modify for wincheck-inference case (#37292)

* for pure fp16

* opt topk

* modify for wincheck-inference case

* modify according to zhouwei's comment

* modify according to zhouwei's comment 2nd time
Co-authored-by: Nzhangkaihuo <zhangkaihuo@baidu.com>

paddle/scripts/paddle_build.bat

@@ -256,11 +256,14 @@ set WITH_GPU=ON
 set WITH_AVX=ON
 set MSVC_STATIC_CRT=ON
 set ON_INFER=ON
+set WITH_TESTING=ON
+set WITH_TENSORRT=ON
+set WITH_INFERENCE_API_TEST=ON
 
 call :cmake || goto cmake_error
 call :build || goto build_error
 call :test_whl_pacakage || goto test_whl_pacakage_error
-:: call :test_unit || goto test_unit_error
+call :test_unit || goto test_unit_error
 ::call :test_inference || goto test_inference_error
 :: call :check_change_of_unittest || goto check_change_of_unittest_error
 goto:success

paddle/scripts/paddle_build.sh

@@ -1227,8 +1227,8 @@ set +x
         is_nightly=''             # indicate whether the case will only run at night
         get_quickly_disable_ut||disable_ut_quickly=''    # indicate whether the case was in quickly disable list
 
-        UT_list=$(ctest -N | awk -F ': ' '{print $2}' | sed '/^$/d' | sed '$d')
-        output=$(python ${PADDLE_ROOT}/tools/parallel_UT_rule.py "${UT_list}")
+        ctest -N | awk -F ': ' '{print $2}' | sed '/^$/d' | sed '$d' > all_ut_list
+        output=$(python ${PADDLE_ROOT}/tools/parallel_UT_rule.py)
         high_parallel_job=$(echo $output | cut -d ";" -f 1)
         secondary_high_parallel_job=$(echo $output | cut -d ";" -f 2)
         third_high_parallel_job=$(echo $output | cut -d ";" -f 3)

tools/parallel_UT_rule.py

@@ -2205,7 +2205,12 @@ def main():
     lowest_high_parallel_job = '^job$'
     non_parallel_job = '^job$'
 
-    test_cases = sys.argv[1]
+    # sys.argv[1] may exceed max_arg_length when busybox run parallel_UT_rule in windows
+    BUILD_DIR = os.getcwd()
+    file_path = os.path.join(BUILD_DIR, 'all_ut_list')
+    with open(file_path, 'r') as f:
+        test_cases = f.read()
+
     test_cases = test_cases.split("\n")
 
     if platform.system() == 'Windows':

tools/windows/get_prec_ut_list.py

@@ -14,6 +14,8 @@
 """To get a list of prec ut """
 
 import sys
+import os
+import platform
 
 
 def get_prec_ut_list(all_test_cases, prec_test_cases):
@@ -24,7 +26,7 @@ def get_prec_ut_list(all_test_cases, prec_test_cases):
     prec_test_cases_list_new = [item.rstrip() for item in prec_test_cases_list]
 
     if len(prec_test_cases) == 0:
-        return "\n".join(all_test_cases_list)
+        return
 
     case_to_run = ['test_prec_ut']
     for case in all_test_cases_list_new:
@@ -32,14 +34,20 @@ def get_prec_ut_list(all_test_cases, prec_test_cases):
             case_to_run.append(case)
         else:
             print("{} will not run in PRECISION_TEST mode.".format(case))
-    for case in case_to_run:
-        print(case)
+
+    with open(file_path, 'w') as f:
+        f.write('\n'.join(case_to_run))
 
 
 if __name__ == '__main__':
     # get prec cases lists
     with open('ut_list', 'r') as f:
         prec_test_cases = f.read()
-    all_test_cases = sys.argv[1]
+
+    # sys.argv[1] may exceed max_arg_length when busybox run parallel_UT_rule in windows
+    BUILD_DIR = os.getcwd()
+    file_path = os.path.join(BUILD_DIR, 'all_ut_list')
+    with open(file_path, 'r') as f:
+        all_test_cases = f.read()
     #prec_test_cases = sys.argv[2]
     get_prec_ut_list(all_test_cases, prec_test_cases)

tools/windows/run_unittests.sh

@@ -157,7 +157,7 @@ long_time_test="^test_gru_op$|\
 if [ ${WITH_GPU:-OFF} == "ON" ];then
     export CUDA_VISIBLE_DEVICES=0
 
-    UT_list=$(ctest -N | awk -F ': ' '{print $2}' | sed '/^$/d' | sed '$d')
+    ctest -N | awk -F ': ' '{print $2}' | sed '/^$/d' | sed '$d' > all_ut_list
     num=$(ctest -N | awk -F ': ' '{print $2}' | sed '/^$/d' | sed '$d' | wc -l)
     echo "Windows 1 card TestCases count is $num"
     if [ ${PRECISION_TEST:-OFF} == "ON" ]; then
@@ -165,19 +165,14 @@ if [ ${WITH_GPU:-OFF} == "ON" ];then
         if [[ -f "ut_list" ]]; then
             echo "PREC length: "`wc -l ut_list`
             precision_cases=`cat ut_list`
+            if [[ "$precision_cases" != "" ]];then
+                python ${PADDLE_ROOT}/tools/windows/get_prec_ut_list.py
+            fi
         fi
     fi
 
-    set +e
-    if [ ${PRECISION_TEST:-OFF} == "ON" ] && [[ "$precision_cases" != "" ]];then
-        UT_list_res=$(python ${PADDLE_ROOT}/tools/windows/get_prec_ut_list.py "$UT_list" )
-        UT_list_prec=$(echo "${UT_list_res}" | grep -v 'PRECISION_TEST')
-        echo "${UT_list_res}" | grep 'PRECISION_TEST'
-        UT_list=$UT_list_prec
-    fi
-    set -e
-
-    output=$(python ${PADDLE_ROOT}/tools/parallel_UT_rule.py "${UT_list}")
+    # sys.argv[1] may exceed max_arg_length when busybox run parallel_UT_rule in windows
+    output=$(python ${PADDLE_ROOT}/tools/parallel_UT_rule.py)
     cpu_parallel_job=$(echo $output | cut -d ";" -f 1)
     tetrad_parallel_job=$(echo $output | cut -d ";" -f 2)
     two_parallel_job=$(echo $output | cut -d ";" -f 3)
@@ -319,6 +314,10 @@ if [ "${WITH_GPU:-OFF}" == "ON" ];then
             echo "========================================"
             exit 8;
         fi
+        if nvcc --version | grep 11.2; then
+            echo "Only test added_ut temporarily when running in CI-Windows-inference of CUDA 11.2."
+            exit 0;
+        fi
     fi
     run_unittest_gpu $cpu_parallel_job 10
     run_unittest_gpu $tetrad_parallel_job 4