sum_op.cc 8.4 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
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif

25 26 27 28 29 30 31 32
namespace paddle {
namespace operators {
using framework::Tensor;

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

33
  void InferShape(framework::InferShapeContext* ctx) const override {
Q
qiaolongfei 已提交
34
    PADDLE_ENFORCE(ctx->HasInputs("X"), "Inputs(X) should not be null");
35

Q
Qiao Longfei 已提交
36 37
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
                   "Output(Out) of SumOp should not be null.");
38 39
    if (ctx->IsRuntime() &&
        ctx->GetOutputsVarType("Out")[0] ==
40
            framework::proto::VarType::LOD_TENSOR_ARRAY) {
41 42
      return;  // skip runtime infershape when is tensor array;
    }
43

44
    auto x_dims = ctx->GetInputsDim("X");
Q
Qiao Longfei 已提交
45
    size_t N = x_dims.size();
46 47 48 49
    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
50

51 52 53 54 55 56 57 58 59 60
    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
61
    }
Q
Qiao Longfei 已提交
62 63
    ctx->SetOutputDim("Out", in_dim);
    ctx->ShareLoD("X", /*->*/ "Out");
64
  }
65 66

 protected:
67
  framework::OpKernelType GetExpectedKernelType(
68 69
      const framework::ExecutionContext& ctx) const override {
    auto x_vars = ctx.MultiInputVar("X");
70 71 72 73 74 75 76 77 78 79 80 81

    framework::LibraryType library{framework::LibraryType::kPlain};
    framework::DataLayout layout{framework::DataLayout::kAnyLayout};

#ifdef PADDLE_WITH_MKLDNN
    if (library == framework::LibraryType::kPlain &&
        platform::CanMKLDNNBeUsed(ctx)) {
      library = framework::LibraryType::kMKLDNN;
      layout = framework::DataLayout::kMKLDNN;
    }
#endif

82
    if (x_vars[0]->IsType<framework::LoDTensor>()) {
83 84 85 86 87 88 89 90 91 92 93 94 95 96 97
      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");

98
      return framework::OpKernelType(
99 100
          static_cast<framework::proto::VarType::Type>(dtype), ctx.GetPlace(),
          layout, library);
101
    } else if (x_vars[0]->IsType<framework::SelectedRows>()) {
102 103 104 105
      for (auto& var : x_vars) {
        auto& value = var->Get<framework::SelectedRows>().value();
        if (value.IsInitialized()) {
          return framework::OpKernelType(framework::ToDataType(value.type()),
106
                                         ctx.device_context(), layout, library);
107 108 109 110
        }
      }
      // if input sparse vars are not initialized, use an default kernel type.
      return framework::OpKernelType(framework::proto::VarType::FP32,
111
                                     ctx.device_context(), layout, library);
112
    } else if (x_vars[0]->IsType<framework::LoDTensorArray>()) {
Y
Yang Yang(Tony) 已提交
113 114 115 116 117
      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()),
118 119
                                           ctx.device_context(), layout,
                                           library);
Y
Yang Yang(Tony) 已提交
120
          }
121 122
        }
      }
Y
Yang Yang(Tony) 已提交
123
      PADDLE_THROW("Cannot find the input data type by all input data");
124 125 126 127
    }
    PADDLE_THROW("Unexpected branch. Input type is %s",
                 x_vars[0]->Type().name());
  }
128 129 130 131
};

class SumOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
132
  void Make() override {
133 134
    AddInput("X", "(vector<Tensor>) The input tensors of sum operator.")
        .AsDuplicable();
135
    AddOutput("Out", "(Tensor) The output tensor of sum operator.").Reuse("X");
136 137 138
    AddAttr<bool>("use_mkldnn",
                  "(bool, default false) Only used in mkldnn kernel")
        .SetDefault(false);
139
    AddComment(R"DOC(
140
Sum operator.
141

142 143
This operators sums the input tensors. All the inputs can carry the
LoD (Level of Details) information. However, the output only shares
144
the LoD information with the first input.
145
)DOC");
146 147 148
  }
};

Q
QI JUN 已提交
149 150
class SumOpVarTypeInference : public framework::VarTypeInference {
 public:
Y
Yu Yang 已提交
151 152
  void operator()(const framework::OpDesc& op_desc,
                  framework::BlockDesc* block) const override {
Q
QI JUN 已提交
153
    auto& inputs = op_desc.Input("X");
154
    auto var_type = framework::proto::VarType::SELECTED_ROWS;
Y
Yang Yang(Tony) 已提交
155 156
    for (auto& name : op_desc.Input("X")) {
      VLOG(10) << name << " "
Y
Yang Yu 已提交
157
               << block->FindRecursiveOrCreateVar(name).GetType();
Y
Yang Yang(Tony) 已提交
158 159
    }

Q
QI JUN 已提交
160 161
    bool any_input_is_lod_tensor = std::any_of(
        inputs.begin(), inputs.end(), [block](const std::string& name) {
Y
Yang Yu 已提交
162
          return block->FindRecursiveOrCreateVar(name).GetType() ==
163
                 framework::proto::VarType::LOD_TENSOR;
Q
QI JUN 已提交
164
        });
165 166

    auto is_tensor_array = [block](const std::string& name) {
Y
Yang Yu 已提交
167
      return block->FindRecursiveOrCreateVar(name).GetType() ==
168
             framework::proto::VarType::LOD_TENSOR_ARRAY;
169 170 171 172 173 174 175 176
    };

    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
177 178 179 180
      if (!all_inputs_are_tensor_array) {
        std::ostringstream os;
        for (auto& each : inputs) {
          os << "    " << each << " type is "
Y
Yang Yu 已提交
181
             << block->FindRecursiveOrCreateVar(each).GetType() << "\n";
Y
Yang Yang(Tony) 已提交
182 183 184 185
        }
        PADDLE_ENFORCE(all_inputs_are_tensor_array,
                       "Not all inputs are tensor array:\n%s", os.str());
      }
186
      var_type = framework::proto::VarType::LOD_TENSOR_ARRAY;
187
    } else if (any_input_is_lod_tensor) {
188
      var_type = framework::proto::VarType::LOD_TENSOR;
Q
QI JUN 已提交
189 190 191
    }

    auto out_var_name = op_desc.Output("Out").front();
Y
Yang Yu 已提交
192
    auto& out_var = block->FindRecursiveOrCreateVar(out_var_name);
Y
Yang Yang(Tony) 已提交
193 194 195
    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
196 197 198
  }
};

199
class SumGradMaker : public framework::GradOpDescMakerBase {
200
 public:
201
  using framework::GradOpDescMakerBase::GradOpDescMakerBase;
202

Y
Yu Yang 已提交
203
  std::vector<std::unique_ptr<framework::OpDesc>> operator()() const override {
204
    auto x_grads = InputGrad("X", false);
Y
Yu Yang 已提交
205
    std::vector<std::unique_ptr<framework::OpDesc>> grad_ops;
206 207 208 209
    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
210
                     auto* grad_op = new framework::OpDesc();
Y
Yu Yang 已提交
211 212 213 214
                     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
215
                     return std::unique_ptr<framework::OpDesc>(grad_op);
216 217
                   });
    return grad_ops;
218 219 220 221 222 223 224
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
225

Q
QI JUN 已提交
226 227
REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker,
                  ops::SumOpVarTypeInference);
228

Q
QI JUN 已提交
229 230 231 232 233
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>);