Merge pull request #16506 from sneaxiy/revert-16424-fix_allocator_bug

Revert "Fix allocator bug"

Author: sneaxiy

paddle/fluid/framework/operator.h

@@ -365,6 +365,9 @@ class ExecutionContext {
     auto shared_allocation = std::shared_ptr<memory::allocation::Allocation>(
         allocation_ptr, deleter);
 
+    PADDLE_ENFORCE(
+        dynamic_cast<platform::TemporaryAllocation*>(allocation_ptr) != nullptr,
+        "The AllocationPtr must be TemporaryAllocation.");
     PADDLE_ENFORCE_GE(allocation_ptr->size(),
                       framework::product(dim) * sizeof(T));
 

paddle/fluid/memory/allocation/CMakeLists.txt

@@ -4,7 +4,6 @@ cc_library(best_fit_allocator SRCS best_fit_allocator.cc DEPS allocator)
 cc_library(locked_allocator SRCS locked_allocator.cc DEPS allocator)
 cc_library(buffered_allocator SRCS buffered_allocator.cc DEPS allocator)
 cc_library(legacy_allocator SRCS legacy_allocator.cc DEPS allocator buddy_allocator profiler)
-cc_library(zero_size_allocator SRCS zero_size_allocator.cc DEPS allocator)
 cc_test(buffered_allocator_test SRCS buffered_allocator_test.cc DEPS best_fit_allocator locked_allocator buffered_allocator cpu_allocator)
 
 if (WITH_GPU)
@@ -38,20 +37,30 @@ else ()
     set(AllocatorFacadeDeps)
 endif()
 
-list(APPEND AllocatorFacadeDeps cpu_allocator locked_allocator best_fit_allocator aligned_allocator auto_increment_allocator conditional_allocator retry_allocator buffered_allocator legacy_allocator zero_size_allocator)
-
 cc_library(aligned_allocator SRCS aligned_allocator.cc DEPS allocator)
 cc_library(auto_increment_allocator SRCS auto_increment_allocator.cc DEPS allocator)
+cc_library(zero_size_allocator SRCS zero_size_allocator.cc DEPS allocator)
 cc_library(conditional_allocator SRCS conditional_allocator.cc DEPS allocator)
-cc_library(allocator_strategy SRCS allocator_strategy.cc DEPS gflags ${AllocatorFacadeDeps})
-cc_library(allocator_facade SRCS allocator_facade.cc DEPS allocator_strategy)
+cc_library(allocator_strategy SRCS allocator_strategy.cc DEPS gflags)
+cc_library(allocator_facade SRCS allocator_facade.cc DEPS
+        ${AllocatorFacadeDeps}
+        cpu_allocator
+        locked_allocator
+        best_fit_allocator
+        aligned_allocator
+        auto_increment_allocator
+        zero_size_allocator
+        conditional_allocator
+        retry_allocator
+        buffered_allocator
+        allocator_strategy
+        legacy_allocator
+        )
 
 nv_test(allocation_and_eigen_test SRCS allocation_and_eigen_test.cu DEPS allocator_facade)
 
 cc_test(retry_allocator_test SRCS retry_allocator_test.cc DEPS retry_allocator best_fit_allocator locked_allocator cpu_allocator)
 
-cc_test(naive_best_fit_allocator_facade_test SRCS naive_best_fit_allocator_facade_test.cc DEPS allocator_facade)
-
 cc_test(allocator_facade_abs_flags_test SRCS allocator_facade_abs_flags_test.cc DEPS allocator_facade)
 
 cc_test(allocator_facade_frac_flags_test SRCS allocator_facade_frac_flags_test.cc DEPS allocator_facade)

paddle/fluid/memory/allocation/aligned_allocator.h

@@ -94,8 +94,6 @@ class AlignedAllocator : public ThinAlignedAllocator {
         underlying_allocator_->Allocate(size + kAlignment, attr);
     return new AlignedAllocation<kAlignment>(std::move(raw_allocation), size);
   }
-
-  void FreeImpl(Allocation* allocation) override { delete allocation; }
 };
 
 }  // namespace allocation

paddle/fluid/memory/allocation/allocator.cc

@@ -27,24 +27,16 @@ bool Allocator::IsAllocThreadSafe() const { return false; }
 
 AllocationPtr Allocator::Allocate(size_t size, Allocator::Attr attr) {
   auto ptr = AllocateImpl(size, attr);
-  ptr->RegisterDecoratedAllocator(this);
+  ptr->set_allocator(this);
   return AllocationPtr(ptr);
 }
 
-void Allocator::FreeImpl(Allocation* allocation) {
-  Allocator* allocator = allocation->TopDecoratedAllocator();
-  allocator->Free(allocation);
-}
-
-void Allocator::Free(Allocation* allocation) {
-  allocation->PopDecoratedAllocator();
-  FreeImpl(allocation);
-}
+void Allocator::Free(Allocation* allocation) { delete allocation; }
 
 const char* BadAlloc::what() const noexcept { return msg_.c_str(); }
 
 void AllocationDeleter::operator()(Allocation* allocation) const {
-  Allocator* allocator = allocation->TopDecoratedAllocator();
+  auto* allocator = allocation->allocator();
   allocator->Free(allocation);
 }
 

paddle/fluid/memory/allocation/allocator.h

@@ -46,56 +46,13 @@ class Allocator;
 // NOTE: this is the base class of Allocation. Each allocator can use its own
 //       allocation object.
 // NOTE: the `Allocation::ptr()` could be nullptr, if the allocation size is 0
-
-/**
- * Allocation is returned by Allocator::Allocate() method.
- *
- * An allocator may be decorated by another allocator. For example, we can
- * decorate
- * a RetryAllocator to any allocator to perform allocation retry when first
- * allocation request fails.
- *
- * Explanations of Allocator design is as follows:
- *
- * Suppose we have an allocator which is decorated by several allocators:
- *
- *   A(1) <- A(2) <- A(3) <- ... <- A(n)
- *
- * , and the public allocator is A(1).
- *
- * The allocation process would be:
- *
- *   A(n).Allocate() -> ... -> A(2).Allocate() -> A(1).Allocate()
- *
- * , and the free process would be:
- *
- *   A(1).Free() -> A(2).Free() -> ... -> A(n).Free()
- *
- * Therefore, we should record the allocator chain when allocating, so
- * that we can free the allocation in the reverse order of allocator chain.
- * The field `decorated_allocators_` is used to record this chain.
- *
- * Another example is that we want to add additional fields in Allocation,
- * e.g., something what is done in AlignedAllocator, etc.
- * In this case, we should declare a derived class of Allocation, which
- * contains an underlying Allocation allocated by the underlying allocator.
- * Therefore, `decorated_allocators_` of the new Allocation object would
- * be a new chain, differing from the underlying Allocation object.
- */
 class Allocation {
  public:
   Allocation(void* ptr, size_t size, platform::Place place)
-      : ptr_(ptr), size_(size), place_(place) {
-    // NOTE(zjl): Since decorated_allocators_ is usually a small vector
-    // We reserve a small buffer to it to prevent frequent heap allocation
-    // Not quite sure whether we need something like gtl vector.
-    decorated_allocators_.reserve(8);
-  }
+      : allocator_(nullptr), ptr_(ptr), size_(size), place_(place) {}
 
   Allocation(const Allocation& o) = delete;
   Allocation& operator=(const Allocation& o) = delete;
-  Allocation(Allocation&& o) = delete;
-  Allocation& operator=(Allocation&& o) = delete;
 
   // Returns the holding pointer.
   // NOTE: For performance consideration, it is better not to make this method
@@ -117,31 +74,17 @@ class Allocation {
 
   const platform::Place& place() const { return place_; }
 
-  virtual ~Allocation();
-
- private:
-  const std::vector<Allocator*>& DecoratedAllocators() const {
-    return decorated_allocators_;
-  }
-
-  inline void RegisterDecoratedAllocator(Allocator* allocator) {
-    decorated_allocators_.push_back(allocator);
-  }
+  Allocator* allocator() { return allocator_; }
 
-  inline void PopDecoratedAllocator() { decorated_allocators_.pop_back(); }
+  void set_allocator(Allocator* allocator) { allocator_ = allocator; }
 
-  inline Allocator* TopDecoratedAllocator() {
-    return decorated_allocators_.back();
-  }
+  virtual ~Allocation();
 
  private:
+  Allocator* allocator_;
   void* ptr_;
   size_t size_;
   platform::Place place_;
-  std::vector<Allocator*> decorated_allocators_;
-
-  friend class Allocator;
-  friend class AllocationDeleter;
 };
 
 using AllocationPtr = std::unique_ptr<Allocation, AllocationDeleter>;
@@ -191,12 +134,9 @@ class Allocator {
   // True if the `Allocate` is thread safe.
   virtual bool IsAllocThreadSafe() const;
 
-  // This function should not be called outside
-  void Free(Allocation* allocation);
-
  protected:
+  virtual void Free(Allocation* allocation);
   virtual Allocation* AllocateImpl(size_t size, Allocator::Attr attr) = 0;
-  virtual void FreeImpl(Allocation* allocation);
 
  private:
   friend class AllocationDeleter;

paddle/fluid/memory/allocation/allocator_facade.cc

@@ -49,17 +49,6 @@ namespace paddle {
 namespace memory {
 namespace allocation {
 
-static inline std::shared_ptr<Allocator> WrapRetryAllocator(
-    std::shared_ptr<Allocator> allocator, int64_t retry_time) {
-  if (retry_time > 0) {
-    auto* retry_allocator =
-        new RetryAllocator(std::move(allocator), retry_time);
-    allocator.reset(retry_allocator);
-  }
-
-  return allocator;
-}
-
 // TODO(yy): Dirty code here. This class should be configurable in runtime.
 class CPUManagedAllocator : public Allocator {
  public:
@@ -123,10 +112,14 @@ class ChunkedAllocator : public Allocator {
   std::shared_ptr<Allocator> CreateAllocatorWithChunk() {
     chunks_.emplace_back(raw_allocator_->Allocate(max_chunk_size_));
     auto* allocation = chunks_.back().get();
-    std::shared_ptr<Allocator> allocator(new LockedAllocator(
-        std::shared_ptr<Allocator>(new BestFitAllocator(allocation))));
+    std::unique_ptr<Allocator> allocator(new LockedAllocator(
+        std::unique_ptr<Allocator>(new BestFitAllocator(allocation))));
 
-    allocator = WrapRetryAllocator(allocator, retry_time_);
+    if (retry_time_ > 0) {
+      auto* retry_allocator =
+          new RetryAllocator(std::move(allocator), retry_time_);
+      allocator.reset(retry_allocator);
+    }
 
     return std::make_shared<AlignedAllocator<64u>>(std::move(allocator));
   }
@@ -197,23 +190,13 @@ class AllocatorFacadePrivate {
   ~AllocatorFacadePrivate() = default;
 
   AllocatorFacadePrivate() {
-    auto strategy = GetAllocatorStrategy();
-    switch (strategy) {
-      case AllocatorStrategy::kLegacy: {
-        InitLegacyAllocator();
-        break;
-      }
-      case AllocatorStrategy::kNaiveBestFit: {
-        InitCPUAllocator();
-        InitCUDAAllocator();
-        InitCUDAPinnedAllocator();
-        WrapZeroSizeAllocator();
-        break;
-      }
-      default: {
-        PADDLE_THROW("Unsupported allocator strategy: %d",
-                     static_cast<int>(strategy));
-      }
+    if (GetAllocatorStrategy() == AllocatorStrategy::kLegacy) {
+      InitLegacyAllocator();
+    } else {
+      InitCPUAllocator();
+      InitCUDAAllocator();
+      InitCUDAPinnedAllocator();
+      WrapZeroSizeAllocator();
     }
   }
 
@@ -271,7 +254,8 @@ AllocatorFacade& AllocatorFacade::Instance() {
 
 std::shared_ptr<Allocation> AllocatorFacade::AllocShared(
     const platform::Place& place, size_t size, Allocator::Attr attr) {
-  return std::shared_ptr<Allocation>(Alloc(place, size, attr));
+  return std::shared_ptr<Allocation>(Alloc(place, size, attr).release(),
+                                     AllocationDeleter());
 }
 
 AllocationPtr AllocatorFacade::Alloc(const platform::Place& place, size_t size,

paddle/fluid/memory/allocation/allocator_strategy.cc

@@ -19,22 +19,16 @@
 DEFINE_string(
     allocator_strategy, "legacy",
     "The allocation strategy. Legacy means the original allocator of Fluid."
-    "naive_best_fit means the experimental best fit allocator. "
-    "allocator. Enum in [legacy, naive_best_fit].");
+    "New means the experimental allocators of Fluid. in [legacy, new]");
 
 namespace paddle {
 namespace memory {
 namespace allocation {
 
 static AllocatorStrategy GetStrategyFromFlag() {
-  if (FLAGS_allocator_strategy == "legacy") {
-    return AllocatorStrategy::kLegacy;
-  } else if (FLAGS_allocator_strategy == "naive_best_fit") {
-    return AllocatorStrategy::kNaiveBestFit;
-  } else {
-    PADDLE_THROW("Unsupported allocator strategy: %s",
-                 FLAGS_allocator_strategy);
-  }
+  return FLAGS_allocator_strategy == "legacy"
+             ? AllocatorStrategy::kLegacy
+             : AllocatorStrategy::kNaiveBestFit;
 }
 
 AllocatorStrategy GetAllocatorStrategy() {

paddle/fluid/memory/allocation/best_fit_allocator.cc

@@ -109,7 +109,7 @@ size_t BestFitAllocator::NumFreeChunks() const {
   }
   return num;
 }
-void BestFitAllocator::FreeImpl(Allocation* allocation) {
+void BestFitAllocator::Free(Allocation* allocation) {
   auto* bf_allocation = dynamic_cast<BestFitAllocation*>(allocation);
   PADDLE_ENFORCE_NOT_NULL(bf_allocation,
                           "The input allocation is not BestFitAllocation.");

paddle/fluid/memory/allocation/best_fit_allocator.h

@@ -119,7 +119,7 @@ class BestFitAllocator : public Allocator {
   void InsertFreeNode(const ListIt& it);
 
  protected:
-  void FreeImpl(Allocation* allocation) override;
+  void Free(Allocation* allocation) override;
   Allocation* AllocateImpl(size_t size, Allocator::Attr attr) override;
 
  private:

paddle/fluid/memory/allocation/buffered_allocator.cc

@@ -22,11 +22,11 @@ namespace paddle {
 namespace memory {
 namespace allocation {
 
-BufferedAllocator::BufferedAllocator(std::shared_ptr<Allocator> allocator)
+BufferedAllocator::BufferedAllocator(std::unique_ptr<Allocator> &&allocator)
     : underlying_allocator_(std::move(allocator)) {
   PADDLE_ENFORCE_NOT_NULL(
       underlying_allocator_,
-      "Underlying allocator of BufferedAllocator must not be null");
+      "Underlying allocator of BufferedAllocator must be unmanaged");
   if (underlying_allocator_->IsAllocThreadSafe()) {
     mtx_.reset(new std::mutex());
   }
@@ -41,35 +41,37 @@ void BufferedAllocator::FreeCache(size_t size) {
   while (!allocations_.empty()) {  // free the largest
     auto it = --allocations_.end();
     cur += it->second->size();
-    underlying_allocator_->Free(it->second.release());
+    delete it->second.release();
     allocations_.erase(it);
     if (cur >= size) return;
   }
 }
 
-bool BufferedAllocator::IsAllocThreadSafe() const { return mtx_ != nullptr; }
-
-void BufferedAllocator::FreeImpl(Allocation *allocation) {
+bool BufferedAllocator::IsAllocThreadSafe() const {
+  return this->underlying_allocator_->IsAllocThreadSafe();
+}
+void BufferedAllocator::Free(Allocation *allocation) {
   platform::LockGuardPtr<std::mutex> guard(mtx_);
   allocations_.emplace(allocation->size(), AllocationPtr(allocation));
 }
-
 Allocation *BufferedAllocator::AllocateImpl(size_t size, Allocator::Attr attr) {
   {
     platform::LockGuardPtr<std::mutex> guard(mtx_);
     auto it = allocations_.lower_bound(size);
     if (it != allocations_.end() && it->first < size * 2) {
       AllocationPtr result(std::move(it->second));
       allocations_.erase(it);
-      return result.release();
+      return new AllocationWithUnderlying(std::move(result));
     }
   }
 
   try {
-    return underlying_allocator_->Allocate(size, attr).release();
+    return new AllocationWithUnderlying(
+        underlying_allocator_->Allocate(size, attr));
   } catch (BadAlloc &) {
     FreeCache(size);
-    return underlying_allocator_->Allocate(size, attr).release();
+    return new AllocationWithUnderlying(
+        underlying_allocator_->Allocate(size, attr));
   }
 }