all: rnn: align computing (per-oc) mask with the library

The i-th bit in the mask corresponds to the i-th dimension.
The dimensions are enumerated from the outermost one, aka C-array style.

Author: Fomenko, Evarist M

examples/cpu_rnn_inference_int8.cpp

@@ -237,7 +237,9 @@ void simple_net() {
     ///
     const float data_shift = 64.;
     const float data_scale = 63.;
-    const int weights_scale_mask = 3; // 11 for last two dimensions of ldigo
+    const int weights_scale_mask = 0
+            + (1 << 3) // bit, indicating the unique scales for `g` dim in `ldigo`
+            + (1 << 4); // bit, indicating the unique scales for `o` dim in `ldigo`
     //[quantize]
     std::vector<float> weights_scales(lstm_n_gates * feature_size);
     // assign halves of vector with arbitrary values

src/cpu/rnn/rnn_reorders.hpp

@@ -125,7 +125,7 @@ struct rnn_weights_reorder_t : public primitive_impl_t {
             if (itag == format_tag::undef) return invalid_arguments;
 
             const int mask = attr->rnn_weights_qparams_.mask_;
-            if (!utils::one_of(mask, 0, 3)) return unimplemented;
+            if (!utils::one_of(mask, 0, 24)) return unimplemented;
 
             auto _pd = new pd_t(
                     engine, attr, src_engine, src_md, dst_engine, dst_md);

tests/benchdnn/rnn/rnn.cpp

@@ -56,7 +56,7 @@ void create_dnnl_rnn_attr(const prb_t &p, dnnl_primitive_attr_t *dnnl_attr) {
 
     if (p.scale_policy == policy_t::PER_OC) {
         DNN_SAFE_V(dnnl_primitive_attr_set_rnn_weights_qparams(
-                *dnnl_attr, p.dic * p.n_gates(), 0x3, p.wei_oc_scales));
+                *dnnl_attr, p.dic * p.n_gates(), 0x18, p.wei_oc_scales));
     } else if (p.scale_policy == policy_t::COMMON && p.wei_scale != 1.) {
         DNN_SAFE_V(dnnl_primitive_attr_set_rnn_weights_qparams(
                 *dnnl_attr, 1, 0, &p.wei_scale));