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_JIT_XLA_TENSOR_H_ 17 #define TENSORFLOW_COMPILER_JIT_XLA_TENSOR_H_ 18 19 #include <memory> 20 21 #include "absl/memory/memory.h" 22 #include "tensorflow/compiler/xla/client/local_client.h" 23 #include "tensorflow/compiler/xla/service/shaped_buffer.h" 24 #include "tensorflow/core/framework/allocator.h" 25 #include "tensorflow/core/framework/device_base.h" 26 #include "tensorflow/core/lib/core/status.h" 27 #include "tensorflow/core/platform/mutex.h" 28 29 namespace tensorflow { 30 31 // The implementation of a Tensor for an XlaDevice. All device tensors are 32 // actually one of these. 33 // 34 // To distinguish between "normal" device tensors and XlaTensors, the raw 35 // pointer data stored in the TensorBuffer is a tagged pointer. 36 class XlaTensor { 37 public: 38 // Downcast from a Tensor to an XlaTensor. Return nullptr if the downcast 39 // fails. 40 static XlaTensor* FromTensor(const Tensor* tensor); 41 42 static bool RefCountIsOne(const Tensor& tensor); 43 44 // Create a DeviceMemoryBase from a Tensor. The Tensor can be an XlaTensor, in 45 // which case the returned value is shaped_buffer()->root_buffer(), or a 46 // normal Tensor in which case the returned value is 47 // {tensor.tensor_data().data(), tensor.tensor_data().size}. 48 static se::DeviceMemoryBase DeviceMemoryFromTensor(const Tensor& tensor); 49 50 // Assign the internal ShapedBuffer to new memory for the given dtype and 51 // shape. If a ShapedBuffer exists already (has_shaped_buffer() == true), it 52 // is replaced and the managed memory deallocated. 53 Status AllocateShapedBuffer(DataType dtype, const xla::Shape& on_host_shape, 54 xla::LocalClient* client, int device_ordinal); 55 56 // Some Tensors can have complex on-device shapes, including tuple shapes. To 57 // manage the memory for these tensors a ShapedBuffer may be required. 58 59 // Return true if this XlaTensor contains a ShapedBuffer. has_shaped_buffer()60 bool has_shaped_buffer() const { return shaped_buffer_ != nullptr; } 61 // Return the contained ShapedBuffer. 62 // REQUIRES: has_shaped_buffer() shaped_buffer()63 const xla::ShapedBuffer& shaped_buffer() const { 64 CHECK(has_shaped_buffer()); 65 return *shaped_buffer_; 66 } shaped_buffer()67 xla::ShapedBuffer& shaped_buffer() { 68 CHECK(has_shaped_buffer()); 69 return *shaped_buffer_; 70 } 71 // Mutates the XlaTensor to set the ShapedBuffer. set_shaped_buffer(xla::ScopedShapedBuffer shaped_buffer)72 void set_shaped_buffer(xla::ScopedShapedBuffer shaped_buffer) { 73 shaped_buffer_ = 74 absl::make_unique<xla::ScopedShapedBuffer>(std::move(shaped_buffer)); 75 } 76 77 // Some tensors on the device may have known values on the host. We use these 78 // in on-demand mode to avoid re-copying values from the device if we know the 79 // host value already. 80 81 // Return true if this XlaTensor contains a host tensor. has_host_tensor()82 bool has_host_tensor() const { return host_tensor_ != nullptr; } 83 // Return the contained host tensor. 84 // REQUIRES: has_host_tensor() host_tensor()85 const Tensor& host_tensor() const { return *host_tensor_; } 86 // Sets the contained host tensor. set_host_tensor(const Tensor & tensor)87 void set_host_tensor(const Tensor& tensor) { 88 host_tensor_.reset(new Tensor(tensor)); 89 } 90 91 // Adds synchronization events to 'stream' that wait for this tensor to be 92 // defined on 'stream'. Does nothing if the tensor is already defined on that 93 // stream. 94 void WaitForDefinitionEventOnStream(se::Stream* stream); 95 96 // (Re)sets the definition event of the tensor to 'event', and promises that 97 // the tensor has already been defined on stream. Removes any previous 98 // definition event or any previous promises about the tensor being defined on 99 // streams. 100 // It is legal to reset the definition event of a tensor when overwriting the 101 // tensor's value (at which point, it is effectively a new tensor once again.) 102 void ResetDefinitionEvent(std::shared_ptr<se::Event> event, 103 se::Stream* stream); 104 105 // Convert from a raw pointer to an XlaTensor, removing the pointer tag. 106 static XlaTensor* FromOpaquePointer(void* ptr); 107 // Convert to a raw pointer from an XlaTensor, adding the pointer tag. 108 static void* ToOpaquePointer(XlaTensor* tensor); 109 110 private: 111 // The optional contained ShapedBuffer. 112 std::unique_ptr<xla::ScopedShapedBuffer> shaped_buffer_; 113 // An optional host tensor value. 114 std::unique_ptr<Tensor> host_tensor_; 115 // An optional event that is triggered when the tensor's content has been 116 // defined. If this event is nullptr, it is assumed that the tensor's content 117 // is always defined. 118 std::shared_ptr<se::Event> definition_event_; 119 // A list of all streams for which the tensor's content is defined for any 120 // newly enqueued command. 121 absl::InlinedVector<se::Stream*, 2> streams_defined_on_ GUARDED_BY(mu_); 122 mutex mu_; 123 }; 124 125 } // namespace tensorflow 126 127 #endif // TENSORFLOW_COMPILER_JIT_XLA_TENSOR_H_ 128