Solve PyTorch Ops compilation error in Windows

.github/workflows/windows.yml

@@ -21,7 +21,7 @@ concurrency:
 
 env:
   PIP_VER: "24.3.1"
-  CUDA_VERSION: "12.1.0"
+  CUDA_VERSION: "12.4.0"
   SRC_DIR: "D:\\a\\open3d\\open3d"
   BUILD_DIR: "C:\\Open3D\\build"
   NPROC: 2
@@ -31,7 +31,7 @@ jobs:
   windows:
     permissions:
       contents: write  # upload
-    runs-on: windows-2019
+    runs-on: windows-2022
     strategy:
       fail-fast: false
       matrix:
@@ -96,6 +96,21 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: 3.8
+
+      - name: Checkout Open3D-ML
+        uses: actions/checkout@v4
+        with:
+          repository: isl-org/Open3D-ML
+          path: open3d_ml
+
+      - name: Install PyTorch requirements
+        run: |
+          if (${{ matrix.BUILD_CUDA_MODULE == 'ON' }}) {
+            python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
+            python -m pip install tensorboard
+          } else {
+            python -m pip install -r open3d_ml/requirements-torch.txt
+          }
 
       - name: Config
         # Move build directory to C: https://github.com/actions/virtual-environments/issues/1341
@@ -107,7 +122,7 @@ jobs:
             ${env:DEVELOPER_BUILD}="ON"
           }
           cmake --version
-          cmake -G "Visual Studio 16 2019" -A x64 `
+          cmake -G "Visual Studio 17 2022" -A x64 `
             -DDEVELOPER_BUILD="${env:DEVELOPER_BUILD}" `
             -DCMAKE_SYSTEM_VERSION="10.0.19041.0" `
             -DBUILD_EXAMPLES=OFF `
@@ -119,6 +134,7 @@ jobs:
             -DBUILD_WEBRTC=${{ env.BUILD_WEBRTC }} `
             -DBUILD_UNIT_TESTS=ON `
             -DBUILD_CUDA_MODULE=${{ matrix.BUILD_CUDA_MODULE }} `
+            -DBUILD_PYTORCH_OPS=ON `
             ${{ env.SRC_DIR }}
 
       - name: Build
@@ -202,7 +218,7 @@ jobs:
           $ErrorActionPreference = 'Stop'
           mkdir build
           cd build
-          cmake -G "Visual Studio 16 2019" -A x64 `
+          cmake -G "Visual Studio 17 2022" -A x64 `
             -DCMAKE_INSTALL_PREFIX="C:\Program Files\Open3D" `
             -DSTATIC_WINDOWS_RUNTIME=${{ matrix.STATIC_RUNTIME }} `
             ..
@@ -214,7 +230,7 @@ jobs:
         run: |
           $ErrorActionPreference = 'Stop'
           python -m pip install -U pip==${{ env.PIP_VER }}
-          python -m pip install -U -c python/requirements_build.txt wheel setuptools
+          python -m pip install -U -r python/requirements_build.txt
       - name: Install Python package
         working-directory: ${{ env.BUILD_DIR }}
         run: |

CMakeLists.txt

@@ -395,6 +395,7 @@ if(BUILD_CUDA_MODULE)
     # https://forums.developer.nvidia.com/t/problems-with-latest-vs2022-update/294150/12
     if (MSVC AND MSVC_VERSION VERSION_LESS_EQUAL "1949")
         # Set this before any CUDA checks
+        add_compile_definitions(_ALLOW_COMPILER_AND_STL_VERSION_MISMATCH)
         set(CMAKE_CUDA_FLAGS "--allow-unsupported-compiler" CACHE STRING "Additional flags for nvcc" FORCE)
         message(WARNING "Using --allow-unsupported-compiler flag for nvcc with MSVC 2022. "
         "Set $Env:NVCC_PREPEND_FLAGS='--allow-unsupported-compiler' if nvcc still fails.")

cpp/open3d/ml/impl/continuous_conv/ContinuousConv.h

@@ -59,6 +59,9 @@ void _CConvComputeFeaturesCPU(TOut* out_features,
 
     memset(out_features, 0, sizeof(TOut) * num_out * out_channels);
 
+    typedef Eigen::Array<TFeat, VECSIZE, Eigen::Dynamic> Matrix;
+    typedef Eigen::Array<TReal, VECSIZE, 3> Matrix3C;
+
     tbb::parallel_for(
             tbb::blocked_range<size_t>(0, num_out, 32),
             [&](const tbb::blocked_range<size_t>& r) {
@@ -72,13 +75,12 @@ void _CConvComputeFeaturesCPU(TOut* out_features,
                         in_channels * spatial_filter_size, range_length);
                 B.setZero();
 
-                typedef Eigen::Array<TFeat, VECSIZE, Eigen::Dynamic> Matrix;
                 Matrix infeat(VECSIZE, in_channels);
 
                 Eigen::Array<TReal, 3, 1> offsets_(offsets[0], offsets[1],
                                                    offsets[2]);
 
-                Eigen::Array<TReal, VECSIZE, 3> inv_extents;
+                Matrix3C inv_extents;
                 if (INDIVIDUAL_EXTENT == false) {
                     if (ISOTROPIC_EXTENT) {
                         inv_extents = 1 / extents[0];

cpp/open3d/ml/impl/continuous_conv/ContinuousConvBackpropFilter.h

@@ -63,6 +63,9 @@ void _CConvBackropFilterCPU(TOut* filter_backprop,
     memset(filter_backprop, 0, sizeof(TOut) * total_filter_size);
     std::mutex filter_backprop_mutex;
 
+    typedef Eigen::Array<TFeat, VECSIZE, Eigen::Dynamic> Matrix;
+    typedef Eigen::Array<TReal, VECSIZE, 3> Matrix3C;
+
     tbb::parallel_for(
             tbb::blocked_range<size_t>(0, num_out, 32),
             [&](const tbb::blocked_range<size_t>& r) {
@@ -74,13 +77,12 @@ void _CConvBackropFilterCPU(TOut* filter_backprop,
                 Eigen::Matrix<TFeat, Eigen::Dynamic, Eigen::Dynamic> C(
                         out_channels, range_length);
 
-                typedef Eigen::Array<TFeat, VECSIZE, Eigen::Dynamic> Matrix;
                 Matrix infeat(VECSIZE, in_channels);
 
                 Eigen::Array<TReal, 3, 1> offsets_(offsets[0], offsets[1],
                                                    offsets[2]);
 
-                Eigen::Array<TReal, VECSIZE, 3> inv_extents;
+                Matrix3C inv_extents;
                 if (INDIVIDUAL_EXTENT == false) {
                     if (ISOTROPIC_EXTENT) {
                         inv_extents = 1 / extents[0];

cpp/open3d/ml/impl/continuous_conv/ContinuousConvTranspose.h

@@ -61,6 +61,9 @@ void _CConvTransposeComputeFeaturesCPU(
 
     memset(out_features, 0, sizeof(TOut) * num_out * out_channels);
 
+    typedef Eigen::Array<TFeat, VECSIZE, Eigen::Dynamic> Matrix;
+    typedef Eigen::Array<TReal, VECSIZE, 3> Matrix3C;
+
     tbb::parallel_for(
             tbb::blocked_range<size_t>(0, num_out, 32),
             [&](const tbb::blocked_range<size_t>& r) {
@@ -70,13 +73,12 @@ void _CConvTransposeComputeFeaturesCPU(
                         in_channels * spatial_filter_size, range_length);
                 B.setZero();
 
-                typedef Eigen::Array<TFeat, VECSIZE, Eigen::Dynamic> Matrix;
                 Matrix infeat(VECSIZE, in_channels);
 
                 Eigen::Array<TReal, 3, 1> offsets_(offsets[0], offsets[1],
                                                    offsets[2]);
 
-                Eigen::Array<TReal, VECSIZE, 3> inv_extents;
+                Matrix3C inv_extents;
                 if (INDIVIDUAL_EXTENT == false) {
                     if (ISOTROPIC_EXTENT) {
                         inv_extents = 1 / extents[0];

cpp/open3d/ml/impl/continuous_conv/ContinuousConvTransposeBackpropFilter.h

@@ -64,6 +64,9 @@ void _CConvTransposeBackpropFilterCPU(TOut* filter_backprop,
            sizeof(TOut) * spatial_filter_size * in_channels * out_channels);
     std::mutex filter_backprop_mutex;
 
+    typedef Eigen::Array<TFeat, VECSIZE, Eigen::Dynamic> Matrix;
+    typedef Eigen::Array<TReal, VECSIZE, 3> Matrix3C;
+
     tbb::parallel_for(
             tbb::blocked_range<size_t>(0, num_out, 32),
             [&](const tbb::blocked_range<size_t>& r) {
@@ -75,13 +78,12 @@ void _CConvTransposeBackpropFilterCPU(TOut* filter_backprop,
                 Eigen::Matrix<TFeat, Eigen::Dynamic, Eigen::Dynamic> C(
                         out_channels, range_length);
 
-                typedef Eigen::Array<TFeat, VECSIZE, Eigen::Dynamic> Matrix;
                 Matrix infeat(VECSIZE, in_channels);
 
                 Eigen::Array<TReal, 3, 1> offsets_(offsets[0], offsets[1],
                                                    offsets[2]);
 
-                Eigen::Array<TReal, VECSIZE, 3> inv_extents;
+                Matrix3C inv_extents;
                 if (INDIVIDUAL_EXTENT == false) {
                     if (ISOTROPIC_EXTENT) {
                         inv_extents = 1 / extents[0];

cpp/open3d/ml/pytorch/CMakeLists.txt

@@ -123,7 +123,8 @@ set(TORCH_OPS_ARCH_DIR
     "${TORCH_OPS_DIR}/$<IF:$<BOOL:${BUILD_CUDA_MODULE}>,cuda,cpu>")
 set_target_properties(open3d_torch_ops PROPERTIES
     LIBRARY_OUTPUT_DIRECTORY "${TORCH_OPS_ARCH_DIR}"
-    ARCHIVE_OUTPUT_DIRECTORY "${TORCH_OPS_ARCH_DIR}")
+    ARCHIVE_OUTPUT_DIRECTORY "${TORCH_OPS_ARCH_DIR}"
+    RUNTIME_OUTPUT_DIRECTORY "${TORCH_OPS_ARCH_DIR}")
 
 # Do not add "lib" prefix
 set_target_properties(open3d_torch_ops PROPERTIES PREFIX "")

cpp/pybind/CMakeLists.txt

@@ -165,8 +165,9 @@ endif()
 # add additional optional compiled modules
 if (BUILD_PYTORCH_OPS)
     list( APPEND COMPILED_MODULE_PATH_LIST $<TARGET_FILE:open3d_torch_ops> )
+    set(DEPENDENCIES_DIR "$<TARGET_FILE_DIR:tbb>")
     add_custom_command( OUTPUT "${CMAKE_BINARY_DIR}/lib/ml/torch/python/ops.py" "${CMAKE_BINARY_DIR}/lib/ml/torch/python/return_types.py"
-            COMMAND ${Python3_EXECUTABLE} generate_torch_ops_wrapper.py --input_ops_py_in "${PYTHON_PACKAGE_SRC_DIR}/open3d/ml/torch/python/ops.py.in" --input_return_types_py_in "${PYTHON_PACKAGE_SRC_DIR}/open3d/ml/torch/python/return_types.py.in" --output_dir "${CMAKE_BINARY_DIR}/lib/ml/torch/python/" --lib $<TARGET_FILE:open3d_torch_ops> --tensorflow_ops_dir "${CMAKE_CURRENT_SOURCE_DIR}/../open3d/ml/tensorflow"
+            COMMAND ${Python3_EXECUTABLE} generate_torch_ops_wrapper.py --input_ops_py_in "${PYTHON_PACKAGE_SRC_DIR}/open3d/ml/torch/python/ops.py.in" --input_return_types_py_in "${PYTHON_PACKAGE_SRC_DIR}/open3d/ml/torch/python/return_types.py.in" --output_dir "${CMAKE_BINARY_DIR}/lib/ml/torch/python/" --lib $<TARGET_FILE:open3d_torch_ops> --tensorflow_ops_dir "${CMAKE_CURRENT_SOURCE_DIR}/../open3d/ml/tensorflow" --dependencies_dir "${DEPENDENCIES_DIR}"
                         DEPENDS open3d_torch_ops
                         WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
                         COMMENT "Generating python ops.py and return_types.py" )

cpp/pybind/generate_torch_ops_wrapper.py

@@ -130,12 +130,22 @@ def main():
                         type=str,
                         required=True,
                         help="This is cpp/open3d/ml/tensorflow")
+    parser.add_argument("--dependencies_dir",
+                        type=str,
+                        required=True,
+                        help="PyTorch dependencies directory")
 
     args = parser.parse_args()
     print(args)
 
+    if sys.platform == "win32":
+        open3d_deps = os.add_dll_directory(args.dependencies_dir)
+
     torch.ops.load_library(args.lib)
 
+    if sys.platform == "win32":
+        open3d_deps.close()
+
     generated_function_strs = ''
     generated_namedtuple_strs = ''
     for schema in torch._C._jit_get_all_schemas():

python/open3d/ml/tf/__init__.py

@@ -11,7 +11,7 @@
 from open3d import _build_config
 
 if not _build_config["Tensorflow_VERSION"]:
-    raise Exception('Open3D was not built with TensorFlow support!')
+    raise ImportError('Open3D was not built with TensorFlow support!')
 
 _o3d_tf_version = _build_config["Tensorflow_VERSION"].split('.')
 if _tf_version.split('.')[:2] != _o3d_tf_version[:2]:

python/open3d/ml/torch/__init__.py

@@ -12,7 +12,7 @@
 from open3d import _build_config
 
 if not _build_config["Pytorch_VERSION"]:
-    raise Exception('Open3D was not built with PyTorch support!')
+    raise ImportError('Open3D was not built with PyTorch support!')
 _o3d_torch_version = _verp(_build_config["Pytorch_VERSION"])
 # Check match with PyTorch version, any patch level is OK
 if _verp(_torch.__version__).release[:2] != _o3d_torch_version.release[:2]: