Optimize the overhead for all_elementwise case in BroadcastKernel. (PaddlePaddle#57247)

Author: Xreki

paddle/phi/kernels/funcs/broadcast_function.h

@@ -74,14 +74,13 @@ struct BroadcastTypeClassifier {
   void InitBroadcastConfigs(const std::vector<const DenseTensor *> &ins,
                             std::vector<DenseTensor *> *outs,
                             int axis) {
+#ifdef PADDLE_WITH_XPU_KP
     const auto dims_simplifier =
         BroadcastDimsSimplifier(ins, (*outs)[0]->dims(), axis);
     if (VLOG_IS_ON(6)) {
       DimsSimplifiedLogger<int64_t>::Log(
           ins, outs, dims_simplifier, "BroadcastKernel");
     }
-
-#ifdef PADDLE_WITH_XPU_KP
     configs[0] = kps::details::BroadcastConfig(dims_simplifier.out_dims,
                                                dims_simplifier.in_dims[0],
                                                dims_simplifier.in_dims[1],
@@ -92,6 +91,12 @@ struct BroadcastTypeClassifier {
                                                dims_simplifier.rank);
 #else
     if (!all_elementwise) {
+      const auto dims_simplifier =
+          BroadcastDimsSimplifier(ins, (*outs)[0]->dims(), axis);
+      if (VLOG_IS_ON(6)) {
+        DimsSimplifiedLogger<int64_t>::Log(
+            ins, outs, dims_simplifier, "BroadcastKernel");
+      }
       for (int i = 0; i < Arity; ++i) {
         // if data shape is[m, n], then you should set data_dim = {n, m}
         // eg: out's shape [3, 45, 1]. then out_dims = {1, 45, 3}