[ET-VK] Only partition copy.default to Vulkan when it is a no-op (#18576)

Summary: Previously, copy.default was unconditionally added to the
Vulkan partition as
an ephemeral op (assumed to always be a no-op removed by
RemoveRedundantOpsTransform). However, copy.default can also represent a
dtype
cast, and blindly partitioning casts to Vulkan causes runtime failures
since
the Vulkan backend has no copy implementation.

Replace the unconditional registration with a custom
are_node_inputs_supported_fn that checks both dtype and shape match
between
src and dst. Only same-dtype, same-shape copies (true no-ops) are
partitioned
to Vulkan; dtype casts remain on CPU.

Authored with Claude.

Test Plan: Exported EfficientNet for Vulkan, which contains copy.default
for dtype casts,
and confirmed it no longer errors during Vulkan compilation. No-op
copies in
other models (e.g. MobileNetV3) still partition correctly.

Reviewers:

Subscribers:

Tasks:

Tags:

---

[ET-VK] Prevent hardswish decomposition by adding to ops_not_to_decompose

Summary: Without this change, aten.hardswish.default decomposes into several
arithmetic ops (add, clamp, mul, div) which prevents the Vulkan conv2d
depthwise fusion pass from recognizing and fusing the activation. Adding
hardswish to ops_not_to_decompose keeps it as a single op so the conv2d
output tile shader can apply the clamp fusion.

Also adds the conv2d_dw_output_tile_3x3_b1x1_clamp shader variant needed
for the fused path.

Authored with Claude.

Test Plan: Exported MobileNetV3 for Vulkan and verified hardswish no longer decomposes.

Reviewers:

Subscribers:

Tasks:

Tags:

---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with
[ReviewStack](https://reviewstack.dev/pytorch/executorch/pull/18576).
* __->__ #18576
* #18575

cc @manuelcandales @digantdesai @cbilgin

Author: SS-JIA

backends/vulkan/op_registry.py

@@ -169,9 +169,6 @@ def update_features_impl(op: OpKey):
         # Guard and assert ops
         torch.ops.aten._assert_scalar.default,
         torch.ops.aten.sym_constrain_range_for_size.default,
-        # copy.default is a no-op when src dtype matches dst dtype; removed by
-        # RemoveRedundantOpsTransform before execution.
-        exir_ops.edge.aten.copy.default,
     ]
 )
 def register_ephemeral_ops():
@@ -181,6 +178,31 @@ def register_ephemeral_ops():
     )
 
 
+def _check_copy_is_noop(node: torch.fx.Node) -> bool:
+    """Only support copy.default when it's a no-op (same dtype and shape)."""
+    src = node.args[1]
+    if not isinstance(src, torch.fx.Node):
+        return False
+    src_val = src.meta.get("val")
+    dst_val = node.meta.get("val")
+    if src_val is None or dst_val is None:
+        return False
+    return src_val.dtype == dst_val.dtype and src_val.shape == dst_val.shape
+
+
+@update_features(
+    [
+        exir_ops.edge.aten.copy.default,
+    ]
+)
+def register_copy_op():
+    return OpFeatures(
+        inputs_storage=utils.ANY_STORAGE,
+        supports_resize=True,
+        are_node_inputs_supported_fn=_check_copy_is_noop,
+    )
+
+
 # =============================================================================
 # UnaryOp.cpp
 # =============================================================================
@@ -193,6 +215,7 @@ def register_ephemeral_ops():
         exir_ops.edge.aten.exp.default,
         exir_ops.edge.aten.gelu.default,
         exir_ops.edge.aten.hardshrink.default,
+        exir_ops.edge.aten.hardswish.default,
         exir_ops.edge.aten.hardtanh.default,
         exir_ops.edge.aten.neg.default,
         exir_ops.edge.aten.relu.default,

backends/vulkan/partitioner/vulkan_partitioner.py

@@ -46,6 +46,7 @@
 
 # pyre-ignore
 ops_not_to_decompose = [
+    torch.ops.aten.hardswish.default,
     torch.ops.aten.upsample_nearest2d.vec,
 ]
 

backends/vulkan/runtime/graph/ops/glsl/conv2d_dw_output_tile.yaml

@@ -28,3 +28,7 @@ conv2d_dw_output_tile:
     - NAME: conv2d_dw_output_tile_3x3_b1x1
       BATCH_SIZE_X: 1
       BATCH_SIZE_Y: 1
+    - NAME: conv2d_dw_output_tile_3x3_b1x1_clamp
+      OPERATOR: clamp(X, A, B)
+      BATCH_SIZE_X: 1
+      BATCH_SIZE_Y: 1