Make a separate target for kernel utils (#10788)

Currently it's built into portable kernels only. Other kernels used it,
and the runner must include portable kernel to avoid undefined symbols.
We need to split these utils out. A fix for
#10677 without BC breaking.

Author: kirklandsign

CMakeLists.txt

@@ -547,6 +547,8 @@ if(EXECUTORCH_BUILD_KERNELS_OPTIMIZED)
   find_package_torch_headers()
 endif()
 
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/kernels/portable/cpu/util)
+
 if(BUILD_EXECUTORCH_PORTABLE_OPS)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/kernels/portable)
 endif()

extension/llm/custom_ops/CMakeLists.txt

@@ -119,7 +119,7 @@ if(EXECUTORCH_BUILD_KERNELS_CUSTOM_AOT)
   else()
     # If no portable_lib, custom_ops_aot_lib still gives the ability to use the
     # ops in PyTorch
-    target_link_libraries(custom_ops_aot_lib PUBLIC executorch_core)
+    target_link_libraries(custom_ops_aot_lib PUBLIC executorch_core kernels_util_all_deps)
   endif()
 
   target_link_libraries(

extension/training/CMakeLists.txt

@@ -25,8 +25,8 @@ target_include_directories(
 
 target_include_directories(extension_training PUBLIC ${EXECUTORCH_ROOT}/..)
 target_compile_options(extension_training PUBLIC ${_common_compile_options})
-target_link_libraries(extension_training executorch_core
-    extension_data_loader extension_module_static extension_tensor extension_flat_tensor)
+target_link_libraries(extension_training executorch_core kernels_util_all_deps
+    extension_data_loader extension_module_static extension_tensor extension_flat_tensor )
 
 
 list(TRANSFORM _train_xor__srcs PREPEND "${EXECUTORCH_ROOT}/")

kernels/optimized/CMakeLists.txt

@@ -64,7 +64,7 @@ add_library(optimized_kernels ${_optimized_kernels__srcs})
 target_include_directories(optimized_kernels PRIVATE ${TORCH_INCLUDE_DIRS} "${EXECUTORCH_ROOT}/third-party/pocketfft")
 target_compile_definitions(optimized_kernels PRIVATE ET_USE_PYTORCH_HEADERS)
 target_link_libraries(
-  optimized_kernels PUBLIC executorch_core cpublas extension_threadpool
+  optimized_kernels PUBLIC executorch_core cpublas extension_threadpool kernels_util_all_deps
 )
 target_compile_options(optimized_kernels PUBLIC ${_common_compile_options})
 # Build a library for _optimized_kernels_srcs

kernels/portable/CMakeLists.txt

@@ -52,7 +52,7 @@ message("Generated files ${gen_command_sources}")
 # Focused on portability and understandability rather than speed.
 #
 add_library(portable_kernels ${_portable_kernels__srcs})
-target_link_libraries(portable_kernels PRIVATE executorch_core)
+target_link_libraries(portable_kernels PRIVATE executorch_core kernels_util_all_deps)
 target_compile_options(portable_kernels PUBLIC ${_common_compile_options})
 
 # Build a library for _portable_kernels__srcs

kernels/portable/cpu/util/CMakeLists.txt

@@ -0,0 +1,30 @@
+# 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.
+
+# Kernel library for portable kernels. Please this file formatted by running:
+# ~~~
+# cmake-format -i CMakeLists.txt
+# ~~~
+
+cmake_minimum_required(VERSION 3.19)
+
+if(NOT CMAKE_CXX_STANDARD)
+  set(CMAKE_CXX_STANDARD 17)
+endif()
+
+if(NOT EXECUTORCH_ROOT)
+  set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../../..)
+endif()
+
+list(TRANSFORM _kernels_util_all_deps__srcs PREPEND "${EXECUTORCH_ROOT}/")
+
+set(_common_compile_options -Wno-deprecated-declarations)
+
+add_library(kernels_util_all_deps ${_kernels_util_all_deps__srcs})
+target_link_libraries(kernels_util_all_deps PRIVATE executorch_core)
+target_include_directories(kernels_util_all_deps PUBLIC ${_common_include_directories})
+target_compile_definitions(kernels_util_all_deps PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
+target_compile_options(kernels_util_all_deps PUBLIC ${_common_compile_options})

kernels/quantized/CMakeLists.txt

@@ -142,7 +142,7 @@ if(NOT CMAKE_GENERATOR STREQUAL "Xcode"
 endif()
 
 add_library(quantized_kernels ${_quantized_kernels__srcs})
-target_link_libraries(quantized_kernels PRIVATE executorch_core)
+target_link_libraries(quantized_kernels PRIVATE executorch_core kernels_util_all_deps)
 target_compile_options(quantized_kernels PUBLIC ${_common_compile_options})
 # Build a library for _quantized_kernels_srcs
 #

tools/cmake/cmake_deps.toml

@@ -56,6 +56,20 @@ excludes = [
 deps = [
   "executorch",
   "executorch_core",
+  "extension_threadpool",
+  "kernels_util_all_deps",
+]
+
+[targets.kernels_util_all_deps]
+buck_targets = [
+  "//kernels/portable/cpu/util:all_deps",
+]
+filters = [
+  ".cpp$",
+]
+deps = [
+  "executorch_core",
+  "extension_threadpool",
 ]
 
 # HACK: prevent reduce_util from also showing up in custom_ops. The
@@ -89,6 +103,7 @@ deps = [
   "executorch",
   "executorch_core",
   "extension_threadpool",
+  "kernels_util_all_deps",
   "optimized_cpublas",
   "portable_kernels",
 ]
@@ -108,6 +123,8 @@ excludes = [
 deps = [
   "executorch",
   "executorch_core",
+  "extension_threadpool",
+  "kernels_util_all_deps",
   "portable_kernels",
 ]
 
@@ -147,6 +164,7 @@ deps = [
   "executorch_core",
   "executorch",
   "extension_threadpool",
+  "kernels_util_all_deps",
   "optimized_cpublas",
   "portable_kernels",
 ]
@@ -288,6 +306,8 @@ excludes = [
 deps = [
   "executorch",
   "executorch_core",
+  "extension_threadpool",
+  "kernels_util_all_deps",
   "portable_kernels",
 ]
 # ---------------------------------- extension end ----------------------------------
@@ -306,6 +326,8 @@ excludes = [
 deps = [
   "executorch",
   "executorch_core",
+  "extension_threadpool",
+  "kernels_util_all_deps",
   "portable_kernels",
   "quantized_kernels",
   "etdump_flatcc",
@@ -340,6 +362,8 @@ excludes = [
 deps = [
   "executorch",
   "executorch_core",
+  "extension_threadpool",
+  "kernels_util_all_deps",
   "portable_kernels",
 ]
 
@@ -380,6 +404,7 @@ deps = [
   "executorch",
   "executorch_core",
   "extension_threadpool",
+  "kernels_util_all_deps",
   "xnnpack_backend",
   "portable_kernels",
   "etdump_flatcc",
@@ -455,6 +480,7 @@ deps = [
   "extension_module",
   "extension_tensor",
   "extension_threadpool",
+  "kernels_util_all_deps",
   "optimized_cpublas",
   "portable_kernels",
   "quantized_kernels",