add xpu controlflow (#51488)

paddle/fluid/operators/controlflow/conditional_block_op.h

@@ -89,6 +89,13 @@ class ConditionalOp : public framework::OperatorBase {
       framework::TensorCopy(*ips[0], platform::CPUPlace(), &cpu_tensor);
       platform::DeviceContextPool::Instance().Get(ips[0]->place())->Wait();
       res = cpu_tensor.data<bool>()[0];
+#endif
+    } else if (platform::is_xpu_place(ips[0]->place())) {
+#ifdef PADDLE_WITH_XPU
+      phi::DenseTensor cpu_tensor;
+      framework::TensorCopy(*ips[0], platform::CPUPlace(), &cpu_tensor);
+      platform::DeviceContextPool::Instance().Get(ips[0]->place())->Wait();
+      res = cpu_tensor.data<bool>()[0];
 #endif
     } else if (platform::is_custom_place(ips[0]->place())) {
 #if defined(PADDLE_WITH_CUSTOM_DEVICE)

paddle/fluid/operators/select_op_helper.h

@@ -40,7 +40,7 @@ inline int GetBranchNumber(const phi::DenseTensor &mask) {
   // when platform::is_gpu_place(mask.place()) is true
   std::unique_ptr<phi::DenseTensor> cpu_mask{new phi::DenseTensor()};
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP) || \
-    defined(PADDLE_WITH_CUSTOM_DEVICE)
+    defined(PADDLE_WITH_CUSTOM_DEVICE) || defined(PADDLE_WITH_XPU)
   framework::TensorCopySync(mask, platform::CPUPlace(), cpu_mask.get());
 #else
   PADDLE_THROW(platform::errors::PreconditionNotMet(