[PATCH, OpenACC, libgomp, v6, stage1] Async-rework update

Chung-Lin Tang chunglin_tang@mentor.com
Mon Feb 25 18:15:00 GMT 2019


Hi Thomas,
I have incorporated all your patches you've included in the last mail (with
some modifications, though pretty minor I think).

The default_async, GOMP_PLUGIN_IF_VERSION, and testsuite changes have all been removed.
We can work on them later as we clarify more things.

Thanks,
Chung-Lin
-------------- next part --------------
Index: libgomp/oacc-async.c
===================================================================
--- libgomp/oacc-async.c	(revision 269183)
+++ libgomp/oacc-async.c	(working copy)
@@ -27,49 +27,162 @@
    <http://www.gnu.org/licenses/>.  */
 
 #include <assert.h>
+#include <string.h>
 #include "openacc.h"
 #include "libgomp.h"
 #include "oacc-int.h"
 
-int
-acc_async_test (int async)
+static struct goacc_thread *
+get_goacc_thread (void)
 {
-  if (!async_valid_p (async))
-    gomp_fatal ("invalid async argument: %d", async);
-
   struct goacc_thread *thr = goacc_thread ();
 
   if (!thr || !thr->dev)
     gomp_fatal ("no device active");
 
-  return thr->dev->openacc.async_test_func (async);
+  return thr;
 }
 
-int
-acc_async_test_all (void)
+static struct gomp_device_descr *
+get_goacc_thread_device (void)
 {
   struct goacc_thread *thr = goacc_thread ();
 
   if (!thr || !thr->dev)
     gomp_fatal ("no device active");
 
-  return thr->dev->openacc.async_test_all_func ();
+  return thr->dev;
 }
 
-void
-acc_wait (int async)
+static int
+validate_async_val (int async)
 {
   if (!async_valid_p (async))
-    gomp_fatal ("invalid async argument: %d", async);
+    gomp_fatal ("invalid async-argument: %d", async);
 
+  if (async == acc_async_sync)
+    return -1;
+
+  if (async == acc_async_noval)
+    return 0;
+
+  if (async >= 0)
+    /* TODO: we reserve 0 for acc_async_noval before we can clarify the
+       semantics of "default_async".  */
+    return 1 + async;
+  else
+    __builtin_unreachable ();
+}
+
+/* Return the asyncqueue to be used for OpenACC async-argument ASYNC.  This
+   might return NULL if no asyncqueue is to be used.  Otherwise, if CREATE,
+   create the asyncqueue if it doesn't exist yet.  */
+
+attribute_hidden struct goacc_asyncqueue *
+lookup_goacc_asyncqueue (struct goacc_thread *thr, bool create, int async)
+{
+  async = validate_async_val (async);
+  if (async < 0)
+    return NULL;
+
+  struct goacc_asyncqueue *ret_aq = NULL;
+  struct gomp_device_descr *dev = thr->dev;
+
+  gomp_mutex_lock (&dev->openacc.async.lock);
+
+  if (!create
+      && (async >= dev->openacc.async.nasyncqueue
+	  || !dev->openacc.async.asyncqueue[async]))
+    goto end;
+
+  if (async >= dev->openacc.async.nasyncqueue)
+    {
+      int diff = async + 1 - dev->openacc.async.nasyncqueue;
+      dev->openacc.async.asyncqueue
+	= gomp_realloc (dev->openacc.async.asyncqueue,
+			sizeof (goacc_aq) * (async + 1));
+      memset (dev->openacc.async.asyncqueue + dev->openacc.async.nasyncqueue,
+	      0, sizeof (goacc_aq) * diff);
+      dev->openacc.async.nasyncqueue = async + 1;
+    }
+
+  if (!dev->openacc.async.asyncqueue[async])
+    {
+      dev->openacc.async.asyncqueue[async] = dev->openacc.async.construct_func ();
+
+      if (!dev->openacc.async.asyncqueue[async])
+	{
+	  gomp_mutex_unlock (&dev->openacc.async.lock);
+	  gomp_fatal ("async %d creation failed", async);
+	}
+      
+      /* Link new async queue into active list.  */
+      goacc_aq_list n = gomp_malloc (sizeof (struct goacc_asyncqueue_list));
+      n->aq = dev->openacc.async.asyncqueue[async];
+      n->next = dev->openacc.async.active;
+      dev->openacc.async.active = n;
+    }
+
+  ret_aq = dev->openacc.async.asyncqueue[async];
+
+ end:
+  gomp_mutex_unlock (&dev->openacc.async.lock);
+  return ret_aq;
+}
+
+/* Return the asyncqueue to be used for OpenACC async-argument ASYNC.  This
+   might return NULL if no asyncqueue is to be used.  Otherwise, create the
+   asyncqueue if it doesn't exist yet.  */
+
+attribute_hidden struct goacc_asyncqueue *
+get_goacc_asyncqueue (int async)
+{
+  struct goacc_thread *thr = get_goacc_thread ();
+  return lookup_goacc_asyncqueue (thr, true, async);
+}
+
+int
+acc_async_test (int async)
+{
   struct goacc_thread *thr = goacc_thread ();
 
   if (!thr || !thr->dev)
     gomp_fatal ("no device active");
 
-  thr->dev->openacc.async_wait_func (async);
+  goacc_aq aq = lookup_goacc_asyncqueue (thr, false, async);
+  if (!aq)
+    return 1;
+  else
+    return thr->dev->openacc.async.test_func (aq);
 }
 
+int
+acc_async_test_all (void)
+{
+  struct goacc_thread *thr = get_goacc_thread ();
+
+  int ret = 1;
+  gomp_mutex_lock (&thr->dev->openacc.async.lock);
+  for (goacc_aq_list l = thr->dev->openacc.async.active; l; l = l->next)
+    if (!thr->dev->openacc.async.test_func (l->aq))
+      {
+	ret = 0;
+	break;
+      }
+  gomp_mutex_unlock (&thr->dev->openacc.async.lock);
+  return ret;
+}
+
+void
+acc_wait (int async)
+{
+  struct goacc_thread *thr = get_goacc_thread ();
+
+  goacc_aq aq = lookup_goacc_asyncqueue (thr, false, async);
+  if (aq && !thr->dev->openacc.async.synchronize_func (aq))
+    gomp_fatal ("wait on %d failed", async);
+}
+
 /* acc_async_wait is an OpenACC 1.0 compatibility name for acc_wait.  */
 #ifdef HAVE_ATTRIBUTE_ALIAS
 strong_alias (acc_wait, acc_async_wait)
@@ -84,23 +197,46 @@ acc_async_wait (int async)
 void
 acc_wait_async (int async1, int async2)
 {
-  struct goacc_thread *thr = goacc_thread ();
+  struct goacc_thread *thr = get_goacc_thread ();
 
-  if (!thr || !thr->dev)
-    gomp_fatal ("no device active");
+  goacc_aq aq1 = lookup_goacc_asyncqueue (thr, false, async1);
+  /* TODO: Is this also correct for acc_async_sync, assuming that in this case,
+     we'll always be synchronous anyways?  */
+  if (!aq1)
+    return;
 
-  thr->dev->openacc.async_wait_async_func (async1, async2);
+  goacc_aq aq2 = lookup_goacc_asyncqueue (thr, true, async2);
+  /* An async queue is always synchronized with itself.  */
+  if (aq1 == aq2)
+    return;
+
+  if (aq2)
+    {
+      if (!thr->dev->openacc.async.serialize_func (aq1, aq2))
+	gomp_fatal ("ordering of async ids %d and %d failed", async1, async2);
+    }
+  else
+    {
+      /* TODO: Local thread synchronization.
+	 Necessary for the "async2 == acc_async_sync" case, or can just skip?  */
+      if (!thr->dev->openacc.async.synchronize_func (aq1))
+	gomp_fatal ("wait on %d failed", async1);
+    }
 }
 
 void
 acc_wait_all (void)
 {
-  struct goacc_thread *thr = goacc_thread ();
+  struct gomp_device_descr *dev = get_goacc_thread_device ();
 
-  if (!thr || !thr->dev)
-    gomp_fatal ("no device active");
+  bool ret = true;
+  gomp_mutex_lock (&dev->openacc.async.lock);
+  for (goacc_aq_list l = dev->openacc.async.active; l; l = l->next)
+    ret &= dev->openacc.async.synchronize_func (l->aq);
+  gomp_mutex_unlock (&dev->openacc.async.lock);
 
-  thr->dev->openacc.async_wait_all_func ();
+  if (!ret)
+    gomp_fatal ("wait all failed");
 }
 
 /* acc_async_wait_all is an OpenACC 1.0 compatibility name for acc_wait_all.  */
@@ -117,13 +253,73 @@ acc_async_wait_all (void)
 void
 acc_wait_all_async (int async)
 {
-  if (!async_valid_p (async))
-    gomp_fatal ("invalid async argument: %d", async);
+  struct goacc_thread *thr = get_goacc_thread ();
 
-  struct goacc_thread *thr = goacc_thread ();
+  goacc_aq waiting_queue = lookup_goacc_asyncqueue (thr, true, async);
 
-  if (!thr || !thr->dev)
-    gomp_fatal ("no device active");
+  bool ret = true;
+  gomp_mutex_lock (&thr->dev->openacc.async.lock);
+  for (goacc_aq_list l = thr->dev->openacc.async.active; l; l = l->next)
+    {
+      if (waiting_queue)
+	ret &= thr->dev->openacc.async.serialize_func (l->aq, waiting_queue);
+      else
+	/* TODO: Local thread synchronization.
+	   Necessary for the "async2 == acc_async_sync" case, or can just skip?  */
+	ret &= thr->dev->openacc.async.synchronize_func (l->aq);
+    }
+  gomp_mutex_unlock (&thr->dev->openacc.async.lock);
 
-  thr->dev->openacc.async_wait_all_async_func (async);
+  if (!ret)
+    gomp_fatal ("wait all async(%d) failed", async);
 }
+
+attribute_hidden void
+goacc_async_free (struct gomp_device_descr *devicep,
+		  struct goacc_asyncqueue *aq, void *ptr)
+{
+  if (!aq)
+    free (ptr);
+  else
+    devicep->openacc.async.queue_callback_func (aq, free, ptr);
+}
+
+/* This function initializes the asyncqueues for the device specified by
+   DEVICEP.  TODO DEVICEP must be locked on entry, and remains locked on
+   return.  */
+
+attribute_hidden void
+goacc_init_asyncqueues (struct gomp_device_descr *devicep)
+{
+  devicep->openacc.async.nasyncqueue = 0;
+  devicep->openacc.async.asyncqueue = NULL;
+  devicep->openacc.async.active = NULL;
+  gomp_mutex_init (&devicep->openacc.async.lock);
+}
+
+/* This function finalizes the asyncqueues for the device specified by DEVICEP.
+   TODO DEVICEP must be locked on entry, and remains locked on return.  */
+
+attribute_hidden bool
+goacc_fini_asyncqueues (struct gomp_device_descr *devicep)
+{
+  bool ret = true;
+  gomp_mutex_lock (&devicep->openacc.async.lock);
+  if (devicep->openacc.async.nasyncqueue > 0)
+    {
+      goacc_aq_list next;
+      for (goacc_aq_list l = devicep->openacc.async.active; l; l = next)
+	{
+	  ret &= devicep->openacc.async.destruct_func (l->aq);
+	  next = l->next;
+	  free (l);
+	}
+      free (devicep->openacc.async.asyncqueue);
+      devicep->openacc.async.nasyncqueue = 0;
+      devicep->openacc.async.asyncqueue = NULL;
+      devicep->openacc.async.active = NULL;
+    }
+  gomp_mutex_unlock (&devicep->openacc.async.lock);
+  gomp_mutex_destroy (&devicep->openacc.async.lock);
+  return ret;
+}
Index: libgomp/oacc-plugin.c
===================================================================
--- libgomp/oacc-plugin.c	(revision 269183)
+++ libgomp/oacc-plugin.c	(working copy)
@@ -30,15 +30,12 @@
 #include "oacc-plugin.h"
 #include "oacc-int.h"
 
+/* This plugin function is now obsolete.  */
 void
-GOMP_PLUGIN_async_unmap_vars (void *ptr, int async)
+GOMP_PLUGIN_async_unmap_vars (void *ptr __attribute__((unused)),
+			      int async __attribute__((unused)))
 {
-  struct target_mem_desc *tgt = ptr;
-  struct gomp_device_descr *devicep = tgt->device_descr;
-
-  devicep->openacc.async_set_async_func (async);
-  gomp_unmap_vars (tgt, true);
-  devicep->openacc.async_set_async_func (acc_async_sync);
+  gomp_fatal ("invalid plugin function");
 }
 
 /* Return the target-specific part of the TLS data for the current thread.  */
Index: libgomp/plugin/cuda/cuda.h
===================================================================
--- libgomp/plugin/cuda/cuda.h	(revision 269183)
+++ libgomp/plugin/cuda/cuda.h	(working copy)
@@ -54,7 +54,11 @@ typedef enum {
   CUDA_ERROR_INVALID_CONTEXT = 201,
   CUDA_ERROR_NOT_FOUND = 500,
   CUDA_ERROR_NOT_READY = 600,
-  CUDA_ERROR_LAUNCH_FAILED = 719
+  CUDA_ERROR_LAUNCH_FAILED = 719,
+  CUDA_ERROR_COOPERATIVE_LAUNCH_TOO_LARGE = 720,
+  CUDA_ERROR_NOT_PERMITTED = 800,
+  CUDA_ERROR_NOT_SUPPORTED = 801,
+  CUDA_ERROR_UNKNOWN = 999
 } CUresult;
 
 typedef enum {
@@ -173,6 +177,8 @@ CUresult cuModuleLoadData (CUmodule *, const void
 CUresult cuModuleUnload (CUmodule);
 CUresult cuOccupancyMaxPotentialBlockSize(int *, int *, CUfunction,
 					  CUoccupancyB2DSize, size_t, int);
+typedef void (*CUstreamCallback)(CUstream, CUresult, void *);
+CUresult cuStreamAddCallback(CUstream, CUstreamCallback, void *, unsigned int);
 CUresult cuStreamCreate (CUstream *, unsigned);
 #define cuStreamDestroy cuStreamDestroy_v2
 CUresult cuStreamDestroy (CUstream);
Index: libgomp/plugin/plugin-nvptx.c
===================================================================
--- libgomp/plugin/plugin-nvptx.c	(revision 269183)
+++ libgomp/plugin/plugin-nvptx.c	(working copy)
@@ -192,175 +192,30 @@ cuda_error (CUresult r)
 static unsigned int instantiated_devices = 0;
 static pthread_mutex_t ptx_dev_lock = PTHREAD_MUTEX_INITIALIZER;
 
-struct cuda_map
+/* NVPTX/CUDA specific definition of asynchronous queues.  */
+struct goacc_asyncqueue
 {
-  CUdeviceptr d;
-  size_t size;
-  bool active;
-  struct cuda_map *next;
+  CUstream cuda_stream;
 };
 
-struct ptx_stream
+struct nvptx_callback
 {
-  CUstream stream;
-  pthread_t host_thread;
-  bool multithreaded;
-  struct cuda_map *map;
-  struct ptx_stream *next;
+  void (*fn) (void *);
+  void *ptr;
+  struct goacc_asyncqueue *aq;
+  struct nvptx_callback *next;
 };
 
 /* Thread-specific data for PTX.  */
 
 struct nvptx_thread
 {
-  struct ptx_stream *current_stream;
+  /* We currently have this embedded inside the plugin because libgomp manages
+     devices through integer target_ids.  This might be better if using an
+     opaque target-specific pointer directly from gomp_device_descr.  */
   struct ptx_device *ptx_dev;
 };
 
-static struct cuda_map *
-cuda_map_create (size_t size)
-{
-  struct cuda_map *map = GOMP_PLUGIN_malloc (sizeof (struct cuda_map));
-
-  assert (map);
-
-  map->next = NULL;
-  map->size = size;
-  map->active = false;
-
-  CUDA_CALL_ERET (NULL, cuMemAlloc, &map->d, size);
-  assert (map->d);
-
-  return map;
-}
-
-static void
-cuda_map_destroy (struct cuda_map *map)
-{
-  if (map->active)
-    /* Possible reasons for the map to be still active:
-       - the associated async kernel might still be running.
-       - the associated async kernel might have finished, but the
-         corresponding event that should trigger the pop_map has not been
-	 processed by event_gc.
-       - the associated sync kernel might have aborted
-
-       The async cases could happen if the user specified an async region
-       without adding a corresponding wait that is guaranteed to be executed
-       (before returning from main, or in an atexit handler).
-       We do not want to deallocate a device pointer that is still being
-       used, so skip it.
-
-       In the sync case, the device pointer is no longer used, but deallocating
-       it using cuMemFree will not succeed, so skip it.
-
-       TODO: Handle this in a more constructive way, by f.i. waiting for streams
-       to finish before de-allocating them (PR88981), or by ensuring the CUDA
-       lib atexit handler is called before rather than after the libgomp plugin
-       atexit handler (PR83795).  */
-    ;
-  else
-    CUDA_CALL_NOCHECK (cuMemFree, map->d);
-
-  free (map);
-}
-
-/* The following map_* routines manage the CUDA device memory that
-   contains the data mapping arguments for cuLaunchKernel.  Each
-   asynchronous PTX stream may have multiple pending kernel
-   invocations, which are launched in a FIFO order.  As such, the map
-   routines maintains a queue of cuLaunchKernel arguments.
-
-   Calls to map_push and map_pop must be guarded by ptx_event_lock.
-   Likewise, calls to map_init and map_fini are guarded by
-   ptx_dev_lock inside GOMP_OFFLOAD_init_device and
-   GOMP_OFFLOAD_fini_device, respectively.  */
-
-static bool
-map_init (struct ptx_stream *s)
-{
-  int size = getpagesize ();
-
-  assert (s);
-
-  s->map = cuda_map_create (size);
-
-  return true;
-}
-
-static bool
-map_fini (struct ptx_stream *s)
-{
-  assert (s->map->next == NULL);
-
-  cuda_map_destroy (s->map);
-
-  return true;
-}
-
-static void
-map_pop (struct ptx_stream *s)
-{
-  struct cuda_map *next;
-
-  assert (s != NULL);
-
-  if (s->map->next == NULL)
-    {
-      s->map->active = false;
-      return;
-    }
-
-  next = s->map->next;
-  cuda_map_destroy (s->map);
-  s->map = next;
-}
-
-static CUdeviceptr
-map_push (struct ptx_stream *s, size_t size)
-{
-  struct cuda_map *map = NULL;
-  struct cuda_map **t;
-
-  assert (s);
-  assert (s->map);
-
-  /* Select an element to push.  */
-  if (s->map->active)
-    map = cuda_map_create (size);
-  else
-    {
-      /* Pop the inactive front element.  */
-      struct cuda_map *pop = s->map;
-      s->map = pop->next;
-      pop->next = NULL;
-
-      if (pop->size < size)
-	{
-	  cuda_map_destroy (pop);
-
-	  map = cuda_map_create (size);
-	}
-      else
-	map = pop;
-    }
-
-  /* Check that the element is as expected.  */
-  assert (map->next == NULL);
-  assert (!map->active);
-
-  /* Mark the element active.  */
-  map->active = true;
-
-  /* Push the element to the back of the list.  */
-  for (t = &s->map; (*t) != NULL; t = &(*t)->next)
-    ;
-  assert (t != NULL && *t == NULL);
-  *t = map;
-
-  return map->d;
-}
-
 /* Target data function launch information.  */
 
 struct targ_fn_launch
@@ -412,22 +267,18 @@ struct ptx_image_data
   struct ptx_image_data *next;
 };
 
+struct ptx_free_block
+{
+  void *ptr;
+  struct ptx_free_block *next;
+};
+
 struct ptx_device
 {
   CUcontext ctx;
   bool ctx_shared;
   CUdevice dev;
-  struct ptx_stream *null_stream;
-  /* All non-null streams associated with this device (actually context),
-     either created implicitly or passed in from the user (via
-     acc_set_cuda_stream).  */
-  struct ptx_stream *active_streams;
-  struct {
-    struct ptx_stream **arr;
-    int size;
-  } async_streams;
-  /* A lock for use when manipulating the above stream list and array.  */
-  pthread_mutex_t stream_lock;
+
   int ord;
   bool overlap;
   bool map;
@@ -445,32 +296,13 @@ struct ptx_device
 
   struct ptx_image_data *images;  /* Images loaded on device.  */
   pthread_mutex_t image_lock;     /* Lock for above list.  */
-  
-  struct ptx_device *next;
-};
 
-enum ptx_event_type
-{
-  PTX_EVT_MEM,
-  PTX_EVT_KNL,
-  PTX_EVT_SYNC,
-  PTX_EVT_ASYNC_CLEANUP
-};
+  struct ptx_free_block *free_blocks;
+  pthread_mutex_t free_blocks_lock;
 
-struct ptx_event
-{
-  CUevent *evt;
-  int type;
-  void *addr;
-  int ord;
-  int val;
-
-  struct ptx_event *next;
+  struct ptx_device *next;
 };
 
-static pthread_mutex_t ptx_event_lock;
-static struct ptx_event *ptx_events;
-
 static struct ptx_device **ptx_devices;
 
 static inline struct nvptx_thread *
@@ -479,193 +311,6 @@ nvptx_thread (void)
   return (struct nvptx_thread *) GOMP_PLUGIN_acc_thread ();
 }
 
-static bool
-init_streams_for_device (struct ptx_device *ptx_dev, int concurrency)
-{
-  int i;
-  struct ptx_stream *null_stream
-    = GOMP_PLUGIN_malloc (sizeof (struct ptx_stream));
-
-  null_stream->stream = NULL;
-  null_stream->host_thread = pthread_self ();
-  null_stream->multithreaded = true;
-  if (!map_init (null_stream))
-    return false;
-
-  ptx_dev->null_stream = null_stream;
-  ptx_dev->active_streams = NULL;
-  pthread_mutex_init (&ptx_dev->stream_lock, NULL);
-
-  if (concurrency < 1)
-    concurrency = 1;
-
-  /* This is just a guess -- make space for as many async streams as the
-     current device is capable of concurrently executing.  This can grow
-     later as necessary.  No streams are created yet.  */
-  ptx_dev->async_streams.arr
-    = GOMP_PLUGIN_malloc (concurrency * sizeof (struct ptx_stream *));
-  ptx_dev->async_streams.size = concurrency;
-
-  for (i = 0; i < concurrency; i++)
-    ptx_dev->async_streams.arr[i] = NULL;
-
-  return true;
-}
-
-static bool
-fini_streams_for_device (struct ptx_device *ptx_dev)
-{
-  free (ptx_dev->async_streams.arr);
-
-  bool ret = true;
-  while (ptx_dev->active_streams != NULL)
-    {
-      struct ptx_stream *s = ptx_dev->active_streams;
-      ptx_dev->active_streams = ptx_dev->active_streams->next;
-
-      ret &= map_fini (s);
-
-      CUresult r = CUDA_CALL_NOCHECK (cuStreamDestroy, s->stream);
-      if (r != CUDA_SUCCESS)
-	{
-	  GOMP_PLUGIN_error ("cuStreamDestroy error: %s", cuda_error (r));
-	  ret = false;
-	}
-      free (s);
-    }
-
-  ret &= map_fini (ptx_dev->null_stream);
-  free (ptx_dev->null_stream);
-  return ret;
-}
-
-/* Select a stream for (OpenACC-semantics) ASYNC argument for the current
-   thread THREAD (and also current device/context).  If CREATE is true, create
-   the stream if it does not exist (or use EXISTING if it is non-NULL), and
-   associate the stream with the same thread argument.  Returns stream to use
-   as result.  */
-
-static struct ptx_stream *
-select_stream_for_async (int async, pthread_t thread, bool create,
-			 CUstream existing)
-{
-  struct nvptx_thread *nvthd = nvptx_thread ();
-  /* Local copy of TLS variable.  */
-  struct ptx_device *ptx_dev = nvthd->ptx_dev;
-  struct ptx_stream *stream = NULL;
-  int orig_async = async;
-
-  /* The special value acc_async_noval (-1) maps (for now) to an
-     implicitly-created stream, which is then handled the same as any other
-     numbered async stream.  Other options are available, e.g. using the null
-     stream for anonymous async operations, or choosing an idle stream from an
-     active set.  But, stick with this for now.  */
-  if (async > acc_async_sync)
-    async++;
-
-  if (create)
-    pthread_mutex_lock (&ptx_dev->stream_lock);
-
-  /* NOTE: AFAICT there's no particular need for acc_async_sync to map to the
-     null stream, and in fact better performance may be obtainable if it doesn't
-     (because the null stream enforces overly-strict synchronisation with
-     respect to other streams for legacy reasons, and that's probably not
-     needed with OpenACC).  Maybe investigate later.  */
-  if (async == acc_async_sync)
-    stream = ptx_dev->null_stream;
-  else if (async >= 0 && async < ptx_dev->async_streams.size
-	   && ptx_dev->async_streams.arr[async] && !(create && existing))
-    stream = ptx_dev->async_streams.arr[async];
-  else if (async >= 0 && create)
-    {
-      if (async >= ptx_dev->async_streams.size)
-	{
-	  int i, newsize = ptx_dev->async_streams.size * 2;
-
-	  if (async >= newsize)
-	    newsize = async + 1;
-
-	  ptx_dev->async_streams.arr
-	    = GOMP_PLUGIN_realloc (ptx_dev->async_streams.arr,
-				   newsize * sizeof (struct ptx_stream *));
-
-	  for (i = ptx_dev->async_streams.size; i < newsize; i++)
-	    ptx_dev->async_streams.arr[i] = NULL;
-
-	  ptx_dev->async_streams.size = newsize;
-	}
-
-      /* Create a new stream on-demand if there isn't one already, or if we're
-	 setting a particular async value to an existing (externally-provided)
-	 stream.  */
-      if (!ptx_dev->async_streams.arr[async] || existing)
-        {
-	  CUresult r;
-	  struct ptx_stream *s
-	    = GOMP_PLUGIN_malloc (sizeof (struct ptx_stream));
-
-	  if (existing)
-	    s->stream = existing;
-	  else
-	    {
-	      r = CUDA_CALL_NOCHECK (cuStreamCreate, &s->stream,
-				     CU_STREAM_DEFAULT);
-	      if (r != CUDA_SUCCESS)
-		{
-		  pthread_mutex_unlock (&ptx_dev->stream_lock);
-		  GOMP_PLUGIN_fatal ("cuStreamCreate error: %s",
-				     cuda_error (r));
-		}
-	    }
-
-	  /* If CREATE is true, we're going to be queueing some work on this
-	     stream.  Associate it with the current host thread.  */
-	  s->host_thread = thread;
-	  s->multithreaded = false;
-
-	  if (!map_init (s))
-	    {
-	      pthread_mutex_unlock (&ptx_dev->stream_lock);
-	      GOMP_PLUGIN_fatal ("map_init fail");
-	    }
-
-	  s->next = ptx_dev->active_streams;
-	  ptx_dev->active_streams = s;
-	  ptx_dev->async_streams.arr[async] = s;
-	}
-
-      stream = ptx_dev->async_streams.arr[async];
-    }
-  else if (async < 0)
-    {
-      if (create)
-	pthread_mutex_unlock (&ptx_dev->stream_lock);
-      GOMP_PLUGIN_fatal ("bad async %d", async);
-    }
-
-  if (create)
-    {
-      assert (stream != NULL);
-
-      /* If we're trying to use the same stream from different threads
-	 simultaneously, set stream->multithreaded to true.  This affects the
-	 behaviour of acc_async_test_all and acc_wait_all, which are supposed to
-	 only wait for asynchronous launches from the same host thread they are
-	 invoked on.  If multiple threads use the same async value, we make note
-	 of that here and fall back to testing/waiting for all threads in those
-	 functions.  */
-      if (thread != stream->host_thread)
-        stream->multithreaded = true;
-
-      pthread_mutex_unlock (&ptx_dev->stream_lock);
-    }
-  else if (stream && !stream->multithreaded
-	   && !pthread_equal (stream->host_thread, thread))
-    GOMP_PLUGIN_fatal ("async %d used on wrong thread", orig_async);
-
-  return stream;
-}
-
 /* Initialize the device.  Return TRUE on success, else FALSE.  PTX_DEV_LOCK
    should be locked on entry and remains locked on exit.  */
 
@@ -677,9 +322,6 @@ nvptx_init (void)
   if (instantiated_devices != 0)
     return true;
 
-  ptx_events = NULL;
-  pthread_mutex_init (&ptx_event_lock, NULL);
-
   if (!init_cuda_lib ())
     return false;
 
@@ -703,6 +345,11 @@ nvptx_attach_host_thread_to_device (int n)
   CUcontext thd_ctx;
 
   r = CUDA_CALL_NOCHECK (cuCtxGetDevice, &dev);
+  if (r == CUDA_ERROR_NOT_PERMITTED)
+    {
+      /* Assume we're in a CUDA callback, just return true.  */
+      return true;
+    }
   if (r != CUDA_SUCCESS && r != CUDA_ERROR_INVALID_CONTEXT)
     {
       GOMP_PLUGIN_error ("cuCtxGetDevice error: %s", cuda_error (r));
@@ -847,8 +494,8 @@ nvptx_open_device (int n)
   ptx_dev->images = NULL;
   pthread_mutex_init (&ptx_dev->image_lock, NULL);
 
-  if (!init_streams_for_device (ptx_dev, async_engines))
-    return NULL;
+  ptx_dev->free_blocks = NULL;
+  pthread_mutex_init (&ptx_dev->free_blocks_lock, NULL);
 
   return ptx_dev;
 }
@@ -859,9 +506,15 @@ nvptx_close_device (struct ptx_device *ptx_dev)
   if (!ptx_dev)
     return true;
 
-  if (!fini_streams_for_device (ptx_dev))
-    return false;
-  
+  for (struct ptx_free_block *b = ptx_dev->free_blocks; b;)
+    {
+      struct ptx_free_block *b_next = b->next;
+      CUDA_CALL (cuMemFree, (CUdeviceptr) b->ptr);
+      free (b);
+      b = b_next;
+    }
+
+  pthread_mutex_destroy (&ptx_dev->free_blocks_lock);
   pthread_mutex_destroy (&ptx_dev->image_lock);
 
   if (!ptx_dev->ctx_shared)
@@ -1041,139 +694,19 @@ link_ptx (CUmodule *module, const struct targ_ptx_
 }
 
 static void
-event_gc (bool memmap_lockable)
-{
-  struct ptx_event *ptx_event = ptx_events;
-  struct ptx_event *async_cleanups = NULL;
-  struct nvptx_thread *nvthd = nvptx_thread ();
-
-  pthread_mutex_lock (&ptx_event_lock);
-
-  while (ptx_event != NULL)
-    {
-      CUresult r;
-      struct ptx_event *e = ptx_event;
-
-      ptx_event = ptx_event->next;
-
-      if (e->ord != nvthd->ptx_dev->ord)
-	continue;
-
-      r = CUDA_CALL_NOCHECK (cuEventQuery, *e->evt);
-      if (r == CUDA_SUCCESS)
-	{
-	  bool append_async = false;
-	  CUevent *te;
-
-	  te = e->evt;
-
-	  switch (e->type)
-	    {
-	    case PTX_EVT_MEM:
-	    case PTX_EVT_SYNC:
-	      break;
-
-	    case PTX_EVT_KNL:
-	      map_pop (e->addr);
-	      break;
-
-	    case PTX_EVT_ASYNC_CLEANUP:
-	      {
-		/* The function gomp_plugin_async_unmap_vars needs to claim the
-		   memory-map splay tree lock for the current device, so we
-		   can't call it when one of our callers has already claimed
-		   the lock.  In that case, just delay the GC for this event
-		   until later.  */
-		if (!memmap_lockable)
-		  continue;
-
-		append_async = true;
-	      }
-	      break;
-	    }
-
-	  CUDA_CALL_NOCHECK (cuEventDestroy, *te);
-	  free ((void *)te);
-
-	  /* Unlink 'e' from ptx_events list.  */
-	  if (ptx_events == e)
-	    ptx_events = ptx_events->next;
-	  else
-	    {
-	      struct ptx_event *e_ = ptx_events;
-	      while (e_->next != e)
-		e_ = e_->next;
-	      e_->next = e_->next->next;
-	    }
-
-	  if (append_async)
-	    {
-	      e->next = async_cleanups;
-	      async_cleanups = e;
-	    }
-	  else
-	    free (e);
-	}
-    }
-
-  pthread_mutex_unlock (&ptx_event_lock);
-
-  /* We have to do these here, after ptx_event_lock is released.  */
-  while (async_cleanups)
-    {
-      struct ptx_event *e = async_cleanups;
-      async_cleanups = async_cleanups->next;
-
-      GOMP_PLUGIN_async_unmap_vars (e->addr, e->val);
-      free (e);
-    }
-}
-
-static void
-event_add (enum ptx_event_type type, CUevent *e, void *h, int val)
-{
-  struct ptx_event *ptx_event;
-  struct nvptx_thread *nvthd = nvptx_thread ();
-
-  assert (type == PTX_EVT_MEM || type == PTX_EVT_KNL || type == PTX_EVT_SYNC
-	  || type == PTX_EVT_ASYNC_CLEANUP);
-
-  ptx_event = GOMP_PLUGIN_malloc (sizeof (struct ptx_event));
-  ptx_event->type = type;
-  ptx_event->evt = e;
-  ptx_event->addr = h;
-  ptx_event->ord = nvthd->ptx_dev->ord;
-  ptx_event->val = val;
-
-  pthread_mutex_lock (&ptx_event_lock);
-
-  ptx_event->next = ptx_events;
-  ptx_events = ptx_event;
-
-  pthread_mutex_unlock (&ptx_event_lock);
-}
-
-static void
 nvptx_exec (void (*fn), size_t mapnum, void **hostaddrs, void **devaddrs,
-	    int async, unsigned *dims, void *targ_mem_desc)
+	    unsigned *dims, void *targ_mem_desc,
+	    CUdeviceptr dp, CUstream stream)
 {
   struct targ_fn_descriptor *targ_fn = (struct targ_fn_descriptor *) fn;
   CUfunction function;
-  CUresult r;
   int i;
-  struct ptx_stream *dev_str;
   void *kargs[1];
-  void *hp;
-  CUdeviceptr dp = 0;
   struct nvptx_thread *nvthd = nvptx_thread ();
   int warp_size = nvthd->ptx_dev->warp_size;
-  const char *maybe_abort_msg = "(perhaps abort was called)";
 
   function = targ_fn->fn;
 
-  dev_str = select_stream_for_async (async, pthread_self (), false, NULL);
-  assert (dev_str == nvthd->current_stream);
-
   /* Initialize the launch dimensions.  Typically this is constant,
      provided by the device compiler, but we must permit runtime
      values.  */
@@ -1361,27 +894,6 @@ nvptx_exec (void (*fn), size_t mapnum, void **host
 			   dims[GOMP_DIM_VECTOR]);
     }
 
-  if (mapnum > 0)
-    {
-      /* This reserves a chunk of a pre-allocated page of memory mapped on both
-	 the host and the device. HP is a host pointer to the new chunk, and DP is
-	 the corresponding device pointer.  */
-      pthread_mutex_lock (&ptx_event_lock);
-      dp = map_push (dev_str, mapnum * sizeof (void *));
-      pthread_mutex_unlock (&ptx_event_lock);
-
-      GOMP_PLUGIN_debug (0, "  %s: prepare mappings\n", __FUNCTION__);
-
-      /* Copy the array of arguments to the mapped page.  */
-      hp = alloca(sizeof(void *) * mapnum);
-      for (i = 0; i < mapnum; i++)
-	((void **) hp)[i] = devaddrs[i];
-
-      /* Copy the (device) pointers to arguments to the device */
-      CUDA_CALL_ASSERT (cuMemcpyHtoD, dp, hp,
-			mapnum * sizeof (void *));
-    }
-
   GOMP_PLUGIN_debug (0, "  %s: kernel %s: launch"
 		     " gangs=%u, workers=%u, vectors=%u\n",
 		     __FUNCTION__, targ_fn->launch->fn, dims[GOMP_DIM_GANG],
@@ -1392,62 +904,14 @@ nvptx_exec (void (*fn), size_t mapnum, void **host
   // num_gangs		nctaid.x
   // num_workers	ntid.y
   // vector length	ntid.x
-
   kargs[0] = &dp;
   CUDA_CALL_ASSERT (cuLaunchKernel, function,
 		    dims[GOMP_DIM_GANG], 1, 1,
 		    dims[GOMP_DIM_VECTOR], dims[GOMP_DIM_WORKER], 1,
-		    0, dev_str->stream, kargs, 0);
+		    0, stream, kargs, 0);
 
-#ifndef DISABLE_ASYNC
-  if (async < acc_async_noval)
-    {
-      r = CUDA_CALL_NOCHECK (cuStreamSynchronize, dev_str->stream);
-      if (r == CUDA_ERROR_LAUNCH_FAILED)
-	GOMP_PLUGIN_fatal ("cuStreamSynchronize error: %s %s\n", cuda_error (r),
-			   maybe_abort_msg);
-      else if (r != CUDA_SUCCESS)
-        GOMP_PLUGIN_fatal ("cuStreamSynchronize error: %s", cuda_error (r));
-    }
-  else
-    {
-      CUevent *e;
-
-      e = (CUevent *)GOMP_PLUGIN_malloc (sizeof (CUevent));
-
-      r = CUDA_CALL_NOCHECK (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
-      if (r == CUDA_ERROR_LAUNCH_FAILED)
-	GOMP_PLUGIN_fatal ("cuEventCreate error: %s %s\n", cuda_error (r),
-			   maybe_abort_msg);
-      else if (r != CUDA_SUCCESS)
-        GOMP_PLUGIN_fatal ("cuEventCreate error: %s", cuda_error (r));
-
-      event_gc (true);
-
-      CUDA_CALL_ASSERT (cuEventRecord, *e, dev_str->stream);
-
-      if (mapnum > 0)
-	event_add (PTX_EVT_KNL, e, (void *)dev_str, 0);
-    }
-#else
-  r = CUDA_CALL_NOCHECK (cuCtxSynchronize, );
-  if (r == CUDA_ERROR_LAUNCH_FAILED)
-    GOMP_PLUGIN_fatal ("cuCtxSynchronize error: %s %s\n", cuda_error (r),
-		       maybe_abort_msg);
-  else if (r != CUDA_SUCCESS)
-    GOMP_PLUGIN_fatal ("cuCtxSynchronize error: %s", cuda_error (r));
-#endif
-
   GOMP_PLUGIN_debug (0, "  %s: kernel %s: finished\n", __FUNCTION__,
 		     targ_fn->launch->fn);
-
-#ifndef DISABLE_ASYNC
-  if (async < acc_async_noval)
-#endif
-    {
-      if (mapnum > 0)
-	map_pop (dev_str);
-    }
 }
 
 void * openacc_get_current_cuda_context (void);
@@ -1462,8 +926,21 @@ nvptx_alloc (size_t s)
 }
 
 static bool
-nvptx_free (void *p)
+nvptx_free (void *p, struct ptx_device *ptx_dev)
 {
+  /* Assume callback context if this is null.  */
+  if (GOMP_PLUGIN_acc_thread () == NULL)
+    {
+      struct ptx_free_block *n
+	= GOMP_PLUGIN_malloc (sizeof (struct ptx_free_block));
+      n->ptr = p;
+      pthread_mutex_lock (&ptx_dev->free_blocks_lock);
+      n->next = ptx_dev->free_blocks;
+      ptx_dev->free_blocks = n;
+      pthread_mutex_unlock (&ptx_dev->free_blocks_lock);
+      return true;
+    }
+
   CUdeviceptr pb;
   size_t ps;
 
@@ -1478,305 +955,6 @@ static bool
   return true;
 }
 
-
-static bool
-nvptx_host2dev (void *d, const void *h, size_t s)
-{
-  CUdeviceptr pb;
-  size_t ps;
-  struct nvptx_thread *nvthd = nvptx_thread ();
-
-  if (!s)
-    return true;
-  if (!d)
-    {
-      GOMP_PLUGIN_error ("invalid device address");
-      return false;
-    }
-
-  CUDA_CALL (cuMemGetAddressRange, &pb, &ps, (CUdeviceptr) d);
-
-  if (!pb)
-    {
-      GOMP_PLUGIN_error ("invalid device address");
-      return false;
-    }
-  if (!h)
-    {
-      GOMP_PLUGIN_error ("invalid host address");
-      return false;
-    }
-  if (d == h)
-    {
-      GOMP_PLUGIN_error ("invalid host or device address");
-      return false;
-    }
-  if ((void *)(d + s) > (void *)(pb + ps))
-    {
-      GOMP_PLUGIN_error ("invalid size");
-      return false;
-    }
-
-#ifndef DISABLE_ASYNC
-  if (nvthd && nvthd->current_stream != nvthd->ptx_dev->null_stream)
-    {
-      CUevent *e = (CUevent *)GOMP_PLUGIN_malloc (sizeof (CUevent));
-      CUDA_CALL (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
-      event_gc (false);
-      CUDA_CALL (cuMemcpyHtoDAsync,
-		 (CUdeviceptr) d, h, s, nvthd->current_stream->stream);
-      CUDA_CALL (cuEventRecord, *e, nvthd->current_stream->stream);
-      event_add (PTX_EVT_MEM, e, (void *)h, 0);
-    }
-  else
-#endif
-    CUDA_CALL (cuMemcpyHtoD, (CUdeviceptr) d, h, s);
-
-  return true;
-}
-
-static bool
-nvptx_dev2host (void *h, const void *d, size_t s)
-{
-  CUdeviceptr pb;
-  size_t ps;
-  struct nvptx_thread *nvthd = nvptx_thread ();
-
-  if (!s)
-    return true;
-  if (!d)
-    {
-      GOMP_PLUGIN_error ("invalid device address");
-      return false;
-    }
-
-  CUDA_CALL (cuMemGetAddressRange, &pb, &ps, (CUdeviceptr) d);
-
-  if (!pb)
-    {
-      GOMP_PLUGIN_error ("invalid device address");
-      return false;
-    }
-  if (!h)
-    {
-      GOMP_PLUGIN_error ("invalid host address");
-      return false;
-    }
-  if (d == h)
-    {
-      GOMP_PLUGIN_error ("invalid host or device address");
-      return false;
-    }
-  if ((void *)(d + s) > (void *)(pb + ps))
-    {
-      GOMP_PLUGIN_error ("invalid size");
-      return false;
-    }
-
-#ifndef DISABLE_ASYNC
-  if (nvthd && nvthd->current_stream != nvthd->ptx_dev->null_stream)
-    {
-      CUevent *e = (CUevent *) GOMP_PLUGIN_malloc (sizeof (CUevent));
-      CUDA_CALL (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
-      event_gc (false);
-      CUDA_CALL (cuMemcpyDtoHAsync,
-		 h, (CUdeviceptr) d, s, nvthd->current_stream->stream);
-      CUDA_CALL (cuEventRecord, *e, nvthd->current_stream->stream);
-      event_add (PTX_EVT_MEM, e, (void *)h, 0);
-    }
-  else
-#endif
-    CUDA_CALL (cuMemcpyDtoH, h, (CUdeviceptr) d, s);
-
-  return true;
-}
-
-static void
-nvptx_set_async (int async)
-{
-  struct nvptx_thread *nvthd = nvptx_thread ();
-  nvthd->current_stream
-    = select_stream_for_async (async, pthread_self (), true, NULL);
-}
-
-static int
-nvptx_async_test (int async)
-{
-  CUresult r;
-  struct ptx_stream *s;
-
-  s = select_stream_for_async (async, pthread_self (), false, NULL);
-  if (!s)
-    return 1;
-
-  r = CUDA_CALL_NOCHECK (cuStreamQuery, s->stream);
-  if (r == CUDA_SUCCESS)
-    {
-      /* The oacc-parallel.c:goacc_wait function calls this hook to determine
-	 whether all work has completed on this stream, and if so omits the call
-	 to the wait hook.  If that happens, event_gc might not get called
-	 (which prevents variables from getting unmapped and their associated
-	 device storage freed), so call it here.  */
-      event_gc (true);
-      return 1;
-    }
-  else if (r == CUDA_ERROR_NOT_READY)
-    return 0;
-
-  GOMP_PLUGIN_fatal ("cuStreamQuery error: %s", cuda_error (r));
-
-  return 0;
-}
-
-static int
-nvptx_async_test_all (void)
-{
-  struct ptx_stream *s;
-  pthread_t self = pthread_self ();
-  struct nvptx_thread *nvthd = nvptx_thread ();
-
-  pthread_mutex_lock (&nvthd->ptx_dev->stream_lock);
-
-  for (s = nvthd->ptx_dev->active_streams; s != NULL; s = s->next)
-    {
-      if ((s->multithreaded || pthread_equal (s->host_thread, self))
-	  && CUDA_CALL_NOCHECK (cuStreamQuery,
-				s->stream) == CUDA_ERROR_NOT_READY)
-	{
-	  pthread_mutex_unlock (&nvthd->ptx_dev->stream_lock);
-	  return 0;
-	}
-    }
-
-  pthread_mutex_unlock (&nvthd->ptx_dev->stream_lock);
-
-  event_gc (true);
-
-  return 1;
-}
-
-static void
-nvptx_wait (int async)
-{
-  struct ptx_stream *s;
-
-  s = select_stream_for_async (async, pthread_self (), false, NULL);
-  if (!s)
-    return;
-
-  CUDA_CALL_ASSERT (cuStreamSynchronize, s->stream);
-
-  event_gc (true);
-}
-
-static void
-nvptx_wait_async (int async1, int async2)
-{
-  CUevent *e;
-  struct ptx_stream *s1, *s2;
-  pthread_t self = pthread_self ();
-
-  s1 = select_stream_for_async (async1, self, false, NULL);
-  if (!s1)
-    return;
-
-  /* The stream that is waiting (rather than being waited for) doesn't
-     necessarily have to exist already.  */
-  s2 = select_stream_for_async (async2, self, true, NULL);
-
-  /* A stream is always synchronized with itself.  */
-  if (s1 == s2)
-    return;
-
-  e = (CUevent *) GOMP_PLUGIN_malloc (sizeof (CUevent));
-
-  CUDA_CALL_ASSERT (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
-
-  event_gc (true);
-
-  CUDA_CALL_ASSERT (cuEventRecord, *e, s1->stream);
-
-  event_add (PTX_EVT_SYNC, e, NULL, 0);
-
-  CUDA_CALL_ASSERT (cuStreamWaitEvent, s2->stream, *e, 0);
-}
-
-static void
-nvptx_wait_all (void)
-{
-  CUresult r;
-  struct ptx_stream *s;
-  pthread_t self = pthread_self ();
-  struct nvptx_thread *nvthd = nvptx_thread ();
-
-  pthread_mutex_lock (&nvthd->ptx_dev->stream_lock);
-
-  /* Wait for active streams initiated by this thread (or by multiple threads)
-     to complete.  */
-  for (s = nvthd->ptx_dev->active_streams; s != NULL; s = s->next)
-    {
-      if (s->multithreaded || pthread_equal (s->host_thread, self))
-	{
-	  r = CUDA_CALL_NOCHECK (cuStreamQuery, s->stream);
-	  if (r == CUDA_SUCCESS)
-	    continue;
-	  else if (r != CUDA_ERROR_NOT_READY)
-	    GOMP_PLUGIN_fatal ("cuStreamQuery error: %s", cuda_error (r));
-
-	  CUDA_CALL_ASSERT (cuStreamSynchronize, s->stream);
-	}
-    }
-
-  pthread_mutex_unlock (&nvthd->ptx_dev->stream_lock);
-
-  event_gc (true);
-}
-
-static void
-nvptx_wait_all_async (int async)
-{
-  struct ptx_stream *waiting_stream, *other_stream;
-  CUevent *e;
-  struct nvptx_thread *nvthd = nvptx_thread ();
-  pthread_t self = pthread_self ();
-
-  /* The stream doing the waiting.  This could be the first mention of the
-     stream, so create it if necessary.  */
-  waiting_stream
-    = select_stream_for_async (async, pthread_self (), true, NULL);
-
-  /* Launches on the null stream already block on other streams in the
-     context.  */
-  if (!waiting_stream || waiting_stream == nvthd->ptx_dev->null_stream)
-    return;
-
-  event_gc (true);
-
-  pthread_mutex_lock (&nvthd->ptx_dev->stream_lock);
-
-  for (other_stream = nvthd->ptx_dev->active_streams;
-       other_stream != NULL;
-       other_stream = other_stream->next)
-    {
-      if (!other_stream->multithreaded
-	  && !pthread_equal (other_stream->host_thread, self))
-	continue;
-
-      e = (CUevent *) GOMP_PLUGIN_malloc (sizeof (CUevent));
-
-      CUDA_CALL_ASSERT (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
-
-      /* Record an event on the waited-for stream.  */
-      CUDA_CALL_ASSERT (cuEventRecord, *e, other_stream->stream);
-
-      event_add (PTX_EVT_SYNC, e, NULL, 0);
-
-      CUDA_CALL_ASSERT (cuStreamWaitEvent, waiting_stream->stream, *e, 0);
-   }
-
-  pthread_mutex_unlock (&nvthd->ptx_dev->stream_lock);
-}
-
 static void *
 nvptx_get_current_cuda_device (void)
 {
@@ -1799,75 +977,6 @@ nvptx_get_current_cuda_context (void)
   return nvthd->ptx_dev->ctx;
 }
 
-static void *
-nvptx_get_cuda_stream (int async)
-{
-  struct ptx_stream *s;
-  struct nvptx_thread *nvthd = nvptx_thread ();
-
-  if (!nvthd || !nvthd->ptx_dev)
-    return NULL;
-
-  s = select_stream_for_async (async, pthread_self (), false, NULL);
-
-  return s ? s->stream : NULL;
-}
-
-static int
-nvptx_set_cuda_stream (int async, void *stream)
-{
-  struct ptx_stream *oldstream;
-  pthread_t self = pthread_self ();
-  struct nvptx_thread *nvthd = nvptx_thread ();
-
-  /* Due to the "null_stream" usage for "acc_async_sync", this cannot be used
-     to change the stream handle associated with "acc_async_sync".  */
-  if (async == acc_async_sync)
-    {
-      GOMP_PLUGIN_debug (0, "Refusing request to set CUDA stream associated"
-			 " with \"acc_async_sync\"\n");
-      return 0;
-    }
-
-  pthread_mutex_lock (&nvthd->ptx_dev->stream_lock);
-
-  /* We have a list of active streams and an array mapping async values to
-     entries of that list.  We need to take "ownership" of the passed-in stream,
-     and add it to our list, removing the previous entry also (if there was one)
-     in order to prevent resource leaks.  Note the potential for surprise
-     here: maybe we should keep track of passed-in streams and leave it up to
-     the user to tidy those up, but that doesn't work for stream handles
-     returned from acc_get_cuda_stream above...  */
-
-  oldstream = select_stream_for_async (async, self, false, NULL);
-
-  if (oldstream)
-    {
-      if (nvthd->ptx_dev->active_streams == oldstream)
-	nvthd->ptx_dev->active_streams = nvthd->ptx_dev->active_streams->next;
-      else
-	{
-	  struct ptx_stream *s = nvthd->ptx_dev->active_streams;
-	  while (s->next != oldstream)
-	    s = s->next;
-	  s->next = s->next->next;
-	}
-
-      CUDA_CALL_ASSERT (cuStreamDestroy, oldstream->stream);
-
-      if (!map_fini (oldstream))
-	GOMP_PLUGIN_fatal ("error when freeing host memory");
-
-      free (oldstream);
-    }
-
-  pthread_mutex_unlock (&nvthd->ptx_dev->stream_lock);
-
-  (void) select_stream_for_async (async, self, true, (CUstream) stream);
-
-  return 1;
-}
-
 /* Plugin entry points.  */
 
 const char *
@@ -2107,6 +1216,23 @@ GOMP_OFFLOAD_alloc (int ord, size_t size)
 {
   if (!nvptx_attach_host_thread_to_device (ord))
     return NULL;
+
+  struct ptx_device *ptx_dev = ptx_devices[ord];
+  struct ptx_free_block *blocks, *tmp;
+
+  pthread_mutex_lock (&ptx_dev->free_blocks_lock);
+  blocks = ptx_dev->free_blocks;
+  ptx_dev->free_blocks = NULL;
+  pthread_mutex_unlock (&ptx_dev->free_blocks_lock);
+
+  while (blocks)
+    {
+      tmp = blocks->next;
+      nvptx_free (blocks->ptr, ptx_dev);
+      free (blocks);
+      blocks = tmp;
+    }
+
   return nvptx_alloc (size);
 }
 
@@ -2114,93 +1240,92 @@ bool
 GOMP_OFFLOAD_free (int ord, void *ptr)
 {
   return (nvptx_attach_host_thread_to_device (ord)
-	  && nvptx_free (ptr));
+	  && nvptx_free (ptr, ptx_devices[ord]));
 }
 
-bool
-GOMP_OFFLOAD_dev2host (int ord, void *dst, const void *src, size_t n)
-{
-  return (nvptx_attach_host_thread_to_device (ord)
-	  && nvptx_dev2host (dst, src, n));
-}
-
-bool
-GOMP_OFFLOAD_host2dev (int ord, void *dst, const void *src, size_t n)
-{
-  return (nvptx_attach_host_thread_to_device (ord)
-	  && nvptx_host2dev (dst, src, n));
-}
-
-bool
-GOMP_OFFLOAD_dev2dev (int ord, void *dst, const void *src, size_t n)
-{
-  struct ptx_device *ptx_dev = ptx_devices[ord];
-  CUDA_CALL (cuMemcpyDtoDAsync, (CUdeviceptr) dst, (CUdeviceptr) src, n,
-				ptx_dev->null_stream->stream);
-  return true;
-}
-
-void (*device_run) (int n, void *fn_ptr, void *vars) = NULL;
-
 void
 GOMP_OFFLOAD_openacc_exec (void (*fn) (void *), size_t mapnum,
 			   void **hostaddrs, void **devaddrs,
-			   int async, unsigned *dims, void *targ_mem_desc)
+			   unsigned *dims, void *targ_mem_desc)
 {
-  nvptx_exec (fn, mapnum, hostaddrs, devaddrs, async, dims, targ_mem_desc);
-}
+  GOMP_PLUGIN_debug (0, "  %s: prepare mappings\n", __FUNCTION__);
 
-void
-GOMP_OFFLOAD_openacc_register_async_cleanup (void *targ_mem_desc, int async)
-{
-  struct nvptx_thread *nvthd = nvptx_thread ();
-  CUevent *e = (CUevent *) GOMP_PLUGIN_malloc (sizeof (CUevent));
+  void **hp = NULL;
+  CUdeviceptr dp = 0;
 
-  CUDA_CALL_ASSERT (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
-  CUDA_CALL_ASSERT (cuEventRecord, *e, nvthd->current_stream->stream);
-  event_add (PTX_EVT_ASYNC_CLEANUP, e, targ_mem_desc, async);
-}
+  if (mapnum > 0)
+    {
+      hp = alloca (mapnum * sizeof (void *));
+      for (int i = 0; i < mapnum; i++)
+	hp[i] = (devaddrs[i] ? devaddrs[i] : hostaddrs[i]);
+      CUDA_CALL_ASSERT (cuMemAlloc, &dp, mapnum * sizeof (void *));
+    }
 
-int
-GOMP_OFFLOAD_openacc_async_test (int async)
-{
-  return nvptx_async_test (async);
-}
+  /* Copy the (device) pointers to arguments to the device (dp and hp might in
+     fact have the same value on a unified-memory system).  */
+  if (mapnum > 0)
+    CUDA_CALL_ASSERT (cuMemcpyHtoD, dp, (void *) hp,
+		      mapnum * sizeof (void *));
 
-int
-GOMP_OFFLOAD_openacc_async_test_all (void)
-{
-  return nvptx_async_test_all ();
-}
+  nvptx_exec (fn, mapnum, hostaddrs, devaddrs, dims, targ_mem_desc,
+	      dp, NULL);
 
-void
-GOMP_OFFLOAD_openacc_async_wait (int async)
-{
-  nvptx_wait (async);
+  CUresult r = CUDA_CALL_NOCHECK (cuStreamSynchronize, NULL);
+  const char *maybe_abort_msg = "(perhaps abort was called)";
+  if (r == CUDA_ERROR_LAUNCH_FAILED)
+    GOMP_PLUGIN_fatal ("cuStreamSynchronize error: %s %s\n", cuda_error (r),
+		       maybe_abort_msg);
+  else if (r != CUDA_SUCCESS)
+    GOMP_PLUGIN_fatal ("cuStreamSynchronize error: %s", cuda_error (r));
+  CUDA_CALL_ASSERT (cuMemFree, dp);
 }
 
-void
-GOMP_OFFLOAD_openacc_async_wait_async (int async1, int async2)
+static void
+cuda_free_argmem (void *ptr)
 {
-  nvptx_wait_async (async1, async2);
+  void **block = (void **) ptr;
+  nvptx_free (block[0], (struct ptx_device *) block[1]);
+  free (block);
 }
 
 void
-GOMP_OFFLOAD_openacc_async_wait_all (void)
+GOMP_OFFLOAD_openacc_async_exec (void (*fn) (void *), size_t mapnum,
+				 void **hostaddrs, void **devaddrs,
+				 unsigned *dims, void *targ_mem_desc,
+				 struct goacc_asyncqueue *aq)
 {
-  nvptx_wait_all ();
-}
+  GOMP_PLUGIN_debug (0, "  %s: prepare mappings\n", __FUNCTION__);
 
-void
-GOMP_OFFLOAD_openacc_async_wait_all_async (int async)
-{
-  nvptx_wait_all_async (async);
-}
+  void **hp = NULL;
+  CUdeviceptr dp = 0;
+  void **block = NULL;
 
-void
-GOMP_OFFLOAD_openacc_async_set_async (int async)
-{
-  nvptx_set_async (async);
+  if (mapnum > 0)
+    {
+      block = (void **) GOMP_PLUGIN_malloc ((mapnum + 2) * sizeof (void *));
+      hp = block + 2;
+      for (int i = 0; i < mapnum; i++)
+	hp[i] = (devaddrs[i] ? devaddrs[i] : hostaddrs[i]);
+      CUDA_CALL_ASSERT (cuMemAlloc, &dp, mapnum * sizeof (void *));
+    }
+
+  /* Copy the (device) pointers to arguments to the device (dp and hp might in
+     fact have the same value on a unified-memory system).  */
+  if (mapnum > 0)
+    {
+      CUDA_CALL_ASSERT (cuMemcpyHtoDAsync, dp, (void *) hp,
+			mapnum * sizeof (void *), aq->cuda_stream);
+      block[0] = (void *) dp;
+
+      struct nvptx_thread *nvthd =
+	(struct nvptx_thread *) GOMP_PLUGIN_acc_thread ();
+      block[1] = (void *) nvthd->ptx_dev;
+    }
+  nvptx_exec (fn, mapnum, hostaddrs, devaddrs, dims, targ_mem_desc,
+	      dp, aq->cuda_stream);
+
+  if (mapnum > 0)
+    GOMP_OFFLOAD_openacc_async_queue_callback (aq, cuda_free_argmem, block);
 }
 
 void *
@@ -2222,7 +1347,6 @@ GOMP_OFFLOAD_openacc_create_thread_data (int ord)
   if (!thd_ctx)
     CUDA_CALL_ASSERT (cuCtxPushCurrent, ptx_dev->ctx);
 
-  nvthd->current_stream = ptx_dev->null_stream;
   nvthd->ptx_dev = ptx_dev;
 
   return (void *) nvthd;
@@ -2246,22 +1370,186 @@ GOMP_OFFLOAD_openacc_cuda_get_current_context (voi
   return nvptx_get_current_cuda_context ();
 }
 
-/* NOTE: This returns a CUstream, not a ptx_stream pointer.  */
-
+/* This returns a CUstream.  */
 void *
-GOMP_OFFLOAD_openacc_cuda_get_stream (int async)
+GOMP_OFFLOAD_openacc_cuda_get_stream (struct goacc_asyncqueue *aq)
 {
-  return nvptx_get_cuda_stream (async);
+  return (void *) aq->cuda_stream;
 }
 
-/* NOTE: This takes a CUstream, not a ptx_stream pointer.  */
+/* This takes a CUstream.  */
+int
+GOMP_OFFLOAD_openacc_cuda_set_stream (struct goacc_asyncqueue *aq, void *stream)
+{
+  if (aq->cuda_stream)
+    {
+      CUDA_CALL_ASSERT (cuStreamSynchronize, aq->cuda_stream);
+      CUDA_CALL_ASSERT (cuStreamDestroy, aq->cuda_stream);
+    }
 
+  aq->cuda_stream = (CUstream) stream;
+  return 1;
+}
+
+struct goacc_asyncqueue *
+GOMP_OFFLOAD_openacc_async_construct (void)
+{
+  CUstream stream = NULL;
+  CUDA_CALL_ERET (NULL, cuStreamCreate, &stream, CU_STREAM_DEFAULT);
+
+  struct goacc_asyncqueue *aq
+    = GOMP_PLUGIN_malloc (sizeof (struct goacc_asyncqueue));
+  aq->cuda_stream = stream;
+  return aq;
+}
+
+bool
+GOMP_OFFLOAD_openacc_async_destruct (struct goacc_asyncqueue *aq)
+{
+  CUDA_CALL_ERET (false, cuStreamDestroy, aq->cuda_stream);
+  free (aq);
+  return true;
+}
+
 int
-GOMP_OFFLOAD_openacc_cuda_set_stream (int async, void *stream)
+GOMP_OFFLOAD_openacc_async_test (struct goacc_asyncqueue *aq)
 {
-  return nvptx_set_cuda_stream (async, stream);
+  CUresult r = CUDA_CALL_NOCHECK (cuStreamQuery, aq->cuda_stream);
+  if (r == CUDA_SUCCESS)
+    return 1;
+  if (r == CUDA_ERROR_NOT_READY)
+    return 0;
+
+  GOMP_PLUGIN_error ("cuStreamQuery error: %s", cuda_error (r));
+  return -1;
 }
 
+bool
+GOMP_OFFLOAD_openacc_async_synchronize (struct goacc_asyncqueue *aq)
+{
+  CUDA_CALL_ERET (false, cuStreamSynchronize, aq->cuda_stream);
+  return true;
+}
+
+bool
+GOMP_OFFLOAD_openacc_async_serialize (struct goacc_asyncqueue *aq1,
+				      struct goacc_asyncqueue *aq2)
+{
+  CUevent e;
+  CUDA_CALL_ERET (false, cuEventCreate, &e, CU_EVENT_DISABLE_TIMING);
+  CUDA_CALL_ERET (false, cuEventRecord, e, aq1->cuda_stream);
+  CUDA_CALL_ERET (false, cuStreamWaitEvent, aq2->cuda_stream, e, 0);
+  return true;
+}
+
+static void
+cuda_callback_wrapper (CUstream stream, CUresult res, void *ptr)
+{
+  if (res != CUDA_SUCCESS)
+    GOMP_PLUGIN_fatal ("%s error: %s", __FUNCTION__, cuda_error (res));
+  struct nvptx_callback *cb = (struct nvptx_callback *) ptr;
+  cb->fn (cb->ptr);
+  free (ptr);
+}
+
+void
+GOMP_OFFLOAD_openacc_async_queue_callback (struct goacc_asyncqueue *aq,
+					   void (*callback_fn)(void *),
+					   void *userptr)
+{
+  struct nvptx_callback *b = GOMP_PLUGIN_malloc (sizeof (*b));
+  b->fn = callback_fn;
+  b->ptr = userptr;
+  b->aq = aq;
+  CUDA_CALL_ASSERT (cuStreamAddCallback, aq->cuda_stream,
+		    cuda_callback_wrapper, (void *) b, 0);
+}
+
+static bool
+cuda_memcpy_sanity_check (const void *h, const void *d, size_t s)
+{
+  CUdeviceptr pb;
+  size_t ps;
+  if (!s)
+    return true;
+  if (!d)
+    {
+      GOMP_PLUGIN_error ("invalid device address");
+      return false;
+    }
+  CUDA_CALL (cuMemGetAddressRange, &pb, &ps, (CUdeviceptr) d);
+  if (!pb)
+    {
+      GOMP_PLUGIN_error ("invalid device address");
+      return false;
+    }
+  if (!h)
+    {
+      GOMP_PLUGIN_error ("invalid host address");
+      return false;
+    }
+  if (d == h)
+    {
+      GOMP_PLUGIN_error ("invalid host or device address");
+      return false;
+    }
+  if ((void *)(d + s) > (void *)(pb + ps))
+    {
+      GOMP_PLUGIN_error ("invalid size");
+      return false;
+    }
+  return true;
+}
+
+bool
+GOMP_OFFLOAD_host2dev (int ord, void *dst, const void *src, size_t n)
+{
+  if (!nvptx_attach_host_thread_to_device (ord)
+      || !cuda_memcpy_sanity_check (src, dst, n))
+    return false;
+  CUDA_CALL (cuMemcpyHtoD, (CUdeviceptr) dst, src, n);
+  return true;
+}
+
+bool
+GOMP_OFFLOAD_dev2host (int ord, void *dst, const void *src, size_t n)
+{
+  if (!nvptx_attach_host_thread_to_device (ord)
+      || !cuda_memcpy_sanity_check (dst, src, n))
+    return false;
+  CUDA_CALL (cuMemcpyDtoH, dst, (CUdeviceptr) src, n);
+  return true;
+}
+
+bool
+GOMP_OFFLOAD_dev2dev (int ord, void *dst, const void *src, size_t n)
+{
+  CUDA_CALL (cuMemcpyDtoDAsync, (CUdeviceptr) dst, (CUdeviceptr) src, n, NULL);
+  return true;
+}
+
+bool
+GOMP_OFFLOAD_openacc_async_host2dev (int ord, void *dst, const void *src,
+				     size_t n, struct goacc_asyncqueue *aq)
+{
+  if (!nvptx_attach_host_thread_to_device (ord)
+      || !cuda_memcpy_sanity_check (src, dst, n))
+    return false;
+  CUDA_CALL (cuMemcpyHtoDAsync, (CUdeviceptr) dst, src, n, aq->cuda_stream);
+  return true;
+}
+
+bool
+GOMP_OFFLOAD_openacc_async_dev2host (int ord, void *dst, const void *src,
+				     size_t n, struct goacc_asyncqueue *aq)
+{
+  if (!nvptx_attach_host_thread_to_device (ord)
+      || !cuda_memcpy_sanity_check (dst, src, n))
+    return false;
+  CUDA_CALL (cuMemcpyDtoHAsync, dst, (CUdeviceptr) src, n, aq->cuda_stream);
+  return true;
+}
+
 /* Adjust launch dimensions: pick good values for number of blocks and warps
    and ensure that number of warps does not exceed CUDA limits as well as GCC's
    own limits.  */
@@ -2360,8 +1648,7 @@ GOMP_OFFLOAD_run (int ord, void *tgt_fn, void *tgt
     CU_LAUNCH_PARAM_END
   };
   r = CUDA_CALL_NOCHECK (cuLaunchKernel, function, teams, 1, 1,
-			 32, threads, 1, 0, ptx_dev->null_stream->stream,
-			 NULL, config);
+			 32, threads, 1, 0, NULL, NULL, config);
   if (r != CUDA_SUCCESS)
     GOMP_PLUGIN_fatal ("cuLaunchKernel error: %s", cuda_error (r));
 
Index: libgomp/plugin/cuda-lib.def
===================================================================
--- libgomp/plugin/cuda-lib.def	(revision 269183)
+++ libgomp/plugin/cuda-lib.def	(working copy)
@@ -42,6 +42,7 @@ CUDA_ONE_CALL (cuModuleLoad)
 CUDA_ONE_CALL (cuModuleLoadData)
 CUDA_ONE_CALL (cuModuleUnload)
 CUDA_ONE_CALL_MAYBE_NULL (cuOccupancyMaxPotentialBlockSize)
+CUDA_ONE_CALL (cuStreamAddCallback)
 CUDA_ONE_CALL (cuStreamCreate)
 CUDA_ONE_CALL (cuStreamDestroy)
 CUDA_ONE_CALL (cuStreamQuery)
Index: libgomp/oacc-mem.c
===================================================================
--- libgomp/oacc-mem.c	(revision 269183)
+++ libgomp/oacc-mem.c	(working copy)
@@ -172,18 +172,11 @@ memcpy_tofrom_device (bool from, void *d, void *h,
       return;
     }
 
-  if (async > acc_async_sync)
-    thr->dev->openacc.async_set_async_func (async);
-
-  bool ret = (from
-	      ? thr->dev->dev2host_func (thr->dev->target_id, h, d, s)
-	      : thr->dev->host2dev_func (thr->dev->target_id, d, h, s));
-
-  if (async > acc_async_sync)
-    thr->dev->openacc.async_set_async_func (acc_async_sync);
-
-  if (!ret)
-    gomp_fatal ("error in %s", libfnname);
+  goacc_aq aq = get_goacc_asyncqueue (async);
+  if (from)
+    gomp_copy_dev2host (thr->dev, aq, h, d, s);
+  else
+    gomp_copy_host2dev (thr->dev, aq, d, h, s, /* TODO: cbuf? */ NULL);
 }
 
 void
@@ -509,17 +502,13 @@ present_create_copy (unsigned f, void *h, size_t s
 
       gomp_mutex_unlock (&acc_dev->lock);
 
-      if (async > acc_async_sync)
-	acc_dev->openacc.async_set_async_func (async);
+      goacc_aq aq = get_goacc_asyncqueue (async);
 
-      tgt = gomp_map_vars (acc_dev, mapnum, &hostaddrs, NULL, &s, &kinds, true,
-			   GOMP_MAP_VARS_OPENACC);
+      tgt = gomp_map_vars_async (acc_dev, aq, mapnum, &hostaddrs, NULL, &s,
+				 &kinds, true, GOMP_MAP_VARS_OPENACC);
       /* Initialize dynamic refcount.  */
       tgt->list[0].key->dynamic_refcount = 1;
 
-      if (async > acc_async_sync)
-	acc_dev->openacc.async_set_async_func (acc_async_sync);
-
       gomp_mutex_lock (&acc_dev->lock);
 
       d = tgt->to_free;
@@ -676,13 +665,9 @@ delete_copyout (unsigned f, void *h, size_t s, int
 
       if (f & FLAG_COPYOUT)
 	{
-	  if (async > acc_async_sync)
-	    acc_dev->openacc.async_set_async_func (async);
-	  acc_dev->dev2host_func (acc_dev->target_id, h, d, s);
-	  if (async > acc_async_sync)
-	    acc_dev->openacc.async_set_async_func (acc_async_sync);
+	  goacc_aq aq = get_goacc_asyncqueue (async);
+	  gomp_copy_dev2host (acc_dev, aq, h, d, s);
 	}
-
       gomp_remove_var (acc_dev, n);
     }
 
@@ -765,17 +750,13 @@ update_dev_host (int is_dev, void *h, size_t s, in
   d = (void *) (n->tgt->tgt_start + n->tgt_offset
 		+ (uintptr_t) h - n->host_start);
 
-  if (async > acc_async_sync)
-    acc_dev->openacc.async_set_async_func (async);
+  goacc_aq aq = get_goacc_asyncqueue (async);
 
   if (is_dev)
-    acc_dev->host2dev_func (acc_dev->target_id, d, h, s);
+    gomp_copy_host2dev (acc_dev, aq, d, h, s, /* TODO: cbuf? */ NULL);
   else
-    acc_dev->dev2host_func (acc_dev->target_id, h, d, s);
+    gomp_copy_dev2host (acc_dev, aq, h, d, s);
 
-  if (async > acc_async_sync)
-    acc_dev->openacc.async_set_async_func (acc_async_sync);
-
   gomp_mutex_unlock (&acc_dev->lock);
 }
 
@@ -805,7 +786,7 @@ acc_update_self_async (void *h, size_t s, int asyn
 
 void
 gomp_acc_insert_pointer (size_t mapnum, void **hostaddrs, size_t *sizes,
-			 void *kinds)
+			 void *kinds, int async)
 {
   struct target_mem_desc *tgt;
   struct goacc_thread *thr = goacc_thread ();
@@ -835,8 +816,9 @@ gomp_acc_insert_pointer (size_t mapnum, void **hos
     }
 
   gomp_debug (0, "  %s: prepare mappings\n", __FUNCTION__);
-  tgt = gomp_map_vars (acc_dev, mapnum, hostaddrs,
-		       NULL, sizes, kinds, true, GOMP_MAP_VARS_OPENACC);
+  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);
   gomp_debug (0, "  %s: mappings prepared\n", __FUNCTION__);
 
   /* Initialize dynamic refcount.  */
@@ -930,7 +912,10 @@ gomp_acc_remove_pointer (void *h, size_t s, bool f
       if (async < acc_async_noval)
 	gomp_unmap_vars (t, true);
       else
-	t->device_descr->openacc.register_async_cleanup_func (t, async);
+	{
+	  goacc_aq aq = get_goacc_asyncqueue (async);
+	  gomp_unmap_vars_async (t, true, aq);
+	}
     }
 
   gomp_mutex_unlock (&acc_dev->lock);
Index: libgomp/oacc-parallel.c
===================================================================
--- libgomp/oacc-parallel.c	(revision 269183)
+++ libgomp/oacc-parallel.c	(working copy)
@@ -217,8 +217,6 @@ GOACC_parallel_keyed (int flags_m, void (*fn) (voi
     }
   va_end (ap);
   
-  acc_dev->openacc.async_set_async_func (async);
-
   if (!(acc_dev->capabilities & GOMP_OFFLOAD_CAP_NATIVE_EXEC))
     {
       k.host_start = (uintptr_t) fn;
@@ -235,44 +233,29 @@ GOACC_parallel_keyed (int flags_m, void (*fn) (voi
   else
     tgt_fn = (void (*)) fn;
 
-  tgt = gomp_map_vars (acc_dev, mapnum, hostaddrs, NULL, sizes, kinds, true,
-		       GOMP_MAP_VARS_OPENACC);
+  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);
-
-  acc_dev->openacc.exec_func (tgt_fn, mapnum, hostaddrs, devaddrs,
-			      async, dims, tgt);
-
-  /* If running synchronously, unmap immediately.  */
-  bool copyfrom = true;
-  if (async_synchronous_p (async))
-    gomp_unmap_vars (tgt, true);
+  if (aq == NULL)
+    {
+      acc_dev->openacc.exec_func (tgt_fn, mapnum, hostaddrs, devaddrs,
+				  dims, tgt);
+      /* If running synchronously, unmap immediately.  */
+      gomp_unmap_vars (tgt, true);
+    }
   else
     {
-      bool async_unmap = false;
-      for (size_t i = 0; i < tgt->list_count; i++)
-	{
-	  splay_tree_key k = tgt->list[i].key;
-	  if (k && k->refcount == 1)
-	    {
-	      async_unmap = true;
-	      break;
-	    }
-	}
-      if (async_unmap)
-	tgt->device_descr->openacc.register_async_cleanup_func (tgt, async);
-      else
-	{
-	  copyfrom = false;
-	  gomp_unmap_vars (tgt, copyfrom);
-	}
+      acc_dev->openacc.async.exec_func (tgt_fn, mapnum, hostaddrs, devaddrs,
+					dims, tgt, aq);
+      gomp_unmap_vars_async (tgt, true, aq);
     }
-
-  acc_dev->openacc.async_set_async_func (acc_async_sync);
 }
 
 /* Legacy entry point, only provide host execution.  */
@@ -383,8 +366,6 @@ GOACC_enter_exit_data (int flags_m, size_t mapnum,
 	finalize = true;
     }
 
-  acc_dev->openacc.async_set_async_func (async);
-
   /* Determine if this is an "acc enter data".  */
   for (i = 0; i < mapnum; ++i)
     {
@@ -437,11 +418,11 @@ GOACC_enter_exit_data (int flags_m, size_t mapnum,
 		{
 		case GOMP_MAP_ALLOC:
 		case GOMP_MAP_FORCE_ALLOC:
-		  acc_create (hostaddrs[i], sizes[i]);
+		  acc_create_async (hostaddrs[i], sizes[i], async);
 		  break;
 		case GOMP_MAP_TO:
 		case GOMP_MAP_FORCE_TO:
-		  acc_copyin (hostaddrs[i], sizes[i]);
+		  acc_copyin_async (hostaddrs[i], sizes[i], async);
 		  break;
 		default:
 		  gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
@@ -452,7 +433,7 @@ GOACC_enter_exit_data (int flags_m, size_t mapnum,
 	  else
 	    {
 	      gomp_acc_insert_pointer (pointer, &hostaddrs[i],
-				       &sizes[i], &kinds[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
@@ -477,17 +458,17 @@ GOACC_enter_exit_data (int flags_m, size_t mapnum,
 		if (acc_is_present (hostaddrs[i], sizes[i]))
 		  {
 		    if (finalize)
-		      acc_delete_finalize (hostaddrs[i], sizes[i]);
+		      acc_delete_finalize_async (hostaddrs[i], sizes[i], async);
 		    else
-		      acc_delete (hostaddrs[i], sizes[i]);
+		      acc_delete_async (hostaddrs[i], sizes[i], async);
 		  }
 		break;
 	      case GOMP_MAP_FROM:
 	      case GOMP_MAP_FORCE_FROM:
 		if (finalize)
-		  acc_copyout_finalize (hostaddrs[i], sizes[i]);
+		  acc_copyout_finalize_async (hostaddrs[i], sizes[i], async);
 		else
-		  acc_copyout (hostaddrs[i], sizes[i]);
+		  acc_copyout_async (hostaddrs[i], sizes[i], async);
 		break;
 	      default:
 		gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
@@ -505,8 +486,6 @@ GOACC_enter_exit_data (int flags_m, size_t mapnum,
 	    i += pointer - 1;
 	  }
       }
-
-  acc_dev->openacc.async_set_async_func (acc_async_sync);
 }
 
 static void
@@ -532,9 +511,10 @@ goacc_wait (int async, int num_waits, va_list *ap)
       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.  */
+	/* 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);
     }
@@ -567,8 +547,6 @@ GOACC_update (int flags_m, size_t mapnum,
       va_end (ap);
     }
 
-  acc_dev->openacc.async_set_async_func (async);
-
   bool update_device = false;
   for (i = 0; i < mapnum; ++i)
     {
@@ -591,6 +569,8 @@ GOACC_update (int flags_m, size_t mapnum,
 		 the value of the allocated device memory in the
 		 previous pointer.  */
 	      *(uintptr_t *) hostaddrs[i] = (uintptr_t)dptr;
+	      /* TODO: verify that we really cannot use acc_update_device_async
+		 here.  */
 	      acc_update_device (hostaddrs[i], sizeof (uintptr_t));
 
 	      /* Restore the host pointer.  */
@@ -608,7 +588,7 @@ GOACC_update (int flags_m, size_t mapnum,
 	  /* Fallthru  */
 	case GOMP_MAP_FORCE_TO:
 	  update_device = true;
-	  acc_update_device (hostaddrs[i], sizes[i]);
+	  acc_update_device_async (hostaddrs[i], sizes[i], async);
 	  break;
 
 	case GOMP_MAP_FROM:
@@ -620,7 +600,7 @@ GOACC_update (int flags_m, size_t mapnum,
 	  /* Fallthru  */
 	case GOMP_MAP_FORCE_FROM:
 	  update_device = false;
-	  acc_update_self (hostaddrs[i], sizes[i]);
+	  acc_update_self_async (hostaddrs[i], sizes[i], async);
 	  break;
 
 	default:
@@ -628,8 +608,6 @@ GOACC_update (int flags_m, size_t mapnum,
 	  break;
 	}
     }
-
-  acc_dev->openacc.async_set_async_func (acc_async_sync);
 }
 
 void
Index: libgomp/oacc-init.c
===================================================================
--- libgomp/oacc-init.c	(revision 269183)
+++ libgomp/oacc-init.c	(working copy)
@@ -309,7 +309,7 @@ acc_shutdown_1 (acc_device_t d)
       if (acc_dev->state == GOMP_DEVICE_INITIALIZED)
         {
 	  devices_active = true;
-	  ret &= acc_dev->fini_device_func (acc_dev->target_id);
+	  ret &= gomp_fini_device (acc_dev);
 	  acc_dev->state = GOMP_DEVICE_UNINITIALIZED;
 	}
       gomp_mutex_unlock (&acc_dev->lock);
@@ -426,8 +426,6 @@ goacc_attach_host_thread_to_device (int ord)
   
   thr->target_tls
     = acc_dev->openacc.create_thread_data_func (ord);
-  
-  acc_dev->openacc.async_set_async_func (acc_async_sync);
 }
 
 /* OpenACC 2.0a (3.2.12, 3.2.13) doesn't specify whether the serialization of
Index: libgomp/oacc-cuda.c
===================================================================
--- libgomp/oacc-cuda.c	(revision 269183)
+++ libgomp/oacc-cuda.c	(working copy)
@@ -30,6 +30,7 @@
 #include "config.h"
 #include "libgomp.h"
 #include "oacc-int.h"
+#include <assert.h>
 
 void *
 acc_get_current_cuda_device (void)
@@ -62,7 +63,11 @@ acc_get_cuda_stream (int async)
     return NULL;
 
   if (thr && thr->dev && thr->dev->openacc.cuda.get_stream_func)
-    return thr->dev->openacc.cuda.get_stream_func (async);
+    {
+      goacc_aq aq = lookup_goacc_asyncqueue (thr, false, async);
+      if (aq)
+	return thr->dev->openacc.cuda.get_stream_func (aq);
+    }
  
   return NULL;
 }
@@ -79,8 +84,23 @@ acc_set_cuda_stream (int async, void *stream)
 
   thr = goacc_thread ();
 
+  int ret = -1;
   if (thr && thr->dev && thr->dev->openacc.cuda.set_stream_func)
-    return thr->dev->openacc.cuda.set_stream_func (async, stream);
+    {
+      goacc_aq aq = get_goacc_asyncqueue (async);
+      /* Due to not using an asyncqueue for "acc_async_sync", this cannot be
+	 used to change the CUDA stream associated with "acc_async_sync".  */
+      if (!aq)
+	{
+	  assert (async == acc_async_sync);
+	  gomp_debug (0, "Refusing request to set CUDA stream associated"
+		      " with \"acc_async_sync\"\n");
+	  return 0;
+	}
+      gomp_mutex_lock (&thr->dev->openacc.async.lock);
+      ret = thr->dev->openacc.cuda.set_stream_func (aq, stream);
+      gomp_mutex_unlock (&thr->dev->openacc.async.lock);
+    }
 
-  return -1;
+  return ret;
 }
Index: libgomp/target.c
===================================================================
--- libgomp/target.c	(revision 269183)
+++ libgomp/target.c	(working copy)
@@ -177,6 +177,22 @@ gomp_device_copy (struct gomp_device_descr *device
     }
 }
 
+static inline void
+goacc_device_copy_async (struct gomp_device_descr *devicep,
+			 bool (*copy_func) (int, void *, const void *, size_t,
+					    struct goacc_asyncqueue *),
+			 const char *dst, void *dstaddr,
+			 const char *src, const void *srcaddr,
+			 size_t size, struct goacc_asyncqueue *aq)
+{
+  if (!copy_func (devicep->target_id, dstaddr, srcaddr, size, aq))
+    {
+      gomp_mutex_unlock (&devicep->lock);
+      gomp_fatal ("Copying of %s object [%p..%p) to %s object [%p..%p) failed",
+		  src, srcaddr, srcaddr + size, dst, dstaddr, dstaddr + size);
+    }
+}
+
 /* Infrastructure for coalescing adjacent or nearly adjacent (in device addresses)
    host to device memory transfers.  */
 
@@ -269,8 +285,9 @@ gomp_to_device_kind_p (int kind)
     }
 }
 
-static void
+attribute_hidden void
 gomp_copy_host2dev (struct gomp_device_descr *devicep,
+		    struct goacc_asyncqueue *aq,
 		    void *d, const void *h, size_t sz,
 		    struct gomp_coalesce_buf *cbuf)
 {
@@ -299,14 +316,23 @@ gomp_copy_host2dev (struct gomp_device_descr *devi
 	    }
 	}
     }
-  gomp_device_copy (devicep, devicep->host2dev_func, "dev", d, "host", h, sz);
+  if (__builtin_expect (aq != NULL, 0))
+    goacc_device_copy_async (devicep, devicep->openacc.async.host2dev_func,
+			     "dev", d, "host", h, sz, aq);
+  else
+    gomp_device_copy (devicep, devicep->host2dev_func, "dev", d, "host", h, sz);
 }
 
-static void
+attribute_hidden void
 gomp_copy_dev2host (struct gomp_device_descr *devicep,
+		    struct goacc_asyncqueue *aq,
 		    void *h, const void *d, size_t sz)
 {
-  gomp_device_copy (devicep, devicep->dev2host_func, "host", h, "dev", d, sz);
+  if (__builtin_expect (aq != NULL, 0))
+    goacc_device_copy_async (devicep, devicep->openacc.async.dev2host_func,
+			     "host", h, "dev", d, sz, aq);
+  else
+    gomp_device_copy (devicep, devicep->dev2host_func, "host", h, "dev", d, sz);
 }
 
 static void
@@ -324,7 +350,8 @@ gomp_free_device_memory (struct gomp_device_descr
    Helper function of gomp_map_vars.  */
 
 static inline void
-gomp_map_vars_existing (struct gomp_device_descr *devicep, splay_tree_key oldn,
+gomp_map_vars_existing (struct gomp_device_descr *devicep,
+			struct goacc_asyncqueue *aq, splay_tree_key oldn,
 			splay_tree_key newn, struct target_var_desc *tgt_var,
 			unsigned char kind, struct gomp_coalesce_buf *cbuf)
 {
@@ -346,7 +373,7 @@ static inline void
     }
 
   if (GOMP_MAP_ALWAYS_TO_P (kind))
-    gomp_copy_host2dev (devicep,
+    gomp_copy_host2dev (devicep, aq,
 			(void *) (oldn->tgt->tgt_start + oldn->tgt_offset
 				  + newn->host_start - oldn->host_start),
 			(void *) newn->host_start,
@@ -364,8 +391,8 @@ get_kind (bool short_mapkind, void *kinds, int idx
 }
 
 static void
-gomp_map_pointer (struct target_mem_desc *tgt, uintptr_t host_ptr,
-		  uintptr_t target_offset, uintptr_t bias,
+gomp_map_pointer (struct target_mem_desc *tgt, struct goacc_asyncqueue *aq,
+		  uintptr_t host_ptr, uintptr_t target_offset, uintptr_t bias,
 		  struct gomp_coalesce_buf *cbuf)
 {
   struct gomp_device_descr *devicep = tgt->device_descr;
@@ -376,7 +403,7 @@ static void
   if (cur_node.host_start == (uintptr_t) NULL)
     {
       cur_node.tgt_offset = (uintptr_t) NULL;
-      gomp_copy_host2dev (devicep,
+      gomp_copy_host2dev (devicep, aq,
 			  (void *) (tgt->tgt_start + target_offset),
 			  (void *) &cur_node.tgt_offset,
 			  sizeof (void *), cbuf);
@@ -398,12 +425,13 @@ static void
      array section.  Now subtract bias to get what we want
      to initialize the pointer with.  */
   cur_node.tgt_offset -= bias;
-  gomp_copy_host2dev (devicep, (void *) (tgt->tgt_start + target_offset),
+  gomp_copy_host2dev (devicep, aq, (void *) (tgt->tgt_start + target_offset),
 		      (void *) &cur_node.tgt_offset, sizeof (void *), cbuf);
 }
 
 static void
-gomp_map_fields_existing (struct target_mem_desc *tgt, splay_tree_key n,
+gomp_map_fields_existing (struct target_mem_desc *tgt,
+			  struct goacc_asyncqueue *aq, splay_tree_key n,
 			  size_t first, size_t i, void **hostaddrs,
 			  size_t *sizes, void *kinds,
 			  struct gomp_coalesce_buf *cbuf)
@@ -423,7 +451,7 @@ static void
       && n2->tgt == n->tgt
       && n2->host_start - n->host_start == n2->tgt_offset - n->tgt_offset)
     {
-      gomp_map_vars_existing (devicep, n2, &cur_node,
+      gomp_map_vars_existing (devicep, aq, n2, &cur_node,
 			      &tgt->list[i], kind & typemask, cbuf);
       return;
     }
@@ -439,8 +467,8 @@ static void
 	      && n2->host_start - n->host_start
 		 == n2->tgt_offset - n->tgt_offset)
 	    {
-	      gomp_map_vars_existing (devicep, n2, &cur_node, &tgt->list[i],
-				      kind & typemask, cbuf);
+	      gomp_map_vars_existing (devicep, aq, n2, &cur_node,
+				      &tgt->list[i], kind & typemask, cbuf);
 	      return;
 	    }
 	}
@@ -451,7 +479,7 @@ static void
 	  && n2->tgt == n->tgt
 	  && n2->host_start - n->host_start == n2->tgt_offset - n->tgt_offset)
 	{
-	  gomp_map_vars_existing (devicep, n2, &cur_node, &tgt->list[i],
+	  gomp_map_vars_existing (devicep, aq, n2, &cur_node, &tgt->list[i],
 				  kind & typemask, cbuf);
 	  return;
 	}
@@ -483,10 +511,12 @@ gomp_map_val (struct target_mem_desc *tgt, void **
   return tgt->tgt_start + tgt->list[i].offset;
 }
 
-attribute_hidden struct target_mem_desc *
-gomp_map_vars (struct gomp_device_descr *devicep, size_t mapnum,
-	       void **hostaddrs, void **devaddrs, size_t *sizes, void *kinds,
-	       bool short_mapkind, enum gomp_map_vars_kind pragma_kind)
+static inline __attribute__((always_inline)) struct target_mem_desc *
+gomp_map_vars_internal (struct gomp_device_descr *devicep,
+			struct goacc_asyncqueue *aq, size_t mapnum,
+			void **hostaddrs, void **devaddrs, size_t *sizes,
+			void *kinds, bool short_mapkind,
+			enum gomp_map_vars_kind pragma_kind)
 {
   size_t i, tgt_align, tgt_size, not_found_cnt = 0;
   bool has_firstprivate = false;
@@ -600,7 +630,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
 	      continue;
 	    }
 	  for (i = first; i <= last; i++)
-	    gomp_map_fields_existing (tgt, n, first, i, hostaddrs,
+	    gomp_map_fields_existing (tgt, aq, n, first, i, hostaddrs,
 				      sizes, kinds, NULL);
 	  i--;
 	  continue;
@@ -645,7 +675,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
       else
 	n = splay_tree_lookup (mem_map, &cur_node);
       if (n && n->refcount != REFCOUNT_LINK)
-	gomp_map_vars_existing (devicep, n, &cur_node, &tgt->list[i],
+	gomp_map_vars_existing (devicep, aq, n, &cur_node, &tgt->list[i],
 				kind & typemask, NULL);
       else
 	{
@@ -756,7 +786,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
 		tgt_size = (tgt_size + align - 1) & ~(align - 1);
 		tgt->list[i].offset = tgt_size;
 		len = sizes[i];
-		gomp_copy_host2dev (devicep,
+		gomp_copy_host2dev (devicep, aq,
 				    (void *) (tgt->tgt_start + tgt_size),
 				    (void *) hostaddrs[i], len, cbufp);
 		tgt_size += len;
@@ -790,7 +820,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
 		    continue;
 		  }
 		for (i = first; i <= last; i++)
-		  gomp_map_fields_existing (tgt, n, first, i, hostaddrs,
+		  gomp_map_fields_existing (tgt, aq, n, first, i, hostaddrs,
 					    sizes, kinds, cbufp);
 		i--;
 		continue;
@@ -810,7 +840,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
 		  cur_node.tgt_offset = gomp_map_val (tgt, hostaddrs, i - 1);
 		if (cur_node.tgt_offset)
 		  cur_node.tgt_offset -= sizes[i];
-		gomp_copy_host2dev (devicep,
+		gomp_copy_host2dev (devicep, aq,
 				    (void *) (n->tgt->tgt_start
 					      + n->tgt_offset
 					      + cur_node.host_start
@@ -831,7 +861,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
 	      k->host_end = k->host_start + sizeof (void *);
 	    splay_tree_key n = splay_tree_lookup (mem_map, k);
 	    if (n && n->refcount != REFCOUNT_LINK)
-	      gomp_map_vars_existing (devicep, n, k, &tgt->list[i],
+	      gomp_map_vars_existing (devicep, aq, n, k, &tgt->list[i],
 				      kind & typemask, cbufp);
 	    else
 	      {
@@ -884,18 +914,19 @@ gomp_map_val (struct target_mem_desc *tgt, void **
 		  case GOMP_MAP_FORCE_TOFROM:
 		  case GOMP_MAP_ALWAYS_TO:
 		  case GOMP_MAP_ALWAYS_TOFROM:
-		    gomp_copy_host2dev (devicep,
+		    gomp_copy_host2dev (devicep, aq,
 					(void *) (tgt->tgt_start
 						  + k->tgt_offset),
 					(void *) k->host_start,
 					k->host_end - k->host_start, cbufp);
 		    break;
 		  case GOMP_MAP_POINTER:
-		    gomp_map_pointer (tgt, (uintptr_t) *(void **) k->host_start,
+		    gomp_map_pointer (tgt, aq,
+				      (uintptr_t) *(void **) k->host_start,
 				      k->tgt_offset, sizes[i], cbufp);
 		    break;
 		  case GOMP_MAP_TO_PSET:
-		    gomp_copy_host2dev (devicep,
+		    gomp_copy_host2dev (devicep, aq,
 					(void *) (tgt->tgt_start
 						  + k->tgt_offset),
 					(void *) k->host_start,
@@ -917,7 +948,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
 			  tgt->list[j].always_copy_from = false;
 			  if (k->refcount != REFCOUNT_INFINITY)
 			    k->refcount++;
-			  gomp_map_pointer (tgt,
+			  gomp_map_pointer (tgt, aq,
 					    (uintptr_t) *(void **) hostaddrs[j],
 					    k->tgt_offset
 					    + ((uintptr_t) hostaddrs[j]
@@ -946,7 +977,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
 		    break;
 		  case GOMP_MAP_FORCE_DEVICEPTR:
 		    assert (k->host_end - k->host_start == sizeof (void *));
-		    gomp_copy_host2dev (devicep,
+		    gomp_copy_host2dev (devicep, aq,
 					(void *) (tgt->tgt_start
 						  + k->tgt_offset),
 					(void *) k->host_start,
@@ -965,7 +996,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
 		    void *tgt_addr = (void *) (tgt->tgt_start + k->tgt_offset);
 		    /* We intentionally do not use coalescing here, as it's not
 		       data allocated by the current call to this function.  */
-		    gomp_copy_host2dev (devicep, (void *) n->tgt_offset,
+		    gomp_copy_host2dev (devicep, aq, (void *) n->tgt_offset,
 					&tgt_addr, sizeof (void *), NULL);
 		  }
 		array++;
@@ -978,7 +1009,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
       for (i = 0; i < mapnum; i++)
 	{
 	  cur_node.tgt_offset = gomp_map_val (tgt, hostaddrs, i);
-	  gomp_copy_host2dev (devicep,
+	  gomp_copy_host2dev (devicep, aq,
 			      (void *) (tgt->tgt_start + i * sizeof (void *)),
 			      (void *) &cur_node.tgt_offset, sizeof (void *),
 			      cbufp);
@@ -989,7 +1020,7 @@ gomp_map_val (struct target_mem_desc *tgt, void **
     {
       long c = 0;
       for (c = 0; c < cbuf.chunk_cnt; ++c)
-	gomp_copy_host2dev (devicep,
+	gomp_copy_host2dev (devicep, aq,
 			    (void *) (tgt->tgt_start + cbuf.chunks[c].start),
 			    (char *) cbuf.buf + (cbuf.chunks[c].start
 						 - cbuf.chunks[0].start),
@@ -1012,7 +1043,27 @@ gomp_map_val (struct target_mem_desc *tgt, void **
   return tgt;
 }
 
-static void
+attribute_hidden struct target_mem_desc *
+gomp_map_vars (struct gomp_device_descr *devicep, size_t mapnum,
+	       void **hostaddrs, void **devaddrs, size_t *sizes, void *kinds,
+	       bool short_mapkind, enum gomp_map_vars_kind pragma_kind)
+{
+  return gomp_map_vars_internal (devicep, NULL, mapnum, hostaddrs, devaddrs,
+				 sizes, kinds, short_mapkind, pragma_kind);
+}
+
+attribute_hidden struct target_mem_desc *
+gomp_map_vars_async (struct gomp_device_descr *devicep,
+		     struct goacc_asyncqueue *aq, size_t mapnum,
+		     void **hostaddrs, void **devaddrs, size_t *sizes,
+		     void *kinds, bool short_mapkind,
+		     enum gomp_map_vars_kind pragma_kind)
+{
+  return gomp_map_vars_internal (devicep, aq, mapnum, hostaddrs, devaddrs,
+				 sizes, kinds, short_mapkind, pragma_kind);
+}
+
+attribute_hidden void
 gomp_unmap_tgt (struct target_mem_desc *tgt)
 {
   /* Deallocate on target the tgt->tgt_start .. tgt->tgt_end region.  */
@@ -1040,12 +1091,24 @@ gomp_remove_var (struct gomp_device_descr *devicep
   return is_tgt_unmapped;
 }
 
+static void
+gomp_unref_tgt (void *ptr)
+{
+  struct target_mem_desc *tgt = (struct target_mem_desc *) ptr;
+
+  if (tgt->refcount > 1)
+    tgt->refcount--;
+  else
+    gomp_unmap_tgt (tgt);
+}
+
 /* Unmap variables described by TGT.  If DO_COPYFROM is true, copy relevant
    variables back from device to host: if it is false, it is assumed that this
    has been done already.  */
 
-attribute_hidden void
-gomp_unmap_vars (struct target_mem_desc *tgt, bool do_copyfrom)
+static inline __attribute__((always_inline)) void
+gomp_unmap_vars_internal (struct target_mem_desc *tgt, bool do_copyfrom,
+			  struct goacc_asyncqueue *aq)
 {
   struct gomp_device_descr *devicep = tgt->device_descr;
 
@@ -1082,7 +1145,7 @@ gomp_remove_var (struct gomp_device_descr *devicep
 
       if ((do_unmap && do_copyfrom && tgt->list[i].copy_from)
 	  || tgt->list[i].always_copy_from)
-	gomp_copy_dev2host (devicep,
+	gomp_copy_dev2host (devicep, aq,
 			    (void *) (k->host_start + tgt->list[i].offset),
 			    (void *) (k->tgt->tgt_start + k->tgt_offset
 				      + tgt->list[i].offset),
@@ -1091,14 +1154,28 @@ gomp_remove_var (struct gomp_device_descr *devicep
 	gomp_remove_var (devicep, k);
     }
 
-  if (tgt->refcount > 1)
-    tgt->refcount--;
+  if (aq)
+    devicep->openacc.async.queue_callback_func (aq, gomp_unref_tgt,
+						(void *) tgt);
   else
-    gomp_unmap_tgt (tgt);
+    gomp_unref_tgt ((void *) tgt);
 
   gomp_mutex_unlock (&devicep->lock);
 }
 
+attribute_hidden void
+gomp_unmap_vars (struct target_mem_desc *tgt, bool do_copyfrom)
+{
+  gomp_unmap_vars_internal (tgt, do_copyfrom, NULL);
+}
+
+attribute_hidden void
+gomp_unmap_vars_async (struct target_mem_desc *tgt, bool do_copyfrom,
+		       struct goacc_asyncqueue *aq)
+{
+  gomp_unmap_vars_internal (tgt, do_copyfrom, aq);
+}
+
 static void
 gomp_update (struct gomp_device_descr *devicep, size_t mapnum, void **hostaddrs,
 	     size_t *sizes, void *kinds, bool short_mapkind)
@@ -1148,9 +1225,10 @@ gomp_update (struct gomp_device_descr *devicep, si
 	    size_t size = cur_node.host_end - cur_node.host_start;
 
 	    if (GOMP_MAP_COPY_TO_P (kind & typemask))
-	      gomp_copy_host2dev (devicep, devaddr, hostaddr, size, NULL);
+	      gomp_copy_host2dev (devicep, NULL, devaddr, hostaddr, size,
+				  NULL);
 	    if (GOMP_MAP_COPY_FROM_P (kind & typemask))
-	      gomp_copy_dev2host (devicep, hostaddr, devaddr, size);
+	      gomp_copy_dev2host (devicep, NULL, hostaddr, devaddr, size);
 	  }
       }
   gomp_mutex_unlock (&devicep->lock);
@@ -1443,9 +1521,24 @@ gomp_init_device (struct gomp_device_descr *device
 				   false);
     }
 
+  /* Initialize OpenACC asynchronous queues.  */
+  goacc_init_asyncqueues (devicep);
+
   devicep->state = GOMP_DEVICE_INITIALIZED;
 }
 
+/* This function finalizes the target device, specified by DEVICEP.  DEVICEP
+   must be locked on entry, and remains locked on return.  */
+
+attribute_hidden bool
+gomp_fini_device (struct gomp_device_descr *devicep)
+{
+  bool ret = goacc_fini_asyncqueues (devicep);
+  ret &= devicep->fini_device_func (devicep->target_id);
+  devicep->state = GOMP_DEVICE_FINALIZED;
+  return ret;
+}
+
 attribute_hidden void
 gomp_unload_device (struct gomp_device_descr *devicep)
 {
@@ -1954,7 +2047,7 @@ gomp_exit_data (struct gomp_device_descr *devicep,
 
 	  if ((kind == GOMP_MAP_FROM && k->refcount == 0)
 	      || kind == GOMP_MAP_ALWAYS_FROM)
-	    gomp_copy_dev2host (devicep, (void *) cur_node.host_start,
+	    gomp_copy_dev2host (devicep, NULL, (void *) cur_node.host_start,
 				(void *) (k->tgt->tgt_start + k->tgt_offset
 					  + cur_node.host_start
 					  - k->host_start),
@@ -2636,20 +2729,20 @@ gomp_load_plugin_for_device (struct gomp_device_de
   if (device->capabilities & GOMP_OFFLOAD_CAP_OPENACC_200)
     {
       if (!DLSYM_OPT (openacc.exec, openacc_exec)
-	  || !DLSYM_OPT (openacc.register_async_cleanup,
-			 openacc_register_async_cleanup)
-	  || !DLSYM_OPT (openacc.async_test, openacc_async_test)
-	  || !DLSYM_OPT (openacc.async_test_all, openacc_async_test_all)
-	  || !DLSYM_OPT (openacc.async_wait, openacc_async_wait)
-	  || !DLSYM_OPT (openacc.async_wait_async, openacc_async_wait_async)
-	  || !DLSYM_OPT (openacc.async_wait_all, openacc_async_wait_all)
-	  || !DLSYM_OPT (openacc.async_wait_all_async,
-			 openacc_async_wait_all_async)
-	  || !DLSYM_OPT (openacc.async_set_async, openacc_async_set_async)
 	  || !DLSYM_OPT (openacc.create_thread_data,
 			 openacc_create_thread_data)
 	  || !DLSYM_OPT (openacc.destroy_thread_data,
-			 openacc_destroy_thread_data))
+			 openacc_destroy_thread_data)
+	  || !DLSYM_OPT (openacc.async.construct, openacc_async_construct)
+	  || !DLSYM_OPT (openacc.async.destruct, openacc_async_destruct)
+	  || !DLSYM_OPT (openacc.async.test, openacc_async_test)
+	  || !DLSYM_OPT (openacc.async.synchronize, openacc_async_synchronize)
+	  || !DLSYM_OPT (openacc.async.serialize, openacc_async_serialize)
+	  || !DLSYM_OPT (openacc.async.queue_callback,
+			 openacc_async_queue_callback)
+	  || !DLSYM_OPT (openacc.async.exec, openacc_async_exec)
+	  || !DLSYM_OPT (openacc.async.dev2host, openacc_async_dev2host)
+	  || !DLSYM_OPT (openacc.async.host2dev, openacc_async_host2dev))
 	{
 	  /* Require all the OpenACC handlers if we have
 	     GOMP_OFFLOAD_CAP_OPENACC_200.  */
@@ -2700,10 +2793,7 @@ gomp_target_fini (void)
       struct gomp_device_descr *devicep = &devices[i];
       gomp_mutex_lock (&devicep->lock);
       if (devicep->state == GOMP_DEVICE_INITIALIZED)
-	{
-	  ret = devicep->fini_device_func (devicep->target_id);
-	  devicep->state = GOMP_DEVICE_FINALIZED;
-	}
+	ret = gomp_fini_device (devicep);
       gomp_mutex_unlock (&devicep->lock);
       if (!ret)
 	gomp_fatal ("device finalization failed");
Index: libgomp/libgomp.h
===================================================================
--- libgomp/libgomp.h	(revision 269183)
+++ libgomp/libgomp.h	(working copy)
@@ -949,25 +949,32 @@ typedef struct acc_dispatch_t
   /* Execute.  */
   __typeof (GOMP_OFFLOAD_openacc_exec) *exec_func;
 
-  /* Async cleanup callback registration.  */
-  __typeof (GOMP_OFFLOAD_openacc_register_async_cleanup)
-    *register_async_cleanup_func;
-
-  /* Asynchronous routines.  */
-  __typeof (GOMP_OFFLOAD_openacc_async_test) *async_test_func;
-  __typeof (GOMP_OFFLOAD_openacc_async_test_all) *async_test_all_func;
-  __typeof (GOMP_OFFLOAD_openacc_async_wait) *async_wait_func;
-  __typeof (GOMP_OFFLOAD_openacc_async_wait_async) *async_wait_async_func;
-  __typeof (GOMP_OFFLOAD_openacc_async_wait_all) *async_wait_all_func;
-  __typeof (GOMP_OFFLOAD_openacc_async_wait_all_async)
-    *async_wait_all_async_func;
-  __typeof (GOMP_OFFLOAD_openacc_async_set_async) *async_set_async_func;
-
   /* Create/destroy TLS data.  */
   __typeof (GOMP_OFFLOAD_openacc_create_thread_data) *create_thread_data_func;
   __typeof (GOMP_OFFLOAD_openacc_destroy_thread_data)
     *destroy_thread_data_func;
+  
+  struct {
+    /* Once created and put into the "active" list, asyncqueues are then never
+       destructed and removed from the "active" list, other than if the TODO
+       device is shut down.  */
+    gomp_mutex_t lock;
+    int nasyncqueue;
+    struct goacc_asyncqueue **asyncqueue;
+    struct goacc_asyncqueue_list *active;
 
+    __typeof (GOMP_OFFLOAD_openacc_async_construct) *construct_func;
+    __typeof (GOMP_OFFLOAD_openacc_async_destruct) *destruct_func;
+    __typeof (GOMP_OFFLOAD_openacc_async_test) *test_func;
+    __typeof (GOMP_OFFLOAD_openacc_async_synchronize) *synchronize_func;
+    __typeof (GOMP_OFFLOAD_openacc_async_serialize) *serialize_func;
+    __typeof (GOMP_OFFLOAD_openacc_async_queue_callback) *queue_callback_func;
+
+    __typeof (GOMP_OFFLOAD_openacc_async_exec) *exec_func;
+    __typeof (GOMP_OFFLOAD_openacc_async_dev2host) *dev2host_func;
+    __typeof (GOMP_OFFLOAD_openacc_async_host2dev) *host2dev_func;
+  } async;
+
   /* NVIDIA target specific routines.  */
   struct {
     __typeof (GOMP_OFFLOAD_openacc_cuda_get_current_device)
@@ -1053,17 +1060,33 @@ enum gomp_map_vars_kind
   GOMP_MAP_VARS_ENTER_DATA
 };
 
-extern void gomp_acc_insert_pointer (size_t, void **, size_t *, void *);
+extern void gomp_acc_insert_pointer (size_t, void **, size_t *, void *, int);
 extern void gomp_acc_remove_pointer (void *, size_t, bool, int, int, int);
 extern void gomp_acc_declare_allocate (bool, size_t, void **, size_t *,
 				       unsigned short *);
+struct gomp_coalesce_buf;
+extern void gomp_copy_host2dev (struct gomp_device_descr *,
+				struct goacc_asyncqueue *, void *, const void *,
+				size_t, struct gomp_coalesce_buf *);
+extern void gomp_copy_dev2host (struct gomp_device_descr *,
+				struct goacc_asyncqueue *, void *, const void *,
+				size_t);
 
 extern struct target_mem_desc *gomp_map_vars (struct gomp_device_descr *,
 					      size_t, void **, void **,
 					      size_t *, void *, bool,
 					      enum gomp_map_vars_kind);
+extern struct target_mem_desc *gomp_map_vars_async (struct gomp_device_descr *,
+						    struct goacc_asyncqueue *,
+						    size_t, void **, void **,
+						    size_t *, void *, bool,
+						    enum gomp_map_vars_kind);
+extern void gomp_unmap_tgt (struct target_mem_desc *);
 extern void gomp_unmap_vars (struct target_mem_desc *, bool);
+extern void gomp_unmap_vars_async (struct target_mem_desc *, bool,
+				   struct goacc_asyncqueue *);
 extern void gomp_init_device (struct gomp_device_descr *);
+extern bool gomp_fini_device (struct gomp_device_descr *);
 extern void gomp_free_memmap (struct splay_tree_s *);
 extern void gomp_unload_device (struct gomp_device_descr *);
 extern bool gomp_remove_var (struct gomp_device_descr *, splay_tree_key);
Index: libgomp/oacc-int.h
===================================================================
--- libgomp/oacc-int.h	(revision 269183)
+++ libgomp/oacc-int.h	(working copy)
@@ -99,6 +99,13 @@ void goacc_restore_bind (void);
 void goacc_lazy_initialize (void);
 void goacc_host_init (void);
 
+void goacc_init_asyncqueues (struct gomp_device_descr *);
+bool goacc_fini_asyncqueues (struct gomp_device_descr *);
+void goacc_async_free (struct gomp_device_descr *, struct goacc_asyncqueue *,
+		       void *);
+struct goacc_asyncqueue *get_goacc_asyncqueue (int);
+struct goacc_asyncqueue *lookup_goacc_asyncqueue (struct goacc_thread *, bool,
+						  int);
 static inline bool
 async_valid_stream_id_p (int async)
 {
Index: libgomp/oacc-host.c
===================================================================
--- libgomp/oacc-host.c	(revision 269183)
+++ libgomp/oacc-host.c	(working copy)
@@ -140,57 +140,91 @@ host_openacc_exec (void (*fn) (void *),
 		   size_t mapnum __attribute__ ((unused)),
 		   void **hostaddrs,
 		   void **devaddrs __attribute__ ((unused)),
-		   int async __attribute__ ((unused)),
-		   unsigned *dims __attribute ((unused)),
+		   unsigned *dims __attribute__ ((unused)),
 		   void *targ_mem_desc __attribute__ ((unused)))
 {
   fn (hostaddrs);
 }
 
 static void
-host_openacc_register_async_cleanup (void *targ_mem_desc __attribute__ ((unused)),
-				     int async __attribute__ ((unused)))
+host_openacc_async_exec (void (*fn) (void *),
+			 size_t mapnum __attribute__ ((unused)),
+			 void **hostaddrs,
+			 void **devaddrs __attribute__ ((unused)),
+			 unsigned *dims __attribute__ ((unused)),
+			 void *targ_mem_desc __attribute__ ((unused)),
+			 struct goacc_asyncqueue *aq __attribute__ ((unused)))
 {
+  fn (hostaddrs);
 }
 
 static int
-host_openacc_async_test (int async __attribute__ ((unused)))
+host_openacc_async_test (struct goacc_asyncqueue *aq __attribute__ ((unused)))
 {
   return 1;
 }
 
-static int
-host_openacc_async_test_all (void)
+static bool
+host_openacc_async_synchronize (struct goacc_asyncqueue *aq
+				__attribute__ ((unused)))
 {
-  return 1;
+  return true;
 }
 
-static void
-host_openacc_async_wait (int async __attribute__ ((unused)))
+static bool
+host_openacc_async_serialize (struct goacc_asyncqueue *aq1
+			      __attribute__ ((unused)),
+			      struct goacc_asyncqueue *aq2
+			      __attribute__ ((unused)))
 {
+  return true;
 }
 
-static void
-host_openacc_async_wait_async (int async1 __attribute__ ((unused)),
-			       int async2 __attribute__ ((unused)))
+static bool
+host_openacc_async_host2dev (int ord __attribute__ ((unused)),
+			     void *dst __attribute__ ((unused)),
+			     const void *src __attribute__ ((unused)),
+			     size_t n __attribute__ ((unused)),
+			     struct goacc_asyncqueue *aq
+			     __attribute__ ((unused)))
 {
+  return true;
 }
 
-static void
-host_openacc_async_wait_all (void)
+static bool
+host_openacc_async_dev2host (int ord __attribute__ ((unused)),
+			     void *dst __attribute__ ((unused)),
+			     const void *src __attribute__ ((unused)),
+			     size_t n __attribute__ ((unused)),
+			     struct goacc_asyncqueue *aq
+			     __attribute__ ((unused)))
 {
+  return true;
 }
 
 static void
-host_openacc_async_wait_all_async (int async __attribute__ ((unused)))
+host_openacc_async_queue_callback (struct goacc_asyncqueue *aq
+				   __attribute__ ((unused)),
+				   void (*callback_fn)(void *)
+				   __attribute__ ((unused)),
+				   void *userptr __attribute__ ((unused)))
 {
 }
 
-static void
-host_openacc_async_set_async (int async __attribute__ ((unused)))
+static struct goacc_asyncqueue *
+host_openacc_async_construct (void)
 {
+  /* Non-NULL 0xffff... value as opaque dummy.  */
+  return (struct goacc_asyncqueue *) -1;
 }
 
+static bool
+host_openacc_async_destruct (struct goacc_asyncqueue *aq
+			     __attribute__ ((unused)))
+{
+  return true;
+}
+
 static void *
 host_openacc_create_thread_data (int ord __attribute__ ((unused)))
 {
@@ -235,19 +269,21 @@ static struct gomp_device_descr host_dispatch =
 
       .exec_func = host_openacc_exec,
 
-      .register_async_cleanup_func = host_openacc_register_async_cleanup,
-
-      .async_test_func = host_openacc_async_test,
-      .async_test_all_func = host_openacc_async_test_all,
-      .async_wait_func = host_openacc_async_wait,
-      .async_wait_async_func = host_openacc_async_wait_async,
-      .async_wait_all_func = host_openacc_async_wait_all,
-      .async_wait_all_async_func = host_openacc_async_wait_all_async,
-      .async_set_async_func = host_openacc_async_set_async,
-
       .create_thread_data_func = host_openacc_create_thread_data,
       .destroy_thread_data_func = host_openacc_destroy_thread_data,
 
+      .async = {
+	.construct_func = host_openacc_async_construct,
+	.destruct_func = host_openacc_async_destruct,
+	.test_func = host_openacc_async_test,
+	.synchronize_func = host_openacc_async_synchronize,
+	.serialize_func = host_openacc_async_serialize,
+	.queue_callback_func = host_openacc_async_queue_callback,
+	.exec_func = host_openacc_async_exec,
+	.dev2host_func = host_openacc_async_dev2host,
+	.host2dev_func = host_openacc_async_host2dev,
+      },
+
       .cuda = {
 	.get_current_device_func = NULL,
 	.get_current_context_func = NULL,
Index: libgomp/libgomp-plugin.h
===================================================================
--- libgomp/libgomp-plugin.h	(revision 269183)
+++ libgomp/libgomp-plugin.h	(working copy)
@@ -53,6 +53,20 @@ enum offload_target_type
   OFFLOAD_TARGET_TYPE_HSA = 7
 };
 
+/* Opaque type to represent plugin-dependent implementation of an
+   OpenACC asynchronous queue.  */
+struct goacc_asyncqueue;
+
+/* Used to keep a list of active asynchronous queues.  */
+struct goacc_asyncqueue_list
+{
+  struct goacc_asyncqueue *aq;
+  struct goacc_asyncqueue_list *next;
+};
+
+typedef struct goacc_asyncqueue *goacc_aq;
+typedef struct goacc_asyncqueue_list *goacc_aq_list;
+
 /* Auxiliary struct, used for transferring pairs of addresses from plugin
    to libgomp.  */
 struct addr_pair
@@ -93,22 +107,31 @@ extern bool GOMP_OFFLOAD_dev2dev (int, void *, con
 extern bool GOMP_OFFLOAD_can_run (void *);
 extern void GOMP_OFFLOAD_run (int, void *, void *, void **);
 extern void GOMP_OFFLOAD_async_run (int, void *, void *, void **, void *);
+
 extern void GOMP_OFFLOAD_openacc_exec (void (*) (void *), size_t, void **,
-				       void **, int, unsigned *, void *);
-extern void GOMP_OFFLOAD_openacc_register_async_cleanup (void *, int);
-extern int GOMP_OFFLOAD_openacc_async_test (int);
-extern int GOMP_OFFLOAD_openacc_async_test_all (void);
-extern void GOMP_OFFLOAD_openacc_async_wait (int);
-extern void GOMP_OFFLOAD_openacc_async_wait_async (int, int);
-extern void GOMP_OFFLOAD_openacc_async_wait_all (void);
-extern void GOMP_OFFLOAD_openacc_async_wait_all_async (int);
-extern void GOMP_OFFLOAD_openacc_async_set_async (int);
+				       void **, unsigned *, void *);
 extern void *GOMP_OFFLOAD_openacc_create_thread_data (int);
 extern void GOMP_OFFLOAD_openacc_destroy_thread_data (void *);
+extern struct goacc_asyncqueue *GOMP_OFFLOAD_openacc_async_construct (void);
+extern bool GOMP_OFFLOAD_openacc_async_destruct (struct goacc_asyncqueue *);
+extern int GOMP_OFFLOAD_openacc_async_test (struct goacc_asyncqueue *);
+extern bool GOMP_OFFLOAD_openacc_async_synchronize (struct goacc_asyncqueue *);
+extern bool GOMP_OFFLOAD_openacc_async_serialize (struct goacc_asyncqueue *,
+						  struct goacc_asyncqueue *);
+extern void GOMP_OFFLOAD_openacc_async_queue_callback (struct goacc_asyncqueue *,
+						       void (*)(void *), void *);
+extern void GOMP_OFFLOAD_openacc_async_exec (void (*) (void *), size_t, void **,
+					     void **, unsigned *, void *,
+					     struct goacc_asyncqueue *);
+extern bool GOMP_OFFLOAD_openacc_async_dev2host (int, void *, const void *, size_t,
+						 struct goacc_asyncqueue *);
+extern bool GOMP_OFFLOAD_openacc_async_host2dev (int, void *, const void *, size_t,
+						 struct goacc_asyncqueue *);
 extern void *GOMP_OFFLOAD_openacc_cuda_get_current_device (void);
 extern void *GOMP_OFFLOAD_openacc_cuda_get_current_context (void);
-extern void *GOMP_OFFLOAD_openacc_cuda_get_stream (int);
-extern int GOMP_OFFLOAD_openacc_cuda_set_stream (int, void *);
+extern void *GOMP_OFFLOAD_openacc_cuda_get_stream (struct goacc_asyncqueue *);
+extern int GOMP_OFFLOAD_openacc_cuda_set_stream (struct goacc_asyncqueue *,
+						 void *);
 
 #ifdef __cplusplus
 }


More information about the Gcc-patches mailing list