Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into dist-table-support-optimizer-regular

paddle/fluid/framework/details/execution_strategy.h

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #pragma once
+#include <cstddef>  // for size_t
 
 namespace paddle {
 namespace framework {
@@ -26,6 +27,7 @@ struct ExecutionStrategy {
   bool allow_op_delay_{false};
   size_t num_iteration_per_drop_scope_{100};
   ExecutorType type_{kDefault};
+  bool dry_run_{false};
 };
 
 }  //  namespace details

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -128,7 +128,9 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
     size_t complete = 0;
     while (op_to_run != nullptr) {
       try {
-        op_to_run->Run(strategy_.use_cuda_);
+        if (LIKELY(!strategy_.dry_run_)) {
+          op_to_run->Run(strategy_.use_cuda_);
+        }
         ++complete;
       } catch (...) {
         exception_.Catch(std::current_exception());

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -211,7 +211,9 @@ void ThreadedSSAGraphExecutor::RunOp(
       if (VLOG_IS_ON(10)) {
         VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
       }
-      op->Run(strategy_.use_cuda_);
+      if (LIKELY(!strategy_.dry_run_)) {
+        op->Run(strategy_.use_cuda_);
+      }
       VLOG(10) << op << " " << op->Name() << " Done ";
       running_ops_--;
       ready_var_q->Extend(op->Outputs());

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -48,7 +48,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
   // Use topological sort algorithm
   FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
 
-  ~ThreadedSSAGraphExecutor() {}
+  ~ThreadedSSAGraphExecutor() final = default;
 
  private:
   void RunOp(const std::shared_ptr<BlockingQueue<VarHandleBase *>> &ready_var_q,

paddle/fluid/framework/parallel_executor.cc

@@ -38,9 +38,20 @@ class ParallelExecutorPrivate {
   explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
       : places_(places) {}
 
+  ~ParallelExecutorPrivate() {
+    if (own_local_scope_) {
+      for (size_t i = 1; i < local_scopes_.size(); ++i) {
+        // Skip the first scope, since it is the global scope.
+        Scope *local_scope = local_scopes_[i];
+        if (global_scope_->HasKid(local_scope)) {
+          global_scope_->DeleteScope(local_scope);
+        }
+      }
+    }
+  }
   std::vector<platform::Place> places_;
   std::vector<Scope *> local_scopes_;
-  Scope *global_scope_;
+  Scope *global_scope_;  // not owned
   std::unique_ptr<details::SSAGraphExecutor> executor_;
 
 #ifdef PADDLE_WITH_CUDA
@@ -306,16 +317,6 @@ ParallelExecutor::~ParallelExecutor() {
   for (auto &p : member_->places_) {
     platform::DeviceContextPool::Instance().Get(p)->Wait();
   }
-
-  if (member_->own_local_scope_) {
-    for (size_t i = 1; i < member_->local_scopes_.size(); ++i) {
-      Scope *local_scope = member_->local_scopes_[i];
-      if (member_->global_scope_->HasKid(local_scope)) {
-        member_->global_scope_->DeleteScope(local_scope);
-      }
-    }
-  }
-
   // member_ must be destructed before gcs_ since the destructor of
   // ReferenceCountOpHandle use raw pointers of gcs_ inside.
   member_.reset();

paddle/fluid/inference/CMakeLists.txt

@@ -1,5 +1,5 @@
 if(WITH_TESTING)
-  include(test.cmake) # some generic cmake funtion for inference
+  include(tests/test.cmake) # some generic cmake funtion for inference
 endif()
 # analysis and tensorrt must be added before creating static library,
 # otherwise, there would be undefined reference to them in static library.

paddle/fluid/inference/tensorrt/convert/conv2d_op.cc

@@ -18,6 +18,21 @@ namespace paddle {
 namespace inference {
 namespace tensorrt {
 
+bool to_skip_merging_optimize(TensorRTEngine* engine_,
+                              const std::vector<int>& filters,
+                              const std::vector<int>& strides,
+                              const std::vector<int>& paddings,
+                              std::string input_name) {
+  if (engine_->itensor_quote_num[input_name] > 0) {
+    return true;
+  }
+  if (filters[0] == 1 && filters[1] == 1 && strides[0] == 1 &&
+      strides[1] == 1 && paddings[0] == 0 && paddings[1] == 0)
+    engine_->itensor_quote_num[input_name] += 1;
+
+  return false;
+}
+
 class Conv2dOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
@@ -31,6 +46,7 @@ class Conv2dOpConverter : public OpConverter {
     PADDLE_ENFORCE_EQ(op_desc.Output("Output").size(), 1);
 
     auto* X = engine_->GetITensor(op_desc.Input("Input").front());
+
     // Declare weights
     auto* Y_v = scope.FindVar(op_desc.Input("Filter").front());
     PADDLE_ENFORCE_NOT_NULL(Y_v);
@@ -83,7 +99,10 @@ class Conv2dOpConverter : public OpConverter {
         std::move(weight_tensor);
     layer->getOutput(0)->setName(output_name.c_str());
     engine_->SetITensor(output_name, layer->getOutput(0));
-    if (test_mode) {
+
+    if (test_mode ||
+        to_skip_merging_optimize(engine_, {filter_h, filter_w}, strides,
+                                 paddings, op_desc.Input("Input").front())) {
       engine_->DeclareOutput(output_name);
     }
   }

paddle/fluid/inference/tensorrt/engine.cc

@@ -133,6 +133,10 @@ void TensorRTEngine::DeclareOutput(const nvinfer1::ILayer *layer, int offset,
   buffer_sizes_[name] = 0;
 }
 
+bool TensorRTEngine::HasDeclared(const std::string &name) {
+  return buffer_sizes_.count(name) > 0;
+}
+
 void TensorRTEngine::DeclareOutput(const std::string &name) {
   PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate output name %s",
                     name);

paddle/fluid/inference/tensorrt/engine.h

@@ -91,6 +91,8 @@ class TensorRTEngine : public EngineBase {
                      const std::string& name);
   // Set the itensor_map_[name] as the network's output, and set its name.
   void DeclareOutput(const std::string& name);
+  // Check if the ITensor has been declared
+  bool HasDeclared(const std::string& name);
 
   // GPU memory address for an ITensor with specific name. One can operate on
   // these memory directly for acceleration, for example, output the converted
@@ -132,6 +134,16 @@ class TensorRTEngine : public EngineBase {
   std::unordered_map<std::string /*name*/, std::unique_ptr<framework::Tensor>>
       weight_map;
 
+  // TODO: (NHZLX)
+  // In the normal case, the paddle-trt exists bug when runing the googlenet.
+  // When there are more than two convolutions of 1 * 1 with the same input, the
+  // paddle-tensorrt will do the merging optimization, which fuse those conv
+  // into
+  // one conv, and then trigger bug. So,  We should use strategy to avoid this
+  // optimization for the time being. This bug will be fixed in the future.
+  std::unordered_map<std::string /*name*/, int /*ITensor_quote_num*/>
+      itensor_quote_num;
+
  private:
   // the max batch size
   int max_batch_;

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -1,5 +1,11 @@
 set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor)
 
+function(download_model install_dir model_name)
+    if (NOT EXISTS ${install_dir})
+        inference_download_and_uncompress(${install_dir} ${INFERENCE_URL} ${model_name})
+    endif()
+endfunction()
+
 function(download_model_and_data install_dir model_name data_name)
     if (NOT EXISTS ${install_dir})
         inference_download_and_uncompress(${install_dir} ${INFERENCE_URL} ${model_name})
@@ -13,6 +19,13 @@ function(inference_analysis_api_test target install_dir filename)
         ARGS --infer_model=${install_dir}/model --infer_data=${install_dir}/data.txt)
 endfunction()
 
+function(inference_analysis_api_test_with_fake_data target install_dir filename model_name)
+    download_model(${install_dir} ${model_name})
+    inference_analysis_test(${target} SRCS ${filename}
+        EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
+        ARGS --infer_model=${install_dir}/model)
+endfunction()
+
 # RNN1
 if(NOT APPLE)
     set(RNN1_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn1")
@@ -61,17 +74,13 @@ inference_analysis_api_test(test_analyzer_seq_conv1 ${SEQ_CONV1_INSTALL_DIR} ana
 # ocr
 set(OCR_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/ocr")
 if (NOT EXISTS ${OCR_INSTALL_DIR})
-    inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz")
+  inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz")
 endif()
 inference_analysis_api_test(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc)
 
 # resnet50
-set(RESNET50_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/resnet50")
-if (NOT EXISTS ${RESNET50_INSTALL_DIR})
-    inference_download_and_uncompress(${RESNET50_INSTALL_DIR} ${INFERENCE_URL} "resnet50_model.tar.gz")
-endif()
-inference_analysis_test(test_analyzer_resnet50 SRCS analyzer_resnet50_tester.cc
-    EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} ARGS --infer_model=${RESNET50_INSTALL_DIR}/model)
+inference_analysis_api_test_with_fake_data(test_analyzer_resnet50 
+  "${INFERENCE_DEMO_INSTALL_DIR}/resnet50" analyzer_resnet50_tester.cc "resnet50_model.tar.gz")
 
 # anakin
 if (WITH_ANAKIN AND WITH_MKL) # only needed in CI

paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc

@@ -30,25 +30,7 @@ void SetConfig(AnalysisConfig *cfg) {
 }
 
 void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
-  PADDLE_ENFORCE_EQ(FLAGS_test_all_data, 0, "Only have single batch of data.");
-
-  PaddleTensor input;
-  // channel=3, height/width=318
-  std::vector<int> shape({FLAGS_batch_size, 3, 318, 318});
-  input.shape = shape;
-  input.dtype = PaddleDType::FLOAT32;
-
-  // fill input data, for profile easily, do not use random data here.
-  size_t size = FLAGS_batch_size * 3 * 318 * 318;
-  input.data.Resize(size * sizeof(float));
-  float *input_data = static_cast<float *>(input.data.data());
-  for (size_t i = 0; i < size; i++) {
-    *(input_data + i) = static_cast<float>(i) / size;
-  }
-
-  std::vector<PaddleTensor> input_slots;
-  input_slots.assign({input});
-  (*inputs).emplace_back(input_slots);
+  SetFakeImageInput(inputs, FLAGS_infer_model);
 }
 
 // Easy for profiling independently.
@@ -61,13 +43,6 @@ void profile(bool use_mkldnn = false) {
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
   TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
-
-  if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
-    PADDLE_ENFORCE_EQ(outputs.size(), 1UL);
-    size_t size = GetSize(outputs[0]);
-    // output is a 512-dimension feature
-    EXPECT_EQ(size, 512 * FLAGS_batch_size);
-  }
 }
 
 TEST(Analyzer_resnet50, profile) { profile(); }
@@ -83,8 +58,7 @@ TEST(Analyzer_resnet50, fuse_statis) {
   auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
   auto fuse_statis = GetFuseStatis(
       static_cast<AnalysisPredictor *>(predictor.get()), &num_ops);
-  ASSERT_TRUE(fuse_statis.count("fc_fuse"));
-  EXPECT_EQ(fuse_statis.at("fc_fuse"), 1);
+  LOG(INFO) << "num_ops: " << num_ops;
 }
 
 // Compare result of NativeConfig and AnalysisConfig

paddle/fluid/inference/tests/api/tester_helper.h

@@ -25,6 +25,7 @@
 #include "paddle/fluid/inference/api/analysis_predictor.h"
 #include "paddle/fluid/inference/api/helper.h"
 #include "paddle/fluid/inference/api/paddle_inference_pass.h"
+#include "paddle/fluid/inference/tests/test_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
 DEFINE_string(infer_model, "", "model path");
@@ -105,6 +106,34 @@ std::unordered_map<std::string, int> GetFuseStatis(PaddlePredictor *predictor,
   return fuse_statis;
 }
 
+void SetFakeImageInput(std::vector<std::vector<PaddleTensor>> *inputs,
+                       const std::string &dirname) {
+  // Set fake_image_data
+  PADDLE_ENFORCE_EQ(FLAGS_test_all_data, 0, "Only have single batch of data.");
+  std::vector<std::vector<int64_t>> feed_target_shapes =
+      GetFeedTargetShapes(dirname, true, "model", "params");
+  int dim1 = feed_target_shapes[0][1];
+  int dim2 = feed_target_shapes[0][2];
+  int dim3 = feed_target_shapes[0][3];
+
+  PaddleTensor input;
+  std::vector<int> shape({FLAGS_batch_size, dim1, dim2, dim3});
+  input.shape = shape;
+  input.dtype = PaddleDType::FLOAT32;
+
+  // fill input data, for profile easily, do not use random data here.
+  size_t size = FLAGS_batch_size * dim1 * dim2 * dim3;
+  input.data.Resize(size * sizeof(float));
+  float *input_data = static_cast<float *>(input.data.data());
+  for (size_t i = 0; i < size; i++) {
+    *(input_data + i) = static_cast<float>(i) / size;
+  }
+
+  std::vector<PaddleTensor> input_slots;
+  input_slots.assign({input});
+  (*inputs).emplace_back(input_slots);
+}
+
 void TestOneThreadPrediction(
     const AnalysisConfig &config,
     const std::vector<std::vector<PaddleTensor>> &inputs,

paddle/fluid/inference/tests/api/trt_models_tester.cc

@@ -93,11 +93,16 @@ void CompareTensorRTWithFluid(int batch_size, std::string model_dirname) {
   }
 }
 
-TEST(trt_models_test, main) {
-  std::vector<std::string> infer_models = {"mobilenet", "resnet50",
-                                           "resnext50"};
-  for (auto &model_dir : infer_models) {
-    CompareTensorRTWithFluid(1, FLAGS_dirname + "/" + model_dir);
-  }
+TEST(trt_models_test, mobilenet) {
+  CompareTensorRTWithFluid(1, FLAGS_dirname + "/mobilenet");
+}
+
+TEST(trt_models_test, resnet50) {
+  CompareTensorRTWithFluid(1, FLAGS_dirname + "/resnet50");
 }
+
+TEST(trt_models_test, resnext50) {
+  CompareTensorRTWithFluid(1, FLAGS_dirname + "/resnext50");
+}
+
 }  // namespace paddle

paddle/fluid/inference/tests/test_helper.h

@@ -18,7 +18,6 @@ limitations under the License. */
 #include <string>
 #include <vector>
 
-#include "paddle/fluid/framework/ir/graph_to_program_pass.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/inference/io.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -94,15 +93,15 @@ void CheckError(const paddle::framework::LoDTensor& output1,
 
 std::unique_ptr<paddle::framework::ProgramDesc> InitProgram(
     paddle::framework::Executor* executor, paddle::framework::Scope* scope,
-    const std::string& dirname, const bool is_combined = false) {
+    const std::string& dirname, const bool is_combined = false,
+    const std::string& prog_filename = "__model_combined__",
+    const std::string& param_filename = "__params_combined__") {
   std::unique_ptr<paddle::framework::ProgramDesc> inference_program;
   if (is_combined) {
     // All parameters are saved in a single file.
     // Hard-coding the file names of program and parameters in unittest.
     // The file names should be consistent with that used in Python API
     //  `fluid.io.save_inference_model`.
-    std::string prog_filename = "__model_combined__";
-    std::string param_filename = "__params_combined__";
     inference_program =
         paddle::inference::Load(executor, scope, dirname + "/" + prog_filename,
                                 dirname + "/" + param_filename);
@@ -115,12 +114,15 @@ std::unique_ptr<paddle::framework::ProgramDesc> InitProgram(
 }
 
 std::vector<std::vector<int64_t>> GetFeedTargetShapes(
-    const std::string& dirname, const bool is_combined = false) {
+    const std::string& dirname, const bool is_combined = false,
+    const std::string& prog_filename = "__model_combined__",
+    const std::string& param_filename = "__params_combined__") {
   auto place = paddle::platform::CPUPlace();
   auto executor = paddle::framework::Executor(place);
   auto* scope = new paddle::framework::Scope();
 
-  auto inference_program = InitProgram(&executor, scope, dirname, is_combined);
+  auto inference_program = InitProgram(&executor, scope, dirname, is_combined,
+                                       prog_filename, param_filename);
   auto& global_block = inference_program->Block(0);
 
   const std::vector<std::string>& feed_target_names =
@@ -136,15 +138,6 @@ std::vector<std::vector<int64_t>> GetFeedTargetShapes(
   return feed_target_shapes;
 }
 
-void Compile(paddle::framework::ProgramDesc* program) {
-  std::unique_ptr<paddle::framework::ir::Graph> g(
-      new paddle::framework::ir::Graph(*program));
-  auto pass = paddle::framework::ir::PassRegistry::Instance().Get(
-      "graph_to_program_pass");
-  pass->SetNotOwned<paddle::framework::ProgramDesc>("program", program);
-  pass->Apply(std::move(g));
-}
-
 template <typename Place, bool CreateVars = true, bool PrepareContext = false>
 void TestInference(const std::string& dirname,
                    const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
@@ -182,7 +175,6 @@ void TestInference(const std::string& dirname,
         paddle::platform::DeviceContextPool::Instance().Get(place));
     inference_program = InitProgram(&executor, scope, dirname, is_combined);
   }
-  Compile(inference_program.get());
 
   // Disable the profiler and print the timing information
   paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,
@@ -261,5 +253,3 @@ void TestInference(const std::string& dirname,
 
   delete scope;
 }
-
-USE_PASS(graph_to_program_pass);

paddle/fluid/operators/distributed/grpc_variable_response.cc

@@ -286,10 +286,10 @@ int GRPCVariableResponse::Parse(Source* source) {
           platform::EnableProfiler(platform::ProfilerState::kCPU);
         } else if (profiling == platform::kDisableProfiler &&
                    platform::IsProfileEnabled()) {
-          // TODO(panyx0718): Should we allow to customize file dir.
           platform::DisableProfiler(
               platform::EventSortingKey::kDefault,
-              string::Sprintf("/tmp/profile_ps_%lld", listener_id));
+              string::Sprintf("%s_%lld", FLAGS_rpc_server_profile_path,
+                              listener_id));
         }
         break;
       }

paddle/fluid/operators/distributed/request_handler_impl.cc

@@ -51,7 +51,6 @@ bool RequestSendHandler::Handle(const std::string& varname,
     // Async
     if (!sync_mode_) {
       VLOG(3) << "async process var: " << varname;
-      rpc_server_->Profiler().OneStep();
       try {
         executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(),
                                       scope);

paddle/fluid/operators/distributed/rpc_server.cc

@@ -20,42 +20,10 @@
 #include "paddle/fluid/operators/distributed/rpc_server.h"
 #include "paddle/fluid/platform/profiler.h"
 
-DEFINE_int32(rpc_server_profile_period, 0,
-             "the period of listen_and_serv to do profile");
-DEFINE_string(rpc_server_profile_path, "/dev/null",
-              "the profile log file path");
-
 namespace paddle {
 namespace operators {
 namespace distributed {
 
-RPCServerProfiler::RPCServerProfiler(int profile_period,
-                                     const std::string& profile_log_path)
-    : profile_period_(profile_period), profile_log_path_(profile_log_path) {
-  step_ = 0;
-}
-
-void RPCServerProfiler::OneStep() {
-  PADDLE_ENFORCE_LE(step_, profile_period_,
-                    "step_ should not be larger then "
-                    "profile_period_");
-  if (profile_period_ <= 0) {
-    return;
-  }
-
-  if (step_ == 0) {
-    auto pf_state = paddle::platform::ProfilerState::kCPU;
-    paddle::platform::EnableProfiler(pf_state);
-  }
-  if (step_ == profile_period_) {
-    paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kTotal,
-                                      profile_log_path_);
-    step_ = 0;
-  } else {
-    step_++;
-  }
-}
-
 void RPCServer::ShutDown() {
   LOG(INFO) << "RPCServer ShutDown ";
   ShutDownImpl();

paddle/fluid/operators/distributed/rpc_server.h

@@ -23,30 +23,14 @@
 
 #include "paddle/fluid/operators/distributed/request_handler.h"
 
-DECLARE_int32(rpc_server_profile_period);
-DECLARE_string(rpc_server_profile_path);
-
 namespace paddle {
 namespace operators {
 namespace distributed {
 
-class RPCServerProfiler {
- public:
-  RPCServerProfiler(int profile_period, const std::string& profile_log_path);
-  void OneStep();
-
- private:
-  const int profile_period_;
-  std::string profile_log_path_;
-  int step_;
-};
-
 class RPCServer {
  public:
   explicit RPCServer(const std::string& address, int client_num)
       : cur_cond_(0),
-        profiler_(FLAGS_rpc_server_profile_period,
-                  FLAGS_rpc_server_profile_path),
         bind_address_(address),
         exit_flag_(false),
         selected_port_(0),
@@ -86,7 +70,6 @@ class RPCServer {
   void Complete();
 
   void ResetBarrierCounter();
-  RPCServerProfiler& Profiler() { return profiler_; }
 
   bool NeedResetAllVars();
 
@@ -101,7 +84,6 @@ class RPCServer {
   std::unordered_map<std::string, int> rpc_cond_map_;
   std::atomic<int> cur_cond_;
   std::condition_variable rpc_cond_;
-  RPCServerProfiler profiler_;
 
  protected:
   std::string bind_address_;

paddle/fluid/operators/distributed/variable_response.cc

@@ -16,6 +16,9 @@
 #include <vector>
 #include "paddle/fluid/operators/distributed/sendrecvop_utils.h"
 
+DEFINE_string(rpc_server_profile_path, "./profile_ps",
+              "the profile log file path");
+
 namespace paddle {
 namespace operators {
 namespace distributed {

paddle/fluid/operators/distributed/variable_response.h

@@ -27,6 +27,8 @@
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/operators/distributed/send_recv.pb.h"
 
+DECLARE_string(rpc_server_profile_path);
+
 namespace paddle {
 namespace operators {
 namespace distributed {

paddle/fluid/operators/listen_and_serv_op.cc

@@ -134,7 +134,6 @@ void ListenAndServOp::RunSyncLoop(
   rpc_service_->ResetBarrierCounter();
 
   while (true) {
-    rpc_service_->Profiler().OneStep();
     // Get from multiple trainers, we don't care about the order in which
     // the gradients arrives, just add suffix 0~n and merge the gradient.
     rpc_service_->SetCond(distributed::kRequestSend);

paddle/fluid/operators/math/cross_entropy.cu

@@ -28,7 +28,7 @@ __device__ __forceinline__ double real_log(double x) { return log(x); }
 
 __device__ __forceinline__ platform::float16 real_log(
     const platform::float16& val) {
-  return static_cast<platform::float16>(hlog(static_cast<half>(val)));
+  return static_cast<platform::float16>(logf(static_cast<float>(val)));
 }
 
 template <typename T>

paddle/fluid/operators/math/fc_compute.h

@@ -36,7 +36,7 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
                                .template Get<jitkernel::VAddReluKernel<T>>(N);
     for (int i = 0; i < M; i++) {
       T* dst = Y + i * N;
-      vaddrelu->Compute(B, dst, dst);
+      vaddrelu->Compute(B, dst, dst, N);
     }
   } else {
     const auto& vadd = jitkernel::KernelPool::Instance()
@@ -47,7 +47,7 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
 #endif
     for (int i = 0; i < M; i++) {
       T* dst = Y + i * N;
-      vadd->Compute(B, dst, dst);
+      vadd->Compute(B, dst, dst, N);
     }
   }
 }

paddle/fluid/operators/math/jit_code.cc

@@ -24,19 +24,29 @@ namespace gen {
 
 using namespace platform::jit;  // NOLINT
 
-bool VMulJitCode::init(int d) {
+bool VVVJitCode::init(int d) {
   // It's not necessary to use avx512 since it would slow down the frequency
   // and this kernel is not compute bound.
   return MayIUse(avx);
 }
 
-void VMulJitCode::generate() {
+void VVVJitCode::generate() {
   // do not need push stack, and do not need save avx512reg if do not use avx512
   int offset = 0;
+  if (with_relu_) {
+    vxorps(ymm_zero, ymm_zero, ymm_zero);
+  }
   for (int i = 0; i < num_ / AVX_FLOAT_BLOCK; ++i) {
     vmovups(ymm_src1, ptr[param1 + offset]);
     vmovups(ymm_src2, ptr[param2 + offset]);
-    vmulps(ymm_dst, ymm_src1, ymm_src2);
+    if (type_ == operand_type::mul) {
+      vmulps(ymm_dst, ymm_src1, ymm_src2);
+    } else if (type_ == operand_type::add) {
+      vaddps(ymm_dst, ymm_src1, ymm_src2);
+    }
+    if (with_relu_) {
+      vmaxps(ymm_dst, ymm_zero, ymm_dst);
+    }
     vmovups(ptr[param3 + offset], ymm_dst);
     offset += sizeof(float) * AVX_FLOAT_BLOCK;
   }
@@ -44,7 +54,14 @@ void VMulJitCode::generate() {
   if (rest >= 4) {
     vmovups(xmm_src1, ptr[param1 + offset]);
     vmovups(xmm_src2, ptr[param2 + offset]);
-    vmulps(xmm_dst, xmm_src1, xmm_src2);
+    if (type_ == operand_type::mul) {
+      vmulps(xmm_dst, xmm_src1, xmm_src2);
+    } else if (type_ == operand_type::add) {
+      vaddps(xmm_dst, xmm_src1, xmm_src2);
+    }
+    if (with_relu_) {
+      vmaxps(xmm_dst, xmm_zero, xmm_dst);
+    }
     vmovups(ptr[param3 + offset], xmm_dst);
     offset += sizeof(float) * 4;
     rest -= 4;
@@ -52,7 +69,14 @@ void VMulJitCode::generate() {
   if (rest >= 2) {
     vmovq(xmm_src1, ptr[param1 + offset]);
     vmovq(xmm_src2, ptr[param2 + offset]);
-    vmulps(xmm_dst, xmm_src1, xmm_src2);
+    if (type_ == operand_type::mul) {
+      vmulps(xmm_dst, xmm_src1, xmm_src2);
+    } else if (type_ == operand_type::add) {
+      vaddps(xmm_dst, xmm_src1, xmm_src2);
+    }
+    if (with_relu_) {
+      vmaxps(xmm_dst, xmm_zero, xmm_dst);
+    }
     vmovq(ptr[param3 + offset], xmm_dst);
     offset += sizeof(float) * 2;
     rest -= 2;
@@ -60,12 +84,18 @@ void VMulJitCode::generate() {
   if (rest > 0) {
     vmovss(xmm_src1, ptr[param1 + offset]);
     vmovss(xmm_src2, ptr[param2 + offset]);
-    vmulss(xmm_dst, xmm_src1, xmm_src2);
+    if (type_ == operand_type::mul) {
+      vmulss(xmm_dst, xmm_src1, xmm_src2);
+    } else if (type_ == operand_type::add) {
+      vaddss(xmm_dst, xmm_src1, xmm_src2);
+    }
+    if (with_relu_) {
+      vmaxps(xmm_dst, xmm_zero, xmm_dst);
+    }
     vmovss(ptr[param3 + offset], xmm_dst);
   }
   ret();
 }
-
 }  // namespace gen
 }  // namespace jitkernel
 }  // namespace math