warp-lang 0.10.1__py3-none-win_amd64.whl → 0.11.0__py3-none-win_amd64.whl

warp/fem/geometry/tetmesh.py

@@ -0,0 +1,840 @@
+from typing import Optional
+
+import warp as wp
+from warp.fem.cache import (
+    TemporaryStore,
+    borrow_temporary,
+    borrow_temporary_like,
+    cached_arg_value,
+)
+from warp.fem.types import (
+    NULL_ELEMENT_INDEX,
+    OUTSIDE,
+    Coords,
+    ElementIndex,
+    Sample,
+    make_free_sample,
+)
+
+from .closest_point import project_on_tet_at_origin
+from .element import Tetrahedron, Triangle
+from .geometry import Geometry
+
+
+@wp.struct
+class TetmeshCellArg:
+    tet_vertex_indices: wp.array2d(dtype=int)
+    positions: wp.array(dtype=wp.vec3)
+
+    # for neighbor cell lookup
+    vertex_tet_offsets: wp.array(dtype=int)
+    vertex_tet_indices: wp.array(dtype=int)
+
+    # for transforming reference gradient
+    deformation_gradients: wp.array(dtype=wp.mat33f)
+
+
+@wp.struct
+class TetmeshSideArg:
+    cell_arg: TetmeshCellArg
+    face_vertex_indices: wp.array(dtype=wp.vec3i)
+    face_tet_indices: wp.array(dtype=wp.vec2i)
+
+
+_mat32 = wp.mat(shape=(3, 2), dtype=float)
+
+
+class Tetmesh(Geometry):
+    """Tetrahedral mesh geometry"""
+
+    dimension = 3
+
+    def __init__(
+        self, tet_vertex_indices: wp.array, positions: wp.array, temporary_store: Optional[TemporaryStore] = None
+    ):
+        """
+        Constructs a tetrahedral mesh.
+
+        Args:
+            tet_vertex_indices: warp array of shape (num_tets, 4) containing vertex indices for each tet
+            positions: warp array of shape (num_vertices, 3) containing 3d position for each vertex
+            temporary_store: shared pool from which to allocate temporary arrays
+        """
+
+        self.tet_vertex_indices = tet_vertex_indices
+        self.positions = positions
+
+        self._face_vertex_indices: wp.array = None
+        self._face_tet_indices: wp.array = None
+        self._vertex_tet_offsets: wp.array = None
+        self._vertex_tet_indices: wp.array = None
+        self._tet_edge_indices: wp.array = None
+        self._edge_count = 0
+        self._build_topology(temporary_store)
+
+        self._deformation_gradients: wp.array = None
+        self._compute_deformation_gradients()
+
+    def cell_count(self):
+        return self.tet_vertex_indices.shape[0]
+
+    def vertex_count(self):
+        return self.positions.shape[0]
+
+    def side_count(self):
+        return self._face_vertex_indices.shape[0]
+
+    def edge_count(self):
+        if self._tet_edge_indices is None:
+            self._compute_tet_edges()
+        return self._edge_count
+
+    def boundary_side_count(self):
+        return self._boundary_face_indices.shape[0]
+
+    def reference_cell(self) -> Tetrahedron:
+        return Tetrahedron()
+
+    def reference_side(self) -> Triangle:
+        return Triangle()
+
+    @property
+    def tet_edge_indices(self) -> wp.array:
+        if self._tet_edge_indices is None:
+            self._compute_tet_edges()
+        return self._tet_edge_indices
+
+    @property
+    def face_tet_indices(self) -> wp.array:
+        return self._face_tet_indices
+
+    @property
+    def face_vertex_indices(self) -> wp.array:
+        return self._face_vertex_indices
+
+    CellArg = TetmeshCellArg
+    SideArg = TetmeshSideArg
+
+    @wp.struct
+    class SideIndexArg:
+        boundary_face_indices: wp.array(dtype=int)
+
+    # Geometry device interface
+
+    @cached_arg_value
+    def cell_arg_value(self, device) -> CellArg:
+        args = self.CellArg()
+
+        args.tet_vertex_indices = self.tet_vertex_indices.to(device)
+        args.positions = self.positions.to(device)
+        args.vertex_tet_offsets = self._vertex_tet_offsets.to(device)
+        args.vertex_tet_indices = self._vertex_tet_indices.to(device)
+        args.deformation_gradients = self._deformation_gradients.to(device)
+
+        return args
+
+    @wp.func
+    def cell_position(args: CellArg, s: Sample):
+        tet_idx = args.tet_vertex_indices[s.element_index]
+        w0 = 1.0 - s.element_coords[0] - s.element_coords[1] - s.element_coords[2]
+        return (
+            w0 * args.positions[tet_idx[0]]
+            + s.element_coords[0] * args.positions[tet_idx[1]]
+            + s.element_coords[1] * args.positions[tet_idx[2]]
+            + s.element_coords[2] * args.positions[tet_idx[3]]
+        )
+
+    @wp.func
+    def cell_deformation_gradient(args: CellArg, s: Sample):
+        return args.deformation_gradients[s.element_index]
+
+    @wp.func
+    def cell_inverse_deformation_gradient(args: CellArg, s: Sample):
+        return wp.inverse(args.deformation_gradients[s.element_index])
+
+    @wp.func
+    def _project_on_tet(args: CellArg, pos: wp.vec3, tet_index: int):
+        p0 = args.positions[args.tet_vertex_indices[tet_index, 0]]
+
+        q = pos - p0
+        e1 = args.positions[args.tet_vertex_indices[tet_index, 1]] - p0
+        e2 = args.positions[args.tet_vertex_indices[tet_index, 2]] - p0
+        e3 = args.positions[args.tet_vertex_indices[tet_index, 3]] - p0
+
+        dist, coords = project_on_tet_at_origin(q, e1, e2, e3)
+        return dist, coords
+
+    @wp.func
+    def cell_lookup(args: CellArg, pos: wp.vec3, guess: Sample):
+        closest_tet = int(NULL_ELEMENT_INDEX)
+        closest_coords = Coords(OUTSIDE)
+        closest_dist = float(1.0e8)
+
+        for v in range(4):
+            vtx = args.tet_vertex_indices[guess.element_index, v]
+            tet_beg = args.vertex_tet_offsets[vtx]
+            tet_end = args.vertex_tet_offsets[vtx + 1]
+
+            for t in range(tet_beg, tet_end):
+                tet = args.vertex_tet_indices[t]
+                dist, coords = Tetmesh._project_on_tet(args, pos, tet)
+                if dist <= closest_dist:
+                    closest_dist = dist
+                    closest_tet = tet
+                    closest_coords = coords
+
+        return make_free_sample(closest_tet, closest_coords)
+
+    @wp.func
+    def cell_measure(args: CellArg, s: Sample):
+        return wp.abs(wp.determinant(args.deformation_gradients[s.element_index])) / 6.0
+
+    @wp.func
+    def cell_measure_ratio(args: CellArg, s: Sample):
+        return 1.0
+
+    @wp.func
+    def cell_normal(args: CellArg, s: Sample):
+        return wp.vec3(0.0)
+
+    @cached_arg_value
+    def side_index_arg_value(self, device) -> SideIndexArg:
+        args = self.SideIndexArg()
+
+        args.boundary_face_indices = self._boundary_face_indices.to(device)
+
+        return args
+
+    @wp.func
+    def boundary_side_index(args: SideIndexArg, boundary_side_index: int):
+        """Boundary side to side index"""
+
+        return args.boundary_face_indices[boundary_side_index]
+
+    @cached_arg_value
+    def side_arg_value(self, device) -> CellArg:
+        args = self.SideArg()
+
+        args.cell_arg = self.cell_arg_value(device)
+        args.face_vertex_indices = self._face_vertex_indices.to(device)
+        args.face_tet_indices = self._face_tet_indices.to(device)
+
+        return args
+
+    @wp.func
+    def side_position(args: SideArg, s: Sample):
+        face_idx = args.face_vertex_indices[s.element_index]
+        return (
+            s.element_coords[0] * args.cell_arg.positions[face_idx[0]]
+            + s.element_coords[1] * args.cell_arg.positions[face_idx[1]]
+            + s.element_coords[2] * args.cell_arg.positions[face_idx[2]]
+        )
+
+    @wp.func
+    def _side_vecs(args: SideArg, side_index: ElementIndex):
+        face_idx = args.face_vertex_indices[side_index]
+        v0 = args.cell_arg.positions[face_idx[0]]
+        v1 = args.cell_arg.positions[face_idx[1]]
+        v2 = args.cell_arg.positions[face_idx[2]]
+
+        return v1 - v0, v2 - v0
+
+    @wp.func
+    def side_deformation_gradient(args: SideArg, s: Sample):
+        e1, e2 = Tetmesh._side_vecs(args, s.element_index)
+        return _mat32(e1, e2)
+
+    @wp.func
+    def side_inner_inverse_deformation_gradient(args: SideArg, s: Sample):
+        cell_index = Tetmesh.side_inner_cell_index(args, s.element_index)
+        return wp.inverse(args.cell_arg.deformation_gradients[cell_index])
+
+    @wp.func
+    def side_outer_inverse_deformation_gradient(args: SideArg, s: Sample):
+        cell_index = Tetmesh.side_outer_cell_index(args, s.element_index)
+        return wp.inverse(args.cell_arg.deformation_gradients[cell_index])
+
+    @wp.func
+    def side_measure(args: SideArg, s: Sample):
+        e1, e2 = Tetmesh._side_vecs(args, s.element_index)
+        return 0.5 * wp.length(wp.cross(e1, e2))
+
+    @wp.func
+    def side_measure_ratio(args: SideArg, s: Sample):
+        inner = Tetmesh.side_inner_cell_index(args, s.element_index)
+        outer = Tetmesh.side_outer_cell_index(args, s.element_index)
+        return Tetmesh.side_measure(args, s) / wp.min(
+            Tetmesh.cell_measure(args.cell_arg, make_free_sample(inner, Coords())),
+            Tetmesh.cell_measure(args.cell_arg, make_free_sample(outer, Coords())),
+        )
+
+    @wp.func
+    def side_normal(args: SideArg, s: Sample):
+        e1, e2 = Tetmesh._side_vecs(args, s.element_index)
+        return wp.normalize(wp.cross(e1, e2))
+
+    @wp.func
+    def side_inner_cell_index(arg: SideArg, side_index: ElementIndex):
+        return arg.face_tet_indices[side_index][0]
+
+    @wp.func
+    def side_outer_cell_index(arg: SideArg, side_index: ElementIndex):
+        return arg.face_tet_indices[side_index][1]
+
+    @wp.func
+    def face_to_tet_coords(args: SideArg, side_index: ElementIndex, tet_index: ElementIndex, side_coords: Coords):
+        fvi = args.face_vertex_indices[side_index]
+
+        tv1 = args.cell_arg.tet_vertex_indices[tet_index, 1]
+        tv2 = args.cell_arg.tet_vertex_indices[tet_index, 2]
+        tv3 = args.cell_arg.tet_vertex_indices[tet_index, 3]
+
+        c1 = float(0.0)
+        c2 = float(0.0)
+        c3 = float(0.0)
+
+        for k in range(3):
+            if tv1 == fvi[k]:
+                c1 = side_coords[k]
+            elif tv2 == fvi[k]:
+                c2 = side_coords[k]
+            elif tv3 == fvi[k]:
+                c3 = side_coords[k]
+
+        return Coords(c1, c2, c3)
+
+    @wp.func
+    def side_inner_cell_coords(args: SideArg, side_index: ElementIndex, side_coords: Coords):
+        inner_cell_index = Tetmesh.side_inner_cell_index(args, side_index)
+        return Tetmesh.face_to_tet_coords(args, side_index, inner_cell_index, side_coords)
+
+    @wp.func
+    def side_outer_cell_coords(args: SideArg, side_index: ElementIndex, side_coords: Coords):
+        outer_cell_index = Tetmesh.side_outer_cell_index(args, side_index)
+        return Tetmesh.face_to_tet_coords(args, side_index, outer_cell_index, side_coords)
+
+    @wp.func
+    def side_from_cell_coords(args: SideArg, side_index: ElementIndex, tet_index: ElementIndex, tet_coords: Coords):
+        fvi = args.face_vertex_indices[side_index]
+
+        tv1 = args.cell_arg.tet_vertex_indices[tet_index, 1]
+        tv2 = args.cell_arg.tet_vertex_indices[tet_index, 2]
+        tv3 = args.cell_arg.tet_vertex_indices[tet_index, 3]
+
+        if tv1 == fvi[0]:
+            c0 = tet_coords[0]
+        elif tv2 == fvi[0]:
+            c0 = tet_coords[1]
+        elif tv3 == fvi[0]:
+            c0 = tet_coords[2]
+        else:
+            c0 = 1.0 - tet_coords[0] - tet_coords[1] - tet_coords[2]
+
+        if tv1 == fvi[1]:
+            c1 = tet_coords[0]
+        elif tv2 == fvi[1]:
+            c1 = tet_coords[1]
+        elif tv3 == fvi[1]:
+            c1 = tet_coords[2]
+        else:
+            c1 = 1.0 - tet_coords[0] - tet_coords[1] - tet_coords[2]
+
+        if tv1 == fvi[2]:
+            c2 = tet_coords[0]
+        elif tv2 == fvi[2]:
+            c2 = tet_coords[1]
+        elif tv3 == fvi[2]:
+            c2 = tet_coords[2]
+        else:
+            c2 = 1.0 - tet_coords[0] - tet_coords[1] - tet_coords[2]
+
+        return wp.select(c0 + c1 + c2 > 0.999, Coords(OUTSIDE), Coords(c0, c1, c2))
+
+    @wp.func
+    def side_to_cell_arg(side_arg: SideArg):
+        return side_arg.cell_arg
+
+    def _build_topology(self, temporary_store: TemporaryStore):
+        from warp.fem.utils import compress_node_indices, masked_indices
+        from warp.utils import array_scan
+
+        device = self.tet_vertex_indices.device
+
+        vertex_tet_offsets, vertex_tet_indices, _, __ = compress_node_indices(
+            self.vertex_count(), self.tet_vertex_indices, temporary_store=temporary_store
+        )
+        self._vertex_tet_offsets = vertex_tet_offsets.detach()
+        self._vertex_tet_indices = vertex_tet_indices.detach()
+
+        vertex_start_face_count = borrow_temporary(temporary_store, dtype=int, device=device, shape=self.vertex_count())
+        vertex_start_face_count.array.zero_()
+        vertex_start_face_offsets = borrow_temporary_like(vertex_start_face_count, temporary_store=temporary_store)
+
+        vertex_face_other_vs = borrow_temporary(
+            temporary_store, dtype=wp.vec2i, device=device, shape=(4 * self.cell_count())
+        )
+        vertex_face_tets = borrow_temporary(temporary_store, dtype=int, device=device, shape=(4 * self.cell_count(), 2))
+
+        # Count face edges starting at each vertex
+        wp.launch(
+            kernel=Tetmesh._count_starting_faces_kernel,
+            device=device,
+            dim=self.cell_count(),
+            inputs=[self.tet_vertex_indices, vertex_start_face_count.array],
+        )
+
+        array_scan(in_array=vertex_start_face_count.array, out_array=vertex_start_face_offsets.array, inclusive=False)
+
+        # Count number of unique edges (deduplicate across faces)
+        vertex_unique_face_count = vertex_start_face_count
+        wp.launch(
+            kernel=Tetmesh._count_unique_starting_faces_kernel,
+            device=device,
+            dim=self.vertex_count(),
+            inputs=[
+                self._vertex_tet_offsets,
+                self._vertex_tet_indices,
+                self.tet_vertex_indices,
+                vertex_start_face_offsets.array,
+                vertex_unique_face_count.array,
+                vertex_face_other_vs.array,
+                vertex_face_tets.array,
+            ],
+        )
+
+        vertex_unique_face_offsets = borrow_temporary_like(vertex_start_face_offsets, temporary_store=temporary_store)
+        array_scan(in_array=vertex_start_face_count.array, out_array=vertex_unique_face_offsets.array, inclusive=False)
+
+        # Get back edge count to host
+        if device.is_cuda:
+            face_count = borrow_temporary(temporary_store, shape=(1,), dtype=int, device="cpu", pinned=True)
+            # Last vertex will not own any edge, so its count will be zero; just fetching last prefix count is ok
+            wp.copy(
+                dest=face_count.array, src=vertex_unique_face_offsets.array, src_offset=self.vertex_count() - 1, count=1
+            )
+            wp.synchronize_stream(wp.get_stream(device))
+            face_count = int(face_count.array.numpy()[0])
+        else:
+            face_count = int(vertex_unique_face_offsets.array.numpy()[self.vertex_count() - 1])
+
+        self._face_vertex_indices = wp.empty(shape=(face_count,), dtype=wp.vec3i, device=device)
+        self._face_tet_indices = wp.empty(shape=(face_count,), dtype=wp.vec2i, device=device)
+
+        boundary_mask = borrow_temporary(temporary_store, shape=(face_count,), dtype=int, device=device)
+
+        # Compress edge data
+        wp.launch(
+            kernel=Tetmesh._compress_faces_kernel,
+            device=device,
+            dim=self.vertex_count(),
+            inputs=[
+                vertex_start_face_offsets.array,
+                vertex_unique_face_offsets.array,
+                vertex_unique_face_count.array,
+                vertex_face_other_vs.array,
+                vertex_face_tets.array,
+                self._face_vertex_indices,
+                self._face_tet_indices,
+                boundary_mask.array,
+            ],
+        )
+
+        vertex_start_face_offsets.release()
+        vertex_unique_face_offsets.release()
+        vertex_unique_face_count.release()
+        vertex_face_other_vs.release()
+        vertex_face_tets.release()
+
+        # Flip normals if necessary
+        wp.launch(
+            kernel=Tetmesh._flip_face_normals,
+            device=device,
+            dim=self.side_count(),
+            inputs=[self._face_vertex_indices, self._face_tet_indices, self.tet_vertex_indices, self.positions],
+        )
+
+        boundary_face_indices, _ = masked_indices(boundary_mask.array)
+        self._boundary_face_indices = boundary_face_indices.detach()
+
+    def _compute_tet_edges(self, temporary_store: Optional[TemporaryStore] = None):
+        from warp.utils import array_scan
+
+        device = self.tet_vertex_indices.device
+
+        vertex_start_edge_count = borrow_temporary(temporary_store, dtype=int, device=device, shape=self.vertex_count())
+        vertex_start_edge_count.array.zero_()
+        vertex_start_edge_offsets = borrow_temporary_like(vertex_start_edge_count, temporary_store=temporary_store)
+
+        vertex_edge_ends = borrow_temporary(temporary_store, dtype=int, device=device, shape=(6 * self.cell_count()))
+
+        # Count face edges starting at each vertex
+        wp.launch(
+            kernel=Tetmesh._count_starting_edges_kernel,
+            device=device,
+            dim=self.cell_count(),
+            inputs=[self.tet_vertex_indices, vertex_start_edge_count.array],
+        )
+
+        array_scan(in_array=vertex_start_edge_count.array, out_array=vertex_start_edge_offsets.array, inclusive=False)
+
+        # Count number of unique edges (deduplicate across faces)
+        vertex_unique_edge_count = vertex_start_edge_count
+        wp.launch(
+            kernel=Tetmesh._count_unique_starting_edges_kernel,
+            device=device,
+            dim=self.vertex_count(),
+            inputs=[
+                self._vertex_tet_offsets,
+                self._vertex_tet_indices,
+                self.tet_vertex_indices,
+                vertex_start_edge_offsets.array,
+                vertex_unique_edge_count.array,
+                vertex_edge_ends.array,
+            ],
+        )
+
+        vertex_unique_edge_offsets = borrow_temporary_like(
+            vertex_start_edge_offsets.array, temporary_store=temporary_store
+        )
+        array_scan(in_array=vertex_start_edge_count.array, out_array=vertex_unique_edge_offsets.array, inclusive=False)
+
+        # Get back edge count to host
+        if device.is_cuda:
+            edge_count = borrow_temporary(temporary_store, shape=(1,), dtype=int, device="cpu", pinned=True)
+            # Last vertex will not own any edge, so its count will be zero; just fetching last prefix count is ok
+            wp.copy(
+                dest=edge_count.array,
+                src=vertex_unique_edge_offsets.array,
+                src_offset=self.vertex_count() - 1,
+                count=1,
+            )
+            wp.synchronize_stream(wp.get_stream(device))
+            self._edge_count = int(edge_count.array.numpy()[0])
+        else:
+            self._edge_count = int(vertex_unique_edge_offsets.array.numpy()[self.vertex_count() - 1])
+
+        self._tet_edge_indices = wp.empty(
+            dtype=int, device=self.tet_vertex_indices.device, shape=(self.cell_count(), 6)
+        )
+
+        # Compress edge data
+        wp.launch(
+            kernel=Tetmesh._compress_edges_kernel,
+            device=device,
+            dim=self.vertex_count(),
+            inputs=[
+                self._vertex_tet_offsets,
+                self._vertex_tet_indices,
+                self.tet_vertex_indices,
+                vertex_start_edge_offsets.array,
+                vertex_unique_edge_offsets.array,
+                vertex_unique_edge_count.array,
+                vertex_edge_ends.array,
+                self._tet_edge_indices,
+            ],
+        )
+
+        vertex_start_edge_offsets.release()
+        vertex_unique_edge_offsets.release()
+        vertex_unique_edge_count.release()
+        vertex_edge_ends.release()
+
+    def _compute_deformation_gradients(self):
+        self._deformation_gradients = wp.empty(dtype=wp.mat33f, device=self.positions.device, shape=(self.cell_count()))
+
+        wp.launch(
+            kernel=Tetmesh._compute_deformation_gradients_kernel,
+            dim=self._deformation_gradients.shape,
+            device=self._deformation_gradients.device,
+            inputs=[self.tet_vertex_indices, self.positions, self._deformation_gradients],
+        )
+
+    @wp.kernel
+    def _count_starting_faces_kernel(
+        tet_vertex_indices: wp.array2d(dtype=int), vertex_start_face_count: wp.array(dtype=int)
+    ):
+        t = wp.tid()
+        for k in range(4):
+            vi = wp.vec3i(
+                tet_vertex_indices[t, k], tet_vertex_indices[t, (k + 1) % 4], tet_vertex_indices[t, (k + 2) % 4]
+            )
+            vm = wp.min(vi)
+
+            for i in range(3):
+                if vm == vi[i]:
+                    wp.atomic_add(vertex_start_face_count, vm, 1)
+
+    @wp.func
+    def _find_face(
+        needle: wp.vec2i,
+        values: wp.array(dtype=wp.vec2i),
+        beg: int,
+        end: int,
+    ):
+        for i in range(beg, end):
+            if values[i] == needle:
+                return i
+
+        return -1
+
+    @wp.kernel
+    def _count_unique_starting_faces_kernel(
+        vertex_tet_offsets: wp.array(dtype=int),
+        vertex_tet_indices: wp.array(dtype=int),
+        tet_vertex_indices: wp.array2d(dtype=int),
+        vertex_start_face_offsets: wp.array(dtype=int),
+        vertex_start_face_count: wp.array(dtype=int),
+        face_other_vs: wp.array(dtype=wp.vec2i),
+        face_tets: wp.array2d(dtype=int),
+    ):
+        v = wp.tid()
+
+        face_beg = vertex_start_face_offsets[v]
+
+        tet_beg = vertex_tet_offsets[v]
+        tet_end = vertex_tet_offsets[v + 1]
+
+        face_cur = face_beg
+
+        for tet in range(tet_beg, tet_end):
+            t = vertex_tet_indices[tet]
+
+            for k in range(4):
+                vi = wp.vec3i(
+                    tet_vertex_indices[t, k], tet_vertex_indices[t, (k + 1) % 4], tet_vertex_indices[t, (k + 2) % 4]
+                )
+                min_v = wp.min(vi)
+
+                if v == min_v:
+                    max_v = wp.max(vi)
+                    mid_v = vi[0] + vi[1] + vi[2] - min_v - max_v
+                    other_v = wp.vec2i(mid_v, max_v)
+
+                    # Check if other_v has been seen
+                    seen_idx = Tetmesh._find_face(other_v, face_other_vs, face_beg, face_cur)
+
+                    if seen_idx == -1:
+                        face_other_vs[face_cur] = other_v
+                        face_tets[face_cur, 0] = t
+                        face_tets[face_cur, 1] = t
+                        face_cur += 1
+                    else:
+                        face_tets[seen_idx, 1] = t
+
+        vertex_start_face_count[v] = face_cur - face_beg
+
+    @wp.kernel
+    def _compress_faces_kernel(
+        vertex_start_face_offsets: wp.array(dtype=int),
+        vertex_unique_face_offsets: wp.array(dtype=int),
+        vertex_unique_face_count: wp.array(dtype=int),
+        uncompressed_face_other_vs: wp.array(dtype=wp.vec2i),
+        uncompressed_face_tets: wp.array2d(dtype=int),
+        face_vertex_indices: wp.array(dtype=wp.vec3i),
+        face_tet_indices: wp.array(dtype=wp.vec2i),
+        boundary_mask: wp.array(dtype=int),
+    ):
+        v = wp.tid()
+
+        start_beg = vertex_start_face_offsets[v]
+        unique_beg = vertex_unique_face_offsets[v]
+        unique_count = vertex_unique_face_count[v]
+
+        for f in range(unique_count):
+            src_index = start_beg + f
+            face_index = unique_beg + f
+
+            face_vertex_indices[face_index] = wp.vec3i(
+                v,
+                uncompressed_face_other_vs[src_index][0],
+                uncompressed_face_other_vs[src_index][1],
+            )
+
+            t0 = uncompressed_face_tets[src_index, 0]
+            t1 = uncompressed_face_tets[src_index, 1]
+            face_tet_indices[face_index] = wp.vec2i(t0, t1)
+            if t0 == t1:
+                boundary_mask[face_index] = 1
+            else:
+                boundary_mask[face_index] = 0
+
+    @wp.kernel
+    def _flip_face_normals(
+        face_vertex_indices: wp.array(dtype=wp.vec3i),
+        face_tet_indices: wp.array(dtype=wp.vec2i),
+        tet_vertex_indices: wp.array2d(dtype=int),
+        positions: wp.array(dtype=wp.vec3),
+    ):
+        e = wp.tid()
+
+        tet = face_tet_indices[e][0]
+
+        tet_vidx = tet_vertex_indices[tet]
+        face_vidx = face_vertex_indices[e]
+
+        tet_centroid = (
+            positions[tet_vidx[0]] + positions[tet_vidx[1]] + positions[tet_vidx[2]] + positions[tet_vidx[3]]
+        ) / 4.0
+
+        v0 = positions[face_vidx[0]]
+        v1 = positions[face_vidx[1]]
+        v2 = positions[face_vidx[2]]
+
+        face_center = (v1 + v0 + v2) / 3.0
+        face_normal = wp.cross(v1 - v0, v2 - v0)
+
+        # if face normal points toward first tet centroid, flip indices
+        if wp.dot(tet_centroid - face_center, face_normal) > 0.0:
+            face_vertex_indices[e] = wp.vec3i(face_vidx[0], face_vidx[2], face_vidx[1])
+
+    @wp.kernel
+    def _count_starting_edges_kernel(
+        tri_vertex_indices: wp.array2d(dtype=int), vertex_start_edge_count: wp.array(dtype=int)
+    ):
+        t = wp.tid()
+        for k in range(3):
+            v0 = tri_vertex_indices[t, k]
+            v1 = tri_vertex_indices[t, (k + 1) % 3]
+
+            if v0 < v1:
+                wp.atomic_add(vertex_start_edge_count, v0, 1)
+            else:
+                wp.atomic_add(vertex_start_edge_count, v1, 1)
+
+        for k in range(3):
+            v0 = tri_vertex_indices[t, k]
+            v1 = tri_vertex_indices[t, 3]
+
+            if v0 < v1:
+                wp.atomic_add(vertex_start_edge_count, v0, 1)
+            else:
+                wp.atomic_add(vertex_start_edge_count, v1, 1)
+
+    @wp.func
+    def _find_edge(
+        needle: int,
+        values: wp.array(dtype=int),
+        beg: int,
+        end: int,
+    ):
+        for i in range(beg, end):
+            if values[i] == needle:
+                return i
+
+        return -1
+
+    @wp.kernel
+    def _count_unique_starting_edges_kernel(
+        vertex_tet_offsets: wp.array(dtype=int),
+        vertex_tet_indices: wp.array(dtype=int),
+        tet_vertex_indices: wp.array2d(dtype=int),
+        vertex_start_edge_offsets: wp.array(dtype=int),
+        vertex_start_edge_count: wp.array(dtype=int),
+        edge_ends: wp.array(dtype=int),
+    ):
+        v = wp.tid()
+
+        edge_beg = vertex_start_edge_offsets[v]
+
+        tet_beg = vertex_tet_offsets[v]
+        tet_end = vertex_tet_offsets[v + 1]
+
+        edge_cur = edge_beg
+
+        for tet in range(tet_beg, tet_end):
+            t = vertex_tet_indices[tet]
+
+            for k in range(3):
+                v0 = tet_vertex_indices[t, k]
+                v1 = tet_vertex_indices[t, (k + 1) % 3]
+
+                if v == wp.min(v0, v1):
+                    other_v = wp.max(v0, v1)
+                    if Tetmesh._find_edge(other_v, edge_ends, edge_beg, edge_cur) == -1:
+                        edge_ends[edge_cur] = other_v
+                        edge_cur += 1
+
+            for k in range(3):
+                v0 = tet_vertex_indices[t, k]
+                v1 = tet_vertex_indices[t, 3]
+
+                if v == wp.min(v0, v1):
+                    other_v = wp.max(v0, v1)
+                    if Tetmesh._find_edge(other_v, edge_ends, edge_beg, edge_cur) == -1:
+                        edge_ends[edge_cur] = other_v
+                        edge_cur += 1
+
+        vertex_start_edge_count[v] = edge_cur - edge_beg
+
+    @wp.kernel
+    def _compress_edges_kernel(
+        vertex_tet_offsets: wp.array(dtype=int),
+        vertex_tet_indices: wp.array(dtype=int),
+        tet_vertex_indices: wp.array2d(dtype=int),
+        vertex_start_edge_offsets: wp.array(dtype=int),
+        vertex_unique_edge_offsets: wp.array(dtype=int),
+        vertex_unique_edge_count: wp.array(dtype=int),
+        uncompressed_edge_ends: wp.array(dtype=int),
+        tet_edge_indices: wp.array2d(dtype=int),
+    ):
+        v = wp.tid()
+
+        uncompressed_beg = vertex_start_edge_offsets[v]
+
+        unique_beg = vertex_unique_edge_offsets[v]
+        unique_count = vertex_unique_edge_count[v]
+
+        tet_beg = vertex_tet_offsets[v]
+        tet_end = vertex_tet_offsets[v + 1]
+
+        for tet in range(tet_beg, tet_end):
+            t = vertex_tet_indices[tet]
+
+            for k in range(3):
+                v0 = tet_vertex_indices[t, k]
+                v1 = tet_vertex_indices[t, (k + 1) % 3]
+
+                if v == wp.min(v0, v1):
+                    other_v = wp.max(v0, v1)
+                    edge_id = (
+                        Tetmesh._find_edge(
+                            other_v, uncompressed_edge_ends, uncompressed_beg, uncompressed_beg + unique_count
+                        )
+                        - uncompressed_beg
+                        + unique_beg
+                    )
+                    tet_edge_indices[t][k] = edge_id
+
+            for k in range(3):
+                v0 = tet_vertex_indices[t, k]
+                v1 = tet_vertex_indices[t, 3]
+
+                if v == wp.min(v0, v1):
+                    other_v = wp.max(v0, v1)
+                    edge_id = (
+                        Tetmesh._find_edge(
+                            other_v, uncompressed_edge_ends, uncompressed_beg, uncompressed_beg + unique_count
+                        )
+                        - uncompressed_beg
+                        + unique_beg
+                    )
+                    tet_edge_indices[t][k + 3] = edge_id
+
+    @wp.kernel
+    def _compute_deformation_gradients_kernel(
+        tet_vertex_indices: wp.array2d(dtype=int),
+        positions: wp.array(dtype=wp.vec3f),
+        transforms: wp.array(dtype=wp.mat33f),
+    ):
+        t = wp.tid()
+
+        p0 = positions[tet_vertex_indices[t, 0]]
+        p1 = positions[tet_vertex_indices[t, 1]]
+        p2 = positions[tet_vertex_indices[t, 2]]
+        p3 = positions[tet_vertex_indices[t, 3]]
+
+        e1 = p1 - p0
+        e2 = p2 - p0
+        e3 = p3 - p0
+
+        transforms[t] = wp.mat33(e1, e2, e3)