Remove MPI comm from GPU PME-PP force transfer initiation

[alexxy/gromacs.git] / src / gromacs / ewald / pme_only.cpp

diff --git a/src/gromacs/ewald/pme_only.cpp b/src/gromacs/ewald/pme_only.cpp
index 50f5b48878452768e338c7739785edc9052b5539..0e67991867eb750d8bb499a27b24887d84236b33 100644 (file)
--- a/src/gromacs/ewald/pme_only.cpp
+++ b/src/gromacs/ewald/pme_only.cpp
@@ -134,6 +134,8 @@ struct gmx_pme_pp
 
     /*! \brief whether GPU direct communications are active for PME-PP transfers */
     bool useGpuDirectComm = false;
+    /*! \brief whether GPU direct communications should send forces directly to remote GPU memory */
+    bool sendForcesDirectToPpGpu = false;
 };
 
 /*! \brief Initialize the PME-only side of the PME <-> PP communication */
@@ -288,6 +290,7 @@ static int gmx_pme_recv_coeffs_coords(struct gmx_pme_t*            pme,
         GMX_ASSERT(!pme_pp->useGpuDirectComm || (pme_pp->pmeForceSenderGpu != nullptr),
                    "The use of GPU direct communication for PME-PP is enabled, "
                    "but the PME GPU force reciever object does not exist");
+        pme_pp->sendForcesDirectToPpGpu = ((cnb.flags & PP_PME_RECVFTOGPU) != 0);
 
         if (cnb.flags & PP_PME_FINISH)
         {
@@ -558,7 +561,8 @@ static void gmx_pme_send_force_vir_ener(const gmx_pme_t& pme,
 
             if (GMX_THREAD_MPI)
             {
-                pme_pp->pmeForceSenderGpu->sendFToPpCudaDirect(receiver.rankId, receiver.numAtoms);
+                pme_pp->pmeForceSenderGpu->sendFToPpCudaDirect(
+                        receiver.rankId, receiver.numAtoms, pme_pp->sendForcesDirectToPpGpu);
             }
             else
             {