Ship the thread pool as its own library so a process has one pool

.ci/scripts/wheel/test_cpp_sdk.py

@@ -36,6 +36,10 @@
     "executorch::runtime::get_backend_class",
 )
 
+# The thread pool accessor. A second definer means a second pool, which
+# oversubscribes the CPU because each pool sizes itself to all cores.
+_THREADPOOL_SYMBOLS = ("executorch::extension::threadpool::get_threadpool",)
+
 # `nm -DC` prints "<hexaddr> <kind> <name>" for a definition and
 # "                 U <name>" for an undefined reference.
 _DEFINED = re.compile(r"^[0-9a-fA-F]+\s+(?P<kind>[A-Za-z])\s+(?P<name>.+)$")
@@ -106,23 +110,33 @@ def _defines_symbol(library: Path, symbol: str) -> bool:
     return False
 
 
-def test_single_backend_registry() -> None:
-    """Exactly one shipped library may define the backend registry."""
+def _assert_single_definer(symbols, what: str) -> None:
+    """Exactly one shipped library may define each of `symbols`."""
     assert shutil.which("nm") is not None, "nm is required to inspect the wheel"
 
     package_dir = _installed_package_dir()
     libraries = _shipped_shared_objects(package_dir)
     assert libraries, f"no shared libraries found under {package_dir}"
 
-    for symbol in _REGISTRY_SYMBOLS:
+    for symbol in symbols:
         definers = [lib for lib in libraries if _defines_symbol(lib, symbol)]
         pretty = [str(lib.relative_to(package_dir)) for lib in definers]
         assert len(definers) == 1, (
             f"expected exactly one library to define {symbol}, found "
             f"{len(definers)}: {pretty}. More than one definition means the "
-            f"process has more than one backend registry."
+            f"process has more than one {what}."
         )
-    print(f"✓ single backend registry across {len(libraries)} shipped libraries")
+    print(f"✓ single {what} across {len(libraries)} shipped libraries")
+
+
+def test_single_backend_registry() -> None:
+    """Exactly one shipped library may define the backend registry."""
+    _assert_single_definer(_REGISTRY_SYMBOLS, "backend registry")
+
+
+def test_single_threadpool() -> None:
+    """Exactly one shipped library may define the thread pool accessor."""
+    _assert_single_definer(_THREADPOOL_SYMBOLS, "thread pool")
 
 
 def test_cpp_consumer(work_dir: Path) -> None:
@@ -224,4 +238,5 @@ def _assert_runs_relocated(consumer, package_dir, work_dir, environment) -> None
 
 def run_tests(work_dir: Path) -> None:
     test_single_backend_registry()
+    test_single_threadpool()
     test_cpp_consumer(work_dir)

extension/threadpool/CMakeLists.txt

@@ -30,13 +30,46 @@ else()
   set(_threadpool_size_flag "EXECUTORCH_THREADPOOL_USE_PERFORMANCE_CORES")
 endif()
 
+# The thread pool is a process-wide singleton held in a function-local static,
+# so every library that links it statically gets its own copy. Build it shared
+# for the wheel, where several extensions are loaded into one interpreter, and
+# keep it static everywhere else so no other build changes.
+if(EXECUTORCH_BUILD_SHARED)
+  set(_threadpool_library_type SHARED)
+else()
+  set(_threadpool_library_type STATIC)
+endif()
+
 add_library(
-  extension_threadpool threadpool.cpp threadpool_guard.cpp thread_parallel.cpp
-                       cpuinfo_utils.cpp
-)
-target_link_libraries(
-  extension_threadpool PUBLIC executorch_core cpuinfo pthreadpool
+  extension_threadpool
+  ${_threadpool_library_type} threadpool.cpp threadpool_guard.cpp
+  thread_parallel.cpp cpuinfo_utils.cpp
 )
+if(EXECUTORCH_BUILD_SHARED)
+  set_target_properties(
+    extension_threadpool
+    PROPERTIES OUTPUT_NAME executorch_threadpool
+               VERSION "${PROJECT_VERSION}"
+               SOVERSION "${PROJECT_VERSION_MAJOR}"
+  )
+  if(NOT APPLE)
+    # Ships beside libexecutorch.so in the wheel's lib/ directory.
+    set_target_properties(
+      extension_threadpool PROPERTIES BUILD_RPATH "$ORIGIN" INSTALL_RPATH
+                                                            "$ORIGIN"
+    )
+  endif()
+  # cpuinfo and pthreadpool are forced static, so bundle them inside this
+  # library instead of making every consumer supply them.
+  executorch_target_whole_archive(extension_threadpool cpuinfo)
+  executorch_target_whole_archive(extension_threadpool pthreadpool)
+  target_link_libraries(extension_threadpool PRIVATE cpuinfo pthreadpool)
+  target_link_libraries(extension_threadpool PUBLIC executorch_shared)
+else()
+  target_link_libraries(
+    extension_threadpool PUBLIC executorch_core cpuinfo pthreadpool
+  )
+endif()
 target_include_directories(
   extension_threadpool PUBLIC ${_common_include_directories}
 )

setup.py

@@ -1114,6 +1114,20 @@ def run(self):  # noqa C901
                     dst=f"executorch/lib/libexecutorch.so.{get_runtime_soname_major()}",
                     dependent_cmake_flags=["EXECUTORCH_BUILD_SHARED"],
                 ),
+                # Install the shared thread pool next to it. It is a separate
+                # library so that a process has one pool rather than one per
+                # component that uses it.
+                BuiltFile(
+                    src_dir="%CMAKE_CACHE_DIR%/extension/threadpool/",
+                    src_name=(
+                        "libexecutorch_threadpool.so." f"{get_runtime_soname_major()}.*"
+                    ),
+                    dst=(
+                        "executorch/lib/libexecutorch_threadpool.so."
+                        f"{get_runtime_soname_major()}"
+                    ),
+                    dependent_cmake_flags=["EXECUTORCH_BUILD_SHARED"],
+                ),
                 # Install the prebuilt pybindings extension wrapper for the runtime,
                 # portable kernels, and a selection of backends. This lets users
                 # load and execute .pte files from python.