| 25 | namespace xla { |
| 26 | |
| 27 | LocalDeviceState::LocalDeviceState(se::StreamExecutor* executor, |
| 28 | LocalClient* client, |
| 29 | bool synchronous_deallocation, |
| 30 | bool asynchronous, bool allow_event_reuse) |
| 31 | : synchronous_deallocation_(synchronous_deallocation), |
| 32 | event_pool_(allow_event_reuse), |
| 33 | compute_semaphore_(/*capacity=*/asynchronous ? 32 : 1), |
| 34 | executor_(executor), |
| 35 | client_(client), |
| 36 | prng_seed_generator_(prng_seed_device_()), |
| 37 | prng_seed_distribution_(std::numeric_limits<int>::min(), |
| 38 | std::numeric_limits<int>::max()) { |
| 39 | compute_stream_ = absl::make_unique<se::Stream>(executor); |
| 40 | host_to_device_stream_ = absl::make_unique<se::Stream>(executor); |
| 41 | callback_stream_ = absl::make_unique<se::Stream>(executor); |
| 42 | compute_stream_->Init(); |
| 43 | host_to_device_stream_->Init(); |
| 44 | callback_stream_->Init(); |
| 45 | device_to_host_streams_.reserve(kNumDeviceToHostStreams); |
| 46 | for (int i = 0; i < kNumDeviceToHostStreams; ++i) { |
| 47 | auto stream = absl::make_unique<se::Stream>(executor); |
| 48 | stream->Init(); |
| 49 | device_to_host_streams_.push_back(std::move(stream)); |
| 50 | } |
| 51 | device_to_device_streams_.reserve(kNumDeviceToDeviceStreams); |
| 52 | for (int i = 0; i < kNumDeviceToDeviceStreams; ++i) { |
| 53 | auto stream = absl::make_unique<se::Stream>(executor); |
| 54 | stream->Init(); |
| 55 | device_to_device_streams_.push_back(std::move(stream)); |
| 56 | } |
| 57 | execute_thread_ = absl::make_unique<WorkerThread>(tensorflow::Env::Default(), |
| 58 | "py_xla_execute"); |
| 59 | callback_thread_ = absl::make_unique<WorkerThread>(tensorflow::Env::Default(), |
| 60 | "py_xla_callback"); |
| 61 | } |
| 62 | |
| 63 | LocalDeviceState::~LocalDeviceState() { |
| 64 | Status status = SynchronizeAllActivity(); |