Make convolution operator fully work with oneDNN v2.4+

src/operator/nn/dnnl/dnnl_convolution.cc

@@ -116,10 +116,18 @@ std::shared_ptr<dnnl::convolution_forward::primitive_desc> GetConvFwdImpl(
           // suboptimal kernel for computation that has the expected memory size requirements
           auto conv_pd =
               std::make_shared<dnnl::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.dnnl_param.quantized &&
-                  conv_pd->weights_desc().get_size() != GetArraySize(weights))) {
+          while (
+              conv_pd->dst_desc().get_size() != GetArraySize(output) ||
+              conv_pd->src_desc().get_size() != GetArraySize(data) ||
+              (!param.dnnl_param.quantized &&
+               conv_pd->weights_desc().get_size() != GetArraySize(weights)) ||
+              // With the upgrade of oneDNN to version 2.4+
+              // tests/python/dnnl/subgraphs/test_conv_subgraph.py::test_pos_conv_add[True-data_shape1]
+              // started failing. Switching away from primitive with weight dnnl::format_tag
+              // ABcd4b16a4b in order to temporarily fix the issue until full fix arrives.
+              // Tracking issue: https://github.com/apache/incubator-mxnet/issues/20826.
+              (param.dnnl_param.quantized && conv_pd->weights_desc().dims()[1] < 4 &&
+               conv_pd->weights_desc().data.padded_dims[1] == 16)) {
             // next_impl() will visit desc and engine, please make sure they are still alive here.
             CHECK(conv_pd->next_impl()) << "No convolution implementation for this request.";
           }

tests/python/dnnl/subgraphs/subgraph_common.py

@@ -42,10 +42,7 @@
   }
 }
 
-DATA_SHAPE=[(64, 4, 10, 10), (4, 4, 24, 24), (1, 16, 32, 32)]
-# Second shape has been temporairly changed from (4, 3, 24, 24) to (4, 4, 24, 24) due to
-# a bug regarding conv+sum fuse with the amount of input channels < 4. It will be reverted
-# as soon as the problem is fixed. Issue: https://github.com/apache/incubator-mxnet/issues/20826.
+DATA_SHAPE=[(64, 4, 10, 10), (4, 3, 24, 24), (1, 16, 32, 32)]
 
 # Helpers
 class RELU6(nn.HybridBlock):