Merge branch 'main' into gh/trivedivivek/66/orig

Author: trivedivivek

.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

@@ -34,6 +34,9 @@ def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
 
         for pad in output_padding:
             if pad != 0:
+                self.reporter.report_reject(
+                    node, "Convolutions with non-zero output padding not implemented."
+                )
                 return False
 
         # Hardware specific constraints
@@ -56,19 +59,33 @@ def _is_node_supported_u55(self, node: fx.Node):
             # Depthwise convolution
             for dim in shape_in[1:]:
                 if not 1 <= dim <= 65536:
+                    self.reporter.report_reject(
+                        node,
+                        f"Depthwise convolution must have CWH <= 65536, got {dim})",
+                    )
                     return False
         else:
             # Convolution
             if not 1 <= C_in <= 65536:
+                self.reporter.report_reject(
+                    node, f"Convolution must have C <= 65536, got {C_in})"
+                )
                 return False
 
         kernel_w = kernel[2]
         kernel_h = kernel[3] if len(kernel) > 3 else 1
         # Kernel condition misses constraint on sum of absolute weights
         if not 1 <= kernel_h <= 64 or not 1 <= kernel_w * kernel_h <= 4096:
+            self.reporter.report_reject(
+                node,
+                f"Convolution needs to have kernel_y<=64, kernel_x*kernel_y<=4096, got kernel ({kernel_w}, {kernel_h})",
+            )
             return False
 
         if not self._stride_condition(node):
+            self.reporter.report_reject(
+                node, "Failed condition on stride, pad and dilation combination."
+            )
             return False
 
         return True

backends/arm/operator_support/convolution_support.py

@@ -54,12 +54,35 @@ def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
         if len(node.args) > 3:
             # Padding case
             if not all(1 <= k <= 8 for k in kernel):
+                self.reporter.report_reject(
+                    node, f"Avgpool2d with padding needs kernel dims < 8, got {kernel}"
+                )
                 return False
         else:
             if not kernel_check(kernel):
+                self.reporter.report_reject(
+                    node,
+                    f"Avgpool2d needs kernel_y < 256, kernel_x*kernel_y<=65536, got {kernel}",
+                )
                 return False
 
-        return dim_check(shape) and shape[0] == 1 and stride_check(stride)
+        if not dim_check(shape):
+            self.reporter.report_reject(
+                node,
+                f"Avgpool2d needs N == 1, rest dims <= 65536, got shape {list(shape)}",
+            )
+            return False
+        if not stride_check(stride):
+            self.reporter.report_reject(
+                node, f"Avgpool2d needs stride <= 3, got {stride}"
+            )
+            return False
+        if not shape[0] == 1:
+            self.reporter.report_reject(
+                node, f"Avgpool2d needs N==1, got N=={shape[0]}"
+            )
+            return False
+        return True
 
 
 @register_tosa_support_check
@@ -82,4 +105,21 @@ def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
         kernel = cast(tuple[int, int], node.args[1])
         stride = cast(tuple[int, int], node.args[2])
 
-        return kernel_check(kernel) and dim_check(shape) and stride_check(stride)
+        if not kernel_check(kernel):
+            self.reporter.report_reject(
+                node,
+                f"Maxpool2d needs kernel_y < 256, kernel_x*kernel_y<=65536, got {kernel}",
+            )
+            return False
+        if not dim_check(shape):
+            self.reporter.report_reject(
+                node,
+                f"Maxpool2d needs N == 1, rest dims <= 65536, got shape {list(shape)}",
+            )
+            return False
+        if not stride_check(stride):
+            self.reporter.report_reject(
+                node, f"Maxpool2d needs stride <= 3, got {stride}"
+            )
+            return False
+        return True

backends/arm/operator_support/pool_2d_support.py

@@ -34,6 +34,9 @@ def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
 
         for dim in dim_list:
             if not 1 <= input_shape[dim] <= 65536:
+                self.reporter.report_reject(
+                    node, f"sum needs dims < 65536, got shape {input_shape}"
+                )
                 return False
 
             # We can't be certain of which dim is the last in memory yet,
@@ -45,7 +48,9 @@ def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
             for length in input_shape[dim + 1 :]:
                 post_R_product *= length
             if not 1 <= pre_R_product <= 65536:
+                self.reporter.report_reject(node, "Failed dim check")
                 return False
             if not 1 <= post_R_product <= 65536:
+                self.reporter.report_reject(node, "Failed dim check")
                 return False
         return True