]> git.ipfire.org Git - thirdparty/gcc.git/blobdiff - libgomp/oacc-parallel.c
Update copyright years.
[thirdparty/gcc.git] / libgomp / oacc-parallel.c
index fa99a2ad1a90778146cfa9ef734247c33e9b70da..4499cd5ddd1fc76890ffcc95eca27ae2e37b0bdf 100644 (file)
@@ -1,4 +1,4 @@
-/* Copyright (C) 2013-2019 Free Software Foundation, Inc.
+/* Copyright (C) 2013-2024 Free Software Foundation, Inc.
 
    Contributed by Mentor Embedded.
 
@@ -28,7 +28,6 @@
 
 #include "openacc.h"
 #include "libgomp.h"
-#include "libgomp_g.h"
 #include "gomp-constants.h"
 #include "oacc-int.h"
 #ifdef HAVE_INTTYPES_H
@@ -48,25 +47,6 @@ _Static_assert (GOACC_FLAGS_UNMARSHAL (GOMP_DEVICE_HOST_FALLBACK)
                "legacy GOMP_DEVICE_HOST_FALLBACK broken");
 
 
-/* Returns the number of mappings associated with the pointer or pset. PSET
-   have three mappings, whereas pointer have two.  */
-
-static int
-find_pointer (int pos, size_t mapnum, unsigned short *kinds)
-{
-  if (pos + 1 >= mapnum)
-    return 0;
-
-  unsigned char kind = kinds[pos+1] & 0xff;
-
-  if (kind == GOMP_MAP_TO_PSET)
-    return 3;
-  else if (kind == GOMP_MAP_POINTER)
-    return 2;
-
-  return 0;
-}
-
 /* Handle the mapping pair that are presented when a
    deviceptr clause is used with Fortran.  */
 
@@ -112,8 +92,6 @@ handle_ftn_pointers (size_t mapnum, void **hostaddrs, size_t *sizes,
     }
 }
 
-static void goacc_wait (int async, int num_waits, va_list *ap);
-
 
 /* Launch a possibly offloaded function with FLAGS.  FN is the host fn
    address.  MAPNUM, HOSTADDRS, SIZES & KINDS  describe the memory
@@ -130,8 +108,6 @@ GOACC_parallel_keyed (int flags_m, void (*fn) (void *),
   va_list ap;
   struct goacc_thread *thr;
   struct gomp_device_descr *acc_dev;
-  struct target_mem_desc *tgt;
-  void **devaddrs;
   unsigned int i;
   struct splay_tree_key_s k;
   splay_tree_key tgt_fn_key;
@@ -152,21 +128,78 @@ GOACC_parallel_keyed (int flags_m, void (*fn) (void *),
   thr = goacc_thread ();
   acc_dev = thr->dev;
 
+  bool profiling_p = GOACC_PROFILING_DISPATCH_P (true);
+
+  acc_prof_info prof_info;
+  if (profiling_p)
+    {
+      thr->prof_info = &prof_info;
+
+      prof_info.event_type = acc_ev_compute_construct_start;
+      prof_info.valid_bytes = _ACC_PROF_INFO_VALID_BYTES;
+      prof_info.version = _ACC_PROF_INFO_VERSION;
+      prof_info.device_type = acc_device_type (acc_dev->type);
+      prof_info.device_number = acc_dev->target_id;
+      prof_info.thread_id = -1;
+      prof_info.async = async;
+      prof_info.async_queue = prof_info.async;
+      prof_info.src_file = NULL;
+      prof_info.func_name = NULL;
+      prof_info.line_no = -1;
+      prof_info.end_line_no = -1;
+      prof_info.func_line_no = -1;
+      prof_info.func_end_line_no = -1;
+    }
+  acc_event_info compute_construct_event_info;
+  if (profiling_p)
+    {
+      compute_construct_event_info.other_event.event_type
+       = prof_info.event_type;
+      compute_construct_event_info.other_event.valid_bytes
+       = _ACC_OTHER_EVENT_INFO_VALID_BYTES;
+      compute_construct_event_info.other_event.parent_construct
+       = acc_construct_parallel;
+      compute_construct_event_info.other_event.implicit = 0;
+      compute_construct_event_info.other_event.tool_info = NULL;
+    }
+  acc_api_info api_info;
+  if (profiling_p)
+    {
+      thr->api_info = &api_info;
+
+      api_info.device_api = acc_device_api_none;
+      api_info.valid_bytes = _ACC_API_INFO_VALID_BYTES;
+      api_info.device_type = prof_info.device_type;
+      api_info.vendor = -1;
+      api_info.device_handle = NULL;
+      api_info.context_handle = NULL;
+      api_info.async_handle = NULL;
+    }
+
+  if (profiling_p)
+    goacc_profiling_dispatch (&prof_info, &compute_construct_event_info,
+                             &api_info);
+
   handle_ftn_pointers (mapnum, hostaddrs, sizes, kinds);
 
   /* Host fallback if "if" clause is false or if the current device is set to
      the host.  */
-  if (flags & GOACC_FLAG_HOST_FALLBACK)
+  if ((flags & GOACC_FLAG_HOST_FALLBACK)
+      /* TODO: a proper pthreads based "multi-core CPU" local device
+        implementation. Currently, this is still the same as host-fallback.  */
+      || (flags & GOACC_FLAG_LOCAL_DEVICE))
     {
+      prof_info.device_type = acc_device_host;
+      api_info.device_type = prof_info.device_type;
       goacc_save_and_set_bind (acc_device_host);
       fn (hostaddrs);
       goacc_restore_bind ();
-      return;
+      goto out_prof;
     }
   else if (acc_device_type (acc_dev->type) == acc_device_host)
     {
       fn (hostaddrs);
-      return;
+      goto out_prof;
     }
 
   /* Default: let the runtime choose.  */
@@ -200,6 +233,13 @@ GOACC_parallel_keyed (int flags_m, void (*fn) (void *),
 
            if (async == GOMP_LAUNCH_OP_MAX)
              async = va_arg (ap, unsigned);
+
+           if (profiling_p)
+             {
+               prof_info.async = async;
+               prof_info.async_queue = prof_info.async;
+             }
+
            break;
          }
 
@@ -233,28 +273,76 @@ GOACC_parallel_keyed (int flags_m, void (*fn) (void *),
   else
     tgt_fn = (void (*)) fn;
 
+  acc_event_info enter_exit_data_event_info;
+  if (profiling_p)
+    {
+      prof_info.event_type = acc_ev_enter_data_start;
+      enter_exit_data_event_info.other_event.event_type
+       = prof_info.event_type;
+      enter_exit_data_event_info.other_event.valid_bytes
+       = _ACC_OTHER_EVENT_INFO_VALID_BYTES;
+      enter_exit_data_event_info.other_event.parent_construct
+       = compute_construct_event_info.other_event.parent_construct;
+      enter_exit_data_event_info.other_event.implicit = 1;
+      enter_exit_data_event_info.other_event.tool_info = NULL;
+      goacc_profiling_dispatch (&prof_info, &enter_exit_data_event_info,
+                               &api_info);
+    }
+
   goacc_aq aq = get_goacc_asyncqueue (async);
 
-  tgt = gomp_map_vars_async (acc_dev, aq, mapnum, hostaddrs, NULL, sizes, kinds,
-                            true, GOMP_MAP_VARS_OPENACC);
-  
-  devaddrs = gomp_alloca (sizeof (void *) * mapnum);
-  for (i = 0; i < mapnum; i++)
-    devaddrs[i] = (void *) (tgt->list[i].key->tgt->tgt_start
-                           + tgt->list[i].key->tgt_offset
-                           + tgt->list[i].offset);
-  if (aq == NULL)
+  struct target_mem_desc *tgt
+    = goacc_map_vars (acc_dev, aq, mapnum, hostaddrs, NULL, sizes, kinds, true,
+                     GOMP_MAP_VARS_TARGET);
+
+  if (profiling_p)
     {
-      acc_dev->openacc.exec_func (tgt_fn, mapnum, hostaddrs, devaddrs,
-                                 dims, tgt);
-      /* If running synchronously, unmap immediately.  */
-      gomp_unmap_vars (tgt, true);
+      prof_info.event_type = acc_ev_enter_data_end;
+      enter_exit_data_event_info.other_event.event_type
+       = prof_info.event_type;
+      goacc_profiling_dispatch (&prof_info, &enter_exit_data_event_info,
+                               &api_info);
     }
+
+  void **devaddrs = (void **) tgt->tgt_start;
+  if (aq == NULL)
+    acc_dev->openacc.exec_func (tgt_fn, mapnum, hostaddrs, devaddrs, dims,
+                               tgt);
   else
+    acc_dev->openacc.async.exec_func (tgt_fn, mapnum, hostaddrs, devaddrs,
+                                     dims, tgt, aq);
+
+  if (profiling_p)
+    {
+      prof_info.event_type = acc_ev_exit_data_start;
+      enter_exit_data_event_info.other_event.event_type = prof_info.event_type;
+      enter_exit_data_event_info.other_event.tool_info = NULL;
+      goacc_profiling_dispatch (&prof_info, &enter_exit_data_event_info,
+                               &api_info);
+    }
+
+  /* If running synchronously (aq == NULL), this will unmap immediately.  */
+  goacc_unmap_vars (tgt, true, aq);
+
+  if (profiling_p)
+    {
+      prof_info.event_type = acc_ev_exit_data_end;
+      enter_exit_data_event_info.other_event.event_type = prof_info.event_type;
+      goacc_profiling_dispatch (&prof_info, &enter_exit_data_event_info,
+                               &api_info);
+    }
+
+ out_prof:
+  if (profiling_p)
     {
-      acc_dev->openacc.async.exec_func (tgt_fn, mapnum, hostaddrs, devaddrs,
-                                       dims, tgt, aq);
-      gomp_unmap_vars_async (tgt, true, aq);
+      prof_info.event_type = acc_ev_compute_construct_end;
+      compute_construct_event_info.other_event.event_type
+       = prof_info.event_type;
+      goacc_profiling_dispatch (&prof_info, &compute_construct_event_info,
+                               &api_info);
+
+      thr->prof_info = NULL;
+      thr->api_info = NULL;
     }
 }
 
@@ -293,230 +381,175 @@ GOACC_data_start (int flags_m, size_t mapnum,
   struct goacc_thread *thr = goacc_thread ();
   struct gomp_device_descr *acc_dev = thr->dev;
 
+  bool profiling_p = GOACC_PROFILING_DISPATCH_P (true);
+
+  acc_prof_info prof_info;
+  if (profiling_p)
+    {
+      thr->prof_info = &prof_info;
+
+      prof_info.event_type = acc_ev_enter_data_start;
+      prof_info.valid_bytes = _ACC_PROF_INFO_VALID_BYTES;
+      prof_info.version = _ACC_PROF_INFO_VERSION;
+      prof_info.device_type = acc_device_type (acc_dev->type);
+      prof_info.device_number = acc_dev->target_id;
+      prof_info.thread_id = -1;
+      prof_info.async = acc_async_sync; /* Always synchronous.  */
+      prof_info.async_queue = prof_info.async;
+      prof_info.src_file = NULL;
+      prof_info.func_name = NULL;
+      prof_info.line_no = -1;
+      prof_info.end_line_no = -1;
+      prof_info.func_line_no = -1;
+      prof_info.func_end_line_no = -1;
+    }
+  acc_event_info enter_data_event_info;
+  if (profiling_p)
+    {
+      enter_data_event_info.other_event.event_type
+       = prof_info.event_type;
+      enter_data_event_info.other_event.valid_bytes
+       = _ACC_OTHER_EVENT_INFO_VALID_BYTES;
+      enter_data_event_info.other_event.parent_construct = acc_construct_data;
+      for (int i = 0; i < mapnum; ++i)
+       if ((kinds[i] & 0xff) == GOMP_MAP_USE_DEVICE_PTR
+           || (kinds[i] & 0xff) == GOMP_MAP_USE_DEVICE_PTR_IF_PRESENT)
+         {
+           /* If there is one such data mapping kind, then this is actually an
+              OpenACC 'host_data' construct.  (GCC maps the OpenACC
+              'host_data' construct to the OpenACC 'data' construct.)  Apart
+              from artificial test cases (such as an OpenACC 'host_data'
+              construct's (implicit) device initialization when there hasn't
+              been any device data be set up before...), there can't really
+              any meaningful events be generated from OpenACC 'host_data'
+              constructs, though.  */
+           enter_data_event_info.other_event.parent_construct
+             = acc_construct_host_data;
+           break;
+         }
+      enter_data_event_info.other_event.implicit = 0;
+      enter_data_event_info.other_event.tool_info = NULL;
+    }
+  acc_api_info api_info;
+  if (profiling_p)
+    {
+      thr->api_info = &api_info;
+
+      api_info.device_api = acc_device_api_none;
+      api_info.valid_bytes = _ACC_API_INFO_VALID_BYTES;
+      api_info.device_type = prof_info.device_type;
+      api_info.vendor = -1;
+      api_info.device_handle = NULL;
+      api_info.context_handle = NULL;
+      api_info.async_handle = NULL;
+    }
+
+  if (profiling_p)
+    goacc_profiling_dispatch (&prof_info, &enter_data_event_info, &api_info);
+
   /* Host fallback or 'do nothing'.  */
   if ((acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
-      || (flags & GOACC_FLAG_HOST_FALLBACK))
+      || (flags & GOACC_FLAG_HOST_FALLBACK)
+      || (flags & GOACC_FLAG_LOCAL_DEVICE))
     {
-      tgt = gomp_map_vars (NULL, 0, NULL, NULL, NULL, NULL, true,
-                          GOMP_MAP_VARS_OPENACC);
+      prof_info.device_type = acc_device_host;
+      api_info.device_type = prof_info.device_type;
+      tgt = goacc_map_vars (NULL, NULL, 0, NULL, NULL, NULL, NULL, true, 0);
       tgt->prev = thr->mapped_data;
       thr->mapped_data = tgt;
 
-      return;
+      goto out_prof;
     }
 
   gomp_debug (0, "  %s: prepare mappings\n", __FUNCTION__);
-  tgt = gomp_map_vars (acc_dev, mapnum, hostaddrs, NULL, sizes, kinds, true,
-                      GOMP_MAP_VARS_OPENACC);
+  tgt = goacc_map_vars (acc_dev, NULL, mapnum, hostaddrs, NULL, sizes, kinds,
+                       true, 0);
   gomp_debug (0, "  %s: mappings prepared\n", __FUNCTION__);
   tgt->prev = thr->mapped_data;
   thr->mapped_data = tgt;
+
+ out_prof:
+  if (profiling_p)
+    {
+      prof_info.event_type = acc_ev_enter_data_end;
+      enter_data_event_info.other_event.event_type = prof_info.event_type;
+      goacc_profiling_dispatch (&prof_info, &enter_data_event_info, &api_info);
+
+      thr->prof_info = NULL;
+      thr->api_info = NULL;
+    }
 }
 
 void
 GOACC_data_end (void)
 {
   struct goacc_thread *thr = goacc_thread ();
+  struct gomp_device_descr *acc_dev = thr->dev;
   struct target_mem_desc *tgt = thr->mapped_data;
 
-  gomp_debug (0, "  %s: restore mappings\n", __FUNCTION__);
-  thr->mapped_data = tgt->prev;
-  gomp_unmap_vars (tgt, true);
-  gomp_debug (0, "  %s: mappings restored\n", __FUNCTION__);
-}
-
-void
-GOACC_enter_exit_data (int flags_m, size_t mapnum,
-                      void **hostaddrs, size_t *sizes, unsigned short *kinds,
-                      int async, int num_waits, ...)
-{
-  int flags = GOACC_FLAGS_UNMARSHAL (flags_m);
-
-  struct goacc_thread *thr;
-  struct gomp_device_descr *acc_dev;
-  bool data_enter = false;
-  size_t i;
-
-  goacc_lazy_initialize ();
-
-  thr = goacc_thread ();
-  acc_dev = thr->dev;
-
-  if ((acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
-      || (flags & GOACC_FLAG_HOST_FALLBACK))
-    return;
+  bool profiling_p = GOACC_PROFILING_DISPATCH_P (true);
 
-  if (num_waits)
+  acc_prof_info prof_info;
+  if (profiling_p)
     {
-      va_list ap;
-
-      va_start (ap, num_waits);
-      goacc_wait (async, num_waits, &ap);
-      va_end (ap);
+      thr->prof_info = &prof_info;
+
+      prof_info.event_type = acc_ev_exit_data_start;
+      prof_info.valid_bytes = _ACC_PROF_INFO_VALID_BYTES;
+      prof_info.version = _ACC_PROF_INFO_VERSION;
+      prof_info.device_type = acc_device_type (acc_dev->type);
+      prof_info.device_number = acc_dev->target_id;
+      prof_info.thread_id = -1;
+      prof_info.async = acc_async_sync; /* Always synchronous.  */
+      prof_info.async_queue = prof_info.async;
+      prof_info.src_file = NULL;
+      prof_info.func_name = NULL;
+      prof_info.line_no = -1;
+      prof_info.end_line_no = -1;
+      prof_info.func_line_no = -1;
+      prof_info.func_end_line_no = -1;
     }
-
-  /* Determine whether "finalize" semantics apply to all mappings of this
-     OpenACC directive.  */
-  bool finalize = false;
-  if (mapnum > 0)
-    {
-      unsigned char kind = kinds[0] & 0xff;
-      if (kind == GOMP_MAP_DELETE
-         || kind == GOMP_MAP_FORCE_FROM)
-       finalize = true;
-    }
-
-  /* Determine if this is an "acc enter data".  */
-  for (i = 0; i < mapnum; ++i)
+  acc_event_info exit_data_event_info;
+  if (profiling_p)
     {
-      unsigned char kind = kinds[i] & 0xff;
-
-      if (kind == GOMP_MAP_POINTER || kind == GOMP_MAP_TO_PSET)
-       continue;
-
-      if (kind == GOMP_MAP_FORCE_ALLOC
-         || kind == GOMP_MAP_FORCE_PRESENT
-         || kind == GOMP_MAP_FORCE_TO
-         || kind == GOMP_MAP_TO
-         || kind == GOMP_MAP_ALLOC)
-       {
-         data_enter = true;
-         break;
-       }
-
-      if (kind == GOMP_MAP_RELEASE
-         || kind == GOMP_MAP_DELETE
-         || kind == GOMP_MAP_FROM
-         || kind == GOMP_MAP_FORCE_FROM)
-       break;
-
-      gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
-                     kind);
+      exit_data_event_info.other_event.event_type
+       = prof_info.event_type;
+      exit_data_event_info.other_event.valid_bytes
+       = _ACC_OTHER_EVENT_INFO_VALID_BYTES;
+      exit_data_event_info.other_event.parent_construct = acc_construct_data;
+      exit_data_event_info.other_event.implicit = 0;
+      exit_data_event_info.other_event.tool_info = NULL;
     }
-
-  /* In c, non-pointers and arrays are represented by a single data clause.
-     Dynamically allocated arrays and subarrays are represented by a data
-     clause followed by an internal GOMP_MAP_POINTER.
-
-     In fortran, scalars and not allocated arrays are represented by a
-     single data clause. Allocated arrays and subarrays have three mappings:
-     1) the original data clause, 2) a PSET 3) a pointer to the array data.
-  */
-
-  if (data_enter)
+  acc_api_info api_info;
+  if (profiling_p)
     {
-      for (i = 0; i < mapnum; i++)
-       {
-         unsigned char kind = kinds[i] & 0xff;
-
-         /* Scan for pointers and PSETs.  */
-         int pointer = find_pointer (i, mapnum, kinds);
-
-         if (!pointer)
-           {
-             switch (kind)
-               {
-               case GOMP_MAP_ALLOC:
-               case GOMP_MAP_FORCE_ALLOC:
-                 acc_create_async (hostaddrs[i], sizes[i], async);
-                 break;
-               case GOMP_MAP_TO:
-               case GOMP_MAP_FORCE_TO:
-                 acc_copyin_async (hostaddrs[i], sizes[i], async);
-                 break;
-               default:
-                 gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
-                             kind);
-                 break;
-               }
-           }
-         else
-           {
-             gomp_acc_insert_pointer (pointer, &hostaddrs[i],
-                                      &sizes[i], &kinds[i], async);
-             /* Increment 'i' by two because OpenACC requires fortran
-                arrays to be contiguous, so each PSET is associated with
-                one of MAP_FORCE_ALLOC/MAP_FORCE_PRESET/MAP_FORCE_TO, and
-                one MAP_POINTER.  */
-             i += pointer - 1;
-           }
-       }
+      thr->api_info = &api_info;
+
+      api_info.device_api = acc_device_api_none;
+      api_info.valid_bytes = _ACC_API_INFO_VALID_BYTES;
+      api_info.device_type = prof_info.device_type;
+      api_info.vendor = -1;
+      api_info.device_handle = NULL;
+      api_info.context_handle = NULL;
+      api_info.async_handle = NULL;
     }
-  else
-    for (i = 0; i < mapnum; ++i)
-      {
-       unsigned char kind = kinds[i] & 0xff;
 
-       int pointer = find_pointer (i, mapnum, kinds);
+  if (profiling_p)
+    goacc_profiling_dispatch (&prof_info, &exit_data_event_info, &api_info);
 
-       if (!pointer)
-         {
-           switch (kind)
-             {
-             case GOMP_MAP_RELEASE:
-             case GOMP_MAP_DELETE:
-               if (acc_is_present (hostaddrs[i], sizes[i]))
-                 {
-                   if (finalize)
-                     acc_delete_finalize_async (hostaddrs[i], sizes[i], async);
-                   else
-                     acc_delete_async (hostaddrs[i], sizes[i], async);
-                 }
-               break;
-             case GOMP_MAP_FROM:
-             case GOMP_MAP_FORCE_FROM:
-               if (finalize)
-                 acc_copyout_finalize_async (hostaddrs[i], sizes[i], async);
-               else
-                 acc_copyout_async (hostaddrs[i], sizes[i], async);
-               break;
-             default:
-               gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
-                           kind);
-               break;
-             }
-         }
-       else
-         {
-           bool copyfrom = (kind == GOMP_MAP_FORCE_FROM
-                            || kind == GOMP_MAP_FROM);
-           gomp_acc_remove_pointer (hostaddrs[i], sizes[i], copyfrom, async,
-                                    finalize, pointer);
-           /* See the above comment.  */
-           i += pointer - 1;
-         }
-      }
-}
+  gomp_debug (0, "  %s: restore mappings\n", __FUNCTION__);
+  thr->mapped_data = tgt->prev;
+  goacc_unmap_vars (tgt, true, NULL);
+  gomp_debug (0, "  %s: mappings restored\n", __FUNCTION__);
 
-static void
-goacc_wait (int async, int num_waits, va_list *ap)
-{
-  while (num_waits--)
+  if (profiling_p)
     {
-      int qid = va_arg (*ap, int);
-
-      /* Waiting on ACC_ASYNC_NOVAL maps to 'wait all'.  */
-      if (qid == acc_async_noval)
-       {
-         if (async == acc_async_sync)
-           acc_wait_all ();
-         else
-           acc_wait_all_async (async);
-         break;
-       }
+      prof_info.event_type = acc_ev_exit_data_end;
+      exit_data_event_info.other_event.event_type = prof_info.event_type;
+      goacc_profiling_dispatch (&prof_info, &exit_data_event_info, &api_info);
 
-      if (acc_async_test (qid))
-       continue;
-
-      if (async == acc_async_sync)
-       acc_wait (qid);
-      else if (qid == async)
-       /* If we're waiting on the same asynchronous queue as we're
-          launching on, the queue itself will order work as
-          required, so there's no need to wait explicitly.  */
-       ;
-      else
-       acc_wait_async (qid, async);
+      thr->prof_info = NULL;
+      thr->api_info = NULL;
     }
 }
 
@@ -534,9 +567,64 @@ GOACC_update (int flags_m, size_t mapnum,
   struct goacc_thread *thr = goacc_thread ();
   struct gomp_device_descr *acc_dev = thr->dev;
 
+  bool profiling_p = GOACC_PROFILING_DISPATCH_P (true);
+
+  acc_prof_info prof_info;
+  if (profiling_p)
+    {
+      thr->prof_info = &prof_info;
+
+      prof_info.event_type = acc_ev_update_start;
+      prof_info.valid_bytes = _ACC_PROF_INFO_VALID_BYTES;
+      prof_info.version = _ACC_PROF_INFO_VERSION;
+      prof_info.device_type = acc_device_type (acc_dev->type);
+      prof_info.device_number = acc_dev->target_id;
+      prof_info.thread_id = -1;
+      prof_info.async = async;
+      prof_info.async_queue = prof_info.async;
+      prof_info.src_file = NULL;
+      prof_info.func_name = NULL;
+      prof_info.line_no = -1;
+      prof_info.end_line_no = -1;
+      prof_info.func_line_no = -1;
+      prof_info.func_end_line_no = -1;
+    }
+  acc_event_info update_event_info;
+  if (profiling_p)
+    {
+      update_event_info.other_event.event_type
+       = prof_info.event_type;
+      update_event_info.other_event.valid_bytes
+       = _ACC_OTHER_EVENT_INFO_VALID_BYTES;
+      update_event_info.other_event.parent_construct = acc_construct_update;
+      update_event_info.other_event.implicit = 0;
+      update_event_info.other_event.tool_info = NULL;
+    }
+  acc_api_info api_info;
+  if (profiling_p)
+    {
+      thr->api_info = &api_info;
+
+      api_info.device_api = acc_device_api_none;
+      api_info.valid_bytes = _ACC_API_INFO_VALID_BYTES;
+      api_info.device_type = prof_info.device_type;
+      api_info.vendor = -1;
+      api_info.device_handle = NULL;
+      api_info.context_handle = NULL;
+      api_info.async_handle = NULL;
+    }
+
+  if (profiling_p)
+    goacc_profiling_dispatch (&prof_info, &update_event_info, &api_info);
+
   if ((acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
       || (flags & GOACC_FLAG_HOST_FALLBACK))
-    return;
+    {
+      prof_info.device_type = acc_device_host;
+      api_info.device_type = prof_info.device_type;
+
+      goto out_prof;
+    }
 
   if (num_waits)
     {
@@ -608,25 +696,20 @@ GOACC_update (int flags_m, size_t mapnum,
          break;
        }
     }
-}
 
-void
-GOACC_wait (int async, int num_waits, ...)
-{
-  if (num_waits)
+ out_prof:
+  if (profiling_p)
     {
-      va_list ap;
+      prof_info.event_type = acc_ev_update_end;
+      update_event_info.other_event.event_type = prof_info.event_type;
+      goacc_profiling_dispatch (&prof_info, &update_event_info, &api_info);
 
-      va_start (ap, num_waits);
-      goacc_wait (async, num_waits, &ap);
-      va_end (ap);
+      thr->prof_info = NULL;
+      thr->api_info = NULL;
     }
-  else if (async == acc_async_sync)
-    acc_wait_all ();
-  else
-    acc_wait_all_async (async);
 }
 
+
 /* Legacy entry point (GCC 5).  */
 
 int
@@ -642,61 +725,3 @@ GOACC_get_thread_num (void)
 {
   return 0;
 }
-
-void
-GOACC_declare (int flags_m, size_t mapnum,
-              void **hostaddrs, size_t *sizes, unsigned short *kinds)
-{
-  int i;
-
-  for (i = 0; i < mapnum; i++)
-    {
-      unsigned char kind = kinds[i] & 0xff;
-
-      if (kind == GOMP_MAP_POINTER || kind == GOMP_MAP_TO_PSET)
-       continue;
-
-      switch (kind)
-       {
-         case GOMP_MAP_FORCE_ALLOC:
-         case GOMP_MAP_FORCE_FROM:
-         case GOMP_MAP_FORCE_TO:
-         case GOMP_MAP_POINTER:
-         case GOMP_MAP_RELEASE:
-         case GOMP_MAP_DELETE:
-           GOACC_enter_exit_data (flags_m, 1, &hostaddrs[i], &sizes[i],
-                                  &kinds[i], GOMP_ASYNC_SYNC, 0);
-           break;
-
-         case GOMP_MAP_FORCE_DEVICEPTR:
-           break;
-
-         case GOMP_MAP_ALLOC:
-           if (!acc_is_present (hostaddrs[i], sizes[i]))
-             GOACC_enter_exit_data (flags_m, 1, &hostaddrs[i], &sizes[i],
-                                    &kinds[i], GOMP_ASYNC_SYNC, 0);
-           break;
-
-         case GOMP_MAP_TO:
-           GOACC_enter_exit_data (flags_m, 1, &hostaddrs[i], &sizes[i],
-                                  &kinds[i], GOMP_ASYNC_SYNC, 0);
-
-           break;
-
-         case GOMP_MAP_FROM:
-           GOACC_enter_exit_data (flags_m, 1, &hostaddrs[i], &sizes[i],
-                                  &kinds[i], GOMP_ASYNC_SYNC, 0);
-           break;
-
-         case GOMP_MAP_FORCE_PRESENT:
-           if (!acc_is_present (hostaddrs[i], sizes[i]))
-             gomp_fatal ("[%p,%ld] is not mapped", hostaddrs[i],
-                         (unsigned long) sizes[i]);
-           break;
-
-         default:
-           assert (0);
-           break;
-       }
-    }
-}