Marching Cubes C++ torch extension

Summary:
Torch C++ extension for Marching Cubes

- Add torch C++ extension for marching cubes. Observe a speed up of ~255x-324x speed up (over varying batch sizes and spatial resolutions)

- Add C++ impl in existing unit-tests.

(Note: this ignores all push blocking failures!)

Reviewed By: kjchalup

Differential Revision: D39590638

fbshipit-source-id: e44d2852a24c2c398e5ea9db20f0dfaa1817e457

pytorch3d/csrc/ext.cpp

@@ -22,6 +22,7 @@
 #include "interp_face_attrs/interp_face_attrs.h"
 #include "iou_box3d/iou_box3d.h"
 #include "knn/knn.h"
+#include "marching_cubes/marching_cubes.h"
 #include "mesh_normal_consistency/mesh_normal_consistency.h"
 #include "packed_to_padded_tensor/packed_to_padded_tensor.h"
 #include "point_mesh/point_mesh_cuda.h"
@@ -94,6 +95,9 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   // 3D IoU
   m.def("iou_box3d", &IoUBox3D);
 
+  // Marching cubes
+  m.def("marching_cubes", &MarchingCubes);
+
   // Pulsar.
 #ifdef PULSAR_LOGGING_ENABLED
   c10::ShowLogInfoToStderr();

pytorch3d/csrc/marching_cubes/marching_cubes.h

@@ -0,0 +1,39 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+#include <torch/extension.h>
+#include <tuple>
+#include <vector>
+#include "utils/pytorch3d_cutils.h"
+
+// Run Marching Cubes algorithm over a batch of volume scalar fields
+// with a pre-defined threshold and return a mesh composed of vertices
+// and faces for the mesh.
+//
+// Args:
+//    vol: FloatTensor of shape (D, H, W) giving a volume
+//    scalar grids.
+//    isolevel: isosurface value to use as the threshoold to determine whether
+//    the points are within a volume.
+//
+// Returns:
+//    vertices: List of N FloatTensors of vertices
+//    faces:    List of N LongTensors of faces
+
+// CPU implementation
+std::tuple<at::Tensor, at::Tensor> MarchingCubesCpu(
+    const at::Tensor& vol,
+    const float isolevel);
+
+// Implementation which is exposed
+inline std::tuple<at::Tensor, at::Tensor> MarchingCubes(
+    const at::Tensor& vol,
+    const float isolevel) {
+  return MarchingCubesCpu(vol.contiguous(), isolevel);
+}

pytorch3d/csrc/marching_cubes/marching_cubes_cpu.cpp

@@ -0,0 +1,115 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <torch/extension.h>
+#include <algorithm>
+#include <array>
+#include <cstring>
+#include <unordered_map>
+#include <vector>
+#include "marching_cubes/marching_cubes_utils.h"
+
+// Cpu implementation for Marching Cubes
+// Args:
+//    vol: a Tensor of size (D, H, W) corresponding to a 3D scalar field
+//    isolevel: the isosurface value to use as the threshold to determine
+//          whether points are within a volume.
+//
+// Returns:
+//    vertices: a float tensor of shape (N, 3) for positions of the mesh
+//    faces: a long tensor of shape (N, 3) for indices of the face vertices
+//
+std::tuple<at::Tensor, at::Tensor> MarchingCubesCpu(
+    const at::Tensor& vol,
+    const float isolevel) {
+  // volume shapes
+  const int D = vol.size(0);
+  const int H = vol.size(1);
+  const int W = vol.size(2);
+
+  // Create tensor accessors
+  auto vol_a = vol.accessor<float, 3>();
+  // vpair_to_edge maps a pair of vertex ids to its corresponding edge id
+  std::unordered_map<std::pair<int, int>, int64_t> vpair_to_edge;
+  // edge_id_to_v maps from an edge id to a vertex position
+  std::unordered_map<int64_t, Vertex> edge_id_to_v;
+  // uniq_edge_id: used to remove redundant edge ids
+  std::unordered_map<int64_t, int64_t> uniq_edge_id;
+  std::vector<int64_t> faces; // store face indices
+  std::vector<Vertex> verts; // store vertex positions
+  // enumerate each cell in the 3d grid
+  for (int z = 0; z < D - 1; z++) {
+    for (int y = 0; y < H - 1; y++) {
+      for (int x = 0; x < W - 1; x++) {
+        Cube cube(x, y, z, vol_a, isolevel);
+        // Cube is entirely in/out of the surface
+        if (_FACE_TABLE[cube.cubeindex][0] == -1) {
+          continue;
+        }
+        // store all boundary vertices that intersect with the edges
+        std::array<Vertex, 12> interp_points;
+        // triangle vertex IDs and positions
+        std::vector<int64_t> tri;
+        std::vector<Vertex> ps;
+
+        // Interpolate the vertices where the surface intersects with the cube
+        for (int j = 0; _FACE_TABLE[cube.cubeindex][j] != -1; j++) {
+          const int e = _FACE_TABLE[cube.cubeindex][j];
+          interp_points[e] = cube.VertexInterp(isolevel, e, vol_a);
+
+          auto vpair = cube.GetVPairFromEdge(e, W, H);
+          if (!vpair_to_edge.count(vpair)) {
+            vpair_to_edge[vpair] = vpair_to_edge.size();
+          }
+
+          int64_t edge = vpair_to_edge[vpair];
+          tri.push_back(edge);
+          ps.push_back(interp_points[e]);
+
+          // Check if the triangle face is degenerate. A triangle face
+          // is degenerate if any of the two verices share the same 3D position
+          if ((j + 1) % 3 == 0 && ps[0] != ps[1] && ps[1] != ps[2] &&
+              ps[2] != ps[0]) {
+            for (int k = 0; k < 3; k++) {
+              int v = tri[k];
+              edge_id_to_v[tri.at(k)] = ps.at(k);
+              if (!uniq_edge_id.count(v)) {
+                uniq_edge_id[v] = verts.size();
+                verts.push_back(edge_id_to_v[v]);
+              }
+              faces.push_back(uniq_edge_id[v]);
+            }
+            tri.clear();
+            ps.clear();
+          }
+        } // endif
+      } // endfor x
+    } // endfor y
+  } // endfor z
+  // Collect returning tensor
+  const int n_vertices = verts.size();
+  const int64_t n_faces = (int64_t)faces.size() / 3;
+  auto vert_tensor = torch::zeros({n_vertices, 3}, torch::kFloat);
+  auto face_tensor = torch::zeros({n_faces, 3}, torch::kInt64);
+
+  auto vert_a = vert_tensor.accessor<float, 2>();
+  for (int i = 0; i < n_vertices; i++) {
+    vert_a[i][0] = verts.at(i).x;
+    vert_a[i][1] = verts.at(i).y;
+    vert_a[i][2] = verts.at(i).z;
+  }
+
+  auto face_a = face_tensor.accessor<int64_t, 2>();
+  for (int64_t i = 0; i < n_faces; i++) {
+    face_a[i][0] = faces.at(i * 3 + 0);
+    face_a[i][1] = faces.at(i * 3 + 1);
+    face_a[i][2] = faces.at(i * 3 + 2);
+  }
+
+  return std::make_tuple(vert_tensor, face_tensor);
+}