conditional_block_op_helper.cc 6.8 KB
Newer Older
Z
Zeng Jinle 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
// Copyright (c) 2019 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/controlflow/conditional_block_op_helper.h"
W
wanghuancoder 已提交
16

Z
Zeng Jinle 已提交
17 18 19
#include <string>
#include <unordered_set>
#include <utility>
W
wanghuancoder 已提交
20

Z
Zeng Jinle 已提交
21 22
#include "paddle/fluid/operators/controlflow/op_variant.h"

W
wanghuancoder 已提交
23 24 25 26 27 28
namespace paddle {
namespace framework {
class ProgramDesc;
}  // namespace framework
}  // namespace paddle

Z
Zeng Jinle 已提交
29 30 31 32 33 34 35 36 37 38
namespace paddle {
namespace operators {

static bool IsMatchedConditionalBlockOpAndConditionalBlockGradOp(
    const OpVariant &fwd_op, const OpVariant &bwd_op) {
  return fwd_op.Outputs().at(ConditionalOp::kScope) ==
         bwd_op.Inputs().at(ConditionalOp::kScope);
}

static void FindAllConditionalBlockAndConditionalBlockGradOp(
39 40
    const framework::ProgramDesc &program, std::vector<OpVariant> *fwd_ops,
    std::vector<OpVariant> *bwd_ops) {
41 42 43 44 45 46
  PADDLE_ENFORCE_GE(
      fwd_ops->size(), bwd_ops->size(),
      platform::errors::InvalidArgument(
          "Size of forward ops must be greater or equal to backward ops. The "
          "number of forward ops is %d and the number of backward ops is %d",
          fwd_ops->size(), bwd_ops->size()));
Z
Zeng Jinle 已提交
47

48 49
  for (size_t i = 1; i < program.Size(); ++i) {
    auto &block = program.Block(i);
Z
Zeng Jinle 已提交
50 51 52 53 54 55 56 57 58 59 60 61
    for (size_t j = 0; j < block.OpSize(); ++j) {
      auto *op = block.Op(j);
      if (op->Type() == "conditional_block") {
        fwd_ops->emplace_back(op);
      } else if (op->Type() == "conditional_block_grad") {
        bwd_ops->emplace_back(op);
      }
    }
  }

  PADDLE_ENFORCE_GE(
      fwd_ops->size(), bwd_ops->size(),
62 63 64 65 66
      platform::errors::InvalidArgument(
          "There are more conditional_block_grad ops than "
          "conditional_block ops in the graph or program. The number of "
          "forward ops is %d and the number of backward ops is %d",
          fwd_ops->size(), bwd_ops->size()));
Z
Zeng Jinle 已提交
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
}

static void SetSkipVarsForConditionalBlockOp(OpVariant *fwd_op,
                                             OpVariant *bwd_op) {
  auto *grad_block = bwd_op->Attr<framework::BlockDesc *>("sub_block");
  auto is_skippable_in_fwd = [grad_block](const std::string &var_name) {
    return var_name != framework::kEmptyVarName &&
           !grad_block->HasVar(var_name);
  };

  std::unordered_set<std::string> forward_skip_vars;
  for (auto *op_desc : grad_block->AllOps()) {
    for (auto &in_arg_name : op_desc->InputArgumentNames()) {
      if (is_skippable_in_fwd(in_arg_name)) {
        forward_skip_vars.insert(in_arg_name);
      }
    }

    for (auto &out_arg_name : op_desc->OutputArgumentNames()) {
      if (is_skippable_in_fwd(out_arg_name)) {
        forward_skip_vars.insert(out_arg_name);
      }
    }
  }

  auto &fwd_attrs = const_cast<framework::AttributeMap &>(fwd_op->Attrs());
  std::vector<std::string> skip_vars_vec(forward_skip_vars.begin(),
                                         forward_skip_vars.end());
  VLOG(2) << "Prepare to skip " << skip_vars_vec.size()
          << " var(s): " << string::join_strings(skip_vars_vec, ' ');
  fwd_attrs[ConditionalOp::kSkipEagerDeletionVars] = std::move(skip_vars_vec);
}

static void PrepareSafeEagerDeletionOnConditionalOpAndConditionalGradOpImpl(
101
    const framework::ProgramDesc &program, std::vector<OpVariant> *ifelse_ops,
Z
Zeng Jinle 已提交
102
    std::vector<OpVariant> *ifelse_grad_ops) {
103 104
  FindAllConditionalBlockAndConditionalBlockGradOp(program, ifelse_ops,
                                                   ifelse_grad_ops);
Z
Zeng Jinle 已提交
105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120

  VLOG(2) << "Found conditional_block op num: " << ifelse_ops->size()
          << ", conditional_block_grad op num: " << ifelse_grad_ops->size();

  if (ifelse_grad_ops->empty()) {
    return;
  }

  std::unordered_set<OpVariant, OpVariant::Hasher> ifelse_op_set(
      ifelse_ops->begin(), ifelse_ops->end());

  for (auto &bwd_op : *ifelse_grad_ops) {
    const OpVariant *matched_fwd_op = nullptr;
    for (auto &fwd_op : ifelse_op_set) {
      if (IsMatchedConditionalBlockOpAndConditionalBlockGradOp(fwd_op,
                                                               bwd_op)) {
121 122 123
        PADDLE_ENFORCE_EQ(matched_fwd_op, nullptr,
                          platform::errors::PreconditionNotMet(
                              "Found multiple matched conditional_block ops."));
Z
Zeng Jinle 已提交
124 125 126 127
        matched_fwd_op = &fwd_op;
      }
    }

128 129 130 131
    PADDLE_ENFORCE_NOT_NULL(
        matched_fwd_op,
        platform::errors::PreconditionNotMet(
            "Cannot find matched forward conditional_block op."));
Z
Zeng Jinle 已提交
132 133 134 135 136 137 138 139

    SetSkipVarsForConditionalBlockOp(const_cast<OpVariant *>(matched_fwd_op),
                                     &bwd_op);
    ifelse_op_set.erase(*matched_fwd_op);
  }
}

void PrepareSafeEagerDeletionOnConditionalOpAndConditionalGradOp(
140
    const framework::ProgramDesc &program, int block_id,
Z
Zeng Jinle 已提交
141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
    const std::vector<std::unique_ptr<framework::OperatorBase>> &all_ops) {
  // If block_id is not 0, returns
  // This is because all conditional_block_ops and conditional_block_grad_ops
  // in the whole program would be processed when block_id is 0 (i.e.
  // when Executor::Run() or ParallelExecutor constructs).

  // What's more, all conditional_block_ops and conditional_block_grad_ops
  // must be processed when block_id is zero. If not, conditional_block_op
  // may run first and erase variables used in conditional_block_grad_op,
  // and in this moment, conditional_block_grad_ops may be not constructed yet.
  if (block_id != 0) return;

  std::vector<OpVariant> fwd_ops, bwd_ops;
  for (auto &op : all_ops) {
    if (op->Type() == "conditional_block") {
      fwd_ops.emplace_back(op.get());
    } else if (op->Type() == "conditional_block_grad") {
      bwd_ops.emplace_back(op.get());
    }
  }

162 163
  PrepareSafeEagerDeletionOnConditionalOpAndConditionalGradOpImpl(
      program, &fwd_ops, &bwd_ops);
Z
Zeng Jinle 已提交
164 165 166
}

void PrepareSafeEagerDeletionOnConditionalOpAndConditionalGradOp(
167
    const framework::ProgramDesc &program,
Z
Zeng Jinle 已提交
168 169 170 171 172 173 174 175 176 177 178 179 180
    const std::vector<framework::OperatorBase *> &ifelse_ops,
    const std::vector<framework::OperatorBase *> &ifelse_grad_ops) {
  std::vector<OpVariant> fwd_ops, bwd_ops;
  fwd_ops.reserve(ifelse_ops.size());
  for (auto *op : ifelse_ops) {
    fwd_ops.emplace_back(op);
  }

  bwd_ops.reserve(ifelse_grad_ops.size());
  for (auto *op : ifelse_grad_ops) {
    bwd_ops.emplace_back(op);
  }

181 182
  PrepareSafeEagerDeletionOnConditionalOpAndConditionalGradOpImpl(
      program, &fwd_ops, &bwd_ops);
Z
Zeng Jinle 已提交
183 184 185 186
}

}  // namespace operators
}  // namespace paddle