Add 16A8W FCNode support with BMM dependency fix

Pull Request resolved: #13801

Add 16A8W quantization support for FCNode operations with BMM dependency fix in ExecutorTorch ARM backend.

This follows the pattern established for linear, mul, sigmoid, tanh, slice, view/transpose, and cat operations, extending int16 support to FCNode operations.

Changes:
- Add INT16 dtype validation support in op_bmm.py
- Add test_addmm_tensor_16a8w_tosa_INT test function
- Enable test_addmm.py in test targets configuration
- Fix BMM dependency for FCNode operations

The 16A8W configuration uses 16-bit activations with 8-bit weights, enabling higher precision for activations while maintaining weight efficiency.
ghstack-source-id: 308986674
@exported-using-ghexport

Differential Revision: [D80512504](https://our.internmc.facebook.com/intern/diff/D80512504/)

Author: Ninja91

backends/arm/operators/op_bmm.py

@@ -55,7 +55,7 @@ def define_node(
         validate_valid_dtype(
             self.target,
             [*inputs, output],
-            [ts.DType.INT8, ts.DType.FP32],
+            [ts.DType.INT8, ts.DType.INT16, ts.DType.FP32],
             output.tosa_spec,
         )
 

backends/arm/test/ops/test_addmm.py

@@ -5,16 +5,23 @@
 
 from typing import Tuple
 
+import pytest
 import torch
+from executorch.backends.arm.quantizer.arm_quantizer import (
+    get_symmetric_a16w8_quantization_config,
+    TOSAQuantizer,
+)
 
-from executorch.backends.arm.test import common
+from executorch.backends.arm.test import common, conftest
 from executorch.backends.arm.test.tester.test_pipeline import (
     EthosU55PipelineINT,
     EthosU85PipelineINT,
     TosaPipelineFP,
     TosaPipelineINT,
     VgfPipeline,
 )
+from executorch.backends.arm.tosa.specification import TosaSpecification
+from executorch.backends.xnnpack.test.tester import Quantize
 
 aten_op = "torch.ops.aten.addmm.default"
 
@@ -182,3 +189,105 @@ def test_addmm_vgf_INT(test_data: input_t1):
         tosa_version="TOSA-1.0+INT",
     )
     pipeline.run()
+
+
+def get_symmetric_a16w8_addmm_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", test_data_suite)
+@pytest.mark.xfail(
+    reason="missing int16 addmm ops support; fails at TOSA reference model with Unsupported operation type or rank. See: https://github.com/pytorch/executorch/issues/13979"
+)
+def test_addmm_16a8w_tosa_INT(test_data: input_t1):
+    """Test addmm (FC layer) operation with 16A8W quantization (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = TosaPipelineINT[input_t1](
+        Addmm(),
+        (*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_addmm_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", test_data_suite)
+@common.XfailIfNoCorstone300
+@pytest.mark.xfail(
+    reason="Vela compilation fails with 'Invalid arguments' for int16 addmm operations"
+)
+def test_addmm_16a8w_u55_INT16(test_data: input_t1):
+    """Test addmm (FC layer) operation with 16A8W quantization on U55 (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = EthosU55PipelineINT[input_t1](
+        Addmm(),
+        (*test_data,),
+        aten_ops=[],
+        exir_ops=[],
+        per_channel_quantization=per_channel_quantization,
+        use_to_edge_transform_and_lower=True,
+        run_on_fvp=True,
+    )
+
+    pipeline.change_args(
+        "quantize",
+        get_symmetric_a16w8_addmm_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", test_data_suite)
+@common.XfailIfNoCorstone320
+@pytest.mark.xfail(
+    reason="Vela compilation fails with 'Invalid arguments' for int16 addmm operations"
+)
+def test_addmm_16a8w_u85_INT16(test_data: input_t1):
+    """Test addmm (FC layer) operation with 16A8W quantization on U85 (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = EthosU85PipelineINT[input_t1](
+        Addmm(),
+        (*test_data,),
+        aten_ops=[],
+        exir_ops=[],
+        per_channel_quantization=per_channel_quantization,
+        use_to_edge_transform_and_lower=True,
+        run_on_fvp=True,
+    )
+
+    pipeline.change_args(
+        "quantize",
+        get_symmetric_a16w8_addmm_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()

backends/arm/test/targets.bzl

@@ -14,6 +14,7 @@ def define_arm_tests():
     # Operators
     test_files += [
         "ops/test_add.py",
+        "ops/test_addmm.py",
         "ops/test_avg_pool2d.py",
         "ops/test_cat.py",
         "ops/test_linear.py",