Optimize IPU Jacobil tile_gather copies.

Splitting the `tile_gather` into 2 parts helps limiting the number
of on-tile copies introduce by the Poplar compiler.

tessellate_ipu/linalg/tile_linalg_jacobi.py

@@ -190,10 +190,16 @@ def tile_rotate_columns(pcols: TileShardedArray, qcols: TileShardedArray) -> Tup
     # Roughtly: pcols move to the right, qcols to the left.
     pcols_indices_new = np.concatenate([pcols_indices[0:1], qcols_indices[0:1], pcols_indices[1:-1]])
     qcols_indices_new = np.concatenate([qcols_indices[1:], pcols_indices[-1:]])
+
     # Move columns around!
-    all_indices = np.concatenate([pcols_indices_new, qcols_indices_new])
-    all_cols_updated = tile_gather(all_cols, all_indices.tolist(), all_cols.tiles)
-    return all_cols_updated[:halfN], all_cols_updated[halfN:]
+    pcols_updated = tile_gather(all_cols, pcols_indices_new.tolist(), pcols.tiles)
+    qcols_updated = tile_gather(all_cols, qcols_indices_new.tolist(), qcols.tiles)
+    return pcols_updated, qcols_updated
+
+    # FIXME: understand why Poplar add a copy with the following code.
+    # all_indices = np.concatenate([pcols_indices_new, qcols_indices_new])
+    # all_cols_updated = tile_gather(all_cols, all_indices.tolist(), all_cols.tiles)
+    # return all_cols_updated[:halfN], all_cols_updated[halfN:]
 
 
 def ipu_jacobi_eigh_body(idx: Array, inputs: Tuple[TileShardedArray, ...]) -> Tuple[TileShardedArray, ...]:
@@ -211,6 +217,12 @@ def ipu_jacobi_eigh_body(idx: Array, inputs: Tuple[TileShardedArray, ...]) -> Tu
     halfN = Apcols.shape[0]
 
     with jax.named_scope("jacobi_eigh"):
+        # with jax.named_scope("Apqcols_rotation"):
+        #     Apcols, Aqcols = tile_rotate_columns(Apcols, Aqcols)
+        # with jax.named_scope("Vpqcols_rotation"):
+        #     Vpcols, Vqcols = tile_rotate_columns(Vpcols, Vqcols)
+        # Apcols, Aqcols, Vpcols, Vqcols = tile_data_barrier(Apcols, Aqcols, Vpcols, Vqcols)
+
         # Sharded constant with p/q indices to ignore in second update stage.
         with jax.named_scope("rotset_index_ignored"):
             rotset_index_ignored = tile_constant_sharded(np.arange(0, halfN, dtype=np.uint32), tiles=Atiles)