fix cinn_instruction_run_op_test when FLAGS_use_system_allocator=True (#47731)

paddle/fluid/operators/cinn/cinn_instruction_run_op_test.cc

@@ -39,92 +39,85 @@ namespace paddle::operators {
 
 using framework::paddle2cinn::CinnCompiler;
 
-TEST(CinnInstructionOpTest, TestWithElementwiseAdd) {
-  paddle::framework::InitDevices();
-  platform::SetNumThreads(1);
-  // cache test graph into CinnCompiler
-  const std::string& test_op_out_name = "cinn_instruction_run_op_out";
-  const std::string& add_op_out_name = "add_op_out";
-  auto compilation_key = CinnCompiler::GetInstance()->AddGraph(
-      CreateOnlyElementwiseAddGraph("x", "y", test_op_out_name));
-
-  // create necessary ops
-  auto cinn_instruction_run_op = paddle::framework::OpRegistry::CreateOp(
-      "cinn_instruction_run",
-      {{"X", {"x", "y"}}},
-      {{"Out", {test_op_out_name}}},
-      {{"cached_index", 0}, {"instruction_index", 0}});
-
-  auto cinn_launch_op = paddle::framework::OpRegistry::CreateOp(
-      "cinn_launch",
-      {{"X", {"x", "y"}}},
-      {{"Out", {test_op_out_name}}},
-      {{"compilation_key", compilation_key}});
-
-  // check case: a compiled object not cached before cinn_launch_op run,
-  // so a cinn_instruction_run_op will throw an error
-  framework::Scope scope;
-  platform::CPUPlace place;
-  InitVariablesWithRandomValue<float>({"x", "y"}, {10, 20}, place, &scope);
-  scope.Var(test_op_out_name)->GetMutable<LoDTensor>();
-  ASSERT_THROW(cinn_instruction_run_op->Run(scope, place),
-               paddle::platform::EnforceNotMet);
-  // run cinn_launch_op firstly to launch the compilation
-  // of the above graph and cache two compiled results
-  // of both type float and int
-  cinn_launch_op->Run(scope, place);
-  scope.EraseVars({"x", "y", test_op_out_name});
-  scope.Var(test_op_out_name)->GetMutable<LoDTensor>();
-  InitVariablesWithRandomValue<int>({"x", "y"}, {30, 40}, place, &scope);
-  cinn_launch_op->Run(scope, place);
-
-  // Run ops and check the computation results
-  auto run_and_check_fn = [&](const platform::Place& place) {
-    framework::Scope scope;
-    scope.Var(test_op_out_name)->GetMutable<LoDTensor>();
-    scope.Var(add_op_out_name)->GetMutable<LoDTensor>();
-    auto elementwise_add_op =
+class TestCinnInstructionRunOp : public ::testing::Test {
+ public:
+  const char* test_op_out_name = "test_op_out";
+  const char* add_op_out_name = "add_op_out";
+  std::unique_ptr<framework::OperatorBase> cinn_launch_op;
+  std::unique_ptr<framework::OperatorBase> cinn_instruction_run_op;
+  std::unique_ptr<framework::OperatorBase> elementwise_add_op;
+
+  void SetUp() override {
+    auto compilation_key = CinnCompiler::GetInstance()->AddGraph(
+        CreateOnlyElementwiseAddGraph("x", "y", test_op_out_name));
+
+    // create necessary ops
+    cinn_launch_op = paddle::framework::OpRegistry::CreateOp(
+        "cinn_launch",
+        {{"X", {"x", "y"}}},
+        {{"Out", {test_op_out_name}}},
+        {{"compilation_key", compilation_key}});
+    cinn_instruction_run_op = paddle::framework::OpRegistry::CreateOp(
+        "cinn_instruction_run",
+        {{"X", {"x", "y"}}},
+        {{"Out", {test_op_out_name}}},
+        {{"cached_index", 0}, {"instruction_index", 0}});
+    elementwise_add_op =
         paddle::framework::OpRegistry::CreateOp("elementwise_add",
                                                 {{"X", {"x"}}, {"Y", {"y"}}},
                                                 {{"Out", {add_op_out_name}}},
                                                 {{}});
+  }
+
+  void Compile(const platform::Place& place) {
+    // check case: a compiled object not cached before cinn_launch_op run,
+    // so a cinn_instruction_run_op will throw an error
+    framework::Scope scope;
+    InitVariablesWithRandomValue<float>({"x", "y"}, {10, 20}, place, &scope);
+    scope.Var(test_op_out_name)->GetMutable<LoDTensor>();
+    ASSERT_THROW(cinn_instruction_run_op->Run(scope, place),
+                 paddle::platform::EnforceNotMet);
+
+    // run cinn_launch_op firstly to launch the compilation
+    // of the above graph and cache two compiled results
+    // of both type float and int
+    cinn_launch_op->Run(scope, place);
+    scope.EraseVars({"x", "y", test_op_out_name});
+    scope.Var(test_op_out_name)->GetMutable<LoDTensor>();
+    InitVariablesWithRandomValue<int>({"x", "y"}, {30, 40}, place, &scope);
+    cinn_launch_op->Run(scope, place);
+  }
 
-    // 1. check on type float
+  void RunAndCheck(const platform::Place& place) {
+    // Run ops and check the computation results
+    framework::Scope scope;
     InitVariablesWithRandomValue<float>({"x", "y"}, {10, 20}, place, &scope);
-    cinn_instruction_run_op->SetAttr("cached_index", 0);
-    cinn_instruction_run_op->Run(scope, place);
+    scope.Var(test_op_out_name)->GetMutable<LoDTensor>();
+    scope.Var(add_op_out_name)->GetMutable<LoDTensor>();
     elementwise_add_op->Run(scope, place);
+    cinn_launch_op->Run(scope, place);
     CompareOpResult<float>(scope.GetVar(test_op_out_name),
                            scope.GetVar(add_op_out_name));
+  }
 
-    // 2. check on type int to indicate cinn_instruction_run op
-    // can mutable data according compiled result
-    scope.EraseVars({"x", "y", test_op_out_name, add_op_out_name});
-    scope.Var(test_op_out_name)->GetMutable<LoDTensor>();
-    scope.Var(add_op_out_name)->GetMutable<LoDTensor>();
+  void TearDown() override { CinnCompiler::GetInstance()->Clear(); }
+};
 
-    InitVariablesWithRandomValue<int>({"x", "y"}, {30, 40}, place, &scope);
-    cinn_instruction_run_op->SetAttr("cached_index", 1);
-    cinn_instruction_run_op->Run(scope, place);
-    // need reconstruct elementwise_add_op to choose a new kernel with type int
-    elementwise_add_op =
-        paddle::framework::OpRegistry::CreateOp("elementwise_add",
-                                                {{"X", {"x"}}, {"Y", {"y"}}},
-                                                {{"Out", {add_op_out_name}}},
-                                                {{}});
-    elementwise_add_op->Run(scope, place);
-    CompareOpResult<int>(scope.GetVar(test_op_out_name),
-                         scope.GetVar(add_op_out_name));
-  };
+TEST_F(TestCinnInstructionRunOp, CPU) {
+  platform::CPUPlace place;
+  Compile(place);
+  RunAndCheck(place);
+  // the second run on the same place is to check the cache logic
+  RunAndCheck(place);
+}
 
-  // CPU
-  run_and_check_fn(platform::CPUPlace());
-  run_and_check_fn(platform::CPUPlace());
 #ifdef PADDLE_WITH_CUDA
-  // GPU
-  run_and_check_fn(platform::CUDAPlace());
-  run_and_check_fn(platform::CUDAPlace());
-#endif
+TEST_F(TestCinnInstructionRunOp, GPU) {
+  platform::CUDAPlace place;
+  Compile(place);
+  RunAndCheck(place);
+  RunAndCheck(place);
 }
+#endif
 
 }  // namespace paddle::operators

paddle/fluid/operators/cinn/cinn_launch_op_test.cc

@@ -49,7 +49,7 @@ using framework::paddle2cinn::CinnCompiler;
 
 class TestCinnLaunchOp : public ::testing::Test {
  public:
-  const char* test_op_out_name = "add_op_out";
+  const char* test_op_out_name = "test_op_out";
   const char* add_op_out_name = "add_op_out";
   std::unique_ptr<framework::OperatorBase> cinn_launch_op;
   std::unique_ptr<framework::OperatorBase> elementwise_add_op;

paddle/fluid/operators/cinn/test_helper.h

@@ -102,6 +102,7 @@ void InitVariablesWithRandomValue(const std::vector<std::string>& var_names,
       tmp_tensor.mutable_data<DataType>(common_ddim, platform::CPUPlace());
   for (const auto& var_name : var_names) {
     auto* tensor = scope->Var(var_name)->GetMutable<LoDTensor>();
+    tensor->mutable_data<DataType>(common_ddim, place);
     for (auto i = 0; i < tensor->numel(); ++i) {
       tmp_data[i] = static_cast<DataType>(dist(engine));
     }
@@ -121,9 +122,9 @@ void CompareOpResult(Variable* test_out, Variable* expected_out) {
   ASSERT_TRUE(expected_tensor.IsInitialized());
   ASSERT_EQ(test_tensor.dims(), expected_tensor.dims());
   const auto* test_data = test_tensor.data<DataType>();
-  const auto* excepted_data = expected_tensor.data<DataType>();
+  const auto* expected_data = expected_tensor.data<DataType>();
   for (auto i = 0; i < expected_tensor.numel(); ++i) {
-    EXPECT_EQ(test_data[i], excepted_data[i]);
+    EXPECT_EQ(test_data[i], expected_data[i]);
   }
 }