[SYCL][NATIVECPU][UR] performance improvments in NativeCPU adapter (intel#17102)

Initial set of performance improvements (less allocation and thread
launches) in Native CPU adapter

Author: uwedolinsky

unified-runtime/source/adapters/native_cpu/enqueue.cpp

@@ -217,19 +217,21 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
       }
       auto numGroups = groups.size();
       auto groupsPerThread = numGroups / numParallelThreads;
-      auto remainder = numGroups % numParallelThreads;
-      for (unsigned thread = 0; thread < numParallelThreads; thread++) {
-        futures.emplace_back(
-            tp.schedule_task([groups, thread, groupsPerThread,
-                              &kernel = *kernel](size_t threadId) {
-              for (unsigned i = 0; i < groupsPerThread; i++) {
-                auto index = thread * groupsPerThread + i;
-                groups[index](threadId, kernel);
-              }
-            }));
+      if (groupsPerThread) {
+        for (unsigned thread = 0; thread < numParallelThreads; thread++) {
+          futures.emplace_back(
+              tp.schedule_task([groups, thread, groupsPerThread,
+                                &kernel = *kernel](size_t threadId) {
+                for (unsigned i = 0; i < groupsPerThread; i++) {
+                  auto index = thread * groupsPerThread + i;
+                  groups[index](threadId, kernel);
+                }
+              }));
+        }
       }
 
       // schedule the remaining tasks
+      auto remainder = numGroups % numParallelThreads;
       if (remainder) {
         futures.emplace_back(
             tp.schedule_task([groups, remainder,
@@ -263,11 +265,12 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
   return UR_RESULT_SUCCESS;
 }
 
-ur_result_t withTimingEvent(ur_command_t command_type, ur_queue_handle_t hQueue,
-                            uint32_t numEventsInWaitList,
-                            const ur_event_handle_t *phEventWaitList,
-                            ur_event_handle_t *phEvent,
-                            const std::function<ur_result_t()> &f) {
+template <class T>
+static inline ur_result_t
+withTimingEvent(ur_command_t command_type, ur_queue_handle_t hQueue,
+                uint32_t numEventsInWaitList,
+                const ur_event_handle_t *phEventWaitList,
+                ur_event_handle_t *phEvent, T &&f) {
   urEventWait(numEventsInWaitList, phEventWaitList);
   ur_event_handle_t event = nullptr;
   if (phEvent) {

unified-runtime/source/adapters/native_cpu/threadpool.hpp

@@ -12,7 +12,6 @@
 #include <condition_variable>
 #include <cstdlib>
 #include <forward_list>
-#include <functional>
 #include <future>
 #include <iterator>
 #include <mutex>
@@ -24,7 +23,7 @@
 
 namespace native_cpu {
 
-using worker_task_t = std::function<void(size_t)>;
+using worker_task_t = std::packaged_task<void(size_t)>;
 
 namespace detail {
 
@@ -63,11 +62,11 @@ class worker_thread {
     m_isRunning.store(true, std::memory_order_release);
   }
 
-  inline void schedule(const worker_task_t &task) {
+  inline void schedule(worker_task_t &&task) {
     {
       std::lock_guard<std::mutex> lock(m_workMutex);
       // Add the task to the queue
-      m_tasks.push(task);
+      m_tasks.emplace(std::move(task));
       ++m_numTasks;
     }
     m_startWorkCondition.notify_one();
@@ -135,9 +134,9 @@ class simple_thread_pool {
     m_isRunning.store(false, std::memory_order_release);
   }
 
-  inline void schedule(const worker_task_t &task) {
+  inline void schedule(worker_task_t &&task) {
     // Schedule the task on the best available worker thread
-    this->best_worker().schedule(task);
+    this->best_worker().schedule(std::move(task));
   }
 
   inline bool is_running() const noexcept {
@@ -201,11 +200,11 @@ template <typename ThreadPoolT> class threadpool_interface {
 
   threadpool_interface() : threadpool() {}
 
-  auto schedule_task(worker_task_t &&task) {
-    auto workerTask = std::make_shared<std::packaged_task<void(size_t)>>(
-        [task](auto &&PH1) { return task(std::forward<decltype(PH1)>(PH1)); });
-    threadpool.schedule([=](size_t threadId) { (*workerTask)(threadId); });
-    return workerTask->get_future();
+  template <class T> std::future<void> schedule_task(T &&task) {
+    auto workerTask = std::packaged_task<void(size_t)>(std::forward<T>(task));
+    auto ret = workerTask.get_future();
+    threadpool.schedule(std::move(workerTask));
+    return ret;
   }
 };