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

… optimizer_base

cmake/cinn.cmake

@@ -60,7 +60,7 @@ if(WITH_MKL)
   add_definitions(-DCINN_WITH_MKL_CBLAS)
 endif()
 if(WITH_MKLDNN)
-  add_definitions(-DCINN_WITH_MKLDNN)
+  add_definitions(-DCINN_WITH_DNNL)
 endif()
 
 if(WITH_GPU)
@@ -326,12 +326,8 @@ set(CINN_LIB "${CINN_LIB_LOCATION}/${CINN_LIB_NAME}")
 # Add CINN's dependencies header files
 ######################################
 
-# Add absl
-set(ABSL_INCLUDE_DIR "${CMAKE_BINARY_DIR}/dist/third_party/absl/include")
-include_directories(${ABSL_INCLUDE_DIR})
-
 # Add isl
-set(ISL_INCLUDE_DIR "${CMAKE_BINARY_DIR}/dist/third_party/isl/include")
+set(ISL_INCLUDE_DIR "${CMAKE_BINARY_DIR}/third_party/install/isl/include")
 include_directories(${ISL_INCLUDE_DIR})
 
 # Add LLVM

cmake/cinn/external/absl.cmake

@@ -1,13 +1,31 @@
 include(ExternalProject)
 
-set(ABSL_SOURCES_DIR ${THIRD_PARTY_PATH}/absl)
+set(ABSL_SOURCES_DIR ${PADDLE_SOURCE_DIR}/third_party/absl)
 set(ABSL_INSTALL_DIR ${THIRD_PARTY_PATH}/install/absl)
-
+set(ABSL_PREFIX_DIR ${THIRD_PARTY_PATH}/absl)
 set(ABSL_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS})
 
 set(ABSL_REPOSITORY "https://github.com/abseil/abseil-cpp.git")
 set(ABSL_TAG "20210324.2")
 
+if(NOT EXISTS ${ABSL_SOURCES_DIR})
+  message(
+    STATUS "Download absl source from ${ABSL_REPOSITORY} to ABSL_SOURCES_DIR")
+  execute_process(COMMAND ${GIT_EXECUTABLE} clone -b ${ABSL_TAG}
+                          ${ABSL_REPOSITORY} ${ABSL_SOURCES_DIR})
+else()
+  # check git tag
+  execute_process(
+    COMMAND ${GIT_EXECUTABLE} -C ${ABSL_SOURCES_DIR} describe --tags
+    OUTPUT_VARIABLE CURRENT_TAG
+    OUTPUT_STRIP_TRAILING_WHITESPACE)
+  if(NOT ${CURRENT_TAG} STREQUAL ${ABSL_TAG})
+    message(STATUS "Checkout absl to ${ABSL_TAG}")
+    execute_process(COMMAND ${GIT_EXECUTABLE} -C ${ABSL_SOURCES_DIR} checkout
+                            -q ${ABSL_TAG})
+  endif()
+endif()
+
 set(OPTIONAL_ARGS
     "-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}"
     "-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}"
@@ -22,9 +40,8 @@ ExternalProject_Add(
   external_absl
   ${EXTERNAL_PROJECT_LOG_ARGS}
   DEPENDS gflags
-  GIT_REPOSITORY ${ABSL_REPOSITORY}
-  GIT_TAG ${ABSL_TAG}
-  PREFIX ${ABSL_SOURCES_DIR}
+  PREFIX ${ABSL_PREFIX_DIR}
+  SOURCE_DIR ${ABSL_SOURCES_DIR}
   UPDATE_COMMAND ""
   CMAKE_ARGS ${OPTIONAL_ARGS}
              -DCMAKE_INSTALL_PREFIX=${ABSL_INSTALL_DIR}

cmake/cinn/external/isl.cmake

@@ -7,25 +7,62 @@ include(ExternalProject)
 # static build
 # CPPFLAGS="-fPIC -DPIC" ./configure --with-gmp-prefix=<gmp-install-path> --with-clang-prefix=<llvm-install-path> --enable-shared=no --enable-static=yes
 
+set(ISL_FILE
+    "isl-6a1760fe.tar.gz"
+    CACHE STRING "" FORCE)
 set(ISL_DOWNLOAD_URL
-    https://paddle-inference-dist.bj.bcebos.com/CINN/isl-6a1760fe.tar.gz)
-set(ISL_MD5 fff10083fb79d394b8a7b7b2089f6183)
+    "https://paddle-inference-dist.bj.bcebos.com/CINN/${ISL_FILE}")
+set(ISL_URL_MD5 fff10083fb79d394b8a7b7b2089f6183)
+set(ISL_DOWNLOAD_DIR ${PADDLE_SOURCE_DIR}/third_party/isl)
+set(ISL_PREFIX_DIR ${THIRD_PARTY_PATH}/isl)
+set(ISL_INSTALL_DIR ${THIRD_PARTY_PATH}/install/isl)
+
+function(download_isl)
+  message(
+    STATUS "Downloading ${ISL_DOWNLOAD_URL} to ${ISL_DOWNLOAD_DIR}/${ISL_FILE}")
+  file(
+    DOWNLOAD ${ISL_DOWNLOAD_URL} ${ISL_DOWNLOAD_DIR}/${ISL_FILE}
+    EXPECTED_MD5 ${ISL_URL_MD5}
+    STATUS ERR)
+  if(ERR EQUAL 0)
+    message(STATUS "Download ${ISL_FILE} success")
+  else()
+    message(
+      FATAL_ERROR
+        "Download failed, error: ${ERR}\n You can try downloading ${ISL_FILE} again"
+    )
+  endif()
+endfunction()
+
+# Download and check isl.
+if(EXISTS ${ISL_DOWNLOAD_DIR}/${ISL_FILE})
+  file(MD5 ${ISL_DOWNLOAD_DIR}/${ISL_FILE} ISL_MD5)
+  if(NOT ISL_MD5 STREQUAL ISL_URL_MD5)
+    # clean build file
+    file(REMOVE_RECURSE ${ISL_PREFIX_DIR})
+    file(REMOVE_RECURSE ${ISL_INSTALL_DIR})
+    download_isl()
+  endif()
+else()
+  download_isl()
+endif()
 
 ExternalProject_Add(
   external_isl
   ${EXTERNAL_PROJECT_LOG_ARGS}
-  URL ${ISL_DOWNLOAD_URL}
-  URL_MD5 ${ISL_MD5}
-  PREFIX ${THIRD_PARTY_PATH}/isl
-  SOURCE_DIR ${THIRD_PARTY_PATH}/install/isl
+  URL ${ISL_DOWNLOAD_DIR}/${ISL_FILE}
+  URL_MD5 ${ISL_URL_MD5}
+  DOWNLOAD_DIR ${ISL_DOWNLOAD_DIR}
+  PREFIX ${ISL_PREFIX_DIR}
+  SOURCE_DIR ${ISL_INSTALL_DIR}
   CONFIGURE_COMMAND ""
   BUILD_COMMAND ""
   UPDATE_COMMAND ""
   INSTALL_COMMAND ""
-  BUILD_BYPRODUCTS ${THIRD_PARTY_PATH}/install/isl/lib/libisl.a)
+  BUILD_BYPRODUCTS ${ISL_INSTALL_DIR}/lib/libisl.a)
 
 add_library(isl STATIC IMPORTED GLOBAL)
 set_property(TARGET isl PROPERTY IMPORTED_LOCATION
-                                 ${THIRD_PARTY_PATH}/install/isl/lib/libisl.a)
+                                 ${ISL_INSTALL_DIR}/lib/libisl.a)
 add_dependencies(isl external_isl)
-include_directories(${THIRD_PARTY_PATH}/install/isl/include)
+include_directories(${ISL_INSTALL_DIR}/include)

cmake/external/mkldnn.cmake

@@ -91,7 +91,7 @@ ExternalProject_Add(
   BUILD_BYPRODUCTS ${BUILD_BYPRODUCTS_ARGS})
 
 message(STATUS "MKLDNN library: ${MKLDNN_LIB}")
-add_definitions(-DPADDLE_WITH_MKLDNN)
+add_definitions(-DPADDLE_WITH_DNNL)
 # copy the real so.0 lib to install dir
 # it can be directly contained in wheel or capi
 if(WIN32)

cmake/external/xpu.cmake

@@ -24,7 +24,7 @@ set(XPU_XFT_LIB_NAME "libxft.so")
 set(XPU_XPTI_LIB_NAME "libxpti.so")
 
 if(NOT DEFINED XPU_BASE_DATE)
-  set(XPU_BASE_DATE "20230807")
+  set(XPU_BASE_DATE "20230810")
 endif()
 set(XPU_XCCL_BASE_VERSION "1.0.53.6")
 if(NOT DEFINED XPU_XFT_BASE_VERSION)

cmake/third_party.cmake

@@ -269,7 +269,7 @@ if(WITH_CINN)
     add_definitions(-DCINN_WITH_MKL_CBLAS)
   endif()
   if(WITH_MKLDNN)
-    add_definitions(-DCINN_WITH_MKLDNN)
+    add_definitions(-DCINN_WITH_DNNL)
   endif()
   include(cmake/cinn/version.cmake)
   if(NOT EXISTS ${CMAKE_BINARY_DIR}/cmake/cinn/config.cmake)

paddle/cinn/hlir/framework/instruction.cc

@@ -365,7 +365,7 @@ void Instruction::Run(
   //   }
 }
 
-std::string Instruction::DumpInstruction() {
+std::string Instruction::DumpInstruction() const {
   std::stringstream ss;
   ss << "Instruction {" << std::endl;
   for (size_t i = 0; i < fn_names_.size(); ++i) {

paddle/cinn/hlir/framework/instruction.h

@@ -132,13 +132,17 @@ class Instruction {
 
   int size() { return fn_ptrs_.size(); }
 
-  std::string DumpInstruction();
+  std::string DumpInstruction() const;
 
-  std::vector<std::vector<std::string>> GetInArgs() { return in_args_; }
-  std::vector<std::vector<std::string>> GetOutArgs() { return out_args_; }
+  const std::vector<std::vector<std::string>>& GetInArgs() const {
+    return in_args_;
+  }
+  const std::vector<std::vector<std::string>>& GetOutArgs() const {
+    return out_args_;
+  }
   void ClearInArgs() { in_args_.clear(); }
   void ClearOutArgs() { out_args_.clear(); }
-  std::vector<std::string> GetFnNames() { return fn_names_; }
+  const std::vector<std::string>& GetFnNames() const { return fn_names_; }
   void AddInArgs(const std::vector<std::string>& in_args) {
     in_args_.push_back(in_args);
   }

paddle/cinn/hlir/framework/program.cc

@@ -140,9 +140,9 @@ void Program::Export(const std::vector<std::string>& persistent_vars,
   int instplaceholder = writeplaceholder(4 * 3, insnum, f);
   int findex = 0;
   for (auto& ins : instrs_) {
-    auto in_args = ins->GetInArgs();
-    auto out_args = ins->GetOutArgs();
-    auto fn_names = ins->GetFnNames();
+    auto& in_args = ins->GetInArgs();
+    auto& out_args = ins->GetOutArgs();
+    auto& fn_names = ins->GetFnNames();
     for (int i = 0; i < fn_names.size(); i++, findex++) {
       std::vector<std::string> all_args(in_args[i].begin(), in_args[i].end());
       all_args.insert(

paddle/cinn/hlir/op/custom_call.cc

@@ -996,7 +996,7 @@ bool RegisteryCustomCallArgsFunc() {
       CustomCallArgsForCudnnPoolBackward);
 #endif
 
-#ifdef CINN_WITH_MKLDNN
+#ifdef CINN_WITH_DNNL
 
 #endif
 

paddle/cinn/hlir/op/nn.cc

@@ -208,7 +208,7 @@ std::shared_ptr<OpStrategy> StrategyForConv2d(
                                        tensor_name,
                                        target);
             } else {
-#ifdef CINN_WITH_MKLDNN
+#ifdef CINN_WITH_DNNL
               out = pe::Conv2d_NCHW_MKLDNN(A.as_tensor_ref(),
                                            B.as_tensor_ref(),
                                            padding[0],
@@ -1897,7 +1897,7 @@ std::shared_ptr<OpStrategy> StrategyForSoftmax(
         std::string tensor_name =
             pack_args[pack_args.size() - 1].operator std::string();
 
-#ifdef CINN_WITH_MKLDNN
+#ifdef CINN_WITH_DNNL
         if (use_mkldnn) {
           out = pe::SoftmaxMKLDNN(A, new_axis, tensor_name);
         } else {

paddle/cinn/hlir/pe/nn.cc

@@ -646,7 +646,7 @@ std::vector<ir::Tensor> Conv2d_NCHWc(const ir::Tensor &input,
   return {packed_out, input_pad};
 }
 
-#ifdef CINN_WITH_MKLDNN
+#ifdef CINN_WITH_DNNL
 std::vector<ir::Tensor> Conv2d_NCHW_MKLDNN(const ir::Tensor &input,
                                            const ir::Tensor &weights,
                                            int pad_h,
@@ -1014,7 +1014,7 @@ std::vector<ir::Tensor> Softmax(const ir::Tensor &A,
   return {out, temp};
 }
 
-#ifdef CINN_WITH_MKLDNN
+#ifdef CINN_WITH_DNNL
 std::vector<ir::Tensor> SoftmaxMKLDNN(const ir::Tensor &A,
                                       int axis,
                                       const std::string &output_name) {

paddle/cinn/hlir/pe/nn.h

@@ -178,7 +178,7 @@ std::vector<ir::Tensor> Conv2d_NCHWc(
     const std::string &output_name = UniqName("T_Conv2d_NCHWc_out"),
     const common::Target &target = common::DefaultHostTarget());
 
-#ifdef CINN_WITH_MKLDNN
+#ifdef CINN_WITH_DNNL
 std::vector<ir::Tensor> Conv2d_NCHW_MKLDNN(
     const ir::Tensor &input,
     const ir::Tensor &weights,
@@ -332,7 +332,7 @@ std::vector<ir::Tensor> Softmax(
     int axis = -1,
     const std::string &output_name = UniqName("T_softmax_out"));
 
-#ifdef CINN_WITH_MKLDNN
+#ifdef CINN_WITH_DNNL
 std::vector<ir::Tensor> SoftmaxMKLDNN(
     const ir::Tensor &A,
     int axis = -1,

paddle/cinn/runtime/cpu/mkldnn_math.h

@@ -15,7 +15,7 @@
 #pragma once
 #include "paddle/cinn/runtime/cinn_runtime.h"
 
-#ifdef CINN_WITH_MKLDNN
+#ifdef CINN_WITH_DNNL
 #include "dnnl.hpp"  // NOLINT
 #endif
 

paddle/cinn/runtime/cpu/use_extern_funcs.h

@@ -20,7 +20,7 @@ CINN_USE_REGISTER(host_intrinsics)
 #ifdef CINN_WITH_MKL_CBLAS
 CINN_USE_REGISTER(mkl_math)
 CINN_USE_REGISTER(cinn_cpu_mkl)
-#ifdef CINN_WITH_MKLDNN
+#ifdef CINN_WITH_DNNL
 CINN_USE_REGISTER(cinn_cpu_mkldnn)
 #endif
 #endif

paddle/fluid/distributed/collective/process_group_custom.cc

@@ -173,7 +173,7 @@ void ProcessGroupCustom::CreateCustomManagerCache(
     phi::DeviceGuard guard(places[i]);
     ccl_comms[i] = CustomCCLCommManager::Create(
         device_type, GetSize(), GetRank(), &ccl_id, new phi::ccl::CCLComm);
-    dev_ctx[i].reset(new CustomDeviceContext(places[i]));
+    dev_ctx[i] = std::make_unique<CustomDeviceContext>(places[i]);
     dev_ctx[i]->SetAllocator(
         &(phi::DeviceContextPool::Instance().Get(places[i])->GetAllocator()));
     dev_ctx[i]->SetHostAllocator(&(

paddle/fluid/distributed/collective/process_group_nccl.cc

@@ -625,7 +625,7 @@ void ProcessGroupNCCL::CreateNCCLManagerCache(
   for (size_t i = 0; i < places.size(); ++i) {
     platform::CUDADeviceGuard guard(places[i]);
 
-    dev_ctx[i].reset(new phi::GPUContext(places[i]));
+    dev_ctx[i] = std::make_unique<phi::GPUContext>(places[i]);
     ncclComm_t nccl_comm;
     NCCL_CHECK(phi::dynload::ncclCommInitRank(
         &nccl_comm, GetSize(), nccl_id, GetRank()));

paddle/fluid/distributed/fleet_executor/dist_model.cc

@@ -374,7 +374,7 @@ void DistModel::InsertCommOp(std::string tmp_var_name,
 }
 
 bool DistModel::PrepareScope() {
-  scope_.reset(new framework::Scope());
+  scope_ = std::make_unique<framework::Scope>();
   return true;
 }
 
@@ -412,7 +412,7 @@ bool DistModel::LoadProgram() {
   fin.close();
   program_proto.ParseFromString(pb_content);
   VLOG(5) << pb_content;
-  program_.reset(new framework::ProgramDesc(program_proto));
+  program_ = std::make_unique<framework::ProgramDesc>(program_proto);
   return true;
 }
 
@@ -469,7 +469,7 @@ bool DistModel::LoadParameters() {
 }
 
 bool DistModel::PrepareFleetExe() {
-  task_node_.reset(new TaskNode(program_.get(), config_.local_rank));
+  task_node_ = std::make_unique<TaskNode>(program_.get(), config_.local_rank);
   // With auto cut, there is no concept of pp, no need to add dependency.
   task_node_->SetType("Compute");
   task_node_->Init();
@@ -487,7 +487,7 @@ bool DistModel::PrepareFleetExe() {
     }
     id_to_rank.insert({i, i});
   }
-  fleet_exe.reset(new FleetExecutor(executor_desc_));
+  fleet_exe = std::make_unique<FleetExecutor>(executor_desc_);
   fleet_exe->Init(carrier_id_,
                   *(program_.get()),
                   scope_.get(),

paddle/fluid/distributed/ps/service/communicator/communicator.cc

@@ -796,8 +796,8 @@ void AsyncCommunicator::InitImpl(const RpcCtxMap &send_varname_to_ctx,
   send_varname_to_ctx_ = std::move(send_varname_to_ctx);
   recv_varname_to_ctx_ = std::move(recv_varname_to_ctx);
   recv_scope_ = std::move(recv_scope);
-  send_scope_.reset(new Scope());
-  xpu_temp_scope_.reset(new Scope());
+  send_scope_ = std::make_unique<Scope>();
+  xpu_temp_scope_ = std::make_unique<Scope>();
   for (auto &iter : send_varname_to_ctx_) {
     auto &ctx = iter.second;
     auto &varnames = ctx.origin_varnames;
@@ -807,7 +807,7 @@ void AsyncCommunicator::InitImpl(const RpcCtxMap &send_varname_to_ctx,
               send_queue_size_);
     }
   }
-  send_threadpool_.reset(new ::ThreadPool(thread_pool_size_));
+  send_threadpool_ = std::make_unique<::ThreadPool>(thread_pool_size_);
 }
 
 AsyncCommunicator::~AsyncCommunicator() {
@@ -1517,7 +1517,7 @@ void FLCommunicator::InitBrpcClient(
                              // before, but no need for Coordinator
   }
   if (coordinator_client_ptr_ == nullptr) {
-    coordinator_client_ptr_.reset(new CoordinatorClient);
+    coordinator_client_ptr_ = std::make_unique<CoordinatorClient>();
   }
   int16_t servers = host_sign_list.size();
   coordinator_client_ptr_->_env = &ps_env_;

paddle/fluid/distributed/ps/table/memory_sparse_geo_table.cc

@@ -138,7 +138,7 @@ int32_t MemorySparseGeoTable::Initialize() {
     shards_task.reset(new ::ThreadPool(1));
   }
 
-  _local_shards.reset(new shard_type[_task_pool_size]);
+  _local_shards.reset(new shard_type[_task_pool_size]);  // NOLINT
   return 0;
 }