sum_op.cc 7.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/operators/sum_op.h"
#include <vector>
Q
QI JUN 已提交
14
#include "paddle/framework/var_type_inference.h"
Y
Yang Yang(Tony) 已提交
15
#include "paddle/operators/detail/safe_ref.h"
16 17 18 19 20 21 22 23 24

namespace paddle {
namespace operators {
using framework::Tensor;

class SumOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

25
  void InferShape(framework::InferShapeContext* ctx) const override {
Q
qiaolongfei 已提交
26
    PADDLE_ENFORCE(ctx->HasInputs("X"), "Inputs(X) should not be null");
27

Q
Qiao Longfei 已提交
28 29
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
                   "Output(Out) of SumOp should not be null.");
30 31 32 33 34
    if (ctx->IsRuntime() &&
        ctx->GetOutputsVarType("Out")[0] ==
            framework::VarDesc::LOD_TENSOR_ARRAY) {
      return;  // skip runtime infershape when is tensor array;
    }
35

36
    auto x_dims = ctx->GetInputsDim("X");
Q
Qiao Longfei 已提交
37
    size_t N = x_dims.size();
Q
qijun 已提交
38
    PADDLE_ENFORCE_GT(N, 1, "Input tensors count should > 1.");
Q
qiaolongfei 已提交
39

40 41 42 43 44 45 46 47 48 49
    framework::DDim in_dim({0});
    for (auto& x_dim : x_dims) {
      if (framework::product(x_dim) == 0) {
        continue;
      }
      if (framework::product(in_dim) == 0) {
        in_dim = x_dim;
      } else {
        PADDLE_ENFORCE_EQ(in_dim, x_dim, "Input tensors must have same shape");
      }
Q
qijun 已提交
50
    }
Q
Qiao Longfei 已提交
51 52
    ctx->SetOutputDim("Out", in_dim);
    ctx->ShareLoD("X", /*->*/ "Out");
53
  }
54 55

 protected:
Y
Yu Yang 已提交
56
  framework::OpKernelType GetKernelType(
57 58 59
      const framework::ExecutionContext& ctx) const override {
    auto x_vars = ctx.MultiInputVar("X");
    if (x_vars[0]->IsType<framework::LoDTensor>()) {
60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76
      int dtype = -1;
      for (auto& x_var : x_vars) {
        auto& lod_tensor = x_var->Get<framework::LoDTensor>();
        if (lod_tensor.numel() == 0) {
          continue;
        }
        if (dtype == -1) {
          dtype = framework::ToDataType(lod_tensor.type());
        } else {
          PADDLE_ENFORCE_EQ(dtype, framework::ToDataType(lod_tensor.type()));
        }
      }
      PADDLE_ENFORCE_NE(dtype, -1,
                        "Sum operator should have at least one tensor");

      return framework::OpKernelType(static_cast<framework::DataType>(dtype),
                                     ctx.device_context());
77
    } else if (x_vars[0]->IsType<framework::SelectedRows>()) {
Y
Yu Yang 已提交
78 79 80 81
      return framework::OpKernelType(
          framework::ToDataType(
              x_vars[0]->Get<framework::SelectedRows>().value().type()),
          ctx.device_context());
82
    } else if (x_vars[0]->IsType<framework::LoDTensorArray>()) {
Y
Yang Yang(Tony) 已提交
83 84 85 86 87 88 89
      for (auto& x_var : x_vars) {
        auto& array = x_var->Get<framework::LoDTensorArray>();
        for (auto& each : array) {
          if (each.numel() != 0) {
            return framework::OpKernelType(framework::ToDataType(each.type()),
                                           ctx.device_context());
          }
90 91
        }
      }
Y
Yang Yang(Tony) 已提交
92
      PADDLE_THROW("Cannot find the input data type by all input data");
93 94 95 96
    }
    PADDLE_THROW("Unexpected branch. Input type is %s",
                 x_vars[0]->Type().name());
  }
97 98 99 100
};

class SumOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Q
Qiao Longfei 已提交
101
  SumOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
102
      : OpProtoAndCheckerMaker(proto, op_checker) {
103 104 105
    AddInput("X", "(vector<Tensor>) The input tensors of sum operator.")
        .AsDuplicable();
    AddOutput("Out", "(Tensor) The output tensor of sum operator.");
106
    AddComment(R"DOC(
107
Sum operator.
108

109 110 111
This operators sums the input tensors. All the inputs can carry the 
LoD (Level of Details) information. However, the output only shares 
the LoD information with the first input.
112
)DOC");
113 114 115
  }
};

Q
QI JUN 已提交
116 117 118 119 120
class SumOpVarTypeInference : public framework::VarTypeInference {
 public:
  void operator()(const framework::OpDescBind& op_desc,
                  framework::BlockDescBind* block) const override {
    auto& inputs = op_desc.Input("X");
121
    auto var_type = framework::VarDesc::SELECTED_ROWS;
Q
QI JUN 已提交
122

Y
Yang Yang(Tony) 已提交
123 124 125 126 127
    for (auto& name : op_desc.Input("X")) {
      VLOG(10) << name << " "
               << block->FindRecursiveOrCreateVar(name)->GetType();
    }

Q
QI JUN 已提交
128 129
    bool any_input_is_lod_tensor = std::any_of(
        inputs.begin(), inputs.end(), [block](const std::string& name) {
Y
Yang Yang(Tony) 已提交
130 131
          return block->FindRecursiveOrCreateVar(name)->GetType() ==
                 framework::VarDesc::LOD_TENSOR;
Q
QI JUN 已提交
132
        });
133 134

    auto is_tensor_array = [block](const std::string& name) {
Y
Yang Yang(Tony) 已提交
135
      return detail::Ref(block->FindRecursiveOrCreateVar(name)).GetType() ==
136 137 138 139 140 141 142 143 144
             framework::VarDesc::LOD_TENSOR_ARRAY;
    };

    bool any_input_is_tensor_array =
        std::any_of(inputs.begin(), inputs.end(), is_tensor_array);
    bool all_inputs_are_tensor_array =
        std::all_of(inputs.begin(), inputs.end(), is_tensor_array);

    if (any_input_is_tensor_array) {
Y
Yang Yang(Tony) 已提交
145 146 147 148 149 150 151 152 153 154
      if (!all_inputs_are_tensor_array) {
        std::ostringstream os;
        for (auto& each : inputs) {
          os << "    " << each << " type is "
             << detail::Ref(block->FindRecursiveOrCreateVar(each)).GetType()
             << "\n";
        }
        PADDLE_ENFORCE(all_inputs_are_tensor_array,
                       "Not all inputs are tensor array:\n%s", os.str());
      }
155 156 157
      var_type = framework::VarDesc::LOD_TENSOR_ARRAY;
    } else if (any_input_is_lod_tensor) {
      var_type = framework::VarDesc::LOD_TENSOR;
Q
QI JUN 已提交
158 159 160
    }

    auto out_var_name = op_desc.Output("Out").front();
Y
Yang Yang(Tony) 已提交
161 162 163 164
    auto& out_var = detail::Ref(block->FindRecursiveOrCreateVar(out_var_name));
    out_var.SetType(var_type);
    auto& in_var = detail::Ref(block->FindVarRecursive(inputs.front()));
    out_var.SetDataType(in_var.GetDataType());
Q
QI JUN 已提交
165 166 167
  }
};

168
class SumGradMaker : public framework::GradOpDescMakerBase {
169
 public:
170
  using framework::GradOpDescMakerBase::GradOpDescMakerBase;
171

Y
Yu Yang 已提交
172 173
  std::vector<std::unique_ptr<framework::OpDescBind>> operator()()
      const override {
174
    auto x_grads = InputGrad("X");
Y
Yu Yang 已提交
175
    std::vector<std::unique_ptr<framework::OpDescBind>> grad_ops;
176 177 178 179
    grad_ops.reserve(x_grads.size());
    auto og = OutputGrad("Out");
    std::transform(x_grads.begin(), x_grads.end(), std::back_inserter(grad_ops),
                   [&og](const std::string& x_grad) {
Y
Yu Yang 已提交
180 181 182 183 184 185
                     auto* grad_op = new framework::OpDescBind();
                     grad_op->SetType("scale");
                     grad_op->SetInput("X", og);
                     grad_op->SetOutput("Out", {x_grad});
                     grad_op->SetAttr("scale", 1.0f);
                     return std::unique_ptr<framework::OpDescBind>(grad_op);
186 187
                   });
    return grad_ops;
188 189 190 191 192 193 194
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
195

Q
QI JUN 已提交
196 197
REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker,
                  ops::SumOpVarTypeInference);
Q
QI JUN 已提交
198 199 200 201 202
REGISTER_OP_CPU_KERNEL(
    sum, ops::SumKernel<paddle::platform::CPUDeviceContext, float>,
    ops::SumKernel<paddle::platform::CPUDeviceContext, double>,
    ops::SumKernel<paddle::platform::CPUDeviceContext, int>,
    ops::SumKernel<paddle::platform::CPUDeviceContext, int64_t>);