Merge remote-tracking branch 'origin/main' into android-etdump

Author: kirklandsign

.buckconfig

@@ -39,6 +39,7 @@
 
 [buck2]
 restarter=true
+file_watcher=notify
 
 [oss]
 folly_cxx_tests = False

.ci/scripts/build-qnn-sdk.sh

@@ -11,17 +11,12 @@ set -o xtrace
 
 build_qnn_backend() {
   echo "Start building qnn backend."
-  export ANDROID_NDK_ROOT=/opt/ndk
-  export QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029
+  export ANDROID_NDK_ROOT=${ANDROID_NDK_ROOT:-/opt/ndk}
+  export QNN_SDK_ROOT=${QNN_SDK_ROOT:-/tmp/qnn/2.28.0.241029}
   export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../.." && pwd)"
 
-  # Workaround to avoid issues around missing flatccrt library (depending on the
-  # number of jobs used), see issue #7300:
-  # Build twice (second time with `--no_clean`) to make sure libflatccrt.a is
-  # available.
-  # TODO: Remove this workaround once the underlying issue is fixed.
-  bash backends/qualcomm/scripts/build.sh --skip_aarch64 --job_number 2 --release || \
-  bash backends/qualcomm/scripts/build.sh --skip_aarch64 --job_number 2 --release --no_clean
+  parallelism=$(( $(nproc) - 1 ))
+  bash backends/qualcomm/scripts/build.sh --skip_aarch64 --job_number ${parallelism} --release
 }
 
 set_up_aot() {

.github/workflows/build-presets.yml

@@ -6,6 +6,8 @@ on:
     branches:
       - main
       - release/*
+    paths:
+      - .github/workflows/build-presets.yml
   workflow_dispatch:
 
 concurrency:
@@ -16,15 +18,51 @@ jobs:
   apple:
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     strategy:
+      fail-fast: false
       matrix:
-        preset: [macos-arm64]
+        preset: [macos-arm64, pybind]
     with:
       job-name: build
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       runner: macos-latest-xlarge
       python-version: 3.12
       submodules: recursive
+      timeout: 90
       script: |
         set -eux
         ${CONDA_RUN} ./install_requirements.sh > /dev/null
         ${CONDA_RUN} cmake --preset ${{ matrix.preset }}
         ${CONDA_RUN} cmake --build cmake-out --parallel
+
+  linux:
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    strategy:
+      fail-fast: false
+      matrix:
+        preset: [pybind]
+        runner: [linux.2xlarge, linux.arm64.2xlarge]
+        docker-image: [executorch-ubuntu-22.04-clang12, executorch-ubuntu-22.04-gcc11-aarch64]
+        # Excluding specific runner + docker image combinations that don't make sense:
+        #   - Excluding the ARM64 gcc image on the x86 runner (linux.2xlarge)
+        #   - Excluding the x86 clang image on the ARM64 runner (linux.arm64.2xlarge)
+        exclude:
+          - runner: linux.2xlarge
+            docker-image: executorch-ubuntu-22.04-gcc11-aarch64
+          - runner: linux.arm64.2xlarge
+            docker-image: executorch-ubuntu-22.04-clang12
+    with:
+      job-name: build
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      runner: ${{ matrix.runner }}
+      docker-image: ${{ matrix.docker-image }}
+      submodules: recursive
+      timeout: 90
+      script: |
+        set -eux
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        ./install_requirements.sh > /dev/null
+        cmake --preset ${{ matrix.preset }}
+        cmake --build cmake-out --parallel

.github/workflows/pull.yml

@@ -343,34 +343,6 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool "${BUILD_TOOL}"
         PYTHON_EXECUTABLE=python bash examples/xnnpack/quantization/test_quantize.sh "${BUILD_TOOL}" mv2
 
-  test-pybind-build-linux:
-    name: test-pybind-build-linux
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    permissions:
-      id-token: write
-      contents: read
-    strategy:
-      fail-fast: false
-    with:
-      runner: linux.2xlarge
-      docker-image: executorch-ubuntu-22.04-clang12
-      submodules: 'recursive'
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      timeout: 90
-      script: |
-        # The generic Linux job chooses to use base env, not the one setup by the image
-        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
-        conda activate "${CONDA_ENV}"
-
-        # build module for executorch.extension.pybindings.portable_lib
-        BUILD_TOOL="cmake"
-        PYTHON_EXECUTABLE=python \
-        CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON" \
-        bash .ci/scripts/setup-linux.sh --build-tool "${BUILD_TOOL}"
-
-        # see if we can import the module successfully
-        python -c "from executorch.extension.pybindings import portable_lib; print('success!')"
-
   test-binary-size-linux-gcc:
     name: test-binary-size-linux-gcc
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main

.github/workflows/trunk.yml

@@ -289,30 +289,6 @@ jobs:
         # Build and test coreml delegate
         PYTHON_EXECUTABLE=python ${CONDA_RUN} bash backends/apple/coreml/scripts/build_all.sh
 
-  test-pybind-build-macos:
-    name: test-pybind-build-macos
-    uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
-    strategy:
-      matrix:
-        include:
-          - build-tool: cmake
-      fail-fast: false
-    with:
-      runner: macos-m1-stable
-      python-version: '3.11'
-      submodules: 'recursive'
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      timeout: 180
-      script: |
-        bash .ci/scripts/setup-conda.sh
-
-        # build module for executorch.extension.pybindings.portable_lib
-        BUILD_TOOL=${{ matrix.build-tool }}
-        CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON" PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh --build-tool "${BUILD_TOOL}"
-
-        # see if we can import the module successfully
-        ${CONDA_RUN} python -c "from executorch.extension.pybindings import portable_lib; print('success!')"
-
   test-static-llama-ane:
     name: test-static-llama-ane
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main

CMakeLists.txt

@@ -48,21 +48,33 @@ project(executorch)
 # MARK: - Start EXECUTORCH_H12025_BUILD_MIGRATION --------------------------------------------------
 
 include(${PROJECT_SOURCE_DIR}/tools/cmake/common/preset.cmake)
+include(${PROJECT_SOURCE_DIR}/tools/cmake/Utils.cmake)
+include(CMakeDependentOption)
+include(ExternalProject)
 
 if(NOT CMAKE_CXX_STANDARD)
   set(CMAKE_CXX_STANDARD 17)
 endif()
 announce_configured_options(CMAKE_CXX_STANDARD)
 
+if(NOT CMAKE_SYSTEM_PROCESSOR)
+  set(CMAKE_SYSTEM_PROCESSOR ${CMAKE_HOST_SYSTEM_PROCESSOR})
+endif()
+announce_configured_options(CMAKE_SYSTEM_PROCESSOR)
+
 if(NOT CMAKE_BUILD_TYPE)
   set(CMAKE_BUILD_TYPE Debug)
 endif()
 announce_configured_options(CMAKE_BUILD_TYPE)
 
+if(NOT PYTHON_EXECUTABLE)
+  resolve_python_executable()
+endif()
+announce_configured_options(PYTHON_EXECUTABLE)
+
 announce_configured_options(CMAKE_CXX_COMPILER_ID)
 announce_configured_options(CMAKE_TOOLCHAIN_FILE)
 announce_configured_options(BUCK2)
-announce_configured_options(PYTHON_EXECUTABLE)
 
 load_build_preset()
 include(${PROJECT_SOURCE_DIR}/tools/cmake/preset/default.cmake)
@@ -72,10 +84,6 @@ print_configured_options()
 
 # MARK: - End EXECUTORCH_H12025_BUILD_MIGRATION ----------------------------------------------------
 
-include(tools/cmake/Utils.cmake)
-include(CMakeDependentOption)
-include(ExternalProject)
-
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 # Setup RPATH.
@@ -251,11 +259,6 @@ if(EXECUTORCH_BUILD_TESTS)
   include(CTest)
 endif()
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-message(STATUS "Using python executable '${PYTHON_EXECUTABLE}'")
-
 # TODO(dbort): Fix these warnings and remove this flag.
 set(_common_compile_options -Wno-deprecated-declarations -fPIC)
 
@@ -579,6 +582,7 @@ if(EXECUTORCH_BUILD_PYBIND)
       ${TORCH_PYTHON_LIBRARY}
       bundled_program
       etdump
+      flatccrt
       executorch
       extension_data_loader
       util

CMakePresets.json

@@ -15,6 +15,7 @@
     },
     {
       "name": "macos-arm64",
+      "displayName": "Build everything buildable on macOS arm64",
       "inherits": ["common"],
       "generator": "Xcode",
       "cacheVariables": {
@@ -28,6 +29,20 @@
         "type": "equals",
         "rhs": "Darwin"
       }
+    },
+    {
+      "name": "pybind",
+      "displayName": "Build pybindings exported in the wheel",
+      "inherits": ["common"],
+      "cacheVariables": {
+        "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/pybind.cmake",
+        "CMAKE_OSX_DEPLOYMENT_TARGET": "10.15"
+      },
+      "condition": {
+        "type": "inList",
+        "string": "${hostSystemName}",
+        "list": ["Darwin", "Linux", "Windows"]
+      }
     }
   ]
 }

backends/apple/mps/CMakeLists.txt

@@ -18,10 +18,6 @@ endif()
 
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 set(_common_compile_options -Wno-deprecated-declarations)
 set(_common_include_directories ${EXECUTORCH_ROOT}/..)
 

backends/arm/_passes/__init__.py

@@ -8,6 +8,7 @@
 from .annotate_channels_last_dim_order_pass import AnnotateChannelsLastDimOrder  # noqa
 from .annotate_decomposed_matmul import AnnotateDecomposedMatmulPass  # noqa
 from .arm_pass import ArmPass  # noqa
+from .broadcast_args_pass import BroadcastArgsPass  # noqa
 from .cast_int64_pass import CastInt64BuffersToInt32Pass  # noqa
 from .cast_to_int32_pass import CastToInt32Pass  # noqa
 from .conv1d_unsqueeze_pass import Conv1dUnsqueezePass  # noqa
@@ -24,6 +25,7 @@
 from .decompose_gelu_pass import DecomposeGeluPass  # noqa
 from .decompose_layernorm_pass import DecomposeLayerNormPass  # noqa
 from .decompose_leaky_relu_pass import DecomposeLeakyReLUPass  # noqa
+from .decompose_linalg_vector_norm_pass import DecomposeLinearVectorNormPass  # noqa
 from .decompose_linear_pass import DecomposeLinearPass  # noqa
 from .decompose_meandim_pass import DecomposeMeanDimPass  # noqa
 from .decompose_ne_pass import DecomposeNotEqualPass  # noqa

backends/arm/_passes/arm_pass_manager.py

@@ -10,6 +10,7 @@
 from executorch.backends.arm._passes import (
     AnnotateChannelsLastDimOrder,
     AnnotateDecomposedMatmulPass,
+    BroadcastArgsPass,
     CastInt64BuffersToInt32Pass,
     CastToInt32Pass,
     ComputeConstantOpsAOT,
@@ -29,6 +30,7 @@
     DecomposeLayerNormPass,
     DecomposeLeakyReLUPass,
     DecomposeLinearPass,
+    DecomposeLinearVectorNormPass,
     DecomposeMeanDimPass,
     DecomposeNotEqualPass,
     DecomposeSelectPass,
@@ -86,6 +88,7 @@ def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(ConvertSplitToSlicePass())
         self.add_pass(ConvertMmToBmmPass())
         self.add_pass(DecomposeLinearPass())
+        self.add_pass(DecomposeLinearVectorNormPass())
         self.add_pass(DecomposeMeanDimPass())
         self.add_pass(ConvertFullLikeToFullPass())
         self.add_pass(ConvertToClampPass())
@@ -102,6 +105,8 @@ def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(RetraceFoldedDtypesPass())
         self.add_pass(UnsqueezeScalarPlaceholdersPass(exported_program))
         self.add_pass(MatchArgRanksPass(exported_program))
+        if self.tosa_spec.is_U55_subset:
+            self.add_pass(BroadcastArgsPass())
         self.add_pass(ComputeConstantOpsAOT(exported_program))
 
         self.add_pass(RemoveClonePass())
@@ -133,6 +138,7 @@ def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(FuseBatchnorm2DPass(exported_program))
         self.add_pass(ConvertMmToBmmPass())
         self.add_pass(DecomposeLinearPass())
+        self.add_pass(DecomposeLinearVectorNormPass())
         self.add_pass(DecomposeLeakyReLUPass())
         self.add_pass(DecomposeBatchNormPass())
         self.add_pass(DecomposeLayerNormPass())
@@ -207,6 +213,7 @@ def transform_for_annotation_pipeline(self, graph_module: GraphModule):
         self.add_pass(DecomposeCosineSimilarityPass())
         self.add_pass(DecomposeDivPass())
         self.add_pass(DecomposeLeakyReLUPass())
+        self.add_pass(DecomposeLinearVectorNormPass())
         self.add_pass(DecomposeSqrtPass())
         self.add_pass(DecomposeSiluPass())
 

backends/arm/_passes/broadcast_args_pass.py

@@ -0,0 +1,63 @@
+# Copyright 2025 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.
+
+from executorch.backends.arm._passes import ArmPass
+
+from executorch.backends.arm._passes.arm_pass_utils import (
+    create_node,
+    get_first_fake_tensor,
+)
+
+from executorch.exir.dialects._ops import ops as exir_ops
+
+from executorch.exir.pass_base import PassResult
+from torch.fx import GraphModule, Node
+
+
+class BroadcastArgsPass(ArmPass):
+    """
+    Pass to manually broadcast arguments by inserting repeats.
+    This is done when more than one arg needs broadcasting.
+    """
+
+    targeted_ops = {
+        exir_ops.edge.aten.add.Tensor,
+        exir_ops.edge.aten.sub.Tensor,
+        # mul is indirectly targeting div as div is decompsed to reciprocal + mul
+        exir_ops.edge.aten.mul.Tensor,
+    }
+
+    def call(self, graph_module: GraphModule) -> PassResult:
+        for node in graph_module.graph.nodes:
+            if node.op != "call_function" or node.target not in self.targeted_ops:
+                continue
+
+            output_shape = get_first_fake_tensor(node).shape
+            nbr_of_broacasts = 0
+            for arg in node.args:
+                if not isinstance(arg, Node):
+                    continue
+
+                shape = get_first_fake_tensor(arg).shape
+                if shape != output_shape:
+                    nbr_of_broacasts += 1
+                if nbr_of_broacasts > 1:
+                    multiples = [
+                        int(output_shape[d] / shape[d])
+                        for d in range(len(output_shape))
+                    ]
+                    with graph_module.graph.inserting_before(node):
+                        repeat = create_node(
+                            graph_module.graph,
+                            exir_ops.edge.aten.repeat.default,
+                            args=(arg, multiples),
+                            kwargs={},
+                            from_node=node,
+                        )
+                        node.replace_input_with(arg, repeat)
+
+        graph_module.recompile()
+        graph_module = super().call(graph_module).graph_module
+        return PassResult(graph_module, True)