Add decomposition of div op for ArmBackend

Implements pass that decomposes aten.div to
aten.reciprocal and aten.mul.
This is done in the Quantizer get quantization annotation
on the decomposed operators.

Add infra for passes in ArmQuantizer

Signed-off-by: Erik Lundell <[email protected]>
Change-Id: Idd1698dc5fc82ab42b68094b405fb3a08804a45e

Author: Erik-Lundell

backends/arm/operators/__init__.py

@@ -15,7 +15,6 @@
     op_cat,
     op_conv2d,
     op_dequant,
-    op_div,
     op_exp,
     op_full,
     op_get_item,

backends/arm/passes/arm_pass_manager.py

@@ -17,10 +17,14 @@
 from executorch.backends.arm.passes.convert_split_to_slice import (
     ConvertSplitToSlicePass,
 )
+from executorch.backends.arm.passes.decompose_div_pass import DecomposeDivPass
 from executorch.backends.arm.passes.meandim_to_averagepool_pass import (
     ConvertMeanDimToAveragePool,
 )
 from executorch.backends.arm.passes.remove_clone_pass import RemoveClonePass
+from executorch.backends.arm.passes.scalars_to_attribute_pass import (
+    ScalarsToAttributePass,
+)
 from executorch.backends.arm.passes.size_adjust_conv2d_pass import SizeAdjustConv2DPass
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from executorch.exir.pass_manager import PassManager
@@ -39,6 +43,7 @@ def transform_to_backend_pipeline(
         self.add_pass(RemoveClonePass())
         self.add_pass(ConvertExpandCopyToRepeatPass())
         self.add_pass(ConvertMeanDimToAveragePool())
+        self.add_pass(DecomposeDivPass())
         self.add_pass(ConvertSplitToSlicePass())
         for spec in compile_spec:
             if spec.key == "permute_memory_format":
@@ -47,3 +52,8 @@ def transform_to_backend_pipeline(
                     self.add_pass(AnnotateChannelsLastDimOrder())
 
         return self._transform(graph_module)
+
+    def transform_for_annotation_pipeline(self, graph_module: torch.fx.GraphModule):
+        self.add_pass(DecomposeDivPass())
+        self.add_pass(ScalarsToAttributePass())
+        return self._transform(graph_module)

backends/arm/passes/decompose_div_pass.py

@@ -0,0 +1,45 @@
+# Copyright 2024 Arm Limited and/or its affiliates.
+# 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 torch
+from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import ExportPass
+
+
+def get_div_decomposition(op) -> tuple:
+    """
+    Returns the the (reciprocal_op, mul_op), where the ops depends on if
+    the div op is in exir_ops torch.ops.aten.
+    """
+    if op == exir_ops.edge.aten.div.Tensor:
+        return (exir_ops.edge.aten.reciprocal.default, exir_ops.edge.aten.mul.Tensor)
+    if op == torch.ops.aten.div.Tensor:
+        return (torch.ops.aten.reciprocal.default, torch.ops.aten.mul.Tensor)
+    raise RuntimeError(f"Can't get div decomposition for op {op}")
+
+
+class DecomposeDivPass(ExportPass):
+    """
+    This pass decomposes div into a mul and a reciprocal node.
+
+    Example:
+        y = div(a,b)
+    Becomes:
+        x = reciprocal(b)
+        y = mul(a,x)
+    """
+
+    def call_operator(self, op, args, kwargs, meta):
+        if op not in (exir_ops.edge.aten.div.Tensor, torch.ops.aten.div.Tensor):
+            return super().call_operator(op, args, kwargs, meta)
+
+        reciprocal_op, mul_op = get_div_decomposition(op)
+
+        numerator = args[0]
+        denominator = args[1]
+        reciprocal = super().call_operator(reciprocal_op, (denominator,), {}, meta)
+
+        return super().call_operator(mul_op, (numerator, reciprocal), {}, meta)

backends/arm/passes/scalars_to_attribute_pass.py

@@ -0,0 +1,69 @@
+# Copyright 2024 Arm Limited and/or its affiliates.
+# 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.
+
+from typing import cast, Union
+
+import torch
+from executorch.backends.arm.tosa_mapping import extract_tensor_meta
+
+from executorch.exir.pass_base import ExportPass, PassResult
+from torch.ao.quantization.fx.utils import get_new_attr_name_with_prefix
+from torch.fx import GraphModule, Node
+
+
+class ScalarsToAttributePass(ExportPass):
+    """
+    For ops in 'targeted_ops', convert inputs that are scalar values
+    to attribute Nodes that output the same value.
+    """
+
+    targeted_ops = [
+        torch.ops.aten.add.Tensor,
+        torch.ops.aten.sub.Tensor,
+        torch.ops.aten.sub_.Tensor,
+        torch.ops.aten.mul.Tensor,
+        torch.ops.aten.div.Tensor,
+    ]
+
+    def call(self, graph_module: GraphModule) -> GraphModule:
+        for n in graph_module.graph.nodes:
+            n = cast(Node, n)
+            if n.op != "call_function" or n.target not in self.targeted_ops:
+                continue
+
+            biggest_rank = 1
+            for arg in n.args:
+                if isinstance(arg, Node):
+                    _, shape, _ = extract_tensor_meta(arg.meta)
+                    biggest_rank = max(biggest_rank, len(shape))
+
+            new_args = []
+            for arg in n.args:
+                if isinstance(arg, Node):
+                    new_args.append(arg)
+                    continue
+
+                prefix = "_tensor_constant_"
+                get_new_attr_name = get_new_attr_name_with_prefix(prefix)
+                tensor_constant_name = get_new_attr_name(graph_module)
+                float_tensor = torch.tensor(
+                    float(cast(Union[int, float], arg))
+                ).reshape((1,) * biggest_rank)
+                graph_module.register_buffer(tensor_constant_name, float_tensor)
+                fake_mode = n.meta["val"].fake_mode
+
+                with graph_module.graph.inserting_before(n):
+                    get_attr_node = graph_module.graph.create_node(
+                        "get_attr", tensor_constant_name, (), {}
+                    )
+                    get_attr_node.meta["val"] = fake_mode.from_tensor(
+                        float_tensor, static_shapes=True
+                    )
+                    new_args.append(get_attr_node)
+            n.args = tuple(new_args)
+
+        graph_module.recompile()
+        return PassResult(graph_module, True)

backends/arm/quantizer/arm_quantizer.py

@@ -19,10 +19,10 @@
 
 import torch
 import torch.nn.functional as F
+from executorch.backends.arm.passes.arm_pass_manager import ArmPassManager
 
 from executorch.backends.arm.quantizer import arm_quantizer_utils
 from executorch.backends.arm.quantizer.arm_quantizer_utils import (
-    convert_scalars_to_attrs,
     mark_nodes_as_annotated,
     propagate_annotation,
 )
@@ -317,7 +317,8 @@ def transform_for_annotation(self, model: GraphModule) -> GraphModule:
         """An initial pass for transforming the graph to prepare it for annotation.
         Currently transforms scalar values to tensor attributes.
         """
-        return convert_scalars_to_attrs(model)
+
+        return ArmPassManager().transform_for_annotation_pipeline(graph_module=model)
 
     def annotate(self, model: GraphModule) -> GraphModule:
         """Performs the quantization annotation on the graph.

backends/arm/quantizer/arm_quantizer_utils.py

@@ -12,12 +12,11 @@
 #
 
 import operator
-from typing import Callable, cast, List, Union
+from typing import Callable, cast, List
 
 import torch
 from executorch.backends.arm.quantizer.quantization_config import QuantizationConfig
 from torch._subclasses import FakeTensor
-from torch.ao.quantization.fx.utils import get_new_attr_name_with_prefix
 
 from torch.ao.quantization.quantizer import (
     QuantizationAnnotation,
@@ -196,42 +195,3 @@ def propagate_annotation(model: GraphModule) -> None:
             output_qspec=shared_qspec,
             _annotated=True,
         )
-
-
-def convert_scalars_to_attrs(model: GraphModule) -> GraphModule:
-    """For ops in 'targeted_ops', convert inputs that are scalar values
-    to attribute Nodes that output the same value.
-    #TODO Seems like this should be a pass.
-    """
-    targeted_ops = [
-        torch.ops.aten.add.Tensor,
-        torch.ops.aten.sub.Tensor,
-        torch.ops.aten.mul.Tensor,
-    ]
-    for n in model.graph.nodes:
-        n = cast(Node, n)
-        if n.op != "call_function" or n.target not in targeted_ops:
-            continue
-        args = list(n.args)
-        new_args = []
-        for i in range(len(args)):
-            if isinstance(args[i], Node):
-                new_args.append(args[i])
-                continue
-            prefix = "_tensor_constant_"
-            get_new_attr_name = get_new_attr_name_with_prefix(prefix)
-            tensor_constant_name = get_new_attr_name(model)
-            float_tensor = torch.tensor(float(cast(Union[int, float], args[i])))
-            model.register_buffer(tensor_constant_name, float_tensor)
-            fake_mode = n.meta["val"].fake_mode
-            with model.graph.inserting_before(n):
-                get_attr_node = model.graph.create_node(
-                    "get_attr", tensor_constant_name, (), {}
-                )
-                get_attr_node.meta["val"] = fake_mode.from_tensor(
-                    float_tensor, static_shapes=True
-                )
-                new_args.append(get_attr_node)
-        n.args = tuple(new_args)
-    model.recompile()
-    return model

backends/arm/quantizer/quantization_annotation/mul_annotator.py

@@ -4,19 +4,15 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-# pyre-unsafe
-
-import itertools
-import operator
 from typing import Callable, List, Optional
 
 import torch
+import torch.fx
 from executorch.backends.arm.quantizer import arm_quantizer_utils
 from executorch.backends.arm.quantizer.quantization_annotation import register_annotator
 from executorch.backends.arm.quantizer.quantization_config import QuantizationConfig
 from torch.ao.quantization.quantizer import QuantizationAnnotation
 from torch.fx import Node
-from torch.fx.passes.utils.source_matcher_utils import get_source_partitions
 
 
 @register_annotator("mul")
@@ -25,14 +21,13 @@ def _annotate_mul(
     quantization_config: QuantizationConfig,
     filter_fn: Optional[Callable[[Node], bool]] = None,
 ) -> Optional[List[List[Node]]]:
-    mul_partitions = get_source_partitions(
-        gm.graph, ["mul", "mul_", operator.mul, torch.mul, operator.imul], filter_fn
-    )
-    mul_partitions = list(itertools.chain.from_iterable(mul_partitions.values()))
+
     annotated_partitions = []
-    for mul_partition in mul_partitions:
-        annotated_partitions.append(mul_partition.nodes)
-        mul_node = mul_partition.output_nodes[0]
+    for node in gm.graph.nodes:
+        if node.target not in (torch.ops.aten.mul.Tensor,):
+            continue
+        mul_node = node
+        annotated_partitions.append([mul_node])
         if arm_quantizer_utils.is_annotated(mul_node):
             continue
 

backends/arm/test/ops/test_div.py

@@ -28,8 +28,8 @@
     ),
     (
         "op_div_rank1_rand",
-        torch.rand(5),
-        torch.rand(5),
+        torch.rand(5) * 5,
+        torch.rand(5) * 5,
         None,
     ),
     (
@@ -70,8 +70,8 @@
     ),
     (
         "op_div_rank4_large_randn",
-        200 * torch.randn(5, 10, 25, 20),
-        torch.rand(5, 10, 25, 20),
+        200 * torch.randn(5, 10, 25, 20) + 1,
+        torch.rand(5, 10, 25, 20) + 1,
         None,
     ),
 ]
@@ -81,26 +81,18 @@ class TestDiv(unittest.TestCase):
     """Tests division"""
 
     class Div(torch.nn.Module):
-        def __init__(
-            self,
-            input_: Union[torch.Tensor, torch.types.Number],
-            other_: Union[torch.Tensor, torch.types.Number],
-            rounding_mode: Optional[str] = None,
-        ):
-            super().__init__()
-            self.rounding_mode = rounding_mode
 
         def forward(
             self,
             input_: Union[torch.Tensor, torch.types.Number],
             other_: Union[torch.Tensor, torch.types.Number],
             rounding_mode: Optional[str] = None,
         ):
-            if self.rounding_mode is None:
+            if rounding_mode is None:
                 return torch.div(input=input_, other=other_)
             else:
                 return torch.div(
-                    input=input_, other=other_, rounding_mode=self.rounding_mode
+                    input=input_, other=other_, rounding_mode=rounding_mode
                 )
 
     def _test_div_tosa_MI_pipeline(
@@ -133,13 +125,15 @@ def _test_div_tosa_BI_pipeline(
             )
             .quantize()
             .export()
-            .check_count({"torch.ops.aten.div.Tensor": 1})
+            .check_count(
+                {"torch.ops.aten.reciprocal.default": 1, "torch.ops.aten.mul.Tensor": 1}
+            )
             .check(["torch.ops.quantized_decomposed"])
             .to_edge()
             .partition()
             .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
             .to_executorch()
-            .run_method_and_compare_outputs(inputs=test_data)
+            .run_method_and_compare_outputs(inputs=test_data, atol=1, rtol=0.1)
         )
 
     def _test_div_u55_BI_pipeline(
@@ -153,7 +147,9 @@ def _test_div_u55_BI_pipeline(
             )
             .quantize()
             .export()
-            .check_count({"torch.ops.aten.div.Tensor": 1})
+            .check_count(
+                {"torch.ops.aten.reciprocal.default": 1, "torch.ops.aten.mul.Tensor": 1}
+            )
             .check(["torch.ops.quantized_decomposed"])
             .to_edge()
             .partition()
@@ -170,14 +166,9 @@ def test_div_tosa_MI(
         rounding_mode: Optional[str] = None,
     ):
         test_data = (input_, other_)
-        self._test_div_tosa_MI_pipeline(
-            self.Div(input_, other_, rounding_mode=rounding_mode), test_data
-        )
+        self._test_div_tosa_MI_pipeline(self.Div(), test_data)
 
-    # Expected to fail since ArmQuantizer cannot quantize a Div layer
-    # TODO(MLETORCH-129)
     @parameterized.expand(test_data_suite)
-    @unittest.expectedFailure
     def test_div_tosa_BI(
         self,
         test_name: str,
@@ -187,12 +178,9 @@ def test_div_tosa_BI(
     ):
 
         test_data = (input_, other_)
-        self._test_div_tosa_BI_pipeline(
-            self.Div(input=input_, other=other_, rounding_mode=rounding_mode), test_data
-        )
+        self._test_div_tosa_BI_pipeline(self.Div(), test_data)
 
-    # Expected to fail since ArmQuantizer cannot quantize a Div layer
-    # TODO(MLETORCH-129)
+    # Fails due to Vela error.
     @parameterized.expand(test_data_suite)
     @unittest.expectedFailure
     def test_div_u55_BI(
@@ -203,6 +191,4 @@ def test_div_u55_BI(
         rounding_mode: Optional[str] = None,
     ):
         test_data = (input_, other_)
-        self._test_div_u55_BI_pipeline(
-            self.Div(input=input_, other=other_, rounding_mode=rounding_mode), test_data
-        )
+        self._test_div_u55_BI_pipeline(self.Div(), test_data)