add xnnpack backend (#2585)

Summary:
Adding support for OSS Dynamically Quantized Linear via XNNPACK for llama models.

This now works for cmake on non-mac devices. Additionally, we modified the test-llama-runner-linux work flow and script to test exporting and running via XNNPACK DQ Linear

Pull Request resolved: #2585

Reviewed By: kirklandsign

Differential Revision: D55229489

Pulled By: mcr229

fbshipit-source-id: 26ffbc6d19567ebc0df3c16389451150cc988d08

Author: mcr229

backends/xnnpack/CMakeLists.txt

@@ -72,26 +72,18 @@ target_include_directories(
   xnnpack_schema INTERFACE ${_xnnpack_schema__include_dir}
                            ${EXECUTORCH_ROOT}/third-party/flatbuffers/include)
 
-set(xnn_executor_runner_libs ${_executor_runner_libs} xnnpack_schema)
-
-list(TRANSFORM _xnnpack_dynamic_quant_utils__srcs PREPEND "${EXECUTORCH_ROOT}/")
-add_library(xnnpack_dynamic_quant_utils ${_xnnpack_dynamic_quant_utils__srcs})
-set_target_properties(xnnpack_dynamic_quant_utils PROPERTIES LINKER_LANGUAGE
-                                                             CXX)
-target_include_directories(xnnpack_dynamic_quant_utils
-                           PUBLIC ${_common_include_directories})
-target_compile_options(xnnpack_dynamic_quant_utils
-                       PUBLIC ${_common_compile_options})
-
-if(ENABLE_DYNAMIC_QUANTIZATION)
-  list(APPEND xnn_executor_runner_libs xnnpack_dynamic_quant_utils)
-endif()
+set(xnnpack_third_party)
 
 include(cmake/Dependencies.cmake)
 
 list(TRANSFORM _xnnpack_backend__srcs PREPEND "${EXECUTORCH_ROOT}/")
-add_library(xnnpack_backend ${_xnnpack_backend__srcs})
-target_link_libraries(xnnpack_backend PRIVATE ${xnn_executor_runner_libs})
+add_library(xnnpack_backend STATIC ${_xnnpack_backend__srcs})
+target_link_libraries(xnnpack_backend
+                      PRIVATE
+                      ${xnnpack_third_party}
+                      executorch
+                      xnnpack_schema)
+
 target_include_directories(xnnpack_backend
                            PUBLIC ${_common_include_directories})
 target_include_directories(xnnpack_backend PUBLIC ${XNNPACK_INCLUDE_DIR})
@@ -108,11 +100,12 @@ if(NOT CMAKE_TOOLCHAIN_FILE MATCHES ".*iOS\.cmake$")
   #
   list(TRANSFORM _xnn_executor_runner__srcs PREPEND "${EXECUTORCH_ROOT}/")
   add_executable(xnn_executor_runner ${_xnn_executor_runner__srcs})
-  target_link_libraries(xnn_executor_runner ${xnn_executor_runner_libs})
+  target_link_libraries(xnn_executor_runner xnnpack_backend gflags)
   target_compile_options(xnn_executor_runner PUBLIC ${_common_compile_options})
-
-  add_library(xnn_executor_runner_lib STATIC ${_xnn_executor_runner__srcs})
-  target_link_libraries(xnn_executor_runner_lib ${xnn_executor_runner_libs})
-  target_compile_options(xnn_executor_runner_lib
-                         PUBLIC ${_common_compile_options})
 endif()
+
+install(
+  TARGETS xnnpack_backend
+  DESTINATION lib
+  INCLUDES
+  DESTINATION ${_common_include_directories})

backends/xnnpack/cmake/Dependencies.cmake

@@ -18,7 +18,7 @@ set(CPUINFO_LIBRARY_TYPE "static" CACHE STRING "")
 set(CPUINFO_LOG_LEVEL "error" CACHE STRING "")
 set(CLOG_SOURCE_DIR "${CPUINFO_SOURCE_DIR}/deps/clog")
 add_subdirectory("${CPUINFO_SOURCE_DIR}")
-list(APPEND xnn_executor_runner_libs cpuinfo)
+list(APPEND xnnpack_third_party cpuinfo)
 
 # --- pthreadpool
 set(PTHREADPOOL_SOURCE_DIR "${THIRD_PARTY_ROOT}/pthreadpool")
@@ -27,14 +27,15 @@ set(PTHREADPOOL_BUILD_BENCHMARKS OFF CACHE BOOL "")
 set(PTHREADPOOL_LIBRARY_TYPE "static" CACHE STRING "")
 set(PTHREADPOOL_ALLOW_DEPRECATED_API ON CACHE BOOL "")
 add_subdirectory("${PTHREADPOOL_SOURCE_DIR}")
-list(APPEND xnn_executor_runner_libs pthreadpool)
+list(APPEND xnnpack_third_party pthreadpool)
 
 # --- XNNPACK
 set(XNNPACK_SOURCE_DIR "${THIRD_PARTY_ROOT}/XNNPACK")
 set(XNNPACK_INCLUDE_DIR "${XNNPACK_SOURCE_DIR}/include")
 set(XNNPACK_LIBRARY_TYPE "static" CACHE STRING "")
 set(XNNPACK_BUILD_BENCHMARKS OFF CACHE BOOL "")
 set(XNNPACK_BUILD_TESTS OFF CACHE BOOL "")
+set(XNNPACK_ENABLE_AVXVNNI OFF CACHE BOOL "")
 add_subdirectory("${XNNPACK_SOURCE_DIR}")
 include_directories(SYSTEM ${XNNPACK_INCLUDE_DIR})
-list(APPEND xnn_executor_runner_libs XNNPACK)
+list(APPEND xnnpack_third_party XNNPACK)

build/executorch-config.cmake

@@ -43,7 +43,8 @@ endif()
 
 set(lib_list
     etdump bundled_program extension_data_loader ${FLATCC_LIB} mpsdelegate
-    qnn_executorch_backend portable_ops_lib extension_module
+    qnn_executorch_backend portable_ops_lib extension_module xnnpack_backend
+    XNNPACK cpuinfo pthreadpool
 )
 foreach(lib ${lib_list})
     # Name of the variable which stores result of the find_library search

examples/models/llama2/CMakeLists.txt

@@ -52,12 +52,23 @@ find_package(executorch CONFIG REQUIRED)
 # llama_runner library
 add_subdirectory(runner)
 
-# XNNPACK pthreadpool cpuinfo
-target_link_libraries(llama_main PUBLIC gflags llama_runner portable_ops_lib)
+
+target_link_libraries(llama_main PUBLIC gflags llama_runner
+                      portable_ops_lib)
 target_link_options(
   llama_main PUBLIC "SHELL:LINKER:--whole-archive \
                     $<TARGET_FILE:portable_ops_lib> \
                     LINKER:--no-whole-archive")
+
+# XNNPACK pthreadpool cpuinfo
+if(TARGET xnnpack_backend)
+  set(xnnpack_backend_libs xnnpack_backend XNNPACK pthreadpool cpuinfo)
+  target_link_libraries(llama_main PUBLIC ${xnnpack_backend_libs})
+  target_link_options(
+    llama_main PUBLIC "SHELL:LINKER:--whole-archive \
+                      $<TARGET_FILE:xnnpack_backend> \
+                      LINKER:--no-whole-archive")
+endif()
 target_compile_options(llama_main PUBLIC ${_common_compile_options})
 
 # Print all summary