Merge branch 'develop' into core_fix_openblas_threads

Dockerfile

@@ -70,7 +70,7 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8
 # specify sphinx version as 1.5.6 and remove -U option for [pip install -U
 # sphinx-rtd-theme] since -U option will cause sphinx being updated to newest
 # version(1.7.1 for now), which causes building documentation failed.
-RUN pip install --upgrade pip==9.0.3 && \
+RUN easy_install -U pip && \
     pip install -U wheel && \
     pip install -U docopt PyYAML sphinx==1.5.6 && \
     pip install sphinx-rtd-theme==0.1.9 recommonmark

benchmark/fluid/mnist.py

@@ -159,6 +159,7 @@ def run_benchmark(model, args):
         paddle.dataset.mnist.train(), batch_size=args.batch_size)
 
     accuracy = fluid.metrics.Accuracy()
+    train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
     iters, num_samples, start_time = 0, 0, time.time()
     for pass_id in range(args.pass_num):
         accuracy.reset()
@@ -175,17 +176,20 @@ def run_benchmark(model, args):
             y_data = np.array(map(lambda x: x[1], data)).astype("int64")
             y_data = y_data.reshape([len(y_data), 1])
 
-            outs = exe.run(
+            outs = train_exe.run(
-                fluid.default_main_program(),
                 feed={"pixel": img_data,
                       "label": y_data},
-                fetch_list=[avg_cost, batch_acc, batch_size_tensor]
+                fetch_list=[
+                    avg_cost.name, batch_acc.name, batch_size_tensor.name
+                ]
             )  # The accuracy is the accumulation of batches, but not the current batch.
-            accuracy.update(value=outs[1], weight=outs[2])
+            accuracy.update(
+                value=np.array(np.mean(outs[1])),
+                weight=np.mean(np.array(outs[2])))
             iters += 1
             num_samples += len(y_data)
-            loss = np.array(outs[0])
+            loss = np.mean(np.array(outs[0]))
-            acc = np.array(outs[1])
+            acc = np.mean(np.array(outs[1]))
             train_losses.append(loss)
             train_accs.append(acc)
             print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %

benchmark/fluid/resnet.py

@@ -241,6 +241,7 @@ def run_benchmark(model, args):
     exe = fluid.Executor(place)
     exe.run(fluid.default_startup_program())
     accuracy = fluid.average.WeightedAverage()
+    train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
     if args.use_fake_data:
         data = train_reader().next()
         image = np.array(map(lambda x: x[0].reshape(dshape), data)).astype(
@@ -264,14 +265,17 @@ def run_benchmark(model, args):
                                      data)).astype('float32')
                 label = np.array(map(lambda x: x[1], data)).astype('int64')
                 label = label.reshape([-1, 1])
-            loss, acc, weight = exe.run(
+            loss, acc, weight = train_exe.run(
-                fluid.default_main_program(),
                 feed={'data': image,
                       'label': label},
-                fetch_list=[avg_cost, batch_acc, batch_size_tensor])
+                fetch_list=[
+                    avg_cost.name, batch_acc.name, batch_size_tensor.name
+                ])
             iters += 1
             num_samples += len(label)
-            accuracy.add(value=acc, weight=weight)
+            accuracy.add(value=np.array(np.mean(acc)), weight=np.mean(weight))
+            loss = np.mean(np.array(loss))
+            acc = np.mean(np.array(acc))
             train_losses.append(loss)
             train_accs.append(acc)
             print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %

benchmark/fluid/vgg.py

@@ -169,6 +169,7 @@ def main():
 
     iters, num_samples, start_time = 0, 0, time.time()
     accuracy = fluid.average.WeightedAverage()
+    train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
     for pass_id in range(args.pass_num):
         accuracy.reset()
         train_accs = []
@@ -184,14 +185,17 @@ def main():
             y_data = np.array(map(lambda x: x[1], data)).astype("int64")
             y_data = y_data.reshape([-1, 1])
 
-            loss, acc, weight = exe.run(
+            loss, acc, weight = train_exe.run(
-                fluid.default_main_program(),
                 feed={"pixel": img_data,
                       "label": y_data},
-                fetch_list=[avg_cost, batch_acc, batch_size_tensor])
+                fetch_list=[
-            accuracy.add(value=acc, weight=weight)
+                    avg_cost.name, batch_acc.name, batch_size_tensor.name
+                ])
+            accuracy.add(value=np.array(np.mean(acc)), weight=np.mean(weight))
             iters += 1
             num_samples += len(y_data)
+            loss = np.mean(np.array(loss))
+            acc = np.mean(np.array(acc))
             print(
                 "Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" %
                 (pass_id, iters, loss, acc)

cmake/external/boost.cmake

@@ -24,7 +24,7 @@ set(BOOST_PROJECT       "extern_boost")
 # So we use 1.41.0 here.
 set(BOOST_VER           "1.41.0")
 set(BOOST_TAR           "boost_1_41_0")
-set(BOOST_URL           "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz")
+set(BOOST_URL           "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz")
 set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
 set(BOOST_DOWNLOAD_DIR  "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
 set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)

cmake/external/eigen.cmake

@@ -21,11 +21,12 @@ else()
     ExternalProject_Add(
         extern_eigen3
         ${EXTERNAL_PROJECT_LOG_ARGS}
-        GIT_REPOSITORY  "https://github.com/RLovelett/eigen.git"
+        GIT_REPOSITORY  "https://github.com/eigenteam/eigen-git-mirror"
         # eigen on cuda9.1 missing header of math_funtions.hpp
         # https://stackoverflow.com/questions/43113508/math-functions-hpp-not-found-when-using-cuda-with-eigen
         GIT_TAG         917060c364181f33a735dc023818d5a54f60e54c
         PREFIX          ${EIGEN_SOURCE_DIR}
+        DOWNLOAD_NAME   "eigen"
         UPDATE_COMMAND  ""
         CONFIGURE_COMMAND ""
         BUILD_COMMAND     ""

cmake/external/mkldnn.cmake

@@ -53,11 +53,9 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "v0.14"
+    GIT_TAG             "db3424ad44901513c03a1ea31ccaacdf633fbe9f"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
-    # Patch MKLDNN to compile with gcc 4.8, the related issue is in intel/mkl-dnn#237.
-    PATCH_COMMAND       ${CMAKE_COMMAND} -E copy_if_different ${CMAKE_CURRENT_SOURCE_DIR}/patches/mkldnn.hpp ${MKLDNN_SOURCES_DIR}/src/extern_mkldnn/include/mkldnn.hpp
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
     CMAKE_ARGS          -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} 
     CMAKE_ARGS          -DMKLROOT=${MKLML_ROOT}

cmake/external/mklml.cmake

@@ -28,7 +28,7 @@ INCLUDE(ExternalProject)
 
 SET(MKLML_PROJECT       "extern_mklml")
 SET(MKLML_VER           "mklml_lnx_2018.0.3.20180406")
-SET(MKLML_URL           "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz")
+SET(MKLML_URL           "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz")
 SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
 SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")

cmake/external/snappy.cmake

@@ -47,8 +47,6 @@ ExternalProject_Add(
                      -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPY_INSTALL_DIR}/lib
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
-    BUILD_COMMAND   make -j8
-    INSTALL_COMMAND make install
 )
 
 add_library(snappy STATIC IMPORTED GLOBAL)

cmake/external/snappystream.cmake

@@ -46,8 +46,6 @@ ExternalProject_Add(
                         -DCMAKE_INSTALL_PREFIX:PATH=${SNAPPYSTREAM_INSTALL_DIR}
                         -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPYSTREAM_INSTALL_DIR}/lib
                         -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
-        BUILD_COMMAND   make -j8
-        INSTALL_COMMAND make install
         DEPENDS snappy
 )
 

cmake/inference_lib.cmake

@@ -98,6 +98,14 @@ elseif (WITH_MKLML)
     )
 endif()
 
+if(WITH_MKLDNN)
+  set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/mkldnn")
+  copy(mkldnn_lib
+    SRCS ${MKLDNN_INC_DIR} ${MKLDNN_SHARED_LIB}
+    DSTS ${dst_dir} ${dst_dir}/lib
+  )
+endif()
+
 if(NOT MOBILE_INFERENCE AND NOT RPI)
   set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/snappy")
   copy(snappy_lib
@@ -148,4 +156,10 @@ copy(string_lib
   DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
 )
 
+set(module "pybind")
+copy(pybind_lib
+  SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h
+  DSTS ${dst_dir}/${module}
+)
+
 add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) 

doc/fluid/design/concepts/functions_operators_layers.md

@@ -40,7 +40,7 @@ template <typename T>
 class FCOp : public OperatorBase {
  public:
   void Run(...) {
-    add(mul(Input<T>("X"), Input<T>("W")), Input<T>("b");
+    add(mul(Input<T>("X"), Input<T>("W")), Input<T>("b"));
   }
 };
 REGISTER_OP(FCOp, "fc");

paddle/CMakeLists.txt

@@ -24,6 +24,6 @@ if(NOT WITH_FLUID_ONLY)
 endif()
 
 add_subdirectory(testing)
-if(NOT MOBILE_INFERENCE AND NOT RPI)
+if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API)
   add_subdirectory(fluid)
 endif()

paddle/fluid/framework/data_device_transform.cc

@@ -36,9 +36,11 @@ void TransDataDevice(const Tensor& in, const platform::Place& dst_place,
   VLOG(3) << "DeviceTransform in, src_place " << in.place()
           << " dst_place: " << dst_place;
   auto* dev_ctx = GetDeviceContext(in.place(), dst_place);
-  dev_ctx->Wait();
+
   TensorCopy(in, dst_place, *dev_ctx, out);
+  if (platform::is_gpu_place(in.place()) && platform::is_cpu_place(dst_place)) {
     dev_ctx->Wait();
+  }
 }
 
 }  // namespace framework

paddle/fluid/framework/data_type.cc

@@ -58,6 +58,7 @@ static DataTypeMap* InitDataTypeMap() {
   RegType(bool, proto::VarType::BOOL);
   RegType(size_t, proto::VarType::SIZE_T);
   RegType(int16_t, proto::VarType::INT16);
+  RegType(uint8_t, proto::VarType::UINT8);
 
 #undef RegType
   return retv;

paddle/fluid/framework/data_type.h

@@ -47,8 +47,14 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
     case proto::VarType::BOOL:
       visitor.template operator()<bool>();
       break;
+    case proto::VarType::UINT8:
+      visitor.template operator()<uint8_t>();
+      break;
+    case proto::VarType::INT16:
+      visitor.template operator()<int16_t>();
+      break;
     default:
-      PADDLE_THROW("Not supported");
+      PADDLE_THROW("Not supported %d", type);
   }
 }
 

paddle/fluid/framework/details/fetch_op_handle.cc

@@ -48,17 +48,18 @@ void FetchOpHandle::RunImpl() {
   WaitInputVarGenerated(platform::CPUPlace());
 
   tensors_.resize(inputs_.size());
-  auto *var_handle = static_cast<VarHandle *>(inputs_[0]);
-  auto &var_name = var_handle->name_;
   platform::CPUPlace cpu;
   auto &scopes = *local_scopes_;
 
-  for (size_t i = 0; i < scopes.size(); ++i) {
+  for (size_t i = 0; i < inputs_.size(); ++i) {
-    auto &scope = scopes[i];
+    auto *var_handle = static_cast<VarHandle *>(inputs_[i]);
-    auto *var =
+    auto &scope = scopes.at(var_handle->scope_idx_);
-        scope->FindVar(kLocalExecScopeName)->Get<Scope *>()->FindVar(var_name);
+    auto *var = scope->FindVar(kLocalExecScopeName)
+                    ->Get<Scope *>()
+                    ->FindVar(var_handle->name_);
     PADDLE_ENFORCE_NOT_NULL(var, "Cannot find variable %s in execution scope",
-                            var_name);
+                            var_handle->name_);
+
     auto &t = var->Get<framework::LoDTensor>();
     if (platform::is_gpu_place(t.place())) {
 #ifdef PADDLE_WITH_CUDA

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -98,7 +98,7 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp(const OpDesc &op,
     return false;
   };
 
-  if (op.Type() == "split") {
+  if (op.Type() == "split" || op.Type() == "split_byref") {
     return checker(op.OutputArgumentNames(), send_op->InputArgumentNames());
   } else if (op.Type() == "concat") {
     return checker(op.InputArgumentNames(), send_op->OutputArgumentNames());

paddle/fluid/framework/details/op_handle_base.h

@@ -70,6 +70,14 @@ class OpHandleBase {
 
   const std::vector<VarHandleBase *> &Inputs() const { return inputs_; }
 
+  size_t NoDupInputSize() const {
+    std::unordered_set<VarHandleBase *> res;
+    for (auto *var : inputs_) {
+      res.emplace(var);
+    }
+    return res.size();
+  }
+
   const std::vector<VarHandleBase *> &Outputs() const { return outputs_; }
 
  protected:

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -174,7 +174,7 @@ void ThreadedSSAGraphExecutor::InsertFetchOps(
 void ThreadedSSAGraphExecutor::InsertPendingOp(
     std::unordered_map<OpHandleBase *, size_t> *pending_ops,
     OpHandleBase *op_instance) const {
-  pending_ops->insert({op_instance, op_instance->Inputs().size()});
+  pending_ops->insert({op_instance, op_instance->NoDupInputSize()});
 }
 
 void ThreadedSSAGraphExecutor::InsertPendingVar(

paddle/fluid/framework/executor.cc

@@ -228,7 +228,8 @@ static bool has_fetch_operators(
 void Executor::Run(const ProgramDesc& program, Scope* scope,
                    std::map<std::string, const LoDTensor*>* feed_targets,
                    std::map<std::string, LoDTensor*>* fetch_targets,
-                   bool create_vars, const std::string& feed_holder_name,
+                   bool create_local_scope, bool create_vars,
+                   const std::string& feed_holder_name,
                    const std::string& fetch_holder_name) {
   platform::RecordBlock b(kProgramId);
   bool has_feed_ops =
@@ -290,8 +291,9 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
   }
 
   auto ctx = Prepare(*copy_program, 0);
-  RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets, create_vars,
+  RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets,
-                     feed_holder_name, fetch_holder_name);
+                     create_local_scope, create_vars, feed_holder_name,
+                     fetch_holder_name);
 }
 
 std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
@@ -366,8 +368,9 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
 void Executor::RunPreparedContext(
     ExecutorPrepareContext* ctx, Scope* scope,
     std::map<std::string, const LoDTensor*>* feed_targets,
-    std::map<std::string, LoDTensor*>* fetch_targets, bool create_vars,
+    std::map<std::string, LoDTensor*>* fetch_targets, bool create_local_scope,
-    const std::string& feed_holder_name, const std::string& fetch_holder_name) {
+    bool create_vars, const std::string& feed_holder_name,
+    const std::string& fetch_holder_name) {
   auto& global_block = ctx->prog_.Block(ctx->block_id_);
 
   PADDLE_ENFORCE(
@@ -387,7 +390,7 @@ void Executor::RunPreparedContext(
     }
   }
 
-  RunPreparedContext(ctx, scope, create_vars, create_vars);
+  RunPreparedContext(ctx, scope, create_local_scope, create_vars);
 
   // obtain the data of fetch_targets from fetch_holder
   for (auto* op : global_block.AllOps()) {

paddle/fluid/framework/executor.h

@@ -57,7 +57,7 @@ class Executor {
   void Run(const ProgramDesc& program, Scope* scope,
            std::map<std::string, const LoDTensor*>* feed_targets,
            std::map<std::string, LoDTensor*>* fetch_targets,
-           bool create_vars = true,
+           bool create_local_scope = true, bool create_vars = true,
            const std::string& feed_holder_name = "feed",
            const std::string& fetch_holder_name = "fetch");
 
@@ -76,6 +76,7 @@ class Executor {
   void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
                           std::map<std::string, const LoDTensor*>* feed_targets,
                           std::map<std::string, LoDTensor*>* fetch_targets,
+                          bool create_local_scope = true,
                           bool create_vars = true,
                           const std::string& feed_holder_name = "feed",
                           const std::string& fetch_holder_name = "fetch");

paddle/fluid/framework/framework.proto

@@ -103,6 +103,7 @@ message VarType {
     FP64 = 6;
     // Tensor<size_t> is used in C++.
     SIZE_T = 19;
+    UINT8 = 20;
 
     // Other types that may need additional descriptions
     LOD_TENSOR = 7;

paddle/fluid/framework/lod_tensor_test.cc

@@ -228,11 +228,12 @@ TEST(LoD, CheckAbsLoD) {
   ASSERT_FALSE(CheckAbsLoD(abs_lod0));
 }
 
-TEST(LoDTensor, RecordIO) {
+template <typename T>
+static void TestRecordIO() {
   LoDTensor tensor;
-  int* tmp = tensor.mutable_data<int>(make_ddim({4, 5}), platform::CPUPlace());
+  T* tmp = tensor.mutable_data<T>(make_ddim({4, 5}), platform::CPUPlace());
   for (int i = 0; i < 20; ++i) {
-    tmp[i] = i;
+    tmp[i] = static_cast<T>(i);
   }
 
   std::stringstream* stream = new std::stringstream();
@@ -247,7 +248,7 @@ TEST(LoDTensor, RecordIO) {
 
   auto assert_tensor_ok = [](const LoDTensor& tensor) {
     for (int i = 0; i < 20; ++i) {
-      ASSERT_EQ(tensor.data<int>()[i], i);
+      ASSERT_EQ(tensor.data<T>()[i], static_cast<T>(i));
     }
   };
 
@@ -265,5 +266,13 @@ TEST(LoDTensor, RecordIO) {
   }
 }
 
+TEST(LoDTensor, RecordIO) {
+  TestRecordIO<int>();
+  TestRecordIO<int16_t>();
+  TestRecordIO<uint8_t>();
+  TestRecordIO<float>();
+  TestRecordIO<double>();
+}
+
 }  // namespace framework
 }  // namespace paddle

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

@@ -49,7 +49,7 @@ class OpConverter {
   // convert fluid block to tensorrt network
   void ConvertBlock(const framework::proto::BlockDesc& block,
                     TensorRTEngine* engine) {
-    for (size_t i = 0; i < block.ops_size(); i++) {
+    for (int i = 0; i < block.ops_size(); i++) {
       const auto& op = block.ops(i);
       OpConverter::Run(op, engine);
     }