Fix the extension compiling issues on win32

Summary:
This commit fixes the two compiling issues on win32:
1. Remove the constexpr in generating protobuf headers.
2. Enforce the "/MT" flag to overwrite the "/MD" always.

.flake8

 [flake8]
 max-line-length = 120
 ignore = E741, # ambiguous variable name
-         F403, # ‘from module import *’ used; unable to detect undefined names
+         F403, # 'from module import *' used; unable to detect undefined names
          F405, # name may be undefined, or defined from star imports: module
          F811, # redefinition of unused name from line N
          F821, # undefined name

dragon/modules/python/CMakeLists.txt

@@ -117,8 +117,13 @@ strip_debug_symbol(dragon_python)
 set(_install_dir ${CMAKE_INSTALL_PREFIX}/include)
 foreach(_file ${MODULE_INCLUDES})
   get_filename_component(_dir ${_file} DIRECTORY)
+  get_filename_component(_ext ${_file} EXT)
   file(RELATIVE_PATH _dir ${PROJECT_SOURCE_DIR} ${_dir})
   file(COPY ${_file} DESTINATION ${_install_dir}/dragon/${_dir})
+  if (${_ext} STREQUAL ".pb.h" AND WIN32)
+    get_filename_component(_name ${_file} NAME)
+    protobuf_remove_constexpr(${_install_dir}/dragon/${_dir}/${_name})
+  endif()
 endforeach()
 file(COPY ${THIRD_PARTY_DIR}/cub/cub DESTINATION ${_install_dir})
 file(COPY ${THIRD_PARTY_DIR}/eigen/Eigen DESTINATION ${_install_dir})

dragon/modules/runtime/CMakeLists.txt

@@ -93,8 +93,13 @@ strip_debug_symbol(dragonrt)
 set(_install_dir ${CMAKE_INSTALL_PREFIX}/include)
 foreach(_file ${MODULE_INCLUDES})
   get_filename_component(_dir ${_file} DIRECTORY)
+  get_filename_component(_ext ${_file} EXT)
   file(RELATIVE_PATH _dir ${PROJECT_SOURCE_DIR} ${_dir})
   file(COPY ${_file} DESTINATION ${_install_dir}/dragon/${_dir})
+  if (${_ext} STREQUAL ".pb.h" AND WIN32)
+    get_filename_component(_name ${_file} NAME)
+    protobuf_remove_constexpr(${_install_dir}/dragon/${_dir}/${_name})
+  endif()
 endforeach()
 file(COPY dragon_runtime.h DESTINATION ${_install_dir}/dragon)
 file(COPY ${THIRD_PARTY_DIR}/cub/cub DESTINATION ${_install_dir})

dragon/python/tools/cpp_extension.py

@@ -206,7 +206,6 @@ class BuildExtension(_build_ext):
                             cflags = self.cflags
                         else:
                             cflags = []
-
                         cflags = COMMON_NVCC_FLAGS + cflags + _get_cuda_arch_flags(cflags)
                         for flag in COMMON_MSVC_FLAGS:
                             cflags = ['-Xcompiler', flag] + cflags
@@ -218,6 +217,9 @@ class BuildExtension(_build_ext):
                         cflags = COMMON_MSVC_FLAGS + self.cflags
                         cmd += cflags
 
+                if '/MD' in cmd:
+                    cmd.remove('/MD')
+
                 return original_spawn(cmd)
 
             try:
@@ -362,7 +364,7 @@ IS_WINDOWS = _sys.platform == 'win32'
 CUDA_HOME = _find_cuda()
 CUDNN_HOME = _os.environ.get('CUDNN_HOME') or _os.environ.get('CUDNN_PATH')
 COMMON_CC_FLAGS = ['-Wno-sign-compare', '-Wno-unused-variable', '-Wno-reorder']
-COMMON_MSVC_FLAGS = ['/EHsc', '/wd4819', '/wd4244', '/wd4251', '/wd4275', '/wd4800', '/wd4996']
+COMMON_MSVC_FLAGS = ['/MT', '/EHsc', '/wd4819', '/wd4244', '/wd4251', '/wd4275', '/wd4800', '/wd4996']
 COMMON_NVCC_FLAGS = ['-w'] if IS_WINDOWS else ['-std=c++14']
 COMMON_LINK_LIBRARIES = ['protobuf'] if IS_WINDOWS else []
 DLLIMPORT_STR = '__declspec(dllimport)' if IS_WINDOWS else ''

dragon/setup.py

@@ -72,7 +72,8 @@ def configure():
     # Copy "torchvision" => "dragon.vm.torchvision"
     shutil.copytree('../torchvision', 'dragon/vm/torchvision')
     # Copy the pre-built libraries.
-    os.makedirs('dragon/lib')
+    if not os.path.exists('dragon/lib'):
+        os.makedirs('dragon/lib')
     for src, dest in find_libraries().items():
         if os.path.exists(src):
             shutil.copy(src, dest)

dragon/utils/device/common_cuda.h

@@ -104,7 +104,7 @@ inline int CUDA_NUM_DEVICES() {
 
 inline int GetCUDADevice() {
   int device_id;
-  cudaGetDevice(&device_id);
+  CUDA_CHECK(cudaGetDevice(&device_id));
   return device_id;
 }
 
@@ -144,7 +144,7 @@ inline bool TENSOR_CORE_AVAILABLE() {
 class CUDADeviceGuard {
  public:
   explicit CUDADeviceGuard(int new_id) {
-    cudaGetDevice(&prev_id_);
+    CUDA_CHECK(cudaGetDevice(&prev_id_));
     if (prev_id_ != new_id) {
       CUDA_CHECK(cudaSetDevice(new_id));
     }

test/dragon/test_autograph.py

@@ -159,14 +159,15 @@ class TestOpSpec(unittest.TestCase):
                 [self.sym1, self.sym1]).shape, None)
 
     def test_broadcast(self):
-        self.assertEqual(dragon.broadcast_to(
-            self.sym1, shape=(1,)).shape, None)
-        self.assertEqual(dragon.broadcast_to(
-            self.sym2, shape=(1, 2)).shape, (1, 2))
-        self.assertEqual(dragon.broadcast_to(
-            self.sym3, shape=(2,)).shape, self.sym3.shape[:-1] + (2,))
-        self.assertEqual(dragon.broadcast_to(
-            self.sym3, shape=(-1, 2, 2)).shape, (1, 2, 2))
+        with dragon.graph_mode():
+            self.assertEqual(dragon.broadcast_to(
+                self.sym1, shape=(1,)).shape, None)
+            self.assertEqual(dragon.broadcast_to(
+                self.sym2, shape=(1, 2)).shape, (1, 2))
+            self.assertEqual(dragon.broadcast_to(
+                self.sym3, shape=(2,)).shape, self.sym3.shape[:-1] + (2,))
+            self.assertEqual(dragon.broadcast_to(
+                self.sym3, shape=(-1, 2, 2)).shape, (1, 2, 2))
 
     def test_cast(self):
         with dragon.graph_mode():