fix #55 windows don't support openmp 4

CMakeLists.txt

@@ -8,6 +8,9 @@ else()
     project(SparseConv LANGUAGES CXX VERSION 1.1)
 endif()
 
+if(WIN32) # true if windows (32 and 64 bit)
+    add_compile_definitions(TV_WINDOWS)
+endif()
 
 set(CMAKE_CXX_EXTENSIONS OFF) # avoid gnu++11 be added to CXX flags
 if(CMAKE_BUILD_TYPE STREQUAL "Debug")
@@ -22,7 +25,6 @@ if (SPCONV_BuildCUDA)
     get_filename_component(CUDA_TOOLKIT_ROOT_DIR "${CUDA_TOOLKIT_ROOT_DIR}" DIRECTORY)
     if(WIN32) # true if windows (32 and 64 bit)
         set(CUDA_LIB_PATH_HINTS "${CUDA_TOOLKIT_ROOT_DIR}/lib/x64")
-        add_compile_definitions(TV_WINDOWS)
     else()
         set(CUDA_LIB_PATH_HINTS "${CUDA_TOOLKIT_ROOT_DIR}/lib64")
     endif()
@@ -45,7 +47,6 @@ if (SPCONV_BuildCUDA)
     set(ALL_INCLUDE ${ALL_INCLUDE} ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
     add_subdirectory(src/cuhash)
 endif()
-
 add_subdirectory(src/spconv)
 add_subdirectory(src/utils)
 

setup.py

@@ -64,6 +64,7 @@ class CMakeBuild(build_ext):
             cmake_args += ['-DCMAKE_LIBRARY_OUTPUT_DIRECTORY_{}={}'.format(cfg.upper(), str(Path(extdir) / "spconv"))]
             # cmake_args += ['-DCMAKE_ARCHIVE_OUTPUT_DIRECTORY_{}={}'.format(cfg.upper(), str(Path(extdir) / "spconv"))]
             cmake_args += ['-DCMAKE_RUNTIME_OUTPUT_DIRECTORY_{}={}'.format(cfg.upper(), str(Path(extdir) / "spconv"))]
+            cmake_args += ["-DCMAKE_WINDOWS_EXPORT_ALL_SYMBOLS=TRUE"]
             if sys.maxsize > 2**32:
                 cmake_args += ['-A', 'x64']
             build_args += ['--', '/m']

src/cuhash/CMakeLists.txt

-add_library(cuhash STATIC hash_functions.cu hash_table.cpp hash_table.cu hash_functions.cpp)
-
+if(WIN32)
+    add_library(cuhash SHARED hash_functions.cu hash_table.cpp hash_table.cu hash_functions.cpp)
+else()
+    add_library(cuhash STATIC hash_functions.cu hash_table.cpp hash_table.cu hash_functions.cpp)
+endif()
 target_include_directories(cuhash PRIVATE ${ALL_INCLUDE} )
 set_property(TARGET cuhash PROPERTY CUDA_STANDARD 14)
 set_property(TARGET cuhash PROPERTY CXX_STANDARD 14)
 set_target_properties(cuhash PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
-set_property(TARGET cuhash PROPERTY POSITION_INDEPENDENT_CODE ON)
+set_target_properties(cuhash PROPERTIES CUDA_RESOLVE_DEVICE_SYMBOLS ON)
+if(NOT WIN32)
+    set_property(TARGET cuhash PROPERTY POSITION_INDEPENDENT_CODE ON)
+endif()
 target_link_libraries(cuhash PRIVATE ${ALL_LIBS})
 install (TARGETS cuhash DESTINATION lib)
 
 if (SPCONV_BuildTests)
     add_executable(cuhash_test main.cc)
-
     target_include_directories(cuhash_test PRIVATE ${ALL_INCLUDE} )
     set_property(TARGET cuhash_test PROPERTY CUDA_STANDARD 14)
     set_property(TARGET cuhash_test PROPERTY CXX_STANDARD 14)

src/spconv/indice.cc

@@ -136,6 +136,8 @@ Index getIndicePairsDeConv(tv::TensorView<const Index> indicesIn,
   return numAct;
 }
 
+
+#ifndef TV_WINDOWS
 template <typename Index, typename IndexGrid, unsigned NDim>
 Index getIndicePairsSubM(tv::TensorView<const Index> indicesIn,
                          tv::TensorView<IndexGrid> gridsOut,
@@ -194,6 +196,58 @@ Index getIndicePairsSubM(tv::TensorView<const Index> indicesIn,
   });
   return numActIn;
 }
+#else 
+template <typename Index, typename IndexGrid, unsigned NDim>
+Index getIndicePairsSubM(tv::TensorView<const Index> indicesIn,
+                         tv::TensorView<IndexGrid> gridsOut,
+                         tv::TensorView<Index> indicePairs,
+                         tv::TensorView<Index> indiceNum,
+                         const Index *const kernelSize,
+                         const Index *const stride, const Index *const padding,
+                         const Index *dilation, const Index *const outSpatialShape) {
+  Index numAct = 0;
+  auto numActIn = indicesIn.dim(0);
+  Index batchIdx = 0;
+  Index spatialVolume = 1;
+#pragma unroll
+  for (int i = 0; i < NDim; ++i) {
+    spatialVolume *= outSpatialShape[i];
+  }
+  Index kernelVolume = 1;
+#pragma unroll
+  for (int i = 0; i < NDim; ++i) {
+    kernelVolume *= kernelSize[i];
+  }
+  Index numValidPoints = 0;
+  // Index validPoints[kernelVolume * (NDim + 1)];
+  std::vector<Index> validPoints_(kernelVolume * (NDim + 1));
+  Index* validPoints = validPoints_.data();
+  Index *pointPtr = nullptr;
+  Index index = 0;
+  for (int j = 0; j < numActIn; ++j) {
+    index = tv::rowArrayIdx<Index, NDim>(indicesIn.data() + j * (NDim + 1) + 1,
+                                         outSpatialShape) +
+            spatialVolume * indicesIn(j, 0);
+    gridsOut[index] = j;
+  }
+  for (int j = 0; j < numActIn; ++j) {
+    numValidPoints = getValidOutPos<Index, NDim>(
+        indicesIn.data() + j * (NDim + 1) + 1, kernelSize, stride, padding,
+        dilation, outSpatialShape, validPoints);
+    for (Index i = 0; i < numValidPoints; ++i) {
+      pointPtr = validPoints + i * (NDim + 1);
+      auto offset = pointPtr[NDim];
+      index = tv::rowArrayIdx<Index, NDim>(pointPtr, outSpatialShape) +
+              spatialVolume * indicesIn(j, 0);
+      if (gridsOut[index] > -1) {
+        indicePairs(offset, 0, indiceNum[offset]) = j;
+        indicePairs(offset, 1, indiceNum[offset]++) = gridsOut[index];
+      }
+    }
+  }
+  return numActIn;
+}
+#endif
 
 namespace functor {
 template <typename Index, typename IndexGrid, unsigned NDim>