Removes PME rank dependency on Non-local X H2D copy when buffer ops are on GPU but...

This was not a problem until now because because !stepWork.haveGpuPmeOnThisRank will always evaluate to true in case of domain decomposition as PME will always be on a separate rank.
But now with PME decomposition this bug gets exposed.

Refs #4125

diff --git a/src/gromacs/mdlib/sim_util.cpp b/src/gromacs/mdlib/sim_util.cpp
index 6ecc338f703ec72a87a248f94d7691cb361f4624..69c380a73f3e9edfefd6dbbf8a249999b60352f6 100644 (file)
--- a/src/gromacs/mdlib/sim_util.cpp
+++ b/src/gromacs/mdlib/sim_util.cpp
@@ -1592,7 +1592,7 @@ void do_force(FILE*                               fplog,
 
             if (stepWork.useGpuXBufferOps)
             {
-                if (!stepWork.haveGpuPmeOnThisRank && !stepWork.useGpuXHalo)
+                if (!stepWork.useGpuXHalo)
                 {
                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
                 }