[Custom OP]Fix problem of custom op unitests on Windows CI (#31114)

* fix some problem of Windows custom op

* fix some problem of Windows custom op

* fix some problem of Windows custom op

paddle/scripts/paddle_build.bat

@@ -123,6 +123,7 @@ where pip
 pip install wheel --user
 pip install -r %work_dir%\python\unittest_py\requirements.txt --user
 pip install -r %work_dir%\python\requirements.txt --user
+
 if %ERRORLEVEL% NEQ 0 (
     echo pip install requirements.txt failed!
     exit /b 7
@@ -234,7 +235,6 @@ set start=%start:~4,10%
 @ECHO ON
 if not defined CUDA_TOOLKIT_ROOT_DIR set CUDA_TOOLKIT_ROOT_DIR=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0
 set PATH=%TENSORRT_ROOT:/=\%\lib;%CUDA_TOOLKIT_ROOT_DIR%\bin;%CUDA_TOOLKIT_ROOT_DIR%\libnvvp;%PATH%
-set CUDA_PATH=%CUDA_TOOLKIT_ROOT_DIR%
 
 rem ------set third_party cache dir------
 
@@ -246,7 +246,11 @@ set /p day_before=< %cache_dir%\day.txt
 if %day_now% NEQ %day_before% (
     echo %day_now% > %cache_dir%\day.txt
     type %cache_dir%\day.txt
-    if %day_now% EQU 20 (
+    if %day_now% EQU 25 (
+        rmdir %cache_dir%\third_party_GPU/ /s/q
+        rmdir %cache_dir%\third_party/ /s/q
+    )
+    if %day_now% EQU 10 (
         rmdir %cache_dir%\third_party_GPU/ /s/q
         rmdir %cache_dir%\third_party/ /s/q
     )
@@ -452,7 +456,7 @@ if "%WITH_GPU%"=="ON" (
 
 :parallel_test_base_gpu
 echo    ========================================
-echo    Running GPU unit tests...
+echo    Running GPU unit tests in parallel way ...
 echo    ========================================
 
 setlocal enabledelayedexpansion
@@ -471,6 +475,7 @@ goto:eof
 echo    ========================================
 echo    Running CPU unit tests in parallel way ...
 echo    ========================================
+
 ctest.exe -E "(%disable_ut_quickly%)" -LE %nightly_label% --output-on-failure -C Release -j 8 --repeat until-pass:4 after-timeout:4
 
 goto:eof
@@ -676,6 +681,7 @@ taskkill /f /im vctip.exe 2>NUL
 taskkill /f /im cvtres.exe 2>NUL
 taskkill /f /im rc.exe 2>NUL
 wmic process where name="op_function_generator.exe" call terminate 2>NUL
+wmic process where name="python.exe" call terminate 2>NUL
 taskkill /f /im python.exe  2>NUL
 echo Windows CI run successfully!
 exit /b 0

python/paddle/fluid/tests/CMakeLists.txt

@@ -10,13 +10,8 @@ add_subdirectory(unittests)
 add_subdirectory(book)
 
 # TODO: support New Custom OP on Mac
-if(LINUX)
+if(NOT APPLE)
   add_subdirectory(custom_op)
 endif()
 
-# Windows CPU machine doesn't have CUDA, can't compile .cu file
-# if(WIN32 AND WITH_GPU)
-#   add_subdirectory(custom_op)
-# endif()
-
 set_tests_properties(test_beam_search_decoder PROPERTIES TIMEOUT 120)

python/paddle/fluid/tests/custom_op/CMakeLists.txt

 # New custom OP can support Windows/Linux now
-# 'test_custom_relu_op_setup/jit' compile .cc and .cu file
-py_test(test_custom_relu_op_setup SRCS test_custom_relu_op_setup.py)
-py_test(test_custom_relu_op_jit SRCS test_custom_relu_op_jit.py)
-
-# Compiling shared library will cost some time, but running process is very fast.
-set_tests_properties(test_custom_relu_op_setup PROPERTIES TIMEOUT 250)
-set_tests_properties(test_custom_relu_op_jit PROPERTIES TIMEOUT 180)
+if(WITH_GPU)
+    # 'test_custom_relu_op_setup/jit' compile .cc and .cu file
+    py_test(test_custom_relu_op_setup SRCS test_custom_relu_op_setup.py)
+    py_test(test_custom_relu_op_jit SRCS test_custom_relu_op_jit.py)
+
+    # Compiling shared library will cost some time, but running process is very fast.
+    set_tests_properties(test_custom_relu_op_setup PROPERTIES TIMEOUT 250)
+    set_tests_properties(test_custom_relu_op_jit PROPERTIES TIMEOUT 180)
+endif()
 
 py_test(test_sysconfig SRCS test_sysconfig.py)
 

python/paddle/utils/cpp_extension/cpp_extension.py

@@ -24,7 +24,7 @@ from setuptools.command.easy_install import easy_install
 from setuptools.command.build_ext import build_ext
 
 from .extension_utils import find_cuda_home, normalize_extension_kwargs, add_compile_flag, bootstrap_context
-from .extension_utils import is_cuda_file, prepare_unix_cflags, prepare_win_cflags, add_std_without_repeat, get_build_directory
+from .extension_utils import is_cuda_file, prepare_unix_cudaflags, prepare_win_cudaflags, add_std_without_repeat, get_build_directory
 from .extension_utils import _import_module_from_library, CustomOpInfo, _write_setup_file, _jit_compile, parse_op_name_from
 from .extension_utils import check_abi_compatibility, log_v, IS_WINDOWS, OS_NAME
 from .extension_utils import use_new_custom_op_load_method, MSVC_COMPILE_FLAGS
@@ -267,7 +267,7 @@ class BuildExtension(build_ext, object):
                     if isinstance(cflags, dict):
                         cflags = cflags['nvcc']
                     else:
-                        cflags = prepare_unix_cflags(cflags)
+                        cflags = prepare_unix_cudaflags(cflags)
                 # cxx compile Cpp source
                 elif isinstance(cflags, dict):
                     cflags = cflags['cxx']
@@ -332,7 +332,7 @@ class BuildExtension(build_ext, object):
                     else:
                         cflags = []
 
-                    cflags = prepare_win_cflags(cflags) + ['--use-local-env']
+                    cflags = prepare_win_cudaflags(cflags) + ['--use-local-env']
                     for flag in MSVC_COMPILE_FLAGS:
                         cflags = ['-Xcompiler', flag] + cflags
                     cmd = [nvcc_cmd, '-c', src, '-o', obj

python/paddle/utils/cpp_extension/extension_utils.py

@@ -41,17 +41,13 @@ IS_WINDOWS = OS_NAME.startswith('win')
 
 MSVC_COMPILE_FLAGS = [
     '/MT', '/wd4819', '/wd4251', '/wd4244', '/wd4267', '/wd4275', '/wd4018',
-    '/wd4190', '/EHsc', '/w', '/DPADDLE_WITH_CUDA', '/DEIGEN_USE_GPU',
-    '/DNDEBUG'
+    '/wd4190', '/EHsc', '/w', '/DGOOGLE_GLOG_DLL_DECL',
+    '/DBOOST_HAS_STATIC_ASSERT', '/DNDEBUG', '/DPADDLE_USE_DSO'
 ]
 
-MSVC_LINK_FLAGS = [
-    '/MACHINE:X64', 'paddle_framework.lib', 'cudadevrt.lib', 'cudart_static.lib'
-]
+MSVC_LINK_FLAGS = ['/MACHINE:X64', 'paddle_framework.lib']
 
-COMMON_NVCC_FLAGS = [
-    '-DPADDLE_WITH_CUDA', '-DEIGEN_USE_GPU', '-DPADDLE_USE_DSO', '-O3'
-]
+COMMON_NVCC_FLAGS = ['-DPADDLE_WITH_CUDA', '-DEIGEN_USE_GPU', '-O3']
 
 GCC_MINI_VERSION = (5, 4, 0)
 # Give warning if using wrong compiler
@@ -216,7 +212,7 @@ class CustomOpInfo:
         return next(reversed(self.op_info_map.items()))
 
 
-def prepare_unix_cflags(cflags):
+def prepare_unix_cudaflags(cflags):
     """
     Prepare all necessary compiled flags for nvcc compiling CUDA files.
     """
@@ -228,13 +224,11 @@ def prepare_unix_cflags(cflags):
     return cflags
 
 
-def prepare_win_cflags(cflags):
+def prepare_win_cudaflags(cflags):
     """
     Prepare all necessary compiled flags for nvcc compiling CUDA files.
     """
-    cflags = COMMON_NVCC_FLAGS + [
-        '-DGOOGLE_GLOG_DLL_DECL', '-DBOOST_HAS_STATIC_ASSERT', '-w'
-    ] + cflags + get_cuda_arch_flags(cflags)
+    cflags = COMMON_NVCC_FLAGS + ['-w'] + cflags + get_cuda_arch_flags(cflags)
 
     return cflags
 
@@ -269,6 +263,7 @@ def normalize_extension_kwargs(kwargs, use_cuda=False):
     # append necessary include dir path of paddle
     include_dirs = kwargs.get('include_dirs', [])
     include_dirs.extend(find_paddle_includes(use_cuda))
+
     kwargs['include_dirs'] = include_dirs
 
     # append necessary lib path of paddle
@@ -282,6 +277,8 @@ def normalize_extension_kwargs(kwargs, use_cuda=False):
         # append link flags
         extra_link_args = kwargs.get('extra_link_args', [])
         extra_link_args.extend(MSVC_LINK_FLAGS)
+        if use_cuda:
+            extra_link_args.extend(['cudadevrt.lib', 'cudart_static.lib'])
         kwargs['extra_link_args'] = extra_link_args
     else:
         # append compile flags
@@ -323,6 +320,7 @@ def find_cuda_home():
                 if six.PY3:
                     nvcc_path = nvcc_path.decode()
                 nvcc_path = nvcc_path.rstrip('\r\n')
+
                 # for example: /usr/local/cuda/bin/nvcc
                 cuda_home = os.path.dirname(os.path.dirname(nvcc_path))
         except:
@@ -368,10 +366,9 @@ def find_paddle_includes(use_cuda=False):
     third_party_dir = os.path.join(paddle_include_dir, 'third_party')
     include_dirs = [paddle_include_dir, third_party_dir]
 
-    #TODO(zhouwei): because eigen need cuda_runtime.h
-    #So, extend cuda_include_dir always
-    cuda_include_dir = find_cuda_includes()
-    include_dirs.extend(cuda_include_dir)
+    if use_cuda:
+        cuda_include_dir = find_cuda_includes()
+        include_dirs.extend(cuda_include_dir)
 
     return include_dirs
 
@@ -400,10 +397,9 @@ def find_paddle_libraries(use_cuda=False):
     # pythonXX/site-packages/paddle/libs
     paddle_lib_dirs = [get_lib()]
 
-    #TODO(zhouwei): because eigen need cuda_runtime.h
-    #So, extend cuda_lib_dir always
-    cuda_lib_dir = find_cuda_libraries()
-    paddle_lib_dirs.extend(cuda_lib_dir)
+    if use_cuda:
+        cuda_lib_dir = find_cuda_libraries()
+        paddle_lib_dirs.extend(cuda_lib_dir)
 
     return paddle_lib_dirs