cpu: rnn: enable matmul impl back for SYCL CPU

Regular, raw CPU pointers can only be used with memory_t objects created for
the classic CPU engine. Use service engine to create such objects.

Author: densamoilov

src/cpu/rnn/ref_rnn.cpp

@@ -830,20 +830,26 @@ template <prop_kind_t aprop, data_type_t src_type, data_type_t weights_type,
 rnn_matmul_sig((_ref_rnn_common_t<aprop, src_type, weights_type,
         acc_type>::execute_matmul)) {
 
-    engine_t *engine = ctx.stream()->engine();
+    // Service engine is just a global classic CPU engine that is used
+    // when it's required to create memory_t objects for classic CPU
+    // engine regardless of the CPU runtime. For example, SYCL CPU engine
+    // cannot be used to create such objects.
+    engine_t *service_engine = get_service_engine();
     constexpr auto mem_flag = memory_flags_t::use_runtime_ptr;
 
+    // a_, b_ and c_ are regular, raw CPU pointers that can only be used with
+    // memory_t objects created for the classic CPU engine.
     std::unique_ptr<memory_t, memory_deleter_t> src_mem;
     CHECK(safe_ptr_assign(src_mem,
-            new memory_t(engine, matmul_prim->pd()->src_md(), mem_flag,
+            new memory_t(service_engine, matmul_prim->pd()->src_md(), mem_flag,
                     (void *)(a_))));
     std::unique_ptr<memory_t, memory_deleter_t> wei_mem;
     CHECK(safe_ptr_assign(wei_mem,
-            new memory_t(engine, matmul_prim->pd()->weights_md(), mem_flag,
-                    (void *)(b_))));
+            new memory_t(service_engine, matmul_prim->pd()->weights_md(),
+                    mem_flag, (void *)(b_))));
     std::unique_ptr<memory_t, memory_deleter_t> dst_mem;
     CHECK(safe_ptr_assign(dst_mem,
-            new memory_t(engine, matmul_prim->pd()->dst_md(), mem_flag,
+            new memory_t(service_engine, matmul_prim->pd()->dst_md(), mem_flag,
                     (void *)(c_))));
 
     exec_args_t matmul_args;

src/cpu/rnn/rnn_utils.hpp

@@ -858,17 +858,14 @@ bool init_conf(rnn_conf_t &rnn, const rnn_desc_t &rd,
     rnn.use_matmul = !rnn.is_brgemm && rnn.is_fwd // TODO: Enable BWD
     // TODO: Below checks are for legacy and a performance study is
     // required to avoid regressions.
-    // TODO: using matmul is disabled for SYCL runtime for now.
-    // Enable it after memory handles issue fix
 #if DNNL_X64
             && IMPLICATION(
                     rnn.is_cell_dt_bf16(), !x64::mayiuse(x64::avx512_core))
             && IMPLICATION(rnn.is_cell_dt_f32() || rnn.is_cell_dt_int8(),
                     x64::mayiuse(x64::avx2)
                             && utils::one_of(rd.cell_kind,
                                     alg_kind::vanilla_gru,
-                                    alg_kind::vanilla_augru))
-            && (DNNL_CPU_RUNTIME != DNNL_RUNTIME_SYCL);
+                                    alg_kind::vanilla_augru));
 #else
             && !rnn.is_cell_dt_f32() && !rnn.is_cell_dt_int8();
 #endif