1 /* Copyright 2015 The TensorFlow Authors. All Rights Reserved. 2 3 Licensed under the Apache License, Version 2.0 (the "License"); 4 you may not use this file except in compliance with the License. 5 You may obtain a copy of the License at 6 7 http://www.apache.org/licenses/LICENSE-2.0 8 9 Unless required by applicable law or agreed to in writing, software 10 distributed under the License is distributed on an "AS IS" BASIS, 11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 See the License for the specific language governing permissions and 13 limitations under the License. 14 ==============================================================================*/ 15 16 #include "tensorflow/core/framework/tracking_allocator.h" 17 18 #include "tensorflow/core/platform/env.h" 19 #include "tensorflow/core/platform/logging.h" 20 21 namespace tensorflow { 22 23 TrackingAllocator::TrackingAllocator(Allocator* allocator, bool track_sizes) 24 : allocator_(allocator), 25 ref_(1), 26 allocated_(0), 27 high_watermark_(0), 28 total_bytes_(0), 29 track_sizes_locally_(track_sizes && !allocator_->TracksAllocationSizes()), 30 next_allocation_id_(0) {} 31 32 void* TrackingAllocator::AllocateRaw( 33 size_t alignment, size_t num_bytes, 34 const AllocationAttributes& allocation_attr) { 35 void* ptr = allocator_->AllocateRaw(alignment, num_bytes, allocation_attr); 36 // If memory is exhausted AllocateRaw returns nullptr, and we should 37 // pass this through to the caller 38 if (nullptr == ptr) { 39 return ptr; 40 } 41 if (allocator_->TracksAllocationSizes()) { 42 size_t allocated_bytes = allocator_->AllocatedSize(ptr); 43 { 44 mutex_lock lock(mu_); 45 allocated_ += allocated_bytes; 46 high_watermark_ = std::max(high_watermark_, allocated_); 47 total_bytes_ += allocated_bytes; 48 allocations_.emplace_back(allocated_bytes, Env::Default()->NowMicros()); 49 ++ref_; 50 } 51 } else if (track_sizes_locally_) { 52 // Call the underlying allocator to try to get the allocated size 53 // whenever possible, even when it might be slow. If this fails, 54 // use the requested size as an approximation. 55 size_t allocated_bytes = allocator_->AllocatedSizeSlow(ptr); 56 allocated_bytes = std::max(num_bytes, allocated_bytes); 57 mutex_lock lock(mu_); 58 next_allocation_id_ += 1; 59 Chunk chunk = {num_bytes, allocated_bytes, next_allocation_id_}; 60 in_use_.emplace(std::make_pair(ptr, chunk)); 61 allocated_ += allocated_bytes; 62 high_watermark_ = std::max(high_watermark_, allocated_); 63 total_bytes_ += allocated_bytes; 64 allocations_.emplace_back(allocated_bytes, Env::Default()->NowMicros()); 65 ++ref_; 66 } else { 67 mutex_lock lock(mu_); 68 total_bytes_ += num_bytes; 69 allocations_.emplace_back(num_bytes, Env::Default()->NowMicros()); 70 ++ref_; 71 } 72 return ptr; 73 } 74 75 void TrackingAllocator::DeallocateRaw(void* ptr) { 76 // freeing a null ptr is a no-op 77 if (nullptr == ptr) { 78 return; 79 } 80 bool should_delete; 81 // fetch the following outside the lock in case the call to 82 // AllocatedSize is slow 83 bool tracks_allocation_sizes = allocator_->TracksAllocationSizes(); 84 size_t allocated_bytes = 0; 85 if (tracks_allocation_sizes) { 86 allocated_bytes = allocator_->AllocatedSize(ptr); 87 } else if (track_sizes_locally_) { 88 mutex_lock lock(mu_); 89 auto itr = in_use_.find(ptr); 90 if (itr != in_use_.end()) { 91 tracks_allocation_sizes = true; 92 allocated_bytes = (*itr).second.allocated_size; 93 in_use_.erase(itr); 94 } 95 } 96 Allocator* allocator = allocator_; 97 { 98 mutex_lock lock(mu_); 99 if (tracks_allocation_sizes) { 100 CHECK_GE(allocated_, allocated_bytes); 101 allocated_ -= allocated_bytes; 102 allocations_.emplace_back(-allocated_bytes, Env::Default()->NowMicros()); 103 } 104 should_delete = UnRef(); 105 } 106 allocator->DeallocateRaw(ptr); 107 if (should_delete) { 108 delete this; 109 } 110 } 111 112 bool TrackingAllocator::TracksAllocationSizes() { 113 return track_sizes_locally_ || allocator_->TracksAllocationSizes(); 114 } 115 116 size_t TrackingAllocator::RequestedSize(const void* ptr) { 117 if (track_sizes_locally_) { 118 mutex_lock lock(mu_); 119 auto it = in_use_.find(ptr); 120 if (it != in_use_.end()) { 121 return (*it).second.requested_size; 122 } 123 return 0; 124 } else { 125 return allocator_->RequestedSize(ptr); 126 } 127 } 128 129 size_t TrackingAllocator::AllocatedSize(const void* ptr) { 130 if (track_sizes_locally_) { 131 mutex_lock lock(mu_); 132 auto it = in_use_.find(ptr); 133 if (it != in_use_.end()) { 134 return (*it).second.allocated_size; 135 } 136 return 0; 137 } else { 138 return allocator_->AllocatedSize(ptr); 139 } 140 } 141 142 int64 TrackingAllocator::AllocationId(const void* ptr) { 143 if (track_sizes_locally_) { 144 mutex_lock lock(mu_); 145 auto it = in_use_.find(ptr); 146 if (it != in_use_.end()) { 147 return (*it).second.allocation_id; 148 } 149 return 0; 150 } else { 151 return allocator_->AllocationId(ptr); 152 } 153 } 154 155 void TrackingAllocator::GetStats(AllocatorStats* stats) { 156 allocator_->GetStats(stats); 157 } 158 159 void TrackingAllocator::ClearStats() { allocator_->ClearStats(); } 160 161 std::tuple<size_t, size_t, size_t> TrackingAllocator::GetSizes() { 162 size_t high_watermark; 163 size_t total_bytes; 164 size_t still_live_bytes; 165 { 166 mutex_lock lock(mu_); 167 high_watermark = high_watermark_; 168 total_bytes = total_bytes_; 169 still_live_bytes = allocated_; 170 } 171 return std::make_tuple(total_bytes, high_watermark, still_live_bytes); 172 } 173 174 gtl::InlinedVector<AllocRecord, 4> TrackingAllocator::GetRecordsAndUnRef() { 175 bool should_delete; 176 gtl::InlinedVector<AllocRecord, 4> allocations; 177 { 178 mutex_lock lock(mu_); 179 allocations.swap(allocations_); 180 should_delete = UnRef(); 181 } 182 if (should_delete) { 183 delete this; 184 } 185 return allocations; 186 } 187 188 gtl::InlinedVector<AllocRecord, 4> TrackingAllocator::GetCurrentRecords() { 189 gtl::InlinedVector<AllocRecord, 4> allocations; 190 { 191 mutex_lock lock(mu_); 192 for (const AllocRecord& alloc : allocations_) { 193 allocations.push_back(alloc); 194 } 195 } 196 return allocations; 197 } 198 199 bool TrackingAllocator::UnRef() { 200 CHECK_GE(ref_, 1); 201 --ref_; 202 return (ref_ == 0); 203 } 204 205 } // end namespace tensorflow 206
