Update on "Add NOSTDINC and NOTORCHINC lintrunner"

Prohibit standard C++ container, aten, and torch includes in core, at least pending further review. Note that non-core directories, such as extension/, test/, and any testing_util/ subdirectories, are exempted.

Differential Revision: [D65541332](https://our.internmc.facebook.com/intern/diff/D65541332/)

[ghstack-poisoned]

Author: swolchok

.github/workflows/ghstack_land.yml

@@ -5,6 +5,7 @@ on:
     branches:
       - 'gh/cccclai/[0-9]+/base'
       - 'gh/dbort/[0-9]+/base'
+      - 'gh/dvorjackz/[0-9]+/base'
       - 'gh/guangy10/[0-9]+/base'
       - 'gh/helunwencser/[0-9]+/base'
       - 'gh/jorgep31415/[0-9]+/base'

backends/arm/TARGETS

@@ -1,3 +1,4 @@
+# @noautodeps
 load("@fbcode_macros//build_defs:python_library.bzl", "python_library")
 
 python_library(

backends/arm/_passes/arm_pass_manager.py

@@ -43,6 +43,7 @@
 from executorch.backends.arm._passes.unsqueeze_scalar_placeholders_pass import (
     UnsqueezeScalarPlaceholdersPass,
 )
+from executorch.backends.xnnpack._passes.remove_getitem_op import RemoveGetItemPass
 from executorch.exir import ExportedProgram
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from executorch.exir.pass_manager import PassManager
@@ -58,6 +59,7 @@ def transform_to_backend_pipeline(
     ):
         """Apply passes before transforming program to backend"""
         self.add_pass(CastInt64ToInt32Pass(exported_program))
+        self.add_pass(RemoveGetItemPass())
         self.add_pass(UnsqueezeScalarPlaceholdersPass(exported_program))
         self.add_pass(SizeAdjustConv2DPass())
         self.add_pass(RemoveClonePass())

backends/arm/arm_backend.py

@@ -13,7 +13,7 @@
 
 import logging
 import os
-from typing import final, List, Optional
+from typing import cast, final, List, Optional
 
 import serializer.tosa_serializer as ts
 from executorch.backends.arm.arm_vela import vela_compile
@@ -31,6 +31,7 @@
 from executorch.exir.backend.backend_details import BackendDetails, PreprocessResult
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from torch.export.exported_program import ExportedProgram
+from torch.fx import Node
 
 # TOSA backend debug functionality
 logger = logging.getLogger(__name__)
@@ -225,6 +226,7 @@ def preprocess(  # noqa: C901
         node_visitors = get_node_visitors(edge_program)
 
         for node in graph_module.graph.nodes:
+            node = cast(Node, node)
             if node.op == "call_function":
                 process_call_function(node, tosa_graph, node_visitors)
             elif node.op == "placeholder":
@@ -236,9 +238,6 @@ def preprocess(  # noqa: C901
                 # any checking of compatibility.
                 dbg_fail(node, tosa_graph, artifact_path)
 
-        # TODO: It would be awesome if this dump could somehow be done on top level and not here.
-        # Problem is that the desc.json has to be created on the tosa_graph object, which we can't
-        # access from top level.
         if artifact_path:
             tag = _get_first_delegation_tag(graph_module)
             dbg_tosa_dump(
@@ -259,6 +258,4 @@ def preprocess(  # noqa: C901
         else:
             raise RuntimeError(f"Unknown format {output_format}")
 
-        # Continueing from above. Can I put tosa_graph into this function?
-        # debug_handle_map = ...
         return PreprocessResult(processed_bytes=binary)

backends/arm/arm_partitioner.py

@@ -55,6 +55,7 @@ def is_node_supported(self, submodules, node: torch.fx.Node) -> bool:
             exir_ops.edge.aten._native_batch_norm_legit_no_training.default,
             exir_ops.edge.aten.native_layer_norm.default,
             exir_ops.edge.aten.avg_pool2d.default,
+            exir_ops.edge.aten.max_pool2d_with_indices.default,
             exir_ops.edge.aten.sigmoid.default,
             exir_ops.edge.aten.mm.default,
             exir_ops.edge.aten.repeat.default,

backends/arm/operators/TARGETS

@@ -1,3 +1,4 @@
+# @noautodeps
 load("@fbcode_macros//build_defs:python_library.bzl", "python_library")
 
 python_library(

backends/arm/operators/__init__.py

@@ -20,6 +20,7 @@
     op_get_item,
     op_hardtanh,
     op_log,
+    op_max_pool2d,
     op_mm,
     op_mul,
     op_permute,

backends/arm/operators/op_max_pool2d.py

@@ -0,0 +1,77 @@
+# Copyright 2024 Arm Limited and/or its affiliates.
+#
+# 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
+from typing import cast, List
+
+import serializer.tosa_serializer as ts
+import torch
+from executorch.backends.arm.operators.node_visitor import (
+    NodeVisitor,
+    register_node_visitor,
+)
+from executorch.backends.arm.tosa_mapping import TosaArg
+from executorch.backends.arm.tosa_utils import get_quant_node_args
+
+from serializer.tosa_serializer import TosaOp
+
+
+@register_node_visitor
+class MaxPool2dVisitor(NodeVisitor):
+    target = "aten.max_pool2d.default"
+
+    def __init__(self, *args):
+        super().__init__(*args)
+
+    def define_node(
+        self,
+        node: torch.fx.Node,
+        tosa_graph: ts.TosaSerializer,
+        inputs: List[TosaArg],
+        output: TosaArg,
+        is_quant_node: bool,
+    ) -> None:
+
+        input_tensor = inputs[0]
+        kernel_size = inputs[1].special
+        stride = inputs[2].special
+
+        try:
+            padding = [*inputs[3].special, *inputs[3].special]
+        except IndexError:
+            padding = [0, 0, 0, 0]
+
+        accumulator_type = input_tensor.dtype
+
+        if is_quant_node:
+            # Accumulator type always is int8 when input tensor is an integer type.
+            accumulator_type = ts.DType.INT8
+
+        # Initilize zero point to zero.
+        input_zp = 0
+        output_zp = 0
+
+        if is_quant_node:
+            input_zp = get_quant_node_args(
+                cast(torch.fx.Node, node.all_input_nodes[0])
+            ).zp
+            output_zp = get_quant_node_args(list(node.users)[0]).zp
+
+        attr = ts.TosaSerializerAttribute()
+        attr.PoolAttribute(
+            kernel=kernel_size,
+            stride=stride,
+            pad=padding,
+            input_zp=input_zp,
+            output_zp=output_zp,
+            accum_dtype=accumulator_type,
+        )
+
+        tosa_graph.addOperator(
+            TosaOp.Op().MAX_POOL2D,
+            [input_tensor.name],
+            [output.name],
+            attr,
+        )

backends/arm/quantizer/arm_quantizer.py

@@ -268,7 +268,6 @@ class ArmQuantizer(Quantizer):
         "sub",
         "mul",
         "mm",
-        "cat",
         "one_to_one",
         "generic",
         "sum",

backends/arm/quantizer/arm_quantizer_utils.py

@@ -144,21 +144,11 @@ def is_share_obs_or_fq_op(op: Callable) -> bool:
         torch.ops.aten.mean.dim,
         torch.ops.aten.permute.default,
         torch.ops.aten.permute_copy.default,
-        torch.ops.aten.squeeze.dim,
-        torch.ops.aten.squeeze.dims,
-        torch.ops.aten.squeeze.default,
-        torch.ops.aten.squeeze_copy.dim,
-        torch.ops.aten.unsqueeze.default,
-        torch.ops.aten.unsqueeze_copy.default,
         # TODO: remove?
         torch.ops.aten.adaptive_avg_pool2d.default,
         torch.ops.aten.avg_pool2d.default,
-        torch.ops.aten.view_copy.default,
-        torch.ops.aten.view.default,
+        torch.ops.aten.max_pool2d.default,
         torch.ops.aten.full.default,
-        torch.ops.aten.slice.Tensor,
-        torch.ops.aten.split.Tensor,
-        torch.ops.aten.split_with_sizes.default,
         torch.ops.aten.flatten.using_ints,
         torch.ops.aten.dropout.default,
         operator.getitem,