convert output to nchw format to align with native version in avx512 mode

test = develop
resolve #16764

paddle/fluid/operators/elementwise/mkldnn/elementwise_mul_mkldnn_op.cc

@@ -130,6 +130,13 @@ class ElementwiseMulMKLDNNKernel : public framework::OpKernel<T> {
 
       z->set_layout(DataLayout::kMKLDNN);
       z->set_format(x->format());
+
+      // convert to nchw format to align with native version
+      using platform::MKLDNNDeviceContext;
+      auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+      const auto& mkldnn_engine = dev_ctx.GetEngine();
+      ReorderInput<T>(const_cast<Tensor*>(z), ctx.GetPlace(), mkldnn_engine,
+                      z->dims().size() == 4);
     } else {
       // Fallback to naive version:
       const bool are_inputs_in_same_format = x->format() == y->format();

python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_mul_mkldnn_op.py

@@ -28,7 +28,8 @@ class TestElementwiseMulMKLDNNOp_BroadcastNCHW16c(ElementwiseMulOp):
         self.y = np.random.rand(1, 16).astype(self.dtype)
 
         self.out = x * self.y.reshape(1, 16, 1, 1)
-        self.out = self.out.transpose(0, 2, 3, 1).reshape(1, 16, 2, 2)
+
+#        self.out = self.out.transpose(0, 2, 3, 1).reshape(1, 16, 2, 2)
 
     def setUp(self):
         super(TestElementwiseMulMKLDNNOp_BroadcastNCHW16c, self).setUp()