Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into tile_op_npu

Author: rainyfly

paddle/fluid/distributed/common/sparse_sharding_merge.h

@@ -21,7 +21,6 @@
 #include <vector>
 
 #include <ThreadPool.h>
-#include "boost/lexical_cast.hpp"
 #include "glog/logging.h"
 #include "paddle/fluid/distributed/common/utils.h"
 #include "paddle/fluid/framework/blocking_queue.h"
@@ -36,8 +35,6 @@ constexpr int Q_SIZE = 10000;
 constexpr int BUCKET = 10;
 constexpr char XEOF[] = "EOF";
 
-using boost::lexical_cast;
-
 inline double GetCurrentUS() {
   struct timeval time;
   gettimeofday(&time, NULL);
@@ -208,8 +205,10 @@ class ShardingMerge {
         for (int x = 0; x < embedding_dim; ++x) {
           float v = 0.0;
           try {
-            v = lexical_cast<float>(values_str[x]);
-          } catch (boost::bad_lexical_cast &e) {
+            v = std::stof(values_str[x]);
+          } catch (std::invalid_argument &e) {
+            VLOG(0) << " get unexpected line: " << line;
+          } catch (std::out_of_range &e) {
             VLOG(0) << " get unexpected line: " << line;
           }
           out->push_back(v);

paddle/fluid/distributed/index_dataset/index_wrapper.cc

@@ -17,8 +17,6 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/framework/io/fs.h"
 
-#include <boost/algorithm/string.hpp>
-#include <boost/lexical_cast.hpp>
 #include "paddle/fluid/distributed/index_dataset/index_wrapper.h"
 
 namespace paddle {
@@ -65,7 +63,7 @@ int TreeIndex::Load(const std::string filename) {
     if (item.key() == ".tree_meta") {
       meta_.ParseFromString(item.value());
     } else {
-      auto code = boost::lexical_cast<uint64_t>(item.key());
+      auto code = std::stoull(item.key());
       IndexNode node;
       node.ParseFromString(item.value());
       PADDLE_ENFORCE_NE(node.id(), 0,

paddle/fluid/distributed/table/common_sparse_table.cc

@@ -15,7 +15,6 @@
 #include "paddle/fluid/distributed/table/common_sparse_table.h"
 #include <sstream>
 
-#include "boost/lexical_cast.hpp"
 #include "glog/logging.h"
 #include "paddle/fluid/platform/enforce.h"
 
@@ -50,8 +49,11 @@ void CommonSparseTable::ProcessALine(const std::vector<std::string>& columns,
       float v = 0.0;
 
       try {
-        v = lexical_cast<float>(va);
-      } catch (boost::bad_lexical_cast& e) {
+        v = std::stof(va);
+      } catch (std::invalid_argument& e) {
+        VLOG(0) << "id: " << id << " get unexpected value: " << va
+                << " and be reset to: 0.0";
+      } catch (std::out_of_range& e) {
         VLOG(0) << "id: " << id << " get unexpected value: " << va
                 << " and be reset to: 0.0";
       }
@@ -131,7 +133,7 @@ int64_t CommonSparseTable::LoadFromText(
 
   while (std::getline(file, line)) {
     auto values = paddle::string::split_string<std::string>(line, "\t");
-    auto id = lexical_cast<uint64_t>(values[0]);
+    auto id = std::stoull(values[0]);
 
     if (id % pserver_num != pserver_id) {
       VLOG(3) << "will not load " << values[0] << " from " << valuepath
@@ -150,10 +152,9 @@ int64_t CommonSparseTable::LoadFromText(
     VALUE* value_instant = block->GetValue(id);
 
     if (values.size() == 5) {
-      value_instant->count_ = lexical_cast<int>(values[1]);
-      value_instant->unseen_days_ = lexical_cast<int>(values[2]);
-      value_instant->is_entry_ =
-          static_cast<bool>(lexical_cast<int>(values[3]));
+      value_instant->count_ = std::stoi(values[1]);
+      value_instant->unseen_days_ = std::stoi(values[2]);
+      value_instant->is_entry_ = static_cast<bool>(std::stoi(values[3]));
     }
 
     std::vector<float*> block_values = block->Get(id, meta.names, meta.dims);

paddle/fluid/distributed/table/common_sparse_table.h

@@ -33,7 +33,6 @@
 #include "paddle/fluid/string/string_helper.h"
 
 #define PSERVER_SAVE_SUFFIX ".shard"
-using boost::lexical_cast;
 
 namespace paddle {
 namespace distributed {

paddle/fluid/distributed/table/ssd_sparse_table.cc

@@ -310,7 +310,7 @@ int64_t SSDSparseTable::LoadFromText(
 
   while (std::getline(file, line)) {
     auto values = paddle::string::split_string<std::string>(line, "\t");
-    auto id = lexical_cast<uint64_t>(values[0]);
+    auto id = std::stoull(values[0]);
 
     if (id % pserver_num != pserver_id) {
       VLOG(3) << "will not load " << values[0] << " from " << valuepath
@@ -329,10 +329,9 @@ int64_t SSDSparseTable::LoadFromText(
     VALUE* value_instant = block->GetValue(id);
 
     if (values.size() == 5) {
-      value_instant->count_ = lexical_cast<int>(values[1]);
-      value_instant->unseen_days_ = lexical_cast<int>(values[2]);
-      value_instant->is_entry_ =
-          static_cast<bool>(lexical_cast<int>(values[3]));
+      value_instant->count_ = std::stoi(values[1]);
+      value_instant->unseen_days_ = std::stoi(values[2]);
+      value_instant->is_entry_ = static_cast<bool>(std::stoi(values[3]));
     }
 
     std::vector<float*> block_values = block->Get(id, meta.names, meta.dims);

paddle/fluid/framework/CMakeLists.txt

@@ -427,6 +427,9 @@ else()
   cc_test(custom_tensor_test SRCS custom_tensor_test.cc DEPS custom_tensor glog)
 endif()
 
+#cc_binary(test_executor SRCS test_executor.cc DEPS executor op_registry ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} )
+#cc_binary(new_executor SRCS new_exec_test.cc DEPS operator op_registry executor ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS} profiler)
+
 set(FLUID_FRAMEWORK_MODULES proto_desc memory lod_tensor executor data_feed_proto layer dynamic_loader custom_operator)
 
 cc_library(paddle_framework DEPS ${FLUID_FRAMEWORK_MODULES})

paddle/fluid/framework/details/CMakeLists.txt

@@ -141,7 +141,7 @@ if(NOT APPLE AND NOT WIN32 AND (WITH_GPU OR WITH_ROCM))
 endif()
 cc_library(build_strategy SRCS build_strategy.cc DEPS pass_builder ${IR_PASS_DEPS})
 cc_test(build_strategy_test SRCS build_strategy_test.cc
-        DEPS build_strategy op_registry op_proto_maker graph)
+        DEPS build_strategy op_registry op_proto_maker graph string_helper)
 
 if (WITH_MKLDNN)
   target_link_libraries(build_strategy mkldnn_placement_pass)

paddle/fluid/framework/distributed_strategy.proto

@@ -183,7 +183,7 @@ message DistributedStrategy {
   optional bool use_hierarchical_allreduce = 15 [ default = false ];
   optional int32 hierarchical_allreduce_inter_nranks = 16 [ default = 1 ];
   optional bool sync_batch_norm = 17 [ default = false ];
-  optional bool fuse_all_reduce_ops = 18 [ default = false ];
+  optional bool fuse_all_reduce_ops = 18 [ default = true ];
   optional int32 fuse_grad_size_in_MB = 19 [ default = 32 ];
   optional float fuse_grad_size_in_TFLOPS = 20 [ default = 50 ];
   optional bool cudnn_exhaustive_search = 21 [ default = false ];

paddle/fluid/framework/fleet/fleet_wrapper.cc

@@ -262,7 +262,7 @@ void FleetWrapper::HeterPushSparseVars(
     int64_t* ids = tensor->data<int64_t>();
     int slot = 0;
     if (dump_slot) {
-      slot = boost::lexical_cast<int>(sparse_key_names[i]);
+      slot = std::stoi(sparse_key_names[i]);
     }
     Variable* g_var = scope.FindVar(sparse_grad_names[i]);
     if (g_var == nullptr) {
@@ -915,12 +915,17 @@ void FleetWrapper::PushSparseVarsWithLabelAsync(
     int slot = 0;
     if (dump_slot) {
       try {
-        slot = boost::lexical_cast<int>(sparse_key_names[i]);
-      } catch (boost::bad_lexical_cast const& e) {
+        slot = std::stoi(sparse_key_names[i]);
+      } catch (std::invalid_argument const& e) {
         PADDLE_THROW(platform::errors::PreconditionNotMet(
             "sparse var's name: %s, doesn't support non-integer type name when "
             "dump_slot=True",
             sparse_key_names[i]));
+      } catch (std::out_of_range const& e) {
+        PADDLE_THROW(platform::errors::PreconditionNotMet(
+            "sparse var's name: %s, integer type name out of range when "
+            "dump_slot=True",
+            sparse_key_names[i]));
       }
     }
     Variable* g_var = scope.FindVar(sparse_grad_names[i]);
@@ -1121,7 +1126,7 @@ void FleetWrapper::PushSparseFromTensorWithLabelAsync(
         data[click_index] = static_cast<float>(fea_labels.at(input_idx));
       }
       if (dump_slot) {
-        int slot = boost::lexical_cast<int>(input_names[index]);
+        int slot = std::stoi(input_names[index]);
         data[0] = static_cast<float>(slot);
       }
       ++input_idx;

paddle/fluid/framework/ir/adaptive_pool2d_convert_global_pass.cc

@@ -60,6 +60,7 @@ AdaptivePool2dConvertGlobalPass::AdaptivePool2dConvertGlobalPass() {
       .IsStringIn({"NHWC", "NCHW"})
       .End()
       .AddAttr("padding_algorithm")
+      .IsOptional()
       .IsStringIn({"EXPLICIT", "SAME", "VALID"})
       .End();
 }