[GitHub] [incubator-mxnet] bartekkuncer commented on a change in pull request #20666: [v1.x][submodule] Upgrade oneDNN to v2.4.4

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

bartekkuncer commented on a change in pull request #20666:
URL: https://github.com/apache/incubator-mxnet/pull/20666#discussion_r792453753



##########
File path: src/operator/nn/mkldnn/mkldnn_convolution.cc
##########
@@ -113,16 +113,28 @@ std::shared_ptr<mkldnn::convolution_forward::primitive_desc> GetConvFwdImpl(
       [&param, &data, &weights, &output, &attr](const mkldnn::convolution_forward::desc& desc) {
         auto engine = CpuEngine::Get()->get_engine();
         try {
-          // MKL-DNN introduced padded formats since 0.15 which require more memory
-          // compared to the actual size of the tensor. Currently, MKL-DNN operators
-          // still reuse memory from memory planning, so here we need to select a
-          // suboptimal kernel for computation that has the expected memory size requirements
+          // MKLDNN introduced padded formats since 0.15 which require more
+          // memory compared to the actual size of the tensor. Currently, MKLDNN
+          // operators still reuse memory from memory planning, so here we need
+          // to select a suboptimal kernel for computation that has the expected
+          // memory size requirements
           auto conv_pd =
-              std::make_shared<mkldnn::convolution_forward::primitive_desc>(desc, attr, engine);
-          while (conv_pd->dst_desc().get_size() != GetArraySize(output) ||
-                 conv_pd->src_desc().get_size() != GetArraySize(data) ||
-                 (!param.mkldnn_param.quantized &&
-                  conv_pd->weights_desc().get_size() != GetArraySize(weights))) {
+              std::make_shared<mkldnn::convolution_forward::primitive_desc>(
+                  desc, attr, engine);
+          while (
+              conv_pd->dst_desc().get_size() != GetArraySize(output) ||
+              conv_pd->src_desc().get_size() != GetArraySize(data) ||
+              (!param.mkldnn_param.quantized &&
+               conv_pd->weights_desc().get_size() != GetArraySize(weights)) ||
+              // With the upgrade of MKLDNN to version 2.4+
+              // tests/python/mkl/test_subgraph.py::test_pos_conv_add
+              // started failing. Switching away from primitive with weight
+              // mkldnn::format_tag ABcd4b16a4b is in place to temporairly fix

Review comment:
       ```suggestion
                 // mkldnn::format_tag ABcd4b16a4b in order to temporarily fix
   ```




-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: commits-unsubscribe@mxnet.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org