Add Intermediate Kernel API for refactor Tensor Lib

paddle/fluid/framework/operator.cc

@@ -23,7 +23,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/data_type_transform.h"
 #include "paddle/fluid/framework/details/nan_inf_utils.h"
 #include "paddle/fluid/framework/op_call_stack.h"
-#include "paddle/fluid/framework/pten_utils.h"
 #include "paddle/fluid/framework/shape_inference.h"
 #include "paddle/fluid/framework/transfer_scope_cache.h"
 #include "paddle/fluid/framework/unused_var_check.h"

paddle/fluid/imperative/prepared_operator.cc

@@ -16,7 +16,6 @@
 
 #include "paddle/fluid/framework/data_type_transform.h"
 #include "paddle/fluid/framework/details/nan_inf_utils.h"
-#include "paddle/fluid/framework/pten_utils.h"
 #include "paddle/fluid/imperative/infer_shape_context.h"
 #include "paddle/pten/common/scalar.h"
 #include "paddle/utils/small_vector.h"

paddle/pten/api/include/creation.h

@@ -14,5 +14,26 @@
 
 #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"
+
+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

paddle/pten/api/include/linalg.h

@@ -15,5 +15,24 @@
 #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"
+
+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

paddle/pten/api/include/manipulation.h

@@ -15,5 +15,25 @@
 #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

paddle/pten/api/include/math.h

@@ -15,5 +15,62 @@ 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"
+
+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

paddle/pten/hapi/lib/utils/tensor_utils.cc

@@ -45,6 +45,7 @@ std::unique_ptr<pten::DenseTensor> MakePtenDenseTensor(
   SetLoD(&meta.lod, src.lod());
   auto shared_storage =
       pten::make_intrusive<SharedStorage>(src.Holder(), src.offset());
+
   return std::make_unique<pten::DenseTensor>(std::move(shared_storage),
                                              std::move(meta));
 }

paddle/pten/kernels/cpu/manipulation.cc

@@ -24,10 +24,9 @@ void Flatten(const CPUContext& dev_ctx,
              int start_axis,
              int stop_axis,
              DenseTensor* out) {
-  auto out_meta = FlattenInferShape(x.meta(), start_axis, stop_axis);
+  auto out_dims = out->dims();
   pten::Copy(dev_ctx, x, out);
-  out->set_lod(out_meta.lod);
-  out->Resize(out_meta.dims);
+  out->Resize(out_dims);
 }
 
 // TODO(yuanrisheng): this kernel is for training and xshape is a Intermediate

paddle/pten/kernels/cuda/manipulation.cu

@@ -24,10 +24,9 @@ void Flatten(const CUDAContext& dev_ctx,
              int start_axis,
              int stop_axis,
              DenseTensor* out) {
-  auto out_meta = FlattenInferShape(x.meta(), start_axis, stop_axis);
+  auto out_dims = out->dims();
   pten::Copy(dev_ctx, x, out);
-  out->set_lod(out_meta.lod);
-  out->Resize(out_meta.dims);
+  out->Resize(out_dims);
 }
 
 // TODO(yuanrisheng): this kernel is for training and xshape is a Intermediate

paddle/pten/tests/CMakeLists.txt

@@ -12,3 +12,4 @@ 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)

paddle/pten/tests/test_dot_api.cc

@@ -21,6 +21,8 @@ limitations under the License. */
 #include "paddle/pten/core/kernel_registry.h"
 #include "paddle/pten/hapi/lib/utils/allocator.h"
 
+#include "paddle/pten/api/include/linalg.h"
+
 PT_DECLARE_MODULE(LinalgCPU);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
@@ -82,3 +84,55 @@ 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);
+}

paddle/pten/tests/test_fill_api.cc

@@ -21,6 +21,8 @@ limitations under the License. */
 #include "paddle/pten/core/kernel_registry.h"
 #include "paddle/pten/hapi/lib/utils/allocator.h"
 
+#include "paddle/pten/api/include/creation.h"
+
 PT_DECLARE_MODULE(CreationCPU);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
@@ -131,3 +133,38 @@ 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);
+  }
+}

paddle/pten/tests/test_flatten_api.cc

@@ -21,6 +21,8 @@ limitations under the License. */
 #include "paddle/pten/core/kernel_registry.h"
 #include "paddle/pten/hapi/lib/utils/allocator.h"
 
+#include "paddle/pten/api/include/manipulation.h"
+
 PT_DECLARE_MODULE(ManipulationCPU);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
@@ -70,3 +72,47 @@ 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);
+}