runtime_graph.cc 7.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 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 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221
// Copyright (c) 2021 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/distributed/fleet_executor/runtime_graph.h"
#include "paddle/fluid/distributed/fleet_executor/task_node.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"

namespace paddle {
namespace distributed {
namespace {

using OperatorBase = RuntimeGraph::OperatorBase;
using OpRole = paddle::framework::OpRole;
using OpRegistry = paddle::framework::OpRegistry;
using ProgramDesc = paddle::framework::ProgramDesc;

bool IsForward(int64_t op_role) {
  return (op_role == static_cast<int64_t>(OpRole::kForward)) ||
         (op_role == (static_cast<int64_t>(OpRole::kForward) |
                      static_cast<int64_t>(OpRole::kLoss)));
}

bool IsLRSched(int64_t op_role) {
  return op_role == static_cast<int64_t>(OpRole::kLRSched);
}

bool IsBackward(int64_t op_role) {
  return (op_role == static_cast<int64_t>(OpRole::kBackward)) ||
         (op_role == (static_cast<int64_t>(OpRole::kBackward) |
                      static_cast<int64_t>(OpRole::kLoss)));
}

bool IsOptimize(int64_t op_role) {
  return op_role == static_cast<int64_t>(OpRole::kOptimize);
}

struct DistCoord {
  int32_t dp_idx;
  int32_t pp_idx;
  int32_t mp_idx;
};

class DistCoordSys final {
 public:
  DistCoordSys(int32_t dp_degree, int32_t pp_degree, int32_t mp_degree)
      : dp_degree_(dp_degree), pp_degree_(pp_degree), mp_degree_(mp_degree) {}
  DistCoord RankToCoord(int64_t rank) const;
  int64_t CoordToRank(const DistCoord& coord) const;

 private:
  DISABLE_COPY_AND_ASSIGN(DistCoordSys);
  bool InvalidCoord(const DistCoord& coord) const;
  int32_t dp_degree_;
  int32_t pp_degree_;
  int32_t mp_degree_;
};

DistCoord DistCoordSys::RankToCoord(int64_t rank) const {
  DistCoord coord;
  coord.mp_idx = rank % mp_degree_;
  rank /= mp_degree_;
  coord.pp_idx = rank % pp_degree_;
  rank /= pp_degree_;
  coord.dp_idx = rank % dp_degree_;
  return coord;
}

int64_t DistCoordSys::CoordToRank(const DistCoord& coord) const {
  if (InvalidCoord(coord)) {
    return -1;
  }
  return coord.dp_idx * pp_degree_ * mp_degree_ + coord.pp_idx * mp_degree_ +
         coord.mp_idx;
}

bool DistCoordSys::InvalidCoord(const DistCoord& coord) const {
  return coord.mp_idx < 0 || coord.mp_idx >= mp_degree_ || coord.pp_idx < 0 ||
         coord.pp_idx >= pp_degree_ || coord.dp_idx < 0 ||
         coord.dp_idx >= dp_degree_;
}

}  // namespace

std::vector<OpRole> RuntimeGraph::functionality_order = {
    OpRole::kLRSched, OpRole::kForward, OpRole::kBackward, OpRole::kOptimize};

RuntimeGraph::RuntimeGraph(const ProgramDesc& program,
                           const FleetExecutorDesc& exe_desc)
    : exe_desc_(exe_desc) {
  if (exe_desc.grain() == "coarse") {
    SplitProgramBasedFunctionality(program);
    AssignTaskToIntercepter();
    FakeDependence();
    FakeRuntimeInfo();
  }
}

void RuntimeGraph::SplitProgramBasedFunctionality(const ProgramDesc& program) {
  for (const auto& op_desc : program.Block(0).AllOps()) {
    ops_.emplace_back(OpRegistry::CreateOp(*op_desc));
  }
  std::unordered_map<int64_t, std::vector<OperatorBase*>> role_to_ops;
  for (const auto& op : ops_) {
    int64_t op_role = op->Attr<int64_t>("op_role");
    OpRole new_op_role;
    if (IsLRSched(op_role)) {
      new_op_role = OpRole::kLRSched;
    } else if (IsForward(op_role)) {
      new_op_role = OpRole::kForward;
    } else if (IsBackward(op_role)) {
      new_op_role = OpRole::kBackward;
    } else if (IsOptimize(op_role)) {
      new_op_role = OpRole::kOptimize;
    } else {
      PADDLE_THROW(platform::errors::PreconditionNotMet(
          "The op %s is None of LRSched, Forward, Backward or Optimize.",
          op->Type()));
    }
    int64_t new_op_role_id = static_cast<int64_t>(new_op_role);
    if (role_to_ops.find(new_op_role_id) == role_to_ops.end()) {
      role_to_ops.insert({new_op_role_id, {}});
    }
    role_to_ops.at(new_op_role_id).emplace_back(op.get());
  }
  int64_t cur_rank = exe_desc_.cur_rank();
  int64_t task_id = cur_rank * functionality_order.size();
  for (std::size_t i = 0; i < functionality_order.size(); ++i) {
    OpRole role = functionality_order[i];
    int64_t role_id = static_cast<int64_t>(role);
    if (role_to_ops.find(role_id) == role_to_ops.end()) {
      task_nodes_.emplace_back(
          TaskNode::CreateEmptyTaskNode(role_id, cur_rank, task_id));
    } else {
      task_nodes_.emplace_back(TaskNode::CreateTaskNode(
          role_id, role_to_ops.at(role_id), cur_rank, task_id));
    }
    ++task_id;
  }
}

void RuntimeGraph::FakeDependence() {
  int64_t cur_rank = exe_desc_.cur_rank();
  DistCoordSys coord_sys(exe_desc_.dp_degree(), exe_desc_.pp_degree(),
                         exe_desc_.mp_degree());
  const auto& coord = coord_sys.RankToCoord(cur_rank);
  DistCoord upstream_coord = coord, downstream_coord = coord;
  upstream_coord.pp_idx -= 1;
  downstream_coord.pp_idx += 1;
  int64_t pp_upstream = coord_sys.CoordToRank(upstream_coord);
  int64_t pp_downstream = coord_sys.CoordToRank(downstream_coord);
  int32_t num_of_functionality = functionality_order.size();
  // lr -> forward -> backward -> optimize
  //         |          |
  // lr -> forward -> backward -> optimize
  for (std::size_t i = 0; i < task_nodes_.size(); ++i) {
    if (i != 0) {
      task_nodes_[i]->AddUpstreamTask(cur_rank * num_of_functionality + i - 1);
    }
    if (i != task_nodes_.size() - 1) {
      task_nodes_[i]->AddDownstreamTask(cur_rank * num_of_functionality + i +
                                        1);
    }
    if (IsForward(task_nodes_[i]->role())) {
      if (pp_upstream != -1) {
        task_nodes_[i]->AddUpstreamTask(pp_upstream * num_of_functionality + i);
      }
      if (pp_downstream != -1) {
        task_nodes_[i]->AddDownstreamTask(pp_downstream * num_of_functionality +
                                          i);
      }
    } else if (IsBackward(task_nodes_[i]->role())) {
      if (pp_downstream != -1) {
        task_nodes_[i]->AddUpstreamTask(pp_downstream * num_of_functionality +
                                        i);
      }
      if (pp_upstream != -1) {
        task_nodes_[i]->AddDownstreamTask(pp_upstream * num_of_functionality +
                                          i);
      }
    }
  }
}

void RuntimeGraph::AssignTaskToIntercepter() {
  for (const auto& task : task_nodes_) {
    int64_t intercepter_id = task->task_id();
    if (intercepter_id_to_node_.find(intercepter_id) !=
        intercepter_id_to_node_.end()) {
      PADDLE_THROW(platform::errors::PreconditionNotMet(
          "Repeated intercepter id: %d", intercepter_id));
    }
    intercepter_id_to_node_.insert({intercepter_id, task.get()});
  }
}

void RuntimeGraph::FakeRuntimeInfo() {
  int64_t nrank = exe_desc_.cluster_info().size();
  int32_t num_of_functionality = functionality_order.size();
  for (int64_t i = 0; i < nrank; ++i) {
    for (int64_t j = 0; j < num_of_functionality; ++j) {
      int64_t intercepter_id = i * num_of_functionality + j;
      intercepter_id_to_rank_.insert({intercepter_id, i});
    }
  }
}

}  // namespace distributed
}  // namespace paddle