Fix compose/decompose of group_conv for CPU backend. (#3947)

* disable group conv if data dilation is not 1s

* Enable group conv decomposition if data or window dilation is enabled. Re-enable a unit-test

* PR fix

* PR fix

* Fix fails

src/ngraph/pass/batch_fusion.cpp

@@ -62,8 +62,8 @@ static bool is_trivial_convolution(std::shared_ptr<op::Convolution> conv)
 {
     Strides stride_1{1, 1};
     CoordinateDiff pad_0{0, 0};
-    return conv->get_window_dilation_strides() == stride_1 ||
-           conv->get_data_dilation_strides() == stride_1 || conv->get_padding_above() == pad_0 ||
+    return conv->get_window_dilation_strides() == stride_1 &&
+           conv->get_data_dilation_strides() == stride_1 && conv->get_padding_above() == pad_0 &&
            conv->get_padding_below() == pad_0;
 }
 
@@ -115,6 +115,7 @@ std::shared_ptr<Node> fuse_group_convolution(const std::shared_ptr<Node>& n)
             NGRAPH_DEBUG << "convolution data's rank isn't equal to 4";
             return {nullptr};
         }
+
         if (!is_trivial_convolution(sconv))
         {
             NGRAPH_DEBUG << arg->get_name() << " isn't trivial convolution";

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -1227,6 +1227,11 @@ void runtime::cpu::CPU_ExternalFunction::register_common_passes(
             return false;
 #endif
         }
+        // GroupConvolution is only supported with MKLDNN
+        else if (auto conv = as_type<ngraph::op::GroupConvolution>(const_cast<Node*>(&node)))
+        {
+            return mkldnn_utils::can_use_mkldnn_conv<ngraph::op::GroupConvolution>(conv);
+        }
 
         if (dex)
         {

src/ngraph/runtime/cpu/unit_test.manifest

@@ -11,9 +11,6 @@ max_3d_to_scalar_int32
 send_recv
 send_recv_ring
 
-# param not supported in CPU backend
-group_conv_data_dilation
-
 # axes input param not supported
 lrn_across_h
 lrn_across_hw