Merge branch 'main' into op_support

Author: Erik-Lundell

.ci/docker/build.sh

@@ -48,6 +48,7 @@ case "${IMAGE_NAME}" in
   executorch-ubuntu-22.04-mediatek-sdk)
     MEDIATEK_SDK=yes
     CLANG_VERSION=12
+    ANDROID_NDK_VERSION=r27b
     ;;
   executorch-ubuntu-22.04-clang12-android)
     LINTRUNNER=""

.ci/scripts/build_llama_android.sh

@@ -60,6 +60,5 @@ build_llama_runner() {
 
     cmake --build cmake-android-out/examples/models/llama -j4 --config Release
 }
-install_flatc_from_source
 install_executorch_and_backend_lib
 build_llama_runner

build/test_ios_ci.sh renamed to .ci/scripts/test_ios_ci.sh

@@ -209,9 +209,14 @@ test_model_with_qnn() {
   EXPORTED_MODEL=$(find "./${EXPORT_SCRIPT}" -type f -name "${MODEL_NAME}*.pte" -print -quit)
 }
 
+# Run CoreML tests.
+#
+# @param should_test If true, build and test the model using the coreml_executor_runner.
 test_model_with_coreml() {
-  if [[ "${BUILD_TOOL}" == "buck2" ]]; then
-    echo "coreml doesn't support buck2."
+  local should_test="$1"
+
+  if [[ "${BUILD_TOOL}" != "cmake" ]]; then
+    echo "coreml only supports cmake."
     exit 1
   fi
 
@@ -229,6 +234,14 @@ test_model_with_coreml() {
     echo "No .pte file found"
     exit 1
   fi
+
+  # Run the model
+  if [ "${should_test}" = true ]; then
+    echo "Testing exported model with coreml_executor_runner..."
+    local out_dir=$(mktemp -d)
+    COREML_EXECUTOR_RUNNER_OUT_DIR="${out_dir}" examples/apple/coreml/scripts/build_executor_runner.sh
+    "${out_dir}/coreml_executor_runner" --model_path "${EXPORTED_MODEL}"
+  fi
 }
 
 test_model_with_mps() {
@@ -247,7 +260,11 @@ elif [[ "${BACKEND}" == *"qnn"* ]]; then
   fi
 elif [[ "${BACKEND}" == *"coreml"* ]]; then
   echo "Testing ${MODEL_NAME} with coreml..."
-  test_model_with_coreml
+  should_test_coreml=false
+  if [[ "${BACKEND}" == *"test"* ]]; then
+    should_test_coreml=true
+  fi
+  test_model_with_coreml "${should_test_coreml}"
   if [[ $? -eq 0 ]]; then
     prepare_artifacts_upload
   fi

.ci/scripts/test_model.sh

@@ -80,25 +80,6 @@ install_pytorch_and_domains() {
   sccache --show-stats || true
 }
 
-install_flatc_from_source() {
-  # NB: This function could be used to install flatbuffer from source
-  pushd third-party/flatbuffers || return
-
-  cmake -G "Unix Makefiles" -DCMAKE_BUILD_TYPE=Release
-  if [ "$(uname)" == "Darwin" ]; then
-    CMAKE_JOBS=$(( $(sysctl -n hw.ncpu) - 1 ))
-  else
-    CMAKE_JOBS=$(( $(nproc) - 1 ))
-  fi
-  cmake --build . -j "${CMAKE_JOBS}"
-
-  # Copy the flatc binary to conda path
-  EXEC_PATH=$(dirname "$(which python)")
-  cp flatc "${EXEC_PATH}"
-
-  popd || return
-}
-
 build_executorch_runner_buck2() {
   # Build executorch runtime with retry as this step is flaky on macos CI
   retry buck2 build //examples/portable/executor_runner:executor_runner
@@ -111,9 +92,14 @@ build_executorch_runner_cmake() {
   mkdir "${CMAKE_OUTPUT_DIR}"
 
   pushd "${CMAKE_OUTPUT_DIR}" || return
+  if [[ $1 == "Debug" ]]; then
+      CXXFLAGS="-fsanitize=address,undefined"
+  else
+      CXXFLAGS=""
+  fi
   # This command uses buck2 to gather source files and buck2 could crash flakily
   # on MacOS
-  retry cmake -DPYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" -DCMAKE_BUILD_TYPE="${1:-Release}" ..
+  CXXFLAGS="$CXXFLAGS" retry cmake -DPYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" -DCMAKE_BUILD_TYPE="${1:-Release}" ..
   popd || return
 
   if [ "$(uname)" == "Darwin" ]; then

.ci/scripts/utils.sh

@@ -14,6 +14,12 @@
             test_base.ModelTest(
                 model=Model.Mv3,
                 backend=Backend.XnnpackQuantizationDelegation,
-            )
+            ),
+            # Enable this once CoreML is suppported out-of-the-box
+            # https://github.com/pytorch/executorch/issues/9019
+            # test_base.ModelTest(
+            #     model=Model.Mv3,
+            #     backend=Backend.CoreMlTest,
+            # )
         ]
     )

.ci/scripts/wheel/test_macos.py

@@ -14,7 +14,7 @@ on:
       - build/build_apple_frameworks.sh
       - build/build_apple_llm_demo.sh
       - build/create_frameworks.sh
-      - build/test_ios_ci.sh
+      - .ci/scripts/test_ios_ci.sh
       - examples/demo-apps/apple_ios/**
       - extension/apple/**
       - extension/benchmark/apple/**
@@ -75,7 +75,7 @@ jobs:
 
         # Build and test iOS Demo App
         PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output \
-        build/test_ios_ci.sh "${ARTIFACTS_DIR_NAME}"
+        .ci/scripts/test_ios_ci.sh "${ARTIFACTS_DIR_NAME}"
 
   # Upload the test demo app to S3
   upload-demo-ios:

.github/workflows/apple.yml

@@ -176,7 +176,7 @@ jobs:
       id-token: write
       contents: read
     with:
-      runner: linux.2xlarge
+      runner: linux.2xlarge.memory
       docker-image: executorch-ubuntu-22.04-arm-sdk
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
@@ -206,7 +206,7 @@ jobs:
       id-token: write
       contents: read
     with:
-      runner: linux.2xlarge
+      runner: linux.2xlarge.memory
       docker-image: executorch-ubuntu-22.04-arm-sdk
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}

.github/workflows/trunk.yml

@@ -460,7 +460,7 @@ endif()
 # tools like `flatc`, along with example executables like `executor_runner` and
 # libraries that it uses, like `gflags`. Disabling this can be helpful when
 # cross-compiling, but some required tools that would have been built need to be
-# provided directly (via, for example, FLATC_EXECUTABLE).
+# provided directly.
 cmake_dependent_option(
   EXECUTORCH_BUILD_HOST_TARGETS "Build host-only targets." ON
   "NOT CMAKE_TOOLCHAIN_IOS" OFF
@@ -471,10 +471,9 @@ cmake_dependent_option(
 #
 cmake_dependent_option(
   EXECUTORCH_BUILD_FLATC "Build the flatc executable." ON
-  "NOT FLATC_EXECUTABLE;EXECUTORCH_BUILD_HOST_TARGETS" OFF
+  "NOT FLATC_EXECUTABLE" OFF
 )
 
-
 set(FLATBUFFERS_BUILD_FLATC OFF CACHE BOOL "")
 set(FLATBUFFERS_BUILD_FLATHASH OFF CACHE BOOL "")
 set(FLATBUFFERS_BUILD_FLATLIB OFF CACHE BOOL "")
@@ -507,6 +506,8 @@ if(EXECUTORCH_BUILD_FLATC)
                -DFLATBUFFERS_BUILD_TESTS=${FLATBUFFERS_BUILD_TESTS}
                -DFLATBUFFERS_INSTALL=${FLATBUFFERS_INSTALL}
                -DCMAKE_CXX_FLAGS="-DFLATBUFFERS_MAX_ALIGNMENT=${FLATBUFFERS_MAX_ALIGNMENT}"
+               # If building for iOS, "unset" these variables to rely on the host (macOS) defaults.
+               $<$<AND:$<BOOL:${CMAKE_TOOLCHAIN_IOS}>,$<BOOL:$<FILTER:${PLATFORM},EXCLUDE,^MAC>>>:-DCMAKE_OSX_SYSROOT=>
     INSTALL_COMMAND ""
     BUILD_BYPRODUCTS <BINARY_DIR>/flatc
   )
@@ -515,6 +516,8 @@ if(EXECUTORCH_BUILD_FLATC)
     # flatbuffers does not use CMAKE_BUILD_TYPE. Internally, the build forces Release
     # config, but from CMake's perspective the build type is always Debug.
     set(FLATC_EXECUTABLE ${BINARY_DIR}/$<CONFIG>/flatc.exe)
+  elseif(CMAKE_GENERATOR STREQUAL "Xcode")
+    set(FLATC_EXECUTABLE ${BINARY_DIR}/$<CONFIG>/flatc)
   else()
     set(FLATC_EXECUTABLE ${BINARY_DIR}/flatc)
   endif()
@@ -528,12 +531,7 @@ if(NOT FLATC_EXECUTABLE)
   find_program(FLATC_EXECUTABLE flatc)
 
   if(NOT FLATC_EXECUTABLE)
-    message(
-      FATAL_ERROR
-        "FLATC_EXECUTABLE must be set when EXECUTORCH_BUILD_FLATC is disabled. "
-        "Note that EXECUTORCH_BUILD_FLATC may be disabled implicitly when "
-        "cross-compiling or when EXECUTORCH_BUILD_HOST_TARGETS is disabled."
-    )
+    message(FATAL_ERROR "FLATC_EXECUTABLE must be set when EXECUTORCH_BUILD_FLATC is disabled.")
   endif()
 endif()
 

CMakeLists.txt

@@ -32,7 +32,6 @@ cmake "$EXECUTORCH_ROOT_PATH" -B"$CMAKE_EXECUTORCH_BUILD_DIR_PATH" \
 -DCMAKE_TOOLCHAIN_FILE="$IOS_TOOLCHAIN_PATH" \
 -DPLATFORM=MAC_UNIVERSAL \
 -DDEPLOYMENT_TARGET=13.0 \
--DFLATC_EXECUTABLE="$(which flatc)" \
 -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=OFF \
 -DEXECUTORCH_BUILD_XNNPACK=OFF \
 -DEXECUTORCH_BUILD_GFLAGS=OFF

backends/apple/coreml/scripts/build_tests.sh

@@ -22,11 +22,11 @@ static inline MTLLanguageVersion getMetalLanguageVersion(const id<MTLDevice>& de
   // MPS Advanced Indexing needs at least Metal 2.0 (support for Argument Buffers and function constants)
   // host_name attribute needs at least Metal 2.2 and ulong needs Metal 2.3 (supported on MacOS 11+)
   MTLLanguageVersion languageVersion = MTLLanguageVersion2_3;
-#if defined(__MAC_13_0)
-  if (macOS13Plus) {
-    languageVersion = MTLLanguageVersion3_0;
+  if (@available(iOS 16, macOS 13, *)) {
+    if (macOS13Plus) {
+      languageVersion = MTLLanguageVersion3_0;
+    }
   }
-#endif
 
   ET_CHECK_MSG([device supportsFamily:MTLGPUFamilyMac2], "Missing Metal support for MTLGPUFamilyMac2");
   return languageVersion;

backends/apple/mps/runtime/MPSDevice.mm

@@ -206,25 +206,32 @@
 
 Error
 MPSGraphBuilder::mpsScatterOp(NodePtr nodePtr) {
-  auto graphNode = nodePtr->mpsnode_union_as_MPSScatter();
-  ET_LOG(
-    Debug, "%s %d: %d",
-    __FUNCTION__, graphNode->input1_id(), graphNode->output_id()
-  );
+  if (@available(iOS 15.4, macOS 12.3, *)) {
+    auto graphNode = nodePtr->mpsnode_union_as_MPSScatter();
+    ET_LOG(
+      Debug, "%s %d: %d",
+      __FUNCTION__, graphNode->input1_id(), graphNode->output_id()
+    );
 
-  int64_t dim = graphNode->dim();
-  MPSGraphTensor* inputTensor = getMPSGraphTensor(graphNode->input1_id());
-  MPSGraphTensor* indicesTensor = getMPSGraphTensor(graphNode->idx_id());
-  MPSGraphTensor* updatesTensor = getMPSGraphTensor(graphNode->src_id());
+    int64_t dim = graphNode->dim();
+    MPSGraphTensor* inputTensor = getMPSGraphTensor(graphNode->input1_id());
+    MPSGraphTensor* indicesTensor = getMPSGraphTensor(graphNode->idx_id());
+    MPSGraphTensor* updatesTensor = getMPSGraphTensor(graphNode->src_id());
 
-  _idToMPSGraphTensor[graphNode->output_id()] =
-    [_mpsGraph scatterAlongAxis:dim
-                 withDataTensor:inputTensor
-                  updatesTensor:updatesTensor
-                  indicesTensor:indicesTensor
-                           mode:MPSGraphScatterModeSet
-                           name:nil];
-  return Error::Ok;
+    _idToMPSGraphTensor[graphNode->output_id()] =
+      [_mpsGraph scatterAlongAxis:dim
+                   withDataTensor:inputTensor
+                    updatesTensor:updatesTensor
+                    indicesTensor:indicesTensor
+                             mode:MPSGraphScatterModeSet
+                             name:nil];
+
+    return Error::Ok;
+  } else {
+    ET_LOG(Error, "MPS: scatter op is not supported on iOS < 15.4 and macOS < 12.3");
+
+    return Error::NotSupported;
+  }
 }
 
 

backends/apple/mps/runtime/operations/IndexingOps.mm

@@ -11,6 +11,7 @@
     pool_2d_support,
     reduce_sum_support,
     right_shift_support,
+    slice_copy_support,
     to_copy_support,
     tosa_supported_operators,
 )

backends/arm/operator_support/__init__.py

@@ -0,0 +1,39 @@
+# 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.
+
+
+import logging
+
+import torch.fx as fx
+from executorch.backends.arm.operator_support.tosa_supported_operators import (
+    register_tosa_support_check,
+    SupportedTOSAOperatorCheck,
+)
+from executorch.backends.arm.tosa_specification import TosaSpecification
+from executorch.backends.arm.tosa_utils import getNodeArgs
+from executorch.exir.dialects._ops import ops as exir_ops
+
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.WARNING)
+
+
+@register_tosa_support_check
+class SliceCopySupported(SupportedTOSAOperatorCheck):
+    targets = [exir_ops.edge.aten.slice_copy.Tensor]
+
+    tosa_specs = [
+        TosaSpecification.create_from_string("TOSA-0.80+BI"),
+        TosaSpecification.create_from_string("TOSA-0.80+MI"),
+    ]
+
+    def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification) -> bool:  # type: ignore[override, misc]
+        if tosa_spec not in self.tosa_specs:
+            return False
+
+        inputs = getNodeArgs(node)
+        if len(inputs) == 5 and (step := inputs[4].number) != 1:
+            logging.warning(f"{node.target} with step size of {step} not supported.")
+            return False
+        return True

backends/arm/operator_support/slice_copy_support.py

@@ -77,7 +77,6 @@ def register_tosa_support_check(checker: Type[SupportedTOSAOperatorCheck]):
 def get_registered_tosa_support_checks(
     tosa_spec: TosaSpecification,
 ) -> list[Type[SupportedTOSAOperatorCheck]]:
-
     if tosa_spec not in _tosa_spec_support:
         raise RuntimeError(
             f"TOSA specification not valid: {tosa_spec} not in {list(_tosa_spec_support.keys())}"
@@ -186,7 +185,6 @@ def is_node_supported(
             exir_ops.edge.aten._softmax.default,
             exir_ops.edge.aten.select_copy.int,
             exir_ops.edge.aten._log_softmax.default,
-            exir_ops.edge.aten.slice_copy.Tensor,
             exir_ops.edge.aten.sub.Tensor,
             exir_ops.edge.aten.tanh.default,
             exir_ops.edge.aten.upsample_nearest2d.vec,

backends/arm/operator_support/tosa_supported_operators.py

@@ -45,6 +45,12 @@ def define_node(
         # Handle int8 (quantized) and int32
         assert inputs[0].dtype in [ts.DType.INT8, ts.DType.INT32]
 
+        dim_order = (
+            inputs[0].dim_order
+            if len(inputs[0].shape) > len(inputs[1].shape)
+            else inputs[1].dim_order
+        )
+
         if inputs[0].dtype == ts.DType.INT8:
             rescaled_inputs, scale_back = tqutils.insert_rescale_ops_to_int32(
                 tosa_graph, inputs, node
@@ -61,13 +67,14 @@ def define_node(
             # output.dtype == ts.DType.INT32
             add_output = output
 
+        input1, input2 = tutils.reshape_for_broadcast(
+            tosa_graph, rescaled_inputs, dim_order
+        )
+
         # Do the INT32 Add
         tosa_graph.addOperator(
             TosaOp.Op().ADD,
-            [
-                rescaled_inputs[0].name,
-                rescaled_inputs[1].name,
-            ],
+            [input1.name, input2.name],
             [add_output.name],
             None,
         )
@@ -108,10 +115,12 @@ def define_node(
             assert inputs[0].dtype == ts.DType.FP32
             assert output.dtype == ts.DType.FP32
 
+            input1, input2 = tutils.reshape_for_broadcast(tosa_graph, inputs)
+
             # MI lowering
             tosa_graph.addOperator(
                 TosaOp.Op().ADD,
-                [inputs[0].name, inputs[1].name],
+                [input1.name, input2.name],
                 [output.name],
                 None,
             )