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_CLIENT_EXECUTABLE_BUILD_OPTIONS_H_ 17 #define TENSORFLOW_COMPILER_XLA_CLIENT_EXECUTABLE_BUILD_OPTIONS_H_ 18 19 #include "absl/strings/string_view.h" 20 #include "absl/types/optional.h" 21 #include "tensorflow/compiler/xla/service/computation_placer.h" 22 #include "tensorflow/compiler/xla/shape.h" 23 #include "tensorflow/compiler/xla/util.h" 24 #include "tensorflow/compiler/xla/xla.pb.h" 25 #include "tensorflow/compiler/xla/xla_data.pb.h" 26 #include "tensorflow/core/platform/threadpool.h" 27 28 namespace stream_executor { 29 30 // Forward-declared to avoid StreamExecutor dependency. 31 class DeviceMemoryAllocator; 32 33 } // namespace stream_executor 34 35 namespace xla { 36 37 // Class containing options for building an LocalExecutable with 38 // LocalClient::Compile. 39 class ExecutableBuildOptions { 40 public: 41 // If set, this is the device to build the computation for. Valid 42 // device_ordinal values are: 0 to # of devices - 1. These values are 43 // identical to the device ordinal values used by StreamExecutor. The built 44 // executable will be executable on any device equivalent to the specified 45 // device as determined by Backend::devices_equivalent(). A value of -1 46 // indicates this option has not been set. 47 ExecutableBuildOptions& set_device_ordinal(int device_ordinal); 48 int device_ordinal() const; 49 50 // If set, this specifies the layout of the result of the computation. If not 51 // set, the service will chose the layout of the result. A Shape is used to 52 // store the layout to accommodate tuple result shapes. A value of nullptr 53 // indicates the option has not been set. 54 ExecutableBuildOptions& set_result_layout(const Shape& shape_with_layout); 55 const Shape* result_layout() const; 56 57 // Expose access to the XLA debug options which will be passed to the 58 // compilation process. has_debug_options()59 bool has_debug_options() const { return debug_options_.has_value(); } debug_options()60 const DebugOptions& debug_options() const { return *debug_options_; } 61 DebugOptions* mutable_debug_options(); 62 63 // If set, this specifies an allocator that can be used to allocate temporary 64 // space on the device during compilation. For example, the compiler might 65 // want to run various algorithms on the device and pick the fastest one -- it 66 // might allocate buffers for use by these algorithms using this allocator. 67 // 68 // This does not need to be the same as the se::DeviceMemoryAllocator passed 69 // when running the executable. 70 ExecutableBuildOptions& set_device_allocator( 71 se::DeviceMemoryAllocator* allocator); 72 se::DeviceMemoryAllocator* device_allocator() const; 73 74 // Returns a string representation of the build options, suitable for 75 // debugging. 76 string ToString() const; 77 78 // The number of replicas of this computation that are to be executed. 79 // Defaults to 1. num_replicas()80 int num_replicas() const { return num_replicas_; } 81 ExecutableBuildOptions& set_num_replicas(int num_replicas); 82 83 // The number of partitions in this computation. Defaults to 1. num_partitions()84 int num_partitions() const { return num_partitions_; } 85 ExecutableBuildOptions& set_num_partitions(int num_partitions); 86 87 // Indicates whether to use SPMD (true) or MPMD (false) partitioning when 88 // num_partitions > 1 and XLA is requested to partition the input program. use_spmd_partitioning()89 bool use_spmd_partitioning() const { return use_spmd_partitioning_; } 90 ExecutableBuildOptions& set_use_spmd_partitioning(bool use_spmd_partitioning); 91 deduplicate_hlo()92 bool deduplicate_hlo() const { return deduplicate_hlo_; } 93 ExecutableBuildOptions& set_deduplicate_hlo(bool deduplicate_hlo); 94 95 // If set, this specifies a static device assignment for the computation. 96 // Otherwise, the computation will be compiled generically and can be run with 97 // any device assignment compatible with the computation's replica and 98 // partition counts. has_device_assignment()99 bool has_device_assignment() const { return device_assignment_.has_value(); } 100 ExecutableBuildOptions& set_device_assignment( 101 const DeviceAssignment& device_assignment); device_assignment()102 const DeviceAssignment& device_assignment() const { 103 CHECK(device_assignment_.has_value()); 104 return device_assignment_.value(); 105 } 106 107 // Whether input and output buffers are aliased if the associated parameter is 108 // passed-through XLA modules without being changed. alias_passthrough_params()109 bool alias_passthrough_params() const { return alias_passthrough_params_; } set_alias_passthrough_params(bool alias_passthrough_params)110 void set_alias_passthrough_params(bool alias_passthrough_params) { 111 alias_passthrough_params_ = alias_passthrough_params; 112 } 113 run_backend_only()114 bool run_backend_only() const { return run_backend_only_; } 115 // By default, XLA builds an executable by invoking standard compilation, i.e, 116 // running Compiler::Compile, or both Compiler::RunHloPasses and 117 // Compiler::RunBackend. When run_backend_only is set to true, XLA builds an 118 // executable by invoking only RunBackend and skip invoking RunHloPasses, 119 // which can be used to compile post-optimizations HLO modules. set_run_backend_only(bool run_backend_only)120 ExecutableBuildOptions& set_run_backend_only(bool run_backend_only) { 121 run_backend_only_ = run_backend_only; 122 return *this; 123 } 124 125 // Thread pool for parallel compilation. compile_thread_pool()126 tensorflow::thread::ThreadPool* compile_thread_pool() const { 127 return compile_thread_pool_; 128 } set_compile_thread_pool(tensorflow::thread::ThreadPool * compile_thread_pool)129 ExecutableBuildOptions& set_compile_thread_pool( 130 tensorflow::thread::ThreadPool* compile_thread_pool) { 131 compile_thread_pool_ = compile_thread_pool; 132 return *this; 133 } 134 135 private: 136 int device_ordinal_ = -1; 137 Shape result_layout_; 138 bool result_layout_set_ = false; 139 absl::optional<DebugOptions> debug_options_; 140 se::DeviceMemoryAllocator* device_allocator_ = nullptr; 141 int num_replicas_ = 1; 142 int num_partitions_ = 1; 143 bool use_spmd_partitioning_ = false; 144 bool deduplicate_hlo_ = false; 145 bool broadcast_replicated_params_ = false; 146 absl::optional<DeviceAssignment> device_assignment_; 147 bool alias_passthrough_params_ = false; 148 bool run_backend_only_ = false; 149 tensorflow::thread::ThreadPool* compile_thread_pool_ = nullptr; 150 }; 151 152 // Creates an ExecutionOptions based on a given ExecutableBuildOptions and 153 // ProgramShape. 154 ExecutionOptions CreateExecutionOptions( 155 const ExecutableBuildOptions& build_options, 156 const ProgramShape* program_shape); 157 158 } // namespace xla 159 160 #endif // TENSORFLOW_COMPILER_XLA_CLIENT_EXECUTABLE_BUILD_OPTIONS_H_ 161