Arm backend: Add where.self (#9869)

Add operator where.self, which maps to the SELECT operator in the TOSA
specification. Tests are also added for the aforementioned operator.

Signed-off-by: Sebastian Larsson <sebastian.larsson@arm.com>

Author: Sebastian-Larsson

backends/arm/_passes/__init__.py

@@ -39,6 +39,7 @@
 from .insert_table_ops import InsertTableOpsPass  # noqa
 from .keep_dims_false_to_squeeze_pass import KeepDimsFalseToSqueezePass  # noqa
 from .match_arg_ranks_pass import MatchArgRanksPass  # noqa
+from .match_where_self_arg_dtype_pass import MatchWhereSelfDtypePass  # noqa
 from .meandim_to_averagepool_pass import ConvertMeanDimToAveragePoolPass  # noqa
 from .mm_to_bmm_pass import ConvertMmToBmmPass  # noqa
 from .remove_clone_pass import RemoveClonePass  # noqa

backends/arm/_passes/arm_pass_manager.py

@@ -40,6 +40,7 @@
     InsertTableOpsPass,
     KeepDimsFalseToSqueezePass,
     MatchArgRanksPass,
+    MatchWhereSelfDtypePass,
     QuantizeOperatorArguments,
     RemoveClonePass,
     ReplaceScalarWithTensorArgPassTOSABI,
@@ -80,6 +81,7 @@ def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(ConvertToClampPass())
         self.add_pass(ConvertMinMaxPass())
         self.add_pass(ConvertAnyDefaultDimDimsPass())
+        self.add_pass(MatchWhereSelfDtypePass())
         if isinstance(self.tosa_spec, Tosa_0_80) and self.tosa_spec.is_U55_subset:
             self.add_pass(CastToInt32Pass())
 
@@ -130,6 +132,7 @@ def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(ConvertToClampPass())
         self.add_pass(ConvertMinMaxPass())
         self.add_pass(ConvertAnyDefaultDimDimsPass())
+        self.add_pass(MatchWhereSelfDtypePass())
 
         self.add_pass(AnnotateDecomposedMatmulPass())
         self.add_pass(QuantizeOperatorArguments())

backends/arm/_passes/match_arg_ranks_pass.py

@@ -49,6 +49,7 @@ def __init__(self, exported_program):
         exir_ops.edge.aten.bitwise_left_shift.Tensor,
         exir_ops.edge.aten.eq.Tensor,
         exir_ops.edge.aten.pow.Tensor_Tensor,
+        exir_ops.edge.aten.where.self,
     ]
 
     def _match_op_rank(self, graph_module, node, arg, max_rank):

backends/arm/_passes/match_where_self_arg_dtype_pass.py

@@ -0,0 +1,95 @@
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# 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.backends.arm._passes.arm_pass_utils import create_node
+from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import ExportPass, PassResult
+
+DTYPE_RANK = {
+    torch.bool: 0,
+    torch.uint8: 1,
+    torch.int8: 2,
+    torch.int16: 3,
+    torch.int32: 4,
+    torch.int64: 5,
+    torch.float16: 6,
+    torch.float32: 7,
+    torch.float64: 8,
+}
+
+
+def get_largest_dtype(dtype_1, dtype_2):
+    """Find the largest dtype."""
+    return dtype_1 if DTYPE_RANK[dtype_1] > DTYPE_RANK[dtype_2] else dtype_2
+
+
+class MatchWhereSelfDtypePass(ExportPass):
+    """Pass to match data types of non-condition input tensors.
+
+    Edge dialect allows different data types for non-condition tensors, while TOSA
+    does not. In cases where they differ a TOSA CAST operator is inserted.
+
+    There is an edge case where one input is `boolean`, which cannot be directly cast
+    to, for example, float32. When this occurs two CAST operators are added to first
+    cast to int8 and then to the correct target data type.
+
+    """
+
+    def call(self, graph_module: torch.fx.GraphModule):
+        modified_graph = False
+        graph = graph_module.graph
+        node_list = graph.find_nodes(
+            op="call_function", target=exir_ops.edge.aten.where.self
+        )
+        for node in node_list:
+            cond, input_, other_ = node.args
+
+            input_dtype = input_.meta["val"].dtype
+            other_dtype = other_.meta["val"].dtype
+            target_dtype = torch.float32
+            if input_dtype != other_dtype:
+                target_dtype = get_largest_dtype(input_dtype, other_dtype)
+
+            for arg in node.args[1:]:
+                arg_dtype = arg.meta["val"].dtype
+
+                if arg_dtype != target_dtype:
+                    if arg_dtype == torch.bool:
+                        # Bool is an edge case which cannot necessarily be directly
+                        # converted to the target data type.
+                        with graph.inserting_after(arg):
+                            replace_node_int8 = create_node(
+                                graph,
+                                exir_ops.edge.dim_order_ops._to_dim_order_copy.default,
+                            )
+                            replace_node_int8.args = (arg,)
+                            replace_node_int8.kwargs = {"dtype": torch.int8}
+
+                        with graph.inserting_after(replace_node_int8):
+                            replace_node_fp32 = create_node(
+                                graph,
+                                exir_ops.edge.dim_order_ops._to_dim_order_copy.default,
+                            )
+                            replace_node_fp32.args = (replace_node_int8,)
+                            replace_node_fp32.kwargs = {"dtype": target_dtype}
+                            node.replace_input_with(arg, replace_node_fp32)
+                    else:
+                        with graph.inserting_after(arg):
+                            replace_node = create_node(
+                                graph,
+                                exir_ops.edge.dim_order_ops._to_dim_order_copy.default,
+                            )
+                            replace_node.args = (arg,)
+                            replace_node.kwargs = {"dtype": target_dtype}
+                            node.replace_input_with(arg, replace_node)
+
+                    modified_graph = True
+
+        if modified_graph:
+            graph_module.recompile()
+            graph_module = super().call(graph_module).graph_module
+
+        return PassResult(graph_module, modified_graph)

backends/arm/operator_support/ethos_u55_support.py

@@ -149,6 +149,7 @@ class EthosU55NotSupported(OperatorSupportBase):
         exir_ops.edge.aten.reflection_pad1d.default,  # REVERSE
         exir_ops.edge.aten.reflection_pad2d.default,  # REVERSE
         exir_ops.edge.aten.reflection_pad3d.default,  # REVERSE
+        exir_ops.edge.aten.where.self,  # SELECT
     ]
 
     def __init__(self, reporter: WhyNoPartitionReporter):

backends/arm/operator_support/tosa_supported_operators.py

@@ -207,6 +207,7 @@ def is_node_supported(
             exir_ops.edge.aten.squeeze_copy.dims,
             exir_ops.edge.aten.pow.Tensor_Scalar,
             exir_ops.edge.aten.pow.Tensor_Tensor,
+            exir_ops.edge.aten.where.self,
             operator.getitem,
             exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
             exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default,

backends/arm/operators/__init__.py

@@ -49,6 +49,7 @@
     op_transpose,
     op_upsample_nearest2d,
     op_view,
+    op_where,
     ops_binary,
     ops_unary,
 )

backends/arm/operators/op_where.py

@@ -0,0 +1,103 @@
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# 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 List, Sequence
+
+import serializer.tosa_serializer as ts  # type: ignore
+
+from executorch.backends.arm.operators.node_visitor import (
+    NodeVisitor,
+    register_node_visitor,
+)
+from executorch.backends.arm.tosa_mapping import TosaArg
+from executorch.backends.arm.tosa_specification import TosaSpecification
+from serializer.tosa_serializer import TosaOp
+from torch.fx import Node
+
+
+def _add_node_to_tosa_graph(
+    tosa_graph: ts.TosaSerializer,
+    inputs: List[TosaArg],
+    output: TosaArg,
+    supported_dtypes: Sequence,
+) -> None:
+    if len(inputs) != 3:
+        raise ValueError(f"aten.where.self expects 3 arguments, got {len(inputs)}")
+
+    if inputs[0].dtype is not ts.DType.BOOL:
+        raise ValueError("Input 0 needs to have dtype BOOL")
+    if inputs[1].dtype != inputs[2].dtype:
+        raise ValueError(
+            "Non-condition tensors must have same data type, got "
+            f"{inputs[1].dtype} and {inputs[2].dtype}"
+        )
+    for input_ in inputs[1:]:
+        if input_.dtype not in supported_dtypes:
+            raise ValueError(
+                f"Input needs to be of torch dtype {supported_dtypes}, got {input_.dtype}"
+            )
+
+    tosa_graph.addOperator(
+        TosaOp.Op().SELECT,
+        [inputs[0].name, inputs[1].name, inputs[2].name],
+        [output.name],
+        None,
+    )
+
+
+@register_node_visitor
+class WhereVisitor_080_BI(NodeVisitor):
+    target = "aten.where.self"
+
+    tosa_specs = [
+        TosaSpecification.create_from_string("TOSA-0.80+BI"),
+    ]
+
+    def __init__(self, *args):
+        super().__init__(*args)
+
+    def define_node(
+        self,
+        node: Node,
+        tosa_graph: ts.TosaSerializer,
+        inputs: List[TosaArg],
+        output: TosaArg,
+    ) -> None:
+
+        bi_supported_dtypes = [
+            ts.DType.INT8,
+            ts.DType.INT16,
+            ts.DType.INT32,
+            ts.DType.BOOL,
+        ]
+        _add_node_to_tosa_graph(tosa_graph, inputs, output, bi_supported_dtypes)
+
+
+@register_node_visitor
+class WhereVisitor_080_MI(WhereVisitor_080_BI):
+
+    tosa_specs = [
+        TosaSpecification.create_from_string("TOSA-0.80+MI"),
+    ]
+
+    def __init__(self, *args):
+        super().__init__(*args)
+
+    def define_node(
+        self,
+        node: Node,
+        tosa_graph: ts.TosaSerializer,
+        inputs: List[TosaArg],
+        output: TosaArg,
+    ) -> None:
+        mi_supported_dtypes = [
+            ts.DType.FP16,
+            ts.DType.FP32,
+            ts.DType.INT8,
+            ts.DType.INT16,
+            ts.DType.INT32,
+            ts.DType.BOOL,
+        ]
+        _add_node_to_tosa_graph(tosa_graph, inputs, output, mi_supported_dtypes)

backends/arm/quantizer/quantization_annotator.py

@@ -238,13 +238,14 @@ def _match_pattern(
     torch.ops.aten.dropout_.default,
     torch.ops.aten.clamp.default,
     torch.ops.aten.clamp.Tensor,
+    torch.ops.aten.where,
     operator.getitem,
 ]
 
 
 def get_quant_properties(  # noqa: C901
     node: Node, gm: torch.fx.GraphModule, quantization_config
-) -> _OpQuantProperties:
+) -> _OpQuantProperties | None:
     input_act_qspec = quantization_config.get_input_act_qspec()
     weight_qspec = quantization_config.get_weight_qspec()
     output_act_qspec = quantization_config.get_output_act_qspec()
@@ -322,6 +323,13 @@ def any_or_hardtanh_min_zero(n: Node):
             ),
         ]
         quant_properties.quant_output = _QuantProperty(0, shared_qspec)  # type: ignore[arg-type]
+    elif node.target in (torch.ops.aten.where.self,):
+        shared_qspec = SharedQuantizationSpec(node.args[1])  # type: ignore[arg-type]
+        quant_properties.quant_inputs = [
+            _QuantProperty(1, shared_qspec),  # type: ignore[arg-type]
+            _QuantProperty(2, shared_qspec),  # type: ignore[arg-type]
+        ]
+        quant_properties.quant_output = _QuantProperty(0, shared_qspec)  # type: ignore[arg-type]
     elif node.target == torch.ops.aten.adaptive_avg_pool2d.default:
         input_qspec = (
             SharedQuantizationSpec(node.args[0])  # type: ignore[arg-type]
@@ -376,16 +384,16 @@ def any_or_hardtanh_min_zero(n: Node):
         quant_properties.quant_output = None
     elif node.target in _parent_shared_qspec:
         if not isinstance(node.args[0], Node):
-            return None  # type: ignore[return-value]
+            return None
 
         if not arm_quantizer_utils.is_output_annotated(node.args[0]):  # type: ignore[attr-defined]
-            return None  # type: ignore[return-value]
+            return None
 
         shared_qspec = SharedQuantizationSpec(node.args[0])
         quant_properties.quant_inputs = [_QuantProperty(0, shared_qspec)]  # type: ignore[arg-type]
         quant_properties.quant_output = _QuantProperty(0, shared_qspec)  # type: ignore[arg-type]
     else:
-        return None  # type: ignore[return-value]
+        return None
 
     # Don't check if operator.getitem is ok for quantization, it's always ok
     if node.target == operator.getitem:
@@ -394,7 +402,7 @@ def any_or_hardtanh_min_zero(n: Node):
     # Check that each inputs/outputs can be quantized properly with the
     # provided quantization properties.
     if not _is_ok_for_quantization(node, quant_properties, gm):
-        return None  # type: ignore[return-value]
+        return None
 
     return quant_properties
 

backends/arm/test/models/test_conformer.py

@@ -31,10 +31,8 @@ class TestConformer(unittest.TestCase):
     # .to_executorch step, i.e. after Arm partitioner.
     ops_after_partitioner = {
         "executorch_exir_dialects_edge__ops_aten_max_default": 1,
-        "executorch_exir_dialects_edge__ops_aten_where_self": 4,
         "torch.ops.aten._assert_scalar.default": 10,
         "torch.ops.aten._local_scalar_dense.default": 1,
-        "torch.ops.higher_order.executorch_call_delegate": 4,
     }
 
     dim = 16