Attempt to resolve the TLS problem (#25390)

* attempt to resolve tls problem, test=develop

* add glibc version check, test=develop

* fix regex, test=develop

* refine get_libc_ver, test=develop

* refine get_libc_ver, test=develop

cmake/external/mkldnn.cmake

@@ -36,28 +36,12 @@ SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/${LIBDIR
 
 INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
 
-IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
-    SET(MKLDNN_DEPENDS   ${MKLML_PROJECT})
-    MESSAGE(STATUS "Build MKLDNN with MKLML ${MKLML_ROOT}")
-ELSE()
-    MESSAGE(STATUS "Build MKLDNN without MKLML")
-ENDIF()
 
 IF(NOT WIN32)
     SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds")
     SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
     SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
     SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
-
-    IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
-    # Force libmkldnn.so to link libiomp5.so (provided by intel mkl) instead of libgomp.so (provided by gcc),
-    # since core_avx.so links libiomp5.so 
-        set(MKLDNN_SHARED_LINKER_FLAG "${CMAKE_SHARED_LINKER_FLAGS} -Wl,--no-as-needed -L${MKLML_LIB_DIR} -liomp5")
-        set(FORBID "-fopenmp")
-    ELSE()
-        set(MKLDNN_SHARED_LINKER_FLAG "${CMAKE_SHARED_LINKER_FLAGS}")
-        set(FORBID "")        
-    ENDIF()
 ELSE()
     SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} /EHsc")
 ENDIF(NOT WIN32)
@@ -91,8 +75,6 @@ ExternalProject_Add(
                         -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
                         -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
                         -DDNNL_BUILD_TESTS=OFF -DDNNL_BUILD_EXAMPLES=OFF
-                        -DCMAKE_SHARED_LINKER_FLAGS=${MKLDNN_SHARED_LINKER_FLAG}
-                        -DCMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS=${FORBID}
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
 )
 if(WIN32)

python/paddle/fluid/core.py

@@ -17,6 +17,8 @@ from __future__ import print_function
 import site
 import sys
 import os
+import warnings
+import platform
 
 core_suffix = 'so'
 if os.name == 'nt':
@@ -62,7 +64,6 @@ def avx_supported():
     """
     Whether current system(Linux, MacOS, Windows) is supported with AVX.
     """
-    import platform
     from .. import compat as cpt
     sysstr = platform.system().lower()
     has_avx = False
@@ -160,6 +161,72 @@ def avx_supported():
         return False
 
 
+def run_shell_command(cmd):
+    import subprocess
+    out, err = subprocess.Popen(
+        cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE,
+        shell=True).communicate()
+    if err:
+        return None
+    else:
+        return out.decode('utf-8')
+
+
+def get_dso_path(core_so, dso_name):
+    if core_so and dso_name:
+        return run_shell_command("ldd %s|grep %s|awk '{print $3}'" %
+                                 (core_so, dso_name)).strip()
+    else:
+        return None
+
+
+def load_dso(dso_absolute_path):
+    if dso_absolute_path:
+        try:
+            from ctypes import cdll
+            cdll.LoadLibrary(dso_absolute_path)
+        except:
+            warnings.warn("Load {} failed".format(dso_absolute_path))
+
+
+def pre_load(dso_name):
+    if has_avx_core:
+        core_so = current_path + os.sep + 'core_avx.' + core_suffix
+    elif has_noavx_core:
+        core_so = current_path + os.sep + 'core_noavx.' + core_suffix
+    else:
+        core_so = None
+    dso_path = get_dso_path(core_so, dso_name)
+    load_dso(dso_path)
+
+
+def get_glibc_ver():
+    return run_shell_command("ldd --version | awk '/ldd/{print $NF}'").strip()
+
+
+def less_than_ver(a, b):
+    import re
+    import operator
+
+    def to_list(s):
+        s = re.sub('(\.0+)+$', '', s)
+        return [int(x) for x in s.split('.')]
+
+    return operator.lt(to_list(a), to_list(b))
+
+
+# NOTE(zhiqiu): An error may occurs when import paddle in linux platform with glibc < 2.22, 
+# the error message of which is "dlopen: cannot load any more object with static TLS".
+# This happens when:
+# (1) the number of dynamic shared librarys (DSO) loaded > 14,
+# (2) after that, load a dynamic shared library (DSO) with static TLS.
+# For paddle, the problem is that 'libgomp' is a DSO with static TLS, and it is loaded after 14 DSOs.
+# So, here is a tricky way to solve the problem by pre load 'libgomp' before 'core_avx.so'.
+# The final solution is to upgrade glibc to > 2.22 on the target system.
+if platform.system().lower() == 'linux' and less_than_ver(get_glibc_ver(),
+                                                          '2.23'):
+    pre_load('libgomp')
+
 load_noavx = False
 
 if avx_supported():