Add torchao kernels to xcframework (pytorch#10963)

### Summary

In accordance with the following
[ticket](pytorch#10694), we want to
enable building the torchao kernel through the main ExecuTorch
CMakeLists.txt. This PR should cover all the necessary steps required to
build low-bit kernels on the iOS app.

#### List of changes
- Provide an optional flag `EXECUTORCH_BUILD_KERNELS_TORCHAO` that
enables building torchao through the CMakeLists.txt.
- Update the apple framework script to include the torchao build.
- Updated the apple framework defaults to include torchao.

### Test plan
- Code has been successfully tested on the iPhone 16 simulator

---------

Co-authored-by: Scott Roy <[email protected]>

Author: hinriksnaer

CMakeLists.txt

@@ -278,6 +278,30 @@ if(EXECUTORCH_BUILD_PTHREADPOOL)
   )
 endif()
 
+if(EXECUTORCH_BUILD_KERNELS_TORCHAO)
+  set(TORCHAO_BUILD_ATEN_OPS OFF)
+  set(TORCHAO_BUILD_EXECUTORCH_OPS ON)
+  set(TORCHAO_BUILD_CPU_AARCH64 ON)
+  set(TORCHAO_ENABLE_ARM_NEON_DOT ON)
+
+  list(APPEND TORCHAO_INCLUDE_DIRS
+    ${EXECUTORCH_ROOT}/backends/xnnpack/third-party/pthreadpool/include
+    ${EXECUTORCH_ROOT}/backends/xnnpack/third-party/cpuinfo/include
+    ${EXECUTORCH_ROOT}/third-party/ao
+  )
+
+  set(EXECUTORCH_INCLUDE_DIRS ${TORCHAO_INCLUDE_DIRS})
+
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/third-party/ao/torchao/experimental)
+  executorch_target_link_options_shared_lib(torchao_ops_executorch)
+  list(APPEND _executorch_kernels torchao_ops_executorch)
+endif()
+
+if(EXECUTORCH_BUILD_TESTS)
+  set(EXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR ON)
+  include(CTest)
+endif()
+
 # TODO(dbort): Fix these warnings and remove this flag.
 set(_common_compile_options -Wno-deprecated-declarations -fPIC)
 

scripts/build_apple_frameworks.sh

@@ -125,6 +125,11 @@ libquantized_kernels.a,\
 libquantized_ops_lib.a,\
 :"
 
+FRAMEWORK_KERNELS_TORCHAO="kernels_torchao:\
+libtorchao_ops_executorch.a,\
+libtorchao_kernels_aarch64.a,\
+:"
+
 usage() {
   echo "Usage: $0 [OPTIONS]"
   echo "Build frameworks for Apple platforms."
@@ -137,6 +142,7 @@ usage() {
   echo "  --mps                Only build the Metal Performance Shaders backend."
   echo "  --optimized          Only build the Optimized kernels."
   echo "  --quantized          Only build the Quantized kernels."
+  echo "  --torchao            Only build the TorchAO kernels."
   echo "  --xnnpack            Only build the XNNPACK backend."
   echo
   exit 0
@@ -154,6 +160,7 @@ set_cmake_options_override() {
       "-DEXECUTORCH_BUILD_MPS=OFF"
       "-DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=OFF"
       "-DEXECUTORCH_BUILD_KERNELS_QUANTIZED=OFF"
+      "-DEXECUTORCH_BUILD_KERNELS_TORCHAO=OFF"
       "-DEXECUTORCH_BUILD_XNNPACK=OFF"
     )
   fi
@@ -184,6 +191,7 @@ for arg in "$@"; do
       --mps) set_cmake_options_override "EXECUTORCH_BUILD_MPS" ;;
       --optimized) set_cmake_options_override "EXECUTORCH_BUILD_KERNELS_OPTIMIZED" ;;
       --quantized) set_cmake_options_override "EXECUTORCH_BUILD_KERNELS_QUANTIZED" ;;
+      --torchao) set_cmake_options_override "EXECUTORCH_BUILD_KERNELS_TORCHAO" ;;
       --xnnpack) set_cmake_options_override "EXECUTORCH_BUILD_XNNPACK" ;;
       *)
         echo -e "\033[31m[error] unknown option: ${arg}\033[0m"
@@ -311,6 +319,7 @@ for mode in "${MODES[@]}"; do
   append_framework_flag "EXECUTORCH_BUILD_KERNELS_LLM" "$FRAMEWORK_KERNELS_LLM" "$mode"
   append_framework_flag "EXECUTORCH_BUILD_KERNELS_OPTIMIZED" "$FRAMEWORK_KERNELS_OPTIMIZED" "$mode"
   append_framework_flag "EXECUTORCH_BUILD_KERNELS_QUANTIZED" "$FRAMEWORK_KERNELS_QUANTIZED" "$mode"
+  append_framework_flag "EXECUTORCH_BUILD_KERNELS_TORCHAO" "$FRAMEWORK_KERNELS_TORCHAO" "$mode"
 
   cd "${OUTPUT_DIR}"
   "$SOURCE_ROOT_DIR"/scripts/create_frameworks.sh "${FRAMEWORK_FLAGS[@]}"

tools/cmake/executorch-config.cmake

@@ -84,6 +84,8 @@ set(optional_lib_list
     quantized_kernels
     quantized_ops_lib
     quantized_ops_aot_lib
+    torchao_ops_executorch
+    torchao_kernels_aarch64
 )
 
 foreach(lib ${optional_lib_list})

tools/cmake/preset/apple_common.cmake

@@ -29,3 +29,4 @@ set_overridable_option(EXECUTORCH_BUILD_EXTENSION_TENSOR ON)
 set_overridable_option(EXECUTORCH_BUILD_KERNELS_LLM ON)
 set_overridable_option(EXECUTORCH_BUILD_KERNELS_OPTIMIZED ON)
 set_overridable_option(EXECUTORCH_BUILD_KERNELS_QUANTIZED ON)
+set_overridable_option(EXECUTORCH_BUILD_KERNELS_TORCHAO ON)