Avoid unnecessary reinterpret_cast of coordinates buffer

The DeviceBuffer<float> is first casted to rvec* and than to void*.
This change eliminates first cast and passes the buffer directly.

Change-Id: I4381ce3a495ef728fd5be45aa3b65be580a6aa49

diff --git a/src/gromacs/ewald/pme_coordinate_receiver_gpu.h b/src/gromacs/ewald/pme_coordinate_receiver_gpu.h
index dfe848159380f441bb2a12be37fb8963773937ef..a4a608b5a25b1dd265ca412eeb64dc44524389f7 100644 (file)
--- a/src/gromacs/ewald/pme_coordinate_receiver_gpu.h
+++ b/src/gromacs/ewald/pme_coordinate_receiver_gpu.h
@@ -66,7 +66,7 @@ public:
      * send coordinates buffer address to PP rank
      * \param[in] d_x   coordinates buffer in GPU memory
      */
-    void sendCoordinateBufferAddressToPpRanks(rvec* d_x);
+    void sendCoordinateBufferAddressToPpRanks(DeviceBuffer<float> d_x);
 
 
     /*! \brief

diff --git a/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cpp b/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cpp
index 35ddc077336c5b1110a43c697fae388cf91fe117..213a88c59edaf5426aaaad7aa2998a4a685e8135 100644 (file)
--- a/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cpp
+++ b/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cpp
@@ -60,9 +60,9 @@ class PmeCoordinateReceiverGpu::Impl
 };
 
 /*!\brief Constructor stub. */
-PmeCoordinateReceiverGpu::PmeCoordinateReceiverGpu(void gmx_unused* pmeStream,
-                                                   MPI_Comm gmx_unused    comm,
-                                                   gmx::ArrayRef<PpRanks> gmx_unused ppRanks) :
+PmeCoordinateReceiverGpu::PmeCoordinateReceiverGpu(void* /* pmeStream */,
+                                                   MPI_Comm /* comm */,
+                                                   gmx::ArrayRef<PpRanks> /* ppRanks */) :
     impl_(nullptr)
 {
     GMX_ASSERT(false,
@@ -73,14 +73,14 @@ PmeCoordinateReceiverGpu::PmeCoordinateReceiverGpu(void gmx_unused* pmeStream,
 PmeCoordinateReceiverGpu::~PmeCoordinateReceiverGpu() = default;
 
 /*!\brief init PME-PP GPU communication stub */
-void PmeCoordinateReceiverGpu::sendCoordinateBufferAddressToPpRanks(rvec gmx_unused* d_x)
+void PmeCoordinateReceiverGpu::sendCoordinateBufferAddressToPpRanks(const DeviceBuffer<float> /* d_x */)
 {
     GMX_ASSERT(false,
                "A CPU stub for PME-PP GPU communication initialization was called instead of the "
                "correct implementation.");
 }
 
-void PmeCoordinateReceiverGpu::launchReceiveCoordinatesFromPpCudaDirect(int gmx_unused ppRank)
+void PmeCoordinateReceiverGpu::launchReceiveCoordinatesFromPpCudaDirect(int /* ppRank */)
 {
     GMX_ASSERT(false,
                "A CPU stub for PME-PP GPU communication was called instead of the correct "

diff --git a/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cu b/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cu
index 3ca69e27a882225bba32590da8afb32f0d2f0501..fd55c832ff2d3dc1c4cf4287766f481113d115d8 100644 (file)
--- a/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cu
+++ b/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cu
@@ -72,7 +72,7 @@ PmeCoordinateReceiverGpu::Impl::Impl(void* pmeStream, MPI_Comm comm, gmx::ArrayR
 
 PmeCoordinateReceiverGpu::Impl::~Impl() = default;
 
-void PmeCoordinateReceiverGpu::Impl::sendCoordinateBufferAddressToPpRanks(rvec* d_x)
+void PmeCoordinateReceiverGpu::Impl::sendCoordinateBufferAddressToPpRanks(const DeviceBuffer<float> d_x)
 {
 
     int ind_start = 0;
@@ -134,7 +134,7 @@ PmeCoordinateReceiverGpu::PmeCoordinateReceiverGpu(void*                  pmeStr
 
 PmeCoordinateReceiverGpu::~PmeCoordinateReceiverGpu() = default;
 
-void PmeCoordinateReceiverGpu::sendCoordinateBufferAddressToPpRanks(rvec* d_x)
+void PmeCoordinateReceiverGpu::sendCoordinateBufferAddressToPpRanks(const DeviceBuffer<float> d_x)
 {
     impl_->sendCoordinateBufferAddressToPpRanks(d_x);
 }

diff --git a/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.h b/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.h
index fe689799e1cb3a4b46d9578786fc0b9cac5c1fc7..84a554725b8e439c9d0ab595d5afafa6b514d28b 100644 (file)
--- a/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.h
+++ b/src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.h
@@ -67,7 +67,7 @@ public:
      * send coordinates buffer address to PP rank
      * \param[in] d_x   coordinates buffer in GPU memory
      */
-    void sendCoordinateBufferAddressToPpRanks(rvec* d_x);
+    void sendCoordinateBufferAddressToPpRanks(const DeviceBuffer<float> d_x);
 
     /*! \brief
      * launch receive of coordinate data from PP rank

diff --git a/src/gromacs/ewald/pme_only.cpp b/src/gromacs/ewald/pme_only.cpp
index be7a977daf1313fbf6f982727bd2de86dde75a8a..a4ec1506569247686254ec41de92fda9fe26e318 100644 (file)
--- a/src/gromacs/ewald/pme_only.cpp
+++ b/src/gromacs/ewald/pme_only.cpp
@@ -423,14 +423,10 @@ static int gmx_pme_recv_coeffs_coords(struct gmx_pme_t*            pme,
                                "but PME run mode is not PmeRunMode::GPU\n");
 
                     // This rank will have its data accessed directly by PP rank, so needs to send the remote addresses.
-                    rvec* d_x = nullptr;
-                    rvec* d_f = nullptr;
-#    if (GMX_GPU == GMX_GPU_CUDA) // avoid invalid cast for OpenCL
-                    d_x = reinterpret_cast<rvec*>(pme_gpu_get_device_x(pme));
-                    d_f = reinterpret_cast<rvec*>(pme_gpu_get_device_f(pme));
-#    endif
-                    pme_pp->pmeCoordinateReceiverGpu->sendCoordinateBufferAddressToPpRanks(d_x);
-                    pme_pp->pmeForceSenderGpu->sendForceBufferAddressToPpRanks(d_f);
+                    pme_pp->pmeCoordinateReceiverGpu->sendCoordinateBufferAddressToPpRanks(
+                            pme_gpu_get_device_x(pme));
+                    pme_pp->pmeForceSenderGpu->sendForceBufferAddressToPpRanks(
+                            reinterpret_cast<rvec*>(pme_gpu_get_device_f(pme)));
                 }
             }