Fix CUDA inter-stream synchronization issue

[alexxy/gromacs.git] / src / mdlib / nbnxn_cuda / nbnxn_cuda_data_mgmt.cu

diff --git a/src/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.cu b/src/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.cu
index a35bfedb818b393dc12c769ccf722fc62eef66be..377dcd50712491c41454bf2551e050e5cbdeb102 100644 (file)
--- a/src/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.cu
+++ b/src/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.cu
@@ -576,8 +576,8 @@ void nbnxn_cuda_init(FILE *fplog,
     /* init events for sychronization (timing disabled for performance reasons!) */
     stat = cudaEventCreateWithFlags(&nb->nonlocal_done, cudaEventDisableTiming);
     CU_RET_ERR(stat, "cudaEventCreate on nonlocal_done failed");
-    stat = cudaEventCreateWithFlags(&nb->misc_ops_done, cudaEventDisableTiming);
-    CU_RET_ERR(stat, "cudaEventCreate on misc_ops_one failed");
+    stat = cudaEventCreateWithFlags(&nb->misc_ops_and_local_H2D_done, cudaEventDisableTiming);
+    CU_RET_ERR(stat, "cudaEventCreate on misc_ops_and_local_H2D_done failed");
 
     /* On GPUs with ECC enabled, cudaStreamSynchronize shows a large overhead
      * (which increases with shorter time/step) caused by a known CUDA driver bug.
@@ -933,8 +933,8 @@ void nbnxn_cuda_free(nbnxn_cuda_ptr_t cu_nb)
 
     stat = cudaEventDestroy(cu_nb->nonlocal_done);
     CU_RET_ERR(stat, "cudaEventDestroy failed on timers->nonlocal_done");
-    stat = cudaEventDestroy(cu_nb->misc_ops_done);
-    CU_RET_ERR(stat, "cudaEventDestroy failed on timers->misc_ops_done");
+    stat = cudaEventDestroy(cu_nb->misc_ops_and_local_H2D_done);
+    CU_RET_ERR(stat, "cudaEventDestroy failed on timers->misc_ops_and_local_H2D_done");
 
     if (cu_nb->bDoTime)
     {