diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h
index 575ce8956d44c1bc672438c815bb23ac2e73f6d2..b75dcade6fccf703b36e705b6171b65d93f0d723 100644
--- a/paddle/fluid/framework/operator.h
+++ b/paddle/fluid/framework/operator.h
@@ -40,7 +40,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/variant.h"
 #include "paddle/utils/flat_hash_map.h"
 
-#include "paddle/pten/api/include/core.h"
+#include "paddle/pten/include/core.h"
 
 namespace paddle {
 namespace framework {
diff --git a/paddle/fluid/framework/pten_utils.h b/paddle/fluid/framework/pten_utils.h
index aab756f50e395aa37942ceb7d42655969661b354..6fe02ad4a4ae4a87c4d053e353da2474186b5d53 100644
--- a/paddle/fluid/framework/pten_utils.h
+++ b/paddle/fluid/framework/pten_utils.h
@@ -25,8 +25,8 @@ limitations under the License. */
 #include "paddle/fluid/imperative/type_defs.h"
 #include "paddle/fluid/platform/macros.h"
 #include "paddle/fluid/platform/place.h"
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/hapi/lib/utils/tensor_utils.h"
+#include "paddle/pten/api/lib/utils/tensor_utils.h"
+#include "paddle/pten/include/core.h"
 #include "paddle/utils/flat_hash_map.h"
 #include "paddle/utils/small_vector.h"
 
diff --git a/paddle/fluid/imperative/prepared_operator.h b/paddle/fluid/imperative/prepared_operator.h
index fab67e87c79480100293904670f1154b6e93c7fc..144f921861f9e1c0d4ace3a2d2ae089425e1e80b 100644
--- a/paddle/fluid/imperative/prepared_operator.h
+++ b/paddle/fluid/imperative/prepared_operator.h
@@ -27,7 +27,7 @@
 #include "paddle/fluid/imperative/layer.h"
 #include "paddle/fluid/imperative/type_defs.h"
 
-#include "paddle/pten/api/include/core.h"
+#include "paddle/pten/include/core.h"
 
 DECLARE_bool(use_mkldnn);
 
diff --git a/paddle/fluid/operators/dot_op.h b/paddle/fluid/operators/dot_op.h
index 6a025fdd9ccc6b6f5b5a0edeeea96ddff671d32a..f6877c57a5c18262b1610baf172a971c8c60c95e 100644
--- a/paddle/fluid/operators/dot_op.h
+++ b/paddle/fluid/operators/dot_op.h
@@ -20,9 +20,9 @@
 #include "paddle/fluid/platform/for_range.h"
 
 // only can include the headers in paddle/pten/api dirs
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/api/include/linalg.h"
-#include "paddle/pten/hapi/lib/utils/tensor_utils.h"
+#include "paddle/pten/api/lib/utils/tensor_utils.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/linalg.h"
 
 namespace paddle {
 namespace operators {
diff --git a/paddle/fluid/operators/fill_any_like_op.h b/paddle/fluid/operators/fill_any_like_op.h
index fc649f42c51a1b479015ab25ace57425615dda15..5e2d06672c3c17d0ff83b8055f10b4afcf507aac 100644
--- a/paddle/fluid/operators/fill_any_like_op.h
+++ b/paddle/fluid/operators/fill_any_like_op.h
@@ -19,8 +19,8 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/pten_utils.h"
 
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/api/include/creation.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/creation.h"
 
 namespace paddle {
 namespace operators {
diff --git a/paddle/fluid/operators/matmul_v2_op.h b/paddle/fluid/operators/matmul_v2_op.h
index 1d361ae45dad9bd6bbb58dfd4e58fadcab9d7e4b..ee95881caa9c53b07dd52e101fb924c7566ac6f6 100644
--- a/paddle/fluid/operators/matmul_v2_op.h
+++ b/paddle/fluid/operators/matmul_v2_op.h
@@ -26,9 +26,9 @@ limitations under the License. */
 #include "paddle/fluid/operators/reduce_ops/reduce_sum_op.h"
 
 // only can include the headers in paddle/pten/api dirs
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/api/include/linalg.h"
-#include "paddle/pten/hapi/lib/utils/tensor_utils.h"
+#include "paddle/pten/api/lib/utils/tensor_utils.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/linalg.h"
 
 #if defined(__NVCC__) || defined(__HIPCC__)
 #include "paddle/fluid/operators/reduce_ops/cub_reduce.h"
diff --git a/paddle/fluid/operators/mean_op.h b/paddle/fluid/operators/mean_op.h
index f909b96c9193c019e4d043c0f974b84a68dcf131..2b398771c5c9689c4243b435336c1bfeca9093ab 100644
--- a/paddle/fluid/operators/mean_op.h
+++ b/paddle/fluid/operators/mean_op.h
@@ -18,9 +18,9 @@ limitations under the License. */
 #include "paddle/fluid/framework/pten_utils.h"
 
 // only can include the headers in paddle/top/api dirs
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/api/include/math.h"
-#include "paddle/pten/hapi/lib/utils/tensor_utils.h"
+#include "paddle/pten/api/lib/utils/tensor_utils.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/math.h"
 
 namespace paddle {
 namespace operators {
diff --git a/paddle/fluid/operators/scale_op.h b/paddle/fluid/operators/scale_op.h
index 0d7113a6f4de90da930e005c04b6400f63c0a444..a75c9fd4fd2450a86fb2e59ac492c897f68d1bb7 100644
--- a/paddle/fluid/operators/scale_op.h
+++ b/paddle/fluid/operators/scale_op.h
@@ -18,9 +18,9 @@ limitations under the License. */
 #include "paddle/fluid/framework/pten_utils.h"
 
 // only can include the headers in paddle/top/api dirs
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/api/include/math.h"
-#include "paddle/pten/hapi/lib/utils/tensor_utils.h"
+#include "paddle/pten/api/lib/utils/tensor_utils.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/math.h"
 
 namespace paddle {
 namespace operators {
diff --git a/paddle/fluid/operators/sign_op.h b/paddle/fluid/operators/sign_op.h
index 0e3036115e3c14fc0feaa00b281871764ba554b4..b7a46cc5467978bd0dbf53b45c1725b5055f682a 100644
--- a/paddle/fluid/operators/sign_op.h
+++ b/paddle/fluid/operators/sign_op.h
@@ -20,8 +20,8 @@ limitations under the License. */
 #include "paddle/fluid/operators/eigen/eigen_function.h"
 
 // only can include the headers in paddle/pten/api dirs
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/api/include/math.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/math.h"
 
 namespace paddle {
 namespace operators {
diff --git a/paddle/pten/CMakeLists.txt b/paddle/pten/CMakeLists.txt
index c1fe2d552af136648b6a416333802cb8c7b24a68..0444fa593c0ac392de49a3b44d8a1a77a96aabd5 100644
--- a/paddle/pten/CMakeLists.txt
+++ b/paddle/pten/CMakeLists.txt
@@ -1,7 +1,6 @@
-# pten api
+# pten (low level) api headers: include
+# pten (high level) api
 add_subdirectory(api)
-# pten high level api
-add_subdirectory(hapi)
 # pten core components
 add_subdirectory(core)
 # pten kernels for diff device
@@ -10,3 +9,13 @@ add_subdirectory(kernels)
 add_subdirectory(infershape)
 # pten tests
 add_subdirectory(tests)
+
+# make an unity target for compile deps
+set(PTEN_DEPS convert_utils dense_tensor kernel_factory kernel_context)
+set(PTEN_DEPS ${PTEN_DEPS} math_cpu linalg_cpu creation_cpu manipulation_cpu)
+set(PTEN_DEPS ${PTEN_DEPS} unary binary)
+if(WITH_GPU OR WITH_ROCM)
+  set(PTEN_DEPS ${PTEN_DEPS} math_cuda linalg_cuda creation_cuda manipulation_cuda)
+endif()
+
+cc_library(pten SRCS all.cc DEPS ${PTEN_DEPS})
diff --git a/paddle/pten/hapi/all.cc b/paddle/pten/all.cc
similarity index 82%
rename from paddle/pten/hapi/all.cc
rename to paddle/pten/all.cc
index 4ea6fabeecf2e52c222be9c22dae70e50880b99a..d8d96e1cd461edf74c2f29cab61f1fbf4191138b 100644
--- a/paddle/pten/hapi/all.cc
+++ b/paddle/pten/all.cc
@@ -12,8 +12,6 @@ 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/hapi/all.h"
+#include "paddle/pten/all.h"
 
-namespace paddle {
-namespace experimental {}  // namespace experimental
-}  // namespace paddle
+namespace pten {}  // namespace pten
diff --git a/paddle/pten/hapi/all.h b/paddle/pten/all.h
similarity index 69%
rename from paddle/pten/hapi/all.h
rename to paddle/pten/all.h
index 1a2a4199e7bf7cf43547e78c11483b2ee0e289d3..e8e41a8c3eabc860799dfd63b33ea8dc2a4c232b 100644
--- a/paddle/pten/hapi/all.h
+++ b/paddle/pten/all.h
@@ -14,9 +14,10 @@ limitations under the License. */
 
 #pragma once
 
-// user apis
-#include "paddle/pten/hapi/include/creation.h"
-#include "paddle/pten/hapi/include/linalg.h"
-#include "paddle/pten/hapi/include/manipulation.h"
-#include "paddle/pten/hapi/include/math.h"
-#include "paddle/pten/hapi/include/tensor.h"
+// develop apis
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/creation.h"
+#include "paddle/pten/include/infershape.h"
+#include "paddle/pten/include/linalg.h"
+#include "paddle/pten/include/manipulation.h"
+#include "paddle/pten/include/math.h"
diff --git a/paddle/pten/api/CMakeLists.txt b/paddle/pten/api/CMakeLists.txt
index 1c107519324e218c31883282b5929e105f6393f8..4b427b3b4a38349f672aa8c643bd06ad7f39d360 100644
--- a/paddle/pten/api/CMakeLists.txt
+++ b/paddle/pten/api/CMakeLists.txt
@@ -1,8 +1,3 @@
-set(PTEN_DEPS convert_utils dense_tensor kernel_factory kernel_context)
-set(PTEN_DEPS ${PTEN_DEPS} math_cpu linalg_cpu creation_cpu manipulation_cpu)
-set(PTEN_DEPS ${PTEN_DEPS} unary binary)
-if(WITH_GPU OR WITH_ROCM)
-  set(PTEN_DEPS ${PTEN_DEPS} math_cuda linalg_cuda creation_cuda manipulation_cuda)
-endif()
+add_subdirectory(lib)
 
-cc_library(pten SRCS all.cc DEPS ${PTEN_DEPS})
+cc_library(pten_hapi SRCS all.cc DEPS linalg_api math_api creation_api)
diff --git a/paddle/pten/api/all.cc b/paddle/pten/api/all.cc
index 0704d6c516fa62db9a7af403fb66cffc3a384abe..c270fbb070689611a76d8fa2dde9ae5016cbd407 100644
--- a/paddle/pten/api/all.cc
+++ b/paddle/pten/api/all.cc
@@ -14,4 +14,6 @@ limitations under the License. */
 
 #include "paddle/pten/api/all.h"
 
-namespace pten {}  // namespace pten
+namespace paddle {
+namespace experimental {}  // namespace experimental
+}  // namespace paddle
diff --git a/paddle/pten/api/all.h b/paddle/pten/api/all.h
index c760960967d9563e06d3c96f099804b01f49f0f2..3a2c7b3fa98fe4adb813e7178bb6ebfad84fbec6 100644
--- a/paddle/pten/api/all.h
+++ b/paddle/pten/api/all.h
@@ -14,10 +14,9 @@ limitations under the License. */
 
 #pragma once
 
-// develop apis
-#include "paddle/pten/api/include/core.h"
+// user apis
 #include "paddle/pten/api/include/creation.h"
-#include "paddle/pten/api/include/infershape.h"
 #include "paddle/pten/api/include/linalg.h"
 #include "paddle/pten/api/include/manipulation.h"
 #include "paddle/pten/api/include/math.h"
+#include "paddle/pten/api/include/tensor.h"
diff --git a/paddle/pten/hapi/include/backend_set.h b/paddle/pten/api/include/backend_set.h
similarity index 100%
rename from paddle/pten/hapi/include/backend_set.h
rename to paddle/pten/api/include/backend_set.h
diff --git a/paddle/pten/api/include/creation.h b/paddle/pten/api/include/creation.h
index 9795d88f81880b9f84491ff4c9ebe9428391510f..755038adb1f71c97496d40dafafcb8fff5a79193 100644
--- a/paddle/pten/api/include/creation.h
+++ b/paddle/pten/api/include/creation.h
@@ -14,26 +14,20 @@
 
 #pragma once
 
-#include "paddle/pten/api/include/infershape.h"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
-#include "paddle/pten/kernels/cpu/creation.h"
-#include "paddle/pten/kernels/cuda/creation.h"
+#include "paddle/pten/api/include/tensor.h"
+#include "paddle/pten/common/data_type.h"
+#include "paddle/pten/common/scalar.h"
 
-namespace pten {
+namespace paddle {
+namespace experimental {
 
-// TODO(YuanRisheng) This function name should be same as User API name.
-// TODO(zyfncg) Automatic code generation
-template <typename T, typename ContextT>
-DenseTensor FillAnyLike(const ContextT& dev_ctx,
-                        const DenseTensor& x,
-                        const Scalar& val) {
-  auto out_meta = UnchangedInferShape(x.meta());
-  const auto allocator =
-      std::make_shared<paddle::experimental::DefaultAllocator>(
-          dev_ctx.GetPlace());
-  pten::DenseTensor dense_out(allocator, out_meta);
-  FillAnyLike<T>(dev_ctx, x, val, &dense_out);
-  return dense_out;
-}
+Tensor full_like(const Tensor& x,
+                 const Scalar& value,
+                 DataType dtype = DataType::UNDEFINED);
 
-}  // namespace pten
+Tensor ones_like(const Tensor& x, DataType dtype = DataType::UNDEFINED);
+
+Tensor zeros_like(const Tensor& x, DataType dtype = DataType::UNDEFINED);
+
+}  // namespace experimental
+}  // namespace paddle
diff --git a/paddle/pten/api/include/linalg.h b/paddle/pten/api/include/linalg.h
index 0d4c7a60fbc1456d0ec4dcc06f1e3af51117e3b6..c28c133018464b635cd9c6b9037d3b12b8e91a5a 100644
--- a/paddle/pten/api/include/linalg.h
+++ b/paddle/pten/api/include/linalg.h
@@ -14,25 +14,17 @@
 
 #pragma once
 
-// See Note: [ How do we organize the kernel directory ]
-#include "paddle/pten/api/include/infershape.h"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
-#include "paddle/pten/kernels/cpu/linalg.h"
-#include "paddle/pten/kernels/cuda/linalg.h"
+#include "paddle/pten/api/include/tensor.h"
 
-namespace pten {
+namespace paddle {
+namespace experimental {
 
-template <typename T, typename ContextT>
-DenseTensor Dot(const ContextT& dev_ctx,
-                const DenseTensor& x,
-                const DenseTensor& y) {
-  auto out_meta = DotInferShape(x.meta(), y.meta());
-  const auto allocator =
-      std::make_shared<paddle::experimental::DefaultAllocator>(
-          dev_ctx.GetPlace());
-  pten::DenseTensor dense_out(allocator, out_meta);
-  Dot<T>(dev_ctx, x, y, &dense_out);
-  return dense_out;
-}
+Tensor dot(const Tensor& x, const Tensor& y);
 
-}  // namespace pten
+Tensor matmul(const Tensor& x,
+              const Tensor& y,
+              bool transpose_x,
+              bool transpose_y);
+
+}  // namespace experimental
+}  // namespace paddle
diff --git a/paddle/pten/api/include/manipulation.h b/paddle/pten/api/include/manipulation.h
index 1f867686a6eb7ee006ed6979939ca2e2e2531ad8..fe8c01cb74b95fe3c6159413df5e1f0e8b618eb1 100644
--- a/paddle/pten/api/include/manipulation.h
+++ b/paddle/pten/api/include/manipulation.h
@@ -1,39 +1,25 @@
-// 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.
+/* 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. */
 
 #pragma once
 
-// See Note: [ How do we organize the kernel directory ]
-#include "paddle/pten/api/include/infershape.h"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
-#include "paddle/pten/kernels/cpu/manipulation.h"
-#include "paddle/pten/kernels/cuda/manipulation.h"
-
-namespace pten {
-
-template <typename T, typename ContextT>
-DenseTensor Flatten(const ContextT& dev_ctx,
-                    const DenseTensor& x,
-                    int start_axis,
-                    int stop_axis) {
-  auto out_meta = FlattenInferShape(x.meta(), start_axis, stop_axis);
-  const auto allocator =
-      std::make_shared<paddle::experimental::DefaultAllocator>(
-          dev_ctx.GetPlace());
-  pten::DenseTensor dense_out(allocator, out_meta);
-  Flatten<T>(dev_ctx, x, start_axis, stop_axis, &dense_out);
-  return dense_out;
-}
-
-}  // namespace pten
+#include "paddle/pten/api/include/tensor.h"
+
+namespace paddle {
+namespace experimental {
+
+Tensor flatten(const Tensor& x, int start_axis, int stop_axis);
+
+}  // namespace experimental
+}  // namespace paddle
diff --git a/paddle/pten/api/include/math.h b/paddle/pten/api/include/math.h
index fa512e8d6db0d7f8dfc28b642ecd91e51066b648..9fea515646d6ec10a08a288e7cb1174728506164 100644
--- a/paddle/pten/api/include/math.h
+++ b/paddle/pten/api/include/math.h
@@ -14,63 +14,14 @@ limitations under the License. */
 
 #pragma once
 
-// See Note: [ How do we organize the kernel directory ]
-#include "paddle/pten/api/include/infershape.h"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
-#include "paddle/pten/kernels/cpu/math.h"
-#include "paddle/pten/kernels/cuda/math.h"
+#include "paddle/pten/api/include/tensor.h"
 
-namespace pten {
+namespace paddle {
+namespace experimental {
 
-template <typename T, typename ContextT>
-DenseTensor Sign(const ContextT& dev_ctx, const DenseTensor& x) {
-  auto out_meta = UnchangedInferShape(x.meta());
-  const auto allocator =
-      std::make_shared<paddle::experimental::DefaultAllocator>(
-          dev_ctx.GetPlace());
-  pten::DenseTensor dense_out(allocator, out_meta);
-  Sign<T>(dev_ctx, x, &dense_out);
-  return dense_out;
-}
+// TODO(chenweihang): add scale API
+// TODO(chenweihang): move mean API into stat.h/cc
+Tensor mean(const Tensor& x);
 
-template <typename T, typename ContextT>
-DenseTensor Mean(const ContextT& dev_ctx, const DenseTensor& x) {
-  auto out_meta = ReductionInferShape(x.meta());
-  const auto allocator =
-      std::make_shared<paddle::experimental::DefaultAllocator>(
-          dev_ctx.GetPlace());
-  pten::DenseTensor dense_out(allocator, out_meta);
-  Mean<T>(dev_ctx, x, &dense_out);
-  return dense_out;
-}
-
-template <typename T, typename ContextT>
-DenseTensor Scale(const ContextT& dev_ctx,
-                  const DenseTensor& x,
-                  float scale,
-                  float bias,
-                  bool bias_after_scale) {
-  auto out_meta = UnchangedInferShape(x.meta());
-  const auto allocator =
-      std::make_shared<paddle::experimental::DefaultAllocator>(
-          dev_ctx.GetPlace());
-  pten::DenseTensor dense_out(allocator, out_meta);
-  Scale<T>(dev_ctx, x, scale, bias, bias_after_scale, &dense_out);
-  return dense_out;
-}
-
-template <typename T, typename ContextT>
-DenseTensor Scale(const ContextT& dev_ctx,
-                  const DenseTensor& x,
-                  const DenseTensor& scale,
-                  float bias,
-                  bool bias_after_scale) {
-  auto out_meta = UnchangedInferShape(x.meta());
-  const auto allocator =
-      std::make_shared<paddle::experimental::DefaultAllocator>(
-          dev_ctx.GetPlace());
-  pten::DenseTensor dense_out(allocator, out_meta);
-  ScaleHost<T>(dev_ctx, x, scale, bias, bias_after_scale, &dense_out);
-  return dense_out;
-}
-}  // namespace pten
+}  // namespace experimental
+}  // namespace paddle
diff --git a/paddle/pten/hapi/include/tensor.h b/paddle/pten/api/include/tensor.h
similarity index 100%
rename from paddle/pten/hapi/include/tensor.h
rename to paddle/pten/api/include/tensor.h
diff --git a/paddle/pten/hapi/lib/CMakeLists.txt b/paddle/pten/api/lib/CMakeLists.txt
similarity index 100%
rename from paddle/pten/hapi/lib/CMakeLists.txt
rename to paddle/pten/api/lib/CMakeLists.txt
diff --git a/paddle/pten/hapi/lib/creation.cc b/paddle/pten/api/lib/creation.cc
similarity index 90%
rename from paddle/pten/hapi/lib/creation.cc
rename to paddle/pten/api/lib/creation.cc
index cda8d24b5e6ad791c8705a9396e51a1e4d367228..893f8b6fbc6f422d139fdb9b14c54a3484bbbe15 100644
--- a/paddle/pten/hapi/lib/creation.cc
+++ b/paddle/pten/api/lib/creation.cc
@@ -12,16 +12,16 @@ 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/hapi/include/creation.h"
+#include "paddle/pten/api/include/creation.h"
 
 #include <memory>
 
 #include "glog/logging.h"
 
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/api/include/infershape.h"
-#include "paddle/pten/hapi/lib/kernel_dispatch.h"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
+#include "paddle/pten/api/lib/kernel_dispatch.h"
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/infershape.h"
 
 namespace paddle {
 namespace experimental {
diff --git a/paddle/pten/hapi/lib/kernel_dispatch.h b/paddle/pten/api/lib/kernel_dispatch.h
similarity index 97%
rename from paddle/pten/hapi/lib/kernel_dispatch.h
rename to paddle/pten/api/lib/kernel_dispatch.h
index f61f3297d6d6c872b244a27319c6de7b16cabff4..567c21eeee9e86757db8842d49afd6ad5cfa02f6 100644
--- a/paddle/pten/hapi/lib/kernel_dispatch.h
+++ b/paddle/pten/api/lib/kernel_dispatch.h
@@ -18,10 +18,10 @@ limitations under the License. */
 #include <string>
 #include <utility>
 
+#include "paddle/pten/api/include/backend_set.h"
+#include "paddle/pten/api/include/tensor.h"
 #include "paddle/pten/common/data_type.h"
 #include "paddle/pten/common/layout.h"
-#include "paddle/pten/hapi/include/backend_set.h"
-#include "paddle/pten/hapi/include/tensor.h"
 
 // TODO(chenweihang): split KernelName, Key, Kernel, Factory into diff files
 #include "paddle/pten/core/convert_utils.h"
diff --git a/paddle/pten/hapi/lib/linalg.cc b/paddle/pten/api/lib/linalg.cc
similarity index 93%
rename from paddle/pten/hapi/lib/linalg.cc
rename to paddle/pten/api/lib/linalg.cc
index 3f13f546ee25e0351c6157f356462fc6d48990ce..587b9cd0f2726fc80d7674a8faa1f222c5e5d2fc 100644
--- a/paddle/pten/hapi/lib/linalg.cc
+++ b/paddle/pten/api/lib/linalg.cc
@@ -12,19 +12,19 @@ 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/hapi/include/linalg.h"
+#include "paddle/pten/api/include/linalg.h"
 
 #include <memory>
 
 #include "glog/logging.h"
 
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/api/include/infershape.h"
+#include "paddle/pten/api/lib/kernel_dispatch.h"
+#include "paddle/pten/api/lib/utils/allocator.h"
 #include "paddle/pten/core/convert_utils.h"
 #include "paddle/pten/core/dense_tensor.h"
 #include "paddle/pten/core/kernel_context.h"
-#include "paddle/pten/hapi/lib/kernel_dispatch.h"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/infershape.h"
 
 namespace paddle {
 namespace experimental {
diff --git a/paddle/pten/hapi/lib/manipulation.cc b/paddle/pten/api/lib/manipulation.cc
similarity index 91%
rename from paddle/pten/hapi/lib/manipulation.cc
rename to paddle/pten/api/lib/manipulation.cc
index fa60bac6d1aed95662ecde2aa26fe2f0a5908bb8..9f071ce8c2d14be013d69e14fa387457b28b27cf 100644
--- a/paddle/pten/hapi/lib/manipulation.cc
+++ b/paddle/pten/api/lib/manipulation.cc
@@ -12,14 +12,14 @@ 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/hapi/include/manipulation.h"
+#include "paddle/pten/api/include/manipulation.h"
 
 #include <memory>
 
 #include "glog/logging.h"
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/hapi/lib/kernel_dispatch.h"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
+#include "paddle/pten/api/lib/kernel_dispatch.h"
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/include/core.h"
 #include "paddle/pten/infershape/unary.h"
 
 namespace paddle {
diff --git a/paddle/pten/hapi/lib/math.cc b/paddle/pten/api/lib/math.cc
similarity index 88%
rename from paddle/pten/hapi/lib/math.cc
rename to paddle/pten/api/lib/math.cc
index 5e4e96d333030370787137a6c29d632e50676c3d..6cb7849e529e038eb8e253749d3876c1bec87029 100644
--- a/paddle/pten/hapi/lib/math.cc
+++ b/paddle/pten/api/lib/math.cc
@@ -12,16 +12,16 @@ 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/hapi/include/math.h"
+#include "paddle/pten/api/include/math.h"
 
 #include <memory>
 
 #include "glog/logging.h"
 
-#include "paddle/pten/api/include/core.h"
-#include "paddle/pten/api/include/infershape.h"
-#include "paddle/pten/hapi/lib/kernel_dispatch.h"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
+#include "paddle/pten/api/lib/kernel_dispatch.h"
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/infershape.h"
 #include "paddle/pten/infershape/unary.h"
 
 namespace paddle {
diff --git a/paddle/pten/hapi/lib/utils/CMakeLists.txt b/paddle/pten/api/lib/utils/CMakeLists.txt
similarity index 100%
rename from paddle/pten/hapi/lib/utils/CMakeLists.txt
rename to paddle/pten/api/lib/utils/CMakeLists.txt
diff --git a/paddle/pten/hapi/lib/utils/allocator.cc b/paddle/pten/api/lib/utils/allocator.cc
similarity index 93%
rename from paddle/pten/hapi/lib/utils/allocator.cc
rename to paddle/pten/api/lib/utils/allocator.cc
index 0c364c97e4d1cdd75edcbf1b9a6a8f44233596e7..e80152431e71292fcd34c96108c191f8c799a220 100644
--- a/paddle/pten/hapi/lib/utils/allocator.cc
+++ b/paddle/pten/api/lib/utils/allocator.cc
@@ -12,7 +12,7 @@ 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/hapi/lib/utils/allocator.h"
+#include "paddle/pten/api/lib/utils/allocator.h"
 
 namespace paddle {
 namespace experimental {
diff --git a/paddle/pten/hapi/lib/utils/allocator.h b/paddle/pten/api/lib/utils/allocator.h
similarity index 100%
rename from paddle/pten/hapi/lib/utils/allocator.h
rename to paddle/pten/api/lib/utils/allocator.h
diff --git a/paddle/pten/hapi/lib/utils/storage.cc b/paddle/pten/api/lib/utils/storage.cc
similarity index 96%
rename from paddle/pten/hapi/lib/utils/storage.cc
rename to paddle/pten/api/lib/utils/storage.cc
index 0682b25c6e0dd7c3c746c4e11ee8f25ebdf4826a..ba26e7f600d60c07df1179fdd1a4ff84d2383b5c 100644
--- a/paddle/pten/hapi/lib/utils/storage.cc
+++ b/paddle/pten/api/lib/utils/storage.cc
@@ -12,7 +12,7 @@ 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/hapi/lib/utils/storage.h"
+#include "paddle/pten/api/lib/utils/storage.h"
 
 namespace paddle {
 namespace experimental {
diff --git a/paddle/pten/hapi/lib/utils/storage.h b/paddle/pten/api/lib/utils/storage.h
similarity index 100%
rename from paddle/pten/hapi/lib/utils/storage.h
rename to paddle/pten/api/lib/utils/storage.h
diff --git a/paddle/pten/hapi/lib/utils/tensor_utils.cc b/paddle/pten/api/lib/utils/tensor_utils.cc
similarity index 98%
rename from paddle/pten/hapi/lib/utils/tensor_utils.cc
rename to paddle/pten/api/lib/utils/tensor_utils.cc
index f7641f424f4910d69c9ca71ddfa7968979a66c81..628fde3a1a4ddb089979356292e23df01f3afb4b 100644
--- a/paddle/pten/hapi/lib/utils/tensor_utils.cc
+++ b/paddle/pten/api/lib/utils/tensor_utils.cc
@@ -12,7 +12,7 @@ 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/hapi/lib/utils/tensor_utils.h"
+#include "paddle/pten/api/lib/utils/tensor_utils.h"
 
 namespace paddle {
 namespace experimental {
diff --git a/paddle/pten/hapi/lib/utils/tensor_utils.h b/paddle/pten/api/lib/utils/tensor_utils.h
similarity index 94%
rename from paddle/pten/hapi/lib/utils/tensor_utils.h
rename to paddle/pten/api/lib/utils/tensor_utils.h
index a2b2688362a4cfffc5a66c055994d021663d3bfb..625d6702f8b6d4814b2fbc52154ed4e4efc6fbdd 100644
--- a/paddle/pten/hapi/lib/utils/tensor_utils.h
+++ b/paddle/pten/api/lib/utils/tensor_utils.h
@@ -19,11 +19,11 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/variable.h"
 
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/api/lib/utils/storage.h"
 #include "paddle/pten/core/convert_utils.h"
 #include "paddle/pten/core/dense_tensor.h"
 #include "paddle/pten/core/kernel_factory.h"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
-#include "paddle/pten/hapi/lib/utils/storage.h"
 
 namespace paddle {
 namespace experimental {
diff --git a/paddle/pten/hapi/lib/utils/tests/CMakeLists.txt b/paddle/pten/api/lib/utils/tests/CMakeLists.txt
similarity index 100%
rename from paddle/pten/hapi/lib/utils/tests/CMakeLists.txt
rename to paddle/pten/api/lib/utils/tests/CMakeLists.txt
diff --git a/paddle/pten/hapi/lib/utils/tests/test_storage.cc b/paddle/pten/api/lib/utils/tests/test_storage.cc
similarity index 95%
rename from paddle/pten/hapi/lib/utils/tests/test_storage.cc
rename to paddle/pten/api/lib/utils/tests/test_storage.cc
index fbbcd2a3ee0e5c2f535103281fe5a1a5cd9fdf59..fc95f95f26e0b0099ad6b8a88529433eb5386b54 100644
--- a/paddle/pten/hapi/lib/utils/tests/test_storage.cc
+++ b/paddle/pten/api/lib/utils/tests/test_storage.cc
@@ -16,8 +16,8 @@ limitations under the License. */
 
 #include "gtest/gtest.h"
 
-#include "paddle/pten/hapi/lib/utils/allocator.h"
-#include "paddle/pten/hapi/lib/utils/storage.h"
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/api/lib/utils/storage.h"
 
 namespace paddle {
 namespace experimental {
diff --git a/paddle/pten/hapi/lib/utils/tests/test_tensor_utils.cc b/paddle/pten/api/lib/utils/tests/test_tensor_utils.cc
similarity index 98%
rename from paddle/pten/hapi/lib/utils/tests/test_tensor_utils.cc
rename to paddle/pten/api/lib/utils/tests/test_tensor_utils.cc
index 56184eec70f26606333ae2e6682b786c1e7dcb5a..fd52b96542c712bcd8449dbaa36d6bcda0149399 100644
--- a/paddle/pten/hapi/lib/utils/tests/test_tensor_utils.cc
+++ b/paddle/pten/api/lib/utils/tests/test_tensor_utils.cc
@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "gtest/gtest.h"
 
-#include "paddle/pten/hapi/lib/utils/tensor_utils.h"
+#include "paddle/pten/api/lib/utils/tensor_utils.h"
 
 namespace paddle {
 namespace experimental {
diff --git a/paddle/pten/hapi/CMakeLists.txt b/paddle/pten/hapi/CMakeLists.txt
deleted file mode 100644
index 4b427b3b4a38349f672aa8c643bd06ad7f39d360..0000000000000000000000000000000000000000
--- a/paddle/pten/hapi/CMakeLists.txt
+++ /dev/null
@@ -1,3 +0,0 @@
-add_subdirectory(lib)
-
-cc_library(pten_hapi SRCS all.cc DEPS linalg_api math_api creation_api)
diff --git a/paddle/pten/hapi/include/creation.h b/paddle/pten/hapi/include/creation.h
deleted file mode 100644
index 6f978be995273e22c4501584b71e5e1ee6dbda20..0000000000000000000000000000000000000000
--- a/paddle/pten/hapi/include/creation.h
+++ /dev/null
@@ -1,33 +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.
-
-#pragma once
-
-#include "paddle/pten/common/data_type.h"
-#include "paddle/pten/common/scalar.h"
-#include "paddle/pten/hapi/include/tensor.h"
-
-namespace paddle {
-namespace experimental {
-
-Tensor full_like(const Tensor& x,
-                 const Scalar& value,
-                 DataType dtype = DataType::UNDEFINED);
-
-Tensor ones_like(const Tensor& x, DataType dtype = DataType::UNDEFINED);
-
-Tensor zeros_like(const Tensor& x, DataType dtype = DataType::UNDEFINED);
-
-}  // namespace experimental
-}  // namespace paddle
diff --git a/paddle/pten/hapi/include/linalg.h b/paddle/pten/hapi/include/linalg.h
deleted file mode 100644
index 6e78b50af11c36b2341df211acd63202680f7f6c..0000000000000000000000000000000000000000
--- a/paddle/pten/hapi/include/linalg.h
+++ /dev/null
@@ -1,30 +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.
-
-#pragma once
-
-#include "paddle/pten/hapi/include/tensor.h"
-
-namespace paddle {
-namespace experimental {
-
-Tensor dot(const Tensor& x, const Tensor& y);
-
-Tensor matmul(const Tensor& x,
-              const Tensor& y,
-              bool transpose_x,
-              bool transpose_y);
-
-}  // namespace experimental
-}  // namespace paddle
diff --git a/paddle/pten/hapi/include/manipulation.h b/paddle/pten/hapi/include/manipulation.h
deleted file mode 100644
index 4622032f5ad545b66c6199a5e454b7c3c004fbbc..0000000000000000000000000000000000000000
--- a/paddle/pten/hapi/include/manipulation.h
+++ /dev/null
@@ -1,25 +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. */
-
-#pragma once
-
-#include "paddle/pten/hapi/include/tensor.h"
-
-namespace paddle {
-namespace experimental {
-
-Tensor flatten(const Tensor& x, int start_axis, int stop_axis);
-
-}  // namespace experimental
-}  // namespace paddle
diff --git a/paddle/pten/hapi/include/math.h b/paddle/pten/hapi/include/math.h
deleted file mode 100644
index db4010c1c14e3aae4b6da5ba1fe7c1a7cee2de82..0000000000000000000000000000000000000000
--- a/paddle/pten/hapi/include/math.h
+++ /dev/null
@@ -1,27 +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. */
-
-#pragma once
-
-#include "paddle/pten/hapi/include/tensor.h"
-
-namespace paddle {
-namespace experimental {
-
-// TODO(chenweihang): add scale API
-// TODO(chenweihang): move mean API into stat.h/cc
-Tensor mean(const Tensor& x);
-
-}  // namespace experimental
-}  // namespace paddle
diff --git a/paddle/pten/api/include/core.h b/paddle/pten/include/core.h
similarity index 100%
rename from paddle/pten/api/include/core.h
rename to paddle/pten/include/core.h
diff --git a/paddle/pten/include/creation.h b/paddle/pten/include/creation.h
new file mode 100644
index 0000000000000000000000000000000000000000..0fb3f905e03fbfb0b9743ff631a02452f0fdd577
--- /dev/null
+++ b/paddle/pten/include/creation.h
@@ -0,0 +1,39 @@
+// 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.
+
+#pragma once
+
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/include/infershape.h"
+#include "paddle/pten/kernels/cpu/creation.h"
+#include "paddle/pten/kernels/cuda/creation.h"
+
+namespace pten {
+
+// TODO(YuanRisheng) This function name should be same as User API name.
+// TODO(zyfncg) Automatic code generation
+template <typename T, typename ContextT>
+DenseTensor FillAnyLike(const ContextT& dev_ctx,
+                        const DenseTensor& x,
+                        const Scalar& val) {
+  auto out_meta = UnchangedInferShape(x.meta());
+  const auto allocator =
+      std::make_shared<paddle::experimental::DefaultAllocator>(
+          dev_ctx.GetPlace());
+  pten::DenseTensor dense_out(allocator, out_meta);
+  FillAnyLike<T>(dev_ctx, x, val, &dense_out);
+  return dense_out;
+}
+
+}  // namespace pten
diff --git a/paddle/pten/api/include/infershape.h b/paddle/pten/include/infershape.h
similarity index 100%
rename from paddle/pten/api/include/infershape.h
rename to paddle/pten/include/infershape.h
diff --git a/paddle/pten/include/linalg.h b/paddle/pten/include/linalg.h
new file mode 100644
index 0000000000000000000000000000000000000000..70eebac5b684154e5dd9d8242aed562391562c38
--- /dev/null
+++ b/paddle/pten/include/linalg.h
@@ -0,0 +1,38 @@
+// 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.
+
+#pragma once
+
+// See Note: [ How do we organize the kernel directory ]
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/include/infershape.h"
+#include "paddle/pten/kernels/cpu/linalg.h"
+#include "paddle/pten/kernels/cuda/linalg.h"
+
+namespace pten {
+
+template <typename T, typename ContextT>
+DenseTensor Dot(const ContextT& dev_ctx,
+                const DenseTensor& x,
+                const DenseTensor& y) {
+  auto out_meta = DotInferShape(x.meta(), y.meta());
+  const auto allocator =
+      std::make_shared<paddle::experimental::DefaultAllocator>(
+          dev_ctx.GetPlace());
+  pten::DenseTensor dense_out(allocator, out_meta);
+  Dot<T>(dev_ctx, x, y, &dense_out);
+  return dense_out;
+}
+
+}  // namespace pten
diff --git a/paddle/pten/include/manipulation.h b/paddle/pten/include/manipulation.h
new file mode 100644
index 0000000000000000000000000000000000000000..236f7c7af956199066954aed5a05161e21614a67
--- /dev/null
+++ b/paddle/pten/include/manipulation.h
@@ -0,0 +1,39 @@
+// 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.
+
+#pragma once
+
+// See Note: [ How do we organize the kernel directory ]
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/include/infershape.h"
+#include "paddle/pten/kernels/cpu/manipulation.h"
+#include "paddle/pten/kernels/cuda/manipulation.h"
+
+namespace pten {
+
+template <typename T, typename ContextT>
+DenseTensor Flatten(const ContextT& dev_ctx,
+                    const DenseTensor& x,
+                    int start_axis,
+                    int stop_axis) {
+  auto out_meta = FlattenInferShape(x.meta(), start_axis, stop_axis);
+  const auto allocator =
+      std::make_shared<paddle::experimental::DefaultAllocator>(
+          dev_ctx.GetPlace());
+  pten::DenseTensor dense_out(allocator, out_meta);
+  Flatten<T>(dev_ctx, x, start_axis, stop_axis, &dense_out);
+  return dense_out;
+}
+
+}  // namespace pten
diff --git a/paddle/pten/include/math.h b/paddle/pten/include/math.h
new file mode 100644
index 0000000000000000000000000000000000000000..0a61a94aa8d69687d86f8f6f0b4f1edd107ebce6
--- /dev/null
+++ b/paddle/pten/include/math.h
@@ -0,0 +1,76 @@
+/* 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. */
+
+#pragma once
+
+// See Note: [ How do we organize the kernel directory ]
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/include/infershape.h"
+#include "paddle/pten/kernels/cpu/math.h"
+#include "paddle/pten/kernels/cuda/math.h"
+
+namespace pten {
+
+template <typename T, typename ContextT>
+DenseTensor Sign(const ContextT& dev_ctx, const DenseTensor& x) {
+  auto out_meta = UnchangedInferShape(x.meta());
+  const auto allocator =
+      std::make_shared<paddle::experimental::DefaultAllocator>(
+          dev_ctx.GetPlace());
+  pten::DenseTensor dense_out(allocator, out_meta);
+  Sign<T>(dev_ctx, x, &dense_out);
+  return dense_out;
+}
+
+template <typename T, typename ContextT>
+DenseTensor Mean(const ContextT& dev_ctx, const DenseTensor& x) {
+  auto out_meta = ReductionInferShape(x.meta());
+  const auto allocator =
+      std::make_shared<paddle::experimental::DefaultAllocator>(
+          dev_ctx.GetPlace());
+  pten::DenseTensor dense_out(allocator, out_meta);
+  Mean<T>(dev_ctx, x, &dense_out);
+  return dense_out;
+}
+
+template <typename T, typename ContextT>
+DenseTensor Scale(const ContextT& dev_ctx,
+                  const DenseTensor& x,
+                  float scale,
+                  float bias,
+                  bool bias_after_scale) {
+  auto out_meta = UnchangedInferShape(x.meta());
+  const auto allocator =
+      std::make_shared<paddle::experimental::DefaultAllocator>(
+          dev_ctx.GetPlace());
+  pten::DenseTensor dense_out(allocator, out_meta);
+  Scale<T>(dev_ctx, x, scale, bias, bias_after_scale, &dense_out);
+  return dense_out;
+}
+
+template <typename T, typename ContextT>
+DenseTensor Scale(const ContextT& dev_ctx,
+                  const DenseTensor& x,
+                  const DenseTensor& scale,
+                  float bias,
+                  bool bias_after_scale) {
+  auto out_meta = UnchangedInferShape(x.meta());
+  const auto allocator =
+      std::make_shared<paddle::experimental::DefaultAllocator>(
+          dev_ctx.GetPlace());
+  pten::DenseTensor dense_out(allocator, out_meta);
+  ScaleHost<T>(dev_ctx, x, scale, bias, bias_after_scale, &dense_out);
+  return dense_out;
+}
+}  // namespace pten
diff --git a/paddle/pten/kernels/cuda/math.cu b/paddle/pten/kernels/cuda/math.cu
index 0ead1f8048bfd6b0252caf495ed7afad73f108bc..cc1a7bef4f18e2d226905182ade5930721eafa2b 100644
--- a/paddle/pten/kernels/cuda/math.cu
+++ b/paddle/pten/kernels/cuda/math.cu
@@ -28,9 +28,9 @@ namespace cub = hipcub;
 
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/float16.h"
+#include "paddle/pten/api/lib/utils/tensor_utils.h"
 #include "paddle/pten/core/convert_utils.h"
 #include "paddle/pten/core/kernel_registry.h"
-#include "paddle/pten/hapi/lib/utils/tensor_utils.h"
 
 namespace pten {
 
diff --git a/paddle/pten/tests/test_copy_api.cc b/paddle/pten/tests/test_copy_api.cc
index fcebe9a310dea87bc3aed8473458655b376effe1..4bf16e7517b7336209a887d4a6843c8bdf77fd9a 100644
--- a/paddle/pten/tests/test_copy_api.cc
+++ b/paddle/pten/tests/test_copy_api.cc
@@ -18,8 +18,8 @@ limitations under the License. */
 #include "paddle/pten/core/kernel_registry.h"
 #include "paddle/pten/kernels/cpu/utils.h"
 
+#include "paddle/pten/api/lib/utils/allocator.h"
 #include "paddle/pten/core/dense_tensor.h"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
 
 PT_DECLARE_MODULE(UtilsCPU);
 
diff --git a/paddle/pten/tests/test_dot_api.cc b/paddle/pten/tests/test_dot_api.cc
index 5401b66544473940938e8fdf55ab86ff580b2463..f79854e557e5a7786eeea48143848a73ceced415 100644
--- a/paddle/pten/tests/test_dot_api.cc
+++ b/paddle/pten/tests/test_dot_api.cc
@@ -15,13 +15,13 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <memory>
 
-#include "paddle/pten/hapi/include/linalg.h"
+#include "paddle/pten/api/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"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
 
-#include "paddle/pten/api/include/linalg.h"
+#include "paddle/pten/include/linalg.h"
 
 PT_DECLARE_MODULE(LinalgCPU);
 
diff --git a/paddle/pten/tests/test_fill_api.cc b/paddle/pten/tests/test_fill_api.cc
index 5a788226086dcc6953357d32fec966493e2d6baa..a83ed76a6f8dfa1da0f0ccedbff993d01d6b2158 100644
--- a/paddle/pten/tests/test_fill_api.cc
+++ b/paddle/pten/tests/test_fill_api.cc
@@ -15,13 +15,13 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <memory>
 
-#include "paddle/pten/hapi/include/creation.h"
+#include "paddle/pten/api/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"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
 
-#include "paddle/pten/api/include/creation.h"
+#include "paddle/pten/include/creation.h"
 
 PT_DECLARE_MODULE(CreationCPU);
 
diff --git a/paddle/pten/tests/test_flatten_api.cc b/paddle/pten/tests/test_flatten_api.cc
index dfb777678a94d05fbbf4ad401aa89cfdfad2decd..ca8ca38f65a3ee4cf68ebfaaf13e7eeacd1c31ea 100644
--- a/paddle/pten/tests/test_flatten_api.cc
+++ b/paddle/pten/tests/test_flatten_api.cc
@@ -15,13 +15,13 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <memory>
 
-#include "paddle/pten/hapi/include/manipulation.h"
+#include "paddle/pten/api/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"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
 
-#include "paddle/pten/api/include/manipulation.h"
+#include "paddle/pten/include/manipulation.h"
 
 PT_DECLARE_MODULE(ManipulationCPU);
 
diff --git a/paddle/pten/tests/test_matmul_api.cc b/paddle/pten/tests/test_matmul_api.cc
index b0579834519aa3ef32e912c3714e3beef9f0fe1a..83a70c905b60ec09c710f6f3047220cfb9675c98 100644
--- a/paddle/pten/tests/test_matmul_api.cc
+++ b/paddle/pten/tests/test_matmul_api.cc
@@ -15,11 +15,11 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <memory>
 
-#include "paddle/pten/hapi/include/linalg.h"
+#include "paddle/pten/api/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"
-#include "paddle/pten/hapi/lib/utils/allocator.h"
 #include "paddle/pten/kernels/cuda/utils.h"
 
 PT_DECLARE_MODULE(LinalgCPU);
diff --git a/paddle/pten/tests/test_mean_api.cc b/paddle/pten/tests/test_mean_api.cc
index b3da90659d005a0d9fbcd38b0cbfa947276ad03c..4dd4222daabe473a499e510675225b25d20d4a46 100644
--- a/paddle/pten/tests/test_mean_api.cc
+++ b/paddle/pten/tests/test_mean_api.cc
@@ -15,13 +15,13 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <memory>
 
-#include "paddle/pten/hapi/include/math.h"
+#include "paddle/pten/api/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/hapi/lib/utils/allocator.h"
 
-#include "paddle/pten/api/include/math.h"
+#include "paddle/pten/include/math.h"
 
 PT_DECLARE_MODULE(MathCPU);
 
diff --git a/paddle/pten/tests/test_scale_api.cc b/paddle/pten/tests/test_scale_api.cc
index 9f80d6d2cc126bb4687626f267d537601bca7ff2..2ecf6d199b55855bdb95f803d333084bcb2dc2b1 100644
--- a/paddle/pten/tests/test_scale_api.cc
+++ b/paddle/pten/tests/test_scale_api.cc
@@ -15,13 +15,13 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <memory>
 
-#include "paddle/pten/hapi/include/math.h"
+#include "paddle/pten/api/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/hapi/lib/utils/allocator.h"
 
-#include "paddle/pten/api/include/math.h"
+#include "paddle/pten/include/math.h"
 
 PT_DECLARE_MODULE(MathCPU);