[new-exec] add standalone executor test (#38101)

* refine test

* add download_program target

* update ut code

* refine code

* disable profiler

* add comments

* refine cmake

* skip coverage ci

paddle/fluid/framework/new_executor/CMakeLists.txt

@@ -12,4 +12,56 @@ cc_library(stream_analyzer SRCS stream_analyzer.cc DEPS ${DEVICE_EVENT_LIBS} glo
 cc_library(interpretercore SRCS interpretercore.cc DEPS workqueue ${DEVICE_EVENT_LIBS} interpretercore_util interpretercore_garbage_collector stream_analyzer event_manager)
 cc_library(standalone_executor SRCS standalone_executor.cc DEPS interpretercore)
 cc_test(workqueue_test SRCS workqueue_test.cc DEPS workqueue)
-# cc_binary(standalone_executor_test SRCS standalone_executor_test.cc DEPS interpretercore standalone_executor operator op_registry executor ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} profiler)
+
+# skip win32 since wget is not installed by default on windows machine.
+# skip COVERAGE_CI since the test runs slowly because of instrumentation.
+if (WITH_TESTING AND NOT WIN32 AND NOT WITH_COVERAGE)
+    add_custom_target(
+        download_program
+        COMMAND wget -nc https://paddle-ci.gz.bcebos.com/new_exec/lm_main_program 
+        COMMAND wget -nc https://paddle-ci.gz.bcebos.com/new_exec/lm_startup_program 
+    )
+    
+    # all operators used in the program
+    set(OPS  
+        fill_constant_op
+        uniform_random_op
+        lookup_table_op
+        transpose_op
+        reshape_op
+        split_op
+        slice_op
+        concat_op
+        matmul_op
+        elementwise_add_op
+        elementwise_mul_op
+        softmax_with_cross_entropy_op
+        reduce_mean_op
+        reduce_sum_op
+        activation_op
+        sum_op
+        elementwise_max_op
+        elementwise_div_op
+        sgd_op
+        squared_l2_norm_op
+        memcpy_h2d_op
+        memcpy_d2h_op)
+    
+    # All deps of the operators above, part of GLOB_OPERATOR_DEPS.
+    set(OP_DEPS 
+        generator
+        softmax
+        selected_rows_functor
+        jit_kernel_helper
+        concat_and_split
+        cross_entropy)
+
+    cc_test(standalone_executor_test SRCS standalone_executor_test.cc DEPS interpretercore standalone_executor operator op_registry executor ${OPS} ${OP_DEPS})
+    set_tests_properties(standalone_executor_test PROPERTIES TIMEOUT 100)
+
+    add_dependencies(standalone_executor_test download_program)
+    if (WITH_PROFILER)
+        target_link_libraries(standalone_executor_test profiler)
+        add_dependencies(standalone_executor_test profiler)
+    endif()
+endif()

paddle/fluid/framework/new_executor/standalone_executor_test.cc

@@ -12,15 +12,11 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <gtest/gtest.h>
+#include <chrono>
 #include <iostream>
 #include <string>
 
-#include <chrono>
-#include <map>
-#include <memory>
-#include <unordered_map>
-#include <vector>
-
 // #include "gperftools/profiler.h"
 
 #include "paddle/fluid/framework/new_executor/standalone_executor.h"
@@ -63,8 +59,12 @@ USE_OP(sgd);
 USE_OP(squared_l2_norm);
 USE_OP(memcpy_h2d);
 USE_OP(memcpy_d2h);
+DECLARE_double(eager_delete_tensor_gb);
+
+namespace paddle {
+namespace framework {
 
-paddle::framework::ProgramDesc load_from_file(const std::string& file_name) {
+ProgramDesc load_from_file(const std::string& file_name) {
   std::ifstream fin(file_name, std::ios::in | std::ios::binary);
   fin.seekg(0, std::ios::end);
   std::string buffer(fin.tellg(), ' ');
@@ -72,18 +72,16 @@ paddle::framework::ProgramDesc load_from_file(const std::string& file_name) {
   fin.read(&buffer[0], buffer.size());
   fin.close();
 
-  paddle::framework::ProgramDesc program_desc(buffer);
+  ProgramDesc program_desc(buffer);
   return program_desc;
 }
 
-int main(int argc, char* argv[]) {
-  paddle::framework::InitDevices();
-  std::cout << "main" << std::endl;
-  int64_t batch_size = std::stoi(argv[1]);
-  paddle::framework::InitDevices();
-  auto place = paddle::platform::CUDAPlace(0);
-  auto test_prog = load_from_file("lm_startup_program");
+TEST(StandaloneExecutor, run) {
+  FLAGS_eager_delete_tensor_gb = 0.1;
+  int64_t batch_size = 20;
 
+  auto place = platform::CUDAPlace(0);
+  auto test_prog = load_from_file("lm_startup_program");
   auto main_prog = load_from_file("lm_main_program");
 
   auto& global_block = main_prog.Block(0);
@@ -103,12 +101,13 @@ int main(int argc, char* argv[]) {
   shape3[0] = batch_size;
   op3->SetAttr("shape", shape3);
 
-  paddle::framework::Scope scope;
-  paddle::framework::StandaloneExecutor exec(place, test_prog, main_prog,
-                                             &scope);
-
+  Scope scope;
+  StandaloneExecutor exec(place, test_prog, main_prog, &scope);
+  exec.Run({}, {}, {});
   auto start = std::chrono::steady_clock::now();
+
   // ProfilerStart("new_executor.prof");
+
   for (size_t i = 0; i < 2320; ++i) {
     if (i % 200 == 0) {
       std::cout << i << std::endl;
@@ -116,11 +115,15 @@ int main(int argc, char* argv[]) {
 
     exec.Run({}, {}, {});
   }
+
   // ProfilerStop();
+
   auto end = std::chrono::steady_clock::now();
   std::chrono::duration<double> diff = end - start;
 
   std::cout << "time cost " << diff.count() << std::endl;
-
-  return 1;
+  ASSERT_LT(diff.count(), 30);
 }
+
+}  // namespace framework
+}  // namespace paddle