Copy hidden gate in GRU op before applying activation

This is a workaround needed because `tanh_in_place` is very slow with
non-contigous inputs. See #192.

src/ops/rnn.rs

@@ -7,8 +7,8 @@ use rten_tensor::{Tensor, TensorView};
 use crate::check_dims;
 use crate::gemm::{GemmExecutor, GemmInputA, GemmInputB};
 use crate::ops::{
-    add_in_place, mul_in_place, sigmoid_in_place, tanh_in_place, InputList, IntoOpResult, OpError,
-    Operator, Output,
+    add_in_place, mul_in_place, sigmoid_in_place, tanh, tanh_in_place, InputList, IntoOpResult,
+    OpError, Operator, Output,
 };
 use crate::tensor_pool::{AutoReturn, TensorPool};
 
@@ -264,6 +264,9 @@ pub fn gru(
                 update_reset_gates.as_dyn_mut(),
                 hidden_scratch_reset_update_gates.as_dyn(),
             );
+
+            // nb. This is slower than it should be because it falls back to
+            // the slow path for non-contiguous tensors.
             sigmoid_in_place(update_reset_gates.as_dyn_mut());
 
             // Combine inputs for hidden gate and apply activation.
@@ -274,11 +277,13 @@ pub fn gru(
 
             let mut hidden_gate = gates.slice_mut::<2, _>((.., gate_range(HIDDEN_GATE)));
             add_in_place(hidden_gate.as_dyn_mut(), hidden_gate_recurrent.as_dyn());
-            tanh_in_place(hidden_gate.as_dyn_mut());
+
+            // Copy the hidden gate because `tanh_in_place` is slow with
+            // non-contiguous tensors.
+            let hidden_gate = tanh(pool, hidden_gate.as_dyn()).auto_return(pool);
 
             // Compute next hidden state
             let mut hidden_item = hidden.slice_mut::<2, _>([dir]);
-            let hidden_gate = gates.slice::<2, _>((.., gate_range(HIDDEN_GATE)));
             let update_gate = gates.slice::<2, _>((.., gate_range(UPDATE_GATE)));
 
             for (hidden, update, hidden_gate) in zip3(