Reimplement constant acceleration groups

[alexxy/gromacs.git] / src / gromacs / mdrun / runner.cpp

diff --git a/src/gromacs/mdrun/runner.cpp b/src/gromacs/mdrun/runner.cpp
index 17db7f419bdfc5a14de39586087c03375584b880..313b4ebb3409d37d90631281a3121e7e7a661bec 100644 (file)
--- a/src/gromacs/mdrun/runner.cpp
+++ b/src/gromacs/mdrun/runner.cpp
@@ -203,8 +203,8 @@ static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& md
 {
     DevelopmentFeatureFlags devFlags;
 
-    devFlags.enableGpuBufferOps =
-            GMX_GPU_CUDA && useGpuForNonbonded && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
+    devFlags.enableGpuBufferOps = (GMX_GPU_CUDA || GMX_GPU_SYCL) && useGpuForNonbonded
+                                  && (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
     devFlags.enableGpuHaloExchange = GMX_MPI && GMX_GPU_CUDA && getenv("GMX_GPU_DD_COMMS") != nullptr;
     devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr) || GMX_FAHCORE;
     devFlags.enableGpuPmePPComm = GMX_MPI && GMX_GPU_CUDA && getenv("GMX_GPU_PME_PP_COMMS") != nullptr;
@@ -739,7 +739,7 @@ static void finish_run(FILE*                     fplog,
         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
     }
 
-    if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
+    if (thisRankHasDuty(cr, DUTY_PP) && haveDDAtomOrdering(*cr))
     {
         print_dd_statistics(cr, inputrec, fplog);
     }
@@ -907,6 +907,7 @@ int Mdrunner::mdrunner()
                     hw_opt.nthreads_tmpi);
             useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(useGpuForNonbonded,
                                                                      pmeTarget,
+                                                                     pmeFftTarget,
                                                                      numAvailableDevices,
                                                                      userGpuTaskAssignment,
                                                                      *hwinfo_,
@@ -937,7 +938,7 @@ int Mdrunner::mdrunner()
         // master and spawned threads joins at the end of this block.
     }
 
-    GMX_RELEASE_ASSERT(ms || simulationCommunicator != MPI_COMM_NULL,
+    GMX_RELEASE_ASSERT(!GMX_MPI || ms || simulationCommunicator != MPI_COMM_NULL,
                        "Must have valid communicator unless running a multi-simulation");
     CommrecHandle crHandle = init_commrec(simulationCommunicator);
     t_commrec*    cr       = crHandle.get();
@@ -969,17 +970,15 @@ int Mdrunner::mdrunner()
     GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
     partialDeserializedTpr.reset(nullptr);
 
-    GMX_RELEASE_ASSERT(
-            !inputrec->useConstantAcceleration,
-            "Linear acceleration has been removed in GROMACS 2022, and was broken for many years "
-            "before that. Use GROMACS 4.5 or earlier if you need this feature.");
-
     // Now the number of ranks is known to all ranks, and each knows
     // the inputrec read by the master rank. The ranks can now all run
     // the task-deciding functions and will agree on the result
     // without needing to communicate.
+    // The LBFGS minimizer, test-particle insertion, normal modes and shell dynamics don't support DD
     const bool useDomainDecomposition =
-            (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == IntegrationAlgorithm::NM));
+            !(inputrec->eI == IntegrationAlgorithm::LBFGS || EI_TPI(inputrec->eI)
+              || inputrec->eI == IntegrationAlgorithm::NM
+              || gmx_mtop_particletype_count(mtop)[ParticleType::Shell] > 0);
 
     // Note that these variables describe only their own node.
     //
@@ -1007,6 +1006,7 @@ int Mdrunner::mdrunner()
                 gpusWereDetected);
         useGpuForPme    = decideWhetherToUseGpusForPme(useGpuForNonbonded,
                                                     pmeTarget,
+                                                    pmeFftTarget,
                                                     userGpuTaskAssignment,
                                                     *hwinfo_,
                                                     *inputrec,
@@ -1108,7 +1108,7 @@ int Mdrunner::mdrunner()
             globalState = std::make_unique<t_state>();
         }
         broadcastStateWithoutDynamics(
-                cr->mpiDefaultCommunicator, DOMAINDECOMP(cr), PAR(cr), globalState.get());
+                cr->mpiDefaultCommunicator, haveDDAtomOrdering(*cr), PAR(cr), globalState.get());
     }
 
     /* A parallel command line option consistency check that we can
@@ -1502,7 +1502,7 @@ int Mdrunner::mdrunner()
 
     if (deviceInfo != nullptr)
     {
-        if (DOMAINDECOMP(cr) && thisRankHasDuty(cr, DUTY_PP))
+        if (runScheduleWork.simulationWork.havePpDomainDecomposition && thisRankHasDuty(cr, DUTY_PP))
         {
             dd_setup_dlb_resource_sharing(cr, deviceId);
         }
@@ -1590,7 +1590,7 @@ int Mdrunner::mdrunner()
     // Enable Peer access between GPUs where available
     // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
     // any of the GPU communication features are active.
-    if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
+    if (haveDDAtomOrdering(*cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
         && (runScheduleWork.simulationWork.useGpuHaloExchange
             || runScheduleWork.simulationWork.useGpuPmePpCommunication))
     {
@@ -1783,7 +1783,7 @@ int Mdrunner::mdrunner()
             }
         }
         // Make the DD reverse topology, now that any vsites that are present are available
-        if (DOMAINDECOMP(cr))
+        if (haveDDAtomOrdering(*cr))
         {
             dd_make_reverse_top(fplog, cr->dd, mtop, vsite.get(), *inputrec, domdecOptions.ddBondedChecking);
         }
@@ -1970,7 +1970,8 @@ int Mdrunner::mdrunner()
                                       ms,
                                       &nrnb,
                                       wcycle.get(),
-                                      fr->bMolPBC);
+                                      fr->bMolPBC,
+                                      &observablesReducerBuilder);
 
         /* Energy terms and groups */
         gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
@@ -2001,7 +2002,7 @@ int Mdrunner::mdrunner()
                                          mdrunOptions.imdOptions,
                                          startingBehavior);
 
-        if (DOMAINDECOMP(cr))
+        if (haveDDAtomOrdering(*cr))
         {
             GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
             /* This call is not included in init_domain_decomposition