Avoid redundant zeroing of output buffer in GatherND

src/ops/gather.rs

@@ -288,7 +288,7 @@ pub fn gather_nd<T: Clone + Default>(
     let out_slice_len = out_shape[out_shape.len() - out_slice_ndim..]
         .iter()
         .product();
-    let mut output = Tensor::<T>::zeros_in(pool, &out_shape);
+    let mut output = Tensor::<T>::uninit_in(pool, &out_shape);
 
     let output_non_batch_dims = output.ndim() - batch_dims;
     let input_non_batch_dims = input.ndim() - batch_dims;
@@ -297,6 +297,7 @@ pub fn gather_nd<T: Clone + Default>(
     // This allows the loop below to rely on index tuples being contiguous.
     let indices = indices.to_contiguous_in(pool).auto_return(pool);
 
+    let mut n_init = 0;
     for (mut output, (input, indices)) in output.inner_iter_dyn_mut(output_non_batch_dims).zip(
         input
             .inner_iter_dyn(input_non_batch_dims)
@@ -313,12 +314,15 @@ pub fn gather_nd<T: Clone + Default>(
                 .map_err(|_| OpError::InvalidValue("Invalid index"))?;
 
             for (out, x) in out_slice.iter_mut().zip(in_slice.iter()) {
-                *out = x.clone();
+                out.write(x.clone());
             }
+            n_init += out_slice.len();
         }
     }
 
-    Ok(output)
+    // Safety: All elements of `output` are initialized.
+    assert!(n_init == output.len());
+    Ok(unsafe { output.assume_init() })
 }
 
 #[derive(Debug)]