Merge pull request #13211 from luotao1/multi-thread2

 add multi-thread ut for ditu-rnn

paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc

@@ -13,13 +13,10 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/ir/attention_lstm_fuse_pass.h"
-
 #include <string>
-
 #include "paddle/fluid/framework/ir/graph_pattern_detector.h"
 #include "paddle/fluid/framework/ir/graph_viz_pass.h"
 #include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/inference/api/helper.h"
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -85,7 +85,7 @@ void GraphPatternDetector::operator()(Graph* graph,
   LOG(INFO) << "detect " << subgraphs.size() << " subgraph matches the pattern";
   int id = 0;
   for (auto& g : subgraphs) {
-    LOG(INFO) << "optimizing #" << id++ << " subgraph";
+    VLOG(3) << "optimizing #" << id++ << " subgraph";
     handler(g, graph);
   }
 }

paddle/fluid/inference/analysis/analyzer.cc

@@ -106,7 +106,6 @@ void Analyzer::Run(Argument* argument) {
     }
   }
   passes.push_back("graph_viz_pass");
-  // Ugly support fluid-to-ir-pass
   argument->Set(kFluidToIrPassesAttr, new std::vector<std::string>(passes));
 
   for (auto& x : data_) {

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -16,6 +16,7 @@
 
 #include <google/protobuf/text_format.h>
 #include <gtest/gtest.h>
+#include <thread>  // NOLINT
 #include "paddle/fluid/framework/ir/fuse_pass_base.h"
 #include "paddle/fluid/framework/ir/pass.h"
 #include "paddle/fluid/inference/analysis/ut_helper.h"
@@ -24,12 +25,12 @@
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/api/paddle_inference_pass.h"
 #include "paddle/fluid/inference/utils/singleton.h"
-#include "paddle/fluid/platform/profiler.h"
 
 DEFINE_string(infer_ditu_rnn_model, "", "model path for ditu RNN");
 DEFINE_string(infer_ditu_rnn_data, "", "data path for ditu RNN");
 DEFINE_int32(batch_size, 10, "batch size.");
 DEFINE_int32(repeat, 1, "Running the inference program repeat times.");
+DEFINE_int32(num_threads, 1, "Running the inference program in multi-threads.");
 
 namespace paddle {
 namespace inference {
@@ -220,39 +221,6 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
   }
 }
 
-std::string DescribeTensor(const PaddleTensor &tensor) {
-  std::stringstream os;
-  os << "Tensor [" << tensor.name << "]\n";
-  os << " - type: ";
-  switch (tensor.dtype) {
-    case PaddleDType::FLOAT32:
-      os << "float32";
-      break;
-    case PaddleDType::INT64:
-      os << "int64";
-      break;
-    default:
-      os << "unset";
-  }
-  os << '\n';
-
-  os << " - shape: " << to_string(tensor.shape) << '\n';
-  os << " - lod: ";
-  for (auto &l : tensor.lod) {
-    os << to_string(l) << "; ";
-  }
-  os << "\n";
-  os << " - data: ";
-
-  int dim = std::accumulate(tensor.shape.begin(), tensor.shape.end(), 1,
-                            [](int a, int b) { return a * b; });
-  for (int i = 0; i < dim; i++) {
-    os << static_cast<float *>(tensor.data.data())[i] << " ";
-  }
-  os << '\n';
-  return os.str();
-}
-
 }  // namespace
 
 const float ditu_rnn_target_data[] = {
@@ -266,11 +234,29 @@ const float ditu_rnn_target_data[] = {
     10.7286, 12.0595, 10.6672, 0,       0,       0,       0,       0,
     93.5771, 3.84641, 0,       0,       0,       0,       0,       0,
     169.426, 0,       0,       0,       0,       0,       0,       0};
+void CompareResult(const std::vector<PaddleTensor> &outputs,
+                   const std::vector<PaddleTensor> &base_outputs) {
+  PADDLE_ENFORCE_GT(outputs.size(), 0);
+  PADDLE_ENFORCE_EQ(outputs.size(), base_outputs.size());
+  for (size_t i = 0; i < outputs.size(); i++) {
+    auto &out = outputs[i];
+    auto &base_out = base_outputs[i];
+    size_t size = std::accumulate(out.shape.begin(), out.shape.end(), 1,
+                                  [](int a, int b) { return a * b; });
+    size_t size1 = std::accumulate(base_out.shape.begin(), base_out.shape.end(),
+                                   1, [](int a, int b) { return a * b; });
+    PADDLE_ENFORCE_EQ(size, size1);
+    PADDLE_ENFORCE_GT(size, 0);
+    float *data = static_cast<float *>(out.data.data());
+    float *base_data = static_cast<float *>(base_out.data.data());
+    for (size_t i = 0; i < size; i++) {
+      EXPECT_NEAR(data[i], base_data[i], 1e-3);
+    }
+  }
+}
 // Test with a really complicate model.
-void TestDituRNNPrediction(const std::string &model_path,
-                           const std::string &data_path, int batch_size,
-                           bool use_analysis, bool activate_ir,
-                           int num_times = 1) {
+void TestDituRNNPrediction(bool use_analysis, bool activate_ir,
+                           int num_threads) {
   AnalysisConfig config;
   config.prog_file = FLAGS_infer_ditu_rnn_model + "/__model__";
   config.param_file = FLAGS_infer_ditu_rnn_model + "/param";
@@ -281,6 +267,8 @@ void TestDituRNNPrediction(const std::string &model_path,
   PADDLE_ENFORCE(config.ir_mode ==
                  AnalysisConfig::IrPassMode::kExclude);  // default
   config.ir_passes.clear();  // Do not exclude any pass.
+  int batch_size = FLAGS_batch_size;
+  int num_times = FLAGS_repeat;
 
   auto base_predictor =
       CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
@@ -288,40 +276,55 @@ void TestDituRNNPrediction(const std::string &model_path,
       CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
           config);
   std::vector<PaddleTensor> input_slots;
-  DataRecord data(data_path, batch_size);
+  DataRecord data(FLAGS_infer_ditu_rnn_data, batch_size);
   // Prepare inputs.
   PrepareInputs(&input_slots, &data, batch_size);
   std::vector<PaddleTensor> outputs, base_outputs;
 
   base_predictor->Run(input_slots, &base_outputs);
 
-  Timer timer;
-  timer.tic();
-  for (int i = 0; i < num_times; i++) {
-    predictor->Run(input_slots, &outputs);
-  }
   LOG(INFO) << "===========profile result===========";
-  LOG(INFO) << "batch_size: " << batch_size << ", repeat: " << num_times
-            << ", latency: " << timer.toc() / num_times << "ms";
-  LOG(INFO) << "=====================================";
-
-  PADDLE_ENFORCE_GT(outputs.size(), 0);
-  PADDLE_ENFORCE_EQ(outputs.size(), base_outputs.size());
-  for (size_t i = 0; i < outputs.size(); i++) {
-    auto &out = outputs[i];
-    auto &base_out = base_outputs[i];
-    size_t size = std::accumulate(out.shape.begin(), out.shape.end(), 1,
-                                  [](int a, int b) { return a * b; });
-    size_t size1 = std::accumulate(base_out.shape.begin(), base_out.shape.end(),
-                                   1, [](int a, int b) { return a * b; });
-    PADDLE_ENFORCE_EQ(size, size1);
-    PADDLE_ENFORCE_GT(size, 0);
-    float *data = static_cast<float *>(out.data.data());
-    float *base_data = static_cast<float *>(base_out.data.data());
-    for (size_t j = 0; j < size; j++) {
-      EXPECT_NEAR(data[j], base_data[j], 1e-3);
+  if (num_threads == 1) {
+    // Prepare inputs.
+    Timer timer;
+    timer.tic();
+    for (int i = 0; i < num_times; i++) {
+      predictor->Run(input_slots, &outputs);
+    }
+    PrintTime(batch_size, num_times, 1, 0, timer.toc() / num_times);
+    CompareResult(outputs, base_outputs);
+  } else {
+    std::vector<std::thread> threads;
+    std::vector<std::unique_ptr<PaddlePredictor>> predictors;
+    // TODO(yanchunwei): Bug here, the analyzer phase can't be parallelled
+    // because AttentionLSTM's hard code nodeid will be damanged.
+    for (int tid = 0; tid < num_threads; ++tid) {
+      predictors.emplace_back(
+          CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
+              config));
+    }
+    for (int tid = 0; tid < num_threads; ++tid) {
+      threads.emplace_back([&, tid]() {
+        // Each thread should have local input_slots and outputs.
+        std::vector<PaddleTensor> input_slots;
+        DataRecord data(FLAGS_infer_ditu_rnn_data, batch_size);
+        PrepareInputs(&input_slots, &data, batch_size);
+        std::vector<PaddleTensor> outputs;
+        Timer timer;
+        timer.tic();
+        for (int i = 0; i < num_times; i++) {
+          predictors[tid]->Run(input_slots, &outputs);
+        }
+        PrintTime(batch_size, num_times, num_threads, tid,
+                  timer.toc() / num_times);
+        CompareResult(outputs, base_outputs);
+      });
+    }
+    for (int i = 0; i < num_threads; ++i) {
+      threads[i].join();
     }
   }
+  LOG(INFO) << "=====================================";
 
   if (use_analysis && activate_ir) {
     AnalysisPredictor *analysis_predictor =
@@ -350,25 +353,26 @@ void TestDituRNNPrediction(const std::string &model_path,
   }
 }
 
-// Directly infer with the original model.
-TEST(Analyzer, DituRNN_without_analysis) {
-  TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
-                        FLAGS_batch_size, false, false, FLAGS_repeat);
+// Inference with analysis and IR, easy for profiling independently.
+TEST(Analyzer, DituRNN) {
+  TestDituRNNPrediction(true, true, FLAGS_num_threads);
 }
 
-// Inference with the original model with the analysis turned on, the analysis
-// module will transform the program to a data flow graph.
-TEST(Analyzer, DituRNN_with_analysis) {
-  LOG(INFO) << "ditu rnn with analysis";
-  TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
-                        FLAGS_batch_size, true, false, FLAGS_repeat);
-}
-
-// Inference with analysis and IR. The IR module will fuse some large kernels.
-TEST(Analyzer, DituRNN_with_analysis_with_IR) {
-  LOG(INFO) << "ditu rnn with analysis and IR fuse";
-  TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
-                        FLAGS_batch_size, true, true, FLAGS_repeat);
+// Other unit-tests of DituRNN, test different options of use_analysis,
+// activate_ir and multi-threads.
+TEST(Analyzer, DituRNN_tests) {
+  int num_threads[2] = {1, 4};
+  for (auto i : num_threads) {
+    // Directly infer with the original model.
+    TestDituRNNPrediction(false, false, i);
+    // Inference with the original model with the analysis turned on, the
+    // analysis
+    // module will transform the program to a data flow graph.
+    TestDituRNNPrediction(true, false, i);
+    // Inference with analysis and IR. The IR module will fuse some large
+    // kernels.
+    TestDituRNNPrediction(true, true, i);
+  }
 }
 
 }  // namespace analysis

paddle/fluid/inference/api/analysis_predictor.cc

@@ -35,7 +35,6 @@ bool AnalysisPredictor::Init(
   } else {
     place_ = paddle::platform::CPUPlace();
   }
-  PADDLE_ENFORCE(!parent_scope);
   if (parent_scope) {
     scope_ = parent_scope;
     sub_scope_ = &(parent_scope->NewScope());

paddle/fluid/inference/api/helper.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <glog/logging.h>
 #include <sys/time.h>
 #include <algorithm>
 #include <numeric>
@@ -88,5 +89,45 @@ static void TensorAssignData(PaddleTensor *tensor,
   }
 }
 
+std::string DescribeTensor(const PaddleTensor &tensor) {
+  std::stringstream os;
+  os << "Tensor [" << tensor.name << "]\n";
+  os << " - type: ";
+  switch (tensor.dtype) {
+    case PaddleDType::FLOAT32:
+      os << "float32";
+      break;
+    case PaddleDType::INT64:
+      os << "int64";
+      break;
+    default:
+      os << "unset";
+  }
+  os << '\n';
+
+  os << " - shape: " << to_string(tensor.shape) << '\n';
+  os << " - lod: ";
+  for (auto &l : tensor.lod) {
+    os << to_string(l) << "; ";
+  }
+  os << "\n";
+  os << " - data: ";
+
+  int dim = std::accumulate(tensor.shape.begin(), tensor.shape.end(), 1,
+                            [](int a, int b) { return a * b; });
+  for (int i = 0; i < dim; i++) {
+    os << static_cast<float *>(tensor.data.data())[i] << " ";
+  }
+  os << '\n';
+  return os.str();
+}
+
+void PrintTime(int batch_size, int repeat, int num_threads, int tid,
+               double latency) {
+  LOG(INFO) << "batch_size: " << batch_size << ", repeat: " << repeat
+            << ", threads: " << num_threads << ", thread id: " << tid
+            << ", latency: " << latency << "ms";
+}
+
 }  // namespace inference
 }  // namespace paddle