From: Szilard Pall <pszilard@cbr.su.se>
Date: Fri, 11 Jan 2013 07:05:04 +0000 (+0100)
Subject: time the GPU force clearing async launch overhead
X-Git-Url: http://biod.pnpi.spb.ru/gitweb/?a=commitdiff_plain;h=381e088ffa7ea3d44f98df424bd0b63b831e1216;p=alexxy%2Fgromacs.git

time the GPU force clearing async launch overhead

Unfortunately, the asynchronous launch of GPU force buffer clearing can
take up to 2% of the total run-time with short iteration times and
many/fast cores/GPU. Timing it will at least remove it form the "Rest".

Change-Id: I397c563ead24d87181de1b03879f164d1a97c2ca
---

diff --git a/src/mdlib/sim_util.c b/src/mdlib/sim_util.c
index eb9f636140..8f211715d9 100644
--- a/src/mdlib/sim_util.c
+++ b/src/mdlib/sim_util.c
@@ -1318,7 +1318,10 @@ void do_force_cutsVERLET(FILE *fplog,t_commrec *cr,
             wallcycle_stop(wcycle,ewcWAIT_GPU_NB_L);
 
             /* now clear the GPU outputs while we finish the step on the CPU */
+
+            wallcycle_start_nocount(wcycle,ewcLAUNCH_GPU_NB);
             nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
+            wallcycle_stop(wcycle,ewcLAUNCH_GPU_NB);
         }
         else
         {