warp-lang 1.7.0__py3-none-manylinux_2_28_x86_64.whl

warp/tests/geometry/test_volume.py

@@ -0,0 +1,1103 @@
+# SPDX-FileCopyrightText: Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import tempfile
+import unittest
+from typing import Any
+
+import numpy as np
+
+import warp as wp
+from warp.tests.unittest_utils import *
+
+
+# float volume tests
+@wp.kernel
+def test_volume_lookup_f(volume: wp.uint64, points: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+    expected = p[0] * p[1] * p[2]
+    if abs(p[0]) > 10.0 or abs(p[1]) > 10.0 or abs(p[2]) > 10.0:
+        expected = 10.0
+
+    i = int(p[0])
+    j = int(p[1])
+    k = int(p[2])
+
+    expect_eq(wp.volume_lookup_f(volume, i, j, k), expected)
+    expect_eq(wp.volume_lookup(volume, i, j, k, dtype=wp.float32), expected)
+
+
+@wp.kernel
+def test_volume_sample_closest_f(volume: wp.uint64, points: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+    i = round(p[0])
+    j = round(p[1])
+    k = round(p[2])
+    expected = i * j * k
+    if abs(i) > 10.0 or abs(j) > 10.0 or abs(k) > 10.0:
+        expected = 10.0
+
+    expect_eq(wp.volume_sample_f(volume, p, wp.Volume.CLOSEST), expected)
+    expect_eq(wp.volume_sample(volume, p, wp.Volume.CLOSEST, dtype=wp.float32), expected)
+
+    q = wp.volume_index_to_world(volume, p)
+    q_inv = wp.volume_world_to_index(volume, q)
+    expect_eq(p, q_inv)
+
+
+@wp.kernel
+def test_volume_sample_linear_f(volume: wp.uint64, points: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+
+    expected = p[0] * p[1] * p[2]
+    if abs(p[0]) > 10.0 or abs(p[1]) > 10.0 or abs(p[2]) > 10.0:
+        return  # not testing against background values
+
+    expect_near(wp.volume_sample_f(volume, p, wp.Volume.LINEAR), expected, 2.0e-4)
+    expect_near(wp.volume_sample(volume, p, wp.Volume.LINEAR, dtype=wp.float32), expected, 2.0e-4)
+
+
+@wp.kernel
+def test_volume_sample_grad_linear_f(volume: wp.uint64, points: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+
+    expected_val = p[0] * p[1] * p[2]
+    expected_gx = p[1] * p[2]
+    expected_gy = p[0] * p[2]
+    expected_gz = p[0] * p[1]
+
+    if abs(p[0]) > 10.0 or abs(p[1]) > 10.0 or abs(p[2]) > 10.0:
+        return  # not testing against background values
+
+    grad = wp.vec3(0.0, 0.0, 0.0)
+    val = wp.volume_sample_grad_f(volume, p, wp.Volume.LINEAR, grad)
+
+    expect_near(val, expected_val, 2.0e-4)
+    expect_near(grad[0], expected_gx, 2.0e-4)
+    expect_near(grad[1], expected_gy, 2.0e-4)
+    expect_near(grad[2], expected_gz, 2.0e-4)
+
+    val = wp.volume_sample_grad(volume, p, wp.Volume.LINEAR, grad, dtype=wp.float32)
+
+    expect_near(val, expected_val, 2.0e-4)
+    expect_near(grad[0], expected_gx, 2.0e-4)
+    expect_near(grad[1], expected_gy, 2.0e-4)
+    expect_near(grad[2], expected_gz, 2.0e-4)
+
+
+@wp.kernel
+def test_volume_sample_local_f_linear_values(
+    volume: wp.uint64, points: wp.array(dtype=wp.vec3), values: wp.array(dtype=wp.float32)
+):
+    tid = wp.tid()
+    p = points[tid]
+    values[tid] = wp.volume_sample_f(volume, p, wp.Volume.LINEAR)
+
+
+@wp.kernel
+def test_volume_sample_grad_local_f_linear_values(
+    volume: wp.uint64, points: wp.array(dtype=wp.vec3), values: wp.array(dtype=wp.float32), case_num: int
+):
+    tid = wp.tid()
+    p = points[tid]
+
+    grad = wp.vec3(0.0, 0.0, 0.0)
+    val = wp.volume_sample_grad_f(volume, p, wp.Volume.LINEAR, grad)
+    if case_num == 0:
+        values[tid] = val
+    elif case_num == 1:
+        values[tid] = grad[0]
+    elif case_num == 2:
+        values[tid] = grad[1]
+    elif case_num == 3:
+        values[tid] = grad[2]
+
+
+@wp.kernel
+def test_volume_sample_world_f_linear_values(
+    volume: wp.uint64, points: wp.array(dtype=wp.vec3), values: wp.array(dtype=wp.float32)
+):
+    tid = wp.tid()
+    q = points[tid]
+    p = wp.volume_world_to_index(volume, q)
+    values[tid] = wp.volume_sample_f(volume, p, wp.Volume.LINEAR)
+
+
+@wp.kernel
+def test_volume_sample_grad_world_f_linear_values(
+    volume: wp.uint64, points: wp.array(dtype=wp.vec3), values: wp.array(dtype=wp.float32), case_num: int
+):
+    tid = wp.tid()
+    q = points[tid]
+    p = wp.volume_world_to_index(volume, q)
+
+    grad = wp.vec3(0.0, 0.0, 0.0)
+    val = wp.volume_sample_grad_f(volume, p, wp.Volume.LINEAR, grad)
+    if case_num == 0:
+        values[tid] = val
+    elif case_num == 1:
+        values[tid] = grad[0]
+    elif case_num == 2:
+        values[tid] = grad[1]
+    elif case_num == 3:
+        values[tid] = grad[2]
+
+
+# vec3f volume tests
+@wp.kernel
+def test_volume_lookup_v(volume: wp.uint64, points: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+    expected = wp.vec3(
+        p[0] + 2.0 * p[1] + 3.0 * p[2], 4.0 * p[0] + 5.0 * p[1] + 6.0 * p[2], 7.0 * p[0] + 8.0 * p[1] + 9.0 * p[2]
+    )
+    if abs(p[0]) > 10.0 or abs(p[1]) > 10.0 or abs(p[2]) > 10.0:
+        expected = wp.vec3(10.8, -4.13, 10.26)
+
+    i = int(p[0])
+    j = int(p[1])
+    k = int(p[2])
+
+    expect_eq(wp.volume_lookup_v(volume, i, j, k), expected)
+    expect_eq(wp.volume_lookup(volume, i, j, k, dtype=wp.vec3), expected)
+
+
+@wp.kernel
+def test_volume_sample_closest_v(volume: wp.uint64, points: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+    i = round(p[0])
+    j = round(p[1])
+    k = round(p[2])
+    expected = wp.vec3(i + 2.0 * j + 3.0 * k, 4.0 * i + 5.0 * j + 6.0 * k, 7.0 * i + 8.0 * j + 9.0 * k)
+    if abs(i) > 10.0 or abs(j) > 10.0 or abs(k) > 10.0:
+        expected = wp.vec3(10.8, -4.13, 10.26)
+
+    expect_eq(wp.volume_sample_v(volume, p, wp.Volume.CLOSEST), expected)
+    expect_eq(wp.volume_sample(volume, p, wp.Volume.CLOSEST, dtype=wp.vec3), expected)
+
+    q = wp.volume_index_to_world(volume, p)
+    q_inv = wp.volume_world_to_index(volume, q)
+    expect_eq(p, q_inv)
+
+
+@wp.kernel
+def test_volume_sample_linear_v(volume: wp.uint64, points: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+
+    expected = wp.vec3(
+        p[0] + 2.0 * p[1] + 3.0 * p[2], 4.0 * p[0] + 5.0 * p[1] + 6.0 * p[2], 7.0 * p[0] + 8.0 * p[1] + 9.0 * p[2]
+    )
+    if abs(p[0]) > 10.0 or abs(p[1]) > 10.0 or abs(p[2]) > 10.0:
+        return  # not testing against background values
+
+    expect_near(wp.volume_sample_v(volume, p, wp.Volume.LINEAR), expected, 2.0e-4)
+    expect_near(wp.volume_sample(volume, p, wp.Volume.LINEAR, dtype=wp.vec3), expected, 2.0e-4)
+
+
+@wp.kernel
+def test_volume_sample_grad_linear_v(volume: wp.uint64, points: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+
+    if abs(p[0]) > 10.0 or abs(p[1]) > 10.0 or abs(p[2]) > 10.0:
+        return  # not testing against background values
+
+    expected_val = wp.vec3(
+        p[0] + 2.0 * p[1] + 3.0 * p[2], 4.0 * p[0] + 5.0 * p[1] + 6.0 * p[2], 7.0 * p[0] + 8.0 * p[1] + 9.0 * p[2]
+    )
+    expected_grad = wp.mat33(1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0)
+
+    grad = wp.mat33(0.0)
+    val = wp.volume_sample_grad(volume, p, wp.Volume.LINEAR, grad, dtype=wp.vec3)
+
+    expect_near(val, expected_val, 2.0e-4)
+    expect_near(grad[0], expected_grad[0], 2.0e-4)
+    expect_near(grad[1], expected_grad[1], 2.0e-4)
+    expect_near(grad[2], expected_grad[2], 2.0e-4)
+
+
+@wp.kernel
+def test_volume_sample_local_v_linear_values(
+    volume: wp.uint64, points: wp.array(dtype=wp.vec3), values: wp.array(dtype=wp.float32)
+):
+    tid = wp.tid()
+    p = points[tid]
+    ones = wp.vec3(1.0, 1.0, 1.0)
+    values[tid] = wp.dot(wp.volume_sample_v(volume, p, wp.Volume.LINEAR), ones)
+
+
+@wp.kernel
+def test_volume_sample_world_v_linear_values(
+    volume: wp.uint64, points: wp.array(dtype=wp.vec3), values: wp.array(dtype=wp.float32)
+):
+    tid = wp.tid()
+    q = points[tid]
+    p = wp.volume_world_to_index(volume, q)
+    ones = wp.vec3(1.0, 1.0, 1.0)
+    values[tid] = wp.dot(wp.volume_sample_v(volume, p, wp.Volume.LINEAR), ones)
+
+
+# int32 volume tests
+@wp.kernel
+def test_volume_lookup_i(volume: wp.uint64, points: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+    i = int(p[0])
+    j = int(p[1])
+    k = int(p[2])
+    expected = i * j * k
+    if abs(i) > 10 or abs(j) > 10 or abs(k) > 10:
+        expected = 10
+
+    expect_eq(wp.volume_lookup_i(volume, i, j, k), expected)
+    expect_eq(wp.volume_lookup(volume, i, j, k, dtype=wp.int32), expected)
+
+
+@wp.kernel
+def test_volume_sample_i(volume: wp.uint64, points: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+    i = round(p[0])
+    j = round(p[1])
+    k = round(p[2])
+    expected = int(i * j * k)
+    if abs(i) > 10.0 or abs(j) > 10.0 or abs(k) > 10.0:
+        expected = 10
+
+    expect_eq(wp.volume_sample_i(volume, p), expected)
+    expect_eq(wp.volume_sample(volume, p, wp.Volume.CLOSEST, dtype=wp.int32), expected)
+
+    q = wp.volume_index_to_world(volume, p)
+    q_inv = wp.volume_world_to_index(volume, q)
+    expect_eq(p, q_inv)
+
+
+# Index/world transformation tests
+@wp.kernel
+def test_volume_index_to_world(
+    volume: wp.uint64,
+    points: wp.array(dtype=wp.vec3),
+    values: wp.array(dtype=wp.float32),
+    grad_values: wp.array(dtype=wp.vec3),
+):
+    tid = wp.tid()
+    p = points[tid]
+    ones = wp.vec3(1.0, 1.0, 1.0)
+    values[tid] = wp.dot(wp.volume_index_to_world(volume, p), ones)
+    grad_values[tid] = wp.volume_index_to_world_dir(volume, ones)
+
+
+@wp.kernel
+def test_volume_world_to_index(
+    volume: wp.uint64,
+    points: wp.array(dtype=wp.vec3),
+    values: wp.array(dtype=wp.float32),
+    grad_values: wp.array(dtype=wp.vec3),
+):
+    tid = wp.tid()
+    p = points[tid]
+    ones = wp.vec3(1.0, 1.0, 1.0)
+    values[tid] = wp.dot(wp.volume_world_to_index(volume, p), ones)
+    grad_values[tid] = wp.volume_world_to_index_dir(volume, ones)
+
+
+# Volume write tests
+@wp.kernel
+def test_volume_store_f(volume: wp.uint64, points: wp.array(dtype=wp.vec3), values: wp.array(dtype=wp.float32)):
+    tid = wp.tid()
+
+    p = points[tid]
+    i = int(p[0])
+    j = int(p[1])
+    k = int(p[2])
+
+    wp.volume_store(volume, i, j, k, float(i + 100 * j + 10000 * k))
+    values[tid] = wp.volume_lookup_f(volume, i, j, k)
+
+
+@wp.kernel
+def test_volume_store_v(volume: wp.uint64, points: wp.array(dtype=wp.vec3), values: wp.array(dtype=wp.vec3)):
+    tid = wp.tid()
+
+    p = points[tid]
+    i = int(p[0])
+    j = int(p[1])
+    k = int(p[2])
+
+    wp.volume_store(volume, i, j, k, p)
+    values[tid] = wp.volume_lookup_v(volume, i, j, k)
+
+
+@wp.kernel
+def test_volume_store_i(volume: wp.uint64, points: wp.array(dtype=wp.vec3), values: wp.array(dtype=wp.int32)):
+    tid = wp.tid()
+
+    p = points[tid]
+    i = int(p[0])
+    j = int(p[1])
+    k = int(p[2])
+
+    wp.volume_store(volume, i, j, k, i + 100 * j + 10000 * k)
+    values[tid] = wp.volume_lookup_i(volume, i, j, k)
+
+
+@wp.kernel
+def test_volume_store_v4(volume: wp.uint64, points: wp.array(dtype=wp.vec3), values: wp.array(dtype=wp.vec4)):
+    tid = wp.tid()
+
+    p = points[tid]
+    i = int(p[0])
+    j = int(p[1])
+    k = int(p[2])
+
+    v = wp.vec4(p[0], p[1], p[2], float(i + 100 * j + 10000 * k))
+
+    wp.volume_store(volume, i, j, k, v)
+
+    values[tid] = wp.volume_lookup(volume, i, j, k, dtype=wp.vec4)
+
+
+devices = get_test_devices()
+rng = np.random.default_rng(101215)
+
+# Note about the test grids:
+# test_grid and test_int32_grid
+#   active region: [-10,10]^3
+#   values: v[i,j,k] = i * j * k
+#   voxel size: 0.25
+#
+# test_vec_grid
+#   active region: [-10,10]^3
+#   values: v[i,j,k] = (i + 2*j + 3*k, 4*i + 5*j + 6*k, 7*i + 8*j + 9*k)
+#   voxel size: 0.25
+#
+# torus
+#   index to world transformation:
+#      [0.1, 0, 0, 0]
+#      [0, 0, 0.1, 0]
+#      [0, 0.1, 0, 0]
+#      [1, 2, 3, 1]
+#   (-90 degrees rotation along X)
+#   voxel size: 0.1
+volume_paths = {
+    "float": os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "assets", "test_grid.nvdb")),
+    "int32": os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "assets", "test_int32_grid.nvdb")),
+    "vec3f": os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "assets", "test_vec_grid.nvdb")),
+    "index": os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "assets", "test_index_grid.nvdb")),
+    "torus": os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "assets", "torus.nvdb")),
+    "float_write": os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "assets", "test_grid.nvdb")),
+}
+
+test_volume_tiles = (
+    np.array([[i, j, k] for i in range(-2, 2) for j in range(-2, 2) for k in range(-2, 2)], dtype=np.int32) * 8
+)
+
+volumes = {}
+for value_type, path in volume_paths.items():
+    volumes[value_type] = {}
+    volume_data = open(path, "rb").read()
+    for device in devices:
+        try:
+            volume = wp.Volume.load_from_nvdb(volume_data, device)
+        except RuntimeError as e:
+            raise RuntimeError(f'Failed to load volume from "{path}" to {device} memory:\n{e}') from e
+
+        volumes[value_type][device.alias] = volume
+
+axis = np.linspace(-1, 1, 3)
+point_grid = np.array([[x, y, z] for x in axis for y in axis for z in axis], dtype=np.float32)
+
+
+def test_volume_sample_linear_f_gradient(test, device):
+    points = rng.uniform(-10.0, 10.0, size=(100, 3))
+    values = wp.array(np.zeros(1), dtype=wp.float32, device=device, requires_grad=True)
+    for test_case in points:
+        uvws = wp.array(test_case, dtype=wp.vec3, device=device, requires_grad=True)
+        xyzs = wp.array(test_case * 0.25, dtype=wp.vec3, device=device, requires_grad=True)
+
+        tape = wp.Tape()
+        with tape:
+            wp.launch(
+                test_volume_sample_local_f_linear_values,
+                dim=1,
+                inputs=[volumes["float"][device.alias].id, uvws, values],
+                device=device,
+            )
+        tape.backward(values)
+
+        x, y, z = test_case
+        grad_expected = np.array([y * z, x * z, x * y])
+        grad_computed = tape.gradients[uvws].numpy()[0]
+        np.testing.assert_allclose(grad_computed, grad_expected, rtol=1e-4)
+
+        tape = wp.Tape()
+        with tape:
+            wp.launch(
+                test_volume_sample_world_f_linear_values,
+                dim=1,
+                inputs=[volumes["float"][device.alias].id, xyzs, values],
+                device=device,
+            )
+        tape.backward(values)
+
+        x, y, z = test_case
+        grad_expected = np.array([y * z, x * z, x * y]) / 0.25
+        grad_computed = tape.gradients[xyzs].numpy()[0]
+        np.testing.assert_allclose(grad_computed, grad_expected, rtol=1e-4)
+
+
+def test_volume_sample_grad_linear_f_gradient(test, device):
+    points = rng.uniform(-10.0, 10.0, size=(100, 3))
+    values = wp.array(np.zeros(1), dtype=wp.float32, device=device, requires_grad=True)
+    for test_case in points:
+        uvws = wp.array(test_case, dtype=wp.vec3, device=device, requires_grad=True)
+        xyzs = wp.array(test_case * 0.25, dtype=wp.vec3, device=device, requires_grad=True)
+
+        for case_num in range(4):
+            tape = wp.Tape()
+            with tape:
+                wp.launch(
+                    test_volume_sample_grad_local_f_linear_values,
+                    dim=1,
+                    inputs=[volumes["float"][device.alias].id, uvws, values, case_num],
+                    device=device,
+                )
+            tape.backward(values)
+
+            x, y, z = test_case
+            grad_computed = tape.gradients[uvws].numpy()[0]
+            if case_num == 0:
+                grad_expected = np.array([y * z, x * z, x * y])
+            elif case_num == 1:
+                grad_expected = np.array([0.0, z, y])
+            elif case_num == 2:
+                grad_expected = np.array([z, 0.0, x])
+            elif case_num == 3:
+                grad_expected = np.array([y, x, 0.0])
+
+            np.testing.assert_allclose(grad_computed, grad_expected, rtol=1e-4)
+            tape.zero()
+
+        for case_num in range(4):
+            tape = wp.Tape()
+            with tape:
+                wp.launch(
+                    test_volume_sample_grad_world_f_linear_values,
+                    dim=1,
+                    inputs=[volumes["float"][device.alias].id, xyzs, values, case_num],
+                    device=device,
+                )
+            tape.backward(values)
+
+            x, y, z = test_case
+            grad_computed = tape.gradients[xyzs].numpy()[0]
+            if case_num == 0:
+                grad_expected = np.array([y * z, x * z, x * y]) / 0.25
+            elif case_num == 1:
+                grad_expected = np.array([0.0, z, y]) / 0.25
+            elif case_num == 2:
+                grad_expected = np.array([z, 0.0, x]) / 0.25
+            elif case_num == 3:
+                grad_expected = np.array([y, x, 0.0]) / 0.25
+
+            np.testing.assert_allclose(grad_computed, grad_expected, rtol=1e-4)
+            tape.zero()
+
+
+def test_volume_sample_linear_v_gradient(test, device):
+    points = rng.uniform(-10.0, 10.0, size=(100, 3))
+    values = wp.zeros(1, dtype=wp.float32, device=device, requires_grad=True)
+    for test_case in points:
+        uvws = wp.array(test_case, dtype=wp.vec3, device=device, requires_grad=True)
+        xyzs = wp.array(test_case * 0.25, dtype=wp.vec3, device=device, requires_grad=True)
+
+        tape = wp.Tape()
+        with tape:
+            wp.launch(
+                test_volume_sample_local_v_linear_values,
+                dim=1,
+                inputs=[volumes["vec3f"][device.alias].id, uvws, values],
+                device=device,
+            )
+        tape.backward(values)
+
+        grad_expected = np.array([12.0, 15.0, 18.0])
+        grad_computed = tape.gradients[uvws].numpy()[0]
+        np.testing.assert_allclose(grad_computed, grad_expected, rtol=1e-4)
+
+        tape = wp.Tape()
+        with tape:
+            wp.launch(
+                test_volume_sample_world_v_linear_values,
+                dim=1,
+                inputs=[volumes["vec3f"][device.alias].id, xyzs, values],
+                device=device,
+            )
+        tape.backward(values)
+
+        grad_expected = np.array([12.0, 15.0, 18.0]) / 0.25
+        grad_computed = tape.gradients[xyzs].numpy()[0]
+        np.testing.assert_allclose(grad_computed, grad_expected, rtol=1e-4)
+
+
+def test_volume_transform_gradient(test, device):
+    values = wp.zeros(1, dtype=wp.float32, device=device, requires_grad=True)
+    grad_values = wp.zeros(1, dtype=wp.vec3, device=device)
+    test_points = rng.uniform(-10.0, 10.0, size=(10, 3))
+    for test_case in test_points:
+        points = wp.array(test_case, dtype=wp.vec3, device=device, requires_grad=True)
+        tape = wp.Tape()
+        with tape:
+            wp.launch(
+                test_volume_index_to_world,
+                dim=1,
+                inputs=[volumes["torus"][device.alias].id, points, values, grad_values],
+                device=device,
+            )
+        tape.backward(values)
+
+        grad_computed = tape.gradients[points].numpy()
+        grad_expected = grad_values.numpy()
+        np.testing.assert_allclose(grad_computed, grad_expected, rtol=1e-4)
+
+        grad_computed = tape.gradients[points].numpy()
+        grad_expected = grad_values.numpy()
+        np.testing.assert_allclose(grad_computed, grad_expected, rtol=1e-4)
+
+
+def test_volume_store(test, device):
+    values_ref = np.array([x + 100 * y + 10000 * z for x, y, z in point_grid])
+    points = wp.array(point_grid, dtype=wp.vec3, device=device)
+    values = wp.empty(len(point_grid), dtype=wp.float32, device=device)
+    wp.launch(
+        test_volume_store_f,
+        dim=len(point_grid),
+        inputs=[volumes["float_write"][device.alias].id, points, values],
+        device=device,
+    )
+
+    values_res = values.numpy()
+    np.testing.assert_equal(values_res, values_ref)
+
+
+def test_volume_allocation_f(test, device):
+    bg_value = -123.0
+    points_np = np.append(point_grid, [[8096, 8096, 8096]], axis=0)
+    values_ref = np.append(np.array([x + 100 * y + 10000 * z for x, y, z in point_grid]), bg_value)
+
+    volume = wp.Volume.allocate(min=[-11, -11, -11], max=[11, 11, 11], voxel_size=0.1, bg_value=bg_value, device=device)
+    points = wp.array(points_np, dtype=wp.vec3, device=device)
+    values = wp.empty(len(points_np), dtype=wp.float32, device=device)
+    wp.launch(test_volume_store_f, dim=len(points_np), inputs=[volume.id, points, values], device=device)
+
+    values_res = values.numpy()
+    np.testing.assert_equal(values_res, values_ref)
+
+
+def test_volume_allocation_v(test, device):
+    bg_value = (-1, 2.0, -3)
+    points_np = np.append(point_grid, [[8096, 8096, 8096]], axis=0)
+    values_ref = np.append(point_grid, [bg_value], axis=0)
+
+    volume = wp.Volume.allocate(min=[-11, -11, -11], max=[11, 11, 11], voxel_size=0.1, bg_value=bg_value, device=device)
+    points = wp.array(points_np, dtype=wp.vec3, device=device)
+    values = wp.empty(len(points_np), dtype=wp.vec3, device=device)
+    wp.launch(test_volume_store_v, dim=len(points_np), inputs=[volume.id, points, values], device=device)
+
+    values_res = values.numpy()
+    np.testing.assert_equal(values_res, values_ref)
+
+
+def test_volume_allocation_i(test, device):
+    bg_value = -123
+    points_np = np.append(point_grid, [[8096, 8096, 8096]], axis=0)
+    values_ref = np.append(np.array([x + 100 * y + 10000 * z for x, y, z in point_grid], dtype=np.int32), bg_value)
+
+    volume = wp.Volume.allocate(min=[-11, -11, -11], max=[11, 11, 11], voxel_size=0.1, bg_value=bg_value, device=device)
+    points = wp.array(points_np, dtype=wp.vec3, device=device)
+    values = wp.empty(len(points_np), dtype=wp.int32, device=device)
+    wp.launch(test_volume_store_i, dim=len(points_np), inputs=[volume.id, points, values], device=device)
+
+    values_res = values.numpy()
+    np.testing.assert_equal(values_res, values_ref)
+
+
+def test_volume_allocation_v4(test, device):
+    bg_value = (-1, 2.0, -3, 5)
+    points_np = np.append(point_grid, [[8096, 8096, 8096]], axis=0)
+
+    w_ref = np.array([x + 100 * y + 10000 * z for x, y, z in point_grid])[:, np.newaxis]
+    values_ref = np.append(np.hstack((point_grid, w_ref)), [bg_value], axis=0)
+
+    volume = wp.Volume.allocate(min=[-11, -11, -11], max=[11, 11, 11], voxel_size=0.1, bg_value=bg_value, device=device)
+    points = wp.array(points_np, dtype=wp.vec3, device=device)
+    values = wp.empty(len(points_np), dtype=wp.vec4, device=device)
+    wp.launch(test_volume_store_v4, dim=len(points_np), inputs=[volume.id, points, values], device=device)
+
+    values_res = values.numpy()
+    np.testing.assert_equal(values_res, values_ref)
+
+
+def test_volume_introspection(test, device):
+    for volume_names in ("float", "vec3f"):
+        with test.subTest(volume_names=volume_names):
+            volume = volumes[volume_names][device.alias]
+            tiles_actual = volume.get_tiles().numpy()
+            tiles_sorted = tiles_actual[np.lexsort(tiles_actual.T[::-1])]
+            voxel_size = np.array(volume.get_voxel_size())
+
+            np.testing.assert_equal(test_volume_tiles, tiles_sorted)
+            np.testing.assert_equal([0.25] * 3, voxel_size)
+
+            voxel_count = volume.get_voxel_count()
+            voxels_actual = volume.get_voxels().numpy()
+            assert voxel_count == voxels_actual.shape[0]
+
+            # Voxel coordinates should be unique
+            voxels_unique = np.unique(voxels_actual, axis=0)
+            assert voxel_count == voxels_unique.shape[0]
+
+            # Get back tiles from voxels, should match get_tiles()
+            voxel_tiles = 8 * (voxels_unique // 8)
+            voxel_tiles_sorted = voxel_tiles[np.lexsort(voxel_tiles.T[::-1])]
+            voxel_tiles_unique = np.unique(voxel_tiles_sorted, axis=0)
+
+            np.testing.assert_equal(voxel_tiles_unique, tiles_sorted)
+
+
+def test_volume_multiple_grids(test, device):
+    volume = volumes["index"][device.alias]
+
+    volume_2 = volume.load_next_grid()
+
+    test.assertIsNotNone(volume_2)
+
+    test.assertNotEqual(volume.id, volume_2.id)
+    test.assertNotEqual(volume.get_voxel_count(), volume_2.get_voxel_count())
+
+    test.assertEqual(volume.get_grid_info().grid_count, volume_2.get_grid_info().grid_count)
+    test.assertEqual(volume.get_grid_info().grid_index + 1, volume_2.get_grid_info().grid_index)
+
+    volume_3 = volume_2.load_next_grid()
+    test.assertIsNone(volume_3)
+
+
+def test_volume_feature_array(test, device):
+    volume = volumes["index"][device.alias]
+
+    test.assertEqual(volume.get_feature_array_count(), 1)
+
+    array = volume.feature_array(0, dtype=wp.uint64)
+    test.assertEqual(array.device, device)
+    test.assertEqual(array.dtype, wp.uint64)
+
+    # fVDB convention, data starts with array ndim + shape
+    np.testing.assert_equal(array.numpy()[0:4], [3, volume.get_voxel_count(), 2, 3])
+
+
+@wp.kernel
+def fill_leaf_values_kernel(volume: wp.uint64, ijk: wp.array2d(dtype=wp.int32), values: wp.array(dtype=Any)):
+    tid = wp.tid()
+
+    i = ijk[tid, 0]
+    j = ijk[tid, 1]
+    k = ijk[tid, 2]
+
+    expect_eq(tid, wp.volume_lookup_index(volume, i, j, k))
+
+    values[tid] = wp.volume_lookup(volume, i, j, k, dtype=values.dtype)
+
+
+@wp.kernel
+def test_volume_sample_index_kernel(
+    volume: wp.uint64,
+    points: wp.array(dtype=wp.vec3),
+    values: wp.array(dtype=Any),
+    background: wp.array(dtype=Any),
+    sampled_values: wp.array(dtype=Any),
+):
+    tid = wp.tid()
+    p = points[tid]
+
+    ref = wp.volume_sample(volume, p, wp.Volume.LINEAR, dtype=values.dtype)
+    sampled_values[tid] = wp.volume_sample_index(volume, p, wp.Volume.LINEAR, values, background[0])
+    expect_eq(sampled_values[tid], ref)
+
+
+@wp.kernel
+def test_volume_sample_grad_index_kernel(
+    volume: wp.uint64,
+    points: wp.array(dtype=wp.vec3),
+    values: wp.array(dtype=Any),
+    background: wp.array(dtype=Any),
+    sampled_values: wp.array(dtype=Any),
+    sampled_grads: wp.array(dtype=Any),
+):
+    tid = wp.tid()
+    p = points[tid]
+
+    ref_grad = sampled_grads.dtype()
+    ref = wp.volume_sample_grad(volume, p, wp.Volume.LINEAR, ref_grad, dtype=values.dtype)
+
+    grad = type(ref_grad)()
+    sampled_values[tid] = wp.volume_sample_grad_index(volume, p, wp.Volume.LINEAR, values, background[0], grad)
+    expect_eq(sampled_values[tid], ref)
+
+    expect_eq(grad[0], ref_grad[0])
+    expect_eq(grad[1], ref_grad[1])
+    expect_eq(grad[2], ref_grad[2])
+    sampled_grads[tid] = grad
+
+
+def test_volume_sample_index(test, device):
+    points = rng.uniform(-10.0, 10.0, size=(100, 3))
+    points[0:10, 0] += 100.0  # ensure some points are over unallocated voxels
+    uvws = wp.array(points, dtype=wp.vec3, device=device)
+
+    bg_values = {
+        "float": 10.0,
+        "vec3f": wp.vec3(10.8, -4.13, 10.26),
+    }
+    grad_types = {
+        "float": wp.vec3,
+        "vec3f": wp.mat33,
+    }
+
+    for volume_names in ("float", "vec3f"):
+        with test.subTest(volume_names=volume_names):
+            volume = volumes[volume_names][device.alias]
+
+            ijk = volume.get_voxels()
+
+            values = wp.empty(shape=volume.get_voxel_count(), dtype=volume.dtype, device=device, requires_grad=True)
+
+            vid = wp.uint64(volume.id)
+            wp.launch(fill_leaf_values_kernel, dim=values.shape, inputs=[vid, ijk, values], device=device)
+
+            sampled_values = wp.empty(shape=points.shape[0], dtype=volume.dtype, device=device, requires_grad=True)
+            background = wp.array([bg_values[volume_names]], dtype=volume.dtype, device=device, requires_grad=True)
+
+            tape = wp.Tape()
+            with tape:
+                wp.launch(
+                    test_volume_sample_index_kernel,
+                    dim=points.shape[0],
+                    inputs=[vid, uvws, values, background, sampled_values],
+                    device=device,
+                )
+
+            sampled_values.grad.fill_(1.0)
+            tape.backward()
+
+            # test adjoint w.r.t voxel and background value arrays
+            # we should have sum(sampled_values) = sum(adj_values * values) + (adj_background * background)
+            sum_sampled_values = np.sum(sampled_values.numpy(), axis=0)
+            sum_values_adj = np.sum(values.numpy() * values.grad.numpy(), axis=0)
+            sum_background_adj = background.numpy()[0] * background.grad.numpy()[0]
+
+            np.testing.assert_allclose(sum_sampled_values, sum_values_adj + sum_background_adj, rtol=1.0e-3)
+
+            tape.reset()
+
+            sampled_grads = wp.empty(
+                shape=points.shape[0], dtype=grad_types[volume_names], device=device, requires_grad=True
+            )
+
+            with tape:
+                wp.launch(
+                    test_volume_sample_grad_index_kernel,
+                    dim=points.shape[0],
+                    inputs=[vid, uvws, values, background, sampled_values, sampled_grads],
+                    device=device,
+                )
+
+            sampled_values.grad.fill_(1.0)
+            tape.backward()
+
+            # we should have sum(sampled_values) = sum(adj_values * values) + (adj_background * background)
+            sum_sampled_values = np.sum(sampled_values.numpy(), axis=0)
+            sum_values_adj = np.sum(values.numpy() * values.grad.numpy(), axis=0)
+            sum_background_adj = background.numpy()[0] * background.grad.numpy()[0]
+            np.testing.assert_allclose(sum_sampled_values, sum_values_adj + sum_background_adj, rtol=1.0e-3)
+
+            tape.zero()
+            sampled_values.grad.fill_(0.0)
+            sampled_grads.grad.fill_(1.0)
+            tape.backward()
+
+            # we should have sum(sampled_grad, axes=(0, -1)) = sum(adj_values * values) + (adj_background * background)
+            sum_sampled_grads = np.sum(np.sum(sampled_grads.numpy(), axis=0), axis=-1)
+            sum_values_adj = np.sum(values.numpy() * values.grad.numpy(), axis=0)
+            sum_background_adj = background.numpy()[0] * background.grad.numpy()[0]
+            np.testing.assert_allclose(sum_sampled_grads, sum_values_adj + sum_background_adj, rtol=1.0e-3)
+
+
+def test_volume_from_numpy(test, device):
+    # Volume.allocate_from_tiles() is only available with CUDA
+    mins = np.array([-3.0, -3.0, -3.0])
+    voxel_size = 0.2
+    maxs = np.array([3.0, 3.0, 3.0])
+    nums = np.ceil((maxs - mins) / (voxel_size)).astype(dtype=int)
+    center = np.array([0.0, 0.0, 0.0])
+    rad = 2.5
+    sphere_sdf_np = np.zeros(tuple(nums))
+    for x in range(nums[0]):
+        for y in range(nums[1]):
+            for z in range(nums[2]):
+                pos = mins + voxel_size * np.array([x, y, z])
+                dis = np.linalg.norm(pos - center)
+                sphere_sdf_np[x, y, z] = dis - rad
+    sphere_vdb = wp.Volume.load_from_numpy(sphere_sdf_np, mins, voxel_size, rad + 3.0 * voxel_size, device=device)
+
+    test.assertNotEqual(sphere_vdb.id, 0)
+
+    sphere_vdb_array = sphere_vdb.array()
+    test.assertEqual(sphere_vdb_array.dtype, wp.uint8)
+    test.assertIsNone(sphere_vdb_array.deleter)
+
+
+def test_volume_from_numpy_3d(test, device):
+    # Volume.allocate_from_tiles() is only available with CUDA
+    mins = np.array([-3.0, -3.0, -3.0])
+    voxel_size = 0.2
+    maxs = np.array([3.0, 3.0, 3.0])
+    nums = np.ceil((maxs - mins) / (voxel_size)).astype(dtype=int)
+    centers = np.array([[-1.0, -1.0, -1.0], [0.0, 0.0, 0.0], [1.0, 1.0, 1.0]])
+    rad = 2.5
+    sphere_sdf_np = np.zeros(tuple(nums) + (3,))
+    for x in range(nums[0]):
+        for y in range(nums[1]):
+            for z in range(nums[2]):
+                for k in range(3):
+                    pos = mins + voxel_size * np.array([x, y, z])
+                    dis = np.linalg.norm(pos - centers[k])
+                    sphere_sdf_np[x, y, z, k] = dis - rad
+    sphere_vdb = wp.Volume.load_from_numpy(
+        sphere_sdf_np, mins, voxel_size, (rad + 3.0 * voxel_size,) * 3, device=device
+    )
+
+    test.assertNotEqual(sphere_vdb.id, 0)
+
+    sphere_vdb_array = sphere_vdb.array()
+    test.assertEqual(sphere_vdb_array.dtype, wp.uint8)
+    test.assertIsNone(sphere_vdb_array.deleter)
+
+
+def test_volume_aniso_transform(test, device):
+    # XY-rotation + z scale
+    transform = [
+        [0, -1, 0],
+        [1, 0, 0],
+        [0, 0, 2],
+    ]
+
+    points = wp.array([[-1, 1, 4]], dtype=float, device=device)
+    volume = wp.Volume.allocate_by_voxels(voxel_points=points, transform=transform, device=device)
+
+    # Check that world points are correctly converted to local space
+    voxels = volume.get_voxels().numpy()
+    assert_np_equal(voxels, [[1, 1, 2]])
+
+    # Check that we retrieve the correct transform from the grid metadata
+    assert_np_equal(volume.get_voxel_size(), [-1, 1, 2])
+    assert_np_equal(transform, np.array(volume.get_grid_info().transform_matrix).reshape(3, 3))
+
+
+def test_volume_write(test, device):
+    codecs = ["none", "zip", "blosc"]
+    try:
+        import blosc  # noqa: F401 I001
+    except ImportError:
+        codecs.pop()
+
+    for volume_name in ("float", "vec3f", "index"):
+        for codec in codecs:
+            with test.subTest(volume_name=volume_name, codec=codec):
+                volume = volumes[volume_name][device.alias]
+                fd, file_path = tempfile.mkstemp(suffix=".nvdb")
+                os.close(fd)
+                try:
+                    volume.save_to_nvdb(file_path, codec=codec)
+                    with open(file_path, "rb") as f:
+                        volume_2 = wp.Volume.load_from_nvdb(f)
+                    next_volume = volume
+                    while next_volume:
+                        np.testing.assert_array_equal(next_volume.array().numpy(), volume_2.array().numpy())
+                        next_volume = next_volume.load_next_grid()
+                        volume_2 = volume_2.load_next_grid()
+
+                finally:
+                    os.remove(file_path)
+
+    with test.subTest(volume_write="unsupported"):
+        volume = volumes["index"][device.alias]
+        volume = volume.load_next_grid()
+
+        fd, file_path = tempfile.mkstemp(suffix=".nvdb")
+        os.close(fd)
+
+        try:
+            with test.assertRaises(RuntimeError):
+                volume.save_to_nvdb(file_path, codec=codec)
+        finally:
+            os.remove(file_path)
+
+
+class TestVolume(unittest.TestCase):
+    def test_volume_new_del(self):
+        # test the scenario in which a volume is created but not initialized before gc
+        instance = wp.Volume.__new__(wp.Volume)
+        instance.__del__()
+
+
+add_function_test(
+    TestVolume, "test_volume_sample_linear_f_gradient", test_volume_sample_linear_f_gradient, devices=devices
+)
+add_function_test(
+    TestVolume, "test_volume_sample_grad_linear_f_gradient", test_volume_sample_grad_linear_f_gradient, devices=devices
+)
+add_function_test(
+    TestVolume, "test_volume_sample_linear_v_gradient", test_volume_sample_linear_v_gradient, devices=devices
+)
+add_function_test(TestVolume, "test_volume_transform_gradient", test_volume_transform_gradient, devices=devices)
+add_function_test(TestVolume, "test_volume_store", test_volume_store, devices=devices)
+add_function_test(
+    TestVolume, "test_volume_allocation_f", test_volume_allocation_f, devices=get_selected_cuda_test_devices()
+)
+add_function_test(
+    TestVolume, "test_volume_allocation_v", test_volume_allocation_v, devices=get_selected_cuda_test_devices()
+)
+add_function_test(
+    TestVolume, "test_volume_allocation_i", test_volume_allocation_i, devices=get_selected_cuda_test_devices()
+)
+add_function_test(
+    TestVolume, "test_volume_allocation_v4", test_volume_allocation_v4, devices=get_selected_cuda_test_devices()
+)
+add_function_test(TestVolume, "test_volume_introspection", test_volume_introspection, devices=devices)
+add_function_test(
+    TestVolume, "test_volume_from_numpy", test_volume_from_numpy, devices=get_selected_cuda_test_devices()
+)
+add_function_test(
+    TestVolume, "test_volume_from_numpy_3d", test_volume_from_numpy_3d, devices=get_selected_cuda_test_devices()
+)
+add_function_test(
+    TestVolume, "test_volume_aniso_transform", test_volume_aniso_transform, devices=get_selected_cuda_test_devices()
+)
+add_function_test(TestVolume, "test_volume_multiple_grids", test_volume_multiple_grids, devices=devices)
+add_function_test(TestVolume, "test_volume_feature_array", test_volume_feature_array, devices=devices)
+add_function_test(TestVolume, "test_volume_sample_index", test_volume_sample_index, devices=devices)
+add_function_test(TestVolume, "test_volume_write", test_volume_write, devices=[wp.get_device("cpu")])
+
+points = {}
+points_jittered = {}
+for device in devices:
+    points_jittered_np = point_grid + rng.uniform(-0.5, 0.5, size=point_grid.shape)
+    points[device.alias] = wp.array(point_grid, dtype=wp.vec3, device=device)
+    points_jittered[device.alias] = wp.array(points_jittered_np, dtype=wp.vec3, device=device)
+
+    add_kernel_test(
+        TestVolume,
+        test_volume_lookup_f,
+        dim=len(point_grid),
+        inputs=[volumes["float"][device.alias].id, points[device.alias]],
+        devices=[device],
+    )
+    add_kernel_test(
+        TestVolume,
+        test_volume_sample_closest_f,
+        dim=len(point_grid),
+        inputs=[volumes["float"][device.alias].id, points_jittered[device.alias]],
+        devices=[device.alias],
+    )
+    add_kernel_test(
+        TestVolume,
+        test_volume_sample_linear_f,
+        dim=len(point_grid),
+        inputs=[volumes["float"][device.alias].id, points_jittered[device.alias]],
+        devices=[device.alias],
+    )
+    add_kernel_test(
+        TestVolume,
+        test_volume_sample_grad_linear_f,
+        dim=len(point_grid),
+        inputs=[volumes["float"][device.alias].id, points_jittered[device.alias]],
+        devices=[device.alias],
+    )
+
+    add_kernel_test(
+        TestVolume,
+        test_volume_lookup_v,
+        dim=len(point_grid),
+        inputs=[volumes["vec3f"][device.alias].id, points[device.alias]],
+        devices=[device.alias],
+    )
+    add_kernel_test(
+        TestVolume,
+        test_volume_sample_closest_v,
+        dim=len(point_grid),
+        inputs=[volumes["vec3f"][device.alias].id, points_jittered[device.alias]],
+        devices=[device.alias],
+    )
+    add_kernel_test(
+        TestVolume,
+        test_volume_sample_linear_v,
+        dim=len(point_grid),
+        inputs=[volumes["vec3f"][device.alias].id, points_jittered[device.alias]],
+        devices=[device.alias],
+    )
+    add_kernel_test(
+        TestVolume,
+        test_volume_sample_grad_linear_v,
+        dim=len(point_grid),
+        inputs=[volumes["vec3f"][device.alias].id, points_jittered[device.alias]],
+        devices=[device.alias],
+    )
+
+    add_kernel_test(
+        TestVolume,
+        test_volume_lookup_i,
+        dim=len(point_grid),
+        inputs=[volumes["int32"][device.alias].id, points[device.alias]],
+        devices=[device.alias],
+    )
+    add_kernel_test(
+        TestVolume,
+        test_volume_sample_i,
+        dim=len(point_grid),
+        inputs=[volumes["int32"][device.alias].id, points_jittered[device.alias]],
+        devices=[device.alias],
+    )
+
+
+if __name__ == "__main__":
+    wp.clear_kernel_cache()
+    unittest.main(verbosity=2)