[OMPT] Second chunk of final OMPT 5.0 interface updates

[protze.joachim]

Summary

The ompt.h file is generated from the OpenMP spec to ensure that the interface is implemented as specified.

The other changes are necessary to update the interface implementation to the final version as published in 5.0.

@AndreyChurbanov for the implicit-task callbacks I'm not sure in which cases an initial task in a teams construct might pass the branch and how to identify the initial task.

Patch partially perpared by @sconvent

[protze.joachim]

runtime/src/ompt-general.cpp

348

This part relies on a meaningful implementation of omp_get_initial_device(). Currently this function is not provided by the OpenMP runtime.
Therefore this patch depends on D55578

[protze.joachim]

• I activated the omp_get_initial_device() call to get the number at runtime.
• I also changed the name of ompt.h to omp-tools.h as defined in the SPEC

[hbae]

runtime/src/include/50/omp-tools.h.var

2

Change ompt.h.var to omp-tools.h.var

runtime/src/kmp_csupport.cpp

514

ompt_task_implicit for the last argument.

runtime/src/kmp_wait_release.h

143

ompt_task_implicit seems to be the right value for the last argument.

[protze.joachim]

Implemented changes by Hansang.

Now I also used the flag -M30% for diff generation, to show the changes in the moved header ompt.h -> omp-tools.h

[protze.joachim]

Is this good to go for 8.0?

[hbae]

Can you also change ompt.h to omp-tools.h in runtime/cmake/LibompExports.cmake:61?
Then, I think it is good to go.

[protze.joachim]

Fix the last mentioning of ompt.h

runtime/cmake/LibompExports.cmake

 
 # Put headers in exports/ directory post build
 add_custom_command(TARGET omp POST_BUILD
   COMMAND ${CMAKE_COMMAND} -E make_directory ${LIBOMP_EXPORTS_CMN_DIR}
   COMMAND ${CMAKE_COMMAND} -E copy omp.h ${LIBOMP_EXPORTS_CMN_DIR}
 )
 if(${LIBOMP_OMPT_SUPPORT})
   add_custom_command(TARGET omp POST_BUILD
-    COMMAND ${CMAKE_COMMAND} -E copy ompt.h ${LIBOMP_EXPORTS_CMN_DIR}
+    COMMAND ${CMAKE_COMMAND} -E copy omp-tools.h ${LIBOMP_EXPORTS_CMN_DIR}
   )
 endif()
 if(${LIBOMP_FORTRAN_MODULES})
   add_custom_command(TARGET libomp-mod POST_BUILD
     COMMAND ${CMAKE_COMMAND} -E make_directory ${LIBOMP_EXPORTS_MOD_DIR}
     COMMAND ${CMAKE_COMMAND} -E copy omp_lib.mod ${LIBOMP_EXPORTS_MOD_DIR}
     COMMAND ${CMAKE_COMMAND} -E copy omp_lib_kinds.mod ${LIBOMP_EXPORTS_MOD_DIR}
   )

runtime/src/CMakeLists.txt

 #
 #//===----------------------------------------------------------------------===//
 #//
 #//                     The LLVM Compiler Infrastructure
 #//
 #// This file is dual licensed under the MIT and the University of Illinois Open
 #// Source Licenses. See LICENSE.txt for details.
 #//
 #//===----------------------------------------------------------------------===//
 #
 
-# Configure omp.h, kmp_config.h and ompt.h if necessary
+# Configure omp.h, kmp_config.h and omp-tools.h if necessary
 configure_file(${LIBOMP_INC_DIR}/omp.h.var omp.h @ONLY)
 configure_file(kmp_config.h.cmake kmp_config.h @ONLY)
 if(${LIBOMP_OMPT_SUPPORT})
-  configure_file(${LIBOMP_INC_DIR}/ompt.h.var ompt.h @ONLY)
+  configure_file(${LIBOMP_INC_DIR}/omp-tools.h.var omp-tools.h @ONLY)
 endif()
 
 # Generate message catalog files: kmp_i18n_id.inc and kmp_i18n_default.inc
 add_custom_command(
   OUTPUT  kmp_i18n_id.inc
   COMMAND ${PERL_EXECUTABLE} ${LIBOMP_TOOLS_DIR}/message-converter.pl --os=${LIBOMP_PERL_SCRIPT_OS}
     --prefix=kmp_i18n --enum=kmp_i18n_id.inc ${LIBOMP_SRC_DIR}/i18n/en_US.txt
   DEPENDS ${LIBOMP_SRC_DIR}/i18n/en_US.txt ${LIBOMP_TOOLS_DIR}/message-converter.pl
else()
   endif()
 endif()
 install(
   FILES
   ${CMAKE_CURRENT_BINARY_DIR}/omp.h
   DESTINATION ${LIBOMP_HEADERS_INSTALL_PATH}
 )
 if(${LIBOMP_OMPT_SUPPORT})
-  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ompt.h DESTINATION ${LIBOMP_HEADERS_INSTALL_PATH})
+  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/omp-tools.h DESTINATION ${LIBOMP_HEADERS_INSTALL_PATH})
+  # install under legacy name ompt.h
+  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/omp-tools.h DESTINATION ${LIBOMP_HEADERS_INSTALL_PATH} RENAME ompt.h)
 endif()
 if(${LIBOMP_FORTRAN_MODULES})
   install(FILES
     ${CMAKE_CURRENT_BINARY_DIR}/omp_lib.h
     ${CMAKE_CURRENT_BINARY_DIR}/omp_lib.mod
     ${CMAKE_CURRENT_BINARY_DIR}/omp_lib_kinds.mod
     DESTINATION ${LIBOMP_HEADERS_INSTALL_PATH}
   )
 endif()
 

runtime/src/include/50/omp-tools.h.var

 /*
- * include/50/ompt.h.var
+ * include/50/omp-tools.h.var
  */
 
 //===----------------------------------------------------------------------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is dual licensed under the MIT and the University of Illinois Open
 // Source Licenses. See LICENSE.txt for details.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef __OMPT__
 #define __OMPT__
 
 /*****************************************************************************
  * system include files
  *****************************************************************************/
 
 #include <stdint.h>
 #include <stddef.h>
 
-
-
 /*****************************************************************************
  * iteration macros
  *****************************************************************************/
 
 #define FOREACH_OMPT_INQUIRY_FN(macro)      \
     macro (ompt_enumerate_states)           \
     macro (ompt_enumerate_mutex_impls)      \
                                             \
#define FOREACH_OMPT_STATE(macro)                                                                \
     /* misc (256..511) */                                                                       \
     macro (ompt_state_idle, 0x100)           /* waiting for work */                              \
     macro (ompt_state_overhead, 0x101)       /* overhead excluding wait states */                \
                                                                                                 \
     /* implementation-specific states (512..) */
 
 
 #define FOREACH_KMP_MUTEX_IMPL(macro)                                                \
-    macro (ompt_mutex_impl_none, 0)        /* unknown implementation */              \
+    macro (kmp_mutex_impl_none, 0)         /* unknown implementation */              \
     macro (kmp_mutex_impl_spin, 1)         /* based on spin */                       \
     macro (kmp_mutex_impl_queuing, 2)      /* based on some fair policy */           \
     macro (kmp_mutex_impl_speculative, 3)  /* based on HW-supported speculation */
 
 #define FOREACH_OMPT_EVENT(macro)                                                                                        \
                                                                                                                          \
     /*--- Mandatory Events ---*/                                                                                         \
     macro (ompt_callback_thread_begin,      ompt_callback_thread_begin_t,       1) /* thread begin                    */ \
#define FOREACH_OMPT_EVENT(macro)                                                                                        \
     macro (ompt_callback_device_load,       ompt_callback_device_load_t,       14) /* device load                     */ \
     macro (ompt_callback_device_unload,     ompt_callback_device_unload_t,     15) /* device unload                   */ \
                                                                                                                          \
     /* Optional Events */                                                                                                \
     macro (ompt_callback_sync_region_wait,  ompt_callback_sync_region_t,       16) /* sync region wait begin or end   */ \
                                                                                                                          \
     macro (ompt_callback_mutex_released,    ompt_callback_mutex_t,             17) /* mutex released                  */ \
                                                                                                                          \
-    macro (ompt_callback_task_dependences,  ompt_callback_task_dependences_t,  18) /* report task dependences         */ \
+    macro (ompt_callback_dependences,       ompt_callback_dependences_t,       18) /* report task dependences         */ \
     macro (ompt_callback_task_dependence,   ompt_callback_task_dependence_t,   19) /* report task dependence          */ \
                                                                                                                          \
     macro (ompt_callback_work,              ompt_callback_work_t,              20) /* task at work begin or end       */ \
                                                                                                                          \
     macro (ompt_callback_master,            ompt_callback_master_t,            21) /* task at master begin or end     */ \
                                                                                                                          \
     macro (ompt_callback_target_map,        ompt_callback_target_map_t,        22) /* target map                      */ \
                                                                                                                          \
#define FOREACH_OMPT_EVENT(macro)                                                                                        \
     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                */
 
+/*****************************************************************************
+ * implementation specific types
+ *****************************************************************************/
 
+typedef enum kmp_mutex_impl_t {
+#define kmp_mutex_impl_macro(impl, code) impl = code,
+    FOREACH_KMP_MUTEX_IMPL(kmp_mutex_impl_macro)
+#undef kmp_mutex_impl_macro
+} kmp_mutex_impl_t;
 
 /*****************************************************************************
- * data types
+ * definitions generated from spec
  *****************************************************************************/
 
-/*---------------------
- * identifiers
- *---------------------*/
+typedef enum ompt_callbacks_t {
+  ompt_callback_thread_begin             = 1,
+  ompt_callback_thread_end               = 2,
+  ompt_callback_parallel_begin           = 3,
+  ompt_callback_parallel_end             = 4,
+  ompt_callback_task_create              = 5,
+  ompt_callback_task_schedule            = 6,
+  ompt_callback_implicit_task            = 7,
+  ompt_callback_target                   = 8,
+  ompt_callback_target_data_op           = 9,
+  ompt_callback_target_submit            = 10,
+  ompt_callback_control_tool             = 11,
+  ompt_callback_device_initialize        = 12,
+  ompt_callback_device_finalize          = 13,
+  ompt_callback_device_load              = 14,
+  ompt_callback_device_unload            = 15,
+  ompt_callback_sync_region_wait         = 16,
+  ompt_callback_mutex_released           = 17,
+  ompt_callback_dependences              = 18,
+  ompt_callback_task_dependence          = 19,
+  ompt_callback_work                     = 20,
+  ompt_callback_master                   = 21,
+  ompt_callback_target_map               = 22,
+  ompt_callback_sync_region              = 23,
+  ompt_callback_lock_init                = 24,
+  ompt_callback_lock_destroy             = 25,
+  ompt_callback_mutex_acquire            = 26,
+  ompt_callback_mutex_acquired           = 27,
+  ompt_callback_nest_lock                = 28,
+  ompt_callback_flush                    = 29,
+  ompt_callback_cancel                   = 30,
+  ompt_callback_reduction                = 31,
+  ompt_callback_dispatch                 = 32
+} ompt_callbacks_t;
+
+typedef enum ompt_record_t {
+  ompt_record_ompt               = 1,
+  ompt_record_native             = 2,
+  ompt_record_invalid            = 3
+} ompt_record_t;
+
+typedef enum ompt_record_native_t {
+  ompt_record_native_info  = 1,
+  ompt_record_native_event = 2
+} ompt_record_native_t;
+
+typedef enum ompt_set_result_t {
+  ompt_set_error            = 0,
+  ompt_set_never            = 1,
+  ompt_set_impossible       = 2,
+  ompt_set_sometimes        = 3,
+  ompt_set_sometimes_paired = 4,
+  ompt_set_always           = 5
+} ompt_set_result_t;
 
 typedef uint64_t ompt_id_t;
-#define ompt_id_none 0
 
-typedef union ompt_data_t {
-  uint64_t value; /* data initialized by runtime to unique id */
-  void *ptr;      /* pointer under tool control */
-} ompt_data_t;
+typedef uint64_t ompt_device_time_t;
 
-static const ompt_data_t ompt_data_none = {0};
+typedef uint64_t ompt_buffer_cursor_t;
 
-typedef uint64_t ompt_wait_id_t;
-static const ompt_wait_id_t omp_wait_id_none = 0;
+typedef enum ompt_thread_t {
+  ompt_thread_initial                 = 1,
+  ompt_thread_worker                  = 2,
+  ompt_thread_other                   = 3,
+  ompt_thread_unknown                 = 4
+} ompt_thread_t;
 
-typedef void ompt_device_t;
+typedef enum ompt_scope_endpoint_t {
+  ompt_scope_begin                    = 1,
+  ompt_scope_end                      = 2
+} ompt_scope_endpoint_t;
 
+typedef enum ompt_dispatch_t {
+  ompt_dispatch_iteration             = 1,
+  ompt_dispatch_section               = 2
+} ompt_dispatch_t;
 
-/*---------------------
- * dependences types
- *---------------------*/
-
-typedef enum ompt_task_dependence_type_t {
-    // a two bit field for the dependence type
-    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;
-    ompt_task_dependence_type_t dependence_type;
-} ompt_task_dependence_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 enum ompt_target_data_op_t {
+  ompt_target_data_alloc                = 1,
+  ompt_target_data_transfer_to_device   = 2,
+  ompt_target_data_transfer_from_device = 3,
+  ompt_target_data_delete               = 4,
+  ompt_target_data_associate            = 5,
+  ompt_target_data_disassociate         = 6
+} ompt_target_data_op_t;
 
-/*****************************************************************************
- * enumerations for thread states and runtime events
- *****************************************************************************/
+typedef enum ompt_work_t {
+  ompt_work_loop               = 1,
+  ompt_work_sections           = 2,
+  ompt_work_single_executor    = 3,
+  ompt_work_single_other       = 4,
+  ompt_work_workshare          = 5,
+  ompt_work_distribute         = 6,
+  ompt_work_taskloop           = 7
+} ompt_work_t;
 
-/*---------------------
- * runtime states
- *---------------------*/
-
-typedef enum {
-#define ompt_state_macro(state, code) state = code,
-    FOREACH_OMPT_STATE(ompt_state_macro)
-#undef ompt_state_macro
-} ompt_state_t;
+typedef enum ompt_mutex_t {
+  ompt_mutex_lock                     = 1,
+  ompt_mutex_test_lock                = 2,
+  ompt_mutex_nest_lock                = 3,
+  ompt_mutex_test_nest_lock           = 4,
+  ompt_mutex_critical                 = 5,
+  ompt_mutex_atomic                   = 6,
+  ompt_mutex_ordered                  = 7
+} ompt_mutex_t;
+
+typedef enum ompt_native_mon_flag_t {
+  ompt_native_data_motion_explicit    = 0x01,
+  ompt_native_data_motion_implicit    = 0x02,
+  ompt_native_kernel_invocation       = 0x04,
+  ompt_native_kernel_execution        = 0x08,
+  ompt_native_driver                  = 0x10,
+  ompt_native_runtime                 = 0x20,
+  ompt_native_overhead                = 0x40,
+  ompt_native_idleness                = 0x80
+} ompt_native_mon_flag_t;
+
+typedef enum ompt_task_flag_t {
+  ompt_task_initial                   = 0x00000001,
+  ompt_task_implicit                  = 0x00000002,
+  ompt_task_explicit                  = 0x00000004,
+  ompt_task_target                    = 0x00000008,
+  ompt_task_undeferred                = 0x08000000,
+  ompt_task_untied                    = 0x10000000,
+  ompt_task_final                     = 0x20000000,
+  ompt_task_mergeable                 = 0x40000000,
+  ompt_task_merged                    = 0x80000000
+} ompt_task_flag_t;
+
+typedef enum ompt_task_status_t {
+  ompt_task_complete      = 1,
+  ompt_task_yield         = 2,
+  ompt_task_cancel        = 3,
+  ompt_task_detach        = 4,
+  ompt_task_early_fulfill = 5,
+  ompt_task_late_fulfill  = 6,
+  ompt_task_switch        = 7
+} ompt_task_status_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_t;
+
+typedef enum ompt_parallel_flag_t {
+  ompt_parallel_invoker_program = 0x00000001,
+  ompt_parallel_invoker_runtime = 0x00000002,
+  ompt_parallel_league          = 0x40000000,
+  ompt_parallel_team            = 0x80000000
+} ompt_parallel_flag_t;
+
+typedef enum ompt_target_map_flag_t {
+  ompt_target_map_flag_to             = 0x01,
+  ompt_target_map_flag_from           = 0x02,
+  ompt_target_map_flag_alloc          = 0x04,
+  ompt_target_map_flag_release        = 0x08,
+  ompt_target_map_flag_delete         = 0x10,
+  ompt_target_map_flag_implicit       = 0x20
+} ompt_target_map_flag_t;
+
+typedef enum ompt_dependence_type_t {
+  ompt_dependence_type_in              = 1,
+  ompt_dependence_type_out             = 2,
+  ompt_dependence_type_inout           = 3,
+  ompt_dependence_type_mutexinoutset   = 4,
+  ompt_dependence_type_source          = 5,
+  ompt_dependence_type_sink            = 6
+} ompt_dependence_type_t;
+
+typedef enum ompt_cancel_flag_t {
+  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
+} ompt_cancel_flag_t;
+
+typedef uint64_t ompt_hwid_t;
+
+typedef uint64_t ompt_wait_id_t;
 
 typedef enum ompt_frame_flag_t {
   ompt_frame_runtime        = 0x00,
   ompt_frame_application    = 0x01,
   ompt_frame_cfa            = 0x10,
   ompt_frame_framepointer   = 0x20,
   ompt_frame_stackaddress   = 0x30
 } ompt_frame_flag_t; 
 
+typedef enum ompt_state_t {
+  ompt_state_work_serial                      = 0x000,
+  ompt_state_work_parallel                    = 0x001,
+  ompt_state_work_reduction                   = 0x002,
+
+  ompt_state_wait_barrier                     = 0x010,
+  ompt_state_wait_barrier_implicit_parallel   = 0x011,
+  ompt_state_wait_barrier_implicit_workshare  = 0x012,
+  ompt_state_wait_barrier_implicit            = 0x013,
+  ompt_state_wait_barrier_explicit            = 0x014,
+
+  ompt_state_wait_taskwait                    = 0x020,
+  ompt_state_wait_taskgroup                   = 0x021,
+
+  ompt_state_wait_mutex                       = 0x040,
+  ompt_state_wait_lock                        = 0x041,
+  ompt_state_wait_critical                    = 0x042,
+  ompt_state_wait_atomic                      = 0x043,
+  ompt_state_wait_ordered                     = 0x044,
+
+  ompt_state_wait_target                      = 0x080,
+  ompt_state_wait_target_map                  = 0x081,
+  ompt_state_wait_target_update               = 0x082,
+
+  ompt_state_idle                             = 0x100,
+  ompt_state_overhead                         = 0x101,
+  ompt_state_undefined                        = 0x102
+} ompt_state_t;
 
-/*---------------------
- * runtime events
- *---------------------*/
-
-typedef enum ompt_callbacks_e{
-#define ompt_event_macro(event, callback, eventid) event = eventid,
-    FOREACH_OMPT_EVENT(ompt_event_macro)
-#undef ompt_event_macro
-} ompt_callbacks_t;
-
-
-/*---------------------
- * set callback results
- *---------------------*/
-typedef enum ompt_set_result_t {
-    ompt_set_error = 0,
-    ompt_set_never = 1,
-    ompt_set_sometimes = 2,
-    ompt_set_sometimes_paired = 3,
-    ompt_set_always = 4
-} ompt_set_result_t;
+typedef uint64_t (*ompt_get_unique_id_t) (void);
 
+typedef uint64_t ompd_size_t;
 
-/*----------------------
- * mutex implementations
- *----------------------*/
-typedef enum kmp_mutex_impl_t {
-#define kmp_mutex_impl_macro(impl, code) impl = code,
-    FOREACH_KMP_MUTEX_IMPL(kmp_mutex_impl_macro)
-#undef kmp_mutex_impl_macro
-} kmp_mutex_impl_t;
+typedef uint64_t ompd_wait_id_t;
 
+typedef uint64_t ompd_addr_t;
+typedef int64_t  ompd_word_t;
+typedef uint64_t ompd_seg_t;
+
+typedef uint64_t ompd_device_t;
+
+typedef uint64_t ompd_thread_id_t;
+
+typedef enum ompd_scope_t {
+  ompd_scope_global = 1,
+  ompd_scope_address_space = 2,
+  ompd_scope_thread = 3,
+  ompd_scope_parallel = 4,
+  ompd_scope_implicit_task = 5,
+  ompd_scope_task = 6
+} ompd_scope_t;
+
+typedef uint64_t ompd_icv_id_t;
+
+typedef enum ompd_rc_t {
+  ompd_rc_ok = 0,
+  ompd_rc_unavailable = 1,
+  ompd_rc_stale_handle = 2,
+  ompd_rc_bad_input = 3,
+  ompd_rc_error = 4,
+  ompd_rc_unsupported = 5,
+  ompd_rc_needs_state_tracking = 6,
+  ompd_rc_incompatible = 7,
+  ompd_rc_device_read_error = 8,
+  ompd_rc_device_write_error = 9,
+  ompd_rc_nomem = 10,
+} ompd_rc_t;
 
-/*****************************************************************************
- * callback signatures
- *****************************************************************************/
-
-/* initialization */
-typedef void (*ompt_interface_fn_t)(void);
+typedef void (*ompt_interface_fn_t) (void);
 
-typedef ompt_interface_fn_t (*ompt_function_lookup_t)(
-    const char *                          /* entry point to look up              */
+typedef ompt_interface_fn_t (*ompt_function_lookup_t) (
+  const char *interface_function_name
 );
 
-/* threads */
-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_t;
+typedef union ompt_data_t {
+  uint64_t value;
+  void *ptr;
+} ompt_data_t;
 
 typedef struct ompt_frame_t {
   ompt_data_t exit_frame;
   ompt_data_t enter_frame;
   int exit_frame_flags;
   int enter_frame_flags;
 } ompt_frame_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_t thread_type,            /* type of thread                      */
-    ompt_data_t *thread_data              /* data of thread                      */
+typedef void (*ompt_callback_t) (void);
+
+typedef void ompt_device_t;
+
+typedef void ompt_buffer_t;
+
+typedef void (*ompt_callback_buffer_request_t) (
+  int device_num,
+  ompt_buffer_t **buffer,
+  size_t *bytes
 );
 
-typedef void (*ompt_callback_thread_end_t) (
-    ompt_data_t *thread_data              /* data of thread                      */
+typedef void (*ompt_callback_buffer_complete_t) (
+  int device_num,
+  ompt_buffer_t *buffer,
+  size_t bytes,
+  ompt_buffer_cursor_t begin,
+  int buffer_owned
 );
 
-typedef void (*ompt_wait_callback_t) (
-    ompt_wait_id_t wait_id                /* wait data                           */
+typedef void (*ompt_finalize_t) (
+  ompt_data_t *tool_data
 );
 
-/* parallel and workshares */
-typedef enum ompt_scope_endpoint_t {
-    ompt_scope_begin = 1,
-    ompt_scope_end   = 2
-} ompt_scope_endpoint_t;
+typedef int (*ompt_initialize_t) (
+  ompt_function_lookup_t lookup,
+  int initial_device_num,
+  ompt_data_t *tool_data
+);
+
+typedef struct ompt_start_tool_result_t {
+  ompt_initialize_t initialize;
+  ompt_finalize_t finalize;
+  ompt_data_t tool_data;
+} ompt_start_tool_result_t;
 
+typedef struct ompt_record_abstract_t {
+  ompt_record_native_t rclass;
+  const char *type;
+  ompt_device_time_t start_time;
+  ompt_device_time_t end_time;
+  ompt_hwid_t hwid;
+} ompt_record_abstract_t;
+
+typedef struct ompt_dependence_t {
+  ompt_data_t variable;
+  ompt_dependence_type_t dependence_type;
+} ompt_dependence_t;
 
-/* implicit task */
-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 actual_parallelism,      /* team size                           */
-    unsigned int index                    /* thread number of calling thread     */
+typedef int (*ompt_enumerate_states_t) (
+  int current_state,
+  int *next_state,
+  const char **next_state_name
 );
 
-typedef void (*ompt_callback_parallel_begin_t) (
-    ompt_data_t *encountering_task_data,         /* data of encountering task           */
-    const ompt_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 */
-    int flag,                                    /* flag for additional information     */
-    const void *codeptr_ra                       /* return address of runtime call      */
+typedef int (*ompt_enumerate_mutex_impls_t) (
+  int current_impl,
+  int *next_impl,
+  const char **next_impl_name
 );
 
-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           */
-    int flag,                             /* flag for additional information     */
-    const void *codeptr_ra                /* return address of runtime call      */
+typedef ompt_set_result_t (*ompt_set_callback_t) (
+  ompt_callbacks_t event,
+  ompt_callback_t callback
 );
 
-/* tasks */
-typedef enum ompt_task_flag_t {
-    ompt_task_initial    = 0x1,
-    ompt_task_implicit   = 0x2,
-    ompt_task_explicit   = 0x4,
-    ompt_task_target     = 0x8,
-    ompt_task_undeferred = 0x8000000,
-    ompt_task_untied     = 0x10000000,
-    ompt_task_final      = 0x20000000,
-    ompt_task_mergeable  = 0x40000000,
-    ompt_task_merged     = 0x80000000
-} ompt_task_flag_t;
+typedef int (*ompt_get_callback_t) (
+  ompt_callbacks_t event,
+  ompt_callback_t *callback
+);
 
-typedef enum ompt_task_status_t {
-    ompt_task_complete = 1,
-    ompt_task_yield    = 2,
-    ompt_task_cancel   = 3,
-    ompt_task_switch   = 4
-} ompt_task_status_t;
+typedef ompt_data_t *(*ompt_get_thread_data_t) (void);
 
-typedef void (*ompt_callback_task_schedule_t) (
-    ompt_data_t *prior_task_data,         /* data of prior task                  */
-    ompt_task_status_t prior_task_status, /* status of prior task                */
-    ompt_data_t *next_task_data           /* data of next task                   */
+typedef int (*ompt_get_num_procs_t) (void);
+
+typedef int (*ompt_get_num_places_t) (void);
+
+typedef int (*ompt_get_place_proc_ids_t) (
+  int place_num,
+  int ids_size,
+  int *ids
 );
 
-typedef void (*ompt_callback_task_create_t) (
-    ompt_data_t *encountering_task_data,         /* data of parent task                 */
-    const ompt_frame_t *encountering_task_frame,  /* frame data for parent task          */
-    ompt_data_t *new_task_data,                  /* data 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      */
+typedef int (*ompt_get_place_num_t) (void);
+
+typedef int (*ompt_get_partition_place_nums_t) (
+  int place_nums_size,
+  int *place_nums
 );
 
-/* task dependences */
-typedef void (*ompt_callback_task_dependences_t) (
-    ompt_data_t *task_data,               /* data of task                        */
-    const ompt_task_dependence_t *deps,   /* dependences of task                 */
-    int ndeps                             /* dependences count of task           */
+typedef int (*ompt_get_proc_id_t) (void);
+
+typedef int (*ompt_get_state_t) (
+  ompt_wait_id_t *wait_id
 );
 
-typedef void (*ompt_callback_task_dependence_t) (
-    ompt_data_t *src_task_data,           /* data of source task                 */
-    ompt_data_t *sink_task_data           /* data of sink task                   */
+typedef int (*ompt_get_parallel_info_t) (
+  int ancestor_level,
+  ompt_data_t **parallel_data,
+  int *team_size
 );
 
-/* target and device */
-typedef enum ompt_target_t {
-    ompt_target = 1,
-    ompt_target_enter_data = 2,
-    ompt_target_exit_data = 3,
-    ompt_target_update = 4
-} ompt_target_t;
+typedef int (*ompt_get_task_info_t) (
+  int ancestor_level,
+  int *flags,
+  ompt_data_t **task_data,
+  ompt_frame_t **task_frame,
+  ompt_data_t **parallel_data,
+  int *thread_num
+);
 
-typedef void (*ompt_callback_target_t) (
-    ompt_target_t kind,
-    ompt_scope_endpoint_t endpoint,
-    uint64_t device_num,
-    ompt_data_t *task_data,
-    ompt_id_t target_id,
-    const void *codeptr_ra
+typedef int (*ompt_get_task_memory_t)(
+  void **addr,
+  size_t *size,
+  int block
 );
 
-typedef enum ompt_target_data_op_t {
-    ompt_target_data_alloc = 1,
-    ompt_target_data_transfer_to_dev = 2,
-    ompt_target_data_transfer_from_dev = 3,
-    ompt_target_data_delete = 4
-} ompt_target_data_op_t;
+typedef int (*ompt_get_target_info_t) (
+  uint64_t *device_num,
+  ompt_id_t *target_id,
+  ompt_id_t *host_op_id
+);
 
-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 *src_addr,
-    int src_device_num,
-    void *dest_addr,
-    int dest_device_num,
-    size_t bytes,
-    const void *codeptr_ra
+typedef int (*ompt_get_num_devices_t) (void);
+
+typedef void (*ompt_finalize_tool_t) (void);
+
+typedef int (*ompt_get_device_num_procs_t) (
+  ompt_device_t *device
 );
 
-typedef void (*ompt_callback_target_submit_t) (
-    ompt_id_t target_id,
-    ompt_id_t host_op_id,
-    unsigned int requested_num_teams
+typedef ompt_device_time_t (*ompt_get_device_time_t) (
+  ompt_device_t *device
 );
 
-typedef void (*ompt_callback_target_map_t) (
-    ompt_id_t target_id,
-    unsigned int nitems,
-    void **host_addr,
-    void **device_addr,
-    size_t *bytes,
-    unsigned int *mapping_flags,
-    const void *codeptr_ra
+typedef double (*ompt_translate_time_t) (
+  ompt_device_t *device,
+  ompt_device_time_t time
 );
 
-typedef void (*ompt_callback_device_initialize_t) (
-    uint64_t device_num,
-    const char *type,
-    ompt_device_t *device,
-    ompt_function_lookup_t lookup,
-    const char *documentation
+typedef ompt_set_result_t (*ompt_set_trace_ompt_t) (
+  unsigned int enable,
+  unsigned int etype
 );
 
-typedef void (*ompt_callback_device_finalize_t) (
-    uint64_t device_num
+typedef ompt_set_result_t (*ompt_set_trace_native_t) (
+  ompt_device_t *device,
+  int enable,
+  int flags
 );
 
-typedef void (*ompt_callback_device_load_t) (
-    uint64_t device_num,
-    const char * filename,
-    int64_t offset_in_file,
-    void * vma_in_file,
-    size_t bytes,
-    void * host_addr,
-    void * device_addr,
-    uint64_t module_id
+typedef int (*ompt_start_trace_t) (
+  ompt_device_t *device,
+  ompt_callback_buffer_request_t request,
+  ompt_callback_buffer_complete_t complete
 );
 
-#define ompt_addr_unknown ((void *) ~0)
+typedef int (*ompt_pause_trace_t) (
+  ompt_device_t *device,
+  int begin_pause
+);
 
-typedef void (*ompt_callback_device_unload_t) (
-    uint64_t device_num,
-    uint64_t module_id
+typedef int (*ompt_flush_trace_t) (
+  ompt_device_t *device
 );
 
-/* control_tool */
-typedef int (*ompt_callback_control_tool_t) (
-    uint64_t command,                     /* command of control call             */
-    uint64_t modifier,                    /* modifier of control call            */
-    void *arg,                            /* argument of control call            */
-    const void *codeptr_ra                /* return address of runtime call      */
+typedef int (*ompt_stop_trace_t) (
+  ompt_device_t *device
 );
 
-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 int (*ompt_advance_buffer_cursor_t) (
+  ompt_device_t *device,
+  ompt_buffer_t *buffer,
+  size_t size,
+  ompt_buffer_cursor_t current,
+  ompt_buffer_cursor_t *next
+);
 
-typedef void (*ompt_callback_mutex_acquire_t) (
-    ompt_mutex_t kind,                    /* mutex kind                          */
-    unsigned int hint,                    /* mutex hint                          */
-    unsigned int impl,                    /* mutex implementation                */
-    ompt_wait_id_t wait_id,                /* id of object being awaited          */
-    const void *codeptr_ra                /* return address of runtime call      */
+typedef ompt_record_t (*ompt_get_record_type_t) (
+  ompt_buffer_t *buffer,
+  ompt_buffer_cursor_t current
 );
 
-typedef void (*ompt_callback_mutex_t) (
-    ompt_mutex_t kind,                    /* mutex kind                          */
-    ompt_wait_id_t wait_id,                /* id of object being awaited          */
-    const void *codeptr_ra                /* return address of runtime call      */
+typedef void *(*ompt_get_record_native_t) (
+  ompt_buffer_t *buffer,
+  ompt_buffer_cursor_t current,
+  ompt_id_t *host_op_id
 );
 
-typedef void (*ompt_callback_nest_lock_t) (
-    ompt_scope_endpoint_t endpoint,       /* endpoint of nested lock             */
-    ompt_wait_id_t wait_id,                /* id of object being awaited          */
-    const void *codeptr_ra                /* return address of runtime call      */
+typedef ompt_record_abstract_t *
+(*ompt_get_record_abstract_t) (
+  void *native_record
 );
 
-typedef void (*ompt_callback_master_t) (
-    ompt_scope_endpoint_t endpoint,       /* endpoint of master region           */
-    ompt_data_t *parallel_data,           /* data of parallel region             */
-    ompt_data_t *task_data,               /* data of task                        */
-    const void *codeptr_ra                /* return address of runtime call      */
+typedef void (*ompt_callback_thread_begin_t) (
+  ompt_thread_t thread_type,
+  ompt_data_t *thread_data
 );
 
-typedef enum ompt_work_t {
-    ompt_work_loop            = 1,
-    ompt_work_sections        = 2,
-    ompt_work_single_executor = 3,
-    ompt_work_single_other    = 4,
-    ompt_work_workshare       = 5,
-    ompt_work_distribute      = 6,
-    ompt_work_taskloop        = 7
-} ompt_work_t;
+typedef struct ompt_record_thread_begin_t {
+  ompt_thread_t thread_type;
+} ompt_record_thread_begin_t;
 
-typedef void (*ompt_callback_work_t) (
-    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                        */
-    uint64_t count,                       /* quantity of work                    */
-    const void *codeptr_ra                /* return address of runtime call      */
+typedef void (*ompt_callback_thread_end_t) (
+  ompt_data_t *thread_data
 );
 
-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_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                        */
-    const void *codeptr_ra                /* return address of runtime call      */
+typedef void (*ompt_callback_parallel_begin_t) (
+  ompt_data_t *encountering_task_data,
+  const ompt_frame_t *encountering_task_frame,
+  ompt_data_t *parallel_data,
+  unsigned int requested_parallelism,
+  int flags,
+  const void *codeptr_ra
 );
 
-typedef enum ompt_cancel_flag_t {
-    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
-} ompt_cancel_flag_t;
+typedef struct ompt_record_parallel_begin_t {
+  ompt_id_t encountering_task_id;
+  ompt_id_t parallel_id;
+  unsigned int requested_parallelism;
+  int flags;
+  const void *codeptr_ra;
+} ompt_record_parallel_begin_t;
 
-typedef void (*ompt_callback_cancel_t) (
-    ompt_data_t *task_data,               /* data of task                        */
-    int flags,                            /* cancel flags                        */
-    const void *codeptr_ra                /* return address of runtime call      */
+typedef void (*ompt_callback_parallel_end_t) (
+  ompt_data_t *parallel_data,
+  ompt_data_t *encountering_task_data,
+  int flags,
+  const void *codeptr_ra
 );
 
-typedef void (*ompt_callback_flush_t) (
-    ompt_data_t *thread_data,             /* data of thread                      */
-    const void *codeptr_ra                /* return address of runtime call      */
+typedef struct ompt_record_parallel_end_t {
+  ompt_id_t parallel_id;
+  ompt_id_t encountering_task_id;
+  int flags;
+  const void *codeptr_ra;
+} ompt_record_parallel_end_t;
+
+typedef void (*ompt_callback_work_t) (
+  ompt_work_t wstype,
+  ompt_scope_endpoint_t endpoint,
+  ompt_data_t *parallel_data,
+  ompt_data_t *task_data,
+  uint64_t count,
+  const void *codeptr_ra
 );
 
-typedef enum ompt_dispatch_t {
-    ompt_dispatch_iteration = 1,
-    ompt_dispatch_section =   2
-} ompt_dispatch_t;
+typedef struct ompt_record_work_t {
+  ompt_work_t wstype;
+  ompt_scope_endpoint_t endpoint;
+  ompt_id_t parallel_id;
+  ompt_id_t task_id;
+  uint64_t count;
+  const void *codeptr_ra;
+} ompt_record_work_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_data_t instance 
 );
 
-/****************************************************************************
- * ompt API
- ***************************************************************************/
+typedef struct ompt_record_dispatch_t {
+  ompt_id_t parallel_id;
+  ompt_id_t task_id;
+  ompt_dispatch_t kind;
+  ompt_data_t instance; 
+} ompt_record_dispatch_t;
 
-#ifdef  __cplusplus
-extern "C" {
-#endif
+typedef void (*ompt_callback_task_create_t) (
+  ompt_data_t *encountering_task_data,
+  const ompt_frame_t *encountering_task_frame,
+  ompt_data_t *new_task_data,
+  int flags,
+  int has_dependences,
+  const void *codeptr_ra
+);
 
-#define OMPT_API_FNTYPE(fn) fn##_t
+typedef struct ompt_record_task_create_t {
+  ompt_id_t encountering_task_id;
+  ompt_id_t new_task_id;
+  int flags;
+  int has_dependences;
+  const void *codeptr_ra;
+} ompt_record_task_create_t;
 
-#define OMPT_API_FUNCTION(return_type, fn, args)  \
-    typedef return_type (*OMPT_API_FNTYPE(fn)) args
+typedef void (*ompt_callback_dependences_t) (
+  ompt_data_t *task_data,
+  const ompt_dependence_t *deps,
+  int ndeps
+);
 
+typedef struct ompt_record_dependences_t {
+  ompt_id_t task_id;
+  ompt_dependence_t dep;
+  int ndeps;
+} ompt_record_dependences_t;
 
+typedef void (*ompt_callback_task_dependence_t) (
+  ompt_data_t *src_task_data,
+  ompt_data_t *sink_task_data
+);
 
-/****************************************************************************
- * INQUIRY FUNCTIONS
- ***************************************************************************/
+typedef struct ompt_record_task_dependence_t {
+  ompt_id_t src_task_id;
+  ompt_id_t sink_task_id;
+} ompt_record_task_dependence_t;
 
-/* state */
-OMPT_API_FUNCTION(ompt_state_t, ompt_get_state, (
-    ompt_wait_id_t *wait_id
-));
+typedef void (*ompt_callback_task_schedule_t) (
+  ompt_data_t *prior_task_data,
+  ompt_task_status_t prior_task_status,
+  ompt_data_t *next_task_data
+);
 
-/* thread */
-OMPT_API_FUNCTION(ompt_data_t*, ompt_get_thread_data, (void));
+typedef struct ompt_record_task_schedule_t {
+  ompt_id_t prior_task_id;
+  ompt_task_status_t prior_task_status;
+  ompt_id_t next_task_id;
+} ompt_record_task_schedule_t;
 
-/* parallel region */
-OMPT_API_FUNCTION(int, ompt_get_parallel_info, (
-    int ancestor_level,
-    ompt_data_t **parallel_data,
-    int *team_size
-));
+typedef void (*ompt_callback_implicit_task_t) (
+  ompt_scope_endpoint_t endpoint,
+  ompt_data_t *parallel_data,
+  ompt_data_t *task_data,
+  unsigned int actual_parallelism,
+  unsigned int index,
+  int flags
+);
 
-/* task */
-OMPT_API_FUNCTION(int, ompt_get_task_info, (
-    int ancestor_level,
-    int *type,
-    ompt_data_t **task_data,
-    ompt_frame_t **task_frame,
-    ompt_data_t **parallel_data,
-    int *thread_num
-));
+typedef struct ompt_record_implicit_task_t {
+  ompt_scope_endpoint_t endpoint;
+  ompt_id_t parallel_id;
+  ompt_id_t task_id;
+  unsigned int actual_parallelism;
+  unsigned int index;
+  int flags;
+} ompt_record_implicit_task_t;
 
-OMPT_API_FUNCTION(int, ompt_get_task_memory, (
-    void **addr,
-    size_t *size,
-    int block
-));
+typedef void (*ompt_callback_master_t) (
+  ompt_scope_endpoint_t endpoint,
+  ompt_data_t *parallel_data,
+  ompt_data_t *task_data,
+  const void *codeptr_ra
+);
 
-/* procs */
-OMPT_API_FUNCTION(int, ompt_get_num_procs, (void));
+typedef struct ompt_record_master_t {
+  ompt_scope_endpoint_t endpoint;
+  ompt_id_t parallel_id;
+  ompt_id_t task_id;
+  const void *codeptr_ra;
+} ompt_record_master_t;
 
-/* places */
-OMPT_API_FUNCTION(int, ompt_get_num_places, (void));
+typedef void (*ompt_callback_sync_region_t) (
+  ompt_sync_region_t kind,
+  ompt_scope_endpoint_t endpoint,
+  ompt_data_t *parallel_data,
+  ompt_data_t *task_data,
+  const void *codeptr_ra
+);
 
-OMPT_API_FUNCTION(int, ompt_get_place_proc_ids, (
-    int place_num,
-    int ids_size,
-    int *ids
-));
+typedef struct ompt_record_sync_region_t {
+  ompt_sync_region_t kind;
+  ompt_scope_endpoint_t endpoint;
+  ompt_id_t parallel_id;
+  ompt_id_t task_id;
+  const void *codeptr_ra;
+} ompt_record_sync_region_t;
 
-OMPT_API_FUNCTION(int, ompt_get_place_num, (void));
+typedef void (*ompt_callback_mutex_acquire_t) (
+  ompt_mutex_t kind,
+  unsigned int hint,
+  unsigned int impl,
+  ompt_wait_id_t wait_id,
+  const void *codeptr_ra
+);
 
-OMPT_API_FUNCTION(int, ompt_get_partition_place_nums, (
-    int place_nums_size,
-    int *place_nums
-));
+typedef struct ompt_record_mutex_acquire_t {
+  ompt_mutex_t kind;
+  unsigned int hint;
+  unsigned int impl;
+  ompt_wait_id_t wait_id;
+  const void *codeptr_ra;
+} ompt_record_mutex_acquire_t;
 
-/* proc_id */
-OMPT_API_FUNCTION(int, ompt_get_proc_id, (void));
+typedef void (*ompt_callback_mutex_t) (
+  ompt_mutex_t kind,
+  ompt_wait_id_t wait_id,
+  const void *codeptr_ra
+);
 
+typedef struct ompt_record_mutex_t {
+  ompt_mutex_t kind;
+  ompt_wait_id_t wait_id;
+  const void *codeptr_ra;
+} ompt_record_mutex_t;
 
-/****************************************************************************
- * INITIALIZATION FUNCTIONS
- ***************************************************************************/
+typedef void (*ompt_callback_nest_lock_t) (
+  ompt_scope_endpoint_t endpoint,
+  ompt_wait_id_t wait_id,
+  const void *codeptr_ra
+);
 
-OMPT_API_FUNCTION(int, ompt_initialize, (
-    ompt_function_lookup_t lookup,
-    ompt_data_t *tool_data
-));
+typedef struct ompt_record_nest_lock_t {
+  ompt_scope_endpoint_t endpoint;
+  ompt_wait_id_t wait_id;
+  const void *codeptr_ra;
+} ompt_record_nest_lock_t;
 
-OMPT_API_FUNCTION(void, ompt_finalize, (
-    ompt_data_t *tool_data
-));
+typedef void (*ompt_callback_flush_t) (
+  ompt_data_t *thread_data,
+  const void *codeptr_ra
+);
 
-typedef struct ompt_start_tool_result_t {
-    ompt_initialize_t initialize;
-    ompt_finalize_t finalize;
-    ompt_data_t tool_data;
-} ompt_start_tool_result_t;
+typedef struct ompt_record_flush_t {
+  const void *codeptr_ra;
+} ompt_record_flush_t;
 
-/* initialization interface to be defined by tool */
-#ifdef _WIN32
-__declspec(dllexport)
-#endif
-ompt_start_tool_result_t * ompt_start_tool(
-    unsigned int omp_version, 
-    const char * runtime_version
+typedef void (*ompt_callback_cancel_t) (
+  ompt_data_t *task_data,
+  int flags,
+  const void *codeptr_ra
 );
 
-typedef void (*ompt_callback_t)(void);
+typedef struct ompt_record_cancel_t {
+  ompt_id_t task_id;
+  int flags;
+  const void *codeptr_ra;
+} ompt_record_cancel_t;
 
-OMPT_API_FUNCTION(int, ompt_set_callback, (
-    ompt_callbacks_t which,
-    ompt_callback_t callback
-));
+typedef void (*ompt_callback_device_initialize_t) (
+  int device_num,
+  const char *type,
+  ompt_device_t *device,
+  ompt_function_lookup_t lookup,
+  const char *documentation
+);
 
-OMPT_API_FUNCTION(int, ompt_get_callback, (
-    ompt_callbacks_t which,
-    ompt_callback_t *callback
-));
+typedef void (*ompt_callback_device_finalize_t) (
+  int device_num
+);
 
+typedef void (*ompt_callback_device_load_t) (
+  int device_num,
+  const char *filename,
+  int64_t offset_in_file,
+  void *vma_in_file,
+  size_t bytes,
+  void *host_addr,
+  void *device_addr,
+  uint64_t module_id
+);
 
+typedef void (*ompt_callback_device_unload_t) (
+  int device_num,
+  uint64_t module_id
+);
 
-/****************************************************************************
- * MISCELLANEOUS FUNCTIONS
- ***************************************************************************/
+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 *src_addr,
+  int src_device_num,
+  void *dest_addr,
+  int dest_device_num,
+  size_t bytes,
+  const void *codeptr_ra
+);
 
-/* state enumeration */
-OMPT_API_FUNCTION(int, ompt_enumerate_states, (
-    int current_state,
-    int *next_state,
-    const char **next_state_name
-));
+typedef struct ompt_record_target_data_op_t {
+  ompt_id_t host_op_id;
+  ompt_target_data_op_t optype;
+  void *src_addr;
+  int src_device_num;
+  void *dest_addr;
+  int dest_device_num;
+  size_t bytes;
+  ompt_device_time_t end_time;
+  const void *codeptr_ra;
+} ompt_record_target_data_op_t;
 
-/* mutex implementation enumeration */
-OMPT_API_FUNCTION(int, ompt_enumerate_mutex_impls, (
-    int current_impl,
-    int *next_impl,
-    const char **next_impl_name
-));
+typedef void (*ompt_callback_target_t) (
+  ompt_target_t kind,
+  ompt_scope_endpoint_t endpoint,
+  int device_num,
+  ompt_data_t *task_data,
+  ompt_id_t target_id,
+  const void *codeptr_ra
+);
+
+typedef struct ompt_record_target_t {
+  ompt_target_t kind;
+  ompt_scope_endpoint_t endpoint;
+  int device_num;
+  ompt_id_t task_id;
+  ompt_id_t target_id;
+  const void *codeptr_ra;
+} ompt_record_target_t;
+
+typedef void (*ompt_callback_target_map_t) (
+  ompt_id_t target_id,
+  unsigned int nitems,
+  void **host_addr,
+  void **device_addr,
+  size_t *bytes,
+  unsigned int *mapping_flags,
+  const void *codeptr_ra
+);
 
-/* get_unique_id */
-OMPT_API_FUNCTION(uint64_t, ompt_get_unique_id, (void));
+typedef struct ompt_record_target_map_t {
+  ompt_id_t target_id;
+  unsigned int nitems;
+  void **host_addr;
+  void **device_addr;
+  size_t *bytes;
+  unsigned int *mapping_flags;
+  const void *codeptr_ra;
+} ompt_record_target_map_t;
 
-/* finalize tool */
-OMPT_API_FUNCTION(void, ompt_finalize_tool, (void)); 
+typedef void (*ompt_callback_target_submit_t) (
+  ompt_id_t target_id,
+  ompt_id_t host_op_id,
+  unsigned int requested_num_teams
+);
+
+typedef struct ompt_record_target_kernel_t {
+  ompt_id_t host_op_id;
+  unsigned int requested_num_teams;
+  unsigned int granted_num_teams;
+  ompt_device_time_t end_time;
+} ompt_record_target_kernel_t;
 
-#ifdef  __cplusplus
-};
-#endif
+typedef int (*ompt_callback_control_tool_t) (
+  uint64_t command,
+  uint64_t modifier,
+  void *arg,
+  const void *codeptr_ra
+);
 
-/****************************************************************************
- * TARGET
- ***************************************************************************/
+typedef struct ompt_record_control_tool_t {
+  uint64_t command;
+  uint64_t modifier;
+  const void *codeptr_ra;
+} ompt_record_control_tool_t;
+
+typedef struct ompd_address_t {
+  ompd_seg_t segment;
+  ompd_addr_t address;
+} ompd_address_t;
+
+typedef struct ompd_frame_info_t {
+  ompd_address_t frame_address;
+  ompd_word_t frame_flag;
+} ompd_frame_info_t;
+
+typedef struct _ompd_aspace_handle ompd_address_space_handle_t;
+typedef struct _ompd_thread_handle ompd_thread_handle_t;
+typedef struct _ompd_parallel_handle ompd_parallel_handle_t;
+typedef struct _ompd_task_handle ompd_task_handle_t;
+
+typedef struct _ompd_aspace_cont ompd_address_space_context_t;
+typedef struct _ompd_thread_cont ompd_thread_context_t;
+
+typedef struct ompd_device_type_sizes_t {
+  uint8_t sizeof_char;
+  uint8_t sizeof_short;
+  uint8_t sizeof_int;
+  uint8_t sizeof_long;
+  uint8_t sizeof_long_long;
+  uint8_t sizeof_pointer;
+} ompd_device_type_sizes_t;
+
+typedef struct ompt_record_ompt_t {
+  ompt_callbacks_t type;
+  ompt_device_time_t time;
+  ompt_id_t thread_id;
+  ompt_id_t target_id;
+  union {
+    ompt_record_thread_begin_t thread_begin;
+    ompt_record_parallel_begin_t parallel_begin;
+    ompt_record_parallel_end_t parallel_end;
+    ompt_record_work_t work;
+    ompt_record_dispatch_t dispatch;
+    ompt_record_task_create_t task_create;
+    ompt_record_dependences_t dependences;
+    ompt_record_task_dependence_t task_dependence;
+    ompt_record_task_schedule_t task_schedule;
+    ompt_record_implicit_task_t implicit_task;
+    ompt_record_master_t master;
+    ompt_record_sync_region_t sync_region;
+    ompt_record_mutex_acquire_t mutex_acquire;
+    ompt_record_mutex_t mutex;
+    ompt_record_nest_lock_t nest_lock;
+    ompt_record_flush_t flush;
+    ompt_record_cancel_t cancel;
+    ompt_record_target_t target;
+    ompt_record_target_data_op_t target_data_op;
+    ompt_record_target_map_t target_map;
+    ompt_record_target_kernel_t target_kernel;
+    ompt_record_control_tool_t control_tool;
+  } record;
+} ompt_record_ompt_t;
+
+typedef ompt_record_ompt_t *(*ompt_get_record_ompt_t) (
+  ompt_buffer_t *buffer,
+  ompt_buffer_cursor_t current
+);
 
- OMPT_API_FUNCTION(int, ompt_get_target_info, (
-    uint64_t *device_num,
-    ompt_id_t *target_id,
-    ompt_id_t *host_op_id
-));
+#define ompt_id_none 0
+#define ompt_data_none {0}
+#define ompt_time_none 0
+#define ompt_hwid_none 0
+#define ompt_addr_none ~0
+#define ompt_mutex_impl_none 0
+#define ompt_wait_id_none 0
 
- OMPT_API_FUNCTION(int, ompt_get_num_devices, (void));
+#define ompd_segment_none 0
 
 #endif /* __OMPT__ */

runtime/src/kmp_barrier.cpp

#if OMPT_OPTIONAL
     }
     if (ompt_enabled.ompt_callback_sync_region) {
       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
           ompt_sync_region_barrier, ompt_scope_end, NULL, task_data, codeptr);
     }
 #endif
     if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) {
       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
-          ompt_scope_end, NULL, task_data, 0, ds_tid);
+          ompt_scope_end, NULL, task_data, 0, ds_tid, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
     }
   }
 #endif
 
   // Early exit for reaping threads releasing forkjoin barrier
   if (TCR_4(__kmp_global.g.g_done)) {
     this_thr->th.th_task_team = NULL;
 

runtime/src/kmp_csupport.cpp

 
 #if OMPT_SUPPORT
   if (ompt_enabled.enabled &&
       this_thr->th.ompt_thread_info.state != ompt_state_overhead) {
     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame = ompt_data_none;
     if (ompt_enabled.ompt_callback_implicit_task) {
       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
           ompt_scope_end, NULL, OMPT_CUR_TASK_DATA(this_thr), 1,
-          OMPT_CUR_TASK_INFO(this_thr)->thread_num);
+          OMPT_CUR_TASK_INFO(this_thr)->thread_num, ompt_task_implicit);
     }
 
     // reset clear the task id only after unlinking the task
     ompt_data_t *parent_task_data;
     __ompt_get_task_info_internal(1, NULL, &parent_task_data, NULL, NULL, NULL);
 
     if (ompt_enabled.ompt_callback_parallel_end) {
       ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(

 #endif
     case locktag_tas:
       return kmp_mutex_impl_spin;
 #if KMP_USE_TSX
     case locktag_hle:
       return kmp_mutex_impl_speculative;
 #endif
     default:
-      return ompt_mutex_impl_none;
+      return kmp_mutex_impl_none;
     }
     ilock = KMP_LOOKUP_I_LOCK(user_lock);
   }
   KMP_ASSERT(ilock);
   switch (ilock->type) {
 #if KMP_USE_TSX
   case locktag_adaptive:
   case locktag_rtm:
     return kmp_mutex_impl_speculative;
 #endif
   case locktag_nested_tas:
     return kmp_mutex_impl_spin;
 #if KMP_USE_FUTEX
   case locktag_nested_futex:
 #endif
   case locktag_ticket:
   case locktag_queuing:
   case locktag_drdpa:
   case locktag_nested_ticket:
   case locktag_nested_queuing:
   case locktag_nested_drdpa:
     return kmp_mutex_impl_queuing;
   default:
-    return ompt_mutex_impl_none;
+    return kmp_mutex_impl_none;
   }
 }
 #else
 // For locks without dynamic binding
 static kmp_mutex_impl_t __ompt_get_mutex_impl_type() {
   switch (__kmp_user_lock_kind) {
   case lk_tas:
     return kmp_mutex_impl_spin;
 #if KMP_USE_FUTEX
   case lk_futex:
 #endif
   case lk_ticket:
   case lk_queuing:
   case lk_drdpa:
     return kmp_mutex_impl_queuing;
 #if KMP_USE_TSX
   case lk_hle:
   case lk_rtm:
   case lk_adaptive:
     return kmp_mutex_impl_speculative;
 #endif
   default:
-    return ompt_mutex_impl_none;
+    return kmp_mutex_impl_none;
   }
 }
 #endif // KMP_USE_DYNAMIC_LOCK
 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
 
 /*!
 @ingroup WORK_SHARING
 @param loc  source location information.

runtime/src/kmp_gsupport.cpp

if (ompt_enabled.enabled) {
     ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
     ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
 
     // implicit task callback
     if (ompt_enabled.ompt_callback_implicit_task) {
       ompt_team_size = __kmp_team_from_gtid(gtid)->t.t_nproc;
       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
           ompt_scope_begin, &(team_info->parallel_data),
-          &(task_info->task_data), ompt_team_size, __kmp_tid_from_gtid(gtid));
+          &(task_info->task_data), ompt_team_size, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
       task_info->thread_num = __kmp_tid_from_gtid(gtid);
     }
     thr->th.ompt_thread_info.state = ompt_state_work_parallel;
   }
 #endif
 }
 
 static void __kmp_GOMP_serialized_parallel(ident_t *loc, kmp_int32 gtid,

runtime/src/kmp_runtime.cpp

if (ompt_enabled.enabled &&
     __ompt_lw_taskteam_link(&lw_taskteam, this_thr, 1);
     // don't use lw_taskteam after linking. content was swaped
 
     /* OMPT implicit task begin */
     implicit_task_data = OMPT_CUR_TASK_DATA(this_thr);
     if (ompt_enabled.ompt_callback_implicit_task) {
       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
           ompt_scope_begin, OMPT_CUR_TEAM_DATA(this_thr),
-          OMPT_CUR_TASK_DATA(this_thr), 1, __kmp_tid_from_gtid(global_tid));
+          OMPT_CUR_TASK_DATA(this_thr), 1, __kmp_tid_from_gtid(global_tid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
       OMPT_CUR_TASK_INFO(this_thr)
           ->thread_num = __kmp_tid_from_gtid(global_tid);
     }
 
     /* OMPT state */
     this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
   }
#if OMPT_SUPPORT
           __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
           // don't use lw_taskteam after linking. content was swaped
 
           /* OMPT implicit task begin */
           implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
           if (ompt_enabled.ompt_callback_implicit_task) {
             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
                 ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
-                implicit_task_data, 1, __kmp_tid_from_gtid(gtid));
+                implicit_task_data, 1, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
             OMPT_CUR_TASK_INFO(master_th)
                 ->thread_num = __kmp_tid_from_gtid(gtid);
           }
 
           /* OMPT state */
           master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
         } else {
           exit_runtime_p = &dummy;

 
 #if OMPT_SUPPORT
         *exit_runtime_p = NULL;
         if (ompt_enabled.enabled) {
           OMPT_CUR_TASK_INFO(master_th)->frame.exit_frame = ompt_data_none;
           if (ompt_enabled.ompt_callback_implicit_task) {
             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
                 ompt_scope_end, NULL, implicit_task_data, 1,
-                OMPT_CUR_TASK_INFO(master_th)->thread_num);
+                OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
           }
           __ompt_lw_taskteam_unlink(master_th);
 
           if (ompt_enabled.ompt_callback_parallel_end) {
             ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
                 OMPT_CUR_TEAM_DATA(master_th), OMPT_CUR_TASK_DATA(master_th),
                 OMPT_INVOKER(call_context), return_address);
           }
#if OMPT_SUPPORT
             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
             // don't use lw_taskteam after linking. content was swaped
 
             task_info = OMPT_CUR_TASK_INFO(master_th);
             exit_runtime_p = &(task_info->frame.exit_frame.ptr);
             if (ompt_enabled.ompt_callback_implicit_task) {
               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
-                  &(task_info->task_data), 1, __kmp_tid_from_gtid(gtid));
+                  &(task_info->task_data), 1, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
               OMPT_CUR_TASK_INFO(master_th)
                   ->thread_num = __kmp_tid_from_gtid(gtid);
             }
 
             /* OMPT state */
             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
           } else {
             exit_runtime_p = &dummy;
#endif
           }
 
 #if OMPT_SUPPORT
           if (ompt_enabled.enabled) {
             exit_runtime_p = NULL;
             if (ompt_enabled.ompt_callback_implicit_task) {
               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
                   ompt_scope_end, NULL, &(task_info->task_data), 1,
-                  OMPT_CUR_TASK_INFO(master_th)->thread_num);
+                  OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
             }
 
             __ompt_lw_taskteam_unlink(master_th);
             if (ompt_enabled.ompt_callback_parallel_end) {
               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
                   OMPT_CUR_TEAM_DATA(master_th), parent_task_data,
                   OMPT_INVOKER(call_context), return_address);
             }
#if OMPT_SUPPORT
             task_info = OMPT_CUR_TASK_INFO(master_th);
             exit_runtime_p = &(task_info->frame.exit_frame.ptr);
 
             /* OMPT implicit task begin */
             implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
             if (ompt_enabled.ompt_callback_implicit_task) {
               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
-                  implicit_task_data, 1, __kmp_tid_from_gtid(gtid));
+                  implicit_task_data, 1, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
               OMPT_CUR_TASK_INFO(master_th)
                   ->thread_num = __kmp_tid_from_gtid(gtid);
             }
 
             /* OMPT state */
             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
           } else {
             exit_runtime_p = &dummy;
#endif
           }
 
 #if OMPT_SUPPORT
           if (ompt_enabled.enabled) {
             *exit_runtime_p = NULL;
             if (ompt_enabled.ompt_callback_implicit_task) {
               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
                   ompt_scope_end, NULL, &(task_info->task_data), 1,
-                  OMPT_CUR_TASK_INFO(master_th)->thread_num);
+                  OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
             }
 
             ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
             __ompt_lw_taskteam_unlink(master_th);
             if (ompt_enabled.ompt_callback_parallel_end) {
               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
                   &ompt_parallel_data, parent_task_data,
                   OMPT_INVOKER(call_context), return_address);

 
 #if OMPT_SUPPORT
   if (ompt_enabled.enabled) {
     ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
     if (ompt_enabled.ompt_callback_implicit_task) {
       int ompt_team_size = team->t.t_nproc;
       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
           ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size,
-          OMPT_CUR_TASK_INFO(master_th)->thread_num);
+          OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
     }
 
     task_info->frame.exit_frame = ompt_data_none;
     task_info->task_data = ompt_data_none;
   }
 #endif
 
   KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,
#if OMPT_SUPPORT
 
   my_task_data =
       &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data);
   my_parallel_data = &(team->t.ompt_team_info.parallel_data);
   if (ompt_enabled.ompt_callback_implicit_task) {
     ompt_team_size = team->t.t_nproc;
     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
         ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size,
-        __kmp_tid_from_gtid(gtid));
+        __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
     OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(gtid);
   }
 #endif
 
   {
     KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
     KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
     rc =
#if OMPT_OPTIONAL
     }
     if (ompt_enabled.ompt_callback_sync_region) {
       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
           ompt_sync_region_barrier, ompt_scope_end, NULL, task_data, codeptr);
     }
 #endif
     if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) {
       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
-          ompt_scope_end, NULL, task_data, 0, ds_tid);
+          ompt_scope_end, NULL, task_data, 0, ds_tid, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
     }
   }
 #endif
 
   KMP_MB(); /* Flush all pending memory write invalidates.  */
   KMP_ASSERT(this_thr->th.th_team == team);
 }
 

runtime/src/kmp_taskdeps.cpp

if (ompt_enabled.enabled) {
     }
 
     new_taskdata->ompt_task_info.frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
   }
 
 #if OMPT_OPTIONAL
   /* OMPT grab all dependences if requested by the tool */
   if (ndeps + ndeps_noalias > 0 &&
-      ompt_enabled.ompt_callback_task_dependences) {
+      ompt_enabled.ompt_callback_dependences) {
     kmp_int32 i;
 
     new_taskdata->ompt_task_info.ndeps = ndeps + ndeps_noalias;
     new_taskdata->ompt_task_info.deps =
-        (ompt_task_dependence_t *)KMP_OMPT_DEPS_ALLOC(
-            thread, (ndeps + ndeps_noalias) * sizeof(ompt_task_dependence_t));
+        (ompt_dependence_t *)KMP_OMPT_DEPS_ALLOC(
+            thread, (ndeps + ndeps_noalias) * sizeof(ompt_dependence_t));
 
     KMP_ASSERT(new_taskdata->ompt_task_info.deps != NULL);
 
     for (i = 0; i < ndeps; i++) {
-      new_taskdata->ompt_task_info.deps[i].variable_addr =
+      new_taskdata->ompt_task_info.deps[i].variable.ptr =
           (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_type =
-            ompt_task_dependence_type_inout;
+            ompt_dependence_type_inout;
       else if (dep_list[i].flags.out)
         new_taskdata->ompt_task_info.deps[i].dependence_type =
-            ompt_task_dependence_type_out;
+            ompt_dependence_type_out;
       else if (dep_list[i].flags.in)
         new_taskdata->ompt_task_info.deps[i].dependence_type =
-            ompt_task_dependence_type_in;
+            ompt_dependence_type_in;
     }
     for (i = 0; i < ndeps_noalias; i++) {
-      new_taskdata->ompt_task_info.deps[ndeps + i].variable_addr =
+      new_taskdata->ompt_task_info.deps[ndeps + i].variable.ptr =
           (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_type =
-            ompt_task_dependence_type_inout;
+            ompt_dependence_type_inout;
       else if (noalias_dep_list[i].flags.out)
         new_taskdata->ompt_task_info.deps[ndeps + i].dependence_type =
-            ompt_task_dependence_type_out;
+            ompt_dependence_type_out;
       else if (noalias_dep_list[i].flags.in)
         new_taskdata->ompt_task_info.deps[ndeps + i].dependence_type =
-            ompt_task_dependence_type_in;
+            ompt_dependence_type_in;
     }
-    ompt_callbacks.ompt_callback(ompt_callback_task_dependences)(
+    ompt_callbacks.ompt_callback(ompt_callback_dependences)(
         &(new_taskdata->ompt_task_info.task_data),
         new_taskdata->ompt_task_info.deps, new_taskdata->ompt_task_info.ndeps);
     /* We can now free the allocated memory for the dependencies */
     /* For OMPD we might want to delay the free until task_end */
     KMP_OMPT_DEPS_FREE(thread, new_taskdata->ompt_task_info.deps);
     new_taskdata->ompt_task_info.deps = NULL;
     new_taskdata->ompt_task_info.ndeps = 0;
   }

runtime/src/kmp_wait_release.h

#if OMPT_OPTIONAL
     if (ompt_enabled.ompt_callback_sync_region) {
       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
           ompt_sync_region_barrier, ompt_scope_end, NULL, tId, codeptr);
     }
 #endif
     if (!KMP_MASTER_TID(ds_tid)) {
       if (ompt_enabled.ompt_callback_implicit_task) {
         ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
-            ompt_scope_end, NULL, tId, 0, ds_tid);
+            ompt_scope_end, NULL, tId, 0, ds_tid, ompt_task_implicit);
       }
       // return to idle state
       this_thr->th.ompt_thread_info.state = ompt_state_idle;
     } else {
       this_thr->th.ompt_thread_info.state = ompt_state_overhead;
     }
   }
 }

runtime/src/ompt-event-specific.h

  +--------------------------------------------------------------------------*/
 
 #define ompt_callback_sync_region_wait_implemented                             \
   ompt_event_MAY_ALWAYS_OPTIONAL
 
 #define ompt_callback_mutex_released_implemented ompt_event_MAY_ALWAYS_OPTIONAL
 
 #if OMP_40_ENABLED
-#define ompt_callback_task_dependences_implemented                             \
+#define ompt_callback_dependences_implemented                             \
   ompt_event_MAY_ALWAYS_OPTIONAL
 #define ompt_callback_task_dependence_implemented ompt_event_MAY_ALWAYS_OPTIONAL
 #else
-#define ompt_callback_task_dependences_implemented ompt_event_UNIMPLEMENTED
+#define ompt_callback_dependences_implemented ompt_event_UNIMPLEMENTED
 #define ompt_callback_task_dependence_implemented ompt_event_UNIMPLEMENTED
 #endif /* OMP_40_ENABLED */
 
 #define ompt_callback_work_implemented ompt_event_MAY_ALWAYS_OPTIONAL
 
 #define ompt_callback_master_implemented ompt_event_MAY_ALWAYS_OPTIONAL
 
 #define ompt_callback_target_map_implemented ompt_event_UNIMPLEMENTED

runtime/src/ompt-general.cpp

fprintf(stderr, "Warning: OMP_TOOL has invalid value \"%s\".\n"
             ompt_env_var);
     break;
   }
 #if OMPT_DEBUG
   printf("ompt_pre_init(): ompt_enabled = %d\n", ompt_enabled);
 #endif
 }
 
+extern "C" int omp_get_initial_device(void);
+
 void ompt_post_init() {
   //--------------------------------------------------
   // Execute the post-initialization logic only once.
   //--------------------------------------------------
   static int ompt_post_initialized = 0;
 
   if (ompt_post_initialized)
     return;
 
   ompt_post_initialized = 1;
 
   //--------------------------------------------------
   // Initialize the tool if so indicated.
   //--------------------------------------------------
   if (ompt_start_tool_result) {
     ompt_enabled.enabled = !!ompt_start_tool_result->initialize(
-        ompt_fn_lookup, &(ompt_start_tool_result->tool_data));
+        ompt_fn_lookup, omp_get_initial_device(), &(ompt_start_tool_result->tool_data));
 
     if (!ompt_enabled.enabled) {
       // tool not enabled, zero out the bitmap, and done
       memset(&ompt_enabled, 0, sizeof(ompt_enabled));
       return;
     }
 
     kmp_info_t *root_thread = ompt_get_thread();
OMPT_API_ROUTINE int ompt_enumerate_mutex_impls(int current_impl,
   }
   return 0;
 }
 
 /*****************************************************************************
  * callbacks
  ****************************************************************************/
 
-OMPT_API_ROUTINE int ompt_set_callback(ompt_callbacks_t which,
+OMPT_API_ROUTINE ompt_set_result_t ompt_set_callback(ompt_callbacks_t which,
                                        ompt_callback_t callback) {
   switch (which) {
 
 #define ompt_event_macro(event_name, callback_type, event_id)                  \
   case event_name:                                                             \
     if (ompt_event_implementation_status(event_name)) {                        \
       ompt_callbacks.ompt_callback(event_name) = (callback_type)callback;      \
       ompt_enabled.event_name = (callback != 0);                               \

 
 OMPT_API_ROUTINE int ompt_get_parallel_info(int ancestor_level,
                                             ompt_data_t **parallel_data,
                                             int *team_size) {
   return __ompt_get_parallel_info_internal(ancestor_level, parallel_data,
                                            team_size);
 }
 
-OMPT_API_ROUTINE ompt_state_t ompt_get_state(ompt_wait_id_t *wait_id) {
-  ompt_state_t thread_state = __ompt_get_state_internal(wait_id);
+OMPT_API_ROUTINE int ompt_get_state(ompt_wait_id_t *wait_id) {
+  int thread_state = __ompt_get_state_internal(wait_id);
 
   if (thread_state == ompt_state_undefined) {
     thread_state = ompt_state_work_serial;
   }
 
   return thread_state;
 }
 

runtime/src/ompt-internal.h

 // Source Licenses. See LICENSE.txt for details.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef __OMPT_INTERNAL_H__
 #define __OMPT_INTERNAL_H__
 
 #include "ompt-event-specific.h"
-#include "ompt.h"
+#include "omp-tools.h"
 
 #define OMPT_VERSION 1
 
 #define _OMP_EXTERN extern "C"
 
 #define OMPT_INVOKER(x)                                                        \
   ((x == fork_context_gnu) ? ompt_parallel_invoker_program                     \
                            : ompt_parallel_invoker_runtime)

 
 typedef struct {
   ompt_frame_t frame;
   ompt_data_t task_data;
   struct kmp_taskdata *scheduling_parent;
   int thread_num;
 #if OMP_40_ENABLED
   int ndeps;
-  ompt_task_dependence_t *deps;
+  ompt_dependence_t *deps;
 #endif /* OMP_40_ENABLED */
 } ompt_task_info_t;
 
 typedef struct {
   ompt_data_t parallel_data;
   void *master_return_address;
 } ompt_team_info_t;
 

runtime/src/ompt-specific.cpp

 //----------------------------------------------------------
 
 void __ompt_thread_assign_wait_id(void *variable) {
   kmp_info_t *ti = ompt_get_thread();
 
   ti->th.ompt_thread_info.wait_id = (ompt_wait_id_t)variable;
 }
 
-ompt_state_t __ompt_get_state_internal(ompt_wait_id_t *omp_wait_id) {
+int __ompt_get_state_internal(ompt_wait_id_t *omp_wait_id) {
   kmp_info_t *ti = ompt_get_thread();
 
   if (ti) {
     if (omp_wait_id)
       *omp_wait_id = ti->th.ompt_thread_info.wait_id;
     return ti->th.ompt_thread_info.state;
   }
   return ompt_state_undefined;

runtime/test/lit.cfg

 config.substitutions.append(("%libomp-run", "%t"))
 config.substitutions.append(("%clangXX", config.test_cxx_compiler))
 config.substitutions.append(("%clang", config.test_c_compiler))
 config.substitutions.append(("%openmp_flags", config.test_openmp_flags))
 config.substitutions.append(("%flags", config.test_flags))
 config.substitutions.append(("%python", '"%s"' % (sys.executable)))
 
 if config.has_ompt:
-    config.substitutions.append(("FileCheck", config.test_filecheck))
+    config.substitutions.append(("FileCheck", "tee %%s.out | %s" % config.test_filecheck))
     config.substitutions.append(("%sort-threads", "sort -n -s"))
     if config.operating_system == 'Windows':
         # No such environment variable on Windows.
         config.substitutions.append(("%preload-tool", "true ||"))
         config.substitutions.append(("%no-as-needed-flag", "-Wl,--no-as-needed"))
     elif config.operating_system == 'Darwin':
         config.substitutions.append(("%preload-tool", "env DYLD_INSERT_LIBRARIES=%T/tool.so"))
         # No such linker flag on Darwin.
         config.substitutions.append(("%no-as-needed-flag", ""))
     else:
         config.substitutions.append(("%preload-tool", "env LD_PRELOAD=%T/tool.so"))
         config.substitutions.append(("%no-as-needed-flag", "-Wl,--no-as-needed"))

runtime/test/ompt/callback.h

 #ifndef _BSD_SOURCE
 #define _BSD_SOURCE
 #endif
 #define _DEFAULT_SOURCE
 #include <stdio.h>
 #ifndef __STDC_FORMAT_MACROS
 #define __STDC_FORMAT_MACROS
 #endif
 #include <inttypes.h>
 #include <omp.h>
-#include <ompt.h>
+#include <omp-tools.h>
 #include "ompt-signal.h"
 
 // Used to detect architecture
 #include "../../src/kmp_platform.h"
 
 static const char* ompt_thread_t_values[] = {
   NULL,
   "ompt_thread_initial",
   "ompt_thread_worker",
   "ompt_thread_other"
 };
 
 static const char* ompt_task_status_t_values[] = {
   NULL,
-  "ompt_task_complete",
-  "ompt_task_yield",
-  "ompt_task_cancel",
-  "ompt_task_others"
+  "ompt_task_complete",       // 1
+  "ompt_task_yield",          // 2
+  "ompt_task_cancel",         // 3
+  "ompt_task_detach",         // 4
+  "ompt_task_early_fulfill",  // 5
+  "ompt_task_late_fulfill",   // 6
+  "ompt_task_switch"          // 7
 };
 static const char* ompt_cancel_flag_t_values[] = {
   "ompt_cancel_parallel",
   "ompt_cancel_sections",
   "ompt_cancel_loop",
   "ompt_cancel_taskgroup",
   "ompt_cancel_activated",
   "ompt_cancel_detected",

 }
 
 static void
 on_ompt_callback_implicit_task(
     ompt_scope_endpoint_t endpoint,
     ompt_data_t *parallel_data,
     ompt_data_t *task_data,
     unsigned int team_size,
-    unsigned int thread_num)
+    unsigned int thread_num,
+    int flags)
 {
   switch(endpoint)
   {
     case ompt_scope_begin:
       if(task_data->ptr)
         printf("%s\n", "0: task_data initially not null");
       task_data->value = ompt_get_unique_id();
       printf("%" PRIu64 ": ompt_event_implicit_task_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", team_size=%" PRIu32 ", thread_num=%" PRIu32 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, team_size, thread_num);
on_ompt_callback_task_schedule(
   printf("%" PRIu64 ": ompt_event_task_schedule: first_task_id=%" PRIu64 ", second_task_id=%" PRIu64 ", prior_task_status=%s=%d\n", ompt_get_thread_data()->value, first_task_data->value, second_task_data->value, ompt_task_status_t_values[prior_task_status], prior_task_status);
   if(prior_task_status == ompt_task_complete)
   {
     printf("%" PRIu64 ": ompt_event_task_end: task_id=%" PRIu64 "\n", ompt_get_thread_data()->value, first_task_data->value);
   }
 }
 
 static void
-on_ompt_callback_task_dependences(
+on_ompt_callback_dependences(
   ompt_data_t *task_data,
-  const ompt_task_dependence_t *deps,
+  const ompt_dependence_t *deps,
   int ndeps)
 {
   printf("%" PRIu64 ": ompt_event_task_dependences: task_id=%" PRIu64 ", deps=%p, ndeps=%d\n", ompt_get_thread_data()->value, task_data->value, (void *)deps, ndeps);
 }
 
 static void
 on_ompt_callback_task_dependence(
   ompt_data_t *first_task_data,
if (ompt_set_callback(name, (ompt_callback_t)f_##name) ==   \
       ompt_set_never)                                         \
     printf("0: Could not register callback '" #name "'\n");   \
 }while(0)
 
 #define register_callback(name) register_callback_t(name, name##_t)
 
 int ompt_initialize(
   ompt_function_lookup_t lookup,
+  int initial_device_num,
   ompt_data_t *tool_data)
 {
   ompt_set_callback = (ompt_set_callback_t) lookup("ompt_set_callback");
   ompt_get_callback = (ompt_get_callback_t) lookup("ompt_get_callback");
   ompt_get_state = (ompt_get_state_t) lookup("ompt_get_state");
   ompt_get_task_info = (ompt_get_task_info_t) lookup("ompt_get_task_info");
   ompt_get_thread_data = (ompt_get_thread_data_t) lookup("ompt_get_thread_data");
   ompt_get_parallel_info = (ompt_get_parallel_info_t) lookup("ompt_get_parallel_info");
int ompt_initialize(
   register_callback_t(ompt_callback_lock_init, ompt_callback_mutex_acquire_t);
   register_callback_t(ompt_callback_lock_destroy, ompt_callback_mutex_t);
   register_callback(ompt_callback_work);
   register_callback(ompt_callback_master);
   register_callback(ompt_callback_parallel_begin);
   register_callback(ompt_callback_parallel_end);
   register_callback(ompt_callback_task_create);
   register_callback(ompt_callback_task_schedule);
-  register_callback(ompt_callback_task_dependences);
+  register_callback(ompt_callback_dependences);
   register_callback(ompt_callback_task_dependence);
   register_callback(ompt_callback_thread_begin);
   register_callback(ompt_callback_thread_end);
   printf("0: NULL_POINTER=%p\n", (void*)NULL);
   return 1; //success
 }
 
 void ompt_finalize(ompt_data_t *tool_data)
 {
   printf("0: ompt_event_runtime_shutdown\n");
 }
 
+#ifdef __cplusplus
+extern "C" {
+#endif
 ompt_start_tool_result_t* ompt_start_tool(
   unsigned int omp_version,
   const char *runtime_version)
 {
   static ompt_start_tool_result_t ompt_start_tool_result = {&ompt_initialize,&ompt_finalize, 0};
   return &ompt_start_tool_result;
 }
+#ifdef __cplusplus
+}
+#endif

runtime/test/ompt/cancel/cancel_taskgroup.c

int main()
   // CHECK: {{^}}0: NULL_POINTER=[[NULL:.*$]]
   // CHECK: {{^}}[[MASTER_ID:[0-9]+]]: ompt_event_master_begin: parallel_id=[[PARALLEL_ID:[0-9]+]], task_id=[[PARENT_TASK_ID:[0-9]+]], codeptr_ra={{0x[0-f]*}}
 
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_create: parent_task_id=[[PARENT_TASK_ID]], parent_task_frame.exit={{0x[0-f]*}}, parent_task_frame.reenter={{0x[0-f]*}}, new_task_id=[[FIRST_TASK_ID:[0-9]+]], codeptr_ra={{0x[0-f]*}}, task_type=ompt_task_explicit=4, has_dependences=no
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_create: parent_task_id=[[PARENT_TASK_ID]], parent_task_frame.exit={{0x[0-f]*}}, parent_task_frame.reenter={{0x[0-f]*}}, new_task_id=[[SECOND_TASK_ID:[0-9]+]], codeptr_ra={{0x[0-f]*}}, task_type=ompt_task_explicit=4, has_dependences=no
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_create: parent_task_id=[[PARENT_TASK_ID]], parent_task_frame.exit={{0x[0-f]*}}, parent_task_frame.reenter={{0x[0-f]*}}, new_task_id=[[THIRD_TASK_ID:[0-9]+]], codeptr_ra={{0x[0-f]*}}, task_type=ompt_task_explicit=4, has_dependences=no
 
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_create: parent_task_id=[[PARENT_TASK_ID]], parent_task_frame.exit={{0x[0-f]*}}, parent_task_frame.reenter={{0x[0-f]*}}, new_task_id=[[CANCEL_TASK_ID:[0-9]+]], codeptr_ra={{0x[0-f]*}}, task_type=ompt_task_explicit|ompt_task_undeferred=134217732, has_dependences=no
-  // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[PARENT_TASK_ID]], second_task_id=[[CANCEL_TASK_ID]], prior_task_status=ompt_task_others=4
+  // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[PARENT_TASK_ID]], second_task_id=[[CANCEL_TASK_ID]], prior_task_status=ompt_task_switch=7
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_cancel: task_data=[[CANCEL_TASK_ID]], flags=ompt_cancel_taskgroup|ompt_cancel_activated=24, codeptr_ra={{0x[0-f]*}}
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[CANCEL_TASK_ID]], second_task_id=[[PARENT_TASK_ID]], prior_task_status=ompt_task_cancel=3
 
   // CHECK-DAG: {{^}}{{[0-9]+}}: ompt_event_cancel: task_data={{[0-9]+}}, flags=ompt_cancel_taskgroup|ompt_cancel_discarded_task=72, codeptr_ra=[[NULL]]
   // CHECK-DAG: {{^}}{{[0-9]+}}: ompt_event_cancel: task_data={{[0-9]+}}, flags=ompt_cancel_taskgroup|ompt_cancel_discarded_task=72, codeptr_ra=[[NULL]]
   
   // CHECK-DAG: {{^}}[[THREAD_ID:[0-9]+]]: ompt_event_thread_begin: thread_type=ompt_thread_worker=2, thread_id=[[THREAD_ID]]
   // CHECK-DAG: {{^}}[[THREAD_ID]]: ompt_event_cancel: task_data={{[0-9]+}}, flags=ompt_cancel_taskgroup|ompt_cancel_detected=40, codeptr_ra={{0x[0-f]*}}
 
   return 0;
 }

runtime/test/ompt/loadtool/tool_available/tool_available.c

int main()
   return 0;
 }
 
 #endif /* CODE */
 
 #ifdef TOOL
 
 #include <stdio.h>
-#include <ompt.h>
+#include <omp-tools.h>
 
 int ompt_initialize(
   ompt_function_lookup_t lookup,
   ompt_data_t* tool_data)
 {
   printf("0: NULL_POINTER=%p\n", (void*)NULL);
   return 1; //success
 }

runtime/test/ompt/loadtool/tool_available_search/tool_available_search.c

  *  -DTOOL -DSECOND_TOOL enables the code for the second tool during compilation
  *  -DTOOL -DTHIRD_TOOL enables the code for the third tool during compilation
  *  -DCODE enables the code for the executable during compilation
  */
 
 #ifdef CODE
 #include "stdio.h"
 #include "omp.h"
-#include "ompt.h"
+#include "omp-tools.h"
 
 int main()
 {
   #pragma omp parallel num_threads(2)
   {
     #pragma omp master
     {
       int result = omp_control_tool(omp_control_tool_start, 0, NULL);
int main()
 
   return 0;
 }
 
 #endif /* CODE */
 
 #ifdef TOOL
 
-#include <ompt.h>
+#include <omp-tools.h>
 #include "stdio.h"
 
 #ifdef SECOND_TOOL
 // The second tool has an implementation of ompt_start_tool that returns NULL
 ompt_start_tool_result_t* ompt_start_tool(
   unsigned int omp_version,
   const char *runtime_version)
 {

runtime/test/ompt/loadtool/tool_not_available/tool_not_available.c

  *  loaded and the code for the OpenMP executable. 
  *  -DTOOL enables the code for the tool during compilation
  *  -DCODE enables the code for the executable during compilation
  */
 
 #ifdef CODE
 #include "stdio.h"
 #include "omp.h"
-#include "ompt.h"
+#include "omp-tools.h"
 
 int main()
 {
   #pragma omp parallel num_threads(2)
   {
     #pragma omp master
     {
       int result = omp_control_tool(omp_control_tool_start, 0, NULL);
int main()
 
   return 0;
 }
 
 #endif /* CODE */
 
 #ifdef TOOL
 
-#include <ompt.h>
+#include <omp-tools.h>
 #include "stdio.h"
 
 ompt_start_tool_result_t* ompt_start_tool(
   unsigned int omp_version,
   const char *runtime_version)
 {
   printf("0: Do not initialize tool\n");
   return NULL;
 }
 #endif /* TOOL */

runtime/test/ompt/parallel/parallel_if0.c

//        print_frame(1);
   // CHECK: {{^}}[[MASTER_ID]]: task level 1: parallel_id=[[IMPLICIT_PARALLEL_ID:[0-9]+]], task_id=[[PARENT_TASK_ID]], exit_frame=[[NULL]], reenter_frame={{0x[0-f]+}}
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_parallel_begin: parent_task_id=[[IMPLICIT_TASK_ID]], parent_task_frame.exit={{0x[0-f]+}}, parent_task_frame.reenter={{0x[0-f]+}}, parallel_id=[[NESTED_PARALLEL_ID:[0-9]+]], requested_team_size=1, codeptr_ra=[[NESTED_RETURN_ADDRESS:0x[0-f]+]]{{[0-f][0-f]}}, invoker=[[PARALLEL_INVOKER]]
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_implicit_task_begin: parallel_id=[[NESTED_PARALLEL_ID]], task_id=[[NESTED_IMPLICIT_TASK_ID:[0-9]+]]
   // CHECK: {{^}}[[MASTER_ID]]: task level 0: parallel_id=[[NESTED_PARALLEL_ID]], task_id=[[NESTED_IMPLICIT_TASK_ID]], exit_frame={{0x[0-f]+}}, reenter_frame=[[NULL]]
   // CHECK: {{^}}[[MASTER_ID]]: task level 1: parallel_id=[[PARALLEL_ID]], task_id=[[IMPLICIT_TASK_ID]], exit_frame={{0x[0-f]+}}, reenter_frame={{0x[0-f]+}}
   // CHECK: {{^}}[[MASTER_ID]]: task level 2: parallel_id=[[IMPLICIT_PARALLEL_ID]], task_id=[[PARENT_TASK_ID]], exit_frame=[[NULL]], reenter_frame={{0x[0-f]+}}
 
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_create: parent_task_id=[[NESTED_IMPLICIT_TASK_ID]], parent_task_frame.exit={{0x[0-f]+}}, parent_task_frame.reenter={{0x[0-f]+}}, new_task_id=[[EXPLICIT_TASK_ID:[0-9]+]]
-  // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[NESTED_IMPLICIT_TASK_ID]], second_task_id=[[EXPLICIT_TASK_ID]], prior_task_status=ompt_task_others=4
+  // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[NESTED_IMPLICIT_TASK_ID]], second_task_id=[[EXPLICIT_TASK_ID]], prior_task_status=ompt_task_switch=7
   // CHECK: {{^}}[[MASTER_ID]]: task level 0: parallel_id=[[NESTED_PARALLEL_ID]], task_id=[[EXPLICIT_TASK_ID]], exit_frame={{0x[0-f]+}}, reenter_frame=[[NULL]]
   // CHECK: {{^}}[[MASTER_ID]]: task level 1: parallel_id=[[NESTED_PARALLEL_ID]], task_id=[[NESTED_IMPLICIT_TASK_ID]], exit_frame={{0x[0-f]+}}, reenter_frame={{0x[0-f]+}}
   // CHECK: {{^}}[[MASTER_ID]]: task level 2: parallel_id=[[PARALLEL_ID]], task_id=[[IMPLICIT_TASK_ID]], exit_frame={{0x[0-f]+}}, reenter_frame={{0x[0-f]+}}
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[EXPLICIT_TASK_ID]], second_task_id=[[NESTED_IMPLICIT_TASK_ID]], prior_task_status=ompt_task_complete=1
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_end: task_id=[[EXPLICIT_TASK_ID]]
 
 
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_implicit_task_end: parallel_id=0, task_id=[[NESTED_IMPLICIT_TASK_ID]]
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_parallel_end: parallel_id=[[NESTED_PARALLEL_ID]], task_id=[[IMPLICIT_TASK_ID]], invoker=[[PARALLEL_INVOKER]]
   // CHECK: {{^}}[[MASTER_ID]]: fuzzy_address={{.*}}[[NESTED_RETURN_ADDRESS]]
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_implicit_task_end: parallel_id=0, task_id=[[IMPLICIT_TASK_ID]]
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_parallel_end: parallel_id=[[PARALLEL_ID]], task_id=[[PARENT_TASK_ID]], invoker=[[PARALLEL_INVOKER]]
   // CHECK: {{^}}[[MASTER_ID]]: fuzzy_address={{.*}}[[RETURN_ADDRESS]]
 
   return 0;
 }

runtime/test/ompt/parallel/serialized.c

//        print_frame(1);
   // CHECK: {{^}}[[MASTER_ID]]: task level 1: parallel_id=[[IMPLICIT_PARALLEL_ID:[0-9]+]], task_id=[[PARENT_TASK_ID]], exit_frame=[[NULL]], reenter_frame={{0x[0-f]+}}
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_parallel_begin: parent_task_id=[[IMPLICIT_TASK_ID]], parent_task_frame.exit={{0x[0-f]+}}, parent_task_frame.reenter={{0x[0-f]+}}, parallel_id=[[NESTED_PARALLEL_ID:[0-9]+]], requested_team_size=1, codeptr_ra=[[INNER_RETURN_ADDRESS:0x[0-f]+]]{{[0-f][0-f]}}, invoker=[[PARALLEL_INVOKER]]
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_implicit_task_begin: parallel_id=[[NESTED_PARALLEL_ID]], task_id=[[NESTED_IMPLICIT_TASK_ID:[0-9]+]]
   // CHECK: {{^}}[[MASTER_ID]]: task level 0: parallel_id=[[NESTED_PARALLEL_ID]], task_id=[[NESTED_IMPLICIT_TASK_ID]], exit_frame={{0x[0-f]+}}, reenter_frame=[[NULL]]
   // CHECK: {{^}}[[MASTER_ID]]: task level 1: parallel_id=[[PARALLEL_ID]], task_id=[[IMPLICIT_TASK_ID]], exit_frame={{0x[0-f]+}}, reenter_frame={{0x[0-f]+}}
   // CHECK: {{^}}[[MASTER_ID]]: task level 2: parallel_id=[[IMPLICIT_PARALLEL_ID]], task_id=[[PARENT_TASK_ID]], exit_frame=[[NULL]], reenter_frame={{0x[0-f]+}}
 
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_create: parent_task_id=[[NESTED_IMPLICIT_TASK_ID]], parent_task_frame.exit={{0x[0-f]+}}, parent_task_frame.reenter={{0x[0-f]+}}, new_task_id=[[EXPLICIT_TASK_ID:[0-9]+]]
-  // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[NESTED_IMPLICIT_TASK_ID]], second_task_id=[[EXPLICIT_TASK_ID]], prior_task_status=ompt_task_others=4
+  // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[NESTED_IMPLICIT_TASK_ID]], second_task_id=[[EXPLICIT_TASK_ID]], prior_task_status=ompt_task_switch=7
   // CHECK: {{^}}[[MASTER_ID]]: task level 0: parallel_id=[[NESTED_PARALLEL_ID]], task_id=[[EXPLICIT_TASK_ID]], exit_frame={{0x[0-f]+}}, reenter_frame=[[NULL]]
   // CHECK: {{^}}[[MASTER_ID]]: task level 1: parallel_id=[[NESTED_PARALLEL_ID]], task_id=[[NESTED_IMPLICIT_TASK_ID]], exit_frame={{0x[0-f]+}}, reenter_frame={{0x[0-f]+}}
   // CHECK: {{^}}[[MASTER_ID]]: task level 2: parallel_id=[[PARALLEL_ID]], task_id=[[IMPLICIT_TASK_ID]], exit_frame={{0x[0-f]+}}, reenter_frame={{0x[0-f]+}}
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[EXPLICIT_TASK_ID]], second_task_id=[[NESTED_IMPLICIT_TASK_ID]], prior_task_status=ompt_task_complete=1
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_end: task_id=[[EXPLICIT_TASK_ID]]
 
 
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_implicit_task_end: parallel_id=0, task_id=[[NESTED_IMPLICIT_TASK_ID]]

runtime/test/ompt/synchronization/barrier/implicit_task_data.c

 // RUN: %libomp-compile-and-run | %sort-threads | FileCheck %s
 // REQUIRES: ompt
 
 // This test checks that values stored in task_data in a barrier_begin event
 // are still present in the corresponding barrier_end event.
 // Therefore, callback implementations different from the ones in callback.h are neccessary.
 // This is a test for an issue reported in 
 // https://github.com/OpenMPToolsInterface/LLVM-openmp/issues/39
 
 #define _BSD_SOURCE
 #include <stdio.h>
 #include <unistd.h>
 #include <inttypes.h>
 #include <omp.h>
-#include <ompt.h>
+#include <omp-tools.h>
 
 static const char* ompt_thread_t_values[] = {
   NULL,
   "ompt_thread_initial",
   "ompt_thread_worker",
   "ompt_thread_other"
 };
 

runtime/test/ompt/tasks/explicit_task.c

-// RUN: %libomp-compile-and-run | %sort-threads | FileCheck %s
+// RUN: %libomp-compile-and-run | %sort-threads | tee %s.out | FileCheck %s
 // REQUIRES: ompt
 // UNSUPPORTED: gcc-4, gcc-5, gcc-6, gcc-7
 #define TEST_NEED_PRINT_FRAME_FROM_OUTLINED_FN
 #include "callback.h"
 #include <omp.h> 
 
 int main()
 {

runtime/test/ompt/tasks/taskyield.c

int main()
   // CHECK: {{^}}[[MASTER_ID:[0-9]+]]: ompt_event_implicit_task_begin: parallel_id={{[0-9]+}}, task_id=[[IMPLICIT_TASK_ID:[0-9]+]], team_size={{[0-9]+}}, thread_num={{[0-9]+}}
 
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_create: parent_task_id={{[0-9]+}}, parent_task_frame.exit={{0x[0-f]+}}, parent_task_frame.reenter={{0x[0-f]+}}, new_task_id=[[WORKER_TASK:[0-9]+]], codeptr_ra={{0x[0-f]+}}, task_type=ompt_task_explicit=4, has_dependences=no
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_create: parent_task_id={{[0-9]+}}, parent_task_frame.exit={{0x[0-f]+}}, parent_task_frame.reenter={{0x[0-f]+}}, new_task_id=[[MAIN_TASK:[0-9]+]], codeptr_ra={{0x[0-f]+}}, task_type=ompt_task_explicit=4, has_dependences=no
 
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[IMPLICIT_TASK_ID]], second_task_id=[[MAIN_TASK]], prior_task_status=ompt_task_yield=2
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_task_schedule: first_task_id=[[MAIN_TASK]], second_task_id=[[IMPLICIT_TASK_ID]], prior_task_status=ompt_task_complete=1
 
-  // CHECK: {{^}}[[THREAD_ID:[0-9]+]]: ompt_event_task_schedule: first_task_id={{[0-9]+}}, second_task_id=[[WORKER_TASK]], prior_task_status=ompt_task_others=4
+  // CHECK: {{^}}[[THREAD_ID:[0-9]+]]: ompt_event_task_schedule: first_task_id={{[0-9]+}}, second_task_id=[[WORKER_TASK]], prior_task_status=ompt_task_switch=7
   // CHECK: {{^}}[[THREAD_ID]]: ompt_event_task_schedule: first_task_id=[[WORKER_TASK]], second_task_id={{[0-9]+}}, prior_task_status=ompt_task_complete=1
 
 
 
 
 
   return 0;
 }