Merge pull request #13 from PaddlePaddle/develop

merge to local

Author: junjun315

paddle/fluid/API.spec

@@ -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'))
@@ -296,7 +296,7 @@ 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'))

paddle/fluid/framework/details/CMakeLists.txt

@@ -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)

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -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(