garbage_collector.cc 4.0 KB
Newer Older
S
sneaxiy 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
// 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 <algorithm>
S
sneaxiy 已提交
16 17 18 19 20
#include <deque>
#include <functional>
#include <memory>
#include <mutex>  // NOLINT
#include <utility>
S
sneaxiy 已提交
21 22 23
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
S
sneaxiy 已提交
24
#include "gflags/gflags.h"
S
sneaxiy 已提交
25 26 27 28 29
#include "paddle/fluid/framework/garbage_collector.h"

namespace paddle {
namespace framework {

S
sneaxiy 已提交
30 31 32 33 34 35 36 37 38
DEFINE_double(
    eager_delete_tensor_gb, -1.0,
    "Memory size threshold (GB) when the garbage collector clear tensors."
    "Disabled when this value is less than 0");

DEFINE_bool(fast_eager_deletion_mode, true,
            "Fast eager deletion mode. If enabled, memory would release "
            "immediately without waiting GPU kernel ends.");

S
sneaxiy 已提交
39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 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 92 93 94 95 96 97 98 99 100 101 102
GarbageCollector::GarbageCollector(const platform::Place &place,
                                   size_t max_memory_size)
    : max_memory_size_((std::max)(max_memory_size, static_cast<size_t>(1))) {
  garbages_.reset(new GarbageQueue());
  dev_ctx_ = platform::DeviceContextPool::Instance().Get(place);
}

CPUGarbageCollector::CPUGarbageCollector(const platform::CPUPlace &place,
                                         size_t max_memory_size)
    : GarbageCollector(place, max_memory_size) {}

void CPUGarbageCollector::ClearCallback(const std::function<void()> &callback) {
  callback();
}

#ifdef PADDLE_WITH_CUDA
UnsafeFastGPUGarbageCollector::UnsafeFastGPUGarbageCollector(
    const platform::CUDAPlace &place, size_t max_memory_size)
    : GarbageCollector(place, max_memory_size) {}

void UnsafeFastGPUGarbageCollector::ClearCallback(
    const std::function<void()> &callback) {
  callback();
}

DefaultStreamGarbageCollector::DefaultStreamGarbageCollector(
    const platform::CUDAPlace &place, size_t max_memory_size)
    : GarbageCollector(place, max_memory_size) {}

void DefaultStreamGarbageCollector::Wait() const {
  static_cast<platform::CUDADeviceContext *>(this->dev_ctx_)
      ->WaitStreamCallback();
}

void DefaultStreamGarbageCollector::ClearCallback(
    const std::function<void()> &callback) {
  static_cast<platform::CUDADeviceContext *>(this->dev_ctx_)
      ->AddStreamCallback(callback);
}

StreamGarbageCollector::StreamGarbageCollector(const platform::CUDAPlace &place,
                                               size_t max_memory_size)
    : GarbageCollector(place, max_memory_size) {
  platform::CUDADeviceGuard guard(place.device);
  PADDLE_ENFORCE(cudaStreamCreate(&stream_));
  callback_manager_.reset(new platform::StreamCallbackManager(stream_));
}

StreamGarbageCollector::~StreamGarbageCollector() {
  auto place = boost::get<platform::CUDAPlace>(this->dev_ctx_->GetPlace());
  platform::CUDADeviceGuard guard(place.device);
  PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
  PADDLE_ENFORCE(cudaStreamDestroy(stream_));
}

cudaStream_t StreamGarbageCollector::stream() const { return stream_; }

void StreamGarbageCollector::Wait() const { callback_manager_->Wait(); }

void StreamGarbageCollector::ClearCallback(
    const std::function<void()> &callback) {
  callback_manager_->AddCallback(callback);
}
#endif
S
sneaxiy 已提交
103 104 105 106 107 108 109 110 111 112 113

void UseGarbageCollectorGFlags() {}

int64_t GetEagerDeletionThreshold() {
  return FLAGS_eager_delete_tensor_gb < 0
             ? -1
             : static_cast<int64_t>(FLAGS_eager_delete_tensor_gb *
                                    (static_cast<int64_t>(1) << 30));
}

bool IsFastEagerDeletionModeEnabled() { return FLAGS_fast_eager_deletion_mode; }
S
sneaxiy 已提交
114 115
}  // namespace framework
}  // namespace paddle