[OMPT] Update types according to TR7

Some types and callback signatures have changed from TR6 to TR7.
Major changes (only adding signatures and stubs):
(-remove idle callback) done by D48362
-add reduction and dispatch callback
-add get_task_memory and finalize_tool runtime entry points
-ompt_invoker_t  becomes ompt_parallel_flag_t
-more types of sync_regions

Patch provided by Simon Convent

Reviewers: hbae, protze.joachim

Differential Revision: https://reviews.llvm.org/D50774

16 files changed, 198 insertions, 129 deletions

diff --git a/openmp/runtime/src/include/50/ompt.h.var b/openmp/runtime/src/include/50/ompt.h.var
index 56c5834a113..24fc90b8c62 100644
--- a/openmp/runtime/src/include/50/ompt.h.var
+++ b/openmp/runtime/src/include/50/ompt.h.var
@@ -38,8 +38,10 @@
                                             \
     macro (ompt_get_parallel_info)          \
     macro (ompt_get_task_info)              \
+    macro (ompt_get_task_memory)            \
     macro (ompt_get_thread_data)            \
     macro (ompt_get_unique_id)              \
+    macro (ompt_finalize_tool)              \
                                             \
     macro(ompt_get_num_procs)               \
     macro(ompt_get_num_places)              \
@@ -151,6 +153,10 @@
     macro (ompt_callback_flush,             ompt_callback_flush_t,             29) /* after executing flush           */ \
                                                                                                                          \
     macro (ompt_callback_cancel,            ompt_callback_cancel_t,            30) /* cancel innermost binding region */ \
+                                                                                                                         \
+    macro (ompt_callback_reduction,         ompt_callback_sync_region_t,       31) /* reduction                       */ \
+                                                                                                                         \
+    macro (ompt_callback_dispatch,          ompt_callback_dispatch_t,          32) /* dispatch of work                */
 
 
 
@@ -191,16 +197,17 @@ typedef struct omp_frame_t {
  * dependences types
  *---------------------*/
 
-typedef enum ompt_task_dependence_flag_t {
+typedef enum ompt_task_dependence_type_t {
     // a two bit field for the dependence type
-    ompt_task_dependence_type_out   = 1,
-    ompt_task_dependence_type_in    = 2,
-    ompt_task_dependence_type_inout = 3,
-} ompt_task_dependence_flag_t;
+    ompt_task_dependence_type_in            = 1,
+    ompt_task_dependence_type_out           = 2,
+    ompt_task_dependence_type_inout         = 3,
+    ompt_task_dependence_type_mutexinoutset = 4
+} ompt_task_dependence_type_t;
 
 typedef struct ompt_task_dependence_t {
     void *variable_addr;
-    unsigned int dependence_flags;
+    ompt_task_dependence_type_t dependence_type;
 } ompt_task_dependence_t;
 
 
@@ -264,20 +271,22 @@ typedef ompt_interface_fn_t (*ompt_function_lookup_t)(
 );
 
 /* threads */
-typedef enum ompt_thread_type_t {
+typedef enum ompt_thread_t {
     ompt_thread_initial = 1, // start the enumeration at 1
     ompt_thread_worker  = 2,
     ompt_thread_other   = 3,
     ompt_thread_unknown = 4
-} ompt_thread_type_t;
+} ompt_thread_t;
 
-typedef enum ompt_invoker_t {
-    ompt_invoker_program = 1,             /* program invokes master task         */
-    ompt_invoker_runtime = 2              /* runtime invokes master task         */
-} ompt_invoker_t;
+typedef enum ompt_parallel_flag_t {
+    ompt_parallel_invoker_program = 0x00000001,   /* program invokes master task */
+    ompt_parallel_invoker_runtime = 0x00000002,   /* runtime invokes master task */
+    ompt_parallel_league = 0x40000000,
+    ompt_parallel_team = 0x80000000
+} ompt_parallel_flag_t;
 
 typedef void (*ompt_callback_thread_begin_t) (
-    ompt_thread_type_t thread_type,       /* type of thread                      */
+    ompt_thread_t thread_type,            /* type of thread                      */
     ompt_data_t *thread_data              /* data of thread                      */
 );
 
@@ -301,8 +310,8 @@ typedef void (*ompt_callback_implicit_task_t) (
     ompt_scope_endpoint_t endpoint,       /* endpoint of implicit task           */
     ompt_data_t *parallel_data,           /* data of parallel region             */
     ompt_data_t *task_data,               /* data of implicit task               */
-    unsigned int team_size,               /* team size                           */
-    unsigned int thread_num               /* thread number of calling thread     */
+    unsigned int actual_parallelism,      /* team size                           */
+    unsigned int index                    /* thread number of calling thread     */
 );
 
 typedef void (*ompt_callback_parallel_begin_t) (
@@ -310,19 +319,19 @@ typedef void (*ompt_callback_parallel_begin_t) (
     const omp_frame_t *encountering_task_frame,  /* frame data of encountering task     */
     ompt_data_t *parallel_data,                  /* data of parallel region             */
     unsigned int requested_team_size,            /* requested number of threads in team */
-    ompt_invoker_t invoker,                      /* invoker of master task              */
+    int flag,                                    /* flag for additional information     */
     const void *codeptr_ra                       /* return address of runtime call      */
 );
 
 typedef void (*ompt_callback_parallel_end_t) (
     ompt_data_t *parallel_data,           /* data of parallel region             */
     ompt_data_t *encountering_task_data,  /* data of encountering task           */
-    ompt_invoker_t invoker,               /* invoker of master task              */ 
+    int flag,                             /* flag for additional information     */
     const void *codeptr_ra                /* return address of runtime call      */
 );
 
 /* tasks */
-typedef enum ompt_task_type_t {
+typedef enum ompt_task_flag_t {
     ompt_task_initial    = 0x1,
     ompt_task_implicit   = 0x2,
     ompt_task_explicit   = 0x4,
@@ -332,13 +341,13 @@ typedef enum ompt_task_type_t {
     ompt_task_final      = 0x20000000,
     ompt_task_mergeable  = 0x40000000,
     ompt_task_merged     = 0x80000000
-} ompt_task_type_t;
+} ompt_task_flag_t;
 
 typedef enum ompt_task_status_t {
     ompt_task_complete = 1,
     ompt_task_yield    = 2,
     ompt_task_cancel   = 3,
-    ompt_task_others   = 4
+    ompt_task_switch   = 4
 } ompt_task_status_t;
 
 typedef void (*ompt_callback_task_schedule_t) (
@@ -351,7 +360,7 @@ typedef void (*ompt_callback_task_create_t) (
     ompt_data_t *encountering_task_data,         /* data of parent task                 */
     const omp_frame_t *encountering_task_frame,  /* frame data for parent task          */
     ompt_data_t *new_task_data,                  /* data of created task                */
-    int type,                                    /* type of created task                */
+    int flag,                                    /* type of created task                */
     int has_dependences,                         /* created task has dependences        */
     const void *codeptr_ra                       /* return address of runtime call      */
 );
@@ -369,15 +378,15 @@ typedef void (*ompt_callback_task_dependence_t) (
 );
 
 /* target and device */
-typedef enum ompt_target_type_t {
+typedef enum ompt_target_t {
     ompt_target = 1,
     ompt_target_enter_data = 2,
     ompt_target_exit_data = 3,
     ompt_target_update = 4
-} ompt_target_type_t;
+} ompt_target_t;
 
 typedef void (*ompt_callback_target_t) (
-    ompt_target_type_t kind,
+    ompt_target_t kind,
     ompt_scope_endpoint_t endpoint,
     uint64_t device_num,
     ompt_data_t *task_data,
@@ -396,14 +405,18 @@ typedef void (*ompt_callback_target_data_op_t) (
     ompt_id_t target_id,
     ompt_id_t host_op_id,
     ompt_target_data_op_t optype,
-    void *host_addr,
-    void *device_addr,
-    size_t bytes
+    void *src_addr,
+    int src_device_num,
+    void *dest_addr,
+    int dest_device_num,
+    size_t bytes,
+    const void *codeptr_ra
 );
 
 typedef void (*ompt_callback_target_submit_t) (
     ompt_id_t target_id,
-    ompt_id_t host_op_id
+    ompt_id_t host_op_id,
+    unsigned int requested_num_teams
 );
 
 typedef void (*ompt_callback_target_map_t) (
@@ -412,7 +425,8 @@ typedef void (*ompt_callback_target_map_t) (
     void **host_addr,
     void **device_addr,
     size_t *bytes,
-    unsigned int *mapping_flags
+    unsigned int *mapping_flags,
+    const void *codeptr_ra
 );
 
 typedef void (*ompt_callback_device_initialize_t) (
@@ -453,32 +467,31 @@ typedef int (*ompt_callback_control_tool_t) (
     const void *codeptr_ra                /* return address of runtime call      */
 );
 
-typedef enum ompt_mutex_kind_t {
-    ompt_mutex           = 0x10,
-    ompt_mutex_lock      = 0x11,
-    ompt_mutex_nest_lock = 0x12,
-    ompt_mutex_critical  = 0x13,
-    ompt_mutex_atomic    = 0x14,
-    ompt_mutex_ordered   = 0x20
-} ompt_mutex_kind_t;
+typedef enum ompt_mutex_t {
+    ompt_mutex_lock      = 1,
+    ompt_mutex_nest_lock = 2,
+    ompt_mutex_critical  = 3,
+    ompt_mutex_atomic    = 4,
+    ompt_mutex_ordered   = 5
+} ompt_mutex_t;
 
 typedef void (*ompt_callback_mutex_acquire_t) (
-    ompt_mutex_kind_t kind,               /* mutex kind                          */
+    ompt_mutex_t kind,                    /* mutex kind                          */
     unsigned int hint,                    /* mutex hint                          */
     unsigned int impl,                    /* mutex implementation                */
-    omp_wait_id_t wait_id,               /* id of object being awaited          */
+    omp_wait_id_t wait_id,                /* id of object being awaited          */
     const void *codeptr_ra                /* return address of runtime call      */
 );
 
 typedef void (*ompt_callback_mutex_t) (
-    ompt_mutex_kind_t kind,               /* mutex kind                          */
-    omp_wait_id_t wait_id,               /* id of object being awaited          */
+    ompt_mutex_t kind,                    /* mutex kind                          */
+    omp_wait_id_t wait_id,                /* id of object being awaited          */
     const void *codeptr_ra                /* return address of runtime call      */
 );
 
 typedef void (*ompt_callback_nest_lock_t) (
     ompt_scope_endpoint_t endpoint,       /* endpoint of nested lock             */
-    omp_wait_id_t wait_id,               /* id of object being awaited          */
+    omp_wait_id_t wait_id,                /* id of object being awaited          */
     const void *codeptr_ra                /* return address of runtime call      */
 );
 
@@ -489,7 +502,7 @@ typedef void (*ompt_callback_master_t) (
     const void *codeptr_ra                /* return address of runtime call      */
 );
 
-typedef enum ompt_work_type_t {
+typedef enum ompt_work_t {
     ompt_work_loop            = 1,
     ompt_work_sections        = 2,
     ompt_work_single_executor = 3,
@@ -497,10 +510,10 @@ typedef enum ompt_work_type_t {
     ompt_work_workshare       = 5,
     ompt_work_distribute      = 6,
     ompt_work_taskloop        = 7
-} ompt_work_type_t;
+} ompt_work_t;
 
 typedef void (*ompt_callback_work_t) (
-    ompt_work_type_t wstype,              /* type of work region                 */
+    ompt_work_t wstype,              /* type of work region                 */
     ompt_scope_endpoint_t endpoint,       /* endpoint of work region             */
     ompt_data_t *parallel_data,           /* data of parallel region             */
     ompt_data_t *task_data,               /* data of task                        */
@@ -508,14 +521,18 @@ typedef void (*ompt_callback_work_t) (
     const void *codeptr_ra                /* return address of runtime call      */
 );
 
-typedef enum ompt_sync_region_kind_t {
-    ompt_sync_region_barrier   = 1,
-    ompt_sync_region_taskwait  = 2,
-    ompt_sync_region_taskgroup = 3
-} ompt_sync_region_kind_t;
+typedef enum ompt_sync_region_t {
+    ompt_sync_region_barrier                  = 1,
+    ompt_sync_region_barrier_implicit         = 2,
+    ompt_sync_region_barrier_explicit         = 3,
+    ompt_sync_region_barrier_implementation   = 4,
+    ompt_sync_region_taskwait                 = 5,
+    ompt_sync_region_taskgroup                = 6,
+    ompt_sync_region_reduction                = 7
+} ompt_sync_region_t;
 
 typedef void (*ompt_callback_sync_region_t) (
-    ompt_sync_region_kind_t kind,         /* kind of sync region                 */
+    ompt_sync_region_t kind,              /* kind of sync region                 */
     ompt_scope_endpoint_t endpoint,       /* endpoint of sync region             */
     ompt_data_t *parallel_data,           /* data of parallel region             */
     ompt_data_t *task_data,               /* data of task                        */
@@ -523,10 +540,10 @@ typedef void (*ompt_callback_sync_region_t) (
 );
 
 typedef enum ompt_cancel_flag_t {
-    ompt_cancel_parallel       = 0x1,
-    ompt_cancel_sections       = 0x2,
-    ompt_cancel_do             = 0x4,
-    ompt_cancel_taskgroup      = 0x8,
+    ompt_cancel_parallel       = 0x01,
+    ompt_cancel_sections       = 0x02,
+    ompt_cancel_loop           = 0x04,
+    ompt_cancel_taskgroup      = 0x08,
     ompt_cancel_activated      = 0x10,
     ompt_cancel_detected       = 0x20,
     ompt_cancel_discarded_task = 0x40
@@ -543,6 +560,18 @@ typedef void (*ompt_callback_flush_t) (
     const void *codeptr_ra                /* return address of runtime call      */
 );
 
+typedef enum ompt_dispatch_t {
+    ompt_dispatch_iteration = 1,
+    ompt_dispatch_section =   2
+} ompt_dispatch_t;
+
+typedef void (*ompt_callback_dispatch_t) (
+    ompt_data_t *parallel_data,
+    ompt_data_t *task_data,
+    ompt_dispatch_t kind,
+    ompt_data_t instance
+);
+
 /****************************************************************************
  * ompt API
  ***************************************************************************/
@@ -587,6 +616,12 @@ OMPT_API_FUNCTION(int, ompt_get_task_info, (
     int *thread_num
 ));
 
+OMPT_API_FUNCTION(int, ompt_get_task_memory, (
+    void **addr,
+    size_t *size,
+    int block
+));
+
 /* procs */
 OMPT_API_FUNCTION(int, ompt_get_num_procs, (void));
 
@@ -615,7 +650,7 @@ OMPT_API_FUNCTION(int, ompt_get_proc_id, (void));
  ***************************************************************************/
 
 OMPT_API_FUNCTION(int, ompt_initialize, (
-    ompt_function_lookup_t ompt_fn_lookup,
+    ompt_function_lookup_t lookup,
     ompt_data_t *tool_data
 ));
 
@@ -673,6 +708,9 @@ OMPT_API_FUNCTION(int, ompt_enumerate_mutex_impls, (
 /* get_unique_id */
 OMPT_API_FUNCTION(uint64_t, ompt_get_unique_id, (void));
 
+/* finalize tool */
+OMPT_API_FUNCTION(void, ompt_finalize_tool, (void)); 
+
 #ifdef  __cplusplus
 };
 #endif
diff --git a/openmp/runtime/src/kmp_barrier.cpp b/openmp/runtime/src/kmp_barrier.cpp
index 3f53d6b6730..e3925cb304d 100644
--- a/openmp/runtime/src/kmp_barrier.cpp
+++ b/openmp/runtime/src/kmp_barrier.cpp
@@ -1223,7 +1223,7 @@ int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
   kmp_info_t *this_thr = __kmp_threads[gtid];
   kmp_team_t *team = this_thr->th.th_team;
   int status = 0;
-#if OMPT_SUPPORT
+#if OMPT_SUPPORT && OMPT_OPTIONAL
   ompt_data_t *my_task_data;
   ompt_data_t *my_parallel_data;
   void *return_address;
@@ -1599,10 +1599,10 @@ void __kmp_join_barrier(int gtid) {
 
   ANNOTATE_BARRIER_BEGIN(&team->t.t_bar);
 #if OMPT_SUPPORT
-  ompt_data_t *my_task_data;
-  ompt_data_t *my_parallel_data;
   if (ompt_enabled.enabled) {
 #if OMPT_OPTIONAL
+    ompt_data_t *my_task_data;
+    ompt_data_t *my_parallel_data;
     void *codeptr = NULL;
     int ds_tid = this_thr->th.th_info.ds.ds_tid;
     if (KMP_MASTER_TID(ds_tid) &&
diff --git a/openmp/runtime/src/kmp_cancel.cpp b/openmp/runtime/src/kmp_cancel.cpp
index 3895f97833e..5ed91673b77 100644
--- a/openmp/runtime/src/kmp_cancel.cpp
+++ b/openmp/runtime/src/kmp_cancel.cpp
@@ -65,7 +65,7 @@ kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 cncl_kind) {
             if (cncl_kind == cancel_parallel)
               type = ompt_cancel_parallel;
             else if (cncl_kind == cancel_loop)
-              type = ompt_cancel_do;
+              type = ompt_cancel_loop;
             else if (cncl_kind == cancel_sections)
               type = ompt_cancel_sections;
             ompt_callbacks.ompt_callback(ompt_callback_cancel)(
@@ -172,7 +172,7 @@ kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
               if (cncl_kind == cancel_parallel)
                 type = ompt_cancel_parallel;
               else if (cncl_kind == cancel_loop)
-                type = ompt_cancel_do;
+                type = ompt_cancel_loop;
               else if (cncl_kind == cancel_sections)
                 type = ompt_cancel_sections;
               ompt_callbacks.ompt_callback(ompt_callback_cancel)(
diff --git a/openmp/runtime/src/kmp_csupport.cpp b/openmp/runtime/src/kmp_csupport.cpp
index 8c8b8eb26f8..51934cbdffc 100644
--- a/openmp/runtime/src/kmp_csupport.cpp
+++ b/openmp/runtime/src/kmp_csupport.cpp
@@ -513,7 +513,7 @@ void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
     if (ompt_enabled.ompt_callback_parallel_end) {
       ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
           &(serial_team->t.ompt_team_info.parallel_data), parent_task_data,
-          ompt_invoker_program, OMPT_LOAD_RETURN_ADDRESS(global_tid));
+          ompt_parallel_invoker_program, OMPT_LOAD_RETURN_ADDRESS(global_tid));
     }
     __ompt_lw_taskteam_unlink(this_thr);
     this_thr->th.ompt_thread_info.state = omp_state_overhead;
@@ -1782,7 +1782,7 @@ void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid) {
 
 #if OMPT_SUPPORT && OMPT_OPTIONAL
   if (ompt_enabled.ompt_callback_work) {
-    ompt_work_type_t ompt_work_type = ompt_work_loop;
+    ompt_work_t ompt_work_type = ompt_work_loop;
     ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
     ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
     // Determine workshare type
diff --git a/openmp/runtime/src/kmp_runtime.cpp b/openmp/runtime/src/kmp_runtime.cpp
index 9078e3a4946..e0c15cdf5ec 100644
--- a/openmp/runtime/src/kmp_runtime.cpp
+++ b/openmp/runtime/src/kmp_runtime.cpp
@@ -1225,7 +1225,8 @@ void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
 
       ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
           &(parent_task_info->task_data), &(parent_task_info->frame),
-          &ompt_parallel_data, team_size, ompt_invoker_program, codeptr);
+          &ompt_parallel_data, team_size, ompt_parallel_invoker_program,
+          codeptr);
     }
   }
 #endif // OMPT_SUPPORT
@@ -3824,7 +3825,7 @@ int __kmp_register_root(int initial_thread) {
 #if OMPT_SUPPORT
   if (!initial_thread && ompt_enabled.enabled) {
 
-    ompt_thread_t *root_thread = ompt_get_thread();
+    kmp_info_t *root_thread = ompt_get_thread();
 
     ompt_set_thread_state(root_thread, omp_state_overhead);
 
diff --git a/openmp/runtime/src/kmp_sched.cpp b/openmp/runtime/src/kmp_sched.cpp
index 67fe02f7944..fc8000e65fd 100644
--- a/openmp/runtime/src/kmp_sched.cpp
+++ b/openmp/runtime/src/kmp_sched.cpp
@@ -68,7 +68,7 @@ static void __kmp_for_static_init(ident_t *loc, kmp_int32 global_tid,
 #if OMPT_SUPPORT && OMPT_OPTIONAL
   ompt_team_info_t *team_info = NULL;
   ompt_task_info_t *task_info = NULL;
-  ompt_work_type_t ompt_work_type = ompt_work_loop;
+  ompt_work_t ompt_work_type = ompt_work_loop;
 
   static kmp_int8 warn = 0;
 
diff --git a/openmp/runtime/src/kmp_taskdeps.cpp b/openmp/runtime/src/kmp_taskdeps.cpp
index 4efd12bac8a..34ea3793d1d 100644
--- a/openmp/runtime/src/kmp_taskdeps.cpp
+++ b/openmp/runtime/src/kmp_taskdeps.cpp
@@ -508,26 +508,26 @@ kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
       new_taskdata->ompt_task_info.deps[i].variable_addr =
           (void *)dep_list[i].base_addr;
       if (dep_list[i].flags.in && dep_list[i].flags.out)
-        new_taskdata->ompt_task_info.deps[i].dependence_flags =
+        new_taskdata->ompt_task_info.deps[i].dependence_type =
             ompt_task_dependence_type_inout;
       else if (dep_list[i].flags.out)
-        new_taskdata->ompt_task_info.deps[i].dependence_flags =
+        new_taskdata->ompt_task_info.deps[i].dependence_type =
             ompt_task_dependence_type_out;
       else if (dep_list[i].flags.in)
-        new_taskdata->ompt_task_info.deps[i].dependence_flags =
+        new_taskdata->ompt_task_info.deps[i].dependence_type =
             ompt_task_dependence_type_in;
     }
     for (i = 0; i < ndeps_noalias; i++) {
       new_taskdata->ompt_task_info.deps[ndeps + i].variable_addr =
           (void *)noalias_dep_list[i].base_addr;
       if (noalias_dep_list[i].flags.in && noalias_dep_list[i].flags.out)
-        new_taskdata->ompt_task_info.deps[ndeps + i].dependence_flags =
+        new_taskdata->ompt_task_info.deps[ndeps + i].dependence_type =
             ompt_task_dependence_type_inout;
       else if (noalias_dep_list[i].flags.out)
-        new_taskdata->ompt_task_info.deps[ndeps + i].dependence_flags =
+        new_taskdata->ompt_task_info.deps[ndeps + i].dependence_type =
             ompt_task_dependence_type_out;
       else if (noalias_dep_list[i].flags.in)
-        new_taskdata->ompt_task_info.deps[ndeps + i].dependence_flags =
+        new_taskdata->ompt_task_info.deps[ndeps + i].dependence_type =
             ompt_task_dependence_type_in;
     }
     ompt_callbacks.ompt_callback(ompt_callback_task_dependences)(
diff --git a/openmp/runtime/src/kmp_tasking.cpp b/openmp/runtime/src/kmp_tasking.cpp
index e1b03f780b5..9028686bea1 100644
--- a/openmp/runtime/src/kmp_tasking.cpp
+++ b/openmp/runtime/src/kmp_tasking.cpp
@@ -473,7 +473,7 @@ static inline void __ompt_task_start(kmp_task_t *task,
                                      kmp_taskdata_t *current_task,
                                      kmp_int32 gtid) {
   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
-  ompt_task_status_t status = ompt_task_others;
+  ompt_task_status_t status = ompt_task_switch;
   if (__kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded) {
     status = ompt_task_yield;
     __kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded = 0;
@@ -1586,7 +1586,7 @@ kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
       // Scheduling back to the parent task.
       __ompt_task_finish(new_task,
                          new_taskdata->ompt_task_info.scheduling_parent,
-                         ompt_task_others);
+                         ompt_task_switch);
       new_taskdata->ompt_task_info.frame.exit_frame = NULL;
     }
   }
@@ -1766,7 +1766,7 @@ static kmp_int32 __kmpc_omp_taskwait_template(ident_t *loc_ref, kmp_int32 gtid,
   return TASK_CURRENT_NOT_QUEUED;
 }
 
-#if OMPT_SUPPORT
+#if OMPT_SUPPORT && OMPT_OPTIONAL
 OMPT_NOINLINE
 static kmp_int32 __kmpc_omp_taskwait_ompt(ident_t *loc_ref, kmp_int32 gtid,
                                           void *frame_address,
@@ -1774,7 +1774,7 @@ static kmp_int32 __kmpc_omp_taskwait_ompt(ident_t *loc_ref, kmp_int32 gtid,
   return __kmpc_omp_taskwait_template<true>(loc_ref, gtid, frame_address,
                                             return_address);
 }
-#endif // OMPT_SUPPORT
+#endif // OMPT_SUPPORT && OMPT_OPTIONAL
 
 // __kmpc_omp_taskwait: Wait until all tasks generated by the current task are
 // complete
diff --git a/openmp/runtime/src/ompt-event-specific.h b/openmp/runtime/src/ompt-event-specific.h
index 305b1764f7b..1fb08444749 100644
--- a/openmp/runtime/src/ompt-event-specific.h
+++ b/openmp/runtime/src/ompt-event-specific.h
@@ -105,4 +105,8 @@
 
 #define ompt_callback_cancel_implemented ompt_event_MAY_ALWAYS_OPTIONAL
 
+#define ompt_callback_reduction_implemented ompt_event_UNIMPLEMENTED
+
+#define ompt_callback_dispatch_implemented ompt_event_UNIMPLEMENTED
+
 #endif
diff --git a/openmp/runtime/src/ompt-general.cpp b/openmp/runtime/src/ompt-general.cpp
index 25ad19d03af..964f6cdd1d8 100644
--- a/openmp/runtime/src/ompt-general.cpp
+++ b/openmp/runtime/src/ompt-general.cpp
@@ -351,7 +351,7 @@ void ompt_post_init() {
       return;
     }
 
-    ompt_thread_t *root_thread = ompt_get_thread();
+    kmp_info_t *root_thread = ompt_get_thread();
 
     ompt_set_thread_state(root_thread, omp_state_overhead);
 
@@ -509,6 +509,12 @@ OMPT_API_ROUTINE int ompt_get_task_info(int ancestor_level, int *type,
                                        task_frame, parallel_data, thread_num);
 }
 
+OMPT_API_ROUTINE int ompt_get_task_memory(void **addr, size_t *size,
+                                          int block) {
+  // stub
+  return 0;
+}
+
 /*****************************************************************************
  * num_procs
  ****************************************************************************/
@@ -678,6 +684,10 @@ OMPT_API_ROUTINE uint64_t ompt_get_unique_id(void) {
   return __ompt_get_unique_id_internal();
 }
 
+OMPT_API_ROUTINE void ompt_finalize_tool(void) {
+  // stub
+}
+
 /*****************************************************************************
  * Target
  ****************************************************************************/
diff --git a/openmp/runtime/src/ompt-internal.h b/openmp/runtime/src/ompt-internal.h
index a8f3bac49a4..6b92eaa5e49 100644
--- a/openmp/runtime/src/ompt-internal.h
+++ b/openmp/runtime/src/ompt-internal.h
@@ -22,7 +22,8 @@
 #define _OMP_EXTERN extern "C"
 
 #define OMPT_INVOKER(x)                                                        \
-  ((x == fork_context_gnu) ? ompt_invoker_program : ompt_invoker_runtime)
+  ((x == fork_context_gnu) ? ompt_parallel_invoker_program                     \
+                           : ompt_parallel_invoker_runtime)
 
 #define ompt_callback(e) e##_callback
 
diff --git a/openmp/runtime/src/ompt-specific.h b/openmp/runtime/src/ompt-specific.h
index 28905bb5dc4..8cf74504a2d 100644
--- a/openmp/runtime/src/ompt-specific.h
+++ b/openmp/runtime/src/ompt-specific.h
@@ -17,25 +17,19 @@
 #include "kmp.h"
 
 /*****************************************************************************
- * types
- ****************************************************************************/
-
-typedef kmp_info_t ompt_thread_t;
-
-/*****************************************************************************
  * forward declarations
  ****************************************************************************/
 
 void __ompt_team_assign_id(kmp_team_t *team, ompt_data_t ompt_pid);
 void __ompt_thread_assign_wait_id(void *variable);
 
-void __ompt_lw_taskteam_init(ompt_lw_taskteam_t *lwt, ompt_thread_t *thr,
+void __ompt_lw_taskteam_init(ompt_lw_taskteam_t *lwt, kmp_info_t *thr,
                              int gtid, ompt_data_t *ompt_pid, void *codeptr);
 
-void __ompt_lw_taskteam_link(ompt_lw_taskteam_t *lwt, ompt_thread_t *thr,
+void __ompt_lw_taskteam_link(ompt_lw_taskteam_t *lwt, kmp_info_t *thr,
                              int on_heap);
 
-void __ompt_lw_taskteam_unlink(ompt_thread_t *thr);
+void __ompt_lw_taskteam_unlink(kmp_info_t *thr);
 
 ompt_team_info_t *__ompt_get_teaminfo(int depth, int *size);
 
@@ -90,16 +84,16 @@ inline void *__ompt_load_return_address(int gtid) {
 // inline functions
 //******************************************************************************
 
-inline ompt_thread_t *ompt_get_thread_gtid(int gtid) {
+inline kmp_info_t *ompt_get_thread_gtid(int gtid) {
   return (gtid >= 0) ? __kmp_thread_from_gtid(gtid) : NULL;
 }
 
-inline ompt_thread_t *ompt_get_thread() {
+inline kmp_info_t *ompt_get_thread() {
   int gtid = __kmp_get_gtid();
   return ompt_get_thread_gtid(gtid);
 }
 
-inline void ompt_set_thread_state(ompt_thread_t *thread, omp_state_t state) {
+inline void ompt_set_thread_state(kmp_info_t *thread, omp_state_t state) {
   thread->th.ompt_thread_info.state = state;
 }
 
diff --git a/openmp/runtime/test/ompt/callback.h b/openmp/runtime/test/ompt/callback.h
index 1ded070e03f..df8304344f3 100755
--- a/openmp/runtime/test/ompt/callback.h
+++ b/openmp/runtime/test/ompt/callback.h
@@ -14,7 +14,7 @@
 // Used to detect architecture
 #include "../../src/kmp_platform.h"
 
-static const char* ompt_thread_type_t_values[] = {
+static const char* ompt_thread_t_values[] = {
   NULL,
   "ompt_thread_initial",
   "ompt_thread_worker",
@@ -31,7 +31,7 @@ static const char* ompt_task_status_t_values[] = {
 static const char* ompt_cancel_flag_t_values[] = {
   "ompt_cancel_parallel",
   "ompt_cancel_sections",
-  "ompt_cancel_do",
+  "ompt_cancel_loop",
   "ompt_cancel_taskgroup",
   "ompt_cancel_activated",
   "ompt_cancel_detected",
@@ -194,7 +194,7 @@ ompt_label_##id:
 
 static void
 on_ompt_callback_mutex_acquire(
-  ompt_mutex_kind_t kind,
+  ompt_mutex_t kind,
   unsigned int hint,
   unsigned int impl,
   omp_wait_id_t wait_id,
@@ -224,7 +224,7 @@ on_ompt_callback_mutex_acquire(
 
 static void
 on_ompt_callback_mutex_acquired(
-  ompt_mutex_kind_t kind,
+  ompt_mutex_t kind,
   omp_wait_id_t wait_id,
   const void *codeptr_ra)
 {
@@ -252,7 +252,7 @@ on_ompt_callback_mutex_acquired(
 
 static void
 on_ompt_callback_mutex_released(
-  ompt_mutex_kind_t kind,
+  ompt_mutex_t kind,
   omp_wait_id_t wait_id,
   const void *codeptr_ra)
 {
@@ -297,7 +297,7 @@ on_ompt_callback_nest_lock(
 
 static void
 on_ompt_callback_sync_region(
-  ompt_sync_region_kind_t kind,
+  ompt_sync_region_t kind,
   ompt_scope_endpoint_t endpoint,
   ompt_data_t *parallel_data,
   ompt_data_t *task_data,
@@ -309,6 +309,9 @@ on_ompt_callback_sync_region(
       switch(kind)
       {
         case ompt_sync_region_barrier:
+        case ompt_sync_region_barrier_implicit:
+        case ompt_sync_region_barrier_explicit:
+        case ompt_sync_region_barrier_implementation:
           printf("%" PRIu64 ": ompt_event_barrier_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
           print_ids(0);
           break;
@@ -318,12 +321,17 @@ on_ompt_callback_sync_region(
         case ompt_sync_region_taskgroup:
           printf("%" PRIu64 ": ompt_event_taskgroup_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
           break;
+        case ompt_sync_region_reduction:
+          break;
       }
       break;
     case ompt_scope_end:
       switch(kind)
       {
         case ompt_sync_region_barrier:
+        case ompt_sync_region_barrier_implicit:
+        case ompt_sync_region_barrier_explicit:
+        case ompt_sync_region_barrier_implementation:
           printf("%" PRIu64 ": ompt_event_barrier_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, codeptr_ra);
           break;
         case ompt_sync_region_taskwait:
@@ -332,6 +340,8 @@ on_ompt_callback_sync_region(
         case ompt_sync_region_taskgroup:
           printf("%" PRIu64 ": ompt_event_taskgroup_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, codeptr_ra);
           break;
+        case ompt_sync_region_reduction:
+          break;
       }
       break;
   }
@@ -339,7 +349,7 @@ on_ompt_callback_sync_region(
 
 static void
 on_ompt_callback_sync_region_wait(
-  ompt_sync_region_kind_t kind,
+  ompt_sync_region_t kind,
   ompt_scope_endpoint_t endpoint,
   ompt_data_t *parallel_data,
   ompt_data_t *task_data,
@@ -351,6 +361,9 @@ on_ompt_callback_sync_region_wait(
       switch(kind)
       {
         case ompt_sync_region_barrier:
+        case ompt_sync_region_barrier_implicit:
+        case ompt_sync_region_barrier_explicit:
+        case ompt_sync_region_barrier_implementation:
           printf("%" PRIu64 ": ompt_event_wait_barrier_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
           break;
         case ompt_sync_region_taskwait:
@@ -359,12 +372,17 @@ on_ompt_callback_sync_region_wait(
         case ompt_sync_region_taskgroup:
           printf("%" PRIu64 ": ompt_event_wait_taskgroup_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
           break;
+        case ompt_sync_region_reduction:
+          break;
       }
       break;
     case ompt_scope_end:
       switch(kind)
       {
         case ompt_sync_region_barrier:
+        case ompt_sync_region_barrier_implicit:
+        case ompt_sync_region_barrier_explicit:
+        case ompt_sync_region_barrier_implementation:
           printf("%" PRIu64 ": ompt_event_wait_barrier_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, codeptr_ra);
           break;
         case ompt_sync_region_taskwait:
@@ -373,6 +391,8 @@ on_ompt_callback_sync_region_wait(
         case ompt_sync_region_taskgroup:
           printf("%" PRIu64 ": ompt_event_wait_taskgroup_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, codeptr_ra);
           break;
+        case ompt_sync_region_reduction:
+          break;
       }
       break;
   }
@@ -398,7 +418,7 @@ on_ompt_callback_cancel(
     first_flag_value = ompt_cancel_flag_t_values[0];
   else if(flags & ompt_cancel_sections)
     first_flag_value = ompt_cancel_flag_t_values[1];
-  else if(flags & ompt_cancel_do)
+  else if(flags & ompt_cancel_loop)
     first_flag_value = ompt_cancel_flag_t_values[2];
   else if(flags & ompt_cancel_taskgroup)
     first_flag_value = ompt_cancel_flag_t_values[3];
@@ -437,7 +457,7 @@ on_ompt_callback_implicit_task(
 
 static void
 on_ompt_callback_lock_init(
-  ompt_mutex_kind_t kind,
+  ompt_mutex_t kind,
   unsigned int hint,
   unsigned int impl,
   omp_wait_id_t wait_id,
@@ -458,7 +478,7 @@ on_ompt_callback_lock_init(
 
 static void
 on_ompt_callback_lock_destroy(
-  ompt_mutex_kind_t kind,
+  ompt_mutex_t kind,
   omp_wait_id_t wait_id,
   const void *codeptr_ra)
 {
@@ -477,7 +497,7 @@ on_ompt_callback_lock_destroy(
 
 static void
 on_ompt_callback_work(
-  ompt_work_type_t wstype,
+  ompt_work_t wstype,
   ompt_scope_endpoint_t endpoint,
   ompt_data_t *parallel_data,
   ompt_data_t *task_data,
@@ -561,29 +581,30 @@ on_ompt_callback_master(
   }
 }
 
-static void
-on_ompt_callback_parallel_begin(
-  ompt_data_t *encountering_task_data,
-  const omp_frame_t *encountering_task_frame,
-  ompt_data_t* parallel_data,
-  uint32_t requested_team_size,
-  ompt_invoker_t invoker,
-  const void *codeptr_ra)
-{
+static void on_ompt_callback_parallel_begin(
+    ompt_data_t *encountering_task_data,
+    const omp_frame_t *encountering_task_frame, ompt_data_t *parallel_data,
+    uint32_t requested_team_size, int flag, const void *codeptr_ra) {
   if(parallel_data->ptr)
     printf("0: parallel_data initially not null\n");
   parallel_data->value = ompt_get_unique_id();
-  printf("%" PRIu64 ": ompt_event_parallel_begin: parent_task_id=%" PRIu64 ", parent_task_frame.exit=%p, parent_task_frame.reenter=%p, parallel_id=%" PRIu64 ", requested_team_size=%" PRIu32 ", codeptr_ra=%p, invoker=%d\n", ompt_get_thread_data()->value, encountering_task_data->value, encountering_task_frame->exit_frame, encountering_task_frame->enter_frame, parallel_data->value, requested_team_size, codeptr_ra, invoker);
+  printf("%" PRIu64 ": ompt_event_parallel_begin: parent_task_id=%" PRIu64
+         ", parent_task_frame.exit=%p, parent_task_frame.reenter=%p, "
+         "parallel_id=%" PRIu64 ", requested_team_size=%" PRIu32
+         ", codeptr_ra=%p, invoker=%d\n",
+         ompt_get_thread_data()->value, encountering_task_data->value,
+         encountering_task_frame->exit_frame,
+         encountering_task_frame->enter_frame, parallel_data->value,
+         requested_team_size, codeptr_ra, flag);
 }
 
-static void
-on_ompt_callback_parallel_end(
-  ompt_data_t *parallel_data,
-  ompt_data_t *encountering_task_data,
-  ompt_invoker_t invoker,
-  const void *codeptr_ra)
-{
-  printf("%" PRIu64 ": ompt_event_parallel_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", invoker=%d, codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, encountering_task_data->value, invoker, codeptr_ra);
+static void on_ompt_callback_parallel_end(ompt_data_t *parallel_data,
+                                          ompt_data_t *encountering_task_data,
+                                          int flag, const void *codeptr_ra) {
+  printf("%" PRIu64 ": ompt_event_parallel_end: parallel_id=%" PRIu64
+         ", task_id=%" PRIu64 ", invoker=%d, codeptr_ra=%p\n",
+         ompt_get_thread_data()->value, parallel_data->value,
+         encountering_task_data->value, flag, codeptr_ra);
 }
 
 static void
@@ -648,13 +669,13 @@ on_ompt_callback_task_dependence(
 
 static void
 on_ompt_callback_thread_begin(
-  ompt_thread_type_t thread_type,
+  ompt_thread_t thread_type,
   ompt_data_t *thread_data)
 {
   if(thread_data->ptr)
     printf("%s\n", "0: thread_data initially not null");
   thread_data->value = ompt_get_unique_id();
-  printf("%" PRIu64 ": ompt_event_thread_begin: thread_type=%s=%d, thread_id=%" PRIu64 "\n", ompt_get_thread_data()->value, ompt_thread_type_t_values[thread_type], thread_type, thread_data->value);
+  printf("%" PRIu64 ": ompt_event_thread_begin: thread_type=%s=%d, thread_id=%" PRIu64 "\n", ompt_get_thread_data()->value, ompt_thread_t_values[thread_type], thread_type, thread_data->value);
 }
 
 static void
diff --git a/openmp/runtime/test/ompt/cancel/cancel_worksharing.c b/openmp/runtime/test/ompt/cancel/cancel_worksharing.c
index db3b16864a1..7fa83b22c27 100644
--- a/openmp/runtime/test/ompt/cancel/cancel_worksharing.c
+++ b/openmp/runtime/test/ompt/cancel/cancel_worksharing.c
@@ -58,9 +58,9 @@ int main()
   // CHECK: {{^}}[[MASTER_ID:[0-9]+]]: ompt_event_task_create: parent_task_id=[[PARENT_TASK_ID:[0-9]+]], parent_task_frame.exit=[[NULL]], parent_task_frame.reenter=[[NULL]], new_task_id=[[TASK_ID:[0-9]+]], codeptr_ra=[[NULL]], task_type=ompt_task_initial=1, has_dependences=no
   
   // cancel for and sections
-  // CHECK: {{^}}[[MASTER_ID]]: ompt_event_cancel: task_data=[[TASK_ID:[0-9]+]], flags=ompt_cancel_do|ompt_cancel_activated=20, codeptr_ra={{0x[0-f]*}}
+  // CHECK: {{^}}[[MASTER_ID]]: ompt_event_cancel: task_data=[[TASK_ID:[0-9]+]], flags=ompt_cancel_loop|ompt_cancel_activated=20, codeptr_ra={{0x[0-f]*}}
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_cancel: task_data=[[TASK_ID:[0-9]+]], flags=ompt_cancel_sections|ompt_cancel_activated=18, codeptr_ra={{0x[0-f]*}}
-  // CHECK: {{^}}[[OTHER_THREAD_ID:[0-9]+]]: ompt_event_cancel: task_data=[[TASK_ID:[0-9]+]], flags=ompt_cancel_do|ompt_cancel_detected=36, codeptr_ra={{0x[0-f]*}}
+  // CHECK: {{^}}[[OTHER_THREAD_ID:[0-9]+]]: ompt_event_cancel: task_data=[[TASK_ID:[0-9]+]], flags=ompt_cancel_loop|ompt_cancel_detected=36, codeptr_ra={{0x[0-f]*}}
   // CHECK: {{^}}[[OTHER_THREAD_ID:[0-9]+]]: ompt_event_cancel: task_data=[[TASK_ID:[0-9]+]], flags=ompt_cancel_sections|ompt_cancel_detected=34, codeptr_ra={{0x[0-f]*}}
 
   return 0;
diff --git a/openmp/runtime/test/ompt/loadtool/tool_available_search/tool_available_search.c b/openmp/runtime/test/ompt/loadtool/tool_available_search/tool_available_search.c
index a6fe8e9b521..0f664888b72 100644
--- a/openmp/runtime/test/ompt/loadtool/tool_available_search/tool_available_search.c
+++ b/openmp/runtime/test/ompt/loadtool/tool_available_search/tool_available_search.c
@@ -70,7 +70,7 @@ ompt_start_tool_result_t* ompt_start_tool(
 
 static void
 on_ompt_callback_thread_begin(
-  ompt_thread_type_t thread_type,
+  ompt_thread_t thread_type,
   ompt_data_t *thread_data)
 {
   printf("0: ompt_event_thread_begin\n");
diff --git a/openmp/runtime/test/ompt/synchronization/barrier/implicit_task_data.c b/openmp/runtime/test/ompt/synchronization/barrier/implicit_task_data.c
index 0824b47bd93..5cfbe006eec 100644
--- a/openmp/runtime/test/ompt/synchronization/barrier/implicit_task_data.c
+++ b/openmp/runtime/test/ompt/synchronization/barrier/implicit_task_data.c
@@ -14,7 +14,7 @@
 #include <omp.h>
 #include <ompt.h>
 
-static const char* ompt_thread_type_t_values[] = {
+static const char* ompt_thread_t_values[] = {
   NULL,
   "ompt_thread_initial",
   "ompt_thread_worker",
@@ -59,18 +59,18 @@ int main()
 
 static void
 on_ompt_callback_thread_begin(
-  ompt_thread_type_t thread_type,
+  ompt_thread_t thread_type,
   ompt_data_t *thread_data)
 {
   if(thread_data->ptr)
     printf("%s\n", "0: thread_data initially not null");
   thread_data->value = ompt_get_unique_id();
-  printf("%" PRIu64 ": ompt_event_thread_begin: thread_type=%s=%d, thread_id=%" PRIu64 "\n", ompt_get_thread_data()->value, ompt_thread_type_t_values[thread_type], thread_type, thread_data->value);
+  printf("%" PRIu64 ": ompt_event_thread_begin: thread_type=%s=%d, thread_id=%" PRIu64 "\n", ompt_get_thread_data()->value, ompt_thread_t_values[thread_type], thread_type, thread_data->value);
 }
 
 static void
 on_ompt_callback_sync_region(
-  ompt_sync_region_kind_t kind,
+  ompt_sync_region_t kind,
   ompt_scope_endpoint_t endpoint,
   ompt_data_t *parallel_data,
   ompt_data_t *task_data,
@@ -92,7 +92,7 @@ on_ompt_callback_sync_region(
 
 static void
 on_ompt_callback_sync_region_wait(
-  ompt_sync_region_kind_t kind,
+  ompt_sync_region_t kind,
   ompt_scope_endpoint_t endpoint,
   ompt_data_t *parallel_data,
   ompt_data_t *task_data,