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_HLO_MODULE_CONFIG_H_ 17 #define TENSORFLOW_COMPILER_XLA_SERVICE_HLO_MODULE_CONFIG_H_ 18 19 #include <string> 20 21 #include "tensorflow/compiler/xla/service/computation_layout.h" 22 #include "tensorflow/compiler/xla/types.h" 23 #include "tensorflow/compiler/xla/xla.pb.h" 24 #include "tensorflow/compiler/xla/xla_data.pb.h" 25 #include "tensorflow/core/lib/gtl/optional.h" 26 27 namespace xla { 28 29 // This class gathers all settings and values which affect the compiled 30 // executable outside of the HLO code itself. This include layouts of inputs and 31 // outputs to the module and settings such as HLO profiling. Together the 32 // HloModule and HloModuleConfig unambiguously determine a particular 33 // executable. 34 class HloModuleConfig { 35 public: 36 // A configuration can be created either with, or without an entry 37 // ComputationLayout. The default ctor creates it without -- in this case 38 // accessing entry_computation_layout will CHECK-fail. The ctor accepting a 39 // ProgramShape creates a computation layout using this shape. 40 HloModuleConfig(); 41 explicit HloModuleConfig(const ProgramShape& program_shape); 42 43 // Checks if this config has an entry computation layout already. has_entry_computation_layout()44 bool has_entry_computation_layout() const { 45 return entry_computation_layout_.has_value(); 46 } 47 48 // Sets the entry computation layout for this config. If the entry computation 49 // layout already exists, it is silently replaced. 50 void SetDefaultComputationLayout(const ProgramShape& program_shape); 51 52 // Returns a constant reference to the layout of the entry computation. 53 // Assumes the layout was set. entry_computation_layout()54 const ComputationLayout& entry_computation_layout() const { 55 CHECK(entry_computation_layout_.has_value()); 56 return *entry_computation_layout_; 57 } 58 59 // Returns a mutable pointer to the layout of the entry computation. Assumes 60 // the layout was set. mutable_entry_computation_layout()61 ComputationLayout* mutable_entry_computation_layout() { 62 CHECK(entry_computation_layout_.has_value()); 63 return &(*entry_computation_layout_); 64 } 65 66 // Sets/returns whether to enable HLO-level profiling. hlo_profiling_enabled()67 bool hlo_profiling_enabled() const { return hlo_profiling_enabled_; } enable_hlo_profiling(bool enabled)68 void enable_hlo_profiling(bool enabled) { hlo_profiling_enabled_ = enabled; } 69 70 // Sets/returns the module seed set during execution. set_seed(uint64 seed)71 void set_seed(uint64 seed) { seed_ = seed; } seed()72 uint64 seed() const { return seed_; } 73 set_replica_count(int64 replica_count)74 void set_replica_count(int64 replica_count) { 75 replica_count_ = replica_count; 76 } replica_count()77 int64 replica_count() const { return replica_count_; } 78 79 // Return a string which unambiguously represents all the fields of this data 80 // structure. Used for generating a cache key for storing the compiled 81 // executable. 82 string compilation_cache_key() const; 83 debug_options()84 const DebugOptions& debug_options() const { return debug_options_; } 85 set_debug_options(const DebugOptions & debug_options)86 void set_debug_options(const DebugOptions& debug_options) { 87 debug_options_ = debug_options; 88 } 89 90 // Sets/returns the number of intra op threads for this module. set_intra_op_parallelism_threads(const int intra_op_parallelism_threads)91 void set_intra_op_parallelism_threads( 92 const int intra_op_parallelism_threads) { 93 intra_op_parallelism_threads_ = intra_op_parallelism_threads; 94 } intra_op_parallelism_threads()95 int64 intra_op_parallelism_threads() const { 96 return intra_op_parallelism_threads_; 97 } 98 99 private: 100 // If you add new members, be sure to update compilation_cache_key. 101 102 tensorflow::gtl::optional<ComputationLayout> entry_computation_layout_; 103 104 // Whether to enable HLO-level profiling. 105 bool hlo_profiling_enabled_ = false; 106 107 // Module/graph-level seed handle. 108 uint64 seed_ = 0; 109 110 // The number of replicas to compile this binary for. 111 int64 replica_count_ = 1; 112 113 // The target maximum parallelism at which to partition HLOs for parallel 114 // execution on the CPU backend. 115 int64 intra_op_parallelism_threads_ = -1; 116 117 DebugOptions debug_options_; 118 }; 119 120 } // namespace xla 121 122 #endif // TENSORFLOW_COMPILER_XLA_SERVICE_HLO_MODULE_CONFIG_H_ 123