[CHERRY-PICK2.1]Remove paddle_custom_op dynamic libraries, and link to FLUID_CORE on windows (#32583) (#32769)

* Remove paddle_custom_op dynamic libraries, change link to FLUID_CORE on windows, and check copy_to

* fix CI

Author: zhwesky2010

paddle/fluid/framework/CMakeLists.txt

@@ -369,36 +369,3 @@ cc_library(paddle_framework DEPS ${FLUID_FRAMEWORK_MODULES})
 if(WITH_TESTING AND TEST selected_rows_test)
   set_tests_properties(selected_rows_test PROPERTIES TIMEOUT 120)
 endif()
-
-##### 2.0 New custom op extension mechanism related #####
-
-# if not deps `layer`, will cause: undefined symbol: _ZN6paddle10imperative7VarBase9name_set_
-if (WIN32)
-  set(PADDLE_CUSTOM_OP_MODULES custom_tensor op_meta_info custom_operator layer)
-
-  set(PADDLE_CUSTOM_OP_SRCS
-      ${CMAKE_CURRENT_SOURCE_DIR}/custom_operator.cc
-      ${CMAKE_CURRENT_SOURCE_DIR}/../extension/src/ext_tensor.cc
-      ${CMAKE_CURRENT_SOURCE_DIR}/../extension/src/ext_op_meta_info.cc
-      ${CMAKE_SOURCE_DIR}/paddle/fluid/imperative/layer.cc)
-  set(PADDLE_CUSTOM_OP_SRCS ${PADDLE_CUSTOM_OP_SRCS} PARENT_SCOPE)
-
-  cc_library(paddle_custom_op_shared
-      SHARED SRCS ${PADDLE_CUSTOM_OP_SRCS} DEPS ${PADDLE_CUSTOM_OP_MODULES})
-
-  get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
-  set_target_properties(paddle_custom_op_shared PROPERTIES OUTPUT_NAME paddle_custom_op)
-  target_link_libraries(paddle_custom_op_shared ${os_dependency_modules})
-
-  if("${CMAKE_GENERATOR}" STREQUAL "Ninja")
-    set(paddle_custom_op_lib_path ${CMAKE_CURRENT_BINARY_DIR})
-  else()
-    set(paddle_custom_op_lib_path ${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE})
-  endif()
-  set(PADDLE_CUSTOM_OP_IMPORT_LIB
-      ${paddle_custom_op_lib_path}/paddle_custom_op.lib
-      CACHE INTERNAL "Paddle custom op import lib")
-  set(PADDLE_CUSTOM_OP_SHARED_LIB
-      ${paddle_custom_op_lib_path}/paddle_custom_op.dll
-      CACHE INTERNAL "Paddle custom op dll")
-endif()

paddle/scripts/paddle_build.bat

@@ -54,14 +54,14 @@ wmic process where name="python.exe" call terminate 2>NUL
 rem ------initialize common variable------
 if not defined GENERATOR set GENERATOR="Visual Studio 15 2017 Win64"
 if not defined BRANCH set BRANCH=develop
-if not defined WITH_TENSORRT set WITH_TENSORRT=ON 
+if not defined WITH_TENSORRT set WITH_TENSORRT=ON
 if not defined TENSORRT_ROOT set TENSORRT_ROOT=D:/TensorRT
 if not defined CUDA_ARCH_NAME set CUDA_ARCH_NAME=Auto
 if not defined WITH_GPU set WITH_GPU=ON
 if not defined WITH_MKL set WITH_MKL=ON
 if not defined WITH_AVX set WITH_AVX=ON
 if not defined WITH_TESTING set WITH_TESTING=ON
-if not defined MSVC_STATIC_CRT set MSVC_STATIC_CRT=OFF
+if not defined MSVC_STATIC_CRT set MSVC_STATIC_CRT=ON
 if not defined WITH_PYTHON set WITH_PYTHON=ON
 if not defined ON_INFER set ON_INFER=ON
 if not defined WITH_INFERENCE_API_TEST set WITH_INFERENCE_API_TEST=ON
@@ -75,6 +75,7 @@ if not defined LOG_LEVEL set LOG_LEVEL=normal
 if not defined PRECISION_TEST set PRECISION_TEST=OFF
 if not defined NIGHTLY_MODE set PRECISION_TEST=OFF
 if not defined retry_times set retry_times=2
+if not defined PYTHON_ROOT set PYTHON_ROOT=C:\Python37
 
 rem -------set cache build directory-----------
 rmdir build\python /s/q
@@ -83,9 +84,6 @@ rmdir build\paddle_inference_install_dir /s/q
 rmdir build\paddle_inference_c_install_dir /s/q
 del build\CMakeCache.txt
 
-: set CI_SKIP_CPP_TEST if only *.py changed
-git diff --name-only %BRANCH% | findstr /V "\.py" || set CI_SKIP_CPP_TEST=ON
-
 if "%WITH_CACHE%"=="OFF" (
     rmdir build /s/q
     goto :mkbuild
@@ -135,58 +133,6 @@ dir .
 dir %cache_dir%
 dir paddle\fluid\pybind\Release
 
-rem ------initialize the python environment------
-if not defined PYTHON_ROOT set PYTHON_ROOT=C:\Python37
-set PYTHON_EXECUTABLE=%PYTHON_ROOT%\python.exe
-set PATH=%PYTHON_ROOT%;%PYTHON_ROOT%\Scripts;%PATH%
-
-rem ToDo: virtual environment can't be deleted safely, some process not exit when task is canceled
-rem Now use system python environment temporarily
-rem %PYTHON_EXECUTABLE% -m pip install virtualenv
-rem %PYTHON_EXECUTABLE% -m virtualenv paddle_winci
-rem call paddle_winci\Scripts\activate.bat
-
-rem ------pre install python requirement----------
-where python
-where pip
-pip install wheel --user
-pip install -r %work_dir%\python\requirements.txt --user
-
-if %ERRORLEVEL% NEQ 0 (
-    echo pip install requirements.txt failed!
-    exit /b 7
-)
-
-rem ------pre install clcache and init config----------
-rem pip install clcache --user
-pip uninstall -y clcache
-:: set USE_CLCACHE to enable clcache
-rem set USE_CLCACHE=1
-:: In some scenarios, CLCACHE_HARDLINK can save one file copy.
-rem set CLCACHE_HARDLINK=1
-:: If it takes more than 1000s to obtain the right to use the cache, an error will be reported
-rem set CLCACHE_OBJECT_CACHE_TIMEOUT_MS=1000000
-:: set maximum cache size to 20G
-rem clcache.exe -M 21474836480
-
-:: install ninja if GENERATOR is Ninja
-if %GENERATOR% == "Ninja" (
-    pip install ninja
-    if %errorlevel% NEQ 0 (
-        echo pip install ninja failed!
-        exit /b 7
-    )
-)
-
-rem ------show summary of current environment----------
-cmake --version
-if "%WITH_GPU%"=="ON" (
-    nvcc --version
-    nvidia-smi
-)
-::python %work_dir%\tools\summary_env.py
-::%cache_dir%\tools\busybox64.exe bash %work_dir%\tools\get_cpu_info.sh
-
 goto :CASE_%1
 
 echo "Usage: paddle_build.bat [OPTION]"
@@ -266,8 +212,10 @@ rem "Other configurations are added here"
 rem :CASE_wincheck_others
 rem call ...
 
+
 rem ---------------------------------------------------------------------------------------------
 :cmake
+@ECHO OFF
 echo    ========================================
 echo    Step 1. Cmake ...
 echo    ========================================
@@ -281,12 +229,52 @@ set PATH=C:\Program Files (x86)\Windows Kits\10\bin\10.0.17763.0\x64;%PATH%
 for /F %%# in ('wmic os get localdatetime^|findstr 20') do set start=%%#
 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
+if not defined CUDA_TOOLKIT_ROOT_DIR set CUDA_TOOLKIT_ROOT_DIR=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2
 set PATH=%TENSORRT_ROOT:/=\%\lib;%CUDA_TOOLKIT_ROOT_DIR%\bin;%CUDA_TOOLKIT_ROOT_DIR%\libnvvp;%PATH%
 
-rem ------set third_party cache dir------
+rem install ninja if GENERATOR is Ninja
+if %GENERATOR% == "Ninja" (
+    pip install ninja
+    if %errorlevel% NEQ 0 (
+        echo pip install ninja failed!
+        exit /b 7
+    )
+)
 
+rem ------show summary of current GPU environment----------
+cmake --version
+if "%WITH_GPU%"=="ON" (
+    nvcc --version
+    nvidia-smi
+)
+
+rem ------initialize the python environment------
+set PYTHON_EXECUTABLE=%PYTHON_ROOT%\python.exe
+set PATH=%PYTHON_ROOT%;%PYTHON_ROOT%\Scripts;%PATH%
+if %WITH_PYTHON% == "OFF" (
+    where python
+    where pip
+    pip install wheel --user
+    pip install -r %work_dir%\python\requirements.txt --user
+    if %ERRORLEVEL% NEQ 0 (
+        echo pip install requirements.txt failed!
+        exit /b 7
+    )
+)
+
+rem ------pre install clcache and init config----------
+rem pip install clcache --user
+pip uninstall -y clcache
+:: set USE_CLCACHE to enable clcache
+rem set USE_CLCACHE=1
+:: In some scenarios, CLCACHE_HARDLINK can save one file copy.
+rem set CLCACHE_HARDLINK=1
+:: If it takes more than 1000s to obtain the right to use the cache, an error will be reported
+rem set CLCACHE_OBJECT_CACHE_TIMEOUT_MS=1000000
+:: set maximum cache size to 20G
+rem clcache.exe -M 21474836480
+
+rem ------set third_party cache dir------
 : clear third party cache every once in a while
 for /F %%# in ('wmic os get localdatetime^|findstr 20') do set datetime=%%#
 set day_now=%datetime:~6,2%
@@ -500,6 +488,10 @@ echo    ========================================
 echo    Step 4. Running unit tests ...
 echo    ========================================
 
+
+: set CI_SKIP_CPP_TEST if only *.py changed
+git diff --name-only %BRANCH% | findstr /V "\.py" || set CI_SKIP_CPP_TEST=ON
+
 pip install -r %work_dir%\python\unittest_py\requirements.txt --user
 if %ERRORLEVEL% NEQ 0 (
     echo pip install unittest requirements.txt failed!

python/CMakeLists.txt

@@ -43,9 +43,20 @@ set(FLUID_DST_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/)
 IF(WIN32)
     # Python would use the .pyd by default under Windows series platform
     set(FLUID_CORE ${FLUID_DST_DIR}/${FLUID_CORE_NAME}.pyd)
-    set(FLUID_NOAVX_CORE ${FLUID_DST_DIR}/core_noavx.pyd)
+    set(FLUID_CORE_LIB ${FLUID_DST_DIR}/${FLUID_CORE_NAME}.lib)
+    
+    add_custom_command(OUTPUT ${FLUID_CORE}
+      COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
+      COMMAND cmake -E copy $<TARGET_LINKER_FILE:paddle_pybind> ${FLUID_CORE_LIB}
+      DEPENDS paddle_pybind)
+
+    set(FLUID_NOAVX_CORE ${FLUID_DST_DIR}/core_noavx.pyd)  
 ELSE()
     set(FLUID_CORE ${FLUID_DST_DIR}/${FLUID_CORE_NAME}.so)
+    add_custom_command(OUTPUT ${FLUID_CORE}
+        COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
+        DEPENDS paddle_pybind)
+
     set(FLUID_NOAVX_CORE ${FLUID_DST_DIR}/core_noavx.so)
 ENDIF()
 
@@ -68,9 +79,6 @@ if(HAS_NOAVX_CORE AND EXISTS "${NOAVX_CORE_FILE}")
   list(APPEND FLUID_CORE_DEPS ${FLUID_NOAVX_CORE})
 endif()
 
-add_custom_command(OUTPUT ${FLUID_CORE}
-        COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
-        DEPENDS paddle_pybind)
 add_custom_target(copy_paddle_pybind ALL DEPENDS ${FLUID_CORE_DEPS})
 
 IF(WIN32)
@@ -84,6 +92,7 @@ ELSE(WIN32)
     COMMAND touch stub.cc
     COMMAND cp -r ${PADDLE_SOURCE_DIR}/python/paddle ${PADDLE_BINARY_DIR}/python
     COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
+    COMMENT "Packing whl packages------>>>"
     DEPENDS copy_paddle_pybind ${FLUID_CORE} framework_py_proto profiler_py_proto ${PY_FILES})
 ENDIF()
 

python/paddle/check_import_scipy.py

@@ -24,6 +24,6 @@ def check_import_scipy(OsName):
             if 'DLL load failed' in print_info:
                 raise ImportError(
                     print_info +
-                    "\nplease download visual C++ Redistributable for vs 2015, https://www.microsoft.com/en-us/download/details.aspx?id=48145"
+                    "\nplease download Visual C++ Redistributable from https://support.microsoft.com/en-us/topic/the-latest-supported-visual-c-downloads-2647da03-1eea-4433-9aff-95f26a218cc0"
                 )
     return

python/paddle/fluid/core.py

@@ -37,7 +37,10 @@
 try:
     if os.name == 'nt':
         third_lib_path = current_path + os.sep + '..' + os.sep + 'libs'
-        os.environ['path'] = third_lib_path + ';' + os.environ['path']
+        # Will load shared library from 'path' on windows
+        os.environ[
+            'path'] = current_path + ';' + third_lib_path + ';' + os.environ[
+                'path']
         sys.path.insert(0, third_lib_path)
         # Note: from python3.8, PATH will not take effect
         # https://github.com/python/cpython/pull/12302
@@ -298,7 +301,7 @@ def to_list(s):
                 "WARNING: AVX is supported on local machine, but you have installed "
                 "paddlepaddle without avx core. Hence, no_avx core which has worse "
                 "preformance will be imported.\nYou could reinstall paddlepaddle by "
-                "'python -m pip install -U paddlepaddle-gpu[==version]' or rebuild "
+                "'python -m pip install --force-reinstall paddlepaddle-gpu[==version]' or rebuild "
                 "paddlepaddle WITH_AVX=ON to get better performance.\n"
                 "The original error is: %s\n" % cpt.get_exception_message(e))
             load_noavx = True
@@ -350,12 +353,19 @@ def to_list(s):
             sys.stderr.write(
                 'Error: Can not import noavx core while this file exists: ' +
                 current_path + os.sep + 'core_noavx.' + core_suffix + '\n')
+        elif avx_supported():
+            sys.stderr.write(
+                "Error: AVX is support on your machine, but you have installed "
+                "paddlepaddle without avx core, you should reinstall paddlepaddle by "
+                "'python -m pip install --force-reinstall paddlepaddle-gpu[==version]\n"
+            )
         else:
             sys.stderr.write(
                 "Error: AVX is not support on your machine, but you have installed "
-                "paddlepaddle with avx core, you should reinstall paddlepaddle by "
-                "'python -m pip install -U paddlepaddle-gpu[==version] -f "
-                "https://paddlepaddle.org.cn/whl/stable_noavx.html'\n")
+                "paddlepaddle without no_avx core, you should reinstall paddlepaddle by "
+                "'python -m pip install --force-reinstall paddlepaddle-gpu[==version] -f "
+                "https://paddlepaddle.org.cn/whl/mkl/stable/noavx.html or "
+                "https://paddlepaddle.org.cn/whl/openblas/stable/noavx.html\n")
         raise e
 
 

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

@@ -1,6 +1,5 @@
-# New custom OP can support Windows/Linux now
-if(WITH_GPU OR APPLE) 
-    # GPU custom op tests: compile both .cc and .cu file
+# New custom OP can support Windows/Linux/Mac now
+if(WITH_GPU OR APPLE)
     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)
     py_test(test_custom_relu_model SRCS test_custom_relu_model.py)

python/paddle/fluid/tests/custom_op/custom_relu_op.cu

@@ -45,8 +45,12 @@ std::vector<paddle::Tensor> relu_cuda_forward(const paddle::Tensor& x) {
   int grid = (numel + block - 1) / block;
   PD_DISPATCH_FLOATING_AND_HALF_TYPES(
       x.type(), "relu_cuda_forward_kernel", ([&] {
+        auto cpu_input = x.copy_to<data_t>(paddle::PlaceType::kCPU);
+        auto gpu_input = cpu_input.copy_to<data_t>(paddle::PlaceType::kGPU);
         relu_cuda_forward_kernel<data_t><<<grid, block, 0, x.stream()>>>(
-            x.data<data_t>(), out.mutable_data<data_t>(x.place()), numel);
+            gpu_input.data<data_t>(),
+            out.mutable_data<data_t>(x.place()),
+            numel);
       }));
 
   return {out};

python/paddle/fluid/tests/custom_op/test_check_abi.py

@@ -64,14 +64,29 @@ def test_wrong_compiler_warning(self):
         # clear environ
         self.del_environ()
         compiler = 'python'  # fake wrong compiler
-        with warnings.catch_warnings(record=True) as error:
-            flag = utils.check_abi_compatibility(compiler, verbose=True)
-            # check return False
-            self.assertFalse(flag)
-            # check Compiler Compatibility WARNING
-            self.assertTrue(len(error) == 1)
-            self.assertTrue(
-                "Compiler Compatibility WARNING" in str(error[0].message))
+        if not utils.IS_WINDOWS:
+            with warnings.catch_warnings(record=True) as error:
+                flag = utils.check_abi_compatibility(compiler, verbose=True)
+                # check return False
+                self.assertFalse(flag)
+                # check Compiler Compatibility WARNING
+                self.assertTrue(len(error) == 1)
+                self.assertTrue(
+                    "Compiler Compatibility WARNING" in str(error[0].message))
+
+    def test_exception_windows(self):
+        # clear environ
+        self.del_environ()
+        compiler = 'fake compiler'  # fake command
+        if utils.IS_WINDOWS:
+            with warnings.catch_warnings(record=True) as error:
+                flag = utils.check_abi_compatibility(compiler, verbose=True)
+                # check return False
+                self.assertFalse(flag)
+                # check ABI Compatibility WARNING
+                self.assertTrue(len(error) == 1)
+                self.assertTrue("Failed to check compiler version for" in
+                                str(error[0].message))
 
     def test_exception_linux(self):
         # clear environ

python/paddle/fluid/tests/custom_op/test_custom_relu_op_jit.py

@@ -105,12 +105,12 @@ def test_exception(self):
                 in str(e))
             if IS_WINDOWS:
                 self.assertTrue(
-                    r"python\paddle\fluid\tests\custom_op\custom_relu_op.cc:47"
-                    in str(e))
+                    r"python\paddle\fluid\tests\custom_op\custom_relu_op.cc" in
+                    str(e))
             else:
                 self.assertTrue(
-                    "python/paddle/fluid/tests/custom_op/custom_relu_op.cc:47"
-                    in str(e))
+                    "python/paddle/fluid/tests/custom_op/custom_relu_op.cc" in
+                    str(e))
         self.assertTrue(caught_exception)
 
         caught_exception = False
@@ -126,7 +126,7 @@ def test_exception(self):
                 "function \"relu_cuda_forward_kernel\" is not implemented for data type `int32_t`"
                 in str(e))
             self.assertTrue(
-                "python/paddle/fluid/tests/custom_op/custom_relu_op.cu:50" in
+                "python/paddle/fluid/tests/custom_op/custom_relu_op.cu" in
                 str(e))
         self.assertTrue(caught_exception)