Switch the GPU buffer ops on when update is on GPU

The update is supported on the GPU only when buffer ops are also
offloaded. This changes the behavior from requiring the
GMX_USE_GPU_BUFFER_OPS to be enabled to it being overriden.

Change-Id: Icdc154daa053f135b0df503697273016a830fb18

diff --git a/admin/builds/gromacs.py b/admin/builds/gromacs.py
index 4fa518f1864865c2b4d442d5e6b9d8d9d9e87590..49007c5a63f125a098c719b7e73bf4b1b005d2e2 100644 (file)
--- a/admin/builds/gromacs.py
+++ b/admin/builds/gromacs.py
@@ -185,7 +185,6 @@ def do_build(context):
 
     # GPU update flag enables GPU update+constraints as well as buffer ops (dependency)
     if context.opts.gpuupdate:
-        context.env.set_env_var('GMX_USE_GPU_BUFFER_OPS', "1")
         context.env.set_env_var('GMX_FORCE_UPDATE_DEFAULT_GPU', "1")
 
     regressiontests_path = context.workspace.get_project_dir(Project.REGRESSIONTESTS)

diff --git a/src/gromacs/mdrun/runner.cpp b/src/gromacs/mdrun/runner.cpp
index bf6532b3e01d38baba2934422e9f4de34bebc198..245d6f9e50e838de17edd51e33221e399b5caf93 100644 (file)
--- a/src/gromacs/mdrun/runner.cpp
+++ b/src/gromacs/mdrun/runner.cpp
@@ -885,6 +885,27 @@ int Mdrunner::mdrunner()
     // and report those features that are enabled.
     const DevelopmentFeatureFlags devFlags = manageDevelopmentFeatures(mdlog, useGpuForNonbonded, useGpuForPme);
 
+    // NOTE: The devFlags need decideWhetherToUseGpusForNonbonded(...) and decideWhetherToUseGpusForPme(...) for overrides,
+    //       decideWhetherToUseGpuForUpdate() needs devFlags for the '-update auto' override, hence the interleaving.
+    // NOTE: When the simulationWork is constructed, the useGpuForUpdate overrides the devFlags.enableGpuBufferOps.
+    try
+    {
+        useGpuForUpdate = decideWhetherToUseGpuForUpdate(devFlags.forceGpuUpdateDefaultOn,
+                                                         useDomainDecomposition,
+                                                         useGpuForPme,
+                                                         useGpuForNonbonded,
+                                                         updateTarget,
+                                                         gpusWereDetected,
+                                                         *inputrec,
+                                                         gmx_mtop_interaction_count(mtop, IF_VSITE) > 0,
+                                                         doEssentialDynamics,
+                                                         gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
+                                                         gmx_mtop_ftype_count(mtop, F_DISRES) > 0,
+                                                         replExParams.exchangeInterval > 0);
+    }
+    GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+
+
     // Build restraints.
     // TODO: hide restraint implementation details from Mdrunner.
     // There is nothing unique about restraints at this point as far as the
@@ -1333,7 +1354,7 @@ int Mdrunner::mdrunner()
         // TODO remove need to pass local stream into GPU halo exchange - Redmine #3093
         if (havePPDomainDecomposition(cr) && prefer1DAnd1PulseDD && is1DAnd1PulseDD(*cr->dd))
         {
-            GMX_RELEASE_ASSERT(devFlags.enableGpuBufferOps, "Must use GMX_GPU_BUFFER_OPS=1 to use GMX_GPU_DD_COMMS=1");
+            GMX_RELEASE_ASSERT(devFlags.enableGpuBufferOps, "Must use GMX_USE_GPU_BUFFER_OPS=1 to use GMX_GPU_DD_COMMS=1");
             void *streamLocal              = Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::Local);
             void *streamNonLocal           = Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::NonLocal);
             void *coordinatesOnDeviceEvent = fr->nbv->get_x_on_device_event();
@@ -1545,21 +1566,6 @@ int Mdrunner::mdrunner()
                             fr->cginfo_mb);
         }
 
-        // Before we start the actual simulator, try if we can run the update task on the GPU.
-        useGpuForUpdate = decideWhetherToUseGpuForUpdate(devFlags.forceGpuUpdateDefaultOn,
-                                                         DOMAINDECOMP(cr),
-                                                         useGpuForPme,
-                                                         useGpuForNonbonded,
-                                                         devFlags.enableGpuBufferOps,
-                                                         updateTarget,
-                                                         gpusWereDetected,
-                                                         *inputrec,
-                                                         mdAtoms->mdatoms()->haveVsites,
-                                                         doEssentialDynamics,
-                                                         gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
-                                                         gmx_mtop_ftype_count(mtop, F_DISRES) > 0,
-                                                         replExParams.exchangeInterval > 0);
-
         const bool inputIsCompatibleWithModularSimulator = ModularSimulator::isInputCompatible(
                     false,
                     inputrec, doRerun, vsite.get(), ms, replExParams,
@@ -1568,8 +1574,21 @@ int Mdrunner::mdrunner()
 
         const bool useModularSimulator = inputIsCompatibleWithModularSimulator && !(getenv("GMX_DISABLE_MODULAR_SIMULATOR") != nullptr);
 
+        // TODO This is not the right place to manage the lifetime of
+        // this data structure, but currently it's the easiest way to
+        // make it work.
+        MdrunScheduleWorkload runScheduleWork;
+        // Also populates the simulation constant workload description.
+        runScheduleWork.simulationWork = createSimulationWorkload(useGpuForNonbonded,
+                                                                  pmeRunMode,
+                                                                  useGpuForBonded,
+                                                                  useGpuForUpdate,
+                                                                  devFlags.enableGpuBufferOps,
+                                                                  devFlags.enableGpuHaloExchange,
+                                                                  devFlags.enableGpuPmePPComm);
+
         std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
-        if (gpusWereDetected && ((useGpuForPme && thisRankHasDuty(cr, DUTY_PME)) || devFlags.enableGpuBufferOps))
+        if (gpusWereDetected && ((useGpuForPme && thisRankHasDuty(cr, DUTY_PME)) || runScheduleWork.simulationWork.useGpuBufferOps))
         {
             const void         *pmeStream      = pme_gpu_get_device_stream(fr->pmedata);
             const void         *localStream    = fr->nbv->gpu_nbv != nullptr ? Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::Local) : nullptr;
@@ -1587,20 +1606,6 @@ int Mdrunner::mdrunner()
             fr->stateGpu = stateGpu.get();
         }
 
-        // TODO This is not the right place to manage the lifetime of
-        // this data structure, but currently it's the easiest way to
-        // make it work.
-        MdrunScheduleWorkload runScheduleWork;
-        // Also populates the simulation constant workload description.
-        runScheduleWork.simulationWork = createSimulationWorkload(useGpuForNonbonded,
-                                                                  pmeRunMode,
-                                                                  useGpuForBonded,
-                                                                  useGpuForUpdate,
-                                                                  devFlags.enableGpuBufferOps,
-                                                                  devFlags.enableGpuHaloExchange,
-                                                                  devFlags.enableGpuPmePPComm);
-
-
         GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
         SimulatorBuilder simulatorBuilder;
 

diff --git a/src/gromacs/taskassignment/decidegpuusage.cpp b/src/gromacs/taskassignment/decidegpuusage.cpp
index 71ca58d1584f1f5647bc5b584a0fcf0190337b81..f40814e23a2085be8e551d5283ca6008a3ab5862 100644 (file)
--- a/src/gromacs/taskassignment/decidegpuusage.cpp
+++ b/src/gromacs/taskassignment/decidegpuusage.cpp
@@ -494,7 +494,6 @@ bool decideWhetherToUseGpuForUpdate(const bool        forceGpuUpdateDefaultOn,
                                     const bool        isDomainDecomposition,
                                     const bool        useGpuForPme,
                                     const bool        useGpuForNonbonded,
-                                    const bool        useGpuForBufferOps,
                                     const TaskTarget  updateTarget,
                                     const bool        gpusWereDetected,
                                     const t_inputrec &inputrec,
@@ -516,10 +515,10 @@ bool decideWhetherToUseGpuForUpdate(const bool        forceGpuUpdateDefaultOn,
     {
         errorMessage += "Domain decomposition is not supported.\n";
     }
-    // Using the GPU-version of update makes sense if forces are already on the GPU, i.e. if at least:
-    // 1. PME is on the GPU (there should be a copy of coordinates on a GPU in rvec format for PME spread).
-    // 2. Non-bonded interactions and buffer ops are on the GPU.
-    if (!(useGpuForPme || (useGpuForNonbonded && useGpuForBufferOps)))
+    // Using the GPU-version of update if:
+    // 1. PME is on the GPU (there should be a copy of coordinates on GPU for PME spread), or
+    // 2. Non-bonded interactions are on the GPU.
+    if (!(useGpuForPme || useGpuForNonbonded))
     {
         errorMessage += "Either PME or short-ranged non-bonded interaction tasks must run on the GPU.\n";
     }

diff --git a/src/gromacs/taskassignment/decidegpuusage.h b/src/gromacs/taskassignment/decidegpuusage.h
index c74ca0a197ce96071b0a9001a8b7152e0fa4f39c..6b8685fa3e1f886e029915a959e6143d798840c7 100644 (file)
--- a/src/gromacs/taskassignment/decidegpuusage.h
+++ b/src/gromacs/taskassignment/decidegpuusage.h
@@ -235,7 +235,6 @@ bool decideWhetherToUseGpusForBonded(bool       useGpuForNonbonded,
  * \param[in]  isDomainDecomposition     Whether there more than one domain.
  * \param[in]  useGpuForPme              Whether GPUs will be used for PME interactions.
  * \param[in]  useGpuForNonbonded        Whether GPUs will be used for nonbonded interactions.
- * \param[in]  useGpuForBufferOps        Whether GPUs will be used for buffer operations.
  * \param[in]  updateTarget              User choice for running simulation on GPU.
  * \param[in]  gpusWereDetected          Whether compatible GPUs were detected on any node.
  * \param[in]  inputrec                  The user input.
@@ -253,7 +252,6 @@ bool decideWhetherToUseGpuForUpdate(bool              forceGpuUpdateDefaultOn,
                                     bool              isDomainDecomposition,
                                     bool              useGpuForPme,
                                     bool              useGpuForNonbonded,
-                                    bool              useGpuForBufferOps,
                                     TaskTarget        updateTarget,
                                     bool              gpusWereDetected,
                                     const t_inputrec &inputrec,

diff --git a/src/gromacs/taskassignment/decidesimulationworkload.cpp b/src/gromacs/taskassignment/decidesimulationworkload.cpp
index 597cf1a81580119c6a48907574c9aa9f43661953..ea82f95f42bccd1eef0cef622a3d343a0204ec28 100644 (file)
--- a/src/gromacs/taskassignment/decidesimulationworkload.cpp
+++ b/src/gromacs/taskassignment/decidesimulationworkload.cpp
@@ -53,7 +53,7 @@ namespace gmx
 SimulationWorkload createSimulationWorkload(bool       useGpuForNonbonded,
                                             PmeRunMode pmeRunMode,
                                             bool       useGpuForBonded,
-                                            bool       useGpuForUpdateConstraints,
+                                            bool       useGpuForUpdate,
                                             bool       useGpuForBufferOps,
                                             bool       useGpuHaloExchange,
                                             bool       useGpuPmePpComm)
@@ -65,8 +65,8 @@ SimulationWorkload createSimulationWorkload(bool       useGpuForNonbonded,
     simulationWorkload.useGpuPme                 = (pmeRunMode == PmeRunMode::GPU || pmeRunMode == PmeRunMode::Mixed);
     simulationWorkload.useGpuPmeFft              = (pmeRunMode == PmeRunMode::Mixed);
     simulationWorkload.useGpuBonded              = useGpuForBonded;
-    simulationWorkload.useGpuUpdate              = useGpuForUpdateConstraints;
-    simulationWorkload.useGpuBufferOps           = useGpuForBufferOps;
+    simulationWorkload.useGpuUpdate              = useGpuForUpdate;
+    simulationWorkload.useGpuBufferOps           = useGpuForBufferOps || useGpuForUpdate;
     simulationWorkload.useGpuHaloExchange        = useGpuHaloExchange;
     simulationWorkload.useGpuPmePpCommunication  = useGpuPmePpComm;
     simulationWorkload.useGpuDirectCommunication = useGpuHaloExchange || useGpuPmePpComm;

diff --git a/src/gromacs/taskassignment/decidesimulationworkload.h b/src/gromacs/taskassignment/decidesimulationworkload.h
index 0e87da68c599e5520b71618064f8b6f55c39d029..b389da53c168a36d98396f004d532f196f800a17 100644 (file)
--- a/src/gromacs/taskassignment/decidesimulationworkload.h
+++ b/src/gromacs/taskassignment/decidesimulationworkload.h
@@ -59,8 +59,8 @@ namespace gmx
  *                               calculations on GPU(s).
  * \param[in] pmeRunMode         Run mode indicating what resource is PME execured on.
  * \param[in] useGpuForBonded    If bonded interactions are calculated on GPU(s).
- * \param[in] useGpuForUpdateConstraints If coordinate update and constraint solving is performed on
- *                                       GPU(s).
+ * \param[in] useGpuForUpdate    If coordinate update and constraint solving is performed on
+ *                               GPU(s).
  * \param[in] useGpuForBufferOps If buffer ops / reduction are calculated on GPU(s).
  * \param[in] useGpuHaloExchange If GPU direct communication is used in halo exchange.
  * \param[in] useGpuPmePpComm    If GPu direct communication is used in PME-PP communication.
@@ -69,7 +69,7 @@ namespace gmx
 SimulationWorkload createSimulationWorkload(bool       useGpuForNonbonded,
                                             PmeRunMode pmeRunMode,
                                             bool       useGpuForBonded,
-                                            bool       useGpuForUpdateConstraints,
+                                            bool       useGpuForUpdate,
                                             bool       useGpuForBufferOps,
                                             bool       useGpuHaloExchange,
                                             bool       useGpuPmePpComm);