Add 16A8W support and test for mul operation

Pull Request resolved: #13795

Add 16A8W quantization support and test for the mul operation in ExecutorTorch ARM backend.

This follows the pattern established for linear operations, extending int16 support to mul operations.

Changes:
- Add INT16 dtype validation support in op_mul.py
- Add test_mul_tensor_16a8w_tosa_INT test function
- Enable test_mul.py in test targets configuration

The 16A8W configuration uses 16-bit activations with 8-bit weights, enabling higher precision for activations while maintaining weight efficiency.
ghstack-source-id: 307682119
@exported-using-ghexport

Differential Revision: [D80510628](https://our.internmc.facebook.com/intern/diff/D80510628/)

Author: Ninja91

backends/arm/operators/op_mul.py

@@ -51,7 +51,7 @@ def define_node(
         validate_valid_dtype(
             self.target,
             [*inputs, output],
-            [ts.DType.INT8, ts.DType.INT32],
+            [ts.DType.INT8, ts.DType.INT16, ts.DType.INT32],
             output.tosa_spec,
         )
 
@@ -80,15 +80,15 @@ def define_node(
                 tosa_spec=self.tosa_spec,
             )
         else:
-            # input[0].dtype == ts.DType.INT32
+            # input[0].dtype == ts.DType.INT16 or ts.DType.INT32
             # Non quantized input, natively support by TOSA.MUL
             input_A_rescaled, input_B_rescaled = inputs[0], inputs[1]
 
         if output.dtype == ts.DType.INT8:
             output_shape = tutils.tosa_shape(output.shape, output.dim_order)
             mul_output = tosa_graph.addIntermediate(output_shape, ts.DType.INT32)
         else:
-            # output.dtype == ts.DType.INT32
+            # output.dtype == ts.DType.INT16 or ts.DType.INT32
             mul_output = output
 
         # Do the INT32 Mul

backends/arm/test/ops/test_mul.py

@@ -8,16 +8,23 @@
 
 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
 
 input_t1 = Tuple[torch.Tensor, torch.Tensor]  # Input x
 aten_op = "torch.ops.aten.mul.Tensor"
@@ -284,3 +291,102 @@ def test_mul_tensor_vgf_INT_int32(test_data: torch.Tensor):
     )
     pipeline.pop_stage("check.quant_nodes")
     pipeline.run()
+
+
+def get_symmetric_a16w8_mul_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)
+def test_mul_tensor_16a8w_tosa_INT(test_data: input_t1):
+    """Test mul operation with 16A8W quantization (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = TosaPipelineINT[input_t1](
+        Mul(),
+        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_mul_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 mul operations. See: https://github.com/pytorch/executorch/issues/13947"
+)
+def test_mul_tensor_16a8w_u55_INT16(test_data: input_t1):
+    """Test mul operation with 16A8W quantization on U55 (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = EthosU55PipelineINT[input_t1](
+        Mul(),
+        test_data(),
+        aten_op,
+        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_mul_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 mul operations. See: https://github.com/pytorch/executorch/issues/13947"
+)
+def test_mul_tensor_16a8w_u85_INT16(test_data: input_t1):
+    """Test mul operation with 16A8W quantization on U85 (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = EthosU85PipelineINT[input_t1](
+        Mul(),
+        test_data(),
+        aten_op,
+        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_mul_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()

backends/arm/test/targets.bzl

@@ -16,6 +16,7 @@ def define_arm_tests():
         "ops/test_add.py",
         "ops/test_avg_pool2d.py",
         "ops/test_linear.py", 
+        "ops/test_mul.py",
         "ops/test_slice.py",
         "ops/test_sigmoid.py",
         "ops/test_tanh.py",