Arm backend: Add sinh decomposition pass and test (pytorch#11848)

Decomposes sinh into other operators/lookup table for MI/BI case.

Signed-off-by: Teo Bergkvist <[email protected]>
Signed-off-by: Emma Kujala <[email protected]>

Author: tbergkvist

backends/arm/_passes/__init__.py

@@ -39,6 +39,7 @@
 from .decompose_round_pass import DecomposeRoundPass  # noqa
 from .decompose_select import DecomposeSelectPass  # noqa
 from .decompose_silu_pass import DecomposeSiluPass  # noqa
+from .decompose_sinh_pass import DecomposeSinhPass  # noqa
 from .decompose_softmax_pass import DecomposeSoftmaxPass  # noqa
 from .decompose_softmax_unstable_pass import DecomposeSoftmaxUnstablePass  # noqa
 from .decompose_sqrt_pass import DecomposeSqrtPass  # noqa

backends/arm/_passes/arm_pass_manager.py

@@ -42,6 +42,7 @@
     DecomposeRoundPass,
     DecomposeSelectPass,
     DecomposeSiluPass,
+    DecomposeSinhPass,
     DecomposeSoftmaxPass,
     DecomposeSoftmaxUnstablePass,
     DecomposeSqrtPass,
@@ -151,6 +152,7 @@ def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(DecomposeSqrtPass())
         self.add_pass(ConvertIntPowToMuls())
         self.add_pass(CastBoolToInt8Pass())
+        self.add_pass(DecomposeSinhPass())
         self.add_pass(ReplaceScalarWithTensorArgPassTOSAMI())
         self.add_pass(DecomposeEmbeddingPass())
         self.add_pass(FuseQuantizedActivationPass())

backends/arm/_passes/decompose_sinh_pass.py

@@ -0,0 +1,53 @@
+# 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
+
+
+# For MI case
+edge_sinh = exir_ops.edge.aten.sinh.default
+
+
+class DecomposeSinhPass(ArmPass):
+    """
+    A decomposition pass that decomposes Sinh operations into a
+    combination of supported TOSA-equivalent operations (MI).
+
+    Supported input ops:
+        - exir_ops.edge.aten.sinh.default
+
+    These are decomposed into exponentials, negation, subtraction,
+        and scalar multiplication.
+    """
+
+    def call_operator(self, op, args, kwargs, meta):
+        if op is not edge_sinh:
+            return super().call_operator(op, args, kwargs, meta)
+
+        x = args
+
+        sub_op, exp_op, neg_op, mul_op = (
+            exir_ops.edge.aten.sub.Tensor,
+            exir_ops.edge.aten.exp.default,
+            exir_ops.edge.aten.neg.default,
+            exir_ops.edge.aten.mul.Scalar,
+        )
+
+        # Exponential 1
+        exp1 = super().call_operator(exp_op, x, {}, meta, updated=True)
+
+        # Exponential 2
+        neg_x = super().call_operator(neg_op, x, {}, meta, updated=True)
+        exp2 = super().call_operator(exp_op, (neg_x,), {}, meta, updated=True)
+
+        # Subtraction
+        sub = super().call_operator(sub_op, (exp1, exp2), {}, meta, updated=True)
+
+        # Multiplication
+        out = super().call_operator(mul_op, (sub, 0.5), {}, meta, updated=True)
+
+        return out

backends/arm/_passes/insert_table_ops.py

@@ -53,6 +53,7 @@ class TableOps:
         exir_ops.edge.aten.tanh.default: torch.tanh,
         exir_ops.edge.aten.hardsigmoid.default: torch.nn.functional.hardsigmoid,
         exir_ops.edge.aten.hardswish.default: torch.nn.functional.hardswish,
+        exir_ops.edge.aten.sinh.default: torch.sinh,
     }
 
     # Targets that must be treated explicitly

backends/arm/operator_support/tosa_supported_operators.py

@@ -243,6 +243,7 @@ def is_node_supported(
             torch.ops.aten.scalar_tensor.default,
             exir_ops.edge.aten.gelu.default,
             exir_ops.edge.aten.alias_copy.default,
+            exir_ops.edge.aten.sinh.default,
         ]
 
         return supported

backends/arm/quantizer/quantization_annotator.py

@@ -213,6 +213,7 @@ def _match_pattern(
     torch.ops.aten.full_like.default,
     torch.ops.aten.pow.Tensor_Scalar,
     torch.ops.aten.gelu.default,
+    torch.ops.aten.sinh.default,
 ]
 
 _one_to_one_shared_input_qspec = [

backends/arm/test/ops/test_sinh.py

@@ -0,0 +1,78 @@
+# 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 Tuple
+
+import torch
+from executorch.backends.arm.test import common
+from executorch.backends.arm.test.tester.test_pipeline import (
+    EthosU55PipelineBI,
+    EthosU85PipelineBI,
+    TosaPipelineBI,
+    TosaPipelineMI,
+)
+
+aten_op = "torch.ops.aten.sinh.default"
+exir_op = "executorch_exir_dialects_edge__ops_aten__sinh_default"
+
+
+input_t1 = Tuple[torch.Tensor]  # Input x
+
+test_data_suite = {
+    # (test_name, test_data)
+    "zeros": torch.zeros(10, 10, 10),
+    "zeros_alt_shape": torch.zeros(10, 3, 5),
+    "ones": torch.ones(10, 10, 10),
+    "rand": torch.rand(10, 10) - 0.5,
+    "rand_alt_shape": torch.rand(10, 3, 5) - 0.5,
+    "randn_pos": torch.randn(10) + 10,
+    "randn_neg": torch.randn(10) - 10,
+    "ramp": torch.arange(-16, 16, 0.2),
+    "large": 100 * torch.ones(1, 1),
+    "small": 0.000001 * torch.ones(1, 1),
+}
+
+
+class Sinh(torch.nn.Module):
+
+    def forward(self, x: torch.Tensor):
+        return torch.sinh(x)
+
+
+@common.parametrize("test_data", test_data_suite)
+def test_sinh_tosa_MI(test_data: Tuple):
+    pipeline = TosaPipelineMI[input_t1](
+        Sinh(),
+        (test_data,),
+        aten_op,
+        exir_op,
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", test_data_suite)
+def test_sinh_tosa_BI(test_data: Tuple):
+    pipeline = TosaPipelineBI[input_t1](
+        Sinh(), (test_data,), aten_op=aten_op, exir_op=exir_op
+    )
+    pipeline.run()
+
+
+@common.XfailIfNoCorstone300
+@common.parametrize("test_data", test_data_suite)
+def test_sinh_u55_BI(test_data: Tuple):
+    pipeline = EthosU55PipelineBI[input_t1](
+        Sinh(), (test_data,), aten_ops=aten_op, exir_ops=exir_op
+    )
+    pipeline.run()
+
+
+@common.XfailIfNoCorstone320
+@common.parametrize("test_data", test_data_suite)
+def test_sinh_u85_BI(test_data: Tuple):
+    pipeline = EthosU85PipelineBI[input_t1](
+        Sinh(), (test_data,), aten_ops=aten_op, exir_ops=exir_op
+    )
+    pipeline.run()