"Reduce GPU PME F",
"Wait GPU NB nonloc.",
"Wait GPU NB local",
+ "Wait GPU state copy",
"NB X/F buffer ops.",
"Vsite spread",
"COM pull force",
"Launch NB GPU tasks",
"Launch Bonded GPU tasks",
"Launch PME GPU tasks",
+ "Launch state copy",
"Ewald F correction",
"NB X buffer ops.",
"NB F buffer ops.",