Add EXECUTORCH_THREADPOOL_SIZE options, default to using only performance cores (#14090)

Summary:
Allow build-time configuration of the thread pool size and default to a performance heuristic.

There are 2 modes that we want to support:
 * Heuristic-based. Choose the number of threads according to a performance heuristic. Use threads equal to the number of detected performance cores, but we can continue to iterate on this by adding fine-grained heuristics for specific chipsets in the future.
 * All cores (threads=cores). This is the current behavior. We need to maintain this as an option for some use cases.

With this PR, the default (for OSS) is to use performance cores. From testing with CV models on ~10 representative devices across the performance spectrum, this gives anywhere from parity with the existing perf to up to a 13x speedup (measured on Pixel 6). Many common devices (S20, S22, iPhone 15 Pro) show a 2-4x speedup.

#### Specifying Threadpool Size

To specify the threadpool size, I've added two preprocessor options (and corresponding CMake options):
* `EXECUTORCH_THREADPOOL_USE_PERFORMANCE_CORES`- Use threads = detected perf cores.
* `EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES` - Use threads = logical cores.


Test Plan:
I've verified that logic functions correctly in OSS by building the executor_runner on M1 Mac and observing the existing logging in cpuinfo_utils.

Measuring MobileNet V3 (exported from examples) on XNNPACK, time to run 100 iterations drops from ~450ms to ~230ms on M1 Pro with this change.

Rollback Plan:

Reviewed By: kimishpatel

Differential Revision: D81965471

Pulled By: GregoryComer

Author: GregoryComer

extension/threadpool/CMakeLists.txt

@@ -20,6 +20,16 @@ if(NOT CMAKE_CXX_STANDARD)
   set(CMAKE_CXX_STANDARD 17)
 endif()
 
+# Threadpool size specifiers. Mutual exclusion is checking in default.cmake.
+# Default to using performance cores if
+# EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES isn't set.
+set(_threadpool_size_flag)
+if(EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES)
+  set(_threadpool_size_flag "EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES")
+else()
+  set(_threadpool_size_flag "EXECUTORCH_THREADPOOL_USE_PERFORMANCE_CORES")
+endif()
+
 add_library(
   extension_threadpool threadpool.cpp threadpool_guard.cpp thread_parallel.cpp
                        cpuinfo_utils.cpp
@@ -36,7 +46,9 @@ target_include_directories(
     $<BUILD_INTERFACE:${EXECUTORCH_ROOT}/backends/xnnpack/third-party/cpuinfo/include>
     $<BUILD_INTERFACE:${EXECUTORCH_ROOT}/backends/xnnpack/third-party/pthreadpool/include>
 )
-target_compile_definitions(extension_threadpool PUBLIC ET_USE_THREADPOOL)
+target_compile_definitions(
+  extension_threadpool PUBLIC ET_USE_THREADPOOL ${_threadpool_size_flag}
+)
 target_compile_options(extension_threadpool PUBLIC ${_common_compile_options})
 
 # Install libraries

extension/threadpool/targets.bzl

@@ -22,6 +22,7 @@ def define_common_targets():
         name = "threadpool_lib",
         srcs = _THREADPOOL_SRCS,
         deps = [
+            ":cpuinfo_utils",
             "//executorch/runtime/core:core",
             "//executorch/runtime/core/portable_type/c10/c10:c10",
         ],

extension/threadpool/test/threadpool_test.cpp

@@ -7,6 +7,7 @@
  */
 
 #include <executorch/extension/threadpool/threadpool.h>
+#include <executorch/runtime/platform/runtime.h>
 
 #include <mutex>
 #include <numeric>
@@ -71,6 +72,8 @@ void run_lambda_with_size(
 } // namespace
 
 TEST(ThreadPoolTest, ParallelAdd) {
+  executorch::runtime::runtime_init();
+
   std::vector<int32_t> a, b, c, c_ref;
   size_t vector_size = 100;
   size_t grain_size = 10;
@@ -111,6 +114,8 @@ TEST(ThreadPoolTest, ParallelAdd) {
 
 // Test parallel reduction where we acquire lock within lambda
 TEST(ThreadPoolTest, ParallelReduce) {
+  executorch::runtime::runtime_init();
+
   std::vector<int32_t> a;
   int32_t c = 0, c_ref = 0;
   size_t vector_size = 100;
@@ -144,6 +149,8 @@ TEST(ThreadPoolTest, ParallelReduce) {
 // Copied from
 // caffe2/aten/src/ATen/test/test_thread_pool_guard.cp
 TEST(TestNoThreadPoolGuard, TestThreadPoolGuard) {
+  executorch::runtime::runtime_init();
+
   auto threadpool_ptr = ::executorch::extension::threadpool::get_pthreadpool();
 
   ASSERT_NE(threadpool_ptr, nullptr);
@@ -173,6 +180,8 @@ TEST(TestNoThreadPoolGuard, TestThreadPoolGuard) {
 }
 
 TEST(TestNoThreadPoolGuard, TestRunWithGuard) {
+  executorch::runtime::runtime_init();
+
   const std::vector<int64_t> array = {1, 2, 3};
 
   auto pool = ::executorch::extension::threadpool::get_threadpool();

extension/threadpool/threadpool.cpp

@@ -6,6 +6,7 @@
  * LICENSE file in the root directory of this source tree.
  */
 
+#include <executorch/extension/threadpool/cpuinfo_utils.h>
 #include <executorch/extension/threadpool/threadpool.h>
 
 #include <algorithm>
@@ -14,9 +15,26 @@
 
 #include <executorch/extension/threadpool/threadpool_guard.h>
 #include <executorch/runtime/platform/assert.h>
+#include <executorch/runtime/platform/runtime.h>
 
 #include <cpuinfo.h>
 
+// At most one mode should be set.
+#if (                                                       \
+    defined(EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES) && \
+    defined(EXECUTORCH_THREADPOOL_USE_PERFORMANCE_CORES))
+#error Multiple \
+            threadpool size specifiers are set.At most one of                \
+    EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES,                             \
+    and EXECUTORCH_THREADPOOL_USE_PERFORMANCE_CORES may be defined.
+#endif
+
+// Default to EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES if no mode is set.
+#if !defined(EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES) && \
+    !defined(EXECUTORCH_THREADPOOL_USE_PERFORMANCE_CORES)
+#define EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES 1
+#endif
+
 namespace executorch::extension::threadpool {
 
 #if !(defined(WIN32))
@@ -96,12 +114,25 @@ void ThreadPool::run(
 // get_threadpool is not thread safe due to leak_corrupted_threadpool
 // Make this part threadsafe: TODO(kimishpatel)
 ThreadPool* get_threadpool() {
+  executorch::runtime::runtime_init();
+
   if (!cpuinfo_initialize()) {
     ET_LOG(Error, "cpuinfo initialization failed");
     return nullptr; // NOLINT(facebook-hte-NullableReturn)
   }
 
-  int num_threads = cpuinfo_get_processors_count();
+  // Choose the number of threads according to the EXECUTORCH_THREADPOOL_
+  // options. See the description in threadpool.h.
+
+#if defined(EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES)
+  // Use threads=cores.
+  static int num_threads = cpuinfo_get_processors_count();
+#else
+  // Set threads equal to the number of performance cores.
+  static int num_threads =
+      ::executorch::extension::cpuinfo::get_num_performant_cores();
+#endif
+
   /*
    * For llvm-tsan, holding limit for the number of locks for a single thread
    * is 63 (because of comparison < 64 instead of <=). pthreadpool's worst

extension/threadpool/threadpool.h

@@ -14,6 +14,22 @@
 
 #include <pthreadpool.h>
 
+/*
+ * Threadpool Options:
+ *
+ * Threadpool size has a sizble affect on performance. By default, the
+ * threadpool will be sized according to the number of performance cores. This
+ * behavior can be overriden with the following build-time options. Note that
+ * these options are mutually exclusive.
+ *
+ * - EXECUTORCH_THREADPOOL_USE_PERFORMANCE_CORES (flag) - Sizes the threadpool
+ * equal to the number of performance cores on the system. This is the default
+ * behavior.
+ * - EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES (flag) - Sizes the threadpool
+ * equal to the number of logical cores on system. This is the historical
+ * behavior.
+ */
+
 namespace executorch::extension::threadpool {
 
 class ThreadPool final {

tools/cmake/preset/default.cmake

@@ -176,6 +176,36 @@ define_overridable_option(
   ${_default_executorch_build_cpuinfo}
 )
 
+# Threadpool size options. At most one can be specified. Note that the default
+# is managed in threadpool.cpp to allow the user to specify an alternate mode
+# without needing to explicitly set the default to off.
+define_overridable_option(
+  EXECUTORCH_THREADPOOL_USE_PERFORMANCE_CORES
+  "Set the number of threads used for CPU parallel computation equal to the number of performant CPU cores."
+  BOOL
+  OFF
+)
+define_overridable_option(
+  EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES
+  "Set the number of threads used for CPU parallel computation equal to the number of logical CPU cores."
+  BOOL
+  OFF
+)
+
+check_required_options_on(
+  IF_ON EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES REQUIRES
+  EXECUTORCH_BUILD_PTHREADPOOL EXECUTORCH_BUILD_CPUINFO
+)
+check_required_options_on(
+  IF_ON EXECUTORCH_THREADPOOL_USE_PERFORMANCE_CORES REQUIRES
+  EXECUTORCH_BUILD_PTHREADPOOL EXECUTORCH_BUILD_CPUINFO
+)
+
+check_conflicting_options_on(
+  IF_ON EXECUTORCH_THREADPOOL_USE_PERFORMANCE_CORES CONFLICTS_WITH
+  EXECUTORCH_THREADPOOL_USE_ALL_LOGICAL_CORES
+)
+
 # TODO(jathu): move this to platform specific presets when created
 set(_default_executorch_build_executor_runner ON)
 if(APPLE AND "${SDK_NAME}" STREQUAL "iphoneos")