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_COST_ANALYSIS_H_ 17 #define TENSORFLOW_COMPILER_XLA_SERVICE_HLO_COST_ANALYSIS_H_ 18 19 #include "absl/types/span.h" 20 #include "tensorflow/compiler/xla/service/dfs_hlo_visitor.h" 21 #include "tensorflow/compiler/xla/service/hlo_computation.h" 22 #include "tensorflow/compiler/xla/service/hlo_instruction.h" 23 #include "tensorflow/compiler/xla/service/hlo_opcode.h" 24 #include "tensorflow/compiler/xla/shape_util.h" 25 #include "tensorflow/compiler/xla/statusor.h" 26 #include "tensorflow/compiler/xla/xla_data.pb.h" 27 #include "tensorflow/core/platform/macros.h" 28 #include "tensorflow/core/platform/types.h" 29 30 namespace xla { 31 32 // HloCostAnalysis traverses an HLO graph and calculates the amount of 33 // computations required for the graph. Each HLO instruction handler provides 34 // the computation cost of the instruction, and the values are accumulated 35 // during the traversal for the entire graph. We treat normal floating point 36 // operations separately from transcendental operations. 37 class HloCostAnalysis : public ConstDfsHloVisitor { 38 public: 39 // Each HLO is associated to a vector of properties with the indices given 40 // below. Sub-classes can add further properties. 41 typedef std::map<string, float> Properties; 42 static constexpr char kFlopsKey[] = "flops"; 43 static constexpr char kTranscendentalsKey[] = "transcendentals"; 44 static constexpr char kBytesAccessedKey[] = "bytes accessed"; 45 static constexpr char kOptimalSecondsKey[] = "optimal_seconds"; 46 47 // shape_size is a function which returns the size in bytes of the top-level 48 // buffer of a shape. 49 using ShapeSizeFunction = std::function<int64(const Shape&)>; 50 explicit HloCostAnalysis(const ShapeSizeFunction& shape_size); 51 52 Status HandleElementwiseUnary(const HloInstruction* hlo) override; 53 Status HandleElementwiseBinary(const HloInstruction* hlo) override; 54 Status HandleConstant(const HloInstruction* constant) override; 55 Status HandleIota(const HloInstruction* iota) override; 56 Status HandleGetTupleElement( 57 const HloInstruction* get_tuple_element) override; 58 Status HandleSelect(const HloInstruction* hlo) override; 59 Status HandleTupleSelect(const HloInstruction* hlo) override; 60 Status HandleCompare(const HloInstruction* compare) override; 61 Status HandleClamp(const HloInstruction* clamp) override; 62 Status HandleReducePrecision(const HloInstruction* hlo) override; 63 Status HandleConcatenate(const HloInstruction* concatenate) override; 64 Status HandleSend(const HloInstruction* send) override; 65 Status HandleSendDone(const HloInstruction* send_done) override; 66 Status HandleRecv(const HloInstruction* recv) override; 67 Status HandleRecvDone(const HloInstruction* recv_done) override; 68 Status HandleConvert(const HloInstruction* convert) override; 69 Status HandleCopy(const HloInstruction* copy) override; 70 Status HandleDomain(const HloInstruction* domain) override; 71 Status HandleDot(const HloInstruction* dot) override; 72 Status HandleConvolution(const HloInstruction* convolution) override; 73 Status HandleFft(const HloInstruction* fft) override; 74 Status HandleTriangularSolve(const HloInstruction* hlo) override; 75 Status HandleCholesky(const HloInstruction* hlo) override; 76 Status HandleAllReduce(const HloInstruction* crs) override; 77 Status HandleAllToAll(const HloInstruction* hlo) override; 78 Status HandleCollectivePermute(const HloInstruction* hlo) override; 79 Status HandleReplicaId(const HloInstruction* hlo) override; 80 Status HandleInfeed(const HloInstruction* infeed) override; 81 Status HandleOutfeed(const HloInstruction* outfeed) override; 82 Status HandleRng(const HloInstruction* random) override; 83 Status HandleReverse(const HloInstruction* reverse) override; 84 Status HandleSort(const HloInstruction* sort) override; 85 Status HandleParameter(const HloInstruction* parameter) override; 86 Status HandleReduce(const HloInstruction* reduce) override; 87 Status HandleBatchNormTraining( 88 const HloInstruction* batch_norm_training) override; 89 Status HandleBatchNormInference( 90 const HloInstruction* batch_norm_inference) override; 91 Status HandleBatchNormGrad(const HloInstruction* batch_norm_grad) override; 92 Status HandleFusion(const HloInstruction* fusion) override; 93 Status HandleCall(const HloInstruction* call) override; 94 Status HandleCustomCall(const HloInstruction* custom_call) override; 95 Status HandleSlice(const HloInstruction* slice) override; 96 Status HandleDynamicSlice(const HloInstruction* dynamic_slice) override; 97 Status HandleDynamicUpdateSlice( 98 const HloInstruction* dynamic_update_slice) override; 99 Status HandleTuple(const HloInstruction* tuple) override; 100 Status HandleMap(const HloInstruction* map) override; 101 Status HandleReduceWindow(const HloInstruction* reduce_window) override; 102 Status HandleSelectAndScatter(const HloInstruction* instruction) override; 103 Status HandleBitcast(const HloInstruction* bitcast) override; 104 Status HandleBroadcast(const HloInstruction* broadcast) override; 105 Status HandlePad(const HloInstruction* pad) override; 106 Status HandleReshape(const HloInstruction* reshape) override; 107 Status HandleAddDependency(const HloInstruction* add_dependency) override; 108 Status HandleAfterAll(const HloInstruction* token) override; 109 Status HandleTranspose(const HloInstruction* transpose) override; 110 Status HandleWhile(const HloInstruction* xla_while) override; 111 Status HandleConditional(const HloInstruction* conditional) override; 112 Status HandleGather(const HloInstruction* gather) override; 113 Status HandleScatter(const HloInstruction* scatter) override; 114 Status HandleGetDimensionSize(const HloInstruction* get_size) override; 115 Status FinishVisit(const HloInstruction* root) override; 116 117 Status Preprocess(const HloInstruction* hlo) override; 118 Status Postprocess(const HloInstruction* hlo) override; 119 120 // Set the rates used to calculate the time taken by the computation. These 121 // need to be set before visiting starts. 122 void set_flops_per_second(float value) { 123 per_second_rates_[kFlopsKey] = value; 124 } 125 void set_transcendentals_per_second(float value) { 126 per_second_rates_[kTranscendentalsKey] = value; 127 } 128 void set_bytes_per_second(float value) { 129 per_second_rates_[kBytesAccessedKey] = value; 130 } 131 132 // Returns properties for the computation. 133 float flop_count() const; 134 float transcendental_count() const; 135 float bytes_accessed() const; 136 float optimal_seconds() const; 137 138 // Returns the respective cost computed for a particular HLO instruction, or 0 139 // if the HLO was not found to have a cost in the analysis. 140 int64 flop_count(const HloInstruction& hlo) const; 141 int64 transcendental_count(const HloInstruction& hlo) const; 142 int64 bytes_accessed(const HloInstruction& hlo) const; 143 float optimal_seconds(const HloInstruction& hlo) const; 144 145 const Properties& properties() const { return properties_sum_; } 146 const float property(const string& key) const { 147 return GetProperty(key, properties()); 148 } 149 150 protected: 151 typedef std::unordered_map<const HloInstruction*, Properties> HloToProperties; 152 153 // An FMA counts as two floating point operations in these analyzes. 154 static constexpr int64 kFmaFlops = 2; 155 156 HloCostAnalysis(const ShapeSizeFunction& shape_size, 157 const Properties& per_second_rates); 158 159 // Returns the properties computed from visiting the computation rooted at the 160 // given hlo. 161 // 162 // The difference between ProcessNestedSubcomputation and 163 // ProcessUnnestedSubcomputation is that we expect to get profile results for 164 // an unnested subcomputation's individual instructions, while we expect that 165 // a nested subcomputation is completely subsumed by its parent. 166 // 167 // For example, subcomputations inside kFusion and kMap are considered nested, 168 // while subcomputations inside kWhile and kConditional are considered 169 // unnested. 170 // 171 // Another way of thinking of this is, kFusion is implemented on the GPU 172 // backend using just one GPU kernel, while kWhile's body is implemented as a 173 // sequence of kernels, one for each HLO therein. Backends don't necessarily 174 // need to follow this same implementation strategy, but we assume they do for 175 // the purposes of this platform-generic cost analysis. 176 StatusOr<Properties> ProcessNestedSubcomputation(HloComputation* computation); 177 StatusOr<Properties> ProcessUnnestedSubcomputation( 178 HloComputation* computation); 179 180 // Utility function to handle all element-wise operations. 181 Status HandleElementwiseOp(const HloInstruction* hlo_instruction); 182 183 // Returns the default value if the key is not present in the 184 // properties. Otherwise, returns the value that the key maps to from the 185 // properties parameter. 186 static float GetProperty(const string& key, const Properties& properties, 187 float default_value = 0.0f); 188 189 // Returns 0.0f if the hlo is not present in hlo_to_properties or if the key 190 // is not present in hlo_to_properties[hlo]. Otherwise, returns the value that 191 // the key maps to in the properties of the given hlo. 192 static float GetPropertyForHlo(const HloInstruction& hlo, const string& key, 193 const HloToProperties& hlo_to_properties); 194 195 // Decorates shape_size_ by returning 0 immediately if the shape does not have 196 // a layout. 197 int64 GetShapeSize(const Shape& shape) const; 198 199 // Function which computes the size of the top-level of a given shape (not 200 // including nested elements, if any). If null then bytes_accessed methods 201 // return an error. 202 const ShapeSizeFunction shape_size_; 203 204 HloToProperties hlo_properties_; 205 206 // If true, the time taken will be computed from the rates for each property 207 // and the total time will be the maximum time, which is the time of the 208 // bottleneck. 209 bool current_should_compute_bottleneck_time_; 210 211 // The properties of the currently visited instruction. A HandleFoo method can 212 // modify these to change the default values computed in Preprocess. 213 Properties current_properties_; 214 215 // The sum of the properties of all HLOs in the computation. 216 Properties properties_sum_; 217 218 // How much of each property can be processed per second. E.g. if the property 219 // is bytes accessed, this is the number of bytes that can be processed per 220 // second. Is empty if no rates have been set. 221 Properties per_second_rates_; 222 223 TF_DISALLOW_COPY_AND_ASSIGN(HloCostAnalysis); 224 }; 225 226 } // namespace xla 227 228 #endif // TENSORFLOW_COMPILER_XLA_SERVICE_HLO_COST_ANALYSIS_H_ 229
