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

… support_mutable_attributes

.clang-tidy

@@ -83,7 +83,7 @@ clang-analyzer-cplusplus.InnerPointer,
 -clang-analyzer-nullability.NullableDereferenced,
 -clang-analyzer-nullability.NullablePassedToNonnull,
 -clang-analyzer-nullability.NullableReturnedFromNonnull,
--clang-analyzer-optin.cplusplus.UninitializedObject,
+clang-analyzer-optin.cplusplus.UninitializedObject,
 -clang-analyzer-optin.cplusplus.VirtualCall,
 -clang-analyzer-optin.mpi.MPI-Checker,
 -clang-analyzer-optin.osx.OSObjectCStyleCast,
@@ -162,15 +162,15 @@ cppcoreguidelines-c-copy-assignment-signature,
 -cppcoreguidelines-pro-type-member-init,
 -cppcoreguidelines-slicing,
 -hicpp-avoid-goto,
--hicpp-exception-baseclass,
+hicpp-exception-baseclass,
 misc-unused-alias-decls,
 misc-unused-using-decls,
 modernize-avoid-bind,
 modernize-avoid-c-arrays,
 -modernize-deprecated-headers,
 -modernize-deprecated-ios-base-aliases,
 modernize-loop-convert,
--modernize-make-shared,
+modernize-make-shared,
 modernize-make-unique,
 -modernize-pass-by-value,
 modernize-raw-string-literal,

.gitignore

@@ -98,3 +98,4 @@ paddle/phi/kernels/fusion/cutlass/conv2d/generated/*
 python/paddle/fluid/incubate/fleet/parameter_server/pslib/ps_pb2.py
 paddle/fluid/ir_adaptor/translator/op_compat_info.cc
 paddle/fluid/pybind/static_op_function.*
+paddle/fluid/pybind/ops_api.cc

CMakeLists.txt

@@ -254,6 +254,7 @@ option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
 option(ON_INFER "Turn on inference optimization and inference-lib generation"
        ON)
 option(WITH_CPP_DIST "Install PaddlePaddle C++ distribution" OFF)
+option(WITH_GFLAGS "Compile PaddlePaddle with gflags support" OFF)
 ################################ Internal Configurations #######################################
 option(WITH_NV_JETSON "Compile PaddlePaddle with NV JETSON" OFF)
 option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler and gperftools"

cmake/cinn.cmake

@@ -183,6 +183,11 @@ if(WITH_MKL)
   endif()
 endif()
 
+if(NOT WITH_GFLAGS)
+  target_link_libraries(cinnapi gflags)
+  add_dependencies(cinnapi gflags)
+endif()
+
 if(WITH_GPU)
   target_link_libraries(
     cinnapi
@@ -237,6 +242,11 @@ function(gen_cinncore LINKTYPE)
     endif()
   endif()
 
+  if(NOT WITH_GFLAGS)
+    target_link_libraries(${CINNCORE_TARGET} gflags)
+    add_dependencies(${CINNCORE_TARGET} gflags)
+  endif()
+
   if(WITH_GPU)
     target_link_libraries(
       ${CINNCORE_TARGET}

cmake/configure.cmake

@@ -201,6 +201,10 @@ if(WITH_DISTRIBUTE)
   add_definitions(-DPADDLE_WITH_DISTRIBUTE)
 endif()
 
+if(WITH_GFLAGS)
+  add_definitions(-DPADDLE_WITH_GFLAGS)
+endif()
+
 if(WITH_PSCORE)
   add_definitions(-DPADDLE_WITH_PSCORE)
 endif()

cmake/external/brpc.cmake

@@ -91,3 +91,16 @@ add_dependencies(brpc extern_brpc)
 add_definitions(-DBRPC_WITH_GLOG)
 
 list(APPEND external_project_dependencies brpc)
+
+set(EXTERNAL_BRPC_DEPS
+    brpc
+    protobuf
+    ssl
+    crypto
+    leveldb
+    glog
+    snappy)
+
+if(NOT WITH_GFLAGS)
+  set(EXTERNAL_BRPC_DEPS ${EXTERNAL_BRPC_DEPS} gflags)
+endif()

cmake/external/gflags.cmake

@@ -102,3 +102,14 @@ if(WIN32)
     set_property(GLOBAL PROPERTY OS_DEPENDENCY_MODULES shlwapi.lib)
   endif()
 endif()
+
+# We have implemented a custom flags tool paddle_flags to replace gflags.
+# User can also choose to use gflags by setting WITH_GFLAGS=ON. But when
+# using paddle_flags, gflags is also needed for other third party libraries
+# including glog and brpc. So we can not remove gflags completely.
+set(flags_dep)
+if(WITH_GFLAGS)
+  list(APPEND flags_dep gflags)
+else()
+  list(APPEND flags_dep paddle_flags)
+endif()

cmake/inference_lib.cmake

@@ -336,11 +336,22 @@ copy(
   inference_lib_dist
   SRCS ${PADDLE_SOURCE_DIR}/paddle/utils/flat_hash_map.h
   DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/utils/)
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/utils/flags.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/utils/)
 copy(
   inference_lib_dist
   SRCS ${PADDLE_SOURCE_DIR}/paddle/extension.h
   DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/)
 
+if(NOT WITH_GFLAGS)
+  copy(
+    inference_lib_dist
+    SRCS ${PADDLE_SOURCE_DIR}/paddle/utils/flags_native.h
+    DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/utils/)
+endif()
+
 # the include path of phi needs to be changed to adapt to inference api path
 add_custom_command(
   TARGET inference_lib_dist

paddle/cinn/hlir/dialect/.gitignore

@@ -1 +1,2 @@
 generated/**
+generated/*

paddle/fluid/distributed/auto_parallel/spmd_rules/rules.h

@@ -19,7 +19,6 @@
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/elementwise_spmd_rule.h"
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/embedding_spmd_rule.h"
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/layer_norm_spmd_rule.h"
-#include "paddle/fluid/distributed/auto_parallel/spmd_rules/matmul_spmd_rule.h"
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/reduction_spmd_rule.h"
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/replicated_spmd_rule.h"
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/reshape_spmd_rule.h"
@@ -32,9 +31,6 @@ namespace paddle {
 namespace distributed {
 namespace auto_parallel {
 
-// matmul rule
-REGISTER_SPMD_RULE(matmul, MatmulSPMDRule);
-
 // reduction rules
 REGISTER_SPMD_RULE(all, ReductionSPMDRule);
 REGISTER_SPMD_RULE(amax, ReductionSPMDRule);

paddle/fluid/distributed/collective/process_group_custom.cc

@@ -27,7 +27,7 @@
 
 constexpr int64_t kWaitBlockTImeout = 10;
 
-DECLARE_bool(use_stream_safe_cuda_allocator);
+PD_DECLARE_bool(use_stream_safe_cuda_allocator);
 
 namespace paddle {
 namespace distributed {

paddle/fluid/distributed/collective/process_group_nccl.cc

@@ -29,7 +29,7 @@
 #include "paddle/phi/core/distributed/comm_context_manager.h"
 
 PHI_DECLARE_bool(nccl_blocking_wait);
-DECLARE_bool(use_stream_safe_cuda_allocator);
+PD_DECLARE_bool(use_stream_safe_cuda_allocator);
 
 // set this flag to `true` and recompile to enable dynamic checks
 constexpr bool FLAGS_enable_nccl_dynamic_check = false;

paddle/fluid/distributed/collective/reducer.cc

@@ -18,7 +18,7 @@
 #include "paddle/phi/backends/device_manager.h"
 #include "paddle/phi/core/flags.h"
 
-DECLARE_bool(use_stream_safe_cuda_allocator);
+PD_DECLARE_bool(use_stream_safe_cuda_allocator);
 PHI_DECLARE_string(allocator_strategy);
 
 namespace paddle {

paddle/fluid/distributed/fleet_executor/CMakeLists.txt

@@ -7,16 +7,7 @@ proto_library(interceptor_message_proto SRCS interceptor_message.proto)
 if(WITH_ARM_BRPC)
   set(BRPC_DEPS arm_brpc snappy phi glog)
 elseif(WITH_DISTRIBUTE AND NOT WITH_PSLIB)
-  set(BRPC_DEPS
-      brpc
-      ssl
-      crypto
-      protobuf
-      zlib
-      leveldb
-      snappy
-      phi
-      glog)
+  set(BRPC_DEPS ${EXTERNAL_BRPC_DEPS} zlib phi)
 else()
   set(BRPC_DEPS "")
 endif()

paddle/fluid/distributed/fleet_executor/carrier.cc

@@ -17,7 +17,6 @@
 #include <algorithm>
 #include <vector>
 
-#include "gflags/gflags.h"
 #include "paddle/fluid/distributed/fleet_executor/global.h"
 #include "paddle/fluid/distributed/fleet_executor/interceptor.h"
 #include "paddle/fluid/distributed/fleet_executor/message_bus.h"
@@ -29,6 +28,7 @@
 #include "paddle/fluid/framework/variable.h"
 #include "paddle/fluid/framework/variable_helper.h"
 #include "paddle/fluid/platform/flags.h"
+#include "paddle/utils/flags.h"
 PADDLE_DEFINE_EXPORTED_bool(
     fleet_executor_with_standalone,
     false,

paddle/fluid/distributed/ps/service/CMakeLists.txt

@@ -3,34 +3,11 @@ set_source_files_properties(${BRPC_SRCS})
 
 if(WITH_HETERPS)
 
-  set(BRPC_DEPS
-      brpc
-      ssl
-      crypto
-      protobuf
-      phi
-      glog
-      zlib
-      leveldb
-      snappy
-      glog
-      device_context
-      rocksdb)
+  set(BRPC_DEPS ${EXTERNAL_BRPC_DEPS} phi zlib device_context rocksdb)
 
 else()
 
-  set(BRPC_DEPS
-      brpc
-      ssl
-      crypto
-      protobuf
-      phi
-      glog
-      zlib
-      leveldb
-      snappy
-      glog
-      device_context)
+  set(BRPC_DEPS ${EXTERNAL_BRPC_DEPS} phi zlib device_context)
 
 endif()
 

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

@@ -34,49 +34,53 @@ class Variable;
 namespace paddle {
 namespace distributed {
 
-DEFINE_int32(pserver_push_dense_merge_limit,
-             12,
-             "limit max push_dense local merge requests");
+PD_DEFINE_int32(pserver_push_dense_merge_limit,
+                12,
+                "limit max push_dense local merge requests");
 
-DEFINE_int32(pserver_push_sparse_merge_limit,
-             12,
-             "limit max push_sparse local merge requests");
+PD_DEFINE_int32(pserver_push_sparse_merge_limit,
+                12,
+                "limit max push_sparse local merge requests");
 
-DEFINE_int32(pserver_pull_dense_limit,
-             12,
-             "limit max push_sparse local merge requests");
+PD_DEFINE_int32(pserver_pull_dense_limit,
+                12,
+                "limit max push_sparse local merge requests");
 
-DEFINE_int32(pserver_async_push_dense_interval_ms,
-             10,
-             "async push_dense to server interval");
+PD_DEFINE_int32(pserver_async_push_dense_interval_ms,
+                10,
+                "async push_dense to server interval");
 
-DEFINE_int32(pserver_async_push_sparse_interval_ms,
-             10,
-             "async push_sparse to server interval");
+PD_DEFINE_int32(pserver_async_push_sparse_interval_ms,
+                10,
+                "async push_sparse to server interval");
 
-DEFINE_bool(pserver_scale_gradient_by_merge,
-            false,
-            "scale dense gradient when merged");
+PD_DEFINE_bool(pserver_scale_gradient_by_merge,
+               false,
+               "scale dense gradient when merged");
 
-DEFINE_int32(pserver_communicate_compress_type,
-             0,
-             "none:0 snappy:1 gzip:2 zlib:3 lz4:4");
+PD_DEFINE_int32(pserver_communicate_compress_type,
+                0,
+                "none:0 snappy:1 gzip:2 zlib:3 lz4:4");
 
-DEFINE_int32(pserver_max_async_call_num,
-             13,
-             "max task num in async_call_server");
+PD_DEFINE_int32(pserver_max_async_call_num,
+                13,
+                "max task num in async_call_server");
 
-DEFINE_int32(pserver_timeout_ms, 500000, "pserver request server timeout_ms");
+PD_DEFINE_int32(pserver_timeout_ms,
+                500000,
+                "pserver request server timeout_ms");
 
-DEFINE_int32(pserver_connect_timeout_ms,
-             10000,
-             "pserver connect server timeout_ms");
+PD_DEFINE_int32(pserver_connect_timeout_ms,
+                10000,
+                "pserver connect server timeout_ms");
 
-DEFINE_int32(pserver_sparse_merge_thread, 1, "pserver sparse merge thread num");
+PD_DEFINE_int32(pserver_sparse_merge_thread,
+                1,
+                "pserver sparse merge thread num");
 
-DEFINE_int32(pserver_sparse_table_shard_num,
-             1000,
-             "sparse table shard for save & load");
+PD_DEFINE_int32(pserver_sparse_table_shard_num,
+                1000,
+                "sparse table shard for save & load");
 
 inline size_t get_sparse_shard(uint32_t shard_num,
                                uint32_t server_num,
@@ -140,7 +144,7 @@ int32_t BrpcPsClient::StartFlClientService(const std::string &self_endpoint) {
 
   if (_fl_server.Start(self_endpoint.c_str(), &options) != 0) {
     VLOG(0) << "fl-ps > StartFlClientService failed. Try again.";
-    auto ip_port = paddle::string::Split(self_endpoint, ':');
+    auto ip_port = ::paddle::string::Split(self_endpoint, ':');
     std::string ip = ip_port[0];
     int port = std::stoi(ip_port[1]);
     std::string int_ip_port = GetIntTypeEndpoint(ip, port);
@@ -202,8 +206,7 @@ int32_t BrpcPsClient::InitializeFlWorker(const std::string &self_endpoint) {
   options.protocol = "baidu_std";
   options.timeout_ms = FLAGS_pserver_timeout_ms;
   options.connection_type = "pooled";
-  options.connect_timeout_ms =
-      paddle::distributed::FLAGS_pserver_connect_timeout_ms;
+  options.connect_timeout_ms = FLAGS_pserver_connect_timeout_ms;
   options.max_retry = 3;
   // 获取 coordinator 列表，并连接
   std::string coordinator_ip_port;
@@ -336,11 +339,11 @@ int32_t BrpcPsClient::Initialize() {
     auto table_id = worker_param.downpour_table_param(i).table_id();
     if (type == PS_DENSE_TABLE) {
       _push_dense_task_queue_map[table_id] =
-          paddle::framework::MakeChannel<DenseAsyncTask *>();
+          ::paddle::framework::MakeChannel<DenseAsyncTask *>();
     }
     if (type == PS_SPARSE_TABLE) {
       _push_sparse_task_queue_map[table_id] =
-          paddle::framework::MakeChannel<SparseAsyncTask *>();
+          ::paddle::framework::MakeChannel<SparseAsyncTask *>();
       _push_sparse_merge_count_map[table_id] = 0;
     }
   }
@@ -446,7 +449,7 @@ std::future<int32_t> BrpcPsClient::PrintTableStat(uint32_t table_id) {
         int ret = 0;
         uint64_t feasign_size = 0;
         uint64_t mf_size = 0;
-        paddle::framework::BinaryArchive ar;
+        ::paddle::framework::BinaryArchive ar;
         auto *closure = reinterpret_cast<DownpourBrpcClosure *>(done);
         for (size_t i = 0; i < request_call_num; ++i) {
           if (closure->check_response(i, PS_PRINT_TABLE_STAT) != 0) {