Fixed openacc async

Author: iomaganaris

coreneuron/gpu/nrn_acc_manager.cpp

@@ -655,7 +655,7 @@ void update_net_receive_buffer(NrnThread* nt) {
             }
         }
     }
-    nrn_pragma_acc(wait(nt->streams[nt->stream_id]))
+    nrn_pragma_acc(wait async(nt->streams[nt->stream_id]))
     nrn_pragma_omp(taskwait)
 }
 

coreneuron/network/netcvode.cpp

@@ -561,15 +561,15 @@ void NetCvode::check_thresh(NrnThread* nt) {  // for default method
             nt->_net_send_buffer[idx] = i;
         }
     }
-    nrn_pragma_acc(wait(nt->streams[nt->stream_id]))
+    nrn_pragma_acc(wait async(nt->streams[nt->stream_id]))
     nt->_net_send_buffer_cnt = net_send_buf_count;
 
     if (nt->compute_gpu && nt->_net_send_buffer_cnt) {
 #ifdef CORENEURON_ENABLE_GPU
         int* nsbuffer = nt->_net_send_buffer;
 #endif
         nrn_pragma_acc(update host(nsbuffer [0:nt->_net_send_buffer_cnt]) async(nt->streams[nt->stream_id]))
-        nrn_pragma_acc(wait(nt->streams[nt->stream_id]))
+        nrn_pragma_acc(wait async(nt->streams[nt->stream_id]))
         nrn_pragma_omp(target update from(nsbuffer [0:nt->_net_send_buffer_cnt]))
     }
 

coreneuron/network/partrans.cpp

@@ -71,7 +71,7 @@ void nrnmpi_v_transfer() {
     for (int tid = 0; tid < nrn_nthread; ++tid) {
         if (nrn_threads[tid].compute_gpu) {
             compute_gpu = true;
-            nrn_pragma_acc(wait(nrn_threads[tid].streams[nrn_threads[tid].stream_id]))
+            nrn_pragma_acc(wait async(nrn_threads[tid].streams[nrn_threads[tid].stream_id]))
             nrn_pragma_omp(taskwait)
         }
         TransferThreadData& ttd = transfer_thread_data_[tid];

coreneuron/permute/cellorder.cpp

@@ -626,7 +626,7 @@ void solve_interleaved2(int ith) {
             }  // serial test mode
 #endif
         }
-        nrn_pragma_acc(wait(nt->streams[nt->stream_id]))
+        nrn_pragma_acc(wait async(nt->streams[nt->stream_id]))
 #ifdef _OPENACC
     }
 #endif
@@ -667,7 +667,7 @@ void solve_interleaved1(int ith) {
         triang_interleaved(nt, icell, icellsize, nstride, stride, lastnode);
         bksub_interleaved(nt, icell, icellsize, nstride, stride, firstnode);
     }
-    nrn_pragma_acc(wait(nt->streams[nt->stream_id]))
+    nrn_pragma_acc(wait async(nt->streams[nt->stream_id]))
 }
 
 void solve_interleaved(int ith) {

coreneuron/sim/fadvance_core.cpp

@@ -319,7 +319,7 @@ void nrncore2nrn_send_values(NrnThread* nth) {
                                    async(nth->streams[nth->stream_id]))
                 nrn_pragma_omp(target update from(gather_i [0:1]) if (nth->compute_gpu))
             }
-            nrn_pragma_acc(wait(nth->streams[nth->stream_id]))
+            nrn_pragma_acc(wait async(nth->streams[nth->stream_id]))
             nrn_pragma_omp(taskwait)
             for (int i = 0; i < tr->n_trajec; ++i) {
                 *(tr->scatter[i]) = *(tr->gather[i]);
@@ -343,7 +343,7 @@ static void* nrn_fixed_step_thread(NrnThread* nth) {
         /*@todo: do we need to update nth->_t on GPU: Yes (Michael, but can
         launch kernel) */
         nrn_pragma_acc(update device(nth->_t) if (nth->compute_gpu) async(nth->streams[nth->stream_id]))
-        nrn_pragma_acc(wait(nth->streams[nth->stream_id]))
+        nrn_pragma_acc(wait async(nth->streams[nth->stream_id]))
         nrn_pragma_omp(target update to(nth->_t) if (nth->compute_gpu))
         fixed_play_continuous(nth);
 
@@ -379,7 +379,7 @@ void* nrn_fixed_step_lastpart(NrnThread* nth) {
     if (nth->ncell) {
         /*@todo: do we need to update nth->_t on GPU */
         nrn_pragma_acc(update device(nth->_t) if (nth->compute_gpu) async(nth->streams[nth->stream_id]))
-        nrn_pragma_acc(wait(nth->streams[nth->stream_id]))
+        nrn_pragma_acc(wait async(nth->streams[nth->stream_id]))
         nrn_pragma_omp(target update to(nth->_t) if (nth->compute_gpu))
         fixed_play_continuous(nth);
         nonvint(nth);

coreneuron/sim/solve_core.cpp

@@ -75,7 +75,7 @@ static void bksub(NrnThread* _nt) {
     }
 
     if (_nt->compute_gpu) {
-        nrn_pragma_acc(wait(_nt->streams[_nt->stream_id]))
+        nrn_pragma_acc(wait async(_nt->streams[_nt->stream_id]))
     }
 }
 }  // namespace coreneuron