Arm backend: Add 16A8W support and test for cat operation (#14217)

This PR was created by the merge bot to help merge the original PR into
the main branch.
ghstack PR number: #13800 by
@Ninja91
^ Please use this as the source of truth for the PR details, comments,
and reviews
ghstack PR base:
https://github.com/pytorch/executorch/tree/gh/Ninja91/13/base
ghstack PR head:
https://github.com/pytorch/executorch/tree/gh/Ninja91/13/head
Merge bot PR base:
https://github.com/pytorch/executorch/tree/gh/Ninja91/12/orig
Merge bot PR head:
https://github.com/pytorch/executorch/tree/gh/Ninja91/13/orig
@diff-train-skip-merge

---------

Co-authored-by: Nitin Jain <[email protected]>

Author: pytorchbot

backends/arm/test/ops/test_cat.py

@@ -8,8 +8,13 @@
 
 from typing import Tuple
 
+import pytest
 import torch
-from executorch.backends.arm.test import common
+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,
@@ -18,6 +23,8 @@
     TosaPipelineINT,
     VgfPipeline,
 )
+from executorch.backends.arm.tosa.specification import TosaSpecification
+from executorch.backends.xnnpack.test.tester import Quantize
 
 input_t1 = Tuple[torch.Tensor]  # Input x
 
@@ -151,3 +158,105 @@ def test_cat_vgf_INT(test_data: Tuple):
         tosa_version="TOSA-1.0+INT",
     )
     pipeline.run()
+
+
+def get_symmetric_a16w8_cat_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", Cat.test_parameters)
+@pytest.mark.xfail(
+    reason="missing int16 cat ops support; fails at TOSA reference model with Unsupported operation type or rank. See: https://github.com/pytorch/executorch/issues/13978"
+)
+def test_cat_16a8w_tosa_INT(test_data: Tuple):
+    """Test cat operation with 16A8W quantization (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = TosaPipelineINT[input_t1](
+        Cat(),
+        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_cat_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", Cat.test_parameters)
+@common.XfailIfNoCorstone300
+@pytest.mark.xfail(
+    reason="Vela compilation fails with 'Invalid arguments' for int16 cat operations"
+)
+def test_cat_16a8w_u55_INT16(test_data: Tuple):
+    """Test cat operation with 16A8W quantization on U55 (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = EthosU55PipelineINT[input_t1](
+        Cat(),
+        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_cat_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", Cat.test_parameters)
+@common.XfailIfNoCorstone320
+@pytest.mark.xfail(
+    reason="Vela compilation fails with 'Invalid arguments' for int16 cat operations"
+)
+def test_cat_16a8w_u85_INT16(test_data: Tuple):
+    """Test cat operation with 16A8W quantization on U85 (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+
+    pipeline = EthosU85PipelineINT[input_t1](
+        Cat(),
+        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_cat_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()

backends/arm/test/targets.bzl

@@ -15,6 +15,7 @@ def define_arm_tests():
     test_files += [
         "ops/test_add.py",
         "ops/test_avg_pool2d.py",
+        "ops/test_cat.py",
         "ops/test_linear.py", 
         "ops/test_mul.py",
         "ops/test_slice.py",