Do not fuse nodes if one node is predecessor of another node in horiz… (#1928)

* Do not fuse nodes if one node is predecessor of another node in horizontal fusion.

* Add dead node check and remove predecessor check in horizontal fusion.

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -1040,7 +1040,7 @@ void runtime::cpu::CPU_ExternalFunction::register_common_passes(ngraph::pass::Ma
 
     pass_manager.register_pass<ngraph::pass::CoreFusion>();
     pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
-    // pass_manager.register_pass<runtime::cpu::pass::CPUHorizontalFusion>();
+    pass_manager.register_pass<runtime::cpu::pass::CPUHorizontalFusion>();
     pass_manager.register_pass<runtime::cpu::pass::CPUCollapseDims>();
 #if defined(NGRAPH_HALIDE)
     pass_manager.register_pass<ngraph::runtime::cpu::pass::HalideSubgraphExtraction>();

src/ngraph/runtime/cpu/pass/cpu_horizontal_fusion.cpp

@@ -61,6 +61,7 @@ void ngraph::runtime::cpu::pass::CPUHorizontalFusion::cpu_conv_horizontal_fusion
         //check if the node has been replaced
         if (conv_bias_root->get_users().empty())
         {
+            NGRAPH_DEBUG << "conv_horizontal_fusion: root node has been replaced\n";
             return false;
         }
 
@@ -75,8 +76,14 @@ void ngraph::runtime::cpu::pass::CPUHorizontalFusion::cpu_conv_horizontal_fusion
 
         for (auto u : m.get_pattern_map()[data_conv]->get_users())
         {
+            if (!is_used(u.get()))
+            {
+                NGRAPH_DEBUG << "conv_horizontal_fusion: dead node\n";
+                continue;
+            }
             if (!pattern::has_class<ngraph::op::ConvolutionBias>()(u))
             {
+                NGRAPH_DEBUG << "conv_horizontal_fusion: not conv_bias node\n";
                 continue;
             }
             if (u->get_argument(0) != m.get_pattern_map()[data_conv])
@@ -96,6 +103,7 @@ void ngraph::runtime::cpu::pass::CPUHorizontalFusion::cpu_conv_horizontal_fusion
             bias_nodes.push_back(u->get_argument(2));
             conv_bias_nodes.push_back(u);
         }
+
         if (conv_bias_nodes.size() <= 1)
         {
             NGRAPH_DEBUG << "conv_horizontal_fusion: need more than one nodes to do fusion\n";

test/cpu_fusion.cpp

@@ -891,7 +891,6 @@ TEST(cpu_fusion, conv_bias_relu_n2c1h2w2_2)
     EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));
 }
 
-#if 0
 TEST(cpu_fusion, conv_horizontal_fusion)
 {
     Shape shape_a{2, 1, 6, 6};
@@ -942,7 +941,6 @@ TEST(cpu_fusion, conv_horizontal_fusion)
     size_t cpu_cb = count_ops_of_type<op::ConvolutionBias>(cpu_f);
     ASSERT_EQ(cpu_cb, 1);
 }
-#endif
 
 // ConvolutionBiasAdd relies on an in-place fused MKLDNN kernel.
 // Need to ensure that it is fused only when in-place buffer allocation is feasible