Merge branch 'embb546_low_performance_on_arm' into development

Author: Marcus Winter

mtapi_c/include/embb/mtapi/c/mtapi.h

@@ -552,8 +552,9 @@ enum mtapi_node_attributes_enum {
                                             the node */
   MTAPI_NODE_MAX_ACTIONS_PER_JOB,      /**< maximum number of actions in a job
                                             allowed by the node */
-  MTAPI_NODE_MAX_PRIORITIES            /**< maximum number of priorities
+  MTAPI_NODE_MAX_PRIORITIES,           /**< maximum number of priorities
                                             allowed by the node */
+  MTAPI_NODE_REUSE_MAIN_THREAD         /**< reuse main thread as worker */
 };
 /** size of the \a MTAPI_NODE_CORE_AFFINITY attribute */
 #define MTAPI_NODE_CORE_AFFINITY_SIZE sizeof(embb_core_set_t)
@@ -577,6 +578,8 @@ enum mtapi_node_attributes_enum {
 #define MTAPI_NODE_MAX_ACTIONS_PER_JOB_SIZE sizeof(mtapi_uint_t)
 /** size of the \a MTAPI_NODE_MAX_PRIORITIES attribute */
 #define MTAPI_NODE_MAX_PRIORITIES_SIZE sizeof(mtapi_uint_t)
+/** size of the \a MTAPI_NODE_REUSE_MAIN_THREAD attribute */
+#define MTAPI_NODE_REUSE_MAIN_THREAD_SIZE sizeof(mtapi_boolean_t)
 
 /* example attribute value */
 #define MTAPI_NODE_TYPE_SMP 1
@@ -688,6 +691,8 @@ struct mtapi_node_attributes_struct {
   mtapi_uint_t max_actions_per_job;    /**< stores
                                             MTAPI_NODE_MAX_ACTIONS_PER_JOB */
   mtapi_uint_t max_priorities;         /**< stores MTAPI_NODE_MAX_PRIORITIES */
+  mtapi_boolean_t reuse_main_thread;   /**< stores
+                                            MTAPI_NODE_REUSE_MAIN_THREAD */
 };
 
 /**

mtapi_c/src/embb_mtapi_thread_context_t.c

@@ -54,6 +54,7 @@ mtapi_boolean_t embb_mtapi_thread_context_initialize_with_node_worker_and_core(
   that->core_num = core_num;
   that->priorities = node->attributes.max_priorities;
   that->is_initialized = MTAPI_FALSE;
+  that->is_main_thread = (worker_index == 0) ? node->attributes.reuse_main_thread : MTAPI_FALSE;
   embb_atomic_store_int(&that->run, 0);
 
   that->queue = (embb_mtapi_task_queue_t**)embb_mtapi_alloc_allocate(
@@ -121,17 +122,29 @@ mtapi_boolean_t embb_mtapi_thread_context_start(
   embb_core_set_add(&core_set, that->core_num);
 
   /* create thread */
-  err = embb_thread_create(&that->thread, &core_set, worker_func, that);
-  if (EMBB_SUCCESS != err) {
-    embb_mtapi_log_error(
-      "embb_mtapi_ThreadContext_initializeWithNodeAndCoreNumber() could not "
-      "create thread %d on core %d\n", that->worker_index, that->core_num);
-    return MTAPI_FALSE;
-  }
-
-  /* wait for worker to come up */
-  while (0 == embb_atomic_load_int(&that->run)) {
-    embb_thread_yield();
+  if (that->is_main_thread) {
+    /* reuse main thread */
+    that->thread = embb_thread_current();
+    err = embb_tss_create(&that->tss_id);
+    if (EMBB_SUCCESS != err) {
+      /* report error to scheduler */
+      embb_atomic_store_int(&that->run, -1);
+      return MTAPI_FALSE;
+    }
+    embb_tss_set(&(that->tss_id), that);
+    embb_atomic_store_int(&that->run, 1);
+  } else {
+    err = embb_thread_create(&that->thread, &core_set, worker_func, that);
+    if (EMBB_SUCCESS != err) {
+      embb_mtapi_log_error(
+        "embb_mtapi_ThreadContext_initializeWithNodeAndCoreNumber() could not "
+        "create thread %d on core %d\n", that->worker_index, that->core_num);
+      return MTAPI_FALSE;
+    }
+    /* wait for worker to come up */
+    while (0 == embb_atomic_load_int(&that->run)) {
+      embb_thread_yield();
+    }
   }
 
   if (0 < embb_atomic_load_int(&that->run)) {
@@ -146,7 +159,9 @@ void embb_mtapi_thread_context_stop(embb_mtapi_thread_context_t* that) {
   if (0 < embb_atomic_load_int(&that->run)) {
     embb_atomic_store_int(&that->run, 0);
     embb_condition_notify_one(&that->work_available);
-    embb_thread_join(&(that->thread), &result);
+    if (MTAPI_FALSE == that->is_main_thread) {
+      embb_thread_join(&(that->thread), &result);
+    }
   }
 }
 
@@ -158,6 +173,9 @@ void embb_mtapi_thread_context_finalize(embb_mtapi_thread_context_t* that) {
   embb_mtapi_log_trace("embb_mtapi_thread_context_finalize() called\n");
 
   if (that->is_initialized) {
+    if (that->is_main_thread) {
+      embb_tss_delete(&that->tss_id);
+    }
     embb_condition_destroy(&that->work_available);
     embb_mutex_destroy(&that->work_available_mutex);
   }

mtapi_c/src/embb_mtapi_thread_context_t.h

@@ -68,6 +68,7 @@ struct embb_mtapi_thread_context_struct {
   embb_atomic_int run;
   mtapi_status_t status;
   mtapi_boolean_t is_initialized;
+  mtapi_boolean_t is_main_thread;
 };
 
 #include <embb_mtapi_thread_context_t_fwd.h>

mtapi_c/src/mtapi_node_attributes_t.c

@@ -52,6 +52,7 @@ void mtapi_nodeattr_init(
     attributes->max_jobs = MTAPI_NODE_MAX_JOBS_DEFAULT;
     attributes->max_actions_per_job = MTAPI_NODE_MAX_ACTIONS_PER_JOB_DEFAULT;
     attributes->max_priorities = MTAPI_NODE_MAX_PRIORITIES_DEFAULT;
+    attributes->reuse_main_thread = MTAPI_FALSE;
 
     embb_core_set_init(&attributes->core_affinity, 1);
     attributes->num_cores = embb_core_set_count(&attributes->core_affinity);
@@ -143,6 +144,11 @@ void mtapi_nodeattr_set(
           &attributes->max_priorities, attribute, attribute_size);
         break;
 
+      case MTAPI_NODE_REUSE_MAIN_THREAD:
+        local_status = embb_mtapi_attr_set_mtapi_boolean_t(
+          &attributes->reuse_main_thread, attribute, attribute_size);
+        break;
+
       default:
         /* attribute unknown */
         local_status = MTAPI_ERR_ATTR_NUM;