use busybox run test on windows openblas (#31728)

* use busybox run test on windows openblas

* fix error

* fix disable_quick and nightly lable issue

* add retry on windows openblas

* fix bug

* use one file to run cpu and gpu tests

* fix with grep warning

* fix syntax error

* change run_unittest to run_unittest_gpu

* Update run_unittests.sh

fix error

paddle/scripts/paddle_build.bat

@@ -52,6 +52,8 @@ if not defined WITH_CACHE set WITH_CACHE=OFF
 if not defined WITH_UNITY_BUILD set WITH_UNITY_BUILD=OFF
 if not defined INFERENCE_DEMO_INSTALL_DIR set INFERENCE_DEMO_INSTALL_DIR=%cache_dir:\=/%/inference_demo
 if not defined LOG_LEVEL set LOG_LEVEL=normal
+if not defined PRECISION_TEST set PRECISION_TEST=OFF
+if not defined NIGHTLY_MODE set PRECISION_TEST=OFF
 
 rem -------set cache build directory-----------
 rmdir build\python /s/q
@@ -501,7 +503,7 @@ setlocal enabledelayedexpansion
 set CUDA_DEVICE_COUNT=1
 set FLAGS_fraction_of_gpu_memory_to_use=0.92
 
-%cache_dir%\tools\busybox64.exe bash %work_dir%\tools\windows\run_unittests.sh %NIGHTLY_MODE% %PRECISION_TEST%
+%cache_dir%\tools\busybox64.exe bash %work_dir%\tools\windows\run_unittests.sh %NIGHTLY_MODE% %PRECISION_TEST% %WITH_GPU%
 
 goto:eof
 
@@ -510,7 +512,7 @@ echo    ========================================
 echo    Running CPU unit tests in parallel way ...
 echo    ========================================
 
-ctest.exe -E "(%disable_ut_quickly%)" -LE %nightly_label% --output-on-failure -C Release -j 8 --repeat until-pass:4 after-timeout:4
+%cache_dir%\tools\busybox64.exe bash %work_dir%\tools\windows\run_unittests.sh %NIGHTLY_MODE% %PRECISION_TEST% %WITH_GPU%
 
 goto:eof
 

tools/windows/run_unittests.sh

@@ -16,6 +16,7 @@ set -e
 set +x
 NIGHTLY_MODE=$1
 PRECISION_TEST=$2
+WITH_GPU=$3
 
 export PADDLE_ROOT="$(cd "$PWD/../" && pwd )"
 if [ ${NIGHTLY_MODE:-OFF} == "ON" ]; then
@@ -204,23 +205,26 @@ long_time_test="^best_fit_allocator_test$|\
 ^test_strided_slice_op$|\
 ^test_transpose_op$"
 
-export FLAGS_call_stack_level=2
-export FLAGS_fraction_of_gpu_memory_to_use=0.92
-export CUDA_VISIBLE_DEVICES=0
+if [ ${WITH_GPU:-OFF} == "ON" ];then
+    export FLAGS_call_stack_level=2
+    export FLAGS_fraction_of_gpu_memory_to_use=0.92
+    export CUDA_VISIBLE_DEVICES=0
 
-UT_list=$(ctest -N | awk -F ': ' '{print $2}' | sed '/^$/d' | sed '$d')
-num=$(ctest -N | awk -F ': ' '{print $2}' | sed '/^$/d' | sed '$d' | wc -l)
-echo "Windows 1 card TestCases count is $num"
-if [ ${PRECISION_TEST:-OFF} == "ON" ]; then
+    UT_list=$(ctest -N | awk -F ': ' '{print $2}' | sed '/^$/d' | sed '$d')
+    num=$(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
         python ${PADDLE_ROOT}/tools/get_pr_ut.py
         if [[ -f "ut_list" ]]; then
+            set +x
             echo "PREC length: "`wc -l ut_list`
             precision_cases=`cat ut_list`
+            set -x
+        fi
     fi
-fi
 
-set +e
-if [ ${PRECISION_TEST:-OFF} == "ON" ] && [[ "$precision_cases" != "" ]];then
+    set +e
+    if [ ${PRECISION_TEST:-OFF} == "ON" ] && [[ "$precision_cases" != "" ]];then
         UT_list_prec=''
         re=$(cat ut_list|awk -F ' ' '{print }' | awk 'BEGIN{ all_str=""}{if (all_str==""){all_str=$1}else{all_str=all_str"$|^"$1}} END{print "^"all_str"$"}')
         for case in $UT_list; do
@@ -236,15 +240,15 @@ if [ ${PRECISION_TEST:-OFF} == "ON" ] && [[ "$precision_cases" != "" ]];then
             fi
         done
         UT_list=$UT_list_prec
-fi
-set -e
-
-output=$(python ${PADDLE_ROOT}/tools/parallel_UT_rule.py "${UT_list}")
-cpu_parallel_job=$(echo $output | cut -d ";" -f 1)
-tetrad_parallel_job=$(echo $output | cut -d ";" -f 2)
-two_parallel_job=$(echo $output | cut -d ";" -f 3)
-non_parallel_job=$(echo $output | cut -d ";" -f 4)
+    fi
+    set -e
 
+    output=$(python ${PADDLE_ROOT}/tools/parallel_UT_rule.py "${UT_list}")
+    cpu_parallel_job=$(echo $output | cut -d ";" -f 1)
+    tetrad_parallel_job=$(echo $output | cut -d ";" -f 2)
+    two_parallel_job=$(echo $output | cut -d ";" -f 3)
+    non_parallel_job=$(echo $output | cut -d ";" -f 4)
+fi
 
 failed_test_lists=''
 tmp_dir=`mktemp -d`
@@ -264,7 +268,13 @@ function collect_failed_tests() {
     set -e
 }
 
-function run_unittest() {
+function run_unittest_cpu() {
+    tmpfile=$tmp_dir/$RANDOM
+    (ctest -E "${disable_ut_quickly}" -LE "${nightly_label}" --output-on-failure -C Release -j 8 | tee $tmpfile) &
+    wait;
+}
+
+function run_unittest_gpu() {
     test_case=$1
     parallel_job=$2
     parallel_level_base=${CTEST_PARALLEL_LEVEL:-1}
@@ -283,7 +293,11 @@ function run_unittest() {
 }
 
 function unittests_retry(){
+    if [ "${WITH_GPU:-OFF}" == "ON" ];then
         parallel_job=1
+    else
+        parallel_job=4
+    fi
     is_retry_execuate=0
     wintest_error=1
     retry_time=3
@@ -334,7 +348,7 @@ function unittests_retry(){
 
 function show_ut_retry_result() {
     if [[ "$is_retry_execuate" != "0" ]];then
-        failed_test_lists_ult=`echo "${failed_test_lists}" | grep -o '[^ ].*$'`
+        failed_test_lists_ult=`echo "${failed_test_lists}"`
         echo "========================================="
         echo "There are more than 10 failed unit tests, so no unit test retry!!!"
         echo "========================================="
@@ -363,10 +377,15 @@ function show_ut_retry_result() {
 }
 
 set +e
-run_unittest $cpu_parallel_job 12
-run_unittest $tetrad_parallel_job 4
-run_unittest $two_parallel_job 2
-run_unittest $non_parallel_job
+
+if [ "${WITH_GPU:-OFF}" == "ON" ];then
+    run_unittest_gpu $cpu_parallel_job 12
+    run_unittest_gpu $tetrad_parallel_job 4
+    run_unittest_gpu $two_parallel_job 2
+    run_unittest_gpu $non_parallel_job
+else
+    run_unittest_cpu
+fi
 collect_failed_tests
 set -e
 rm -f $tmp_dir/*