warp-lang 1.4.2__py3-none-manylinux2014_aarch64.whl → 1.5.1__py3-none-manylinux2014_aarch64.whl

warp/examples/tile/example_tile_mlp.py

@@ -0,0 +1,375 @@
+# Copyright (c) 2024 NVIDIA CORPORATION.  All rights reserved.
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+###########################################################################
+# Example Image Multilayer Perceptron (MLP)
+#
+# Shows how to train a coordinate-based MLP on an image to predict the RGB
+# color at a given input position. By default, a positional encoding is
+# applied to the input coordinates to improve the ability of the MLP to
+# represent higher-frequency content. This can be disabled by passing the
+# '--no_encoding' option.
+#
+# References:
+#   Ben Mildenhall et al. 2021. NeRF: representing scenes
+#   as neural radiance fields for view synthesis. Commun. ACM 65, 1
+#   (January 2022), 99–106. https://doi.org/10.1145/3503250
+#
+###########################################################################
+
+import math
+import os
+
+import numpy as np
+from PIL import Image
+
+import warp as wp
+import warp.examples
+import warp.optim
+
+rng = np.random.default_rng(45)
+
+
+def create_layer(dim_in, dim_hid, dtype=float):
+    w = rng.uniform(-1.0 / np.sqrt(dim_in), 1.0 / np.sqrt(dim_in), (dim_hid, dim_in))
+    b = rng.uniform(-1.0 / np.sqrt(dim_in), 1.0 / np.sqrt(dim_in), (dim_hid, 1))
+
+    weights = wp.array(w, dtype=dtype, requires_grad=True)
+    bias = wp.array(b, dtype=dtype, requires_grad=True)
+
+    return (weights, bias)
+
+
+def create_array(dim_in, dim_hid, dtype=float):
+    s = rng.uniform(-1.0 / np.sqrt(dim_in), 1.0 / np.sqrt(dim_in), (dim_hid, dim_in))
+    a = wp.array(s, dtype=dtype, requires_grad=True)
+
+    return a
+
+
+# number of frequencies for the positional encoding
+NUM_FREQ = wp.constant(8)
+
+DIM_IN = wp.constant(4 * NUM_FREQ)  # sin,cos for both x,y at each frequenecy
+DIM_HID = 32
+DIM_OUT = 3
+
+# threads per-block
+NUM_THREADS = 32
+
+IMG_WIDTH = 512
+IMG_HEIGHT = 512
+
+BATCH_SIZE = min(1024, int((IMG_WIDTH * IMG_HEIGHT) / 8))
+
+# dtype for our weights and bias matrices
+dtype = wp.float16
+
+
+@wp.func
+def relu(x: dtype):
+    return wp.max(x, dtype(0.0))
+
+
+@wp.kernel
+def compute(
+    indices: wp.array(dtype=int),
+    weights_0: wp.array2d(dtype=dtype),
+    bias_0: wp.array2d(dtype=dtype),
+    weights_1: wp.array2d(dtype=dtype),
+    bias_1: wp.array2d(dtype=dtype),
+    weights_2: wp.array2d(dtype=dtype),
+    bias_2: wp.array2d(dtype=dtype),
+    weights_3: wp.array2d(dtype=dtype),
+    bias_3: wp.array2d(dtype=dtype),
+    reference: wp.array2d(dtype=float),
+    loss: wp.array1d(dtype=float),
+    out: wp.array2d(dtype=float),
+):
+    # batch indices
+    linear = indices[wp.tid()]
+
+    row = linear / IMG_WIDTH
+    col = linear % IMG_WIDTH
+
+    # normalize input coordinates to [-1, 1]
+    x = (float(row) / float(IMG_WIDTH) - 0.5) * 2.0
+    y = (float(col) / float(IMG_HEIGHT) - 0.5) * 2.0
+
+    local = wp.vector(dtype=dtype, length=DIM_IN)
+
+    # construct positional encoding
+    for s in range(NUM_FREQ):
+        scale = wp.pow(2.0, float(s)) * wp.pi
+
+        # x-coord
+        local[s * 4 + 0] = dtype(wp.sin(x * scale))
+        local[s * 4 + 1] = dtype(wp.cos(x * scale))
+        # y-coord
+        local[s * 4 + 2] = dtype(wp.sin(y * scale))
+        local[s * 4 + 3] = dtype(wp.cos(y * scale))
+
+    # tile feature vectors across the block, returns [dim(f), NUM_THREADS]
+    f = wp.tile(local)
+
+    # input layer
+    w0 = wp.tile_load(weights_0, 0, 0, m=DIM_HID, n=DIM_IN)
+    b0 = wp.tile_load(bias_0, 0, 0, m=DIM_HID, n=1)
+    z = wp.tile_map(relu, wp.tile_matmul(w0, f) + wp.tile_broadcast(b0, m=DIM_HID, n=NUM_THREADS))
+
+    # hidden layer
+    w1 = wp.tile_load(weights_1, 0, 0, m=DIM_HID, n=DIM_HID)
+    b1 = wp.tile_load(bias_1, 0, 0, m=DIM_HID, n=1)
+    z = wp.tile_map(relu, wp.tile_matmul(w1, z) + wp.tile_broadcast(b1, m=DIM_HID, n=NUM_THREADS))
+
+    w2 = wp.tile_load(weights_2, 0, 0, m=DIM_HID, n=DIM_HID)
+    b2 = wp.tile_load(bias_2, 0, 0, m=DIM_HID, n=1)
+    z = wp.tile_map(relu, wp.tile_matmul(w2, z) + wp.tile_broadcast(b2, m=DIM_HID, n=NUM_THREADS))
+
+    # output layer
+    w3 = wp.tile_load(weights_3, 0, 0, m=DIM_OUT, n=DIM_HID)
+    b3 = wp.tile_load(bias_3, 0, 0, m=DIM_OUT, n=1)
+    o = wp.tile_map(relu, wp.tile_matmul(w3, z) + wp.tile_broadcast(b3, m=DIM_OUT, n=NUM_THREADS))
+
+    # untile back to SIMT
+    output = wp.untile(o)
+
+    # compute error
+    error = wp.vec3(
+        float(output[0]) - reference[0, linear],
+        float(output[1]) - reference[1, linear],
+        float(output[2]) - reference[2, linear],
+    )
+
+    # write MSE loss
+    if loss:
+        wp.atomic_add(loss, 0, wp.length_sq(error) / float(3 * BATCH_SIZE))
+
+    #  write image output
+    if out:
+        for i in range(DIM_OUT):
+            out[i, linear] = float(output[i])
+
+
+class Example:
+    def __init__(self, train_iters):
+        self.weights_0, self.bias_0 = create_layer(DIM_IN, DIM_HID, dtype=dtype)
+        self.weights_1, self.bias_1 = create_layer(DIM_HID, DIM_HID, dtype=dtype)
+        self.weights_2, self.bias_2 = create_layer(DIM_HID, DIM_HID, dtype=dtype)
+        self.weights_3, self.bias_3 = create_layer(DIM_HID, DIM_OUT, dtype=dtype)
+
+        # reference
+        reference_path = os.path.join(wp.examples.get_asset_directory(), "pixel.jpg")
+        with Image.open(reference_path) as im:
+            reference_image = np.asarray(im.resize((IMG_WIDTH, IMG_HEIGHT)).convert("RGB")) / 255.0
+        self.reference = wp.array(reference_image.reshape(IMG_WIDTH * IMG_HEIGHT, 3).T, dtype=float)
+
+        # create randomized batch indices
+        indices = np.arange(0, IMG_WIDTH * IMG_HEIGHT, dtype=np.int32)
+        rng.shuffle(indices)
+        self.indices = wp.array(indices)
+
+        self.num_batches = int((IMG_WIDTH * IMG_HEIGHT) / BATCH_SIZE)
+        self.max_iters = train_iters
+        self.max_epochs = max(1, int(self.max_iters / self.num_batches))
+
+    def train_warp(self):
+        params = [
+            self.weights_0,
+            self.bias_0,
+            self.weights_1,
+            self.bias_1,
+            self.weights_2,
+            self.bias_2,
+            self.weights_3,
+            self.bias_3,
+        ]
+
+        optimizer_grads = [p.grad.flatten() for p in params]
+        optimizer_inputs = [p.flatten() for p in params]
+        optimizer = warp.optim.Adam(optimizer_inputs, lr=0.01)
+
+        loss = wp.zeros(1, dtype=float, requires_grad=True)
+        output = create_array(IMG_WIDTH * IMG_HEIGHT, DIM_OUT)
+
+        # capture graph for whole epoch
+        wp.capture_begin()
+
+        for b in range(0, IMG_WIDTH * IMG_HEIGHT, BATCH_SIZE):
+            loss.zero_()
+
+            with wp.Tape() as tape:
+                wp.launch(
+                    compute,
+                    dim=[BATCH_SIZE],
+                    inputs=[
+                        self.indices[b : b + BATCH_SIZE],
+                        self.weights_0,
+                        self.bias_0,
+                        self.weights_1,
+                        self.bias_1,
+                        self.weights_2,
+                        self.bias_2,
+                        self.weights_3,
+                        self.bias_3,
+                        self.reference,
+                        loss,
+                        None,
+                    ],
+                    block_dim=NUM_THREADS,
+                )
+
+            tape.backward(loss)
+            optimizer.step(optimizer_grads)
+            tape.zero()
+
+        graph = wp.capture_end()
+
+        with wp.ScopedTimer("Training"):
+            for i in range(self.max_epochs):
+                with wp.ScopedTimer("Epoch"):
+                    wp.capture_launch(graph)
+                    print(f"Epoch: {i} Loss: {loss.numpy()}")
+
+        # evaluate full image
+        wp.launch(
+            compute,
+            dim=[IMG_WIDTH * IMG_HEIGHT],
+            inputs=[
+                self.indices,
+                self.weights_0,
+                self.bias_0,
+                self.weights_1,
+                self.bias_1,
+                self.weights_2,
+                self.bias_2,
+                self.weights_3,
+                self.bias_3,
+                self.reference,
+                loss,
+                output,
+            ],
+            block_dim=NUM_THREADS,
+        )
+
+        self.save_image("example_tile_mlp.jpg", output.numpy())
+
+    def train_torch(self):
+        import torch as tc
+
+        weights_0 = tc.nn.Parameter(wp.to_torch(self.weights_0))
+        weights_1 = tc.nn.Parameter(wp.to_torch(self.weights_1))
+        weights_2 = tc.nn.Parameter(wp.to_torch(self.weights_2))
+        weights_3 = tc.nn.Parameter(wp.to_torch(self.weights_3))
+
+        bias_0 = tc.nn.Parameter(wp.to_torch(self.bias_0))
+        bias_1 = tc.nn.Parameter(wp.to_torch(self.bias_1))
+        bias_2 = tc.nn.Parameter(wp.to_torch(self.bias_2))
+        bias_3 = tc.nn.Parameter(wp.to_torch(self.bias_3))
+
+        indices = wp.to_torch(self.indices)
+        reference = wp.to_torch(self.reference)
+
+        optimizer = tc.optim.Adam(
+            [weights_0, bias_0, weights_1, bias_1, weights_2, bias_2, weights_3, bias_3],
+            capturable=True,
+            lr=0.0001,
+            betas=(0.9, 0.95),
+            eps=1.0e-6,
+        )
+
+        # generate frequency space encoding of pixels
+        # based on their linear index in the image
+        def encode(linear):
+            row = (linear // IMG_WIDTH).float()
+            col = (linear % IMG_WIDTH).float()
+
+            x = (row / float(IMG_WIDTH) - 0.5) * 2.0
+            y = (col / float(IMG_HEIGHT) - 0.5) * 2.0
+
+            encoding = tc.zeros((NUM_FREQ * 4, len(linear)), dtype=tc.float16, device="cuda")
+
+            for s in range(NUM_FREQ):
+                scale = math.pow(2.0, float(s)) * math.pi
+
+                # Directly write the computed values into the encoding tensor
+                encoding[s * 4 + 0, :] = tc.sin(scale * x)
+                encoding[s * 4 + 1, :] = tc.cos(scale * x)
+                encoding[s * 4 + 2, :] = tc.sin(scale * y)
+                encoding[s * 4 + 3, :] = tc.cos(scale * y)
+
+            return encoding
+
+        stream = tc.cuda.Stream()
+        graph = tc.cuda.CUDAGraph()
+
+        # warm-up
+        with tc.cuda.stream(stream):
+            f = tc.rand((NUM_FREQ * 4, BATCH_SIZE), dtype=tc.float16, device="cuda")
+            z = tc.relu(weights_0 @ f + bias_0)
+            z = tc.relu(weights_1 @ z + bias_1)
+            z = tc.relu(weights_2 @ z + bias_2)
+            z = tc.relu(weights_3 @ z + bias_3)
+            ref = tc.rand((3, BATCH_SIZE), dtype=tc.float16, device="cuda")
+            loss = tc.mean((z - ref) ** 2)
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+
+        with tc.cuda.graph(graph):
+            for b in range(0, IMG_WIDTH * IMG_HEIGHT, BATCH_SIZE):
+                linear = indices[b : b + BATCH_SIZE]
+
+                f = encode(linear)
+
+                z = tc.relu(weights_0 @ f + bias_0)
+                z = tc.relu(weights_1 @ z + bias_1)
+                z = tc.relu(weights_2 @ z + bias_2)
+                z = tc.relu(weights_3 @ z + bias_3)
+
+                ref = reference[:, linear]
+                loss = tc.mean((z - ref) ** 2)
+
+                optimizer.zero_grad()
+                loss.backward()
+                optimizer.step()
+
+        with wp.ScopedTimer("Training (Torch)"):
+            for _i in range(self.max_epochs):
+                with wp.ScopedTimer("Epoch"):
+                    graph.replay()
+
+                    print(loss)
+
+        f = encode(tc.arange(0, IMG_WIDTH * IMG_HEIGHT))
+        z = tc.relu(weights_0 @ f + bias_0)
+        z = tc.relu(weights_1 @ z + bias_1)
+        z = tc.relu(weights_2 @ z + bias_2)
+        z = tc.relu(weights_3 @ z + bias_3)
+
+        self.save_image("example_tile_mlp_torch.jpg", z.detach().cpu().numpy())
+
+    def save_image(self, name, output):
+        predicted_image = output.T.reshape(IMG_WIDTH, IMG_HEIGHT, 3)
+        predicted_image = (predicted_image * 255).astype(np.uint8)
+
+        predicted_image_pil = Image.fromarray(predicted_image)
+        predicted_image_pil.save(name)
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument("--train_iters", type=int, default=20000, help="Total number of training iterations.")
+
+    args = parser.parse_known_args()[0]
+
+    with wp.ScopedDevice("cuda:0"):
+        example = Example(args.train_iters)
+        example.train_warp()
+        # example.train_torch()

warp/fem/__init__.py

@@ -24,14 +24,17 @@ from .geometry import (
     LinearGeometryPartition,
     Nanogrid,
     Quadmesh2D,
+    Quadmesh3D,
     Tetmesh,
     Trimesh2D,
+    Trimesh3D,
 )
 from .integrate import integrate, interpolate
 from .operator import (
     D,
     at_node,
     average,
+    cells,
     curl,
     deformation_gradient,
     degree,
@@ -50,6 +53,9 @@ from .operator import (
     normal,
     outer,
     position,
+    to_cell_side,
+    to_inner_cell,
+    to_outer_cell,
 )
 from .polynomial import Polynomial
 from .quadrature import ExplicitQuadrature, NodalQuadrature, PicQuadrature, Quadrature, RegularQuadrature
@@ -65,6 +71,8 @@ from .space import (
     SpaceTopology,
     SymmetricTensorMapper,
     make_collocated_function_space,
+    make_contravariant_function_space,
+    make_covariant_function_space,
     make_polynomial_basis_space,
     make_polynomial_space,
     make_space_partition,

warp/fem/cache.py

@@ -6,6 +6,7 @@ from copy import copy
 from typing import Any, Callable, Dict, Optional, Tuple, Union
 
 import warp as wp
+from warp.fem.operator import Integrand
 
 _kernel_cache = {}
 _struct_cache = {}
@@ -186,7 +187,7 @@ class ExpandStarredArgumentStruct(ast.NodeTransformer):
 
 
 def get_integrand_function(
-    integrand: "warp.fem.operator.Integrand",  # noqa: F821
+    integrand: Integrand,
     suffix: str,
     func=None,
     annotations=None,
@@ -208,27 +209,30 @@ def get_integrand_function(
 
 
 def get_integrand_kernel(
-    integrand: "warp.fem.operator.Integrand",  # noqa: F821
+    integrand: Integrand,
     suffix: str,
     kernel_fn: Optional[Callable] = None,
     kernel_options: Dict[str, Any] = None,
     code_transformers=None,
 ):
-    if kernel_options is None:
-        kernel_options = {}
+    options = integrand.module.options.copy()
+    options.update(integrand.kernel_options)
+    if kernel_options is not None:
+        options.update(kernel_options)
 
-    key = _make_key(integrand.func, suffix, use_qualified_name=True)
+    kernel_key = _make_key(integrand.func, suffix, use_qualified_name=True)
+    opts_key = "".join([f"{k}:{v}" for k, v in sorted(options.items())])
+    cache_key = kernel_key + opts_key
 
-    if key not in _kernel_cache:
+    if cache_key not in _kernel_cache:
         if kernel_fn is None:
             return None
 
         module = wp.get_module(f"{integrand.module.name}.{integrand.name}")
-        module.options = copy(integrand.module.options)
-        module.options.update(kernel_options)
-
-        _kernel_cache[key] = wp.Kernel(func=kernel_fn, key=key, module=module, code_transformers=code_transformers)
-    return _kernel_cache[key]
+        _kernel_cache[cache_key] = wp.Kernel(
+            func=kernel_fn, key=kernel_key, module=module, code_transformers=code_transformers, options=options
+        )
+    return _kernel_cache[cache_key]
 
 
 def cached_arg_value(func: Callable):
@@ -478,7 +482,7 @@ def borrow_temporary(
     if temporary_store is None:
         temporary_store = TemporaryStore._default_store
 
-    if temporary_store is None:
+    if temporary_store is None or (requires_grad and wp.context.runtime.tape is not None):
         return Temporary(
             array=wp.empty(shape=shape, dtype=dtype, pinned=pinned, device=device, requires_grad=requires_grad)
         )

warp/fem/dirichlet.py

@@ -1,7 +1,7 @@
 from typing import Any, Optional
 
 import warp as wp
-from warp.fem.utils import array_axpy, symmetric_eigenvalues_qr
+from warp.fem.linalg import array_axpy, symmetric_eigenvalues_qr
 from warp.sparse import BsrMatrix, bsr_assign, bsr_axpy, bsr_copy, bsr_mm, bsr_mv
 from warp.types import type_is_matrix, type_length
 

warp/fem/domain.py

@@ -1,4 +1,4 @@
-from typing import Optional, Union
+from typing import Any, Optional, Set, Union
 
 import warp as wp
 import warp.codegen
@@ -11,6 +11,7 @@ from warp.fem.geometry import (
     GeometryPartition,
     WholeGeometryPartition,
 )
+from warp.fem.operator import Operator
 from warp.fem.types import ElementKind
 
 GeometryOrPartition = Union[Geometry, GeometryPartition]
@@ -94,6 +95,10 @@ class GeometryDomain:
     element_lookup: wp.Function
     """Device function returning the sample point corresponding to a world position"""
 
+    def notify_operator_usage(self, ops: Set[Operator]):
+        """Makes the Domain aware that the operators `ops` will be applied"""
+        pass
+
 
 class Cells(GeometryDomain):
     """A Domain containing all cells of the geometry or geometry partition"""
@@ -160,6 +165,17 @@ class Cells(GeometryDomain):
     def element_lookup(self) -> wp.Function:
         return self.geometry.cell_lookup
 
+    @property
+    def domain_cell_arg(self) -> wp.Function:
+        return Cells._identity_fn
+
+    def cell_domain(self):
+        return self
+
+    @wp.func
+    def _identity_fn(x: Any):
+        return x
+
 
 class Sides(GeometryDomain):
     """A Domain containing all (interior and boundary) sides of the geometry or geometry partition"""
@@ -225,6 +241,33 @@ class Sides(GeometryDomain):
     def element_normal(self) -> wp.Function:
         return self.geometry.side_normal
 
+    @property
+    def element_inner_cell_index(self) -> wp.Function:
+        return self.geometry.side_inner_cell_index
+
+    @property
+    def element_outer_cell_index(self) -> wp.Function:
+        return self.geometry.side_outer_cell_index
+
+    @property
+    def element_inner_cell_coords(self) -> wp.Function:
+        return self.geometry.side_inner_cell_coords
+
+    @property
+    def element_outer_cell_coords(self) -> wp.Function:
+        return self.geometry.side_outer_cell_coords
+
+    @property
+    def cell_to_element_coords(self) -> wp.Function:
+        return self.geometry.side_from_cell_coords
+
+    @property
+    def domain_cell_arg(self) -> wp.Function:
+        return self.geometry.side_to_cell_arg
+
+    def cell_domain(self):
+        return Cells(self.geometry_partition)
+
 
 class BoundarySides(Sides):
     """A Domain containing boundary sides of the geometry or geometry partition"""

warp/fem/field/__init__.py

@@ -6,8 +6,7 @@ from warp.fem.space import FunctionSpace, SpacePartition, SpaceRestriction, make
 from .field import DiscreteField, FieldLike, GeometryField, ImplicitField, NonconformingField, SpaceField, UniformField
 from .nodal_field import NodalField
 from .restriction import FieldRestriction
-from .test import TestField
-from .trial import TrialField
+from .virtual import LocalTestField, LocalTrialField, TestField, TrialField
 
 
 def make_restriction(

warp/fem/field/field.py

@@ -1,10 +1,10 @@
-from typing import Any, Dict, Optional
+from typing import Any, Dict, Optional, Set
 
 import warp as wp
 from warp.fem import cache
 from warp.fem.domain import GeometryDomain, Sides
 from warp.fem.geometry import DeformedGeometry, Geometry
-from warp.fem.operator import integrand
+from warp.fem.operator import Operator, integrand
 from warp.fem.space import FunctionSpace, SpacePartition
 from warp.fem.types import NULL_ELEMENT_INDEX, ElementKind, Sample
 
@@ -48,32 +48,32 @@ class FieldLike:
         return False
 
     @staticmethod
-    def eval_inner(args: "ElementEvalArg", s: "Sample"):  # noqa: F821
+    def eval_inner(args: "ElementEvalArg", s: Sample):  # noqa: F821
         """Device function evaluating the inner field value at a sample point"""
         raise NotImplementedError
 
     @staticmethod
-    def eval_grad_inner(args: "ElementEvalArg", s: "Sample"):  # noqa: F821
+    def eval_grad_inner(args: "ElementEvalArg", s: Sample):  # noqa: F821
         """Device function evaluating the inner field gradient at a sample point"""
         raise NotImplementedError
 
     @staticmethod
-    def eval_div_inner(args: "ElementEvalArg", s: "Sample"):  # noqa: F821
+    def eval_div_inner(args: "ElementEvalArg", s: Sample):  # noqa: F821
         """Device function evaluating the inner field divergence at a sample point"""
         raise NotImplementedError
 
     @staticmethod
-    def eval_outer(args: "ElementEvalArg", s: "Sample"):  # noqa: F821
+    def eval_outer(args: "ElementEvalArg", s: Sample):  # noqa: F821
         """Device function evaluating the outer field value at a sample point"""
         raise NotImplementedError
 
     @staticmethod
-    def eval_grad_outer(args: "ElementEvalArg", s: "Sample"):  # noqa: F821
+    def eval_grad_outer(args: "ElementEvalArg", s: Sample):  # noqa: F821
         """Device function evaluating the outer field gradient at a sample point"""
         raise NotImplementedError
 
     @staticmethod
-    def eval_div_outer(args: "ElementEvalArg", s: "Sample"):  # noqa: F821
+    def eval_div_outer(args: "ElementEvalArg", s: Sample):  # noqa: F821
         """Device function evaluating the outer field divergence at a sample point"""
         raise NotImplementedError
 
@@ -82,6 +82,10 @@ class FieldLike:
         """Polynomial degree of the field is applicable, or hint for determination of interpolation order"""
         raise NotImplementedError
 
+    def notify_operator_usage(self, ops: Set[Operator]):
+        """Makes the Domain aware that the operators `ops` will be applied"""
+        pass
+
 
 class GeometryField(FieldLike):
     """Base class for fields defined over a geometry"""
@@ -97,12 +101,12 @@ class GeometryField(FieldLike):
         raise NotImplementedError
 
     @staticmethod
-    def eval_reference_grad_inner(args: "ElementEvalArg", s: "Sample"):  # noqa: F821
+    def eval_reference_grad_inner(args: "ElementEvalArg", s: Sample):  # noqa: F821
         """Device function evaluating the inner field gradient with respect to reference element coordinates at a sample point"""
         raise NotImplementedError
 
     @staticmethod
-    def eval_reference_grad_outer(args: "ElementEvalArg", s: "Sample"):  # noqa: F821
+    def eval_reference_grad_outer(args: "ElementEvalArg", s: Sample):  # noqa: F821
         """Device function evaluating the outer field gradient with respect to reference element coordinates at a sample point"""
         raise NotImplementedError
 
@@ -128,6 +132,9 @@ class SpaceField(GeometryField):
         self._space = space
         self._space_partition = space_partition
 
+        self.gradient_valid = self.space.gradient_valid
+        self.divergence_valid = self.space.divergence_valid
+
     @property
     def geometry(self) -> Geometry:
         return self._space.geometry
@@ -156,17 +163,22 @@ class SpaceField(GeometryField):
     def dof_dtype(self) -> type:
         return self.space.dof_dtype
 
-    def gradient_valid(self) -> bool:
-        """Whether gradient operator can be computed. Only for scalar and vector fields as higher-order tensors are not support yet"""
-        return not wp.types.type_is_matrix(self.dtype)
+    @property
+    def gradient_dtype(self):
+        """Return type of the gradient operator. Assumes self.gradient_valid()"""
+        if wp.types.type_is_vector(self.dtype):
+            return cache.cached_mat_type(
+                shape=(wp.types.type_length(self.dtype), self.geometry.dimension),
+                dtype=wp.types.type_scalar_type(self.dtype),
+            )
+        return cache.cached_vec_type(length=self.geometry.dimension, dtype=wp.types.type_scalar_type(self.dtype))
 
-    def divergence_valid(self) -> bool:
-        """Whether divergence of this field can be computed. Only for vector and tensor fields with same dimension as embedding geometry"""
+    @property
+    def divergence_dtype(self):
+        """Return type of the divergence operator. Assumes self.gradient_valid()"""
         if wp.types.type_is_vector(self.dtype):
-            return wp.types.type_length(self.dtype) == self.space.geometry.dimension
-        if wp.types.type_is_matrix(self.dtype):
-            return self.dtype._shape_[0] == self.space.geometry.dimension
-        return False
+            return wp.types.type_scalar_type(self.dtype)
+        return cache.cached_vec_type(length=self.dtype._shape_[1], dtype=wp.types.type_scalar_type(self.dtype))
 
     def _make_eval_degree(self):
         ORDER = self.space.ORDER