merge develop branch. test=develop

Author: dzhwinter

CMakeLists.txt

@@ -62,7 +62,6 @@ option(WITH_DISTRIBUTE  "Compile with distributed support"              OFF)
 option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
 option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen"            OFF)
 option(WITH_ARM_FP16    "Use half precision support on armv8.2-a cpu"   OFF)
-option(WITH_FAST_BUNDLE_TEST    "Bundle tests that can be run in a single process together to reduce launch overhead"   OFF)
 option(WITH_CONTRIB     "Compile the third-party contributation"        OFF)
 option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
 option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)

cmake/configure.cmake

@@ -50,11 +50,7 @@ if(NOT WITH_PROFILER)
 endif(NOT WITH_PROFILER)
 
 if(NOT CMAKE_CROSSCOMPILING)
-    if(WITH_AVX AND AVX512F_FOUND)
-        set(SIMD_FLAG ${AVX512F_FLAG})
-    elseif(WITH_AVX AND AVX2_FOUND)
-        set(SIMD_FLAG ${AVX2_FLAG})
-    elseif(WITH_AVX AND AVX_FOUND)
+    if(WITH_AVX AND AVX_FOUND)
         set(SIMD_FLAG ${AVX_FLAG})
     elseif(SSE3_FOUND)
         set(SIMD_FLAG ${SSE3_FLAG})

cmake/simd.cmake

@@ -89,7 +89,9 @@ CHECK_CXX_SOURCE_RUNS("
 #include <immintrin.h>
 int main()
 {
-    __m512i a = _mm512_undefined_epi32();
+    __m512i a = _mm512_set_epi32 (-1, 2, -3, 4, -1, 2, -3, 4,
+                                  13, -5, 6, -7, 9, 2, -6, 3);
+    __m512i result = _mm512_abs_epi32 (a);
     return 0;
 }" AVX512F_FOUND)
 

paddle/fluid/API.spec

@@ -67,8 +67,8 @@ paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size',
 paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type', 'is_test'], varargs=None, keywords=None, defaults=(False,))
 paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None))
 paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
-paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
-paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
+paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
+paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
 paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False))
 paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
@@ -103,7 +103,7 @@ paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 's
 paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
-paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
+paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode'], varargs=None, keywords=None, defaults=(False, -100, False))
 paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1))
@@ -174,9 +174,11 @@ paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None
 paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None))
 paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
 paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None))
+paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
 paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))

paddle/fluid/framework/details/CMakeLists.txt

@@ -1,5 +1,6 @@
 cc_library(var_handle SRCS var_handle.cc DEPS place framework_proto node)
 cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context lod_tensor)
+cc_library(op_graph_view SRCS op_graph_view.cc DEPS op_handle_base)
 cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
 cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
 cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
@@ -30,20 +31,25 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_
 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(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
+
+if (WITH_GPU)
   cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS 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 graph graph_helper pass)
 endif()
 
+cc_library(sequential_execution_pass SRCS sequential_execution_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)
 
-if(WITH_GPU)
-  cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto reference_count_pass)
-else()
-  cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto)
+set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass) 
+if (WITH_GPU)
+  list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
 endif()
 
+cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
+
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
         simple_threadpool device_context)
 

paddle/fluid/framework/details/build_strategy.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
+#include "paddle/fluid/framework/details/sequential_execution_pass.h"
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/graph_viz_pass.h"
 
@@ -27,6 +28,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
  public:
   explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy)
       : ir::PassBuilder(), strategy_(strategy) {
+    if (strategy_.enable_sequential_execution_) {
+      AppendPass("sequential_execution_pass");
+    }
+
     // Add a graph viz pass to record a graph.
     if (!strategy_.debug_graphviz_path_.empty()) {
       auto viz_pass = AppendPass("graph_viz_pass");
@@ -64,6 +69,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
 
     // Verify that the graph is correct for multi-device executor.
     AppendPass("multi_devices_check_pass");
+
+    if (strategy_.remove_unnecessary_lock_) {
+      AppendPass("modify_op_lock_and_record_event_pass");
+    }
   }
 
  private:
@@ -110,6 +119,11 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
       pass->Erase("nccl_ctxs");
       pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
 #endif
+    } else if (pass->Type() == "sequential_execution_pass") {
+      pass->Erase(kAllOpDescs);
+      pass->Set<const std::vector<OpDesc *>>(
+          kAllOpDescs,
+          new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
     }
     graph = pass->Apply(std::move(graph));
   }
@@ -125,3 +139,5 @@ USE_PASS(multi_batch_merge_pass);
 USE_PASS(multi_devices_pass);
 USE_PASS(multi_devices_check_pass);
 USE_PASS(multi_devices_print_pass);
+USE_PASS(sequential_execution_pass);
+USE_PASS(modify_op_lock_and_record_event_pass);

paddle/fluid/framework/details/build_strategy.h

@@ -69,8 +69,12 @@ struct BuildStrategy {
 
   bool enable_data_balance_{false};
 
+  bool enable_sequential_execution_{false};
+
   bool fuse_broadcast_op_{false};
 
+  bool remove_unnecessary_lock_{false};
+
   // User normally doesn't need to call this API.
   // The PassBuilder allows for more customized insert, remove of passes
   // from python side.

paddle/fluid/framework/details/computation_op_handle.cc

@@ -29,9 +29,15 @@ ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope,
 void ComputationOpHandle::RunImpl() {
   WaitInputVarGenerated(place_);
 
-  this->RunAndRecordEvent([this] {
+  auto run_func = [this]() {
     op_->Run(*scope_->FindVar(kLocalExecScopeName)->Get<Scope *>(), place_);
-  });
+  };
+
+  if (is_lock_and_record_event_free_) {
+    run_func();
+  } else {
+    this->RunAndRecordEvent(run_func);
+  }
 }
 
 bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) {

paddle/fluid/framework/details/computation_op_handle.h

@@ -36,6 +36,8 @@ struct ComputationOpHandle : public OpHandleBase {
 
   const platform::Place &GetPlace() const { return place_; }
 
+  void SetLockAndRecordEventFree(bool b) { is_lock_and_record_event_free_ = b; }
+
  protected:
   void RunImpl() override;
 
@@ -45,6 +47,7 @@ struct ComputationOpHandle : public OpHandleBase {
   std::unique_ptr<OperatorBase> op_;
   Scope *scope_;
   platform::Place place_;
+  bool is_lock_and_record_event_free_{false};
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc

@@ -0,0 +1,59 @@
+// 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/modify_op_lock_and_record_event_pass.h"
+#include "paddle/fluid/framework/details/computation_op_handle.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/details/op_graph_view.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+static bool IsLockAndRecordEventFreeComputationOpHandle(
+    ComputationOpHandle *op, const OpGraphView &graph_view) {
+  if (!platform::is_gpu_place(op->GetPlace())) return false;
+  for (auto &pending_op : graph_view.PendingOps(op)) {
+    auto *tmp = dynamic_cast<ComputationOpHandle *>(pending_op);
+    if (tmp == nullptr || !(tmp->GetPlace() == op->GetPlace())) {
+      return false;
+    }
+  }
+  return true;
+}
+
+std::unique_ptr<ir::Graph> ModifyOpLockAndRecordEventPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> ir_graph) const {
+  auto &all_ops = ir_graph->Get<GraphOps>(kGraphOps);
+  OpGraphView graph_view(all_ops);
+  for (auto &op : all_ops) {
+    auto *compute_op = dynamic_cast<ComputationOpHandle *>(op.get());
+    if (compute_op == nullptr) continue;
+    bool is_lock_and_record_event_free =
+        IsLockAndRecordEventFreeComputationOpHandle(compute_op, graph_view);
+    compute_op->SetLockAndRecordEventFree(is_lock_and_record_event_free);
+    if (is_lock_and_record_event_free) {
+      VLOG(10) << "Set is_lock_and_record_event_free be true in op "
+               << compute_op->DebugString();
+    }
+  }
+  return ir_graph;
+}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(modify_op_lock_and_record_event_pass,
+              paddle::framework::details::ModifyOpLockAndRecordEventPass);