Merge pull request #12885 from luotao1/test_ditu_rnn

enhance test_analyzer to profile ditu inference demo

paddle/fluid/framework/ir/graph_pattern_detecter_tester.cc

@@ -163,8 +163,8 @@ TEST(GraphPatternDetecter, MultiSubgraph) {
   // 3. Detect op2 -> var2 -> op4
   // 4. Detect op2 -> var3 -> op5
   // But 2 and 3 and 4 overlapped, so keep 2, so the final choices are 1 and 2
-  ASSERT_GE(count, 1UL);
-  ASSERT_LE(count, 2UL);
+  ASSERT_GE(count, 1);
+  ASSERT_LE(count, 2);
 }
 
 }  // namespace ir

paddle/fluid/framework/selected_rows.cc

@@ -139,7 +139,7 @@ int64_t SelectedRows::AutoGrownIndex(int64_t key, bool auto_grown) {
     }
     auto write_iter = id_to_index_.find(key);
     if (write_iter == id_to_index_.end()) {
-      size_t row_num = rows_.size();
+      int row_num = rows_.size();
       if (row_num == value_->dims()[0]) {
         rwlock_->UNLock();
         PADDLE_THROW("selected rows is full, then length exceed %d", row_num);
@@ -182,7 +182,7 @@ void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value,
     PADDLE_ENFORCE_EQ(value_width, value->numel() / value->dims()[0],
                       "output tensor should have the same shape with table "
                       "except the dims[0].");
-    for (size_t i = 0; i < ids.numel(); ++i) {
+    for (int i = 0; i < ids.numel(); ++i) {
       int64_t index = AutoGrownIndex(ids.data<int64_t>()[i], auto_grown);
       framework::VisitDataType(
           framework::ToDataType(value_->type()),

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -23,6 +23,8 @@
 
 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.");
 
 namespace paddle {
 namespace inference {
@@ -92,7 +94,7 @@ struct DataRecord {
   size_t batch_iter{0};
   size_t batch_size{1};
   DataRecord() = default;
-  DataRecord(const std::string &path, int batch_size = 1)
+  explicit DataRecord(const std::string &path, int batch_size = 1)
       : batch_size(batch_size) {
     Load(path);
   }
@@ -165,7 +167,6 @@ struct DataRecord {
 };
 void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
                    int batch_size) {
-  // DataRecord data(FLAGS_datapath, batch_size);
   PaddleTensor lod_attention_tensor, init_zero_tensor, lod_tensor_tensor,
       week_tensor, minute_tensor;
   lod_attention_tensor.name = "data_lod_attention";
@@ -174,28 +175,33 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
   week_tensor.name = "week";
   minute_tensor.name = "minute";
   auto one_batch = data->NextBatch();
-  // clang-format off
-  std::vector<int> rnn_link_data_shape
-      ({static_cast<int>(one_batch.rnn_link_data.size()), static_cast<int>(one_batch.rnn_link_data.front().size())});
+  std::vector<int> rnn_link_data_shape(
+      {static_cast<int>(one_batch.rnn_link_data.size()),
+       static_cast<int>(one_batch.rnn_link_data.front().size())});
   lod_attention_tensor.shape.assign({1, 2});
   lod_attention_tensor.lod.assign({one_batch.lod1, one_batch.lod2});
   init_zero_tensor.shape.assign({batch_size, 15});
   init_zero_tensor.lod.assign({one_batch.lod3});
   lod_tensor_tensor.shape = rnn_link_data_shape;
   lod_tensor_tensor.lod.assign({one_batch.lod1});
-  week_tensor.shape.assign({(int) one_batch.rnn_week_datas.size(), (int) one_batch.rnn_week_datas.front().size()});
+  // clang-format off
+  week_tensor.shape.assign(
+      {static_cast<int>(one_batch.rnn_week_datas.size()),
+       static_cast<int>(one_batch.rnn_week_datas.front().size())});
   week_tensor.lod.assign({one_batch.lod3});
-  minute_tensor.shape.assign({(int) one_batch.rnn_minute_datas.size(),
-                              (int) one_batch.rnn_minute_datas.front().size()});
+  minute_tensor.shape.assign(
+      {static_cast<int>(one_batch.rnn_minute_datas.size()),
+       static_cast<int>(one_batch.rnn_minute_datas.front().size())});
   minute_tensor.lod.assign({one_batch.lod3});
+  // clang-format on
   // assign data
-  TensorAssignData(&lod_attention_tensor, std::vector<std::vector<float>>({{0, 0}}));
+  TensorAssignData(&lod_attention_tensor,
+                   std::vector<std::vector<float>>({{0, 0}}));
   std::vector<float> tmp_zeros(batch_size * 15, 0.);
   TensorAssignData(&init_zero_tensor, {tmp_zeros});
   TensorAssignData(&lod_tensor_tensor, one_batch.rnn_link_data);
   TensorAssignData(&week_tensor, one_batch.rnn_week_datas);
   TensorAssignData(&minute_tensor, one_batch.rnn_minute_datas);
-  // clang-format on
   // Set inputs.
   auto init_zero_tensor1 = init_zero_tensor;
   init_zero_tensor1.name = "hidden_init";
@@ -231,12 +237,9 @@ std::string DescribeTensor(const PaddleTensor &tensor) {
   os << "\n";
   os << " - data: ";
 
-  // clang-format off
-  int dim = std::accumulate(tensor.shape.begin(),
-                            tensor.shape.end(),
-                            1,
-                            [](int a, int b) { return a * b; });  // clang-format on
-  for (size_t i = 0; i < dim; i++) {
+  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';
@@ -300,13 +303,16 @@ void TestDituRNNPrediction(const std::string &model_path,
   for (int i = 0; i < num_times; i++) {
     predictor->Run(input_slots, &outputs);
   }
-  LOG(INFO) << "time/batch: " << timer.toc() / num_times;
+  LOG(INFO) << "===========profile result===========";
+  LOG(INFO) << "batch_size: " << batch_size << ", repeat: " << num_times
+            << ", latency: " << timer.toc() / num_times << "ms";
+  LOG(INFO) << "=====================================";
 
   for (auto &out : outputs) {
     size_t size = std::accumulate(out.shape.begin(), out.shape.end(), 1,
                                   [](int a, int b) { return a * b; });
     float *data = static_cast<float *>(out.data.data());
-    for (int i = 0;
+    for (size_t i = 0;
          i < std::min(sizeof(ditu_rnn_target_data) / sizeof(float), size);
          i++) {
       EXPECT_NEAR(data[i], ditu_rnn_target_data[i], 1e-3);
@@ -336,7 +342,7 @@ TEST(Analyzer, SupportIRPass) {
 // Directly infer with the original model.
 TEST(Analyzer, DituRNN_without_analysis) {
   TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
-                        10, false, false);
+                        FLAGS_batch_size, false, false, FLAGS_repeat);
 }
 
 // Inference with the original model with the analysis turned on, the analysis
@@ -344,14 +350,14 @@ TEST(Analyzer, DituRNN_without_analysis) {
 TEST(Analyzer, DituRNN_with_analysis) {
   LOG(INFO) << "ditu rnn with analysis";
   TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
-                        10, true, false, 1);
+                        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,
-                        10, true, true, 1);
+                        FLAGS_batch_size, true, true, FLAGS_repeat);
 }
 
 }  // namespace analysis

paddle/fluid/operators/sampling_id_op.h

@@ -54,7 +54,7 @@ class SamplingIdKernel : public framework::OpKernel<T> {
         static_cast<T>(context.Attr<float>("max")));
 
     std::vector<T> ids(batch_size);
-    for (size_t i = 0; i < batch_size; ++i) {
+    for (int i = 0; i < batch_size; ++i) {
       T r = dist(engine);
       int idx = width - 1;
       for (int j = 0; j < width; ++j) {

paddle/scripts/paddle_build.sh

@@ -116,7 +116,6 @@ function cmake_gen() {
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
         -DWITH_CONTRIB=${WITH_CONTRIB:-ON}
         -DWITH_ANAKIN=${WITH_ANAKIN:-OFF}
-        -DWITH_INFERENCE_DEMO=${WITH_INFERENCE_DEMO:-ON}
         -DPY_VERSION=${PY_VERSION:-2.7}
     ========================================
 EOF
@@ -146,7 +145,6 @@ EOF
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
         -DWITH_CONTRIB=${WITH_CONTRIB:-ON} \
         -DWITH_ANAKIN=${WITH_ANAKIN:-OFF} \
-        -DWITH_INFERENCE_DEMO=${WITH_INFERENCE_DEMO:-ON} \
         -DPY_VERSION=${PY_VERSION:-2.7}
 }