From 71f9b9a3120357c5edd06377f75425fe4b04ccaf Mon Sep 17 00:00:00 2001 From: minqiyang Date: Tue, 8 Jan 2019 16:36:00 +0800 Subject: [PATCH] Accelerate little models test=release/1.2 --- paddle/fluid/framework/CMakeLists.txt | 2 +- .../framework/details/execution_strategy.h | 2 +- .../scope_buffered_ssa_graph_executor.cc | 33 ++---- .../scope_buffered_ssa_graph_executor.h | 24 ++++ paddle/fluid/framework/rw_lock.h | 103 ++++++------------ paddle/fluid/framework/scope.cc | 51 +++++---- paddle/fluid/framework/scope.h | 20 +++- 7 files changed, 121 insertions(+), 114 deletions(-) diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index 9f5631b87cb..303b1c47af1 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -80,7 +80,7 @@ cc_test(variable_test SRCS variable_test.cc) cc_library(threadpool SRCS threadpool.cc DEPS enforce) cc_test(threadpool_test SRCS threadpool_test.cc DEPS threadpool) -cc_library(scope SRCS scope.cc DEPS glog threadpool) +cc_library(scope SRCS scope.cc DEPS glog threadpool xxhash) cc_test(scope_test SRCS scope_test.cc DEPS scope) cc_library(data_device_transform SRCS data_device_transform.cc DEPS tensor) diff --git a/paddle/fluid/framework/details/execution_strategy.h b/paddle/fluid/framework/details/execution_strategy.h index 15c496130c2..37b07e57363 100644 --- a/paddle/fluid/framework/details/execution_strategy.h +++ b/paddle/fluid/framework/details/execution_strategy.h @@ -25,7 +25,7 @@ struct ExecutionStrategy { size_t num_threads_{0}; bool use_cuda_{true}; bool allow_op_delay_{false}; - size_t num_iteration_per_drop_scope_{100}; + size_t num_iteration_per_drop_scope_{1}; ExecutorType type_{kDefault}; bool dry_run_{false}; }; diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc index 499246a9856..eb18a6fa6b7 100644 --- a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc @@ -67,35 +67,26 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run( } platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr); - drop_scope_counter_ += 1; + ++drop_scope_counter_; -#ifdef PADDLE_WITH_CUDA - const std::string gc_name = "garbage_collector"; - DeviceGarbageCollectorMap *gc = - Graph().Has(gc_name) ? &(Graph().Get(gc_name)) - : nullptr; -#endif + bool stream_end = false; + if (!fetch_tensors.empty()) { + WaitComputationalStreams(); + stream_end = true; + } - if (!fetch_tensors.empty() || - drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) { - drop_scope_counter_ = 0; - // Wait All computational streams - for (auto p : places_) { - platform::DeviceContextPool::Instance().Get(p)->Wait(); -#ifdef PADDLE_WITH_CUDA - if (gc != nullptr && platform::is_gpu_place(p)) { - auto gpu_place = boost::get(p); - auto &gc_at_place = gc->at(gpu_place.device); - gc_at_place->Wait(); - gc_at_place->Reset(); - } -#endif + if (drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) { + if (!stream_end) { + WaitComputationalStreams(); } + for (auto &scope : local_scopes_) { auto &local_scope = *scope->Var(details::kLocalExecScopeName)->GetMutable(); scope->DeleteScope(local_scope); } + + drop_scope_counter_ = 0; } if (eptr) { std::rethrow_exception(eptr); diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h index 5e87e0bf50b..4131cbb3ea0 100644 --- a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h @@ -47,6 +47,30 @@ class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor { FeedFetchList Run(const std::vector& fetch_tensors) override; + private: + inline void WaitComputationalStreams() { +#ifdef PADDLE_WITH_CUDA + const std::string gc_name = "garbage_collector"; + DeviceGarbageCollectorMap* gc = + Graph().Has(gc_name) + ? &(Graph().Get(gc_name)) + : nullptr; +#endif + + // Wait All computational streams + for (auto p : places_) { + platform::DeviceContextPool::Instance().Get(p)->Wait(); +#ifdef PADDLE_WITH_CUDA + if (gc != nullptr && platform::is_gpu_place(p)) { + auto gpu_place = boost::get(p); + auto& gc_at_place = gc->at(gpu_place.device); + gc_at_place->Wait(); + gc_at_place->Reset(); + } +#endif + } + } + private: size_t drop_scope_counter_{0}; diff --git a/paddle/fluid/framework/rw_lock.h b/paddle/fluid/framework/rw_lock.h index dbf00f3a79f..f8aa87519a2 100644 --- a/paddle/fluid/framework/rw_lock.h +++ b/paddle/fluid/framework/rw_lock.h @@ -16,7 +16,9 @@ limitations under the License. */ #if !defined(_WIN32) #include -#endif // !_WIN32 +#else +#include // NOLINT +#endif // !_WIN32 #include "paddle/fluid/platform/enforce.h" @@ -29,17 +31,17 @@ struct RWLock { ~RWLock() { pthread_rwlock_destroy(&lock_); } - void RDLock() { + inline void RDLock() { PADDLE_ENFORCE_EQ(pthread_rwlock_rdlock(&lock_), 0, "acquire read lock failed"); } - void WRLock() { + inline void WRLock() { PADDLE_ENFORCE_EQ(pthread_rwlock_wrlock(&lock_), 0, "acquire write lock failed"); } - void UNLock() { + inline void UNLock() { PADDLE_ENFORCE_EQ(pthread_rwlock_unlock(&lock_), 0, "unlock failed"); } @@ -51,81 +53,46 @@ struct RWLock { // https://stackoverflow.com/questions/7125250/making-pthread-rwlock-wrlock-recursive // In windows, rw_lock seems like a hack. Use empty object and do nothing. struct RWLock { - void RDLock() {} - void WRLock() {} - void UNLock() {} + // FIXME(minqiyang): use mutex here to do fake lock + inline void RDLock() { mutex_.lock(); } + + inline void WRLock() { mutex_.lock(); } + + inline void UNLock() { mutex_.unlock(); } + + private: + std::mutex mutex_; }; #endif -class RWLockGuard { +class AutoWRLock { public: - enum Status { kUnLock, kWRLock, kRDLock }; - - RWLockGuard(RWLock* rw_lock, Status init_status) - : lock_(rw_lock), status_(Status::kUnLock) { - switch (init_status) { - case Status::kRDLock: { - RDLock(); - break; - } - case Status::kWRLock: { - WRLock(); - break; - } - case Status::kUnLock: { - break; - } - } - } + explicit AutoWRLock(RWLock* rw_lock) : lock_(rw_lock) { Lock(); } - void WRLock() { - switch (status_) { - case Status::kUnLock: { - lock_->WRLock(); - status_ = Status::kWRLock; - break; - } - case Status::kWRLock: { - break; - } - case Status::kRDLock: { - PADDLE_THROW( - "Please unlock read lock first before invoking write lock."); - break; - } - } - } + ~AutoWRLock() { UnLock(); } - void RDLock() { - switch (status_) { - case Status::kUnLock: { - lock_->RDLock(); - status_ = Status::kRDLock; - break; - } - case Status::kRDLock: { - break; - } - case Status::kWRLock: { - PADDLE_THROW( - "Please unlock write lock first before invoking read lock."); - break; - } - } - } + private: + inline void Lock() { lock_->WRLock(); } - void UnLock() { - if (status_ != Status::kUnLock) { - lock_->UNLock(); - status_ = Status::kUnLock; - } - } + inline void UnLock() { lock_->UNLock(); } + + private: + RWLock* lock_; +}; + +class AutoRDLock { + public: + explicit AutoRDLock(RWLock* rw_lock) : lock_(rw_lock) { Lock(); } + + ~AutoRDLock() { UnLock(); } + + private: + inline void Lock() { lock_->RDLock(); } - ~RWLockGuard() { UnLock(); } + inline void UnLock() { lock_->UNLock(); } private: RWLock* lock_; - Status status_; }; } // namespace framework diff --git a/paddle/fluid/framework/scope.cc b/paddle/fluid/framework/scope.cc index 0d261dd7ccc..e92e9beb01a 100644 --- a/paddle/fluid/framework/scope.cc +++ b/paddle/fluid/framework/scope.cc @@ -43,9 +43,15 @@ DEFINE_double( // the mutex will cause serious performance issue. // So the mutex is disabled when `ON_INFER`. #ifdef PADDLE_ON_INFERENCE -#define SCOPE_LOCK_GUARD +#define SCOPE_KIDS_READER_LOCK +#define SCOPE_KIDS_WRITER_LOCK +#define SCOPE_VARS_READER_LOCK +#define SCOPE_VARS_WRITER_LOCK #else -#define SCOPE_LOCK_GUARD std::lock_guard lock(mutex_); +#define SCOPE_KIDS_READER_LOCK AutoRDLock auto_lock(&kids_lock_); +#define SCOPE_KIDS_WRITER_LOCK AutoWRLock auto_lock(&kids_lock_); +#define SCOPE_VARS_READER_LOCK AutoRDLock auto_lock(&vars_lock_); +#define SCOPE_VARS_WRITER_LOCK AutoWRLock auto_lock(&vars_lock_); #endif namespace paddle { @@ -61,64 +67,69 @@ int64_t GetEagerDeletionThreshold() { Scope::~Scope() { DropKids(); } Scope& Scope::NewScope() const { - SCOPE_LOCK_GUARD - kids_.push_back(new Scope(this)); - return *kids_.back(); + Scope* child = new Scope(this); + { + SCOPE_KIDS_WRITER_LOCK + kids_.push_back(child); + } + return *child; } Variable* Scope::Var(const std::string& name) { - SCOPE_LOCK_GUARD + SCOPE_VARS_WRITER_LOCK return VarInternal(name); } Variable* Scope::Var(std::string* name) { - SCOPE_LOCK_GUARD auto new_name = string::Sprintf("%p.%d", this, vars_.size()); if (name != nullptr) { *name = new_name; } + SCOPE_VARS_WRITER_LOCK return VarInternal(new_name); } Variable* Scope::FindVar(const std::string& name) const { - SCOPE_LOCK_GUARD + SCOPE_VARS_READER_LOCK return FindVarInternal(name); } Variable* Scope::FindLocalVar(const std::string& name) const { - SCOPE_LOCK_GUARD + SCOPE_VARS_READER_LOCK return FindVarLocally(name); } const Scope* Scope::FindScope(const Variable* var) const { - SCOPE_LOCK_GUARD + SCOPE_VARS_READER_LOCK return FindScopeInternal(var); } void Scope::DropKids() { - SCOPE_LOCK_GUARD + SCOPE_KIDS_WRITER_LOCK for (Scope* s : kids_) delete s; kids_.clear(); } bool Scope::HasKid(const Scope* scope) const { - SCOPE_LOCK_GUARD + SCOPE_KIDS_READER_LOCK auto it = std::find(this->kids_.begin(), this->kids_.end(), scope); return it != this->kids_.end(); } std::vector Scope::LocalVarNames() const { - SCOPE_LOCK_GUARD std::vector known_vars; - known_vars.reserve(this->vars_.size()); - for (auto& p : vars_) { - known_vars.emplace_back(p.first); + { + SCOPE_VARS_READER_LOCK + known_vars.reserve(this->vars_.size()); + for (auto& p : vars_) { + known_vars.emplace_back(p.first); + } } return known_vars; } void Scope::DeleteScope(Scope* scope) const { - SCOPE_LOCK_GUARD + SCOPE_KIDS_WRITER_LOCK auto it = std::find(this->kids_.begin(), this->kids_.end(), scope); PADDLE_ENFORCE(it != this->kids_.end(), "%p Cannot find %p as kid scope", this, scope); @@ -132,8 +143,8 @@ void Scope::DeleteScope(Scope* scope) const { } void Scope::EraseVars(const std::vector& var_names) { - SCOPE_LOCK_GUARD std::set var_set(var_names.begin(), var_names.end()); + SCOPE_VARS_WRITER_LOCK for (auto it = vars_.begin(); it != vars_.end();) { if (var_set.find(it->first) != var_set.end()) { it = vars_.erase(it); @@ -145,12 +156,12 @@ void Scope::EraseVars(const std::vector& var_names) { void Scope::Rename(const std::string& origin_name, const std::string& new_name) const { - SCOPE_LOCK_GUARD + SCOPE_VARS_WRITER_LOCK RenameInternal(origin_name, new_name); } std::string Scope::Rename(const std::string& origin_name) const { - SCOPE_LOCK_GUARD + SCOPE_VARS_WRITER_LOCK auto new_name = string::Sprintf("%p.%d", this, vars_.size()); RenameInternal(origin_name, new_name); return new_name; diff --git a/paddle/fluid/framework/scope.h b/paddle/fluid/framework/scope.h index 1901ffbe57e..9dc79c7b40c 100644 --- a/paddle/fluid/framework/scope.h +++ b/paddle/fluid/framework/scope.h @@ -14,12 +14,18 @@ limitations under the License. */ #pragma once +extern "C" { +#include +} + #include -#include // NOLINT +#include #include #include +#include #include +#include "paddle/fluid/framework/rw_lock.h" #include "paddle/fluid/framework/variable.h" #include "paddle/fluid/platform/macros.h" @@ -94,7 +100,14 @@ class Scope { std::string Rename(const std::string& origin_name) const; protected: - mutable std::unordered_map> vars_; + struct KeyHasher { + std::size_t operator()(const std::string& key) const { + return XXH32(key.c_str(), key.size(), 1); + } + }; + + mutable std::unordered_map, KeyHasher> + vars_; private: // Call Scope::NewScope for a sub-scope. @@ -123,7 +136,8 @@ class Scope { DISABLE_COPY_AND_ASSIGN(Scope); private: - mutable std::mutex mutex_; + mutable RWLock kids_lock_; + mutable RWLock vars_lock_; }; // Generate some debug string about the inherience structure of scope, quite -- GitLab