Workaround: Building from source in Docker results in build error, requires `--std=c++17`

[MathijsdeBoer]

🐛 Bugs / Unexpected behaviors

Trying to setup a docker container with Pytorch3D, using Nvidia's Pytorch images, I kept running into a strange build error, which spat out an immense amount of red text. Digging through I found some references to certain features in the source code requiring the C++17 standard, which apparently had not been set.

To work around this issue I've adjusted the setup.py file on lines 52 and 77 to build with --std=c++17 instead, and builds are now successful. I haven't yet had a chance to test whether the GPU acceleration works, but at least it'll build in a consistent manner.

Instructions To Reproduce the Issue:

Please include the following (depending on what the issue is):

1. Any changes you made (git diff) or code you wrote

   
   diff --git a/setup.py b/setup.py
   index 54e6283..a5aecf7 100755
   --- a/setup.py
   +++ b/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": ["-std=c++14"]}
• extra_compile_args = {"cxx": ["-std=c++17"]} define_macros = [] include_dirs = [extensions_dir]

@@ -74,7 +74,7 @@ def get_extensions(): "-D__CUDA_NO_HALF2_OPERATORS__", ] if os.name != "nt":

• nvcc_args.append("-std=c++14")
• nvcc_args.append("-std=c++17") if cub_home is None: prefix = os.environ.get("CONDA_PREFIX", None) if prefix is not None and os.path.isdir(prefix + "/include/cub"):

Dockerfile:

FROM nvcr.io/nvidia/pytorch:23.09-py3

RUN apt-get update \
    && apt-get upgrade -y\
    && apt-get install -y wget \
    build-essential \
    ca-certificates \
    git \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/* \
    && gcc --version

COPY ./pytorch3d /pytorch3d
RUN cd /pytorch3d \
    && pip install .

2. The exact command(s) you ran: In order to create the following file setup:

   ./
   |- pytorch3d (cloned from git, tag v0.7.4)
   |- Dockerfile (referenced above)

   $ git clone --depth 1 --branch v0.7.4 https://github.com/facebookresearch/pytorch3d.git
   $ docker build -t pytorch3d:latest .

3. What you observed (including the full logs):

   The full error log is way too long to include here, in fact it didn't even fit in my terminal buffer.

I am not in the loop enough to know why the setup.py calls out the c++14 standard, while the C++ code seems to require c++17 features, nor do I know if this is kept at c++14 for a particular reason. But if this is all it takes to get a successful build in a Docker image, I'll post it here too in case it ends up being useful for someone else.

[bottler]

setup.py is now updated to use c++17.