nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
{
nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
+
+ /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
+ * also sharing texture references. To keep the code simple, we don't
+ * treat texture references as shared resources, but this means that
+ * the coulomb_tab texture ref will get updated by multiple threads.
+ * Hence, to ensure that the non-bonded kernels don't start before all
+ * texture binding operations are finished, we need to wait for all ranks
+ * to arrive here before continuing.
+ *
+ * Note that we could omit this barrier if GPUs are not shared (or
+ * texture objects are used), but as this is initialization code, there
+ * is not point in complicating things.
+ */
+#ifdef GMX_THREAD_MPI
+ if (PAR(cr))
+ {
+ gmx_barrier(cr);
+ }
+#endif /* GMX_THREAD_MPI */
}
else
{