memcpy_op.cc 6.3 KB
Newer Older
J
JZ-LIANG 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
/* Copyright (c) 2020 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 "paddle/fluid/operators/memcpy_op.h"

#include <string>

namespace paddle {
namespace framework {
class OpDesc;
22 23 24
class InferShapeContext;
template <typename T>
class EmptyGradOpMaker;
J
JZ-LIANG 已提交
25 26 27 28 29 30 31 32 33 34 35
}  // namespace framework
namespace imperative {
class OpBase;
}  // namespace imperative
}  // namespace paddle

namespace paddle {
namespace operators {

class MemcpyOp : public framework::OperatorWithKernel {
 public:
36 37
  MemcpyOp(const std::string &type,
           const framework::VariableNameMap &inputs,
J
JZ-LIANG 已提交
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54
           const framework::VariableNameMap &outputs,
           const framework::AttributeMap &attrs)
      : OperatorWithKernel(type, inputs, outputs, attrs) {}

  void InferShape(framework::InferShapeContext *ctx) const override {
    auto type = ctx->GetInputsVarType("X")[0];
    if (type == framework::proto::VarType::SELECTED_ROWS ||
        type == framework::proto::VarType::LOD_TENSOR) {
      ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
      if (type == framework::proto::VarType::LOD_TENSOR) {
        ctx->ShareLoD("X", /*->*/ "Out");
      }
    }
  }

 protected:
  framework::OpKernelType GetKernelTypeForVar(
55 56
      const std::string &var_name,
      const framework::Tensor &tensor,
J
JZ-LIANG 已提交
57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
      const framework::OpKernelType &expected_kernel_type) const override {
    return framework::OpKernelType(expected_kernel_type.data_type_,
                                   expected_kernel_type.place_,
                                   tensor.layout());
  }

  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext &ctx) const override {
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "X"),
        ctx.device_context());
  }
};

class MemcpyInferVarType : public framework::VarTypeInference {
 public:
  void operator()(framework::InferVarTypeContext *ctx) const override {
    ctx->SyncTypeAndDataType("X", "Out");
  }
};

class MemcpyKernel {
 public:
  void operator()(const framework::ExecutionContext &ctx) const {
    auto *x = ctx.InputVar("X");
    if (x == nullptr) {
      return;
    }
    PADDLE_ENFORCE_EQ(
86 87
        ctx.HasOutput("Out"),
        true,
J
JZ-LIANG 已提交
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105
        platform::errors::NotFound("Output(Out) of memcpy_op is not found."));
    auto *out = ctx.OutputVar("Out");
    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
    auto &dev_ctx = *pool.Get(ctx.GetPlace());
    auto dst_place_type = ctx.Attr<int>("dst_place_type");
    framework::VisitVarType(*x, MemcpyFunctor(out, dev_ctx, dst_place_type));
  }
};

class MemcpyOpProtoMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("X", "(LoDTensor) The input variable ");
    AddOutput("Out",
              "(LoDTensor) The type of output "
              "is the same as input X.");
    AddAttr<int>("dst_place_type",
                 "Determine the dst place of tensor copy. "
106 107 108
                 "By Now it ONLY support CUDAPlace <-> CUDAPinnedPlace or "
                 "NPUPlace <-> CPUPlace. "
                 "Other place type is Unimplemented and will cause ERROR."
J
JZ-LIANG 已提交
109 110 111
                 "0: dst is on CPUPlace. "
                 "1: dst is on CUDAPlace. "
                 "2: dst is on CUDAPinnedPlace. "
112 113
                 "3: dst is on XPUPlace. "
                 "4: dst is on NPUPlace. ");
J
JZ-LIANG 已提交
114 115
    AddComment(R"DOC(
    Memcpy Operator.
116 117
    By now, it ONLY supports the memcopy between CUDAPinnedPlace <-> CUDAPlace or 
    NPUPlace <-> CPUPlace, and used as an internal op by Recompute-Offload.
J
JZ-LIANG 已提交
118 119 120 121 122 123 124 125 126 127 128 129 130 131
    You would have to update it if you want other more capacities.

Out = X,  when type in [LoDTensor]
raise error if the type is not listed above.
)DOC");
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OPERATOR(
132 133 134 135
    memcpy,
    ops::MemcpyOp,
    ops::MemcpyOpProtoMaker,
    ops::MemcpyInferVarType,
J
JZ-LIANG 已提交
136 137 138
    paddle::framework::EmptyGradOpMaker<paddle::framework::OpDesc>,
    paddle::framework::EmptyGradOpMaker<paddle::imperative::OpBase>);

139 140 141 142 143 144 145 146 147 148 149 150
REGISTER_OP_CPU_KERNEL_FUNCTOR(memcpy,
                               float,
                               ops::MemcpyKernel,
                               double,
                               ops::MemcpyKernel,
                               int,
                               ops::MemcpyKernel,
                               int64_t,
                               ops::MemcpyKernel,
                               bool,
                               ops::MemcpyKernel,
                               plat::float16,
J
JZ-LIANG 已提交
151 152
                               ops::MemcpyKernel);

153
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
154 155 156 157 158 159 160 161 162 163 164 165
REGISTER_OP_CUDA_KERNEL_FUNCTOR(memcpy,
                                float,
                                ops::MemcpyKernel,
                                double,
                                ops::MemcpyKernel,
                                int,
                                ops::MemcpyKernel,
                                int64_t,
                                ops::MemcpyKernel,
                                bool,
                                ops::MemcpyKernel,
                                plat::float16,
J
JZ-LIANG 已提交
166 167
                                ops::MemcpyKernel);
#endif
168 169

#ifdef PADDLE_WITH_ASCEND_CL
170 171 172 173 174 175 176 177 178 179 180 181
REGISTER_OP_NPU_KERNEL_FUNCTOR(memcpy,
                               float,
                               ops::MemcpyKernel,
                               double,
                               ops::MemcpyKernel,
                               int,
                               ops::MemcpyKernel,
                               int64_t,
                               ops::MemcpyKernel,
                               bool,
                               ops::MemcpyKernel,
                               plat::float16,
182 183
                               ops::MemcpyKernel);
#endif