Arm backend: Do not partition view_copy

backends/arm/test/ops/test_mean_dim.py

@@ -356,14 +356,12 @@ def forward(self, tensor: torch.Tensor, keepdim: bool):
         return tensor.mean()
 
     test_data_suite: dict[str, Callable[[], mean_input_t]] = {
-        "rank1": lambda: (
-            torch.rand(
-                1,
-            ),
+        "rank_2": lambda: (
+            torch.rand(1, 2),
             False,
         ),
-        "rank2": lambda: (torch.rand(5, 5), True),
-        "rank4": lambda: (torch.rand(5, 1, 10, 1), False),
+        "rank_2_keepdim": lambda: (torch.rand(5, 5), True),
+        "rank_4": lambda: (torch.rand(5, 1, 10, 1), False),
     }
 
 

backends/arm/test/ops/test_sum.py

@@ -66,7 +66,7 @@ def test_sum_dim_intlist_tosa_INT(test_data: input_t1):
 
 @common.parametrize("test_data", Sum.test_parameters)
 @common.XfailIfNoCorstone300
-def test_view_u55_INT_1_0(test_data: Tuple):
+def test_sum_u55_INT_1_0(test_data: Tuple):
     pipeline = EthosU55PipelineINT[input_t1](
         Sum(),
         test_data(),
@@ -78,7 +78,7 @@ def test_view_u55_INT_1_0(test_data: Tuple):
 
 @common.parametrize("test_data", Sum.test_parameters)
 @common.XfailIfNoCorstone320
-def test_view_u85_INT_1_0(test_data: Tuple):
+def test_sum_u85_INT_1_0(test_data: Tuple):
     pipeline = EthosU85PipelineINT[input_t1](
         Sum(),
         test_data(),
@@ -122,7 +122,7 @@ def test_sum_dim_intlist_vgf_INT(test_data: input_t1):
 
 
 @common.parametrize("test_data", reject_inputs)
-def test_view_u55_INT_failure_set(test_data: Tuple):
+def test_sum_u55_INT_failure_set(test_data: Tuple):
     pipeline = EthosU55PipelineINT[input_t1](
         Sum(),
         test_data(),

backends/arm/test/ops/test_unflatten.py

@@ -29,7 +29,10 @@ def __init__(self, dim: int, sizes: Tuple[int, ...]):
         self.sizes = sizes
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
-        return torch.unflatten(x, self.dim, self.sizes)
+        unflatten_op = torch.unflatten(x, self.dim, self.sizes)
+        # Because we treat a single view as a no compute operation and therefore do not partition it,
+        # we want to provide a mul op to verify that it does indeed get partitioned when bundled with another op.
+        return unflatten_op * unflatten_op
 
     test_data: dict[str, test_data_t] = {
         "rand_3d_batch3": (lambda: (Unflatten(1, (-1, 2)), (torch.rand(3, 4, 4),))),

backends/arm/test/ops/test_view.py

@@ -65,7 +65,10 @@ def __init__(self, new_shape):
         self.new_shape = new_shape
 
     def forward(self, x: torch.Tensor):
-        return x.view(self.new_shape)
+        view_op = x.view(self.new_shape)
+        # Because we treat a single view as a no compute operation and therefore do not partition it,
+        # we want to provide a mul op to verify that it does indeed get partitioned when bundled with another op.
+        return view_op * view_op
 
 
 @common.parametrize("test_data", View.needs_transpose_tests)
@@ -139,7 +142,7 @@ def test_view_u55_INT_not_delegated(test_data: Tuple):
         View(new_shape),
         (test_tensor,),
         {"executorch_exir_dialects_edge__ops_aten_view_copy": 1},
-        n_expected_delegates=0,
+        n_expected_delegates=1,
         quantize=True,
         u55_subset=True,
     )

backends/arm/tosa/partitioner.py

@@ -115,6 +115,22 @@ def is_noop_expand(node: torch.fx.node.Node) -> bool:
     return all(m == 1 for m in multiples) and not changes_rank
 
 
+def is_view_copy(node: torch.fx.node.Node) -> bool:
+    """Return True if node is a ``view_copy``.
+
+    view_copy can be regarded as a no-compute op.
+
+    Args:
+        node (torch.fx.Node): FX node to inspect.
+
+    Returns:
+        bool: True if the node targets ``aten.view_copy.default``; otherwise,
+        False.
+
+    """
+    return node.target == exir_ops.edge.aten.view_copy.default
+
+
 def is_partitioned(
     node: torch.fx.Node,
     tag: str,
@@ -267,16 +283,18 @@ def _tag_module(  # noqa
                             del node.meta["delegation_tag"]
                             break
 
-            is_noop_partition = all(
+            # Check whether the partition contains only no-op or non-computational ops. Such partitions don't make sense to delegate, and in the worst case may be optimized away during lowering, which can break compilation."
+            is_nocompute_partition = all(
                 is_noop_clone(node)
                 or is_noop_alias_copy(node)
                 or is_noop_expand(node)
                 or is_noop_to_dim_order_copy(node)
+                or is_view_copy(node)
                 or node.target in Q_OPS
                 or node.target in DQ_OPS
                 for node in partition.nodes
             )
-            if is_noop_partition:
+            if is_nocompute_partition:
                 reject_partition(
                     "Partition contained only ops which are removed in the TOSA lowering, leading to an empty partition.",
                     partition,