Merge branch 'develop' of github.com:PaddlePaddle/Paddle into dist_test_word2vec

Author: Yancey0623

benchmark/fluid/args.py

@@ -122,5 +122,9 @@ def parse_args():
         type=str,
         default="",
         help='Directory that contains all the training recordio files.')
+    parser.add_argument(
+        '--use_inference_transpiler',
+        action='store_true',
+        help='If set, uses inference transpiler to optimize the program.')
     args = parser.parse_args()
     return args

benchmark/fluid/fluid_benchmark.py

@@ -131,6 +131,11 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
     exe = fluid.Executor(place)
     exe.run(startup_prog)
 
+    # Use inference_transpiler to speedup
+    if args.use_inference_transpiler:
+        t = fluid.InferenceTranspiler()
+        t.transpile(infer_prog, place)
+
     if not args.use_reader_op:
         feed_var_list = [
             var for var in train_prog.global_block().vars.itervalues()

cmake/external/anakin.cmake

@@ -26,13 +26,15 @@ function(fetch_include_recursively root_dir)
     endforeach()
 endfunction()
 
-# download library
-message(STATUS "Download Anakin library from ${ANAKIN_LIBRARY_URL}")
-execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
-execute_process(COMMAND bash -c "rm -rf ${ANAKIN_INSTALL_DIR}/*")
-execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; wget -q ${ANAKIN_LIBRARY_URL}")
-execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
-execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; tar xzf anakin_release_simple.tar.gz")
+if (NOT EXISTS "${ANAKIN_INSTALL_DIR}")
+    # download library
+    message(STATUS "Download Anakin library from ${ANAKIN_LIBRARY_URL}")
+    execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
+    execute_process(COMMAND bash -c "rm -rf ${ANAKIN_INSTALL_DIR}/*")
+    execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; wget -q ${ANAKIN_LIBRARY_URL}")
+    execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
+    execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; tar xzf anakin_release_simple.tar.gz")
+endif()
 
 if (WITH_ANAKIN)
     message(STATUS "Anakin for inference is enabled")

cmake/inference_lib.cmake

@@ -149,21 +149,33 @@ copy(memory_lib
   DSTS ${dst_dir}/${module} ${dst_dir}/${module}/detail
 )
 
-set(module "inference")
-copy(inference_lib DEPS paddle_fluid_shared paddle_fluid
-  SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
-  DSTS ${dst_dir}/${module} ${dst_dir}/${module}
-)
+set(inference_deps paddle_fluid_shared paddle_fluid)
 
 if(WITH_CONTRIB)
-   set(contrib_dst_dir "${FLUID_INSTALL_DIR}/contrib/inference")
-   copy(contrib_inference_lib DEPS paddle_inference_api
+    message(STATUS "installing contrib")
+    set(contrib_dst_dir "${FLUID_INSTALL_DIR}/contrib/inference")
+    if (WITH_ANAKIN)
+        copy(contrib_anakin_inference_lib DEPS paddle_inference_api inference_anakin_api
+            SRCS
+            ${PADDLE_BINARY_DIR}/paddle/contrib/inference/libinference_anakin_api* # compiled anakin api
+            ${PADDLE_BINARY_DIR}/third_party/install/anakin/*.tar.gz # anakin release
+            DSTS ${contrib_dst_dir}/anakin ${contrib_dst_dir}/anakin)
+        list(APPEND inference_deps contrib_anakin_inference_lib)
+   endif()
+
+  copy(contrib_inference_lib DEPS paddle_inference_api
         SRCS ${PADDLE_SOURCE_DIR}/paddle/contrib/inference/paddle_inference_api.h
         ${PADDLE_BINARY_DIR}/paddle/contrib/inference/libpaddle_inference_api.*
-        DSTS ${contrib_dst_dir} ${contrib_dst_dir}
-   )
+        DSTS ${contrib_dst_dir} ${contrib_dst_dir})
+  list(APPEND inference_deps contrib_inference_lib)
 endif()
 
+set(module "inference")
+copy(inference_lib DEPS ${inference_deps}
+  SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module}
+)
+
 set(module "platform")
 copy(platform_lib DEPS profiler_py_proto
   SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/dynload/*.h ${src_dir}/${module}/details/*.h

paddle/contrib/inference/CMakeLists.txt

@@ -18,7 +18,7 @@ if(APPLE)
 endif(APPLE)
 
 
-set(inference_deps paddle_inference_api paddle_fluid_api)
+set(inference_deps paddle_inference_api paddle_fluid_api paddle_inference_tensorrt_subgraph_engine)
 
 function(inference_api_test TARGET_NAME)
     if (WITH_TESTING)
@@ -50,13 +50,24 @@ cc_test(test_paddle_inference_api
 inference_api_test(test_paddle_inference_api_impl
                     ARGS test_word2vec test_image_classification)
 
+if(WITH_GPU AND TENSORRT_FOUND)
+cc_library(paddle_inference_tensorrt_subgraph_engine
+        SRCS paddle_inference_api_tensorrt_subgraph_engine.cc
+        DEPS paddle_inference_api analysis tensorrt_engine paddle_inference_api paddle_fluid_api)
+
+inference_api_test(test_paddle_inference_api_tensorrt_subgraph_engine ARGS test_word2vec)
+endif()
+
 if (WITH_ANAKIN AND WITH_TESTING) # only needed in CI
     # Due to Anakin do not have official library releases and the versions of protobuf and cuda do not match Paddle's,
     # so anakin library will not be merged to our official inference library. To use anakin prediction API, one need to
     # compile the libinference_anakin_api.a and compile with anakin.so.
-    nv_library(inference_anakin_api SHARED SRCS paddle_inference_api.cc paddle_inference_api_anakin_engine.cc)
+    nv_library(inference_anakin_api SRCS paddle_inference_api.cc paddle_inference_api_anakin_engine.cc)
+    nv_library(inference_anakin_api_shared SHARED SRCS paddle_inference_api.cc paddle_inference_api_anakin_engine.cc)
     target_compile_options(inference_anakin_api BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
+    target_compile_options(inference_anakin_api_shared BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
     target_link_libraries(inference_anakin_api anakin anakin_saber_common)
+    target_link_libraries(inference_anakin_api_shared anakin anakin_saber_common)
     cc_test(inference_anakin_test SRCS paddle_inference_api_anakin_engine_tester.cc
                                   ARGS --model=${ANAKIN_INSTALL_DIR}/mobilenet_v2.anakin.bin
                                   DEPS inference_anakin_api)

paddle/contrib/inference/demo/CMakeLists.txt

@@ -15,6 +15,11 @@
 
 inference_api_test(simple_on_word2vec ARGS test_word2vec)
 
+option(WITH_INFERENCE_DEMO "Compile with Inference demo" OFF)
+if(NOT WITH_INFERENCE_DEMO)
+  return()
+endif()
+
 set(DEMO_INSTALL_DIR "${PADDLE_BINARY_DIR}/inference_demo")
 set(URL_ROOT http://paddlemodels.bj.bcebos.com/inference-vis-demos%2F)
 

paddle/contrib/inference/high_level_api_cn.md

@@ -0,0 +1,87 @@
+# Paddle 预测 API
+
+为了更简单方便的预测部署，Fluid 提供了一套高层 API 用来隐藏底层不同的优化实现。
+
+预测库包含:
+
+- 头文件 `paddle_inference_api.h` 定义了所有的接口
+- 库文件`libpaddle_fluid.so` 或 `libpaddle_fluid.a`
+- 库文件 `libpaddle_inference_api.so` 或 `libpaddle_inference_api.a`
+
+下面是详细的一些 API 概念介绍
+
+## PaddleTensor
+
+PaddleTensor 定义了预测最基本的输入输出的数据格式，其定义是
+
+```c++
+struct PaddleTensor {
+  std::string name;  // variable name.
+  std::vector<int> shape;
+  PaddleBuf data;  // blob of data.
+  PaddleDType dtype;
+};
+```
+
+- `name` 用于指定输入数据对应的 模型中variable 的名字 （暂时没有用，但会在后续支持任意 target 时启用）
+- `shape` 表示一个 Tensor 的 shape
+- `data`  数据以连续内存的方式存储在`PaddleBuf` 中，`PaddleBuf` 可以接收外面的数据或者独立`malloc`内存，详细可以参考头文件中相关定义。
+- `dtype` 表示 Tensor 的数据类型
+
+## engine
+
+高层 API 底层有多种优化实现，我们称之为 engine，目前有三种 engine
+
+- 原生 engine，由 paddle 原生的 forward operator 组成，可以天然支持所有paddle 训练出的模型，
+- Anakin engine，封装了 [Anakin](https://github.com/PaddlePaddle/Anakin) ，在某些模型上性能不错，但只能接受自带模型格式，无法支持所有 paddle 模型，
+- TensorRT mixed engine，用子图的方式支持了 [TensorRT](https://developer.nvidia.com/tensorrt) ，支持所有paddle 模型，并自动切割部分计算子图到 TensorRT 上加速（WIP）
+
+其实现为
+
+```c++
+enum class PaddleEngineKind {
+  kNative = 0,       // Use the native Fluid facility.
+  kAnakin,           // Use Anakin for inference.
+  kAutoMixedTensorRT // Automatically mixing TensorRT with the Fluid ops.
+};
+```
+
+## 预测部署过程
+
+总体上分为以下步骤
+
+1. 用合适的配置创建 `PaddlePredictor`
+2. 创建输入用的 `PaddleTensor`，传入到 `PaddlePredictor` 中
+3. 获取输出的 `PaddleTensor` ，将结果取出
+
+下面完整演示一个简单的模型，部分细节代码隐去
+
+```c++
+#include "paddle_inference_api.h"
+
+// 创建一个 config，并修改相关设置
+paddle::NativeConfig config;
+config.model_dir = "xxx";
+config.use_gpu = false;
+// 创建一个原生的 PaddlePredictor
+auto predictor =
+      paddle::CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+// 创建输入 tensor
+int64_t data[4] = {1, 2, 3, 4};
+paddle::PaddleTensor tensor{.name = "",
+                            .shape = std::vector<int>({4, 1}),
+                            .data = PaddleBuf(data, sizeof(data)),
+                            .dtype = PaddleDType::INT64};
+// 创建输出 tensor，输出 tensor 的内存可以复用
+std::vector<paddle::PaddleTensor> outputs;
+// 执行预测
+CHECK(predictor->Run(slots, &outputs));
+// 获取 outputs ...
+```
+
+编译时，联编 `libpaddle_fluid.a/.so` 和 `libpaddle_inference_api.a/.so` 便可。 
+
+## 详细代码参考
+
+- [inference demos](./demo)
+- [复杂单线程/多线程例子](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/contrib/inference/test_paddle_inference_api_impl.cc)

paddle/contrib/inference/paddle_inference_api.h

@@ -73,12 +73,12 @@ struct PaddleTensor {
 };
 
 enum class PaddleEngineKind {
-  kNative = 0,  // Use the native Fluid facility.
-  kAnakin,      // Use Anakin for inference.
+  kNative = 0,         // Use the native Fluid facility.
+  kAnakin,             // Use Anakin for inference.
+  kAutoMixedTensorRT,  // Automatically mix Fluid with TensorRT.
   // TODO(Superjomn) support following engines latter.
   // kTensorRT,           // Use TensorRT for inference.
   // kAutoMixedAnakin,    // Automatically mix Fluid with Anakin.
-  // kAutoMixedTensorRT,  // Automatically mix Fluid with TensorRT.
 };
 
 /*
@@ -130,6 +130,11 @@ struct AnakinConfig : public PaddlePredictor::Config {
   int max_batch_size{-1};
 };
 
+struct TensorRTConfig : public NativeConfig {
+  // Determine whether a subgraph will be executed by TRT.
+  int min_subgraph_size{1};
+};
+
 // A factory to help create different predictors.
 //
 // FOR EXTENSION DEVELOPER:

paddle/contrib/inference/paddle_inference_api_impl.cc

@@ -89,6 +89,7 @@ bool NativePaddlePredictor::Init(
     LOG(ERROR) << "fail to load inference model.";
     return false;
   }
+
   ctx_ = executor_->Prepare(*inference_program_, 0);
   executor_->CreateVariables(
       *inference_program_, sub_scope_ ? sub_scope_ : scope_.get(), 0);
@@ -119,6 +120,7 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
     return false;
   }
   for (size_t i = 0; i < feed_target_names_.size(); ++i) {
+    VLOG(4) << "setting " << i << "-th target";
     feed_targets[feed_target_names_[i]] = &feeds[i];
   }
   // get fetch variable
@@ -130,14 +132,16 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
   }
   // Run the inference program
   // if share variables, we need not create variables
+  VLOG(4) << "Run prepared context";
   executor_->RunPreparedContext(
       ctx_.get(),
       sub_scope_ != nullptr ? sub_scope_ : scope_.get(),
       &feed_targets,
       &fetch_targets,
       false /* don't create variable eatch time */);
+  VLOG(4) << "Finish prepared context";
   if (!GetFetch(fetchs, output_data)) {
-    LOG(ERROR) << "fail to get fetchs";
+    LOG(ERROR) << "fail to get fetches";
     return false;
   }
   VLOG(3) << "predict cost: " << timer.toc() << "ms";

paddle/contrib/inference/paddle_inference_api_impl.h

@@ -44,7 +44,7 @@ class NativePaddlePredictor : public PaddlePredictor {
 
   ~NativePaddlePredictor() override;
 
- private:
+ protected:
   bool SetFeed(const std::vector<PaddleTensor> &input_datas,
                std::vector<framework::LoDTensor> *feeds);
   bool GetFetch(const std::vector<framework::LoDTensor> &fetchs,