Fix windows compiling (#16230)

test=develop

paddle/fluid/framework/ir/CMakeLists.txt

@@ -102,7 +102,9 @@ cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS g
 cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
 cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DEPS seqpool_concat_fuse_pass framework_proto)
 cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
-cc_test(test_sync_batch_norm_pass SRCS sync_batch_norm_pass_tester.cc DEPS sync_batch_norm_pass)
+if(NOT WIN32)
+    cc_test(test_sync_batch_norm_pass SRCS sync_batch_norm_pass_tester.cc DEPS sync_batch_norm_pass)
+endif()
 cc_test(test_cpu_quantize_squash_pass SRCS cpu_quantize_squash_pass_tester.cc DEPS cpu_quantize_squash_pass naive_executor)
 if (WITH_MKLDNN)
     cc_test(test_depthwise_conv_mkldnn_pass SRCS mkldnn/depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)

paddle/fluid/operators/CMakeLists.txt

@@ -58,8 +58,10 @@ if (WITH_GPU)
         op_library(conv_fusion_op)
         file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(conv2d_fusion);\n")
     endif()
-    op_library(sync_batch_norm_op)
-    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(sync_batch_norm);\n")
+    if (NOT WIN32)
+        op_library(sync_batch_norm_op)
+        file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(sync_batch_norm);\n")
+    endif()
 else()
     op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
 endif()

paddle/fluid/platform/device_context.cc

@@ -316,7 +316,9 @@ CUDADeviceContext::~CUDADeviceContext() {
   eigen_stream_.reset();
   eigen_device_.reset();
   PADDLE_ENFORCE(cudaStreamDestroy(stream_));
+#if !defined(_WIN32)
   PADDLE_ENFORCE(dynload::ncclCommDestroy(nccl_comm_));
+#endif
 }
 
 Place CUDADeviceContext::GetPlace() const { return place_; }

paddle/fluid/platform/device_context.h

@@ -265,11 +265,13 @@ class CUDADeviceContext : public DeviceContext {
   /*! \brief  Return cuda stream in the device context. */
   cudaStream_t stream() const;
 
+#if !defined(_WIN32)
   /*! \brief  Return nccl communicators. */
   ncclComm_t nccl_comm() const { return nccl_comm_; }
 
   /*! \brief  Set nccl communicators. */
   void set_nccl_comm(ncclComm_t comm) { nccl_comm_ = comm; }
+#endif
 
   template <typename Callback>
   void RecordEvent(cudaEvent_t ev, Callback callback) {
@@ -295,12 +297,14 @@ class CUDADeviceContext : public DeviceContext {
   std::unique_ptr<CublasHandleHolder> cublas_handle_;
   std::unique_ptr<CublasHandleHolder> cublas_tensor_core_handle_;
 
+#if !defined(_WIN32)
   // NCCL communicator (single process version) for NCCL collective operations.
   // NCCL collective operations provides fast collectives over multiple GPUs
   // both within and across nodes.
   // But, this collectives is used for collectives over multiple GPUs within
   // nodes.
   ncclComm_t nccl_comm_{nullptr};
+#endif
 
   int compute_capability_;
   int runtime_version_;