buffered_allocator.cc 6.4 KB
Newer Older
S
sneaxiy 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
// 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/memory/allocation/buffered_allocator.h"
#include <algorithm>
#include <limits>
#include <utility>

namespace paddle {
namespace memory {
namespace allocation {

BufferedAllocator::BufferedAllocator(std::unique_ptr<Allocator>&& allocator) {
  std::vector<size_t> division_plan(8 * sizeof(size_t));
  for (size_t i = 0; i < 8 * sizeof(size_t); ++i) {
    division_plan[i] = (static_cast<size_t>(1) << i);
  }
  InitAndEnforceCheck(std::move(allocator), division_plan);
}

BufferedAllocator::BufferedAllocator(std::unique_ptr<Allocator>&& allocator,
                                     const std::vector<size_t>& division_plan) {
  InitAndEnforceCheck(std::move(allocator), division_plan);
}

37 38 39
BufferedAllocator::~BufferedAllocator() { FlushImpl(); }

void BufferedAllocator::FlushImpl() {
S
sneaxiy 已提交
40 41 42 43
  for (auto& v : allocations_) {
    for (auto& pair : v) {
      underlying_allocator_->FreeUniquePtr(std::move(pair.second));
    }
44 45 46 47 48 49 50 51 52 53
    v.clear();
  }
}

void BufferedAllocator::Flush() {
  if (mtx_) {
    std::lock_guard<std::mutex> lock(*mtx_);
    FlushImpl();
  } else {
    FlushImpl();
S
sneaxiy 已提交
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
  }
}

void BufferedAllocator::InitAndEnforceCheck(
    std::unique_ptr<Allocator>&& allocator,
    const std::vector<size_t>& division_plan) {
  underlying_allocator_.reset(
      dynamic_cast<UnmanagedAllocator*>(allocator.release()));
  PADDLE_ENFORCE_NOT_NULL(
      underlying_allocator_,
      "Underlying allocator of BufferedAllocator must be unmanaged");
  if (underlying_allocator_->IsAllocThreadSafe()) {
    mtx_.reset(new std::mutex());
  }
  constexpr size_t kMax = std::numeric_limits<size_t>::max();
  if (division_plan.empty()) {
    division_plan_.assign({0, kMax});
  } else {
    auto from = division_plan.front() == 0 ? division_plan.begin() + 1
                                           : division_plan.begin();
    auto to = division_plan.back() == kMax ? division_plan.end() - 1
                                           : division_plan.end();
    division_plan_.reserve(to - from + 2);
    division_plan_.push_back(0);
    division_plan_.insert(division_plan_.end(), from, to);
    division_plan_.push_back(kMax);
    for (size_t i = 1; i < division_plan_.size(); ++i) {
      PADDLE_ENFORCE_LT(division_plan_[i - 1], division_plan_[i],
                        "Division plan must be strictly sorted");
    }
  }
  allocations_.resize(division_plan_.size() - 1);
}

void BufferedAllocator::InsertAllocationImpl(
    std::unique_ptr<Allocation>&& allocation) {
  auto size = allocation->size();
  auto idx = GetListIndex(size);
92
  allocations_[idx].emplace(size, std::move(allocation));
S
sneaxiy 已提交
93 94 95 96 97 98 99 100 101 102 103 104
}

void BufferedAllocator::InsertAllocation(
    std::unique_ptr<Allocation>&& allocation) {
  if (mtx_) {
    std::lock_guard<std::mutex> lock(*mtx_);
    InsertAllocationImpl(std::move(allocation));
  } else {
    InsertAllocationImpl(std::move(allocation));
  }
}

105 106
bool BufferedAllocator::Match(size_t actual_size, size_t requested_size) {
  return (actual_size >> 1) < requested_size;
S
sneaxiy 已提交
107 108 109 110 111 112 113 114 115 116 117 118 119 120
}

size_t BufferedAllocator::GetListIndex(size_t size) {
  auto it =
      std::upper_bound(division_plan_.begin(), division_plan_.end(), size);
  return static_cast<size_t>(it - division_plan_.begin()) - 1;
}

std::unique_ptr<Allocation> BufferedAllocator::RemoveAllocationImpl(
    size_t size) {
  auto idx = GetListIndex(size);
  auto& allocation_map = allocations_[idx];
  auto it = allocation_map.lower_bound(size);
  // Only remove allocation whose size is not more than twice of requested size
121 122 123 124 125 126 127 128
  if (it != allocation_map.end()) {
    if (Match(it->second->size(), size)) {
      auto ret = std::move(it->second);
      allocation_map.erase(it);
      return ret;
    } else {
      return nullptr;
    }
S
sneaxiy 已提交
129
  } else {
130 131 132 133 134 135 136 137 138 139 140 141 142
    while (++idx < allocations_.size() && Match(division_plan_[idx], size)) {
      auto& allocation_map = allocations_[idx];
      if (!allocation_map.empty()) {
        auto it = allocation_map.begin();
        if (Match(it->second->size(), size)) {
          auto ret = std::move(it->second);
          allocation_map.erase(it);
          return ret;
        } else {
          return nullptr;
        }
      }
    }
S
sneaxiy 已提交
143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200
    return nullptr;
  }
}

std::unique_ptr<Allocation> BufferedAllocator::RemoveAllocation(size_t size) {
  if (mtx_) {
    std::lock_guard<std::mutex> lock(*mtx_);
    return RemoveAllocationImpl(size);
  } else {
    return RemoveAllocationImpl(size);
  }
}

std::unique_ptr<Allocation> BufferedAllocator::Allocate(size_t size,
                                                        Allocator::Attr attr) {
  auto ret = RemoveAllocation(size);
  if (!ret) {
    try {
      return underlying_allocator_->Allocate(size, attr);
    } catch (BadAlloc&) {
      // if allocation failed, try to free some memorys from buffers
      FreeAllocations(size);
      return underlying_allocator_->Allocate(size, attr);
    }
  }
  return ret;
}

void BufferedAllocator::FreeAllocationsImpl(size_t size) {
  if (UNLIKELY(size == 0)) return;
  size_t cur = 0;
  for (auto& alloc_map : allocations_) {
    // use reverse iterator to free large allocations first
    while (!alloc_map.empty()) {
      auto it = --(alloc_map.end());
      cur += it->second->size();
      underlying_allocator_->FreeUniquePtr(std::move(it->second));
      alloc_map.erase(it);
      if (cur >= size) return;
    }
  }
}

void BufferedAllocator::FreeAllocations(size_t size) {
  if (mtx_) {
    std::lock_guard<std::mutex> lock(*mtx_);
    FreeAllocationsImpl(size);
  } else {
    FreeAllocationsImpl(size);
  }
}

void BufferedAllocator::FreeUniquePtr(std::unique_ptr<Allocation> allocation) {
  InsertAllocation(std::move(allocation));
}

bool BufferedAllocator::IsAllocThreadSafe() const { return mtx_ != nullptr; }

201 202 203 204
const std::vector<size_t>& BufferedAllocator::GetDivisionPlan() const {
  return division_plan_;
}

S
sneaxiy 已提交
205 206 207
}  // namespace allocation
}  // namespace memory
}  // namespace paddle