cherry-pick 22509. test=develop test=release/1.7 (#22527)

[cherry-pick] #22509

支持不依赖nccl进行编译。

多卡下，如果没有打开WITH_NCCL开关编译，多卡不能通信，则只能选择一张卡使用

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -20,7 +20,7 @@
 #include "paddle/fluid/platform/gpu_info.h"
 #include "paddle/fluid/platform/profiler.h"
 
-#ifdef PADDLE_WITH_CUDA
+#ifdef PADDLE_WITH_NCCL
 DECLARE_bool(sync_nccl_allreduce);
 #endif
 

paddle/fluid/framework/details/build_strategy.cc

@@ -303,7 +303,7 @@ ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
                                 const std::string &loss_var_name,
                                 const std::vector<Scope *> &local_scopes,
                                 const size_t &nranks,
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
                                 const bool use_cuda,
                                 platform::NCCLCommunicator *nccl_ctxs) const {
 #else
@@ -326,7 +326,7 @@ ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
       pass->Erase(kNRanks);
       pass->Set<size_t>(kNRanks, new size_t(nranks));
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
       platform::NCCLCommunicator *nctx = use_cuda ? nccl_ctxs : nullptr;
       pass->Erase(kNCCLCtxs);
       pass->SetNotOwned<platform::NCCLCommunicator>(kNCCLCtxs, nctx);
@@ -339,7 +339,7 @@ ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
       pass->Erase(kLocalScopes);
       pass->SetNotOwned<const std::vector<Scope *>>(kLocalScopes,
                                                     &local_scopes);
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
       platform::NCCLCommunicator *nctx = use_cuda ? nccl_ctxs : nullptr;
       pass->Erase(kNCCLCtxs);
       pass->SetNotOwned<platform::NCCLCommunicator>(kNCCLCtxs, nctx);
@@ -354,7 +354,7 @@ ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
       LOG(INFO) << "set enable_sequential_execution:"
                 << enable_sequential_execution_;
     } else if (pass->Type() == "all_reduce_deps_pass") {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
       platform::NCCLCommunicator *nctx = use_cuda ? nccl_ctxs : nullptr;
       pass->Erase(kNCCLCtxs);
       pass->SetNotOwned<platform::NCCLCommunicator>(kNCCLCtxs, nctx);

paddle/fluid/framework/details/build_strategy.h

@@ -26,7 +26,7 @@
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -162,7 +162,7 @@ struct BuildStrategy {
                    const std::string &loss_var_name,
                    const std::vector<Scope *> &local_scopes,
                    const size_t &nranks,
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
                    const bool use_cuda,
                    platform::NCCLCommunicator *nccl_ctxs) const;
 #else

paddle/fluid/framework/details/reduce_op_handle_test.cc

@@ -36,7 +36,7 @@ struct TestReduceOpHandle {
   std::vector<p::Place> gpu_list_;
   std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
 #endif
 
@@ -44,7 +44,7 @@ struct TestReduceOpHandle {
     for (size_t j = 0; j < ctxs_.size(); ++j) {
       ctxs_[j]->Wait();
     }
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
     if (nccl_ctxs_) {
       nccl_ctxs_->WaitAll();
     }
@@ -54,7 +54,7 @@ struct TestReduceOpHandle {
   void InitCtxOnGpu(bool use_gpu) {
     use_gpu_ = use_gpu;
     if (use_gpu) {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
       int count = p::GetCUDADeviceCount();
       if (count <= 1) {
         LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
@@ -78,7 +78,7 @@ struct TestReduceOpHandle {
         gpu_list_.push_back(p);
         ctxs_.emplace_back(new p::CPUDeviceContext(p));
       }
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
       nccl_ctxs_.reset(nullptr);
 #endif
     }
@@ -99,14 +99,14 @@ struct TestReduceOpHandle {
 
     nodes.emplace_back(new ir::Node("node"));
     if (use_gpu_) {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
       op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_,
                                           gpu_list_, nccl_ctxs_.get()));
 #else
       PADDLE_THROW("CUDA is not support.");
 #endif
     } else {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
       op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_,
                                           gpu_list_, nccl_ctxs_.get()));
 #else

paddle/fluid/framework/device_worker.h

@@ -38,7 +38,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/platform/timer.h"
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -257,7 +257,7 @@ class DownpourWorker : public HogwildWorker {
   std::unordered_map<uint64_t, std::unordered_set<uint64_t>> feasign_set_;
 };
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 using ScopeQueue = operators::reader::BlockingQueue<Scope*>;
 
 class SyncFunctor {

paddle/fluid/framework/fleet/nccl_wrapper.cc

@@ -24,7 +24,7 @@ std::shared_ptr<NCCLWrapper> NCCLWrapper::s_instance_ = NULL;
 bool NCCLWrapper::is_initialized_ = false;
 
 void NCCLWrapper::InitNCCL() {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
   PADDLE_ENFORCE(platform::dynload::ncclCommInitRank(
       &(nccl_info_.comm_), nccl_info_.global_ranks_, nccl_info_.nccl_id_,
       nccl_info_.my_global_rank_));
@@ -33,14 +33,14 @@ void NCCLWrapper::InitNCCL() {
 }
 
 void NCCLWrapper::SetNCCLId(const NCCLInfo& nccl_info) {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
   nccl_info_.nccl_id_ = nccl_info.nccl_id_;
 #endif
   return;
 }
 
 NCCLInfo NCCLWrapper::GetNCCLId() {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
   PADDLE_ENFORCE(platform::dynload::ncclGetUniqueId(&(nccl_info_.nccl_id_)));
 #endif
   return nccl_info_;
@@ -48,7 +48,7 @@ NCCLInfo NCCLWrapper::GetNCCLId() {
 
 void NCCLWrapper::SetRankInfo(const int local_rank, const int global_rank,
                               const int ranks) {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
   nccl_info_.local_rank_ = local_rank;
   nccl_info_.my_global_rank_ = global_rank;
   nccl_info_.global_ranks_ = ranks;
@@ -60,7 +60,7 @@ void NCCLWrapper::SetRankInfo(const int local_rank, const int global_rank,
 
 void NCCLWrapper::SyncVar(const int root_rank, const Scope& scope,
                           const std::vector<std::string>& var_names) {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
   for (auto& name : var_names) {
     auto var = scope.FindVar(name);
     LoDTensor* tensor = var->GetMutable<LoDTensor>();

paddle/fluid/framework/fleet/nccl_wrapper.h

@@ -24,7 +24,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/variable_helper.h"
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/dynload/nccl.h"
 #endif
 #include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
@@ -41,7 +41,7 @@ class NCCLInfo {
   int local_rank_;
   int global_ranks_;
   int my_global_rank_;
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
   ncclUniqueId nccl_id_;
   ncclComm_t comm_;
   cudaStream_t stream_;

paddle/fluid/framework/parallel_executor.cc

@@ -247,7 +247,7 @@ class ParallelExecutorPrivate {
 
   std::unordered_map<std::string, bool> is_persistable_;
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
   platform::NCCLCommunicator *nccl_ctxs_{nullptr};
 #endif
   bool own_local_scope_;
@@ -427,6 +427,16 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
   }
 #endif
 
+#if defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_NCCL)
+  PADDLE_ENFORCE_EQ(
+      places.size(), 1,
+      platform::errors::PermissionDenied(
+          "Your machine has multiple cards, "
+          "but the WITH_NCCL option is not turned on during compilation, "
+          "and you cannot use multi-card training or prediction. "
+          "Please recompile and turn on the WITH_NCCL option."));
+#endif
+
   LOG(INFO) << string::Sprintf(
       "The Program will be executed on %s using ParallelExecutor, %lu "
       "cards are used, so %lu programs are executed in parallel.",
@@ -516,7 +526,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
   // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
   // ncclOp
   std::vector<ir::Graph *> async_graphs(places.size());
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
   if (member_->build_strategy_.async_mode_) {
     VLOG(3) << "use local async mode";
     graph = member_->build_strategy_.Apply(

paddle/fluid/framework/trainer.h

@@ -110,7 +110,7 @@ class DistMultiTrainer : public MultiTrainer {
   int dump_file_num_;
 };
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 class PipelineTrainer : public TrainerBase {
  public:
   PipelineTrainer() {}

paddle/fluid/framework/var_type_traits.cc

@@ -21,7 +21,7 @@
 #include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
 #include "paddle/fluid/platform/macros.h"
 #ifdef PADDLE_WITH_CUDA
-#ifndef _WIN32
+#if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif

paddle/fluid/framework/var_type_traits.h

@@ -24,7 +24,7 @@
 #include "paddle/fluid/platform/place.h"
 #ifdef PADDLE_WITH_CUDA
 #include <cudnn.h>
-#ifndef _WIN32
+#if defined(PADDLE_WITH_NCCL)
 #include <nccl.h>
 #endif
 #endif
@@ -34,7 +34,7 @@ namespace paddle {
 
 namespace platform {
 #ifdef PADDLE_WITH_CUDA
-#ifndef _WIN32
+#if defined(PADDLE_WITH_NCCL)
 class Communicator;
 class NCCLCommunicator;
 #endif
@@ -140,7 +140,7 @@ using VarTypeRegistry = detail::VarTypeRegistryImpl<
     LoDTensorArray, platform::PlaceList, ReaderHolder, std::string, Scope *,
     operators::reader::LoDTensorBlockingQueueHolder,
 #ifdef PADDLE_WITH_CUDA
-#ifndef _WIN32
+#if defined(PADDLE_WITH_NCCL)
     ncclUniqueId, platform::Communicator, platform::NCCLCommunicator,
 #endif
     operators::CudnnRNNCache,

paddle/fluid/framework/var_type_traits_test.cc

@@ -24,7 +24,7 @@
 #include "paddle/fluid/framework/var_type_traits.h"
 #include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
 #ifdef PADDLE_WITH_CUDA
-#ifndef _WIN32
+#if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif

paddle/fluid/imperative/tests/nccl_context_test.cc

@@ -29,7 +29,7 @@ imperative::ParallelStrategy GetStrategy(int local_rank) {
   return strategy;
 }
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 void BcastNCCLId(int local_rank, ncclUniqueId *nccl_id) {
   auto strategy = GetStrategy(local_rank);
   platform::CUDAPlace gpu(local_rank);

paddle/fluid/operators/collective/c_allgather_op.cu.cc

@@ -16,7 +16,7 @@ limitations under the License. */
 
 #include <memory>
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/collective_helper.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
@@ -28,7 +28,7 @@ template <typename T>
 class CAllGatherOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
     auto in = ctx.Input<framework::Tensor>("X");
     auto out = ctx.Output<framework::Tensor>("Out");
     ncclDataType_t dtype = platform::ToNCCLDataType(in->type());

paddle/fluid/operators/collective/c_allreduce_op.h

@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/collective_helper.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
@@ -58,7 +58,7 @@ template <ReduceType red_type, typename T>
 class CAllReduceOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
     auto in = ctx.Input<framework::Tensor>("X");
     auto out = ctx.Output<framework::Tensor>("Out");
 

paddle/fluid/operators/collective/c_broadcast_op.cu.cc

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/collective/c_broadcast_op.h"
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/collective_helper.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
@@ -26,7 +26,7 @@ template <typename T>
 class CBroadcastOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
     auto x = ctx.Input<framework::LoDTensor>("X");
     auto out = ctx.Output<framework::LoDTensor>("Out");
     int numel = x->numel();

paddle/fluid/operators/collective/c_comm_init_all_op.cc

@@ -11,7 +11,7 @@ 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. */
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 #include <nccl.h>
 #endif
 #include <stdint.h>
@@ -25,7 +25,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/operators/distributed/distributed.h"
 #include "paddle/fluid/operators/distributed/request_handler_impl.h"
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/collective_helper.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
@@ -52,7 +52,7 @@ class CCommInitAllOp : public framework::OperatorBase {
     PADDLE_ENFORCE_EQ(is_gpu_place(place), true,
                       "CCommInitAllOp can run on gpu place only.");
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
     std::vector<int> devices = Attr<std::vector<int>>("devices");
     if (devices.empty()) {
       devices = platform::GetSelectedDevices();

paddle/fluid/operators/collective/c_comm_init_op.cc

@@ -11,7 +11,7 @@ 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. */
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 #include <nccl.h>
 #endif
 #include <stdint.h>
@@ -24,7 +24,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/operators/distributed/distributed.h"
 #include "paddle/fluid/operators/distributed/request_handler_impl.h"
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/collective_helper.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
@@ -46,7 +46,7 @@ class CCommInitOp : public framework::OperatorBase {
 
     auto var = scope.FindVar(Input("X"));
     PADDLE_ENFORCE_NOT_NULL(var);
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#if defined(PADDLE_WITH_NCCL)
     ncclUniqueId* nccl_id = var->GetMutable<ncclUniqueId>();
 
     int nranks = Attr<int>("nranks");