Add OSS CI for the CPU backend test suite flow (#24252)

.ci/scripts/test_backend.sh

@@ -53,6 +53,10 @@ if [[ "$FLOW" == *qnn* ]]; then
     EXTRA_BUILD_ARGS+=" -DEXECUTORCH_BUILD_QNN=ON -DQNN_SDK_ROOT=$QNN_SDK_ROOT -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON"
 fi
 
+if [[ "$FLOW" == *cpu* ]]; then
+    EXTRA_BUILD_ARGS+=" -DEXECUTORCH_BUILD_CPU=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON"
+fi
+
 if [[ "$FLOW" == *vulkan* ]]; then
     # CPU runners can have Mesa ICDs installed without a usable hardware GPU.
     source .ci/scripts/setup-vulkan-linux-deps.sh "swiftshader"

.github/workflows/_test_backend.yml

@@ -4,7 +4,7 @@ on:
   workflow_call:
     inputs:
       backend:
-        description: 'Backend to test (xnnpack, coreml, vulkan, qnn, openvino)'
+        description: 'Backend to test (xnnpack, coreml, vulkan, qnn, openvino, cpu)'
         required: true
         type: string
       flows:

.github/workflows/test-backend-cpu.yml

@@ -0,0 +1,31 @@
+name: Test CPU Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    branches:
+      - main
+      - release/*
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
+  cancel-in-progress: true
+
+jobs:
+  test-cpu:
+    uses: ./.github/workflows/_test_backend.yml
+    permissions:
+      id-token: write
+      contents: read
+    with:
+      backend: cpu
+      flows: '["cpu_fp32"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-linux: true
+      docker-image: ci-image:executorch-ubuntu-22.04-gcc11

CMakeLists.txt

@@ -787,7 +787,7 @@ endif()
 if(EXECUTORCH_BUILD_NATIVE)
   string(APPEND _executorch_backend_dependencies
          "find_dependency(libzip CONFIG)\n"
-         "find_dependency(nlohmann_json CONFIG)\n" "find_dependency(Threads)\n"
+         "find_dependency(nlohmann_json CONFIG)\n"
   )
 endif()
 if(EXECUTORCH_BUILD_NATIVE_GGML)
@@ -797,7 +797,10 @@ if(EXECUTORCH_BUILD_NATIVE_GGML)
     "  add_library(executorch_ggml ALIAS ggml::ggml)\n" "endif()\n"
   )
 endif()
-if(EXECUTORCH_BUILD_VULKAN)
+if(EXECUTORCH_BUILD_NATIVE
+   OR EXECUTORCH_BUILD_CPU
+   OR EXECUTORCH_BUILD_VULKAN
+)
   string(APPEND _executorch_backend_dependencies "find_dependency(Threads)\n")
 endif()
 configure_file(
@@ -1417,6 +1420,10 @@ if(EXECUTORCH_BUILD_PYBIND)
     endif()
   endif()
 
+  if(EXECUTORCH_BUILD_CPU)
+    list(APPEND _dep_libs cpu_backend)
+  endif()
+
   if(EXECUTORCH_BUILD_VULKAN)
     list(APPEND _dep_libs vulkan_backend)
   endif()
@@ -1624,10 +1631,15 @@ if(EXECUTORCH_BUILD_VULKAN)
   list(APPEND _executorch_backends vulkan_backend vulkan_schema)
 endif()
 
-if(EXECUTORCH_BUILD_NATIVE)
+if(EXECUTORCH_BUILD_NATIVE OR EXECUTORCH_BUILD_CPU)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/native)
 endif()
 
+if(EXECUTORCH_BUILD_CPU)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/cpu)
+  list(APPEND _executorch_backends cpu_backend)
+endif()
+
 if(EXECUTORCH_BUILD_WEBGPU)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/webgpu)
   list(APPEND _executorch_backends webgpu_backend)

backends/cpu/CMakeLists.txt

@@ -4,56 +4,147 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-add_library(
-  cpu_runtime runtime/KernelProvider.cpp runtime/CPUPlan.cpp
-              runtime/op_helpers/Convolution.cpp
-)
-target_link_libraries(
-  cpu_runtime
-  PUBLIC native_runtime
-  PRIVATE extension_threadpool
+cmake_minimum_required(VERSION 3.24)
+
+if(NOT EXECUTORCH_ROOT)
+  set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../..)
+endif()
+
+include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
+
+if(EXECUTORCH_BUILD_NATIVE_CPU)
+  add_library(
+    cpu_runtime runtime/KernelProvider.cpp runtime/CPUPlan.cpp
+                runtime/op_helpers/Convolution.cpp
+  )
+  target_link_libraries(
+    cpu_runtime
+    PUBLIC native_runtime
+    PRIVATE extension_threadpool
+  )
+  if(EXECUTORCH_BUILD_SHARED)
+    target_link_libraries(cpu_runtime PUBLIC executorch_shared)
+  else()
+    target_link_libraries(cpu_runtime PUBLIC executorch_core)
+  endif()
+
+  add_library(cpu_reference runtime/providers/reference/ReferenceProvider.cpp)
+  target_link_libraries(cpu_reference PUBLIC cpu_runtime)
+
+  add_library(cpu_xnnpack runtime/providers/xnnpack/XNNPACKProvider.cpp)
+  target_link_libraries(
+    cpu_xnnpack
+    PUBLIC cpu_runtime
+    PRIVATE extension_threadpool
+  )
+  if(EXECUTORCH_BUILD_SHARED)
+    target_link_libraries(cpu_xnnpack PRIVATE xnnpack_shared)
+  else()
+    target_link_libraries(cpu_xnnpack PRIVATE XNNPACK)
+  endif()
+
+  foreach(target cpu_runtime cpu_reference cpu_xnnpack)
+    add_library(executorch::${target} ALIAS ${target})
+    target_compile_features(${target} PUBLIC cxx_std_20)
+    target_compile_options(
+      ${target} PRIVATE $<$<CXX_COMPILER_ID:MSVC>:/EHsc>
+                        $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-fexceptions>
+    )
+  endforeach()
+
+  install(
+    TARGETS cpu_runtime cpu_reference cpu_xnnpack
+    EXPORT ExecuTorchTargets
+    ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
+  )
+  install(
+    DIRECTORY runtime/
+    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/executorch/backends/cpu/runtime
+    FILES_MATCHING
+    PATTERN "*.h"
+    PATTERN "test" EXCLUDE
+  )
+endif()
+
+if(NOT EXECUTORCH_BUILD_CPU)
+  return()
+endif()
+
+# The provider composition that cpu_backend() in targets.bzl generates for the
+# default :cpu_backend target.
+set(_cpu_provider_registry ${CMAKE_CURRENT_BINARY_DIR}/provider_registry.cpp)
+file(
+  GENERATE
+  OUTPUT ${_cpu_provider_registry}
+  CONTENT
+    "#include <executorch/backends/cpu/runtime/KernelProvider.h>
+#include <executorch/backends/cpu/runtime/providers/xnnpack/XNNPACKProvider.h>
+#include <executorch/backends/cpu/runtime/providers/reference/ReferenceProvider.h>
+namespace executorch::backends::cpu {
+namespace {
+RuntimeConfiguration runtime_configuration() {
+static const ProviderFactory factories[] = {create_xnnpack_provider,create_reference_provider};
+return {{factories, sizeof(factories) / sizeof(factories[0])}, {}, false};
+}
+const bool registered = (register_runtime_configuration(runtime_configuration), true);
+}
+}
+"
 )
+
+list(TRANSFORM _cpu_backend__srcs PREPEND "${EXECUTORCH_ROOT}/")
 if(EXECUTORCH_BUILD_SHARED)
-  target_link_libraries(cpu_runtime PUBLIC executorch_shared)
+  set(_cpu_backend_library_type SHARED)
 else()
-  target_link_libraries(cpu_runtime PUBLIC executorch_core)
+  set(_cpu_backend_library_type STATIC)
 endif()
-
-add_library(cpu_reference runtime/providers/reference/ReferenceProvider.cpp)
-target_link_libraries(cpu_reference PUBLIC cpu_runtime)
-
-add_library(cpu_xnnpack runtime/providers/xnnpack/XNNPACKProvider.cpp)
+add_library(
+  cpu_backend ${_cpu_backend_library_type} ${_cpu_backend__srcs}
+              ${_cpu_provider_registry}
+)
 target_link_libraries(
-  cpu_xnnpack
-  PUBLIC cpu_runtime
-  PRIVATE extension_threadpool
+  cpu_backend PRIVATE native_program extension_threadpool
+                      optimized_native_cpu_ops_lib
 )
 if(EXECUTORCH_BUILD_SHARED)
-  target_link_libraries(cpu_xnnpack PRIVATE xnnpack_shared)
+  set_target_properties(
+    cpu_backend PROPERTIES OUTPUT_NAME executorch_backend_cpu
+  )
+  executorch_target_soname_policy(cpu_backend)
+  executorch_target_link_shared_runtime(cpu_backend)
+  # The XNNPACK delegate library already bundles XNNPACK, so resolve against it
+  # rather than shipping a second copy.
+  target_link_libraries(
+    cpu_backend
+    PUBLIC executorch_shared
+    PRIVATE xnnpack_backend
+  )
+  target_include_directories(
+    cpu_backend
+    PRIVATE $<TARGET_PROPERTY:XNNPACK,INTERFACE_INCLUDE_DIRECTORIES>
+  )
+  executorch_target_shipped_runtime_path(cpu_backend)
 else()
-  target_link_libraries(cpu_xnnpack PRIVATE XNNPACK)
+  target_link_libraries(
+    cpu_backend
+    PUBLIC executorch_core
+    PRIVATE XNNPACK xnnpack-microkernels-prod
+  )
 endif()
-
-foreach(target cpu_runtime cpu_reference cpu_xnnpack)
-  add_library(executorch::${target} ALIAS ${target})
-  target_compile_features(${target} PUBLIC cxx_std_20)
-  target_compile_options(
-    ${target} PRIVATE $<$<CXX_COMPILER_ID:MSVC>:/EHsc>
+target_compile_features(cpu_backend PRIVATE cxx_std_20)
+target_compile_options(
+  cpu_backend PRIVATE $<$<CXX_COMPILER_ID:MSVC>:/EHsc>
                       $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-fexceptions>
-  )
-endforeach()
+)
+set_target_properties(cpu_backend PROPERTIES POSITION_INDEPENDENT_CODE ON)
+executorch_target_link_options_shared_lib(cpu_backend)
 
 install(
-  TARGETS cpu_runtime cpu_reference cpu_xnnpack
+  TARGETS cpu_backend
   EXPORT ExecuTorchTargets
   ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
   LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
   RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
 )
-install(
-  DIRECTORY runtime/
-  DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/executorch/backends/cpu/runtime
-  FILES_MATCHING
-  PATTERN "*.h"
-  PATTERN "test" EXCLUDE
-)

backends/cpu/README.md

@@ -37,6 +37,13 @@ The shared `KernelProvider` contract, `CPUPlan`, and `CpuBackend` remain in
 [`targets.bzl`](targets.bzl).
 
 The Buck `cpu_backend` target links the baseline XNNPACK and Reference providers.
+The CMake `cpu_backend` target (`-DEXECUTORCH_BUILD_CPU=ON`, which requires
+`-DEXECUTORCH_BUILD_XNNPACK=ON` and `-DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON`)
+links the same composition, including the full optimized and portable kernel
+library for the Reference provider. Selective kernel builds are not supported.
+The CPU backend is disabled by default, including in Python wheels. Enable it
+for an editable install with
+`CMAKE_ARGS="-DEXECUTORCH_BUILD_CPU=ON" ./install_executorch.sh --editable`.
 Applications compose providers with [`cpu_backend()`](cpu_backend.bzl).
 
 `CPUPlan` retains selection, regions, storage requirements and prepared executables

backends/cpu/targets.bzl

@@ -1,15 +1,12 @@
 load("@fbsource//xplat/executorch/backends/xnnpack/third-party:third_party_libs.bzl", "third_party_dep")
+load("@fbsource//xplat/executorch/build:build_variables.bzl", "CPU_RUNTIME_BUCK_SRCS")
 load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
 load(":cpu_backend.bzl", "cpu_backend", "cpu_provider")
 
 def define_common_targets():
     runtime.cxx_library(
         name = "cpu_runtime",
-        srcs = [
-            "runtime/KernelProvider.cpp",
-            "runtime/CPUPlan.cpp",
-            "runtime/op_helpers/Convolution.cpp",
-        ],
+        srcs = CPU_RUNTIME_BUCK_SRCS,
         exported_headers = [
             "runtime/KernelProvider.h",
             "runtime/CPUPlan.h",

backends/native/CMakeLists.txt

@@ -11,9 +11,14 @@ if(NOT DEFINED CMAKE_POSITION_INDEPENDENT_CODE)
 endif()
 
 find_package(Threads REQUIRED)
-find_package(libzip CONFIG REQUIRED)
-if(NOT TARGET nlohmann_json::nlohmann_json)
-  find_package(nlohmann_json CONFIG REQUIRED)
+
+# A CPU backend build without EXECUTORCH_BUILD_NATIVE only needs the graph and
+# the program reader, which have no third-party dependencies.
+if(EXECUTORCH_BUILD_NATIVE)
+  find_package(libzip CONFIG REQUIRED)
+  if(NOT TARGET nlohmann_json::nlohmann_json)
+    find_package(nlohmann_json CONFIG REQUIRED)
+  endif()
 endif()
 
 add_subdirectory(runtime)
@@ -45,6 +50,6 @@ if(EXECUTORCH_BUILD_NATIVE_VISION_RUNNER OR EXECUTORCH_BUILD_NATIVE_LLM_RUNNER)
   add_subdirectory(examples)
 endif()
 
-if(EXECUTORCH_BUILD_TESTS)
+if(EXECUTORCH_BUILD_NATIVE AND EXECUTORCH_BUILD_TESTS)
   add_subdirectory(test)
 endif()