warp-lang 1.6.0__py3-none-manylinux2014_aarch64.whl → 1.6.1__py3-none-manylinux2014_aarch64.whl

warp/tests/test_launch.py

@@ -46,6 +46,12 @@ def kernel4d(a: wp.array(dtype=int, ndim=4)):
     wp.expect_eq(a[i, j, k, l], i * dim_y * dim_z * dim_w + j * dim_z * dim_w + k * dim_w + l)
 
 
+@wp.kernel
+def square_kernel(input: wp.array(dtype=float), output: wp.array(dtype=float)):
+    i = wp.tid()
+    output[i] = input[i] * input[i]
+
+
 def test1d(test, device):
     a = np.arange(0, dim_x).reshape(dim_x)
 
@@ -98,8 +104,19 @@ def kernel_cmd(params: Params, i: int, f: float, v: wp.vec3, m: wp.mat33, out: w
 
 
 def test_launch_cmd(test, device):
+    """Tests recording and executing a kernel launch command.
+
+    Verifies that:
+    - A kernel can be recorded as a command without immediate execution
+    - The recorded command can be launched later
+    - Parameters are correctly passed to the kernel
+    - Output matches expected results for both immediate and delayed launches
+
+    Args:
+        test: Test context
+        device: Device to run the test on
+    """
     n = 1
-
     ref = np.arange(0, n)
     out = wp.zeros(n, dtype=int, device=device)
 
@@ -274,12 +291,62 @@ def test_launch_cmd_empty(test, device):
     assert_np_equal(out.numpy(), ref)
 
 
+def test_launch_cmd_adjoint(test, device):
+    """Test recording an adjoint launch with record_cmd=True."""
+    input_arr = wp.array([1.0, 2.0, 3.0], dtype=float, requires_grad=True, device=device)
+    output_arr = wp.empty_like(input_arr)
+
+    output_arr.grad.fill_(1.0)
+
+    cmd = wp.launch(
+        square_kernel,
+        dim=input_arr.size,
+        inputs=[input_arr, output_arr],
+        adj_inputs=[None, None],
+        adjoint=True,
+        device=device,
+        record_cmd=True,
+    )
+
+    cmd.launch()
+
+    assert_np_equal(input_arr.grad.numpy(), np.array([2.0, 4.0, 6.0]))
+
+
+def test_launch_cmd_adjoint_empty(test, device):
+    """Test constructing a Launch object for an adjoint kernel."""
+    input_arr = wp.array([1.0, 2.0, 3.0], dtype=float, requires_grad=True, device=device)
+    output_arr = wp.empty_like(input_arr)
+    output_arr.grad.fill_(1.0)
+
+    cmd = wp.Launch(square_kernel, device, adjoint=True)
+    cmd.set_param_by_name("input", input_arr)
+    cmd.set_param_by_name("output", output_arr)
+    cmd.set_dim(input_arr.size)
+    cmd.launch()
+
+    assert_np_equal(input_arr.grad.numpy(), np.array([2.0, 4.0, 6.0]))
+
+    # Now update the launch object's parameters with arrays of different sizes and values
+    # and check that the adjoints are correctly computed
+    input_arr_updated = wp.array([4.0, 5.0, 6.0, 7.0], dtype=float, device=device)
+    input_arr_updated_grad = wp.zeros_like(input_arr_updated)
+
+    output_arr_updated = wp.empty_like(input_arr_updated)
+    output_arr_updated_grad = wp.full_like(output_arr_updated, 1.0)
+
+    cmd.set_param_by_name("input", input_arr_updated)
+    cmd.set_param_by_name("output", output_arr_updated)
+    cmd.set_param_by_name("input", input_arr_updated_grad, adjoint=True)
+    cmd.set_param_by_name("output", output_arr_updated_grad, adjoint=True)
+    cmd.set_dim(input_arr_updated.size)
+    cmd.launch()
+
+    assert_np_equal(input_arr_updated_grad.numpy(), np.array([8.0, 10.0, 12.0, 14.0]))
+
+
 @wp.kernel
-def kernel_mul(
-    values: wp.array(dtype=int),
-    coeff: int,
-    out: wp.array(dtype=int),
-):
+def kernel_mul(values: wp.array(dtype=int), coeff: int, out: wp.array(dtype=int)):
     tid = wp.tid()
     out[tid] = values[tid] * coeff
 
@@ -301,28 +368,10 @@ def test_launch_tuple_args(test, device):
     )
     assert_np_equal(out.numpy(), np.array((0, 3, 6, 9)))
 
-    wp.launch(
-        kernel_mul,
-        dim=len(values),
-        inputs=(
-            values,
-            coeff,
-            out,
-        ),
-        device=device,
-    )
+    wp.launch(kernel_mul, dim=len(values), inputs=(values, coeff, out), device=device)
     assert_np_equal(out.numpy(), np.array((0, 3, 6, 9)))
 
-    wp.launch(
-        kernel_mul,
-        dim=len(values),
-        outputs=(
-            values,
-            coeff,
-            out,
-        ),
-        device=device,
-    )
+    wp.launch(kernel_mul, dim=len(values), outputs=(values, coeff, out), device=device)
     assert_np_equal(out.numpy(), np.array((0, 3, 6, 9)))
 
 
@@ -343,6 +392,8 @@ add_function_test(TestLaunch, "test_launch_cmd_set_param", test_launch_cmd_set_p
 add_function_test(TestLaunch, "test_launch_cmd_set_ctype", test_launch_cmd_set_ctype, devices=devices)
 add_function_test(TestLaunch, "test_launch_cmd_set_dim", test_launch_cmd_set_dim, devices=devices)
 add_function_test(TestLaunch, "test_launch_cmd_empty", test_launch_cmd_empty, devices=devices)
+add_function_test(TestLaunch, "test_launch_cmd_adjoint", test_launch_cmd_adjoint, devices=devices)
+add_function_test(TestLaunch, "test_launch_cmd_adjoint_empty", test_launch_cmd_adjoint_empty, devices=devices)
 
 add_function_test(TestLaunch, "test_launch_tuple_args", test_launch_tuple_args, devices=devices)
 

warp/tests/test_mat.py

@@ -384,6 +384,77 @@ def test_negation(test, device, dtype, register_kernels=False):
                     idx = idx + 1
 
 
+def test_matmul(test, device, dtype, register_kernels=False):
+    rng = np.random.default_rng(123)
+
+    tol = {
+        np.float16: 5.0e-3,
+        np.float32: 1.0e-6,
+        np.float64: 1.0e-12,
+    }.get(dtype, 0)
+
+    wptype = wp.types.np_dtype_to_warp_type[np.dtype(dtype)]
+    mat22 = wp.types.matrix(shape=(2, 2), dtype=wptype)
+    mat33 = wp.types.matrix(shape=(3, 3), dtype=wptype)
+    mat23 = wp.types.matrix(shape=(2, 3), dtype=wptype)
+    mat32 = wp.types.matrix(shape=(3, 2), dtype=wptype)
+    mat44 = wp.types.matrix(shape=(4, 4), dtype=wptype)
+
+    output_select_kernel = get_select_kernel(wptype)
+
+    def check_mat_mul(
+        i23: wp.array(dtype=mat23),
+        i32: wp.array(dtype=mat32),
+        i44: wp.array(dtype=mat44),
+        o22: wp.array(dtype=mat22),
+        o33: wp.array(dtype=mat33),
+        o44: wp.array(dtype=mat44),
+    ):
+        i = wp.tid()
+        o22[i] = i23[i] @ i32[i]
+        o33[i] = i32[i] @ i23[i]
+        o44[i] = i44[i] @ i44[i]
+
+    kernel = getkernel(check_mat_mul, suffix=dtype.__name__)
+
+    if register_kernels:
+        return
+
+    test_adj = dtype in np_float_types
+
+    i23 = wp.array(randvals(rng, [1, 2, 3], dtype), dtype=mat23, requires_grad=test_adj, device=device)
+    i32 = wp.array(randvals(rng, [1, 3, 2], dtype), dtype=mat32, requires_grad=test_adj, device=device)
+    i44 = wp.array(randvals(rng, [1, 4, 4], dtype), dtype=mat44, requires_grad=test_adj, device=device)
+    o22 = wp.array(randvals(rng, [1, 2, 2], dtype), dtype=mat22, requires_grad=test_adj, device=device)
+    o33 = wp.array(randvals(rng, [1, 3, 3], dtype), dtype=mat33, requires_grad=test_adj, device=device)
+    o44 = wp.array(randvals(rng, [1, 4, 4], dtype), dtype=mat44, requires_grad=test_adj, device=device)
+
+    tape = wp.Tape()
+    with tape:
+        wp.launch(
+            kernel,
+            dim=1,
+            inputs=[i23, i32, i44],
+            outputs=[o22, o33, o44],
+            device=device,
+        )
+
+    assert_np_equal(o22.numpy(), i23.numpy() @ i32.numpy(), tol=tol)
+    assert_np_equal(o33.numpy(), i32.numpy() @ i23.numpy(), tol=tol)
+    assert_np_equal(o44.numpy(), i44.numpy() @ i44.numpy(), tol=tol)
+
+    if test_adj:
+        o22.grad.assign([np.eye(2)])
+        o33.grad.assign([np.eye(3)])
+        o44.grad.assign([np.eye(4)])
+
+        tape.backward()
+
+        assert_np_equal(i23.grad.numpy(), 2.0 * i32.numpy().T, tol=tol)
+        assert_np_equal(i32.grad.numpy(), 2.0 * i23.numpy().T, tol=tol)
+        assert_np_equal(i44.grad.numpy(), 2.0 * i44.numpy().T, tol=tol)
+
+
 def test_subtraction(test, device, dtype, register_kernels=False):
     rng = np.random.default_rng(123)
 
@@ -874,7 +945,7 @@ def test_svd(test, device, dtype, register_kernels=False):
     tol = {
         np.float16: 1.0e-3,
         np.float32: 1.0e-6,
-        np.float64: 1.0e-6,
+        np.float64: 1.0e-12,
     }.get(dtype, 0)
 
     wptype = wp.types.np_dtype_to_warp_type[np.dtype(dtype)]
@@ -1765,6 +1836,9 @@ for dtype in np_signed_int_types + np_float_types:
     add_function_test_register_kernel(
         TestMat, f"test_subtraction_{dtype.__name__}", test_subtraction, devices=devices, dtype=dtype
     )
+    add_function_test_register_kernel(
+        TestMat, f"test_matmul_{dtype.__name__}", test_matmul, devices=devices, dtype=dtype
+    )
 
 add_function_test(
     TestMat,

warp/tests/test_overwrite.py

@@ -1,6 +1,7 @@
 import contextlib
 import io
 import unittest
+from typing import Any
 
 import numpy as np
 
@@ -164,18 +165,18 @@ def test_kernel_writeread_kernel_write(test, device):
 
 
 @wp.func
-def read_func(a: wp.array(dtype=float), idx: int):
+def read_func(a: wp.array(dtype=Any), idx: int):
     x = a[idx]
     return x
 
 
 @wp.func
-def read_return_func(b: wp.array(dtype=float), idx: int):
+def read_return_func(b: wp.array(dtype=Any), idx: int):
     return 1.0, b[idx]
 
 
 @wp.func
-def write_func(c: wp.array(dtype=float), idx: int):
+def write_func(c: wp.array(dtype=Any), idx: int):
     c[idx] = 1.0
 
 

warp/tests/test_tile_load.py

@@ -133,6 +133,49 @@ def test_tile_load(kernel, ndim):
     return test
 
 
+@wp.kernel
+def tile_load_unaligned_kernel(
+    input: wp.array2d(dtype=float),
+    output: wp.array2d(dtype=float),
+):
+    t = wp.tile_load(input, shape=(TILE_M, TILE_N), offset=(1, 1), storage="shared")
+    wp.tile_store(output, t, offset=(1, 1))
+
+
+def test_tile_load_unaligned(test, device):
+    rng = np.random.default_rng(42)
+
+    shape = [TILE_M + 1, TILE_N + 1]
+
+    input = wp.array(rng.random(shape), dtype=float, requires_grad=True, device=device)
+    output = wp.zeros(shape, dtype=float, device=device)
+
+    with wp.Tape() as tape:
+        wp.launch_tiled(
+            tile_load_unaligned_kernel,
+            dim=[1],
+            inputs=[input, output],
+            block_dim=TILE_DIM,
+            device=device,
+        )
+
+    # first row and column should be zero
+    assert_np_equal(output.numpy()[0, :], np.zeros(TILE_N + 1))
+    assert_np_equal(output.numpy()[:, 0], np.zeros(TILE_M + 1))
+
+    # check output elements
+    assert_np_equal(output.numpy()[1:, 1:], input.numpy()[1:, 1:])
+
+    output.grad = wp.ones_like(output)
+    tape.backward()
+
+    expected_grad = np.ones_like(input.grad.numpy())
+    expected_grad[0, :] = 0.0
+    expected_grad[:, 0] = 0.0
+
+    assert_np_equal(input.grad.numpy(), expected_grad)
+
+
 # ----------------------------------------------------------------------------------------
 
 TILE_SIZE = 4
@@ -336,7 +379,7 @@ add_function_test(TestTileLoad, "test_tile_load_1d", test_tile_load(tile_load_1d
 add_function_test(TestTileLoad, "test_tile_load_2d", test_tile_load(tile_load_2d_kernel, 2), devices=devices)
 add_function_test(TestTileLoad, "test_tile_load_3d", test_tile_load(tile_load_3d_kernel, 3), devices=devices)
 add_function_test(TestTileLoad, "test_tile_load_4d", test_tile_load(tile_load_4d_kernel, 4), devices=devices)
-
+add_function_test(TestTileLoad, "test_tile_load_unaligned", test_tile_load_unaligned, devices=devices)
 
 add_function_test(TestTileLoad, "test_tile_extract_1d", test_tile_extract(tile_extract_1d_kernel, 1), devices=devices)
 add_function_test(TestTileLoad, "test_tile_extract_2d", test_tile_extract(tile_extract_2d_kernel, 2), devices=devices)

warp/thirdparty/unittest_parallel.py

@@ -554,6 +554,9 @@ def initialize_test_process(lock, shared_index, args, temp_dir):
             wp.config.kernel_cache_dir = cache_root_dir
 
             wp.build.clear_kernel_cache()
+        elif "WARP_CACHE_ROOT" in os.environ:
+            # Using a shared cache for all test processes
+            wp.config.kernel_cache_dir = os.path.join(os.getenv("WARP_CACHE_ROOT"), wp.config.version)
 
 
 if __name__ == "__main__":  # pragma: no cover

warp/types.py

@@ -1768,77 +1768,75 @@ class array(Array):
             dtype_ndim = 0
             scalar_dtype = dtype
 
-        if hasattr(data, "__cuda_array_interface__"):
-            try:
-                # Performance note: try first, ask questions later
-                device = warp.context.runtime.get_device(device)
-            except Exception:
-                # Fallback to using the public API for retrieving the device,
-                # which takes take of initializing Warp if needed.
-                device = warp.context.get_device(device)
-
-            if device.is_cuda:
-                desc = data.__cuda_array_interface__
-                data_shape = desc.get("shape")
-                data_strides = desc.get("strides")
-                data_dtype = np.dtype(desc.get("typestr"))
-                data_ptr = desc.get("data")[0]
-
-                if dtype == Any:
-                    dtype = np_dtype_to_warp_type[data_dtype]
-
-                data_ndim = len(data_shape)
-
-                # determine whether the input needs reshaping
-                target_npshape = None
-                if shape is not None:
-                    target_npshape = (*shape, *dtype_shape)
-                elif dtype_ndim > 0:
-                    # prune inner dimensions of length 1
-                    while data_ndim > 1 and data_shape[-1] == 1:
-                        data_shape = data_shape[:-1]
-                    # if the inner dims don't match exactly, check if the innermost dim is a multiple of type length
-                    if data_ndim < dtype_ndim or data_shape[-dtype_ndim:] != dtype_shape:
-                        if data_shape[-1] == dtype._length_:
-                            target_npshape = (*data_shape[:-1], *dtype_shape)
-                        elif data_shape[-1] % dtype._length_ == 0:
-                            target_npshape = (*data_shape[:-1], data_shape[-1] // dtype._length_, *dtype_shape)
+        try:
+            # Performance note: try first, ask questions later
+            device = warp.context.runtime.get_device(device)
+        except Exception:
+            # Fallback to using the public API for retrieving the device,
+            # which takes take of initializing Warp if needed.
+            device = warp.context.get_device(device)
+
+        if device.is_cuda and hasattr(data, "__cuda_array_interface__"):
+            desc = data.__cuda_array_interface__
+            data_shape = desc.get("shape")
+            data_strides = desc.get("strides")
+            data_dtype = np.dtype(desc.get("typestr"))
+            data_ptr = desc.get("data")[0]
+
+            if dtype == Any:
+                dtype = np_dtype_to_warp_type[data_dtype]
+
+            if data_strides is None:
+                data_strides = strides_from_shape(data_shape, dtype)
+
+            data_ndim = len(data_shape)
+
+            # determine whether the input needs reshaping
+            target_npshape = None
+            if shape is not None:
+                target_npshape = (*shape, *dtype_shape)
+            elif dtype_ndim > 0:
+                # prune inner dimensions of length 1
+                while data_ndim > 1 and data_shape[-1] == 1:
+                    data_shape = data_shape[:-1]
+                # if the inner dims don't match exactly, check if the innermost dim is a multiple of type length
+                if data_ndim < dtype_ndim or data_shape[-dtype_ndim:] != dtype_shape:
+                    if data_shape[-1] == dtype._length_:
+                        target_npshape = (*data_shape[:-1], *dtype_shape)
+                    elif data_shape[-1] % dtype._length_ == 0:
+                        target_npshape = (*data_shape[:-1], data_shape[-1] // dtype._length_, *dtype_shape)
+                    else:
+                        if dtype_ndim == 1:
+                            raise RuntimeError(
+                                f"The inner dimensions of the input data are not compatible with the requested vector type {warp.context.type_str(dtype)}: expected an inner dimension that is a multiple of {dtype._length_}"
+                            )
                         else:
-                            if dtype_ndim == 1:
-                                raise RuntimeError(
-                                    f"The inner dimensions of the input data are not compatible with the requested vector type {warp.context.type_str(dtype)}: expected an inner dimension that is a multiple of {dtype._length_}"
-                                )
-                            else:
-                                raise RuntimeError(
-                                    f"The inner dimensions of the input data are not compatible with the requested matrix type {warp.context.type_str(dtype)}: expected inner dimensions {dtype._shape_} or a multiple of {dtype._length_}"
-                                )
-
-                if target_npshape is None:
-                    target_npshape = data_shape if shape is None else shape
-
-                # determine final shape and strides
-                if dtype_ndim > 0:
-                    # make sure the inner dims are contiguous for vector/matrix types
-                    scalar_size = type_size_in_bytes(dtype._wp_scalar_type_)
-                    inner_contiguous = data_strides[-1] == scalar_size
-                    if inner_contiguous and dtype_ndim > 1:
-                        inner_contiguous = data_strides[-2] == scalar_size * dtype_shape[-1]
-
-                    shape = target_npshape[:-dtype_ndim] or (1,)
-                    strides = data_strides if shape == data_shape else strides_from_shape(shape, dtype)
-                else:
-                    shape = target_npshape or (1,)
-                    strides = data_strides if shape == data_shape else strides_from_shape(shape, dtype)
+                            raise RuntimeError(
+                                f"The inner dimensions of the input data are not compatible with the requested matrix type {warp.context.type_str(dtype)}: expected inner dimensions {dtype._shape_} or a multiple of {dtype._length_}"
+                            )
+
+            if target_npshape is None:
+                target_npshape = data_shape if shape is None else shape
+
+            # determine final shape and strides
+            if dtype_ndim > 0:
+                # make sure the inner dims are contiguous for vector/matrix types
+                scalar_size = type_size_in_bytes(dtype._wp_scalar_type_)
+                inner_contiguous = data_strides[-1] == scalar_size
+                if inner_contiguous and dtype_ndim > 1:
+                    inner_contiguous = data_strides[-2] == scalar_size * dtype_shape[-1]
+
+                shape = target_npshape[:-dtype_ndim] or (1,)
+                strides = data_strides if shape == data_shape else strides_from_shape(shape, dtype)
+            else:
+                shape = target_npshape or (1,)
+                strides = data_strides if shape == data_shape else strides_from_shape(shape, dtype)
 
-                self._init_from_ptr(data_ptr, dtype, shape, strides, None, device, False, None)
+            self._init_from_ptr(data_ptr, dtype, shape, strides, None, device, False, None)
 
-                # keep a ref to the source data to keep allocation alive
-                self._ref = data
-                return
-            else:
-                raise RuntimeError(
-                    f"Trying to construct a Warp array from data argument's __cuda_array_interface__ but {device} is not CUDA-capable"
-                )
+            # keep a ref to the source data to keep allocation alive
+            self._ref = data
+            return
 
         # convert input data to ndarray (handles lists, tuples, etc.) and determine dtype
         if dtype == Any:

{warp_lang-1.6.0.dist-info → warp_lang-1.6.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: warp-lang
-Version: 1.6.0
+Version: 1.6.1
 Summary: A Python framework for high-performance simulation and graphics programming
 Author-email: NVIDIA Corporation <mmacklin@nvidia.com>
 License: NVIDIA Software License
@@ -78,12 +78,24 @@ the `pip install` command, e.g.
 
 | Platform        | Install Command                                                                                                               |
 | --------------- | ----------------------------------------------------------------------------------------------------------------------------- |
-| Linux aarch64   | `pip install https://github.com/NVIDIA/warp/releases/download/v1.6.0/warp_lang-1.6.0+cu11-py3-none-manylinux2014_aarch64.whl` |
-| Linux x86-64    | `pip install https://github.com/NVIDIA/warp/releases/download/v1.6.0/warp_lang-1.6.0+cu11-py3-none-manylinux2014_x86_64.whl`  |
-| Windows x86-64  | `pip install https://github.com/NVIDIA/warp/releases/download/v1.6.0/warp_lang-1.6.0+cu11-py3-none-win_amd64.whl`             |
+| Linux aarch64   | `pip install https://github.com/NVIDIA/warp/releases/download/v1.6.1/warp_lang-1.6.1+cu11-py3-none-manylinux2014_aarch64.whl` |
+| Linux x86-64    | `pip install https://github.com/NVIDIA/warp/releases/download/v1.6.1/warp_lang-1.6.1+cu11-py3-none-manylinux2014_x86_64.whl`  |
+| Windows x86-64  | `pip install https://github.com/NVIDIA/warp/releases/download/v1.6.1/warp_lang-1.6.1+cu11-py3-none-win_amd64.whl`             |
 
 The `--force-reinstall` option may need to be used to overwrite a previous installation.
 
+### Nightly Builds
+
+Nightly builds of Warp from the `main` branch are available on the [NVIDIA Package Index](https://pypi.nvidia.com/warp-lang/).
+
+To install the latest nightly build, use the following command:
+
+```text
+pip install -U --pre warp-lang --extra-index-url=https://pypi.nvidia.com/
+```
+
+Note that the nightly builds are built with the CUDA 12 runtime and are not published for macOS.
+
 ### CUDA Requirements
 
 * Warp packages built with CUDA Toolkit 11.x require NVIDIA driver 470 or newer.
@@ -300,25 +312,13 @@ python -m warp.tests
             <td><a href="https://github.com/NVIDIA/warp/tree/main/warp/examples/optim/example_inverse_kinematics.py"><img src="https://media.githubusercontent.com/media/NVIDIA/warp/refs/heads/main/docs/img/examples/optim_inverse_kinematics.png"></a></td>
             <td><a href="https://github.com/NVIDIA/warp/tree/main/warp/examples/optim/example_spring_cage.py"><img src="https://media.githubusercontent.com/media/NVIDIA/warp/refs/heads/main/docs/img/examples/optim_spring_cage.png"></a></td>
             <td><a href="https://github.com/NVIDIA/warp/tree/main/warp/examples/optim/example_trajectory.py"><img src="https://media.githubusercontent.com/media/NVIDIA/warp/refs/heads/main/docs/img/examples/optim_trajectory.png"></a></td>
-            <td><a href="https://github.com/NVIDIA/warp/tree/main/warp/examples/optim/example_walker.py"><img src="https://media.githubusercontent.com/media/NVIDIA/warp/refs/heads/main/docs/img/examples/optim_walker.png"></a></td>
+            <td><a href="https://github.com/NVIDIA/warp/tree/main/warp/examples/optim/example_softbody_properties.py"><img src="https://media.githubusercontent.com/media/NVIDIA/warp/refs/heads/main/docs/img/examples/optim_softbody_properties.png"></a></td>
         </tr>
         <tr>
             <td align="center">inverse kinematics</td>
             <td align="center">spring cage</td>
             <td align="center">trajectory</td>
-            <td align="center">walker</td>
-        </tr>
-        <tr>
-            <td><a href="https://github.com/NVIDIA/warp/tree/main/warp/examples/optim/example_softbody_properties.py"><img src="https://media.githubusercontent.com/media/NVIDIA/warp/refs/heads/main/docs/img/examples/optim_softbody_properties.png"></a></td>
-            <td></td>
-            <td></td>
-            <td></td>
-        </tr>
-        <tr>
             <td align="center">soft body properties</td>
-            <td align="center"></td>
-            <td align="center"></td>
-            <td align="center"></td>
         </tr>
     </tbody>
 </table>
@@ -378,6 +378,23 @@ python -m warp.tests
     </tbody>
 </table>
 
+### warp/examples/tile
+
+<table>
+    <tbody>
+        <tr>
+            <td><a href="https://github.com/NVIDIA/warp/tree/main/warp/examples/tile/example_tile_mlp.py"><img src="https://media.githubusercontent.com/media/NVIDIA/warp/refs/heads/main/docs/img/examples/tile_mlp.png"></a></td>
+            <td><a href="https://github.com/NVIDIA/warp/tree/main/warp/examples/tile/example_tile_nbody.py"><img src="https://media.githubusercontent.com/media/NVIDIA/warp/refs/heads/main/docs/img/examples/tile_nbody.png"></a></td>
+            <td><a href="https://github.com/NVIDIA/warp/tree/main/warp/examples/tile/example_tile_walker.py"><img src="https://media.githubusercontent.com/media/NVIDIA/warp/refs/heads/main/docs/img/examples/tile_walker.png"></a></td>
+        </tr>
+        <tr>
+            <td align="center">mlp</td>
+            <td align="center">nbody</td>
+            <td align="center">walker</td>
+        </tr>
+    </tbody>
+</table>
+
 ## Building
 
 For developers who want to build the library themselves, the following tools are required: