benchdnn: graph: skip unimplemented GenIndex op on gpu

tests/benchdnn/graph/graph.cpp

@@ -404,7 +404,10 @@ void skip_unimplemented_ops(const dnnl::graph::partition &partition,
 
     // A list of ops that don't have DNNL backend support so far.
     static const std::vector<std::string> unimplemented_ops {"Pow"};
-
+    // A list of ops that don't have DNNL backend support so far on GPU.
+    static const std::vector<std::string> unimplemented_ops_gpu {"GenIndex"};
+    const auto &eng = get_graph_engine();
+    bool is_gpu = eng.get_kind() == dnnl::engine::kind::gpu;
     // For an unsupported partition, retrieve all operation IDs, find a
     // correspondent operation kind in a deserialized_graph and match it against
     // a list of known unsupported ops.
@@ -423,6 +426,21 @@ void skip_unimplemented_ops(const dnnl::graph::partition &partition,
             res->reason = skip_reason::case_not_supported;
             return;
         }
+
+        if (is_gpu) {
+            const bool has_unimplemented_op_gpu = std::any_of(
+                    unimplemented_ops_gpu.begin(), unimplemented_ops_gpu.end(),
+                    [&dg_op_kind](const std::string &kind) {
+                        return dg_op_kind == kind;
+                    });
+            if (has_unimplemented_op_gpu) {
+                BENCHDNN_PRINT(2, "[INFO]: Unimplemented op on GPU: %s.\n",
+                        dg_op_kind.c_str());
+                res->state = SKIPPED;
+                res->reason = skip_reason::case_not_supported;
+                return;
+            }
+        }
     }
 }