warp-lang 1.4.1__py3-none-manylinux2014_aarch64.whl → 1.5.0__py3-none-manylinux2014_aarch64.whl

warp/tests/unittest_suites.py

@@ -99,8 +99,11 @@ def default_suite(test_loader: unittest.TestLoader = unittest.defaultTestLoader)
     from warp.tests.test_closest_point_edge_edge import TestClosestPointEdgeEdgeMethods
     from warp.tests.test_codegen import TestCodeGen
     from warp.tests.test_codegen_instancing import TestCodeGenInstancing
+    from warp.tests.test_collision import TestCollision
+    from warp.tests.test_coloring import TestColoring
     from warp.tests.test_compile_consts import TestConstants
     from warp.tests.test_conditional import TestConditional
+    from warp.tests.test_context import TestContext
     from warp.tests.test_copy import TestCopy
     from warp.tests.test_ctypes import TestCTypes
     from warp.tests.test_dense import TestDense
@@ -115,7 +118,7 @@ def default_suite(test_loader: unittest.TestLoader = unittest.defaultTestLoader)
     )
     from warp.tests.test_fabricarray import TestFabricArray
     from warp.tests.test_fast_math import TestFastMath
-    from warp.tests.test_fem import TestFem, TestFemShapeFunctions
+    from warp.tests.test_fem import TestFem, TestFemShapeFunctions, TestFemUtilities
     from warp.tests.test_fp16 import TestFp16
     from warp.tests.test_func import TestFunc
     from warp.tests.test_future_annotations import TestFutureAnnotations
@@ -127,6 +130,7 @@ def default_suite(test_loader: unittest.TestLoader = unittest.defaultTestLoader)
     from warp.tests.test_import import TestImport
     from warp.tests.test_indexedarray import TestIndexedArray
     from warp.tests.test_intersect import TestIntersect
+    from warp.tests.test_iter import TestIter
     from warp.tests.test_jax import TestJax
     from warp.tests.test_large import TestLarge
     from warp.tests.test_launch import TestLaunch
@@ -170,9 +174,14 @@ def default_suite(test_loader: unittest.TestLoader = unittest.defaultTestLoader)
     from warp.tests.test_sparse import TestSparse
     from warp.tests.test_spatial import TestSpatial
     from warp.tests.test_special_values import TestSpecialValues
+    from warp.tests.test_static import TestStatic
     from warp.tests.test_streams import TestStreams
     from warp.tests.test_struct import TestStruct
     from warp.tests.test_tape import TestTape
+    from warp.tests.test_tile import TestTile
+    from warp.tests.test_tile_mathdx import TestTileMathDx
+    from warp.tests.test_tile_reduce import TestTileReduce
+    from warp.tests.test_tile_shared_memory import TestTileSharedMemory
     from warp.tests.test_torch import TestTorch
     from warp.tests.test_transient_module import TestTransientModule
     from warp.tests.test_triangle_closest_point import TestTriangleClosestPoint
@@ -199,8 +208,11 @@ def default_suite(test_loader: unittest.TestLoader = unittest.defaultTestLoader)
         TestClosestPointEdgeEdgeMethods,
         TestCodeGen,
         TestCodeGenInstancing,
-        TestConstants,
+        TestCollision,
+        TestColoring,
         TestConditional,
+        TestConstants,
+        TestContext,
         TestCopy,
         TestCTypes,
         TestDense,
@@ -215,6 +227,7 @@ def default_suite(test_loader: unittest.TestLoader = unittest.defaultTestLoader)
         TestFastMath,
         TestFem,
         TestFemShapeFunctions,
+        TestFemUtilities,
         TestFp16,
         TestFunc,
         TestFutureAnnotations,
@@ -226,6 +239,7 @@ def default_suite(test_loader: unittest.TestLoader = unittest.defaultTestLoader)
         TestImport,
         TestIndexedArray,
         TestIntersect,
+        TestIter,
         TestJax,
         TestLarge,
         TestLaunch,
@@ -269,9 +283,14 @@ def default_suite(test_loader: unittest.TestLoader = unittest.defaultTestLoader)
         TestSparse,
         TestSpatial,
         TestSpecialValues,
+        TestStatic,
         TestStreams,
         TestStruct,
         TestTape,
+        TestTile,
+        TestTileMathDx,
+        TestTileReduce,
+        TestTileSharedMemory,
         TestTorch,
         TestTransientModule,
         TestTriangleClosestPoint,
@@ -329,6 +348,7 @@ def kit_suite(test_loader: unittest.TestLoader = unittest.defaultTestLoader):
     from warp.tests.test_rounding import TestRounding
     from warp.tests.test_runlength_encode import TestRunlengthEncode
     from warp.tests.test_sparse import TestSparse
+    from warp.tests.test_static import TestStatic
     from warp.tests.test_streams import TestStreams
     from warp.tests.test_tape import TestTape
     from warp.tests.test_transient_module import TestTransientModule
@@ -374,6 +394,7 @@ def kit_suite(test_loader: unittest.TestLoader = unittest.defaultTestLoader):
         TestRounding,
         TestRunlengthEncode,
         TestSparse,
+        TestStatic,
         TestStreams,
         TestTape,
         TestTransientModule,

warp/tests/unittest_utils.py

@@ -232,6 +232,10 @@ def create_test_func(func, device, check_output, **kwargs):
         else:
             func(self, device, **kwargs)
 
+    # Copy the __unittest_expecting_failure__ attribute from func to test_func
+    if hasattr(func, "__unittest_expecting_failure__"):
+        test_func.__unittest_expecting_failure__ = func.__unittest_expecting_failure__
+
     return test_func
 
 

warp/types.py

@@ -12,9 +12,10 @@ import ctypes
 import inspect
 import struct
 import zlib
-from typing import Any, Callable, Generic, List, NamedTuple, Optional, Sequence, Tuple, TypeVar, Union
+from typing import Any, Callable, Generic, List, Literal, NamedTuple, Optional, Sequence, Tuple, TypeVar, Union
 
 import numpy as np
+import numpy.typing as npt
 
 import warp
 
@@ -100,8 +101,10 @@ def vector(length, dtype):
 
         if dtype is bool:
             _type_ = ctypes.c_bool
-        elif dtype in [Scalar, Float]:
+        elif dtype in (Scalar, Float):
             _type_ = ctypes.c_float
+        elif dtype is Int:
+            _type_ = ctypes.c_int
         else:
             _type_ = dtype._type_
 
@@ -289,8 +292,10 @@ def matrix(shape, dtype):
 
         if dtype is bool:
             _type_ = ctypes.c_bool
-        elif dtype in [Scalar, Float]:
+        elif dtype in (Scalar, Float):
             _type_ = ctypes.c_float
+        elif dtype is Int:
+            _type_ = ctypes.c_int
         else:
             _type_ = dtype._type_
 
@@ -991,43 +996,6 @@ vector_types = (
     spatial_matrixd,
 )
 
-atomic_vector_types = (
-    vec2i,
-    vec2ui,
-    vec2l,
-    vec2ul,
-    vec2h,
-    vec2f,
-    vec2d,
-    vec3i,
-    vec3ui,
-    vec3l,
-    vec3ul,
-    vec3h,
-    vec3f,
-    vec3d,
-    vec4i,
-    vec4ui,
-    vec4l,
-    vec4ul,
-    vec4h,
-    vec4f,
-    vec4d,
-    mat22h,
-    mat22f,
-    mat22d,
-    mat33h,
-    mat33f,
-    mat33d,
-    mat44h,
-    mat44f,
-    mat44d,
-    quath,
-    quatf,
-    quatd,
-)
-atomic_types = float_types + (int32, uint32, int64, uint64) + atomic_vector_types
-
 np_dtype_to_warp_type = {
     # Numpy scalar types
     np.bool_: bool,
@@ -1076,6 +1044,14 @@ warp_type_to_np_dtype = {
     float64: np.float64,
 }
 
+non_atomic_types = (
+    int8,
+    uint8,
+    int16,
+    uint16,
+    int64,
+)
+
 
 def dtype_from_numpy(numpy_dtype):
     """Return the Warp dtype corresponding to a NumPy dtype."""
@@ -1337,6 +1313,8 @@ def type_typestr(dtype):
 def type_repr(t):
     if is_array(t):
         return str(f"array(ndim={t.ndim}, dtype={t.dtype})")
+    if is_tile(t):
+        return str(f"tile(dtype={t.dtype}, m={t.M}, n={t.N})")
     if type_is_vector(t):
         return str(f"vector(length={t._shape_[0]}, dtype={t._wp_scalar_type_})")
     if type_is_matrix(t):
@@ -1448,6 +1426,8 @@ def scalars_equal(a, b, match_generic):
 
 def types_equal(a, b, match_generic=False):
     if match_generic:
+        # Special cases to interpret the types listed in `int_tuple_type_hints`
+        # as generic hints that accept any integer types.
         if a in int_tuple_type_hints and isinstance(b, Sequence):
             a_length = int_tuple_type_hints[a]
             if (a_length == -1 or a_length == len(b)) and all(
@@ -1466,6 +1446,24 @@ def types_equal(a, b, match_generic=False):
             if a_length is None or b_length is None or a_length == b_length:
                 return True
 
+    a_origin = warp.codegen.get_type_origin(a)
+    b_origin = warp.codegen.get_type_origin(b)
+    if a_origin is tuple and b_origin is tuple:
+        a_args = warp.codegen.get_type_args(a)
+        b_args = warp.codegen.get_type_args(b)
+        if len(a_args) == len(b_args) and all(
+            scalars_equal(x, y, match_generic=match_generic) for x, y in zip(a_args, b_args)
+        ):
+            return True
+    elif a_origin is tuple and isinstance(b, Sequence):
+        a_args = warp.codegen.get_type_args(a)
+        if len(a_args) == len(b) and all(scalars_equal(x, y, match_generic=match_generic) for x, y in zip(a_args, b)):
+            return True
+    elif b_origin is tuple and isinstance(a, Sequence):
+        b_args = warp.codegen.get_type_args(b)
+        if len(b_args) == len(a) and all(scalars_equal(x, y, match_generic=match_generic) for x, y in zip(b_args, a)):
+            return True
+
     # convert to canonical types
     if a == float:
         a = float32
@@ -1488,13 +1486,16 @@ def types_equal(a, b, match_generic=False):
 
         return True
 
-    if is_array(a) and type(a) is type(b):
+    if is_array(a) and type(a) is type(b) and types_equal(a.dtype, b.dtype, match_generic=match_generic):
         return True
 
     # match NewStructInstance and Struct dtype
     if getattr(a, "cls", "a") is getattr(b, "cls", "b"):
         return True
 
+    if is_tile(a) and is_tile(b):
+        return True
+
     return scalars_equal(a, b, match_generic)
 
 
@@ -1581,6 +1582,23 @@ def array_ctype_from_interface(interface: dict, dtype=None, owner=None):
 
 
 class array(Array):
+    """A fixed-size multi-dimensional array containing values of the same type.
+
+    Attributes:
+        dtype (DType): The data type of the array.
+        ndim (int): The number of array dimensions.
+        size (int): The number of items in the array.
+        capacity (int): The amount of memory in bytes allocated for this array.
+        shape (Tuple[int]): Dimensions of the array.
+        strides (Tuple[int]): Number of bytes in each dimension between successive elements of the array.
+        ptr (int): Pointer to underlying memory allocation backing the array.
+        device (Device): The device where the array's memory allocation resides.
+        pinned (bool): Indicates whether the array was allocated in pinned host memory.
+        is_contiguous (bool): Indicates whether this array has a contiguous memory layout.
+        deleter (Callable[[int, int], None]): A function to be called when the array is deleted,
+            taking two arguments: pointer and size. If ``None``, then no function is called.
+    """
+
     # member attributes available during code-gen (e.g.: d = array.shape[0])
     # (initialized when needed)
     _vars = None
@@ -1592,21 +1610,21 @@ class array(Array):
 
     def __init__(
         self,
-        data=None,
-        dtype: DType = Any,
-        shape=None,
-        strides=None,
-        length=None,
-        ptr=None,
-        capacity=None,
+        data: Optional[Union[List, Tuple, npt.NDArray]] = None,
+        dtype: Union[DType, Any] = Any,
+        shape: Optional[Tuple[int, ...]] = None,
+        strides: Optional[Tuple[int, ...]] = None,
+        length: Optional[int] = None,
+        ptr: Optional[int] = None,
+        capacity: Optional[int] = None,
         device=None,
-        pinned=False,
-        copy=True,
-        owner=False,  # deprecated - pass deleter instead
-        deleter=None,
-        ndim=None,
-        grad=None,
-        requires_grad=False,
+        pinned: bool = False,
+        copy: bool = True,
+        owner: bool = False,  # deprecated - pass deleter instead
+        deleter: Optional[Callable[[int, int], None]] = None,
+        ndim: Optional[int] = None,
+        grad: Optional[array] = None,
+        requires_grad: bool = False,
     ):
         """Constructs a new Warp array object
 
@@ -1628,20 +1646,24 @@ class array(Array):
         are taken into account and no memory is allocated for the array.
 
         Args:
-            data (Union[list, tuple, ndarray]): An object to construct the array from, can be a Tuple, List, or generally any type convertible to an np.array
-            dtype (Union): One of the available `data types <#data-types>`_, such as :class:`warp.float32`, :class:`warp.mat33`, or a custom `struct <#structs>`_. If dtype is ``Any`` and data is an ndarray, then it will be inferred from the array data type
-            shape (tuple): Dimensions of the array
-            strides (tuple): Number of bytes in each dimension between successive elements of the array
-            length (int): Number of elements of the data type (deprecated, users should use `shape` argument)
-            ptr (uint64): Address of an external memory address to alias (data should be None)
-            capacity (int): Maximum size in bytes of the ptr allocation (data should be None)
+            data: An object to construct the array from, can be a Tuple, List, or generally any type convertible to an np.array
+            dtype: One of the available `data types <#data-types>`_, such as :class:`warp.float32`, :class:`warp.mat33`, or a custom `struct <#structs>`_. If dtype is ``Any`` and data is an ndarray, then it will be inferred from the array data type
+            shape: Dimensions of the array
+            strides: Number of bytes in each dimension between successive elements of the array
+            length: Number of elements of the data type (deprecated, users should use ``shape`` argument)
+            ptr: Address of an external memory address to alias (``data`` should be ``None``)
+            capacity: Maximum size in bytes of the ``ptr`` allocation (``data`` should be ``None``)
             device (Devicelike): Device the array lives on
-            copy (bool): Whether the incoming data will be copied or aliased, this is only possible when the incoming `data` already lives on the device specified and types match
-            owner (bool): Should the array object try to deallocate memory when it is deleted (deprecated, pass `deleter` if you wish to transfer ownership to Warp)
-            deleter (Callable): Function to be called when deallocating the array, taking two arguments, pointer and size
-            requires_grad (bool): Whether or not gradients will be tracked for this array, see :class:`warp.Tape` for details
-            grad (array): The gradient array to use
-            pinned (bool): Whether to allocate pinned host memory, which allows asynchronous host-device transfers (only applicable with device="cpu")
+            copy: Whether the incoming ``data`` will be copied or aliased. Aliasing requires that
+                the incoming ``data`` already lives on the ``device`` specified and the data types match.
+            owner: Whether the array will try to deallocate the underlying memory when it is deleted
+                (deprecated, pass ``deleter`` if you wish to transfer ownership to Warp)
+            deleter: Function to be called when the array is deleted, taking two arguments: pointer and size
+            requires_grad: Whether or not gradients will be tracked for this array, see :class:`warp.Tape` for details
+            grad: The array in which to accumulate gradients in the backward pass. If ``None`` and ``requires_grad`` is ``True``,
+                then a gradient array will be allocated automatically.
+            pinned: Whether to allocate pinned host memory, which allows asynchronous host–device transfers
+                (only applicable with ``device="cpu"``)
 
         """
 
@@ -2963,6 +2985,116 @@ def array_type_id(a):
         raise ValueError("Invalid array type")
 
 
+# tile expression objects
+class Tile:
+    alignment = 16
+
+    def __init__(self, dtype, M, N, op=None, storage="register", layout="rowmajor", strides=None, owner=True):
+        self.dtype = type_to_warp(dtype)
+        self.M = M
+        self.N = N
+        self.op = op
+        self.storage = storage
+        self.layout = layout
+
+        if strides is None:
+            if layout == "rowmajor":
+                self.strides = (N, 1)
+            elif layout == "colmajor":
+                self.strides = (1, M)
+        else:
+            self.strides = strides
+
+        self.owner = owner
+
+    # generates C-type string
+    def ctype(self):
+        from warp.codegen import Var
+
+        if self.storage == "register":
+            return f"wp::tile_register_t<{Var.type_to_ctype(self.dtype)},{self.M},{self.N}>"
+        elif self.storage == "shared":
+            return f"wp::tile_shared_t<{Var.type_to_ctype(self.dtype)},{self.M},{self.N},{self.strides[0]}, {self.strides[1]}, {'true' if self.owner else 'false'}>"
+        else:
+            raise RuntimeError(f"Unrecognized tile storage type {self.storage}")
+
+    # generates C-initializer string
+    def cinit(self, requires_grad=False):
+        from warp.codegen import Var
+
+        if self.storage == "register":
+            return self.ctype() + "(0.0)"
+        elif self.storage == "shared":
+            if self.owner:
+                # allocate new shared memory tile
+                return f"wp::tile_alloc_empty<{Var.type_to_ctype(self.dtype)},{self.M},{self.N},{'true' if requires_grad else 'false'}>()"
+            else:
+                # tile will be initialized by another call, e.g.: tile_transpose()
+                return "NULL"
+
+    # return total tile size in bytes
+    def size_in_bytes(self):
+        num_bytes = self.align(type_size_in_bytes(self.dtype) * self.M * self.N)
+        return num_bytes
+
+    # align tile size to natural boundary, default 16-bytes
+    def align(self, bytes):
+        return ((bytes + self.alignment - 1) // self.alignment) * self.alignment
+
+
+class TileZeros(Tile):
+    def __init__(self, dtype, M, N, storage="register"):
+        Tile.__init__(self, dtype, M, N, op="zeros", storage=storage)
+
+
+class TileRange(Tile):
+    def __init__(self, dtype, start, stop, step, storage="register"):
+        self.start = start
+        self.stop = stop
+        self.step = step
+
+        M = 1
+        N = int((stop - start) / step)
+
+        Tile.__init__(self, dtype, M, N, op="arange", storage=storage)
+
+
+class TileConstant(Tile):
+    def __init__(self, dtype, M, N):
+        Tile.__init__(self, dtype, M, N, op="constant", storage="register")
+
+
+class TileLoad(Tile):
+    def __init__(self, array, M, N, storage="register"):
+        Tile.__init__(self, array.dtype, M, N, op="load", storage=storage)
+
+
+class TileUnaryMap(Tile):
+    def __init__(self, t, storage="register"):
+        Tile.__init__(self, t.dtype, t.M, t.N, op="unary_map", storage=storage)
+
+        self.t = t
+
+
+class TileBinaryMap(Tile):
+    def __init__(self, a, b, storage="register"):
+        Tile.__init__(self, a.dtype, a.M, a.N, op="binary_map", storage=storage)
+
+        self.a = a
+        self.b = b
+
+
+class TileShared(Tile):
+    def __init__(self, t):
+        Tile.__init__(self, t.dtype, t.M, t.N, "shared", storage="shared")
+
+        self.t = t
+
+
+def is_tile(t):
+    return isinstance(t, Tile)
+
+
 class Bvh:
     def __new__(cls, *args, **kwargs):
         instance = super(Bvh, cls).__new__(cls)
@@ -3544,9 +3676,9 @@ class Volume:
             grid_data = bytearray()
             while grid_data_offset < file_end:
                 chunk_size = struct.unpack("<Q", data[grid_data_offset : grid_data_offset + 8])[0]
-                grid_data += zlib.decompress(data[grid_data_offset + 8 :])
-                grid_data_offset += 8 + chunk_size
-
+                grid_data_offset += 8
+                grid_data += zlib.decompress(data[grid_data_offset : grid_data_offset + chunk_size])
+                grid_data_offset += chunk_size
         elif codec == 2:  # blosc compression
             try:
                 import blosc
@@ -3558,8 +3690,9 @@ class Volume:
             grid_data = bytearray()
             while grid_data_offset < file_end:
                 chunk_size = struct.unpack("<Q", data[grid_data_offset : grid_data_offset + 8])[0]
-                grid_data += blosc.decompress(data[grid_data_offset + 8 :])
-                grid_data_offset += 8 + chunk_size
+                grid_data_offset += 8
+                grid_data += blosc.decompress(data[grid_data_offset : grid_data_offset + chunk_size])
+                grid_data_offset += chunk_size
         else:
             raise RuntimeError(f"Unsupported codec code: {codec}")
 
@@ -3570,6 +3703,139 @@ class Volume:
         data_array = array(np.frombuffer(grid_data, dtype=np.byte), device=device)
         return cls(data_array)
 
+    def save_to_nvdb(self, path, codec: Literal["none", "zip", "blosc"] = "none"):
+        """Serialize the Volume into a NanoVDB (.nvdb) file.
+
+        Args:
+            path: File path to save.
+            codec: Compression codec used
+                "none" - no compression
+                "zip" - ZIP compression
+                "blosc" - BLOSC compression, requires the blosc module to be installed
+        """
+
+        codec_dict = {"none": 0, "zip": 1, "blosc": 2}
+
+        class FileHeader(ctypes.Structure):
+            _fields_ = [
+                ("magic", ctypes.c_uint64),
+                ("version", ctypes.c_uint32),
+                ("gridCount", ctypes.c_uint16),
+                ("codec", ctypes.c_uint16),
+            ]
+
+        class FileMetaData(ctypes.Structure):
+            _fields_ = [
+                ("gridSize", ctypes.c_uint64),
+                ("fileSize", ctypes.c_uint64),
+                ("nameKey", ctypes.c_uint64),
+                ("voxelCount", ctypes.c_uint64),
+                ("gridType", ctypes.c_uint32),
+                ("gridClass", ctypes.c_uint32),
+                ("worldBBox", ctypes.c_double * 6),
+                ("indexBBox", ctypes.c_uint32 * 6),
+                ("voxelSize", ctypes.c_double * 3),
+                ("nameSize", ctypes.c_uint32),
+                ("nodeCount", ctypes.c_uint32 * 4),
+                ("tileCount", ctypes.c_uint32 * 3),
+                ("codec", ctypes.c_uint16),
+                ("padding", ctypes.c_uint16),
+                ("version", ctypes.c_uint32),
+            ]
+
+        class GridData(ctypes.Structure):
+            _fields_ = [
+                ("magic", ctypes.c_uint64),
+                ("checksum", ctypes.c_uint64),
+                ("version", ctypes.c_uint32),
+                ("flags", ctypes.c_uint32),
+                ("gridIndex", ctypes.c_uint32),
+                ("gridCount", ctypes.c_uint32),
+                ("gridSize", ctypes.c_uint64),
+                ("gridName", ctypes.c_char * 256),
+                ("map", ctypes.c_byte * 264),
+                ("worldBBox", ctypes.c_double * 6),
+                ("voxelSize", ctypes.c_double * 3),
+                ("gridClass", ctypes.c_uint32),
+                ("gridType", ctypes.c_uint32),
+                ("blindMetadataOffset", ctypes.c_int64),
+                ("blindMetadataCount", ctypes.c_uint32),
+                ("data0", ctypes.c_uint32),
+                ("data1", ctypes.c_uint64),
+                ("data2", ctypes.c_uint64),
+            ]
+
+        NVDB_MAGIC = 0x304244566F6E614E
+        NVDB_VERSION = 32 << 21 | 3 << 10 | 3
+
+        try:
+            codec_int = codec_dict[codec]
+        except KeyError as err:
+            raise RuntimeError(f"Unsupported codec requested: {codec}") from err
+
+        if codec_int == 2:
+            try:
+                import blosc
+            except ImportError as err:
+                raise RuntimeError(
+                    f"blosc compression was requested, but Python module could not be imported: {err}"
+                ) from err
+
+        data = self.array().numpy()
+        grid_data = GridData.from_buffer(data)
+
+        if grid_data.gridIndex > 0:
+            raise RuntimeError(
+                "Saving of aliased Volumes is not supported. Use `save_to_nvdb` on the original volume, before any `load_next_grid` calls."
+            )
+
+        file_header = FileHeader(NVDB_MAGIC, NVDB_VERSION, grid_data.gridCount, codec_int)
+
+        grid_data_offset = 0
+        all_file_meta_data = []
+        for i in range(file_header.gridCount):
+            if i > 0:
+                grid_data = GridData.from_buffer(data[grid_data_offset : grid_data_offset + 672])
+            current_grid_data = data[grid_data_offset : grid_data_offset + grid_data.gridSize]
+            if codec_int == 1:  # zip compression
+                compressed_data = zlib.compress(current_grid_data)
+                compressed_size = len(compressed_data)
+            elif codec_int == 2:  # blosc compression
+                compressed_data = blosc.compress(current_grid_data)
+                compressed_size = len(compressed_data)
+            else:  # no compression
+                compressed_data = current_grid_data
+                compressed_size = grid_data.gridSize
+
+            file_meta_data = FileMetaData()
+            file_meta_data.gridSize = grid_data.gridSize
+            file_meta_data.fileSize = compressed_size
+            file_meta_data.gridType = grid_data.gridType
+            file_meta_data.gridClass = grid_data.gridClass
+            file_meta_data.worldBBox = grid_data.worldBBox
+            file_meta_data.voxelSize = grid_data.voxelSize
+            file_meta_data.nameSize = len(grid_data.gridName) + 1  # including the closing 0x0
+            file_meta_data.codec = codec_int
+            file_meta_data.version = NVDB_VERSION
+
+            grid_data_offset += file_meta_data.gridSize
+
+            all_file_meta_data.append((file_meta_data, grid_data.gridName, compressed_data))
+
+        with open(path, "wb") as nvdb:
+            nvdb.write(file_header)
+            for file_meta_data, grid_name, _ in all_file_meta_data:
+                nvdb.write(file_meta_data)
+                nvdb.write(grid_name + b"\x00")
+
+            for file_meta_data, _, compressed_data in all_file_meta_data:
+                if codec_int > 0:
+                    chunk_size = struct.pack("<Q", file_meta_data.fileSize)
+                    nvdb.write(chunk_size)
+                nvdb.write(compressed_data)
+
+        return path
+
     @classmethod
     def load_from_address(cls, grid_ptr: int, buffer_size: int = 0, device=None) -> Volume:
         """

warp/utils.py

@@ -18,6 +18,7 @@ import numpy as np
 import warp as wp
 import warp.context
 import warp.types
+from warp.context import Devicelike
 
 warnings_seen = set()
 
@@ -554,7 +555,27 @@ def mem_report():  # pragma: no cover
 
 
 class ScopedDevice:
-    def __init__(self, device):
+    """A context manager to temporarily change the current default device.
+
+    For CUDA devices, this context manager makes the device's CUDA context
+    current and restores the previous CUDA context on exit. This is handy when
+    running Warp scripts as part of a bigger pipeline because it avoids any side
+    effects of changing the CUDA context in the enclosed code.
+
+    Attributes:
+        device (Device): The device that will temporarily become the default
+          device within the context.
+        saved_device (Device): The previous default device. This is restored as
+          the default device on exiting the context.
+    """
+
+    def __init__(self, device: Devicelike):
+        """Initializes the context manager with a device.
+
+        Args:
+            device: The device that will temporarily become the default device
+              within the context.
+        """
         self.device = wp.get_device(device)
 
     def __enter__(self):