Use c10 version of half/bfloat16 in executorch (#8320)

Pull Request resolved: #7040

Pull Request resolved: pytorch/pytorch#144111

Accomplished by importing relevant files from c10 into
executorch/runtime/core/portable_type/c10, and then using `using` in
the top-level ExecuTorch headers. This approach should keep the
ExecuTorch build hermetic for embedded use cases. In the future, we
should add a CI job to ensure the c10 files stay identical to the
PyTorch ones.
ghstack-source-id: 265190624
@exported-using-ghexport

Differential Revision: [D66106969](https://our.internmc.facebook.com/intern/diff/D66106969/)

Co-authored-by: Github Executorch <[email protected]>

Author: pytorchbot

.lintrunner.toml

@@ -78,6 +78,8 @@ exclude_patterns = [
     # File contains @generated
     'extension/llm/custom_ops/spinquant/fast_hadamard_transform_special.h',
     'extension/llm/custom_ops/spinquant/test/fast_hadamard_transform_special_unstrided_cpu.h',
+    # Want to be able to keep c10 in sync with PyTorch core.
+    'runtime/core/portable_type/c10/**',
 ]
 command = [
     'python',
@@ -261,6 +263,8 @@ exclude_patterns = [
     'extension/**',
     'kernels/optimized/**',
     'runtime/core/exec_aten/**',
+    # Want to be able to keep c10 in sync with PyTorch core.
+    'runtime/core/portable_type/c10/**',
     'runtime/executor/tensor_parser_aten.cpp',
     'scripts/**',
     'test/**',

CMakeLists.txt

@@ -373,7 +373,7 @@ if(NOT "${_repo_dir_name}" STREQUAL "executorch")
       "fix for this restriction."
   )
 endif()
-set(_common_include_directories ${CMAKE_CURRENT_SOURCE_DIR}/..)
+set(_common_include_directories ${CMAKE_CURRENT_SOURCE_DIR}/.. ${CMAKE_CURRENT_SOURCE_DIR}/runtime/core/portable_type)
 
 #
 # The `_<target>_srcs` lists are defined by including ${EXECUTORCH_SRCS_FILE}.
@@ -556,6 +556,7 @@ endif()
 target_include_directories(
   executorch_core PUBLIC ${_common_include_directories}
 )
+target_compile_definitions(executorch_core PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
 target_compile_options(executorch_core PUBLIC ${_common_compile_options})
 if(MAX_KERNEL_NUM)
   target_compile_definitions(
@@ -576,6 +577,7 @@ if(EXECUTORCH_BUILD_PYBIND AND APPLE)
   target_include_directories(
     executorch_core_shared PUBLIC ${_common_include_directories}
   )
+  target_compile_definitions(executorch_core_shared PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
   target_compile_options(
     executorch_core_shared PUBLIC ${_common_compile_options}
   )
@@ -596,6 +598,7 @@ endif()
 add_library(executorch ${_executorch__srcs})
 target_link_libraries(executorch PRIVATE executorch_core)
 target_include_directories(executorch PUBLIC ${_common_include_directories})
+target_compile_definitions(executorch PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
 target_compile_options(executorch PUBLIC ${_common_compile_options})
 target_link_options_shared_lib(executorch)
 
@@ -629,6 +632,12 @@ endif()
 
 # Install `executorch` library as well as `executorch-config.cmake` under
 # ${CMAKE_INSTALL_PREFIX}/
+install(DIRECTORY runtime/core/  DESTINATION include/executorch/runtime/core FILES_MATCHING PATTERN "*.h")
+install(DIRECTORY runtime/kernel/  DESTINATION include/executorch/runtime/kernel FILES_MATCHING PATTERN "*.h")
+install(DIRECTORY runtime/platform/  DESTINATION include/executorch/runtime/platform FILES_MATCHING PATTERN "*.h")
+install(DIRECTORY extension/kernel_util/  DESTINATION include/executorch/extension/kernel_util FILES_MATCHING PATTERN "*.h")
+install(DIRECTORY extension/tensor/  DESTINATION include/executorch/extension/tensor FILES_MATCHING PATTERN "*.h")
+install(DIRECTORY extension/threadpool/  DESTINATION include/executorch/extension/threadpool FILES_MATCHING PATTERN "*.h")
 install(
   TARGETS executorch executorch_core
   DESTINATION lib
@@ -792,6 +801,8 @@ if(EXECUTORCH_BUILD_PYBIND)
   target_include_directories(
     util PUBLIC ${_common_include_directories} ${TORCH_INCLUDE_DIRS}
   )
+  target_compile_definitions(util PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
+
   target_compile_options(util PUBLIC ${_pybind_compile_options})
   target_link_libraries(util PRIVATE torch c10 executorch extension_tensor)
 

backends/apple/coreml/CMakeLists.txt

@@ -134,6 +134,8 @@ target_include_directories(
   coremldelegate PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/runtime/util
 )
 target_include_directories(coremldelegate PRIVATE ${EXECUTORCH_ROOT}/..)
+target_include_directories(coremldelegate PRIVATE ${EXECUTORCH_ROOT}/runtime/core/portable_type)
+target_compile_definitions(coremldelegate PRIVATE C10_USING_CUSTOM_GENERATED_MACROS)
 target_link_libraries(coremldelegate PRIVATE executorch_core)
 
 if(EXECUTORCH_BUILD_DEVTOOLS)

backends/apple/coreml/runtime/workspace/executorchcoreml.xcodeproj/project.pbxproj

@@ -830,6 +830,7 @@
 				GCC_OPTIMIZATION_LEVEL = 0;
 				GCC_PREPROCESSOR_DEFINITIONS = (
 					"DEBUG=1",
+                                        "C10_USING_CUSTOM_GENERATED_MACROS",
 					"$(inherited)",
 				);
 				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
@@ -911,6 +912,7 @@
 				DEVELOPMENT_TEAM = "";
 				GCC_PREPROCESSOR_DEFINITIONS = (
 					"DEBUG=1",
+                                        "C10_USING_CUSTOM_GENERATED_MACROS",
 					"ET_EVENT_TRACER_ENABLED=1",
 					"$(inherited)",
 				);
@@ -920,6 +922,7 @@
 					"$(SRCROOT)/../kvstore",
 					"$(SRCROOT)/../inmemoryfs",
 					"$(SRCROOT)/../include",
+					"$(SRCROOT)/../include/executorch/runtime/core/portable_type",
 					"$(SRCROOT)/../sdk",
 					"$(SRCROOT)/../util",
 					"$(SRCROOT)/../../third-party/nlohmann_json/single_include",
@@ -951,6 +954,7 @@
 					"$(SRCROOT)/../kvstore",
 					"$(SRCROOT)/../inmemoryfs",
 					"$(SRCROOT)/../include",
+					"$(SRCROOT)/../include/executorch/runtime/core/portable_type",
 					"$(SRCROOT)/../sdk",
 					"$(SRCROOT)/../util",
 					"$(SRCROOT)/../../third-party/nlohmann_json/single_include",

backends/arm/CMakeLists.txt

@@ -14,7 +14,8 @@ endif()
 
 include(${EXECUTORCH_ROOT}/build/Utils.cmake)
 
-set(_common_include_directories ${EXECUTORCH_ROOT}/..)
+set(_common_include_directories ${EXECUTORCH_ROOT}/.. ${EXECUTORCH_ROOT}/runtime/core/portable_type)
+add_compile_definitions(C10_USING_CUSTOM_GENERATED_MACROS)
 
 # Third-party folder and Ethos-U driver inclued
 set(THIRD_PARTY_ROOT "${CMAKE_CURRENT_SOURCE_DIR}/third-party")

backends/qualcomm/CMakeLists.txt

@@ -54,6 +54,7 @@ add_custom_command(
 )
 
 add_compile_options("-Wall" "-Werror" "-Wno-sign-compare")
+add_compile_definitions(C10_USING_CUSTOM_GENERATED_MACROS)
 
 # GNU emit wanring for ignored attributes Unfortunately, we use [[maybe_unused]]
 # which can be ignored by GNU. So we make it a warning, not an error in GNU.
@@ -73,6 +74,7 @@ endif()
 include_directories(
   BEFORE ${_common_include_directories} ${QNN_SDK_ROOT}/include/QNN
   ${EXECUTORCH_SOURCE_DIR}/third-party/flatbuffers/include
+  ${EXECUTORCH_SOURCE_DIR}/runtime/core/portable_type
 )
 
 set(_qnn_schema__srcs

backends/xnnpack/CMakeLists.txt

@@ -119,7 +119,7 @@ target_include_directories(
 target_compile_options(xnnpack_backend PUBLIC ${_common_compile_options})
 target_link_options_shared_lib(xnnpack_backend)
 
-list(APPEND xnn_executor_runner_libs xnnpack_backend)
+list(APPEND xnn_executor_runner_libs xnnpack_backend executorch)
 
 # ios can only build library but not binary
 if(NOT CMAKE_TOOLCHAIN_FILE MATCHES ".*(iOS|ios\.toolchain)\.cmake$")

build/build_apple_frameworks.sh

@@ -200,6 +200,20 @@ check_command "$BUCK2"
   //extension/tensor: \
 | rsync -av --files-from=- "$SOURCE_ROOT_DIR" "$HEADERS_PATH/executorch"
 
+# HACK: XCFrameworks don't appear to support exporting any build
+# options, but we need the following:
+# - runtime/core/portable/type/c10 reachable with `#include <c10/...>`
+# - exported -DC10_USING_CUSTOM_GENERATED_MACROS compiler flag
+# So, just patch our generated framework to do that.
+sed -i '' '1i\
+#define C10_USING_CUSTOM_GENERATED_MACROS
+' $HEADERS_PATH/executorch/runtime/core/portable_type/c10/macros/Macros.h
+sed -i '' '1i\
+#define C10_USING_CUSTOM_GENERATED_MACROS
+' $HEADERS_PATH/executorch/runtime/core/portable_type/c10/macros/Export.h
+ln -s $HEADERS_PATH/executorch/runtime/core/portable_type/c10 "$HEADERS_PATH/"
+
+
 cp "$SOURCE_ROOT_DIR/extension/apple/ExecuTorch/Exported/"*.h "$HEADERS_PATH/executorch"
 cp "$SOURCE_ROOT_DIR/extension/apple/ExecuTorch/Exported/"*.modulemap "$HEADERS_PATH"
 

build/executorch-config.cmake

@@ -26,20 +26,21 @@
 
 cmake_minimum_required(VERSION 3.19)
 
-set(_root "${CMAKE_CURRENT_LIST_DIR}/../..")
+set(_root "${CMAKE_CURRENT_LIST_DIR}/../../..")
 set(required_lib_list executorch executorch_core portable_kernels)
 set(EXECUTORCH_LIBRARIES)
-set(EXECUTORCH_INCLUDE_DIRS ${_root})
+set(EXECUTORCH_INCLUDE_DIRS ${_root}/include ${_root}/include/executorch/runtime/core/portable_type ${_root}/lib)
 foreach(lib ${required_lib_list})
   set(lib_var "LIB_${lib}")
   add_library(${lib} STATIC IMPORTED)
   find_library(
     ${lib_var} ${lib}
-    HINTS "${_root}"
+    HINTS "${_root}/lib"
     CMAKE_FIND_ROOT_PATH_BOTH
   )
   set_target_properties(${lib} PROPERTIES IMPORTED_LOCATION "${${lib_var}}")
-  target_include_directories(${lib} INTERFACE ${_root})
+  target_compile_definitions(${lib} INTERFACE C10_USING_CUSTOM_GENERATED_MACROS)
+  target_include_directories(${lib} INTERFACE ${_root}/include ${_root}/include/executorch/runtime/core/portable_type ${_root}/lib)
   list(APPEND EXECUTORCH_LIBRARIES ${lib})
 endforeach()
 
@@ -93,7 +94,7 @@ foreach(lib ${lib_list})
   set(lib_var "LIB_${lib}")
   find_library(
     ${lib_var} ${lib}
-    HINTS "${_root}"
+    HINTS "${_root}/lib"
     CMAKE_FIND_ROOT_PATH_BOTH
   )
   if(NOT ${lib_var})
@@ -109,7 +110,7 @@ foreach(lib ${lib_list})
       add_library(${lib} STATIC IMPORTED)
     endif()
     set_target_properties(${lib} PROPERTIES IMPORTED_LOCATION "${${lib_var}}")
-    target_include_directories(${lib} INTERFACE ${_root})
+    target_include_directories(${lib} INTERFACE ${_root}/include ${_root}/include/executorch/runtime/core/portable_type ${_root}/lib)
     list(APPEND EXECUTORCH_LIBRARIES ${lib})
   endif()
 endforeach()

examples/apple/coreml/executor_runner/coreml_executor_runner.xcodeproj/project.pbxproj

@@ -214,6 +214,7 @@
 				GCC_OPTIMIZATION_LEVEL = 0;
 				GCC_PREPROCESSOR_DEFINITIONS = (
 					"DEBUG=1",
+                                	"C10_USING_CUSTOM_GENERATED_MACROS",
 					"$(inherited)",
 				);
 				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
@@ -271,6 +272,10 @@
 				ENABLE_USER_SCRIPT_SANDBOXING = YES;
 				GCC_C_LANGUAGE_STANDARD = gnu17;
 				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_PREPROCESSOR_DEFINITIONS = (
+                                	"C10_USING_CUSTOM_GENERATED_MACROS",
+					"$(inherited)",
+				);
 				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
 				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
 				GCC_WARN_UNDECLARED_SELECTOR = YES;
@@ -291,7 +296,10 @@
 				CODE_SIGN_STYLE = Automatic;
 				DEVELOPMENT_TEAM = "";
 				ENABLE_HARDENED_RUNTIME = YES;
-				HEADER_SEARCH_PATHS = "$(SRCROOT)/include";
+				HEADER_SEARCH_PATHS = (
+                                	"$(SRCROOT)/include",
+                                	"$(SRCROOT)/include/executorch/runtime/core/portable_type",
+                                );
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				LIBRARY_SEARCH_PATHS = (
 					"$(SRCROOT)/libraries",
@@ -310,7 +318,10 @@
 				CODE_SIGN_STYLE = Automatic;
 				DEVELOPMENT_TEAM = "";
 				ENABLE_HARDENED_RUNTIME = YES;
-				HEADER_SEARCH_PATHS = "$(SRCROOT)/include";
+				HEADER_SEARCH_PATHS = (
+                                	"$(SRCROOT)/include",
+                                	"$(SRCROOT)/include/executorch/runtime/core/portable_type",
+                                );
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				LIBRARY_SEARCH_PATHS = (
 					"$(SRCROOT)/libraries",

examples/arm/executor_runner/CMakeLists.txt

@@ -382,8 +382,9 @@ target_link_options( arm_executor_runner PUBLIC LINKER:-Map=arm_executor_runner.
 
 # ET headers and generated headers includes
 target_include_directories(
-  arm_executor_runner PRIVATE ${ET_INCLUDE_PATH} ${CMAKE_CURRENT_BINARY_DIR}
+  arm_executor_runner PRIVATE ${ET_INCLUDE_PATH} ${ET_DIR_PATH}/runtime/core/portable_type ${CMAKE_CURRENT_BINARY_DIR}
 )
+target_compile_definitions(arm_executor_runner PRIVATE C10_USING_CUSTOM_GENERATED_MACROS)
 
 if(SEMIHOSTING)
   target_compile_definitions(arm_executor_runner PUBLIC SEMIHOSTING)