tico 0.1.0.dev250411__py3-none-any.whl

tico/passes/lower_to_resize_nearest_neighbor.py

@@ -0,0 +1,249 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    import torch.fx
+from typing import Optional
+
+import torch
+from torch.export import ExportedProgram
+
+from tico.serialize.circle_mapping import extract_shape
+from tico.utils import logging
+from tico.utils.errors import NotYetSupportedError
+from tico.utils.passes import PassBase, PassResult
+from tico.utils.trace_decorators import trace_graph_diff_on_pass
+from tico.utils.validate_args_kwargs import IndexArgs, UpsampleNearest2DVecArgs
+
+
+@trace_graph_diff_on_pass
+class LowerToResizeNearestNeighbor(PassBase):
+    """
+    This pass lowers `aten.index` and `aten.upsample_nearest2d.vec` to `circle_custom.resize_nearest_neighbor` when it is possible.
+
+    Until torch 2.7, `torch.nn.functional.interpolate` is converted to `aten.index` op.
+
+        [EXAMPLE]
+        class InterpolateDouble(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x):
+                return torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+
+            def get_example_inputs(self):
+                return (torch.randn(1, 2, 3, 4),)
+
+        [EXPORTED GRAPH]
+        [constants]
+            _prop_tensor_constant0 = tensor([0, 0, 1, 1, 2, 2, 3, 3]
+            _prop_tensor_constant1 = tensor([[0], [0], [1], [1], [2], [2]])
+
+        [graph]
+            %_prop_tensor_constant0 : [num_users=1] = placeholder[target=_prop_tensor_constant0]
+            %_prop_tensor_constant1 : [num_users=1] = placeholder[target=_prop_tensor_constant1]
+            %x : [num_users=1] = placeholder[target=x]
+            %_to_copy : [num_users=1] = call_function[target=torch.ops.aten._to_copy.default](args = (%x,), kwargs = {dtype: torch.float32})
+            %index : [num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%_to_copy, [None, None, %_prop_tensor_constant1, %_prop_tensor_constant0]), kwargs = {})
+            %_to_copy_3 : [num_users=1] = call_function[target=torch.ops.aten._to_copy.default](args = (%index,), kwargs = {dtype: torch.float32})
+            return (_to_copy_3,)
+
+        [BEFORE PASS]
+        input - aten.index - output
+
+        [AFTER PASS]
+        input - aten.permute(NCHW_to_NHWC) - circle_custom.resize_nearest_neighbor - aten.permute(NHWC_to_NCHW) - output
+
+    Since torch 2.8, `torch.nn.functional.interpolate` is converted to aten.upsample_nearest2d.vec` op.
+    """
+
+    def __init__(self):
+        super().__init__()
+
+    def convert_index_to_resize_nearest_neighbor(
+        self, exported_program, node
+    ) -> Optional[torch.fx.Node]:
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+
+        args = IndexArgs(*node.args, **node.kwargs)
+        input_tensor = args.input
+        indices = args.indices
+
+        # Only support 4-D tensor
+        if len(indices) != 4:
+            return None
+        # indices = [None, None, H index, W index]
+        N, C, H, W = indices
+        if N != None or C != None:
+            return None
+        if not isinstance(H, torch.fx.Node):
+            return None
+        if not isinstance(W, torch.fx.Node):
+            return None
+        constants_dict = exported_program.constants
+        if (H.name not in constants_dict) or (W.name not in constants_dict):
+            return None
+        H_index, W_index = constants_dict[H.name], constants_dict[W.name]
+        input_tensor_shape = extract_shape(input_tensor)
+        input_tensor_H, input_tensor_W = (
+            input_tensor_shape[2],
+            input_tensor_shape[3],
+        )
+        if H_index.size()[0] % input_tensor_H != 0:
+            return None
+        scale_factor = int(H_index.size()[0] / input_tensor_H)
+        # H and W should be resized with same ratio.
+        if scale_factor != W_index.size()[0] / input_tensor_W:
+            return None
+        expected_H_index = []
+        expected_W_index = []
+        # Please refer to above `_prop_tensor_constant1` constant in the example.
+        for i in range(input_tensor_H):
+            expected_H_index += [[i]] * scale_factor
+        # Please refer to above `_prop_tensor_constant0` constant in the example.
+        for i in range(input_tensor_W):
+            expected_W_index += [i] * scale_factor
+        if not torch.all(
+            torch.eq(H_index, torch.tensor(expected_H_index))
+        ) or not torch.all(torch.eq(W_index, torch.tensor(expected_W_index))):
+            return None
+        expected_shape = [
+            input_tensor_shape[0],
+            input_tensor_shape[1],
+            len(expected_H_index),
+            len(expected_W_index),
+        ]
+        assert expected_shape == list(extract_shape(node))
+
+        with graph.inserting_before(node):
+            nchw_to_nhwc = graph.call_function(
+                torch.ops.aten.permute.default, args=(input_tensor, [0, 2, 3, 1])
+            )
+            resize_nearest_neighbor = graph.call_function(
+                torch.ops.circle_custom.resize_nearest_neighbor,
+                args=(nchw_to_nhwc, [len(expected_H_index), len(expected_W_index)]),
+            )
+            nhwc_to_nchw = graph.call_function(
+                torch.ops.aten.permute.default,
+                args=(resize_nearest_neighbor, [0, 3, 1, 2]),
+            )
+            # Not set meta for propagating replacing node's meta.
+            node.replace_all_uses_with(nhwc_to_nchw, propagate_meta=True)
+
+        return resize_nearest_neighbor
+
+    def convert_upsample_nearest2d_to_resize_nearest_neighbor(
+        self, exported_program, node
+    ) -> Optional[torch.fx.Node]:
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+
+        args = UpsampleNearest2DVecArgs(*node.args, **node.kwargs)
+        input_tensor = args.input
+        output_size = args.output_size
+        scale_factors = args.scale_factors
+
+        input_tensor_shape = extract_shape(input_tensor)
+        input_tensor_H, input_tensor_W = (
+            input_tensor_shape[2],
+            input_tensor_shape[3],
+        )
+
+        if output_size is not None:
+            raise NotYetSupportedError("output_size is not supported yet")
+
+        if scale_factors is None:
+            raise NotYetSupportedError("scale_factors is None")
+        # TODO Support output_size case. Currently only scale_factors case is supported.
+
+        assert (
+            isinstance(scale_factors[0], float)
+            and isinstance(scale_factors[1], float)
+            and scale_factors[0] > 0
+            and scale_factors[1] > 0
+        )
+
+        def close_enough(x, y, epsilon=1e-10):
+            return abs(x - y) < epsilon
+
+        expected_H = int(input_tensor_H * scale_factors[0])
+        if not close_enough(expected_H, input_tensor_H * scale_factors[0]):
+            raise NotYetSupportedError(
+                f"Cannot support input_tensor_H ({input_tensor_H}) with scaling factor ({scale_factors[0]})"
+            )
+
+        expected_W = int(input_tensor_W * scale_factors[1])
+        if not close_enough(expected_W, input_tensor_W * scale_factors[1]):
+            raise NotYetSupportedError(
+                f"Cannot support input_tensor_W ({input_tensor_W}) with scaling factor ({scale_factors[1]})"
+            )
+
+        with graph.inserting_before(node):
+            nchw_to_nhwc = graph.call_function(
+                torch.ops.aten.permute.default, args=(input_tensor, [0, 2, 3, 1])
+            )
+            resize_nearest_neighbor = graph.call_function(
+                torch.ops.circle_custom.resize_nearest_neighbor,
+                args=(nchw_to_nhwc, [expected_H, expected_W]),
+            )
+            nhwc_to_nchw = graph.call_function(
+                torch.ops.aten.permute.default,
+                args=(resize_nearest_neighbor, [0, 3, 1, 2]),
+            )
+            # Not set meta for propagating replacing node's meta.
+            node.replace_all_uses_with(nhwc_to_nchw, propagate_meta=True)
+            return resize_nearest_neighbor
+
+    def call(self, exported_program: ExportedProgram) -> PassResult:
+        logger = logging.getLogger(__name__)
+
+        modified = False
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+        for node in graph.nodes:
+            if not node.op == "call_function":
+                continue
+
+            if node.target not in [
+                torch.ops.aten.index.Tensor,
+                torch.ops.aten.upsample_nearest2d.vec,
+            ]:
+                continue
+
+            resize_nearest_neighbor = None
+            if node.target == torch.ops.aten.index.Tensor:
+                resize_nearest_neighbor = self.convert_index_to_resize_nearest_neighbor(
+                    exported_program, node
+                )
+            elif node.target == torch.ops.aten.upsample_nearest2d.vec:
+                resize_nearest_neighbor = (
+                    self.convert_upsample_nearest2d_to_resize_nearest_neighbor(
+                        exported_program, node
+                    )
+                )
+
+            if resize_nearest_neighbor:
+                modified = True
+                logger.debug(
+                    f"{node.name} is replaced with {resize_nearest_neighbor.name} operator"
+                )
+
+        graph.eliminate_dead_code()
+        graph.lint()
+        graph_module.recompile()
+
+        return PassResult(modified)

tico/passes/lower_to_slice.py

@@ -0,0 +1,112 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    import torch.fx
+import torch
+from torch.export import ExportedProgram
+
+from tico.passes import ops
+
+from tico.serialize.circle_graph import extract_shape
+from tico.utils import logging
+from tico.utils.passes import PassBase, PassResult
+from tico.utils.trace_decorators import trace_const_diff_on_pass
+from tico.utils.validate_args_kwargs import SelectCopyIntArgs
+
+
+@trace_const_diff_on_pass
+class LowerToSlice(PassBase):
+    """
+    This pass lowers aten.ops.select/selct_copy.int to aten.ops.slice.
+    We support only when it is index in args, which is a constant tensor.
+    Since the index in node'args isn't constant tensor, we can't support converting the below op list yet.
+    - torch.ops.aten.index_select.default
+    - torch.ops.aten.embedding.default
+    - torch.ops.aten.index.Tensor
+
+    [before]
+            input (tensor, dim, *index)
+                |
+            select
+                |
+            output
+
+    [after]
+
+            input (tensor, dim, *index)
+                |
+            slice (input=tensor, dim=dim, start=index, end=index+1, step=1)
+                |
+            reshape (input=slice_copy, size=select_shape)
+                |
+            output
+    """
+
+    def __init__(self):
+        super().__init__()
+
+    def call(self, exported_program: ExportedProgram) -> PassResult:
+        logger = logging.getLogger(__name__)
+
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+        modified = False
+        for node in graph.nodes:
+            if not node.op == "call_function":
+                continue
+
+            if not node.target in ops.aten.select:
+                continue
+
+            args = SelectCopyIntArgs(*node.args, **node.kwargs)
+            input = args.input
+            dim = args.dim
+            index = args.index
+
+            input_shape = extract_shape(input)
+            if dim < 0:
+                dim = dim % len(input_shape)
+
+            start = index
+            end = index + 1
+            step = 1
+            slice_copy_args = (input, dim, start, end, step)
+
+            with graph.inserting_after(node):
+                # slice
+                slice_node = graph.call_function(
+                    torch.ops.aten.slice.Tensor, args=slice_copy_args
+                )
+                node_shape = extract_shape(node)
+            with graph.inserting_after(slice_node):
+                # reshape
+                reshape_args = (slice_node, list(node_shape))
+                reshape_node = graph.call_function(
+                    torch.ops.aten.reshape.default, args=reshape_args
+                )
+                node.replace_all_uses_with(reshape_node, propagate_meta=False)
+
+            modified = True
+            logger.debug(
+                f"{node.name} is replaced with {slice_node.name} and {reshape_node.name} operators"
+            )
+
+        graph.eliminate_dead_code()
+        graph.lint()
+        graph_module.recompile()
+
+        return PassResult(modified)

tico/passes/merge_consecutive_cat.py

@@ -0,0 +1,82 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from torch.export import ExportedProgram
+
+from tico.passes import ops
+from tico.utils import logging
+from tico.utils.passes import PassBase, PassResult
+from tico.utils.trace_decorators import trace_graph_diff_on_pass
+from tico.utils.validate_args_kwargs import CatArgs
+
+
+@trace_graph_diff_on_pass
+class MergeConsecutiveCat(PassBase):
+    """
+    This pass merges consecutive `aten.cat` operators when they can be merged into single operator.
+    """
+
+    def __init__(self):
+        super().__init__()
+
+    def call(self, exported_program: ExportedProgram) -> PassResult:
+        logger = logging.getLogger(__name__)
+
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+        modified = False
+        for cat in graph.nodes:
+            if not cat.op == "call_function":
+                continue
+
+            if not cat.target in ops.aten.cat:
+                continue
+
+            args = CatArgs(*cat.args, **cat.kwargs)  # type: ignore[arg-type]
+            inputs = args.tensors
+            dim = args.dim
+
+            new_inputs = []
+            for prev_cat in inputs:
+                new_inputs.append(prev_cat)
+                if not prev_cat.op == "call_function":
+                    continue
+
+                if not prev_cat.target in ops.aten.cat:
+                    continue
+
+                prev_args = CatArgs(*prev_cat.args, **prev_cat.kwargs)  # type: ignore[arg-type]
+                prev_inputs = prev_args.tensors
+                prev_dim = prev_args.dim
+
+                if not prev_dim == dim:
+                    continue
+
+                new_inputs.pop()
+                for prev_input in prev_inputs:
+                    new_inputs.append(prev_input)
+
+            if len(new_inputs) > len(inputs):
+                cat.args = (new_inputs, dim)
+
+                modified = True
+                logger.debug(
+                    f"Consecutive cat nodes before {cat.name} are merged into {cat.name}"
+                )
+
+        graph.eliminate_dead_code()
+        graph.lint()
+        graph_module.recompile()
+
+        return PassResult(modified)

tico/passes/ops.py

@@ -0,0 +1,75 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+
+"""
+This module contains Op lists used for finding the target Ops in passes.
+The module is introduced to reduce duplicate codes.
+It should be guaranteed that Ops in the same list have the same input/output signature.
+"""
+
+
+class AtenOps:
+    def __init__(self):
+        # In alphabetical order
+        self.add = [torch.ops.aten.add.Tensor]
+        self.alias = [torch.ops.aten.alias.default, torch.ops.aten.alias_copy.default]
+        self.cat = [torch.ops.aten.cat.default]
+        self.clamp = [torch.ops.aten.clamp.default, torch.ops.aten.clamp.Tensor]
+        self.clone = [torch.ops.aten.clone.default]
+        self.conv2d = [
+            torch.ops.aten.conv2d.default,
+            torch.ops.aten.conv2d.padding,
+        ]
+        self.conv1d = [
+            torch.ops.aten.conv1d.default,
+            torch.ops.aten.conv1d.padding,
+        ]
+        self.detach = [
+            torch.ops.aten.detach_.default,
+            torch.ops.aten.detach.default,
+        ]
+        self.expand = [
+            torch.ops.aten.expand.default,
+            torch.ops.aten.expand_copy.default,
+        ]
+        self.index_select = [torch.ops.aten.index_select.default]
+        self.mean = [torch.ops.aten.mean.dim]
+        self.mul_scalar = [torch.ops.aten.mul.Scalar]
+        self.mul_tensor = [torch.ops.aten.mul.Tensor]
+        self.permute = [torch.ops.aten.permute.default]
+        self.reshape = [torch.ops.aten.reshape.default]
+        self.select = [torch.ops.aten.select_copy.int, torch.ops.aten.select.int]
+        self.slice = [torch.ops.aten.slice.Tensor, torch.ops.aten.slice_copy.Tensor]
+        self.softmax = [torch.ops.aten._softmax.default]
+        self.squeeze = [torch.ops.aten.squeeze.dims, torch.ops.aten.squeeze_copy.dims]
+        self.to_copy = [
+            torch.ops.aten._to_copy.default,
+            torch.ops.aten.to.dtype,
+            torch.ops.aten.to.dtype_layout,
+        ]
+        self.unsqueeze = [
+            torch.ops.aten.unsqueeze.default,
+            torch.ops.aten.unsqueeze_copy.default,
+        ]
+        self.view = [
+            torch.ops.aten.view,
+            torch.ops.aten.view.default,
+            torch.ops.aten.view_copy.default,
+        ]
+
+
+aten = AtenOps()

tico/passes/remove_nop.py

@@ -0,0 +1,85 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    import torch.fx
+import torch
+from torch.export import ExportedProgram
+
+from tico.passes import ops
+from tico.utils import logging
+from tico.utils.passes import PassBase, PassResult
+from tico.utils.trace_decorators import trace_graph_diff_on_pass
+
+
+@trace_graph_diff_on_pass
+class RemoveNop(PassBase):
+    """
+    Let's remove noops by propagation.
+    """
+
+    target_ops = (
+        [
+            torch.ops.prims.view_of.default,
+        ]
+        + ops.aten.alias
+        + ops.aten.clone
+        + ops.aten.detach
+        + [torch.ops.aten.lift_fresh_copy.default]
+    )
+
+    def __init__(self):
+        super().__init__()
+
+    def call(self, exported_program: ExportedProgram) -> PassResult:
+        logger = logging.getLogger(__name__)
+
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+        modified = False
+        for node in graph.nodes:
+            if not node.op == "call_function":
+                continue
+
+            if not node.target in RemoveNop.target_ops:
+                continue
+            # TODO Consider memory format
+            if node.target in ops.aten.clone and "memory_format" in node.kwargs:
+                if node.kwargs["memory_format"] not in [
+                    torch.preserve_format,
+                    # Converting non-contiguous layout to contiguous only updates
+                    # strides of tensor. This is not visible on circle, so we can
+                    # safely ignore this operation.
+                    torch.contiguous_format,
+                ]:
+                    continue
+
+            assert len(node.args) == 1
+
+            src = node.args[0]
+            assert isinstance(src, torch.fx.Node)
+
+            with graph.inserting_after(node):
+                node.replace_all_uses_with(src, propagate_meta=False)
+
+            modified = True
+            logger.debug(f"{node.name} is replaced with {src}")
+
+        graph.eliminate_dead_code()
+        graph.lint()
+        graph_module.recompile()
+
+        return PassResult(modified)

tico/passes/remove_redundant_assert_nodes.py

@@ -0,0 +1,50 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from torch.export import ExportedProgram
+
+from tico.utils.passes import PassBase, PassResult
+from tico.utils.trace_decorators import trace_graph_diff_on_pass
+
+
+assert_node_targets = [
+    torch.ops.aten._assert_tensor_metadata.default,
+]
+
+
+@trace_graph_diff_on_pass
+class RemoveRedundantAssertionNodes(PassBase):
+    """
+    This removes redundant assertion nodes.
+    - `aten.assert_tensor_meta.default`
+    """
+
+    def __init__(self):
+        super().__init__()
+
+    def call(self, exported_program: ExportedProgram) -> PassResult:
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+        modified = False
+        for node in graph.nodes:
+            if node.op == "call_function" and node.target in assert_node_targets:
+                graph.erase_node(node)
+                modified = True
+
+        graph.eliminate_dead_code()
+        graph.lint()
+        graph_module.recompile()
+
+        return PassResult(modified)