Update

[ghstack-poisoned]

Author: GregoryComer

backends/arm/TARGETS

@@ -21,6 +21,19 @@ python_library(
         "//executorch/exir/dialects:lib",
     ],
 )
+python_library(
+    name = "common",
+    srcs = [
+        "common/__init__.py",
+        "common/debug.py",
+    ],
+    deps = [
+        "fbsource//third-party/tosa_tools/v0.80/serialization_lib/python/serializer:serializer",
+        "fbsource//third-party/tosa_tools/v1.00/serialization_lib/python/serializer:serializer",
+        "//caffe2:torch",
+        "//executorch/exir:lib",
+    ],
+)
 python_library(
     name = "arm_partitioner",
     srcs = [

backends/arm/_passes/TARGETS

@@ -4,6 +4,7 @@ python_library(
     name = "passes",
     srcs = glob(["*.py"]),
     deps = [
+        "//executorch/backends/arm:common",
         "//executorch/backends/arm:constants",
         "//executorch/backends/arm:tosa_quant_utils",
         "//executorch/backends/arm:tosa_utils",

backends/arm/_passes/__init__.py

@@ -35,6 +35,7 @@
 from .decompose_cosine_similarity_pass import DecomposeCosineSimilarityPass  # noqa
 from .decompose_div_pass import DecomposeDivPass  # noqa
 from .decompose_embedding_pass import DecomposeEmbeddingPass  # noqa  # noqa
+from .decompose_expm1_pass import DecomposeExpm1Pass  # noqa
 from .decompose_gelu_pass import DecomposeGeluPass  # noqa
 from .decompose_glu_pass import DecomposeGluPass  # noqa
 from .decompose_grouped_conv import DecomposeGroupedConv  # noqa

backends/arm/_passes/arm_pass_manager.py

@@ -40,6 +40,7 @@
     DecomposeCosineSimilarityPass,
     DecomposeDivPass,
     DecomposeEmbeddingPass,
+    DecomposeExpm1Pass,
     DecomposeGeluPass,
     DecomposeGluPass,
     DecomposeGroupedConv,
@@ -164,6 +165,7 @@ def _tosa_INT_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         return self._transform(exported_program.graph_module)
 
     def _tosa_FP_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
+        self.add_pass(DecomposeExpm1Pass())
         self.add_pass(DecomposeMaskedFill())
         self.add_pass(DecomposeRoundPass())
         self.add_pass(DecomposeAcoshPass())

backends/arm/_passes/decompose_expm1_pass.py

@@ -0,0 +1,135 @@
+# 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 executorch.backends.arm._passes import ArmPass
+from executorch.exir.dialects._ops import ops as exir_ops
+
+
+edge_expm1_ops = (exir_ops.edge.aten.expm1.default,)  # MI case
+
+
+def _get_expm1_decomposition(op) -> tuple:
+    """
+    Returns the decomposition of the given aten.expm1 operation into
+    its equivalent TOSA-supported operations
+
+    This handles both edge dialect ops and core PyTorch ops. The decomposition strategy
+    is:
+        expm1(x) → where(and(ge(x, -0.35), le(x, 0.35)), {taylor_series_expansion}, (exp(x)-1))
+
+    where {taylor_series_expansion} = x + (x^2/2) + (x^3/6) + (x^4/24)
+
+    Returns:
+        A tuple (op_pow, op_div, op_add, op_exp, op_sub, op_ge, op_where, op_le, op_and)
+        corresponding to the appropriate operator overloads for the input op.
+
+    Raises:
+        RuntimeError: If the provided operator is not a supported elu variant.
+    """
+    if op in edge_expm1_ops:
+        return (
+            exir_ops.edge.aten.pow.Tensor_Scalar,
+            exir_ops.edge.aten.div.Scalar,
+            exir_ops.edge.aten.add.Tensor,
+            exir_ops.edge.aten.exp.default,
+            exir_ops.edge.aten.sub.Scalar,
+            exir_ops.edge.aten.ge.Scalar,
+            exir_ops.edge.aten.where.self,
+            exir_ops.edge.aten.le.Scalar,
+            exir_ops.edge.aten.logical_and.default,
+        )
+
+    raise RuntimeError(f"Can't get expm1 decomposition for op {op}")
+
+
+class DecomposeExpm1Pass(ArmPass):
+    """
+    A transformation pass that decomposes unsupported 'aten.expm1' operations
+    into a combination of supported TOSA-equivalent operations.
+
+    Since TOSA does not provide a native expm1 operator, this pass rewrites:
+        expm1(x) →  where(and(ge(x, -0.35), le(x, 0.35)), {taylor_series_expansion}, (exp(x)-1))
+    where {taylor_series_expansion} = x + (x^2/2) + (x^3/6) + (x^4/24)
+
+    Supported input ops:
+        - exir_ops.edge.aten.expm1.default(x)
+
+    These are replaced with:
+        - exir_ops.edge.aten.pow.Tensor_Scalar,
+        - exir_ops.edge.aten.div.Scalar,
+        - exir_ops.edge.aten.add.Tensor,
+        - exir_ops.edge.aten.exp.default,
+        - exir_ops.edge.aten.sub.Scalar,
+        - exir_ops.edge.aten.ge.Scalar,
+        - exir_ops.edge.aten.where.self,
+        - exir_ops.edge.aten.le.Scalar,
+        - exir_ops.edge.aten.logical_and.default
+    """
+
+    def call_operator(self, op, args, kwargs, meta):
+        if op not in edge_expm1_ops:
+            return super().call_operator(op, args, kwargs, meta, updated=False)
+
+        (
+            op_pow,
+            op_div,
+            op_add,
+            op_exp,
+            op_sub,
+            op_ge,
+            op_where,
+            op_le,
+            op_and,
+        ) = _get_expm1_decomposition(op)
+
+        input = args[0]
+
+        cutlo = -0.35
+        cuthi = 0.35
+
+        taylor_term_2_numerator = super().call_operator(
+            op_pow, (input, 2), {}, meta, updated=False
+        )
+        taylor_term_3_numerator = super().call_operator(
+            op_pow, (input, 3), {}, meta, updated=False
+        )
+        taylor_term_4_numerator = super().call_operator(
+            op_pow, (input, 4), {}, meta, updated=False
+        )
+
+        taylor_term_2 = super().call_operator(
+            op_div, (taylor_term_2_numerator, 2), {}, meta, updated=False
+        )
+        taylor_term_3 = super().call_operator(
+            op_div, (taylor_term_3_numerator, 6), {}, meta, updated=False
+        )
+        taylor_term_4 = super().call_operator(
+            op_div, (taylor_term_4_numerator, 24), {}, meta, updated=False
+        )
+
+        add_terms_1_2 = super().call_operator(
+            op_add, (input, taylor_term_2), {}, meta, updated=False
+        )
+        add_term_3 = super().call_operator(
+            op_add, (add_terms_1_2, taylor_term_3), {}, meta, updated=False
+        )
+        taylor_expansion = super().call_operator(
+            op_add, (add_term_3, taylor_term_4), {}, meta, updated=False
+        )
+
+        decomp_exp = super().call_operator(op_exp, (input,), {}, meta, updated=False)
+        decomp_sub = super().call_operator(
+            op_sub, (decomp_exp, 1.0), {}, meta, updated=False
+        )
+
+        ge = super().call_operator(op_ge, (input, cutlo), {}, meta, updated=False)
+        le = super().call_operator(op_le, (input, cuthi), {}, meta, updated=False)
+
+        cond_and = super().call_operator(op_and, (ge, le), {}, meta, updated=False)
+        where = super().call_operator(
+            op_where, (cond_and, taylor_expansion, decomp_sub), {}, meta, updated=True
+        )
+
+        return where

backends/arm/_passes/insert_table_ops.py

@@ -43,6 +43,7 @@ class TableOps:
         exir_ops.edge.aten.ceil.default: torch.ceil,
         exir_ops.edge.aten.erf.default: torch.erf,
         exir_ops.edge.aten.exp.default: torch.exp,
+        exir_ops.edge.aten.expm1.default: torch.expm1,
         exir_ops.edge.aten.floor.default: torch.floor,
         exir_ops.edge.aten.log.default: torch.log,
         exir_ops.edge.aten.reciprocal.default: torch.reciprocal,

backends/arm/arm_backend.py

@@ -217,13 +217,6 @@ def is_vgf(compile_spec: List[CompileSpec]) -> bool:
     return False
 
 
-def get_tosa_spec(compile_spec: List[CompileSpec]) -> TosaSpecification:
-    for spec in compile_spec:
-        if spec.key == "tosa_spec":
-            return TosaSpecification.create_from_string(spec.value.decode())
-    raise ValueError("Could not find TOSA version in CompileSpec")
-
-
 def get_intermediate_path(compile_spec: List[CompileSpec]) -> Optional[str]:
     for spec in compile_spec:
         if spec.key == "debug_artifact_path":

backends/arm/operator_support/tosa_supported_operators.py

@@ -179,6 +179,7 @@ def is_node_supported(
             exir_ops.edge.aten.eq.Scalar,
             exir_ops.edge.aten.erf.default,
             exir_ops.edge.aten.exp.default,
+            exir_ops.edge.aten.expm1.default,
             exir_ops.edge.aten.log.default,
             exir_ops.edge.aten.linear.default,
             exir_ops.edge.aten.split_with_sizes_copy.default,

backends/arm/process_node.py

@@ -14,7 +14,7 @@
 from executorch.backends.arm.operators.node_visitor import NodeVisitor
 from executorch.backends.arm.tosa_mapping import TosaArg
 from executorch.backends.arm.tosa_specification import TosaSpecification
-from executorch.backends.arm.tosa_utils import getNodeArgs, tosa_shape
+from executorch.backends.arm.tosa_utils import tosa_shape
 from torch._export.utils import (
     get_buffer,
     get_lifted_tensor_constant,
@@ -33,7 +33,10 @@ def process_call_function(
     tosa_spec: TosaSpecification,
 ):
     # Unpack arguments and convert
-    inputs = getNodeArgs(node, tosa_spec)
+    try:
+        inputs = [TosaArg(arg, tosa_spec) for arg in node.args]
+    except ValueError as e:
+        raise ValueError(f"Failed processing args to op:\n{node}") from e
 
     # Convert output (this node itself)
     try:

backends/arm/quantizer/arm_quantizer.py

@@ -20,12 +20,11 @@
 from executorch.backends.arm._passes import ArmPassManager
 
 from executorch.backends.arm.quantizer import QuantizationConfig
-from executorch.backends.arm.tosa_specification import TosaSpecification
+from executorch.backends.arm.tosa_specification import get_tosa_spec, TosaSpecification
 
 from .arm_quantizer_utils import is_annotated, mark_node_as_annotated
 from .quantization_annotator import annotate_graph
 from executorch.backends.arm.arm_backend import (
-    get_tosa_spec,
     is_ethosu,
     is_vgf,
 )  # usort: skip