diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt
index d0cbd7801ddaffd4b77e521c8e208d4a8b14ebf2..19a7b4389b7d16f1ac7646c620ebb3741a888585 100644
--- a/paddle/fluid/framework/CMakeLists.txt
+++ b/paddle/fluid/framework/CMakeLists.txt
@@ -399,7 +399,7 @@ cc_library(save_load_util SRCS save_load_util.cc DEPS tensor scope layer)
 cc_test(save_load_util_test SRCS save_load_util_test.cc DEPS save_load_util tensor scope layer)
 cc_library(generator SRCS generator.cc DEPS enforce place)
 
-cc_library(pten_utils SRCS pten_utils.cc DEPS lod_tensor selected_rows place pten var_type_traits pten_hapi_utils op_info)
+cc_library(pten_utils SRCS pten_utils.cc DEPS lod_tensor selected_rows place pten var_type_traits pten_api_utils op_info)
 
 # Get the current working branch
 execute_process(
diff --git a/paddle/pten/api/CMakeLists.txt b/paddle/pten/api/CMakeLists.txt
index 4b427b3b4a38349f672aa8c643bd06ad7f39d360..387da3bc68f120f7ec3efb9977e1c99c859ac2bd 100644
--- a/paddle/pten/api/CMakeLists.txt
+++ b/paddle/pten/api/CMakeLists.txt
@@ -1,3 +1,3 @@
 add_subdirectory(lib)
 
-cc_library(pten_hapi SRCS all.cc DEPS linalg_api math_api creation_api)
+cc_library(pten_api SRCS all.cc DEPS linalg_api math_api creation_api manipulation_api)
diff --git a/paddle/pten/api/lib/utils/CMakeLists.txt b/paddle/pten/api/lib/utils/CMakeLists.txt
index c89ef812846adb337c2abbd6b55854cec3ca0b97..34f8da94c661151494715e8819c002fb5594ecb6 100644
--- a/paddle/pten/api/lib/utils/CMakeLists.txt
+++ b/paddle/pten/api/lib/utils/CMakeLists.txt
@@ -1,4 +1 @@
-add_subdirectory(tests)
-
-cc_library(pten_hapi_utils SRCS allocator.cc storage.cc tensor_utils.cc DEPS tensor_base convert_utils
-dense_tensor lod_tensor selected_rows place var_type_traits)
+cc_library(pten_api_utils SRCS allocator.cc storage.cc tensor_utils.cc DEPS tensor_base convert_utils dense_tensor lod_tensor selected_rows place var_type_traits)
diff --git a/paddle/pten/api/lib/utils/tests/CMakeLists.txt b/paddle/pten/api/lib/utils/tests/CMakeLists.txt
deleted file mode 100644
index 8ac30a1fa6909a3e6cf9a404ab4a6cc9663cc487..0000000000000000000000000000000000000000
--- a/paddle/pten/api/lib/utils/tests/CMakeLists.txt
+++ /dev/null
@@ -1,2 +0,0 @@
-cc_test(test_framework_storage SRCS test_storage.cc DEPS pten_hapi_utils)
-cc_test(test_framework_tensor_utils SRCS test_tensor_utils.cc DEPS pten_hapi_utils)
diff --git a/paddle/pten/tests/CMakeLists.txt b/paddle/pten/tests/CMakeLists.txt
index 3d2da6a5afdd1a6b780074d6b272a93346c898f8..9946821581c3af86c4f645954c0b927429665b1c 100644
--- a/paddle/pten/tests/CMakeLists.txt
+++ b/paddle/pten/tests/CMakeLists.txt
@@ -1,15 +1,4 @@
+add_subdirectory(api)
+add_subdirectory(common)
 add_subdirectory(core)
-add_subdirectory(utils)
-
-cc_test(pten_backend_test SRCS backend_test.cc DEPS gtest)
-cc_test(pten_data_layout_test SRCS data_layout_test.cc DEPS gtest)
-cc_test(pten_data_type_test SRCS data_type_test.cc DEPS gtest)
-cc_test(dense_tensor_test SRCS dense_tensor_test.cc DEPS dense_tensor)
-cc_test(kernel_factory_test SRCS kernel_factory_test.cc DEPS kernel_factory)
-cc_test(test_mean_api SRCS test_mean_api.cc DEPS math_api pten_hapi_utils)
-cc_test(test_dot_api SRCS test_dot_api.cc DEPS linalg_api pten_hapi_utils)
-cc_test(test_matmul_api SRCS test_matmul_api.cc DEPS linalg_api pten_hapi_utils)
-cc_test(test_fill_api SRCS test_fill_api.cc DEPS creation_api pten_hapi_utils)
-cc_test(test_copy_api SRCS test_copy_api.cc DEPS utils_cpu pten_hapi_utils)
-cc_test(test_flatten_api SRCS test_flatten_api.cc DEPS utils_cpu manipulation_api pten_hapi_utils)
-cc_test(test_scale_api SRCS test_scale_api.cc DEPS math_api pten_hapi_utils)
+add_subdirectory(kernels)
diff --git a/paddle/pten/tests/api/CMakeLists.txt b/paddle/pten/tests/api/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..2c6bd9c45d18a7f740376470b88278e5f3b7a2b1
--- /dev/null
+++ b/paddle/pten/tests/api/CMakeLists.txt
@@ -0,0 +1,7 @@
+cc_test(test_mean_api SRCS test_mean_api.cc DEPS pten_api pten_api_utils)
+cc_test(test_dot_api SRCS test_dot_api.cc DEPS pten_api pten_api_utils)
+cc_test(test_matmul_api SRCS test_matmul_api.cc DEPS pten_api pten_api_utils)
+cc_test(test_fill_api SRCS test_fill_api.cc DEPS pten_api pten_api_utils)
+cc_test(test_flatten_api SRCS test_flatten_api.cc DEPS pten_api pten_api_utils)
+cc_test(test_framework_storage SRCS test_storage.cc DEPS pten_api_utils)
+cc_test(test_framework_tensor_utils SRCS test_tensor_utils.cc DEPS pten_api_utils)
diff --git a/paddle/pten/tests/test_dot_api.cc b/paddle/pten/tests/api/test_dot_api.cc
similarity index 58%
rename from paddle/pten/tests/test_dot_api.cc
rename to paddle/pten/tests/api/test_dot_api.cc
index f79854e557e5a7786eeea48143848a73ceced415..2c7b65f98135be8b464badcfa0114d9c98e0d52d 100644
--- a/paddle/pten/tests/test_dot_api.cc
+++ b/paddle/pten/tests/api/test_dot_api.cc
@@ -21,8 +21,6 @@ limitations under the License. */
 #include "paddle/pten/core/dense_tensor.h"
 #include "paddle/pten/core/kernel_registry.h"
 
-#include "paddle/pten/include/linalg.h"
-
 PT_DECLARE_MODULE(LinalgCPU);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
@@ -84,55 +82,3 @@ TEST(API, dot) {
   ASSERT_NEAR(expect_result[1], actual_result1, 1e-6f);
   ASSERT_NEAR(expect_result[2], actual_result2, 1e-6f);
 }
-
-// TODO(YuanRisheng) This unitest should be created in other file.
-//                   It is convenient to make compilation decoupling.
-TEST(DEV_API, dot) {
-  // 1. create tensor
-  const auto alloc = std::make_shared<paddle::experimental::DefaultAllocator>(
-      paddle::platform::CPUPlace());
-  pten::DenseTensor dense_x(alloc,
-                            pten::DenseTensorMeta(pten::DataType::FLOAT32,
-                                                  framework::make_ddim({3, 10}),
-                                                  pten::DataLayout::NCHW));
-  auto* dense_x_data = dense_x.mutable_data<float>();
-
-  pten::DenseTensor dense_y(alloc,
-                            pten::DenseTensorMeta(pten::DataType::FLOAT32,
-                                                  framework::make_ddim({3, 10}),
-                                                  pten::DataLayout::NCHW));
-  auto* dense_y_data = dense_y.mutable_data<float>();
-
-  float sum[3] = {0.0, 0.0, 0.0};
-  for (size_t i = 0; i < 3; ++i) {
-    for (size_t j = 0; j < 10; ++j) {
-      dense_x_data[i * 10 + j] = (i * 10 + j) * 1.0;
-      dense_y_data[i * 10 + j] = (i * 10 + j) * 1.0;
-      sum[i] += (i * 10 + j) * (i * 10 + j) * 1.0;
-    }
-  }
-
-  paddle::platform::DeviceContextPool& pool =
-      paddle::platform::DeviceContextPool::Instance();
-  auto* dev_ctx = pool.Get(paddle::platform::CPUPlace());
-
-  // 2. test API
-  auto out = pten::Dot<float>(
-      *(static_cast<paddle::platform::CPUDeviceContext*>(dev_ctx)),
-      dense_x,
-      dense_y);
-
-  // 3. check result
-  ASSERT_EQ(out.dims().size(), 2);
-  ASSERT_EQ(out.dims()[0], 3);
-  ASSERT_EQ(out.meta().type, pten::DataType::FLOAT32);
-  ASSERT_EQ(out.meta().layout, pten::DataLayout::NCHW);
-
-  auto expect_result = sum;
-  auto actual_result0 = out.data<float>()[0];
-  auto actual_result1 = out.data<float>()[1];
-  auto actual_result2 = out.data<float>()[2];
-  ASSERT_NEAR(expect_result[0], actual_result0, 1e-6f);
-  ASSERT_NEAR(expect_result[1], actual_result1, 1e-6f);
-  ASSERT_NEAR(expect_result[2], actual_result2, 1e-6f);
-}
diff --git a/paddle/pten/tests/test_fill_api.cc b/paddle/pten/tests/api/test_fill_api.cc
similarity index 78%
rename from paddle/pten/tests/test_fill_api.cc
rename to paddle/pten/tests/api/test_fill_api.cc
index a83ed76a6f8dfa1da0f0ccedbff993d01d6b2158..cbac4c541f519bdd1d95b07508749596987bb6d9 100644
--- a/paddle/pten/tests/test_fill_api.cc
+++ b/paddle/pten/tests/api/test_fill_api.cc
@@ -21,8 +21,6 @@ limitations under the License. */
 #include "paddle/pten/core/dense_tensor.h"
 #include "paddle/pten/core/kernel_registry.h"
 
-#include "paddle/pten/include/creation.h"
-
 PT_DECLARE_MODULE(CreationCPU);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
@@ -133,38 +131,3 @@ TEST(API, ones_like) {
     ASSERT_EQ(actual_result[i], 1);
   }
 }
-
-TEST(DEV_API, fill_any_like) {
-  // 1. create tensor
-  const auto alloc = std::make_shared<paddle::experimental::DefaultAllocator>(
-      paddle::platform::CPUPlace());
-  pten::DenseTensor dense_x(alloc,
-                            pten::DenseTensorMeta(pten::DataType::FLOAT32,
-                                                  framework::make_ddim({3, 2}),
-                                                  pten::DataLayout::NCHW));
-  auto* dense_x_data = dense_x.mutable_data<float>();
-  dense_x_data[0] = 0;
-  float val = 1.0;
-
-  paddle::platform::DeviceContextPool& pool =
-      paddle::platform::DeviceContextPool::Instance();
-  auto* dev_ctx = pool.Get(paddle::platform::CPUPlace());
-
-  // 2. test API
-  auto out = pten::FillAnyLike<float>(
-      *(static_cast<paddle::platform::CPUDeviceContext*>(dev_ctx)),
-      dense_x,
-      val);
-
-  // 3. check result
-  ASSERT_EQ(out.dims().size(), 2);
-  ASSERT_EQ(out.dims()[0], 3);
-  ASSERT_EQ(out.numel(), 6);
-  ASSERT_EQ(out.meta().type, pten::DataType::FLOAT32);
-  ASSERT_EQ(out.meta().layout, pten::DataLayout::NCHW);
-
-  auto* actual_result = out.data<float>();
-  for (auto i = 0; i < 6; i++) {
-    ASSERT_NEAR(actual_result[i], val, 1e-6f);
-  }
-}
diff --git a/paddle/pten/tests/test_flatten_api.cc b/paddle/pten/tests/api/test_flatten_api.cc
similarity index 62%
rename from paddle/pten/tests/test_flatten_api.cc
rename to paddle/pten/tests/api/test_flatten_api.cc
index ca8ca38f65a3ee4cf68ebfaaf13e7eeacd1c31ea..3701c358c667e1c96851f5b5d18ac72101411057 100644
--- a/paddle/pten/tests/test_flatten_api.cc
+++ b/paddle/pten/tests/api/test_flatten_api.cc
@@ -21,8 +21,6 @@ limitations under the License. */
 #include "paddle/pten/core/dense_tensor.h"
 #include "paddle/pten/core/kernel_registry.h"
 
-#include "paddle/pten/include/manipulation.h"
-
 PT_DECLARE_MODULE(ManipulationCPU);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
@@ -72,47 +70,3 @@ TEST(API, flatten) {
   }
   ASSERT_EQ(value_equal, true);
 }
-
-TEST(DEV_API, flatten) {
-  // 1. create tensor
-  const auto alloc = std::make_shared<paddle::experimental::DefaultAllocator>(
-      paddle::platform::CPUPlace());
-  pten::DenseTensor dense_x(
-      alloc,
-      pten::DenseTensorMeta(pten::DataType::FLOAT32,
-                            framework::make_ddim({3, 2, 2, 3}),
-                            pten::DataLayout::NCHW));
-  auto* dense_x_data = dense_x.mutable_data<float>();
-
-  for (int i = 0; i < dense_x.numel(); i++) {
-    dense_x_data[i] = i;
-  }
-  int start_axis = 1, stop_axis = 2;
-  paddle::platform::DeviceContextPool& pool =
-      paddle::platform::DeviceContextPool::Instance();
-  auto* dev_ctx = pool.Get(paddle::platform::CPUPlace());
-
-  // 2. test API
-  auto out = pten::Flatten<float>(
-      *(static_cast<paddle::platform::CPUDeviceContext*>(dev_ctx)),
-      dense_x,
-      start_axis,
-      stop_axis);
-
-  // 3. check result
-  std::vector<int> expect_shape = {3, 4, 3};
-  ASSERT_EQ(out.dims()[0], expect_shape[0]);
-  ASSERT_EQ(out.dims()[1], expect_shape[1]);
-  ASSERT_EQ(out.dims()[2], expect_shape[2]);
-  ASSERT_EQ(out.numel(), 36);
-  ASSERT_EQ(out.meta().type, pten::DataType::FLOAT32);
-  ASSERT_EQ(out.meta().layout, pten::DataLayout::NCHW);
-
-  bool value_equal = true;
-  auto* dense_out_data = out.data<float>();
-  for (int i = 0; i < dense_x.numel(); i++) {
-    if (std::abs(dense_x_data[i] - dense_out_data[i]) > 1e-6f)
-      value_equal = false;
-  }
-  ASSERT_EQ(value_equal, true);
-}
diff --git a/paddle/pten/tests/test_matmul_api.cc b/paddle/pten/tests/api/test_matmul_api.cc
similarity index 100%
rename from paddle/pten/tests/test_matmul_api.cc
rename to paddle/pten/tests/api/test_matmul_api.cc
diff --git a/paddle/pten/tests/test_mean_api.cc b/paddle/pten/tests/api/test_mean_api.cc
similarity index 64%
rename from paddle/pten/tests/test_mean_api.cc
rename to paddle/pten/tests/api/test_mean_api.cc
index 4dd4222daabe473a499e510675225b25d20d4a46..d772e58659507a7d098fdba2b785a62fab019eac 100644
--- a/paddle/pten/tests/test_mean_api.cc
+++ b/paddle/pten/tests/api/test_mean_api.cc
@@ -21,8 +21,6 @@ limitations under the License. */
 #include "paddle/pten/core/dense_tensor.h"
 #include "paddle/pten/core/kernel_registry.h"
 
-#include "paddle/pten/include/math.h"
-
 PT_DECLARE_MODULE(MathCPU);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
@@ -69,36 +67,3 @@ TEST(API, mean) {
   auto actual_result = dense_out->data<float>()[0];
   ASSERT_NEAR(expect_result, actual_result, 1e-6f);
 }
-
-TEST(DEV_API, mean) {
-  // 1. create tensor
-  const auto alloc = std::make_shared<paddle::experimental::DefaultAllocator>(
-      paddle::platform::CPUPlace());
-  pten::DenseTensor dense_x(alloc,
-                            pten::DenseTensorMeta(pten::DataType::FLOAT32,
-                                                  framework::make_ddim({3, 4}),
-                                                  pten::DataLayout::NCHW));
-  auto* dense_x_data = dense_x.mutable_data<float>();
-
-  float sum = 0.0;
-  for (size_t i = 0; i < 12; ++i) {
-    dense_x_data[i] = i * 1.0;
-    sum += i * 1.0;
-  }
-  paddle::platform::DeviceContextPool& pool =
-      paddle::platform::DeviceContextPool::Instance();
-  auto* dev_ctx = pool.Get(paddle::platform::CPUPlace());
-  // 2. test API
-  auto out = pten::Mean<float>(
-      *(static_cast<paddle::platform::CPUDeviceContext*>(dev_ctx)), dense_x);
-
-  // 3. check result
-  ASSERT_EQ(out.dims().size(), 1);
-  ASSERT_EQ(out.numel(), 1);
-  ASSERT_EQ(out.meta().type, pten::DataType::FLOAT32);
-  ASSERT_EQ(out.meta().layout, pten::DataLayout::NCHW);
-
-  auto expect_result = sum / 12;
-  auto actual_result = out.data<float>()[0];
-  ASSERT_NEAR(expect_result, actual_result, 1e-6f);
-}
diff --git a/paddle/pten/api/lib/utils/tests/test_storage.cc b/paddle/pten/tests/api/test_storage.cc
similarity index 100%
rename from paddle/pten/api/lib/utils/tests/test_storage.cc
rename to paddle/pten/tests/api/test_storage.cc
diff --git a/paddle/pten/api/lib/utils/tests/test_tensor_utils.cc b/paddle/pten/tests/api/test_tensor_utils.cc
similarity index 100%
rename from paddle/pten/api/lib/utils/tests/test_tensor_utils.cc
rename to paddle/pten/tests/api/test_tensor_utils.cc
diff --git a/paddle/pten/tests/common/CMakeLists.txt b/paddle/pten/tests/common/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..c0a5414d53e478bd6bea4d7eff82113681814e1c
--- /dev/null
+++ b/paddle/pten/tests/common/CMakeLists.txt
@@ -0,0 +1,3 @@
+cc_test(pten_test_backend SRCS test_backend.cc DEPS gtest)
+cc_test(pten_test_data_layout SRCS test_data_layout.cc DEPS gtest)
+cc_test(pten_test_data_type SRCS test_data_type.cc DEPS gtest)
diff --git a/paddle/pten/tests/backend_test.cc b/paddle/pten/tests/common/test_backend.cc
similarity index 100%
rename from paddle/pten/tests/backend_test.cc
rename to paddle/pten/tests/common/test_backend.cc
index 2bae2cd4171650a3616c5dc8605eb064fe7616d2..1c17c881ed24f23148525072301c5eba3b671125 100644
--- a/paddle/pten/tests/backend_test.cc
+++ b/paddle/pten/tests/common/test_backend.cc
@@ -12,11 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/pten/common/backend.h"
-
 #include <gtest/gtest.h>
 #include <iostream>
 
+#include "paddle/pten/common/backend.h"
+
 TEST(Backend, OStream) {
   std::ostringstream oss;
   oss << pten::Backend::UNDEFINED;
diff --git a/paddle/pten/tests/data_layout_test.cc b/paddle/pten/tests/common/test_data_layout.cc
similarity index 99%
rename from paddle/pten/tests/data_layout_test.cc
rename to paddle/pten/tests/common/test_data_layout.cc
index efa19670f25be16c6b9d597e5ebd9cef0453f8c4..7fe1b6c2ffd2be88afa2e6978f6c3e9be10745f1 100644
--- a/paddle/pten/tests/data_layout_test.cc
+++ b/paddle/pten/tests/common/test_data_layout.cc
@@ -15,6 +15,7 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <iostream>
 #include <sstream>
+
 #include "paddle/pten/common/layout.h"
 
 TEST(DataLayout, OStream) {
diff --git a/paddle/pten/tests/data_type_test.cc b/paddle/pten/tests/common/test_data_type.cc
similarity index 100%
rename from paddle/pten/tests/data_type_test.cc
rename to paddle/pten/tests/common/test_data_type.cc
index bcdef84040523cd103a4f95d4dd9320e858303ea..28d58858bb42c8a5d778fa5ce9fa24a0a171bce4 100644
--- a/paddle/pten/tests/data_type_test.cc
+++ b/paddle/pten/tests/common/test_data_type.cc
@@ -12,12 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/pten/common/data_type.h"
-
 #include <gtest/gtest.h>
 #include <iostream>
 #include <sstream>
 
+#include "paddle/pten/common/data_type.h"
+
 TEST(DataType, OStream) {
   std::ostringstream oss;
   oss << pten::DataType::UNDEFINED;
diff --git a/paddle/pten/tests/core/CMakeLists.txt b/paddle/pten/tests/core/CMakeLists.txt
index 34888200b2f1918ee76428c274fa8c0b9216a9ac..b25439cfe25279b3db97afe90643624fe014a62a 100644
--- a/paddle/pten/tests/core/CMakeLists.txt
+++ b/paddle/pten/tests/core/CMakeLists.txt
@@ -1,3 +1,6 @@
 cc_test(test_allocator SRCS test_allocator.cc DEPS tensor_base)
 cc_test(test_storage SRCS test_storage.cc DEPS tensor_base)
 cc_test(test_dense_tensor SRCS test_dense_tensor.cc DEPS dense_tensor)
+cc_test(test_intrusive_ptr SRCS test_intrusive_ptr.cc)
+cc_test(test_type_info SRCS test_type_info.cc)
+cc_test(test_kernel_factory SRCS test_kernel_factory.cc DEPS kernel_factory)
diff --git a/paddle/pten/tests/utils/test_intrusive_ptr.cc b/paddle/pten/tests/core/test_intrusive_ptr.cc
similarity index 100%
rename from paddle/pten/tests/utils/test_intrusive_ptr.cc
rename to paddle/pten/tests/core/test_intrusive_ptr.cc
diff --git a/paddle/pten/tests/kernel_factory_test.cc b/paddle/pten/tests/core/test_kernel_factory.cc
similarity index 100%
rename from paddle/pten/tests/kernel_factory_test.cc
rename to paddle/pten/tests/core/test_kernel_factory.cc
diff --git a/paddle/pten/tests/utils/test_type_info.cc b/paddle/pten/tests/core/test_type_info.cc
similarity index 100%
rename from paddle/pten/tests/utils/test_type_info.cc
rename to paddle/pten/tests/core/test_type_info.cc
diff --git a/paddle/pten/tests/dense_tensor_test.cc b/paddle/pten/tests/dense_tensor_test.cc
deleted file mode 100644
index e74917263dafb5b16f7f1f50ed17c721868b23f8..0000000000000000000000000000000000000000
--- a/paddle/pten/tests/dense_tensor_test.cc
+++ /dev/null
@@ -1,20 +0,0 @@
-/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include "paddle/pten/core/dense_tensor.h"
-
-#include <gtest/gtest.h>
-
-namespace framework = paddle::framework;
-using DDim = paddle::framework::DDim;
diff --git a/paddle/pten/tests/kernels/CMakeLists.txt b/paddle/pten/tests/kernels/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..b0dc29de5214079157e93c5b67c25698b885af74
--- /dev/null
+++ b/paddle/pten/tests/kernels/CMakeLists.txt
@@ -0,0 +1,6 @@
+cc_test(test_copy_dev_api SRCS test_copy_dev_api.cc DEPS pten pten_api_utils)
+cc_test(test_dot_dev_api SRCS test_dot_dev_api.cc DEPS pten pten_api_utils)
+cc_test(test_fill_dev_api SRCS test_fill_dev_api.cc DEPS pten pten_api_utils)
+cc_test(test_flatten_dev_api SRCS test_flatten_dev_api.cc DEPS pten pten_api_utils)
+cc_test(test_mean_dev_api SRCS test_mean_dev_api.cc DEPS pten pten_api_utils)
+cc_test(test_scale_dev_api SRCS test_scale_dev_api.cc DEPS pten pten_api_utils)
diff --git a/paddle/pten/tests/test_copy_api.cc b/paddle/pten/tests/kernels/test_copy_dev_api.cc
similarity index 99%
rename from paddle/pten/tests/test_copy_api.cc
rename to paddle/pten/tests/kernels/test_copy_dev_api.cc
index 4bf16e7517b7336209a887d4a6843c8bdf77fd9a..dadbcf098dd17a667a1039dcc4e184f35651b92e 100644
--- a/paddle/pten/tests/test_copy_api.cc
+++ b/paddle/pten/tests/kernels/test_copy_dev_api.cc
@@ -29,7 +29,7 @@ using DDim = paddle::framework::DDim;
 // TODO(YuanRisheng): This TEST file need to be refactored after 'copy' realized
 // in
 // 'paddle/api',
-TEST(API, copy) {
+TEST(DEV_API, copy) {
   // 1. create tensor
   const auto alloc = std::make_shared<paddle::experimental::DefaultAllocator>(
       paddle::platform::CPUPlace());
diff --git a/paddle/pten/tests/kernels/test_dot_dev_api.cc b/paddle/pten/tests/kernels/test_dot_dev_api.cc
new file mode 100644
index 0000000000000000000000000000000000000000..897c49f4f255e544222ea200e67603517921e851
--- /dev/null
+++ b/paddle/pten/tests/kernels/test_dot_dev_api.cc
@@ -0,0 +1,81 @@
+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include <memory>
+
+#include "paddle/pten/include/linalg.h"
+
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/core/kernel_registry.h"
+
+PT_DECLARE_MODULE(LinalgCPU);
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+PT_DECLARE_MODULE(LinalgCUDA);
+#endif
+
+namespace framework = paddle::framework;
+using DDim = paddle::framework::DDim;
+
+TEST(DEV_API, dot) {
+  // 1. create tensor
+  const auto alloc = std::make_shared<paddle::experimental::DefaultAllocator>(
+      paddle::platform::CPUPlace());
+  pten::DenseTensor dense_x(alloc,
+                            pten::DenseTensorMeta(pten::DataType::FLOAT32,
+                                                  framework::make_ddim({3, 10}),
+                                                  pten::DataLayout::NCHW));
+  auto* dense_x_data = dense_x.mutable_data<float>();
+
+  pten::DenseTensor dense_y(alloc,
+                            pten::DenseTensorMeta(pten::DataType::FLOAT32,
+                                                  framework::make_ddim({3, 10}),
+                                                  pten::DataLayout::NCHW));
+  auto* dense_y_data = dense_y.mutable_data<float>();
+
+  float sum[3] = {0.0, 0.0, 0.0};
+  for (size_t i = 0; i < 3; ++i) {
+    for (size_t j = 0; j < 10; ++j) {
+      dense_x_data[i * 10 + j] = (i * 10 + j) * 1.0;
+      dense_y_data[i * 10 + j] = (i * 10 + j) * 1.0;
+      sum[i] += (i * 10 + j) * (i * 10 + j) * 1.0;
+    }
+  }
+
+  paddle::platform::DeviceContextPool& pool =
+      paddle::platform::DeviceContextPool::Instance();
+  auto* dev_ctx = pool.Get(paddle::platform::CPUPlace());
+
+  // 2. test API
+  auto out = pten::Dot<float>(
+      *(static_cast<paddle::platform::CPUDeviceContext*>(dev_ctx)),
+      dense_x,
+      dense_y);
+
+  // 3. check result
+  ASSERT_EQ(out.dims().size(), 2);
+  ASSERT_EQ(out.dims()[0], 3);
+  ASSERT_EQ(out.meta().type, pten::DataType::FLOAT32);
+  ASSERT_EQ(out.meta().layout, pten::DataLayout::NCHW);
+
+  auto expect_result = sum;
+  auto actual_result0 = out.data<float>()[0];
+  auto actual_result1 = out.data<float>()[1];
+  auto actual_result2 = out.data<float>()[2];
+  ASSERT_NEAR(expect_result[0], actual_result0, 1e-6f);
+  ASSERT_NEAR(expect_result[1], actual_result1, 1e-6f);
+  ASSERT_NEAR(expect_result[2], actual_result2, 1e-6f);
+}
diff --git a/paddle/pten/tests/kernels/test_fill_dev_api.cc b/paddle/pten/tests/kernels/test_fill_dev_api.cc
new file mode 100644
index 0000000000000000000000000000000000000000..8546396e18a9ac6a2771a6d1f61658727e336463
--- /dev/null
+++ b/paddle/pten/tests/kernels/test_fill_dev_api.cc
@@ -0,0 +1,66 @@
+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include <memory>
+
+#include "paddle/pten/include/creation.h"
+
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/core/kernel_registry.h"
+
+PT_DECLARE_MODULE(CreationCPU);
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+PT_DECLARE_MODULE(CreationCUDA);
+#endif
+
+namespace framework = paddle::framework;
+using DDim = paddle::framework::DDim;
+
+TEST(DEV_API, fill_any_like) {
+  // 1. create tensor
+  const auto alloc = std::make_shared<paddle::experimental::DefaultAllocator>(
+      paddle::platform::CPUPlace());
+  pten::DenseTensor dense_x(alloc,
+                            pten::DenseTensorMeta(pten::DataType::FLOAT32,
+                                                  framework::make_ddim({3, 2}),
+                                                  pten::DataLayout::NCHW));
+  auto* dense_x_data = dense_x.mutable_data<float>();
+  dense_x_data[0] = 0;
+  float val = 1.0;
+
+  paddle::platform::DeviceContextPool& pool =
+      paddle::platform::DeviceContextPool::Instance();
+  auto* dev_ctx = pool.Get(paddle::platform::CPUPlace());
+
+  // 2. test API
+  auto out = pten::FillAnyLike<float>(
+      *(static_cast<paddle::platform::CPUDeviceContext*>(dev_ctx)),
+      dense_x,
+      val);
+
+  // 3. check result
+  ASSERT_EQ(out.dims().size(), 2);
+  ASSERT_EQ(out.dims()[0], 3);
+  ASSERT_EQ(out.numel(), 6);
+  ASSERT_EQ(out.meta().type, pten::DataType::FLOAT32);
+  ASSERT_EQ(out.meta().layout, pten::DataLayout::NCHW);
+
+  auto* actual_result = out.data<float>();
+  for (auto i = 0; i < 6; i++) {
+    ASSERT_NEAR(actual_result[i], val, 1e-6f);
+  }
+}
diff --git a/paddle/pten/tests/kernels/test_flatten_dev_api.cc b/paddle/pten/tests/kernels/test_flatten_dev_api.cc
new file mode 100644
index 0000000000000000000000000000000000000000..16e6aa5acdd987c7e2ffe63bf2c10774ba0f84f5
--- /dev/null
+++ b/paddle/pten/tests/kernels/test_flatten_dev_api.cc
@@ -0,0 +1,75 @@
+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include <memory>
+
+#include "paddle/pten/include/manipulation.h"
+
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/core/kernel_registry.h"
+
+PT_DECLARE_MODULE(ManipulationCPU);
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+PT_DECLARE_MODULE(ManipulationCUDA);
+#endif
+
+namespace framework = paddle::framework;
+using DDim = paddle::framework::DDim;
+
+TEST(DEV_API, flatten) {
+  // 1. create tensor
+  const auto alloc = std::make_shared<paddle::experimental::DefaultAllocator>(
+      paddle::platform::CPUPlace());
+  pten::DenseTensor dense_x(
+      alloc,
+      pten::DenseTensorMeta(pten::DataType::FLOAT32,
+                            framework::make_ddim({3, 2, 2, 3}),
+                            pten::DataLayout::NCHW));
+  auto* dense_x_data = dense_x.mutable_data<float>();
+
+  for (int i = 0; i < dense_x.numel(); i++) {
+    dense_x_data[i] = i;
+  }
+  int start_axis = 1, stop_axis = 2;
+  paddle::platform::DeviceContextPool& pool =
+      paddle::platform::DeviceContextPool::Instance();
+  auto* dev_ctx = pool.Get(paddle::platform::CPUPlace());
+
+  // 2. test API
+  auto out = pten::Flatten<float>(
+      *(static_cast<paddle::platform::CPUDeviceContext*>(dev_ctx)),
+      dense_x,
+      start_axis,
+      stop_axis);
+
+  // 3. check result
+  std::vector<int> expect_shape = {3, 4, 3};
+  ASSERT_EQ(out.dims()[0], expect_shape[0]);
+  ASSERT_EQ(out.dims()[1], expect_shape[1]);
+  ASSERT_EQ(out.dims()[2], expect_shape[2]);
+  ASSERT_EQ(out.numel(), 36);
+  ASSERT_EQ(out.meta().type, pten::DataType::FLOAT32);
+  ASSERT_EQ(out.meta().layout, pten::DataLayout::NCHW);
+
+  bool value_equal = true;
+  auto* dense_out_data = out.data<float>();
+  for (int i = 0; i < dense_x.numel(); i++) {
+    if (std::abs(dense_x_data[i] - dense_out_data[i]) > 1e-6f)
+      value_equal = false;
+  }
+  ASSERT_EQ(value_equal, true);
+}
diff --git a/paddle/pten/tests/kernels/test_mean_dev_api.cc b/paddle/pten/tests/kernels/test_mean_dev_api.cc
new file mode 100644
index 0000000000000000000000000000000000000000..43f9371d72a2f64d2f32d85460ad47048179765c
--- /dev/null
+++ b/paddle/pten/tests/kernels/test_mean_dev_api.cc
@@ -0,0 +1,64 @@
+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include <memory>
+
+#include "paddle/pten/include/math.h"
+
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/core/kernel_registry.h"
+
+PT_DECLARE_MODULE(MathCPU);
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+PT_DECLARE_MODULE(MathCUDA);
+#endif
+
+namespace framework = paddle::framework;
+using DDim = paddle::framework::DDim;
+
+TEST(DEV_API, mean) {
+  // 1. create tensor
+  const auto alloc = std::make_shared<paddle::experimental::DefaultAllocator>(
+      paddle::platform::CPUPlace());
+  pten::DenseTensor dense_x(alloc,
+                            pten::DenseTensorMeta(pten::DataType::FLOAT32,
+                                                  framework::make_ddim({3, 4}),
+                                                  pten::DataLayout::NCHW));
+  auto* dense_x_data = dense_x.mutable_data<float>();
+
+  float sum = 0.0;
+  for (size_t i = 0; i < 12; ++i) {
+    dense_x_data[i] = i * 1.0;
+    sum += i * 1.0;
+  }
+  paddle::platform::DeviceContextPool& pool =
+      paddle::platform::DeviceContextPool::Instance();
+  auto* dev_ctx = pool.Get(paddle::platform::CPUPlace());
+  // 2. test API
+  auto out = pten::Mean<float>(
+      *(static_cast<paddle::platform::CPUDeviceContext*>(dev_ctx)), dense_x);
+
+  // 3. check result
+  ASSERT_EQ(out.dims().size(), 1);
+  ASSERT_EQ(out.numel(), 1);
+  ASSERT_EQ(out.meta().type, pten::DataType::FLOAT32);
+  ASSERT_EQ(out.meta().layout, pten::DataLayout::NCHW);
+
+  auto expect_result = sum / 12;
+  auto actual_result = out.data<float>()[0];
+  ASSERT_NEAR(expect_result, actual_result, 1e-6f);
+}
diff --git a/paddle/pten/tests/test_scale_api.cc b/paddle/pten/tests/kernels/test_scale_dev_api.cc
similarity index 98%
rename from paddle/pten/tests/test_scale_api.cc
rename to paddle/pten/tests/kernels/test_scale_dev_api.cc
index 2ecf6d199b55855bdb95f803d333084bcb2dc2b1..f1764e803250a7179e969c9c63ddaefa3148266c 100644
--- a/paddle/pten/tests/test_scale_api.cc
+++ b/paddle/pten/tests/kernels/test_scale_dev_api.cc
@@ -15,14 +15,12 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <memory>
 
-#include "paddle/pten/api/include/math.h"
+#include "paddle/pten/include/math.h"
 
 #include "paddle/pten/api/lib/utils/allocator.h"
 #include "paddle/pten/core/dense_tensor.h"
 #include "paddle/pten/core/kernel_registry.h"
 
-#include "paddle/pten/include/math.h"
-
 PT_DECLARE_MODULE(MathCPU);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
diff --git a/paddle/pten/tests/utils/CMakeLists.txt b/paddle/pten/tests/utils/CMakeLists.txt
deleted file mode 100644
index e9974819322517b92c9c9b44e2fe3d00b91480d4..0000000000000000000000000000000000000000
--- a/paddle/pten/tests/utils/CMakeLists.txt
+++ /dev/null
@@ -1,2 +0,0 @@
-cc_test(test_intrusive_ptr SRCS test_intrusive_ptr.cc)
-cc_test(test_type_info SRCS test_type_info.cc)