Tweak the bonded GPU energy copy/wait operations

Move the launch of device to host copy earlier and add wait cycle
counter around the operation that may block and lead to time spent
waiting rather than just a launch.

Change-Id: I241001fc0b60a8ae994f173bc3b38f0c4bbd8fff

diff --git a/src/gromacs/mdlib/sim_util.cpp b/src/gromacs/mdlib/sim_util.cpp
index e5b269f6c8cbd4a400102c52698eda2634610095..5205830b050fc72a978fc1cf109e321336de9d56 100644 (file)
--- a/src/gromacs/mdlib/sim_util.cpp
+++ b/src/gromacs/mdlib/sim_util.cpp
@@ -1385,6 +1385,13 @@ static void do_force_cutsVERLET(FILE *fplog,
         nbnxn_gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat,
                                  flags, eatLocal, ppForceWorkload->haveGpuBondedWork);
         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
+
+        wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
+        if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
+        {
+            fr->gpuBonded->launchEnergyTransfer();
+        }
+        wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
     }
 
@@ -1699,27 +1706,29 @@ static void do_force_cutsVERLET(FILE *fplog,
         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
     }
 
-    /* Do the nonbonded GPU (or emulation) force buffer reduction
-     * on the non-alternating path. */
-    if (bUseOrEmulGPU && !alternateGpuWait)
-    {
-        nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs.get(), eatLocal,
-                                       nbv->nbat, f, wcycle);
-    }
-
     if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
     {
+        wallcycle_start(wcycle, ewcWAIT_GPU_BONDED);
+        // in principle this should be included in the DD balancing region,
+        // but generally it is infrequent so we'll omit it for the sake of
+        // simpler code
+        fr->gpuBonded->accumulateEnergyTerms(enerd);
+        wallcycle_stop(wcycle, ewcWAIT_GPU_BONDED);
+
         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
-        fr->gpuBonded->launchEnergyTransfer();
-        fr->gpuBonded->accumulateEnergyTerms(enerd);
-        // TODO The clearing call could come later in the
-        // force-calculation sequence.
         fr->gpuBonded->clearEnergies();
         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
     }
 
+    /* Do the nonbonded GPU (or emulation) force buffer reduction
+     * on the non-alternating path. */
+    if (bUseOrEmulGPU && !alternateGpuWait)
+    {
+        nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs.get(), eatLocal,
+                                       nbv->nbat, f, wcycle);
+    }
     if (DOMAINDECOMP(cr))
     {
         dd_force_flop_stop(cr->dd, nrnb);

diff --git a/src/gromacs/timing/wallcycle.cpp b/src/gromacs/timing/wallcycle.cpp
index 99da970cdaeb8c43a61b1831c7ce0502004e0aca..92ce8cb343cb6269564c9e1c1225618fb4f81efb 100644 (file)
--- a/src/gromacs/timing/wallcycle.cpp
+++ b/src/gromacs/timing/wallcycle.cpp
@@ -109,7 +109,7 @@ static const char *wcn[ewcNR] =
     "PME redist. X/F", "PME spread", "PME gather", "PME 3D-FFT", "PME 3D-FFT Comm.", "PME solve LJ", "PME solve Elec",
     "PME wait for PP", "Wait + Recv. PME F",
     "Wait PME GPU spread", "PME 3D-FFT", "PME solve", /* the strings for FFT/solve are repeated here for mixed mode counters */
-    "Wait PME GPU gather", "Reduce GPU PME F",
+    "Wait PME GPU gather", "Wait Bonded GPU", "Reduce GPU PME F",
     "Wait GPU NB nonloc.", "Wait GPU NB local", "NB X/F buffer ops.",
     "Vsite spread", "COM pull force", "AWH",
     "Write traj.", "Update", "Constraints", "Comm. energies",

diff --git a/src/gromacs/timing/wallcycle.h b/src/gromacs/timing/wallcycle.h
index 123feda178a9a00da382ad58ccd5fe00d9eb6366..9468a0b60afd9fa6ef00973c3909bfb6c72eec42 100644 (file)
--- a/src/gromacs/timing/wallcycle.h
+++ b/src/gromacs/timing/wallcycle.h
@@ -55,7 +55,7 @@ enum {
     ewcPME_REDISTXF, ewcPME_SPREAD, ewcPME_GATHER, ewcPME_FFT, ewcPME_FFTCOMM, ewcLJPME, ewcPME_SOLVE,
     ewcPMEWAITCOMM, ewcPP_PMEWAITRECVF,
     ewcWAIT_GPU_PME_SPREAD, ewcPME_FFT_MIXED_MODE, ewcPME_SOLVE_MIXED_MODE,
-    ewcWAIT_GPU_PME_GATHER, ewcPME_GPU_F_REDUCTION,
+    ewcWAIT_GPU_PME_GATHER, ewcWAIT_GPU_BONDED, ewcPME_GPU_F_REDUCTION,
     ewcWAIT_GPU_NB_NL, ewcWAIT_GPU_NB_L, ewcNB_XF_BUF_OPS,
     ewcVSITESPREAD, ewcPULLPOT, ewcAWH,
     ewcTRAJ, ewcUPDATE, ewcCONSTR, ewcMoveE, ewcROT, ewcROTadd, ewcSWAP, ewcIMD,