Package request: Open3D

[rowanG077]

Project description Open3D is a modern fully featured 3D processing(Pointclouds, meshes etc) environment. I tried to package it myself but quickly ran into my limited nix knowledge. :(

Metadata

• homepage URL: http://www.open3d.org/
• source URL: https://github.com/intel-isl/Open3D
• license: mit
• platforms: linux, OSX

[rowanG077]

The entire day I have attempted to package it. I'm stuck on trying to resolve an include for ``. Open3D depends on filament which depends on google test which depends on libcxxabi which apparently depends on libcxx. But that last dependency link goes wrong and I don't know how to fix it.

I have cloned the open3D repo added the nix files and patched some things. It's located here: https://github.com/rowanG077/Open3D/tree/v0.12.0-nix

Any help would be greatly appreciated!

[rowanG077]

I have isolated the libcxxabi include issue to the following expression:

clangStdenv.mkDerivation {
    name = "filament";
    src = fetchFromGitHub {
      owner = "intel-isl";
      repo = "filament";
      rev = "13ad8e25289cb173a4f8e71e85cb4e0d026eacdc";
      sha256 = "1pq7cfws4k8havcb589mw8smwgl3hnkw7ignrmw8n0s2wp2zjgnv";
      fetchSubmodules = true;
    };

    buildInputs = [ python ninja xorg.libXi libGLU cmake libcxx libcxxabi ];
  };

[stale[bot]]

I marked this as stale due to inactivity. → More info

[aaronchall]

I hope this will get support in the future. it looks like a pretty exciting new project: http://www.open3d.org/

[lapp0]

@rowanG077 this expression works for me. Would you be willing to clean it up so it can be packaged?

(borrows some code from https://github.com/mipmip/nixos/blob/45dea7930eafa18b31424be0e2d850e7a3c7504a/pkgs/open3d/default.nix)

        pkgs = nixpkgs.legacyPackages.${system};
        pkgsUnfree = import nixpkgs {
          inherit system;
          config = { allowUnfree = true; };
        };
        pkgsUnfreeNoCheck = import nixpkgs {
          inherit system;
          config = { allowUnfree = true; doCheck = false; };
        };
        python = pkgs.python39;
        pythonPackages = python.pkgs;

        pytorch-cuda = pkgsUnfreeNoCheck.python39Packages.pytorch.override {
          cudaSupport = true;
        };
        libtorch-cuda = pkgs.libtorch-bin.override { cudaSupport = true; };

        open3d = pythonPackages.buildPythonPackage rec {
          pname = "open3d";
          version = "0.14.1";
          format = "wheel";

          src = pythonPackages.fetchPypi {
            inherit pname version format;
            dist = "cp39";
            python = "cp39";
            abi = "cp39";
            platform = "manylinux_2_27_x86_64";
            sha256 = "sha256-rgk+O4FDXIIQbCs5zpXR5lV8x7cCekiQfaoD/83JwCA=";
          };

          patchPhase = ''
            ${pkgs.unzip}/bin/unzip ./dist/open3d-0.14.1-cp39-cp39-manylinux_2_27_x86_64.whl -d tmp
            rm ./dist/open3d-0.14.1-cp39-cp39-manylinux_2_27_x86_64.whl
            sed -i 's/sklearn/scikit-learn/g' tmp/open3d-0.14.1.dist-info/METADATA
            cd tmp
            ${pkgs.zip}/bin/zip -0 -r ../dist/open3d-0.14.1-cp39-cp39-manylinux_2_27_x86_64.whl ./*
            cd ../
          '';

          nativeBuildInputs = [
            pkgs.autoPatchelfHook
          ];

          buildInputs = with pkgs; [
            stdenv.cc.cc.lib
            libusb.out
            pytorch-cuda
            libtensorflow-bin
            pkgsUnfree.cudaPackages.cudatoolkit_11.lib
            libtorch-cuda
            libGL
          ];

          propagatedBuildInputs = with pythonPackages; [
            jupyterlab
            jupyter-packaging
            ipywidgets
            tqdm
            pyyaml
            pandas
            plyfile
            scipy
            scikitlearn
            numpy
            addict
            matplotlib
          ];
        };

[rowanG077]

@lapp0 Fun fact that linked expression was written by me a while back :), see here: https://gist.github.com/rowanG077/7f46a0807c108fe4b582067f237ae867. Just didn't care to upstream it. The reason for it is because it uses the precompiled wheels. Which I don't really like.

[breakds]

@rowanG077 Are you still working on packaging open3d? I found that I am in need of packaging it as well. Ideally I think it should be built from source with one c++ library and one python library. But packing the wheel can also work for now.

May I ask what is your current solution to this? I am willing to join the effort if you are still on it. Thanks!

[breakds]

Not stale.

[rowanG077]

@breakds I'm still interested. But going to "correct" path will take a lot of time I don't have right now. My current solution is that I have packaged the wheel locally.

[breakds]

@rowanG077 Thank you! Do you have an updated wheel solution, or is the 0.12 gist the most up to date?

Meanwhile I'll make an attempt to package it from the source, but since I am less experienced, I might put problems / questioins here.

[rowanG077]

@breakds yes 0.12 is the most uptodate.

[pfmephisto]

I am trying to get Open3D to work with python3.10 and I have gotten this far.

I seem to have trouble with the libraries that I mention under autoPatchelfIgnoreMissingDeps If I include mesa as a dependency most of them are found but at run time I get an error that it can not find the swrast driver, which seems to me that it is not using the GPU but trying to rely on a fallback solution. If I leave mesa out my issue now becomes that libglapi can not be found. So I have the impression that actually something with libGL is off, but I'm not sure what.

Which ever way I keep tying things something always does not seem to work properly.

Did anyone manage to get it to work or has any suggestions.

{ config, pkgs, lib, fetchPypi, fetchurl, ... }:
let
  unstable = import <nixos-unstable> {
    config.allowUnfree = true;
    config.cudaSupport = true;
  };

  my_python = unstable.pkgs.python3.override {
    packageOverrides = self: super: {

      nbformat = super.buildPythonPackage rec {

        pname = "nbformat";
        version = "5.7.0";
        format = "wheel";
        src = pkgs.fetchurl {
          url = "https://files.pythonhosted.org/packages/5c/9f/957655d02f43b8bff77e6da08c94472b1229c13e7455bbd662163c9b78c0/nbformat-5.7.0-py3-none-any.whl";
          sha256 = "sha256-GwXsLFUsLxrcdF9O3c4erIyp/9Wbuf2FnoJ+qgMTGfk=";
        };

        buildInputs = with super; [
          fastjsonschema
          jupyter_core
          jsonschema
        ];

        propagatedBuildInputs = with super; [
          fastjsonschema
          jsonschema
        ];

      };
    };
  };

  ml_python_packages = ps: with ps; [
    pip
    ipykernel
    torch-bin
    tensorflowWithoutCuda
    gpustat
    plotly
    scipy
    numpy
    dash
    werkzeug
    nbformat
    configargparse
    ipywidgets
    addict
    pillow
    matplotlib
    pandas
    pyyaml
    scikit-learn
    tqdm
    pyquaternion
    (
      buildPythonPackage rec {
        pname = "open3d";
        version = "0.17.0";
        format = "wheel";

        src = unstable.python3Packages.fetchPypi {
          inherit pname version format;
          sha256 = "sha256-PcMAaXMgu2iCRsXQn2gQRYFcMyIlaFc/GWSy11ZDFlc=";
          dist = "cp310";
          python = "cp310";
          abi = "cp310";
          platform = "manylinux_2_27_x86_64";
        };

        nativeBuildInputs = [
          unstable.autoPatchelfHook
        ];

        autoPatchelfIgnoreMissingDeps = [
          "libLLVM-10.so.1"
          #"libglapi.so.0"
          #"libdrm.so.2"
          #"libexpat.so.1"
          #"libXfixes.so.3"
          #"libXxf86vm.so.1"
        ];

        buildInputs = with unstable; [
          mesa
          zstd
          stdenv.cc.cc.lib
          libusb.out
          glibc
          libGL
          cudaPackages.cudatoolkit
          pkgs.libtorch-bin
          libtensorflow

        ];

        propagatedBuildInputs =  with unstable; [
          (my_python.withPackages(
            ps: with ps; [
              numpy
              dash
              werkzeug
              nbformat
              configargparse
              ipywidgets
              addict
              pillow
              matplotlib
              pandas
              pyyaml
              scikit-learn
              tqdm
              pyquaternion
              #pywinpty
          ]))
        ];

      })

    ];

in
  {

    nixpkgs.config = {
      allowUnfree = true;
      cudaSupport = true;
    };

    # List packages installed in system profile. To search, run:
    environment.systemPackages = with unstable.pkgs; [
      cudaPackages.cudatoolkit
          cudaPackages.cudnn
          conda
          wget
          nvtop
          ctop
          (my_python.withPackages ml_python_packages)
    ];

}

[rowanG077]

libLLVM-10.so.1 can be fixed by symlinking:

postInstall = ''
    ln -s "${llvm_10.lib}/lib/libLLVM-10.so" "$out/lib/libLLVM-10.so.1"
''

It's an ubuntu-ism. Adding xorg.libXfixes, xorg.libXxf86vm, expat and libdrm fixes some more dependencies you marked.

Regarding libglapi I had the same problem while upgrading.

[pfmephisto]

@rowanG077 Thanks,

yah that brings me down to libglapi.so.0 libdrm.so.2 libexpat.so.1 libXfixes.so.3 and libXxf86vm.so.1 missing.

Those in turn seem to be dependencies of kms_swrast_dri.so libEGL.so.1 libgallium_dri.so libGL.so.1 and swrast_dri.so

Could it be that there are some libGL libraries that also simply just need to be remapped?

Is there a way to explore what is inside a package e.g. start a nix-shell and then explore the directories it adds? Then I could try to have a look and see what I find.

[rowanG077]

@pfmephisto

Yeah, honestly opengl/opencl/nvidia/cuda etc are all a mess in nix. I don't know how it all fits together.

You can use this command to get the nix store path of a package (mesa used here an example):

nix --extra-experimental-features nix-command --extra-experimental-features flakes eval -f '<nixpkgs>' --raw 'mesa'

You can look around in that folder to see what certain package bring into scope.

[ShaddyDC]

Hi, did any of you make any progress packaging the latest version?

[rowanG077]

Tagging @pfmephisto, you are probably interested :).

Yes. I have something working. I have found no way to get pytorch to output the split cuda libraries. I tried various things but building it takes ages so I stopped trying since I didn't really need t anyway. See the pytorchSplitCuda which is unused.

{ stdenv, lib, buildPythonPackage, fetchPypi, addict
, numpy, matplotlib, ipywidgets, pandas, pyyaml, tqdm
, autoPatchelfHook, llvm_10, xorg, libtorch-bin, libtensorflow, libusb
, cudaPackages, libGL, nbformat, dash, configargparse, pyquaternion, scikit-learn
, expat, pytorchWithCuda }:

let
  nbformat-570 = nbformat.overridePythonAttrs (old: rec {
    version = "5.7.0";

    JUPYTER_PLATFORM_DIRS=1;
    src = fetchPypi {
      inherit version;
      pname = old.pname;
      hash = "sha256-HUdgwVwaBCae9crzdb6LmN0vaW5eueYD7Cvwkfmw0/M=";
    };
  });

  pytorchSplitCuda = pytorchWithCuda.overridePythonAttrs (old: rec {
    preConfigure = old.preConfigure + ''
      export BUILD_SPLIT_CUDA=1
    '';
  });
in buildPythonPackage rec {
    pname = "open3d";
    version = "0.17.0";
    format = "wheel";

    src = fetchPypi {
      inherit pname version format;
      sha256 = "sha256-PcMAaXMgu2iCRsXQn2gQRYFcMyIlaFc/GWSy11ZDFlc=";
      dist = "cp310";
      python = "cp310";
      abi = "cp310";
      platform = "manylinux_2_27_x86_64";
    };

  nativeBuildInputs = [
    autoPatchelfHook
  ];

  autoPatchelfIgnoreMissingDeps = [
    "libtorch_cuda_cpp.so"
    "libtorch_cuda_cu.so"
  ];

  # Fix llvm library name to match Ubuntu name
  # Remove libraries that actually should be provided by environment
  postInstall = ''
    ln -s "${llvm_10.lib}/lib/libLLVM-10.so" "$out/lib/libLLVM-10.so.1"

    rm $out/lib/python3.10/site-packages/open3d/libGL.so.1
    rm $out/lib/python3.10/site-packages/open3d/swrast_dri.so
    rm $out/lib/python3.10/site-packages/open3d/libgallium_dri.so
    rm $out/lib/python3.10/site-packages/open3d/kms_swrast_dri.so
    rm $out/lib/python3.10/site-packages/open3d/libEGL.so.1
  '';

  buildInputs = [
    stdenv.cc.cc.lib
    libusb.out
    libGL
    cudaPackages.cudatoolkit
    libtorch-bin
    libtensorflow
    expat
    xorg.libXfixes
    nbformat
  ];

  propagatedBuildInputs = [
    # py deps
    numpy
    dash
    configargparse
    scikit-learn
    ipywidgets
    addict
    matplotlib
    pandas
    pyyaml
    tqdm
    pyquaternion
    pytorchWithCuda
  ];
}

[pfmephisto]

I ended up switching to Ubuntu (I know 😅), but I need something that worked and did not have more time to debug the derivation.

[OladapoAjala]

https://github.com/NixOS/nixpkgs/issues/115218#issuecomment-1578342266

@rowanG077 is this package available in a git repo on your profile?

Could you also help with pointers to convert your code to a flake version?

[breakds]

@OladapoAjala This is what I have based on @rowanG077's original packaging: https://github.com/nixvital/ml-pkgs/blob/main/pkgs/open3d/default.nix

[OladapoAjala]

Thanks @breakds , ooks like Open3d is currently unavailable from the ml-pkgs. Can you confirm that the current open3d you have works or it still needs some fine tuning? @breakds

[breakds]

Thanks @breakds , ooks like Open3d is currently unavailable from the ml-pkgs. Can you confirm that the current open3d you have works or it still needs some fine tuning? @breakds

Right, sorry that I left it broken. It should be fixed now in the main branch. However, like @rowanG077 has mentioned there are still some dependencies for patchelf not being found. I tested it just by loading a .pcd file which works.

[spagnoloG]

Hello I have a similar problem now. On nixos 23.05 flake that @rowanG077 provided worked. But once i updated to nixos 23.11 I get segmentaiton fault on my system. (it worked before update)

Here is my flake.nix:

``` { description = "Obss python development environment"; inputs = { nixpkgs.url = "github:NixOS/nixpkgs/nixos-23.05"; flake-utils.url = "github:numtide/flake-utils"; }; outputs = { self, nixpkgs, flake-utils, ... }@inputs: flake-utils.lib.eachDefaultSystem (system: let pkgs = import nixpkgs { inherit system; config = { allowUnfree = true; }; }; fetchPypi = pkgs.python3Packages.fetchPypi; customPython = pkgs.python3.override { packageOverrides = self: super: { opencv4 = super.opencv4.override { enableGtk2 = true; gtk2 = pkgs.gtk2; enableFfmpeg = true; }; }; }; nbformat-570 = pkgs.python3Packages.nbformat.overridePythonAttrs (old: rec { version = "5.7.0"; JUPYTER_PLATFORM_DIRS = 1; src = fetchPypi { inherit version; pname = old.pname; hash = "sha256-HUdgwVwaBCae9crzdb6LmN0vaW5eueYD7Cvwkfmw0/M="; }; }); # Define a custom Python packages set with the nbformat override, version collision! customPythonPackages = pkgs.python3Packages.override { overrides = self: super: { nbformat = nbformat-570; }; }; open3d = pkgs.python3Packages.buildPythonPackage rec { pname = "open3d"; version = "0.17.0"; format = "wheel"; src = pkgs.python3Packages.fetchPypi { pname = "open3d"; version = "0.17.0"; format = "wheel"; sha256 = "sha256-PcMAaXMgu2iCRsXQn2gQRYFcMyIlaFc/GWSy11ZDFlc="; dist = "cp310"; python = "cp310"; abi = "cp310"; platform = "manylinux_2_27_x86_64"; }; nativeBuildInputs = [ pkgs.autoPatchelfHook ]; autoPatchelfIgnoreMissingDeps = [ "libtorch_cuda_cpp.so" "libtorch_cuda_cu.so" "libtorch_cuda.so" "libc10_cuda.so" ]; buildInputs = with pkgs; [ cudaPackages.cudatoolkit stdenv.cc.cc.lib libusb.out libGL cudaPackages.cudatoolkit libtorch-bin libtensorflow expat xorg.libXfixes mesa xorg.libX11 xorg.libXfixes ]; propagatedBuildInputs = with customPythonPackages; [ nbformat-570 numpy dash configargparse scikit-learn ipywidgets addict matplotlib pandas pyyaml tqdm pyquaternion ]; postInstall = '' ln -s "${pkgs.llvm_10.lib}/lib/libLLVM-10.so" "$out/lib/libLLVM-10.so.1" rm $out/lib/python3.10/site-packages/open3d/libGL.so.1 rm $out/lib/python3.10/site-packages/open3d/swrast_dri.so rm $out/lib/python3.10/site-packages/open3d/libgallium_dri.so rm $out/lib/python3.10/site-packages/open3d/kms_swrast_dri.so rm $out/lib/python3.10/site-packages/open3d/libEGL.so.1 ''; }; pythonEnv = customPython.withPackages (ps: with ps; [ matplotlib numpy scipy requests opencv4 flake8 black open3d ]); in { devShell = pkgs.mkShell { buildInputs = [ pythonEnv pkgs.python3Packages.pip pkgs.gtk2 pkgs.ffmpeg ]; shellHook = '' if [ ! -d ./.venv ]; then python3 -m venv .venv fi source .venv/bin/activate pip install pre-commit pre-commit autoupdate echo "Welcome to the Python development environment." ''; }; }); } ```

Finished processing files...
Segmentation fault (core dumped)

If I update the nixpkgs.url to 23.11 I get the following error:

error: builder for '/nix/store/crqckibld9wkq7lfmvqz7gykd1aqwxfb-python3.11-nbformat-5.7.0.drv' failed with exit code 2;
       last 10 log lines:
       > ERROR tests/v4/nbexamples.py - DeprecationWarning: Importing ErrorTree directly from the jsonschema packag...
       > ERROR tests/v4/test_convert.py - DeprecationWarning: Importing ErrorTree directly from the jsonschema packag...
       > ERROR tests/v4/test_convert.py - DeprecationWarning: Importing ErrorTree directly from the jsonschema packag...
       > ERROR tests/v4/test_json.py - DeprecationWarning: Importing ErrorTree directly from the jsonschema packag...
       > ERROR tests/v4/test_json.py - DeprecationWarning: Importing ErrorTree directly from the jsonschema packag...
       > ERROR tests/v4/test_validate.py - DeprecationWarning: Importing ErrorTree directly from the jsonschema packag...
       > ERROR tests/v4/test_validate.py - DeprecationWarning: Importing ErrorTree directly from the jsonschema packag...
       > !!!!!!!!!!!!!!!!!!! Interrupted: 20 errors during collection !!!!!!!!!!!!!!!!!!!
       > ============================== 20 errors in 3.65s ==============================
       > /nix/store/wr08yanv2bjrphhi5aai12hf2qz5kvic-stdenv-linux/setup: line 1559: pop_var_context: head of shell_variables not a function context
       For full logs, run 'nix log /nix/store/crqckibld9wkq7lfmvqz7gykd1aqwxfb-python3.11-nbformat-5.7.0.drv'.

I am a nixos newbie and would gladly accept any suggestions how to make it work. Thanks in advance :)

[spagnoloG]

Finally made it work :) :partying_face: :partying_face: on nixos 23.11 and open3d version 0.18. Here is flake.nix if anyone is also struggling to make it work :)

{
  description = "Opend3d developmen environment";

  inputs = {
    nixpkgs.url = "github:NixOS/nixpkgs/nixos-23.11";
    flake-utils.url = "github:numtide/flake-utils";
  };

  outputs = { self, nixpkgs, flake-utils, ... }@inputs:
    flake-utils.lib.eachDefaultSystem (system:
      let
        pkgs = import nixpkgs {
          inherit system;
          config = { allowUnfree = true; };
        };

        fetchPypi = pkgs.python311Packages.fetchPypi;

        customPython = pkgs.python311.override {
          packageOverrides = self: super: {
            opencv4 = super.opencv4.override {
              enableGtk2 = true;
              gtk2 = pkgs.gtk2;
              enableFfmpeg = true;
            };
          };
        };

        open3d = pkgs.python311Packages.buildPythonPackage rec {
          pname = "open3d";
          version = "0.18.0";
          format = "wheel";

          src = pkgs.python311Packages.fetchPypi {
            pname = "open3d";
            version = "0.18.0";
            format = "wheel";
            sha256 = "sha256-jj0dGQCo9NlW9oGcJGx4CBclubCIj4VJ0qeknI2qEwM=";
            dist = "cp311";
            python = "cp311";
            abi = "cp311";
            platform = "manylinux_2_27_x86_64";
          };

          nativeBuildInputs = [ pkgs.autoPatchelfHook ];
          autoPatchelfIgnoreMissingDeps = [
            "libtorch_cuda_cpp.so"
            "libtorch_cuda_cu.so"
            "libtorch_cuda.so"
            "libc10_cuda.so"
          ];

          buildInputs = with pkgs; [
            cudaPackages.cudatoolkit
            stdenv.cc.cc.lib
            libusb.out
            libGL
            cudaPackages.cudatoolkit
            libtorch-bin
            libtensorflow
            expat
            xorg.libXfixes
            mesa
            xorg.libX11
            xorg.libXfixes
          ];

          propagatedBuildInputs = with pkgs.python311Packages; [
            nbformat
            numpy
            dash
            configargparse
            scikit-learn
            ipywidgets
            addict
            matplotlib
            pandas
            pyyaml
            tqdm
            pyquaternion
          ];

          postInstall = ''
            ln -s "${pkgs.llvm_10.lib}/lib/libLLVM-10.so" "$out/lib/libLLVM-10.so.1"

            rm $out/lib/python3.11/site-packages/open3d/libGL.so.1
            rm $out/lib/python3.11/site-packages/open3d/swrast_dri.so
            rm $out/lib/python3.11/site-packages/open3d/libgallium_dri.so
            rm $out/lib/python3.11/site-packages/open3d/kms_swrast_dri.so
            rm $out/lib/python3.11/site-packages/open3d/libEGL.so.1
          '';

        };

        pythonEnv = customPython.withPackages (ps:
          with ps; [
            matplotlib
            numpy
            scipy
            requests
            opencv4
            flake8
            black
            open3d
          ]);

      in {
        devShell = pkgs.mkShell {
          buildInputs =
            [ pythonEnv pkgs.python311Packages.pip pkgs.gtk2 pkgs.ffmpeg ];

          shellHook = ''
            echo "Welcome to the Python development environment."
          '';
        };
      });
}

[purepani]

This currently doesn't work on unstable due to https://github.com/NixOS/nixpkgs/issues/283954 removing it. Is there a way to get this built on unstable, or will upstream need to be patched for that to be possible?