README_WINDOWS.md

Compiling ADOP on Windows

Important: The Windows setup may not be working in all future commits, it was originally written for commit a433698 and updated based on the setup used in TRIPS. It is also not tested as well as the Ubuntu setup, so prefer using that if issues arise.

Install Instructions Windows

Software Requirements:

• VS2022

• CUDA 11.8

• Cudnn (copy into 11.8 folder as per install instructions) (we used version 8.9.7)

• conda (we used Anaconda3)

  [Start VS2022 once for CUDA integration setup]

Clone Repo

git clone [email protected]:darglein/ADOP.git
cd ADOP/
git submodule update --init --recursive --jobs 0

Setup Environment

conda update -n base -c defaults conda

conda create -y -n adop python=3.9.7

conda activate adop

conda install -y cmake=3.26.4
conda install -y -c intel mkl=2024.0.0
conda install -y -c intel mkl-static=2024.0.0
conda install openmp=8.0.1 -c conda-forge

Install libtorch:

• Download: https://download.pytorch.org/libtorch/cu116/libtorch-win-shared-with-deps-1.13.1%2Bcu116.zip
• Unzip
• Copy into ADOP/External

Folder structure should look like:

ADOP/
    External/
        libtorch/
            bin/
            cmake/
            include/
            lib/
            ...
        saiga/
        ...
    src/
    ...

Compile

cmake -Bbuild -DCMAKE_CUDA_COMPILER="$ENV:CUDA_PATH\bin\nvcc.exe" -DCMAKE_PREFIX_PATH=".\External\libtorch" -DCONDA_P_PATH="$ENV:CONDA_PREFIX" -DCUDA_P_PATH="$ENV:CUDA_PATH" -DCMAKE_BUILD_TYPE=RelWithDebInfo .

cmake --build build --config RelWithDebInfo -j

Run Instructions Windows

Executable Paths on Windows need the build version added in the run path. You can start the viewer with:

./build/bin/RelWithDebInfo/adop_viewer.exe  --scene_dir scenes/tt_train

and the training with:

./build/bin/RelWithDebInfo/adop_train.exe --config configs/train_boat.ini