[GitHub] ZhennanQin commented on a change in pull request #12530: Implement mkldnn convolution fusion and quantization.

[GitBox <gi...@apache.org>]

ZhennanQin commented on a change in pull request #12530: Implement mkldnn convolution fusion and quantization.
URL: https://github.com/apache/incubator-mxnet/pull/12530#discussion_r219376419
 
 

 ##########
 File path: src/operator/nn/mkldnn/mkldnn_convolution.cc
 ##########
 @@ -257,32 +300,59 @@ void MKLDNNConvolutionForward(const nnvm::NodeAttrs& attrs, const OpContext &ctx
       // This asks the engine to change the layout of the weight array after
       // it's used.
       weight.Reorder2DefaultAsync();
-    weight_mem = GetWeights(weight, fwd.fwd_pd.weights_primitive_desc(), param.num_group);
+    weight_mem = GetWeights(weight, fwd->fwd_pd.weights_primitive_desc(),
+                            param.conv_param.num_group);
   } else {
     // 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);
+      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.MKLDNNDataReorderAsync(fwd->fwd_pd.weights_primitive_desc());
     } else {
       weight_mem = weight.GetMKLDNNData();
-      CHECK(weight_mem->get_primitive_desc() == fwd.fwd_pd.weights_primitive_desc());
+      CHECK(weight_mem->get_primitive_desc() == fwd->fwd_pd.weights_primitive_desc());
     }
   }
-  auto out_mem = CreateMKLDNNMem(out_data[conv::kOut], fwd.fwd_pd.dst_primitive_desc(),
-                                 req[conv::kOut]);
+  mkldnn_output_t out_mem;
+  if (param.mkldnn_param.with_sum) {
+    out_mem = mkldnn_output_t(
+        OutDataOp::Noop,
+        const_cast<mkldnn::memory *>(out_data[conv::kOut].GetMKLDNNDataReorder(
+            fwd->fwd_pd.dst_primitive_desc())));
 
 Review comment:
   Actually, out_data always has same memory description with fwd->fwd_pd, so we should use GetMKLDNNData instead of GetMKLDNNDataReorder. Then out_mem will always be the memory in out_data.

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services