Eliminate unnecessary ScatterND (#2422)

Identify ScatterND(data, indices, updates) that can be replaced by
Identity(updates). This is generated by the translation of `x[:, ...] =
y` in PyTorch. The specific pattern is that the updated indices take the
form [[0], ..., [S-1]] for the first dimension, where S is the size of
the first dimension of the updated-data tensor. In effect, the
scatter-update ends up being an assignment of a new value to the entire
tensor.

---------

Signed-off-by: Ganesan Ramalingam <[email protected]>

Author: gramalingam

onnxscript/rewriter/__init__.py

@@ -8,6 +8,7 @@
     "pattern",
     "rewrite",
     "RewritePass",
+    "MatchResult",
 ]
 
 import onnx
@@ -21,7 +22,9 @@
     collapse_slices,
     no_op,
     pattern,
+    redundant_scatter_nd,
 )
+from onnxscript.rewriter._basics import MatchResult
 
 _ModelProtoOrIr = TypeVar("_ModelProtoOrIr", onnx.ModelProto, ir.Model)
 _DEFAULT_REWRITE_RULES: tuple[pattern.RewriteRule, ...] = (
@@ -30,6 +33,7 @@
     *cast_constant_of_shape.rules.rules,
     *collapse_slices.rules.rules,
     *basic_rules.basic_optimization_rules().rules,
+    *redundant_scatter_nd.rules.rules,
 )
 
 

onnxscript/rewriter/redundant_scatter_nd.py

@@ -0,0 +1,65 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+"""Rewrite rule to eliminate redundant ScatterND operations.
+
+Identify ScatterND(data, indices, updates) that can be replaced by Identity(updates).
+This is generated by the translation of `x[:, ...] = y` in PyTorch.
+The specific pattern is that the updated indices take the form [[0], ..., [S-1]] for the first dimension,
+where S is the size of the first dimension of the updated-data tensor.
+In effect, the scatter-update ends up being an assignment of a new value to the entire tensor.
+"""
+
+from __future__ import annotations
+
+import onnx_ir as ir
+
+import onnxscript.rewriter
+from onnxscript.rewriter import _ir_utils as ir_utils
+from onnxscript.rewriter import pattern as orp
+
+
+def fail(*args):
+    return onnxscript.rewriter.MatchResult().fail(*args)
+
+
+class ScatterAll(orp.RewriteRuleClassBase):
+    def pattern(self, op, data, axis, transposed_data, updates):
+        # Construct update-indices spanning an entire axis:
+        shape = op.Shape(data, start=0)
+        dim = op.Gather(shape, axis, axis=0)
+        full_range = op.Range(0, dim, 1)
+        full_range_2d = op.Unsqueeze(full_range, [-1])
+        # The update is applied to the data transposed to bring the updated axis to the front:
+        return op.ScatterND(transposed_data, full_range_2d, updates, reduction="none")
+
+    def check(self, context, data, axis, transposed_data, **_):
+        # Check that updated-indices represent the full range of the first dimension of the transposed data.
+        # That is: check that the data.shape[axis] matches transposed_data.shape[0].
+        axis_value = ir_utils.get_singleton_value(axis)
+        if not isinstance(axis_value, int):
+            return fail("Axis value must be a constant integer.", axis)
+        shape: ir.Shape | None = data.shape
+        if shape is None:
+            return fail("Data shape is not statically known.", data)
+        updated_dim_value = shape[axis_value]
+        transposed_data_shape: ir.Shape | None = transposed_data.shape
+        if transposed_data_shape is None:
+            return fail("Transposed data shape is not statically known.", transposed_data)
+        actual_dim_value = transposed_data_shape[0]
+        if updated_dim_value != actual_dim_value:
+            # The first dimension of the transposed data does not match the updated dimension,
+            # so we cannot apply this rule.
+            return fail(
+                "The first dimension of the transposed data does not match the updated dimension.",
+                data,
+                transposed_data,
+            )
+        return True
+
+    def rewrite(self, op, updates, **_):
+        return op.Identity(updates)
+
+
+rule = ScatterAll.rule()
+
+rules = orp.RewriteRuleSet([rule])

onnxscript/rewriter/redundant_scatter_nd_test.py

@@ -0,0 +1,70 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+# ruff: noqa: F821
+
+import unittest
+
+import numpy as np
+import onnx_ir as ir
+import onnxruntime
+from onnx_ir.passes.common import CheckerPass, ShapeInferencePass
+
+import onnxscript.optimizer
+from onnxscript import FLOAT, script
+from onnxscript import opset18 as op
+from onnxscript.rewriter import redundant_scatter_nd
+
+shape_inference = ShapeInferencePass()
+onnx_check = CheckerPass(True)
+
+
+class RedundantScatterNdTest(unittest.TestCase):
+    def test_redundant_scatter_nd(self):
+        @script()
+        def model_script(
+            data: FLOAT[8, "N", 16], updates: FLOAT[8, "N", 16]
+        ) -> FLOAT[8, "N", 16]:
+            # Construct update-indices spanning an entire axis:
+            axis = op.Constant(value_int=1)
+            shape = op.Shape(data, start=0)
+            dim = op.Gather(shape, axis, axis=0)
+            full_range = op.Range(0, dim, 1)
+            full_range_2d = op.Unsqueeze(full_range, [-1])
+            # The update is applied to the data transposed to bring the updated axis to the front:
+            transposed_data = op.Transpose(data, perm=[1, 0, 2])
+            transposed_updates = op.Transpose(updates, perm=[1, 0, 2])
+            scattered = op.ScatterND(
+                transposed_data, full_range_2d, transposed_updates, reduction="none"
+            )
+            # Transpose the result back to the original shape:
+            output = op.Transpose(scattered, perm=[1, 0, 2])
+            return output
+
+        input_model_proto = model_script.to_model_proto()
+        model = ir.serde.deserialize_model(input_model_proto)
+        onnx_check(model)
+        shape_inference(model)
+        onnxscript.optimizer.fold_constants(model)
+        count = redundant_scatter_nd.rules.apply_to_model(model)
+        self.assertEqual(count, 1)
+        onnx_check(model)
+        optimized_model_proto = ir.serde.serialize_model(model)
+        # Test that both models are equivalent:
+        inputs = {
+            "data": np.random.rand(8, 4, 16).astype(np.float32),
+            "updates": np.random.rand(8, 4, 16).astype(np.float32),
+        }
+        session = onnxruntime.InferenceSession(
+            input_model_proto.SerializeToString(), providers=["CPUExecutionProvider"]
+        )
+        outputs = session.run(None, inputs)
+        optimized_session = onnxruntime.InferenceSession(
+            optimized_model_proto.SerializeToString(), providers=["CPUExecutionProvider"]
+        )
+        optimized_outputs = optimized_session.run(None, inputs)
+        for output, optimized_output in zip(outputs, optimized_outputs):
+            np.testing.assert_allclose(output, optimized_output, rtol=1e-6, atol=1e-6)
+
+
+if __name__ == "__main__":
+    unittest.main()