windows phi dynamic (#74950)

* phi  dynamic

* fix

* fix

* 2

* fix

* 829

* test

* test

* test

* test

* disable ut

* disable ut

* disable ut

* fluid test

Author: swgu98

.github/workflows/_Windows-GPU.yml

@@ -39,7 +39,7 @@ jobs:
       WITH_UNITY_BUILD: "ON"
       WITH_TPCACHE: "ON"
       WITH_SCCACHE: "ON"
-      WITH_SHARED_PHI: "OFF"
+      WITH_SHARED_PHI: "ON"
       GIT_PR_ID: ${{ github.event.pull_request.number }}
       WITH_TESTING: "ON"
       PRECISION_TEST: "OFF"

.github/workflows/_Windows-Inference.yml

@@ -39,7 +39,7 @@ jobs:
       WITH_UNITY_BUILD: "ON"
       WITH_TPCACHE: "ON"
       WITH_SCCACHE: "ON"
-      WITH_SHARED_PHI: "OFF"
+      WITH_SHARED_PHI: "ON"
       GIT_PR_ID: ${{ github.event.pull_request.number }}
       WITH_TESTING: "ON"
       PRECISION_TEST: "OFF"

.github/workflows/_Windows-OPENBLAS.yml

@@ -41,7 +41,7 @@ jobs:
       WITH_CACHE: "OFF"
       WITH_TPCACHE: "ON"
       WITH_SCCACHE: "ON"
-      WITH_SHARED_PHI: "OFF"
+      WITH_SHARED_PHI: "ON"
       FLAGS_enable_eager_mode: 1
       GIT_PR_ID: ${{ github.event.pull_request.number }}
       WITH_TESTING: "ON"

CMakeLists.txt

@@ -156,9 +156,12 @@ if(WIN32)
   # re-runs CMake to regenerate the build system when the target specification source changes.
   set(CMAKE_SUPPRESS_REGENERATION OFF)
   set(CMAKE_STATIC_LIBRARY_PREFIX lib)
-  set(WITH_SHARED_PHI
-      OFF
-      CACHE BOOL "Disable WITH_SHARED_PHI when compiling PADDLE ON WIN32" FORCE)
+  # set(WITH_SHARED_PHI
+  #     ON
+  #     CACHE
+  #       BOOL
+  #       "Disable WITH_SHARED_PHI when compiling PADDLE ON WIN32 with static library"
+  #       FORCE)
 
   set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} /bigobj")
   set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /bigobj")

ci/windows/build.bat

@@ -208,7 +208,7 @@ if "%WITH_TESTING%"=="ON" (
 )
 
 cd /d %work_dir%\%BUILD_DIR%
-echo cmake .. -G %GENERATOR% --trace-expand -DCMAKE_BUILD_TYPE=Release -DWITH_AVX=%WITH_AVX% -DWITH_GPU=%WITH_GPU% -DWITH_MKL=%WITH_MKL% ^
+echo cmake .. -G %GENERATOR% -DCMAKE_BUILD_TYPE=Release -DWITH_AVX=%WITH_AVX% -DWITH_GPU=%WITH_GPU% -DWITH_MKL=%WITH_MKL% ^
 -DWITH_TESTING=%WITH_TESTING% -DWITH_PYTHON=%WITH_PYTHON% -DPYTHON_EXECUTABLE=%PYTHON_EXECUTABLE% -DON_INFER=%ON_INFER% ^
 -DWITH_INFERENCE_API_TEST=%WITH_INFERENCE_API_TEST% -DTHIRD_PARTY_PATH=%THIRD_PARTY_PATH% ^
 -DINFERENCE_DEMO_INSTALL_DIR=%INFERENCE_DEMO_INSTALL_DIR% -DWITH_STATIC_LIB=%WITH_STATIC_LIB% ^
@@ -217,7 +217,18 @@ echo cmake .. -G %GENERATOR% --trace-expand -DCMAKE_BUILD_TYPE=Release -DWITH_AV
 -DCUDA_TOOLKIT_ROOT_DIR="%CUDA_TOOLKIT_ROOT_DIR%" -DNEW_RELEASE_ALL=%NEW_RELEASE_ALL% -DNEW_RELEASE_PYPI=%NEW_RELEASE_PYPI% ^
 -DNEW_RELEASE_JIT=%NEW_RELEASE_JIT% -DWITH_ONNXRUNTIME=%WITH_ONNXRUNTIME% -DWITH_CPP_TEST=%WITH_CPP_TEST% ^
 -DWIN_UNITTEST_LEVEL=%WIN_UNITTEST_LEVEL% -DWITH_NIGHTLY_BUILD=%WITH_NIGHTLY_BUILD% -DWITH_PIP_CUDA_LIBRARIES=%WITH_PIP_CUDA_LIBRARIES% ^
--DWITH_SCCACHE=%WITH_SCCACHE% >> %work_dir%\win_cmake.sh
+-DWITH_SCCACHE=%WITH_SCCACHE% -DWITH_SHARED_PHI=%WITH_SHARED_PHI% >> %work_dir%\win_cmake.sh
+
+echo cmake .. -G %GENERATOR% -DCMAKE_BUILD_TYPE=Release -DWITH_AVX=%WITH_AVX% -DWITH_GPU=%WITH_GPU% -DWITH_MKL=%WITH_MKL% ^
+-DWITH_TESTING=%WITH_TESTING% -DWITH_PYTHON=%WITH_PYTHON% -DPYTHON_EXECUTABLE=%PYTHON_EXECUTABLE% -DON_INFER=%ON_INFER% ^
+-DWITH_INFERENCE_API_TEST=%WITH_INFERENCE_API_TEST% -DTHIRD_PARTY_PATH=%THIRD_PARTY_PATH% ^
+-DINFERENCE_DEMO_INSTALL_DIR=%INFERENCE_DEMO_INSTALL_DIR% -DWITH_STATIC_LIB=%WITH_STATIC_LIB% ^
+-DWITH_TENSORRT=%WITH_TENSORRT% -DTENSORRT_ROOT="%TENSORRT_ROOT%" -DMSVC_STATIC_CRT=%MSVC_STATIC_CRT% ^
+-DWITH_UNITY_BUILD=%WITH_UNITY_BUILD% -DCUDA_ARCH_NAME=%CUDA_ARCH_NAME% -DCUDA_ARCH_BIN=%CUDA_ARCH_BIN% -DCUB_PATH=%THIRD_PARTY_HOME%/cub ^
+-DCUDA_TOOLKIT_ROOT_DIR="%CUDA_TOOLKIT_ROOT_DIR%" -DNEW_RELEASE_ALL=%NEW_RELEASE_ALL% -DNEW_RELEASE_PYPI=%NEW_RELEASE_PYPI% ^
+-DNEW_RELEASE_JIT=%NEW_RELEASE_JIT% -DWITH_ONNXRUNTIME=%WITH_ONNXRUNTIME% -DWITH_CPP_TEST=%WITH_CPP_TEST% ^
+-DWIN_UNITTEST_LEVEL=%WIN_UNITTEST_LEVEL% -DWITH_NIGHTLY_BUILD=%WITH_NIGHTLY_BUILD% -DWITH_PIP_CUDA_LIBRARIES=%WITH_PIP_CUDA_LIBRARIES% ^
+-DWITH_SCCACHE=%WITH_SCCACHE% -DWITH_SHARED_PHI=%WITH_SHARED_PHI%
 
 cmake .. -G %GENERATOR% -DCMAKE_BUILD_TYPE=Release -DWITH_AVX=%WITH_AVX% -DWITH_GPU=%WITH_GPU% -DWITH_MKL=%WITH_MKL% ^
 -DWITH_TESTING=%WITH_TESTING% -DWITH_PYTHON=%WITH_PYTHON% -DPYTHON_EXECUTABLE=%PYTHON_EXECUTABLE% -DON_INFER=%ON_INFER% ^
@@ -228,7 +239,7 @@ cmake .. -G %GENERATOR% -DCMAKE_BUILD_TYPE=Release -DWITH_AVX=%WITH_AVX% -DWITH_
 -DCUDA_TOOLKIT_ROOT_DIR="%CUDA_TOOLKIT_ROOT_DIR%" -DNEW_RELEASE_ALL=%NEW_RELEASE_ALL% -DNEW_RELEASE_PYPI=%NEW_RELEASE_PYPI% ^
 -DNEW_RELEASE_JIT=%NEW_RELEASE_JIT% -DWITH_ONNXRUNTIME=%WITH_ONNXRUNTIME% -DWITH_CPP_TEST=%WITH_CPP_TEST% ^
 -DWIN_UNITTEST_LEVEL=%WIN_UNITTEST_LEVEL% -DWITH_NIGHTLY_BUILD=%WITH_NIGHTLY_BUILD% -DWITH_PIP_CUDA_LIBRARIES=%WITH_PIP_CUDA_LIBRARIES% ^
--DWITH_SCCACHE=%WITH_SCCACHE%
+-DWITH_SCCACHE=%WITH_SCCACHE% -DWITH_SHARED_PHI=%WITH_SHARED_PHI%
 goto:eof
 
 :cmake_error

cmake/external/cccl.cmake

@@ -17,8 +17,8 @@ include_directories(${CCCL_INCLUDE_DIR})
 
 file(TO_NATIVE_PATH ${PADDLE_SOURCE_DIR}/patches/cccl/util_device.cuh.patch
      native_src)
-set(CCCL_PATCH_COMMAND git checkout -- . && git checkout ${CCCL_TAG} && patch
-                       -p1 -Nd ${CCCL_SOURCE_DIR} < ${native_src})
+set(CCCL_PATCH_COMMAND git checkout -- . && git checkout ${CCCL_TAG} && git
+                       apply ${native_src})
 
 ExternalProject_Add(
   extern_cccl

cmake/generic.cmake

@@ -513,45 +513,45 @@ function(cc_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS ARGS)
     cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}"
                           "${multiValueArgs}" ${ARGN})
-    if(WIN32)
-      # NOTE(zhiqiu): on windows platform, the symbols should be exported
-      # explicitly by __declspec(dllexport), however, there are several
-      # symbols not exported, and link error occurs.
-      # so, the tests are not built against dynamic libraries now.
-      cc_test_old(
-        ${TARGET_NAME}
-        SRCS
-        ${cc_test_SRCS}
-        DEPS
-        ${cc_test_DEPS}
-        ARGS
-        ${cc_test_ARGS})
-    else()
-      list(LENGTH cc_test_SRCS len)
-      # message("cc_test_SRCS ${cc_test_SRCS}")
-      # message("cc_test_ARGS ${cc_test_ARGS}")
-
-      if(${len} GREATER 1)
-        message(
-          SEND_ERROR
-            "The number source file of cc_test should be 1, but got ${len}, the source files are: ${cc_test_SRCS}"
-        )
-      endif()
+    # if(WIN32)
+    #   # NOTE(zhiqiu): on windows platform, the symbols should be exported
+    #   # explicitly by __declspec(dllexport), however, there are several
+    #   # symbols not exported, and link error occurs.
+    #   # so, the tests are not built against dynamic libraries now.
+    #   cc_test_old(
+    #     ${TARGET_NAME}
+    #     SRCS
+    #     ${cc_test_SRCS}
+    #     DEPS
+    #     ${cc_test_DEPS}
+    #     ARGS
+    #     ${cc_test_ARGS})
+    # else()
+    list(LENGTH cc_test_SRCS len)
+    # message("cc_test_SRCS ${cc_test_SRCS}")
+    # message("cc_test_ARGS ${cc_test_ARGS}")
+
+    if(${len} GREATER 1)
+      message(
+        SEND_ERROR
+          "The number source file of cc_test should be 1, but got ${len}, the source files are: ${cc_test_SRCS}"
+      )
+    endif()
 
-      list(LENGTH cc_test_ARGS len_arg)
-      if(len_arg GREATER_EQUAL 1)
-        set_property(GLOBAL PROPERTY "${TARGET_NAME}_ARGS" "${cc_test_ARGS}")
-        #message("${TARGET_NAME}_ARGS arg ${arg}")
-      endif()
+    list(LENGTH cc_test_ARGS len_arg)
+    if(len_arg GREATER_EQUAL 1)
+      set_property(GLOBAL PROPERTY "${TARGET_NAME}_ARGS" "${cc_test_ARGS}")
+      #message("${TARGET_NAME}_ARGS arg ${arg}")
+    endif()
 
-      get_property(test_srcs GLOBAL PROPERTY TEST_SRCS)
-      set(test_srcs ${test_srcs} "${CMAKE_CURRENT_SOURCE_DIR}/${cc_test_SRCS}")
-      set_property(GLOBAL PROPERTY TEST_SRCS "${test_srcs}")
+    get_property(test_srcs GLOBAL PROPERTY TEST_SRCS)
+    set(test_srcs ${test_srcs} "${CMAKE_CURRENT_SOURCE_DIR}/${cc_test_SRCS}")
+    set_property(GLOBAL PROPERTY TEST_SRCS "${test_srcs}")
 
-      get_property(test_names GLOBAL PROPERTY TEST_NAMES)
-      set(test_names ${test_names} ${TARGET_NAME})
-      set_property(GLOBAL PROPERTY TEST_NAMES "${test_names}")
-    endif()
+    get_property(test_names GLOBAL PROPERTY TEST_NAMES)
+    set(test_names ${test_names} ${TARGET_NAME})
+    set_property(GLOBAL PROPERTY TEST_NAMES "${test_names}")
+    # endif()
   endif()
 endfunction()
 

cmake/inference_lib.cmake

@@ -288,6 +288,11 @@ copy(
   DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/lib)
 
 if(WIN32)
+  set(paddle_phi_libs ${PADDLE_BINARY_DIR}/paddle/phi/phi*)
+  copy(
+    inference_lib_dist
+    SRCS ${paddle_phi_libs}
+    DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/lib)
   if(WITH_STATIC_LIB)
     set(paddle_inference_lib
         $<TARGET_FILE_DIR:paddle_inference>/libpaddle_inference.lib

paddle/common/flags.cc

@@ -1799,6 +1799,15 @@ PHI_DEFINE_EXPORTED_string(
     "",
     "Remove some redundant information when printing the pir program");
 
+#ifdef _WIN32
+PHI_DEFINE_EXPORTED_string(
+    flagcx_dir,  // NOLINT
+    "",
+    "Specify path for loading libflagcx.so. For instance, "
+    "For instance, /usr/local/flagcx/lib. If default, "
+    "dlopen will search flagcx from LD_LIBRARY_PATH");
+#endif
+
 /**
  * ProcessGroupNCCL related FLAG
  * Name: enable_async_trace

paddle/common/flags.h

@@ -52,10 +52,10 @@
 #define PD_DECLARE_string(name) DECLARE_string(name)
 #endif
 
-#define PD_DECLARE_VARIABLE(type, name)     \
-  namespace paddle_flags {                  \
-  extern PHI_IMPORT_FLAG type FLAGS_##name; \
-  }                                         \
+#define PD_DECLARE_VARIABLE(type, name)        \
+  namespace paddle_flags {                     \
+  extern COMMON_IMPORT_FLAG type FLAGS_##name; \
+  }                                            \
   using paddle_flags::FLAGS_##name
 
 #define COMMON_DECLARE_VARIABLE(type, name)    \
@@ -358,16 +358,16 @@ PADDLE_API ExportedFlagInfoMap* GetMutableExportedFlagInfoMap();
     int Touch() const { return 0; }                                           \
   };                                                                          \
   static __PaddleRegisterFlag_##__name __PaddleRegisterFlag_instance##__name; \
-  int TouchPaddleFlagRegister_##__name() {                                    \
+  PADDLE_API int TouchPaddleFlagRegister_##__name() {                         \
     return __PaddleRegisterFlag_instance##__name.Touch();                     \
   }                                                                           \
   static_assert(std::is_same<__PaddleRegisterFlag_##__name,                   \
                              ::__PaddleRegisterFlag_##__name>::value,         \
                 "FLAGS should define in global namespace")
 
-#define PADDLE_FORCE_LINK_FLAG(__name)           \
-  extern int TouchPaddleFlagRegister_##__name(); \
-  UNUSED static int __paddle_use_flag_##__name = \
+#define PADDLE_FORCE_LINK_FLAG(__name)                      \
+  PADDLE_API extern int TouchPaddleFlagRegister_##__name(); \
+  UNUSED static int __paddle_use_flag_##__name =            \
       TouchPaddleFlagRegister_##__name()
 
 #define PHI_DEFINE_EXPORTED_bool(name, default_value, doc) \