[XNNPACK] Lower depthwise transposed convs as polyphase depthwise convs + depth_to_space

backends/transforms/test/test_create_delete_constant_placeholder.py

@@ -13,6 +13,8 @@
     duplicate_constant_node,
 )
 from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import PassResult
+from executorch.exir.pass_manager import ExportedProgramPassManager
 from torch.export import export
 from torch.export.graph_signature import InputKind
 
@@ -247,3 +249,59 @@ def test_create_delete_non_persistent_buffer():
 
 def test_create_delete_constant_tensor():
     _test_create_delete(InputKind.CONSTANT_TENSOR)
+
+
+def _test_delete_inside_a_pass(kind: InputKind, persistent_buffer: bool = None):
+    """Backend passes keep a reference to the program they were built from,
+    while ExportedProgramPassManager runs them on a shallow copy of it. The
+    deletion has to edit the shared signature in place, or the copy that the
+    manager verifies and returns still lists the deleted placeholder."""
+    module = EmptyNetwork()
+    exported_program = export(module, args=module.test_data, strict=True)
+    exported_program = to_edge(exported_program).exported_program()
+    graph = exported_program.graph_module.graph
+
+    const_name = "test_node"
+    input_node = list(graph.nodes)[0]
+    with graph.inserting_before(input_node):
+        const_node = create_constant_placeholder(
+            exp_program=exported_program,
+            graph=graph,
+            kind=kind,
+            name=const_name,
+            data=torch.ones(1),
+            persistent_buffer=persistent_buffer,
+        )
+    assert len(exported_program.graph_signature.input_specs) == 2
+
+    def delete_unused_constant(graph_module: torch.fx.GraphModule) -> PassResult:
+        delete_constant_placeholder(exported_program, const_node)
+        return PassResult(graph_module, True)
+
+    result = ExportedProgramPassManager([delete_unused_constant])(exported_program)
+    transformed = result.exported_program
+    signature = transformed.graph_signature
+
+    assert len(signature.input_specs) == 1
+    assert const_name not in signature.inputs_to_parameters
+    assert const_name not in signature.inputs_to_buffers
+    assert const_name not in signature.inputs_to_lifted_tensor_constants
+    assert len(transformed.state_dict) == 0
+    assert len(transformed.constants) == 0
+    assert transformed.module()(torch.zeros(1)) == 0
+
+
+def test_delete_inside_a_pass_parameter():
+    _test_delete_inside_a_pass(InputKind.PARAMETER)
+
+
+def test_delete_inside_a_pass_persistent_buffer():
+    _test_delete_inside_a_pass(InputKind.BUFFER, True)
+
+
+def test_delete_inside_a_pass_non_persistent_buffer():
+    _test_delete_inside_a_pass(InputKind.BUFFER, False)
+
+
+def test_delete_inside_a_pass_constant_tensor():
+    _test_delete_inside_a_pass(InputKind.CONSTANT_TENSOR)

backends/transforms/utils.py

@@ -266,16 +266,10 @@ def delete_constant_placeholder(exp_program: ExportedProgram, node: torch.fx.Nod
             f"Cannot delete input node {node.name} since it is not a parameter, a buffer, nor a lifted tensor constant."
         )
 
-    # Remove input from graph signature
-    input_specs = [
-        spec
-        for spec in exp_program.graph_signature.input_specs
-        if spec.arg.name != node.name
-    ]
-    new_graph_signature = ExportGraphSignature(
-        input_specs, exp_program.graph_signature.output_specs
-    )
-    exp_program._graph_signature = new_graph_signature
+    # Edit the spec list in place: ExportedProgramPassManager runs passes on a
+    # shallow copy of the program, which would not see a replaced signature.
+    input_specs = exp_program.graph_signature.input_specs
+    input_specs[:] = [spec for spec in input_specs if spec.arg.name != node.name]
 
     # Copies made by duplicate_constant_node keep the original target, so keep
     # the tensor while another placeholder still refers to it.

backends/xnnpack/BUCK

@@ -45,3 +45,14 @@ fbcode_target(_kind = runtime.python_library,
         "//executorch/backends/xnnpack/utils:xnnpack_utils",
     ],
 )
+
+fbcode_target(_kind = runtime.python_library,
+    name = "custom_ops_lib",
+    srcs = [
+        "custom_ops_lib.py",
+    ],
+    visibility = ["PUBLIC"],
+    deps = [
+        "//caffe2:torch",
+    ],
+)

backends/xnnpack/_passes/BUCK

@@ -11,6 +11,7 @@ fbcode_target(_kind = runtime.python_library,
         "//executorch/backends/transforms:addmm_mm_to_linear",
         "//executorch/backends/transforms:remove_clone_ops",
         "//executorch/backends/transforms:lib",
+        "//executorch/backends/xnnpack:custom_ops_lib",
         "//executorch/backends/xnnpack/partition:partitioner_graphs",
         "//executorch/backends/xnnpack/serialization:xnnpack_schema",
         "//executorch/backends/xnnpack/utils:xnnpack_utils",

backends/xnnpack/_passes/__init__.py

@@ -26,6 +26,9 @@
 )
 from executorch.backends.xnnpack._passes.decompose_batch_norm import DecomposeBatchNorm
 from executorch.backends.xnnpack._passes.decompose_cat import DecomposeConcatenate
+from executorch.backends.xnnpack._passes.decompose_depthwise_conv_transpose import (
+    DecomposeDepthwiseConvTransposePass,
+)
 from executorch.backends.xnnpack._passes.fuse_activation_pass import FuseActivationPass
 from executorch.backends.xnnpack._passes.fuse_batch_norm import FuseBatchNormPass
 from executorch.backends.xnnpack._passes.fuse_pad_into_conv_pass import (
@@ -85,6 +88,7 @@ def __init__(
                 ConstPropPass,
                 FuseBatchNormPass,
                 DecomposeBatchNorm,
+                DecomposeDepthwiseConvTransposePass,
                 FuseActivationPass,
                 DecomposeConcatenate,
                 RemoveGetItemPass,

backends/xnnpack/_passes/channels_last_tagged_reshape_pass.py

@@ -9,6 +9,7 @@
 from enum import Enum
 from typing import List, Optional, Tuple
 
+import executorch.backends.xnnpack.custom_ops_lib  # noqa: F401  registers et_xnnpack ops
 import torch
 from executorch.backends.xnnpack._passes.xnnpack_pass import XNNPACKPass
 from executorch.backends.xnnpack.utils.quant_utils import (
@@ -62,6 +63,7 @@ class ChannelsLastTaggedReshapePass(XNNPACKPass):
         exir_ops.edge.aten.convolution.default,
         exir_ops.edge.aten.upsample_bilinear2d.vec,
         exir_ops.edge.aten.upsample_nearest2d.vec,
+        exir_ops.edge.et_xnnpack.depth_to_space_2d.default,
         exir_ops.edge.aten.mean.dim,
         exir_ops.edge.aten.max_pool2d.default,
         exir_ops.edge.aten.amax.default,

backends/xnnpack/_passes/decompose_depthwise_conv_transpose.py

@@ -0,0 +1,230 @@
+# Copyright (c) 2026 Software Mansion
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from typing import cast, List, Optional, Tuple
+
+import executorch.backends.xnnpack.custom_ops_lib  # noqa: F401  registers et_xnnpack ops
+import torch
+from executorch.backends.transforms.utils import (
+    create_constant_placeholder,
+    delete_constant_placeholder,
+    get_param_tensor,
+    is_param_node,
+)
+from executorch.backends.xnnpack._passes.xnnpack_pass import XNNPACKPass
+from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import PassResult
+
+
+class DecomposeDepthwiseConvTransposePass(XNNPACKPass):
+    """
+    Lower a depthwise transposed conv (groups == in_channels == out_channels)
+    with stride s > 1 into s*s depthwise convs on the low-resolution input, a
+    channel concatenation and a block-major depth_to_space_2d.
+
+    XNNPACK runs transposed convs through its subconvolution IGEMM path, one
+    group at a time; with one channel per group that is far slower than its
+    depthwise conv microkernels. Polyphase decomposition: output phase a of a
+    stride-s transposed conv is an ordinary stride-1 correlation of the input
+    with every s-th tap of the kernel,
+
+        y[s*m + a] = sum_t x[m - t] * w[s*t + a + p],   0 <= s*t + a + p < k,
+
+    per axis. The s*s phase outputs concatenated along channels in (ay, ax)
+    order are exactly the block-major input depth_to_space_2d expects. Each
+    phase reads the input through a (possibly asymmetric) zero pad.
+
+    Only applied when the transposed conv is a clean x s upscale, i.e.
+    (H - 1)*s - 2p + k + output_padding == s*H on both axes with the same s,
+    dilation is 1 and every phase's taps stay inside the input; any other
+    transposed conv is left to XNNPACK's own node.
+    """
+
+    def call(self, graph_module: torch.fx.GraphModule):
+        graph = graph_module.graph
+        # Plan every conv before rewriting any: the nodes created by
+        # _decompose have no meta["val"] until the retrace below.
+        plans = []
+        for node in graph.nodes:
+            if (
+                node.op == "call_function"
+                and node.target == exir_ops.edge.aten.convolution.default
+            ):
+                plan = self._plan(node)
+                if plan is not None:
+                    plans.append((node, plan))
+        for node, plan in plans:
+            self._decompose(graph, node, *plan)
+
+        modified = bool(plans)
+        if modified:
+            graph_module.recompile()
+            graph_module = super().call(graph_module).graph_module
+        return PassResult(graph_module, modified)
+
+    # (t_min, t_max) per phase for one axis, None when the axis
+    # does not decompose.
+    @staticmethod
+    def _axis_phases(
+        size: int, kernel: int, stride: int, padding: int, output_padding: int
+    ) -> Optional[List[Tuple[int, int]]]:
+        if stride < 2:
+            return None
+        if (size - 1) * stride - 2 * padding + kernel + output_padding != size * stride:
+            return None
+        phases = []
+        for a in range(stride):
+            # integer t with 0 <= stride*t + a + padding <= kernel - 1
+            t_min = math.ceil(-(a + padding) / stride)
+            t_max = math.floor((kernel - 1 - a - padding) / stride)
+            if t_min > 0 or t_max < 0:
+                return None
+            phases.append((t_min, t_max))
+        return phases
+
+    def _plan(self, node: torch.fx.Node):
+        (
+            x,
+            weight,
+            _bias,
+            stride,
+            padding,
+            dilation,
+            transposed,
+            output_padding,
+            groups,
+        ) = node.args
+        if not transposed:
+            return None
+        if not is_param_node(self.exported_program, weight):
+            return None
+        x_shape = list(x.meta["val"].shape)
+        w_shape = list(weight.meta["val"].shape)
+        if len(x_shape) != 4 or len(w_shape) != 4:
+            return None
+        channels = x_shape[1]
+        # transposed conv weight is (in, out / groups, kh, kw)
+        if groups != channels or w_shape[0] != channels or w_shape[1] != 1:
+            return None
+        if any(d != 1 for d in dilation):
+            return None
+        # ATen broadcasts a single spatial value over both spatial dims.
+        stride, padding, output_padding = (
+            p * 2 if len(p) == 1 else p
+            for p in cast(List[List[int]], [stride, padding, output_padding])
+        )
+        if stride[0] != stride[1]:
+            return None
+        phases_y = self._axis_phases(
+            x_shape[2], w_shape[2], stride[0], padding[0], output_padding[0]
+        )
+        phases_x = self._axis_phases(
+            x_shape[3], w_shape[3], stride[1], padding[1], output_padding[1]
+        )
+        if phases_y is None or phases_x is None:
+            return None
+        return stride[0], padding, phases_y, phases_x
+
+    def _decompose(
+        self,
+        graph: torch.fx.Graph,
+        node: torch.fx.Node,
+        s: int,
+        padding: List[int],
+        phases_y: List[Tuple[int, int]],
+        phases_x: List[Tuple[int, int]],
+    ) -> None:
+        x, weight_node, bias_node = node.args[0], node.args[1], node.args[2]
+        groups = node.args[8]
+        weight = get_param_tensor(self.exported_program, weight_node)
+        assert weight is not None
+        weight_spec = next(
+            spec
+            for spec in self.exported_program.graph_signature.input_specs
+            if spec.arg.name == weight_node.name
+        )
+
+        phase_convs = []
+        for ay, (ty_min, ty_max) in enumerate(phases_y):
+            for ax, (tx_min, tx_max) in enumerate(phases_x):
+                # taps of phase (ay, ax) as a stride-1 correlation kernel:
+                # W[q] = w[s * (t_max - q) + a + p]
+                iy = torch.tensor(
+                    [
+                        s * (ty_max - q) + ay + padding[0]
+                        for q in range(ty_max - ty_min + 1)
+                    ]
+                )
+                ix = torch.tensor(
+                    [
+                        s * (tx_max - q) + ax + padding[1]
+                        for q in range(tx_max - tx_min + 1)
+                    ]
+                )
+                phase_weight = weight[:, :, iy][:, :, :, ix].contiguous()
+                # The phase taps depend on s and padding too, so a weight
+                # shared by differently configured deconvs must not reuse
+                # the placeholders create_constant_placeholder finds by name.
+                phase_name = (
+                    f"{weight_node.name}_s{s}_p{padding[0]}x{padding[1]}"
+                    f"_phase_{ay}_{ax}"
+                )
+                with graph.inserting_before(weight_node):
+                    phase_weight_node = create_constant_placeholder(
+                        exp_program=self.exported_program,
+                        graph=graph,
+                        kind=weight_spec.kind,
+                        persistent_buffer=weight_spec.persistent,
+                        name=phase_name,
+                        data=phase_weight,
+                    )
+                # create_constant_placeholder only sets meta["val"]; keep the
+                # external-constant tag so the phase weights follow the
+                # original weight into its .ptd file.
+                if "custom" in weight_node.meta:
+                    phase_weight_node.meta["custom"] = dict(weight_node.meta["custom"])
+                # (left, right, top, bottom)
+                pad = [tx_max, -tx_min, ty_max, -ty_min]
+                with graph.inserting_before(node):
+                    conv_input = x
+                    if any(pad):
+                        conv_input = graph.create_node(
+                            "call_function",
+                            exir_ops.edge.aten.constant_pad_nd.default,
+                            (x, pad, 0.0),
+                        )
+                    phase_convs.append(
+                        graph.create_node(
+                            "call_function",
+                            exir_ops.edge.aten.convolution.default,
+                            (
+                                conv_input,
+                                phase_weight_node,
+                                bias_node,
+                                [1, 1],
+                                [0, 0],
+                                [1, 1],
+                                False,
+                                [0, 0],
+                                groups,
+                            ),
+                        )
+                    )
+
+        with graph.inserting_before(node):
+            cat = graph.create_node(
+                "call_function", exir_ops.edge.aten.cat.default, (phase_convs, 1)
+            )
+            d2s = graph.create_node(
+                "call_function",
+                exir_ops.edge.et_xnnpack.depth_to_space_2d.default,
+                (cat, s),
+            )
+        node.replace_all_uses_with(d2s)
+        graph.erase_node(node)
+        if len(weight_node.users) == 0:
+            delete_constant_placeholder(self.exported_program, weight_node)