1 /* Copyright 2018 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 #ifndef TENSORFLOW_COMPILER_XLA_SERVICE_GPU_XFEED_QUEUE_H_ 17 #define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_XFEED_QUEUE_H_ 18 19 #include <deque> 20 #include <functional> 21 #include <utility> 22 #include <vector> 23 24 #include "absl/base/thread_annotations.h" 25 #include "tensorflow/core/platform/logging.h" 26 27 namespace xla { 28 namespace gpu { 29 30 // TODO(b/30467474) Once GPU outfeed implementation settles, consider 31 // folding back the cpu and gpu outfeed implementations into a generic 32 // one if possible. 33 34 // Manages a thread-safe queue of buffers. 35 template <typename BufferType> 36 class XfeedQueue { 37 public: 38 // Adds a tree of buffers to the queue. The individual buffers correspond to 39 // the elements of a tuple and may be nullptr if the buffer is a tuple index 40 // buffer. EnqueueDestination(BufferType buffers)41 void EnqueueDestination(BufferType buffers) { 42 absl::MutexLock l(&mu_); 43 enqueued_buffers_.push_back(std::move(buffers)); 44 enqueue_cv_.Signal(); 45 46 EnqueueHook(); 47 } 48 49 // Blocks until the queue is non-empty, then returns the buffer at the head of 50 // the queue. BlockingGetNextDestination()51 BufferType BlockingGetNextDestination() { 52 for (const auto& callback : before_get_next_dest_callbacks_) { 53 callback(); 54 } 55 56 bool became_empty; 57 BufferType current_buffer; 58 { 59 absl::MutexLock l(&mu_); 60 while (enqueued_buffers_.empty()) { 61 enqueue_cv_.Wait(&mu_); 62 } 63 current_buffer = std::move(enqueued_buffers_.front()); 64 enqueued_buffers_.pop_front(); 65 DequeueHook(); 66 became_empty = enqueued_buffers_.empty(); 67 } 68 if (became_empty) { 69 for (const auto& callback : on_empty_callbacks_) { 70 callback(); 71 } 72 } 73 return current_buffer; 74 } 75 RegisterOnEmptyCallback(std::function<void ()> callback)76 void RegisterOnEmptyCallback(std::function<void()> callback) { 77 on_empty_callbacks_.push_back(std::move(callback)); 78 } RegisterBeforeGetNextDestinationCallback(std::function<void ()> callback)79 void RegisterBeforeGetNextDestinationCallback( 80 std::function<void()> callback) { 81 before_get_next_dest_callbacks_.push_back(std::move(callback)); 82 } 83 ~XfeedQueue()84 virtual ~XfeedQueue() {} 85 86 protected: DequeueHook()87 virtual void DequeueHook() ABSL_EXCLUSIVE_LOCKS_REQUIRED(this->mu_) {} EnqueueHook()88 virtual void EnqueueHook() ABSL_EXCLUSIVE_LOCKS_REQUIRED(this->mu_) {} 89 90 absl::Mutex mu_; 91 92 // The queue of trees of buffers. Buffer* queue contents are not owned. 93 std::deque<BufferType> enqueued_buffers_ ABSL_GUARDED_BY(mu_); 94 95 private: 96 // Condition variable that is signaled every time a buffer is enqueued. 97 absl::CondVar enqueue_cv_; 98 99 // List of callbacks which will be called when 'enqueued_buffers_' becomes 100 // empty. 101 std::vector<std::function<void()>> on_empty_callbacks_; 102 103 // List of callbacks which will be called before BlockingGetNextDestination() 104 // is called. This lets you e.g. call EnqueueDestination() for each call to 105 // BlockingGetNextDestination(). 106 std::vector<std::function<void()>> before_get_next_dest_callbacks_; 107 }; 108 109 // Like XfeedQueue but with a maximum capacity. Clients can call 110 // `BlockUntilEnqueueSlotAvailable` to block until there are fewer than 111 // `max_pending_xfeeds_` capacity pending infeed items. 112 // 113 // We introduce a separate `BlockUntilEnqueueSlotAvailable` (as opposed to 114 // overriding `EnqueueDestination` to block) because we want to block before we 115 // copy the buffer to GPU memory, in order to bound the memory consumption due 116 // to pending infeeds. 117 template <typename BufferType> 118 class BlockingXfeedQueue : public XfeedQueue<BufferType> { 119 public: BlockingXfeedQueue(int max_pending_xfeeds)120 explicit BlockingXfeedQueue(int max_pending_xfeeds) 121 : max_pending_xfeeds_(max_pending_xfeeds) {} 122 BlockUntilEnqueueSlotAvailable()123 void BlockUntilEnqueueSlotAvailable() { 124 absl::MutexLock l{&this->mu_}; 125 while (pending_buffers_ + this->enqueued_buffers_.size() >= 126 max_pending_xfeeds_) { 127 VLOG(2) << "Capacity " 128 << (pending_buffers_ + this->enqueued_buffers_.size()) 129 << " >= max capacity " << max_pending_xfeeds_; 130 dequeue_cv_.Wait(&this->mu_); 131 } 132 133 pending_buffers_++; 134 } 135 136 protected: EnqueueHook()137 void EnqueueHook() ABSL_EXCLUSIVE_LOCKS_REQUIRED(this->mu_) override { 138 pending_buffers_--; 139 } 140 DequeueHook()141 void DequeueHook() ABSL_EXCLUSIVE_LOCKS_REQUIRED(this->mu_) override { 142 dequeue_cv_.Signal(); 143 } 144 145 private: 146 const int max_pending_xfeeds_; 147 148 // Condition variable that is signaled every time a buffer is dequeued. 149 absl::CondVar dequeue_cv_; 150 151 // Keeps track of the number of buffers reserved but not added to 152 // enqueued_buffers_. 153 int pending_buffers_ ABSL_GUARDED_BY(this->mu_) = 0; 154 }; 155 156 } // namespace gpu 157 } // namespace xla 158 159 #endif // TENSORFLOW_COMPILER_XLA_SERVICE_GPU_XFEED_QUEUE_H_ 160