Optimize aten::min/max.dim with TopK op

onnxscript/rewriter/__init__.py

@@ -41,6 +41,7 @@
     _collapse_slices,
     _fuse_batchnorm,
     _fuse_pad_into_conv,
+    _fuse_reduce_arg_to_topk,
     _fuse_relus_clips,
     _min_max_to_clip,
     _no_op,
@@ -61,6 +62,7 @@
     *_fuse_pad_into_conv.rules,
     *_fuse_batchnorm.rules,
     *_remove_optional_bias.rules,
+    *_fuse_reduce_arg_to_topk.rules,
 )
 
 

onnxscript/rewriter/rules/common/_fuse_reduce_arg_to_topk.py

@@ -0,0 +1,272 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+"""Fuses Reduce{Max,Min} and Arg{Max,Min} patterns into TopK.
+
+Supported transformations:
+- ReduceMax(X, axes=[axis], keepdims=k) + ArgMax(X, axis=axis, keepdims=k) → TopK(X, k=1, axis=axis, largest=1) [+ Squeeze if k=0]
+- ReduceMin(X, axes=[axis], keepdims=k) + ArgMin(X, axis=axis, keepdims=k) → TopK(X, k=1, axis=axis, largest=0) [+ Squeeze if k=0]
+
+Supports both ONNX opset versions:
+    - Opset 13-17: Reduce{Max,Min} with axes as an attribute
+    - Opset 18+: Reduce{Max,Min} with axes as a second input
+
+Constraints:
+    - Both nodes must operate on the same input X.
+    - Both nodes must target the same axis.
+    - Both nodes must have the same keepdims attribute value.
+    - The Reduce node must operate on a single axis (len(axes) == 1).
+    - For opset 18+, the Reduce node's axes input must be a constant.
+"""
+
+from __future__ import annotations
+
+from abc import abstractmethod
+
+import numpy as np
+import onnx_ir as ir
+
+from onnxscript.rewriter._basics import MatchResult
+from onnxscript.rewriter._rewrite_rule import RewriteRuleClassBase, RewriteRuleSet
+
+
+class _FuseReduceArgToTopKBase(RewriteRuleClassBase):
+    """Base class for fusing Reduce{Max,Min} + Arg{Max,Min} into TopK.
+
+    This base class contains the common logic for checking and rewriting patterns where
+    a Reduce operation and its corresponding Arg operation can be replaced with a single
+    TopK operation.
+
+    Subclasses must implement:
+        - pattern(): Define the specific Reduce and Arg operations to match
+        - largest: Property returning 1 for Max operations, 0 for Min operations
+    """
+
+    @property
+    @abstractmethod
+    def largest(self) -> int:
+        """Return 1 for Max operations (largest elements), 0 for Min operations (smallest elements)."""
+
+    @staticmethod
+    def _normalize_axis(axis: int, rank: int | None) -> int:
+        """Normalize a potentially negative axis to a positive axis index.
+
+        Args:
+            axis: The axis to normalize (can be negative).
+            rank: The rank of the tensor, or None if unknown.
+
+        Returns:
+            The normalized axis (non-negative if rank is known and axis was negative).
+        """
+        if rank is not None and axis < 0:
+            return axis + rank
+        return axis
+
+    def check(self, context, reduce_val, arg_idx, **_) -> MatchResult:
+        """Check if Reduce and Arg operations can be safely fused into TopK.
+
+        Conditions:
+            - Both nodes must have the same keepdims attribute.
+            - The Reduce node must operate on a single axis.
+            - Both nodes must operate on the same axis.
+            - The Arg node must not use select_last_index=1 (TopK doesn't support this).
+
+        Args:
+            context: The rewrite context (unused).
+            reduce_val: The output of the Reduce operation (ReduceMax/ReduceMin).
+            arg_idx: The output of the Arg operation (ArgMax/ArgMin).
+
+        Returns:
+            MatchResult: Success if the pattern can be fused, Failure otherwise.
+        """
+        del context
+        check_result = MatchResult()
+
+        reduce_node = reduce_val.producer()
+        arg_node = arg_idx.producer()
+
+        # Step 1: Get keepdims attribute from both nodes
+        reduce_keepdims_attr = reduce_node.attributes.get("keepdims")
+        arg_keepdims_attr = arg_node.attributes.get("keepdims")
+
+        # ONNX default: keepdims = 1 for both Reduce and Arg operations
+        reduce_keepdims = (
+            reduce_keepdims_attr.as_int() if reduce_keepdims_attr is not None else 1
+        )
+        arg_keepdims = arg_keepdims_attr.as_int() if arg_keepdims_attr is not None else 1
+
+        # Step 2: Check if keepdims match
+        if reduce_keepdims != arg_keepdims:
+            return check_result.fail(
+                f"keepdims mismatch: {reduce_node.op_type} has {reduce_keepdims}, "
+                f"{arg_node.op_type} has {arg_keepdims}."
+            )
+
+        # Step 3: Get axes from Reduce operation
+        # In opset 18+, axes is an input; in opset 13-17, it's an attribute
+        reduce_axes_attr = reduce_node.attributes.get("axes")
+
+        if reduce_axes_attr is not None:
+            # Opset 13-17: axes is an attribute
+            try:
+                axes_list = list(reduce_axes_attr.as_ints())
+            except Exception:
+                return check_result.fail(f"Cannot parse {reduce_node.op_type} axes attribute.")
+        elif len(reduce_node.inputs) >= 2 and reduce_node.inputs[1] is not None:
+            # Opset 18+: axes is the second input
+            axes_input = reduce_node.inputs[1]
+            axes_tensor = ir.convenience.get_const_tensor(axes_input)
+            if axes_tensor is None:
+                return check_result.fail(
+                    f"{reduce_node.op_type} axes input is not a constant."
+                )
+            try:
+                axes_array = axes_tensor.numpy()
+                axes_list = axes_array.tolist() if axes_array.ndim > 0 else [int(axes_array)]
+            except Exception:
+                return check_result.fail(f"Cannot parse {reduce_node.op_type} axes input.")
+        else:
+            return check_result.fail(
+                f"{reduce_node.op_type} axes not found (neither attribute nor input)."
+            )
+
+        # Step 4: Check that Reduce operates on exactly one axis
+        if len(axes_list) != 1:
+            return check_result.fail(
+                f"{reduce_node.op_type} must operate on a single axis, got {len(axes_list)} axes."
+            )
+
+        reduce_axis = axes_list[0]
+
+        # Step 5: Get axis from Arg operation
+        # ONNX default: axis = 0 for ArgMax/ArgMin
+        arg_axis_attr = arg_node.attributes.get("axis")
+        arg_axis = arg_axis_attr.as_int() if arg_axis_attr is not None else 0
+
+        # Step 6: Check select_last_index attribute (if present)
+        # TopK always returns the first occurrence in case of ties
+        select_last_index_attr = arg_node.attributes.get("select_last_index")
+        if select_last_index_attr is not None and select_last_index_attr.as_int() != 0:
+            return check_result.fail(
+                f"{arg_node.op_type} has select_last_index=1, which is not supported by TopK."
+            )
+
+        # Step 7: Normalize axes if rank is known (handle negative indices)
+        input_x = reduce_node.inputs[0]
+        rank = len(input_x.shape) if input_x.shape is not None else None
+
+        if self._normalize_axis(reduce_axis, rank) != self._normalize_axis(arg_axis, rank):
+            return check_result.fail(
+                f"Axis mismatch: {reduce_node.op_type} operates on axis {reduce_axis}, "
+                f"{arg_node.op_type} operates on axis {arg_axis}."
+            )
+
+        return check_result
+
+    def rewrite(self, op, x, reduce_val, arg_idx):
+        """Rewrite the matched pattern with TopK (and optionally Squeeze).
+
+        Args:
+            op: The operation builder.
+            x: The input to both Reduce and Arg operations.
+            reduce_val: The output of the Reduce operation.
+            arg_idx: The output of the Arg operation.
+
+        Returns:
+            Tuple of (values, indices) matching the original outputs.
+        """
+        # Step 1: Get the nodes
+        arg_node = arg_idx.producer()
+
+        # Step 2: Extract necessary attributes with ONNX default values
+        axis_attr = arg_node.attributes.get("axis")
+        keepdims_attr = arg_node.attributes.get("keepdims")
+
+        axis = axis_attr.as_int() if axis_attr is not None else 0
+        keepdims = keepdims_attr.as_int() if keepdims_attr is not None else 1
+
+        # Step 2b: Normalize axis (convert negative to positive) if rank is known
+        if axis < 0 and x.shape is not None:
+            axis = len(x.shape) + axis
+
+        # Step 3: Create K constant
+        k_constant = op.Constant(value=ir.tensor(np.array([1], dtype=np.int64)))
+
+        # Step 4: Create TopK node
+        topk_values, topk_indices = op.TopK(
+            x,
+            k_constant,
+            axis=axis,
+            largest=self.largest,
+            sorted=1,
+            _outputs=2,
+        )
+
+        # Step 5: Handle keepdims=0 case
+        if keepdims == 0:
+            # TopK output always keeps the dimension (just makes it size 1)
+            # We need to squeeze it to match the original Reduce/Arg behavior
+            axes_constant = op.Constant(value=ir.tensor(np.array([axis], dtype=np.int64)))
+
+            new_values = op.Squeeze(topk_values, axes_constant)
+            new_indices = op.Squeeze(topk_indices, axes_constant)
+        else:
+            new_values = topk_values
+            new_indices = topk_indices
+
+        return new_values, new_indices
+
+
+class FuseReduceMaxArgMaxToTopK(_FuseReduceArgToTopKBase):
+    """Replaces ReduceMax + ArgMax with TopK(largest=1).
+
+    Transformation:
+        ReduceMax(X, axes=[axis], keepdims=k) + ArgMax(X, axis=axis, keepdims=k)
+        → TopK(X, k=1, axis=axis, largest=1) [+ Squeeze if k=0]
+
+    When keepdims=0, the output of TopK is squeezed to match the original output shapes.
+    """
+
+    @property
+    def largest(self) -> int:
+        return 1  # TopK returns largest elements
+
+    def pattern(self, op, x):
+        """Define the pattern to match: ReduceMax and ArgMax on the same input.
+
+        Note: For opset 18+, ReduceMax has a second input for axes, which we allow
+        but will validate in check() to ensure it's a constant.
+        """
+        reduce_val = op.ReduceMax(x, _allow_other_inputs=True, _outputs=["reduce_val"])
+        arg_idx = op.ArgMax(x, _outputs=["arg_idx"])
+        return reduce_val, arg_idx
+
+
+class FuseReduceMinArgMinToTopK(_FuseReduceArgToTopKBase):
+    """Replaces ReduceMin + ArgMin with TopK(largest=0).
+
+    Transformation:
+        ReduceMin(X, axes=[axis], keepdims=k) + ArgMin(X, axis=axis, keepdims=k)
+        → TopK(X, k=1, axis=axis, largest=0) [+ Squeeze if k=0]
+
+    When keepdims=0, the output of TopK is squeezed to match the original output shapes.
+    """
+
+    @property
+    def largest(self) -> int:
+        return 0  # TopK returns smallest elements
+
+    def pattern(self, op, x):
+        """Define the pattern to match: ReduceMin and ArgMin on the same input.
+
+        Note: For opset 18+, ReduceMin has a second input for axes, which we allow
+        but will validate in check() to ensure it's a constant.
+        """
+        reduce_val = op.ReduceMin(x, _allow_other_inputs=True, _outputs=["reduce_val"])
+        arg_idx = op.ArgMin(x, _outputs=["arg_idx"])
+        return reduce_val, arg_idx
+
+
+reduce_max_argmax_to_topk_rule = FuseReduceMaxArgMaxToTopK().rule()
+reduce_min_argmin_to_topk_rule = FuseReduceMinArgMinToTopK().rule()
+
+rules = RewriteRuleSet([reduce_max_argmax_to_topk_rule, reduce_min_argmin_to_topk_rule])