solve mklml memory leak on windows (#24015)

* solve mklml memory leak on windows

test=develop

* remove unused msvcr120.dll

test=develop

cmake/external/mklml.cmake

@@ -21,14 +21,12 @@ SET(MKLML_INC_DIR       ${MKLML_ROOT}/include)
 SET(MKLML_LIB_DIR       ${MKLML_ROOT}/lib)
 SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
 
-SET(TIME_VERSION "2019.0.1.20181227")
 IF(WIN32)
-    SET(MKLML_VER "mklml_win_${TIME_VERSION}" CACHE STRING "" FORCE)
+    SET(MKLML_VER "mklml_win_2019.0.5.20190502" CACHE STRING "" FORCE)
     SET(MKLML_URL "https://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
     SET(MKLML_LIB                 ${MKLML_LIB_DIR}/mklml.lib)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5md.lib)
     SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/mklml.dll)
-    SET(MKLML_SHARED_LIB_DEPS     ${MKLML_LIB_DIR}/msvcr120.dll)
     SET(MKLML_SHARED_IOMP_LIB     ${MKLML_LIB_DIR}/libiomp5md.dll)
 ELSE()
     #TODO(intel-huying):

cmake/inference_lib.cmake

@@ -61,9 +61,9 @@ function(copy_part_of_thrid_party TARGET DST)
         if(WIN32)
             copy(${TARGET}
                     SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_SHARED_LIB}
-                    ${MKLML_SHARED_LIB_DEPS} ${MKLML_SHARED_IOMP_LIB} ${MKLML_INC_DIR}
+                    ${MKLML_SHARED_IOMP_LIB} ${MKLML_INC_DIR}
                     DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}/lib
-                    ${dst_dir}/lib ${dst_dir}/lib ${dst_dir})
+                    ${dst_dir}/lib ${dst_dir})
         else()
             copy(${TARGET}
                     SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}

paddle/fluid/inference/api/analysis_predictor.cc

@@ -323,13 +323,11 @@ bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
 #ifdef PADDLE_WITH_MKLDNN
   if (config_.use_mkldnn_) MkldnnPostReset();
 #endif
-#if defined(PADDLE_WITH_MKLML) && defined(_LINUX)
+#if defined(PADDLE_WITH_MKLML)
   // Frees unused memory allocated by the Intel® MKL Memory Allocator to
   // avoid memory leak. See:
   // https://software.intel.com/en-us/mkl-developer-reference-c-mkl-free-buffers
   platform::dynload::MKL_Free_Buffers();
-// We don't support windows since MKL_Free_Buffers is not in
-// mklml_win_2019.0.1.20181227.zip. We will upgrade mklml_win version later.
 #endif
   return true;
 }
@@ -710,13 +708,11 @@ bool AnalysisPredictor::ZeroCopyRun() {
   // recover the cpu_math_library_num_threads to 1, in order to avoid thread
   // conflict when integrating it into deployment service.
   paddle::platform::SetNumThreads(1);
-#if defined(PADDLE_WITH_MKLML) && defined(_LINUX)
+#if defined(PADDLE_WITH_MKLML)
   // Frees unused memory allocated by the Intel® MKL Memory Allocator to
   // avoid memory leak. See:
   // https://software.intel.com/en-us/mkl-developer-reference-c-mkl-free-buffers
   platform::dynload::MKL_Free_Buffers();
-// We don't support windows since MKL_Free_Buffers is not in
-// mklml_win_2019.0.1.20181227.zip. We will upgrade mklml_win version later.
 #endif
   return true;
 }

paddle/fluid/inference/api/demo_ci/CMakeLists.txt

@@ -165,7 +165,6 @@ if(WIN32)
     add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
           COMMAND ${CMAKE_COMMAND} -E copy ${MATH_LIB_PATH}/lib/mklml.dll ${CMAKE_BINARY_DIR}/Release
           COMMAND ${CMAKE_COMMAND} -E copy ${MATH_LIB_PATH}/lib/libiomp5md.dll ${CMAKE_BINARY_DIR}/Release
-          COMMAND ${CMAKE_COMMAND} -E copy ${MATH_LIB_PATH}/lib/msvcr120.dll ${CMAKE_BINARY_DIR}/Release
           COMMAND ${CMAKE_COMMAND} -E copy ${MKLDNN_PATH}/lib/mkldnn.dll  ${CMAKE_BINARY_DIR}/Release
     )
   else()

paddle/fluid/pybind/CMakeLists.txt

@@ -69,7 +69,6 @@ if(WITH_PYTHON)
     if(${CBLAS_PROVIDER} STREQUAL MKLML)
       add_custom_command(TARGET op_function_generator POST_BUILD
             COMMAND ${CMAKE_COMMAND} -E copy ${MKLML_SHARED_LIB} ${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}
-            COMMAND ${CMAKE_COMMAND} -E copy ${MKLML_SHARED_LIB_DEPS} ${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}
             COMMAND ${CMAKE_COMMAND} -E copy ${MKLML_SHARED_IOMP_LIB} ${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}
           )
     else(${CBLAS_PROVIDER} STREQUAL EXTERN_OPENBLAS)

python/setup.py.in

@@ -251,10 +251,6 @@ if '${WITH_MKL}' == 'ON':
     shutil.copy('${MKLML_SHARED_LIB}', libs_path)
     shutil.copy('${MKLML_SHARED_IOMP_LIB}', libs_path)
     package_data['paddle.libs']+=[('libmklml_intel' if os.name != 'nt' else 'mklml') + ext_name, ('libiomp5' if os.name != 'nt' else 'libiomp5md') + ext_name]
-    # mklml has dependency to vs runtime library
-    if os.name == 'nt':
-        shutil.copy('${MKLML_SHARED_LIB_DEPS}', libs_path)
-        package_data['paddle.libs'] += ['msvcr120.dll']
 else:
     if os.name == 'nt':
         # copy the openblas.dll