Build optimized_portable_kernels if threadpool is enabled

build/cmake_deps.toml

@@ -89,6 +89,7 @@ deps = [
   "executorch",
   "executorch_core",
   "extension_threadpool",
+  "optimized_cpublas",
   "portable_kernels",
 ]
 
@@ -146,6 +147,7 @@ deps = [
   "executorch_core",
   "executorch",
   "extension_threadpool",
+  "optimized_cpublas",
   "portable_kernels",
 ]
 # ---------------------------------- core end ----------------------------------
@@ -413,6 +415,7 @@ excludes = [
 deps = [
   "executorch",
   "executorch_core",
+  "optimized_cpublas",
   "optimized_kernels",
   "extension_threadpool",
   "reduce_util",
@@ -452,6 +455,7 @@ deps = [
   "extension_data_loader",
   "extension_module",
   "extension_threadpool",
+  "optimized_cpublas",
   "portable_kernels",
   "quantized_kernels",
   "xnnpack_backend",

build/executorch-config.cmake

@@ -89,6 +89,7 @@ set(lib_list
     pthreadpool
     vulkan_backend
     optimized_kernels
+    optimized_portable_kernels
     cpublas
     eigen_blas
     optimized_ops_lib
@@ -132,7 +133,26 @@ endforeach()
 # target_compile_options/target_compile_definitions for everything.
 if(TARGET cpublas)
   set_target_properties(
-    cpublas PROPERTIES INTERFACE_LINK_LIBRARIES extension_threadpool
+    cpublas PROPERTIES INTERFACE_LINK_LIBRARIES
+                       "extension_threadpool;eigen_blas"
+  )
+endif()
+if(TARGET optimized_kernels)
+  set_target_properties(
+    optimized_kernels PROPERTIES INTERFACE_LINK_LIBRARIES
+                                 "executorch_core;cpublas;extension_threadpool"
+  )
+endif()
+if(TARGET optimized_native_cpu_ops_lib)
+  if(TARGET optimized_portable_kernels)
+    set(_maybe_optimized_portable_kernels_lib optimized_portable_kernels)
+  else()
+    set(_maybe_optimized_portable_kernels_lib portable_kernels)
+  endif()
+  set_target_properties(
+    optimized_native_cpu_ops_lib
+    PROPERTIES INTERFACE_LINK_LIBRARIES
+               "optimized_kernels;${_maybe_optimized_portable_kernels_lib}"
   )
 endif()
 if(TARGET extension_threadpool)

configurations/CMakeLists.txt

@@ -47,12 +47,17 @@ if(EXECUTORCH_BUILD_KERNELS_OPTIMIZED)
   message("Generated files ${gen_command_sources}")
 
   # optimized_native_cpu_ops_lib: Register optimized op kernels into the runtime
+  if(TARGET optimized_portable_kernels)
+    set(_optimized_native_cpu_ops_lib_portable_kernels_lib optimized_portable_kernels)
+  else()
+    set(_optimized_native_cpu_ops_lib_portable_kernels_lib portable_kernels)
+  endif()
   gen_operators_lib(
     LIB_NAME
     "optimized_native_cpu_ops_lib"
     KERNEL_LIBS
-    portable_kernels
     optimized_kernels
+    ${_optimized_native_cpu_ops_lib_portable_kernels_lib}
     DEPS
     executorch
   )

extension/android/CMakeLists.txt

@@ -84,10 +84,6 @@ if(TARGET optimized_native_cpu_ops_lib)
     APPEND
     link_libraries
     optimized_native_cpu_ops_lib
-    optimized_kernels
-    portable_kernels
-    cpublas
-    eigen_blas
   )
   target_link_options_shared_lib(optimized_native_cpu_ops_lib)
 else()

kernels/portable/CMakeLists.txt

@@ -63,6 +63,22 @@ gen_operators_lib(
   LIB_NAME "portable_ops_lib" KERNEL_LIBS portable_kernels DEPS executorch
 )
 
+# Portable kernels support optional parallelization (and, in the
+# future, perhaps other performance features). If support is present,
+# produce an optimized version.
+set(BUILD_OPTIMIZED_PORTABLE_KERNELS EXECUTORCH_BUILD_PTHREADPOOL)
+
+if(BUILD_OPTIMIZED_PORTABLE_KERNELS)
+  add_library(optimized_portable_kernels ${_portable_kernels__srcs})
+  target_link_libraries(optimized_portable_kernels PRIVATE executorch)
+  target_link_libraries(optimized_portable_kernels PUBLIC extension_threadpool)
+  target_compile_options(optimized_portable_kernels PUBLIC ${_common_compile_options})
+  install(
+    TARGETS optimized_portable_kernels
+    DESTINATION lib
+  )
+endif()
+
 install(
   TARGETS portable_kernels portable_ops_lib
   DESTINATION lib

kernels/test/CMakeLists.txt

@@ -23,11 +23,11 @@ foreach(kernel ${_kernels})
       "${CMAKE_CURRENT_BINARY_DIR}/include/${kernel}/executorch/kernels/test"
   )
   set(_wrapper_path "${_wrapper_dir}/FunctionHeaderWrapper.h")
+  set(_functions_include "#include <executorch/kernels/${kernel}/Functions.h>")
   add_custom_command(
     OUTPUT "${_wrapper_path}"
     COMMAND mkdir -p ${_wrapper_dir}
-    COMMAND echo "#include <executorch/kernels/${kernel}/Functions.h>" >
-            "${_wrapper_path}"
+    COMMAND echo ${_functions_include} > "${_wrapper_path}"
     DEPENDS
       "${CMAKE_CURRENT_BINARY_DIR}/include/${kernel}/executorch/kernels/${kernel}/Functions.h"
       "${CMAKE_CURRENT_BINARY_DIR}/include/${kernel}/executorch/kernels/${kernel}/NativeFunctions.h"
@@ -53,7 +53,17 @@ foreach(kernel ${_kernels})
     COMMENT "Generating ${_wrapper_dir}/supported_features.cpp and header"
     VERBATIM
   )
-
+  if(${kernel} STREQUAL "optimized")
+    set(_kernel_ops_lib "optimized_native_cpu_ops_lib")
+    set(_kernel_ops_lib_path
+        "${CMAKE_CURRENT_BINARY_DIR}/../../configurations/optimized_native_cpu_ops_lib"
+    )
+  else()
+    set(_kernel_ops_lib "${kernel}_ops_lib")
+    set(_kernel_ops_lib_path
+        "${CMAKE_CURRENT_BINARY_DIR}/../../kernels/${kernel}/${kernel}_ops_lib"
+    )
+  endif()
   add_custom_command(
     OUTPUT
       "${CMAKE_CURRENT_BINARY_DIR}/include/${kernel}/executorch/kernels/${kernel}/Functions.h"
@@ -63,10 +73,9 @@ foreach(kernel ${_kernels})
       mkdir -p
       "${CMAKE_CURRENT_BINARY_DIR}/include/${kernel}/executorch/kernels/${kernel}/"
     COMMAND
-      cp
-      "${CMAKE_CURRENT_BINARY_DIR}/../../kernels/${kernel}/${kernel}_ops_lib/*.h"
+      cp "${_kernel_ops_lib_path}/*.h"
       "${CMAKE_CURRENT_BINARY_DIR}/include/${kernel}/executorch/kernels/${kernel}/"
-    DEPENDS "${kernel}_ops_lib"
+    DEPENDS ${_kernel_ops_lib}
   )
 endforeach()
 
@@ -280,23 +289,28 @@ set(_optimized_kernels_test_sources
     ${CMAKE_CURRENT_BINARY_DIR}/include/optimized/executorch/kernels/test/supported_features.cpp
 )
 
+if(TARGET optimized_portable_kernels)
+  list(APPEND _optimized_kernels_test_sources ${all_test_sources})
+  list(REMOVE_DUPLICATES _optimized_kernels_test_sources)
+endif()
+
 et_cxx_test(
   optimized_kernels_test
   SOURCES
   ${_optimized_kernels_test_sources}
   EXTRA_LIBS
   cpuinfo
   extension_threadpool
-  optimized_kernels
-  optimized_ops_lib
-  portable_kernels
+  optimized_native_cpu_ops_lib
   pthreadpool
   eigen_blas
 )
 add_dependencies(optimized_kernels_test generate_wrapper)
 target_include_directories(
-  optimized_kernels_test PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/include/optimized"
-                                 "${CMAKE_INSTALL_PREFIX}/include"
+  optimized_kernels_test
+  PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/include/optimized"
+          "${CMAKE_CURRENT_BINARY_DIR}/include/portable"
+          "${CMAKE_INSTALL_PREFIX}/include"
 )
 
 if(TARGET quantized_kernels)