inplace_op_pass.cc

// Copyright (c) 2018 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/framework/details/inplace_op_pass.h"
#include <algorithm>
#include <deque>
#include <iterator>
#include <stack>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/details/memory_optimize_pass.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/op_info.h"

// NOTE(dzhwinter): inplace means one op output variable reuse the input space.
// By our design, one operator only can read its input(const Variable),
// write its output(non-const Variable). If one operator is inplaced, means
// user have chance to write the space before reading happens.
// Especially when some optimize code writing style is applied.
//
//
// /* wrong case in operator */
// /*In this case, a larger allocation is allocated, input content is lost*/
// const Tensor* in = ctx.Input<Tensor>("In")
// Tensor* out = ctx.Output<Tensor>("Out");
// auto* out_ptr = out->mutable_data<T>(ctx.GetPlace());
// out_ptr[0] = 0;  // input contect is overwrited.

// NOTE(dzhwinter):
// Only for backward compacity and stable. if enable_inplace_whitelist is turn
// on.
// only the ops in whitelist will be use inplace strategy.
// if not, all the op will be inplaced if it registered with InplaceClass
DEFINE_bool(
    enable_inplace_whitelist, false,
    "If this option turns on, only these op in whitelist can be inplaced."
    "If it turns off, all of the running op can be candidate of inplaced op."
    "Such as scale, elementwise_add"
    "By default, it's turned off");

DECLARE_string(memory_optimize_debug);

// clang-format off
const std::string kInplacedOpWhiteList[] = { // NOLINT
    "sigmoid",
    "exp",
    "relu",
    "tanh",
    "sqrt",
    "ceil",
    "floor",
    "reciprocal",
    "relu6",
    "soft_relu",
    "hard_sigmoid",
    "batch_norm",
    "batch_norm_grad",
    "sum",
    "sum_grad",
    "scale",
    "reshape",
    "elementwise_add",
    "elementwise_add_grad",
};
// clang-format on

namespace paddle {
namespace framework {
namespace details {

static inline ir::Node* GetNextCascadeInplacedVar(ir::Node* var) {
  // if next op is inplaced, then return the output var
  // otherwise return nullptr
  PADDLE_ENFORCE(var && var->IsVar() && !var->IsCtrlVar());
  ir::Node* inplaced_var = nullptr;
  for (auto* next_op : var->outputs) {
    for (auto* output : next_op->outputs) {
      if (output->IsVar() && !output->IsCtrlVar() &&
          output->Name() == var->Name()) {
        inplaced_var = output;
      }
    }
  }
  return inplaced_var;
}

static inline ir::Node* GetPrevCascadeInplacedVar(ir::Node* var) {
  PADDLE_ENFORCE(var && var->IsVar() && !var->IsCtrlVar());
  if (var->inputs.empty()) return nullptr;
  auto* prev_op = var->inputs.at(0);
  auto input_it = std::find_if(prev_op->inputs.begin(), prev_op->inputs.end(),
                               [&](ir::Node* node) {
                                 if (node->IsVar() && !node->IsCtrlVar() &&
                                     node->Name() == var->Name()) {
                                   return true;
                                 } else {
                                   return false;
                                 }
                               });
  return input_it == prev_op->inputs.end() ? nullptr : *input_it;
}

InplacePass::InplacePass() : Pass() {
  if (FLAGS_enable_inplace_whitelist) {
    for (auto& s : kInplacedOpWhiteList) {
      whitelist_.emplace(s);
    }
  }
}

void InplacePass::InitSSAGraphNodes() const {
  std::unordered_map<std::string, std::unordered_set<ir::Node*>> all_vars;
  for (auto* op : view_.AllOps()) {
    for (auto* node : op->inputs) {
      if (!node->IsVar() || node->IsCtrlVar()) continue;
      if (all_vars[node->Name()].count(node) == 0) {
        all_vars[node->Name()].emplace(node);
        var_nodes_[node->Name()].emplace_back(node);
      }
    }
    for (auto* node : op->outputs) {
      if (!node->IsVar() || node->IsCtrlVar()) continue;
      if (all_vars[node->Name()].count(node) == 0) {
        all_vars[node->Name()].emplace(node);
        var_nodes_[node->Name()].emplace_back(node);
      }
    }
  }
}

std::unique_ptr<ir::Graph> InplacePass::ApplyImpl(
    std::unique_ptr<ir::Graph> graph) const {
  var_nodes_.clear();
  view_.Build(graph.get());
  InitSSAGraphNodes();

  for (auto* op : view_.AllOps()) {
    if (FLAGS_enable_inplace_whitelist && !whitelist_.count(op->Name()))
      continue;
    TryInplaceOpInputOutput(op, graph.get());
  }
  graph->ResolveHazard(var_nodes_);

  return graph;
}

void InplacePass::InplaceModifyDesc(const std::string& var,
                                    const std::string& cache_var,
                                    const size_t& idx) const {
  for (size_t i = idx; i < view_.AllOps().size(); ++i) {
    ir::Node* op = view_.AllOps()[i];
    PADDLE_ENFORCE(op->IsOp() && op->Op());
    auto* op_desc = op->Op();
    op_desc->RenameInput(var, cache_var);
    op_desc->RenameOutput(var, cache_var);
    if (op_desc->Block()->HasVar(var)) op_desc->Block()->RemoveVar(var);
    op_desc->Flush();
  }
}

const NodeSwapQueue InplacePass::TryInplaceModifyVar(
    const std::string& var, const std::string& cache_var, const size_t& idx,
    ir::Graph* graph) const {
  PADDLE_ENFORCE(var_nodes_[var].size() >= 1 &&
                 var_nodes_[var].at(0)->Var() != nullptr);
  std::unique_ptr<VarDesc> var_desc(new VarDesc(*var_nodes_[var].at(0)->Var()));
  var_desc->SetName(cache_var);

  NodeSwapQueue swap_nodes;

  for (size_t i = idx; i < view_.AllOps().size(); ++i) {
    auto* op = view_.AllOps()[i];

    // redirect the input to the latest version of cache_var
    for (auto* node : op->inputs) {
      if (node->Name() == var) {
        ir::Node* cache_node = graph->CreateVarNode(var_desc.get());

        // swap node to cache_node
        cache_node->outputs.insert(cache_node->outputs.end(),
                                   node->outputs.begin(), node->outputs.end());
        PADDLE_ENFORCE(node->inputs.size() == 1 && node->inputs[0]->IsOp());
        auto* prev_op = node->inputs[0];
        std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node,
                     cache_node);
        cache_node->inputs.emplace_back(prev_op);
        for (auto* next_op : node->outputs) {
          std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
                       cache_node);
        }

        swap_nodes.emplace_back(std::make_pair(node, cache_node));
      }
    }

    // if we need to rename the output,
    // always create a newer version of cache_var
    for (auto* node : op->outputs) {
      if (node->Name() == var) {
        ir::Node* cache_node = graph->CreateVarNode(var_desc.get());
        // swap node to cache node
        cache_node->outputs.insert(cache_node->outputs.end(),
                                   node->outputs.begin(), node->outputs.end());
        cache_node->inputs.emplace_back(op);
        std::replace(op->outputs.begin(), op->outputs.end(), node, cache_node);
        for (auto* next_op : node->outputs) {
          std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
                       cache_node);
        }

        swap_nodes.emplace_back(std::make_pair(node, cache_node));
      }
    }
  }

  return swap_nodes;
}

void InplacePass::CommitModify(const NodeSwapQueue& swap_nodes,
                               ir::Graph* graph) const {
  for (auto& pair : swap_nodes) {
    auto *node = pair.first, *cache_node = pair.second;
    const std::string var = node->Name(), cache_var = cache_node->Name();
    var_nodes_[cache_var].emplace_back(cache_node);
    graph->RemoveNode(node);
    auto& nodes = var_nodes_.at(var);
    // release unused var in graph. Because python side memory optimize
    // may reused the var in same name, so we only clear the var node
    // after current inplaced index.
    nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end());
  }
}

void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
                                 ir::Graph* graph) const {
  for (auto& pair : nodes) {
    auto *node = pair.first, *cache_node = pair.second;
    const std::string var = node->Name(), cache_var = cache_node->Name();
    auto* prev_op = node->inputs[0];
    std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), cache_node,
                 node);
    for (auto* next_op : node->outputs) {
      std::replace(next_op->inputs.begin(), next_op->inputs.end(), cache_node,
                   node);
    }
    graph->RemoveNode(cache_node);
  }
}

void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
                                          ir::Graph* graph) const {
  VLOG(4) << "Try to inplace op " << op->Name();
  PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
                 "op_desc is nullptr");
  // some pre-requirments need to meet if the op want to inplaced.

  auto* op_desc = op->Op();
  auto& infer_inplace =
      OpInfoMap::Instance().Get(op_desc->Type()).infer_inplace_;

  // 1. infer_inplace_ is registered.
  if (!static_cast<bool>(infer_inplace)) return;
  PADDLE_ENFORCE(static_cast<bool>(infer_inplace),
                 "%s's infer_inplace has not been registered", op_desc->Type());

  auto* block = op_desc->Block();
  auto in_to_outs = infer_inplace(*op_desc, block);

  auto& all_ops = view_.AllOps();
  auto cursor = std::find(all_ops.begin(), all_ops.end(), op);
  size_t idx = std::distance(all_ops.begin(), cursor);

  for (auto& pair : in_to_outs) {
    auto& in_var_name = pair.first;
    auto& out_var_name = pair.second;
    auto* in_node = view_.GetNodeByName(in_var_name, op->inputs);
    auto* out_node = view_.GetNodeByName(out_var_name, op->outputs);

    // 2. there is no external pending op on the input node
    if (view_.PendingOpsOnVar(in_node).size() > 1) {
      VLOG(4) << string::Sprintf(
          "Skiped pair %s => %s. %s input has external dependency."
          "inplace such pair will overwrite the memory.",
          out_var_name, in_var_name, op->Name());
      continue;
    }

    // 3. if output has been memory optimize by python(fluid.memory_optmize()).
    // this candidate  can not be inplaced. Will be deprecated in the future.
    if (view_.InSkipSet(out_node->Name())) {
      VLOG(4) << string::Sprintf(
          "Skiped %s => %s reused previous memory block in python memory "
          "optmize,"
          "it inplace may generate a circle",
          out_var_name, in_var_name, op->Name());
      continue;
    }

    // Debug Interface. Which would be skipped by the pass.
    if (out_node->Name() == FLAGS_memory_optimize_debug) {
      VLOG(3) << "Skiped var by force. FLAGS_memory_optimize_debug="
              << out_node->Name();
      continue;
    }

    // NOTE(dzhwinter):
    // two stage commit of inplaced process. if after inplace happens generate a
    // circle,
    // then withdraw the changes. Otherwise, safely add the node.
    auto swap_nodes =
        TryInplaceModifyVar(out_var_name, in_var_name, idx, graph);

    if (!ir::HasCircle(*graph)) {
      VLOG(3) << string::Sprintf("!!! %s,  %s => %s inplaced", op->Name(),
                                 out_var_name, in_var_name);
      InplaceModifyDesc(out_var_name, in_var_name, idx);
      CommitModify(swap_nodes, graph);
    } else {
      VLOG(3) << string::Sprintf(
          "Skiped pair %s => %s, inplace will generate a circle. withdraw %s",
          out_var_name, in_var_name, op->Name());
      WithdrawModify(swap_nodes, graph);
    }
  }
}

ir::Node* GraphView::GetNodeByName(const std::string& name,
                                   const std::vector<ir::Node*>& nodes) const {
  // nodes should be op->inputs/outputs
  // node in same node do have different name.
  std::unordered_set<std::string> nodes_in_op;
  bool has_dup_node =
      std::all_of(nodes.begin(), nodes.end(), [&nodes_in_op](ir::Node* node) {
        if (!node->IsVar() || node->IsCtrlVar() || node->Var() == nullptr) {
          if (nodes_in_op.count(node->Name())) return true;
          nodes_in_op.emplace(node->Name());
        }
        return false;
      });
  PADDLE_ENFORCE(has_dup_node == false, "nodes has same name!");
  ir::Node* node = nullptr;
  for (auto* it : nodes) {
    if (!it->IsVar() || it->IsCtrlVar() || it->Var() == nullptr) continue;
    if (it->Name() == name) {
      node = it;
      break;
    }
  }
  PADDLE_ENFORCE(node != nullptr,
                 string::Sprintf("Not found var %s in nodes!", name));
  return node;
}

std::vector<ir::Node*> GraphView::PendingOpsOnVar(ir::Node* node) {
  // get the pending ops depends on same var node.
  // because node also maybe a inplaced variable, so need to backtrack all the
  // previous inplaced vars.
  std::vector<ir::Node*> pending_ops;
  ir::Node* p = node;
  while (p != nullptr) {
    pending_ops.insert(pending_ops.end(), p->outputs.begin(), p->outputs.end());
    p = GetPrevCascadeInplacedVar(p);
  }
  return pending_ops;
}

void GraphView::Build(ir::Graph* g) {
  // track the var nodes in correct order.
  // Because we insert some new created node. Which may have data race between
  // nodes.
  // resolve data harzards depends on the var nodes in right order.
  ops_ = SortOpLikeDescOrder(*g);

  // 1. track the nodes which reused previous node in Python memory optimize.
  // these node can not be inplaced, otherwise may generate a circle in graph.
  std::unordered_set<std::string> all_vars;
  for (auto& node : g->Nodes()) {
    if (node->IsVar()) continue;
    for (auto& out : node->outputs) {
      if (out->IsCtrlVar() || out->Var() == nullptr) continue;
      if (all_vars.count(out->Name())) {
        dup_nodes_.emplace(out->Name());
      } else {
        all_vars.emplace(out->Name());
      }
    }
  }

  // 2. track the nodes which used by parameter server.
  // these node can not be inplaced, otherwise trainer
  // pserver can not find each other name.
  auto update_skip_set = [&](ir::Node* node) {
    for (auto& in : node->inputs) {
      if (in->IsVar() && in->Var() != nullptr) dup_nodes_.emplace(in->Name());
    }
    for (auto& out : node->outputs) {
      if (out->IsVar() && out->Var() != nullptr)
        dup_nodes_.emplace(out->Name());
    }
  };
  for (auto& node : g->Nodes()) {
    if (!node->IsOp()) continue;
    if (node->Name() == "send") update_skip_set(node);
    if (node->Name() == "recv") update_skip_set(node);
    if (node->Name() == "prefetch") update_skip_set(node);
  }
}

const std::vector<ir::Node*>& GraphView::AllOps() { return ops_; }

bool GraphView::InSkipSet(const std::string& var) const {
  return dup_nodes_.count(var);
}

}  // namespace details
}  // namespace framework
}  // namespace paddle

REGISTER_PASS(inplace_pass, paddle::framework::details::InplacePass);