Merge branch 'main' into export-D85704977

Author: Gasoonjia

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-e6f766c7d750d40603eee3f66c5915bac606b3ea
+556fc09a9f67f24ca5591ec049c5d0c347c5f62a

.github/workflows/trunk.yml

@@ -347,7 +347,7 @@ jobs:
         elif [[ ${{ matrix.os}} == "zephyr-preset" ]]; then
           setup_script_args="--target-toolchain zephyr"
           toolchain_prefix=arm-zephyr-eabi-
-          threshold="135240" # 132 KiB
+          threshold="135656" # 132 KiB
           toolchain_cmake=examples/zephyr/x86_64-linux-arm-zephyr-eabi-gcc.cmake
         else
           echo "Fail unsupport OS selection ${{ matrix.os }}"

backends/arm/_passes/convert_expand_copy_to_repeat.py

@@ -20,6 +20,7 @@
 
 
 def calculate_multiples(args):
+    """Returns expand args converted to repeat args, and whether the expand changes the rank"""
     input_node_or_tensor = args[0]
 
     if isinstance(input_node_or_tensor, torch.fx.node.Node):
@@ -45,7 +46,7 @@ def calculate_multiples(args):
         multiples[i] if multiples[i] != -1 and extended_shape[i] == 1 else 1
         for i in range(expanded_rank)
     ]
-    return multiples
+    return multiples, expanded_rank != len(input_shape)
 
 
 class ConvertExpandCopyToRepeatPass(ArmPass):
@@ -62,9 +63,9 @@ def call_operator(self, op, args, kwargs, meta):
         if op != self.expand_copy:
             return super().call_operator(op, args, kwargs, meta)
 
-        multiples = calculate_multiples(args)
+        multiples, changes_rank = calculate_multiples(args)
 
-        if all((x == 1 for x in multiples)):
+        if all((x == 1 for x in multiples)) and not changes_rank:
             # All dimensions/repetitions occur only once. Remove node
             # altogether since it's in practice just a copy.
             logger.warning("Found redundant expand node (no-op). Removing it.")

backends/arm/_passes/decompose_embedding_pass.py

@@ -17,7 +17,6 @@
 from .arm_pass_utils import create_node, get_first_fake_tensor
 
 logger = logging.getLogger(__name__)
-logger.setLevel(logging.WARNING)
 
 
 class DecomposeEmbeddingPass(ArmPass):

backends/arm/_passes/insert_int32_casts_after_int64_placeholders.py

@@ -36,6 +36,8 @@ class InsertInt32CastsAfterInt64PlaceholdersPass(ArmPass):
     # Key: op overload; Value: zero-based indices of positional args that must be i64.
     I64_INPUT_ARG_POSITIONS = {
         torch.ops.aten.one_hot.default: (0,),
+        torch.ops.aten.index_copy_.default: (2,),
+        torch.ops.aten.index_copy.default: (2,),
     }
 
     def _insert_callsite_i32_to_i64_casts(self, graph_module: torch.fx.GraphModule):

backends/arm/ethosu/backend.py

@@ -63,7 +63,7 @@ def _compile_tosa_flatbuffer(
         binary = vela_compile(
             tosa_flatbuffer,
             compile_flags,
-            verbose=logger.getEffectiveLevel() == logging.INFO,
+            verbose=logger.getEffectiveLevel() <= logging.INFO,
             intermediate_path=compile_spec.get_intermediate_path(),
         )
         return binary

backends/arm/operator_support/right_shift_support.py

@@ -48,5 +48,5 @@ def is_node_tosa_supported(
         """
         # TODO MLETORCH-525 Remove warning
         if tosa_spec.is_U55_subset:
-            logging.warning(f"{node.target} may introduce one-off errors.")
+            logger.warning(f"{node.target} may introduce one-off errors.")
         return True

backends/arm/operator_support/slice_copy_support.py

@@ -2,7 +2,11 @@
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
+"""Declare operator support for ``aten.slice_copy`` in TOSA.
 
+Support slicing with unit step only; emit a warning and reject otherwise.
+
+"""
 
 import logging
 
@@ -19,19 +23,29 @@
 
 @register_tosa_support_check
 class SliceCopySupported(SupportedTOSAOperatorCheck):
+    """Provide TOSA support check for ``aten.slice_copy``."""
+
     targets = [exir_ops.edge.aten.slice_copy.Tensor]
 
     tosa_specs = [
         TosaSpecification.create_from_string("TOSA-1.0+INT"),
         TosaSpecification.create_from_string("TOSA-1.0+FP"),
     ]
 
-    def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification) -> bool:  # type: ignore[override, misc]
+    def is_node_tosa_supported(
+        self, node: fx.Node, tosa_spec: TosaSpecification
+    ) -> bool:  # type: ignore[override, misc]
+        """Return True if the node is supported by TOSA.
+
+        Accept slice_copy when the step is 1 (or unspecified). Warn and reject
+        non-unit step sizes.
+
+        """
         if tosa_spec not in self.tosa_specs:
             return False
 
         args = node.args
         if len(args) == 5 and (step := args[4]) != 1:
-            logging.warning(f"{node.target} with step size of {step} not supported.")
+            logger.warning(f"{node.target} with step size of {step} not supported.")
             return False
         return True

backends/arm/test/conftest.py

@@ -3,10 +3,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-import logging
 import os
 import random
-import sys
 from typing import Any
 
 import pytest
@@ -29,8 +27,6 @@ def pytest_configure(config):
     if config.option.arm_run_tosa_version:
         pytest._test_options["tosa_version"] = config.option.arm_run_tosa_version
 
-    logging.basicConfig(level=logging.INFO, stream=sys.stdout)
-
 
 def pytest_collection_modifyitems(config, items):
     pass

backends/arm/test/misc/test_debug_feats.py

@@ -23,7 +23,6 @@
 )
 from executorch.backends.test.harness.stages import StageType
 
-
 input_t1 = Tuple[torch.Tensor]  # Input x
 
 
@@ -261,14 +260,14 @@ def test_dump_tosa_debug_tosa(test_data: input_t1):
 
 
 @common.parametrize("test_data", Linear.inputs)
-def test_dump_tosa_ops(caplog, test_data: input_t1):
+def test_dump_tosa_ops(capsys, test_data: input_t1):
     aten_ops: list[str] = []
     exir_ops: list[str] = []
     pipeline = TosaPipelineINT[input_t1](Linear(), test_data, aten_ops, exir_ops)
     pipeline.pop_stage("run_method_and_compare_outputs")
     pipeline.dump_operator_distribution("to_edge_transform_and_lower")
     pipeline.run()
-    assert "TOSA operators:" in caplog.text
+    assert "TOSA operators:" in capsys.readouterr().out
 
 
 class Add(torch.nn.Module):
@@ -282,12 +281,15 @@ def forward(self, x):
 
 @common.parametrize("test_data", Add.inputs)
 @common.XfailIfNoCorstone300
-def test_fail_dump_tosa_ops(caplog, test_data: input_t1):
+def test_fail_dump_tosa_ops(capsys, test_data: input_t1):
     aten_ops: list[str] = []
     exir_ops: list[str] = []
     pipeline = EthosU55PipelineINT[input_t1](
         Add(), test_data, aten_ops, exir_ops, use_to_edge_transform_and_lower=True
     )
     pipeline.dump_operator_distribution("to_edge_transform_and_lower")
     pipeline.run()
-    assert "Can not get operator distribution for Vela command stream." in caplog.text
+    assert (
+        "Can not get operator distribution for Vela command stream."
+        in capsys.readouterr().out
+    )