add details. test=develop

revert-15661-fix-cpu-broadcast
dzhwinter 6 years ago
parent bec68fa0b3
commit 04e9776aef

@ -21,12 +21,13 @@ function(CheckCompilerCXX11Flag)
if (${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 3.3)
message(FATAL_ERROR "Unsupported Clang version. Clang >= 3.3 required.")
endif()
endif()
endif()
endif()
endfunction()
CheckCompilerCXX11Flag()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64")
# safe_set_flag
#
# Set a compile flag only if compiler is support

@ -54,8 +54,6 @@ cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph grap
cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass)
cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info)
cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
cc_library(memory_early_delete_pass SRCS memory_early_delete_pass.cc DEPS memory_optimize_pass computation_op_handle scale_loss_grad_op_handle rpc_op_handle
all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle)
cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper)
cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass)
@ -67,13 +65,11 @@ cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_he
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass memory_early_delete_pass inplace_op_pass)
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass inplace_op_pass)
if (WITH_GPU)
list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
endif()
cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph)
cc_test(memory_optimize_pass_test SRCS memory_optimize_pass_test.cc memory_optimize_pass.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry pass)
cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry)
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope

@ -206,8 +206,6 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
new std::vector<OpDesc *>(main_program.Block(0).AllOps());
graph->Set<const std::vector<OpDesc *>>(kAllOpDescs,
all_op_descs); // take ownership
graph->Set<GraphNodePool>(kGraphNodePool,
new GraphNodePool); // take ownership
pass->Erase(kAllOpDescs);
pass->SetNotOwned<const std::vector<OpDesc *>>(kAllOpDescs, all_op_descs);

@ -77,9 +77,6 @@ struct BuildStrategy {
bool fuse_relu_depthwise_conv_{false};
bool memory_optimize_{false};
bool memory_early_delete_{false};
// TODO(dzhwinter):
// make enable_inplace, memory_optimize_
// memory_early_delete_ true by default

@ -171,16 +171,15 @@ void InplacePass::InplaceModifyDesc(const std::string& var,
}
}
const SSANodePair InplacePass::TryInplaceModifyVar(const std::string& var,
const std::string& cache_var,
const size_t& idx,
ir::Graph* graph) const {
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);
SSANodePair swap_nodes;
NodeSwapQueue swap_nodes;
for (size_t i = idx; i < view_.AllOps().size(); ++i) {
auto* op = view_.AllOps()[i];
@ -230,7 +229,7 @@ const SSANodePair InplacePass::TryInplaceModifyVar(const std::string& var,
return swap_nodes;
}
void InplacePass::CommitModify(const SSANodePair& 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;
@ -245,7 +244,7 @@ void InplacePass::CommitModify(const SSANodePair& swap_nodes,
}
}
void InplacePass::WithdrawModify(const SSANodePair& nodes,
void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
ir::Graph* graph) const {
for (auto& pair : nodes) {
auto *node = pair.first, *cache_node = pair.second;

@ -56,7 +56,8 @@ class GraphView {
std::map<ir::Node*, std::unordered_set<ir::Node*>> adj_list_;
};
typedef std::vector<std::pair<ir::Node*, ir::Node*>> SSANodePair;
// swap pairs in sequence
typedef std::vector<std::pair<ir::Node*, ir::Node*>> NodeSwapQueue;
class InplacePass : public ir::Pass {
public:
InplacePass();
@ -68,14 +69,14 @@ class InplacePass : public ir::Pass {
void InitSSAGraphNodes() const;
private:
const SSANodePair TryInplaceModifyVar(const std::string& var,
const std::string& cache_var,
const size_t& idx,
ir::Graph* graph) const;
const NodeSwapQueue TryInplaceModifyVar(const std::string& var,
const std::string& cache_var,
const size_t& idx,
ir::Graph* graph) const;
void CommitModify(const SSANodePair&, ir::Graph* graph) const;
void CommitModify(const NodeSwapQueue&, ir::Graph* graph) const;
void WithdrawModify(const SSANodePair& nodes, ir::Graph* graph) const;
void WithdrawModify(const NodeSwapQueue& nodes, ir::Graph* graph) const;
void InplaceModifyDesc(const std::string& in_var, const std::string& out_var,
const size_t& idx) const;

@ -1,117 +0,0 @@
// 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/memory_early_delete_pass.h"
#include <queue>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
static ComputationOpHandle* FindNextComputationOpHandle(VarHandle* var_in) {
std::queue<VarHandleBase*> queue;
queue.push(var_in);
do {
auto* var = queue.front();
queue.pop();
for (auto* op : var->PendingOps()) {
auto* compute_op = dynamic_cast<ComputationOpHandle*>(op);
if (compute_op != nullptr && compute_op->GetPlace() == var_in->place()) {
return compute_op;
}
for (auto* out_var : op->Outputs()) {
queue.push(out_var);
}
}
} while (!queue.empty());
return nullptr;
}
std::unique_ptr<ir::Graph> MemoryEarlyDeletePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto& graph_pool = Get<GraphNodePool>(kGraphNodePool);
auto& gcs = Get<GarbageCollectorMap>(kGarbageCollector);
std::unordered_map<std::string, std::unordered_set<OpDesc*>> unlived_vars;
unlived_vars.reserve(graph_pool.size());
for (auto& pair : graph_pool) {
unlived_vars.insert(std::make_pair(pair.first, pair.second));
}
auto compare_and_insert_early_delete_op = [&](
OpHandleBase* op, const std::vector<VarHandleBase*>& vars) {
if (unlived_vars.empty()) return;
// unlived vars can be deleted after the last used op has finished.
auto* compute_op = dynamic_cast<ComputationOpHandle*>(op);
const auto& places = Get<std::vector<platform::Place>>(kAllPlaces);
for (auto& var : vars) {
auto* var_handle = dynamic_cast<VarHandle*>(var);
auto var_name = var->Node()->Name();
auto& var_place = var_handle->place();
if (unlived_vars.count(var_name) == 0) continue;
if (!unlived_vars[var_name].empty()) {
if (compute_op != nullptr &&
unlived_vars[var_name].count(compute_op->Node()->Op()) != 0) {
unlived_vars[var_name].erase(compute_op->Node()->Op());
}
continue;
}
if (var_handle == nullptr || !var_handle->Node()->IsVar() ||
var_handle->Node()->IsCtrlVar())
continue;
// shameless copyed from reference count pass.
if (compute_op == nullptr) {
// use next computation op scope
compute_op = FindNextComputationOpHandle(var_handle);
}
auto* early_delete_node =
graph->CreateEmptyNode("early_delete", ir::Node::Type::kOperation);
GarbageCollector* gc = gcs.at(places[compute_op->GetScopeIdx()]).get();
auto* early_delete_handle = new EarlyDeleteOpHandle(
early_delete_node, compute_op->GetScope(), var_place, {var_name}, gc);
if (compute_op->Outputs().empty()) {
auto* dep_var = new DummyVarHandle(graph->CreateControlDepVar());
compute_op->AddOutput(dep_var);
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
}
early_delete_handle->AddInput(compute_op->Outputs().front());
VLOG(5) << "Add early delete op " << var_name << " to Operator"
<< compute_op->Name();
}
};
auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
for (auto& op : all_ops) {
compare_and_insert_early_delete_op(op, op->Inputs());
compare_and_insert_early_delete_op(op, op->Outputs());
}
return graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(memory_early_delete_pass,
paddle::framework::details::MemoryEarlyDeletePass)
.RequireGraphAttr(paddle::framework::details::kGraphNodePool)
.RequireGraphAttr(paddle::framework::details::kGarbageCollector);

@ -1,32 +0,0 @@
// 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.
#pragma once
#include "paddle/fluid/framework/details/early_delete_op_handle.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
class MemoryEarlyDeletePass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace details
} // namespace framework
} // namespace paddle

File diff suppressed because it is too large Load Diff

@ -17,6 +17,8 @@
#include <iostream>
#include <iterator>
#include <list>
#include <map>
#include <set>
#include <string>
#include <utility>
#include <vector>
@ -27,41 +29,41 @@ namespace paddle {
namespace framework {
namespace details {
constexpr char kFetchedVars[] = "fetched_vars";
constexpr char kGraphNodePool[] = "graph_node_pool";
constexpr char kAllOpDescs[] = "all_op_descs";
// NOTE(dzh): Variable and the operators use the var.
// for early delete pass.
// Because analysis var pass build base on ir::Node, which maybe released
// or modified between passes, so we use OpDesc* to mark ops.
using GraphNodePool = std::vector<
std::pair<std::string /*var node*/, std::unordered_set<OpDesc*> /* ops */>>;
std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph);
// NOTE(dzh): by default, it sort node in ascend order(by node bytes size).
// in fluid, -1 means the batch_size is determined in runtime.
// the node batch_size equal -1 always ranking in the front than the node not.
// NOTE(dzh): A ordered set for node reuse in memory optimize.
// the orderedset sort node in ascend order(by node bytes size).
// in fluid, -1 means the batch_size, which is determined in runtime.
// So the reuse happens between nodes who's batch_size both are -1
// simultaneously or not.
//
// sort rule:
// rule 0 : smaller node ranking in front.
// rule 1 : batch_size equal -1 ranking in the front than the node not.
//
// For example,
// node0[-1, 1] node1[-1, 1, 1], node2[1,1], node3[1,1024], ..
// O(1) insert, delete
class OrderedNodeList {
public:
using NodePair = std::pair<ir::Node*, std::unordered_set<ir::Node*>>;
using Iter = typename std::list<NodePair>::iterator;
using ConstIter = typename std::list<NodePair>::const_iterator;
void Insert(ir::Node* var, ir::Node* op);
class OrderedSet {
public:
// nodes with same name exists in pool.
using NodeVector = std::vector<ir::Node*>;
using Iter = typename std::list<NodeVector>::iterator;
using ConstIter = typename std::list<NodeVector>::const_iterator;
void Insert(ir::Node* var);
void Erase(ir::Node* var);
void Erase(const std::string& var);
bool Has(ir::Node* var) { return mark_table_.count(var->Name()); }
bool Has(const std::string& var) { return mark_table_.count(var); }
ir::Node* NodeMatch(ir::Node* var) const;
bool Has(ir::Node* var) const;
void Clear() {
mark_table_.clear();
nodes_.clear();
}
// find the bestfit shape node block with var.
ir::Node* FindBestFitNode(ir::Node* var) const;
// map store non-const iterator, can not promise const
int GetIndex(ir::Node* var);
int GetNodeIndexInPool(ir::Node* var);
// pool all node to string
std::string ToString() const;
@ -69,18 +71,54 @@ class OrderedNodeList {
Iter end() { return nodes_.end(); }
ConstIter begin() const { return nodes_.begin(); }
ConstIter end() const { return nodes_.end(); }
size_t size() const { return nodes_.size(); }
void Clear() {
mark_table_.clear();
nodes_.clear();
}
size_t size() const { return nodes_.size(); }
private:
// for searching.
std::unordered_map<std::string, Iter> mark_table_;
// node swap pairs. var -> ops dep var
std::list<NodePair> nodes_;
// node pool
std::list<NodeVector> nodes_;
};
class ControlFlowGraph {
public:
ControlFlowGraph() = default;
// IR Graph
explicit ControlFlowGraph(const ir::Graph& graph);
void LiveVariableAnalysis();
void RenameVarInCFGGraph(const std::string& old_node,
const std::string& new_node, int begin_idx);
const std::set<std::string> LiveIn(ir::Node* op) const;
const std::set<std::string> LiveOut(ir::Node* op) const;
const std::set<std::string> Use(ir::Node* op) const;
const std::vector<ir::Node*> Ops() const;
std::vector<ir::Node*>& Ops();
// for ssa-graph nodes
ir::Node* GetNodeByName(const std::string& name, ir::Node* op) const;
private:
void BuildCFGGraph();
void ConnectNodes();
using NodeListMap = std::unordered_map<ir::Node*, std::set<ir::Node*>>;
using VarSetMap = std::map<ir::Node*, std::set<std::string>>;
// successors ops use the output variables.
NodeListMap successors_;
// predecessors ops generated input variables.
NodeListMap predecessors_;
// variables lived before run current op.
VarSetMap live_in_;
// variables lived after run current op.
VarSetMap live_out_;
VarSetMap uses_; // op inputs
VarSetMap defs_; // op outputs
std::vector<ir::Node*> ops_; // op sequence by topology sort
};
// valid a tensor can be reuse or not
@ -93,15 +131,24 @@ bool NodeCanReused(const VarDesc& node);
bool OpHasSubBlock(OpDesc* desc);
// node memory size in bytes
size_t NodeSizeInBytes(ir::Node* n);
size_t NodeSize(ir::Node* n);
// node memory size in bytes
size_t NodeSizeInBytes(const VarDesc&);
size_t NodeSize(const VarDesc&);
std::string DebugString(ir::Node* var);
// NOTE(dzhwinter)
// after node reuse, the replaced node shape is
// different with its VarDesc. So need to find the
// correct VarDesc in Block.
VarDesc* FindVarDescInBlock(ir::Node* n);
static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) {
return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() &&
op1->Outputs() == op2->Outputs();
}
template <typename Container, typename Callback>
class FilterVariableImpl {
public:

File diff suppressed because it is too large Load Diff

@ -32,20 +32,15 @@
namespace paddle {
namespace framework {
namespace details {
constexpr char kAllOpDescs[] = "all_op_descs";
std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph);
class ControlFlowGraph;
class MemoryOptimizePass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
private:
// fill the variable map(var_nodes) by version.
void InitSSAGraphNodes() const;
private:
// update program descs
void RenameVarInGraphDesc(const std::string& var,
const std::string& cache_var, size_t idx) const;
@ -62,7 +57,7 @@ class MemoryOptimizePass : public ir::Pass {
private:
// Reuse Node Pool, Owned.
mutable OrderedNodeList pool_;
mutable OrderedSet pool_;
// controlflow Graph
mutable std::unique_ptr<ControlFlowGraph> cfg_;
// skip set
@ -71,45 +66,6 @@ class MemoryOptimizePass : public ir::Pass {
mutable std::map<std::string, std::vector<ir::Node*>> var_nodes_;
};
class ControlFlowGraph {
public:
ControlFlowGraph() = default;
// For IR Graph in parallelexecutor
explicit ControlFlowGraph(const ir::Graph& graph);
void LiveVariableAnalysis();
void RenameVarInCFGGraph(const std::string& old_node,
const std::string& new_node, int begin_idx);
const std::set<std::string> LiveIn(ir::Node* op) const;
const std::set<std::string> LiveOut(ir::Node* op) const;
const std::set<std::string> Use(ir::Node* op) const;
const std::vector<ir::Node*> Ops() const;
std::vector<ir::Node*>& Ops();
// for ssa-graph nodes
ir::Node* GetNodeFromVarName(const std::string& name, ir::Node* op) const;
private:
void BuildCFGGraph();
void ConnectNodes();
using NodeListMap = std::unordered_map<ir::Node*, std::set<ir::Node*>>;
using VarSetMap = std::map<ir::Node*, std::set<std::string>>;
// successors ops use the output variables.
NodeListMap successors_;
// predecessors ops generated input variables.
NodeListMap predecessors_;
// variables lived before run current op.
VarSetMap live_in_;
// variables lived after run current op.
VarSetMap live_out_;
VarSetMap uses_; // op inputs
VarSetMap defs_; // op outputs
std::vector<ir::Node*> ops_; // op sequence by topology sort
};
} // namespace details
} // namespace framework
} // namespace paddle

File diff suppressed because it is too large Load Diff

@ -17,6 +17,7 @@
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle {

@ -21,8 +21,6 @@ namespace paddle {
namespace framework {
namespace details {
constexpr char kAllOpDescs[] = "all_op_descs";
class SequentialExecutionPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(

@ -69,7 +69,7 @@ class InplaceInToOut : public InplaceOpInference {
bool TryInplaceInputOutput(const VarDesc& in, const VarDesc& out) const {
return in.Name() != out.Name() && details::NodeCanReused(in) &&
details::NodeCanReused(out) &&
details::NodeSizeInBytes(out) <= details::NodeSizeInBytes(in);
details::NodeSize(out) <= details::NodeSize(in);
}
};

@ -171,14 +171,6 @@ std::unique_ptr<ir::Graph> ParallelExecutorPrivate::PrepareGCAndRefCnts(
eager_deletion_pass->SetNotOwned(details::kAllPlaces, &places_);
graph = eager_deletion_pass->Apply(std::move(graph));
VLOG(10) << "EagerDeletionPass Applied";
if (build_strategy_.memory_early_delete_) {
auto early_delete_pass =
ir::PassRegistry::Instance().Get("memory_early_delete_pass");
early_delete_pass->SetNotOwned(details::kGarbageCollector, &gcs_);
graph = early_delete_pass->Apply(std::move(graph));
}
VLOG(10) << "MemoryEarlyDeletePass Applied.";
}
return graph;
@ -288,6 +280,8 @@ ParallelExecutor::ParallelExecutor(
graphs.push_back(std::move(graph));
#endif
auto max_memory_size = GetEagerDeletionThreshold();
VLOG(10) << "Eager Deletion Threshold "
<< static_cast<float>(max_memory_size) / (1 << 30);
if (max_memory_size >= 0) {
for (size_t i = 0; i < graphs.size(); ++i) {
graphs[i] = member_->PrepareGCAndRefCnts(
@ -506,6 +500,5 @@ ParallelExecutor::~ParallelExecutor() {
} // namespace framework
} // namespace paddle
USE_PASS(memory_early_delete_pass);
USE_PASS(reference_count_pass);
USE_PASS(eager_deletion_pass);

@ -22,11 +22,7 @@ limitations under the License. */
#include "paddle/fluid/framework/threadpool.h"
#include "paddle/fluid/string/printf.h"
DEFINE_bool(benchmark, false,
"Doing memory benchmark. It will make deleting scope synchronized, "
"and add some memory usage logs."
"Default cuda is asynchronous device, set to True will"
"force op run in synchronous mode.");
DECLARE_bool(benchmark);
DEFINE_bool(
eager_delete_scope, true,

@ -36,6 +36,7 @@ DEFINE_bool(init_allocated_mem, false,
"that initializing the allocated memory with a small value "
"during unit testing.");
DECLARE_double(fraction_of_gpu_memory_to_use);
DECLARE_bool(benchmark);
namespace paddle {
namespace memory {
@ -198,7 +199,7 @@ void *Alloc<platform::CUDAPlace>(const platform::CUDAPlace &place,
<< string::HumanReadableSize(Used<platform::CUDAPlace>(place));
platform::SetDeviceId(cur_dev);
} else {
if (VLOG_IS_ON(3)) {
if (FLAGS_benchmark) {
allocation::GPUMemMonitor.Add(place.device, size);
}
if (FLAGS_init_allocated_mem) {
@ -216,7 +217,7 @@ void Free<platform::CUDAPlace>(const platform::CUDAPlace &place, void *p,
size_t size) {
#ifdef PADDLE_WITH_CUDA
GetGPUBuddyAllocator(place.device)->Free(p);
if (VLOG_IS_ON(3)) {
if (FLAGS_benchmark) {
allocation::GPUMemMonitor.Minus(place.device, size);
}
#else

@ -14,6 +14,12 @@ limitations under the License. */
#include "paddle/fluid/platform/place.h"
DEFINE_bool(benchmark, false,
"Doing memory benchmark. It will make deleting scope synchronized, "
"and add some memory usage logs."
"Default cuda is asynchronous device, set to True will"
"force op run in synchronous mode.");
namespace paddle {
namespace platform {

@ -1092,10 +1092,6 @@ All parameter, weight, gradient are variables in Paddle.
"is_distribution",
[](const BuildStrategy &self) { return self.is_distribution_; },
[](BuildStrategy &self, bool b) { self.is_distribution_ = b; })
.def_property(
"memory_early_delete",
[](const BuildStrategy &self) { return self.memory_early_delete_; },
[](BuildStrategy &self, bool b) { self.memory_early_delete_ = b; })
.def_property(
"enable_inplace",
[](const BuildStrategy &self) { return self.enable_inplace_; },

@ -148,7 +148,8 @@ class ParallelExecutor(object):
else framework.default_main_program()
# FIXME(dzhwinter): enable_inplace should be after memory_optimize
# if turn on python memory optimize, turn off the inplace_pass.
build_strategy.enable_inplace = False if main._is_mem_optimized else True
if build_strategy.enable_inplace is None:
build_strategy.enable_inplace = False if main._is_mem_optimized else True
scope = scope if scope is not None else executor.global_scope()
if share_vars_from and not isinstance(share_vars_from,

Loading…
Cancel
Save