Add 16A8W support and test for add operation

Pull Request resolved: #13789

Add 16A8W quantization support and comprehensive tests for the add operation in ExecutorTorch ARM backend targeting Ethos U55 and U85 NPUs.

This follows the pattern established for linear operations, extending int16 support to add operations with hardware-specific testing.

Changes:
- Add INT16 dtype validation support in op_add.py
- Add test_add_tensor_16a8w_tosa_INT test function with U55/U85 pipeline support
- Add U55 and U85 specific 16A8W tests with proper xfail decorators
- Fix U55/U85 test parameter usage (remove unsupported tosa_extensions, clean quantizer function calls)
- Update xfail reasons to consistent 'Vela compilation fails with Invalid arguments' pattern

ghstack-source-id: 306430209
ghstack-source-id: 306430209
@exported-using-ghexport

Differential Revision: [D80510463](https://our.internmc.facebook.com/intern/diff/D80510463/)

Author: Ninja91

backends/arm/operators/op_add.py

@@ -47,10 +47,14 @@ def define_node(
 
         validate_num_inputs(self.target, inputs, 2)
         validate_same_dtype(self.target, [*inputs, output], ts)
+        valid_dtypes = []
+        if self.tosa_spec.support_integer():
+            valid_dtypes.extend([ts.DType.INT8, ts.DType.INT16, ts.DType.INT32])
+
         validate_valid_dtype(
             self.target,
             [*inputs, output],
-            [ts.DType.INT8, ts.DType.INT32],
+            valid_dtypes,
             output.tosa_spec,
         )
 

backends/arm/test/ops/test_add.py

@@ -10,6 +10,10 @@
 import pytest
 import torch
 from executorch.backends.arm.quantizer import arm_quantizer
+from executorch.backends.arm.quantizer.arm_quantizer import (
+    get_symmetric_a16w8_quantization_config,
+    TOSAQuantizer,
+)
 from executorch.backends.arm.test import common, conftest
 from executorch.backends.arm.test.tester.test_pipeline import (
     EthosU55PipelineINT,
@@ -216,3 +220,105 @@ def test_add_tensor_vgf_INT(test_data: input_t1):
         tosa_version="TOSA-1.0+INT",
     )
     pipeline.run()
+
+
+def get_symmetric_a16w8_add_quantizer(per_channel_quantization=False):
+    tosa_version = conftest.get_option("tosa_version")
+    tosa_profiles = {
+        "1.0": TosaSpecification.create_from_string("TOSA-1.0+INT+int16"),
+    }
+
+    quantizer = TOSAQuantizer(tosa_profiles[tosa_version])
+    quantizer.set_global(
+        get_symmetric_a16w8_quantization_config(is_per_channel=per_channel_quantization)
+    )
+
+    return Quantize(
+        quantizer,
+        get_symmetric_a16w8_quantization_config(
+            is_per_channel=per_channel_quantization
+        ),
+    )
+
+
+@common.parametrize("test_data", Add.test_data)
+@pytest.mark.xfail(
+    reason="missing int16 add ops support; fails at TOSA reference model with Unsupported operation type or rank"
+)
+def test_add_tensor_16a8w_tosa_INT(test_data: input_t1):
+    """Test add operation with 16A8W quantization (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = TosaPipelineINT[input_t1](
+        Add(),
+        test_data(),
+        aten_op,
+        exir_op=[],
+        per_channel_quantization=per_channel_quantization,
+        use_to_edge_transform_and_lower=True,
+        tosa_extensions=["int16"],
+    )
+
+    pipeline.change_args(
+        "quantize",
+        get_symmetric_a16w8_add_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", Add.test_data)
+@common.XfailIfNoCorstone300
+@pytest.mark.xfail(
+    reason="Vela compilation fails with 'Invalid arguments' for int16 add operations"
+)
+def test_add_tensor_16a8w_u55_INT16(test_data: input_t1):
+    """Test add operation with 16A8W quantization on U55 (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = EthosU55PipelineINT[input_t1](
+        Add(),
+        test_data(),
+        aten_op,
+        exir_op,
+        per_channel_quantization=per_channel_quantization,
+        use_to_edge_transform_and_lower=True,
+        run_on_fvp=True,
+    )
+
+    pipeline.change_args(
+        "quantize",
+        get_symmetric_a16w8_add_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", Add.test_data)
+@common.XfailIfNoCorstone320
+@pytest.mark.xfail(
+    reason="Vela compilation fails with 'Invalid arguments' for int16 add operations"
+)
+def test_add_tensor_16a8w_u85_INT16(test_data: input_t1):
+    """Test add operation with 16A8W quantization on U85 (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = EthosU85PipelineINT[input_t1](
+        Add(),
+        test_data(),
+        aten_op,
+        exir_op,
+        per_channel_quantization=per_channel_quantization,
+        use_to_edge_transform_and_lower=True,
+        run_on_fvp=True,
+    )
+
+    pipeline.change_args(
+        "quantize",
+        get_symmetric_a16w8_add_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()

backends/arm/test/targets.bzl

@@ -13,6 +13,7 @@ def define_arm_tests():
 
     # Operators
     test_files += [
+        "ops/test_add.py",
         "ops/test_avg_pool2d.py",
         "ops/test_linear.py", 
         "ops/test_slice.py",