Split simulationWork.useGpuBufferOps into separate x and f flags

[alexxy/gromacs.git] / src / gromacs / mdtypes / simulation_workload.h

diff --git a/src/gromacs/mdtypes/simulation_workload.h b/src/gromacs/mdtypes/simulation_workload.h
index 281c1fc6eb7cd4f23cbdce5b8c39f75a7d5fdafd..d871be32de47fcc88bad175c640312c393cadffd 100644 (file)
--- a/src/gromacs/mdtypes/simulation_workload.h
+++ b/src/gromacs/mdtypes/simulation_workload.h
@@ -47,18 +47,19 @@ namespace gmx
 {
 
 /*! \libinternal
- * \brief Describes work done on this domain that may change per-step.
+ * \brief Describes work done on this domain by the current rank that may change per-step.
  *
  * This work description is based on the SimulationWorkload in the context of the
  * current particle interactions assigned to this domain as well as other
  * factors that may change during the lifetime of a domain.
  *
+ * Note that unlike the other workload descriptors, these flags are also used on
+ * dedicated PME ranks, hence the content is rank-specific (at least when it
+ * comes to flags related to PME).
+ *
  * Note that the contents of an object of this type is valid for
  * a single step and it is expected to be set at the beginning each step.
  *
- * The initial set of flags map the legacy force flags to boolean flags;
- * these have the role of directing per-step compute tasks undertaken by a PP rank.
- *
  */
 class StepWorkload
 {
@@ -77,6 +78,8 @@ public:
     bool computeEnergy = false;
     //! Whether (any) forces need to be computed this step, not only energies
     bool computeForces = false;
+    //! Whether only the MTS combined force buffers are needed and not the separate normal force buffer.
+    bool useOnlyMtsCombinedForceBuffer = false;
     //! Whether nonbonded forces need to be computed this step
     bool computeNonbondedForces = false;
     //! Whether listed forces need to be computed this step
@@ -97,8 +100,10 @@ public:
     bool useGpuXHalo = false;
     //! Whether GPU forces halo exchange is active this step
     bool useGpuFHalo = false;
-    //! Whether GPU PME work is compute this step (can be false also on fast steps with MTS)
+    //! Whether GPU PME work is computed on the current rank this step (can be false on PP-only ranks or on fast steps with MTS)
     bool haveGpuPmeOnThisRank = false;
+    //! Whether to combine the forces for multiple time stepping before the halo exchange
+    bool combineMtsForcesBeforeHaloExchange = false;
 };
 
 /*! \libinternal
@@ -130,6 +135,13 @@ public:
 
     //! Whether the current nstlist step-range Free energy work on the CPU.
     bool haveFreeEnergyWork = false;
+    //! Whether the CPU force buffer has contributions to local atoms that need to be reduced on the GPU (with DD).
+    // This depends on whether there are CPU-based force tasks
+    // or when DD is active the halo exchange has resulted in contributions
+    // from the non-local part.
+    bool haveLocalForceContribInCpuBuffer = false;
+    //! Whether the CPU force buffer has contributions to nonlocal atoms that need to be reduced on the GPU (with DD).
+    bool haveNonLocalForceContribInCpuBuffer = false;
 };
 
 /*! \libinternal
@@ -149,7 +161,7 @@ class SimulationWorkload
 public:
     //! Whether to compute nonbonded pair interactions
     bool computeNonbonded = false;
-    //! Wether nonbonded pair forces are to be computed at slow MTS steps only
+    //! Whether nonbonded pair forces are to be computed at slow MTS steps only
     bool computeNonbondedAtMtsLevel1 = false;
     //! Whether total dipole needs to be computed
     bool computeMuTot = false;
@@ -167,16 +179,28 @@ public:
     bool useGpuBonded = false;
     //! If update and constraint solving is performed on GPU.
     bool useGpuUpdate = false;
-    //! If buffer operations are performed on GPU.
-    bool useGpuBufferOps = false;
+    //! If X buffer operations are performed on GPU.
+    bool useGpuXBufferOps = false;
+    //! If F buffer operations are performed on GPU.
+    bool useGpuFBufferOps = false;
+    //! If PP domain decomposition is active.
+    bool havePpDomainDecomposition = false;
+    //! If domain decomposition halo exchange is performed on CPU (in CPU-only runs or with staged GPU communication).
+    bool useCpuHaloExchange = false;
     //! If domain decomposition halo exchange is performed on GPU.
     bool useGpuHaloExchange = false;
+    //! If separate PME rank(s) are used.
+    bool haveSeparatePmeRank = false;
+    //! If PP-PME communication is done purely on CPU (in CPU-only runs or with staged GPU communication).
+    bool useCpuPmePpCommunication = false;
     //! If direct PP-PME communication between GPU is used.
     bool useGpuPmePpCommunication = false;
     //! If direct GPU-GPU communication is enabled.
     bool useGpuDirectCommunication = false;
     //! If there is an Ewald surface (dipole) term to compute
     bool haveEwaldSurfaceContribution = false;
+    //! Whether to use multiple time stepping
+    bool useMts = false;
 };
 
 class MdrunScheduleWorkload