[OpenMP] Optimized stream selection by scheduling data mapping for the same target region into a same stream

[tianshilei1992]

Summary

This patch introduces two things for offloading:

1. Asynchronous data transferring: those functions are suffix with _async. They have one more argument compared with their synchronous counterparts: __tgt_async_info*, which is a new struct that only has one field, void *Identifier. This struct is for information exchange between different asynchronous operations. It can be used for stream selection, like in this case, or operation synchronization, which is also used. We may expect more usages in the future.
2. Optimization of stream selection for data mapping. Previous implementation was using asynchronous device memory transfer but synchronizing after each memory transfer. Actually, if we say kernel A needs four memory copy to device and two memory copy back to host, then we can schedule these seven operations (four H2D, two D2H, and one kernel launch) into a same stream and just need synchronization after memory copy from device to host. In this way, we can save a huge overhead compared with synchronization after each operation.

[lildmh]

openmp/libomptarget/include/omptarget.h

119

Do you need a clang patch to generate code for this structure and new APIs?

openmp/libomptarget/plugins/cuda/src/rtl.cpp

912

It seems no synchronization by default, not aligned with the original behavior

[tianshilei1992]

openmp/libomptarget/include/omptarget.h

119

Currently no. But of course this information can be used by Clang in the future if needed.

openmp/libomptarget/plugins/cuda/src/rtl.cpp

912

The assumption here is: if this function is called at the end of target function, Identifier must be a valid pointer because anyhow during kernel launch it is assigned with a valid CUstream.

[jdoerfert]

openmp/libomptarget/include/omptarget.h

119

We'll target the new API also/mainly by rewriting old API calls in OpenMPOpt

openmp/libomptarget/plugins/cuda/src/rtl.cpp

912

It seems wrong to call this with Identifier = nullptr, doesn't it? If so we should make it an assert.

916

Why reinterpret_Cast? Is a CUstream a pointer? If not, why do we move objects around?

[tianshilei1992]

openmp/libomptarget/plugins/cuda/src/rtl.cpp

912

Yeah, seems right. "If there is nothing to be synchronized, what do users expect to synchronize?" Yes, should do an assert here.

916

Yep.

typedef struct CUstream_st *CUstream;                     /**< CUDA stream */

[jdoerfert]

openmp/libomptarget/include/omptarget.h

118

I don't like Identifier as name. Maybe execution queue? Or the name we used for streams before? Also, add a comment what it is and why it is a void *

openmp/libomptarget/plugins/cuda/src/rtl.cpp

320

Nit: I know the style in here is mixed but I would remove the braces around single statements, especially return.

916

then we should be able to just use a static cast, shouldn't we? Maybe also add a static_assert just to be safe. Also other places.

openmp/libomptarget/src/device.h

188

Remove the old versions if we don't expose them to the outside. Let's not accumulate clutter. Similarly for other functions.

openmp/libomptarget/src/omptarget.cpp

646

Move this to the struct definition. It's C++ after all.

[tianshilei1992]

openmp/libomptarget/include/omptarget.h

118

The basic idea is, this structure is device independent. For CUDA device, we use it as CUstream so yes, we can call it stream. However, we don't know how it can be used in other platforms, and that's why I use a "weird" word Identifier and a void * here.

openmp/libomptarget/plugins/cuda/src/rtl.cpp

320

Speaking of this, is it required in LLVM that single-statement should not be wrapped into braces? It is often good practice to have these braces even just with one statement.

916

Correct me if I'm wrong. Compiler will generate code for the cast if using static_cast, but it will not for reinterpret_cast. My thought is, since it is already in CUDA plugin, we can make sure that this void * is actually a CUstream so we can directly reinterpret it to avoid one cast instruction.

openmp/libomptarget/src/device.h

188

Do we need to make AsyncInfoPtr with default value?

openmp/libomptarget/src/omptarget.cpp

646

Yep. Will do that.

[jdoerfert]

openmp/libomptarget/include/omptarget.h

118

We need a better word and documentation of the member. Also initialize it here

openmp/libomptarget/plugins/cuda/src/rtl.cpp

320

Coding style says no braces but the runtime is unfortunately not in accordance...

916

It could but not for such a pointer 2 pointer cast on a "normal" architecture. It's just that I don't see reinterpret_cast often, almost 1:4 against static_cast, but I'm fine with leaving it.

openmp/libomptarget/src/device.h

188

Not if we can modify all callers (which we can if the symbols is not exported). If we can't we should add the new API calls as you have.

[tianshilei1992]

openmp/libomptarget/include/omptarget.h

118

Sure. Do you have any suggestion for the name?

openmp/libomptarget/plugins/cuda/src/rtl.cpp

320

Okay. Will change that.

openmp/libomptarget/src/device.h

188

Got you. Will do that.

[jdoerfert]

openmp/libomptarget/include/omptarget.h

118

IssueQueue?

[tianshilei1992]

Update code according to review

[jdoerfert]

openmp/libomptarget/src/rtl.h

39

Similar to my last comment, why do we need async versions here? Let them be the default (w/o the _async).

[tianshilei1992]

Removed all synchronous version and make asynchronous as default

[tianshilei1992]

Added comments on why we need conditional statements in __tgt_rtl_data_retrieve and __tgt_rtl_data_submit

[jdoerfert]

I think this is good. Do you happen to have some runtime results?

[tianshilei1992]

In D77005#1961961, @jdoerfert wrote:

I think this is good. Do you happen to have some runtime results?

Not yet. Just got accel-1.3 and didn't have a chance to execute them yet. Will do maybe tomorrow.

[tianshilei1992]

Fixed an issue that passing arguments to dataSubmit in a wrong order

[jdoerfert]

One nit, otherwise LGTM.

openmp/libomptarget/plugins/cuda/src/rtl.cpp

740

Style: Remove the else case, there is a return before. Same below.

[tianshilei1992]

openmp/libomptarget/plugins/cuda/src/rtl.cpp

740

Sorry, didn't get your point. The return above only applies when rc != OFFLOAD_SUCCESS.

[tianshilei1992]

openmp/libomptarget/plugins/cuda/src/rtl.cpp

740

Okay, will do that.

[tianshilei1992]

Remove else after return

[jdoerfert]

LGTM

openmp/libomptarget/include/omptarget.h

 
 /// This struct contains the offload entries identified by the target runtime
 struct __tgt_target_table {
   __tgt_offload_entry *EntriesBegin; // Begin of the table with all the entries
   __tgt_offload_entry
       *EntriesEnd; // End of the table with all the entries (non inclusive)
 };
 
+/// This struct contains information exchanged between different asynchronous
+/// operations for device-dependent optimization and potential synchronization
+struct __tgt_async_info {
+  // A pointer to a queue-like structure where offloading operations are issued.
+  // We assume to use this structure to do synchronization. In CUDA backend, it
+  // is CUstream.
+  void *Queue = nullptr;
+};
+
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 int omp_get_num_devices(void);
 int omp_get_initial_device(void);
 void *omp_target_alloc(size_t size, int device_num);
 void omp_target_free(void *device_ptr, int device_num);

openmp/libomptarget/include/omptargetplugin.h

 // is allocated by omp_target_alloc() API). Return address of the
 // allocated data on the target that will be used by libomptarget.so to
 // initialize the target data mapping structures. These addresses are
 // used to generate a table of target variables to pass to
 // __tgt_rtl_run_region(). The __tgt_rtl_data_alloc() returns NULL in
 // case an error occurred on the target device.
 void *__tgt_rtl_data_alloc(int32_t ID, int64_t Size, void *HostPtr);
 
-// Pass the data content to the target device using the target address.
-// In case of success, return zero. Otherwise, return an error code.
+// Pass the data content to the target device using the target address. If
+// AsyncInfoPtr is nullptr, it is synchronous; otherwise it is asynchronous.
+// However, AsyncInfoPtr may be ignored on some platforms, like x86_64. In that
+// case, it is synchronous. In case of success, return zero. Otherwise, return
+// an error code.
 int32_t __tgt_rtl_data_submit(int32_t ID, void *TargetPtr, void *HostPtr,
-                              int64_t Size);
+                              int64_t Size, __tgt_async_info *AsyncInfoPtr);
 
-// Retrieve the data content from the target device using its address.
-// In case of success, return zero. Otherwise, return an error code.
+// Retrieve the data content from the target device using its address. If
+// AsyncInfoPtr is nullptr, it is synchronous; otherwise it is asynchronous.
+// However, AsyncInfoPtr may be ignored on some platforms, like x86_64. In that
+// case, it is synchronous. In case of success, return zero. Otherwise, return
+// an error code.
 int32_t __tgt_rtl_data_retrieve(int32_t ID, void *HostPtr, void *TargetPtr,
-                                int64_t Size);
+                                int64_t Size, __tgt_async_info *AsyncInfoPtr);
 
 // De-allocate the data referenced by target ptr on the device. In case of
 // success, return zero. Otherwise, return an error code.
 int32_t __tgt_rtl_data_delete(int32_t ID, void *TargetPtr);
 
 // Transfer control to the offloaded entry Entry on the target device.
 // Args and Offsets are arrays of NumArgs size of target addresses and
 // offsets. An offset should be added to the target address before passing it
-// to the outlined function on device side. In case of success, return zero.
-// Otherwise, return an error code.
+// to the outlined function on device side. If AsyncInfoPtr is nullptr, it is
+// synchronous; otherwise it is asynchronous. However, AsyncInfoPtr may be
+// ignored on some platforms, like x86_64. In that case, it is synchronous. In
+// case of success, return zero. Otherwise, return an error code.
 int32_t __tgt_rtl_run_target_region(int32_t ID, void *Entry, void **Args,
-                                    ptrdiff_t *Offsets, int32_t NumArgs);
+                                    ptrdiff_t *Offsets, int32_t NumArgs,
+                                    __tgt_async_info *AsyncInfoPtr);
 
 // Similar to __tgt_rtl_run_target_region, but additionally specify the
-// number of teams to be created and a number of threads in each team.
+// number of teams to be created and a number of threads in each team. If
+// AsyncInfoPtr is nullptr, it is synchronous; otherwise it is asynchronous.
+// However, AsyncInfoPtr may be ignored on some platforms, like x86_64. In that
+// case, it is synchronous.
 int32_t __tgt_rtl_run_target_team_region(int32_t ID, void *Entry, void **Args,
                                          ptrdiff_t *Offsets, int32_t NumArgs,
                                          int32_t NumTeams, int32_t ThreadLimit,
-                                         uint64_t loop_tripcount);
+                                         uint64_t loop_tripcount,
+                                         __tgt_async_info *AsyncInfoPtr);
+
+// Device synchronization. In case of success, return zero. Otherwise, return an
+// error code.
+int32_t __tgt_rtl_synchronize(int32_t ID, __tgt_async_info *AsyncInfoPtr);
 
 #ifdef __cplusplus
 }
 #endif
 
 #endif // _OMPTARGETPLUGIN_H_

openmp/libomptarget/plugins/cuda/src/rtl.cpp

for (auto &ctx : Contexts)
           CUDA_ERR_STRING(err);
         }
       }
   }
 };
 
 static RTLDeviceInfoTy DeviceInfo;
 
+namespace {
+CUstream selectStream(int32_t Id, __tgt_async_info *AsyncInfo) {
+  if (!AsyncInfo)
+    return DeviceInfo.getNextStream(Id);
+
+  if (!AsyncInfo->Queue)
+    AsyncInfo->Queue = DeviceInfo.getNextStream(Id);
+
+  return reinterpret_cast<CUstream>(AsyncInfo->Queue);
+}
+
+int32_t dataRetrieve(int32_t DeviceId, void *HstPtr, void *TgtPtr, int64_t Size,
+                     __tgt_async_info *AsyncInfoPtr) {
+  assert(AsyncInfoPtr && "AsyncInfoPtr is nullptr");
+  // Set the context we are using.
+  CUresult err = cuCtxSetCurrent(DeviceInfo.Contexts[DeviceId]);
+  if (err != CUDA_SUCCESS) {
+    DP("Error when setting CUDA context\n");
+    CUDA_ERR_STRING(err);
+    return OFFLOAD_FAIL;
+  }
+
+  CUstream Stream = selectStream(DeviceId, AsyncInfoPtr);
+
+  err = cuMemcpyDtoHAsync(HstPtr, (CUdeviceptr)TgtPtr, Size, Stream);
+  if (err != CUDA_SUCCESS) {
+    DP("Error when copying data from device to host. Pointers: host = " DPxMOD
+       ", device = " DPxMOD ", size = %" PRId64 "\n",
+       DPxPTR(HstPtr), DPxPTR(TgtPtr), Size);
+    CUDA_ERR_STRING(err);
+    return OFFLOAD_FAIL;
+  }
+
+  return OFFLOAD_SUCCESS;
+}
+
+int32_t dataSubmit(int32_t DeviceId, void *TgtPtr, void *HstPtr, int64_t Size,
+                   __tgt_async_info *AsyncInfoPtr) {
+  assert(AsyncInfoPtr && "AsyncInfoPtr is nullptr");
+  // Set the context we are using.
+  CUresult err = cuCtxSetCurrent(DeviceInfo.Contexts[DeviceId]);
+  if (err != CUDA_SUCCESS) {
+    DP("Error when setting CUDA context\n");
+    CUDA_ERR_STRING(err);
+    return OFFLOAD_FAIL;
+  }
+
+  CUstream Stream = selectStream(DeviceId, AsyncInfoPtr);
+
+  err = cuMemcpyHtoDAsync((CUdeviceptr)TgtPtr, HstPtr, Size, Stream);
+  if (err != CUDA_SUCCESS) {
+    DP("Error when copying data from host to device. Pointers: host = " DPxMOD
+       ", device = " DPxMOD ", size = %" PRId64 "\n",
+       DPxPTR(HstPtr), DPxPTR(TgtPtr), Size);
+    CUDA_ERR_STRING(err);
+    return OFFLOAD_FAIL;
+  }
+
+  return OFFLOAD_SUCCESS;
+}
+} // namespace
+
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 int32_t __tgt_rtl_is_valid_binary(__tgt_device_image *image) {
   return elf_check_machine(image, 190); // EM_CUDA = 190.
 }
 
if (err != CUDA_SUCCESS) {
     return NULL;
   }
 
   void *vptr = (void *)ptr;
   return vptr;
 }
 
 int32_t __tgt_rtl_data_submit(int32_t device_id, void *tgt_ptr, void *hst_ptr,
-    int64_t size) {
-  // Set the context we are using.
-  CUresult err = cuCtxSetCurrent(DeviceInfo.Contexts[device_id]);
-  if (err != CUDA_SUCCESS) {
-    DP("Error when setting CUDA context\n");
-    CUDA_ERR_STRING(err);
-    return OFFLOAD_FAIL;
-  }
-
-  CUstream &Stream = DeviceInfo.getNextStream(device_id);
+                              int64_t size, __tgt_async_info *async_info_ptr) {
+  if (async_info_ptr) {
+    return dataSubmit(device_id, hst_ptr, tgt_ptr, size, async_info_ptr);
+  } else {
+    __tgt_async_info async_info;
 
-  err = cuMemcpyHtoDAsync((CUdeviceptr)tgt_ptr, hst_ptr, size, Stream);
-  if (err != CUDA_SUCCESS) {
-    DP("Error when copying data from host to device. Pointers: host = " DPxMOD
-       ", device = " DPxMOD ", size = %" PRId64 "\n",
-       DPxPTR(hst_ptr), DPxPTR(tgt_ptr), size);
-    CUDA_ERR_STRING(err);
+    const int32_t rc =
+        dataSubmit(device_id, hst_ptr, tgt_ptr, size, &async_info);
+    if (rc != OFFLOAD_SUCCESS) {
       return OFFLOAD_FAIL;
     }
 
-  err = cuStreamSynchronize(Stream);
-  if (err != CUDA_SUCCESS) {
-    DP("Error when synchronizing async data transfer from host to device. "
-       "Pointers: host = " DPxMOD ", device = " DPxMOD ", size = %" PRId64 "\n",
-       DPxPTR(hst_ptr), DPxPTR(tgt_ptr), size);
-    CUDA_ERR_STRING(err);
-    return OFFLOAD_FAIL;
+    return __tgt_rtl_synchronize(device_id, &async_info);
   }
-
-  return OFFLOAD_SUCCESS;
 }
 
 int32_t __tgt_rtl_data_retrieve(int32_t device_id, void *hst_ptr, void *tgt_ptr,
-    int64_t size) {
-  // Set the context we are using.
-  CUresult err = cuCtxSetCurrent(DeviceInfo.Contexts[device_id]);
-  if (err != CUDA_SUCCESS) {
-    DP("Error when setting CUDA context\n");
-    CUDA_ERR_STRING(err);
+                                int64_t size,
+                                __tgt_async_info *async_info_ptr) {
+  if (async_info_ptr) {
+    return dataRetrieve(device_id, hst_ptr, tgt_ptr, size, async_info_ptr);
+  } else {
+    __tgt_async_info async_info;
+    const int32_t rc =
+        dataRetrieve(device_id, hst_ptr, tgt_ptr, size, &async_info);
+    if (rc != OFFLOAD_SUCCESS) {
       return OFFLOAD_FAIL;
     }
-
-  CUstream &Stream = DeviceInfo.getNextStream(device_id);
-
-  err = cuMemcpyDtoHAsync(hst_ptr, (CUdeviceptr)tgt_ptr, size, Stream);
-  if (err != CUDA_SUCCESS) {
-    DP("Error when copying data from device to host. Pointers: host = " DPxMOD
-       ", device = " DPxMOD ", size = %" PRId64 "\n",
-       DPxPTR(hst_ptr), DPxPTR(tgt_ptr), size);
-    CUDA_ERR_STRING(err);
-    return OFFLOAD_FAIL;
+    return __tgt_rtl_synchronize(device_id, &async_info);
   }
-
-  err = cuStreamSynchronize(Stream);
-  if (err != CUDA_SUCCESS) {
-    DP("Error when synchronizing async data transfer from device to host. "
-       "Pointers: host = " DPxMOD ", device = " DPxMOD ", size = %" PRId64 "\n",
-       DPxPTR(hst_ptr), DPxPTR(tgt_ptr), size);
-    CUDA_ERR_STRING(err);
-    return OFFLOAD_FAIL;
-  }
-
-  return OFFLOAD_SUCCESS;
 }
 
 int32_t __tgt_rtl_data_delete(int32_t device_id, void *tgt_ptr) {
   // Set the context we are using.
   CUresult err = cuCtxSetCurrent(DeviceInfo.Contexts[device_id]);
   if (err != CUDA_SUCCESS) {
     DP("Error when setting CUDA context\n");
     CUDA_ERR_STRING(err);
     return OFFLOAD_FAIL;
   }
 
   err = cuMemFree((CUdeviceptr)tgt_ptr);
   if (err != CUDA_SUCCESS) {
     DP("Error when freeing CUDA memory\n");
     CUDA_ERR_STRING(err);
     return OFFLOAD_FAIL;
   }
   return OFFLOAD_SUCCESS;
 }
 
 int32_t __tgt_rtl_run_target_team_region(int32_t device_id, void *tgt_entry_ptr,
-    void **tgt_args, ptrdiff_t *tgt_offsets, int32_t arg_num, int32_t team_num,
-    int32_t thread_limit, uint64_t loop_tripcount) {
+                                         void **tgt_args,
+                                         ptrdiff_t *tgt_offsets,
+                                         int32_t arg_num, int32_t team_num,
+                                         int32_t thread_limit,
+                                         uint64_t loop_tripcount,
+                                         __tgt_async_info *async_info) {
   // Set the context we are using.
   CUresult err = cuCtxSetCurrent(DeviceInfo.Contexts[device_id]);
   if (err != CUDA_SUCCESS) {
     DP("Error when setting CUDA context\n");
     CUDA_ERR_STRING(err);
     return OFFLOAD_FAIL;
   }
 
if (team_num <= 0) {
     cudaBlocksPerGrid = team_num;
     DP("Using requested number of teams %d\n", team_num);
   }
 
   // Run on the device.
   DP("Launch kernel with %d blocks and %d threads\n", cudaBlocksPerGrid,
      cudaThreadsPerBlock);
 
-  CUstream &Stream = DeviceInfo.getNextStream(device_id);
-
+  CUstream Stream = selectStream(device_id, async_info);
   err = cuLaunchKernel(KernelInfo->Func, cudaBlocksPerGrid, 1, 1,
                        cudaThreadsPerBlock, 1, 1, 0 /*bytes of shared memory*/,
                        Stream, &args[0], 0);
   if (err != CUDA_SUCCESS) {
     DP("Device kernel launch failed!\n");
     CUDA_ERR_STRING(err);
     return OFFLOAD_FAIL;
   }
 
   DP("Launch of entry point at " DPxMOD " successful!\n",
       DPxPTR(tgt_entry_ptr));
 
-  CUresult sync_err = cuStreamSynchronize(Stream);
-  if (sync_err != CUDA_SUCCESS) {
-    DP("Kernel execution error at " DPxMOD "!\n", DPxPTR(tgt_entry_ptr));
-    CUDA_ERR_STRING(sync_err);
-    return OFFLOAD_FAIL;
-  } else {
-    DP("Kernel execution at " DPxMOD " successful!\n", DPxPTR(tgt_entry_ptr));
-  }
-
   return OFFLOAD_SUCCESS;
 }
 
 int32_t __tgt_rtl_run_target_region(int32_t device_id, void *tgt_entry_ptr,
-    void **tgt_args, ptrdiff_t *tgt_offsets, int32_t arg_num) {
+                                    void **tgt_args, ptrdiff_t *tgt_offsets,
+                                    int32_t arg_num,
+                                    __tgt_async_info *async_info) {
   // use one team and the default number of threads.
   const int32_t team_num = 1;
   const int32_t thread_limit = 0;
   return __tgt_rtl_run_target_team_region(device_id, tgt_entry_ptr, tgt_args,
-      tgt_offsets, arg_num, team_num, thread_limit, 0);
+                                          tgt_offsets, arg_num, team_num,
+                                          thread_limit, 0, async_info);
+}
+
+int32_t __tgt_rtl_synchronize(int32_t device_id, __tgt_async_info *async_info) {
+  assert(async_info && "async_info is nullptr");
+  assert(async_info->Queue && "async_info->Queue is nullptr");
+
+  CUstream Stream = reinterpret_cast<CUstream>(async_info->Queue);
+  CUresult Err = cuStreamSynchronize(Stream);
+  if (Err != CUDA_SUCCESS) {
+    DP("Error when synchronizing stream. stream = " DPxMOD
+       ", async info ptr = " DPxMOD "\n",
+       DPxPTR(Stream), DPxPTR(async_info));
+    CUDA_ERR_STRING(Err);
+    return OFFLOAD_FAIL;
+  }
+  return OFFLOAD_SUCCESS;
 }
 
 #ifdef __cplusplus
 }
 #endif

openmp/libomptarget/plugins/exports

 VERS1.0 {
   global:
     __tgt_rtl_is_valid_binary;
     __tgt_rtl_number_of_devices;
     __tgt_rtl_init_requires;
     __tgt_rtl_init_device;
     __tgt_rtl_load_binary;
     __tgt_rtl_data_alloc;
     __tgt_rtl_data_submit;
     __tgt_rtl_data_retrieve;
     __tgt_rtl_data_delete;
     __tgt_rtl_run_target_team_region;
     __tgt_rtl_run_target_region;
+    __tgt_rtl_synchronize;
   local:
     *;
 };

openmp/libomptarget/plugins/generic-elf-64bit/src/rtl.cpp

 }
 
 void *__tgt_rtl_data_alloc(int32_t device_id, int64_t size, void *hst_ptr) {
   void *ptr = malloc(size);
   return ptr;
 }
 
 int32_t __tgt_rtl_data_submit(int32_t device_id, void *tgt_ptr, void *hst_ptr,
-                              int64_t size) {
+                              int64_t size, __tgt_async_info *) {
   memcpy(tgt_ptr, hst_ptr, size);
   return OFFLOAD_SUCCESS;
 }
 
 int32_t __tgt_rtl_data_retrieve(int32_t device_id, void *hst_ptr, void *tgt_ptr,
-                                int64_t size) {
+                                int64_t size, __tgt_async_info *) {
   memcpy(hst_ptr, tgt_ptr, size);
   return OFFLOAD_SUCCESS;
 }
 
 int32_t __tgt_rtl_data_delete(int32_t device_id, void *tgt_ptr) {
   free(tgt_ptr);
   return OFFLOAD_SUCCESS;
 }
 
-int32_t __tgt_rtl_run_target_team_region(int32_t device_id, void *tgt_entry_ptr,
-    void **tgt_args, ptrdiff_t *tgt_offsets, int32_t arg_num, int32_t team_num,
-    int32_t thread_limit, uint64_t loop_tripcount /*not used*/) {
+int32_t __tgt_rtl_run_target_team_region(
+    int32_t device_id, void *tgt_entry_ptr, void **tgt_args,
+    ptrdiff_t *tgt_offsets, int32_t arg_num, int32_t team_num,
+    int32_t thread_limit, uint64_t loop_tripcount /*not used*/,
+    __tgt_async_info *async_info /*not used*/) {
   // ignore team num and thread limit.
 
   // Use libffi to launch execution.
   ffi_cif cif;
 
   // All args are references.
   std::vector<ffi_type *> args_types(arg_num, &ffi_type_pointer);
   std::vector<void *> args(arg_num);
int32_t __tgt_rtl_run_target_team_region(
 
   void (*entry)(void);
   *((void**) &entry) = tgt_entry_ptr;
   ffi_call(&cif, entry, NULL, &args[0]);
   return OFFLOAD_SUCCESS;
 }
 
 int32_t __tgt_rtl_run_target_region(int32_t device_id, void *tgt_entry_ptr,
-    void **tgt_args, ptrdiff_t *tgt_offsets, int32_t arg_num) {
+                                    void **tgt_args, ptrdiff_t *tgt_offsets,
+                                    int32_t arg_num,
+                                    __tgt_async_info *async_info_ptr) {
   // use one team and one thread.
   return __tgt_rtl_run_target_team_region(device_id, tgt_entry_ptr, tgt_args,
-      tgt_offsets, arg_num, 1, 1, 0);
+                                          tgt_offsets, arg_num, 1, 1, 0,
+                                          async_info_ptr);
+}
+
+int32_t __tgt_rtl_synchronize(int32_t device_id,
+                              __tgt_async_info *async_info_ptr) {
+  return OFFLOAD_SUCCESS;
 }
 
 #ifdef __cplusplus
 }
 #endif

openmp/libomptarget/src/api.cpp

if (src_device == omp_get_initial_device() &&
       dst_device == omp_get_initial_device()) {
     DP("copy from host to host\n");
     const void *p = memcpy(dstAddr, srcAddr, length);
     if (p == NULL)
       rc = OFFLOAD_FAIL;
   } else if (src_device == omp_get_initial_device()) {
     DP("copy from host to device\n");
     DeviceTy& DstDev = Devices[dst_device];
-    rc = DstDev.data_submit(dstAddr, srcAddr, length);
+    rc = DstDev.data_submit(dstAddr, srcAddr, length, nullptr);
   } else if (dst_device == omp_get_initial_device()) {
     DP("copy from device to host\n");
     DeviceTy& SrcDev = Devices[src_device];
-    rc = SrcDev.data_retrieve(dstAddr, srcAddr, length);
+    rc = SrcDev.data_retrieve(dstAddr, srcAddr, length, nullptr);
   } else {
     DP("copy from device to device\n");
     void *buffer = malloc(length);
     DeviceTy& SrcDev = Devices[src_device];
     DeviceTy& DstDev = Devices[dst_device];
-    rc = SrcDev.data_retrieve(buffer, srcAddr, length);
+    rc = SrcDev.data_retrieve(buffer, srcAddr, length, nullptr);
     if (rc == OFFLOAD_SUCCESS)
-      rc = DstDev.data_submit(dstAddr, buffer, length);
+      rc = DstDev.data_submit(dstAddr, buffer, length, nullptr);
     free(buffer);
   }
 
   DP("omp_target_memcpy returns %d\n", rc);
   return rc;
 }
 
 EXTERN int omp_target_memcpy_rect(void *dst, void *src, size_t element_size,

openmp/libomptarget/src/device.h

 #include <map>
 #include <mutex>
 #include <vector>
 
 // Forward declarations.
 struct RTLInfoTy;
 struct __tgt_bin_desc;
 struct __tgt_target_table;
+struct __tgt_async_info;
 
 /// Map between host data and target data.
 struct HostDataToTargetTy {
   uintptr_t HstPtrBase; // host info.
   uintptr_t HstPtrBegin;
   uintptr_t HstPtrEnd; // non-inclusive.
 
   uintptr_t TgtPtrBegin; // target info.
int deallocTgtPtr(void *TgtPtrBegin, int64_t Size, bool ForceDelete,
                     bool HasCloseModifier = false);
   int associatePtr(void *HstPtrBegin, void *TgtPtrBegin, int64_t Size);
   int disassociatePtr(void *HstPtrBegin);
 
   // calls to RTL
   int32_t initOnce();
   __tgt_target_table *load_binary(void *Img);
 
-  int32_t data_submit(void *TgtPtrBegin, void *HstPtrBegin, int64_t Size);
-  int32_t data_retrieve(void *HstPtrBegin, void *TgtPtrBegin, int64_t Size);
+  // Asynchronous data transfer. When AsyncInfoPtr is nullptr, the transfer will
+  // be synchronous.
+  int32_t data_submit(void *TgtPtrBegin, void *HstPtrBegin, int64_t Size,
+                      __tgt_async_info *AsyncInfoPtr);
+  int32_t data_retrieve(void *HstPtrBegin, void *TgtPtrBegin, int64_t Size,
+                        __tgt_async_info *AsyncInfoPtr);
 
   int32_t run_region(void *TgtEntryPtr, void **TgtVarsPtr,
-      ptrdiff_t *TgtOffsets, int32_t TgtVarsSize);
+                     ptrdiff_t *TgtOffsets, int32_t TgtVarsSize,
+                     __tgt_async_info *AsyncInfo);
   int32_t run_team_region(void *TgtEntryPtr, void **TgtVarsPtr,
-      ptrdiff_t *TgtOffsets, int32_t TgtVarsSize, int32_t NumTeams,
-      int32_t ThreadLimit, uint64_t LoopTripCount);
+                          ptrdiff_t *TgtOffsets, int32_t TgtVarsSize,
+                          int32_t NumTeams, int32_t ThreadLimit,
+                          uint64_t LoopTripCount, __tgt_async_info *AsyncInfo);
 
 private:
   // Call to RTL
   void init(); // To be called only via DeviceTy::initOnce()
 };
 
 /// Map between Device ID (i.e. openmp device id) and its DeviceTy.
 typedef std::vector<DeviceTy> DevicesTy;
 extern DevicesTy Devices;
 
 extern bool device_is_ready(int device_num);
 
 #endif

openmp/libomptarget/src/device.cpp

 // Load binary to device.
 __tgt_target_table *DeviceTy::load_binary(void *Img) {
   RTL->Mtx.lock();
   __tgt_target_table *rc = RTL->load_binary(RTLDeviceID, Img);
   RTL->Mtx.unlock();
   return rc;
 }
 
-// Submit data to device.
+// Submit data to device
 int32_t DeviceTy::data_submit(void *TgtPtrBegin, void *HstPtrBegin,
-    int64_t Size) {
-  return RTL->data_submit(RTLDeviceID, TgtPtrBegin, HstPtrBegin, Size);
+                              int64_t Size, __tgt_async_info *AsyncInfoPtr) {
+
+  return RTL->data_submit(RTLDeviceID, TgtPtrBegin, HstPtrBegin, Size,
+                          AsyncInfoPtr);
 }
 
-// Retrieve data from device.
+// Retrieve data from device
 int32_t DeviceTy::data_retrieve(void *HstPtrBegin, void *TgtPtrBegin,
-    int64_t Size) {
-  return RTL->data_retrieve(RTLDeviceID, HstPtrBegin, TgtPtrBegin, Size);
+                                int64_t Size, __tgt_async_info *AsyncInfoPtr) {
+  return RTL->data_retrieve(RTLDeviceID, HstPtrBegin, TgtPtrBegin, Size,
+                            AsyncInfoPtr);
 }
 
 // Run region on device
 int32_t DeviceTy::run_region(void *TgtEntryPtr, void **TgtVarsPtr,
-    ptrdiff_t *TgtOffsets, int32_t TgtVarsSize) {
+                             ptrdiff_t *TgtOffsets, int32_t TgtVarsSize,
+                             __tgt_async_info *AsyncInfo) {
   return RTL->run_region(RTLDeviceID, TgtEntryPtr, TgtVarsPtr, TgtOffsets,
-      TgtVarsSize);
+                         TgtVarsSize, AsyncInfo);
 }
 
 // Run team region on device.
 int32_t DeviceTy::run_team_region(void *TgtEntryPtr, void **TgtVarsPtr,
-    ptrdiff_t *TgtOffsets, int32_t TgtVarsSize, int32_t NumTeams,
-    int32_t ThreadLimit, uint64_t LoopTripCount) {
+                                  ptrdiff_t *TgtOffsets, int32_t TgtVarsSize,
+                                  int32_t NumTeams, int32_t ThreadLimit,
+                                  uint64_t LoopTripCount,
+                                  __tgt_async_info *AsyncInfo) {
   return RTL->run_team_region(RTLDeviceID, TgtEntryPtr, TgtVarsPtr, TgtOffsets,
-      TgtVarsSize, NumTeams, ThreadLimit, LoopTripCount);
+                              TgtVarsSize, NumTeams, ThreadLimit, LoopTripCount,
+                              AsyncInfo);
 }
 
 /// Check whether a device has an associated RTL and initialize it if it's not
 /// already initialized.
 bool device_is_ready(int device_num) {
   DP("Checking whether device %d is ready.\n", device_num);
   // Devices.size() can only change while registering a new
   // library, so try to acquire the lock of RTLs' mutex.

openmp/libomptarget/src/omptarget.cpp

int CheckDeviceAndCtors(int64_t device_id) {
   return OFFLOAD_SUCCESS;
 }
 
 static int32_t member_of(int64_t type) {
   return ((type & OMP_TGT_MAPTYPE_MEMBER_OF) >> 48) - 1;
 }
 
 /// Internal function to do the mapping and transfer the data to the device
-int target_data_begin(DeviceTy &Device, int32_t arg_num,
-    void **args_base, void **args, int64_t *arg_sizes, int64_t *arg_types) {
+int target_data_begin(DeviceTy &Device, int32_t arg_num, void **args_base,
+                      void **args, int64_t *arg_sizes, int64_t *arg_types,
+                      __tgt_async_info *async_info_ptr) {
   // process each input.
   for (int32_t i = 0; i < arg_num; ++i) {
     // Ignore private variables and arrays - there is no mapping for them.
     if ((arg_types[i] & OMP_TGT_MAPTYPE_LITERAL) ||
         (arg_types[i] & OMP_TGT_MAPTYPE_PRIVATE))
       continue;
 
     void *HstPtrBegin = args[i];
if (arg_types[i] & OMP_TGT_MAPTYPE_TO) {
           if (parent_rc == 1) {
             copy = true;
           }
         }
       }
 
       if (copy && !IsHostPtr) {
         DP("Moving %" PRId64 " bytes (hst:" DPxMOD ") -> (tgt:" DPxMOD ")\n",
-            data_size, DPxPTR(HstPtrBegin), DPxPTR(TgtPtrBegin));
-        int rt = Device.data_submit(TgtPtrBegin, HstPtrBegin, data_size);
+        int rt = Device.data_submit(TgtPtrBegin, HstPtrBegin, data_size,
+                                    async_info_ptr);
         if (rt != OFFLOAD_SUCCESS) {
           DP("Copying data to device failed.\n");
           return OFFLOAD_FAIL;
         }
       }
     }
 
     if (arg_types[i] & OMP_TGT_MAPTYPE_PTR_AND_OBJ && !IsHostPtr) {
       DP("Update pointer (" DPxMOD ") -> [" DPxMOD "]\n",
           DPxPTR(Pointer_TgtPtrBegin), DPxPTR(TgtPtrBegin));
       uint64_t Delta = (uint64_t)HstPtrBegin - (uint64_t)HstPtrBase;
       void *TgtPtrBase = (void *)((uint64_t)TgtPtrBegin - Delta);
       int rt = Device.data_submit(Pointer_TgtPtrBegin, &TgtPtrBase,
-          sizeof(void *));
+                                  sizeof(void *), async_info_ptr);
       if (rt != OFFLOAD_SUCCESS) {
         DP("Copying data to device failed.\n");
         return OFFLOAD_FAIL;
       }
       // create shadow pointers for this entry
       Device.ShadowMtx.lock();
       Device.ShadowPtrMap[Pointer_HstPtrBegin] = {HstPtrBase,
           Pointer_TgtPtrBegin, TgtPtrBase};
       Device.ShadowMtx.unlock();
     }
   }
 
   return OFFLOAD_SUCCESS;
 }
 
 /// Internal function to undo the mapping and retrieve the data from the device.
 int target_data_end(DeviceTy &Device, int32_t arg_num, void **args_base,
-    void **args, int64_t *arg_sizes, int64_t *arg_types) {
+                    void **args, int64_t *arg_sizes, int64_t *arg_types,
+                    __tgt_async_info *async_info_ptr) {
   // process each input.
   for (int32_t i = arg_num - 1; i >= 0; --i) {
     // Ignore private variables and arrays - there is no mapping for them.
     // Also, ignore the use_device_ptr directive, it has no effect here.
     if ((arg_types[i] & OMP_TGT_MAPTYPE_LITERAL) ||
         (arg_types[i] & OMP_TGT_MAPTYPE_PRIVATE))
       continue;
 
if ((arg_types[i] & OMP_TGT_MAPTYPE_FROM) || DelEntry) {
             }
           }
         }
 
         if ((DelEntry || Always || CopyMember) &&
             !(RTLs->RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY &&
               TgtPtrBegin == HstPtrBegin)) {
           DP("Moving %" PRId64 " bytes (tgt:" DPxMOD ") -> (hst:" DPxMOD ")\n",
-              data_size, DPxPTR(TgtPtrBegin), DPxPTR(HstPtrBegin));
-          int rt = Device.data_retrieve(HstPtrBegin, TgtPtrBegin, data_size);
+          int rt = Device.data_retrieve(HstPtrBegin, TgtPtrBegin, data_size,
+                                        async_info_ptr);
           if (rt != OFFLOAD_SUCCESS) {
             DP("Copying data from device failed.\n");
             return OFFLOAD_FAIL;
           }
         }
       }
 
       // If we copied back to the host a struct/array containing pointers, we
if (RTLs->RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY &&
       DP("hst data:" DPxMOD " unified and shared, becomes a noop\n",
          DPxPTR(HstPtrBegin));
       continue;
     }
 
     if (arg_types[i] & OMP_TGT_MAPTYPE_FROM) {
       DP("Moving %" PRId64 " bytes (tgt:" DPxMOD ") -> (hst:" DPxMOD ")\n",
           arg_sizes[i], DPxPTR(TgtPtrBegin), DPxPTR(HstPtrBegin));
-      int rt = Device.data_retrieve(HstPtrBegin, TgtPtrBegin, MapSize);
+      int rt = Device.data_retrieve(HstPtrBegin, TgtPtrBegin, MapSize, nullptr);
       if (rt != OFFLOAD_SUCCESS) {
         DP("Copying data from device failed.\n");
         return OFFLOAD_FAIL;
       }
 
       uintptr_t lb = (uintptr_t) HstPtrBegin;
       uintptr_t ub = (uintptr_t) HstPtrBegin + MapSize;
       Device.ShadowMtx.lock();
if (arg_types[i] & OMP_TGT_MAPTYPE_FROM) {
         *ShadowHstPtrAddr = it->second.HstPtrVal;
       }
       Device.ShadowMtx.unlock();
     }
 
     if (arg_types[i] & OMP_TGT_MAPTYPE_TO) {
       DP("Moving %" PRId64 " bytes (hst:" DPxMOD ") -> (tgt:" DPxMOD ")\n",
           arg_sizes[i], DPxPTR(HstPtrBegin), DPxPTR(TgtPtrBegin));
-      int rt = Device.data_submit(TgtPtrBegin, HstPtrBegin, MapSize);
+      int rt = Device.data_submit(TgtPtrBegin, HstPtrBegin, MapSize, nullptr);
       if (rt != OFFLOAD_SUCCESS) {
         DP("Copying data to device failed.\n");
         return OFFLOAD_FAIL;
       }
 
       uintptr_t lb = (uintptr_t) HstPtrBegin;
       uintptr_t ub = (uintptr_t) HstPtrBegin + MapSize;
       Device.ShadowMtx.lock();
       for (ShadowPtrListTy::iterator it = Device.ShadowPtrMap.begin();
           it != Device.ShadowPtrMap.end(); ++it) {
         void **ShadowHstPtrAddr = (void**) it->first;
         if ((uintptr_t) ShadowHstPtrAddr < lb)
           continue;
         if ((uintptr_t) ShadowHstPtrAddr >= ub)
           break;
         DP("Restoring original target pointer value " DPxMOD " for target "
             "pointer " DPxMOD "\n", DPxPTR(it->second.TgtPtrVal),
             DPxPTR(it->second.TgtPtrAddr));
         rt = Device.data_submit(it->second.TgtPtrAddr,
-            &it->second.TgtPtrVal, sizeof(void *));
+            &it->second.TgtPtrVal, sizeof(void *), nullptr);
         if (rt != OFFLOAD_SUCCESS) {
           DP("Copying data to device failed.\n");
           Device.ShadowMtx.unlock();
           return OFFLOAD_FAIL;
         }
       }
       Device.ShadowMtx.unlock();
     }
int target(int64_t device_id, void *host_ptr, int32_t arg_num,
   // get target table.
   TrlTblMtx->lock();
   assert(TM->Table->TargetsTable.size() > (size_t)device_id &&
          "Not expecting a device ID outside the table's bounds!");
   __tgt_target_table *TargetTable = TM->Table->TargetsTable[device_id];
   TrlTblMtx->unlock();
   assert(TargetTable && "Global data has not been mapped\n");
 
+  __tgt_async_info AsyncInfo;
+
   // Move data to device.
   int rc = target_data_begin(Device, arg_num, args_base, args, arg_sizes,
-      arg_types);
+                             arg_types, &AsyncInfo);
   if (rc != OFFLOAD_SUCCESS) {
     DP("Call to target_data_begin failed, abort target.\n");
     return OFFLOAD_FAIL;
   }
 
   std::vector<void *> tgt_args;
   std::vector<ptrdiff_t> tgt_offsets;
 
if (!(arg_types[i] & OMP_TGT_MAPTYPE_TARGET_PARAM)) {
             TgtPtrBegin == HstPtrBegin) {
           DP("Unified memory is active, no need to map lambda captured"
              "variable (" DPxMOD ")\n", DPxPTR(HstPtrVal));
           continue;
         }
         DP("Update lambda reference (" DPxMOD ") -> [" DPxMOD "]\n",
            DPxPTR(Pointer_TgtPtrBegin), DPxPTR(TgtPtrBegin));
         int rt = Device.data_submit(TgtPtrBegin, &Pointer_TgtPtrBegin,
-                                    sizeof(void *));
+                                    sizeof(void *), &AsyncInfo);
         if (rt != OFFLOAD_SUCCESS) {
           DP("Copying data to device failed.\n");
           return OFFLOAD_FAIL;
         }
       }
       continue;
     }
     void *HstPtrBegin = args[i];
#ifdef OMPTARGET_DEBUG
       DP("Allocated %" PRId64 " bytes of target memory at " DPxMOD " for "
           "%sprivate array " DPxMOD " - pushing target argument " DPxMOD "\n",
           arg_sizes[i], DPxPTR(TgtPtrBegin),
           (arg_types[i] & OMP_TGT_MAPTYPE_TO ? "first-" : ""),
           DPxPTR(HstPtrBegin), DPxPTR(TgtPtrBase));
 #endif
       // If first-private, copy data from host
       if (arg_types[i] & OMP_TGT_MAPTYPE_TO) {
-        int rt = Device.data_submit(TgtPtrBegin, HstPtrBegin, arg_sizes[i]);
+        int rt = Device.data_submit(TgtPtrBegin, HstPtrBegin, arg_sizes[i],
+                                    &AsyncInfo);
         if (rt != OFFLOAD_SUCCESS) {
-          DP ("Copying data to device failed, failed.\n");
+          DP("Copying data to device failed, failed.\n");
           return OFFLOAD_FAIL;
         }
       }
     } else if (arg_types[i] & OMP_TGT_MAPTYPE_PTR_AND_OBJ) {
       TgtPtrBegin = Device.getTgtPtrBegin(HstPtrBase, sizeof(void *), IsLast,
           false, IsHostPtr);
       TgtBaseOffset = 0; // no offset for ptrs.
       DP("Obtained target argument " DPxMOD " from host pointer " DPxMOD " to "
#endif
   TblMapMtx->unlock();
 
   // Launch device execution.
   DP("Launching target execution %s with pointer " DPxMOD " (index=%d).\n",
       TargetTable->EntriesBegin[TM->Index].name,
       DPxPTR(TargetTable->EntriesBegin[TM->Index].addr), TM->Index);
   if (IsTeamConstruct) {
     rc = Device.run_team_region(TargetTable->EntriesBegin[TM->Index].addr,
-        &tgt_args[0], &tgt_offsets[0], tgt_args.size(), team_num,
-        thread_limit, ltc);
+                                &tgt_args[0], &tgt_offsets[0], tgt_args.size(),
+                                team_num, thread_limit, ltc, &AsyncInfo);
   } else {
     rc = Device.run_region(TargetTable->EntriesBegin[TM->Index].addr,
-        &tgt_args[0], &tgt_offsets[0], tgt_args.size());
+                           &tgt_args[0], &tgt_offsets[0], tgt_args.size(),
+                           &AsyncInfo);
   }
   if (rc != OFFLOAD_SUCCESS) {
     DP ("Executing target region abort target.\n");
     return OFFLOAD_FAIL;
   }
 
   // Deallocate (first-)private arrays
   for (auto it : fpArrays) {
     int rt = Device.RTL->data_delete(Device.RTLDeviceID, it);
     if (rt != OFFLOAD_SUCCESS) {
       DP("Deallocation of (first-)private arrays failed.\n");
       return OFFLOAD_FAIL;
     }
   }
 
   // Move data from device.
   int rt = target_data_end(Device, arg_num, args_base, args, arg_sizes,
-      arg_types);
+                           arg_types, &AsyncInfo);
   if (rt != OFFLOAD_SUCCESS) {
     DP("Call to target_data_end failed, abort targe.\n");
     return OFFLOAD_FAIL;
   }
 
-  return OFFLOAD_SUCCESS;
+  return Device.RTL->synchronize(device_id, &AsyncInfo);
 }

openmp/libomptarget/src/private.h

 #ifndef _OMPTARGET_PRIVATE_H
 #define _OMPTARGET_PRIVATE_H
 
 #include <omptarget.h>
 
 #include <cstdint>
 
 extern int target_data_begin(DeviceTy &Device, int32_t arg_num,
-    void **args_base, void **args, int64_t *arg_sizes, int64_t *arg_types);
+                             void **args_base, void **args, int64_t *arg_sizes,
+                             int64_t *arg_types,
+                             __tgt_async_info *async_info_ptr = nullptr);
 
 extern int target_data_end(DeviceTy &Device, int32_t arg_num, void **args_base,
-    void **args, int64_t *arg_sizes, int64_t *arg_types);
+                           void **args, int64_t *arg_sizes, int64_t *arg_types,
+                           __tgt_async_info *async_info_ptr = nullptr);
 
 extern int target_data_update(DeviceTy &Device, int32_t arg_num,
     void **args_base, void **args, int64_t *arg_sizes, int64_t *arg_types);
 
 extern int target(int64_t device_id, void *host_ptr, int32_t arg_num,
     void **args_base, void **args, int64_t *arg_sizes, int64_t *arg_types,
     int32_t team_num, int32_t thread_limit, int IsTeamConstruct);
 

openmp/libomptarget/src/rtl.h

 struct __tgt_bin_desc;
 
 struct RTLInfoTy {
   typedef int32_t(is_valid_binary_ty)(void *);
   typedef int32_t(number_of_devices_ty)();
   typedef int32_t(init_device_ty)(int32_t);
   typedef __tgt_target_table *(load_binary_ty)(int32_t, void *);
   typedef void *(data_alloc_ty)(int32_t, int64_t, void *);
-  typedef int32_t(data_submit_ty)(int32_t, void *, void *, int64_t);
-  typedef int32_t(data_retrieve_ty)(int32_t, void *, void *, int64_t);
+  typedef int32_t(data_submit_ty)(int32_t, void *, void *, int64_t,
+                                  __tgt_async_info *);
+  typedef int32_t(data_retrieve_ty)(int32_t, void *, void *, int64_t,
+                                    __tgt_async_info *);
   typedef int32_t(data_delete_ty)(int32_t, void *);
-  typedef int32_t(run_region_ty)(int32_t, void *, void **, ptrdiff_t *,
-                                 int32_t);
+  typedef int32_t(run_region_ty)(int32_t, void *, void **, ptrdiff_t *, int32_t,
+                                 __tgt_async_info *);
   typedef int32_t(run_team_region_ty)(int32_t, void *, void **, ptrdiff_t *,
-                                      int32_t, int32_t, int32_t, uint64_t);
+                                      int32_t, int32_t, int32_t, uint64_t,
+                                      __tgt_async_info *);
   typedef int64_t(init_requires_ty)(int64_t);
+  typedef int64_t(synchronize_ty)(int64_t, __tgt_async_info *);
 
   int32_t Idx = -1;             // RTL index, index is the number of devices
                                 // of other RTLs that were registered before,
                                 // i.e. the OpenMP index of the first device
                                 // to be registered with this RTL.
   int32_t NumberOfDevices = -1; // Number of devices this RTL deals with.
 
   void *LibraryHandler = nullptr;
#endif
   load_binary_ty *load_binary = nullptr;
   data_alloc_ty *data_alloc = nullptr;
   data_submit_ty *data_submit = nullptr;
   data_retrieve_ty *data_retrieve = nullptr;
   data_delete_ty *data_delete = nullptr;
   run_region_ty *run_region = nullptr;
   run_team_region_ty *run_team_region = nullptr;
   init_requires_ty *init_requires = nullptr;
+  synchronize_ty *synchronize = nullptr;
 
   // Are there images associated with this RTL.
   bool isUsed = false;
 
   // Mutex for thread-safety when calling RTL interface functions.
   // It is easier to enforce thread-safety at the libomptarget level,
   // so that developers of new RTLs do not have to worry about it.
   std::mutex Mtx;
#endif
     data_alloc = r.data_alloc;
     data_submit = r.data_submit;
     data_retrieve = r.data_retrieve;
     data_delete = r.data_delete;
     run_region = r.run_region;
     run_team_region = r.run_team_region;
     init_requires = r.init_requires;
     isUsed = r.isUsed;
+    synchronize = r.synchronize;
   }
 };
 
 /// RTLs identified in the system.
 class RTLsTy {
 private:
   // Mutex-like object to guarantee thread-safety and unique initialization
   // (i.e. the library attempts to load the RTLs (plugins) only once).

openmp/libomptarget/src/rtl.cpp

if (!(*((void**) &R.data_delete) = dlsym(
               dynlib_handle, "__tgt_rtl_data_delete")))
       continue;
     if (!(*((void**) &R.run_region) = dlsym(
               dynlib_handle, "__tgt_rtl_run_target_region")))
       continue;
     if (!(*((void**) &R.run_team_region) = dlsym(
               dynlib_handle, "__tgt_rtl_run_target_team_region")))
       continue;
+    if (!(*((void**) &R.synchronize) = dlsym(
+              dynlib_handle, "__tgt_rtl_synchronize")))
+      continue;
 
     // Optional functions
     *((void**) &R.init_requires) = dlsym(
         dynlib_handle, "__tgt_rtl_init_requires");
 
     // No devices are supported by this RTL?
     if (!(R.NumberOfDevices = R.number_of_devices())) {
       DP("No devices supported in this RTL\n");