Update

[ghstack-poisoned]

Author: swolchok

.ci/scripts/setup-openvino.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and 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.
+
+set -ex
+
+# shellcheck source=/dev/null
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+git clone https://github.com/openvinotoolkit/openvino.git
+cd openvino && git checkout releases/2025/1
+git submodule update --init --recursive
+sudo ./install_build_dependencies.sh
+mkdir build && cd build
+cmake .. -DCMAKE_BUILD_TYPE=Release -DENABLE_PYTHON=ON
+make -j$(nproc)
+
+cd ..
+cmake --install build --prefix dist
+
+source dist/setupvars.sh
+cd ../backends/openvino
+pip install -r requirements.txt
+cd scripts
+./openvino_build.sh --enable_python

.ci/scripts/test_openvino.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and 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.
+
+set -ex
+
+# shellcheck source=/dev/null
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+source openvino/dist/setupvars.sh
+cd backends/openvino/tests
+python test_runner.py --test_type ops
+python test_runner.py --test_type models

.github/workflows/pull.yml

@@ -736,3 +736,25 @@ jobs:
         conda activate "${CONDA_ENV}"
 
         # placeholder for mediatek to add more tests
+
+  test-openvino-linux:
+    name: test-openvino-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-gcc9
+      submodules: 'true'
+      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}"
+
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-openvino.sh
+        PYTHON_EXECUTABLE=python bash .ci/scripts/test_openvino.sh

.lintrunner.toml

@@ -299,12 +299,14 @@ include_patterns = [
     # TODO(https://github.com/pytorch/executorch/issues/7441): Gradually start enabling all folders.
     # 'backends/**/*.py',
     'backends/arm/**/*.py',
+    'backends/openvino/**/*.py',
     'build/**/*.py',
     'codegen/**/*.py',
     # 'devtools/**/*.py',
     'devtools/visualization/**/*.py',
     'docs/**/*.py',
     # 'examples/**/*.py',
+    'examples/openvino/**/*.py',
     # 'exir/**/*.py',
     # 'extension/**/*.py',
     'kernels/**/*.py',

CMakeLists.txt

@@ -1,6 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
-# Copyright 2024-2025 Arm Limited and/or its affiliates.
 # All rights reserved.
+# Copyright 2024-2025 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.
@@ -161,7 +161,7 @@ if(OPTIMIZE_SIZE)
   set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -Os")
 else()
   # -O2: Moderate opt.
-  set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -O2")
+  set(CMAKE_CXX_FLAGS_RELEASE "-O2 ${CMAKE_CXX_FLAGS_RELEASE}")
 endif()
 
 option(EXECUTORCH_BUILD_ANDROID_JNI "Build Android JNI" OFF)
@@ -204,6 +204,8 @@ option(EXECUTORCH_BUILD_MPS "Build the MPS backend" OFF)
 
 option(EXECUTORCH_BUILD_NEURON "Build the backends/mediatek directory" OFF)
 
+option(EXECUTORCH_BUILD_OPENVINO "Build the Openvino backend" OFF)
+
 option(EXECUTORCH_BUILD_PYBIND "Build the Python Bindings" OFF)
 
 option(EXECUTORCH_BUILD_QNN "Build the Qualcomm backend" OFF)
@@ -715,6 +717,10 @@ if(EXECUTORCH_BUILD_NEURON)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/mediatek)
 endif()
 
+if(EXECUTORCH_BUILD_OPENVINO)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/openvino)
+endif()
+
 if(EXECUTORCH_BUILD_QNN)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/qualcomm)
 endif()
@@ -817,6 +823,10 @@ if(EXECUTORCH_BUILD_PYBIND)
     list(APPEND _dep_libs mpsdelegate)
   endif()
 
+  if(EXECUTORCH_BUILD_OPENVINO)
+    list(APPEND _dep_libs openvino_backend)
+  endif()
+
   if(EXECUTORCH_BUILD_XNNPACK)
     # need to explicitly specify XNNPACK and microkernels-prod
     # here otherwise uses XNNPACK and microkernel-prod symbols from libtorch_cpu

README.md

@@ -29,6 +29,7 @@ Platform Support:
   - Arm
   - Cadence
   - MediaTek
+  - OpenVINO
   - Qualcomm
   - Vulkan
   - XNNPACK

backends/apple/coreml/TARGETS

@@ -72,7 +72,7 @@ runtime.cxx_python_extension(
     headers = glob([
         "runtime/inmemoryfs/**/*.hpp",
     ]),
-    base_module = "",
+    base_module = "executorch.backends.apple.coreml",
     compiler_flags = [
         "-std=c++17",
     ],

backends/arm/_passes/__init__.py

@@ -7,7 +7,8 @@
 from . import arm_pass_utils  # noqa
 from .annotate_channels_last_dim_order_pass import AnnotateChannelsLastDimOrder  # noqa
 from .annotate_decomposed_matmul import AnnotateDecomposedMatmulPass  # noqa
-from .cast_int64_pass import CastInt64ToInt32Pass  # noqa
+from .cast_int64_pass import CastInt64BuffersToInt32Pass  # noqa
+from .cast_to_int32_pass import CastToInt32Pass  # noqa
 from .conv1d_unsqueeze_pass import Conv1dUnsqueezePass  # noqa
 from .convert_any_default_dim_dims_pass import ConvertAnyDefaultDimDimsPass  # noqa
 from .convert_expand_copy_to_repeat import ConvertExpandCopyToRepeatPass  # noqa

backends/arm/_passes/arm_pass_manager.py

@@ -10,7 +10,8 @@
 from executorch.backends.arm._passes import (
     AnnotateChannelsLastDimOrder,
     AnnotateDecomposedMatmulPass,
-    CastInt64ToInt32Pass,
+    CastInt64BuffersToInt32Pass,
+    CastToInt32Pass,
     ComputeConstantOpsAOT,
     Conv1dUnsqueezePass,
     ConvertAnyDefaultDimDimsPass,
@@ -80,6 +81,8 @@ def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(ConvertToClampPass())
         self.add_pass(ConvertMinMaxPass())
         self.add_pass(ConvertAnyDefaultDimDimsPass())
+        if isinstance(self.tosa_spec, Tosa_0_80) and self.tosa_spec.is_U55_subset:
+            self.add_pass(CastToInt32Pass())
 
         self.add_pass(ReplaceScalarWithTensorArgPass())
         self.add_pass(AnnotateDecomposedMatmulPass())
@@ -94,7 +97,7 @@ def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(SizeAdjustConv2DPass())
         self.add_pass(ConvertExpandCopyToRepeatPass())
         self.add_pass(UnsqueezeBeforeRepeatPass())
-        self.add_pass(CastInt64ToInt32Pass(exported_program))
+        self.add_pass(CastInt64BuffersToInt32Pass(exported_program))
         self.add_pass(KeepDimsFalseToSqueezePass())
         self.add_pass(Conv1dUnsqueezePass(exported_program))
         self.add_pass(DecomposeSelectPass())
@@ -141,7 +144,7 @@ def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(SizeAdjustConv2DPass())
         self.add_pass(ConvertExpandCopyToRepeatPass())
         self.add_pass(UnsqueezeBeforeRepeatPass())
-        self.add_pass(CastInt64ToInt32Pass(exported_program))
+        self.add_pass(CastInt64BuffersToInt32Pass(exported_program))
         self.add_pass(KeepDimsFalseToSqueezePass())
         self.add_pass(Conv1dUnsqueezePass(exported_program))
         self.add_pass(DecomposeSelectPass())

backends/arm/_passes/cast_int64_pass.py

@@ -15,13 +15,13 @@
 logger.setLevel(logging.WARNING)
 
 
-class CastInt64ToInt32Pass(ExportPass):
+class CastInt64BuffersToInt32Pass(ExportPass):
     """
     Cast int64 buffers to int32 if the int64 data is in int32 range.
     """
 
     def __init__(self, exported_program: torch.export.ExportedProgram):
-        super(CastInt64ToInt32Pass, self).__init__()
+        super(CastInt64BuffersToInt32Pass, self).__init__()
         self.exported_program = exported_program
 
     def _assert_within_int32(self, tensor: torch.Tensor, node: torch.fx.Node):