1 /* Copyright 2017 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_CPU_CPU_COMPILER_H_ 17 #define TENSORFLOW_COMPILER_XLA_SERVICE_CPU_CPU_COMPILER_H_ 18 19 #include <memory> 20 21 #include "absl/types/span.h" 22 #include "llvm/Target/TargetMachine.h" 23 #include "tensorflow/compiler/tf2xla/cpu_function_runtime.h" 24 #include "tensorflow/compiler/xla/service/cpu/target_machine_features.h" 25 #include "tensorflow/compiler/xla/service/executable.h" 26 #include "tensorflow/compiler/xla/service/hlo_module.h" 27 #include "tensorflow/compiler/xla/service/llvm_compiler.h" 28 #include "tensorflow/compiler/xla/statusor.h" 29 #include "tensorflow/core/platform/macros.h" 30 #include "tensorflow/core/platform/stream_executor_no_cuda.h" 31 32 namespace xla { 33 namespace cpu { 34 35 // This class wraps the configurability options that LLVM exposes including: the 36 // target triple, the target cpu and the target features. It also includes the 37 // desired linkage name for the computation entry point. 38 class CpuAotCompilationOptions : public AotCompilationOptions { 39 public: 40 // Relocation models available for compilation. 41 enum class RelocationModel { 42 // Corresponds to the -fno-pic compiler option. 43 Static, 44 // Corresponds to the -fpic compiler option. 45 SmallPic, 46 // Corresponds to the -fPIC compiler option. 47 BigPic, 48 // Corresponds to the -fpie compiler option. 49 SmallPie, 50 // Corresponds to the -fPIE compiler option. 51 BigPie 52 }; 53 54 CpuAotCompilationOptions(string triple, string cpu_name, string features, 55 string entry_point_name, 56 RelocationModel relocation_model); 57 ~CpuAotCompilationOptions() override; 58 59 se::Platform::Id PlatformId() const override; 60 61 // The triple used for compilation, similar to clang's -target flag. triple()62 const string& triple() const { return triple_; } 63 // The CPU name used for compilation, similar to clang's -mcpu flag. cpu_name()64 const string& cpu_name() const { return cpu_name_; } 65 // The target features used for compilation ("+avx2", "+neon", etc). features()66 const string& features() const { return features_; } 67 // The name to be used for the compiled code's entry point. entry_point_name()68 const string& entry_point_name() const { return entry_point_name_; } 69 // The relocation model used for compilation. relocation_model()70 RelocationModel relocation_model() const { return relocation_model_; } 71 72 private: 73 const string triple_; 74 const string cpu_name_; 75 const string features_; 76 const string entry_point_name_; 77 const RelocationModel relocation_model_; 78 }; 79 80 class CpuAotCompilationResult : public AotCompilationResult { 81 public: 82 CpuAotCompilationResult( 83 ObjectFileData object_file_data, 84 std::vector<::tensorflow::cpu_function_runtime::BufferInfo> buffer_infos, 85 int64 result_buffer_index, 86 std::unique_ptr<HloProfilePrinterData> hlo_profile_printer_data); 87 ~CpuAotCompilationResult(); 88 hlo_profile_printer_data()89 HloProfilePrinterData* hlo_profile_printer_data() const { 90 return hlo_profile_printer_data_.get(); 91 } 92 object_file_data()93 const ObjectFileData& object_file_data() const { return object_file_data_; } 94 const std::vector<::tensorflow::cpu_function_runtime::BufferInfo>& buffer_infos()95 buffer_infos() const { 96 return buffer_infos_; 97 } result_buffer_index()98 int64 result_buffer_index() const { return result_buffer_index_; } 99 100 private: 101 // Contains the compiled computation: an object file. 102 const ObjectFileData object_file_data_; 103 104 // A list of BufferInfo objects describing the buffers used by the XLA 105 // computation. 106 const std::vector<::tensorflow::cpu_function_runtime::BufferInfo> 107 buffer_infos_; 108 109 // Contains which buffer index into |buffer_sizes| was designated to the 110 // result of the computation. This buffer should be passed into the output 111 // parameter when calling the compiled computation. 112 const int64 result_buffer_index_; 113 114 // Contains an instance of HloProfilePrinterData if HLO profiling is enabled, 115 // otherwise is nullptr. 116 std::unique_ptr<HloProfilePrinterData> hlo_profile_printer_data_; 117 }; 118 119 // CPU-targeting implementation of the XLA Compiler interface. 120 // 121 // The compiler translates XLA HLO code into LLVM IR and uses LLVM's JIT 122 // infrastructure to create an executable "blob" that can then be returned 123 // wrapped in CpuExecutable and actually invoked. 124 class CpuCompiler : public LLVMCompiler { 125 public: 126 CpuCompiler(); ~CpuCompiler()127 ~CpuCompiler() override {} 128 129 // Bring in 130 // StatusOr<std::vector<std::unique_ptr<Executable>>> Compile( 131 // std::vector<std::unique_ptr<HloModule>> modules, 132 // std::vector<std::vector<se::StreamExecutor*>> 133 // stream_execs) 134 using LLVMCompiler::Compile; 135 136 StatusOr<std::unique_ptr<HloModule>> RunHloPasses( 137 std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec, 138 DeviceMemoryAllocator* device_allocator) override; 139 140 StatusOr<std::unique_ptr<Executable>> RunBackend( 141 std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec, 142 DeviceMemoryAllocator* device_allocator) override; 143 144 StatusOr<std::vector<std::unique_ptr<AotCompilationResult>>> 145 CompileAheadOfTime(std::unique_ptr<HloModuleGroup> module_group, 146 const AotCompilationOptions& options) override; 147 148 se::Platform::Id PlatformId() const override; 149 150 HloCostAnalysis::ShapeSizeFunction ShapeSizeBytesFunction() const override; 151 152 private: 153 // Initialize the LLVM target. 154 static void InitializeLLVMTarget(); 155 156 // Runs the HLO passes which are necessary for both optimizations and 157 // correctness. 158 Status RunHloPasses(HloModule* module, bool is_aot_compile, 159 llvm::TargetMachine* target_machine); 160 161 // Runs HLO passes up to and including layout assignment. 162 Status RunHloPassesThroughLayoutAssn( 163 HloModule* module, bool /*is_aot_compile*/, 164 LLVMTargetMachineFeatures* target_machine_features); 165 166 // Runs HLO passes after layout assignment. 167 Status RunHloPassesAfterLayoutAssn( 168 HloModule* module, bool is_aot_compile, 169 LLVMTargetMachineFeatures* target_machine_features); 170 171 TF_DISALLOW_COPY_AND_ASSIGN(CpuCompiler); 172 }; 173 174 } // namespace cpu 175 } // namespace xla 176 177 #endif // TENSORFLOW_COMPILER_XLA_SERVICE_CPU_CPU_COMPILER_H_ 178