m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

merge develop

Browse Source 
test=develop
move-code
sneaxiy 6 years ago
parent c20db6357b c64d959343
commit 2d92b6be98
175 changed files with 8243 additions and 2019 deletions
Show Stats Download Patch File Download Diff File

42
paddle/fluid/API.spec
Unescape View File

@ -56,7 +56,7 @@ paddle.fluid.io.save_persistables (ArgSpec(args=['executor', 'dirname', 'main_pr
paddle.fluid.io.load_vars (ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', '0a5308f496632ab1ec3ba1f1377e6f95'))
paddle.fluid.io.load_params (ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)), ('document', '41779819cef32f2246e83aebc5a002e2'))
paddle.fluid.io.load_persistables (ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)), ('document', '28df5bfe26ca7a077f91156abb0fe6d2'))
paddle.fluid.io.save_inference_model (ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename', 'export_for_deployment'], varargs=None, keywords=None, defaults=(None, None, None, True)), ('document', '582d87b8df75a5a639a107db8ff86f9c'))
paddle.fluid.io.save_inference_model (ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename', 'export_for_deployment'], varargs=None, keywords=None, defaults=(None, None, None, True)), ('document', '70f4f53f13572436ac72d1c8b5efeb9d'))
paddle.fluid.io.load_inference_model (ArgSpec(args=['dirname', 'executor', 'model_filename', 'params_filename', 'pserver_endpoints'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '7a5255386075dac3c75b7058254fcdcb'))
paddle.fluid.initializer.ConstantInitializer.__init__ (ArgSpec(args=['self', 'value', 'force_cpu'], varargs=None, keywords=None, defaults=(0.0, False)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.initializer.UniformInitializer.__init__ (ArgSpec(args=['self', 'low', 'high', 'seed'], varargs=None, keywords=None, defaults=(-1.0, 1.0, 0)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@ -109,7 +109,7 @@ paddle.fluid.layers.reduce_prod (ArgSpec(args=['input', 'dim', 'keep_dim', 'name
paddle.fluid.layers.sequence_first_step (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', '2b290d3d77882bfe9bb8d331cac8cdd3'))
paddle.fluid.layers.sequence_last_step (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', 'c16a892f44f7fe71bfa5afc32d3f34ce'))
paddle.fluid.layers.sequence_slice (ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'fdcea0e8b5bc7d8d4b1b072c521014e6'))
paddle.fluid.layers.dropout (ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name', 'dropout_implementation'], varargs=None, keywords=None, defaults=(False, None, None, 'downgrade_in_infer')), ('document', 'dc7042734c6d8b8ce97321f017f01d6f'))
paddle.fluid.layers.dropout (ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name', 'dropout_implementation'], varargs=None, keywords=None, defaults=(False, None, None, 'downgrade_in_infer')), ('document', 'f1dd22f7351f7f9853212958e0d8aa7a'))
paddle.fluid.layers.split (ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None)), ('document', '652625345c2acb900029c78cc75f8aa6'))
paddle.fluid.layers.ctc_greedy_decoder (ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'ebbf2adbd79683dc93db03454dfa18c2'))
paddle.fluid.layers.edit_distance (ArgSpec(args=['input', 'label', 'normalized', 'ignored_tokens'], varargs=None, keywords=None, defaults=(True, None)), ('document', '97f0262f97602644c83142789d784571'))
@ -205,7 +205,7 @@ paddle.fluid.layers.maxout (ArgSpec(args=['x', 'groups', 'name'], varargs=None,
paddle.fluid.layers.space_to_depth (ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '5f207ae10589ebe38a63575ef6ff8e1e'))
paddle.fluid.layers.affine_grid (ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '51def402b8910e163cbace9d0c0526ed'))
paddle.fluid.layers.sequence_reverse (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '77a6d80aa5551ca70324fc975c44507f'))
paddle.fluid.layers.affine_channel (ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None)), ('document', '2f46f1ff39a13ab00857e7b9f44b2fa7'))
paddle.fluid.layers.affine_channel (ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name', 'act'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None, None)), ('document', 'ab84fdc6dc60f3ad9aa397e6007e3bf9'))
paddle.fluid.layers.similarity_focus (ArgSpec(args=['input', 'axis', 'indexes', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '70e3b5182a18b40b47ecabd7c8490a35'))
paddle.fluid.layers.hash (ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None)), ('document', '9bb77f8dc002dd2ce75d4769eaaf5007'))
paddle.fluid.layers.grid_sampler (ArgSpec(args=['x', 'grid', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd256cba1c41a5ed92ce3f31e24a2ca6d'))
@ -255,6 +255,7 @@ paddle.fluid.layers.reverse (ArgSpec(args=['x', 'axis'], varargs=None, keywords=
paddle.fluid.layers.has_inf (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '8f8c0306117ea441f20dcbbdba1f0ecc'))
paddle.fluid.layers.has_nan (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '2e53e83127dbfd86e7098bdfe9a549e8'))
paddle.fluid.layers.isfinite (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '0a437011c3906079fd8947ed3e52d292'))
paddle.fluid.layers.range (ArgSpec(args=['start', 'end', 'step', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '2ec937ede953ded2fdff2675883900bb'))
paddle.fluid.layers.While.__init__ (ArgSpec(args=['self', 'cond', 'is_test', 'name'], varargs=None, keywords=None, defaults=(False, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.layers.While.block (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.layers.Switch.__init__ (ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@ -277,7 +278,7 @@ paddle.fluid.layers.DynamicRNN.block (ArgSpec(args=['self'], varargs=None, keywo
paddle.fluid.layers.DynamicRNN.memory (ArgSpec(args=['self', 'init', 'shape', 'value', 'need_reorder', 'dtype'], varargs=None, keywords=None, defaults=(None, None, 0.0, False, 'float32')), ('document', 'b9174d4e91505b0c8ecc193eb51e248d'))
paddle.fluid.layers.DynamicRNN.output (ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None), ('document', 'b439a176a3328de8a75bdc5c08eece4a'))
paddle.fluid.layers.DynamicRNN.static_input (ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None), ('document', 'f29ad2478b6b2ad4f413d2936a331ea0'))
paddle.fluid.layers.DynamicRNN.step_input (ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None), ('document', '169d694d2224f62b4f3afdc3dbc19e95'))
paddle.fluid.layers.DynamicRNN.step_input (ArgSpec(args=['self', 'x', 'level'], varargs=None, keywords=None, defaults=(0,)), ('document', '7568c5ac7622a10288d3307a94134655'))
paddle.fluid.layers.DynamicRNN.update_memory (ArgSpec(args=['self', 'ex_mem', 'new_mem'], varargs=None, keywords=None, defaults=None), ('document', '5d83987da13b98363d6a807a52d8024f'))
paddle.fluid.layers.StaticRNN.__init__ (ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.layers.StaticRNN.memory (ArgSpec(args=['self', 'init', 'shape', 'batch_ref', 'init_value', 'init_batch_dim_idx', 'ref_batch_dim_idx'], varargs=None, keywords=None, defaults=(None, None, None, 0.0, 0, 1)), ('document', 'c24e368e23afac1ed91a78a639d7a9c7'))
@ -296,7 +297,7 @@ paddle.fluid.layers.tanh (ArgSpec(args=['x', 'name'], varargs=None, keywords=Non
paddle.fluid.layers.atan (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '3a46e0b5f9ce82348406478e610f14c9'))
paddle.fluid.layers.tanh_shrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '1e521554b9fdda9061ec6d306f0709b7'))
paddle.fluid.layers.softshrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '9eef31597bbafa2bd49691e072296e13'))
paddle.fluid.layers.sqrt (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '072a8541e0f632366bba10f67cb0db27'))
paddle.fluid.layers.sqrt (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e9e27491c39ac74d0b1ffe506aec0ebb'))
paddle.fluid.layers.abs (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '64650ac42cf82e9920cb0b172b1d29fd'))
paddle.fluid.layers.ceil (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c75d67dc5fe28f68e4cfffead4f698ad'))
paddle.fluid.layers.floor (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '647b16c5da5ef909649ae02abb434973'))
@ -376,23 +377,9 @@ paddle.fluid.contrib.Calibrator.__init__ (ArgSpec(args=['self'], varargs='args',
paddle.fluid.contrib.Calibrator.sample_data (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '3b8c85ca1e2cf753cc8c90a6c6992958'))
paddle.fluid.contrib.Calibrator.save_int8_model (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.reader.ctr_reader.ctr_reader (ArgSpec(args=['feed_dict', 'file_type', 'file_format', 'dense_slot_index', 'sparse_slot_index', 'capacity', 'thread_num', 'batch_size', 'file_list', 'slots', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'b2ebf3de2a6ef1af2c3b88d2db7591ab'))
paddle.fluid.contrib.build_compressor (ArgSpec(args=['place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'config'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.CompressPass.__init__ (ArgSpec(args=['self', 'place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'program_exe'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.CompressPass.add_strategy (ArgSpec(args=['self', 'strategy'], varargs=None, keywords=None, defaults=None), ('document', '3bf6010b6f47d3c86df0ec8957be95e0'))
paddle.fluid.contrib.CompressPass.apply (ArgSpec(args=['self', 'graph'], varargs=None, keywords=None, defaults=None), ('document', 'a92bf85d4b59bd4f2ac1706d7c4899a6'))
paddle.fluid.contrib.ImitationGraph.__init__ (ArgSpec(args=['self', 'program'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.ImitationGraph.all_parameters (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.SensitivePruneStrategy.__init__ (ArgSpec(args=['self', 'pruner', 'start_epoch', 'end_epoch', 'delta_rate', 'acc_loss_threshold', 'sensitivities'], varargs=None, keywords=None, defaults=(None, 0, 10, 0.2, 0.2, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.SensitivePruneStrategy.on_batch_begin (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.SensitivePruneStrategy.on_batch_end (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.SensitivePruneStrategy.on_compress_begin (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.SensitivePruneStrategy.on_compress_end (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.SensitivePruneStrategy.on_epoch_begin (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.SensitivePruneStrategy.on_epoch_end (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.MagnitudePruner.__init__ (ArgSpec(args=['self', 'threshold'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.MagnitudePruner.prune (ArgSpec(args=['self', 'param', 'threshold'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.contrib.RatioPruner.__init__ (ArgSpec(args=['self', 'ratios'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e7a81a325b296a9ca502ee5adb4fc85d'))
paddle.fluid.contrib.RatioPruner.prune (ArgSpec(args=['self', 'param', 'ratio'], varargs=None, keywords=None, defaults=(None,)), ('document', '358cbf2978c91028fb96a195a9884645'))
paddle.fluid.contrib.Compressor.__init__ (ArgSpec(args=['self', 'place', 'scope', 'train_program', 'train_reader', 'train_feed_list', 'train_fetch_list', 'eval_program', 'eval_reader', 'eval_feed_list', 'eval_fetch_list', 'teacher_programs', 'checkpoint_path', 'train_optimizer', 'distiller_optimizer'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, [], './checkpoints', None, None)), ('document', '31ae143830c9bf6b43547dd546c5ba80'))
paddle.fluid.contrib.Compressor.config (ArgSpec(args=['self', 'config_file'], varargs=None, keywords=None, defaults=None), ('document', '780d9c007276ccbb95b292400d7807b0'))
paddle.fluid.contrib.Compressor.run (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', 'c6e43d6a078d307672283c1f36e04fe9'))
paddle.fluid.contrib.load_persistables_for_increment (ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var', 'lookup_table_var_path'], varargs=None, keywords=None, defaults=None), ('document', '2ab36d4f7a564f5f65e455807ad06c67'))
paddle.fluid.contrib.load_persistables_for_inference (ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var_name'], varargs=None, keywords=None, defaults=None), ('document', '59066bac9db0ac6ce414d05780b7333f'))
paddle.fluid.contrib.convert_dist_to_sparse_program (ArgSpec(args=['program'], varargs=None, keywords=None, defaults=None), ('document', '74c39c595dc70d6be2f16d8e462d282b'))
@ -432,48 +419,59 @@ paddle.fluid.nets.img_conv_group (ArgSpec(args=['input', 'conv_num_filter', 'poo
paddle.fluid.optimizer.SGDOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'regularization', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.SGDOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.SGDOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.SGDOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.SGDOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.MomentumOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'momentum', 'use_nesterov', 'regularization', 'name'], varargs=None, keywords=None, defaults=(False, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.MomentumOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.MomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.MomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.MomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.AdagradOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name', 'initial_accumulator_value'], varargs=None, keywords=None, defaults=(1e-06, None, None, 0.0)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.AdagradOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.AdagradOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.AdagradOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.AdagradOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.AdamOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name', 'lazy_mode'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None, False)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.AdamOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.AdamOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.AdamOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.AdamOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.AdamaxOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.AdamaxOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.AdamaxOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.AdamaxOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.AdamaxOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.DecayedAdagradOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'decay', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.95, 1e-06, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.DecayedAdagradOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.DecayedAdagradOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.DecayedAdagradOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.DecayedAdagradOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.FtrlOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'l1', 'l2', 'lr_power', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.0, 0.0, -0.5, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.FtrlOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.FtrlOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.FtrlOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.FtrlOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.RMSPropOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'rho', 'epsilon', 'momentum', 'centered', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.95, 1e-06, 0.0, False, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.RMSPropOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.RMSPropOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.RMSPropOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.RMSPropOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.AdadeltaOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'epsilon', 'rho', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1e-06, 0.95, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.AdadeltaOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.AdadeltaOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.AdadeltaOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.AdadeltaOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.ModelAverage.__init__ (ArgSpec(args=['self', 'average_window_rate', 'min_average_window', 'max_average_window', 'regularization', 'name'], varargs=None, keywords=None, defaults=(10000, 10000, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.ModelAverage.apply (ArgSpec(args=['self', 'executor', 'need_restore'], varargs=None, keywords=None, defaults=(True,)), ('document', '46234a5470590feb336346f70a3db715'))
paddle.fluid.optimizer.ModelAverage.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.ModelAverage.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.ModelAverage.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.ModelAverage.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.ModelAverage.restore (ArgSpec(args=['self', 'executor'], varargs=None, keywords=None, defaults=None), ('document', '18db9c70be9c4dd466f9844457b21bfe'))
paddle.fluid.optimizer.LarsMomentumOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'momentum', 'lars_coeff', 'lars_weight_decay', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.0005, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.LarsMomentumOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.LarsMomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.LarsMomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.LarsMomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.backward.append_backward (ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '1a79bd7d10ae54ca763ec81bca36ba24'))
paddle.fluid.regularizer.L1DecayRegularizer.__init__ (ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))

15
paddle/fluid/framework/details/CMakeLists.txt
Unescape View File

@ -9,6 +9,7 @@ cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place
cc_library(multi_devices_helper SRCS multi_devices_helper.cc DEPS graph graph_helper)
cc_library(multi_devices_graph_print_pass SRCS multi_devices_graph_print_pass.cc DEPS multi_devices_helper)
cc_library(multi_devices_graph_check_pass SRCS multi_devices_graph_check_pass.cc DEPS multi_devices_helper)
cc_library(alloc_continuous_space_for_grad_pass SRCS alloc_continuous_space_for_grad_pass.cc DEPS graph graph_helper)
cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)
@ -22,6 +23,8 @@ endif()
if(WITH_GPU)
nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda variable_visitor)
nv_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda variable_visitor)
if(WITH_DISTRIBUTE)
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim dynload_cuda selected_rows_functor sendrecvop_rpc)
@ -35,6 +38,8 @@ if(WITH_GPU)
else()
cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
variable_visitor)
cc_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
variable_visitor)
if(WITH_DISTRIBUTE)
cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim selected_rows_functor sendrecvop_rpc)
@ -46,9 +51,7 @@ else()
cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
endif()
cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_base scope lod_tensor)
cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
if(WITH_GPU)
cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info)
@ -69,7 +72,9 @@ cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS grap
cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle fused_broadcast_op_handle)
cc_library(fuse_all_reduce_op_pass SRCS fuse_all_reduce_op_pass.cc DEPS graph graph_helper fused_all_reduce_op_handle)
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass inplace_op_pass)
if (WITH_GPU)
@ -98,5 +103,5 @@ cc_library(build_strategy SRCS build_strategy.cc DEPS
graph_viz_pass multi_devices_graph_pass
multi_devices_graph_print_pass multi_devices_graph_check_pass
fuse_elewise_add_act_pass multi_batch_merge_pass
fuse_relu_depthwise_conv_pass
memory_optimize_pass lock_free_optimize_pass)
fuse_relu_depthwise_conv_pass
memory_optimize_pass lock_free_optimize_pass alloc_continuous_space_for_grad_pass fuse_all_reduce_op_pass)

4
paddle/fluid/framework/details/all_reduce_op_handle.cc
Unescape View File

@ -11,9 +11,8 @@
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <algorithm>
#include "paddle/fluid/framework/details/all_reduce_op_handle.h"
#include <algorithm>
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/reduce_and_gather.h"
#include "paddle/fluid/framework/details/variable_visitor.h"
@ -56,6 +55,7 @@ void AllReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name());
WaitInputVarGenerated();
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(

393
paddle/fluid/framework/details/alloc_continuous_space_for_grad_pass.cc
View File

File diff suppressed because it is too large Load Diff

2
paddle/fluid/framework/details/broadcast_op_handle.h
Unescape View File

@ -57,7 +57,7 @@ struct BroadcastOpHandle : public OpHandleBase {
std::string Name() const override;
bool IsMultiDeviceTransfer() override { return false; };
bool IsMultiDeviceTransfer() override { return true; };
protected:
void RunImpl() override;

90
paddle/fluid/framework/details/build_strategy.cc
Unescape View File

@ -46,7 +46,16 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
public:
explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy)
: ir::PassBuilder(), strategy_(strategy) {
// Add a graph viz pass to record a graph.
if (!strategy_.debug_graphviz_path_.empty()) {
auto viz_pass = AppendPass("graph_viz_pass");
const std::string graph_path = string::Sprintf(
"%s%s", strategy_.debug_graphviz_path_.c_str(), "_original_graph");
viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
}
if (strategy_.enable_sequential_execution_) {
VLOG(10) << "Add sequential_execution_pass";
AppendPass("sequential_execution_pass");
}
@ -57,6 +66,7 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// Add op fusion.
if (strategy.fuse_relu_depthwise_conv_) {
VLOG(10) << "Add fuse_relu_depthwise_conv_pass";
AppendPass("fuse_relu_depthwise_conv_pass");
}
@ -68,29 +78,30 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// Add automatically inplace.
if (strategy_.enable_inplace_) {
VLOG(10) << "Add inplace_pass";
AppendPass("inplace_pass");
}
if (strategy.fuse_elewise_add_act_ops_) {
VLOG(10) << "Add fuse_elewise_add_act_pass";
AppendPass("fuse_elewise_add_act_pass");
}
// for single card training, fuse_all_reduce_ops is unnecessary.
// alloc_continuous_space_for_grad_pass should be before of MultiDevPass.
if (strategy.fuse_all_reduce_ops_) {
VLOG(10) << "Add alloc_continuous_space_for_grad_pass";
AppendPass("alloc_continuous_space_for_grad_pass");
}
// Add a graph viz pass to record a graph.
if (!strategy_.debug_graphviz_path_.empty()) {
if (!strategy.debug_graphviz_path_.empty()) {
auto viz_pass = AppendPass("graph_viz_pass");
const std::string graph_path = string::Sprintf(
"%s%s", strategy_.debug_graphviz_path_.c_str(), "_original_graph");
"%s%s", strategy.debug_graphviz_path_.c_str(), "_fused_graph");
viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
}
if (strategy.fuse_elewise_add_act_ops_) {
auto fuse_elewise_add_act_pass = AppendPass("fuse_elewise_add_act_pass");
// Add a graph viz pass to record a graph.
if (!strategy.debug_graphviz_path_.empty()) {
auto viz_pass = AppendPass("graph_viz_pass");
const std::string graph_path = string::Sprintf(
"%s%s", strategy.debug_graphviz_path_.c_str(), "_fused_graph");
viz_pass->Set<std::string>("graph_viz_path",
new std::string(graph_path));
}
}
CollectiveContext *context = CollectiveContext::GetInstance();
context->endpoints_ = strategy_.trainers_endpoints_;
context->trainer_id_ = strategy_.trainer_id_;
@ -108,11 +119,19 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// A side-effect of that, memory optimize cannot forsee the fetched vars
// , so fetchlist should be set persistable before call the Run interface.
if (strategy.memory_optimize_) {
auto memory_optimize_pass = AppendPass("memory_optimize_pass");
VLOG(10) << "Add memory_optimize_pass";
AppendPass("memory_optimize_pass");
}
AppendMultiDevPass(strategy);
if (strategy.fuse_all_reduce_ops_) {
// NOTE: fuse_all_reduce_ops will count the number of all_reduce operator
// first, if the number is zero, fuse_all_reduce_ops will do nothing.
VLOG(10) << "Add fuse_all_reduce_op_pass";
AppendPass("fuse_all_reduce_op_pass");
}
// Add a graph print pass to record a graph with device info.
if (!strategy_.debug_graphviz_path_.empty()) {
auto multi_devices_print_pass = AppendPass("multi_devices_print_pass");
@ -128,28 +147,34 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// Verify that the graph is correct for multi-device executor.
AppendPass("multi_devices_check_pass");
if (VLOG_IS_ON(2)) {
AppendPass("all_reduce_deps_pass");
}
if (SeqOnlyAllReduceOps(strategy)) {
VLOG(10) << "Add all_reduce_deps_pass";
AppendPass("all_reduce_deps_pass");
}
if (strategy_.remove_unnecessary_lock_) {
VLOG(10) << "Add modify_op_lock_and_record_event_pass";
AppendPass("modify_op_lock_and_record_event_pass");
}
}
// Convert graph to run on multi-devices.
void AppendMultiDevPass(const BuildStrategy &strategy) {
ir::Pass *multi_devices_pass;
ir::Pass *multi_devices_pass = nullptr;
if (strategy_.is_distribution_) {
VLOG(3) << "multi device parameter server mode";
VLOG(10) << "Add dist_multi_devices_pass";
multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
} else {
if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
VLOG(3) << "multi devices collective mode with allreduce";
VLOG(10) << "Add all_reduce_mode_multi_devices_pass";
multi_devices_pass =
AppendPass("allreduce_mode_multi_devices_pass").get();
AppendPass("all_reduce_mode_multi_devices_pass").get();
} else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
VLOG(3) << "multi deivces collective mode with reduce";
VLOG(10) << "Add reduce_mode_multi_devices_pass";
multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
} else {
PADDLE_THROW("Unknown reduce strategy.");
@ -206,9 +231,26 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
pass->Erase("nccl_ctxs");
pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
pass->Erase(kNCCLCtxs);
pass->SetNotOwned<platform::NCCLContextMap>(kNCCLCtxs, nctx);
#endif
} else if (pass->Type() == "fuse_all_reduce_op_pass") {
pass->Erase(kPlaces);
pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
pass->Erase(kLocalScopes);
pass->SetNotOwned<const std::vector<Scope *>>(kLocalScopes,
&local_scopes);
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
pass->Erase(kNCCLCtxs);
pass->SetNotOwned<platform::NCCLContextMap>(kNCCLCtxs, nctx);
#endif
} else if (pass->Type() == "alloc_continuous_space_for_grad_pass") {
pass->Erase(kPlaces);
pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
pass->Erase(kLocalScopes);
pass->SetNotOwned<const std::vector<Scope *>>(kLocalScopes,
&local_scopes);
} else if (pass->Type() == "sequential_execution_pass") {
LOG(INFO) << "set enable_sequential_execution:"
<< enable_sequential_execution_;
@ -239,7 +281,7 @@ USE_PASS(fuse_elewise_add_act_pass);
USE_PASS(graph_viz_pass);
USE_PASS(multi_batch_merge_pass);
USE_PASS(reduce_mode_multi_devices_pass);
USE_PASS(allreduce_mode_multi_devices_pass);
USE_PASS(all_reduce_mode_multi_devices_pass);
USE_PASS(dist_multi_devices_pass);
USE_PASS(multi_devices_check_pass);
USE_PASS(multi_devices_print_pass);
@ -249,4 +291,6 @@ USE_PASS(all_reduce_deps_pass);
USE_PASS(modify_op_lock_and_record_event_pass);
USE_PASS(inplace_pass);
USE_PASS(lock_free_optimize_pass);
USE_PASS(alloc_continuous_space_for_grad_pass);
USE_PASS(graph_to_program_pass);
USE_PASS(fuse_all_reduce_op_pass);

3
paddle/fluid/framework/details/build_strategy.h
Unescape View File

@ -16,6 +16,7 @@
#include <memory>
#include <string>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/ir/pass_builder.h"
@ -75,6 +76,8 @@ struct BuildStrategy {
bool fuse_elewise_add_act_ops_{false};
bool fuse_all_reduce_ops_{false};
bool fuse_relu_depthwise_conv_{false};
bool sync_batch_norm_{false};

154
paddle/fluid/framework/details/data_balance_op_handle.cc
Unescape View File

@ -1,154 +0,0 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/data_balance_op_handle.h"
#include <algorithm>
#include "paddle/fluid/framework/details/container_cast.h"
namespace paddle {
namespace framework {
namespace details {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
DataBalanceOpHandle::DataBalanceOpHandle(
ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs)
: OpHandleBase(node), local_scopes_(local_scopes), places_(places) {
if (ctxs) {
for (auto &p : places_) {
this->SetDeviceContext(p, ctxs->DevCtx(p));
}
}
}
#else
DataBalanceOpHandle::DataBalanceOpHandle(
ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places)
: OpHandleBase(node), local_scopes_(local_scopes), places_(places) {}
#endif
std::string DataBalanceOpHandle::Name() const { return "data balance"; }
std::vector<std::array<int, 3>> DataBalanceOpHandle::GetBalancePlan(
const std::vector<int> &device_sizes) {
int device_num = device_sizes.size();
int total_size = 0;
int empty_num = 0;
std::vector<std::array<int, 2>> size_device_vec;
size_device_vec.reserve(device_num);
for (int i = 0; i < device_num; ++i) {
if (device_sizes[i] == 0) {
++empty_num;
}
total_size += device_sizes[i];
size_device_vec.push_back({{device_sizes[i], i}});
}
std::vector<std::array<int, 3>> res;
if (empty_num == 0) {
// No need to do data balance.
return res;
}
if (total_size < device_num) {
// No enough data.
PADDLE_THROW_EOF();
}
std::sort(size_device_vec.begin(), size_device_vec.end(),
[](const std::array<int, 2> &a, const std::array<int, 2> &b) {
return a[0] > b[0];
});
int expected_device_size = total_size / device_num;
int src_idx = 0;
for (int dst_idx = device_num - empty_num; dst_idx < device_num; ++dst_idx) {
if (size_device_vec[src_idx][0] <= expected_device_size) {
++src_idx;
PADDLE_ENFORCE_LT(
src_idx, device_num - empty_num,
"In current srategy an empty tensor should not be copy source.");
}
size_device_vec[src_idx][0] -= expected_device_size;
size_device_vec[dst_idx][0] += expected_device_size;
res.push_back({{size_device_vec[src_idx][1], size_device_vec[dst_idx][1],
expected_device_size}});
}
return res;
}
void DataBalanceOpHandle::RunImpl() {
PADDLE_ENFORCE_GT(places_.size(), 1UL,
"Data balance can only be enabled when the number of "
"places to run larger than 1.");
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE(in_var_handles.size() % places_.size() == 0);
PADDLE_ENFORCE_EQ(
in_var_handles.size(), out_var_handles.size(),
"The NoDummyInputSize and NoDummyOutputSize should be equal.");
int data_num = in_var_handles.size() / places_.size();
WaitInputVarGenerated();
std::vector<std::vector<LoDTensor *>> lod_tensors(data_num);
std::vector<int> device_sizes;
for (int i = 0; i < static_cast<int>(in_var_handles.size()); ++i) {
PADDLE_ENFORCE_EQ(in_var_handles[i]->name(), out_var_handles[i]->name(),
"The name of input and output should be equal.");
int place_idx = i / data_num;
int data_idx = i % data_num;
auto *local_scope =
local_scopes_[place_idx]->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto *tensor_var = local_scope->FindVar(in_var_handles[i]->name());
PADDLE_ENFORCE(tensor_var->IsType<LoDTensor>());
auto *tensor = tensor_var->GetMutable<LoDTensor>();
lod_tensors[data_idx].push_back(tensor);
int ins_size =
tensor->lod().empty() ? tensor->dims()[0] : tensor->NumElements();
if (data_idx == 0) {
device_sizes.emplace_back(ins_size);
} else {
PADDLE_ENFORCE_EQ(
ins_size, device_sizes.at(place_idx),
"All data on the same device shall have the same batch size.");
}
}
const auto &balance_plan = GetBalancePlan(device_sizes);
for (const auto &trans : balance_plan) {
for (int data_idx = 0; data_idx < data_num; ++data_idx) {
LoDTensor *src_tensor = lod_tensors[data_idx][trans[0]];
LoDTensor *dst_tensor = lod_tensors[data_idx][trans[1]];
int trans_ins_size = trans[2];
LoD src_lod = src_tensor->lod();
int src_ins_size =
src_lod.empty() ? src_tensor->dims()[0] : src_tensor->NumElements();
int cut_point = src_ins_size - trans_ins_size;
if (!src_lod.empty()) {
for (auto &level : src_lod) {
cut_point = level[cut_point];
}
}
TensorCopySync(src_tensor->Slice(cut_point, src_tensor->dims()[0]),
dst_tensor->place(), dst_tensor);
src_tensor->ShareDataWith(src_tensor->Slice(0, cut_point));
if (!src_lod.empty()) {
dst_tensor->set_lod(SliceInLevel(
src_lod, 0, src_ins_size - trans_ins_size, src_ins_size));
src_tensor->set_lod(
SliceInLevel(src_lod, 0, 0, src_ins_size - trans_ins_size));
}
}
}
}
} // namespace details
} // namespace framework
} // namespace paddle

59
paddle/fluid/framework/details/data_balance_op_handle.h
Unescape View File

@ -1,59 +0,0 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "paddle/fluid/platform/nccl_helper.h"
#endif
namespace paddle {
namespace framework {
namespace details {
struct DataBalanceOpHandle : public OpHandleBase {
public:
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
DataBalanceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs);
#else
DataBalanceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places);
#endif
std::string Name() const override;
bool IsMultiDeviceTransfer() override { return false; };
protected:
void RunImpl() override;
private:
// std::vector<(src_dev_id, dst_dev_id, trans_size)>
std::vector<std::array<int, 3>> GetBalancePlan(
const std::vector<int> &batch_size_per_device);
const std::vector<Scope *> local_scopes_;
const std::vector<platform::Place> places_;
};
} // namespace details
} // namespace framework
} // namespace paddle

2
paddle/fluid/framework/details/fetch_op_handle.cc
Unescape View File

@ -82,6 +82,8 @@ void FetchOpHandle::WaitInputVarGenerated(const platform::Place &place) {
}
}
bool FetchOpHandle::IsMultiDeviceTransfer() { return true; }
std::string FetchOpHandle::Name() const { return "Fetch"; }
} // namespace details

2
paddle/fluid/framework/details/fetch_op_handle.h
Unescape View File

@ -39,6 +39,8 @@ struct FetchOpHandle : public OpHandleBase {
std::string Name() const override;
bool IsMultiDeviceTransfer() override;
protected:
void RunImpl() override;

195
paddle/fluid/framework/details/fuse_all_reduce_op_pass.cc
Unescape View File

@ -0,0 +1,195 @@
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <algorithm>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/all_reduce_op_handle.h"
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/fused_all_reduce_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
class FuseAllReduceOpPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override {
ir::Graph &result = *graph;
auto &places = Get<const std::vector<platform::Place>>(kPlaces);
auto &local_scopes = Get<const std::vector<Scope *>>(kLocalScopes);
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
auto *nccl_ctxs = &Get<platform::NCCLContextMap>(kNCCLCtxs);
#endif
std::unordered_set<std::string> grads;
auto &params_grads = result.Get<ParamsAndGrads>(kParamsAndGrads);
size_t num_of_all_reduce = params_grads.size();
grads.reserve(num_of_all_reduce);
for (auto p_g : params_grads) {
grads.insert(p_g.second);
}
size_t num_place = places.size();
std::unordered_map<std::string, ir::Node *> all_reduce_ops;
all_reduce_ops.reserve(grads.size());
for (auto &node : result.Nodes()) {
if (node->IsOp()) {
PADDLE_ENFORCE(node->IsWrappedBy<OpHandleBase>());
auto *all_reduce_op_handle =
dynamic_cast<AllReduceOpHandle *>(&node->Wrapper<OpHandleBase>());
if (all_reduce_op_handle) {
auto inputs = DynamicCast<VarHandle>(all_reduce_op_handle->Inputs());
PADDLE_ENFORCE_EQ(inputs.size(), num_place);
// The inputs' name should be the same.
auto &grad_name = inputs[0]->name();
for (size_t i = 1; i < inputs.size(); ++i) {
PADDLE_ENFORCE_EQ(inputs[i]->name(), grad_name,
"The input name should be the same.");
}
PADDLE_ENFORCE_NE(grads.count(grad_name), static_cast<size_t>(0));
all_reduce_ops.emplace(grad_name, node);
}
}
}
VLOG(10) << "Find all_reduce_ops: " << all_reduce_ops.size();
if (all_reduce_ops.size() == 0) {
return std::move(graph);
}
PADDLE_ENFORCE_EQ(all_reduce_ops.size(), grads.size(),
"The number of all_reduce OpHandle is not equal to the "
"number of grads. Maybe some gradients are sparse type, "
"it is not supported currently.");
VLOG(10) << "Insert fused_all_reduce";
auto &group_grads_params =
graph->Get<GroupGradsAndParams>(kGroupGradsAndParams);
for (auto &group_g_p : group_grads_params) {
size_t group_size = group_g_p.size();
PADDLE_ENFORCE_GT(group_size, static_cast<size_t>(0));
std::vector<ir::Node *> group_all_reduce_ops;
group_all_reduce_ops.reserve(group_size);
for (auto &g_p : group_g_p) {
group_all_reduce_ops.emplace_back(all_reduce_ops.at(g_p.first));
}
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
InsertFusedAllReduce(places, local_scopes, group_size,
group_all_reduce_ops, nccl_ctxs, &result);
#else
InsertFusedAllReduce(places, local_scopes, group_size,
group_all_reduce_ops, &result);
#endif
}
return std::move(graph);
}
void InsertFusedAllReduce(const std::vector<platform::Place> &places,
const std::vector<Scope *> &local_scopes,
const size_t num_of_all_reduce,
const std::vector<ir::Node *> &all_reduce_ops,
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const platform::NCCLContextMap *nccl_ctxs,
#endif
ir::Graph *result) const {
std::vector<VarHandleBase *> inputs;
std::vector<VarHandleBase *> outputs;
for (auto &op : all_reduce_ops) {
auto &op_handle = op->Wrapper<OpHandleBase>();
inputs.insert(inputs.end(), op_handle.Inputs().begin(),
op_handle.Inputs().end());
// Remove output
for_each(op_handle.Inputs().begin(), op_handle.Inputs().end(),
[&op_handle](VarHandleBase *var_handle) {
var_handle->RemoveOutput(&op_handle, op_handle.Node());
});
outputs.insert(outputs.end(), op_handle.Outputs().begin(),
op_handle.Outputs().end());
// Remove Input
for_each(
op_handle.Outputs().begin(), op_handle.Outputs().end(),
[](VarHandleBase *var_handle) { var_handle->ClearGeneratedOp(); });
result->RemoveNode(op_handle.Node());
}
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
CreateFusedAllReduceOp(inputs, outputs, num_of_all_reduce, places,
local_scopes, nccl_ctxs, result);
#else
CreateFusedAllReduceOp(inputs, outputs, num_of_all_reduce, places,
local_scopes, result);
#endif
}
private:
void CreateFusedAllReduceOp(const std::vector<VarHandleBase *> &inputs,
const std::vector<VarHandleBase *> &outputs,
const size_t num_of_all_reduce,
const std::vector<platform::Place> &places,
const std::vector<Scope *> &local_scopes,
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const platform::NCCLContextMap *nccl_ctxs,
#endif
ir::Graph *result) const {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
auto *op_handle = new FusedAllReduceOpHandle(
result->CreateEmptyNode("fused_all_reduce", ir::Node::Type::kOperation),
local_scopes, places, num_of_all_reduce, nccl_ctxs);
#else
auto *op_handle = new FusedAllReduceOpHandle(
result->CreateEmptyNode("fused_all_reduce", ir::Node::Type::kOperation),
local_scopes, places, num_of_all_reduce);
#endif
for (auto in : inputs) {
op_handle->AddInput(in);
}
for (auto out : outputs) {
op_handle->AddOutput(out);
}
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
if (!nccl_ctxs) {
SetCommunicationContext(places, op_handle);
}
#else
SetCommunicationContext(places, op_handle);
#endif
}
void SetCommunicationContext(const std::vector<platform::Place> &places,
FusedAllReduceOpHandle *op_handle) const {
for (size_t i = 0; i < places.size(); ++i) {
op_handle->SetDeviceContext(
places[i], platform::DeviceContextPool::Instance().Get(places[i]));
}
}
};
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(fuse_all_reduce_op_pass,
paddle::framework::details::FuseAllReduceOpPass);

51
paddle/fluid/framework/details/fuse_vars_op_handle.cc
Unescape View File

@ -1,51 +0,0 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/fuse_vars_op_handle.h"
namespace paddle {
namespace framework {
namespace details {
void FuseVarsOpHandle::RunImpl() {
WaitInputVarGenerated(place_);
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(in_var_handles.size(), 0UL);
PADDLE_ENFORCE_EQ(out_var_handles.size() - 1, inputs_numel_.size(), "");
auto scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto out_var_handle = out_var_handles[0];
auto out_var = scope->Var(out_var_handle->name());
auto out_tensor = out_var->GetMutable<LoDTensor>();
out_tensor->Resize({total_numel_}).mutable_data(this->place_, type_);
int64_t s = 0;
for (size_t i = 1; i < out_var_handles.size(); ++i) {
auto out_name = out_var_handles[i]->name();
auto out_t = scope->Var(out_name)->GetMutable<LoDTensor>();
auto numel = this->inputs_numel_.at(out_name);
out_t->ShareDataWith(out_tensor->Slice(s, s + numel));
s += numel;
}
this->RunAndRecordEvent([] {});
}
std::string FuseVarsOpHandle::Name() const { return "fuse vars"; }
} // namespace details
} // namespace framework
} // namespace paddle

65
paddle/fluid/framework/details/fuse_vars_op_handle.h
Unescape View File

@ -1,65 +0,0 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <map>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/device_context.h"
namespace paddle {
namespace framework {
namespace details {
struct FuseVarsOpHandle : public OpHandleBase {
public:
FuseVarsOpHandle(ir::Node *node, Scope *local_scope,
const platform::Place &place,
const std::unordered_map<std::string, int64_t> &inputs_numel,
const proto::VarType::Type var_type)
: OpHandleBase(node),
local_scope_(local_scope),
place_(place),
inputs_numel_(inputs_numel),
type_(var_type) {
total_numel_ = 0;
for (auto in_numel : inputs_numel) {
PADDLE_ENFORCE_GT(in_numel.second, 0);
total_numel_ += in_numel.second;
}
}
std::string Name() const override;
bool IsMultiDeviceTransfer() override { return false; };
protected:
void RunImpl() override;
private:
Scope *local_scope_;
const platform::Place place_;
const std::unordered_map<std::string, int64_t> inputs_numel_;
const proto::VarType::Type type_;
int64_t total_numel_;
};
} // namespace details
} // namespace framework
} // namespace paddle

249
paddle/fluid/framework/details/fused_all_reduce_op_handle.cc
Unescape View File

@ -0,0 +1,249 @@
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/fused_all_reduce_op_handle.h"
#include <algorithm>
#include <utility>
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/reduce_and_gather.h"
#include "paddle/fluid/framework/details/variable_visitor.h"
#include "paddle/fluid/platform/profiler.h"
DEFINE_bool(skip_fused_all_reduce_check, false, "");
namespace paddle {
namespace framework {
namespace details {
typedef std::vector<std::vector<std::pair<std::string, const LoDTensor *>>>
GradientAndLoDTensor;
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
FusedAllReduceOpHandle::FusedAllReduceOpHandle(
ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places, const size_t num_of_all_reduce,
const platform::NCCLContextMap *ctxs)
: OpHandleBase(node),
local_scopes_(local_scopes),
places_(places),
num_of_all_reduce_(num_of_all_reduce),
nccl_ctxs_(ctxs) {
if (nccl_ctxs_) {
for (auto &p : places_) {
this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
}
}
PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
}
#else
FusedAllReduceOpHandle::FusedAllReduceOpHandle(
ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places, const size_t num_of_all_reduce)
: OpHandleBase(node),
local_scopes_(local_scopes),
places_(places),
num_of_all_reduce_(num_of_all_reduce) {
PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
}
#endif
void FusedAllReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name());
VLOG(4) << this->DebugString();
WaitInputVarGenerated();
// The input: grad0(dev0), grad0(dev1), grad1(dev0), grad1(dev1)...
// The output: grad0(dev0), grad0(dev1), grad1(dev0), grad1(dev1)...
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
size_t place_num = places_.size();
PADDLE_ENFORCE_EQ(
in_var_handles.size(), place_num * num_of_all_reduce_,
"The NoDummyInputSize should be equal to the number of places.");
PADDLE_ENFORCE_EQ(
in_var_handles.size(), out_var_handles.size(),
"The NoDummyInputSize and NoDummyOutputSize should be equal.");
GradientAndLoDTensor grads_tensor;
grads_tensor.resize(place_num);
int64_t numel = -1;
auto dtype = static_cast<framework::proto::VarType::Type>(0);
for (size_t scope_idx = 0; scope_idx < local_scopes_.size(); ++scope_idx) {
auto &g_tensor = grads_tensor.at(scope_idx);
g_tensor.reserve(num_of_all_reduce_);
GetGradLoDTensor(scope_idx, in_var_handles, out_var_handles, &g_tensor);
int64_t element_num = 0;
framework::proto::VarType::Type ele_dtype =
static_cast<framework::proto::VarType::Type>(0);
GetDTypeAndNumel(g_tensor, &ele_dtype, &element_num);
if (numel == -1) {
numel = element_num;
}
if (dtype == static_cast<framework::proto::VarType::Type>(0)) {
dtype = ele_dtype;
PADDLE_ENFORCE_NE(ele_dtype,
static_cast<framework::proto::VarType::Type>(0));
}
PADDLE_ENFORCE_EQ(ele_dtype, dtype);
// Check whether the address space is contiguous.
std::sort(
g_tensor.begin(), g_tensor.end(),
[](const std::pair<std::string, const LoDTensor *> &grad1,
const std::pair<std::string, const LoDTensor *> &grad2) -> bool {
return grad1.second->data<void>() < grad2.second->data<void>();
});
for (size_t k = 1; k < g_tensor.size(); ++k) {
const void *cur_address = g_tensor.at(k - 1).second->data<void>();
int64_t len = g_tensor.at(k - 1).second->numel();
auto offset = len * framework::SizeOfType(dtype);
void *infer_next_address = reinterpret_cast<void *>(
reinterpret_cast<uintptr_t>(cur_address) + offset);
const void *next_address = g_tensor.at(k).second->data<void>();
VLOG(10) << string::Sprintf(
"Input[%d](%s) address: 0X%02x, Input[%d](%s) address: 0X%02x, Infer "
"input[%d] address: 0X%02x. The offset: %d",
k - 1, g_tensor.at(k - 1).first, cur_address, g_tensor.at(k).first, k,
next_address, k, infer_next_address, offset);
PADDLE_ENFORCE_EQ(infer_next_address, next_address,
"The address is not consistent.");
}
}
if (!FLAGS_skip_fused_all_reduce_check) {
for (size_t scope_idx = 0; scope_idx < place_num; ++scope_idx) {
for (size_t j = 1; j < num_of_all_reduce_; ++j) {
PADDLE_ENFORCE_EQ(grads_tensor.at(0).at(j).first,
grads_tensor.at(scope_idx).at(j).first);
}
}
}
std::vector<const void *> lod_tensor_data;
for (size_t scope_idx = 0; scope_idx < place_num; ++scope_idx) {
auto data = grads_tensor.at(scope_idx).at(0).second->data<void>();
lod_tensor_data.emplace_back(data);
}
if (platform::is_gpu_place(places_[0])) {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
int nccl_dtype = platform::ToNCCLDataType(dtype);
std::vector<std::function<void()>> all_reduce_calls;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto &p = places_[i];
void *buffer = const_cast<void *>(lod_tensor_data.at(i));
int dev_id = boost::get<platform::CUDAPlace>(p).device;
auto &nccl_ctx = nccl_ctxs_->at(dev_id);
auto stream = nccl_ctx.stream();
auto comm = nccl_ctx.comm_;
all_reduce_calls.emplace_back([=] {
PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
buffer, buffer, numel, static_cast<ncclDataType_t>(nccl_dtype),
ncclSum, comm, stream));
});
}
this->RunAndRecordEvent([&] {
if (all_reduce_calls.size() == 1UL) {
// Do not use NCCLGroup when manage NCCL by per thread per device
all_reduce_calls[0]();
} else {
platform::NCCLGroupGuard guard;
for (auto &call : all_reduce_calls) {
call();
}
}
});
#else
PADDLE_THROW("Not compiled with CUDA");
#endif
} else {
// Special handle CPU only Operator's gradient. Like CRF
auto grad_name = grads_tensor.at(0).at(0).first;
auto &trg = *this->local_scopes_[0]
->FindVar(kLocalExecScopeName)
->Get<Scope *>()
->FindVar(grad_name)
->GetMutable<framework::LoDTensor>();
// Reduce All data to trg in CPU
ReduceBufferData func(lod_tensor_data, trg.data<void>(), numel);
VisitDataType(trg.type(), func);
for (size_t i = 1; i < local_scopes_.size(); ++i) {
auto &scope =
*local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto &p = places_[i];
auto *var = scope.FindVar(grad_name);
auto *dev_ctx = dev_ctxes_.at(p);
size_t size = numel * SizeOfType(trg.type());
RunAndRecordEvent(p, [&trg, var, dev_ctx, p, size] {
auto dst_ptr = var->GetMutable<framework::LoDTensor>()->data<void>();
platform::CPUPlace cpu_place;
memory::Copy(cpu_place, dst_ptr, cpu_place, trg.data<void>(), size);
});
}
}
}
void FusedAllReduceOpHandle::GetGradLoDTensor(
const size_t &scope_idx, const std::vector<VarHandle *> &in_var_handles,
const std::vector<VarHandle *> &out_var_handles,
std::vector<std::pair<std::string, const LoDTensor *>> *grad_tensor) const {
auto *local_scope =
local_scopes_.at(scope_idx)->FindVar(kLocalExecScopeName)->Get<Scope *>();
size_t place_num = places_.size();
for (size_t j = 0; j < in_var_handles.size(); j += place_num) {
auto var_name = in_var_handles[j]->name();
PADDLE_ENFORCE_EQ(var_name, out_var_handles[j]->name());
auto &lod_tensor = local_scope->FindVar(var_name)->Get<LoDTensor>();
PADDLE_ENFORCE_EQ(lod_tensor.place(), places_.at(scope_idx));
grad_tensor->emplace_back(std::make_pair(var_name, &lod_tensor));
}
}
void FusedAllReduceOpHandle::GetDTypeAndNumel(
const std::vector<std::pair<std::string, const LoDTensor *>> &grad_tensor,
proto::VarType::Type *dtype, int64_t *numel) const {
*numel = 0;
for (size_t i = 0; i < grad_tensor.size(); ++i) {
// Get element number
int64_t len = grad_tensor.at(i).second->numel();
PADDLE_ENFORCE_GT(len, 0);
*numel += len;
// Get dtype
auto ele_type = grad_tensor.at(i).second->type();
if (i == 0) {
*dtype = ele_type;
}
PADDLE_ENFORCE_EQ(ele_type, *dtype);
}
}
std::string FusedAllReduceOpHandle::Name() const { return "fused_all_reduce"; }
} // namespace details
} // namespace framework
} // namespace paddle

76
paddle/fluid/framework/details/fused_all_reduce_op_handle.h
Unescape View File

@ -0,0 +1,76 @@
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <string>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "paddle/fluid/platform/nccl_helper.h"
#endif
namespace paddle {
namespace framework {
namespace details {
struct FusedAllReduceOpHandle : public OpHandleBase {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
FusedAllReduceOpHandle(ir::Node *node,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const size_t num_of_all_reduce,
const platform::NCCLContextMap *ctxs);
#else
FusedAllReduceOpHandle(ir::Node *node,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const size_t num_of_all_reduce);
#endif
std::string Name() const override;
// Delay and buffer nccl_all_reduce together can significantly increase
// performance. Disable this feature by returning false.
bool IsMultiDeviceTransfer() override { return true; };
protected:
void RunImpl() override;
private:
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
size_t num_of_all_reduce_;
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const platform::NCCLContextMap *nccl_ctxs_;
#endif
// Check the dtype of the input
void GetDTypeAndNumel(
const std::vector<std::pair<std::string, const LoDTensor *>> &g_tensor,
proto::VarType::Type *dtype, int64_t *total_num) const;
// Get gradient's name and LoDTensor
void GetGradLoDTensor(const size_t &scope_idx,
const std::vector<VarHandle *> &in_var_handles,
const std::vector<VarHandle *> &out_var_handles,
std::vector<std::pair<std::string, const LoDTensor *>>
*grad_tensor) const;
};
} // namespace details
} // namespace framework
} // namespace paddle

60
paddle/fluid/framework/details/multi_devices_graph_pass.cc
Unescape View File

@ -11,18 +11,19 @@
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
#include <algorithm>
#include <fstream>
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/details/all_reduce_op_handle.h"
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/data_balance_op_handle.h"
#include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
#include "paddle/fluid/framework/details/reduce_op_handle.h"
#include "paddle/fluid/framework/details/rpc_op_handle.h"
#include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
@ -134,21 +135,26 @@ void AddOutputToLeafOps(ir::Graph *graph) {
}
} // namespace
void MultiDevSSAGraphBuilderBase::CheckGraph(const ir::Graph &graph) const {}
void MultiDevSSAGraphBuilderBase::Init() const {
all_vars_.clear();
loss_var_name_ = Get<const std::string>(kLossVarName);
VLOG(10) << "Init MultiDevSSAGraphBuilder, loss name: " << loss_var_name_;
places_ = Get<const std::vector<platform::Place>>(kPlaces);
local_scopes_ = Get<const std::vector<Scope *>>(kLocalScopes);
strategy_ = Get<const BuildStrategy>(kStrategy);
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
nccl_ctxs_ = &Get<platform::NCCLContextMap>("nccl_ctxs");
nccl_ctxs_ = &Get<platform::NCCLContextMap>(kNCCLCtxs);
#endif
PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
}
std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
Init();
CheckGraph(*graph);
std::vector<ir::Node *> sorted_ops = SortOperations(*graph);
auto nodes = graph->ReleaseNodes();
@ -166,7 +172,6 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
result.Set(kGraphOps, new GraphOps);
bool is_forwarding = true;
bool insert_collection_ops = NeedCollectiveOps();
for (ir::Node *node : sorted_ops) {
if (DealWithSpecialOp(&result, node)) {
@ -185,8 +190,8 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
CreateComputationalOps(&result, node, places_.size());
}
// Insert collection ops
if (!is_forwarding && insert_collection_ops) {
// Insert collective ops if nranks > 1
if (!is_forwarding && Get<size_t>(kNRanks) > 1) {
try {
bool is_bk_op =
static_cast<bool>(boost::get<int>(node->Op()->GetAttr(
@ -200,13 +205,13 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
boost::get<std::vector<std::string>>(node->Op()->GetNullableAttr(
OpProtoAndCheckerMaker::OpRoleVarAttrName()));
PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
for (size_t i = 0; i < backward_vars.size(); i += 2) {
auto &p_name = backward_vars[i];
auto &g_name = backward_vars[i + 1];
VLOG(10) << "Bcast " << g_name << " for parameter " << p_name;
InsertCollectiveOp(&result, p_name, g_name);
if (NeedCollectiveForGrad(g_name, sorted_ops)) {
InsertCollectiveOp(&result, p_name, g_name);
}
}
} catch (boost::bad_get e) {
}
@ -226,6 +231,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
* Only variables should be the leaves of graph.
*/
AddOutputToLeafOps(&result);
result.Erase(kGraphOps);
return graph;
}
@ -258,6 +264,11 @@ void MultiDevSSAGraphBuilderBase::InsertScaleLossGradOp(
}
}
bool MultiDevSSAGraphBuilderBase::DealWithSpecialOp(ir::Graph *result,
ir::Node *node) const {
return false;
}
std::vector<ir::Node *> MultiDevSSAGraphBuilderBase::SortOperations(
const ir::Graph &graph) const {
return ir::TopologySortOperations(graph);
@ -271,8 +282,20 @@ bool MultiDevSSAGraphBuilderBase::UseGPU() const {
return use_gpu;
}
bool MultiDevSSAGraphBuilderBase::NeedCollectiveOps() const {
return Get<size_t>(kNRanks) > 1;
bool MultiDevSSAGraphBuilderBase::NeedCollectiveForGrad(
const std::string &grad_name, std::vector<ir::Node *> ops) const {
// if we have allreduce_op for current gradient variable in the graph,
// then we don't need to add allreduce_op_handle for this gradient
// NOTE: This is for the case that all gradients should add collective ops
for (auto *node : ops) {
if (node->Op()->Type() != "allreduce") continue;
for (auto in_name : node->Op()->InputArgumentNames()) {
if (in_name == grad_name) {
return false;
}
}
}
return true;
}
void MultiDevSSAGraphBuilderBase::CreateOpHandleIOs(ir::Graph *result,
@ -496,20 +519,17 @@ VarHandle *MultiDevSSAGraphBuilderBase::CreateReduceOp(ir::Graph *result,
}
bool MultiDevSSAGraphBuilderBase::IsScaleLossOp(ir::Node *node) const {
return boost::get<int>(
return !loss_var_name_.empty() && node->Op() &&
boost::get<int>(
node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) ==
(static_cast<int>(OpRole::kBackward) |
static_cast<int>(OpRole::kLoss)) &&
!loss_var_name_.empty(); // If loss_var is empty. This is test mode
static_cast<int>(OpRole::kLoss));
}
bool MultiDevSSAGraphBuilderBase::IsSparseGradient(
const std::string &og) const {
PADDLE_ENFORCE(all_vars_.count(og) != 0);
if (all_vars_.at(og)->GetType() == proto::VarType::SELECTED_ROWS) {
return true;
}
return false;
return all_vars_.at(og)->GetType() == proto::VarType::SELECTED_ROWS;
}
void AllReduceSSAGraphBuilder::InsertCollectiveOp(
@ -995,7 +1015,7 @@ static int MultiDevSSAGraphBuilderRegister(const std::string &builder_mode) {
REGISTER_MULTI_DEVICES_PASS(reduce_mode_multi_devices_pass,
paddle::framework::details::ReduceSSAGraphBuilder);
REGISTER_MULTI_DEVICES_PASS(
allreduce_mode_multi_devices_pass,
all_reduce_mode_multi_devices_pass,
paddle::framework::details::AllReduceSSAGraphBuilder);
REGISTER_MULTI_DEVICES_PASS(dist_multi_devices_pass,
paddle::framework::details::DistSSAGraphBuilder);

20
paddle/fluid/framework/details/multi_devices_graph_pass.h
Unescape View File

@ -14,7 +14,10 @@
#pragma once
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
@ -31,12 +34,6 @@ namespace framework {
class Scope;
namespace details {
constexpr char kLossVarName[] = "loss_var_name";
constexpr char kPlaces[] = "places";
constexpr char kLocalScopes[] = "local_scopes";
constexpr char kStrategy[] = "strategy";
constexpr char kNRanks[] = "nranks";
class MultiDevSSAGraphBuilderBase : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
@ -44,18 +41,21 @@ class MultiDevSSAGraphBuilderBase : public ir::Pass {
virtual void Init() const;
virtual void CheckGraph(const ir::Graph &graph) const;
virtual std::vector<ir::Node *> SortOperations(const ir::Graph &graph) const;
virtual void InsertCollectiveOp(ir::Graph *result, const std::string &p_name,
const std::string &g_name) const = 0;
virtual bool DealWithSpecialOp(ir::Graph *result, ir::Node *node) const = 0;
virtual bool DealWithSpecialOp(ir::Graph *result, ir::Node *node) const;
virtual void InsertPostprocessOps(ir::Graph *result) const = 0;
bool UseGPU() const;
bool NeedCollectiveOps() const;
bool NeedCollectiveForGrad(const std::string &grad_name,
std::vector<ir::Node *> ops) const;
bool IsScaleLossOp(ir::Node *node) const;
@ -109,10 +109,6 @@ class AllReduceSSAGraphBuilder : public MultiDevSSAGraphBuilderBase {
virtual void InsertCollectiveOp(ir::Graph *result, const std::string &p_name,
const std::string &g_name) const;
virtual bool DealWithSpecialOp(ir::Graph *result, ir::Node *node) const {
return false;
}
virtual void InsertPostprocessOps(ir::Graph *result) const {}
};

23
paddle/fluid/framework/details/multi_devices_helper.h
Unescape View File

@ -16,6 +16,9 @@
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
@ -44,6 +47,26 @@ const char kGraphVars[] = "vars";
typedef std::unordered_set<VarHandleBase *> GraphDepVars;
const char kGraphDepVars[] = "dep_vars";
constexpr char kNCCLCtxs[] = "nccl_ctxs";
constexpr char kLossVarName[] = "loss_var_name";
constexpr char kPlaces[] = "places";
constexpr char kLocalScopes[] = "local_scopes";
constexpr char kStrategy[] = "strategy";
constexpr char kNRanks[] = "nranks";
typedef std::unordered_set<std::string> FusedVars;
constexpr char kFusedVars[] = "fused_vars";
typedef std::vector<std::pair<std::string, std::string>> ParamsAndGrads;
constexpr char kParamsAndGrads[] = "params_grads";
typedef std::vector<std::vector<std::pair<std::string, std::string>>>
GroupGradsAndParams;
constexpr char kGroupGradsAndParams[] = "group_grads_params";
constexpr char kFusedVarNamePrefix[] = "@FUSEDVAR@";
} // namespace details
} // namespace framework
} // namespace paddle

71
paddle/fluid/framework/details/op_handle_base.cc
Unescape View File

@ -13,6 +13,7 @@
// limitations under the License.
#include "paddle/fluid/framework/details/op_handle_base.h"
#include <map>
#include <unordered_set>
namespace paddle {
namespace framework {
@ -41,15 +42,42 @@ OpHandleBase::~OpHandleBase() {
void OpHandleBase::Run(bool use_cuda) {
#ifdef PADDLE_WITH_CUDA
if (events_.empty() && use_cuda) {
if (events_.empty() && use_cuda && dev_ctxes_.size() > 0) {
for (auto &p : dev_ctxes_) {
int dev_id = boost::get<platform::CUDAPlace>(p.first).device;
PADDLE_ENFORCE(cudaSetDevice(dev_id));
PADDLE_ENFORCE(
cudaEventCreateWithFlags(&events_[dev_id], cudaEventDisableTiming));
}
if (IsMultiDeviceTransfer() && dev_ctxes_.size() > 0) {
for (auto &out_var : outputs_) {
auto *out_var_handle = dynamic_cast<VarHandle *>(out_var);
if (out_var_handle) {
int dev_id =
boost::get<platform::CUDAPlace>(out_var_handle->place()).device;
out_var_handle->SetGenerateEvent(events_[dev_id]);
}
}
} else {
PADDLE_ENFORCE_EQ(dev_ctxes_.size(), 1UL,
"%s should have only one dev_ctx.", Name());
auto &place = dev_ctxes_.begin()->first;
int dev_id = boost::get<platform::CUDAPlace>(place).device;
for (auto &out_var : outputs_) {
auto *out_var_handle = dynamic_cast<VarHandle *>(out_var);
if (out_var_handle) {
PADDLE_ENFORCE(
platform::is_same_place(place, out_var_handle->place()),
"The place of input(%s) is not consistent with the "
"place of current op(%s).",
out_var_handle->Name(), Name());
out_var_handle->SetGenerateEvent(events_[dev_id]);
}
}
}
}
#else
PADDLE_ENFORCE(!use_cuda);
#endif
@ -93,17 +121,48 @@ void OpHandleBase::AddOutput(VarHandleBase *out) {
void OpHandleBase::WaitInputVarGenerated() {
for (auto in_var : inputs_) {
if (NeedWait(in_var)) {
for (auto &pair : dev_ctxes_) {
in_var->GeneratedOp()->RecordWaitEventOnCtx(pair.second);
// Dummy Variable is used to represent dependencies between operators, so
// there doesn't add event for it.
auto *in_var_handle = dynamic_cast<VarHandle *>(in_var);
if (in_var_handle) {
auto &place = in_var_handle->place();
if (platform::is_gpu_place(place)) {
#ifdef PADDLE_WITH_CUDA
auto stream =
static_cast<platform::CUDADeviceContext *>(dev_ctxes_.at(place))
->stream();
PADDLE_ENFORCE(
cudaStreamWaitEvent(stream, in_var_handle->GetEvent(), 0));
#else
PADDLE_THROW("Doesn't compile the GPU.");
#endif
}
// There are nothing to do when the place is CPUPlace.
}
}
}
}
void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) {
for (auto *in : inputs_) {
if (NeedWait(in)) {
in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(place));
for (auto in_var : inputs_) {
if (NeedWait(in_var)) {
// Dummy Variable is used to represent dependencies between operators, so
// there doesn't add event for it.
auto *in_var_handle = dynamic_cast<VarHandle *>(in_var);
if (in_var_handle) {
if (platform::is_gpu_place(in_var_handle->place())) {
#ifdef PADDLE_WITH_CUDA
auto stream = static_cast<platform::CUDADeviceContext *>(
dev_ctxes_.at(in_var_handle->place()))
->stream();
PADDLE_ENFORCE(
cudaStreamWaitEvent(stream, in_var_handle->GetEvent(), 0));
#else
PADDLE_THROW("Doesn't compile the GPU.");
#endif
}
// There are nothing to do when the place is CPUPlace.
}
}
}
}

25
paddle/fluid/framework/details/reduce_and_gather.h
Unescape View File

@ -53,6 +53,31 @@ struct ReduceLoDTensor {
}
};
struct ReduceBufferData {
const std::vector<const void *> &src_data_;
void *dst_data_;
int64_t numel_;
ReduceBufferData(const std::vector<const void *> &src, void *dst,
int64_t numel)
: src_data_(src), dst_data_(dst), numel_(numel) {}
template <typename T>
void apply() const {
T *dst_data = reinterpret_cast<T *>(dst_data_);
for (size_t i = 0; i < src_data_.size(); ++i) {
auto srd_data = reinterpret_cast<const T *>(src_data_[i]);
VLOG(10) << "dst: " << dst_data_ << ", " << srd_data;
if (srd_data == dst_data_) {
continue;
}
std::transform(srd_data, srd_data + numel_, dst_data, dst_data,
[](T a, T b) -> T { return a + b; });
}
}
};
inline void GatherLocalSelectedRows(
const std::vector<const SelectedRows *> &src_selecte_rows_,
const std::vector<platform::Place> &in_places,

152
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
View File

File diff suppressed because it is too large Load Diff

24
paddle/fluid/framework/details/threaded_ssa_graph_executor.h
Unescape View File

@ -15,18 +15,20 @@
#pragma once
#include <deque>
#include <functional>
#include <list>
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include <functional>
#include "ThreadPool.h" // ThreadPool in thrird party
#include "paddle/fluid/framework/blocking_queue.h"
#include "paddle/fluid/framework/details/exception_holder.h"
#include "paddle/fluid/framework/details/execution_strategy.h"
#include "paddle/fluid/framework/details/fetch_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/ssa_graph_executor.h"
#include "paddle/fluid/framework/ir/graph.h"
@ -36,6 +38,12 @@ class Scope;
namespace details {
struct OpDependentData {
std::unordered_map<OpHandleBase *, size_t> pending_ops_;
std::unordered_set<VarHandleBase *> pending_vars_;
std::unordered_set<OpHandleBase *> ready_ops_;
};
class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
public:
ThreadedSSAGraphExecutor(const ExecutionStrategy &strategy,
@ -57,29 +65,35 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
private:
ir::Graph *graph_;
std::unique_ptr<::ThreadPool> pool_;
::ThreadPool prepare_pool_;
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
platform::DeviceContextPool fetch_ctxs_;
ExceptionHolder exception_holder_;
std::atomic<int> running_ops_;
void InsertPendingOp(std::unordered_map<OpHandleBase *, size_t> *pending_ops,
OpHandleBase *op_instance) const;
void InsertPendingVar(std::unordered_set<VarHandleBase *> *pending_vars,
BlockingQueue<VarHandleBase *> *ready_vars,
std::unordered_set<VarHandleBase *> *ready_vars,
VarHandleBase *var) const;
void InsertFetchOps(const std::vector<std::string> &fetch_tensors,
std::vector<FetchOpHandle *> *fetch_ops,
std::unordered_set<VarHandleBase *> *fetch_dependencies,
std::unordered_set<OpHandleBase *> *ready_ops,
std::unordered_map<OpHandleBase *, size_t> *pending_ops,
std::unordered_set<VarHandleBase *> *pending_vars,
BlockingQueue<VarHandleBase *> *ready_vars,
FeedFetchList *fetch_data);
void PrepareOpDeps();
void CopyOpDeps();
private:
std::future<std::unique_ptr<OpDependentData>> op_deps_futures_;
ExecutionStrategy strategy_;
std::unique_ptr<OpDependentData> op_deps_;
// use std::list because clear(), push_back, and for_each are O(1)
std::list<std::future<void>> run_op_futures_;
};

27
paddle/fluid/framework/details/var_handle.h
Unescape View File

@ -43,6 +43,7 @@ struct VarHandleBase {
virtual ~VarHandleBase();
virtual std::string DebugString() const = 0;
virtual const std::string& Name() const = 0;
void AddInput(OpHandleBase* in, ir::Node* node) {
node_->inputs.clear();
@ -95,8 +96,6 @@ struct VarHandleBase {
//
// NOTE: runtime variables have place.
struct VarHandle : public VarHandleBase {
explicit VarHandle(ir::Node* node) : VarHandleBase(node) {}
virtual ~VarHandle();
std::string DebugString() const override;
@ -109,6 +108,20 @@ struct VarHandle : public VarHandleBase {
name_(std::move(name)),
place_(std::move(place)) {}
#ifdef PADDLE_WITH_CUDA
bool HasEvent() { return has_event_; }
const cudaEvent_t& GetEvent() {
PADDLE_ENFORCE(HasEvent(), "The event is not set.");
return event_;
}
void SetGenerateEvent(const cudaEvent_t& event) {
has_event_ = true;
event_ = event;
}
#endif
// version field currently is not used, however, just store the version to
// debug easily.
private:
@ -116,6 +129,11 @@ struct VarHandle : public VarHandleBase {
size_t scope_idx_;
std::string name_;
platform::Place place_;
#ifdef PADDLE_WITH_CUDA
// Only when this event is triggered, var is generated.
cudaEvent_t event_;
bool has_event_{false};
#endif
public:
bool IsTheSameVar(const VarHandle& o) const {
@ -125,6 +143,7 @@ struct VarHandle : public VarHandleBase {
size_t version() const { return version_; }
size_t scope_idx() const { return scope_idx_; }
const std::string& Name() const override { return name_; }
const std::string& name() const { return name_; }
const platform::Place& place() const { return place_; }
};
@ -136,6 +155,10 @@ struct DummyVarHandle : public VarHandleBase {
virtual ~DummyVarHandle();
std::string DebugString() const override;
public:
const std::string& Name() const override { return name_; }
std::string name_{"DummyVar"};
};
} // namespace details

2
paddle/fluid/framework/ir/cpu_quantize_pass.cc
Unescape View File

@ -224,8 +224,8 @@ std::unique_ptr<ir::Graph> CPUQuantizePass::ApplyImpl(
PADDLE_ENFORCE(param_scope());
QuantizeConv(graph.get(), false /* with_residual_data */);
QuantizeConv(graph.get(), true /* with_residual_data */);
QuantizeConv(graph.get());
QuantizePool(graph.get());
return graph;

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save