split PR

test=develop

Author: sneaxiy

paddle/fluid/framework/CMakeLists.txt

@@ -202,8 +202,6 @@ cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
 
 cc_test(tuple_test SRCS tuple_test.cc )
 
-cc_test(inlined_vector_test SRCS inlined_vector_test.cc)
-
 if (NOT WIN32)
 cc_test(rw_lock_test SRCS rw_lock_test.cc)
 endif (NOT WIN32)

paddle/fluid/framework/inlined_vector.h

@@ -3,18 +3,9 @@ cc_library(cpu_allocator SRCS cpu_allocator.cc DEPS allocator)
 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(multi_bin_buffered_allocator SRCS multi_bin_buffered_allocator.cc DEPS allocator gflags)
 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)
-cc_test(multi_bin_buffered_allocator_test SRCS multi_bin_buffered_allocator_test.cc DEPS best_fit_allocator locked_allocator multi_bin_buffered_allocator cpu_allocator)
-
-cc_library(auto_growth_best_fit_allocator SRCS auto_growth_best_fit_allocator.cc DEPS allocator)
-cc_test(auto_growth_best_fit_allocator_test SRCS auto_growth_best_fit_allocator_test.cc DEPS cpu_allocator auto_growth_best_fit_allocator)
-
-if (NOT WIN32)
-  cc_test(test_multi_bin_buffered_allocator_division_plan SRCS test_multi_bin_buffered_allocator_division_plan.cc DEPS multi_bin_buffered_allocator)
-endif()
 
 if (WITH_GPU)
   nv_library(cuda_allocator SRCS cuda_allocator.cc DEPS allocator cuda_device_guard)
@@ -47,7 +38,7 @@ 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 multi_bin_buffered_allocator auto_growth_best_fit_allocator legacy_allocator zero_size_allocator)
+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)
@@ -59,8 +50,8 @@ nv_test(allocation_and_eigen_test SRCS allocation_and_eigen_test.cu DEPS allocat
 
 cc_test(retry_allocator_test SRCS retry_allocator_test.cc DEPS retry_allocator best_fit_allocator locked_allocator cpu_allocator)
 
-cc_test(allocator_facade_test SRCS allocator_facade_test.cc DEPS allocator_facade)
-
 cc_test(naive_best_fit_allocator_facade_test SRCS naive_best_fit_allocator_facade_test.cc DEPS allocator_facade)
 
-cc_test(auto_growth_best_fit_allocator_facade_test SRCS auto_growth_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/framework/inlined_vector_test.cc

@@ -17,7 +17,6 @@
 #include <string>
 #include <utility>
 #include <vector>
-#include "paddle/fluid/framework/inlined_vector.h"
 #include "paddle/fluid/platform/place.h"
 
 namespace paddle {
@@ -50,7 +49,9 @@ class Allocator;
 class Allocation {
  public:
   Allocation(void* ptr, size_t size, platform::Place place)
-      : ptr_(ptr), size_(size), place_(place) {}
+      : ptr_(ptr), size_(size), place_(place) {
+    decorated_allocators_.reserve(8);
+  }
 
   Allocation(const Allocation& o) = delete;
   Allocation& operator=(const Allocation& o) = delete;
@@ -80,8 +81,8 @@ class Allocation {
   virtual ~Allocation();
 
  private:
-  std::vector<Allocator*> DecoratedAllocators() const {
-    return static_cast<std::vector<Allocator*>>(decorated_allocators_);
+  const std::vector<Allocator*>& DecoratedAllocators() const {
+    return decorated_allocators_;
   }
 
   inline void RegisterDecoratedAllocator(Allocator* allocator) {
@@ -98,7 +99,7 @@ class Allocation {
   void* ptr_;
   size_t size_;
   platform::Place place_;
-  framework::InlinedVector<Allocator*, 8> decorated_allocators_;
+  std::vector<Allocator*> decorated_allocators_;
 
   friend class Allocator;
   friend class AllocationDeleter;

paddle/fluid/memory/allocation/CMakeLists.txt

@@ -22,14 +22,12 @@
 #include "paddle/fluid/memory/allocation/aligned_allocator.h"
 #include "paddle/fluid/memory/allocation/allocator_facade.h"
 #include "paddle/fluid/memory/allocation/allocator_strategy.h"
-#include "paddle/fluid/memory/allocation/auto_growth_best_fit_allocator.h"
 #include "paddle/fluid/memory/allocation/auto_increment_allocator.h"
 #include "paddle/fluid/memory/allocation/best_fit_allocator.h"
 #include "paddle/fluid/memory/allocation/conditional_allocator.h"
 #include "paddle/fluid/memory/allocation/cpu_allocator.h"
 #include "paddle/fluid/memory/allocation/legacy_allocator.h"
 #include "paddle/fluid/memory/allocation/locked_allocator.h"
-#include "paddle/fluid/memory/allocation/multi_bin_buffered_allocator.h"
 #include "paddle/fluid/memory/allocation/retry_allocator.h"
 #include "paddle/fluid/memory/allocation/zero_size_allocator.h"
 #include "paddle/fluid/platform/cpu_info.h"
@@ -47,24 +45,18 @@ DEFINE_int64(
     "The retry time (milliseconds) when allocator fails "
     "to allocate memory. No retry if this value is not greater than 0");
 
-DEFINE_bool(enable_buffered_allocator, false, "Enable buffered_allocator");
-
 namespace paddle {
 namespace memory {
 namespace allocation {
 
-static inline std::shared_ptr<Allocator> WrapRetryAndBufferedAllocator(
-    std::shared_ptr<Allocator> allocator, int64_t retry_time,
-    bool enable_buffered) {
+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);
   }
 
-  if (enable_buffered) {
-    allocator.reset(new MultiBinBufferedAllocator(allocator));
-  }
   return allocator;
 }
 
@@ -134,8 +126,7 @@ class ChunkedAllocator : public Allocator {
     std::shared_ptr<Allocator> allocator(new LockedAllocator(
         std::shared_ptr<Allocator>(new BestFitAllocator(allocation))));
 
-    allocator = WrapRetryAndBufferedAllocator(allocator, retry_time_,
-                                              FLAGS_enable_buffered_allocator);
+    allocator = WrapRetryAllocator(allocator, retry_time_);
 
     return std::make_shared<AlignedAllocator<4096>>(std::move(allocator));
   }
@@ -219,13 +210,6 @@ class AllocatorFacadePrivate {
         WrapZeroSizeAllocator();
         break;
       }
-      case AllocatorStrategy::kAutoGrowthBestFit: {
-        InitAutoGrowthCPUAllocator();
-        InitAutoGrowthCUDAAllocator();
-        InitAutoGrowthCUDAPinnedAllocator();
-        WrapZeroSizeAllocator();
-        break;
-      }
       default: {
         PADDLE_THROW("Unsupported allocator strategy: %d",
                      static_cast<int>(strategy));
@@ -234,39 +218,6 @@ class AllocatorFacadePrivate {
   }
 
  private:
-  void InitAutoGrowthCPUAllocator() {
-    auto cpu_allocator = std::make_shared<AlignedAllocator<4096>>(
-        std::make_shared<CPUAllocator>());
-    allocators_[platform::CPUPlace()] =
-        std::make_shared<AutoGrowthBestFitAllocator>(
-            cpu_allocator, platform::CpuMaxChunkSize(), 4096);
-  }
-
-  void InitAutoGrowthCUDAAllocator() {
-#ifdef PADDLE_WITH_CUDA
-    int dev_cnt = platform::GetCUDADeviceCount();
-    for (int dev_id = 0; dev_id < dev_cnt; ++dev_id) {
-      auto cuda_allocator = std::make_shared<AlignedAllocator<4096>>(
-          std::make_shared<CUDAAllocator>(platform::CUDAPlace(dev_id)));
-      auto allocator = std::make_shared<AutoGrowthBestFitAllocator>(
-          cuda_allocator, platform::GpuMaxChunkSize(), 4096);
-
-      allocators_[platform::CUDAPlace(dev_id)] = WrapRetryAndBufferedAllocator(
-          allocator, FLAGS_gpu_allocator_retry_time, false);
-    }
-#endif
-  }
-
-  void InitAutoGrowthCUDAPinnedAllocator() {
-#ifdef PADDLE_WITH_CUDA
-    auto cuda_pinned_allocator = std::make_shared<AlignedAllocator<4096>>(
-        std::make_shared<CPUPinnedAllocator>());
-    allocators_[platform::CUDAPinnedPlace()] =
-        std::make_shared<AutoGrowthBestFitAllocator>(
-            cuda_pinned_allocator, platform::CUDAPinnedMaxChunkSize(), 4096);
-#endif
-  }
-
   void InitLegacyAllocator() {
     std::vector<platform::Place> places{platform::CPUPlace()};
 #ifdef PADDLE_WITH_CUDA

paddle/fluid/memory/allocation/allocator.h

@@ -20,8 +20,7 @@ DEFINE_string(
     allocator_strategy, "legacy",
     "The allocation strategy. Legacy means the original allocator of Fluid."
     "naive_best_fit means the experimental best fit allocator. "
-    "auto_growth_best_fit means the experimental auto growth best fit "
-    "allocator. Enum in [legacy, naive_best_fit, auto_growth_best_fit].");
+    "allocator. Enum in [legacy, naive_best_fit].");
 
 namespace paddle {
 namespace memory {
@@ -32,8 +31,6 @@ static AllocatorStrategy GetStrategyFromFlag() {
     return AllocatorStrategy::kLegacy;
   } else if (FLAGS_allocator_strategy == "naive_best_fit") {
     return AllocatorStrategy::kNaiveBestFit;
-  } else if (FLAGS_allocator_strategy == "auto_growth_best_fit") {
-    return AllocatorStrategy::kAutoGrowthBestFit;
   } else {
     PADDLE_THROW("Unsupported allocator strategy: %s",
                  FLAGS_allocator_strategy);

paddle/fluid/memory/allocation/allocator_facade.cc

@@ -18,7 +18,7 @@ namespace paddle {
 namespace memory {
 namespace allocation {
 
-enum class AllocatorStrategy { kLegacy, kNaiveBestFit, kAutoGrowthBestFit };
+enum class AllocatorStrategy { kLegacy, kNaiveBestFit };
 
 extern AllocatorStrategy GetAllocatorStrategy();