Utilize StreamSafeCUDAAllocator to support fast GC in new executor (#37642)

* fix reshape move storage error

* remove needless set type

* alloc tensor by shared storage

* Utilize StreamSafeCUDAAllocator to support fast GC in new executor

* Fix compile error for Windows and ROCm

* Fix compile error for Windows

* Modify UT stream_safe_cuda_alloc_test

* Modify UT stream_safe_cuda_alloc_test

* Rewrite fast GC

* Rewrite fast GC

* Fix compile error for BOOST_GET_CONST

* Fix compile error for BOOST_GET_CONST

* Changes default stream for StreamSafeCUDAAllocator

* Fix a small CI error

* Remove some redundant code

* Fix conflict

* Fix compile error for ROCm

* Fix Windoes CI error

* Fix CI error

* Remove some unnecessary code

* Fix CI error

* Add UT for fast GC

* Fix CI error

* add device-agnostic stream class

* add stream.h

* fix ut

* fix cpu compile

* Use RWLock in GetAllocator

* Fix CI error
Co-authored-by: NChen Weihang <chenweihang@baidu.com>
Co-authored-by: Nzhiqiu <chenqiuliang@baidu.com>

paddle/fluid/framework/new_executor/CMakeLists.txt

 set(INTERPRETERCORE_DEPS op_registry device_context scope framework_proto data_feed_proto heter_service_proto trainer_desc_proto glog 
 lod_rank_table fs shell fleet_wrapper heter_wrapper ps_gpu_wrapper box_wrapper lodtensor_printer feed_fetch_method
-graph_to_program_pass variable_helper timer monitor nan_inf_utils)
+graph_to_program_pass variable_helper timer monitor nan_inf_utils interpretercore_event_garbage_collector)
+
+if(WITH_GPU)
+list(APPEND INTERPRETERCORE_DEPS interpretercore_fast_garbage_collector)
+endif()
 
 add_subdirectory(workqueue)
 
 cc_library(data_transfer SRCS data_transfer.cc DEPS enforce scope glog)
 cc_library(new_executor_defs SRCS new_executor_defs.cc DEPS enforce glog scope)
-cc_library(interpretercore_garbage_collector SRCS interpretercore_garbage_collector.cc DEPS workqueue ${DEVICE_EVENT_LIBS} executor_gc_helper)
+cc_library(interpretercore_garbage_collector SRCS interpretercore_garbage_collector.cc DEPS garbage_collector)
+cc_library(interpretercore_event_garbage_collector SRCS interpretercore_event_garbage_collector.cc DEPS interpretercore_garbage_collector)
 cc_library(interpretercore_util SRCS interpretercore_util.cc DEPS ${INTERPRETERCORE_DEPS} workqueue new_executor_defs data_transfer)
 cc_library(event_manager SRCS event_manager.cc DEPS ${DEVICE_EVENT_LIBS} glog new_executor_defs)
 cc_library(stream_analyzer SRCS stream_analyzer.cc DEPS ${DEVICE_EVENT_LIBS} glog device_context new_executor_defs)
-cc_library(interpretercore SRCS interpretercore.cc DEPS workqueue ${DEVICE_EVENT_LIBS} interpretercore_util interpretercore_garbage_collector stream_analyzer event_manager)
+cc_library(interpretercore SRCS interpretercore.cc DEPS workqueue ${DEVICE_EVENT_LIBS} interpretercore_util interpretercore_event_garbage_collector stream_analyzer event_manager)
 cc_library(standalone_executor SRCS standalone_executor.cc DEPS interpretercore)
 
+if(WITH_GPU OR WITH_ROCM)
+    if(WITH_GPU)
+        nv_library(interpretercore_fast_garbage_collector SRCS interpretercore_fast_garbage_collector.cc DEPS interpretercore_garbage_collector)
+    elseif(WITH_ROCM)
+        hip_library(interpretercore_fast_garbage_collector SRCS interpretercore_fast_garbage_collector.cc DEPS interpretercore_garbage_collector)
+    endif()
+    
+    target_link_libraries(interpretercore interpretercore_fast_garbage_collector)
+endif()
+
+# cc_binary(standalone_executor_test SRCS standalone_executor_test.cc DEPS interpretercore standalone_executor operator op_registry executor ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} profiler)
 # skip win32 since wget is not installed by default on windows machine.
 # skip COVERAGE_CI since the test runs slowly because of instrumentation.
 if (WITH_TESTING AND NOT WIN32 AND NOT WITH_COVERAGE AND NOT "$ENV{CI_SKIP_CPP_TEST}" STREQUAL "ON")

paddle/fluid/framework/new_executor/interpretercore.cc

@@ -16,10 +16,15 @@
 #include <unordered_set>
 #include "paddle/fluid/framework/details/nan_inf_utils.h"
 #include "paddle/fluid/framework/details/share_tensor_buffer_functor.h"
+#include "paddle/fluid/framework/new_executor/interpretercore_event_garbage_collector.h"
 #include "paddle/fluid/framework/new_executor/interpretercore_util.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/platform/profiler.h"
 
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+#include "paddle/fluid/framework/new_executor/interpretercore_fast_garbage_collector.h"
+#endif
+
 PADDLE_DEFINE_EXPORTED_bool(new_executor_use_inplace, true,
                             "Use inplace in new executor");
 PADDLE_DEFINE_EXPORTED_bool(new_executor_use_local_scope, true,
@@ -28,6 +33,8 @@ PADDLE_DEFINE_EXPORTED_bool(new_executor_use_local_scope, true,
 
 DECLARE_bool(check_nan_inf);
 DECLARE_bool(benchmark);
+DECLARE_bool(fast_eager_deletion_mode);
+DECLARE_bool(use_stream_safe_cuda_allocator);
 
 constexpr const char* kExceptionCaught = "ExceptionCaught";
 constexpr const char* kTaskCompletion = "TaskCompletion";
@@ -37,6 +44,10 @@ namespace framework {
 // NOTE(Aurelius84): Need a better strategy to determine it.
 static constexpr size_t kHostNumThreads = 4;
 
+bool IsInterpretercoreFastGCEnabled() {
+  return FLAGS_fast_eager_deletion_mode && FLAGS_use_stream_safe_cuda_allocator;
+}
+
 InterpreterCore::InterpreterCore(const platform::Place& place,
                                  const BlockDesc& block,
                                  VariableScope* global_scope)
@@ -47,7 +58,16 @@ InterpreterCore::InterpreterCore(const platform::Place& place,
   is_build_ = false;
   async_work_queue_.reset(
       new interpreter::AsyncWorkQueue(kHostNumThreads, &main_thread_blocker_));
-  gc_.reset(new InterpreterCoreGarbageCollector());
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+  if (IsInterpretercoreFastGCEnabled()) {
+    gc_ = std::make_unique<InterpreterCoreFastGarbageCollector>();
+  } else {
+    gc_ = std::make_unique<InterpreterCoreEventGarbageCollector>();
+  }
+#else
+  gc_ = std::make_unique<InterpreterCoreEventGarbageCollector>();
+#endif
 
   exception_notifier_ = main_thread_blocker_.RegisterEvent(kExceptionCaught);
   completion_notifier_ = main_thread_blocker_.RegisterEvent(kTaskCompletion);
@@ -513,7 +533,10 @@ void InterpreterCore::RunInstructionAsync(size_t instr_id) {
 
     try {
       RunInstruction(instr_node);
-      // GC infomation
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+      RecordStreamForGC(instr_node);
+#endif
       CheckGC(instr_node);
     } catch (platform::EnforceNotMet& ex) {
       framework::InsertCallStackInfo(op->Type(), op->Attrs(), &ex);
@@ -552,6 +575,99 @@ void InterpreterCore::RunInstructionAsync(size_t instr_id) {
   }
 }
 
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+void InterpreterCore::RecordStreamForGC(const Instruction& instr) {
+  if (!IsInterpretercoreFastGCEnabled() ||
+      instr.KernelType() != OpFuncType::kQueueAsync) {
+    return;
+  }
+
+  gpuStream_t stream = reinterpret_cast<const platform::CUDADeviceContext&>(
+                           instr.DeviceContext())
+                           .stream();
+  auto TensorRecordStream = [&stream](Tensor& tensor) {
+    auto allocation = tensor.Holder();
+    if (allocation == nullptr) {
+      return;
+    }
+
+    const platform::Place& place = allocation->place();
+    if (platform::is_gpu_place(place)) {
+      memory::RecordStream(allocation, stream);
+    } else if (platform::is_cuda_pinned_place(place)) {
+      // TODO(Ruibiao): Here should do something to make sure that the tensor is
+      // not freed until the H2D copies done. However, simplely launch a CUDA
+      // runtime callback to the H2D stream may lead a high performance
+      // overhead. As all the cases we meet in H2D are copies from CPUPlace at
+      // present, we just log a WARNING here. A better design is required.
+      LOG(WARNING) << "Copy data from a CUDAPinned tensor in an asynchronous "
+                      "manner may lead a data inconsistent";
+    } else {
+      // memory copies involve CPUPlace are always synchronous, so just do
+      // nothing here
+    }
+  };
+
+  /* NOTE(Ruibiao)：Cross-stream tensor synchronization is required only when
+   * all the following conditions are satisfied:
+   * 1. The tensor will be GC after running the instruction, i.e., in
+   * instr.GCCheckVars.
+   * 2. The stream which initializes this tensor is different from the stream
+   * which the instruction run in.
+   * 3. The tensor is the instruction's input, cause we assume that instruction
+   * will initialize all output tensors with its running stream.
+   * 4. In the OP function of this instruction, the tensor is an input of a
+   * async CUDA kernel.
+   *
+   * Here we only process the first condition, because:
+   * 1. Since the RecordStream function will directly return when the recored
+   * stream is equal to the owning stream, recording a stream same as which
+   * initialized this tensor has less time overhead. Conversely, it may take
+   * more time if we try to extract those cross-stream input vars from
+   * instr.GCCheckVars.
+   * 2. Now the instruction has no idea of which vars involving async running in
+   * OP function, and thus we can not recognize condition 4. It should be
+   * supported later.
+   */
+  for (int var_id : instr.GCCheckVars()) {
+    VLOG(4) << "GC sync " << global_scope_->GetNameById(var_id) << " "
+            << global_scope_->VarDesc(var_id);
+
+    // persistable var will be ignore while GC
+    if (global_scope_->VarDesc(var_id) &&
+        global_scope_->VarDesc(var_id)->Persistable()) {
+      continue;
+    }
+
+    paddle::framework::Variable* var = global_scope_->Var(var_id);
+    if (var == nullptr) {
+      continue;
+    }
+
+    if (var->IsType<LoDTensor>()) {
+      TensorRecordStream(*(var->GetMutable<LoDTensor>()));
+    } else if (var->IsType<
+                   operators::reader::
+                       OrderedMultiDeviceLoDTensorBlockingQueueHolder>()) {
+      // do nothing
+    } else if (var->IsType<SelectedRows>()) {
+      TensorRecordStream(*(var->GetMutable<SelectedRows>()->mutable_value()));
+    } else if (var->IsType<LoDTensorArray>()) {
+      auto* tensor_arr = var->GetMutable<LoDTensorArray>();
+      for (auto& tensor : *tensor_arr) {
+        TensorRecordStream(tensor);
+      }
+    } else if (var->IsType<std::vector<Scope*>>()) {
+      // do nothing
+    } else {
+      PADDLE_THROW(platform::errors::Unimplemented(
+          "The variable(%s) is not supported in eager deletion.",
+          framework::ToTypeName(var->Type())));
+    }
+  }
+}
+#endif
+
 void InterpreterCore::CheckGC(const Instruction& instr) {
   size_t instr_id = instr.Id();
   auto& var_scope = *global_scope_;
@@ -570,9 +686,22 @@ void InterpreterCore::CheckGC(const Instruction& instr) {
     if (is_ready) {
       VLOG(6) << "Async delete variable with name : "
               << var_scope.GetNameById(var_id);
-      gc_->Add(var_scope.Var(var_id), gc_event_.at(instr_id),
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+      if (IsInterpretercoreFastGCEnabled()) {
+        static_cast<InterpreterCoreFastGarbageCollector*>(gc_.get())->Add(
+            var_scope.Var(var_id));
+
+      } else {
+        static_cast<InterpreterCoreEventGarbageCollector*>(gc_.get())->Add(
+            var_scope.Var(var_id), gc_event_.at(instr_id),
             &instr.DeviceContext());
       }
+#else
+      static_cast<InterpreterCoreEventGarbageCollector*>(gc_.get())->Add(
+          var_scope.Var(var_id), gc_event_.at(instr_id),
+          &instr.DeviceContext());
+#endif
+    }
   }
 }
 

paddle/fluid/framework/new_executor/interpretercore.h

@@ -72,6 +72,10 @@ class InterpreterCore {
                const std::vector<framework::LoDTensor>& feed_tensors,
                bool prepare_feed);
 
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+  void RecordStreamForGC(const Instruction& instr);
+#endif
+
   void CheckGC(const Instruction& instr);
 
   void RunInstructionAsync(size_t instr_id);

paddle/fluid/framework/new_executor/interpretercore_event_garbage_collector.cc

+// 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/framework/new_executor/interpretercore_event_garbage_collector.h"
+
+#if !defined(_WIN32)
+#include <sched.h>
+#else
+#define NOMINMAX
+#include <windows.h>
+#endif  // !_WIN32
+
+namespace paddle {
+namespace framework {
+
+InterpreterCoreEventGarbageCollector::InterpreterCoreEventGarbageCollector() {
+  WorkQueueOptions options(/*num_threads*/ 1, /*allow_spinning*/ true,
+                           /*track_task*/ false);
+  queue_ = CreateSingleThreadedWorkQueue(options);
+}
+
+InterpreterCoreEventGarbageCollector::~InterpreterCoreEventGarbageCollector() {
+  queue_.reset(nullptr);
+}
+
+void InterpreterCoreEventGarbageCollector::Add(
+    Garbage garbage, platform::DeviceEvent& event,
+    const platform::DeviceContext* ctx) {
+  if (!garbage) {
+    return;
+  }
+
+  if (max_memory_size_ <= 1) {
+    Free(garbage, event, ctx);
+  } else {
+    std::unique_ptr<GarbageQueue> pending_delete_garbages;
+    {  // lock guard
+      std::lock_guard<memory::SpinLock> guard(spinlock_);
+      cur_memory_size_ += garbage->size();
+      garbages_->push_back(std::move(garbage));
+
+      if (cur_memory_size_ >= max_memory_size_) {
+        cur_memory_size_ = 0;
+        pending_delete_garbages = std::move(garbages_);
+        garbages_ = std::make_unique<GarbageQueue>();
+      }
+    }
+  }
+}
+
+void InterpreterCoreEventGarbageCollector::Add(
+    Variable* var, platform::DeviceEvent& event,
+    const platform::DeviceContext* ctx) {
+  if (UNLIKELY(max_memory_size_ < 0) || var == nullptr) {
+    return;
+  }
+
+  if (var->IsType<LoDTensor>()) {
+    Add(var->GetMutable<LoDTensor>()->MoveMemoryHolder(), event, ctx);
+  } else if (var->IsType<
+                 operators::reader::
+                     OrderedMultiDeviceLoDTensorBlockingQueueHolder>()) {
+    // TODO(xiongkun03) in old executor, this type of variable is not support
+    // eager deletion. so we just leave it here ?
+  } else if (var->IsType<LoDRankTable>()) {
+    // TODO(xiongkun03) in old executor, this type of variable is not support
+    // eager deletion. so we just leave it here ?
+  } else if (var->IsType<SelectedRows>()) {
+    Add(var->GetMutable<SelectedRows>()->mutable_value()->MoveMemoryHolder(),
+        event, ctx);
+    var->GetMutable<SelectedRows>()->mutable_rows()->clear();
+  } else if (var->IsType<LoDTensorArray>()) {
+    auto* tensor_arr = var->GetMutable<LoDTensorArray>();
+    for (auto& t : *tensor_arr) {
+      Add(t.MoveMemoryHolder(), event, ctx);
+    }
+  } else if (var->IsType<std::vector<Scope*>>()) {
+    // NOTE(@xiongkun03) conditional_op / while_op will create a STEP_SCOPE
+    // refer to executor.cc to see what old garbage collector does.
+    // do nothing, because the sub scope will be deleted by sub-executor.
+  } else {
+    PADDLE_THROW(platform::errors::Unimplemented(
+        "The variable(%s) is not supported in eager deletion.",
+        framework::ToTypeName(var->Type())));
+  }
+}
+
+void InterpreterCoreEventGarbageCollector::Free(
+    GarbageQueue* garbages, platform::DeviceEvent& event,
+    const platform::DeviceContext* ctx) {
+  event.Record(ctx);
+  event.SetFininshed();  // Only for CPU Event
+  queue_->AddTask([ container = garbages, event = &event ]() {
+    while (!event->Query()) {
+#if defined(_WIN32)
+      SleepEx(50, FALSE);
+#else
+      sched_yield();
+#endif
+      continue;
+    }
+    delete container;
+  });
+}
+
+void InterpreterCoreEventGarbageCollector::Free(
+    Garbage& garbage, platform::DeviceEvent& event,
+    const platform::DeviceContext* ctx) {
+  event.Record(ctx);
+  event.SetFininshed();  // Only for CPU Event
+  queue_->AddTask([ container = garbage, event = &event ]() {
+    while (!event->Query()) {
+#if defined(_WIN32)
+      SleepEx(50, FALSE);
+#else
+      sched_yield();
+#endif
+      continue;
+    }
+  });
+}
+
+}  // namespace framework
+}  // namespace paddle
\ No newline at end of file

paddle/fluid/framework/new_executor/interpretercore_event_garbage_collector.h

+// 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.
+#pragma once
+
+#include <queue>
+#include "paddle/fluid/framework/new_executor/interpretercore_garbage_collector.h"
+#include "paddle/fluid/framework/new_executor/workqueue/workqueue.h"
+
+namespace paddle {
+namespace framework {
+
+class InterpreterCoreEventGarbageCollector
+    : public InterpreterCoreGarbageCollector {
+ public:
+  InterpreterCoreEventGarbageCollector();
+  ~InterpreterCoreEventGarbageCollector();
+
+  virtual void Add(Variable* var, platform::DeviceEvent& event,
+                   const platform::DeviceContext* ctx) override;
+
+ private:
+  void Add(Garbage garbage, platform::DeviceEvent& event,
+           const platform::DeviceContext* ctx);
+  void Free(GarbageQueue* garbages, platform::DeviceEvent& event,
+            const platform::DeviceContext* ctx);
+  void Free(Garbage& garbage, platform::DeviceEvent& event,
+            const platform::DeviceContext* ctx);
+
+  std::unique_ptr<WorkQueue> queue_;
+  paddle::memory::SpinLock spinlock_;
+};
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/new_executor/interpretercore_fast_garbage_collector.cc

+// 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/framework/new_executor/interpretercore_fast_garbage_collector.h"
+
+namespace paddle {
+namespace framework {
+
+void InterpreterCoreFastGarbageCollector::Add(Variable* var) {
+  if (UNLIKELY(max_memory_size_ < 0) || var == nullptr) {
+    return;
+  }
+
+  if (var->IsType<LoDTensor>()) {
+    Add(var->GetMutable<LoDTensor>()->MoveMemoryHolder());
+  } else if (var->IsType<
+                 operators::reader::
+                     OrderedMultiDeviceLoDTensorBlockingQueueHolder>()) {
+    // TODO(xiongkun03) in old executor, this type of variable is not support
+    // eager deletion. so we just leave it here ?
+  } else if (var->IsType<LoDRankTable>()) {
+    // TODO(xiongkun03) in old executor, this type of variable is not support
+    // eager deletion. so we just leave it here ?
+  } else if (var->IsType<SelectedRows>()) {
+    Add(var->GetMutable<SelectedRows>()->mutable_value()->MoveMemoryHolder());
+    var->GetMutable<SelectedRows>()->mutable_rows()->clear();
+  } else if (var->IsType<LoDTensorArray>()) {
+    auto* tensor_arr = var->GetMutable<LoDTensorArray>();
+    for (auto& t : *tensor_arr) {
+      Add(t.MoveMemoryHolder());
+    }
+  } else if (var->IsType<std::vector<Scope*>>()) {
+    // NOTE(@xiongkun03) conditional_op / while_op will create a STEP_SCOPE
+    // refer to executor.cc to see what old garbage collector does.
+    // do nothing, because the sub scope will be deleted by sub-executor.
+  } else {
+    PADDLE_THROW(platform::errors::Unimplemented(
+        "The variable(%s) is not supported in eager deletion.",
+        framework::ToTypeName(var->Type())));
+  }
+}
+
+void InterpreterCoreFastGarbageCollector::Add(Garbage garbage) {
+  if (!garbage) {
+    return;
+  }
+
+  if (max_memory_size_ > 1) {
+    std::unique_ptr<GarbageQueue> pending_delete_garbages;
+    {  // lock guard
+      std::lock_guard<memory::SpinLock> guard(spinlock_);
+      cur_memory_size_ += garbage->size();
+      garbages_->push_back(std::move(garbage));
+
+      if (cur_memory_size_ >= max_memory_size_) {
+        cur_memory_size_ = 0;
+        pending_delete_garbages = std::move(garbages_);
+        garbages_ = std::make_unique<GarbageQueue>();
+      }
+    }
+  }
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/new_executor/interpretercore_fast_garbage_collector.h

+// 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.
+#pragma once
+
+#include "paddle/fluid/framework/new_executor/interpretercore_garbage_collector.h"
+
+namespace paddle {
+namespace framework {
+
+class InterpreterCoreFastGarbageCollector
+    : public InterpreterCoreGarbageCollector {
+ public:
+  virtual void Add(Variable* var) override;
+
+ private:
+  void Add(Garbage garbage);
+};
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/new_executor/interpretercore_garbage_collector.cc

@@ -19,114 +19,23 @@ namespace paddle {
 namespace framework {
 
 InterpreterCoreGarbageCollector::InterpreterCoreGarbageCollector() {
-  garbages_.reset(new GarbageQueue());
-  max_memory_size_ = static_cast<size_t>(GetEagerDeletionThreshold());
+  garbages_ = std::make_unique<GarbageQueue>();
+  max_memory_size_ = static_cast<int64_t>(GetEagerDeletionThreshold());
   cur_memory_size_ = 0;
-
-  WorkQueueOptions options(/*num_threads*/ 1, /*allow_spinning*/ true,
-                           /*track_task*/ false);
-  queue_ = CreateSingleThreadedWorkQueue(options);
-}
-
-InterpreterCoreGarbageCollector::~InterpreterCoreGarbageCollector() {
-  queue_.reset(nullptr);
 }
 
-void InterpreterCoreGarbageCollector::Add(
-    std::shared_ptr<memory::Allocation> garbage,
-    paddle::platform::DeviceEvent& event, const platform::DeviceContext* ctx) {
-  if (max_memory_size_ <= 1) {
-    Free(garbage, event, ctx);
-  } else {
-    if (!garbage) return;
-    GarbageQueue* garbage_ptr = nullptr;
-    {
-      std::lock_guard<paddle::memory::SpinLock> guard(spinlock_);
-      cur_memory_size_ += garbage->size();
-      garbages_->push_back(std::move(garbage));
-
-      if (cur_memory_size_ >= max_memory_size_) {
-        cur_memory_size_ = 0;
-        garbage_ptr = garbages_.release();
-        garbages_.reset(new GarbageQueue());
-      }
-    }
-    if (garbage_ptr) {
-      Free(garbage_ptr, event, ctx);
-    }
-  }
+void InterpreterCoreGarbageCollector::Add(Variable* var) {
+  PADDLE_THROW(
+      platform::errors::Unimplemented("Not allowed to call the member function "
+                                      "of InterpreterCoreGarbageCollector"));
 }
 
-void InterpreterCoreGarbageCollector::Add(paddle::framework::Variable* var,
-                                          paddle::platform::DeviceEvent& event,
+void InterpreterCoreGarbageCollector::Add(Variable* var,
+                                          platform::DeviceEvent& event,
                                           const platform::DeviceContext* ctx) {
-  if (!var) {
-    return;
-  }
-
-  if (var->IsType<LoDTensor>()) {
-    Add(var->GetMutable<LoDTensor>()->MoveMemoryHolder(), event, ctx);
-  } else if (var->IsType<
-                 operators::reader::
-                     OrderedMultiDeviceLoDTensorBlockingQueueHolder>()) {
-    // TODO(xiongkun03) in old executor, this type of variable is not support
-    // eager deletion. so we just leave it here ?
-  } else if (var->IsType<LoDRankTable>()) {
-    // TODO(xiongkun03) in old executor, this type of variable is not support
-    // eager deletion. so we just leave it here ?
-  } else if (var->IsType<SelectedRows>()) {
-    Add(var->GetMutable<SelectedRows>()->mutable_value()->MoveMemoryHolder(),
-        event, ctx);
-    var->GetMutable<SelectedRows>()->mutable_rows()->clear();
-  } else if (var->IsType<LoDTensorArray>()) {
-    auto* tensor_arr = var->GetMutable<LoDTensorArray>();
-    for (auto& t : *tensor_arr) {
-      Add(t.MoveMemoryHolder(), event, ctx);
-    }
-  } else if (var->IsType<std::vector<Scope*>>()) {
-    // NOTE(@xiongkun03) conditional_op / while_op will create a STEP_SCOPE
-    // refer to executor.cc to see what old garbage collector does.
-    // do nothing, because the sub scope will be deleted by sub-executor.
-  } else {
-    PADDLE_THROW(platform::errors::Unimplemented(
-        "The variable(%s) is not supported in eager deletion.",
-        framework::ToTypeName(var->Type())));
-  }
-}
-
-void InterpreterCoreGarbageCollector::Free(GarbageQueue* garbages,
-                                           paddle::platform::DeviceEvent& event,
-                                           const platform::DeviceContext* ctx) {
-  event.Record(ctx);
-  event.SetFininshed();  // Only for CPU Event
-  queue_->AddTask([ container = garbages, event = &event ]() {
-    while (!event->Query()) {
-#if defined(_WIN32)
-      SleepEx(50, FALSE);
-#else
-      sched_yield();
-#endif
-      continue;
-    }
-    delete container;
-  });
-}
-
-void InterpreterCoreGarbageCollector::Free(
-    std::shared_ptr<memory::Allocation>& garbage,
-    paddle::platform::DeviceEvent& event, const platform::DeviceContext* ctx) {
-  event.Record(ctx);
-  event.SetFininshed();  // Only for CPU Event
-  queue_->AddTask([ container = garbage, event = &event ]() {
-    while (!event->Query()) {
-#if defined(_WIN32)
-      SleepEx(50, FALSE);
-#else
-      sched_yield();
-#endif
-      continue;
-    }
-  });
+  PADDLE_THROW(
+      platform::errors::Unimplemented("Not allowed to call the member function "
+                                      "of InterpreterCoreGarbageCollector"));
 }
 
 }  // namespace framework

paddle/fluid/framework/new_executor/interpretercore_garbage_collector.h

@@ -13,54 +13,31 @@
 // limitations under the License.
 #pragma once
 
-#if !defined(_WIN32)
-#include <sched.h>
-#else
-#define NOMINMAX
-#include <windows.h>
-#endif  // !_WIN32
-
 #include <queue>
-#include <vector>
-
-#include "paddle/fluid/framework/new_executor/workqueue/workqueue.h"
 #include "paddle/fluid/memory/allocation/spin_lock.h"
 #include "paddle/fluid/platform/device_event.h"
 
 namespace paddle {
 namespace framework {
 
-using GarbageQueue = std::deque<std::shared_ptr<memory::Allocation>>;
+using Garbage = std::shared_ptr<memory::Allocation>;
+using GarbageQueue = std::deque<Garbage>;
+
 class InterpreterCoreGarbageCollector {
  public:
   InterpreterCoreGarbageCollector();
-
-  ~InterpreterCoreGarbageCollector();
-
-  void Add(std::shared_ptr<memory::Allocation> garbage,  // NOLINT
-           paddle::platform::DeviceEvent& event,         // NOLINT
-           const platform::DeviceContext* ctx);
-
-  void Add(paddle::framework::Variable* var,
-           paddle::platform::DeviceEvent& event,  // NOLINT
+  virtual ~InterpreterCoreGarbageCollector(){};
+  virtual void Add(Variable* var);
+  virtual void Add(Variable* var, platform::DeviceEvent& event,
                    const platform::DeviceContext* ctx);
-
   DISABLE_COPY_AND_ASSIGN(InterpreterCoreGarbageCollector);
 
- private:
-  void Free(GarbageQueue* garbages,
-            paddle::platform::DeviceEvent& event,  // NOLINT
-            const platform::DeviceContext* ctx);
-
-  void Free(std::shared_ptr<memory::Allocation>& garbage,  // NOLINT
-            paddle::platform::DeviceEvent& event,          // NOLINT
-            const platform::DeviceContext* ctx);
-
+ protected:
   std::unique_ptr<GarbageQueue> garbages_;
-  size_t max_memory_size_;
-  size_t cur_memory_size_;
-  std::unique_ptr<WorkQueue> queue_;
-  paddle::memory::SpinLock spinlock_;
+  int64_t max_memory_size_;
+  int64_t cur_memory_size_;
+  memory::SpinLock spinlock_;
 };
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/tensor.cc

@@ -98,7 +98,9 @@ void* Tensor::mutable_data(const platform::Place& place,
 
   /* some versions of boost::variant don't have operator!= */
   if (holder_ == nullptr || !(holder_->place() == place) ||
-      holder_->size() < size + offset_) {
+      holder_->size() < size + offset_ ||
+      !(platform::is_gpu_place(place) &&
+        memory::InSameStream(holder_, stream))) {
     holder_.reset();
     holder_ = memory::AllocShared(place, size, stream);
     offset_ = 0;

paddle/fluid/memory/allocation/allocator_facade.cc

@@ -26,6 +26,7 @@
 #include "paddle/fluid/platform/place.h"
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+#include <shared_mutex>
 #include "paddle/fluid/memory/allocation/cuda_allocator.h"
 #include "paddle/fluid/memory/allocation/pinned_allocator.h"
 #include "paddle/fluid/memory/allocation/stream_safe_cuda_allocator.h"
@@ -151,11 +152,12 @@ class AllocatorFacadePrivate {
         }
 #endif
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
-        if (FLAGS_use_stream_safe_cuda_allocator) {
-          LOG(WARNING) << "FLAGS_use_stream_safe_cuda_allocator is invalid for "
-                          "naive_best_fit strategy";
-          FLAGS_use_stream_safe_cuda_allocator = false;
-        }
+        PADDLE_ENFORCE_EQ(
+            FLAGS_use_stream_safe_cuda_allocator, false,
+            paddle::platform::errors::Unimplemented(
+                "StreamSafeCUDAAllocator is only implemented for auto_growth "
+                "strategy, not support naive_best_fit strategy"));
+
         for (int dev_id = 0; dev_id < platform::GetGPUDeviceCount(); ++dev_id) {
           InitNaiveBestFitCUDAAllocator(platform::CUDAPlace(dev_id));
         }
@@ -185,9 +187,6 @@ class AllocatorFacadePrivate {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
         allow_free_idle_chunk_ = allow_free_idle_chunk;
         if (FLAGS_use_stream_safe_cuda_allocator) {
-          default_streams_ =
-              std::vector<gpuStream_t>(platform::GetGPUDeviceCount(), nullptr);
-          // TODO(Ruibiao): Support multi-stream allocator for other strategies
           for (int dev_id = 0; dev_id < platform::GetGPUDeviceCount();
                ++dev_id) {
             InitStreamSafeCUDAAllocator(platform::CUDAPlace(dev_id), nullptr);
@@ -232,11 +231,11 @@ class AllocatorFacadePrivate {
         }
 #endif
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
-        if (FLAGS_use_stream_safe_cuda_allocator) {
-          LOG(WARNING) << "FLAGS_use_stream_safe_cuda_allocator is invalid for "
-                          "thread_local strategy";
-          FLAGS_use_stream_safe_cuda_allocator = false;
-        }
+        PADDLE_ENFORCE_EQ(
+            FLAGS_use_stream_safe_cuda_allocator, false,
+            paddle::platform::errors::Unimplemented(
+                "StreamSafeCUDAAllocator is only implemented for auto_growth "
+                "strategy, not support thread_local strategy"));
 
         for (int dev_id = 0; dev_id < platform::GetGPUDeviceCount(); ++dev_id) {
           InitThreadLocalCUDAAllocator(platform::CUDAPlace(dev_id));
@@ -282,50 +281,45 @@ class AllocatorFacadePrivate {
   }
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+  bool HasCUDAAllocator(const platform::CUDAPlace& place,
+                        const gpuStream_t& stream) {
+    auto it = cuda_allocators_.find(place);
+    if (it == cuda_allocators_.end()) {
+      return false;
+    }
+    const std::map<gpuStream_t, std::shared_ptr<Allocator>>& allocator_map =
+        it->second;
+    return allocator_map.find(stream) != allocator_map.end();
+  }
+
   const std::shared_ptr<Allocator>& GetAllocator(
       const platform::CUDAPlace& place, const gpuStream_t& stream,
       bool create_if_not_found = false) {
-    auto place_it = cuda_allocators_.find(place);
-    PADDLE_ENFORCE_NE(place_it, cuda_allocators_.end(),
+    {  // shared_lock_guard
+      std::shared_lock<std::shared_timed_mutex> lock_guard(
+          cuda_allocator_mutex_);
+      if (LIKELY(HasCUDAAllocator(place, stream))) {
+        return cuda_allocators_[place][stream];
+      } else {
+        PADDLE_ENFORCE_NE(create_if_not_found, false,
                           platform::errors::NotFound(
-                          "No allocator found for the place %s", place));
+                              "No allocator found for stream %s in place %s "
+                              "with create_if_not_found = false",
+                              stream, place));
+      }
+    }
 
-    const std::map<gpuStream_t, std::shared_ptr<Allocator>>& allocator_map =
-        place_it->second;
-    auto stream_it = allocator_map.find(stream);
-    if (stream_it == allocator_map.end()) {
-      if (create_if_not_found) {
+    {  // unique_lock_guard
+      std::unique_lock<std::shared_timed_mutex> lock_guard(
+          cuda_allocator_mutex_);
       InitStreamSafeCUDAAllocator(place, stream);
       return cuda_allocators_[place][stream];
-      } else {
-        PADDLE_THROW(platform::errors::NotFound(
-            "No allocator found for stream %s in place %s", stream, place));
-      }
     }
-    return stream_it->second;
   }
 
-  const gpuStream_t& GetDefaultStream(const platform::CUDAPlace& place) {
-    int dev_id = place.GetDeviceId();
-    gpuStream_t& default_stream = default_streams_[dev_id];
-    if (UNLIKELY(default_stream == nullptr)) {
-      /* NOTE(Ruibiao): Here if we set default_stream by code " default_stream =
-       * platform::stream::get_current_stream(place.GetDeviceId())->raw_stream()
-       * ", then it will be fail to make target 'jit_kernel_benchmark', says a
-       * undefined reference to `paddle::platform::DeviceContextPool::Get(
-       * paddle::platform::Place const&)' in function
-       * `paddle::platform::stream::get_current_stream(int)'. However, target
-       * allocator_facade will not be affected. It seems a circular dependency
-       * problem between 'cuda_stream' and 'device_context' that causes this
-       * strange bug.
-       */
-      platform::DeviceContextPool& pool =
-          platform::DeviceContextPool::Instance();
-      default_stream =
-          static_cast<platform::CUDADeviceContext*>(pool.Get(place))->stream();
-      InitStreamSafeCUDAAllocator(place, default_stream);
-    }
-    return default_stream;
+  gpuStream_t GetDefaultStream(const platform::CUDAPlace& place) {
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    return static_cast<platform::CUDADeviceContext*>(pool.Get(place))->stream();
   }
 
   void RecordStream(std::shared_ptr<Allocation> allocation,
@@ -443,18 +437,12 @@ class AllocatorFacadePrivate {
             "Only support auto-growth strategey for StreamSafeCUDAAllocator, "
             "the allocator strategy %d is unsupported for multi-stream",
             static_cast<int>(strategy_)));
-    VLOG(9) << "Init CUDA allocator for stream " << stream << " in place " << p;
-    std::lock_guard<SpinLock> lock_guard(cuda_allocators_lock_);
-    try {
-      GetAllocator(p, stream);
-      VLOG(9) << "Other thread had build a allocator for stream " << stream
-              << " in place " << p;
-    } catch (platform::EnforceNotMet&) {
+    if (LIKELY(!HasCUDAAllocator(p, stream))) {
+      VLOG(8) << "Init CUDA allocator for stream " << stream << " in place "
+              << p;
       InitAutoGrowthCUDAAllocator(p, stream);
       WrapStreamSafeCUDAAllocator(p, stream);
       WrapCUDARetryAllocator(p, stream, FLAGS_gpu_allocator_retry_time);
-    } catch (...) {
-      throw;
     }
   }
 
@@ -618,7 +606,7 @@ class AllocatorFacadePrivate {
 
   void WrapStreamSafeCUDAAllocator(platform::CUDAPlace p, gpuStream_t stream) {
     const std::shared_ptr<Allocator>& underlying_allocator =
-        GetAllocator(p, stream);
+        cuda_allocators_[p][stream];
     cuda_allocators_[p][stream] = std::make_shared<StreamSafeCUDAAllocator>(
         underlying_allocator, p, stream);
   }
@@ -629,7 +617,7 @@ class AllocatorFacadePrivate {
         retry_time, 0,
         platform::errors::InvalidArgument(
             "Retry time should be larger than 0, but got %d", retry_time));
-    std::shared_ptr<Allocator> allocator = GetAllocator(p, stream);
+    std::shared_ptr<Allocator> allocator = cuda_allocators_[p][stream];
     allocator = std::make_shared<RetryAllocator>(allocator, retry_time);
   }
 
@@ -784,8 +772,7 @@ class AllocatorFacadePrivate {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   // a standalone CUDA allocator to support multi-stream GC in new executor
   CUDAAllocatorMap cuda_allocators_;
-  std::vector<gpuStream_t> default_streams_;
-  SpinLock cuda_allocators_lock_;
+  std::shared_timed_mutex cuda_allocator_mutex_;
 #ifdef PADDLE_WITH_CUDA
   std::unordered_map<CUDAGraphID, std::unique_ptr<AllocatorFacadePrivate>>
       cuda_graph_allocator_map_;
@@ -903,6 +890,31 @@ std::shared_ptr<Allocation> AllocatorFacade::AllocShared(
 #endif
 }
 
+bool AllocatorFacade::InSameStream(
+    const std::shared_ptr<Allocation>& allocation,
+    const platform::Stream& stream) {
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+  PADDLE_ENFORCE_EQ(
+      FLAGS_use_stream_safe_cuda_allocator, true,
+      platform::errors::Unimplemented(
+          "StreamSafeCUDAAllocator is disabled, you should not call this "
+          "multi-stream 'InSameStream' function. To enable it, you can enter"
+          "'export FLAGS_use_stream_safe_cuda_allocator=true' in the "
+          "terminal."));
+
+#ifdef PADDLE_WITH_CUDA
+  if (UNLIKELY(platform::CUDAGraph::IsCapturing())) {
+    PADDLE_THROW(platform::errors::Unavailable(
+        "Not allow to use StreamSafeCUDAAllocator with CUDAGraphAllocator"));
+  }
+#endif
+  gpuStream_t s = reinterpret_cast<gpuStream_t>(stream.id());
+  return s == GetStream(allocation);
+#else
+  PADDLE_THROW(platform::errors::PreconditionNotMet("Not compiled with GPU."));
+#endif
+}
+
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 AllocationPtr AllocatorFacade::Alloc(const platform::Place& place, size_t size,
                                      const gpuStream_t& stream) {

paddle/fluid/memory/allocation/allocator_facade.h

@@ -61,6 +61,10 @@ class AllocatorFacade {
   std::shared_ptr<Allocation> AllocShared(const platform::Place& place,
                                           size_t size,
                                           const platform::Stream& stream);
+
+  bool InSameStream(const std::shared_ptr<Allocation>& allocation,
+                    const platform::Stream& stream);
+
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   // TODO(zhiqiu): change gpuStream_t to platform::Stream if needed.
   AllocationPtr Alloc(const platform::Place& place, size_t size,

paddle/fluid/memory/allocation/stream_safe_cuda_allocator.cc

@@ -118,6 +118,7 @@ bool StreamSafeCUDAAllocator::IsAllocThreadSafe() const { return true; }
 
 Allocation* StreamSafeCUDAAllocator::AllocateImpl(size_t size) {
   ProcessUnfreedAllocations();
+  VLOG(8) << "Try allocate " << size << " bytes";
   AllocationPtr underlying_allocation;
   try {
     underlying_allocation = underlying_allocator_->Allocate(size);
@@ -150,10 +151,12 @@ void StreamSafeCUDAAllocator::FreeImpl(Allocation* allocation) {
                               "StreamSafeCUDAAllocation*",
                               allocation));
   VLOG(8) << "Try free allocation " << stream_safe_cuda_allocation->ptr();
+  std::lock_guard<SpinLock> lock_guard(unfreed_allocation_lock_);
   if (stream_safe_cuda_allocation->CanBeFreed()) {
+    VLOG(9) << "Directly delete allocation";
     delete stream_safe_cuda_allocation;
   } else {
-    std::lock_guard<SpinLock> lock_guard(unfreed_allocation_lock_);
+    VLOG(9) << "Put into unfreed_allocation list";
     unfreed_allocations_.emplace_back(stream_safe_cuda_allocation);
   }
 }

paddle/fluid/memory/malloc.cc

@@ -41,6 +41,12 @@ std::shared_ptr<Allocation> AllocShared(const platform::Place& place,
                                                              stream);
 }
 
+bool InSameStream(const std::shared_ptr<Allocation>& allocation,
+                  const platform::Stream& stream) {
+  return allocation::AllocatorFacade::Instance().InSameStream(allocation,
+                                                              stream);
+}
+
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 AllocationPtr Alloc(const platform::CUDAPlace& place, size_t size,
                     const gpuStream_t& stream) {

paddle/fluid/memory/malloc.h

@@ -45,6 +45,9 @@ extern std::shared_ptr<Allocation> AllocShared(const platform::Place& place,
                                                size_t size,
                                                const platform::Stream& stream);
 
+extern bool InSameStream(const std::shared_ptr<Allocation>& allocation,
+                         const platform::Stream& stream);
+
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 extern AllocationPtr Alloc(const platform::CUDAPlace& place, size_t size,
                            const gpuStream_t& stream);

python/paddle/fluid/tests/unittests/interpreter/CMakeLists.txt

@@ -2,7 +2,8 @@ file(GLOB TEST_INTERP_CASES RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
 string(REPLACE ".py" "" TEST_INTERP_CASES "${TEST_INTERP_CASES}")
 
 foreach(target ${TEST_INTERP_CASES})
-  py_test_modules(${target} MODULES ${target})
+  py_test_modules(${target} MODULES ${target} ENVS FLAGS_allocator_strategy=auto_growth FLAGS_use_stream_safe_cuda_allocator=true FLAGS_fast_eager_deletion_mode=false FLAGS_eager_delete_tensor_gb=0)
+  py_test_modules(${target}_non_eager_deletion MODULES ${target} ENVS FLAGS_allocator_strategy=auto_growth FLAGS_use_stream_safe_cuda_allocator=true FLAGS_fast_eager_deletion_mode=false FLAGS_eager_delete_tensor_gb=0.000001) 
+  py_test_modules(${target}_fast_gc MODULES ${target} ENVS FLAGS_allocator_strategy=auto_growth FLAGS_use_stream_safe_cuda_allocator=true FLAGS_fast_eager_deletion_mode=true FLAGS_eager_delete_tensor_gb=0)
+  py_test_modules(${target}_fast_gc_non_eager_deletion MODULES ${target} ENVS FLAGS_allocator_strategy=auto_growth FLAGS_use_stream_safe_cuda_allocator=true FLAGS_fast_eager_deletion_mode=true FLAGS_eager_delete_tensor_gb=0.000001)
 endforeach()
-
-set_tests_properties(test_standalone_executor PROPERTIES ENVIRONMENT FLAGS_use_stream_safe_cuda_allocator=true)