Make PyTorch3D C++17 incompatible again :(

Summary: D38919607 (c4545a7) and D38858887 (06cbba2) were premature, turns out CUDA 10.2 doesn't support C++17.

Reviewed By: bottler

Differential Revision: D39156205

fbshipit-source-id: 5e2e84cc4a57d1113a915166631651d438540d56

Author: Krzysztof Chalupka

pytorch3d/csrc/iou_box3d/iou_box3d.cu

@@ -29,10 +29,10 @@ __global__ void IoUBox3DKernel(
   const size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
   const size_t stride = gridDim.x * blockDim.x;
 
-  std::array<FaceVerts, NUM_TRIS> box1_tris{};
-  std::array<FaceVerts, NUM_TRIS> box2_tris{};
-  std::array<FaceVerts, NUM_PLANES> box1_planes{};
-  std::array<FaceVerts, NUM_PLANES> box2_planes{};
+  FaceVerts box1_tris[NUM_TRIS];
+  FaceVerts box2_tris[NUM_TRIS];
+  FaceVerts box1_planes[NUM_PLANES];
+  FaceVerts box2_planes[NUM_PLANES];
 
   for (size_t i = tid; i < N * M; i += stride) {
     const size_t n = i / M; // box1 index

pytorch3d/csrc/iou_box3d/iou_utils.cuh

@@ -175,7 +175,7 @@ __device__ inline float3 FaceNormal(
   auto normal = float3();
   auto maxDist = -1;
   for (auto v1 = vertices.begin(); v1 != vertices.end() - 1; ++v1) {
-    for (auto v2 = std::next(v1); v2 != vertices.end(); ++v2) {
+    for (auto v2 = v1 + 1; v2 != vertices.end(); ++v2) {
       const auto v1ToCenter = *v1 - faceCenter;
       const auto v2ToCenter = *v2 - faceCenter;
       const auto dist = norm(cross(v1ToCenter, v2ToCenter));
@@ -472,8 +472,10 @@ __device__ inline bool IsCoplanarTriTri(
   const bool check1 = abs(dot(tri1_n, tri2_n)) > 1 - dEpsilon;
 
   // Compute most distant points
-  const auto [v1m, v2m] =
+  const auto v1mAndv2m =
       ArgMaxVerts({tri1.v0, tri1.v1, tri1.v2}, {tri2.v0, tri2.v1, tri2.v2});
+  const auto v1m = std::get<0>(v1mAndv2m);
+  const auto v2m = std::get<1>(v1mAndv2m);
 
   float3 n12m = v1m - v2m;
   n12m = n12m / fmaxf(norm(n12m), kEpsilon);
@@ -506,8 +508,10 @@ __device__ inline bool IsCoplanarTriPlane(
   const bool check1 = abs(dot(nt, normal)) > 1 - dEpsilon;
 
   // Compute most distant points
-  const auto [v1m, v2m] = ArgMaxVerts(
+  const auto v1mAndv2m = ArgMaxVerts(
       {tri.v0, tri.v1, tri.v2}, {plane.v0, plane.v1, plane.v2, plane.v3});
+  const auto v1m = std::get<0>(v1mAndv2m);
+  const auto v2m = std::get<1>(v1mAndv2m);
 
   float3 n12m = v1m - v2m;
   n12m = n12m / fmaxf(norm(n12m), kEpsilon);

setup.py

@@ -49,7 +49,7 @@ def get_extensions():
     source_cuda = glob.glob(os.path.join(extensions_dir, "**", "*.cu"), recursive=True)
     extension = CppExtension
 
-    extra_compile_args = {"cxx": []}
+    extra_compile_args = {"cxx": ["-std=c++14"]}
     define_macros = []
     include_dirs = [extensions_dir]
 
@@ -73,6 +73,8 @@ def get_extensions():
             "-D__CUDA_NO_HALF_CONVERSIONS__",
             "-D__CUDA_NO_HALF2_OPERATORS__",
         ]
+        if os.name != "nt":
+            nvcc_args.append("-std=c++14")
         if cub_home is None:
             prefix = os.environ.get("CONDA_PREFIX", None)
             if prefix is not None and os.path.isdir(prefix + "/include/cub"):