Split simulationWork.useGpuBufferOps into separate x and f flags

[alexxy/gromacs.git] / src / gromacs / taskassignment / decidesimulationworkload.cpp

diff --git a/src/gromacs/taskassignment/decidesimulationworkload.cpp b/src/gromacs/taskassignment/decidesimulationworkload.cpp
index f3a3a5cdec19e4a36ee699cfc6bd1323de9804c8..a6a98046241c7c93c8d5c69c6aa079ba1b0741bd 100644 (file)
--- a/src/gromacs/taskassignment/decidesimulationworkload.cpp
+++ b/src/gromacs/taskassignment/decidesimulationworkload.cpp
@@ -73,11 +73,18 @@ SimulationWorkload createSimulationWorkload(const t_inputrec& inputrec,
     simulationWorkload.useGpuNonbonded = useGpuForNonbonded;
     simulationWorkload.useCpuPme       = (pmeRunMode == PmeRunMode::CPU);
     simulationWorkload.useGpuPme = (pmeRunMode == PmeRunMode::GPU || pmeRunMode == PmeRunMode::Mixed);
-    simulationWorkload.useGpuPmeFft    = (pmeRunMode == PmeRunMode::Mixed);
-    simulationWorkload.useGpuBonded    = useGpuForBonded;
-    simulationWorkload.useGpuUpdate    = useGpuForUpdate;
-    simulationWorkload.useGpuBufferOps = (devFlags.enableGpuBufferOps || useGpuForUpdate)
-                                         && !simulationWorkload.computeNonbondedAtMtsLevel1;
+    simulationWorkload.useGpuPmeFft = (pmeRunMode == PmeRunMode::Mixed);
+    simulationWorkload.useGpuBonded = useGpuForBonded;
+    simulationWorkload.useGpuUpdate = useGpuForUpdate;
+    simulationWorkload.useGpuXBufferOps =
+            (devFlags.enableGpuBufferOps || useGpuForUpdate) && !inputrec.useMts;
+    simulationWorkload.useGpuFBufferOps =
+            (devFlags.enableGpuBufferOps || useGpuForUpdate) && !inputrec.useMts;
+    if (simulationWorkload.useGpuXBufferOps || simulationWorkload.useGpuFBufferOps)
+    {
+        GMX_ASSERT(simulationWorkload.useGpuNonbonded,
+                   "Can only offload X/F buffer ops if nonbonded computation is also offloaded");
+    }
     simulationWorkload.havePpDomainDecomposition = havePpDomainDecomposition;
     simulationWorkload.useCpuHaloExchange        = havePpDomainDecomposition && !useGpuDirectHalo;
     simulationWorkload.useGpuHaloExchange        = useGpuDirectHalo;