1 /* Copyright 2019 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_CORE_GRAPPLER_OPTIMIZERS_ARITHMETIC_OPTIMIZER_TEST_UTILS_H_ 17 #define TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_ARITHMETIC_OPTIMIZER_TEST_UTILS_H_ 18 19 #include "tensorflow/core/grappler/optimizers/arithmetic_optimizer.h" 20 #include "tensorflow/core/grappler/optimizers/common_subgraph_elimination.h" 21 #include "tensorflow/core/grappler/optimizers/constant_folding.h" 22 #include "tensorflow/core/grappler/optimizers/model_pruner.h" 23 #include "tensorflow/core/grappler/utils/grappler_test.h" 24 #include "tensorflow/core/lib/core/status_test_util.h" 25 26 namespace tensorflow { 27 namespace grappler { 28 29 class ArithmeticOptimizerTest : public GrapplerTest { 30 protected: 31 // Optimize a graph using optimizer and prune all the nodes that no 32 // longer have any output consumers. OptimizeAndPrune(GraphOptimizer * optimizer,GrapplerItem * item,GraphDef * output)33 void OptimizeAndPrune(GraphOptimizer* optimizer, GrapplerItem* item, 34 GraphDef* output) { 35 TF_EXPECT_OK(optimizer->Optimize(nullptr, *item, output)); 36 item->graph.Swap(output); 37 output->Clear(); 38 TF_EXPECT_OK(ModelPruner().Optimize(nullptr, *item, output)); 39 } 40 41 // Run optimizer twice to make sure the rewrite is idempotent. DedupAndOptimizeTwiceAndPrune(GraphOptimizer * optimizer,GrapplerItem * item,GraphDef * output)42 void DedupAndOptimizeTwiceAndPrune(GraphOptimizer* optimizer, 43 GrapplerItem* item, GraphDef* output) { 44 TF_EXPECT_OK(CommonSubgraphElimination().Optimize(nullptr, *item, output)); 45 item->graph.Swap(output); 46 output->Clear(); 47 TF_EXPECT_OK(optimizer->Optimize(nullptr, *item, output)); 48 item->graph.Swap(output); 49 output->Clear(); 50 TF_EXPECT_OK(optimizer->Optimize(nullptr, *item, output)); 51 item->graph.Swap(output); 52 output->Clear(); 53 TF_EXPECT_OK(ModelPruner().Optimize(nullptr, *item, output)); 54 } 55 56 // Run optimizer twice to make sure the rewrite is idempotent. OptimizeTwice(GraphOptimizer * optimizer,GrapplerItem * item,GraphDef * output)57 void OptimizeTwice(GraphOptimizer* optimizer, GrapplerItem* item, 58 GraphDef* output) { 59 TF_EXPECT_OK(optimizer->Optimize(nullptr, *item, output)); 60 item->graph.Swap(output); 61 output->Clear(); 62 TF_EXPECT_OK(optimizer->Optimize(nullptr, *item, output)); 63 } 64 65 // Run optimizer twice to make sure the rewrite is idempotent. 66 // Optionally run a constant folding pass before pruning. 67 void OptimizeTwiceAndPrune(GraphOptimizer* optimizer, GrapplerItem* item, 68 GraphDef* output, bool const_folding = false) { 69 TF_EXPECT_OK(optimizer->Optimize(nullptr, *item, output)); 70 71 item->graph.Swap(output); 72 output->Clear(); 73 TF_EXPECT_OK(optimizer->Optimize(nullptr, *item, output)); 74 75 if (const_folding) { 76 item->graph.Swap(output); 77 output->Clear(); 78 TF_EXPECT_OK(ConstantFolding(/*cpu_device=*/nullptr) 79 .Optimize(nullptr, *item, output)); 80 } 81 82 item->graph.Swap(output); 83 output->Clear(); 84 TF_EXPECT_OK(ModelPruner().Optimize(nullptr, *item, output)); 85 } 86 DisableAddToAddNCombining(ArithmeticOptimizer * optimizer)87 void DisableAddToAddNCombining(ArithmeticOptimizer* optimizer) { 88 optimizer->options_.combine_add_to_addn = false; 89 } 90 EnableOnlyAddToAddNCombining(ArithmeticOptimizer * optimizer)91 void EnableOnlyAddToAddNCombining(ArithmeticOptimizer* optimizer) { 92 DisableAllStages(optimizer); 93 optimizer->options_.combine_add_to_addn = true; 94 } 95 EnableOnlyFoldConjugateIntoTranspose(ArithmeticOptimizer * optimizer)96 void EnableOnlyFoldConjugateIntoTranspose(ArithmeticOptimizer* optimizer) { 97 DisableAllStages(optimizer); 98 optimizer->options_.fold_conjugate_into_transpose = true; 99 } 100 EnableOnlyFoldMultipleIntoConv(ArithmeticOptimizer * optimizer)101 void EnableOnlyFoldMultipleIntoConv(ArithmeticOptimizer* optimizer) { 102 DisableAllStages(optimizer); 103 optimizer->options_.fold_multiply_into_conv = true; 104 } 105 EnableOnlyFoldTransposeIntoMatMul(ArithmeticOptimizer * optimizer)106 void EnableOnlyFoldTransposeIntoMatMul(ArithmeticOptimizer* optimizer) { 107 DisableAllStages(optimizer); 108 optimizer->options_.fold_transpose_into_matmul = true; 109 } 110 EnableOnlyHoistCommonFactor(ArithmeticOptimizer * optimizer)111 void EnableOnlyHoistCommonFactor(ArithmeticOptimizer* optimizer) { 112 DisableAllStages(optimizer); 113 optimizer->options_.hoist_common_factor_out_of_aggregation = true; 114 } 115 EnableOnlyMinimizeBroadcasts(ArithmeticOptimizer * optimizer)116 void EnableOnlyMinimizeBroadcasts(ArithmeticOptimizer* optimizer) { 117 DisableAllStages(optimizer); 118 optimizer->options_.minimize_broadcasts = true; 119 } 120 EnableOnlyRemoveIdentityTranspose(ArithmeticOptimizer * optimizer)121 void EnableOnlyRemoveIdentityTranspose(ArithmeticOptimizer* optimizer) { 122 DisableAllStages(optimizer); 123 optimizer->options_.remove_identity_transpose = true; 124 } 125 EnableOnlyRemoveInvolution(ArithmeticOptimizer * optimizer)126 void EnableOnlyRemoveInvolution(ArithmeticOptimizer* optimizer) { 127 DisableAllStages(optimizer); 128 optimizer->options_.remove_involution = true; 129 } 130 EnableOnlyRemoveRedundantBitcast(ArithmeticOptimizer * optimizer)131 void EnableOnlyRemoveRedundantBitcast(ArithmeticOptimizer* optimizer) { 132 DisableAllStages(optimizer); 133 optimizer->options_.remove_redundant_bitcast = true; 134 } 135 EnableOnlyRemoveRedundantCast(ArithmeticOptimizer * optimizer)136 void EnableOnlyRemoveRedundantCast(ArithmeticOptimizer* optimizer) { 137 DisableAllStages(optimizer); 138 optimizer->options_.remove_redundant_cast = true; 139 } 140 EnableOnlyReduceUpsamplingDims(ArithmeticOptimizer * optimizer)141 void EnableOnlyReduceUpsamplingDims(ArithmeticOptimizer* optimizer) { 142 DisableAllStages(optimizer); 143 optimizer->options_.reduce_upsampling_dims = true; 144 } 145 EnableOnlyRemoveRedundantReshape(ArithmeticOptimizer * optimizer)146 void EnableOnlyRemoveRedundantReshape(ArithmeticOptimizer* optimizer) { 147 DisableAllStages(optimizer); 148 optimizer->options_.remove_redundant_reshape = true; 149 } 150 EnableOnlyRemoveNegation(ArithmeticOptimizer * optimizer)151 void EnableOnlyRemoveNegation(ArithmeticOptimizer* optimizer) { 152 DisableAllStages(optimizer); 153 optimizer->options_.remove_negation = true; 154 } 155 EnableOnlyReorderCastAndTranspose(ArithmeticOptimizer * optimizer)156 void EnableOnlyReorderCastAndTranspose(ArithmeticOptimizer* optimizer) { 157 DisableAllStages(optimizer); 158 optimizer->options_.reorder_cast_like_and_value_preserving = true; 159 } 160 EnableOnlyReplaceMulWithBroadcastByTile(ArithmeticOptimizer * optimizer)161 void EnableOnlyReplaceMulWithBroadcastByTile(ArithmeticOptimizer* optimizer) { 162 DisableAllStages(optimizer); 163 optimizer->options_.replace_mul_with_tile = true; 164 } 165 EnableOnlyReplaceMulWithSquare(ArithmeticOptimizer * optimizer)166 void EnableOnlyReplaceMulWithSquare(ArithmeticOptimizer* optimizer) { 167 DisableAllStages(optimizer); 168 optimizer->options_.replace_mul_with_square = true; 169 } 170 EnableOnlyReplacePackWithTileReshape(ArithmeticOptimizer * optimizer)171 void EnableOnlyReplacePackWithTileReshape(ArithmeticOptimizer* optimizer) { 172 DisableAllStages(optimizer); 173 optimizer->options_.replace_pack_with_tile_reshape = true; 174 } 175 EnableOnlyHoistCWiseUnaryChains(ArithmeticOptimizer * optimizer)176 void EnableOnlyHoistCWiseUnaryChains(ArithmeticOptimizer* optimizer) { 177 DisableAllStages(optimizer); 178 optimizer->options_.hoist_cwise_unary_chains = true; 179 } 180 EnableOnlySqrtDivToRsqrtMul(ArithmeticOptimizer * optimizer)181 void EnableOnlySqrtDivToRsqrtMul(ArithmeticOptimizer* optimizer) { 182 DisableAllStages(optimizer); 183 optimizer->options_.convert_sqrt_div_to_rsqrt_mul = true; 184 } 185 EnableOnlyLogSoftmax(ArithmeticOptimizer * optimizer)186 void EnableOnlyLogSoftmax(ArithmeticOptimizer* optimizer) { 187 DisableAllStages(optimizer); 188 optimizer->options_.convert_log_softmax = true; 189 } 190 EnableOnlyConvertPow(ArithmeticOptimizer * optimizer)191 void EnableOnlyConvertPow(ArithmeticOptimizer* optimizer) { 192 DisableAllStages(optimizer); 193 optimizer->options_.convert_pow = true; 194 } 195 EnableOnlyFuseSquaredDiff(ArithmeticOptimizer * optimizer)196 void EnableOnlyFuseSquaredDiff(ArithmeticOptimizer* optimizer) { 197 DisableAllStages(optimizer); 198 optimizer->options_.fuse_squared_diff = true; 199 } 200 EnableOnlyRemoveIdempotent(ArithmeticOptimizer * optimizer)201 void EnableOnlyRemoveIdempotent(ArithmeticOptimizer* optimizer) { 202 DisableAllStages(optimizer); 203 optimizer->options_.remove_idempotent = true; 204 } 205 EnableOnlyRemoveLogicalNot(ArithmeticOptimizer * optimizer)206 void EnableOnlyRemoveLogicalNot(ArithmeticOptimizer* optimizer) { 207 DisableAllStages(optimizer); 208 optimizer->options_.remove_logical_not = true; 209 } 210 EnableOnlySimplifyAggregation(ArithmeticOptimizer * optimizer)211 void EnableOnlySimplifyAggregation(ArithmeticOptimizer* optimizer) { 212 DisableAllStages(optimizer); 213 optimizer->options_.simplify_aggregation = true; 214 } 215 EnableOnlyLog1p(ArithmeticOptimizer * optimizer)216 void EnableOnlyLog1p(ArithmeticOptimizer* optimizer) { 217 DisableAllStages(optimizer); 218 optimizer->options_.convert_log1p = true; 219 } 220 EnableOnlyOptimizeMaxOrMinOfMonotonic(ArithmeticOptimizer * optimizer)221 void EnableOnlyOptimizeMaxOrMinOfMonotonic(ArithmeticOptimizer* optimizer) { 222 DisableAllStages(optimizer); 223 optimizer->options_.optimize_max_or_min_of_monotonic = true; 224 } 225 EnableOnlyExpm1(ArithmeticOptimizer * optimizer)226 void EnableOnlyExpm1(ArithmeticOptimizer* optimizer) { 227 DisableAllStages(optimizer); 228 optimizer->options_.convert_expm1 = true; 229 } 230 EnableOnlyUnaryOpsComposition(ArithmeticOptimizer * optimizer)231 void EnableOnlyUnaryOpsComposition(ArithmeticOptimizer* optimizer) { 232 DisableAllStages(optimizer); 233 optimizer->options_.unary_ops_composition = true; 234 } 235 EnableOnlyRemoveStackSliceSameAxis(ArithmeticOptimizer * optimizer)236 void EnableOnlyRemoveStackSliceSameAxis(ArithmeticOptimizer* optimizer) { 237 DisableAllStages(optimizer); 238 optimizer->options_.remove_stack_slice_same_axis = true; 239 } 240 EnableOnlySimplifyEmbeddingLookup(ArithmeticOptimizer * optimizer)241 void EnableOnlySimplifyEmbeddingLookup(ArithmeticOptimizer* optimizer) { 242 DisableAllStages(optimizer); 243 optimizer->options_.simplify_embedding_lookup = true; 244 } 245 EnableOnlyRemoveCastIntoSegmentReduction(ArithmeticOptimizer * optimizer)246 void EnableOnlyRemoveCastIntoSegmentReduction( 247 ArithmeticOptimizer* optimizer) { 248 DisableAllStages(optimizer); 249 optimizer->options_.remove_cast_into_segment_reduction = true; 250 } 251 252 private: DisableAllStages(ArithmeticOptimizer * optimizer)253 void DisableAllStages(ArithmeticOptimizer* optimizer) { 254 ArithmeticOptimizer::ArithmeticOptimizerOptions options; 255 options.dedup_computations = false; 256 options.combine_add_to_addn = false; 257 options.convert_sqrt_div_to_rsqrt_mul = false; 258 options.convert_pow = false; 259 options.convert_log1p = false; 260 options.optimize_max_or_min_of_monotonic = false; 261 options.fold_conjugate_into_transpose = false; 262 options.fold_multiply_into_conv = false; 263 options.fold_transpose_into_matmul = false; 264 options.hoist_common_factor_out_of_aggregation = false; 265 options.hoist_cwise_unary_chains = false; 266 options.minimize_broadcasts = false; 267 options.remove_identity_transpose = false; 268 options.remove_involution = false; 269 options.remove_idempotent = false; 270 options.remove_redundant_bitcast = false; 271 options.remove_redundant_cast = false; 272 options.remove_redundant_reshape = false; 273 options.remove_negation = false; 274 options.remove_logical_not = false; 275 options.reorder_cast_like_and_value_preserving = false; 276 options.replace_mul_with_tile = false; 277 options.replace_mul_with_square = false; 278 options.simplify_aggregation = false; 279 options.unary_ops_composition = false; 280 options.simplify_embedding_lookup = false; 281 options.remove_cast_into_segment_reduction = false; 282 optimizer->options_ = options; 283 } 284 }; 285 286 } // end namespace grappler 287 } // end namespace tensorflow 288 289 #endif // TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_ARITHMETIC_OPTIMIZER_TEST_UTILS_H_ 290