refine paddle inference api (#26774)

* refine paddle inference api
Co-authored-by: Nnhzlx <nhzlx.dragon@gmail.com>

paddle/fluid/inference/api/analysis_config.cc

@@ -15,7 +15,6 @@
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/api/paddle_analysis_config.h"
-#include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/api/paddle_pass_builder.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/gpu_info.h"
@@ -103,8 +102,8 @@ AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
                                   // params_file_ fields.
 
   CP_MEMBER(opt_cache_dir_);
-  prog_file_ = std::move(other.prog_file_);
-  params_file_ = std::move(other.params_file_);
+  CP_MEMBER(prog_file_);
+  CP_MEMBER(params_file_);
 
   CP_MEMBER(use_fc_padding_);
   // GPU related.

paddle/fluid/inference/api/analysis_predictor.cc

@@ -32,7 +32,6 @@
 #include "paddle/fluid/inference/analysis/helper.h"
 #include "paddle/fluid/inference/analysis/passes/memory_optimize_pass.h"
 #include "paddle/fluid/inference/api/helper.h"
-#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/memory/memcpy.h"
@@ -517,6 +516,8 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
 template <>
 std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
     AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig &config) {
+  // TODO(NHZlX): Should add the link to the doc of
+  // paddle_infer::CreatePredictor<paddle_infer::Config>
   if (config.glog_info_disabled()) {
     FLAGS_logtostderr = 1;
     FLAGS_minloglevel = 2;  // GLOG_ERROR
@@ -1058,3 +1059,122 @@ USE_TRT_CONVERTER(skip_layernorm);
 USE_TRT_CONVERTER(slice);
 USE_TRT_CONVERTER(scale);
 #endif
+
+namespace paddle_infer {
+
+void Tensor::Reshape(const std::vector<int> &shape) { tensor_->Reshape(shape); }
+
+std::vector<int> Tensor::shape() const { return tensor_->shape(); }
+
+void Tensor::SetLoD(const std::vector<std::vector<size_t>> &x) {
+  return tensor_->SetLoD(x);
+}
+
+std::vector<std::vector<size_t>> Tensor::lod() const { return tensor_->lod(); }
+
+const std::string &Tensor::name() const { return tensor_->name(); }
+
+DataType Tensor::type() const { return tensor_->type(); }
+
+Predictor::Predictor(const Config &config) {
+  const_cast<Config *>(&config)->SwitchUseFeedFetchOps(false);
+  // The second parameter indicates that the discard log is not printed
+  predictor_ = paddle::CreatePaddlePredictor<
+      Config, paddle::PaddleEngineKind::kAnalysis>(config);
+}
+
+std::vector<std::string> Predictor::GetInputNames() {
+  return predictor_->GetInputNames();
+}
+
+std::unique_ptr<Tensor> Predictor::GetInputHandle(const std::string &name) {
+  auto zero_copy_tensor = predictor_->GetInputTensor(name);
+  std::unique_ptr<Tensor> tensor(new Tensor(std::move(zero_copy_tensor)));
+  return tensor;
+}
+
+std::vector<std::string> Predictor::GetOutputNames() {
+  return predictor_->GetOutputNames();
+}
+
+std::unique_ptr<Tensor> Predictor::GetOutputHandle(const std::string &name) {
+  auto zero_copy_tensor = predictor_->GetOutputTensor(name);
+  std::unique_ptr<Tensor> tensor(new Tensor(std::move(zero_copy_tensor)));
+  return tensor;
+}
+
+bool Predictor::Run() { return predictor_->ZeroCopyRun(); }
+
+std::unique_ptr<Predictor> Predictor::Clone() {
+  auto analysis_pred = predictor_->Clone();
+  std::unique_ptr<Predictor> pred(new Predictor(std::move(analysis_pred)));
+  return pred;
+}
+
+void Predictor::ClearIntermediateTensor() {
+  predictor_->ClearIntermediateTensor();
+}
+
+int GetNumBytesOfDataType(DataType dtype) {
+  switch (dtype) {
+    case DataType::FLOAT32:
+      return sizeof(float);
+    case DataType::INT64:
+      return sizeof(int64_t);
+    case DataType::INT32:
+      return sizeof(int32_t);
+    case DataType::UINT8:
+      return sizeof(uint8_t);
+    default:
+      assert(false);
+      return -1;
+  }
+}
+
+std::string GetVersion() { return paddle::get_version(); }
+
+std::string UpdateDllFlag(const char *name, const char *value) {
+  return paddle::UpdateDllFlag(name, value);
+}
+
+}  // namespace paddle_infer
+
+namespace paddle_infer {
+std::shared_ptr<Predictor> CreatePredictor(const Config &config) {  // NOLINT
+  std::shared_ptr<Predictor> predictor(new Predictor(config));
+  return predictor;
+}
+
+namespace services {
+PredictorPool::PredictorPool(const Config &config, size_t size) {
+  PADDLE_ENFORCE_GE(
+      size, 1UL,
+      paddle::platform::errors::InvalidArgument(
+          "The predictor pool size should be greater than 1, but it's (%d)",
+          size));
+  Config copy_config(config);
+  main_pred_.reset(new Predictor(config));
+  for (size_t i = 0; i < size - 1; i++) {
+    if (config.tensorrt_engine_enabled()) {
+      Config config_tmp(copy_config);
+      preds_.push_back(
+          std::move(std::unique_ptr<Predictor>(new Predictor(config_tmp))));
+    } else {
+      preds_.push_back(std::move(main_pred_->Clone()));
+    }
+  }
+}
+
+Predictor *PredictorPool::Retrive(size_t idx) {
+  PADDLE_ENFORCE_LT(
+      idx, preds_.size() + 1,
+      paddle::platform::errors::InvalidArgument(
+          "There are (%d) predictors in the pool, but the idx is (%d)", idx,
+          preds_.size() + 1));
+  if (idx == 0) {
+    return main_pred_.get();
+  }
+  return preds_[idx - 1].get();
+}
+}  // namespace services
+}  // namespace paddle_infer

paddle/fluid/inference/api/api.cc

@@ -112,6 +112,12 @@ void PaddleBuf::Free() {
   }
 }
 
+NativeConfig::NativeConfig() {
+  LOG(WARNING) << "The paddle::NativeConfig interface is going to be "
+                  "deprecated in the next release, plase use the latest "
+                  "paddle_infer::Config instead.";
+}
+
 std::string get_version() {
   std::stringstream ss;
   ss << "version: " << framework::paddle_version() << "\n";

paddle/fluid/inference/api/api_impl.cc

@@ -15,6 +15,7 @@ limitations under the License. */
 #include <glog/logging.h>
 #include <algorithm>
 #include <map>
+#include <memory>
 #include <set>
 #include <sstream>
 #include <string>
@@ -25,6 +26,7 @@ limitations under the License. */
 #include "paddle/fluid/inference/api/api_impl.h"
 #include "paddle/fluid/inference/api/details/reset_tensor_array.h"
 #include "paddle/fluid/inference/api/helper.h"
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -311,6 +313,8 @@ bool NativePaddlePredictor::GetFetch(std::vector<PaddleTensor> *outputs,
 template <>
 std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
     NativeConfig, PaddleEngineKind::kNative>(const NativeConfig &config) {
+  // TODO(NHZlX): Should add the link to the doc of
+  // paddle_infer::CreatePredictor<paddle_infer::Config>
   VLOG(3) << "create NativePaddlePredictor";
   if (config.use_gpu) {
     // 1. GPU memory

paddle/fluid/inference/api/paddle_api.h

@@ -347,6 +347,7 @@ class PD_INFER_DECL PaddlePredictor {
 /// place of inference, etc.)
 ///
 struct PD_INFER_DECL NativeConfig : public PaddlePredictor::Config {
+  NativeConfig();
   /// GPU related fields.
   bool use_gpu{false};
   int device{0};
@@ -421,7 +422,8 @@ enum class PaddleEngineKind {
 };
 
 template <typename ConfigT, PaddleEngineKind engine>
-std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT& config);
+PD_INFER_DECL std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(
+    const ConfigT& config);
 
 template <>
 PD_INFER_DECL std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
@@ -437,6 +439,4 @@ PD_INFER_DECL std::string get_version();
 
 PD_INFER_DECL std::string UpdateDllFlag(const char* name, const char* value);
 
-PD_INFER_DECL std::shared_ptr<framework::Cipher> MakeCipher(
-    const std::string& config_file);
 }  // namespace paddle

paddle/fluid/inference/api/paddle_inference_api.h

@@ -22,9 +22,124 @@ limitations under the License. */
 #pragma once
 
 #include <cassert>
+#include <map>
 #include <memory>
 #include <string>
+#include <utility>
 #include <vector>
 
 #include "paddle_analysis_config.h"  // NOLINT
 #include "paddle_api.h"              // NOLINT
+
+namespace paddle_infer {
+using DataType = paddle::PaddleDType;
+using PlaceType = paddle::PaddlePlace;
+using PrecisionType = paddle::AnalysisConfig::Precision;
+using Config = paddle::AnalysisConfig;
+
+class PD_INFER_DECL Tensor {
+ public:
+  // Can only be created by predictor->GetInputHandle(cosnt std::string& name)
+  // or predictor->GetOutputHandle(cosnt std::string& name)
+  Tensor() = delete;
+  explicit Tensor(std::unique_ptr<paddle::ZeroCopyTensor>&& tensor)
+      : tensor_(std::move(tensor)) {}
+  void Reshape(const std::vector<int>& shape);
+
+  template <typename T>
+  void CopyFromCpu(const T* data);
+
+  // should add the place
+  template <typename T>
+  T* mutable_data(PlaceType place);
+
+  template <typename T>
+  void CopyToCpu(T* data);
+
+  template <typename T>
+  T* data(PlaceType* place, int* size) const;
+
+  void SetLoD(const std::vector<std::vector<size_t>>& x);
+  std::vector<std::vector<size_t>> lod() const;
+
+  DataType type() const;
+
+  std::vector<int> shape() const;
+  const std::string& name() const;
+
+ private:
+  std::unique_ptr<paddle::ZeroCopyTensor> tensor_;
+};
+
+class PD_INFER_DECL Predictor {
+ public:
+  Predictor() = default;
+  ~Predictor() {}
+  // Use for clone
+  explicit Predictor(std::unique_ptr<paddle::PaddlePredictor>&& pred)
+      : predictor_(std::move(pred)) {}
+
+  explicit Predictor(const Config& config);
+
+  std::vector<std::string> GetInputNames();
+  std::unique_ptr<Tensor> GetInputHandle(const std::string& name);
+
+  bool Run();
+
+  std::vector<std::string> GetOutputNames();
+  std::unique_ptr<Tensor> GetOutputHandle(const std::string& name);
+
+  std::unique_ptr<Predictor> Clone();
+  void ClearIntermediateTensor();
+
+ private:
+  std::unique_ptr<paddle::PaddlePredictor> predictor_;
+};
+
+PD_INFER_DECL std::shared_ptr<Predictor> CreatePredictor(
+    const Config& config);  // NOLINT
+PD_INFER_DECL int GetNumBytesOfDataType(DataType dtype);
+
+PD_INFER_DECL std::string GetVersion();
+PD_INFER_DECL std::string UpdateDllFlag(const char* name, const char* value);
+
+template <typename T>
+void Tensor::CopyFromCpu(const T* data) {
+  tensor_->copy_from_cpu<T>(data);
+}
+
+template <typename T>
+void Tensor::CopyToCpu(T* data) {
+  return tensor_->copy_to_cpu<T>(data);
+}
+
+template <typename T>
+T* Tensor::mutable_data(PlaceType place) {
+  return tensor_->mutable_data<T>(place);
+}
+
+template <typename T>
+T* Tensor::data(PlaceType* place, int* size) const {
+  return tensor_->data<T>(place, size);
+}
+
+}  // namespace paddle_infer
+
+namespace paddle_infer {
+namespace services {
+
+class PD_INFER_DECL PredictorPool {
+ public:
+  PredictorPool() = delete;
+  PredictorPool(const PredictorPool&) = delete;
+  PredictorPool& operator=(const PredictorPool&) = delete;
+
+  explicit PredictorPool(const Config& config, size_t size = 1);
+  Predictor* Retrive(size_t idx);
+
+ private:
+  std::shared_ptr<Predictor> main_pred_;
+  std::vector<std::unique_ptr<Predictor>> preds_;
+};
+}  // namespace services
+}  // namespace paddle_infer

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

@@ -515,3 +515,9 @@ if(WITH_MKLDNN)
 inference_analysis_test(test_analyzer_capi_ner SRCS analyzer_capi_ner_tester.cc 
         EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} paddle_fluid_c
         ARGS --infer_model=${CHINESE_NER_INSTALL_DIR}/model)
+
+if(WITH_GPU)
+  inference_analysis_test(paddle_infer_api_test SRCS paddle_infer_api_test.cc
+        EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
+        ARGS --infer_model=${RESNET50_MODEL_DIR})
+endif()

paddle/fluid/inference/tests/api/lite_resnet50_test.cc

@@ -72,3 +72,59 @@ TEST(AnalysisPredictor, use_gpu) {
 
 }  // namespace inference
 }  // namespace paddle
+
+namespace paddle_infer {
+
+TEST(Predictor, use_gpu) {
+  std::string model_dir = FLAGS_infer_model + "/" + "model";
+  Config config;
+  config.EnableUseGpu(100, 0);
+  config.SetModel(model_dir + "/model", model_dir + "/params");
+  config.EnableLiteEngine(PrecisionType::kFloat32);
+
+  auto predictor = CreatePredictor(config);
+  const int batch = 1;
+  const int channel = 3;
+  const int height = 318;
+  const int width = 318;
+  const int input_num = batch * channel * height * width;
+  std::vector<float> input(input_num, 1);
+
+  auto input_names = predictor->GetInputNames();
+  auto input_t = predictor->GetInputHandle(input_names[0]);
+
+  input_t->Reshape({1, 3, 318, 318});
+  input_t->CopyFromCpu(input.data());
+  predictor->Run();
+
+  auto output_names = predictor->GetOutputNames();
+  auto output_t = predictor->GetOutputHandle(output_names[0]);
+  std::vector<int> output_shape = output_t->shape();
+  size_t out_num = std::accumulate(output_shape.begin(), output_shape.end(), 1,
+                                   std::multiplies<int>());
+
+  std::vector<float> out_data;
+  out_data.resize(out_num);
+  output_t->CopyToCpu(out_data.data());
+
+  const std::vector<float> truth_values = {
+      127.780396f, 738.16656f,  1013.2264f,  -438.17206f, 366.4022f,
+      927.66187f,  736.2241f,   -633.68567f, -329.92737f, -430.15637f,
+      -633.0639f,  -146.54858f, -1324.2804f, -1349.3661f, -242.67671f,
+      117.44864f,  -801.7251f,  -391.51495f, -404.8202f,  454.16132f,
+      515.48206f,  -133.03114f, 69.293076f,  590.09753f,  -1434.6917f,
+      -1070.8903f, 307.0744f,   400.52573f,  -316.12177f, -587.1265f,
+      -161.05742f, 800.3663f,   -96.47157f,  748.708f,    868.17645f,
+      -447.9403f,  112.73656f,  1127.1992f,  47.43518f,   677.7219f,
+      593.1881f,   -336.4011f,  551.3634f,   397.82474f,  78.39835f,
+      -715.4006f,  405.96988f,  404.25684f,  246.01978f,  -8.430191f,
+      131.36617f,  -648.0528f};
+
+  float* data_o = out_data.data();
+  for (size_t j = 0; j < out_num; j += 10) {
+    EXPECT_NEAR((data_o[j] - truth_values[j / 10]) / truth_values[j / 10], 0.,
+                10e-5);
+  }
+}
+
+}  // namespace paddle_infer

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

+/* Copyright (c) 2018 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 <cuda_runtime.h>
+#include <gflags/gflags.h>
+#include <glog/logging.h>
+#include <gtest/gtest.h>
+#include <cstring>
+#include <numeric>
+
+#include "paddle/fluid/inference/tests/api/trt_test_helper.h"
+
+namespace paddle_infer {
+
+TEST(Predictor, use_gpu) {
+  LOG(INFO) << GetVersion();
+  UpdateDllFlag("conv_workspace_size_limit", "4000");
+  std::string model_dir = FLAGS_infer_model + "/model";
+  Config config;
+  config.SetModel(model_dir + "/model", model_dir + "/params");
+  config.EnableUseGpu(100, 0);
+
+  auto predictor = CreatePredictor(config);
+  auto pred_clone = predictor->Clone();
+
+  std::vector<int> in_shape = {1, 3, 318, 318};
+  int in_num = std::accumulate(in_shape.begin(), in_shape.end(), 1,
+                               [](int &a, int &b) { return a * b; });
+
+  std::vector<float> input(in_num, 0);
+
+  auto input_names = predictor->GetInputNames();
+  auto input_t = predictor->GetInputHandle(input_names[0]);
+
+  input_t->Reshape(in_shape);
+  input_t->CopyFromCpu(input.data());
+  predictor->Run();
+
+  auto output_names = predictor->GetOutputNames();
+  auto output_t = predictor->GetOutputHandle(output_names[0]);
+  std::vector<int> output_shape = output_t->shape();
+  int out_num = std::accumulate(output_shape.begin(), output_shape.end(), 1,
+                                std::multiplies<int>());
+
+  std::vector<float> out_data;
+  out_data.resize(out_num);
+  output_t->CopyToCpu(out_data.data());
+  predictor->ClearIntermediateTensor();
+}
+
+TEST(PredictorPool, basic) {
+  LOG(INFO) << GetVersion();
+  UpdateDllFlag("conv_workspace_size_limit", "4000");
+  std::string model_dir = FLAGS_infer_model + "/model";
+  Config config;
+  config.SetModel(model_dir + "/model", model_dir + "/params");
+  config.EnableUseGpu(100, 0);
+
+  services::PredictorPool pred_pool(config, 4);
+  auto pred = pred_pool.Retrive(2);
+
+  std::vector<int> in_shape = {1, 3, 318, 318};
+  int in_num = std::accumulate(in_shape.begin(), in_shape.end(), 1,
+                               [](int &a, int &b) { return a * b; });
+  std::vector<float> input(in_num, 0);
+
+  auto in_names = pred->GetInputNames();
+  auto input_t = pred->GetInputHandle(in_names[0]);
+  input_t->name();
+  input_t->Reshape(in_shape);
+  input_t->CopyFromCpu(input.data());
+  pred->Run();
+  auto out_names = pred->GetOutputNames();
+  auto output_t = pred->GetOutputHandle(out_names[0]);
+  auto out_type = output_t->type();
+  LOG(INFO) << GetNumBytesOfDataType(out_type);
+  if (out_type == DataType::FLOAT32) {
+    PlaceType place;
+    int size;
+    output_t->data<float>(&place, &size);
+  }
+}
+
+}  // namespace paddle_infer

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

@@ -41,7 +41,7 @@ TEST(AnalysisPredictor, use_gpu) {
   SetFakeImageInput(&inputs_all, model_dir, false, "__model__", "");
 
   std::vector<PaddleTensor> outputs;
-  for (auto& input : inputs_all) {
+  for (auto &input : inputs_all) {
     ASSERT_TRUE(predictor->Run(input, &outputs));
     predictor->ClearIntermediateTensor();
   }
@@ -49,3 +49,27 @@ TEST(AnalysisPredictor, use_gpu) {
 
 }  // namespace inference
 }  // namespace paddle
+
+namespace paddle_infer {
+TEST(PredictorPool, use_gpu) {
+  std::string model_dir = FLAGS_infer_model + "/" + "mobilenet";
+  Config config;
+  config.EnableUseGpu(100, 0);
+  config.SetModel(model_dir);
+  config.EnableTensorRtEngine();
+  services::PredictorPool pred_pool(config, 1);
+
+  auto predictor = pred_pool.Retrive(0);
+  auto input_names = predictor->GetInputNames();
+  auto input_t = predictor->GetInputHandle(input_names[0]);
+  std::vector<int> in_shape = {1, 3, 224, 224};
+  int in_num = std::accumulate(in_shape.begin(), in_shape.end(), 1,
+                               [](int &a, int &b) { return a * b; });
+
+  std::vector<float> input(in_num, 0);
+  input_t->Reshape(in_shape);
+  input_t->CopyFromCpu(input.data());
+  predictor->Run();
+}
+
+}  // namespace paddle_infer

paddle/fluid/pybind/inference_api.cc

@@ -206,9 +206,9 @@ void BindInferenceApi(py::module *m) {
   BindMkldnnQuantizerConfig(m);
 #endif
   m->def("create_paddle_predictor",
-         &paddle::CreatePaddlePredictor<AnalysisConfig>);
+         &paddle::CreatePaddlePredictor<AnalysisConfig>, py::arg("config"));
   m->def("create_paddle_predictor",
-         &paddle::CreatePaddlePredictor<NativeConfig>);
+         &paddle::CreatePaddlePredictor<NativeConfig>, py::arg("config"));
   m->def("paddle_dtype_size", &paddle::PaddleDtypeSize);
   m->def("paddle_tensor_to_bytes", &SerializePDTensorToBytes);
 }