
+        new._annotations = annotations
+        return new
+
     def __copy__(self) -> Self:
         return type(self)(
             string=self.string,

masque/ref.py

@@ -86,6 +86,26 @@ class Ref(
         self.repetition = repetition
         self.annotations = annotations if annotations is not None else {}
 
+    @classmethod
+    def _from_raw(
+            cls,
+            *,
+            offset: NDArray[numpy.float64],
+            rotation: float,
+            mirrored: bool,
+            scale: float,
+            repetition: Repetition | None,
+            annotations: annotations_t | None,
+            ) -> Self:
+        new = cls.__new__(cls)
+        new._offset = offset
+        new._rotation = rotation % (2 * pi)
+        new._scale = scale
+        new._mirrored = mirrored
+        new._repetition = repetition
+        new._annotations = annotations
+        return new
+
     def __copy__(self) -> 'Ref':
         new = Ref(
             offset=self.offset.copy(),

masque/repetition.py

@@ -113,6 +113,22 @@ class Grid(Repetition):
         self.a_count = a_count
         self.b_count = b_count
 
+    @classmethod
+    def _from_raw(
+            cls: type[GG],
+            *,
+            a_vector: NDArray[numpy.float64],
+            a_count: int,
+            b_vector: NDArray[numpy.float64],
+            b_count: int,
+            ) -> GG:
+        new = cls.__new__(cls)
+        new._a_vector = a_vector
+        new._b_vector = b_vector
+        new._a_count = int(a_count)
+        new._b_count = int(b_count)
+        return new
+
     @classmethod
     def aligned(
             cls: type[GG],

masque/shapes/__init__.py

@@ -11,6 +11,7 @@ from .shape import (
 
 from .polygon import Polygon as Polygon
 from .poly_collection import PolyCollection as PolyCollection
+from .rect_collection import RectCollection as RectCollection
 from .circle import Circle as Circle
 from .ellipse import Ellipse as Ellipse
 from .arc import Arc as Arc

masque/shapes/arc.py

@@ -159,20 +159,7 @@ class Arc(PositionableImpl, Shape):
             rotation: float = 0,
             repetition: Repetition | None = None,
             annotations: annotations_t = None,
-            raw: bool = False,
             ) -> None:
-        if raw:
-            assert isinstance(radii, numpy.ndarray)
-            assert isinstance(angles, numpy.ndarray)
-            assert isinstance(offset, numpy.ndarray)
-            self._radii = radii
-            self._angles = angles
-            self._width = width
-            self._offset = offset
-            self._rotation = rotation
-            self._repetition = repetition
-            self._annotations = annotations
-        else:
         self.radii = radii
         self.angles = angles
         self.width = width
@@ -181,6 +168,28 @@ class Arc(PositionableImpl, Shape):
         self.repetition = repetition
         self.annotations = annotations
 
+    @classmethod
+    def _from_raw(
+            cls,
+            *,
+            radii: NDArray[numpy.float64],
+            angles: NDArray[numpy.float64],
+            width: float,
+            offset: NDArray[numpy.float64],
+            rotation: float,
+            annotations: annotations_t = None,
+            repetition: Repetition | None = None,
+            ) -> 'Arc':
+        new = cls.__new__(cls)
+        new._radii = radii
+        new._angles = angles
+        new._width = width
+        new._offset = offset
+        new._rotation = rotation % (2 * pi)
+        new._repetition = repetition
+        new._annotations = annotations
+        return new
+
     def __deepcopy__(self, memo: dict | None = None) -> 'Arc':
         memo = {} if memo is None else memo
         new = copy.copy(self)

masque/shapes/circle.py

@@ -50,20 +50,28 @@ class Circle(PositionableImpl, Shape):
             offset: ArrayLike = (0.0, 0.0),
             repetition: Repetition | None = None,
             annotations: annotations_t = None,
-            raw: bool = False,
             ) -> None:
-        if raw:
-            assert isinstance(offset, numpy.ndarray)
-            self._radius = radius
-            self._offset = offset
-            self._repetition = repetition
-            self._annotations = annotations
-        else:
         self.radius = radius
         self.offset = offset
         self.repetition = repetition
         self.annotations = annotations
 
+    @classmethod
+    def _from_raw(
+            cls,
+            *,
+            radius: float,
+            offset: NDArray[numpy.float64],
+            annotations: annotations_t = None,
+            repetition: Repetition | None = None,
+            ) -> 'Circle':
+        new = cls.__new__(cls)
+        new._radius = radius
+        new._offset = offset
+        new._repetition = repetition
+        new._annotations = annotations
+        return new
+
     def __deepcopy__(self, memo: dict | None = None) -> 'Circle':
         memo = {} if memo is None else memo
         new = copy.copy(self)

masque/shapes/ellipse.py

@@ -95,23 +95,31 @@ class Ellipse(PositionableImpl, Shape):
             rotation: float = 0,
             repetition: Repetition | None = None,
             annotations: annotations_t = None,
-            raw: bool = False,
             ) -> None:
-        if raw:
-            assert isinstance(radii, numpy.ndarray)
-            assert isinstance(offset, numpy.ndarray)
-            self._radii = radii
-            self._offset = offset
-            self._rotation = rotation
-            self._repetition = repetition
-            self._annotations = annotations
-        else:
         self.radii = radii
         self.offset = offset
         self.rotation = rotation
         self.repetition = repetition
         self.annotations = annotations
 
+    @classmethod
+    def _from_raw(
+            cls,
+            *,
+            radii: NDArray[numpy.float64],
+            offset: NDArray[numpy.float64],
+            rotation: float,
+            annotations: annotations_t = None,
+            repetition: Repetition | None = None,
+            ) -> Self:
+        new = cls.__new__(cls)
+        new._radii = radii
+        new._offset = offset
+        new._rotation = rotation % pi
+        new._repetition = repetition
+        new._annotations = annotations
+        return new
+
     def __deepcopy__(self, memo: dict | None = None) -> Self:
         memo = {} if memo is None else memo
         new = copy.copy(self)

masque/shapes/path.py

@@ -201,20 +201,9 @@ class Path(Shape):
             rotation: float = 0,
             repetition: Repetition | None = None,
             annotations: annotations_t = None,
-            raw: bool = False,
             ) -> None:
         self._cap_extensions = None     # Since .cap setter might access it
 
-        if raw:
-            assert isinstance(vertices, numpy.ndarray)
-            assert isinstance(cap_extensions, numpy.ndarray) or cap_extensions is None
-            self._vertices = vertices
-            self._repetition = repetition
-            self._annotations = annotations
-            self._width = width
-            self._cap = cap
-            self._cap_extensions = cap_extensions
-        else:
         self.vertices = vertices
         self.repetition = repetition
         self.annotations = annotations
@@ -229,6 +218,26 @@ class Path(Shape):
         if numpy.any(offset):
             self.translate(offset)
 
+    @classmethod
+    def _from_raw(
+            cls,
+            *,
+            vertices: NDArray[numpy.float64],
+            width: float,
+            cap: PathCap,
+            cap_extensions: NDArray[numpy.float64] | None = None,
+            annotations: annotations_t = None,
+            repetition: Repetition | None = None,
+            ) -> Self:
+        new = cls.__new__(cls)
+        new._vertices = vertices
+        new._width = width
+        new._cap = cap
+        new._cap_extensions = cap_extensions
+        new._repetition = repetition
+        new._annotations = annotations
+        return new
+
     def __deepcopy__(self, memo: dict | None = None) -> 'Path':
         memo = {} if memo is None else memo
         new = copy.copy(self)

masque/shapes/poly_collection.py

@@ -34,7 +34,7 @@ class PolyCollection(Shape):
     _vertex_lists: NDArray[numpy.float64]
     """ 2D NDArray ((N+M+...) x 2)  of vertices `[[xa0, ya0], [xa1, ya1], ..., [xb0, yb0], [xb1, yb1], ... ]` """
 
-    _vertex_offsets: NDArray[numpy.intp]
+    _vertex_offsets: NDArray[numpy.integer[Any]]
     """ 1D NDArray specifying the starting offset for each polygon """
 
     @property
@@ -45,7 +45,7 @@ class PolyCollection(Shape):
         return self._vertex_lists
 
     @property
-    def vertex_offsets(self) -> NDArray[numpy.intp]:
+    def vertex_offsets(self) -> NDArray[numpy.integer[Any]]:
         """
         Starting offset (in `vertex_lists`) for each polygon
         """
@@ -63,7 +63,7 @@ class PolyCollection(Shape):
                 chain(self._vertex_offsets[1:], [self._vertex_lists.shape[0]]),
                 strict=True,
                 ):
-            yield slice(ii, ff)
+            yield slice(int(ii), int(ff))
 
     @property
     def polygon_vertices(self) -> Iterator[NDArray[numpy.float64]]:
@@ -100,16 +100,7 @@ class PolyCollection(Shape):
             rotation: float = 0.0,
             repetition: Repetition | None = None,
             annotations: annotations_t = None,
-            raw: bool = False,
             ) -> None:
-        if raw:
-            assert isinstance(vertex_lists, numpy.ndarray)
-            assert isinstance(vertex_offsets, numpy.ndarray)
-            self._vertex_lists = vertex_lists
-            self._vertex_offsets = vertex_offsets
-            self._repetition = repetition
-            self._annotations = annotations
-        else:
         self._vertex_lists = numpy.asarray(vertex_lists, dtype=float)
         self._vertex_offsets = numpy.asarray(vertex_offsets, dtype=numpy.intp)
         self.repetition = repetition
@@ -119,6 +110,22 @@ class PolyCollection(Shape):
         if numpy.any(offset):
             self.translate(offset)
 
+    @classmethod
+    def _from_raw(
+            cls,
+            *,
+            vertex_lists: NDArray[numpy.float64],
+            vertex_offsets: NDArray[numpy.integer[Any]],
+            annotations: annotations_t = None,
+            repetition: Repetition | None = None,
+            ) -> Self:
+        new = cls.__new__(cls)
+        new._vertex_lists = vertex_lists
+        new._vertex_offsets = vertex_offsets
+        new._repetition = repetition
+        new._annotations = annotations
+        return new
+
     def __deepcopy__(self, memo: dict | None = None) -> Self:
         memo = {} if memo is None else memo
         new = copy.copy(self)
@@ -132,7 +139,7 @@ class PolyCollection(Shape):
         return (
             type(self) is type(other)
             and numpy.array_equal(self._vertex_lists, other._vertex_lists)
-            and numpy.array_equal(self._vertex_offsets, other._vertex_offsets)
+            and numpy.array_equal(self.vertex_offsets, other.vertex_offsets)
             and self.repetition == other.repetition
             and annotations_eq(self.annotations, other.annotations)
             )
@@ -215,11 +222,11 @@ class PolyCollection(Shape):
 
         # TODO: normalize mirroring?
 
-        return ((type(self), rotated_vertices.data.tobytes() + self._vertex_offsets.tobytes()),
+        return ((type(self), rotated_vertices.data.tobytes() + self.vertex_offsets.tobytes()),
                 (offset, scale / norm_value, rotation, False),
                 lambda: PolyCollection(
                     vertex_lists=rotated_vertices * norm_value,
-                    vertex_offsets=self._vertex_offsets.copy(),
+                    vertex_offsets=self.vertex_offsets.copy(),
                     ),
                 )
 

masque/shapes/polygon.py

@@ -115,14 +115,7 @@ class Polygon(Shape):
             rotation: float = 0.0,
             repetition: Repetition | None = None,
             annotations: annotations_t = None,
-            raw: bool = False,
             ) -> None:
-        if raw:
-            assert isinstance(vertices, numpy.ndarray)
-            self._vertices = vertices
-            self._repetition = repetition
-            self._annotations = annotations
-        else:
         self.vertices = vertices
         self.repetition = repetition
         self.annotations = annotations
@@ -131,6 +124,20 @@ class Polygon(Shape):
         if numpy.any(offset):
             self.translate(offset)
 
+    @classmethod
+    def _from_raw(
+            cls,
+            *,
+            vertices: NDArray[numpy.float64],
+            annotations: annotations_t = None,
+            repetition: Repetition | None = None,
+            ) -> Self:
+        new = cls.__new__(cls)
+        new._vertices = vertices
+        new._repetition = repetition
+        new._annotations = annotations
+        return new
+
     def __deepcopy__(self, memo: dict | None = None) -> 'Polygon':
         memo = {} if memo is None else memo
         new = copy.copy(self)

masque/shapes/rect_collection.py

@@ -0,0 +1,249 @@
+from typing import Any, cast, Self
+from collections.abc import Iterator
+import copy
+import functools
+
+import numpy
+from numpy import pi
+from numpy.typing import NDArray, ArrayLike
+
+from . import Shape, normalized_shape_tuple
+from .polygon import Polygon
+from ..error import PatternError
+from ..repetition import Repetition
+from ..utils import annotations_lt, annotations_eq, rep2key, annotations_t
+
+
+def _normalize_rects(rects: ArrayLike) -> NDArray[numpy.float64]:
+    arr = numpy.asarray(rects, dtype=float)
+    if arr.ndim != 2 or arr.shape[1] != 4:
+        raise PatternError('Rectangles must be an Nx4 array of [xmin, ymin, xmax, ymax]')
+    if numpy.any(arr[:, 0] > arr[:, 2]) or numpy.any(arr[:, 1] > arr[:, 3]):
+        raise PatternError('Rectangles must satisfy xmin <= xmax and ymin <= ymax')
+    if arr.shape[0] <= 1:
+        return arr
+    order = numpy.lexsort((arr[:, 3], arr[:, 2], arr[:, 1], arr[:, 0]))
+    return arr[order]
+
+
+def _renormalize_rects_in_place(rects: NDArray[numpy.float64]) -> None:
+    x0 = numpy.minimum(rects[:, 0], rects[:, 2])
+    x1 = numpy.maximum(rects[:, 0], rects[:, 2])
+    y0 = numpy.minimum(rects[:, 1], rects[:, 3])
+    y1 = numpy.maximum(rects[:, 1], rects[:, 3])
+    rects[:, 0] = x0
+    rects[:, 1] = y0
+    rects[:, 2] = x1
+    rects[:, 3] = y1
+
+
+@functools.total_ordering
+class RectCollection(Shape):
+    """
+    A collection of axis-aligned rectangles, stored as an Nx4 array of
+      `[xmin, ymin, xmax, ymax]` rows.
+    """
+    __slots__ = (
+        '_rects',
+        '_repetition', '_annotations',
+        )
+
+    _rects: NDArray[numpy.float64]
+
+    @property
+    def rects(self) -> NDArray[numpy.float64]:
+        return self._rects
+
+    @rects.setter
+    def rects(self, val: ArrayLike) -> None:
+        self._rects = _normalize_rects(val)
+
+    @property
+    def offset(self) -> NDArray[numpy.float64]:
+        return numpy.zeros(2)
+
+    @offset.setter
+    def offset(self, val: ArrayLike) -> None:
+        if numpy.any(val):
+            raise PatternError('RectCollection offset is forced to (0, 0)')
+
+    def set_offset(self, val: ArrayLike) -> Self:
+        if numpy.any(val):
+            raise PatternError('RectCollection offset is forced to (0, 0)')
+        return self
+
+    def translate(self, offset: ArrayLike) -> Self:
+        delta = numpy.asarray(offset, dtype=float).reshape(2)
+        self._rects[:, [0, 2]] += delta[0]
+        self._rects[:, [1, 3]] += delta[1]
+        return self
+
+    def __init__(
+            self,
+            rects: ArrayLike,
+            *,
+            offset: ArrayLike = (0.0, 0.0),
+            rotation: float = 0.0,
+            repetition: Repetition | None = None,
+            annotations: annotations_t = None,
+            ) -> None:
+        self.rects = rects
+        self.repetition = repetition
+        self.annotations = annotations
+        if rotation:
+            self.rotate(rotation)
+        if numpy.any(offset):
+            self.translate(offset)
+
+    @classmethod
+    def _from_raw(
+            cls,
+            *,
+            rects: NDArray[numpy.float64],
+            annotations: annotations_t = None,
+            repetition: Repetition | None = None,
+            ) -> Self:
+        new = cls.__new__(cls)
+        new._rects = rects
+        new._repetition = repetition
+        new._annotations = annotations
+        return new
+
+    @property
+    def polygon_vertices(self) -> Iterator[NDArray[numpy.float64]]:
+        for rect in self._rects:
+            xmin, ymin, xmax, ymax = rect
+            yield numpy.array([
+                [xmin, ymin],
+                [xmin, ymax],
+                [xmax, ymax],
+                [xmax, ymin],
+                ], dtype=float)
+
+    def __deepcopy__(self, memo: dict | None = None) -> Self:
+        memo = {} if memo is None else memo
+        new = copy.copy(self)
+        new._rects = self._rects.copy()
+        new._repetition = copy.deepcopy(self._repetition, memo)
+        new._annotations = copy.deepcopy(self._annotations)
+        return new
+
+    def _sorted_rects(self) -> NDArray[numpy.float64]:
+        if self._rects.shape[0] <= 1:
+            return self._rects
+        order = numpy.lexsort((self._rects[:, 3], self._rects[:, 2], self._rects[:, 1], self._rects[:, 0]))
+        return self._rects[order]
+
+    def __eq__(self, other: Any) -> bool:
+        return (
+            type(self) is type(other)
+            and numpy.array_equal(self._sorted_rects(), other._sorted_rects())
+            and self.repetition == other.repetition
+            and annotations_eq(self.annotations, other.annotations)
+            )
+
+    def __lt__(self, other: Shape) -> bool:
+        if type(self) is not type(other):
+            if repr(type(self)) != repr(type(other)):
+                return repr(type(self)) < repr(type(other))
+            return id(type(self)) < id(type(other))
+
+        other = cast('RectCollection', other)
+        self_rects = self._sorted_rects()
+        other_rects = other._sorted_rects()
+        if not numpy.array_equal(self_rects, other_rects):
+            min_len = min(self_rects.shape[0], other_rects.shape[0])
+            eq_mask = self_rects[:min_len] != other_rects[:min_len]
+            eq_lt = self_rects[:min_len] < other_rects[:min_len]
+            eq_lt_masked = eq_lt[eq_mask]
+            if eq_lt_masked.size > 0:
+                return bool(eq_lt_masked.flat[0])
+            return self_rects.shape[0] < other_rects.shape[0]
+        if self.repetition != other.repetition:
+            return rep2key(self.repetition) < rep2key(other.repetition)
+        return annotations_lt(self.annotations, other.annotations)
+
+    def to_polygons(
+            self,
+            num_vertices: int | None = None,      # unused  # noqa: ARG002
+            max_arclen: float | None = None,      # unused  # noqa: ARG002
+            ) -> list[Polygon]:
+        return [
+            Polygon(
+                vertices=vertices,
+                repetition=copy.deepcopy(self.repetition),
+                annotations=copy.deepcopy(self.annotations),
+                )
+            for vertices in self.polygon_vertices
+            ]
+
+    def get_bounds_single(self) -> NDArray[numpy.float64] | None:
+        if self._rects.size == 0:
+            return None
+        mins = self._rects[:, :2].min(axis=0)
+        maxs = self._rects[:, 2:].max(axis=0)
+        return numpy.vstack((mins, maxs))
+
+    def rotate(self, theta: float) -> Self:
+        quarter_turns = int(numpy.rint(theta / (pi / 2)))
+        if not numpy.isclose(theta, quarter_turns * (pi / 2)):
+            raise PatternError('RectCollection only supports Manhattan rotations')
+        turns = quarter_turns % 4
+        if turns == 0 or self._rects.size == 0:
+            return self
+
+        corners = numpy.stack((
+            self._rects[:, [0, 1]],
+            self._rects[:, [0, 3]],
+            self._rects[:, [2, 3]],
+            self._rects[:, [2, 1]],
+            ), axis=1)
+        flat = corners.reshape(-1, 2)
+        if turns == 1:
+            rotated = numpy.column_stack((-flat[:, 1], flat[:, 0]))
+        elif turns == 2:
+            rotated = -flat
+        else:
+            rotated = numpy.column_stack((flat[:, 1], -flat[:, 0]))
+        corners = rotated.reshape(corners.shape)
+        self._rects[:, 0] = corners[:, :, 0].min(axis=1)
+        self._rects[:, 1] = corners[:, :, 1].min(axis=1)
+        self._rects[:, 2] = corners[:, :, 0].max(axis=1)
+        self._rects[:, 3] = corners[:, :, 1].max(axis=1)
+        return self
+
+    def mirror(self, axis: int = 0) -> Self:
+        if axis not in (0, 1):
+            raise PatternError('Axis must be 0 or 1')
+        if axis == 0:
+            self._rects[:, [1, 3]] *= -1
+        else:
+            self._rects[:, [0, 2]] *= -1
+        _renormalize_rects_in_place(self._rects)
+        return self
+
+    def scale_by(self, c: float) -> Self:
+        self._rects *= c
+        _renormalize_rects_in_place(self._rects)
+        return self
+
+    def normalized_form(self, norm_value: float) -> normalized_shape_tuple:
+        rects = self._sorted_rects()
+        centers = 0.5 * (rects[:, :2] + rects[:, 2:])
+        offset = centers.mean(axis=0)
+        zeroed = rects.copy()
+        zeroed[:, [0, 2]] -= offset[0]
+        zeroed[:, [1, 3]] -= offset[1]
+        normed = zeroed / norm_value
+        return (
+            (type(self), normed.data.tobytes()),
+            (offset, 1.0, 0.0, False),
+            lambda: RectCollection(rects=normed * norm_value),
+            )
+
+    def __repr__(self) -> str:
+        if self._rects.size == 0:
+            return '<RectCollection r0>'
+        centers = 0.5 * (self._rects[:, :2] + self._rects[:, 2:])
+        centroid = centers.mean(axis=0)
+        return f'<RectCollection centroid {centroid} r{self._rects.shape[0]}>'

masque/shapes/text.py

@@ -73,18 +73,7 @@ class Text(PositionableImpl, RotatableImpl, Shape):
             mirrored: bool = False,
             repetition: Repetition | None = None,
             annotations: annotations_t = None,
-            raw: bool = False,
             ) -> None:
-        if raw:
-            assert isinstance(offset, numpy.ndarray)
-            self._offset = offset
-            self._string = string
-            self._height = height
-            self._rotation = rotation
-            self._mirrored = mirrored
-            self._repetition = repetition
-            self._annotations = annotations
-        else:
         self.offset = offset
         self.string = string
         self.height = height
@@ -94,6 +83,30 @@ class Text(PositionableImpl, RotatableImpl, Shape):
         self.annotations = annotations
         self.font_path = font_path
 
+    @classmethod
+    def _from_raw(
+            cls,
+            *,
+            string: str,
+            height: float,
+            font_path: str,
+            offset: NDArray[numpy.float64],
+            rotation: float,
+            mirrored: bool,
+            annotations: annotations_t = None,
+            repetition: Repetition | None = None,
+            ) -> Self:
+        new = cls.__new__(cls)
+        new._offset = offset
+        new._string = string
+        new._height = height
+        new._rotation = rotation % (2 * pi)
+        new._mirrored = mirrored
+        new._repetition = repetition
+        new._annotations = annotations
+        new.font_path = font_path
+        return new
+
     def __deepcopy__(self, memo: dict | None = None) -> Self:
         memo = {} if memo is None else memo
         new = copy.copy(self)

masque/test/test_file_roundtrip.py

@@ -5,7 +5,7 @@ from numpy.testing import assert_allclose
 
 from ..pattern import Pattern
 from ..library import Library
-from ..shapes import Path as MPath, Circle, Polygon
+from ..shapes import Path as MPath, Circle, Polygon, RectCollection
 from ..repetition import Grid, Arbitrary
 
 def create_test_library(for_gds: bool = False) -> Library:
@@ -150,3 +150,30 @@ def test_oasis_full_roundtrip(tmp_path: Path) -> None:
     assert poly.repetition is not None
     assert isinstance(poly.repetition, Grid)
     assert poly.repetition.a_count == 5
+
+
+def test_gdsii_rect_collection_roundtrip(tmp_path: Path) -> None:
+    from ..file import gdsii
+
+    lib = Library()
+    pat = Pattern()
+    pat.shapes[(5, 0)].append(
+        RectCollection(
+            rects=[[0, 0, 10, 5], [20, -5, 30, 10]],
+            annotations={'1': ['rects']},
+            )
+        )
+    lib['rects'] = pat
+
+    gds_file = tmp_path / 'rect_collection.gds'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    read_lib, _ = gdsii.readfile(gds_file)
+    polys = read_lib['rects'].shapes[(5, 0)]
+
+    assert len(polys) == 2
+    assert all(isinstance(poly, Polygon) for poly in polys)
+    assert_allclose(polys[0].vertices, [[0, 0], [0, 5], [10, 5], [10, 0]])
+    assert_allclose(polys[1].vertices, [[20, -5], [20, 10], [30, 10], [30, -5]])
+    assert polys[0].annotations == {'1': ['rects']}
+    assert polys[1].annotations == {'1': ['rects']}

masque/test/test_gdsii_arrow.py

@@ -0,0 +1,603 @@
+from pathlib import Path
+import subprocess
+import sys
+import textwrap
+
+import klamath
+import numpy
+import pytest
+
+pytest.importorskip('pyarrow')
+
+from .. import Ref, Label, PatternError
+from ..library import Library
+from ..pattern import Pattern
+from ..repetition import Grid
+from ..shapes import Path as MPath, Polygon, PolyCollection, RectCollection
+from ..file import gdsii, gdsii_arrow
+from ..file.gdsii_perf import write_fixture
+
+
+if not gdsii_arrow.is_available():
+    pytest.skip('klamath_rs_ext shared library is not available', allow_module_level=True)
+
+
+def _annotations_key(annotations: dict[str, list[object]] | None) -> tuple[tuple[str, tuple[object, ...]], ...] | None:
+    if not annotations:
+        return None
+    return tuple(sorted((key, tuple(values)) for key, values in annotations.items()))
+
+
+def _coord_key(values: object) -> tuple[int, ...] | tuple[tuple[int, int], ...]:
+    arr = numpy.rint(numpy.asarray(values, dtype=float)).astype(int)
+    if arr.ndim == 1:
+        return tuple(arr.tolist())
+    return tuple(tuple(row.tolist()) for row in arr)
+
+
+def _canonical_polygon_key(vertices: object) -> tuple[tuple[int, int], ...]:
+    arr = numpy.rint(numpy.asarray(vertices, dtype=float)).astype(int)
+    rows = [tuple(tuple(row.tolist()) for row in numpy.roll(arr, -shift, axis=0)) for shift in range(arr.shape[0])]
+    rev = arr[::-1]
+    rows.extend(tuple(tuple(row.tolist()) for row in numpy.roll(rev, -shift, axis=0)) for shift in range(rev.shape[0]))
+    return min(rows)
+
+
+def _shape_key(shape: object, layer: tuple[int, int]) -> list[tuple[object, ...]]:
+    if isinstance(shape, MPath):
+        cap_extensions = None if shape.cap_extensions is None else _coord_key(shape.cap_extensions)
+        return [(
+            'path',
+            layer,
+            _coord_key(shape.vertices),
+            _coord_key(shape.offset),
+            int(round(float(shape.width))),
+            shape.cap.name,
+            cap_extensions,
+            _annotations_key(shape.annotations),
+            )]
+
+    keys = []
+    for poly in shape.to_polygons():
+        keys.append((
+            'polygon',
+            layer,
+            _canonical_polygon_key(poly.vertices),
+            _coord_key(poly.offset),
+            _annotations_key(poly.annotations),
+            ))
+    return keys
+
+
+def _ref_keys(target: str, ref: object) -> list[tuple[object, ...]]:
+    keys = []
+    for transform in ref.as_transforms():
+        keys.append((
+            target,
+            _coord_key(transform[:2]),
+            round(float(transform[2]), 8),
+            round(float(transform[4]), 8),
+            bool(int(round(float(transform[3])))),
+            _annotations_key(ref.annotations),
+            ))
+    return keys
+
+
+def _label_key(layer: tuple[int, int], label: object) -> tuple[object, ...]:
+    return (
+        layer,
+        label.string,
+        _coord_key(label.offset),
+        _annotations_key(label.annotations),
+        )
+
+
+def _pattern_summary(pattern: Pattern) -> dict[str, object]:
+    shape_keys: list[tuple[object, ...]] = []
+    for layer, shapes in pattern.shapes.items():
+        for shape in shapes:
+            shape_keys.extend(_shape_key(shape, layer))
+
+    ref_keys: list[tuple[object, ...]] = []
+    for target, refs in pattern.refs.items():
+        for ref in refs:
+            ref_keys.extend(_ref_keys(target, ref))
+
+    label_keys = [
+        _label_key(layer, label)
+        for layer, labels in pattern.labels.items()
+        for label in labels
+        ]
+
+    return {
+        'shapes': sorted(shape_keys),
+        'refs': sorted(ref_keys),
+        'labels': sorted(label_keys),
+        }
+
+
+def _library_summary(lib: Library) -> dict[str, dict[str, object]]:
+    return {name: _pattern_summary(pattern) for name, pattern in lib.items()}
+
+
+def _make_arrow_test_library() -> Library:
+    lib = Library()
+
+    leaf = Pattern()
+    leaf.polygon((1, 0), vertices=[[0, 0], [10, 0], [10, 10], [0, 10]], annotations={'1': ['leaf-poly']})
+    leaf.polygon((2, 0), vertices=[[40, 0], [50, 0], [50, 10], [40, 10]])
+    leaf.polygon((1, 0), vertices=[[20, 0], [30, 0], [30, 10], [20, 10]])
+    leaf.polygon((1, 0), vertices=[[80, 0], [90, 0], [90, 10], [80, 10]])
+    leaf.polygon((2, 0), vertices=[[60, 0], [70, 0], [70, 10], [60, 10]], annotations={'18': ['leaf-poly-2']})
+    leaf.label((10, 0), string='LEAF', offset=(3, 4), annotations={'10': ['leaf-label']})
+    lib['leaf'] = leaf
+
+    child = Pattern()
+    child.path(
+        (2, 0),
+        vertices=[[0, 0], [15, 5], [30, 5]],
+        width=6,
+        cap=MPath.Cap.SquareCustom,
+        cap_extensions=(2, 4),
+        annotations={'2': ['child-path']},
+        )
+    child.label((11, 0), string='CHILD', offset=(7, 8), annotations={'11': ['child-label']})
+    child.ref('leaf', offset=(100, 200), rotation=numpy.pi / 2, mirrored=True, scale=1.25, annotations={'12': ['child-ref']})
+    lib['child'] = child
+
+    sibling = Pattern()
+    sibling.polygon((3, 0), vertices=[[0, 0], [5, 0], [5, 6], [0, 6]])
+    sibling.label((12, 0), string='SIB', offset=(1, 2), annotations={'13': ['sib-label']})
+    sibling.ref(
+        'leaf',
+        offset=(-50, 60),
+        repetition=Grid(a_vector=(20, 0), a_count=3, b_vector=(0, 30), b_count=2),
+        annotations={'14': ['sib-ref']},
+        )
+    lib['sibling'] = sibling
+
+    fanout = Pattern()
+    fanout.ref('leaf', offset=(0, 0))
+    fanout.ref('child', offset=(10, 0), mirrored=True, rotation=numpy.pi / 6, scale=1.1)
+    fanout.ref('leaf', offset=(20, 0))
+    fanout.ref('leaf', offset=(30, 0), repetition=Grid(a_vector=(5, 0), a_count=2, b_vector=(0, 7), b_count=3))
+    fanout.ref('child', offset=(40, 0), mirrored=True, rotation=numpy.pi / 4, scale=1.2,
+               repetition=Grid(a_vector=(9, 0), a_count=2, b_vector=(0, 11), b_count=2))
+    fanout.ref('leaf', offset=(50, 0), repetition=Grid(a_vector=(6, 0), a_count=3, b_vector=(0, 8), b_count=2))
+    fanout.ref('leaf', offset=(60, 0), annotations={'19': ['fanout-sref']})
+    fanout.ref('child', offset=(70, 0), repetition=Grid(a_vector=(4, 0), a_count=2, b_vector=(0, 5), b_count=2),
+               annotations={'20': ['fanout-aref']})
+    lib['fanout'] = fanout
+
+    top = Pattern()
+    top.ref('child', offset=(500, 600), annotations={'15': ['top-child-ref']})
+    top.ref('sibling', offset=(-100, 50), rotation=numpy.pi, annotations={'16': ['top-sibling-ref']})
+    top.ref('fanout', offset=(250, -75))
+    top.label((13, 0), string='TOP', offset=(0, 0), annotations={'17': ['top-label']})
+    lib['top'] = top
+
+    return lib
+
+
+def _write_invalid_path_type_fixture(path: Path) -> None:
+    with path.open('wb') as stream:
+        header = klamath.library.FileHeader(
+            name=b'test',
+            user_units_per_db_unit=1.0,
+            meters_per_db_unit=1e-9,
+        )
+        header.write(stream)
+        elem = klamath.elements.Path(
+            layer=(1, 0),
+            path_type=3,
+            width=10,
+            extension=(0, 0),
+            xy=numpy.array([[0, 0], [10, 0]], dtype=numpy.int32),
+            properties={},
+        )
+        klamath.library.write_struct(stream, name=b'top', elements=[elem])
+        klamath.records.ENDLIB.write(stream, None)
+
+
+def test_gdsii_arrow_matches_gdsii_readfile(tmp_path: Path) -> None:
+    lib = _make_arrow_test_library()
+    gds_file = tmp_path / 'arrow_roundtrip.gds'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    canonical_lib, canonical_info = gdsii.readfile(gds_file)
+    arrow_lib, arrow_info = gdsii_arrow.readfile(gds_file)
+
+    assert canonical_info == arrow_info
+    assert _library_summary(canonical_lib) == _library_summary(arrow_lib)
+
+
+def test_gdsii_arrow_matches_gdsii_readfile_for_gzipped_file(tmp_path: Path) -> None:
+    lib = _make_arrow_test_library()
+    gds_file = tmp_path / 'arrow_roundtrip.gds.gz'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    canonical_lib, canonical_info = gdsii.readfile(gds_file)
+    arrow_lib, arrow_info = gdsii_arrow.readfile(gds_file)
+
+    assert canonical_info == arrow_info
+    assert _library_summary(canonical_lib) == _library_summary(arrow_lib)
+
+
+def test_gdsii_arrow_readfile_arrow_returns_native_payload(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'many_cells_native.gds'
+    manifest = write_fixture(gds_file, preset='many_cells', scale=0.001)
+
+    libarr, info = gdsii_arrow.readfile_arrow(gds_file)
+
+    assert info['name'] == manifest.library_name
+    assert libarr['lib_name'].as_py() == manifest.library_name
+    assert len(libarr['cells']) == manifest.cells
+    assert 0 < len(libarr['layers']) <= manifest.layers
+
+
+def test_gdsii_arrow_readfile_arrow_reads_gzipped_file(tmp_path: Path) -> None:
+    lib = _make_arrow_test_library()
+    gds_file = tmp_path / 'native_payload.gds.gz'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    libarr, info = gdsii_arrow.readfile_arrow(gds_file)
+
+    assert info['name'] == 'masque-klamath'
+    assert libarr['lib_name'].as_py() == 'masque-klamath'
+    assert len(libarr['cells']) == len(lib)
+    assert len(libarr['layers']) > 0
+
+
+def test_gdsii_arrow_removed_raw_mode_arg(tmp_path: Path) -> None:
+    lib = _make_arrow_test_library()
+    gds_file = tmp_path / 'removed_raw_mode.gds'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    libarr, _ = gdsii_arrow.readfile_arrow(gds_file)
+
+    with pytest.raises(TypeError):
+        gdsii_arrow.readfile(gds_file, raw_mode=False)
+
+    with pytest.raises(TypeError):
+        gdsii_arrow.read_arrow(libarr, raw_mode=False)
+
+
+def test_gdsii_arrow_invalid_input_raises_klamath_error(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'invalid.gds'
+    gds_file.write_bytes(b'not-a-gds')
+
+    script = textwrap.dedent(f"""
+        from masque.file import gdsii_arrow
+        try:
+            gdsii_arrow.readfile({str(gds_file)!r})
+        except Exception as exc:
+            print(type(exc).__module__)
+            print(type(exc).__qualname__)
+            print(exc)
+        else:
+            raise SystemExit('expected gdsii_arrow.readfile() to fail')
+        """)
+    result = subprocess.run([sys.executable, '-c', script], capture_output=True, text=True, check=False)
+
+    assert result.returncode == 0, result.stderr
+    assert 'klamath.basic' in result.stdout
+    assert 'KlamathError' in result.stdout
+
+
+def test_gdsii_arrow_reads_small_perf_fixture(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'many_cells_smoke.gds'
+    manifest = write_fixture(gds_file, preset='many_cells', scale=0.001)
+
+    lib, info = gdsii_arrow.readfile(gds_file)
+
+    assert info['name'] == manifest.library_name
+    assert len(lib) == manifest.cells
+    assert 'TOP' in lib
+    assert sum(len(refs) for refs in lib['TOP'].refs.values()) > 0
+
+
+def test_gdsii_arrow_degenerate_aref_decodes_as_single_transform(tmp_path: Path) -> None:
+    lib = Library()
+    leaf = Pattern()
+    leaf.polygon((1, 0), vertices=[[0, 0], [5, 0], [5, 5], [0, 5]])
+    lib['leaf'] = leaf
+
+    top = Pattern()
+    top.ref('leaf', offset=(100, 200), repetition=Grid(a_vector=(7, 0), a_count=1, b_vector=(0, 9), b_count=1))
+    lib['top'] = top
+
+    gds_file = tmp_path / 'degenerate_aref.gds'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    canonical_lib, _ = gdsii.readfile(gds_file)
+    arrow_lib, _ = gdsii_arrow.readfile(gds_file)
+    assert _library_summary(arrow_lib) == _library_summary(canonical_lib)
+
+    decoded_ref = arrow_lib['top'].refs['leaf'][0]
+    assert decoded_ref.repetition is None
+
+
+def test_gdsii_arrow_plain_srefs_decode_without_arbitrary(tmp_path: Path) -> None:
+    lib = _make_arrow_test_library()
+    gds_file = tmp_path / 'plain_srefs.gds'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    arrow_lib, _ = gdsii_arrow.readfile(gds_file)
+    fanout = arrow_lib['fanout']
+
+    plain_leaf_refs = [
+        ref
+        for ref in fanout.refs['leaf']
+        if ref.annotations is None and ref.repetition is None
+        ]
+    assert len(plain_leaf_refs) == 2
+    assert all(type(ref.repetition) is not Grid for ref in plain_leaf_refs)
+
+
+def test_gdsii_arrow_degenerate_aref_schema_normalizes_to_sref(tmp_path: Path) -> None:
+    lib = Library()
+    leaf = Pattern()
+    leaf.polygon((1, 0), vertices=[[0, 0], [5, 0], [5, 5], [0, 5]])
+    lib['leaf'] = leaf
+
+    top = Pattern()
+    top.ref('leaf', offset=(100, 200), repetition=Grid(a_vector=(7, 0), a_count=1, b_vector=(0, 9), b_count=1))
+    lib['top'] = top
+
+    gds_file = tmp_path / 'degenerate_aref_schema.gds'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    libarr = gdsii_arrow._read_to_arrow(gds_file)[0]
+    cells = libarr['cells'].values
+    cell_ids = cells.field('id').to_numpy()
+    cell_names = libarr['cell_names'].as_py()
+    top_index = next(ii for ii, cell_id in enumerate(cell_ids) if cell_names[cell_id] == 'top')
+
+    srefs = cells.field('srefs')[top_index].as_py()
+    arefs = cells.field('arefs')[top_index].as_py()
+
+    assert len(srefs) == 1
+    assert len(arefs) == 0
+    assert cell_names[srefs[0]['target']] == 'leaf'
+
+
+def test_gdsii_arrow_boundary_batch_schema(tmp_path: Path) -> None:
+    lib = _make_arrow_test_library()
+    gds_file = tmp_path / 'arrow_batches.gds'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    libarr = gdsii_arrow._read_to_arrow(gds_file)[0]
+    cells = libarr['cells'].values
+    cell_ids = cells.field('id').to_numpy()
+    cell_names = libarr['cell_names'].as_py()
+    layer_table = [
+        ((int(layer) >> 16) & 0xFFFF, int(layer) & 0xFFFF)
+        for layer in libarr['layers'].values.to_numpy()
+        ]
+
+    leaf_index = next(ii for ii, cell_id in enumerate(cell_ids) if cell_names[cell_id] == 'leaf')
+
+    rect_batches = cells.field('rect_batches')[leaf_index].as_py()
+    boundary_batches = cells.field('boundary_batches')[leaf_index].as_py()
+    boundary_props = cells.field('boundary_props')[leaf_index].as_py()
+
+    assert len(rect_batches) == 2
+    assert len(boundary_batches) == 0
+    assert len(boundary_props) == 2
+
+    rects_by_layer = {tuple(layer_table[entry['layer']]): entry for entry in rect_batches}
+    assert rects_by_layer[(1, 0)]['rects'] == [20, 0, 30, 10, 80, 0, 90, 10]
+    assert rects_by_layer[(2, 0)]['rects'] == [40, 0, 50, 10]
+
+    props_by_layer = {tuple(layer_table[entry['layer']]): entry for entry in boundary_props}
+    assert sorted(props_by_layer) == [(1, 0), (2, 0)]
+    assert props_by_layer[(1, 0)]['properties'][0]['value'] == 'leaf-poly'
+    assert props_by_layer[(2, 0)]['properties'][0]['value'] == 'leaf-poly-2'
+
+
+def test_gdsii_arrow_rect_batch_schema_for_mixed_layer(tmp_path: Path) -> None:
+    lib = Library()
+    top = Pattern()
+    top.shapes[(1, 0)].append(RectCollection(rects=[[0, 0, 10, 10], [20, 0, 30, 10], [40, 0, 50, 10], [60, 0, 70, 10]]))
+    top.polygon((1, 0), vertices=[[80, 0], [85, 10], [90, 0]])
+    top.polygon((1, 0), vertices=[[100, 0], [105, 10], [110, 0]])
+    lib['top'] = top
+
+    gds_file = tmp_path / 'arrow_rect_batches.gds'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    libarr = gdsii_arrow._read_to_arrow(gds_file)[0]
+    cells = libarr['cells'].values
+    cell_ids = cells.field('id').to_numpy()
+    cell_names = libarr['cell_names'].as_py()
+    layer_table = [
+        ((int(layer) >> 16) & 0xFFFF, int(layer) & 0xFFFF)
+        for layer in libarr['layers'].values.to_numpy()
+        ]
+    top_index = next(ii for ii, cell_id in enumerate(cell_ids) if cell_names[cell_id] == 'top')
+
+    rect_batches = cells.field('rect_batches')[top_index].as_py()
+    boundary_batches = cells.field('boundary_batches')[top_index].as_py()
+
+    assert len(rect_batches) == 1
+    assert tuple(layer_table[rect_batches[0]['layer']]) == (1, 0)
+    assert rect_batches[0]['rects'] == [
+        0, 0, 10, 10,
+        20, 0, 30, 10,
+        40, 0, 50, 10,
+        60, 0, 70, 10,
+        ]
+
+    assert len(boundary_batches) == 1
+    assert tuple(layer_table[boundary_batches[0]['layer']]) == (1, 0)
+    assert boundary_batches[0]['vertex_offsets'] == [0, 3]
+
+
+def test_gdsii_arrow_ref_schema(tmp_path: Path) -> None:
+    lib = _make_arrow_test_library()
+    gds_file = tmp_path / 'arrow_ref_batches.gds'
+    gdsii.writefile(lib, gds_file, meters_per_unit=1e-9)
+
+    libarr = gdsii_arrow._read_to_arrow(gds_file)[0]
+    cells = libarr['cells'].values
+    cell_ids = cells.field('id').to_numpy()
+    cell_names = libarr['cell_names'].as_py()
+
+    fanout_index = next(ii for ii, cell_id in enumerate(cell_ids) if cell_names[cell_id] == 'fanout')
+
+    srefs = cells.field('srefs')[fanout_index].as_py()
+    arefs = cells.field('arefs')[fanout_index].as_py()
+    sref_props = cells.field('sref_props')[fanout_index].as_py()
+    aref_props = cells.field('aref_props')[fanout_index].as_py()
+
+    sref_target_ids = [entry['target'] for entry in srefs]
+    sref_targets = [cell_names[target] for target in sref_target_ids]
+    assert sorted(sref_targets) == ['child', 'leaf', 'leaf']
+    assert sref_target_ids == sorted(sref_target_ids)
+    sref_by_target = {}
+    for entry in srefs:
+        sref_by_target.setdefault(cell_names[entry['target']], []).append(entry)
+    assert [entry['invert_y'] for entry in sref_by_target['child']] == [True]
+    assert [entry['scale'] for entry in sref_by_target['child']] == pytest.approx([1.1])
+    assert len(sref_by_target['leaf']) == 2
+
+    aref_target_ids = [entry['target'] for entry in arefs]
+    aref_targets = [cell_names[target] for target in aref_target_ids]
+    assert sorted(aref_targets) == ['child', 'leaf', 'leaf']
+    assert aref_target_ids == sorted(aref_target_ids)
+    aref_by_target = {}
+    for entry in arefs:
+        aref_by_target.setdefault(cell_names[entry['target']], []).append(entry)
+    assert [entry['invert_y'] for entry in aref_by_target['child']] == [True]
+    assert [entry['scale'] for entry in aref_by_target['child']] == pytest.approx([1.2])
+    assert len(aref_by_target['leaf']) == 2
+
+    assert len(sref_props) == 1
+    assert cell_names[sref_props[0]['target']] == 'leaf'
+    assert sref_props[0]['properties'][0]['value'] == 'fanout-sref'
+
+    assert len(aref_props) == 1
+    assert cell_names[aref_props[0]['target']] == 'child'
+    assert aref_props[0]['properties'][0]['value'] == 'fanout-aref'
+
+
+def test_gdsii_arrow_invalid_path_type_matches_gdsii(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'invalid_path_type.gds'
+    _write_invalid_path_type_fixture(gds_file)
+
+    with pytest.raises(PatternError, match='Unrecognized path type: 3'):
+        gdsii.readfile(gds_file)
+
+    with pytest.raises(PatternError, match='Unrecognized path type: 3'):
+        gdsii_arrow.readfile(gds_file)
+
+
+def test_raw_ref_grid_label_constructors_match_public() -> None:
+    raw_grid = Grid._from_raw(
+        a_vector=numpy.array([20, 0]),
+        a_count=3,
+        b_vector=numpy.array([0, 30]),
+        b_count=2,
+        )
+    public_grid = Grid(a_vector=(20, 0), a_count=3, b_vector=(0, 30), b_count=2)
+    assert raw_grid == public_grid
+
+    raw_poly = Polygon._from_raw(
+        vertices=numpy.array([[0.0, 0.0], [5.0, 0.0], [5.0, 5.0], [0.0, 5.0]]),
+        annotations={'1': ['poly']},
+        )
+    public_poly = Polygon(
+        vertices=[[0, 0], [5, 0], [5, 5], [0, 5]],
+        annotations={'1': ['poly']},
+        )
+    assert raw_poly == public_poly
+
+    raw_poly_collection = PolyCollection._from_raw(
+        vertex_lists=numpy.array([
+            [0.0, 0.0], [2.0, 0.0], [2.0, 2.0],
+            [10.0, 10.0], [12.0, 10.0], [12.0, 12.0],
+            ]),
+        vertex_offsets=numpy.array([0, 3], dtype=numpy.uint32),
+        annotations={'2': ['pc']},
+        )
+    public_poly_collection = PolyCollection(
+        vertex_lists=[[0, 0], [2, 0], [2, 2], [10, 10], [12, 10], [12, 12]],
+        vertex_offsets=[0, 3],
+        annotations={'2': ['pc']},
+        )
+    assert raw_poly_collection == public_poly_collection
+    assert [tuple(s.indices(len(raw_poly_collection.vertex_lists))) for s in raw_poly_collection.vertex_slices] == [(0, 3, 1), (3, 6, 1)]
+
+    raw_rect_collection = RectCollection._from_raw(
+        rects=numpy.array([[10.0, 10.0, 12.0, 12.0], [0.0, 0.0, 5.0, 5.0]]),
+        annotations={'3': ['rects']},
+        )
+    public_rect_collection = RectCollection(
+        rects=[[0, 0, 5, 5], [10, 10, 12, 12]],
+        annotations={'3': ['rects']},
+        )
+    assert raw_rect_collection == public_rect_collection
+
+    raw_ref_empty = Ref._from_raw(
+        offset=numpy.array([100, 200]),
+        rotation=numpy.pi / 2,
+        mirrored=False,
+        scale=1.0,
+        repetition=None,
+        annotations=None,
+        )
+    public_ref_empty = Ref(
+        offset=(100, 200),
+        rotation=numpy.pi / 2,
+        mirrored=False,
+        scale=1.0,
+        repetition=None,
+        annotations=None,
+        )
+    assert raw_ref_empty.annotations is None
+    assert raw_ref_empty == public_ref_empty
+
+    raw_ref = Ref._from_raw(
+        offset=numpy.array([100, 200]),
+        rotation=numpy.pi / 2,
+        mirrored=True,
+        scale=1.25,
+        repetition=raw_grid,
+        annotations={'12': ['child-ref']},
+        )
+    public_ref = Ref(
+        offset=(100, 200),
+        rotation=numpy.pi / 2,
+        mirrored=True,
+        scale=1.25,
+        repetition=public_grid,
+        annotations={'12': ['child-ref']},
+        )
+    assert raw_ref == public_ref
+    assert numpy.array_equal(raw_ref.as_transforms(), public_ref.as_transforms())
+
+    raw_label_empty = Label._from_raw(
+        'LEAF',
+        offset=numpy.array([3, 4]),
+        annotations=None,
+        )
+    public_label_empty = Label(
+        'LEAF',
+        offset=(3, 4),
+        annotations=None,
+        )
+    assert raw_label_empty.annotations is None
+    assert raw_label_empty == public_label_empty
+
+    raw_label = Label._from_raw(
+        'LEAF',
+        offset=numpy.array([3, 4]),
+        annotations={'10': ['leaf-label']},
+        )
+    public_label = Label(
+        'LEAF',
+        offset=(3, 4),
+        annotations={'10': ['leaf-label']},
+        )
+    assert raw_label == public_label
+    assert numpy.array_equal(raw_label.get_bounds_single(), public_label.get_bounds_single())

masque/test/test_gdsii_lazy_arrow.py

@@ -0,0 +1,334 @@
+from pathlib import Path
+import subprocess
+import sys
+import textwrap
+
+import klamath
+import numpy
+import pytest
+
+pytest.importorskip('pyarrow')
+
+from .. import PatternError
+from ..library import Library
+from ..pattern import Pattern
+from ..repetition import Grid
+from ..file import gdsii, gdsii_lazy_arrow
+from ..file.gdsii_perf import write_fixture
+
+
+if not gdsii_lazy_arrow.is_available():
+    pytest.skip('klamath_rs_ext shared library is not available', allow_module_level=True)
+
+
+def _make_small_library() -> Library:
+    lib = Library()
+
+    leaf = Pattern()
+    leaf.polygon((1, 0), vertices=[[0, 0], [10, 0], [10, 5], [0, 5]])
+    lib['leaf'] = leaf
+
+    mid = Pattern()
+    mid.ref('leaf', offset=(10, 20))
+    mid.ref('leaf', offset=(40, 0), repetition=Grid(a_vector=(12, 0), a_count=2, b_vector=(0, 9), b_count=2))
+    lib['mid'] = mid
+
+    top = Pattern()
+    top.ref('mid', offset=(100, 200))
+    lib['top'] = top
+    return lib
+
+
+def _make_complex_ref_library() -> Library:
+    lib = Library()
+
+    leaf = Pattern()
+    leaf.polygon((1, 0), vertices=[[0, 0], [10, 0], [10, 10], [0, 10]])
+    lib['leaf'] = leaf
+
+    child = Pattern()
+    child.ref('leaf', offset=(100, 200), rotation=numpy.pi / 2, mirrored=True, scale=1.25)
+    lib['child'] = child
+
+    sibling = Pattern()
+    sibling.ref(
+        'leaf',
+        offset=(-50, 60),
+        repetition=Grid(a_vector=(20, 0), a_count=3, b_vector=(0, 30), b_count=2),
+        )
+    lib['sibling'] = sibling
+
+    fanout = Pattern()
+    fanout.ref('leaf', offset=(0, 0))
+    fanout.ref('child', offset=(10, 0), mirrored=True, rotation=numpy.pi / 6, scale=1.1)
+    fanout.ref('leaf', offset=(30, 0), repetition=Grid(a_vector=(5, 0), a_count=2, b_vector=(0, 7), b_count=3))
+    fanout.ref(
+        'child',
+        offset=(40, 0),
+        mirrored=True,
+        rotation=numpy.pi / 4,
+        scale=1.2,
+        repetition=Grid(a_vector=(9, 0), a_count=2, b_vector=(0, 11), b_count=2),
+        )
+    lib['fanout'] = fanout
+
+    top = Pattern()
+    top.ref('child', offset=(500, 600))
+    top.ref('sibling', offset=(-100, 50), rotation=numpy.pi)
+    top.ref('fanout', offset=(250, -75))
+    lib['top'] = top
+
+    return lib
+
+
+def _write_invalid_path_type_fixture(path: Path) -> None:
+    with path.open('wb') as stream:
+        header = klamath.library.FileHeader(
+            name=b'test',
+            user_units_per_db_unit=1.0,
+            meters_per_db_unit=1e-9,
+        )
+        header.write(stream)
+        elem = klamath.elements.Path(
+            layer=(1, 0),
+            path_type=3,
+            width=10,
+            extension=(0, 0),
+            xy=numpy.array([[0, 0], [10, 0]], dtype=numpy.int32),
+            properties={},
+        )
+        klamath.library.write_struct(stream, name=b'top', elements=[elem])
+        klamath.records.ENDLIB.write(stream, None)
+
+
+def _transform_rows_key(values: numpy.ndarray) -> tuple[tuple[object, ...], ...]:
+    arr = numpy.asarray(values, dtype=float)
+    arr = numpy.atleast_2d(arr)
+    rows = [
+        (
+            round(float(row[0]), 8),
+            round(float(row[1]), 8),
+            round(float(row[2]), 8),
+            bool(int(round(float(row[3])))),
+            round(float(row[4]), 8),
+            )
+        for row in arr
+        ]
+    return tuple(sorted(rows))
+
+
+def _local_refs_key(refs: dict[str, list[numpy.ndarray]]) -> dict[str, tuple[tuple[object, ...], ...]]:
+    return {
+        parent: _transform_rows_key(numpy.concatenate(transforms))
+        for parent, transforms in refs.items()
+        }
+
+
+def _global_refs_key(refs: dict[tuple[str, ...], numpy.ndarray]) -> dict[tuple[str, ...], tuple[tuple[object, ...], ...]]:
+    return {
+        path: _transform_rows_key(transforms)
+        for path, transforms in refs.items()
+        }
+
+
+def test_gdsii_lazy_arrow_loads_perf_fixture(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'many_cells_lazy.gds'
+    manifest = write_fixture(gds_file, preset='many_cells', scale=0.001)
+
+    lib, info = gdsii_lazy_arrow.readfile(gds_file)
+
+    assert info['name'] == manifest.library_name
+    assert len(lib) == manifest.cells
+    assert lib.top() == 'TOP'
+    assert 'TOP' in lib.child_graph(dangling='ignore')
+
+
+def test_gdsii_lazy_arrow_local_and_global_refs(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'refs.gds'
+    src = _make_small_library()
+    gdsii.writefile(src, gds_file, meters_per_unit=1e-9, library_name='lazy-refs')
+
+    lib, _ = gdsii_lazy_arrow.readfile(gds_file)
+
+    local = lib.find_refs_local('leaf')
+    assert set(local) == {'mid'}
+    assert sum(arr.shape[0] for arr in local['mid']) == 5
+
+    global_refs = lib.find_refs_global('leaf')
+    assert {path for path in global_refs} == {('top', 'mid', 'leaf')}
+    assert global_refs[('top', 'mid', 'leaf')].shape[0] == 5
+
+
+def test_gdsii_lazy_arrow_ref_queries_match_eager_reader(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'complex_refs.gds'
+    src = _make_complex_ref_library()
+    gdsii.writefile(src, gds_file, meters_per_unit=1e-9, library_name='lazy-complex-refs')
+
+    eager, _ = gdsii.readfile(gds_file)
+    lazy, _ = gdsii_lazy_arrow.readfile(gds_file)
+
+    for name in ('leaf', 'child'):
+        assert _local_refs_key(lazy.find_refs_local(name)) == _local_refs_key(eager.find_refs_local(name))
+        assert _global_refs_key(lazy.find_refs_global(name)) == _global_refs_key(eager.find_refs_global(name))
+
+
+def test_gdsii_lazy_arrow_invalid_input_raises_klamath_error(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'invalid.gds'
+    gds_file.write_bytes(b'not-a-gds')
+
+    script = textwrap.dedent(f"""
+        from masque.file import gdsii_lazy_arrow
+        try:
+            gdsii_lazy_arrow.readfile({str(gds_file)!r})
+        except Exception as exc:
+            print(type(exc).__module__)
+            print(type(exc).__qualname__)
+            print(exc)
+        else:
+            raise SystemExit('expected gdsii_lazy_arrow.readfile() to fail')
+        """)
+    result = subprocess.run([sys.executable, '-c', script], capture_output=True, text=True, check=False)
+
+    assert result.returncode == 0, result.stderr
+    assert 'klamath.basic' in result.stdout
+    assert 'KlamathError' in result.stdout
+
+
+def test_gdsii_lazy_arrow_invalid_path_type_raises_pattern_error(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'invalid_path_type.gds'
+    _write_invalid_path_type_fixture(gds_file)
+
+    lib, _ = gdsii_lazy_arrow.readfile(gds_file)
+
+    with pytest.raises(PatternError, match='Unrecognized path type: 3'):
+        lib['top']
+
+
+def test_gdsii_lazy_arrow_untouched_write_is_copy_through(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'copy_source.gds'
+    src = _make_small_library()
+    gdsii.writefile(src, gds_file, meters_per_unit=1e-9, library_name='copy-through')
+
+    lib, info = gdsii_lazy_arrow.readfile(gds_file)
+    out_file = tmp_path / 'copy_out.gds'
+    gdsii_lazy_arrow.writefile(
+        lib,
+        out_file,
+        meters_per_unit=info['meters_per_unit'],
+        logical_units_per_unit=info['logical_units_per_unit'],
+        library_name=info['name'],
+    )
+
+    assert out_file.read_bytes() == gds_file.read_bytes()
+
+
+def test_gdsii_lazy_arrow_gzipped_copy_through(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'copy_source.gds.gz'
+    src = _make_small_library()
+    gdsii.writefile(src, gds_file, meters_per_unit=1e-9, library_name='copy-through-gz')
+
+    lib, info = gdsii_lazy_arrow.readfile(gds_file)
+    out_file = tmp_path / 'copy_out.gds.gz'
+    gdsii_lazy_arrow.writefile(
+        lib,
+        out_file,
+        meters_per_unit=info['meters_per_unit'],
+        logical_units_per_unit=info['logical_units_per_unit'],
+        library_name=info['name'],
+    )
+
+    assert out_file.read_bytes() == gds_file.read_bytes()
+
+
+def test_gdsii_lazy_overlay_merge_and_write(tmp_path: Path) -> None:
+    base_a = Library()
+    leaf_a = Pattern()
+    leaf_a.polygon((1, 0), vertices=[[0, 0], [8, 0], [8, 8], [0, 8]])
+    base_a['leaf'] = leaf_a
+    top_a = Pattern()
+    top_a.ref('leaf', offset=(0, 0))
+    base_a['top_a'] = top_a
+
+    base_b = Library()
+    leaf_b = Pattern()
+    leaf_b.polygon((2, 0), vertices=[[0, 0], [5, 0], [5, 5], [0, 5]])
+    base_b['leaf'] = leaf_b
+    top_b = Pattern()
+    top_b.ref('leaf', offset=(20, 30))
+    base_b['top_b'] = top_b
+
+    gds_a = tmp_path / 'a.gds'
+    gds_b = tmp_path / 'b.gds'
+    gdsii.writefile(base_a, gds_a, meters_per_unit=1e-9, library_name='overlay')
+    gdsii.writefile(base_b, gds_b, meters_per_unit=1e-9, library_name='overlay')
+
+    lib_a, _ = gdsii_lazy_arrow.readfile(gds_a)
+    lib_b, _ = gdsii_lazy_arrow.readfile(gds_b)
+
+    overlay = gdsii_lazy_arrow.OverlayLibrary()
+    overlay.add_source(lib_a)
+    rename_map = overlay.add_source(lib_b, rename_theirs=lambda lib, name: lib.get_name(name))
+    renamed_leaf = rename_map['leaf']
+
+    assert rename_map == {'leaf': renamed_leaf}
+    assert renamed_leaf != 'leaf'
+    assert len(lib_a._cache) == 0
+    assert len(lib_b._cache) == 0
+
+    overlay.move_references('leaf', renamed_leaf)
+
+    out_file = tmp_path / 'overlay_out.gds'
+    gdsii_lazy_arrow.writefile(overlay, out_file)
+
+    roundtrip, _ = gdsii.readfile(out_file)
+    assert set(roundtrip.keys()) == {'leaf', renamed_leaf, 'top_a', 'top_b'}
+    assert 'top_b' in roundtrip
+    assert list(roundtrip['top_b'].refs.keys()) == [renamed_leaf]
+
+
+def test_gdsii_writer_accepts_overlay_library(tmp_path: Path) -> None:
+    gds_file = tmp_path / 'overlay_source.gds'
+    src = _make_small_library()
+    gdsii.writefile(src, gds_file, meters_per_unit=1e-9, library_name='overlay-src')
+
+    lib, info = gdsii_lazy_arrow.readfile(gds_file)
+
+    overlay = gdsii_lazy_arrow.OverlayLibrary()
+    overlay.add_source(lib)
+    overlay.rename('leaf', 'leaf_copy', move_references=True)
+
+    out_file = tmp_path / 'overlay_via_eager_writer.gds'
+    gdsii.writefile(
+        overlay,
+        out_file,
+        meters_per_unit=info['meters_per_unit'],
+        logical_units_per_unit=info['logical_units_per_unit'],
+        library_name=info['name'],
+    )
+
+    roundtrip, _ = gdsii.readfile(out_file)
+    assert set(roundtrip.keys()) == {'leaf_copy', 'mid', 'top'}
+    assert list(roundtrip['mid'].refs.keys()) == ['leaf_copy']
+
+
+def test_svg_writer_uses_detached_materialized_copy(tmp_path: Path) -> None:
+    pytest.importorskip('svgwrite')
+    from ..file import svg
+    from ..shapes import Path as MPath
+
+    gds_file = tmp_path / 'svg_source.gds'
+    src = _make_small_library()
+    src['top'].path((3, 0), vertices=[[0, 0], [0, 20]], width=4)
+    gdsii.writefile(src, gds_file, meters_per_unit=1e-9, library_name='svg-src')
+
+    lib, _ = gdsii_lazy_arrow.readfile(gds_file)
+    top_pat = lib['top']
+    assert list(top_pat.refs.keys()) == ['mid']
+    assert any(isinstance(shape, MPath) for shape in top_pat.shapes[(3, 0)])
+
+    svg_path = tmp_path / 'lazy.svg'
+    svg.writefile(lib, 'top', str(svg_path))
+
+    assert svg_path.exists()
+    assert list(top_pat.refs.keys()) == ['mid']
+    assert any(isinstance(shape, MPath) for shape in top_pat.shapes[(3, 0)])

masque/test/test_gdsii_perf.py

@@ -0,0 +1,24 @@
+from dataclasses import asdict
+import json
+from pathlib import Path
+
+from ..file import gdsii
+from ..file.gdsii_perf import fixture_manifest, write_fixture
+
+
+def test_gdsii_perf_fixture_smoke(tmp_path: Path) -> None:
+    output = tmp_path / 'many_cells.gds'
+    manifest = write_fixture(output, preset='many_cells', scale=0.002)
+    expected = fixture_manifest(output, preset='many_cells', scale=0.002)
+
+    assert output.exists()
+    assert manifest == expected
+
+    sidecar = json.loads(output.with_suffix('.gds.json').read_text())
+    assert sidecar == asdict(manifest)
+
+    read_lib, info = gdsii.readfile(output)
+    assert info['name'] == manifest.library_name
+    assert len(read_lib) == manifest.cells
+    assert 'TOP' in read_lib
+    assert len(read_lib['TOP'].refs) > 0

masque/test/test_raw_constructors.py

@@ -0,0 +1,97 @@
+import numpy
+from numpy import pi
+from numpy.testing import assert_allclose
+
+from ..shapes import Arc, Circle, Ellipse, Path, Text
+
+
+def test_circle_raw_constructor_matches_public() -> None:
+    raw = Circle._from_raw(
+        radius=5.0,
+        offset=numpy.array([1.0, 2.0]),
+        annotations={'1': ['circle']},
+        )
+    public = Circle(
+        radius=5.0,
+        offset=(1.0, 2.0),
+        annotations={'1': ['circle']},
+        )
+    assert raw == public
+
+
+def test_ellipse_raw_constructor_matches_public() -> None:
+    raw = Ellipse._from_raw(
+        radii=numpy.array([3.0, 5.0]),
+        offset=numpy.array([1.0, 2.0]),
+        rotation=5 * pi / 2,
+        annotations={'2': ['ellipse']},
+        )
+    public = Ellipse(
+        radii=(3.0, 5.0),
+        offset=(1.0, 2.0),
+        rotation=5 * pi / 2,
+        annotations={'2': ['ellipse']},
+        )
+    assert raw == public
+
+
+def test_arc_raw_constructor_matches_public() -> None:
+    raw = Arc._from_raw(
+        radii=numpy.array([10.0, 6.0]),
+        angles=numpy.array([0.0, pi / 2]),
+        width=2.0,
+        offset=numpy.array([1.0, 2.0]),
+        rotation=5 * pi / 2,
+        annotations={'3': ['arc']},
+        )
+    public = Arc(
+        radii=(10.0, 6.0),
+        angles=(0.0, pi / 2),
+        width=2.0,
+        offset=(1.0, 2.0),
+        rotation=5 * pi / 2,
+        annotations={'3': ['arc']},
+        )
+    assert raw == public
+
+
+def test_path_raw_constructor_matches_public() -> None:
+    raw = Path._from_raw(
+        vertices=numpy.array([[0.0, 0.0], [10.0, 0.0], [10.0, 5.0]]),
+        width=2.0,
+        cap=Path.Cap.SquareCustom,
+        cap_extensions=numpy.array([1.0, 3.0]),
+        annotations={'4': ['path']},
+        )
+    public = Path(
+        vertices=((0.0, 0.0), (10.0, 0.0), (10.0, 5.0)),
+        width=2.0,
+        cap=Path.Cap.SquareCustom,
+        cap_extensions=(1.0, 3.0),
+        annotations={'4': ['path']},
+        )
+    assert raw == public
+    assert raw.cap_extensions is not None
+    assert_allclose(raw.cap_extensions, [1.0, 3.0])
+
+
+def test_text_raw_constructor_matches_public() -> None:
+    raw = Text._from_raw(
+        string='RAW',
+        height=12.0,
+        font_path='font.otf',
+        offset=numpy.array([1.0, 2.0]),
+        rotation=5 * pi / 2,
+        mirrored=True,
+        annotations={'5': ['text']},
+        )
+    public = Text(
+        string='RAW',
+        height=12.0,
+        font_path='font.otf',
+        offset=(1.0, 2.0),
+        rotation=5 * pi / 2,
+        mirrored=True,
+        annotations={'5': ['text']},
+        )
+    assert raw == public

masque/test/test_rect_collection.py

@@ -0,0 +1,70 @@
+import copy
+
+import numpy
+import pytest
+from numpy.testing import assert_allclose, assert_equal
+
+from ..error import PatternError
+from ..shapes import Polygon, RectCollection
+
+
+def test_rect_collection_init_and_to_polygons() -> None:
+    rects = RectCollection([[10, 10, 12, 12], [0, 0, 5, 5]])
+    assert_equal(rects.rects, [[0, 0, 5, 5], [10, 10, 12, 12]])
+
+    polys = rects.to_polygons()
+    assert len(polys) == 2
+    assert all(isinstance(poly, Polygon) for poly in polys)
+    assert_equal(polys[0].vertices, [[0, 0], [0, 5], [5, 5], [5, 0]])
+
+
+def test_rect_collection_rejects_invalid_rects() -> None:
+    with pytest.raises(PatternError):
+        RectCollection([[0, 0, 1]])
+    with pytest.raises(PatternError):
+        RectCollection([[5, 0, 1, 2]])
+    with pytest.raises(PatternError):
+        RectCollection([[0, 5, 1, 2]])
+
+
+def test_rect_collection_raw_constructor_matches_public() -> None:
+    raw = RectCollection._from_raw(
+        rects=numpy.array([[10.0, 10.0, 12.0, 12.0], [0.0, 0.0, 5.0, 5.0]]),
+        annotations={'1': ['rects']},
+        )
+    public = RectCollection(
+        [[0, 0, 5, 5], [10, 10, 12, 12]],
+        annotations={'1': ['rects']},
+        )
+    assert raw == public
+    assert_equal(raw.get_bounds_single(), [[0, 0], [12, 12]])
+
+
+def test_rect_collection_manhattan_transforms() -> None:
+    rects = RectCollection([[0, 0, 2, 4], [10, 20, 12, 22]])
+
+    mirrored = copy.deepcopy(rects).mirror(1)
+    assert_equal(mirrored.rects, [[-2, 0, 0, 4], [-12, 20, -10, 22]])
+
+    scaled = copy.deepcopy(rects).scale_by(-2)
+    assert_equal(scaled.rects, [[-4, -8, 0, 0], [-24, -44, -20, -40]])
+
+    rotated = copy.deepcopy(rects).rotate(numpy.pi / 2)
+    assert_equal(rotated.rects, [[-4, 0, 0, 2], [-22, 10, -20, 12]])
+
+
+def test_rect_collection_non_manhattan_rotation_raises() -> None:
+    rects = RectCollection([[0, 0, 2, 4]])
+    with pytest.raises(PatternError, match='Manhattan rotations'):
+        rects.rotate(numpy.pi / 4)
+
+
+def test_rect_collection_normalized_form_rebuild_is_independent() -> None:
+    rects = RectCollection([[0, 0, 2, 4], [10, 20, 12, 22]])
+    _intrinsic, extrinsic, rebuild = rects.normalized_form(2)
+
+    clone = rebuild()
+    clone.rects[:] = [[1, 1, 2, 2], [3, 3, 4, 4]]
+
+    assert_allclose(extrinsic[0], [6, 11.5])
+    assert_equal(rects.rects, [[0, 0, 2, 4], [10, 20, 12, 22]])

masque/utils/comparisons.py

@@ -9,7 +9,15 @@ def annotation2key(aaa: int | float | str) -> tuple[bool, Any]:
     return (isinstance(aaa, str), aaa)
 
 
+def _normalized_annotations(annotations: annotations_t) -> annotations_t:
+    if not annotations:
+        return None
+    return annotations
+
+
 def annotations_lt(aa: annotations_t, bb: annotations_t) -> bool:
+    aa = _normalized_annotations(aa)
+    bb = _normalized_annotations(bb)
     if aa is None:
         return bb is not None
     elif bb is None:            # noqa: RET505
@@ -36,6 +44,8 @@ def annotations_lt(aa: annotations_t, bb: annotations_t) -> bool:
 
 
 def annotations_eq(aa: annotations_t, bb: annotations_t) -> bool:
+    aa = _normalized_annotations(aa)
+    bb = _normalized_annotations(bb)
     if aa is None:
         return bb is None
     elif bb is None:            # noqa: RET505

pyproject.toml

@@ -44,7 +44,7 @@ dependencies = [
 
 [dependency-groups]
 dev = [
-    "pytest",
+    "masque[arrow]",
     "masque[oasis]",
     "masque[dxf]",
     "masque[svg]",
@@ -52,6 +52,7 @@ dev = [
     "masque[text]",
     "masque[manhattanize]",
     "masque[manhattanize_slow]",
+    "masque[boolean]",
     "matplotlib>=3.10.8",
     "pytest>=9.0.2",
     "ruff>=0.15.5",
@@ -66,6 +67,7 @@ build-backend = "hatchling.build"
 path = "masque/__init__.py"
 
 [project.optional-dependencies]
+arrow = ["pyarrow", "cffi"]
 oasis = ["fatamorgana~=0.11"]
 dxf = ["ezdxf~=1.4"]
 svg = ["svgwrite"]
@@ -121,4 +123,3 @@ mypy_path = "stubs"
 python_version = "3.11"
 strict = false
 check_untyped_defs = true
-