[Cherry-pick] Add C-API for fluid inference api (#20259)

* Add capi for fluid inference api (#20092)

* add capi for fluid inference api, including AnalysisConfig, AnalysisPredictor, PaddleBuf, PaddleTensor, ZeroCopyTensor

* add dll to inference capi (#20180)

* add dll to inference capi, test=develop

* add if win32 in cmakelists, test=develop

paddle/fluid/inference/CMakeLists.txt

@@ -36,6 +36,7 @@ else(WIN32)
 endif(WIN32)
 
 add_subdirectory(api)
+add_subdirectory(capi)
 
 if(WITH_MKLDNN)
 	set(mkldnn_quantizer_src ${CMAKE_CURRENT_SOURCE_DIR}/api/mkldnn_quantizer.cc)

paddle/fluid/inference/capi/CMakeLists.txt

+
+cc_library(pd_config SRCS pd_config.cc)
+cc_library(pd_predictor SRCS pd_predictor.cc)
+cc_library(pd_tensor SRCS pd_tensor.cc)
+cc_library(pd_c_api SRCS c_api.cc)
+cc_library(paddle_fluid_c SRCS c_api.cc DEPS paddle_fluid pd_config pd_predictor pd_tensor pd_c_api)
+cc_library(paddle_fluid_c_shared SHARED SRCS c_api.cc DEPS paddle_fluid pd_config pd_predictor pd_tensor pd_c_api)
+set_target_properties(paddle_fluid_c_shared PROPERTIES OUTPUT_NAME paddle_fluid_c)
+if(WIN32)
+    target_link_libraries(paddle_fluid_c_shared shlwapi.lib)
+endif()

paddle/fluid/inference/capi/c_api.cc

+// Copyright (c) 2019 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 "paddle/fluid/inference/capi/c_api.h"
+#include <algorithm>
+#include <vector>
+#include "paddle/fluid/inference/capi/c_api_internal.h"
+
+using paddle::ConvertToPaddleDType;
+using paddle::ConvertToPDDataType;
+using paddle::ConvertToACPrecision;
+
+extern "C" {
+
+PD_PaddleBuf* PD_NewPaddleBuf() { return new PD_PaddleBuf; }
+
+void PD_DeletePaddleBuf(PD_PaddleBuf* buf) {
+  if (buf) {
+    delete buf;
+    buf = nullptr;
+  }
+}
+
+void PD_PaddleBufResize(PD_PaddleBuf* buf, size_t length) {
+  buf->buf.Resize(length);
+}
+
+void PD_PaddleBufReset(PD_PaddleBuf* buf, void* data, size_t length) {
+  buf->buf.Reset(data, length);
+}
+
+bool PD_PaddleBufEmpty(PD_PaddleBuf* buf) { return buf->buf.empty(); }
+
+void* PD_PaddleBufData(PD_PaddleBuf* buf) { return buf->buf.data(); }
+
+size_t PD_PaddleBufLength(PD_PaddleBuf* buf) { return buf->buf.length(); }
+
+}  // extern "C"
+
+namespace paddle {
+paddle::PaddleDType ConvertToPaddleDType(PD_DataType dtype) {
+  switch (dtype) {
+    case PD_FLOAT32:
+      return PD_PaddleDType::FLOAT32;
+    case PD_INT32:
+      return PD_PaddleDType::INT32;
+    case PD_INT64:
+      return PD_PaddleDType::INT64;
+    case PD_UINT8:
+      return PD_PaddleDType::UINT8;
+    default:
+      CHECK(false) << "Unsupport dtype.";
+      return PD_PaddleDType::FLOAT32;
+  }
+}
+
+PD_DataType ConvertToPDDataType(PD_PaddleDType dtype) {
+  switch (dtype) {
+    case PD_PaddleDType::FLOAT32:
+      return PD_DataType::PD_FLOAT32;
+    case PD_PaddleDType::INT32:
+      return PD_DataType::PD_INT32;
+    case PD_PaddleDType::INT64:
+      return PD_DataType::PD_INT64;
+    case PD_PaddleDType::UINT8:
+      return PD_DataType::PD_UINT8;
+    default:
+      CHECK(false) << "Unsupport dtype.";
+      return PD_DataType::PD_UNKDTYPE;
+  }
+}
+
+PD_ACPrecision ConvertToACPrecision(Precision dtype) {
+  switch (dtype) {
+    case Precision::kFloat32:
+      return PD_ACPrecision::kFloat32;
+    case Precision::kInt8:
+      return PD_ACPrecision::kInt8;
+    case Precision::kHalf:
+      return PD_ACPrecision::kHalf;
+    default:
+      CHECK(false) << "Unsupport precision.";
+      return PD_ACPrecision::kFloat32;
+  }
+}
+}  // namespace paddle

paddle/fluid/inference/capi/c_api.h

+// Copyright (c) 2019 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.
+
+#pragma once
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+
+#if defined(_WIN32)
+#ifdef PADDLE_ON_INFERENCE
+#define PADDLE_CAPI_EXPORT __declspec(dllexport)
+#else
+#define PADDLE_CAPI_EXPORT __declspec(dllimport)
+#endif  // PADDLE_ON_INFERENCE
+#else
+#define PADDLE_CAPI_EXPORT __attribute__((visibility("default")))
+#endif  // _WIN32
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum PD_DataType { PD_FLOAT32, PD_INT32, PD_INT64, PD_UINT8, PD_UNKDTYPE };
+typedef struct PD_PaddleBuf PD_PaddleBuf;
+typedef struct PD_AnalysisConfig PD_AnalysisConfig;
+
+typedef struct PD_ZeroCopyData {
+  char* name = new char[50];
+  void* data;
+  PD_DataType dtype;
+  int* shape;
+  int shape_size;
+} PD_ZeroCopyData;
+typedef struct InTensorShape {
+  char* name;
+  int* tensor_shape;
+  int shape_size;
+} InTensorShape;
+
+PADDLE_CAPI_EXPORT extern PD_PaddleBuf* PD_NewPaddleBuf();
+
+PADDLE_CAPI_EXPORT extern void PD_DeletePaddleBuf(PD_PaddleBuf* buf);
+
+PADDLE_CAPI_EXPORT extern void PD_PaddleBufResize(PD_PaddleBuf* buf,
+                                                  size_t length);
+
+PADDLE_CAPI_EXPORT extern void PD_PaddleBufReset(PD_PaddleBuf* buf, void* data,
+                                                 size_t length);
+
+PADDLE_CAPI_EXPORT extern bool PD_PaddleBufEmpty(PD_PaddleBuf* buf);
+
+PADDLE_CAPI_EXPORT extern void* PD_PaddleBufData(PD_PaddleBuf* buf);
+
+PADDLE_CAPI_EXPORT extern size_t PD_PaddleBufLength(PD_PaddleBuf* buf);
+
+// PaddleTensor
+typedef struct PD_Tensor PD_Tensor;
+
+PADDLE_CAPI_EXPORT extern PD_Tensor* PD_NewPaddleTensor();
+
+PADDLE_CAPI_EXPORT extern void PD_DeletePaddleTensor(PD_Tensor* tensor);
+
+PADDLE_CAPI_EXPORT extern void PD_SetPaddleTensorName(PD_Tensor* tensor,
+                                                      char* name);
+
+PADDLE_CAPI_EXPORT extern void PD_SetPaddleTensorDType(PD_Tensor* tensor,
+                                                       PD_DataType dtype);
+
+PADDLE_CAPI_EXPORT extern void PD_SetPaddleTensorData(PD_Tensor* tensor,
+                                                      PD_PaddleBuf* buf);
+
+PADDLE_CAPI_EXPORT extern void PD_SetPaddleTensorShape(PD_Tensor* tensor,
+                                                       int* shape, int size);
+
+PADDLE_CAPI_EXPORT extern const char* PD_GetPaddleTensorName(
+    const PD_Tensor* tensor);
+
+PADDLE_CAPI_EXPORT extern PD_DataType PD_GetPaddleTensorDType(
+    const PD_Tensor* tensor);
+
+PADDLE_CAPI_EXPORT extern PD_PaddleBuf* PD_GetPaddleTensorData(
+    const PD_Tensor* tensor);
+
+PADDLE_CAPI_EXPORT extern int* PD_GetPaddleTensorShape(const PD_Tensor* tensor,
+                                                       int** size);
+
+// AnalysisPredictor
+PADDLE_CAPI_EXPORT extern bool PD_PredictorRun(const PD_AnalysisConfig* config,
+                                               PD_Tensor* inputs, int in_size,
+                                               PD_Tensor* output_data,
+                                               int** out_size, int batch_size);
+
+PADDLE_CAPI_EXPORT extern bool PD_PredictorZeroCopyRun(
+    const PD_AnalysisConfig* config, PD_ZeroCopyData* inputs, int in_size,
+    PD_ZeroCopyData* output, int** out_size);
+
+// AnalysisConfig
+enum Precision { kFloat32 = 0, kInt8, kHalf };
+
+PADDLE_CAPI_EXPORT extern PD_AnalysisConfig* PD_NewAnalysisConfig();
+
+PADDLE_CAPI_EXPORT extern void PD_DeleteAnalysisConfig(
+    PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_SetModel(PD_AnalysisConfig* config,
+                                           const char* model_dir,
+                                           const char* params_path = NULL);
+
+PADDLE_CAPI_EXPORT
+extern void PD_SetProgFile(PD_AnalysisConfig* config, const char* x);
+
+PADDLE_CAPI_EXPORT extern void PD_SetParamsFile(PD_AnalysisConfig* config,
+                                                const char* x);
+
+PADDLE_CAPI_EXPORT extern void PD_SetOptimCacheDir(PD_AnalysisConfig* config,
+                                                   const char* opt_cache_dir);
+
+PADDLE_CAPI_EXPORT extern const char* PD_ModelDir(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern const char* PD_ProgFile(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern const char* PD_ParamsFile(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_EnableUseGpu(
+    PD_AnalysisConfig* config, uint64_t memory_pool_init_size_mb,
+    int device_id = 0);
+
+PADDLE_CAPI_EXPORT extern void PD_DisableGpu(PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern bool PD_UseGpu(const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern int PD_GpuDeviceId(const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern int PD_MemoryPoolInitSizeMb(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern float PD_FractionOfGpuMemoryForPool(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_EnableCUDNN(PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern bool PD_CudnnEnabled(const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_SwitchIrOptim(PD_AnalysisConfig* config,
+                                                bool x = true);
+
+PADDLE_CAPI_EXPORT extern bool PD_IrOptim(const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_SwitchUseFeedFetchOps(
+    PD_AnalysisConfig* config, bool x = true);
+
+PADDLE_CAPI_EXPORT extern bool PD_UseFeedFetchOpsEnabled(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_SwitchSpecifyInputNames(
+    PD_AnalysisConfig* config, bool x = true);
+
+PADDLE_CAPI_EXPORT extern bool PD_SpecifyInputName(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_EnableTensorRtEngine(
+    PD_AnalysisConfig* config, int workspace_size = 1 << 20,
+    int max_batch_size = 1, int min_subgraph_size = 3,
+    Precision precision = Precision::kFloat32, bool use_static = false,
+    bool use_calib_mode = false);
+
+PADDLE_CAPI_EXPORT extern bool PD_TensorrtEngineEnabled(
+    const PD_AnalysisConfig* config);
+
+typedef struct PD_MaxInputShape {
+  char* name;
+  int* shape;
+  int shape_size;
+} PD_MaxInputShape;
+
+PADDLE_CAPI_EXPORT extern void PD_EnableAnakinEngine(
+    PD_AnalysisConfig* config, int max_batch_size = 1,
+    PD_MaxInputShape* max_input_shape = NULL, int max_input_shape_size = 0,
+    int min_subgraph_size = 6, Precision precision = Precision::kFloat32,
+    bool auto_config_layout = false, char** passes_filter = NULL,
+    int passes_filter_size = 0, char** ops_filter = NULL,
+    int ops_filter_size = 0);
+
+PADDLE_CAPI_EXPORT extern bool PD_AnakinEngineEnabled(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_SwitchIrDebug(PD_AnalysisConfig* config,
+                                                bool x = true);
+
+PADDLE_CAPI_EXPORT extern void PD_EnableNgraph(PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern bool PD_NgraphEnabled(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_EnableMKLDNN(PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_SetMkldnnCacheCapacity(
+    PD_AnalysisConfig* config, int capacity);
+
+PADDLE_CAPI_EXPORT extern bool PD_MkldnnEnabled(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_SetCpuMathLibraryNumThreads(
+    PD_AnalysisConfig* config, int cpu_math_library_num_threads);
+
+PADDLE_CAPI_EXPORT extern int PD_CpuMathLibraryNumThreads(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_EnableMkldnnQuantizer(
+    PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern bool PD_MkldnnQuantizerEnabled(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_SetModelBuffer(PD_AnalysisConfig* config,
+                                                 const char* prog_buffer,
+                                                 size_t prog_buffer_size,
+                                                 const char* params_buffer,
+                                                 size_t params_buffer_size);
+
+PADDLE_CAPI_EXPORT extern bool PD_ModelFromMemory(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_EnableMemoryOptim(PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern bool PD_MemoryOptimEnabled(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_EnableProfile(PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern bool PD_ProfileEnabled(
+    const PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern void PD_SetInValid(PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern bool PD_IsValid(const PD_AnalysisConfig* config);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif

paddle/fluid/inference/capi/c_api_internal.h

+// Copyright (c) 2019 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.
+
+#pragma once
+
+#include <memory>
+#include "paddle/fluid/inference/api/paddle_analysis_config.h"
+#include "paddle/fluid/inference/api/paddle_api.h"
+#include "paddle/fluid/platform/enforce.h"
+
+using PD_PaddleDType = paddle::PaddleDType;
+using PD_ACPrecision = paddle::AnalysisConfig::Precision;
+
+struct PD_AnalysisConfig {
+  paddle::AnalysisConfig config;
+};
+
+struct PD_Tensor {
+  paddle::PaddleTensor tensor;
+};
+
+struct PD_PaddleBuf {
+  paddle::PaddleBuf buf;
+};
+
+namespace paddle {
+paddle::PaddleDType ConvertToPaddleDType(PD_DataType dtype);
+
+PD_DataType ConvertToPDDataType(PD_PaddleDType dtype);
+
+PD_ACPrecision ConvertToACPrecision(Precision dtype);
+}

paddle/fluid/inference/capi/pd_config.cc

+// Copyright (c) 2019 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 <algorithm>
+#include <limits>
+#include <map>
+#include <memory>
+#include <string>
+#include <vector>
+#include "paddle/fluid/inference/capi/c_api.h"
+#include "paddle/fluid/inference/capi/c_api_internal.h"
+
+using paddle::ConvertToPaddleDType;
+using paddle::ConvertToPDDataType;
+using paddle::ConvertToACPrecision;
+
+extern "C" {
+
+PD_AnalysisConfig* PD_NewAnalysisConfig() { return new PD_AnalysisConfig; }  //
+
+void PD_DeleteAnalysisConfig(PD_AnalysisConfig* config) {
+  if (config) {
+    delete config;
+    config = nullptr;
+  }
+}
+
+void PD_SetModel(PD_AnalysisConfig* config, const char* model_dir,
+                 const char* params_path) {
+  LOG(INFO) << model_dir;
+  PADDLE_ENFORCE_NOT_NULL(config);
+  LOG(INFO) << std::string(model_dir);
+  if (!params_path) {
+    config->config.SetModel(std::string(model_dir));
+  } else {
+    config->config.SetModel(std::string(model_dir), std::string(params_path));
+  }
+}
+
+void PD_SetProgFile(PD_AnalysisConfig* config, const char* x) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SetProgFile(std::string(x));
+}
+
+void PD_SetParamsFile(PD_AnalysisConfig* config, const char* x) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SetParamsFile(std::string(x));
+}
+
+void PD_SetOptimCacheDir(PD_AnalysisConfig* config, const char* opt_cache_dir) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SetOptimCacheDir(std::string(opt_cache_dir));
+}
+
+const char* PD_ModelDir(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.model_dir().c_str();
+}
+
+const char* PD_ProgFile(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.prog_file().c_str();
+}
+
+const char* PD_ParamsFile(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.params_file().c_str();
+}
+
+void PD_EnableUseGpu(PD_AnalysisConfig* config,
+                     uint64_t memory_pool_init_size_mb, int device_id) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.EnableUseGpu(memory_pool_init_size_mb, device_id);
+}
+
+void PD_DisableGpu(PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.DisableGpu();
+}
+
+bool PD_UseGpu(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.use_gpu();
+}
+
+int PD_GpuDeviceId(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.gpu_device_id();
+}
+
+int PD_MemoryPoolInitSizeMb(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.memory_pool_init_size_mb();
+}
+
+float PD_FractionOfGpuMemoryForPool(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.fraction_of_gpu_memory_for_pool();
+}
+
+void PD_EnableCUDNN(PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.EnableCUDNN();
+}
+
+bool PD_CudnnEnabled(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.cudnn_enabled();
+}
+
+void PD_SwitchIrOptim(PD_AnalysisConfig* config, bool x) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SwitchIrOptim(x);
+}
+
+bool PD_IrOptim(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.ir_optim();
+}
+
+void PD_SwitchUseFeedFetchOps(PD_AnalysisConfig* config, bool x) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SwitchUseFeedFetchOps(x);
+}
+
+bool PD_UseFeedFetchOpsEnabled(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.use_feed_fetch_ops_enabled();
+}
+
+void PD_SwitchSpecifyInputNames(PD_AnalysisConfig* config, bool x) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SwitchSpecifyInputNames(x);
+}
+
+bool PD_SpecifyInputName(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.specify_input_name();
+}
+
+void PD_EnableTensorRtEngine(PD_AnalysisConfig* config, int workspace_size,
+                             int max_batch_size, int min_subgraph_size,
+                             Precision precision, bool use_static,
+                             bool use_calib_mode) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.EnableTensorRtEngine(
+      workspace_size, max_batch_size, min_subgraph_size,
+      paddle::ConvertToACPrecision(precision), use_static, use_calib_mode);
+}
+
+bool PD_TensorrtEngineEnabled(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.tensorrt_engine_enabled();
+}
+
+void PD_EnableAnakinEngine(PD_AnalysisConfig* config, int max_batch_size,
+                           PD_MaxInputShape* max_input_shape,
+                           int max_input_shape_size, int min_subgraph_size,
+                           Precision precision, bool auto_config_layout,
+                           char** passes_filter, int passes_filter_size,
+                           char** ops_filter, int ops_filter_size) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  std::map<std::string, std::vector<int>> mis;
+  if (max_input_shape) {
+    for (int i = 0; i < max_input_shape_size; ++i) {
+      std::vector<int> tmp_shape;
+      tmp_shape.assign(
+          max_input_shape[i].shape,
+          max_input_shape[i].shape + max_input_shape[i].shape_size);
+      mis[std::string(max_input_shape[i].name)] = std::move(tmp_shape);
+    }
+  }
+  std::vector<std::string> pf;
+  std::vector<std::string> of;
+  if (passes_filter) {
+    pf.assign(passes_filter, passes_filter + passes_filter_size);
+  }
+  if (ops_filter) {
+    of.assign(ops_filter, ops_filter + ops_filter_size);
+  }
+
+  config->config.EnableAnakinEngine(max_batch_size, mis, min_subgraph_size,
+                                    paddle::ConvertToACPrecision(precision),
+                                    auto_config_layout, pf, of);
+}
+
+bool PD_AnakinEngineEnabled(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.anakin_engine_enabled();
+}
+
+void PD_SwitchIrDebug(PD_AnalysisConfig* config, bool x) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SwitchIrDebug(x);
+}
+
+void PD_EnableNgraph(PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.EnableNgraph();
+}
+
+bool PD_NgraphEnabled(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.ngraph_enabled();
+}
+
+void PD_EnableMKLDNN(PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.EnableMKLDNN();
+}
+
+void PD_SetMkldnnCacheCapacity(PD_AnalysisConfig* config, int capacity) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SetMkldnnCacheCapacity(capacity);
+}
+
+bool PD_MkldnnEnabled(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.mkldnn_enabled();
+}
+
+void PD_SetCpuMathLibraryNumThreads(PD_AnalysisConfig* config,
+                                    int cpu_math_library_num_threads) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SetCpuMathLibraryNumThreads(cpu_math_library_num_threads);
+}
+
+int PD_CpuMathLibraryNumThreads(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.cpu_math_library_num_threads();
+}
+
+void PD_EnableMkldnnQuantizer(PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.EnableMkldnnQuantizer();
+}
+
+bool PD_MkldnnQuantizerEnabled(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.mkldnn_quantizer_enabled();
+}
+
+void PD_SetModelBuffer(PD_AnalysisConfig* config, const char* prog_buffer,
+                       size_t prog_buffer_size, const char* params_buffer,
+                       size_t params_buffer_size) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SetModelBuffer(prog_buffer, prog_buffer_size, params_buffer,
+                                params_buffer_size);
+}
+
+bool PD_ModelFromMemory(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.model_from_memory();
+}
+
+void PD_EnableMemoryOptim(PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.EnableMemoryOptim();
+}
+
+bool PD_MemoryOptimEnabled(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.enable_memory_optim();
+}
+
+void PD_EnableProfile(PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.EnableProfile();
+}
+
+bool PD_ProfileEnabled(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.profile_enabled();
+}
+
+void PD_SetInValid(PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  config->config.SetInValid();
+}
+
+bool PD_IsValid(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config);
+  return config->config.is_valid();
+}
+}  // extern "C"

paddle/fluid/inference/capi/pd_predictor.cc

+// Copyright (c) 2019 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 <algorithm>
+#include <map>
+#include <numeric>
+#include <vector>
+#include "paddle/fluid/inference/capi/c_api.h"
+#include "paddle/fluid/inference/capi/c_api_internal.h"
+
+using paddle::ConvertToPaddleDType;
+using paddle::ConvertToPDDataType;
+using paddle::ConvertToACPrecision;
+
+extern "C" {
+
+bool PD_PredictorRun(const PD_AnalysisConfig* config, PD_Tensor* inputs,
+                     int in_size, PD_Tensor* output_data, int** out_size,
+                     int batch_size) {
+  auto predictor = paddle::CreatePaddlePredictor(config->config);
+  std::vector<paddle::PaddleTensor> in;
+  for (int i = 0; i < in_size; ++i) {
+    in.emplace_back(inputs->tensor);
+  }
+  std::vector<paddle::PaddleTensor> out;
+  if (predictor->Run(in, &out, batch_size)) {
+    int osize = out.size();
+    for (int i = 0; i < osize; ++i) {
+      output_data[i].tensor = out[i];
+    }
+    *out_size = &osize;
+    return true;
+  }
+  return false;
+}
+
+bool PD_PredictorZeroCopyRun(const PD_AnalysisConfig* config,
+                             PD_ZeroCopyData* inputs, int in_size,
+                             PD_ZeroCopyData* output, int** out_size) {
+  auto predictor = paddle::CreatePaddlePredictor(config->config);
+  auto input_names = predictor->GetInputNames();
+  PADDLE_ENFORCE_EQ(
+      input_names.size(), in_size,
+      "The number of input and the number of model's input must match. ");
+  for (int i = 0; i < in_size; ++i) {
+    auto input_t = predictor->GetInputTensor(inputs[i].name);
+    std::vector<int> tensor_shape;
+    tensor_shape.assign(inputs[i].shape,
+                        inputs[i].shape + inputs[i].shape_size);
+    input_t->Reshape(tensor_shape);
+    switch (inputs[i].dtype) {
+      case PD_FLOAT32:
+        input_t->copy_from_cpu(static_cast<float*>(inputs[i].data));
+        break;
+      case PD_INT32:
+        input_t->copy_from_cpu(static_cast<int32_t*>(inputs[i].data));
+        break;
+      case PD_INT64:
+        input_t->copy_from_cpu(static_cast<int64_t*>(inputs[i].data));
+        break;
+      case PD_UINT8:
+        input_t->copy_from_cpu(static_cast<uint8_t*>(inputs[i].data));
+        break;
+      default:
+        CHECK(false) << "Unsupport data type.";
+        break;
+    }
+  }
+  CHECK(predictor->ZeroCopyRun());
+  auto output_names = predictor->GetOutputNames();
+  int osize = output_names.size();
+  *out_size = &osize;
+  output = new PD_ZeroCopyData[osize];
+  for (int i = 0; i < osize; ++i) {
+    LOG(INFO) << 1;
+    output[i].name = new char[output_names[i].length() + 1];
+    snprintf(output[i].name, output_names[i].length() + 1, "%s",
+             output_names[i].c_str());
+    auto output_t = predictor->GetOutputTensor(output_names[i]);
+    output[i].dtype = ConvertToPDDataType(output_t->type());
+    std::vector<int> output_shape = output_t->shape();
+    output[i].shape = new int[output_shape.size()];
+    output[i].shape = output_shape.data();
+    output[i].shape_size = output_shape.size();
+    switch (output[i].dtype) {
+      case PD_FLOAT32: {
+        std::vector<float> out_data;
+        int out_num = std::accumulate(output_shape.begin(), output_shape.end(),
+                                      1, std::multiplies<int>());
+        out_data.resize(out_num);
+        output_t->copy_to_cpu(out_data.data());
+        output[i].data = static_cast<void*>(out_data.data());
+      } break;
+      case PD_INT32: {
+        std::vector<int32_t> out_data;
+        int out_num = std::accumulate(output_shape.begin(), output_shape.end(),
+                                      1, std::multiplies<int>());
+        out_data.resize(out_num);
+        output_t->copy_to_cpu(out_data.data());
+        output[i].data = static_cast<void*>(out_data.data());
+      } break;
+      case PD_INT64: {
+        std::vector<int64_t> out_data;
+        int out_num = std::accumulate(output_shape.begin(), output_shape.end(),
+                                      1, std::multiplies<int>());
+        out_data.resize(out_num);
+        output_t->copy_to_cpu(out_data.data());
+        output[i].data = static_cast<void*>(out_data.data());
+      } break;
+      case PD_UINT8: {
+        std::vector<uint8_t> out_data;
+        int out_num = std::accumulate(output_shape.begin(), output_shape.end(),
+                                      1, std::multiplies<int>());
+        out_data.resize(out_num);
+        output_t->copy_to_cpu(out_data.data());
+        output[i].data = static_cast<void*>(out_data.data());
+      } break;
+      default:
+        CHECK(false) << "Unsupport data type.";
+        break;
+    }
+  }
+  return true;
+}
+}  // extern "C"

paddle/fluid/inference/capi/pd_tensor.cc

+// Copyright (c) 2019 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 <algorithm>
+#include <vector>
+#include "paddle/fluid/inference/capi/c_api.h"
+#include "paddle/fluid/inference/capi/c_api_internal.h"
+
+using paddle::ConvertToPaddleDType;
+using paddle::ConvertToPDDataType;
+using paddle::ConvertToACPrecision;
+
+extern "C" {
+// PaddleTensor
+PD_Tensor* PD_NewPaddleTensor() { return new PD_Tensor; }
+
+void PD_DeletePaddleTensor(PD_Tensor* tensor) {
+  if (tensor) {
+    delete tensor;
+    tensor = nullptr;
+  }
+}
+
+void PD_SetPaddleTensorName(PD_Tensor* tensor, char* name) {
+  tensor->tensor.name = std::string(name);
+}
+
+void PD_SetPaddleTensorDType(PD_Tensor* tensor, PD_DataType dtype) {
+  tensor->tensor.dtype = paddle::ConvertToPaddleDType(dtype);
+}
+
+void PD_SetPaddleTensorData(PD_Tensor* tensor, PD_PaddleBuf* buf) {
+  tensor->tensor.data = buf->buf;
+}
+
+void PD_SetPaddleTensorShape(PD_Tensor* tensor, int* shape, int size) {
+  tensor->tensor.shape.assign(shape, shape + size);
+}
+
+const char* PD_GetPaddleTensorName(const PD_Tensor* tensor) {
+  return tensor->tensor.name.c_str();
+}
+
+PD_DataType PD_GetPaddleTensorDType(const PD_Tensor* tensor) {
+  return ConvertToPDDataType(tensor->tensor.dtype);
+}
+
+PD_PaddleBuf* PD_GetPaddleTensorData(const PD_Tensor* tensor) {
+  PD_PaddleBuf* ret = PD_NewPaddleBuf();
+  ret->buf = tensor->tensor.data;
+  return ret;
+}
+
+int* PD_GetPaddleTensorShape(const PD_Tensor* tensor, int** size) {
+  std::vector<int> shape = tensor->tensor.shape;
+  int s = shape.size();
+  *size = &s;
+  return shape.data();
+}
+
+}  // extern "C"

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

@@ -283,4 +283,29 @@ if(WITH_GPU AND TENSORRT_FOUND)
     inference_analysis_test(trt_cascade_rcnn_test SRCS trt_cascade_rcnn_test.cc
             EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
             ARGS --infer_model=${TRT_MODEL_INSTALL_DIR}/trt_inference_test_models)
+    inference_analysis_test(test_analyzer_capi_gpu SRCS analyzer_capi_gpu_tester.cc
+            EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} paddle_fluid_c
+            ARGS --infer_model=${TRT_MODEL_INSTALL_DIR}/trt_inference_test_models)
+endif()
+
+set(CAPI_MODEL_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/capi_tests_models")
+if (NOT EXISTS ${CAPI_MODEL_INSTALL_DIR})
+    inference_download_and_uncompress(${CAPI_MODEL_INSTALL_DIR} ${INFERENCE_URL}/tensorrt_test "trt_inference_test_models.tar.gz")
 endif()
+inference_analysis_test(test_analyzer_capi SRCS analyzer_capi_tester.cc
+            EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} paddle_fluid_c
+            ARGS --infer_model=${CAPI_MODEL_INSTALL_DIR}/trt_inference_test_models)
+
+set(CAPI_MODEL_INSTALL_PD_DIR "${INFERENCE_DEMO_INSTALL_DIR}/capi_mobilenet")
+if (NOT EXISTS ${CAPI_MODEL_INSTALL_PD_DIR})
+    inference_download_and_uncompress(${CAPI_MODEL_INSTALL_PD_DIR} "http://paddlemodels.bj.bcebos.com/" "inference-vis-demos%2Fmobilenet.tar.gz")
+endif()
+inference_analysis_test(test_analyzer_capi_pd_tensor SRCS analyzer_capi_pd_tensor_tester.cc
+            EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} paddle_fluid_c
+            ARGS --infer_model=${CAPI_MODEL_INSTALL_PD_DIR}/model)
+
+if(WITH_MKLDNN)
+  inference_analysis_test(test_analyzer_capi_int SRCS analyzer_capi_int_tester.cc
+            EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} paddle_fluid_c
+            ARGS --infer_model=${INT8_DATA_DIR}/resnet50/model)
+ endif()

paddle/fluid/inference/tests/api/analyzer_capi_gpu_tester.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 <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+#include "paddle/fluid/inference/capi/c_api.h"
+#include "paddle/fluid/inference/tests/api/tester_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+TEST(PD_AnalysisConfig, use_gpu) {
+  std::string model_dir = FLAGS_infer_model + "/mobilenet";
+  PD_AnalysisConfig *config = PD_NewAnalysisConfig();
+
+  PD_DisableGpu(config);
+  PD_SetCpuMathLibraryNumThreads(config, 10);
+  int num_thread = PD_CpuMathLibraryNumThreads(config);
+  CHECK(10 == num_thread) << "NO";
+  PD_SwitchUseFeedFetchOps(config, false);
+  PD_SwitchSpecifyInputNames(config, true);
+  PD_SwitchIrDebug(config, true);
+  PD_SetModel(config, model_dir.c_str());
+  PD_SetOptimCacheDir(config, (FLAGS_infer_model + "/OptimCacheDir").c_str());
+  const char *model_dir_ = PD_ModelDir(config);
+  LOG(INFO) << model_dir_;
+  PD_EnableUseGpu(config, 100, 0);
+  bool use_gpu = PD_UseGpu(config);
+  CHECK(use_gpu) << "NO";
+  int device = PD_GpuDeviceId(config);
+  CHECK(0 == device) << "NO";
+  int init_size = PD_MemoryPoolInitSizeMb(config);
+  CHECK(100 == init_size) << "NO";
+  float frac = PD_FractionOfGpuMemoryForPool(config);
+  LOG(INFO) << frac;
+  PD_EnableCUDNN(config);
+  bool cudnn = PD_CudnnEnabled(config);
+  CHECK(cudnn) << "NO";
+  PD_SwitchIrOptim(config, true);
+  bool ir_optim = PD_IrOptim(config);
+  CHECK(ir_optim) << "NO";
+  PD_EnableTensorRtEngine(config);
+  bool trt_enable = PD_TensorrtEngineEnabled(config);
+  CHECK(trt_enable) << "NO";
+  PD_EnableNgraph(config);
+  bool ngraph_enable = PD_NgraphEnabled(config);
+  LOG(INFO) << ngraph_enable << " Ngraph";
+  PD_EnableMemoryOptim(config);
+  bool memory_optim_enable = PD_MemoryOptimEnabled(config);
+  CHECK(memory_optim_enable) << "NO";
+  PD_EnableProfile(config);
+  bool profiler_enable = PD_ProfileEnabled(config);
+  CHECK(profiler_enable) << "NO";
+  PD_SetInValid(config);
+  bool is_valid = PD_IsValid(config);
+  CHECK(!is_valid) << "NO";
+  PD_DeleteAnalysisConfig(config);
+}
+
+TEST(PD_AnalysisConfig, trt_int8) {
+  std::string model_dir = FLAGS_infer_model + "/mobilenet";
+  PD_AnalysisConfig *config = PD_NewAnalysisConfig();
+  PD_EnableUseGpu(config, 100, 0);
+  PD_EnableTensorRtEngine(config, 1 << 20, 1, 3, Precision::kInt8, false, true);
+  bool trt_enable = PD_TensorrtEngineEnabled(config);
+  CHECK(trt_enable) << "NO";
+  PD_DeleteAnalysisConfig(config);
+}
+
+TEST(PD_AnalysisConfig, trt_fp16) {
+  std::string model_dir = FLAGS_infer_model + "/mobilenet";
+  PD_AnalysisConfig *config = PD_NewAnalysisConfig();
+  PD_EnableUseGpu(config, 100, 0);
+  PD_EnableTensorRtEngine(config, 1 << 20, 1, 3, Precision::kHalf, false,
+                          false);
+  bool trt_enable = PD_TensorrtEngineEnabled(config);
+  CHECK(trt_enable) << "NO";
+  PD_DeleteAnalysisConfig(config);
+}
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tests/api/analyzer_capi_int_tester.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 <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <typeinfo>
+#include <vector>
+#include "paddle/fluid/inference/capi/c_api.h"
+#include "paddle/fluid/inference/tests/api/tester_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+template <typename T>
+void zero_copy_run() {
+  std::string model_dir = FLAGS_infer_model;
+  PD_AnalysisConfig *config = PD_NewAnalysisConfig();
+  PD_DisableGpu(config);
+  PD_SetCpuMathLibraryNumThreads(config, 10);
+  PD_SwitchUseFeedFetchOps(config, false);
+  PD_SwitchSpecifyInputNames(config, true);
+  PD_SwitchIrDebug(config, true);
+  PD_SetModel(config, model_dir.c_str());  //, params_file1.c_str());
+  bool use_feed_fetch = PD_UseFeedFetchOpsEnabled(config);
+  CHECK(!use_feed_fetch) << "NO";
+  bool specify_input_names = PD_SpecifyInputName(config);
+  CHECK(specify_input_names) << "NO";
+
+  const int batch_size = 1;
+  const int channels = 3;
+  const int height = 224;
+  const int width = 224;
+  T input[batch_size * channels * height * width] = {0};
+  int shape[4] = {batch_size, channels, height, width};
+  int shape_size = 4;
+  int in_size = 2;
+  int *out_size;
+  PD_ZeroCopyData *inputs = new PD_ZeroCopyData[2];
+  PD_ZeroCopyData *outputs = new PD_ZeroCopyData;
+  inputs[0].data = static_cast<void *>(input);
+  std::string nm = typeid(T).name();
+  if ("f" == nm) {
+    inputs[0].dtype = PD_FLOAT32;
+  } else if ("i" == nm) {
+    inputs[0].dtype = PD_INT32;
+  } else if ("x" == nm) {
+    inputs[0].dtype = PD_INT64;
+  } else if ("h" == nm) {
+    inputs[0].dtype = PD_UINT8;
+  } else {
+    CHECK(false) << "Unsupport dtype. ";
+  }
+  inputs[0].name = new char[6];
+  inputs[0].name[0] = 'i';
+  inputs[0].name[1] = 'm';
+  inputs[0].name[2] = 'a';
+  inputs[0].name[3] = 'g';
+  inputs[0].name[4] = 'e';
+  inputs[0].name[5] = '\0';
+  inputs[0].shape = shape;
+  inputs[0].shape_size = shape_size;
+
+  int *label = new int[1];
+  label[0] = 0;
+  inputs[1].data = static_cast<void *>(label);
+  inputs[1].dtype = PD_INT64;
+  inputs[1].name = new char[6];
+  inputs[1].name[0] = 'l';
+  inputs[1].name[1] = 'a';
+  inputs[1].name[2] = 'b';
+  inputs[1].name[3] = 'e';
+  inputs[1].name[4] = 'l';
+  inputs[1].name[5] = '\0';
+  int label_shape[2] = {1, 1};
+  int label_shape_size = 2;
+  inputs[1].shape = label_shape;
+  inputs[1].shape_size = label_shape_size;
+
+  PD_PredictorZeroCopyRun(config, inputs, in_size, outputs, &out_size);
+}
+
+TEST(PD_ZeroCopyRun, zero_copy_run) {
+  // zero_copy_run<int32_t>();
+  // zero_copy_run<int64_t>();
+  zero_copy_run<float>();
+}
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tests/api/analyzer_capi_pd_tensor_tester.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 <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <fstream>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <vector>
+#include "paddle/fluid/inference/capi/c_api.h"
+#include "paddle/fluid/inference/tests/api/tester_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+void PD_run() {
+  PD_AnalysisConfig* config = PD_NewAnalysisConfig();
+  std::string prog_file = FLAGS_infer_model + "/__model__";
+  std::string params_file = FLAGS_infer_model + "/__params__";
+  PD_SetModel(config, prog_file.c_str(), params_file.c_str());
+  PD_SetProgFile(config, prog_file.c_str());
+  PD_SetParamsFile(config, params_file.c_str());
+  LOG(INFO) << PD_ProgFile(config);
+  LOG(INFO) << PD_ParamsFile(config);
+  PD_Tensor* input = PD_NewPaddleTensor();
+  PD_PaddleBuf* buf = PD_NewPaddleBuf();
+  LOG(INFO) << "PaddleBuf empty: " << PD_PaddleBufEmpty(buf);
+  int batch = 1;
+  int channel = 3;
+  int height = 300;
+  int width = 300;
+  int shape[4] = {batch, channel, height, width};
+  int shape_size = 4;
+  float* data = new float[batch * channel * height * width];
+  PD_PaddleBufReset(buf, static_cast<void*>(data),
+                    sizeof(float) * (batch * channel * height * width));
+
+  char name[6] = {'i', 'm', 'a', 'g', 'e', '\0'};
+  PD_SetPaddleTensorName(input, name);
+  PD_SetPaddleTensorDType(input, PD_FLOAT32);
+  PD_SetPaddleTensorShape(input, shape, shape_size);
+  PD_SetPaddleTensorData(input, buf);
+
+  PD_Tensor* out_data = PD_NewPaddleTensor();
+  int* out_size;
+  PD_PredictorRun(config, input, 1, out_data, &out_size, 1);
+  LOG(INFO) << *out_size;
+  LOG(INFO) << PD_GetPaddleTensorName(out_data);
+  LOG(INFO) << PD_GetPaddleTensorDType(out_data);
+  PD_PaddleBuf* b = PD_GetPaddleTensorData(out_data);
+  LOG(INFO) << PD_PaddleBufLength(b);
+  float* result = static_cast<float*>(PD_PaddleBufData(b));
+  LOG(INFO) << *result;
+  PD_PaddleBufResize(b, 500);
+  PD_DeletePaddleTensor(input);
+  int* size;
+  PD_GetPaddleTensorShape(out_data, &size);
+  PD_DeletePaddleBuf(buf);
+}
+
+TEST(PD_Tensor, PD_run) { PD_run(); }
+
+TEST(PD_Tensor, int32) {
+  PD_Tensor* input = PD_NewPaddleTensor();
+  PD_SetPaddleTensorDType(input, PD_INT32);
+  LOG(INFO) << PD_GetPaddleTensorDType(input);
+}
+
+TEST(PD_Tensor, int64) {
+  PD_Tensor* input = PD_NewPaddleTensor();
+  PD_SetPaddleTensorDType(input, PD_INT64);
+  LOG(INFO) << PD_GetPaddleTensorDType(input);
+}
+
+TEST(PD_Tensor, int8) {
+  PD_Tensor* input = PD_NewPaddleTensor();
+  PD_SetPaddleTensorDType(input, PD_UINT8);
+  LOG(INFO) << PD_GetPaddleTensorDType(input);
+}
+
+std::string read_file(std::string filename) {
+  std::ifstream file(filename);
+  return std::string((std::istreambuf_iterator<char>(file)),
+                     std::istreambuf_iterator<char>());
+}
+
+void buffer_run() {
+  PD_AnalysisConfig* config = PD_NewAnalysisConfig();
+  std::string prog_file = FLAGS_infer_model + "/__model__";
+  std::string params_file = FLAGS_infer_model + "/__params__";
+
+  std::string prog_str = read_file(prog_file);
+  std::string params_str = read_file(params_file);
+
+  PD_SetModelBuffer(config, prog_str.c_str(), prog_str.size(),
+                    params_str.c_str(), params_str.size());
+  LOG(INFO) << PD_ProgFile(config);
+  LOG(INFO) << PD_ParamsFile(config);
+  CHECK(PD_ModelFromMemory(config)) << "NO";
+
+  PD_Tensor* input = PD_NewPaddleTensor();
+  PD_PaddleBuf* buf = PD_NewPaddleBuf();
+  LOG(INFO) << "PaddleBuf empty: " << PD_PaddleBufEmpty(buf);
+  int batch = 1;
+  int channel = 3;
+  int height = 300;
+  int width = 300;
+  int shape[4] = {batch, channel, height, width};
+  int shape_size = 4;
+  float* data = new float[batch * channel * height * width];
+  PD_PaddleBufReset(buf, static_cast<void*>(data),
+                    sizeof(float) * (batch * channel * height * width));
+
+  char name[6] = {'i', 'm', 'a', 'g', 'e', '\0'};
+  PD_SetPaddleTensorName(input, name);
+  PD_SetPaddleTensorDType(input, PD_FLOAT32);
+  PD_SetPaddleTensorShape(input, shape, shape_size);
+  PD_SetPaddleTensorData(input, buf);
+
+  PD_Tensor* out_data = PD_NewPaddleTensor();
+  int* out_size;
+  PD_PredictorRun(config, input, 1, out_data, &out_size, 1);
+  LOG(INFO) << *out_size;
+  LOG(INFO) << PD_GetPaddleTensorName(out_data);
+  LOG(INFO) << PD_GetPaddleTensorDType(out_data);
+  PD_PaddleBuf* b = PD_GetPaddleTensorData(out_data);
+  LOG(INFO) << PD_PaddleBufLength(b);
+  float* result = static_cast<float*>(PD_PaddleBufData(b));
+  LOG(INFO) << *result;
+  PD_PaddleBufResize(b, 500);
+  PD_DeletePaddleTensor(input);
+  PD_DeletePaddleBuf(buf);
+}
+
+TEST(SetModelBuffer, read) { buffer_run(); }
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tests/api/analyzer_capi_tester.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 <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <typeinfo>
+#include <vector>
+#include "paddle/fluid/inference/capi/c_api.h"
+#include "paddle/fluid/inference/tests/api/tester_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+template <typename T>
+void zero_copy_run() {
+  std::string model_dir = FLAGS_infer_model + "/mobilenet";
+  PD_AnalysisConfig *config = PD_NewAnalysisConfig();
+  PD_DisableGpu(config);
+  PD_SetCpuMathLibraryNumThreads(config, 10);
+  PD_SwitchUseFeedFetchOps(config, false);
+  PD_SwitchSpecifyInputNames(config, true);
+  PD_SwitchIrDebug(config, true);
+  PD_SetModel(config, model_dir.c_str());  //, params_file1.c_str());
+  bool use_feed_fetch = PD_UseFeedFetchOpsEnabled(config);
+  CHECK(!use_feed_fetch) << "NO";
+  bool specify_input_names = PD_SpecifyInputName(config);
+  CHECK(specify_input_names) << "NO";
+
+  const int batch_size = 1;
+  const int channels = 3;
+  const int height = 224;
+  const int width = 224;
+  T input[batch_size * channels * height * width] = {0};
+
+  int shape[4] = {batch_size, channels, height, width};
+  int shape_size = 4;
+  int in_size = 1;
+  int *out_size;
+  PD_ZeroCopyData *inputs = new PD_ZeroCopyData;
+  PD_ZeroCopyData *outputs = new PD_ZeroCopyData;
+  inputs->data = static_cast<void *>(input);
+  std::string nm = typeid(T).name();
+  if ("f" == nm) {
+    inputs->dtype = PD_FLOAT32;
+  } else if ("i" == nm) {
+    inputs->dtype = PD_INT32;
+  } else if ("x" == nm) {
+    inputs->dtype = PD_INT64;
+  } else if ("h" == nm) {
+    inputs->dtype = PD_UINT8;
+  } else {
+    CHECK(false) << "Unsupport dtype. ";
+  }
+  inputs->name = new char[2];
+  inputs->name[0] = 'x';
+  inputs->name[1] = '\0';
+  LOG(INFO) << inputs->name;
+  inputs->shape = shape;
+  inputs->shape_size = shape_size;
+
+  PD_PredictorZeroCopyRun(config, inputs, in_size, outputs, &out_size);
+}
+
+TEST(PD_ZeroCopyRun, zero_copy_run) { zero_copy_run<float>(); }
+
+#ifdef PADDLE_WITH_MKLDNN
+TEST(PD_AnalysisConfig, profile_mkldnn) {
+  std::string model_dir = FLAGS_infer_model + "/mobilenet";
+  PD_AnalysisConfig *config = PD_NewAnalysisConfig();
+  PD_DisableGpu(config);
+  PD_SetCpuMathLibraryNumThreads(config, 10);
+  PD_SwitchUseFeedFetchOps(config, false);
+  PD_SwitchSpecifyInputNames(config, true);
+  PD_SwitchIrDebug(config, true);
+  PD_EnableMKLDNN(config);
+  bool mkldnn_enable = PD_MkldnnEnabled(config);
+  CHECK(mkldnn_enable) << "NO";
+  PD_EnableMkldnnQuantizer(config);
+  bool quantizer_enable = PD_MkldnnQuantizerEnabled(config);
+  CHECK(quantizer_enable) << "NO";
+  PD_SetMkldnnCacheCapacity(config, 0);
+  PD_SetModel(config, model_dir.c_str());
+  PD_EnableAnakinEngine(config);
+  bool anakin_enable = PD_AnakinEngineEnabled(config);
+  LOG(INFO) << anakin_enable;
+  PD_DeleteAnalysisConfig(config);
+}
+#endif
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle