X-Git-Url: http://biod.pnpi.spb.ru/gitweb/?a=blobdiff_plain;f=src%2Fgromacs%2Fewald%2Fpme.h;h=f8653e20dda06838fb2647cc432ca8f23eb8ae56;hb=c088e63019ebc68760d35c3bc916015864aa8e89;hp=93145d4e1e9807b6aa92197e912eff175e9a2c6f;hpb=f59f6ec6e04f4a3732876c2c76250b8444fdcb69;p=alexxy%2Fgromacs.git

diff --git a/src/gromacs/ewald/pme.h b/src/gromacs/ewald/pme.h
index 93145d4e1e..f8653e20dd 100644
--- a/src/gromacs/ewald/pme.h
+++ b/src/gromacs/ewald/pme.h
@@ -119,6 +119,7 @@ private:
     std::vector<std::string> reasons_;
 };
 
+class PmeCoordinateReceiverGpu;
 } // namespace gmx
 
 enum
@@ -373,17 +374,24 @@ GPU_FUNC_QUALIFIER void pme_gpu_prepare_computation(gmx_pme_t*     GPU_FUNC_ARGU
 /*! \brief
  * Launches first stage of PME on GPU - spreading kernel.
  *
- * \param[in] pme                The PME data structure.
- * \param[in] xReadyOnDevice     Event synchronizer indicating that the coordinates
- * are ready in the device memory; nullptr allowed only on separate PME ranks.
- * \param[in] wcycle             The wallclock counter.
- * \param[in] lambdaQ            The Coulomb lambda of the current state of the
- * system. Only used if FEP of Coulomb is active.
+ * \param[in] pme                            The PME data structure.
+ * \param[in] xReadyOnDevice                 Event synchronizer indicating that the coordinates
+ *                                           are ready in the device memory; nullptr allowed only
+ *                                           on separate PME ranks.
+ * \param[in] wcycle                         The wallclock counter.
+ * \param[in] lambdaQ                        The Coulomb lambda of the current state of the
+ *                                           system. Only used if FEP of Coulomb is active.
+ * \param[in] useGpuDirectComm               Whether direct GPU PME-PP communication is active
+ * \param[in]  pmeCoordinateReceiverGpu      Coordinate receiver object, which must be valid when
+ *                                           direct GPU PME-PP communication is active
  */
-GPU_FUNC_QUALIFIER void pme_gpu_launch_spread(gmx_pme_t*            GPU_FUNC_ARGUMENT(pme),
-                                              GpuEventSynchronizer* GPU_FUNC_ARGUMENT(xReadyOnDevice),
-                                              gmx_wallcycle*        GPU_FUNC_ARGUMENT(wcycle),
-                                              real GPU_FUNC_ARGUMENT(lambdaQ)) GPU_FUNC_TERM;
+GPU_FUNC_QUALIFIER void pme_gpu_launch_spread(
+        gmx_pme_t*                     GPU_FUNC_ARGUMENT(pme),
+        GpuEventSynchronizer*          GPU_FUNC_ARGUMENT(xReadyOnDevice),
+        gmx_wallcycle*                 GPU_FUNC_ARGUMENT(wcycle),
+        real                           GPU_FUNC_ARGUMENT(lambdaQ),
+        const bool                     GPU_FUNC_ARGUMENT(useGpuDirectComm),
+        gmx::PmeCoordinateReceiverGpu* GPU_FUNC_ARGUMENT(pmeCoordinateReceiverGpu)) GPU_FUNC_TERM;
 
 /*! \brief
  * Launches middle stages of PME (FFT R2C, solving, FFT C2R) either on GPU or on CPU, depending on the run mode.