Merge release-5-0 into master

[alexxy/gromacs.git] / src / gromacs / timing / wallcycle.c

diff --git a/src/gromacs/timing/wallcycle.c b/src/gromacs/timing/wallcycle.c
index 27eb275cd0e088e521f9d693ae731a6f188980c5..3b0d90663631b29e7e15b6844e771ef653781a02 100644 (file)
--- a/src/gromacs/timing/wallcycle.c
+++ b/src/gromacs/timing/wallcycle.c
@@ -99,7 +99,7 @@ static const char *wcn[ewcNR] =
     "DD comm. bounds", "Vsite constr.", "Send X to PME", "Neighbor search", "Launch GPU ops.",
     "Comm. coord.", "Born radii", "Force", "Wait + Comm. F", "PME mesh",
     "PME redist. X/F", "PME spread/gather", "PME 3D-FFT", "PME 3D-FFT Comm.", "PME solve LJ", "PME solve Elec",
-    "PME wait for PP", "Wait + Recv. PME F", "Wait GPU nonlocal", "Wait GPU local", "NB X/F buffer ops.",
+    "PME wait for PP", "Wait + Recv. PME F", "Wait GPU nonlocal", "Wait GPU local", "Wait GPU loc. est.", "NB X/F buffer ops.",
     "Vsite spread", "COM pull force",
     "Write traj.", "Update", "Constraints", "Comm. energies",
     "Enforced rotation", "Add rot. forces", "Coordinate swapping", "IMD", "Test"
@@ -388,6 +388,12 @@ void wallcycle_sum(t_commrec *cr, gmx_wallcycle_t wc)
 
     wcc = wc->wcc;
 
+    /* The GPU wait estimate counter is used for load balancing only
+     * and will mess up the total due to double counting: clear it.
+     */
+    wcc[ewcWAIT_GPU_NB_L_EST].n = 0;
+    wcc[ewcWAIT_GPU_NB_L_EST].c = 0;
+
     for (i = 0; i < ewcNR; i++)
     {
         if (is_pme_counter(i) || (i == ewcRUN && cr->duty == DUTY_PME))