[ROCm] Fix 3D tensor perf degradation with NHWC format (#2175)

Co-author: @doru1004

Author: jerrymannil

aten/src/ATen/native/cuda/Reduce.cuh

@@ -1159,7 +1159,8 @@ ReduceConfig setReduceConfig(const TensorIterator& iter){
       config.ctas_per_output = div_up(num_mp, 2);
     else if (config.ctas_per_output < 16)
       config.ctas_per_output = 1;
-    if (iter.ndim() == 3 && !reduction_on_fastest_striding_dimension)
+    bool is_channel_last = iter.tensor_base(1).is_contiguous(at::MemoryFormat::ChannelsLast);
+    if (iter.ndim() == 3 && !reduction_on_fastest_striding_dimension && !is_channel_last)
       config.ctas_per_output = 4;
 #endif
     if (config.ctas_per_output > 1) {