Merge remote-tracking branch 'ups/develop' into feature/op/fusion_gru

Author: tensor-tang

cmake/external/anakin.cmake

@@ -2,6 +2,11 @@ if (NOT WITH_ANAKIN)
   return()
 endif()
 
+option(ANAKIN_ENABLE_OP_TIMER      "Get more detailed information with Anakin op time"        OFF)
+if(ANAKIN_ENABLE_OP_TIMER)
+  add_definitions(-DPADDLE_ANAKIN_ENABLE_OP_TIMER)
+endif()
+
 INCLUDE(ExternalProject)
 set(ANAKIN_SOURCE_DIR  ${THIRD_PARTY_PATH}/anakin)
 # the anakin install dir is only default one now
@@ -11,23 +16,34 @@ set(ANAKIN_LIBRARY     ${ANAKIN_INSTALL_DIR})
 set(ANAKIN_SHARED_LIB  ${ANAKIN_LIBRARY}/libanakin.so)
 set(ANAKIN_SABER_LIB   ${ANAKIN_LIBRARY}/libanakin_saber_common.so)
 
-# TODO(luotao): ANAKIN_MODLE_URL will move to demo ci later.
-set(ANAKIN_MODLE_URL "http://paddle-inference-dist.bj.bcebos.com/mobilenet_v2.anakin.bin")
+# TODO(luotao): ANAKIN_MODLE_URL etc will move to demo ci later.
+set(INFERENCE_URL "http://paddle-inference-dist.bj.bcebos.com")
+set(ANAKIN_MODLE_URL "${INFERENCE_URL}/mobilenet_v2.anakin.bin")
+set(ANAKIN_RNN_MODLE_URL "${INFERENCE_URL}/anakin_test%2Fditu_rnn.anakin2.model.bin")
+set(ANAKIN_RNN_DATA_URL "${INFERENCE_URL}/anakin_test%2Fditu_rnn_data.txt")
 execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_SOURCE_DIR}")
-execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_MODLE_URL}")
+execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_MODLE_URL} -N")
+execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_RNN_MODLE_URL} -N")
+execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_RNN_DATA_URL} -N")
 
 include_directories(${ANAKIN_INCLUDE})
 include_directories(${ANAKIN_INCLUDE}/saber/)
+include_directories(${ANAKIN_INCLUDE}/saber/core/)
+include_directories(${ANAKIN_INCLUDE}/saber/funcs/impl/x86/)
+include_directories(${ANAKIN_INCLUDE}/saber/funcs/impl/cuda/base/cuda_c/)
 
 set(ANAKIN_COMPILE_EXTRA_FLAGS
     -Wno-error=unused-but-set-variable -Wno-unused-but-set-variable
     -Wno-error=unused-variable -Wno-unused-variable
     -Wno-error=format-extra-args -Wno-format-extra-args
-    -Wno-error=comment -Wno-comment
-    -Wno-error=format -Wno-format
+    -Wno-error=comment -Wno-comment 
+    -Wno-error=format -Wno-format 
+    -Wno-error=maybe-uninitialized -Wno-maybe-uninitialized
     -Wno-error=switch -Wno-switch
     -Wno-error=return-type -Wno-return-type
     -Wno-error=non-virtual-dtor -Wno-non-virtual-dtor
+    -Wno-error=ignored-qualifiers
+    -Wno-ignored-qualifiers
     -Wno-sign-compare
     -Wno-reorder
     -Wno-error=cpp)
@@ -38,7 +54,7 @@ ExternalProject_Add(
     DEPENDS             ${MKLML_PROJECT}
     # Anakin codes error on Intel(R) Xeon(R) Gold 5117 CPU, temporary do not compile avx512 related code.
     GIT_REPOSITORY      "https://github.com/luotao1/Anakin"
-    GIT_TAG             "bcf17aabe7921ceb7bce591244b4f9dce7dba5c8"
+    GIT_TAG             "211d1fc5d813d70c0c14072f9083cf25f40940ea"
     PREFIX              ${ANAKIN_SOURCE_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DUSE_GPU_PLACE=YES
@@ -48,6 +64,7 @@ ExternalProject_Add(
                         -DMKLML_ROOT=${THIRD_PARTY_PATH}/install/mklml
                         -DCUDNN_ROOT=${CUDNN_ROOT}
                         -DCUDNN_INCLUDE_DIR=${CUDNN_INCLUDE_DIR}
+                        -DENABLE_OP_TIMER=${ANAKIN_ENABLE_OP_TIMER}
                         ${EXTERNAL_OPTIONAL_ARGS}
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${ANAKIN_INSTALL_DIR}
 )

doc/fluid/dev/new_op_cn.md

@@ -119,10 +119,29 @@ $$Out = scale*X$$
 
 这个例子有`AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。
 
+### 定义GradProtoMaker类
+每个Op的必须有一个对应的GraProtoMaker，若未定制对应前向Op的GradProtoMaker，fluid提供了DefaultGradProtoMaker，默认注册会使用全部输入输出，包括Input, Output, Output@Grad等，使用不需要的变量的会造成显存浪费。
+下面示例定义了ScaleOp的GradProtoMaker。
+
+```cpp
+class ScaleGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto *grad_op = new framework::OpDesc();
+    grad_op->SetType("scale");
+    grad_op->SetInput("X", OutputGrad("Out"));
+    grad_op->SetOutput("Out", InputGrad("X"));
+    grad_op->SetAttr("scale", GetAttr("scale"));
+    return std::unique_ptr<framework::OpDesc>(grad_op);
+  }
+};
+```
 
 ### 定义Operator类
 
-下面的点实现了MulOp的定义：
+下面实现了MulOp的定义：
 
 ```cpp
 class MulOp : public framework::OperatorWithKernel {
@@ -334,3 +353,83 @@ ctest -R test_mul_op
 - 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OPERATOR(B, ...)`等，这将会导致单元测试出错。
 - 如果Op没有实现CUDA Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
 - 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。
+
+### PADDLE_ENFORCE使用注意
+
+实现Op时检查数据的合法性需要使用PADDLE_ENFORCE以及PADDLE_ENFORCE_EQ等宏定义，基本格式如下：
+
+```
+PADDLE_ENFORCE(表达式, 错误提示信息)
+PADDLE_ENFORCE_EQ(比较对象A, 比较对象B, 错误提示信息)
+```
+
+如果表达式为真，或者比较对象A=B，则检查通过，否则会终止程序运行，向用户反馈相应的错误提示信息。
+为了确保提示友好易懂，开发者需要注意其使用方法。
+
+#### 总体原则
+
+任何使用了PADDLE_ENFORCE与PADDLE_ENFORCE_**检查的地方，必须有详略得当的备注解释！**错误提示信息**不能为空！
+
+#### 提示信息书写标准
+
+1. [required] 哪里错了？为什么错了？
+    - 例如：`ValueError: Mismatched label shape`
+2. [optional] 期望的输入是什么样的？实际的输入是怎样的？
+    - 例如：`Expected labels dimension=1. Received 4.`
+3. [optional] 能否给出修改意见？
+    - 例如：`Suggested Fix:If your classifier expects one-hot encoding label,check your n_classes argument to the estimatorand/or the shape of your label.Otherwise, check the shape of your label.`
+
+如果并非必要或者简洁的描述即可表达清楚以上要点，根据情况书写亦可。
+
+##### FAQ 典型问题
+
+1. 无报错信息或报错信息过于简单，不能给用户提供有效的提示！
+
+问题示例1 ：未写提示信息
+```
+PADDLE_ENFORCE(ctx->HasInput("X"), "");
+```
+问题示例2 ：提示信息过于简单
+```
+PADDLE_ENFORCE(i != nullptr, "I must be set"); // I是什么？
+```
+
+2. 在报错信息中使用开发人员定义的变量缩写，不易理解！
+
+问题示例：
+```
+PADDLE_ENFORCE(forward_pd != nullptr,
+                    "Fail to find eltwise_fwd_pd in device context");  //eltwise_fwd_pd用户可能看不懂
+```
+
+3. OP内部调用非法接口：Op内部如果出现Output = ShareDataWith(Input) 
+问题示例：
+```cpp
+auto *out = ctx.Output<framework::LoDTensor>("Out");
+auto *in = ctx.Input<framework::LoDTensor>("X");
+out->ShareDataWith(*in);
+```
+Op内部如果出现Output = ShareDataWith(Input)，相当于operator图的中有一条隐藏边，连接了Input和Output，这条边无法在图分析中表达，引发基于图优化的错误。
+
+4. OP实现的性能实践
+调用了eigen的broadcast, chop等操作，性能会比手写cuda kernel差几倍以上。此时cpu的实现可以复用eigen，gpu实现可以实现cuda kernel.
+
+
+#### OP InferShape检查提示信息特别说明
+
+- 检查输入输出变量，请统一遵循以下格式
+`Input(变量名) of OP名 operator should not be null.`  
+
+正确示例：
+```
+PADDLE_ENFORCE(ctx->HasInput("Input"),
+                        "Input(Input) of LSTMP operator should not be null.");
+```
+
+- 反向Op的输入输出检查，要写明反向Op的名字
+
+正确示例：
+```
+PADDLE_ENFORCE(ctx->HasInput("X"),
+                        "Input(X) of LoDResetGrad opreator should not be null.");
+```

paddle/fluid/framework/ir/graph_helper.cc

@@ -104,7 +104,7 @@ std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
       for (auto &adj_n : var->inputs) {
         PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
         adj_list[n].insert(adj_n);
-        VLOG(3) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
+        VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
                 << " -> " << n->Name() << reinterpret_cast<void *>(n)
                 << "  via " << var->Name() << reinterpret_cast<void *>(var);
       }

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -22,7 +22,7 @@ function (inference_analysis_test TARGET)
     if(WITH_TESTING)
         set(options "")
         set(oneValueArgs "")
-        set(multiValueArgs SRCS)
+        set(multiValueArgs SRCS EXTRA_DEPS)
         cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
 
         set(mem_opt "")
@@ -31,22 +31,43 @@ function (inference_analysis_test TARGET)
         endif()
         cc_test(${TARGET}
                 SRCS "${analysis_test_SRCS}"
-                DEPS analysis graph fc_fuse_pass graph_viz_pass infer_clean_graph_pass graph_pattern_detecter pass
+                DEPS analysis graph fc_fuse_pass graph_viz_pass infer_clean_graph_pass graph_pattern_detecter pass ${analysis_test_EXTRA_DEPS}
                 ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt})
         set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec)
     endif(WITH_TESTING)
 endfunction(inference_analysis_test)
 
-cc_test(test_analyzer SRCS analyzer_tester.cc DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis
-		  # ir
-		  fc_fuse_pass
-		  graph_viz_pass
-		  infer_clean_graph_pass
-		  graph_pattern_detecter
-		  pass
-        ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model)
-#set_tests_properties(test_analyzer PROPERTIES DEPENDS test_word2vec)
-#inference_api_test(test_analyzer SRC analyzer_tester.cc ARGS test_word2vec)
+set(DITU_RNN_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/ditu_rnn_fluid%2Fmodel.tar.gz")
+set(DITU_RNN_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/ditu_rnn_fluid%2Fdata.txt.tar.gz")
+set(DITU_INSTALL_DIR "${THIRD_PARTY_PATH}/install/ditu_rnn" CACHE PATH "Ditu RNN model and data root." FORCE)
+set(DITU_RNN_MODEL ${DITU_INSTALL_DIR}/model)
+set(DITU_RNN_DATA ${DITU_INSTALL_DIR}/data.txt)
+
+function (inference_download_and_uncompress target url gz_filename)
+    message(STATUS "Download inference test stuff ${gz_filename} from ${url}")
+    execute_process(COMMAND bash -c "mkdir -p ${DITU_INSTALL_DIR}")
+    execute_process(COMMAND bash -c "cd ${DITU_INSTALL_DIR} && wget -q ${url}")
+    execute_process(COMMAND bash -c "cd ${DITU_INSTALL_DIR} && tar xzf ${gz_filename}")
+    message(STATUS "finish downloading ${gz_filename}")
+endfunction(inference_download_and_uncompress)
+
+if (NOT EXISTS ${DITU_INSTALL_DIR})
+    inference_download_and_uncompress(ditu_rnn_model ${DITU_RNN_MODEL_URL} "ditu_rnn_fluid%2Fmodel.tar.gz")
+    inference_download_and_uncompress(ditu_rnn_data ${DITU_RNN_DATA_URL} "ditu_rnn_fluid%2Fdata.txt.tar.gz")
+endif()
+
+inference_analysis_test(test_analyzer SRCS analyzer_tester.cc
+    EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis
+		# ir
+		fc_fuse_pass
+		graph_viz_pass
+		infer_clean_graph_pass
+		graph_pattern_detecter
+        infer_clean_graph_pass
+		pass
+    ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model
+        --infer_ditu_rnn_model=${DITU_INSTALL_DIR}/model
+        --infer_ditu_rnn_data=${DITU_INSTALL_DIR}/data.txt)
 
 inference_analysis_test(test_data_flow_graph SRCS data_flow_graph_tester.cc)
 inference_analysis_test(test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc)

paddle/fluid/inference/analysis/analyzer.cc

@@ -23,8 +23,6 @@
 #include "paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h"
 #include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h"
 
-namespace paddle {
-
 DEFINE_bool(IA_enable_tensorrt_subgraph_engine, false,
             "Enable subgraph to TensorRT engine for acceleration");
 
@@ -35,6 +33,7 @@ DEFINE_string(IA_graphviz_log_root, "./",
 
 DEFINE_string(IA_output_storage_path, "", "optimized model output path");
 
+namespace paddle {
 namespace inference {
 namespace analysis {
 

paddle/fluid/inference/analysis/analyzer.h

@@ -39,15 +39,14 @@ limitations under the License. */
 #include "paddle/fluid/inference/analysis/pass.h"
 #include "paddle/fluid/inference/analysis/pass_manager.h"
 
-namespace paddle {
-
 // TODO(Superjomn) add a definition flag like PADDLE_WITH_TENSORRT and hide this
 // flag if not available.
 DECLARE_bool(IA_enable_tensorrt_subgraph_engine);
 DECLARE_string(IA_graphviz_log_root);
 DECLARE_string(IA_output_storage_path);
 DECLARE_bool(IA_enable_ir);
 
+namespace paddle {
 namespace inference {
 namespace analysis {