fix dynamic link of xpu library (#47434)

* refine comments,test=kunlun

* link xpu lib, test=kunlun

* add sleep for test, test=kunlun

* merge develop, fix compile, test=kunlun

* remove debug code, test=kunlun

* add dependency to avoid potential concurrency error, test=kunlun

paddle/CMakeLists.txt

@@ -49,18 +49,15 @@ endif()
 
 list(LENGTH test_names len)
 if(${len} GREATER_EQUAL 1)
-  message("Total tests: ${len}")
+  message("Total cpp tests using dynamic link: ${len}")
   math(EXPR stop "${len} - 1")
   foreach(idx RANGE ${stop})
     if(WITH_TESTING)
       list(GET test_srcs ${idx} test_src)
       list(GET test_names ${idx} test_name)
       get_property(test_arg GLOBAL PROPERTY "${test_name}_ARGS")
-      message("add test ${test_name}")
+      # message("add test ${test_name}")
       add_executable(${test_name} ${test_src})
-      # target_link_libraries(
-      #   ${test_name}
-      #   ${CMAKE_BINARY_DIR}/paddle/fluid/pybind/libpaddle${POSTFIX})
       target_link_libraries(${test_name} $<TARGET_LINKER_FILE:${paddle_lib}>)
       target_link_libraries(${test_name} paddle_gtest_main_new)
       add_dependencies(${test_name} ${paddle_lib} paddle_gtest_main_new)
@@ -75,6 +72,9 @@ if(${len} GREATER_EQUAL 1)
         target_link_libraries(${test_name}
                               "-Wl,-rpath,$<TARGET_FILE_DIR:${paddle_lib}>")
       endif()
+      if(WITH_XPU)
+        target_link_libraries(${test_name} xpulib)
+      endif()
       if(NOT
          ("${test_name}" STREQUAL "c_broadcast_op_npu_test"
           OR "${test_name}" STREQUAL "c_allreduce_sum_op_npu_test"

paddle/fluid/operators/collective/c_comm_init_op.cc

@@ -57,11 +57,9 @@ class CCommInitOp : public framework::OperatorBase {
     using CommContext = platform::NCCLCommContext;
 #elif defined(PADDLE_WITH_XPU_BKCL)
     using UniqueId = BKCLUniqueId;
-    using Place = platform::XPUPlace;
     using CommContext = platform::BKCLCommContext;
 #elif defined(PADDLE_WITH_CNCL)
     using UniqueId = cnclCliqueId;
-    using Place = platform::MLUPlace;
     using CommContext = platform::CNCLCommContext;
 #else
     PADDLE_THROW(platform::errors::PreconditionNotMet(

python/CMakeLists.txt

@@ -48,6 +48,16 @@ set(FLUID_CORE_DEPS ${FLUID_CORE})
 
 add_custom_target(copy_libpaddle ALL DEPENDS ${FLUID_CORE_DEPS})
 
+# NOTE(zhiqiu): WHY?
+# In `setup.py.in`, some dynamic libraries (eg, libxpuapi.so) are modified using
+# patchelf. In rare cases, if the a linker is linking that dynamic library for
+# some executables at the same time, a `file not recognized, file truncated`
+# error may occur, result in the compilation error.
+# So, add dependency to force the cpp tests built before running `setup.py.in`.
+if(WITH_TESTING)
+  add_dependencies(copy_libpaddle build_tests)
+endif()
+
 if(WIN32)
   add_custom_command(
     OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp