use oneDNN GRU INT8 optimized version

use faster data format

cmake/external/mkldnn.cmake

@@ -20,7 +20,7 @@ SET(MKLDNN_SOURCE_DIR     ${THIRD_PARTY_PATH}/mkldnn/src/extern_mkldnn)
 SET(MKLDNN_INSTALL_DIR    ${THIRD_PARTY_PATH}/install/mkldnn)
 SET(MKLDNN_INC_DIR        "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE)
 SET(MKLDNN_REPOSITORY     https://github.com/oneapi-src/oneDNN.git)
-SET(MKLDNN_TAG            64a48f9565aa72f6359917b3406328075a409939)
+SET(MKLDNN_TAG            361725600224f41b7347a1c6bee9b04d1e6c14d7)
 
 # Introduce variables:
 # * CMAKE_INSTALL_LIBDIR

paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc

@@ -95,7 +95,7 @@ class GRUMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::gru_forward> {
 
       // Create memory descriptors
       auto input_md = MKLDNNMemDesc({Ti, N, IC}, MKLDNNGetDataType<T>(),
-                                    MKLDNNMemoryFormat::any);
+                                    MKLDNNMemoryFormat::ntc);
       auto weight_x_md =
           MKLDNNMemDesc({L, D, IC, G, OC}, weights_dt, MKLDNNMemoryFormat::any);
       auto weight_h_md =
@@ -103,7 +103,7 @@ class GRUMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::gru_forward> {
       auto bias_md = MKLDNNMemDesc({L, D, G, OC}, MKLDNNGetDataType<float>(),
                                    MKLDNNMemoryFormat::ldgo);
       auto hidden_md = MKLDNNMemDesc({Ti, N, OC}, MKLDNNGetDataType<T_out>(),
-                                     MKLDNNMemoryFormat::any);
+                                     MKLDNNMemoryFormat::ntc);
       auto h0_md = MKLDNNMemDesc({L, D, N, OC}, MKLDNNGetDataType<T>(),
                                  MKLDNNMemoryFormat::ldnc);