sum_op.cc 7.8 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6 7 8 9 10 11
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. */

Y
Yi Wang 已提交
12
#include "paddle/fluid/operators/sum_op.h"
13

14 15
#include <algorithm>
#include <string>
16
#include <vector>
17

Y
Yi Wang 已提交
18 19
#include "paddle/fluid/framework/var_type_inference.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
20 21 22 23 24 25 26 27 28

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

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

29
  void InferShape(framework::InferShapeContext* ctx) const override {
Q
qiaolongfei 已提交
30
    PADDLE_ENFORCE(ctx->HasInputs("X"), "Inputs(X) should not be null");
31

Q
Qiao Longfei 已提交
32 33
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
                   "Output(Out) of SumOp should not be null.");
34 35
    if (ctx->IsRuntime() &&
        ctx->GetOutputsVarType("Out")[0] ==
36
            framework::proto::VarType::LOD_TENSOR_ARRAY) {
37 38
      return;  // skip runtime infershape when is tensor array;
    }
39

40
    auto x_dims = ctx->GetInputsDim("X");
Q
Qiao Longfei 已提交
41
    size_t N = x_dims.size();
42 43 44 45
    PADDLE_ENFORCE_GT(N, 0, "Input tensors count should > 0.");
    if (N == 1) {
      VLOG(3) << "Warning: sum have only one input, may waste memory";
    }
Q
qiaolongfei 已提交
46

47 48 49 50 51 52 53 54 55 56
    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
57
    }
Q
Qiao Longfei 已提交
58 59
    ctx->SetOutputDim("Out", in_dim);
    ctx->ShareLoD("X", /*->*/ "Out");
60
  }
61 62

 protected:
63
  framework::OpKernelType GetExpectedKernelType(
64 65 66
      const framework::ExecutionContext& ctx) const override {
    auto x_vars = ctx.MultiInputVar("X");
    if (x_vars[0]->IsType<framework::LoDTensor>()) {
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81
      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");

82
      return framework::OpKernelType(
83 84
          static_cast<framework::proto::VarType::Type>(dtype),
          ctx.device_context());
85
    } else if (x_vars[0]->IsType<framework::SelectedRows>()) {
86 87 88 89 90 91 92 93 94 95
      for (auto& var : x_vars) {
        auto& value = var->Get<framework::SelectedRows>().value();
        if (value.IsInitialized()) {
          return framework::OpKernelType(framework::ToDataType(value.type()),
                                         ctx.device_context());
        }
      }
      // if input sparse vars are not initialized, use an default kernel type.
      return framework::OpKernelType(framework::proto::VarType::FP32,
                                     ctx.device_context());
96
    } else if (x_vars[0]->IsType<framework::LoDTensorArray>()) {
Y
Yang Yang(Tony) 已提交
97 98 99 100 101 102 103
      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());
          }
104 105
        }
      }
Y
Yang Yang(Tony) 已提交
106
      PADDLE_THROW("Cannot find the input data type by all input data");
107 108 109 110
    }
    PADDLE_THROW("Unexpected branch. Input type is %s",
                 x_vars[0]->Type().name());
  }
111 112 113 114
};

class SumOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
115
  void Make() override {
116 117 118
    AddInput("X", "(vector<Tensor>) The input tensors of sum operator.")
        .AsDuplicable();
    AddOutput("Out", "(Tensor) The output tensor of sum operator.");
119
    AddComment(R"DOC(
120
Sum operator.
121

122 123
This operators sums the input tensors. All the inputs can carry the
LoD (Level of Details) information. However, the output only shares
124
the LoD information with the first input.
125
)DOC");
126 127 128
  }
};

Q
QI JUN 已提交
129 130
class SumOpVarTypeInference : public framework::VarTypeInference {
 public:
Y
Yu Yang 已提交
131 132
  void operator()(const framework::OpDesc& op_desc,
                  framework::BlockDesc* block) const override {
Q
QI JUN 已提交
133
    auto& inputs = op_desc.Input("X");
134
    auto var_type = framework::proto::VarType::SELECTED_ROWS;
Q
QI JUN 已提交
135

Y
Yang Yang(Tony) 已提交
136 137
    for (auto& name : op_desc.Input("X")) {
      VLOG(10) << name << " "
Y
Yang Yu 已提交
138
               << block->FindRecursiveOrCreateVar(name).GetType();
Y
Yang Yang(Tony) 已提交
139 140
    }

Q
QI JUN 已提交
141 142
    bool any_input_is_lod_tensor = std::any_of(
        inputs.begin(), inputs.end(), [block](const std::string& name) {
Y
Yang Yu 已提交
143
          return block->FindRecursiveOrCreateVar(name).GetType() ==
144
                 framework::proto::VarType::LOD_TENSOR;
Q
QI JUN 已提交
145
        });
146 147

    auto is_tensor_array = [block](const std::string& name) {
Y
Yang Yu 已提交
148
      return block->FindRecursiveOrCreateVar(name).GetType() ==
149
             framework::proto::VarType::LOD_TENSOR_ARRAY;
150 151 152 153 154 155 156 157
    };

    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
158 159 160 161
      if (!all_inputs_are_tensor_array) {
        std::ostringstream os;
        for (auto& each : inputs) {
          os << "    " << each << " type is "
Y
Yang Yu 已提交
162
             << block->FindRecursiveOrCreateVar(each).GetType() << "\n";
Y
Yang Yang(Tony) 已提交
163 164 165 166
        }
        PADDLE_ENFORCE(all_inputs_are_tensor_array,
                       "Not all inputs are tensor array:\n%s", os.str());
      }
167
      var_type = framework::proto::VarType::LOD_TENSOR_ARRAY;
168
    } else if (any_input_is_lod_tensor) {
169
      var_type = framework::proto::VarType::LOD_TENSOR;
Q
QI JUN 已提交
170 171 172
    }

    auto out_var_name = op_desc.Output("Out").front();
Y
Yang Yu 已提交
173
    auto& out_var = block->FindRecursiveOrCreateVar(out_var_name);
Y
Yang Yang(Tony) 已提交
174 175 176
    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
177 178 179
  }
};

180
class SumGradMaker : public framework::GradOpDescMakerBase {
181
 public:
182
  using framework::GradOpDescMakerBase::GradOpDescMakerBase;
183

Y
Yu Yang 已提交
184
  std::vector<std::unique_ptr<framework::OpDesc>> operator()() const override {
185
    auto x_grads = InputGrad("X", false);
Y
Yu Yang 已提交
186
    std::vector<std::unique_ptr<framework::OpDesc>> grad_ops;
187 188 189 190
    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
191
                     auto* grad_op = new framework::OpDesc();
Y
Yu Yang 已提交
192 193 194 195
                     grad_op->SetType("scale");
                     grad_op->SetInput("X", og);
                     grad_op->SetOutput("Out", {x_grad});
                     grad_op->SetAttr("scale", 1.0f);
Y
Yu Yang 已提交
196
                     return std::unique_ptr<framework::OpDesc>(grad_op);
197 198
                   });
    return grad_ops;
199 200 201 202 203 204 205
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
206

Q
QI JUN 已提交
207 208
REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker,
                  ops::SumOpVarTypeInference);
Q
QI JUN 已提交
209 210 211 212 213
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>);