From 83d6a5ad2c79c6551ab4c5c0512a5d72a1e1d163 Mon Sep 17 00:00:00 2001
From: Sing_chan <51314274+betterpig@users.noreply.github.com>
Date: Mon, 11 Apr 2022 11:38:59 +0800
Subject: [PATCH] =?UTF-8?q?=E3=80=90cherry-pick=202.3=E3=80=91modify=20inf?=
 =?UTF-8?q?erence=20model=20test=20build=20method=20to=20support=20multi?=
 =?UTF-8?q?=20version=20(#41503)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* change inference demo_test build method to ninja to choose visual studio version automaticly

* notest;test=windows_ci_inference

* set cuda of demo_ci by arg,fix bug of ninja compile,test=document_fix;test=windows_ci;test=windows_ci_inference

* fix bug;test=document_fix;test=windows_ci;test=windows_ci_inference

* fix bug;test=document_fix;test=windows_ci_inference"

* set lib_path according to generator
---
 .../inference/api/demo_ci/CMakeLists.txt      | 30 ++++++++++-------
 paddle/fluid/inference/api/demo_ci/run.sh     | 33 +++++++++++--------
 .../inference/tests/infer_ut/CMakeLists.txt   | 33 ++++++++++++-------
 .../infer_ut/external-cmake/gtest-cpp.cmake   |  6 ++--
 paddle/fluid/inference/tests/infer_ut/run.sh  | 14 ++++----
 paddle/scripts/paddle_build.bat               |  7 ++--
 6 files changed, 73 insertions(+), 50 deletions(-)

diff --git a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt
index df98a7b05cf..c02fcd07813 100644
--- a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt
+++ b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt
@@ -84,13 +84,15 @@ if(WITH_GPU)
   if(NOT WIN32)
     set(CUDA_LIB "/usr/local/cuda/lib64/" CACHE STRING "CUDA Library")
   else()
-    if(NOT DEFINED CUDA_LIB)
+    set(CUDA_LIB "" CACHE STRING "CUDA_LIB")
+    if("${TENSORRT_ROOT}" STREQUAL "")
       if(DEFINED ENV{CUDA_PATH})
         set(CUDA_LIB "$ENV{CUDA_PATH}\\lib\\x64")
       else()
         set(CUDA_LIB "C:\\Program\ Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.2\\lib\\x64")
       endif()
     endif()
+    message(STATUS "Current CUDA lib path: ${CUDA_LIB}")
   endif(NOT WIN32)
 endif()
 
@@ -208,41 +210,47 @@ endif()
 add_executable(${DEMO_NAME} ${DEMO_NAME}.cc)
 target_link_libraries(${DEMO_NAME} ${DEPS})
 if(WIN32)
+  if("${CMAKE_GENERATOR}" MATCHES "Ninja")
+    set(LIB_PATH ${CMAKE_BINARY_DIR})
+  else()
+    set(LIB_PATH ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE})
+  endif()
+
   if(USE_TENSORRT)
     add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
             COMMAND ${CMAKE_COMMAND} -E copy ${TENSORRT_LIB_DIR}/nvinfer${CMAKE_SHARED_LIBRARY_SUFFIX}
-              ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
+              ${LIB_PATH}
             COMMAND ${CMAKE_COMMAND} -E copy ${TENSORRT_LIB_DIR}/nvinfer_plugin${CMAKE_SHARED_LIBRARY_SUFFIX}
-              ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
+              ${LIB_PATH}
     )
     if(${TENSORRT_MAJOR_VERSION} GREATER_EQUAL 7)
       add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
               COMMAND ${CMAKE_COMMAND} -E copy ${TENSORRT_LIB_DIR}/myelin64_1${CMAKE_SHARED_LIBRARY_SUFFIX}
-                ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE})
+                ${LIB_PATH})
     endif()
   endif()
   if(WITH_MKL)
     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 ${MKLDNN_PATH}/lib/mkldnn.dll  ${CMAKE_BINARY_DIR}/Release
+          COMMAND ${CMAKE_COMMAND} -E copy ${MATH_LIB_PATH}/lib/mklml.dll ${LIB_PATH}
+          COMMAND ${CMAKE_COMMAND} -E copy ${MATH_LIB_PATH}/lib/libiomp5md.dll ${LIB_PATH}
+          COMMAND ${CMAKE_COMMAND} -E copy ${MKLDNN_PATH}/lib/mkldnn.dll  ${LIB_PATH}
     )
   else()
     add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
-          COMMAND ${CMAKE_COMMAND} -E copy ${OPENBLAS_LIB_PATH}/lib/openblas.dll ${CMAKE_BINARY_DIR}/Release
+          COMMAND ${CMAKE_COMMAND} -E copy ${OPENBLAS_LIB_PATH}/lib/openblas.dll ${LIB_PATH}
     )
   endif()
   if(WITH_ONNXRUNTIME)
     add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
     COMMAND ${CMAKE_COMMAND} -E copy ${PADDLE_LIB_THIRD_PARTY_PATH}onnxruntime/lib/onnxruntime.dll
-      ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
+      ${LIB_PATH}
     COMMAND ${CMAKE_COMMAND} -E copy ${PADDLE_LIB_THIRD_PARTY_PATH}paddle2onnx/lib/paddle2onnx.dll
-      ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
+      ${LIB_PATH}
     )
   endif()
   if(NOT WITH_STATIC_LIB)
       add_custom_command(TARGET ${DEMO_NAME} POST_BUILD 
-        COMMAND ${CMAKE_COMMAND} -E copy "${PADDLE_LIB}/paddle/lib/paddle_inference.dll" ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
+        COMMAND ${CMAKE_COMMAND} -E copy "${PADDLE_LIB}/paddle/lib/paddle_inference.dll" ${LIB_PATH}
       )
   endif()
 endif()
diff --git a/paddle/fluid/inference/api/demo_ci/run.sh b/paddle/fluid/inference/api/demo_ci/run.sh
index 2c0945cd5b3..290c547c986 100755
--- a/paddle/fluid/inference/api/demo_ci/run.sh
+++ b/paddle/fluid/inference/api/demo_ci/run.sh
@@ -23,6 +23,7 @@ USE_TENSORRT=$5
 TENSORRT_ROOT_DIR=$6 # TensorRT root dir, default to /usr
 WITH_ONNXRUNTIME=$7
 MSVC_STATIC_CRT=$8
+CUDA_LIB=$9/lib/x64
 inference_install_dir=${PADDLE_ROOT}/build/paddle_inference_install_dir
 WIN_DETECT=$(echo `uname` | grep "Win") # detect current platform
 
@@ -112,16 +113,18 @@ for WITH_STATIC_LIB in ON OFF; do
       continue
     fi
     # -----simple_on_word2vec on windows-----
-    cmake .. -G "Visual Studio 15 2017" -A x64 -T host=x64 -DPADDLE_LIB=${inference_install_dir} \
+    cmake .. -GNinja -DPADDLE_LIB=${inference_install_dir} \
       -DWITH_MKL=$TURN_ON_MKL \
       -DDEMO_NAME=simple_on_word2vec \
       -DWITH_GPU=$TEST_GPU_CPU \
       -DWITH_STATIC_LIB=$WITH_STATIC_LIB \
       -DMSVC_STATIC_CRT=$MSVC_STATIC_CRT \
-      -DWITH_ONNXRUNTIME=$WITH_ONNXRUNTIME
-    msbuild  /maxcpucount /property:Configuration=Release cpp_inference_demo.sln
+      -DWITH_ONNXRUNTIME=$WITH_ONNXRUNTIME \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DCUDA_LIB="$CUDA_LIB"
+    ninja
     for use_gpu in $use_gpu_list; do
-      Release/simple_on_word2vec.exe \
+      ./simple_on_word2vec.exe \
         --dirname=$DATA_DIR/word2vec/word2vec.inference.model \
         --use_gpu=$use_gpu
       if [ $? -ne 0 ]; then
@@ -132,17 +135,19 @@ for WITH_STATIC_LIB in ON OFF; do
 
     # -----vis_demo on windows-----
     rm -rf *
-    cmake .. -G "Visual Studio 15 2017" -A x64 -T host=x64 -DPADDLE_LIB=${inference_install_dir} \
+    cmake .. -GNinja -DPADDLE_LIB=${inference_install_dir} \
       -DWITH_MKL=$TURN_ON_MKL \
       -DDEMO_NAME=vis_demo \
       -DWITH_GPU=$TEST_GPU_CPU \
       -DWITH_STATIC_LIB=$WITH_STATIC_LIB \
       -DMSVC_STATIC_CRT=$MSVC_STATIC_CRT \
-      -DWITH_ONNXRUNTIME=$WITH_ONNXRUNTIME
-    msbuild  /maxcpucount /property:Configuration=Release cpp_inference_demo.sln
+      -DWITH_ONNXRUNTIME=$WITH_ONNXRUNTIME \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DCUDA_LIB="$CUDA_LIB"
+    ninja
     for use_gpu in $use_gpu_list; do
       for vis_demo_name in $vis_demo_list; do
-        Release/vis_demo.exe \
+        ./vis_demo.exe \
           --modeldir=$DATA_DIR/$vis_demo_name/model \
           --data=$DATA_DIR/$vis_demo_name/data.txt \
           --refer=$DATA_DIR/$vis_demo_name/result.txt \
@@ -153,11 +158,11 @@ for WITH_STATIC_LIB in ON OFF; do
         fi
       done
     done
-
+    
     # --------tensorrt mobilenet on windows------
     if [ $USE_TENSORRT == ON -a $TEST_GPU_CPU == ON ]; then
       rm -rf *
-      cmake .. -G "Visual Studio 15 2017" -A x64 -T host=x64 -DPADDLE_LIB=${inference_install_dir} \
+      cmake .. -GNinja -DPADDLE_LIB=${inference_install_dir} \
         -DWITH_MKL=$TURN_ON_MKL \
         -DDEMO_NAME=trt_mobilenet_demo \
         -DWITH_GPU=$TEST_GPU_CPU \
@@ -165,9 +170,11 @@ for WITH_STATIC_LIB in ON OFF; do
         -DMSVC_STATIC_CRT=$MSVC_STATIC_CRT \
         -DUSE_TENSORRT=$USE_TENSORRT \
         -DTENSORRT_ROOT=$TENSORRT_ROOT_DIR \
-        -DWITH_ONNXRUNTIME=$WITH_ONNXRUNTIME
-      msbuild  /maxcpucount /property:Configuration=Release cpp_inference_demo.sln
-      Release/trt_mobilenet_demo.exe \
+        -DWITH_ONNXRUNTIME=$WITH_ONNXRUNTIME \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DCUDA_LIB="$CUDA_LIB"
+      ninja
+      ./trt_mobilenet_demo.exe \
         --modeldir=$DATA_DIR/mobilenet/model \
         --data=$DATA_DIR/mobilenet/data.txt \
         --refer=$DATA_DIR/mobilenet/result.txt 
diff --git a/paddle/fluid/inference/tests/infer_ut/CMakeLists.txt b/paddle/fluid/inference/tests/infer_ut/CMakeLists.txt
index f376cbd4fb3..5c17e2d62d3 100644
--- a/paddle/fluid/inference/tests/infer_ut/CMakeLists.txt
+++ b/paddle/fluid/inference/tests/infer_ut/CMakeLists.txt
@@ -83,13 +83,15 @@ if(WITH_GPU)
   if(NOT WIN32)
     set(CUDA_LIB "/usr/local/cuda/lib64/" CACHE STRING "CUDA Library")
   else()
-    if(NOT DEFINED CUDA_LIB)
+    set(CUDA_LIB "" CACHE STRING "CUDA_LIB")
+    if("${TENSORRT_ROOT}" STREQUAL "")
       if(DEFINED ENV{CUDA_PATH})
         set(CUDA_LIB "$ENV{CUDA_PATH}\\lib\\x64")
       else()
         set(CUDA_LIB "C:\\Program\ Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.2\\lib\\x64")
       endif()
     endif()
+    message(STATUS "Current CUDA lib path: ${CUDA_LIB}")
   endif(NOT WIN32)
 endif()
 
@@ -236,47 +238,54 @@ if(WITH_GTEST)
   include_directories(${GTEST_INSTALL_DIR}/include)
   add_dependencies(${DEMO_NAME} thirdparty_gtest)
   IF(WIN32)
-    target_link_libraries(${DEMO_NAME} ${GTEST_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}/gtest${CMAKE_STATIC_LIBRARY_SUFFIX})
+    target_link_libraries(${DEMO_NAME} ${GTEST_LIBRARIES})
   ELSE()
     target_link_libraries(${DEMO_NAME} ${GTEST_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}/libgtest${CMAKE_STATIC_LIBRARY_SUFFIX})
   ENDIF(WIN32)
 endif()
 if(WIN32)
+  if("${CMAKE_GENERATOR}" MATCHES "Ninja")
+    set(LIB_PATH ${CMAKE_BINARY_DIR})
+  else()
+    set(LIB_PATH ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE})
+  endif()
+
   if(USE_TENSORRT)
     add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
             COMMAND ${CMAKE_COMMAND} -E copy ${TENSORRT_LIB_DIR}/nvinfer${CMAKE_SHARED_LIBRARY_SUFFIX}
-              ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
+              ${LIB_PATH}
             COMMAND ${CMAKE_COMMAND} -E copy ${TENSORRT_LIB_DIR}/nvinfer_plugin${CMAKE_SHARED_LIBRARY_SUFFIX}
-              ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
+              ${LIB_PATH}
     )
     if(${TENSORRT_MAJOR_VERSION} EQUAL 7)
       add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
               COMMAND ${CMAKE_COMMAND} -E copy ${TENSORRT_LIB_DIR}/myelin64_1${CMAKE_SHARED_LIBRARY_SUFFIX}
-                ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE})
+                ${LIB_PATH})
     endif()
   endif()
   if(WITH_MKL)
+    message("LIB_PATH IS ${LIB_PATH}")
     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 ${MKLDNN_PATH}/lib/mkldnn.dll  ${CMAKE_BINARY_DIR}/Release
+          COMMAND ${CMAKE_COMMAND} -E copy ${MATH_LIB_PATH}/lib/mklml.dll ${LIB_PATH}
+          COMMAND ${CMAKE_COMMAND} -E copy ${MATH_LIB_PATH}/lib/libiomp5md.dll ${LIB_PATH}
+          COMMAND ${CMAKE_COMMAND} -E copy ${MKLDNN_PATH}/lib/mkldnn.dll  ${LIB_PATH}
     )
   else()
     add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
-          COMMAND ${CMAKE_COMMAND} -E copy ${OPENBLAS_LIB_PATH}/lib/openblas.dll ${CMAKE_BINARY_DIR}/Release
+          COMMAND ${CMAKE_COMMAND} -E copy ${OPENBLAS_LIB_PATH}/lib/openblas.dll ${LIB_PATH}
     )
   endif()
   if(WITH_ONNXRUNTIME)
     add_custom_command(TARGET ${DEMO_NAME} POST_BUILD
     COMMAND ${CMAKE_COMMAND} -E copy ${PADDLE_LIB_THIRD_PARTY_PATH}onnxruntime/lib/onnxruntime.dll
-      ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
+      ${LIB_PATH}
     COMMAND ${CMAKE_COMMAND} -E copy ${PADDLE_LIB_THIRD_PARTY_PATH}paddle2onnx/lib/paddle2onnx.dll
-      ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
+      ${LIB_PATH}
     )
   endif()
   if(NOT WITH_STATIC_LIB)
       add_custom_command(TARGET ${DEMO_NAME} POST_BUILD 
-        COMMAND ${CMAKE_COMMAND} -E copy "${PADDLE_LIB}/paddle/lib/paddle_inference.dll" ${CMAKE_BINARY_DIR}/${CMAKE_BUILD_TYPE}
+        COMMAND ${CMAKE_COMMAND} -E copy "${PADDLE_LIB}/paddle/lib/paddle_inference.dll" ${LIB_PATH}
       )
   endif()
 endif()
diff --git a/paddle/fluid/inference/tests/infer_ut/external-cmake/gtest-cpp.cmake b/paddle/fluid/inference/tests/infer_ut/external-cmake/gtest-cpp.cmake
index 3e83a21e386..b38984314ec 100644
--- a/paddle/fluid/inference/tests/infer_ut/external-cmake/gtest-cpp.cmake
+++ b/paddle/fluid/inference/tests/infer_ut/external-cmake/gtest-cpp.cmake
@@ -8,10 +8,12 @@ set(GTEST_REPOSITORY     https://github.com/google/googletest.git)
 set(GTEST_TAG            release-1.8.1)
 INCLUDE_DIRECTORIES(${GTEST_INCLUDE_DIR})
 IF(WIN32)
+    # if use CMAKE_INSTALL_LIBDIR, the path of lib actually is install/gtest/lib/gtest.lib but GTEST_LIBRARIES 
+    # is install/gtest/gtest.lib
     set(GTEST_LIBRARIES
-        "${GTEST_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}/gtest.lib" CACHE FILEPATH "gtest libraries." FORCE)
+        "${GTEST_INSTALL_DIR}/lib/gtest.lib" CACHE FILEPATH "gtest libraries." FORCE)
     set(GTEST_MAIN_LIBRARIES
-        "${GTEST_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}/gtest_main.lib" CACHE FILEPATH "gtest main libraries." FORCE)
+        "${GTEST_INSTALL_DIR}/lib/gtest_main.lib" CACHE FILEPATH "gtest main libraries." FORCE)
 ELSE()
     set(GTEST_LIBRARIES
         "${GTEST_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}/libgtest.a" CACHE FILEPATH "gtest libraries." FORCE)
diff --git a/paddle/fluid/inference/tests/infer_ut/run.sh b/paddle/fluid/inference/tests/infer_ut/run.sh
index 8123d378500..331608a2cbc 100755
--- a/paddle/fluid/inference/tests/infer_ut/run.sh
+++ b/paddle/fluid/inference/tests/infer_ut/run.sh
@@ -22,6 +22,7 @@ DATA_DIR=$4 # dataset
 TENSORRT_ROOT_DIR=$5 # TensorRT ROOT dir, default to /usr/local/TensorRT
 WITH_ONNXRUNTIME=$6
 MSVC_STATIC_CRT=$7
+CUDA_LIB=$8/lib/x64
 inference_install_dir=${PADDLE_ROOT}/build/paddle_inference_install_dir
 EXIT_CODE=0 # init default exit code
 WIN_DETECT=$(echo `uname` | grep "Win") # detect current platform
@@ -135,7 +136,7 @@ function compile_test() {
     cd ${build_dir}
     TEST_NAME=$1
     if [ $WIN_DETECT != "" ]; then
-        cmake .. -G "Visual Studio 15 2017" -A x64 -T host=x64 -DPADDLE_LIB=${inference_install_dir} \
+        cmake .. -GNinja -DPADDLE_LIB=${inference_install_dir} \
              -DWITH_MKL=$TURN_ON_MKL \
              -DDEMO_NAME=${TEST_NAME} \
              -DWITH_GPU=$TEST_GPU_CPU \
@@ -146,8 +147,9 @@ function compile_test() {
              -DWITH_GTEST=ON \
              -DCMAKE_CXX_FLAGS='/std:c++17' \
              -DCMAKE_BUILD_TYPE=Release \
-             -DWITH_ONNXRUNTIME=$WITH_ONNXRUNTIME
-        msbuild /maxcpucount /property:Configuration=Release ALL_BUILD.vcxproj
+             -DWITH_ONNXRUNTIME=$WITH_ONNXRUNTIME \
+             -DCUDA_LIB="$CUDA_LIB"
+        ninja
     else
         cmake .. -DPADDLE_LIB=${inference_install_dir} \
                  -DWITH_MKL=$TURN_ON_MKL \
@@ -171,11 +173,7 @@ mkdir -p ${log_dir}
 cd ${build_dir}
 rm -rf *
 
-if [ $WIN_DETECT != "" ]; then
-    exe_dir=${build_dir}/Release
-else
-    exe_dir=${build_dir}
-fi;
+exe_dir=${build_dir}
 
 printf "${YELLOW} start test_resnet50 ${NC} \n";
 compile_test "test_resnet50"
diff --git a/paddle/scripts/paddle_build.bat b/paddle/scripts/paddle_build.bat
index f9ab3f606bf..cc55ea82df6 100644
--- a/paddle/scripts/paddle_build.bat
+++ b/paddle/scripts/paddle_build.bat
@@ -255,7 +255,6 @@ set MSVC_STATIC_CRT=ON
 set ON_INFER=ON
 set WITH_TENSORRT=ON
 set WITH_INFERENCE_API_TEST=ON
-set vcvars64_dir="D:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\Build\vcvars64.bat"
 
 call :cmake || goto cmake_error
 call :build || goto build_error
@@ -711,7 +710,7 @@ echo cmake .. -G %GENERATOR% -DCMAKE_BUILD_TYPE=Release -DWITH_AVX=%WITH_AVX% -D
 -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% ^
+-DWITH_UNITY_BUILD=%WITH_UNITY_BUILD% -DCUDA_ARCH_NAME=%CUDA_ARCH_NAME% -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% >> %work_dir%\win_cmake.sh
 
@@ -753,7 +752,7 @@ for /F %%i in ("%libsize%") do (
 )
 
 cd /d %work_dir%\paddle\fluid\inference\api\demo_ci
-%cache_dir%\tools\busybox64.exe bash run.sh %work_dir:\=/% %WITH_MKL% %WITH_GPU% %cache_dir:\=/%/inference_demo %WITH_TENSORRT% %TENSORRT_ROOT% %WITH_ONNXRUNTIME% %MSVC_STATIC_CRT%
+%cache_dir%\tools\busybox64.exe bash run.sh %work_dir:\=/% %WITH_MKL% %WITH_GPU% %cache_dir:\=/%/inference_demo %WITH_TENSORRT% %TENSORRT_ROOT% %WITH_ONNXRUNTIME% %MSVC_STATIC_CRT% "%CUDA_TOOLKIT_ROOT_DIR%"
 goto:eof
 
 :test_inference_error
@@ -784,7 +783,7 @@ echo    Step 7. Testing fluid library with infer_ut for inference ...
 echo    ========================================
 
 cd /d %work_dir%\paddle\fluid\inference\tests\infer_ut
-%cache_dir%\tools\busybox64.exe bash run.sh %work_dir:\=/% %WITH_MKL% %WITH_GPU% %cache_dir:\=/%/inference_demo %TENSORRT_ROOT% %WITH_ONNXRUNTIME% %MSVC_STATIC_CRT%
+%cache_dir%\tools\busybox64.exe bash run.sh %work_dir:\=/% %WITH_MKL% %WITH_GPU% %cache_dir:\=/%/inference_demo %TENSORRT_ROOT% %WITH_ONNXRUNTIME% %MSVC_STATIC_CRT% "%CUDA_TOOLKIT_ROOT_DIR%"
 goto:eof
 
 :test_inference_ut_error
-- 
GitLab