Merge branch 'main' into where_cleanup

Author: zingo

.ci/docker/ci_commit_pins/buck2.txt

@@ -1 +1 @@
-2024-12-16
+2025-05-06

.github/workflows/apple.yml

@@ -5,6 +5,8 @@ on:
     branches:
       - main
       - release/*
+    tags:
+      - ciflow/trunk/*
   pull_request:
     paths:
       - .ci/scripts/setup-ios.sh

CMakeLists.txt

@@ -44,6 +44,14 @@
 
 cmake_minimum_required(VERSION 3.24)
 project(executorch)
+
+# MARK: - Start EXECUTORCH_H12025_BUILD_MIGRATION --------------------------------------------------
+
+include(${PROJECT_SOURCE_DIR}/tools/cmake/common/preset.cmake)
+include(${PROJECT_SOURCE_DIR}/tools/cmake/preset/default.cmake)
+
+# MARK: - End EXECUTORCH_H12025_BUILD_MIGRATION ----------------------------------------------------
+
 include(tools/cmake/Utils.cmake)
 include(CMakeDependentOption)
 
@@ -96,9 +104,6 @@ set(EXECUTORCH_PAL_DEFAULT
           "Which PAL default implementation to use: one of {posix, minimal}"
 )
 
-option(EXECUTORCH_ENABLE_LOGGING "Build with ET_LOG_ENABLED"
-       ${_default_release_disabled_options}
-)
 if(NOT EXECUTORCH_ENABLE_LOGGING)
   # Avoid pulling in the logging strings, which can be large. Note that this
   # will set the compiler flag for all targets in this directory, and for all

backends/arm/_passes/annotate_decomposed_matmul.py

@@ -1,13 +1,12 @@
 # Copyright 2024-2025 Arm Limited and/or its affiliates.
-# All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
 # pyre-unsafe
 
 import itertools
-
+import operator
 from typing import List
 
 import torch
@@ -22,7 +21,7 @@
 
 class AnnotateDecomposedMatmulPass(ExportPass):
     """
-    torch.matmul can be decomposed in many ways, for instance:
+    torch.matmul and it's equivalent operator @ can be decomposed in many ways, for instance:
     dq -> matmul -> q can become
     dq -> repeat -> view -> bmm -> view -> dq which makes quantization folding
     difficult. This helper function find all matmul partitions and annotate its
@@ -50,6 +49,7 @@ def call(self, graph_module: GraphModule) -> PassResult:
             graph_module.graph,
             [
                 torch.matmul,
+                operator.matmul,
             ],
             None,
         )

backends/arm/operator_support/pool_2d_support.py

@@ -54,8 +54,11 @@ def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
         kernel = cast(tuple[int, int], node.args[1])
         stride = cast(tuple[int, int], node.args[2])
         if len(node.args) > 3:
+            padding = cast(tuple[int, int], node.args[3])
             # Padding case
-            if not all(1 <= k <= 8 for k in kernel):
+            if not all(1 <= k <= 8 for k in kernel) and not all(
+                v == 0 for v in padding
+            ):
                 self.reporter.report_reject(
                     node, f"Avgpool2d with padding needs kernel dims < 8, got {kernel}"
                 )

backends/arm/operator_support/tosa_supported_operators.py

@@ -335,6 +335,7 @@ def _is_matmul_node_supported(
                 graph_module.graph,
                 [
                     torch.matmul,
+                    operator.matmul,
                 ],
                 None,
             )
@@ -385,7 +386,7 @@ def is_node_supported(
         ):
             source_fn_stack: tuple[typing.Any] = node.meta.get("source_fn_stack", [])
             if len(source_fn_stack) > 0:
-                if source_fn_stack[-1][1] in (torch.matmul,):
+                if source_fn_stack[-1][1] in (torch.matmul, operator.matmul):
                     return self._is_matmul_node_supported(submodules, node)
 
         elif node.target in (exir_ops.edge.aten.max_pool2d_with_indices.default,):

backends/arm/operators/op_max_pool2d.py

@@ -23,6 +23,24 @@
 from executorch.backends.arm.tosa_specification import TosaSpecification
 
 
+# Similarly to Conv2d, the TOSA spec requires that following is exactly divisible:
+# `(input + 2 * pad - kernel_size) / stride`
+# PyTorch however, does not require this, so as needed, we must adjust the padding.
+def adjust_pad_if_needed(
+    input_size: int, kernel_size: int, stride: int, pad: int
+) -> int:
+    if pad == 0:
+        return pad
+
+    mod_remainder = (input_size + 2 * pad - kernel_size) % stride
+
+    # No need to adjust
+    if mod_remainder == 0:
+        return pad
+
+    return pad - mod_remainder
+
+
 @register_node_visitor
 class MaxPool2dVisitor_0_80(NodeVisitor):
     target = "aten.max_pool2d.default"
@@ -61,6 +79,20 @@ def define_node(
         except IndexError:
             pad_size_list = [0, 0, 0, 0]
 
+        # Adjust the padding as necessary
+        pad_size_list[1] = adjust_pad_if_needed(
+            input_tensor.shape[2],
+            kernel_size[0],
+            stride[0],
+            pad_size_list[1],
+        )
+        pad_size_list[3] = adjust_pad_if_needed(
+            input_tensor.shape[3],
+            kernel_size[1],
+            stride[1],
+            pad_size_list[3],
+        )
+
         accumulator_type = output.dtype
 
         # Initilize zero point to zero.
@@ -131,6 +163,20 @@ def define_node(
         except IndexError:
             pad_size_list = [0, 0, 0, 0]
 
+        # Adjust the padding as necessary
+        pad_size_list[1] = adjust_pad_if_needed(
+            input_tensor.shape[2],
+            kernel_size[0],
+            stride[0],
+            pad_size_list[1],
+        )
+        pad_size_list[3] = adjust_pad_if_needed(
+            input_tensor.shape[3],
+            kernel_size[1],
+            stride[1],
+            pad_size_list[3],
+        )
+
         attr = ts.TosaSerializerAttribute()
         attr.MaxPool2dAttribute(
             kernel=kernel_size, stride=stride, pad=pad_size_list, nan_mode=1

backends/arm/operators/op_slice.py

@@ -68,8 +68,16 @@ def define_node(
         end_index = _fixup_end(end, shape, dim)
         size = end_index - start_index
 
-        assert size > 0
-        assert size <= shape[dim]
+        if size <= 0:
+            raise ValueError(
+                f"The calculated slice size must be positive. Got {size=} "
+                f"with {start_index=} and {end_index=}."
+            )
+        if size > shape[dim]:
+            raise ValueError(
+                f"The calculated slice size cannot be greater than the dimension size"
+                f". Got {size=} and {shape[dim]=}."
+            )
 
         # Convert aten args to Tosa's start and size attributes and in TOSA dim order.
         attr = ts.TosaSerializerAttribute()
@@ -122,8 +130,16 @@ def define_node(
         end_index = _fixup_end(end, shape, dim)
         size = end_index - start_index
 
-        assert size > 0
-        assert size <= shape[dim]
+        if size <= 0:
+            raise ValueError(
+                f"The calculated slice size must be positive. Got {size=} "
+                f"with {start_index=} and {end_index=}."
+            )
+        if size > shape[dim]:
+            raise ValueError(
+                f"The calculated slice size cannot be greater than the dimension size"
+                f". Got {size=} and {shape[dim]=}."
+            )
 
         # Convert aten args to Tosa's start and size shape_t tensors and in TOSA dim order.
         starts = [

backends/arm/scripts/parse_test_names.py

@@ -5,7 +5,15 @@
 from executorch.exir.dialects.edge.spec.utils import SAMPLE_INPUT
 
 # Add edge ops which we lower but which are not included in exir/dialects/edge/edge.yaml here.
-CUSTOM_EDGE_OPS = ["linspace.default", "eye.default"]
+CUSTOM_EDGE_OPS = [
+    "linspace.default",
+    "eye.default",
+    "hardsigmoid.default",
+    "hardswish.default",
+    "linear.default",
+    "maximum.default",
+    "adaptive_avg_pool2d.default",
+]
 ALL_EDGE_OPS = SAMPLE_INPUT.keys() | CUSTOM_EDGE_OPS
 
 # Add all targets and TOSA profiles we support here.

backends/arm/test/common.py

@@ -259,17 +259,15 @@ def decorator_func(func):
                     raise RuntimeError(
                         "xfail info needs to be str, or tuple[str, type[Exception]]"
                     )
-                pytest_param = pytest.param(
-                    test_parameters,
-                    id=id,
-                    marks=pytest.mark.xfail(
-                        reason=reason, raises=raises, strict=strict
-                    ),
+                # Set up our fail marker
+                marker = (
+                    pytest.mark.xfail(reason=reason, raises=raises, strict=strict),
                 )
             else:
-                pytest_param = pytest.param(test_parameters, id=id)
-            pytest_testsuite.append(pytest_param)
+                marker = ()
 
+            pytest_param = pytest.param(test_parameters, id=id, marks=marker)
+            pytest_testsuite.append(pytest_param)
         return pytest.mark.parametrize(arg_name, pytest_testsuite)(func)
 
     return decorator_func