split fill_constant & fill_constant_batch_size_like, enable and add uts (#3023)

lite/core/arena/framework.cc

@@ -76,7 +76,7 @@ void TestCase::PrepareInputsForInstruction() {
 
       const auto* inst_type = Type::GetTensorTy(TARGET(kHost));
       CHECK(scope_->FindVar(var));
-      const auto* shared_tensor = scope_->FindTensor((var));
+      const auto* shared_tensor = scope_->FindTensor(var);
       if (!TargetCompatibleTo(*inst_type, *param_type->type)) {
         /// Create a tensor in the instruction's scope, alloc memory and then
         /// copy data there.

lite/kernels/arm/CMakeLists.txt

@@ -83,7 +83,6 @@ add_kernel(conditional_block_compute_arm ARM extra SRCS conditional_block_comput
 add_kernel(collect_fpn_proposals_compute_arm ARM extra SRCS collect_fpn_proposals_compute.cc DEPS ${lite_kernel_deps} math_arm)
 add_kernel(distribute_fpn_proposals_compute_arm ARM extra SRCS distribute_fpn_proposals_compute.cc DEPS ${lite_kernel_deps} math_arm)
 
-
 # for OCR specific
 add_kernel(gru_unit_compute_arm ARM extra SRCS gru_unit_compute.cc DEPS ${lite_kernel_deps} math_arm)
 add_kernel(gru_compute_arm ARM extra SRCS gru_compute.cc DEPS ${lite_kernel_deps} math_arm)
@@ -100,6 +99,7 @@ add_kernel(write_to_array_compute_arm ARM extra SRCS write_to_array_compute.cc D
 add_kernel(read_from_array_compute_arm ARM extra SRCS read_from_array_compute.cc DEPS ${lite_kernel_deps} math_arm)
 add_kernel(beam_search_compute_arm ARM extra SRCS beam_search_compute.cc DEPS ${lite_kernel_deps} math_arm)
 add_kernel(fill_constant_compute_arm ARM basic SRCS fill_constant_compute.cc DEPS ${lite_kernel_deps} math_arm)
+add_kernel(fill_constant_batch_size_like_compute_arm ARM basic SRCS fill_constant_batch_size_like_compute.cc DEPS ${lite_kernel_deps} math_arm)
 add_kernel(lod_reset_compute_arm ARM extra SRCS lod_reset_compute.cc DEPS ${lite_kernel_deps} math_arm)
 add_kernel(is_empty_compute_arm ARM extra SRCS is_empty_compute.cc DEPS ${lite_kernel_deps} math_arm)
 add_kernel(lstm_arm ARM extra SRCS lstm_compute.cc DEPS ${lite_kernel_deps} math_arm)

lite/kernels/arm/fill_constant_batch_size_like_compute.cc

+// Copyright (c) 2019 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 "lite/kernels/arm/fill_constant_batch_size_like_compute.h"
+
+namespace paddle {
+namespace lite {
+namespace kernels {
+namespace arm {
+
+void FillConstantBatchSizeLikeCompute::Run() {
+  auto& param = *param_.get_mutable<param_t>();
+  auto& context = ctx_->As<ARMContext>();
+
+  if (param.dtype == static_cast<int32_t>(lite::core::FluidType::FP32)) {
+    auto data = param.out->template mutable_data<float>();
+    for (int i = 0; i < param.out->numel(); i++) {
+      data[i] = param.value;
+    }
+  } else if (param.dtype ==
+             static_cast<int32_t>(lite::core::FluidType::INT32)) {
+    auto data = param.out->template mutable_data<int32_t>();
+    for (int i = 0; i < param.out->numel(); i++) {
+      data[i] = param.value;
+    }
+  } else if (param.dtype == static_cast<int32_t>(lite::core::FluidType::INT8)) {
+    auto data = param.out->template mutable_data<int8_t>();
+    for (int i = 0; i < param.out->numel(); i++) {
+      data[i] = param.value;
+    }
+  } else {
+    LOG(FATAL) << "not supported dtype " << param.dtype;
+  }
+}
+
+}  // namespace arm
+}  // namespace kernels
+}  // namespace lite
+}  // namespace paddle
+
+REGISTER_LITE_KERNEL(
+    fill_constant_batch_size_like,
+    kARM,
+    kFloat,
+    kNCHW,
+    paddle::lite::kernels::arm::FillConstantBatchSizeLikeCompute,
+    def)
+    .BindInput("Input", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kAny))})
+    .BindOutput("Out", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kAny))})
+    .Finalize();

lite/kernels/arm/fill_constant_batch_size_like_compute.h

+// Copyright (c) 2019 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 "lite/core/kernel.h"
+#include "lite/core/op_registry.h"
+
+namespace paddle {
+namespace lite {
+namespace kernels {
+namespace arm {
+
+class FillConstantBatchSizeLikeCompute
+    : public KernelLite<TARGET(kARM), PRECISION(kFloat)> {
+ public:
+  using param_t = operators::FillConstantBatchSizeLikeParam;
+
+  void Run() override;
+
+  ~FillConstantBatchSizeLikeCompute() {}
+};
+
+}  // namespace arm
+}  // namespace kernels
+}  // namespace lite
+}  // namespace paddle

lite/kernels/arm/fill_constant_compute.cc

@@ -12,118 +12,37 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "lite/core/kernel.h"
-#include "lite/core/op_registry.h"
+#include "lite/kernels/arm/fill_constant_compute.h"
 
 namespace paddle {
 namespace lite {
 namespace kernels {
 namespace arm {
 
-class FillConstantCompute : public KernelLite<TARGET(kARM), PRECISION(kAny)> {
- public:
-  using param_t = operators::FillConstantParam;
+void FillConstantCompute::Run() {
+  auto& param = *param_.get_mutable<param_t>();
+  auto& context = ctx_->As<ARMContext>();
 
-  inline DDimLite GetShape(const param_t& param) {
-    // 1. shape is a Tensor
-    if (param.shape_tensor != nullptr) {
-      auto* shape_tensor = param.shape_tensor;
-      auto* shape_data = shape_tensor->data<int>();
-      auto vec_shape =
-          std::vector<int64_t>(shape_data, shape_data + shape_tensor->numel());
-      return DDimLite(vec_shape);
+  if (param.dtype == static_cast<int32_t>(lite::core::FluidType::FP32)) {
+    auto data = param.out->template mutable_data<float>();
+    for (int i = 0; i < param.out->numel(); i++) {
+      data[i] = param.value;
     }
-
-    // 2. shape is a list/tuple containing Tensor
-    auto shape_tensor_list = param.shape_tensor_list;
-    if (shape_tensor_list.size() > 0) {
-      std::vector<int64_t> vec_shape;
-      for (size_t i = 0; i < shape_tensor_list.size(); ++i) {
-        auto tensor = shape_tensor_list[i];
-        vec_shape.push_back(*tensor->data<int>());
-      }
-      return DDimLite(vec_shape);
+  } else if (param.dtype ==
+             static_cast<int32_t>(lite::core::FluidType::INT32)) {
+    auto data = param.out->template mutable_data<int32_t>();
+    for (int i = 0; i < param.out->numel(); i++) {
+      data[i] = param.value;
     }
-
-    // 3. shape is a list/tuple without containing Tensor
-    auto vec_shape = param.shape;
-    return DDimLite(vec_shape);
-  }
-
-  void PrepareForRun() override {
-    auto& param = *param_.get_mutable<param_t>();
-    auto outdims = GetShape(param);
-    param.Out->Resize(outdims);
-  }
-
-  void Run() override {
-    auto& param = *param_.get_mutable<param_t>();
-    auto& context = ctx_->As<ARMContext>();
-
-    if (param.dtype == static_cast<int32_t>(lite::core::FluidType::FP32)) {
-      auto data = param.Out->template mutable_data<float>();
-      for (int i = 0; i < param.Out->numel(); i++) {
-        data[i] = param.value;
-      }
-    } else if (param.dtype ==
-               static_cast<int32_t>(lite::core::FluidType::INT32)) {
-      auto data = param.Out->template mutable_data<int32_t>();
-      for (int i = 0; i < param.Out->numel(); i++) {
-        data[i] = param.value;
-      }
-    } else if (param.dtype ==
-               static_cast<int32_t>(lite::core::FluidType::INT8)) {
-      auto data = param.Out->template mutable_data<int8_t>();
-      for (int i = 0; i < param.Out->numel(); i++) {
-        data[i] = param.value;
-      }
-    } else {
-      LOG(FATAL) << "not supported dtype " << param.dtype;
+  } else if (param.dtype == static_cast<int32_t>(lite::core::FluidType::INT8)) {
+    auto data = param.out->template mutable_data<int8_t>();
+    for (int i = 0; i < param.out->numel(); i++) {
+      data[i] = param.value;
     }
+  } else {
+    LOG(FATAL) << "not supported dtype " << param.dtype;
   }
-
-  virtual ~FillConstantCompute() = default;
-};
-
-class FillConstantBatchLikeCompute
-    : public KernelLite<TARGET(kARM), PRECISION(kAny)> {
- public:
-  using param_t = operators::FillConstantBatchLikeParam;
-
-  void Run() override {
-    auto& param = *param_.get_mutable<param_t>();
-    auto& context = ctx_->As<ARMContext>();
-
-    if (param.input->lod().size() && param.input_dim_idx == 0) {
-      auto odims = param.out->dims();
-      odims[param.output_dim_idx] = param.input->lod().back().size() - 1;
-      param.out->Resize(odims);
-    }
-
-    if (param.dtype == static_cast<int32_t>(lite::core::FluidType::FP32)) {
-      auto data = param.out->template mutable_data<float>();
-      for (int i = 0; i < param.out->numel(); i++) {
-        data[i] = param.value;
-      }
-    } else if (param.dtype ==
-               static_cast<int32_t>(lite::core::FluidType::INT32)) {
-      auto data = param.out->template mutable_data<int32_t>();
-      for (int i = 0; i < param.out->numel(); i++) {
-        data[i] = param.value;
-      }
-    } else if (param.dtype ==
-               static_cast<int32_t>(lite::core::FluidType::INT8)) {
-      auto data = param.out->template mutable_data<int8_t>();
-      for (int i = 0; i < param.out->numel(); i++) {
-        data[i] = param.value;
-      }
-    } else {
-      LOG(FATAL) << "not supported dtype " << param.dtype;
-    }
-  }
-
-  virtual ~FillConstantBatchLikeCompute() = default;
-};
+}
 
 }  // namespace arm
 }  // namespace kernels
@@ -133,23 +52,13 @@ class FillConstantBatchLikeCompute
 // float
 REGISTER_LITE_KERNEL(fill_constant,
                      kARM,
-                     kAny,
+                     kFloat,
                      kNCHW,
                      paddle::lite::kernels::arm::FillConstantCompute,
                      def)
-    .BindInput("X", {LiteType::GetTensorTy(TARGET(kARM))})
     .BindInput("ShapeTensor",
                {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kInt32))})
     .BindInput("ShapeTensorList",
                {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kInt32))})
     .BindOutput("Out", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kAny))})
     .Finalize();
-REGISTER_LITE_KERNEL(fill_constant_batch_size_like,
-                     kARM,
-                     kAny,
-                     kNCHW,
-                     paddle::lite::kernels::arm::FillConstantBatchLikeCompute,
-                     def)
-    .BindInput("Input", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kAny))})
-    .BindOutput("Out", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kAny))})
-    .Finalize();

lite/kernels/arm/fill_constant_compute.h

+// Copyright (c) 2019 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 "lite/core/kernel.h"
+#include "lite/core/op_registry.h"
+
+namespace paddle {
+namespace lite {
+namespace kernels {
+namespace arm {
+
+class FillConstantCompute : public KernelLite<TARGET(kARM), PRECISION(kFloat)> {
+ public:
+  using param_t = operators::FillConstantParam;
+
+  void Run() override;
+
+  ~FillConstantCompute() {}
+};
+
+}  // namespace arm
+}  // namespace kernels
+}  // namespace lite
+}  // namespace paddle

lite/kernels/x86/fill_constant_batch_size_like_compute.h

@@ -34,8 +34,8 @@ class FillConstantBatchSizeLikeCompute
   void Run() override {
     auto& param = *param_.get_mutable<param_t>();
     auto& ctx = ctx_->As<X86Context>();
-    auto* out = param.Out;
-    auto* in = param.Input;
+    auto* out = param.out;
+    auto* in = param.input;
     if (in->lod().size() && param.input_dim_idx == 0) {
       // set the correct batch size for the LoDTensor.
       auto odims = out->dims();

lite/kernels/x86/fill_constant_batch_size_like_compute_test.cc

@@ -56,8 +56,8 @@ TEST(fill_constant_batch_size_like_x86, run_test) {
 
   FillConstantBatchSizeLikeCompute<float> fill_constant_batch_size_like;
   operators::FillConstantBatchSizeLikeParam param;
-  param.Input = &input;
-  param.Out = &out;
+  param.input = &input;
+  param.out = &out;
   std::vector<int> shape{-1, 132, 7};
   float value = 3.5;
   param.shape = shape;

lite/operators/CMakeLists.txt

@@ -20,6 +20,7 @@ add_operator(box_coder_op_lite basic SRCS box_coder_op.cc DEPS ${op_DEPS})
 add_operator(multiclass_nms_op_lite basic SRCS multiclass_nms_op.cc DEPS ${op_DEPS})
 add_operator(mean_op basic SRCS mean_op.cc DEPS ${op_DEPS})
 add_operator(fill_constant_op basic SRCS fill_constant_op.cc DEPS ${op_DEPS})
+add_operator(fill_constant_batch_size_like_op basic SRCS fill_constant_batch_size_like_op.cc DEPS ${op_DEPS})
 add_operator(shuffle_channel_op basic SRCS shuffle_channel_op.cc DEPS ${op_DEPS})
 add_operator(yolo_box_op basic SRCS yolo_box_op.cc DEPS ${op_DEPS})
 add_operator(interpolate_op basic SRCS interpolate_op.cc DEPS ${op_DEPS})

lite/operators/fill_constant_batch_size_like_op.cc

+// Copyright (c) 2019 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 "lite/operators/fill_constant_batch_size_like_op.h"
+#include "lite/core/op_registry.h"
+
+namespace paddle {
+namespace lite {
+namespace operators {
+
+bool FillConstantBatchSizeLikeOp::CheckShape() const {
+  CHECK(param_.out);
+  CHECK(param_.input);
+  CHECK_GT(param_.shape.size(), 0);
+  CHECK_GE(param_.input_dim_idx, 0);
+  CHECK_GE(param_.output_dim_idx, 0);
+  return true;
+}
+
+bool FillConstantBatchSizeLikeOp::InferShape() const {
+  std::vector<int64_t> output_dim{param_.shape.begin(), param_.shape.end()};
+  if (param_.input_dim_idx == 0 && !param_.input->lod().empty()) {
+    output_dim[param_.output_dim_idx] = param_.input->lod().back().size() - 1;
+  } else {
+    output_dim[param_.output_dim_idx] =
+        param_.input->dims()[param_.input_dim_idx];
+  }
+  param_.out->Resize(output_dim);
+  return true;
+}
+
+bool FillConstantBatchSizeLikeOp::AttachImpl(const cpp::OpDesc& opdesc,
+                                             lite::Scope* scope) {
+  auto out_name = opdesc.Output("Out").front();
+  auto input_name = opdesc.Input("Input").front();
+
+  param_.out = GetMutableVar<lite::Tensor>(scope, out_name);
+  param_.input = GetMutableVar<lite::Tensor>(scope, input_name);
+  param_.dtype = opdesc.GetAttr<int>("dtype");
+  param_.shape = opdesc.GetAttr<std::vector<int>>("shape");
+  if (opdesc.HasAttr("value")) {
+    param_.value = opdesc.GetAttr<float>("value");
+  }
+  if (opdesc.HasAttr("input_dim_idx")) {
+    param_.input_dim_idx = opdesc.GetAttr<int>("input_dim_idx");
+  }
+  if (opdesc.HasAttr("output_dim_idx")) {
+    param_.output_dim_idx = opdesc.GetAttr<int>("output_dim_idx");
+  }
+
+  return true;
+}
+
+}  // namespace operators
+}  // namespace lite
+}  // namespace paddle
+
+REGISTER_LITE_OP(fill_constant_batch_size_like,
+                 paddle::lite::operators::FillConstantBatchSizeLikeOp);

lite/operators/fill_constant_batch_size_like_op.h

@@ -45,6 +45,6 @@ class FillConstantBatchSizeLikeOp : public OpLite {
   mutable FillConstantBatchSizeLikeParam param_;
 };
 
-} /* namespace operators */
-} /* namespace lite */
-} /* namespace paddle */
+}  // namespace operators
+}  // namespace lite
+}  // namespace paddle

lite/operators/fill_constant_op.cc

@@ -12,129 +12,69 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "lite/core/op_lite.h"
+#include "lite/operators/fill_constant_op.h"
 #include "lite/core/op_registry.h"
 
 namespace paddle {
 namespace lite {
 namespace operators {
 
-class FillConstantOp : public OpLite {
- public:
-  explicit FillConstantOp(const std::string& type) : OpLite(type) {}
-
-  bool CheckShape() const override {
-    CHECK_OR_FALSE(param_.Out);
-    return true;
-  }
-
-  bool InferShape() const override {
-    lite::Tensor* shape_tensor_ = param_.shape_tensor;
-    if (param_.shape.empty() && shape_tensor_ != nullptr) {
-      param_.Out->Resize(shape_tensor_->dims());
-      return true;
-    }
-
-    param_.Out->Resize(param_.shape);
-    return true;
-  }
-
-  bool AttachImpl(const cpp::OpDesc& opdesc, lite::Scope* scope) override {
-    auto Out_name = opdesc.Output("Out").front();
-
-    param_.Out = GetMutableVar<lite::Tensor>(scope, Out_name);
-    param_.dtype = opdesc.GetAttr<int>("dtype");
-    param_.shape = opdesc.GetAttr<std::vector<int64_t>>("shape");
-    param_.value = opdesc.GetAttr<float>("value");
-    param_.force_cpu = opdesc.GetAttr<bool>("force_cpu");
-    param_.shape_tensor = nullptr;
-    param_.shape_tensor_list = {};
-
-    std::vector<std::string> input_arg_names = opdesc.InputArgumentNames();
-    if (opdesc.HasInput("ShapeTensor") &&
-        !opdesc.Input("ShapeTensor").empty()) {
-      auto args = opdesc.Input("ShapeTensor");
-      auto* var = scope->FindVar(args.front());
-      param_.shape_tensor = var->GetMutable<lite::Tensor>();
+bool FillConstantOp::CheckShape() const {
+  CHECK(param_.out);
+  return true;
+}
+
+bool FillConstantOp::InferShape() const {
+  std::vector<int64_t> out_shape;
+  auto shape_tensor = param_.shape_tensor;
+  auto shape_tensor_list = param_.shape_tensor_list;
+  if (shape_tensor != nullptr) {
+    auto shape_tensor_data = shape_tensor->data<int>();
+    for (int i = 0; i < shape_tensor->numel(); i++) {
+      out_shape.push_back(shape_tensor_data[i]);
     }
-    if (opdesc.HasAttr("ShapeTensorList")) {
-      auto args = opdesc.Input("ShapeTensorList");
-      auto* var = scope->FindVar(args.front());
-      param_.shape_tensor_list =
-          *(var->GetMutable<std::vector<lite::Tensor*>>());
+  } else if (!shape_tensor_list.empty()) {
+    for (int i = 0; i < shape_tensor_list.size(); i++) {
+      out_shape.push_back(shape_tensor_list[i]->data<int>()[0]);
     }
-    return true;
+  } else if (!param_.shape.empty()) {
+    out_shape = param_.shape;
+  } else {
+    LOG(FATAL) << "no valid out_shape. Must set one of shape_tensor, or "
+                  "shape_tensor_list, or shape.";
   }
 
-  void AttachKernel(KernelBase* kernel) override { kernel->SetParam(param_); }
-
-  std::string DebugString() const override { return "fill_constant"; }
-
- private:
-  mutable operators::FillConstantParam param_;
-};
+  param_.out->Resize(out_shape);
+  return true;
+}
 
-class FillConstantBatchLikeOp : public OpLite {
- public:
-  explicit FillConstantBatchLikeOp(const std::string& type) : OpLite(type) {}
+bool FillConstantOp::AttachImpl(const cpp::OpDesc& opdesc, lite::Scope* scope) {
+  auto out_name = opdesc.Output("Out").front();
 
-  bool CheckShape() const override {
-    CHECK_OR_FALSE(param_.out);
-    CHECK_OR_FALSE(param_.input);
-    CHECK_GT_OR_FALSE(param_.shape.size(), 0);
-    CHECK_GE_OR_FALSE(param_.input_dim_idx, 0);
-    CHECK_GE_OR_FALSE(param_.output_dim_idx, 0);
-    return true;
+  param_.out = GetMutableVar<lite::Tensor>(scope, out_name);
+  param_.dtype = opdesc.GetAttr<int>("dtype");
+  if (opdesc.HasAttr("shape")) {
+    param_.shape = opdesc.GetAttr<std::vector<int64_t>>("shape");
   }
+  param_.value = opdesc.GetAttr<float>("value");
+  param_.force_cpu = opdesc.GetAttr<bool>("force_cpu");
 
-  bool InferShape() const override {
-    auto output_dim = param_.shape;
-    output_dim[param_.output_dim_idx] =
-        param_.input->dims()[param_.input_dim_idx];
-    param_.out->Resize(output_dim);
-    return true;
+  if (opdesc.HasInput("ShapeTensor") && !opdesc.Input("ShapeTensor").empty()) {
+    auto shape_tensor_name = opdesc.Input("ShapeTensor").front();
+    param_.shape_tensor = GetMutableVar<lite::Tensor>(scope, shape_tensor_name);
   }
-
-  bool AttachImpl(const cpp::OpDesc& opdesc, lite::Scope* scope) override {
-    auto Out_name = opdesc.Output("Out").front();
-    auto In_name = opdesc.Input("Input").front();
-
-    param_.out = GetMutableVar<lite::Tensor>(scope, Out_name);
-    param_.input = GetMutableVar<lite::Tensor>(scope, In_name);
-    param_.dtype = opdesc.GetAttr<int>("dtype");
-    auto shape = opdesc.GetAttr<std::vector<int>>("shape");
-    std::vector<int64_t> outshape;
-    for (auto i : shape) {
-      outshape.push_back(i);
-    }
-    param_.shape = outshape;
-    if (opdesc.HasAttr("value")) {
-      param_.value = opdesc.GetAttr<float>("value");
-    }
-    if (opdesc.HasAttr("input_dim_idx")) {
-      param_.input_dim_idx = opdesc.GetAttr<int>("input_dim_idx");
-    }
-    if (opdesc.HasAttr("output_dim_idx")) {
-      param_.output_dim_idx = opdesc.GetAttr<int>("output_dim_idx");
+  if (opdesc.HasInput("ShapeTensorList") &&
+      !opdesc.Input("ShapeTensorList").empty()) {
+    for (auto shape_tensor_name : opdesc.Input("ShapeTensorList")) {
+      param_.shape_tensor_list.push_back(
+          GetMutableVar<lite::Tensor>(scope, shape_tensor_name));
     }
-
-    return true;
   }
-
-  void AttachKernel(KernelBase* kernel) override { kernel->SetParam(param_); }
-
-  std::string DebugString() const override {
-    return "fill_constant_batch_size_like";
-  }
-
- private:
-  mutable operators::FillConstantBatchLikeParam param_;
-};
+  return true;
+}
 
 }  // namespace operators
 }  // namespace lite
 }  // namespace paddle
 
 REGISTER_LITE_OP(fill_constant, paddle::lite::operators::FillConstantOp);
-REGISTER_LITE_OP(fill_constant_batch_size_like,
-                 paddle::lite::operators::FillConstantBatchLikeOp);

lite/operators/fill_constant_op.h

+// Copyright (c) 2019 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 <string>
+#include <vector>
+#include "lite/core/op_lite.h"
+#include "lite/core/scope.h"
+#include "lite/utils/all.h"
+
+namespace paddle {
+namespace lite {
+namespace operators {
+
+class FillConstantOp : public OpLite {
+ public:
+  FillConstantOp() {}
+
+  explicit FillConstantOp(const std::string &op_type) : OpLite(op_type) {}
+
+  bool CheckShape() const override;
+
+  bool InferShape() const override;
+
+  bool AttachImpl(const cpp::OpDesc &opdesc, lite::Scope *scope) override;
+
+  void AttachKernel(KernelBase *kernel) override { kernel->SetParam(param_); }
+  std::string DebugString() const override { return "fill_constant"; }
+
+ private:
+  mutable FillConstantParam param_;
+};
+
+}  // namespace operators
+}  // namespace lite
+}  // namespace paddle

lite/operators/op_params.h

@@ -418,35 +418,26 @@ struct MeanGradParam {
 struct FillConstantParam {
   int dtype{static_cast<int>(VarDescAPI::VarDataType::FP32)};
   std::vector<int64_t> shape{};
-  lite::Tensor* shape_tensor;
+  lite::Tensor* shape_tensor{nullptr};
   std::vector<lite::Tensor*> shape_tensor_list{};
 
-  float value{0.0f};
-  // useless for x86, keep it for compatibility
-  bool force_cpu{false};
-  lite::Tensor* Out{};
-};
-struct FillConstantBatchLikeParam {
-  int dtype{static_cast<int>(VarDescAPI::VarDataType::FP32)};
-  std::vector<int64_t> shape{};
   float value{0.0f};
   // useless for x86, keep it for compatibility
   bool force_cpu{false};
   lite::Tensor* out{};
-  const lite::Tensor* input{};
-  int input_dim_idx{0};
-  int output_dim_idx{0};
 };
 
 struct FillConstantBatchSizeLikeParam {
-  lite::Tensor* Input;
-  lite::Tensor* Out;
+  const lite::Tensor* input{nullptr};
+  lite::Tensor* out{nullptr};
 
-  std::vector<int> shape;
+  std::vector<int> shape{};
   int input_dim_idx{0};
   int output_dim_idx{0};
   int dtype{static_cast<int>(VarDescAPI::VarDataType::FP32)};
   float value{0.0f};
+  // useless for x86, keep it for compatibility
+  bool force_cpu{false};
 };
 
 //

lite/tests/kernels/CMakeLists.txt

@@ -34,6 +34,8 @@ if((NOT LITE_WITH_OPENCL AND NOT LITE_WITH_FPGA AND NOT LITE_WITH_BM) AND (LITE_
     lite_cc_test(test_kernel_mul_compute SRCS mul_compute_test.cc DEPS arena_framework ${xpu_kernels} ${npu_kernels} ${x86_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
     lite_cc_test(test_kernel_batch_norm_compute SRCS batch_norm_compute_test.cc DEPS arena_framework ${xpu_kernels} ${npu_kernels} ${x86_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
     lite_cc_test(test_kernel_pool_compute SRCS pool_compute_test.cc DEPS arena_framework ${xpu_kernels} ${npu_kernels} ${x86_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
+    lite_cc_test(test_kernel_fill_constant_compute SRCS fill_constant_compute_test.cc DEPS arena_framework ${xpu_kernels} ${npu_kernels} ${x86_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
+    lite_cc_test(test_kernel_fill_constant_batch_size_like_compute SRCS fill_constant_batch_size_like_compute_test.cc DEPS arena_framework ${xpu_kernels} ${npu_kernels} ${x86_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
 
 if(LITE_BUILD_EXTRA)
     lite_cc_test(test_gru_unit SRCS gru_unit_test.cc DEPS arena_framework ${xpu_kernels} ${npu_kernels} ${x86_kernels} ${bm_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})

lite/tests/kernels/fill_constant_batch_size_like_compute_test.cc

+// Copyright (c) 2019 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 "lite/api/paddle_use_kernels.h"
+#include "lite/api/paddle_use_ops.h"
+#include "lite/core/arena/framework.h"
+#include "lite/tests/utils/fill_data.h"
+
+namespace paddle {
+namespace lite {
+
+class FillConstantBatchSizeLikeComputeTester : public arena::TestCase {
+ protected:
+  // common attributes for this op.
+  std::string input_ = "input";
+  std::string out_ = "out";
+  DDim in_dims_{};
+  LoD in_lod_{};
+  std::vector<int> shape_{};
+  float value_{0.f};
+  int input_dim_idx_{0};
+  int output_dim_idx_{0};
+  int dtype_{static_cast<int>(VarDescAPI::VarDataType::FP32)};
+
+ public:
+  FillConstantBatchSizeLikeComputeTester(
+      const Place& place,
+      const std::string& alias,
+      DDim in_dims,
+      LoD in_lod,
+      std::vector<int> shape,
+      float value = 0.f,
+      int input_dim_idx = 0,
+      int output_dim_idx = 0,
+      int dtype = static_cast<int>(VarDescAPI::VarDataType::FP32))
+      : TestCase(place, alias),
+        in_dims_(in_dims),
+        in_lod_(in_lod),
+        shape_(shape),
+        value_(value),
+        input_dim_idx_(input_dim_idx),
+        output_dim_idx_(output_dim_idx),
+        dtype_(dtype) {}
+
+  void RunBaseline(Scope* scope) override {
+    auto* out = scope->NewTensor(out_);
+    auto* input = scope->FindTensor(input_);
+    std::vector<int64_t> out_shape{shape_.begin(), shape_.end()};
+    if (input_dim_idx_ == 0 && !input->lod().empty()) {
+      out_shape[output_dim_idx_] = input->lod().back().size() - 1;
+    } else {
+      out_shape[output_dim_idx_] = input->dims()[input_dim_idx_];
+    }
+    out->Resize(out_shape);
+
+    auto* output_data = out->mutable_data<float>();
+    for (int i = 0; i < out->numel(); i++) {
+      output_data[i] = value_;
+    }
+  }
+
+  void PrepareOpDesc(cpp::OpDesc* op_desc) {
+    op_desc->SetType("fill_constant_batch_size_like");
+    op_desc->SetInput("Input", {input_});
+    op_desc->SetOutput("Out", {out_});
+    op_desc->SetAttr("shape", shape_);
+    op_desc->SetAttr("value", value_);
+    op_desc->SetAttr("input_dim_idx", input_dim_idx_);
+    op_desc->SetAttr("output_dim_idx", output_dim_idx_);
+    op_desc->SetAttr("dtype", dtype_);
+  }
+
+  void PrepareData() override {
+    std::vector<float> din(in_dims_.production());
+    fill_data_rand(din.data(), -1.f, 1.f, in_dims_.production());
+    SetCommonTensor(input_, in_dims_, din.data(), in_lod_);
+
+    SetPrecisionType(out_, PRECISION(kFloat));
+  }
+};
+
+void TestFillConstantBatchSizeLike(Place place, float abs_error) {
+  for (auto input_dim_idx : {0, 1, 2}) {
+    for (auto output_dim_idx : {0, 1, 2}) {
+      std::unique_ptr<arena::TestCase> tester(
+          new FillConstantBatchSizeLikeComputeTester(place,
+                                                     "def",
+                                                     DDim{{5, 4, 3}},
+                                                     {},
+                                                     {2, 3, 4},
+                                                     0.f,
+                                                     input_dim_idx,
+                                                     output_dim_idx));
+      arena::Arena arena(std::move(tester), place, abs_error);
+      arena.TestPrecision();
+    }
+  }
+}
+
+void TestFillConstantBatchSizeLikeLod(Place place, float abs_error) {
+  for (auto lod : std::vector<LoD>{{{0, 1, 4, 5}}, {{0, 2, 4}, {0, 1, 4, 5}}}) {
+    std::unique_ptr<arena::TestCase> tester(
+        new FillConstantBatchSizeLikeComputeTester(
+            place, "def", DDim{{5, 4, 3}}, lod, {2, 3, 4}, 0.f));
+    arena::Arena arena(std::move(tester), place, abs_error);
+    arena.TestPrecision();
+  }
+}
+
+void TestFillConstantBatchSizeLikeValue(Place place, float abs_error) {
+  std::vector<float> values{-1., 3.5};
+  for (auto value : values) {
+    std::unique_ptr<arena::TestCase> tester(
+        new FillConstantBatchSizeLikeComputeTester(
+            place, "def", DDim{{5, 4, 3}}, {}, {2, 3}, value));
+    arena::Arena arena(std::move(tester), place, abs_error);
+    arena.TestPrecision();
+  }
+}
+
+TEST(fill_constant_batch_size_like, precision) {
+  LOG(INFO) << "test fill_constant_batch_size_like op";
+  Place place;
+  float abs_error = 1e-5;
+#ifdef LITE_WITH_ARM
+  place = TARGET(kARM);
+#else
+  return;
+#endif
+
+  TestFillConstantBatchSizeLike(place, abs_error);
+  TestFillConstantBatchSizeLikeLod(place, abs_error);
+  TestFillConstantBatchSizeLikeValue(place, abs_error);
+}
+
+}  // namespace lite
+}  // namespace paddle

lite/tests/kernels/fill_constant_compute_test.cc

@@ -24,60 +24,56 @@ class FillConstantComputeTester : public arena::TestCase {
  protected:
   // common attributes for this op.
   std::string out_ = "out";
-  int dtype_{static_cast<int>(VarDescAPI::VarDataType::FP32)};
+  std::string shape_tensor_ = "shape_tensor";
+  std::vector<std::string> shape_tensor_list_{};
+
   std::vector<int64_t> shape_{};
-  std::string shape_tensor_ = "ShapeTensor";
-  std::vector<std::string> shape_tensor_list_;
+  float value_{0.0f};
+  int dtype_{static_cast<int>(VarDescAPI::VarDataType::FP32)};
+
   bool is_use_shape_tensor_{false};
   bool is_use_shape_tensor_list_{false};
-
-  float value_{0.0f};
   // useless for x86, keep it for compatibility
   bool force_cpu_{false};
-  // DDim shape_tensor_data{{5, 3}};
-  std::vector<int32_t> shape_tensor_data;
-  DDim shape_test{{1, 2}};
 
  public:
   FillConstantComputeTester(const Place& place,
                             const std::string& alias,
                             std::vector<int64_t> shape,
-                            const bool is_use_shape_tensor,
-                            const bool is_use_shape_tensor_list,
                             float value,
-                            bool force_cpu)
-      : TestCase(place, alias) {
-    shape_ = shape;
-    value_ = value;
-    force_cpu_ = force_cpu;
-    is_use_shape_tensor_ = is_use_shape_tensor;
-    is_use_shape_tensor_list_ = is_use_shape_tensor_list;
-
-    for (int i = 0; i < shape_test.size(); i++) {
-      shape_tensor_data.push_back(i + 1);
+                            int dtype,
+                            const bool is_use_shape_tensor = false,
+                            const bool is_use_shape_tensor_list = false)
+      : TestCase(place, alias),
+        shape_(shape),
+        value_(value),
+        dtype_(dtype),
+        is_use_shape_tensor_(is_use_shape_tensor),
+        is_use_shape_tensor_list_(is_use_shape_tensor_list) {
+    if (is_use_shape_tensor_list) {
+      for (int i = 0; i < shape.size(); i++) {
+        shape_tensor_list_.push_back(shape_tensor_ + std::to_string(i));
+      }
     }
   }
 
   void RunBaseline(Scope* scope) override {
     auto* out = scope->NewTensor(out_);
-    DDim output_dims{shape_};
+    std::vector<int64_t> out_shape;
     if (is_use_shape_tensor_) {
-      auto* temp_shape = scope->FindTensor(shape_tensor_);
-      auto* shape_data = temp_shape->data<int>();
-      auto vec_shape =
-          std::vector<int64_t>(shape_data, shape_data + temp_shape->numel());
-      output_dims.ConstructFrom(vec_shape);
-    }
-    if (is_use_shape_tensor_list_) {
-      std::vector<int64_t> vec_shape;
+      auto* shape_tensor = scope->FindTensor(shape_tensor_);
+      auto* shape_tensor_data = shape_tensor->data<int>();
+      out_shape = std::vector<int64_t>(
+          shape_tensor_data, shape_tensor_data + shape_tensor->numel());
+    } else if (is_use_shape_tensor_list_) {
       for (int i = 0; i < shape_tensor_list_.size(); i++) {
-        auto* temp_shape = scope->FindTensor(shape_tensor_list_[i]);
-        vec_shape.push_back(*temp_shape->data<int>());
+        auto* shape_tensor = scope->FindTensor(shape_tensor_list_[i]);
+        out_shape.push_back(shape_tensor->data<int>()[0]);
       }
-
-      output_dims.ConstructFrom(vec_shape);
+    } else {
+      out_shape = shape_;
     }
-    out->Resize(output_dims);
+    out->Resize(out_shape);
 
     auto* output_data = out->mutable_data<float>();
     for (int i = 0; i < out->numel(); i++) {
@@ -86,92 +82,105 @@ class FillConstantComputeTester : public arena::TestCase {
   }
 
   void PrepareOpDesc(cpp::OpDesc* op_desc) {
-    LOG(INFO) << "PrepareOpDesc";
-
     op_desc->SetType("fill_constant");
-    op_desc->SetAttr("dtype", dtype_);
-    op_desc->SetAttr("shape", shape_);
-    op_desc->SetAttr("value", value_);
-    op_desc->SetAttr("force_cpu", force_cpu_);
     if (is_use_shape_tensor_) {
       op_desc->SetInput("ShapeTensor", {shape_tensor_});
-    }
-    if (is_use_shape_tensor_list_) {
-      // std::vector<std::string> shape_tensor_list_;
-      for (int i = 0; i < shape_test.size(); ++i) {
-        shape_tensor_list_.push_back("shape_tensor_list_" + std::to_string(i));
-      }
-      op_desc->SetInput("ShapeTensorList", {shape_tensor_list_});
+    } else if (is_use_shape_tensor_list_) {
+      op_desc->SetInput("ShapeTensorList", shape_tensor_list_);
+    } else {
+      op_desc->SetAttr("shape", shape_);
     }
     op_desc->SetOutput("Out", {out_});
+    op_desc->SetAttr("dtype", dtype_);
+    op_desc->SetAttr("value", value_);
+    op_desc->SetAttr("force_cpu", force_cpu_);
   }
 
   void PrepareData() override {
     if (is_use_shape_tensor_) {
-      // std::vector<int64_t> temp = x_dims_.data();
-      // int64_t* data = temp.data();
-      SetCommonTensor(shape_tensor_, shape_test, shape_tensor_data.data());
+      std::vector<int> dshape_tensor(shape_.begin(), shape_.end());
+      SetCommonTensor(shape_tensor_,
+                      DDim({static_cast<int64_t>(shape_.size())}),
+                      dshape_tensor.data());
     }
     if (is_use_shape_tensor_list_) {
-      Scope& scope_ = this->scope();
-      for (int i = 0; i < shape_test.size(); ++i) {
-        auto* tensor =
-            scope_.NewTensor("shape_tensor_list_" + std::to_string(i));
-        tensor->Resize(DDim({1}));
-        auto* d = tensor->mutable_data<int>();
-        d[0] = shape_tensor_data[i];
+      for (int i = 0; i < shape_.size(); ++i) {
+        std::vector<int> dshape_tensor{static_cast<int>(shape_[i])};
+        SetCommonTensor(shape_tensor_list_[i], DDim({1}), dshape_tensor.data());
       }
     }
+    SetPrecisionType(out_, PRECISION(kFloat));
   }
 };
 
-TEST(fill_constant, precision) {
-  LOG(INFO) << "test fill_constant op, kARM";
-#ifdef LITE_WITH_ARM
-  Place place(TARGET(kARM));
-  std::vector<int64_t> shape{1, 2};
-
-  for (int dtype : {static_cast<int>(VarDescAPI::VarDataType::INT32)}) {
-    for (float value : {1, 2}) {
-      for (bool is_use_shape_tensor_list : {false, true}) {
-        for (bool is_use_shape_tensor : {false, true}) {
-          if (is_use_shape_tensor && is_use_shape_tensor_list) break;
-          LOG(INFO) << "value:" << value
-                    << ", is_use_shape_tensor:" << is_use_shape_tensor
-                    << ", is_use_shape_tensor_list:"
-                    << is_use_shape_tensor_list;
-
-          std::unique_ptr<arena::TestCase> tester(
-              new FillConstantComputeTester(place,
-                                            "def",
-                                            shape,
-                                            is_use_shape_tensor,
-                                            is_use_shape_tensor_list,
-                                            value,
-                                            false));
-          arena::Arena arena(std::move(tester), place, 2e-5);
-          arena.TestPrecision();
-        }
-      }
-    }
+void TestFillConstantShape(Place place, float abs_error) {
+  std::vector<std::vector<int64_t>> out_shapes{
+      {2, 3, 4, 5}, {2, 3, 4}, {3, 4}, {4}};
+  for (auto out_shape : out_shapes) {
+    std::unique_ptr<arena::TestCase> tester(new FillConstantComputeTester(
+        place,
+        "def",
+        out_shape,
+        1.f,
+        static_cast<int>(VarDescAPI::VarDataType::FP32)));
+    arena::Arena arena(std::move(tester), place, abs_error);
+    arena.TestPrecision();
   }
-#endif
+}
 
-#ifdef LITE_WITH_X86
-  Place place(TARGET(kX86));
-  LOG(INFO) << "test concate op, x86";
-  for (int axis : {1, 2}) {
-    for (bool is_use_axis_tensor : {false, true}) {
-      LOG(INFO) << "axis:" << axis
-                << ", is_use_axis_tensor:" << is_use_axis_tensor;
-      std::unique_ptr<arena::TestCase> tester(
-          new ConcateComputeTester(place, "def", axis, is_use_axis_tensor));
-      arena::Arena arena(std::move(tester), place, 2e-5);
-      arena.TestPrecision();
-    }
+void TestFillConstantValue(Place place, float abs_error) {
+  std::vector<float> values{-1., 0., 3.5};
+  for (auto value : values) {
+    std::unique_ptr<arena::TestCase> tester(new FillConstantComputeTester(
+        place,
+        "def",
+        {2, 3},
+        value,
+        static_cast<int>(VarDescAPI::VarDataType::FP32)));
+    arena::Arena arena(std::move(tester), place, abs_error);
+    arena.TestPrecision();
   }
+}
 
+void TestFillConstantShapeTensor(Place place, float abs_error) {
+  std::unique_ptr<arena::TestCase> tester(new FillConstantComputeTester(
+      place,
+      "def",
+      {2, 3, 4},
+      1.f,
+      static_cast<int>(VarDescAPI::VarDataType::FP32),
+      true));
+  arena::Arena arena(std::move(tester), place, abs_error);
+  arena.TestPrecision();
+}
+
+void TestFillConstantShapeTensorList(Place place, float abs_error) {
+  std::unique_ptr<arena::TestCase> tester(new FillConstantComputeTester(
+      place,
+      "def",
+      {2, 3, 4},
+      1.f,
+      static_cast<int>(VarDescAPI::VarDataType::FP32),
+      false,
+      true));
+  arena::Arena arena(std::move(tester), place, abs_error);
+  arena.TestPrecision();
+}
+
+TEST(fill_constant, precision) {
+  LOG(INFO) << "test fill_constant op";
+  Place place;
+  float abs_error = 1e-5;
+#ifdef LITE_WITH_ARM
+  place = TARGET(kARM);
+#else
+  return;
 #endif
+
+  TestFillConstantShape(place, abs_error);
+  TestFillConstantValue(place, abs_error);
+  TestFillConstantShapeTensor(place, abs_error);
+  TestFillConstantShapeTensorList(place, abs_error);
 }
 
 }  // namespace lite