Fix mkldnn backend when using naive engine

src/operator/nn/mkldnn/mkldnn_convolution.cc

@@ -411,11 +411,12 @@ void MKLDNNConvolutionForwardFullFeature(const MKLDNNConvFullParam &param,
     // For inference, we want to reorder the weight array so we don't need to
     // reorder data every time.
     if (weight.IsDefaultData()) {
-      weight_mem = GetWeights(weight, fwd->fwd_pd.weights_primitive_desc(),
-                              param.conv_param.num_group);
       // We also need to modify the layout on the original weight array. The
       // data conversion happens after the weight array is used.
       weight.MKLDNNDataReorderAsync(fwd->fwd_pd.weights_primitive_desc());
+      weight_mem = GetWeights(weight, fwd->fwd_pd.weights_primitive_desc(),
+                              param.conv_param.num_group);
+
     } else {
       weight_mem = weight.GetMKLDNNData();
       CHECK(weight_mem->get_primitive_desc() == fwd->fwd_pd.weights_primitive_desc());

src/operator/nn/mkldnn/mkldnn_deconvolution.cc

@@ -262,10 +262,10 @@ void MKLDNNDeconvForward::SetDataHandle(const DeconvolutionParam& param,
     // For inference, we want to reorder the weight array so we don't need to
     // reorder data every time.
     if (weight.IsDefaultData()) {
-      weight_mem = GetWeights(weight, fwd_pd.weights_primitive_desc(), param.num_group);
       // We also need to modify the layout on the original weight array. The
       // data conversion happens after the weight array is used.
       const_cast<NDArray&>(weight).MKLDNNDataReorderAsync(fwd_pd.weights_primitive_desc());
+      weight_mem = GetWeights(weight, fwd_pd.weights_primitive_desc(), param.num_group);
     } else {
       weight_mem = weight.GetMKLDNNData();
       CHECK(weight_mem->get_primitive_desc() == fwd_pd.weights_primitive_desc());

src/operator/nn/mkldnn/mkldnn_fully_connected.cc

@@ -253,8 +253,8 @@ void MKLDNNFCForwardFullFeature(const MKLDNNFCFullParam &full_param,
     weight_mem = GetWeights(weight, fwd->fwd_pd.weights_primitive_desc(), 1);
   } else {
     if (weight.IsDefaultData()) {
-      weight_mem = GetWeights(weight, fwd->fwd_pd.weights_primitive_desc(), 1);
       weight.MKLDNNDataReorderAsync(fwd->fwd_pd.weights_primitive_desc());
+      weight_mem = GetWeights(weight, fwd->fwd_pd.weights_primitive_desc(), 1);
     } else {
       weight_mem = weight.GetMKLDNNData();
       CHECK(weight_mem->get_primitive_desc() == fwd->fwd_pd.weights_primitive_desc());

src/operator/quantization/mkldnn/mkldnn_quantized_conv.cc

@@ -52,10 +52,10 @@ static void MKLDNNQuantizedConvForward(const nnvm::NodeAttrs& attrs,
   // For inference, we want to reorder the weight array so we don't need to
   // reorder data every time.
   if (weight.IsDefaultData()) {
-    weight_mem = GetWeights(weight, fwd.fwd_pd.weights_primitive_desc(), param.num_group);
     // We also need to modify the layout on the original weight array. The
     // data conversion happens after the weight array is used.
     weight.MKLDNNDataReorderAsync(fwd.fwd_pd.weights_primitive_desc());
+    weight_mem = GetWeights(weight, fwd.fwd_pd.weights_primitive_desc(), param.num_group);
   } else {
     weight_mem = weight.GetMKLDNNData();
     CHECK(weight_mem->get_primitive_desc() == fwd.fwd_pd.weights_primitive_desc());

src/operator/quantization/mkldnn/mkldnn_quantized_fully_connected.cc

@@ -93,8 +93,8 @@ void MKLDNNQuantizedFullyConnectedForward(const nnvm::NodeAttrs &attrs,
   const mkldnn::memory *weight_mem = nullptr;
 
   if (weight.IsDefaultData()) {
-    weight_mem = GetWeights(weight, fwd.fwd_pd.weights_primitive_desc(), 1);
     weight.MKLDNNDataReorderAsync(fwd.fwd_pd.weights_primitive_desc());
+    weight_mem = GetWeights(weight, fwd.fwd_pd.weights_primitive_desc(), 1);
   } else {
     weight_mem = weight.GetMKLDNNData();
     CHECK(weight_mem->get_primitive_desc() == fwd.fwd_pd.weights_primitive_desc());