From ab4c4737ecbd227443699c9706737db30e7f1871 Mon Sep 17 00:00:00 2001
From: Nitin Jain <jainnitin@meta.com>
Date: Thu, 28 Aug 2025 23:42:55 -0700
Subject: [PATCH] Add 16A8W support for view and transpose operations

Add 16A8W quantization support for view and transpose operations in ExecutorTorch ARM backend.

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

Changes:
- Add INT16 dtype validation support in op_transpose.py
- Add test_view_tensor_16a8w_tosa_INT test function
- Enable test_view.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.

Differential Revision: [D80511313](https://our.internmc.facebook.com/intern/diff/D80511313/)

[ghstack-poisoned]
---
 backends/arm/test/ops/test_view.py | 111 ++++++++++++++++++++++++++++-
 backends/arm/test/targets.bzl      |   1 +
 2 files changed, 111 insertions(+), 1 deletion(-)

diff --git a/backends/arm/test/ops/test_view.py b/backends/arm/test/ops/test_view.py
index 0aa6f9a0245..bb1e36fd133 100644
--- a/backends/arm/test/ops/test_view.py
+++ b/backends/arm/test/ops/test_view.py
@@ -9,9 +9,14 @@
 
 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,
@@ -20,6 +25,8 @@
     TosaPipelineINT,
     VgfPipeline,
 )
+from executorch.backends.arm.tosa_specification import TosaSpecification
+from executorch.backends.xnnpack.test.tester import Quantize
 
 aten_op = "torch.ops.aten.view.default"
 
@@ -147,3 +154,105 @@ def test_view_u85_INT(test_data: Tuple):
         exir_ops=[],
     )
     pipeline.run()
+
+
+def get_symmetric_a16w8_view_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", View.needs_transpose_tests)
+def test_view_16a8w_tosa_INT(test_data: Tuple):
+    """Test view operation with 16A8W quantization (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+    test_tensor, new_shape = test_data()
+
+    pipeline = TosaPipelineINT[input_t1](
+        View(new_shape),
+        (test_tensor,),
+        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_view_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", View.needs_transpose_tests)
+@common.XfailIfNoCorstone300
+@pytest.mark.xfail(
+    reason="Vela compilation fails with 'Invalid arguments' for int16 view operations"
+)
+def test_view_16a8w_u55_INT16(test_data: Tuple):
+    """Test view operation with 16A8W quantization on U55 (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+    test_tensor, new_shape = test_data()
+
+    pipeline = EthosU55PipelineINT[input_t1](
+        View(new_shape),
+        (test_tensor,),
+        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_view_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_data", View.needs_transpose_tests)
+@common.XfailIfNoCorstone320
+@pytest.mark.xfail(
+    reason="Vela compilation fails with 'Invalid arguments' for int16 view operations"
+)
+def test_view_16a8w_u85_INT16(test_data: Tuple):
+    """Test view operation with 16A8W quantization on U85 (16-bit activations, 8-bit weights)"""
+    per_channel_quantization = False
+    test_tensor, new_shape = test_data()
+
+    pipeline = EthosU85PipelineINT[input_t1](
+        View(new_shape),
+        (test_tensor,),
+        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_view_quantizer(
+            per_channel_quantization=per_channel_quantization
+        ),
+    )
+    pipeline.run()
diff --git a/backends/arm/test/targets.bzl b/backends/arm/test/targets.bzl
index b438e556cca..5714039d134 100644
--- a/backends/arm/test/targets.bzl
+++ b/backends/arm/test/targets.bzl
@@ -20,6 +20,7 @@ def define_arm_tests():
         "ops/test_slice.py",
         "ops/test_sigmoid.py",
         "ops/test_tanh.py",
+        "ops/test_view.py",
         "ops/test_cos.py",
     ]