Merge remote-tracking branch 'ups/develop' into refine/op/fc

CMakeLists.txt

@@ -204,12 +204,11 @@ include(external/snappy)    # download snappy
 include(external/snappystream)
 include(external/threadpool)
 
+set(WITH_ANAKIN OFF CACHE STRING "Disable Anakin first, will add it later." FORCE)
 if(WITH_GPU)
     include(cuda)
     include(tensorrt)
     include(external/anakin)
-else()
-  set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when GPU is set." FORCE)
 endif()
 
 include(cudnn)              # set cudnn libraries, must before configure

cmake/configure.cmake

@@ -97,6 +97,14 @@ if(WITH_GPU)
         endif()
         include_directories(${TENSORRT_INCLUDE_DIR})
     endif()
+    if(WITH_ANAKIN)
+        if(${CUDA_VERSION_MAJOR} VERSION_LESS 8)
+            message(FATAL_ERROR "Anakin needs CUDA >= 8.0 to compile")
+        endif()
+        if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
+            message(FATAL_ERROR "Anakin needs CUDNN >= 7.0 to compile")
+        endif()
+    endif()
 elseif(WITH_AMD_GPU)
     add_definitions(-DPADDLE_WITH_HIP)
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__")

cmake/external/anakin.cmake

@@ -2,10 +2,22 @@ if (NOT WITH_ANAKIN)
   return()
 endif()
 
-set(ANAKIN_INSTALL_DIR "${THIRD_PARTY_PATH}/install/anakin" CACHE PATH
-  "Anakin install path." FORCE)
-set(ANAKIN_INCLUDE "${ANAKIN_INSTALL_DIR}" CACHE STRING "root of Anakin header files")
-set(ANAKIN_LIBRARY "${ANAKIN_INSTALL_DIR}" CACHE STRING "path of Anakin library")
+INCLUDE(ExternalProject)
+set(ANAKIN_SOURCE_DIR  ${THIRD_PARTY_PATH}/anakin)
+# the anakin install dir is only default one now
+set(ANAKIN_INSTALL_DIR ${THIRD_PARTY_PATH}/anakin/src/extern_anakin/output)
+set(ANAKIN_INCLUDE     ${ANAKIN_INSTALL_DIR})
+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")
+execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_SOURCE_DIR}")
+execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_MODLE_URL}")
+
+include_directories(${ANAKIN_INCLUDE})
+include_directories(${ANAKIN_INCLUDE}/saber/)
 
 set(ANAKIN_COMPILE_EXTRA_FLAGS 
     -Wno-error=unused-but-set-variable -Wno-unused-but-set-variable
@@ -20,36 +32,33 @@ set(ANAKIN_COMPILE_EXTRA_FLAGS
     -Wno-reorder 
     -Wno-error=cpp)
 
-set(ANAKIN_LIBRARY_URL "https://github.com/pangge/Anakin/releases/download/Version0.1.0/anakin.tar.gz")
-
-# A helper function used in Anakin, currently, to use it, one need to recursively include
-# nearly all the header files.
-function(fetch_include_recursively root_dir)
-    if (IS_DIRECTORY ${root_dir})
-        include_directories(${root_dir})
-    endif()
-
-    file(GLOB ALL_SUB RELATIVE ${root_dir} ${root_dir}/*)
-    foreach(sub ${ALL_SUB})
-        if (IS_DIRECTORY ${root_dir}/${sub})
-            fetch_include_recursively(${root_dir}/${sub})
-        endif()
-    endforeach()
-endfunction()
-
-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 --no-check-certificate -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.tar.gz")
-endif()
+ExternalProject_Add(
+    extern_anakin
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    # TODO(luotao): use PaddlePaddle/Anakin later
+    GIT_REPOSITORY      "https://github.com/luotao1/Anakin"
+    GIT_TAG             "3957ae9263eaa0b1986758dac60a88852afb09be"
+    PREFIX              ${ANAKIN_SOURCE_DIR}
+    UPDATE_COMMAND      ""
+    CMAKE_ARGS          -DUSE_GPU_PLACE=YES
+                        -DUSE_X86_PLACE=YES
+                        -DBUILD_WITH_UNIT_TEST=NO
+                        -DPROTOBUF_ROOT=${THIRD_PARTY_PATH}/install/protobuf
+                        -DMKLML_ROOT=${THIRD_PARTY_PATH}/install/mklml
+                        -DCUDNN_ROOT=${CUDNN_ROOT}
+                        ${EXTERNAL_OPTIONAL_ARGS}
+    CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${ANAKIN_INSTALL_DIR}
+)
 
-if (WITH_ANAKIN)
-    message(STATUS "Anakin for inference is enabled")
-    message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}")
-    fetch_include_recursively(${ANAKIN_INCLUDE})
-    link_directories(${ANAKIN_LIBRARY})
-endif()
+message(STATUS "Anakin for inference is enabled")
+message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}")
+
+add_library(anakin_shared SHARED IMPORTED GLOBAL)
+set_property(TARGET anakin_shared PROPERTY IMPORTED_LOCATION ${ANAKIN_SHARED_LIB})
+add_dependencies(anakin_shared extern_anakin protobuf mklml)
+
+add_library(anakin_saber SHARED IMPORTED GLOBAL)
+set_property(TARGET anakin_saber PROPERTY IMPORTED_LOCATION ${ANAKIN_SABER_LIB})
+add_dependencies(anakin_saber extern_anakin protobuf mklml)
+
+list(APPEND external_project_dependencies anakin_shared anakin_saber)

cmake/inference_lib.cmake

@@ -143,7 +143,7 @@ if (WITH_ANAKIN AND WITH_GPU)
     copy(anakin_inference_lib DEPS paddle_inference_api inference_anakin_api
         SRCS
         ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api
-        ${PADDLE_BINARY_DIR}/third_party/install/anakin/*.tar.gz # anakin release
+        ${ANAKIN_INSTALL_DIR} # anakin release
         DSTS ${dst_dir}/inference/anakin ${dst_dir}/inference/anakin)
      list(APPEND inference_deps anakin_inference_lib)
 endif()

doc/fluid/api/executor.rst

@@ -38,11 +38,3 @@ _switch_scope
 ..  autofunction:: paddle.fluid.executor._switch_scope
     :noindex:
 
-.. _api_fluid_executor_fetch_var:
-
-fetch_var
----------
-
-..  autofunction:: paddle.fluid.executor.fetch_var
-    :noindex:
-

doc/fluid/api/fluid.rst

@@ -106,22 +106,6 @@ _switch_scope
 ..  autofunction:: paddle.fluid._switch_scope
     :noindex:
 
-.. _api_fluid_fetch_var:
-
-fetch_var
----------
-
-..  autofunction:: paddle.fluid.fetch_var
-    :noindex:
-
-.. _api_fluid_Go:
-
-Go
---
-
-..  autoclass:: paddle.fluid.Go
-    :members:
-    :noindex:
 
 .. _api_fluid_make_channel:
 

paddle/fluid/API.spec

@@ -6,7 +6,7 @@ paddle.fluid.Program.create_block ArgSpec(args=['self', 'parent_idx'], varargs=N
 paddle.fluid.Program.current_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Program.get_desc ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Program.global_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.Program.inference_optimize ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Program.inference_optimize ArgSpec(args=['self', 'export_for_deployment'], varargs=None, keywords=None, defaults=(True,))
 paddle.fluid.Program.list_vars ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Program.optimized_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
 paddle.fluid.Program.parse_from_string ArgSpec(args=['binary_str'], varargs=None, keywords=None, defaults=None)
@@ -18,6 +18,9 @@ paddle.fluid.Operator.all_attrs ArgSpec(args=['self'], varargs=None, keywords=No
 paddle.fluid.Operator.attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.attr_type ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.block_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Operator.block_attr_id ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Operator.blocks_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.Operator.blocks_attr_ids ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.has_attr ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.has_kernel ArgSpec(args=['self', 'op_type'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Operator.input ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=None)
@@ -34,21 +37,10 @@ paddle.fluid.default_main_program ArgSpec(args=[], varargs=None, keywords=None,
 paddle.fluid.program_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
 paddle.fluid.get_var ArgSpec(args=['name', 'program'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.Executor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.Executor.as_lodtensor ArgSpec(args=['self', 'data'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Executor.close ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Executor.run ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False))
 paddle.fluid.global_scope ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
 paddle.fluid.scope_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
-paddle.fluid.fetch_var ArgSpec(args=['name', 'scope', 'return_numpy'], varargs=None, keywords=None, defaults=(None, True))
-paddle.fluid.Go.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.Go.construct_go_op ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.make_channel ArgSpec(args=['dtype', 'capacity'], varargs=None, keywords=None, defaults=(0,))
-paddle.fluid.channel_send ArgSpec(args=['channel', 'value', 'is_copy'], varargs=None, keywords=None, defaults=(False,))
-paddle.fluid.channel_recv ArgSpec(args=['channel', 'return_value'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.channel_close ArgSpec(args=['channel'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.Select.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.Select.case ArgSpec(args=['self', 'channel_action_fn', 'channel', 'value', 'is_copy'], varargs=None, keywords=None, defaults=(False,))
-paddle.fluid.Select.default ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Trainer.__init__ ArgSpec(args=['self', 'train_func', 'optimizer_func', 'param_path', 'place', 'parallel', 'checkpoint_config'], varargs=None, keywords=None, defaults=(None, None, False, None))
 paddle.fluid.Trainer.save_params ArgSpec(args=['self', 'param_path'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.Trainer.stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
@@ -62,20 +54,16 @@ paddle.fluid.CheckpointConfig.__init__ ArgSpec(args=['self', 'checkpoint_dir', '
 paddle.fluid.Inferencer.__init__ ArgSpec(args=['self', 'infer_func', 'param_path', 'place', 'parallel'], varargs=None, keywords=None, defaults=(None, False))
 paddle.fluid.Inferencer.infer ArgSpec(args=['self', 'inputs', 'return_numpy'], varargs=None, keywords=None, defaults=(True,))
 paddle.fluid.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.DistributeTranspiler.create_splited_vars ArgSpec(args=['self', 'source_var', 'block', 'tag'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.DistributeTranspiler.get_trainer_program ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True))
 paddle.fluid.InferenceTranspiler.__init__ 
-paddle.fluid.InferenceTranspiler.fuse_batch_norm ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.InferenceTranspiler.fuse_relu_mkldnn ArgSpec(args=['self', 'program'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.InferenceTranspiler.transpile ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.memory_optimize ArgSpec(args=['input_program', 'skip_opt_set', 'print_log', 'level'], varargs=None, keywords=None, defaults=(None, False, 0))
 paddle.fluid.release_memory ArgSpec(args=['input_program', 'skip_opt_set'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.DistributeTranspilerConfig.__init__ 
 paddle.fluid.ParallelExecutor.__init__ ArgSpec(args=['self', 'use_cuda', 'loss_name', 'main_program', 'share_vars_from', 'exec_strategy', 'build_strategy', 'num_trainers', 'trainer_id'], varargs=None, keywords='kwargs', defaults=(None, None, None, None, None, 1, 0))
-paddle.fluid.ParallelExecutor.bcast_params ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.ParallelExecutor.run ArgSpec(args=['self', 'fetch_list', 'feed', 'feed_dict', 'return_numpy'], varargs=None, keywords=None, defaults=(None, None, True))
 paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ExecutionStrategy) -> None
 paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.GradientScaleStrategy, arg0: int) -> None
@@ -89,7 +77,7 @@ paddle.fluid.io.save_persistables ArgSpec(args=['executor', 'dirname', 'main_pro
 paddle.fluid.io.load_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.io.load_params ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.io.load_persistables ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
-paddle.fluid.io.save_inference_model ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename'], varargs=None, keywords=None, defaults=(None, None, None))
+paddle.fluid.io.save_inference_model ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename', 'export_for_deployment'], varargs=None, keywords=None, defaults=(None, None, None, True))
 paddle.fluid.io.load_inference_model ArgSpec(args=['dirname', 'executor', 'model_filename', 'params_filename'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.io.get_inference_program ArgSpec(args=['target_vars', 'main_program'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.initializer.ConstantInitializer.__init__ ArgSpec(args=['self', 'value', 'force_cpu'], varargs=None, keywords=None, defaults=(0.0, False))
@@ -338,14 +326,11 @@ paddle.fluid.contrib.BeamSearchDecoder.read_array ArgSpec(args=['self', 'init',
 paddle.fluid.contrib.BeamSearchDecoder.update_array ArgSpec(args=['self', 'array', 'value'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.contrib.memory_usage ArgSpec(args=['program', 'batch_size'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.transpiler.DistributeTranspiler.create_splited_vars ArgSpec(args=['self', 'source_var', 'block', 'tag'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.get_startup_program ArgSpec(args=['self', 'endpoint', 'pserver_program'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.get_trainer_program ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.transpile ArgSpec(args=['self', 'trainer_id', 'program', 'pservers', 'trainers', 'sync_mode'], varargs=None, keywords=None, defaults=(None, '127.0.0.1:6174', 1, True))
 paddle.fluid.transpiler.InferenceTranspiler.__init__ 
-paddle.fluid.transpiler.InferenceTranspiler.fuse_batch_norm ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.transpiler.InferenceTranspiler.fuse_relu_mkldnn ArgSpec(args=['self', 'program'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.InferenceTranspiler.transpile ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.transpiler.memory_optimize ArgSpec(args=['input_program', 'skip_opt_set', 'print_log', 'level'], varargs=None, keywords=None, defaults=(None, False, 0))
 paddle.fluid.transpiler.release_memory ArgSpec(args=['input_program', 'skip_opt_set'], varargs=None, keywords=None, defaults=(None,))

paddle/fluid/framework/op_desc.cc

@@ -238,7 +238,20 @@ Attribute OpDesc::GetNullableAttr(const std::string &name) const {
   }
 }
 
-int OpDesc::GetBlockAttr(const std::string &name) const {
+std::vector<int> OpDesc::GetBlocksAttrIds(const std::string &name) const {
+  auto it = attrs_.find(name);
+  PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name);
+  auto blocks = boost::get<std::vector<BlockDesc *>>(it->second);
+
+  std::vector<int> ids;
+  for (auto n : blocks) {
+    ids.push_back(n->ID());
+  }
+
+  return ids;
+}
+
+int OpDesc::GetBlockAttrId(const std::string &name) const {
   auto it = attrs_.find(name);
   PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name);
   return boost::get<BlockDesc *>(it->second)->ID();

paddle/fluid/framework/op_desc.h

@@ -83,7 +83,9 @@ class OpDesc {
 
   Attribute GetNullableAttr(const std::string &name) const;
 
-  int GetBlockAttr(const std::string &name) const;
+  int GetBlockAttrId(const std::string &name) const;
+
+  std::vector<int> GetBlocksAttrIds(const std::string &name) const;
 
   void Rename(const std::string &old_name, const std::string &new_name);
 

paddle/fluid/framework/program_desc.cc

@@ -58,7 +58,7 @@ ProgramDesc::ProgramDesc(const ProgramDesc &o) {
       for (const std::string &attr_name : op->AttrNames()) {
         if (op->GetAttrType(attr_name) == proto::AttrType::BLOCK) {
           int sub_block_id =
-              o.Block(block_id).Op(op_id)->GetBlockAttr(attr_name);
+              o.Block(block_id).Op(op_id)->GetBlockAttrId(attr_name);
           op->SetBlockAttr(attr_name, MutableBlock(sub_block_id));
         }
       }

paddle/fluid/inference/analysis/analyzer.cc

@@ -44,13 +44,13 @@ class DfgPassManagerImpl final : public DfgPassManager {
     if (FLAGS_inference_analysis_enable_tensorrt_subgraph_engine) {
       auto trt_teller = [&](const Node* node) {
         std::unordered_set<std::string> teller_set(
-            {"elementwise_add", "mul", "conv2d", "pool2d", "relu"});
+            {"elementwise_add", "mul", "conv2d", "pool2d", "relu", "softmax"});
         if (!node->IsFunction()) return false;
 
         const auto* func = static_cast<const Function*>(node);
-        if (teller_set.count(func->func_type()))
+        if (teller_set.count(func->func_type())) {
           return true;
-        else {
+        } else {
           return false;
         }
       };

paddle/fluid/inference/api/CMakeLists.txt

@@ -45,7 +45,6 @@ endfunction(inference_api_test)
 
 cc_library(paddle_inference_api SRCS api.cc api_impl.cc DEPS lod_tensor)
 
-
 cc_test(test_paddle_inference_api
         SRCS api_tester.cc
         DEPS paddle_inference_api)
@@ -62,22 +61,18 @@ inference_api_test(test_api_tensorrt_subgraph_engine SRC api_tensorrt_subgraph_e
 endif()
 
 if (WITH_ANAKIN) # 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.
-    fetch_include_recursively(${ANAKIN_INCLUDE})
     # compile the libinference_anakin_api.a and anakin.so.
-    nv_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc)
-    nv_library(inference_anakin_api_shared SHARED SRCS api.cc 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 ${ANAKIN_COMPILE_EXTRA_FLAGS})
-    target_link_libraries(inference_anakin_api anakin anakin_saber_common)
-    target_link_libraries(inference_anakin_api_shared anakin anakin_saber_common)
+    nv_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber)
+    #nv_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc DEPS anakin)
+    function(anakin_target target_name)
+      target_compile_options(${target_name} BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
+    endfunction()
+    anakin_target(inference_anakin_api)
+    #anakin_target(inference_anakin_api_shared)
     if (WITH_TESTING)
-        # this test is unstable, disable it first.
-        #cc_test(inference_anakin_test SRCS api_anakin_engine_tester.cc
-                                  #ARGS --model=${ANAKIN_INSTALL_DIR}/mobilenet_v2.anakin.bin
-                                  #DEPS inference_anakin_api_shared)
-        #target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
-     endif(WITH_TESTING)
+        cc_test(inference_anakin_test SRCS api_anakin_engine_tester.cc
+                ARGS --model=${ANAKIN_SOURCE_DIR}/mobilenet_v2.anakin.bin
+                DEPS inference_anakin_api dynload_cuda SERIAL)
+        target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
+    endif(WITH_TESTING)
 endif()

paddle/fluid/inference/api/paddle_inference_api.h

@@ -45,7 +45,7 @@ class PaddleBuf {
   PaddleBuf(void* data, size_t length)
       : data_(data), length_(length), memory_owned_{false} {}
   // Own memory.
-  PaddleBuf(size_t length)
+  explicit PaddleBuf(size_t length)
       : data_(new char[length]), length_(length), memory_owned_(true) {}
   // Resize to `length` bytes.
   void Resize(size_t length);

paddle/fluid/inference/tensorrt/convert/CMakeLists.txt

@@ -1,7 +1,7 @@
 # Add TRT tests
 nv_library(tensorrt_converter
   SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc
-activation_op.cc
+activation_op.cc softmax_op.cc
   DEPS tensorrt_engine operator scope framework_proto op_registry)
 
 nv_test(test_op_converter SRCS test_op_converter.cc DEPS
@@ -21,3 +21,6 @@ nv_test(test_trt_pool2d_op SRCS test_pool2d_op.cc pool2d_op.cc
 
 nv_test(test_trt_elementwise_op SRCS test_elementwise_op.cc elementwise_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine elementwise_add_op SERIAL)
+
+nv_test(test_trt_softmax_op SRCS test_softmax_op.cc softmax_op.cc
+        DEPS ${FLUID_CORE_MODULES} tensorrt_engine softmax_op SERIAL)

paddle/fluid/inference/tensorrt/convert/softmax_op.cc

@@ -0,0 +1,49 @@
+/* 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 "paddle/fluid/inference/tensorrt/convert/op_converter.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+/*
+ * SoftMaxOp, ISoftMaxLayer in TRT. This Layer doesn't has weights.
+ */
+class SoftMaxOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope, bool test_mode) override {
+    VLOG(4)
+        << "convert a fluid softmax op to tensorrt softmax layer without bias";
+    framework::OpDesc op_desc(op, nullptr);
+    // Declare inputs
+    auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]);
+    auto* layer = TRT_ENGINE_ADD_LAYER(engine_, SoftMax,
+                                       *const_cast<nvinfer1::ITensor*>(input1));
+
+    auto output_name = op_desc.Output("Out")[0];
+    engine_->SetITensor(output_name, layer->getOutput(0));
+    if (test_mode) {
+      engine_->DeclareOutput(output_name);
+    }
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+USE_OP(softmax);
+REGISTER_TRT_OP_CONVERTER(softmax, SoftMaxOpConverter);

paddle/fluid/inference/tensorrt/convert/test_softmax_op.cc

@@ -0,0 +1,49 @@
+/* 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 <gtest/gtest.h>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+TEST(SoftMaxOpConverter, main) {
+  framework::Scope scope;
+  std::unordered_set<std::string> parameters;
+  TRTConvertValidation validator(8, parameters, scope, 1000);
+
+  std::vector<int> tensor_shape{8, 10};
+  validator.DeclInputVar("softmax-X", tensor_shape,
+                         nvinfer1::DimsCHW(10, 1, 1));
+  validator.DeclOutputVar("softmax-Out", nvinfer1::DimsCHW(10, 1, 1));
+
+  // Prepare Op description
+  framework::OpDesc desc;
+  desc.SetType("softmax");
+  desc.SetInput("X", {"softmax-X"});
+  desc.SetOutput("Out", {"softmax-Out"});
+
+  LOG(INFO) << "set OP";
+  validator.SetOp(*desc.Proto());
+  LOG(INFO) << "execute";
+
+  validator.Execute(3);
+}
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+USE_OP(softmax);

paddle/fluid/inference/tensorrt/convert/ut_helper.h

@@ -79,6 +79,12 @@ class TRTConvertValidation {
   }
 
   // Declare a Variable as input with random initialization.
+  void DeclInputVar(const std::string& name, const std::vector<int> tensor_dims,
+                    const nvinfer1::Dims& trt_dims) {
+    DeclVar(name, tensor_dims);
+    engine_->DeclareInput(name, nvinfer1::DataType::kFLOAT, trt_dims);
+  }
+
   void DeclInputVar(const std::string& name, const nvinfer1::Dims& dims) {
     DeclVar(name, dims);
     // Declare TRT inputs.
@@ -94,12 +100,18 @@ class TRTConvertValidation {
     DeclVar(name, dims);
   }
 
-  // Declare a variable in a fluid Scope.
-  void DeclVar(const std::string& name, const nvinfer1::Dims& dims,
-               bool is_param = false) {
+  void DeclVar(const std::string& name, const std::vector<int> dim_vec) {
     platform::CPUPlace place;
     platform::CPUDeviceContext ctx(place);
 
+    auto* x = scope_.Var(name);
+    auto* x_tensor = x->GetMutable<framework::LoDTensor>();
+    x_tensor->Resize(framework::make_ddim(dim_vec));
+    RandomizeTensor(x_tensor, place, ctx);
+  }
+  // Declare a variable in a fluid Scope.
+  void DeclVar(const std::string& name, const nvinfer1::Dims& dims,
+               bool is_param = false) {
     // Init Fluid tensor.
     std::vector<int> dim_vec(dims.d, dims.d + dims.nbDims);
     // There is no batchsize in ITensor's shape, but We should add it to
@@ -107,10 +119,8 @@ class TRTConvertValidation {
     // if_add_batch_ flag is true, add the max batchsize to dim_vec.
     if (is_param != true && if_add_batch_ == true)
       dim_vec.insert(dim_vec.begin(), max_batch_size_);
-    auto* x = scope_.Var(name);
-    auto* x_tensor = x->GetMutable<framework::LoDTensor>();
-    x_tensor->Resize(framework::make_ddim(dim_vec));
-    RandomizeTensor(x_tensor, place, ctx);
+
+    DeclVar(name, dim_vec);
   }
 
   void SetOp(const framework::proto::OpDesc& desc) {

paddle/fluid/operators/CMakeLists.txt

@@ -235,7 +235,12 @@ else()
 endif()
 
 op_library(cross_entropy_op DEPS cross_entropy)
-op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
+if(WITH_GPU)
+  op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax cub)
+else()
+  op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
+endif()
+
 op_library(softmax_op DEPS softmax)
 op_library(sequence_softmax_op DEPS softmax)
 if (WITH_GPU AND TENSORRT_FOUND)
@@ -273,9 +278,9 @@ op_library(squeeze_op DEPS reshape_op)
 op_library(extract_rows_op DEPS memory)
 op_library(flatten_op DEPS reshape_op)
 
-
 if (WITH_GPU)
     op_library(conv_op DEPS vol2col depthwise_conv im2col)
+    op_library(layer_norm_op DEPS cub)
 else()
     op_library(conv_op DEPS vol2col im2col)
 endif()

paddle/fluid/operators/crop_op.cc

@@ -1,4 +1,4 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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.
@@ -188,6 +188,7 @@ namespace ops = paddle::operators;
 REGISTER_OPERATOR(crop, ops::CropOp, ops::CropOpMaker,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(crop_grad, ops::CropOpGrad);
-REGISTER_OP_CPU_KERNEL(crop, ops::CropKernel<float>);
+REGISTER_OP_CPU_KERNEL(
+    crop, ops::CropKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     crop_grad, ops::CropGradKernel<paddle::platform::CPUDeviceContext, float>);

paddle/fluid/operators/crop_op.cu

@@ -1,4 +1,4 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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.
@@ -16,6 +16,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/crop_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_CUDA_KERNEL(crop, ops::CropKernel<float>);
+REGISTER_OP_CUDA_KERNEL(
+    crop, ops::CropKernel<paddle::platform::CUDADeviceContext, float>);
 REGISTER_OP_CUDA_KERNEL(
     crop_grad, ops::CropGradKernel<paddle::platform::CUDADeviceContext, float>);

paddle/fluid/operators/crop_op.h

@@ -1,4 +1,4 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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.
@@ -58,32 +58,74 @@ static std::vector<int> GetOffsets(const framework::ExecutionContext& ctx) {
   return res;
 }
 
-template <typename T>
+template <typename DeviceContext, typename T, size_t D>
+void CropFunction(const framework::ExecutionContext& context) {
+  auto* x = context.Input<Tensor>("X");
+  auto* out = context.Output<Tensor>("Out");
+  auto out_dims = out->dims();
+  if (out_dims[0] == -1) {
+    out_dims[0] = x->dims()[0];
+  }
+  out->mutable_data<T>(out_dims, context.GetPlace());
+  auto x_stride = framework::stride(x->dims());
+  auto out_stride = framework::stride(out->dims());
+  auto offsets = GetOffsets(context);
+  int64_t offset = 0;
+  for (size_t i = 0; i < offsets.size(); ++i) {
+    offset += (x_stride[i] * offsets[i]);
+  }
+
+  auto x_tensor = EigenTensor<T, D>::From(*x);
+  auto out_tensor = EigenTensor<T, D>::From(*out);
+  Eigen::array<int, D> e_offsets;
+  Eigen::array<int, D> e_shape;
+  for (size_t i = 0; i < D; ++i) {
+    e_offsets[i] = offsets[i];
+    e_shape[i] = out->dims()[i];
+  }
+  auto& place =
+      *context.template device_context<DeviceContext>().eigen_device();
+  out_tensor.device(place) = x_tensor.slice(e_offsets, e_shape);
+}
+
+template <typename DeviceContext, typename T>
 class CropKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* x = context.Input<Tensor>("X");
-    auto* out = context.Output<Tensor>("Out");
-    const T* x_data = x->data<T>();
-    T* out_data = out->mutable_data<T>(context.GetPlace());
-    auto x_stride = framework::stride(x->dims());
-    auto out_stride = framework::stride(out->dims());
-    auto offsets = GetOffsets(context);
-    int64_t offset = 0;
-    for (size_t i = 0; i < offsets.size(); ++i) {
-      offset += (x_stride[i] * offsets[i]);
+    int rank = context.Input<Tensor>("X")->dims().size();
+    switch (rank) {
+      case 1:
+        CropFunction<DeviceContext, T, 1>(context);
+        break;
+      case 2:
+        CropFunction<DeviceContext, T, 2>(context);
+        break;
+      case 3:
+        CropFunction<DeviceContext, T, 3>(context);
+        break;
+      case 4:
+        CropFunction<DeviceContext, T, 4>(context);
+        break;
+      case 5:
+        CropFunction<DeviceContext, T, 5>(context);
+        break;
+      case 6:
+        CropFunction<DeviceContext, T, 6>(context);
+        break;
+      default:
+        PADDLE_THROW(
+            "CropOp only support tensors with no more than 6 dimensions.");
     }
-    StridedMemcpy<T>(context.device_context(), x_data + offset, x_stride,
-                     out->dims(), out_stride, out_data);
   }
 };
 
 template <typename DeviceContext, typename T, size_t D>
 void CropGradFunction(const framework::ExecutionContext& context) {
   auto* d_x = context.Output<Tensor>(framework::GradVarName("X"));
+  auto* x = context.Input<Tensor>("X");
   if (d_x != nullptr) {
     auto* d_out = context.Input<Tensor>(framework::GradVarName("Out"));
-    d_x->mutable_data<T>(context.GetPlace());
+    d_x->mutable_data<T>(x->dims(), context.GetPlace());
     auto offsets = GetOffsets(context);
     Eigen::array<std::pair<int, int>, D> paddings;
     for (size_t i = 0; i < D; ++i) {

paddle/fluid/operators/elementwise_add_op.cu

@@ -16,60 +16,6 @@ limitations under the License. */
 #include "paddle/fluid/operators/elementwise_add_op.h"
 #include "paddle/fluid/platform/float16.h"
 
-namespace paddle {
-namespace operators {
-
-template <typename T>
-__global__ void ElementwiseAddCUDAKernel(const T *x, const T *y, T *z, int n,
-                                         int post, int size) {
-  int idx_x = threadIdx.x + blockIdx.x * blockDim.x;
-  if (idx_x < size) {
-    int idx_y = idx_x / post - (idx_x / (n * post)) * n;
-    z[idx_x] = x[idx_x] + y[idx_y];
-  }
-}
-
-template <typename T>
-class ElementwiseAddKernel<platform::CUDADeviceContext, T>
-    : public framework::OpKernel<T> {
- public:
-  void Compute(const framework::ExecutionContext &ctx) const override {
-    using Tensor = framework::Tensor;
-
-    const auto x = ctx.Input<Tensor>("X");
-    const auto y = ctx.Input<Tensor>("Y");
-    auto z = ctx.Output<Tensor>("Out");
-    auto *z_data = z->mutable_data<T>(ctx.GetPlace());
-
-    auto &device = *(ctx.cuda_device_context().eigen_device());
-    const framework::DDim &x_dim = x->dims();
-    framework::DDim y_dim = y->dims();
-    int size = x->numel();
-    if (x_dim == y_dim) {
-      auto dim = framework::make_ddim({size});
-      auto z_eigen = framework::EigenTensor<T, 1>::From(*z, dim);
-      auto x_eigen = framework::EigenTensor<T, 1>::From(*x, dim);
-      auto y_eigen = framework::EigenTensor<T, 1>::From(*y, dim);
-      z_eigen.device(device) = x_eigen + y_eigen;
-    } else {
-      int axis = ctx.Attr<int>("axis");
-      axis = (axis == -1 ? x_dim.size() - y_dim.size() : axis);
-      y_dim = trim_trailing_singular_dims(y_dim);
-      axis = (y_dim.size() == 0) ? x_dim.size() : axis;
-      int pre, n, post;
-      get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post);
-      int threads = 512;
-      int grids = (size + threads - 1) / threads;
-      auto stream = ctx.cuda_device_context().stream();
-      ElementwiseAddCUDAKernel<T><<<grids, threads, 0, stream>>>(
-          x->data<T>(), y->data<T>(), z_data, n, post, size);
-    }
-  }
-};
-
-}  // namespace operators
-}  // namespace paddle
-
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 

paddle/fluid/operators/elementwise_add_op.h

@@ -144,41 +144,16 @@ class ElementwiseAddGradKernel : public framework::OpKernel<T> {
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
     auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    // skip out, x, y
+    auto* out = dout;
+    auto *x = dout, *y = dout;
 
-    if (dx != nullptr) {
-      // In fact, we can just share memory, but it may cause a bug of memory
-      // optimizer
-      // dx->ShareDataWith(*dout);
-      framework::TensorCopy(*dout, ctx.GetPlace(),
-                            ctx.template device_context<DeviceContext>(), dx);
-    }
-
-    if (dy == nullptr) return;
-
-    const framework::DDim& x_dim = dout->dims();
-    framework::DDim y_dim = dy->dims();
-    if (x_dim == y_dim) {
-      // dy->ShareDataWith(*dout);
-      framework::TensorCopy(*dout, ctx.GetPlace(),
-                            ctx.template device_context<DeviceContext>(), dy);
+    if (platform::is_cpu_place(ctx.GetPlace()) && dx != nullptr &&
+        dy != nullptr && (dx->dims() == dy->dims())) {
+      elementwise_add_grad<DeviceContext, T>(ctx, x, y, out, dout, dx, dy);
     } else {
-      dy->mutable_data<T>(ctx.GetPlace());
-      // Perform reduction to dout to calculate dy
-      int axis = ctx.Attr<int>("axis");
-      axis = (axis == -1 ? x_dim.size() - y_dim.size() : axis);
-      y_dim = trim_trailing_singular_dims(y_dim);
-      axis = (y_dim.size() == 0) ? x_dim.size() : axis;
-
-      auto& device =
-          *(ctx.template device_context<DeviceContext>().eigen_device());
-      int pre, n, post;
-      get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post);
-      auto eigen_dout = framework::EigenTensor<T, 3>::From(
-          *dout, framework::make_ddim({pre, n, post}));
-      auto eigen_dy =
-          framework::EigenTensor<T, 1>::From(*dy, framework::make_ddim({n}));
-      eigen_dy.device(device) = eigen_dout.sum(
-          framework::EigenDim<2>::From(framework::make_ddim({0, 2})));
+      default_elementwise_add_grad<DeviceContext, T>(ctx, x, y, out, dout, dx,
+                                                     dy);
     }
   }
 };

paddle/fluid/pybind/protobuf.cc

@@ -301,7 +301,8 @@ void BindOpDesc(pybind11::module *m) {
              std::string ser(seriralized);
              self.SetAttr(name, ser);
            })
-      .def("block_attr", &pd::OpDesc::GetBlockAttr)
+      .def("block_attr_id", &pd::OpDesc::GetBlockAttrId)
+      .def("blocks_attr_ids", &pd::OpDesc::GetBlocksAttrIds)
       .def("check_attrs", &pd::OpDesc::CheckAttrs)
       .def("infer_shape", &pd::OpDesc::InferShape)
       .def("infer_var_type", &pd::OpDesc::InferVarType)