sendrecvop_utils.cc 7.8 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
G
gongweibao 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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
15
#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
Y
Yi Wang 已提交
16

17
#include <sys/time.h>
Y
Yi Wang 已提交
18 19
#include <thread>  // NOLINT

20 21 22 23 24
#include "google/protobuf/io/coded_stream.h"
#include "google/protobuf/io/zero_copy_stream.h"
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/operators/detail/bytebuffer_stream.h"
#include "paddle/fluid/operators/detail/proto_encoder_helper.h"
25
#include "paddle/fluid/operators/detail/variable_response.h"
G
gongweibao 已提交
26 27 28 29 30

namespace paddle {
namespace operators {
namespace detail {

31 32
void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
                           const platform::DeviceContext& ctx,
Y
Yancey1989 已提交
33 34
                           ::grpc::ByteBuffer* msg,
                           const std::string& out_name) {
35 36 37 38 39 40 41 42
  using VarMsg = sendrecv::VariableMessage;
  // When using GPU, need to free the copied CPU buffer
  // when the ByteBuffer destroies
  // TODO(typhoonzero): add unref here, if we have dependent
  // parallelism execution, need to know when to free the tensor.
  DestroyCallback destroy_callback = [](void* backing) {};

  void* buf = malloc(1024);
Y
Yancey 已提交
43
  void* payload = nullptr;
44
  size_t payload_size;
Y
Yi Wang 已提交
45
  ProtoEncodeHelper e(static_cast<char*>(buf), 1024);
46 47 48 49 50 51 52
  e.WriteString(VarMsg::kVarnameFieldNumber, name);
  if (var->IsType<framework::LoDTensor>()) {
    e.WriteUint64(VarMsg::kTypeFieldNumber, 0);
  } else if (var->IsType<framework::SelectedRows>()) {
    e.WriteUint64(VarMsg::kTypeFieldNumber, 1);
  }

Y
Yancey1989 已提交
53 54 55
  if (!out_name.empty()) {
    e.WriteString(VarMsg::kOutVarnameFieldNumber, out_name);
  }
56 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
  switch (framework::ToVarType(var->Type())) {
    case framework::proto::VarType_Type_LOD_TENSOR: {
      auto tensor = var->Get<framework::LoDTensor>();
      e.WriteUint64(VarMsg::kDataTypeFieldNumber,
                    framework::ToDataType(tensor.type()));
      for (auto& dim : framework::vectorize(tensor.dims())) {
        e.WriteUint64(VarMsg::kDimsFieldNumber, dim);
      }
      auto lod = tensor.lod();  // std::vector<Vector<size_t>>
      if (lod.size() > 0) {
        e.WriteUint64(VarMsg::kLodLevelFieldNumber, lod.size());

        for (auto& each : lod) {
          e.WriteVarlengthBeginning(VarMsg::kLodFieldNumber,
                                    2 +      // tag + varintlength of submessage
                                        1 +  // kLodDataFieldNumber
                                        each.size());
          // auto copied from GPU
          for (auto& d : each) {
            e.WriteUint64(VarMsg::LodData::kLodDataFieldNumber, d);
          }
        }
      }
      if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
        PADDLE_ENFORCE(platform::is_gpu_place(tensor.place()));
        platform::CPUPlace cpu;
        auto& gpu_dev_ctx =
            static_cast<const platform::CUDADeviceContext&>(ctx);
T
typhoonzero 已提交
85
        auto copy_size = tensor.numel() * framework::SizeOfType(tensor.type());
86
        payload = memory::Alloc(cpu, copy_size);
87

88 89 90 91
        memory::Copy(cpu, payload,
                     boost::get<platform::CUDAPlace>(tensor.place()),
                     reinterpret_cast<const void*>(tensor.data<void>()),
                     copy_size, gpu_dev_ctx.stream());
92
        ctx.Wait();
93 94 95 96
        destroy_callback = [](void* backing) {
          platform::CPUPlace cpu;
          memory::Free(cpu, backing);
        };
97

98 99 100 101
#endif
      } else {
        payload = tensor.data<void>();
      }
T
typhoonzero 已提交
102
      payload_size = tensor.numel() * framework::SizeOfType(tensor.type());
103 104 105 106 107 108 109 110 111 112 113
      e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size);
    } break;
    case framework::proto::VarType_Type_SELECTED_ROWS: {
      // TODO(typhoonzero): selectedrows implement should not use unique_ptr
      auto* slr = var->GetMutable<framework::SelectedRows>();
      e.WriteUint64(VarMsg::kDataTypeFieldNumber,
                    framework::ToDataType(slr->value().type()));
      for (auto& dim : framework::vectorize(slr->value().dims())) {
        e.WriteUint64(VarMsg::kDimsFieldNumber, dim);
      }
      e.WriteUint64(VarMsg::kLodLevelFieldNumber, 0);
114
      e.WriteUint64(VarMsg::kSlrHeightFieldNumber, slr->height());
115 116 117 118 119 120
      auto* tensor = slr->mutable_value();
      if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
        platform::CPUPlace cpu;
        auto& gpu_dev_ctx =
            static_cast<const platform::CUDADeviceContext&>(ctx);
T
typhoonzero 已提交
121 122
        auto copy_size =
            tensor->numel() * framework::SizeOfType(tensor->type());
123 124 125 126 127 128 129
        payload = memory::Alloc(cpu, copy_size);
        memory::Copy(cpu, payload,
                     boost::get<platform::CUDAPlace>(tensor->place()),
                     reinterpret_cast<const void*>(tensor->data<void>()),
                     copy_size, gpu_dev_ctx.stream());
        ctx.Wait();
        destroy_callback = [](void* backing) {
130 131
          platform::CPUPlace cpu;
          memory::Free(cpu, backing);
132 133 134 135 136
        };
#endif
      } else {
        payload = slr->mutable_value()->data<void>();
      }
T
typhoonzero 已提交
137
      payload_size = tensor->numel() * framework::SizeOfType(tensor->type());
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
      e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size);
    } break;
    default:
      PADDLE_THROW("Serialize does not support type: %s",
                   typeid(var->Type()).name());
      break;
  }
  // steal reference of tensor data
  ::grpc::Slice slices[4];  // metadata, tensor, rows meta, rows
  int num_slices = 2;       // only SelectedRows have rows buffer
  slices[0] = ::grpc::Slice(e.size());
  memcpy(const_cast<uint8_t*>(slices[0].begin()), e.data(), e.size());
  slices[1] = ::grpc::Slice(
      grpc_slice_new_with_user_data(payload, payload_size, destroy_callback,
                                    static_cast<char*>(payload)),
      ::grpc::Slice::STEAL_REF);

  if (framework::ToVarType(var->Type()) ==
      framework::proto::VarType_Type_SELECTED_ROWS) {
    auto* slr = var->GetMutable<framework::SelectedRows>();

Y
Yi Wang 已提交
159
    ProtoEncodeHelper e2(static_cast<char*>(buf), 128);
160 161
    // NOTE: rows is of type int64_t
    size_t rows_memory_size =
T
typhoonzero 已提交
162
        slr->rows().size() * framework::SizeOfType(typeid(int64_t));
163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186
    e2.WriteVarlengthBeginning(VarMsg::kRowsFieldNumber, rows_memory_size);
    slices[2] = ::grpc::Slice(e2.size());
    memcpy(const_cast<uint8_t*>(slices[2].begin()), e2.data(), e2.size());

    slices[3] = ::grpc::Slice(
        grpc_slice_new_with_user_data(
            const_cast<void*>(
                reinterpret_cast<const void*>(slr->rows().data())),
            rows_memory_size,
            [](void* backing) {
              // TODO(typhoonzero): add unref here, same as above.
            },
            const_cast<char*>(
                reinterpret_cast<const char*>(slr->rows().data()))),
        ::grpc::Slice::STEAL_REF);
    num_slices = 4;
  }

  ::grpc::ByteBuffer tmp(&slices[0], num_slices);
  msg->Swap(&tmp);
}

void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg,
                               const platform::DeviceContext& ctx,
187
                               const framework::Scope* scope,
Y
Yi Wang 已提交
188
                               framework::Variable** var) {
189
  operators::detail::VariableResponse resp(false, scope, &ctx);
190
  PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!");
Y
Yi Wang 已提交
191
  *var = resp.GetVar();
192 193
}

G
gongweibao 已提交
194 195
}  // namespace detail
}  // namespace operators
Y
Yancey 已提交
196
}  // namespace paddle