add mkl multi-thread test cases in PR-CI-INFERENCE (#34946)

* add mkl multi-thread test cases

* fix codestyle

* fix codestyle & enable ernie mkl test

paddle/fluid/inference/tests/infer_ut/run.sh

@@ -77,6 +77,12 @@ for model_name in $clas_download_list; do
     download $url_prefix $model_name
 done
 
+nlp_download_list='ernie_text_cls'
+for model_name in $nlp_download_list; do
+    url_prefix="https://paddle-qa.bj.bcebos.com/inference_model/2.1.1/nlp"
+    download $url_prefix $model_name
+done
+
 # compile and run test
 cd $current_dir
 mkdir -p build
@@ -144,6 +150,26 @@ if [ $USE_TENSORRT == ON -a $TEST_GPU_CPU == ON ]; then
     fi
 fi
 
+# ---------gpu ernie_text_cls on linux---------
+if [ $USE_TENSORRT == ON -a $TEST_GPU_CPU == ON ]; then
+    cmake .. -DPADDLE_LIB=${inference_install_dir} \
+        -DWITH_MKL=$TURN_ON_MKL \
+        -DDEMO_NAME=test_ernie_text_cls \
+        -DWITH_GPU=$TEST_GPU_CPU \
+        -DWITH_STATIC_LIB=OFF \
+        -DUSE_TENSORRT=$USE_TENSORRT \
+        -DTENSORRT_ROOT=$TENSORRT_ROOT_DIR \
+        -DWITH_GTEST=ON
+    make -j$(nproc)
+    ./test_ernie_text_cls \
+        --modeldir=$DATA_DIR/ernie_text_cls/ernie_text_cls \
+        --gtest_output=xml:test_ernie_text_cls.xml
+    if [ $? -ne 0 ]; then
+        echo "test_ernie_text_cls runs failed" >> ${current_dir}/build/test_summary.txt
+        EXIT_CODE=1
+    fi
+fi
+
 if [[ -f ${current_dir}/build/test_summary.txt ]];then
   echo "=====================test summary======================"
   cat ${current_dir}/build/test_summary.txt

paddle/fluid/inference/tests/infer_ut/test_det_mv3_db.cc

@@ -123,7 +123,7 @@ TEST(test_det_mv3_db, multi_thread2_trt_fp32_dynamic_shape_bz2) {
                   FLAGS_modeldir + "/inference.pdiparams");
   config.EnableUseGpu(100, 0);
   config.EnableTensorRtEngine(
-      1 << 20, 2, 3, paddle_infer::PrecisionType::kFloat32, false, false);
+      1 << 20, 2, 3, paddle_infer::PrecisionType::kFloat32, true, false);
   PrepareDynamicShape(&config, 4);
   // get groudtruth by disbale ir
   paddle_infer::services::PredictorPool pred_pool_no_ir(config_no_ir, 1);
@@ -149,6 +149,50 @@ TEST(test_det_mv3_db, multi_thread2_trt_fp32_dynamic_shape_bz2) {
   std::cout << "finish multi-thread test" << std::endl;
 }
 
+TEST(test_det_mv3_db, multi_thread2_mkl_fp32_bz2) {
+  int thread_num = 2;  // thread > 2 may OOM
+  // init input data
+  std::map<std::string, paddle::test::Record> my_input_data_map;
+  my_input_data_map["x"] = PrepareInput(2, 640);
+  // init output data
+  std::map<std::string, paddle::test::Record> infer_output_data,
+      truth_output_data;
+  // prepare groudtruth config
+  paddle_infer::Config config, config_no_ir;
+  config_no_ir.SetModel(FLAGS_modeldir + "/inference.pdmodel",
+                        FLAGS_modeldir + "/inference.pdiparams");
+  config_no_ir.SwitchIrOptim(false);
+  // prepare inference config
+  config.SetModel(FLAGS_modeldir + "/inference.pdmodel",
+                  FLAGS_modeldir + "/inference.pdiparams");
+  config.DisableGpu();
+  config.EnableMKLDNN();
+  config.SetMkldnnCacheCapacity(10);
+  config.SetCpuMathLibraryNumThreads(10);
+  // get groudtruth by disbale ir
+  paddle_infer::services::PredictorPool pred_pool_no_ir(config_no_ir, 1);
+  SingleThreadPrediction(pred_pool_no_ir.Retrive(0), &my_input_data_map,
+                         &truth_output_data, 1);
+
+  // get infer results from multi threads
+  std::vector<std::thread> threads;
+  services::PredictorPool pred_pool(config, thread_num);
+  for (int i = 0; i < thread_num; ++i) {
+    threads.emplace_back(paddle::test::SingleThreadPrediction,
+                         pred_pool.Retrive(i), &my_input_data_map,
+                         &infer_output_data, 2);
+  }
+
+  // thread join & check outputs
+  for (int i = 0; i < thread_num; ++i) {
+    LOG(INFO) << "join tid : " << i;
+    threads[i].join();
+    CompareRecord(&truth_output_data, &infer_output_data, 1e-4);
+  }
+
+  std::cout << "finish multi-thread test" << std::endl;
+}
+
 }  // namespace paddle_infer
 
 int main(int argc, char** argv) {

paddle/fluid/inference/tests/infer_ut/test_ernie_text_cls.cc

+// Copyright (c) 2021 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 "test_suite.h"  // NOLINT
+
+DEFINE_string(modeldir, "", "Directory of the inference model.");
+
+namespace paddle_infer {
+
+template <typename T>
+T cRandom(int min, int max) {
+  unsigned int seed = 100;
+  return (min +
+          static_cast<T>(max * rand_r(&seed) / static_cast<T>(RAND_MAX + 1)));
+}
+
+std::map<std::string, paddle::test::Record> PrepareInput(int batch_size) {
+  // init input data
+  int digit_length = 115;
+  paddle::test::Record input_ids, segment_ids;
+  int input_num = batch_size * digit_length;
+  std::vector<int64_t> input_data(input_num, 1);
+  std::vector<int64_t> segment_data(input_num, 0);
+  srand((unsigned)time(NULL));
+  for (int x = 0; x < input_data.size(); x++) {
+    input_data[x] = cRandom<int>(1, 100);
+  }
+  input_ids.data = std::vector<float>(input_data.begin(), input_data.end());
+  input_ids.shape = std::vector<int>{batch_size, digit_length};
+  input_ids.type = paddle::PaddleDType::INT64;
+
+  segment_ids.data =
+      std::vector<float>(segment_data.begin(), segment_data.end());
+  segment_ids.shape = std::vector<int>{batch_size, digit_length};
+  segment_ids.type = paddle::PaddleDType::INT64;
+
+  std::map<std::string, paddle::test::Record> my_input_data_map;
+  my_input_data_map.insert({"input_ids", input_ids});
+  my_input_data_map.insert({"token_type_ids", segment_ids});
+
+  return my_input_data_map;
+}
+
+TEST(test_ernie_text_cls, analysis_gpu_bz2_buffer) {
+  // init input data
+  auto my_input_data_map = PrepareInput(2);
+  // init output data
+  std::map<std::string, paddle::test::Record> infer_output_data,
+      truth_output_data;
+  // prepare groudtruth config
+  paddle_infer::Config config, config_no_ir;
+  config_no_ir.SetModel(FLAGS_modeldir + "/inference.pdmodel",
+                        FLAGS_modeldir + "/inference.pdiparams");
+  config_no_ir.SwitchIrOptim(false);
+
+  // prepare inference config from buffer
+  std::string prog_file = FLAGS_modeldir + "/inference.pdmodel";
+  std::string params_file = FLAGS_modeldir + "/inference.pdiparams";
+  std::string prog_str = paddle::test::read_file(prog_file);
+  std::string params_str = paddle::test::read_file(params_file);
+  config.SetModelBuffer(prog_str.c_str(), prog_str.size(), params_str.c_str(),
+                        params_str.size());
+  // get groudtruth by disbale ir
+  paddle_infer::services::PredictorPool pred_pool_no_ir(config_no_ir, 1);
+  SingleThreadPrediction(pred_pool_no_ir.Retrive(0), &my_input_data_map,
+                         &truth_output_data, 1);
+  // get infer results
+  paddle_infer::services::PredictorPool pred_pool(config, 1);
+  SingleThreadPrediction(pred_pool.Retrive(0), &my_input_data_map,
+                         &infer_output_data);
+  // check outputs
+  CompareRecord(&truth_output_data, &infer_output_data);
+  std::cout << "finish test" << std::endl;
+}
+
+TEST(test_ernie_text_cls, multi_thread4_mkl_fp32_bz2) {
+  int thread_num = 4;
+  // init input data
+  auto my_input_data_map = PrepareInput(2);
+  // init output data
+  std::map<std::string, paddle::test::Record> infer_output_data,
+      truth_output_data;
+  // prepare groudtruth config
+  paddle_infer::Config config, config_no_ir;
+  config_no_ir.SetModel(FLAGS_modeldir + "/inference.pdmodel",
+                        FLAGS_modeldir + "/inference.pdiparams");
+  config.DisableGpu();
+  config_no_ir.SwitchIrOptim(false);
+  // prepare inference config
+  config.SetModel(FLAGS_modeldir + "/inference.pdmodel",
+                  FLAGS_modeldir + "/inference.pdiparams");
+  config.DisableGpu();
+  config.EnableMKLDNN();
+  config.SetMkldnnCacheCapacity(10);
+  config.SetCpuMathLibraryNumThreads(10);
+  // get groudtruth by disbale ir
+  paddle_infer::services::PredictorPool pred_pool_no_ir(config_no_ir, 1);
+  SingleThreadPrediction(pred_pool_no_ir.Retrive(0), &my_input_data_map,
+                         &truth_output_data, 1);
+
+  // get infer results from multi threads
+  std::vector<std::thread> threads;
+  services::PredictorPool pred_pool(config, thread_num);
+  for (int i = 0; i < thread_num; ++i) {
+    threads.emplace_back(paddle::test::SingleThreadPrediction,
+                         pred_pool.Retrive(i), &my_input_data_map,
+                         &infer_output_data, 2);
+  }
+
+  // thread join & check outputs
+  for (int i = 0; i < thread_num; ++i) {
+    LOG(INFO) << "join tid : " << i;
+    threads[i].join();
+    CompareRecord(&truth_output_data, &infer_output_data);
+  }
+
+  std::cout << "finish multi-thread test" << std::endl;
+}
+
+}  // namespace paddle_infer
+
+int main(int argc, char** argv) {
+  ::testing::InitGoogleTest(&argc, argv);
+  ::google::ParseCommandLineFlags(&argc, &argv, true);
+  return RUN_ALL_TESTS();
+}

paddle/fluid/inference/tests/infer_ut/test_suite.h

@@ -37,6 +37,12 @@ class Record {
   paddle::PaddleDType type;
 };
 
+std::string read_file(std::string filename) {
+  std::ifstream file(filename);
+  return std::string((std::istreambuf_iterator<char>(file)),
+                     std::istreambuf_iterator<char>());
+}
+
 void SingleThreadPrediction(paddle_infer::Predictor *predictor,
                             std::map<std::string, Record> *input_data_map,
                             std::map<std::string, Record> *output_data_map,
@@ -44,14 +50,37 @@ void SingleThreadPrediction(paddle_infer::Predictor *predictor,
   // prepare input tensor
   auto input_names = predictor->GetInputNames();
   for (const auto & [ key, value ] : *input_data_map) {
-    auto input_tensor = predictor->GetInputHandle(key);
-    input_tensor->Reshape(value.shape);
-    input_tensor->CopyFromCpu(value.data.data());
+    switch (value.type) {
+      case paddle::PaddleDType::INT64: {
+        std::vector<int64_t> input_value =
+            std::vector<int64_t>(value.data.begin(), value.data.end());
+        auto input_tensor = predictor->GetInputHandle(key);
+        input_tensor->Reshape(value.shape);
+        input_tensor->CopyFromCpu(input_value.data());
+        break;
+      }
+      case paddle::PaddleDType::INT32: {
+        std::vector<int32_t> input_value =
+            std::vector<int32_t>(value.data.begin(), value.data.end());
+        auto input_tensor = predictor->GetInputHandle(key);
+        input_tensor->Reshape(value.shape);
+        input_tensor->CopyFromCpu(input_value.data());
+        break;
+      }
+      case paddle::PaddleDType::FLOAT32: {
+        std::vector<float> input_value =
+            std::vector<float>(value.data.begin(), value.data.end());
+        auto input_tensor = predictor->GetInputHandle(key);
+        input_tensor->Reshape(value.shape);
+        input_tensor->CopyFromCpu(input_value.data());
+        break;
+      }
+    }
   }
 
   // inference
   for (size_t i = 0; i < repeat_times; ++i) {
-    CHECK(predictor->Run());
+    ASSERT_TRUE(predictor->Run());
   }
 
   // get output data to Record
@@ -112,8 +141,8 @@ void CompareRecord(std::map<std::string, Record> *truth_output_data,
     size_t numel = value.data.size() / sizeof(float);
     EXPECT_EQ(value.data.size(), truth_record.data.size());
     for (size_t i = 0; i < numel; ++i) {
-      CHECK_LT(fabs(value.data.data()[i] - truth_record.data.data()[i]),
-               epislon);
+      ASSERT_LT(fabs(value.data.data()[i] - truth_record.data.data()[i]),
+                epislon);
     }
   }
 }