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 // Contains utilities for launching compiled XLA kernels for a KernelContext. 17 18 #ifndef TENSORFLOW_COMPILER_JIT_XLA_LAUNCH_UTIL_H_ 19 #define TENSORFLOW_COMPILER_JIT_XLA_LAUNCH_UTIL_H_ 20 21 #include "tensorflow/compiler/jit/xla_compilation_cache.h" 22 #include "tensorflow/compiler/jit/xla_tensor.h" 23 #include "tensorflow/compiler/tf2xla/xla_compiler.h" 24 #include "tensorflow/compiler/xla/client/local_client.h" 25 #include "tensorflow/compiler/xla/service/shaped_buffer.h" 26 #include "tensorflow/core/framework/allocation_description.pb.h" 27 #include "tensorflow/core/framework/resource_var.h" 28 #include "tensorflow/core/framework/tensor.h" 29 #include "tensorflow/core/framework/types.h" 30 #include "tensorflow/core/lib/core/status.h" 31 #include "tensorflow/core/lib/gtl/array_slice.h" 32 #include "tensorflow/core/platform/thread_annotations.h" 33 #include "tensorflow/stream_executor/device_memory_allocator.h" 34 35 namespace tensorflow { 36 37 // Snapshot of resource variables for a TF kernel invocation, mapping from 38 // parameter number to values at execution time. If the resource variable is not 39 // initialized, the value will not be present. 40 using ResourceVarsSnapshot = absl::flat_hash_map<int, absl::optional<Tensor>>; 41 42 // Information about the state of a variable passed as input to the _XlaCompile 43 // and _XlaRun operators. Unlocks the resource variable and decrements its 44 // refcount on destruction. 45 class VariableInfo { 46 public: 47 explicit VariableInfo(int index, absl::string_view name, Var* var); 48 VariableInfo(VariableInfo&& other); 49 50 VariableInfo& operator=(VariableInfo&& other); 51 52 VariableInfo(const VariableInfo&) = delete; 53 VariableInfo& operator=(const VariableInfo&) = delete; 54 55 // The index of the DT_RESOURCE input to the _XlaCompile/_XlaRun operator. 56 // Note that the indices can be different between _XlaCompile and _XlaRun. index()57 int index() const { return index_; } 58 59 // A pointer to the resource variable. May be null if this VariableInfo is 60 // "empty", i.e. it does not track a resource variable. var()61 Var* var() const { return var_; } 62 63 // Returns the variable name. name()64 absl::string_view name() const { return name_; } 65 66 // Returns true if the resource variable lock was successfully acquired by 67 // this thread. lock_held()68 bool lock_held() const { return lock_held_; } set_lock_held()69 void set_lock_held() { lock_held_ = true; } 70 71 ~VariableInfo(); 72 73 private: 74 int index_; 75 std::string name_; 76 Var* var_; 77 78 // We can't use a optional<mutex_lock> here because it confuses the compiler's 79 // thread safety analysis. Instead we use a boolean flag and release the lock 80 // in the VariableInfo destructor. 81 bool lock_held_ = false; 82 }; 83 84 // Creates a list of updated resource variables. 85 xla::StatusOr<std::vector<VariableInfo>> GatherVariableInfo( 86 OpKernelContext* ctx, 87 const XlaCompiler::CompilationResult& compilation_result, 88 int missing_ctx_input_prefix); 89 90 // Takes a snapshot of the values of resource variable arguments, whose indices 91 // are specified in `variable_indices` argument. We snapshot tensors that back 92 // resource variables since concurrent updates may modify the shape, and it is 93 // important that the shapes used for compilation match the true shapes of the 94 // buffers. 95 // 96 // We snapshot the entire set of resource variables as one atomic operation. 97 // This models Read->* dependencies between resource variable operations. See 98 // jit/resource_operation_safety_analysis for details. 99 Status SnapshotResourceVariables(OpKernelContext* ctx, 100 absl::Span<const int> variable_indices, 101 absl::Span<VariableInfo const> variable_infos, 102 ResourceVarsSnapshot* result); 103 104 // Acquires the mutexes for all the variables in `variables` using a 105 // deadlock-safe protocol (acquire the mutexes in increasing-address order). 106 // 107 // `variables` is allowed to contain instances that don't track a resource 108 // variable (i.e. variables[i].var() can be null for some i). 109 Status LockVariables(absl::Span<VariableInfo> variables) 110 TF_EXCLUSIVE_LOCK_FUNCTION(); 111 112 // Returns a vector of VariableInfo instances for the resource variable inputs, 113 // given that *all* inputs are in `inputs`. The input indices for the resource 114 // variable inputs are in `variable_indices`. 115 Status GetVariableInfosFromInputs(ResourceMgr* rm, DeviceBase* dev, 116 absl::Span<const Tensor* const> inputs, 117 absl::Span<const int> variable_indices, 118 std::vector<VariableInfo>* result); 119 120 // Returns pointers to inputs stored in `ctx`. 121 std::vector<const Tensor*> InputsFromContext(OpKernelContext* ctx); 122 123 // Helper class to perform the marshalling of TensorFlow inputs and outputs to 124 // ShapedBuffers suitable for passing to an XLA computation. 125 class XlaComputationLaunchContext { 126 public: 127 // Create a new launch context. 'allocate_xla_tensors' is true if allocated 128 // output tensors and variables are always XlaTensors. If false they are 129 // assumed to be "normal" device pointers. 130 // If 'use_multiple_streams' is true, tensors may be defined and used on 131 // multiple streams and so se::Events must be defined and waited for. If 132 // 'use_multiple_streams' is true, 'allocate_xla_tensors' must also be true 133 // because we track inter-stream dependencies through events inside XlaTensor 134 // objects. 135 XlaComputationLaunchContext(xla::LocalClient* client, 136 se::DeviceMemoryAllocator* xla_allocator, 137 int device_ordinal, bool allocate_xla_tensors, 138 bool use_multiple_streams); 139 140 // Builds a XlaCompiler::Argument vector from the arguments to an XlaLaunch 141 // op. 142 // Precondition: variables in `variable_args` are locked. 143 static xla::StatusOr<std::vector<XlaCompiler::Argument>> 144 BuildXlaCompilerArguments(absl::Span<int const> must_be_constant_idxs, 145 absl::Span<const Tensor* const> inputs, 146 absl::Span<VariableInfo const> variable_args, 147 Device* device); 148 149 // Add all inputs within `ctx` as XLA arguments (returned by arguments()). 150 // `variables` is a map from TensorFlow argument number to resource variable. 151 // 152 // Assumes that the first `missing_ctx_input_prefix` inputs to the kernel are 153 // missing and adjusts input indices accordingly. All elements in kernel's 154 // input_mapping must be greater than or equal to `missing_ctx_input_prefix` 155 // (in other words, no inputs actually required by the kernel can be missing). 156 xla::StatusOr<std::vector<xla::ExecutionInput>> PopulateInputs( 157 OpKernelContext* ctx, 158 const XlaCompiler::CompilationResult* compilation_result, 159 const std::map<int, const Tensor*>& resource_vars, 160 int missing_ctx_input_prefix, 161 const xla::HloInputOutputAliasConfig& input_output_alias); 162 163 // Given the XLA output in `output`, populate all outputs of `ctx`. Also 164 // writes out the resource variable updates. 165 // 166 // Updates to all resource variables are written in a single atomic operation. 167 // This models *->Write dependencies between resource variable operations. 168 // See jit/resource_operation_safety_analysis for details. 169 // 170 // 171 // Assumes that the first `missing_ctx_input_prefix` inputs to the 172 // compilation_result are missing and adjusts input indices accordingly. 173 Status PopulateOutputs( 174 OpKernelContext* ctx, 175 const XlaCompiler::CompilationResult* compilation_result, 176 xla::ScopedShapedBuffer output, int missing_ctx_input_prefix, 177 absl::Span<VariableInfo> variable_infos, 178 const xla::HloInputOutputAliasConfig& input_output_alias, 179 const std::map<int, const Tensor*>& resource_vars); 180 181 private: 182 xla::LocalClient* client_; 183 se::DeviceMemoryAllocator* xla_allocator_; 184 bool allocate_xla_tensors_; 185 bool use_multiple_streams_; 186 int device_ordinal_; 187 }; 188 189 // A simple TensorBuffer implementation that allows us to create Tensors that 190 // take ownership of pre-allocated memory. 191 class XlaTensorBuffer : public TensorBuffer { 192 public: XlaTensorBuffer(const void * ptr,size_t expected_size,size_t actual_size,Allocator * allocator)193 XlaTensorBuffer(const void* ptr, size_t expected_size, size_t actual_size, 194 Allocator* allocator) 195 : TensorBuffer(const_cast<void*>(ptr)), 196 expected_size_(expected_size), 197 actual_size_(actual_size), 198 allocator_(allocator) {} 199 ~XlaTensorBuffer()200 ~XlaTensorBuffer() override { 201 if (data()) { 202 allocator_->DeallocateRaw(data()); 203 } 204 } 205 size()206 size_t size() const override { return expected_size_; } 207 root_buffer()208 TensorBuffer* root_buffer() override { return this; } 209 FillAllocationDescription(AllocationDescription * proto)210 void FillAllocationDescription(AllocationDescription* proto) const override { 211 proto->set_allocated_bytes(actual_size_); 212 } 213 214 private: 215 size_t expected_size_; 216 size_t actual_size_; 217 Allocator* allocator_; 218 }; 219 220 } // namespace tensorflow 221 222 #endif // TENSORFLOW_COMPILER_JIT_XLA_LAUNCH_UTIL_H_ 223