Index: ompi/mpi/c/exscan.c
===================================================================
--- ompi/mpi/c/exscan.c	(revision 31741)
+++ ompi/mpi/c/exscan.c	(working copy)
@@ -12,6 +12,7 @@
  *                         All rights reserved.
  * Copyright (c) 2013      Los Alamos National Security, LLC.  All rights
  *                         reserved.
+ * Copyright (c) 2014      NVIDIA Corporation.  All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -21,6 +22,10 @@
 
 #include "ompi_config.h"
 
+#if OPAL_CUDA_SUPPORT
+#include "opal/datatype/opal_convertor.h"
+#include "opal/datatype/opal_datatype_cuda.h"
+#endif /* OPAL_CUDA_SUPPORT */
 #include "ompi/mpi/c/bindings.h"
 #include "ompi/runtime/params.h"
 #include "ompi/communicator/communicator.h"
@@ -44,6 +49,12 @@
                MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) 
 {
     int err;
+#if OPAL_CUDA_SUPPORT
+    ptrdiff_t true_lb, true_extent, lb, extent;
+    char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2;
+    const char *sbuf2;
+    size_t bufsize;
+#endif /* OPAL_CUDA_SUPPORT */
 
     MEMCHECKER(
         memchecker_datatype(datatype);
@@ -87,10 +98,50 @@
     /* Invoke the coll component to perform the back-end operation */
 
     OBJ_RETAIN(op);
+
+#if OPAL_CUDA_SUPPORT
+    ompi_datatype_get_extent(datatype, &lb, &extent);
+    ompi_datatype_get_true_extent(datatype, &true_lb, &true_extent);
+    bufsize = true_extent + (ptrdiff_t)(count - 1) * extent;
+    if ((MPI_IN_PLACE != sendbuf) && (opal_cuda_check_bufs((char *)sendbuf, NULL))) {
+        sbuf1 = (char*)malloc(bufsize);
+        if (NULL == sbuf1) {
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        opal_cuda_memcpy_sync(sbuf1, sendbuf, bufsize);
+        sbuf2 = sendbuf; /* save away original buffer */
+        sendbuf = sbuf1 - lb;
+    }
+
+    if (opal_cuda_check_bufs(recvbuf, NULL)) {
+        rbuf1 = (char*)malloc(bufsize);
+        if (NULL == rbuf1) {
+            if (NULL != sbuf1) free(sbuf1);
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        opal_cuda_memcpy_sync(rbuf1, recvbuf, bufsize);
+        rbuf2 = recvbuf; /* save away original buffer */
+        recvbuf = rbuf1 - lb;
+    }
+#endif /* OPAL_CUDA_SUPPORT */
+
     /* XXX -- CONST -- do not cast away const -- update mca/coll */
     err = comm->c_coll.coll_exscan((void *) sendbuf, recvbuf, count,
                                    datatype, op, comm,
                                    comm->c_coll.coll_exscan_module);
+
+#if OPAL_CUDA_SUPPORT
+    if (NULL != sbuf1) {
+        free(sbuf1);
+        sendbuf = sbuf2;
+    }
+    if (NULL != rbuf1) {
+        recvbuf = rbuf2;
+        opal_cuda_memcpy_sync(recvbuf, rbuf1, bufsize);
+        free(rbuf1);
+    }
+#endif /* OPAL_CUDA_SUPPORT */
+
     OBJ_RELEASE(op);
     OMPI_ERRHANDLER_RETURN(err, comm, err, FUNC_NAME);
 }
Index: ompi/mpi/c/reduce_scatter_block.c
===================================================================
--- ompi/mpi/c/reduce_scatter_block.c	(revision 31741)
+++ ompi/mpi/c/reduce_scatter_block.c	(working copy)
@@ -14,6 +14,7 @@
  * Copyright (c) 2012      Oak Ridge National Labs. All rights reserved.
  * Copyright (c) 2013      Los Alamos National Security, LLC.  All rights
  *                         reserved.
+ * Copyright (c) 2014      NVIDIA Corporation.  All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -23,6 +24,10 @@
 #include "ompi_config.h"
 #include <stdio.h>
 
+#if OPAL_CUDA_SUPPORT
+#include "opal/datatype/opal_convertor.h"
+#include "opal/datatype/opal_datatype_cuda.h"
+#endif /* OPAL_CUDA_SUPPORT */
 #include "ompi/mpi/c/bindings.h"
 #include "ompi/runtime/params.h"
 #include "ompi/communicator/communicator.h"
@@ -46,6 +51,12 @@
                              MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
 {
     int err;
+#if OPAL_CUDA_SUPPORT
+    ptrdiff_t true_lb, true_extent, lb, extent;
+    char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2;
+    const char *sbuf2;
+    size_t sbufsize, rbufsize;
+#endif /* OPAL_CUDA_SUPPORT */
 
     MEMCHECKER(
         int rank;
@@ -98,10 +109,51 @@
     /* Invoke the coll component to perform the back-end operation */
 
     OBJ_RETAIN(op);
+
+#if OPAL_CUDA_SUPPORT
+    ompi_datatype_get_extent(datatype, &lb, &extent);
+    ompi_datatype_get_true_extent(datatype, &true_lb, &true_extent);
+    sbufsize = (true_extent + (ptrdiff_t)(recvcount - 1) * extent) * ompi_comm_size(comm);
+    rbufsize = true_extent + (ptrdiff_t)(recvcount - 1) * extent;
+    if ((MPI_IN_PLACE != sendbuf) && (opal_cuda_check_bufs((char *)sendbuf, NULL))) {
+        sbuf1 = (char*)malloc(sbufsize);
+        if (NULL == sbuf1) {
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        opal_cuda_memcpy_sync(sbuf1, sendbuf, sbufsize);
+        sbuf2 = sendbuf; /* save away original buffer */
+        sendbuf = sbuf1 - lb;
+    }
+
+    if (opal_cuda_check_bufs(recvbuf, NULL)) {
+        rbuf1 = (char*)malloc(rbufsize);
+        if (NULL == rbuf1) {
+            if (NULL != sbuf1) free(sbuf1);
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        opal_cuda_memcpy_sync(rbuf1, recvbuf, rbufsize);
+        rbuf2 = recvbuf; /* save away original buffer */
+        recvbuf = rbuf1 - lb;
+    }
+#endif /* OPAL_CUDA_SUPPORT */
+
     /* XXX -- CONST -- do not cast away const -- update mca/coll */
     err = comm->c_coll.coll_reduce_scatter_block((void *) sendbuf, recvbuf, recvcount,
                                                  datatype, op, comm,
                                                  comm->c_coll.coll_reduce_scatter_block_module);
+
+#if OPAL_CUDA_SUPPORT
+    if (NULL != sbuf1) {
+        free(sbuf1);
+        sendbuf = sbuf2;
+    }
+    if (NULL != rbuf1) {
+        recvbuf = rbuf2;
+        opal_cuda_memcpy_sync(recvbuf, rbuf1, rbufsize);
+        free(rbuf1);
+    }
+#endif /* OPAL_CUDA_SUPPORT */
+
     OBJ_RELEASE(op);
     OMPI_ERRHANDLER_RETURN(err, comm, err, FUNC_NAME);
 }
Index: ompi/mpi/c/scan.c
===================================================================
--- ompi/mpi/c/scan.c	(revision 31741)
+++ ompi/mpi/c/scan.c	(working copy)
@@ -13,6 +13,7 @@
  * Copyright (c) 2006      Cisco Systems, Inc.  All rights reserved.
  * Copyright (c) 2013      Los Alamos National Security, LLC.  All rights
  *                         reserved.
+ * Copyright (c) 2014      NVIDIA Corporation.  All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -22,6 +23,10 @@
 #include "ompi_config.h"
 #include <stdio.h>
 
+#if OPAL_CUDA_SUPPORT
+#include "opal/datatype/opal_convertor.h"
+#include "opal/datatype/opal_datatype_cuda.h"
+#endif /* OPAL_CUDA_SUPPORT */
 #include "ompi/mpi/c/bindings.h"
 #include "ompi/runtime/params.h"
 #include "ompi/communicator/communicator.h"
@@ -45,6 +50,12 @@
              MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) 
 {
     int err;
+#if OPAL_CUDA_SUPPORT
+    ptrdiff_t true_lb, true_extent, lb, extent;
+    char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2;
+    const char *sbuf2;
+    size_t bufsize;
+#endif /* OPAL_CUDA_SUPPORT */
 
     MEMCHECKER(
         memchecker_datatype(datatype);
@@ -101,10 +112,50 @@
     /* Call the coll component to actually perform the allgather */
 
     OBJ_RETAIN(op);
+
+#if OPAL_CUDA_SUPPORT
+    ompi_datatype_get_extent(datatype, &lb, &extent);
+    ompi_datatype_get_true_extent(datatype, &true_lb, &true_extent);
+    bufsize = true_extent + (ptrdiff_t)(count - 1) * extent;
+    if ((MPI_IN_PLACE != sendbuf) && (opal_cuda_check_bufs((char *)sendbuf, NULL))) {
+        sbuf1 = (char*)malloc(bufsize);
+        if (NULL == sbuf1) {
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        opal_cuda_memcpy_sync(sbuf1, sendbuf, bufsize);
+        sbuf2 = sendbuf; /* save away original buffer */
+        sendbuf = sbuf1 - lb;
+    }
+
+    if (opal_cuda_check_bufs(recvbuf, NULL)) {
+        rbuf1 = (char*)malloc(bufsize);
+        if (NULL == rbuf1) {
+            if (NULL != sbuf1) free(sbuf1);
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        opal_cuda_memcpy_sync(rbuf1, recvbuf, bufsize);
+        rbuf2 = recvbuf; /* save away original buffer */
+        recvbuf = rbuf1 - lb;
+    }
+#endif /* OPAL_CUDA_SUPPORT */
+
     /* XXX -- CONST -- do not cast away const -- update mca/coll */
     err = comm->c_coll.coll_scan((void *) sendbuf, recvbuf, count,
                                  datatype, op, comm,
                                  comm->c_coll.coll_scan_module);
+
+#if OPAL_CUDA_SUPPORT
+    if (NULL != sbuf1) {
+        free(sbuf1);
+        sendbuf = sbuf2;
+    }
+    if (NULL != rbuf1) {
+        recvbuf = rbuf2;
+        opal_cuda_memcpy_sync(recvbuf, rbuf1, bufsize);
+        free(rbuf1);
+    }
+#endif /* OPAL_CUDA_SUPPORT */
+
     OBJ_RELEASE(op);
     OMPI_ERRHANDLER_RETURN(err, comm, err, FUNC_NAME);
 }
Index: ompi/mpi/c/allreduce.c
===================================================================
--- ompi/mpi/c/allreduce.c	(revision 31741)
+++ ompi/mpi/c/allreduce.c	(working copy)
@@ -12,6 +12,7 @@
  *                         All rights reserved.
  * Copyright (c) 2013      Los Alamos National Security, LLC.  All rights
  *                         reserved.
+ * Copyright (c) 2014      NVIDIA Corporation.  All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -22,6 +23,10 @@
 #include "ompi_config.h"
 #include <stdio.h>
 
+#if OPAL_CUDA_SUPPORT
+#include "opal/datatype/opal_convertor.h"
+#include "opal/datatype/opal_datatype_cuda.h"
+#endif /* OPAL_CUDA_SUPPORT */
 #include "ompi/mpi/c/bindings.h"
 #include "ompi/runtime/params.h"
 #include "ompi/communicator/communicator.h"
@@ -45,6 +50,12 @@
                   MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) 
 {
     int err;
+#if OPAL_CUDA_SUPPORT
+    ptrdiff_t true_lb, true_extent, lb, extent;
+    char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2;
+    const char *sbuf2;
+    size_t bufsize;
+#endif /* OPAL_CUDA_SUPPORT */
 
     MEMCHECKER(
         memchecker_datatype(datatype);
@@ -105,10 +116,50 @@
     /* Invoke the coll component to perform the back-end operation */
 
     OBJ_RETAIN(op);
+
+#if OPAL_CUDA_SUPPORT
+    ompi_datatype_get_extent(datatype, &lb, &extent);
+    ompi_datatype_get_true_extent(datatype, &true_lb, &true_extent);
+    bufsize = true_extent + (ptrdiff_t)(count - 1) * extent;
+    if ((MPI_IN_PLACE != sendbuf) && (opal_cuda_check_bufs((char *)sendbuf, NULL))) {
+        sbuf1 = (char*)malloc(bufsize);
+        if (NULL == sbuf1) {
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        opal_cuda_memcpy_sync(sbuf1, sendbuf, bufsize);
+        sbuf2 = sendbuf; /* save away original buffer */
+        sendbuf = sbuf1 - lb;
+    }
+
+    if (opal_cuda_check_bufs(recvbuf, NULL)) {
+        rbuf1 = (char*)malloc(bufsize);
+        if (NULL == rbuf1) {
+            if (NULL != sbuf1) free(sbuf1);
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        opal_cuda_memcpy_sync(rbuf1, recvbuf, bufsize);
+        rbuf2 = recvbuf; /* save away original buffer */
+        recvbuf = rbuf1 - lb;
+    }
+#endif /* OPAL_CUDA_SUPPORT */
+
     /* XXX -- CONST -- do not cast away const -- update mca/coll */
     err = comm->c_coll.coll_allreduce((void *) sendbuf, recvbuf, count,
                                       datatype, op, comm,
                                       comm->c_coll.coll_allreduce_module);
+
+#if OPAL_CUDA_SUPPORT
+    if (NULL != sbuf1) {
+        free(sbuf1);
+        sendbuf = sbuf2;
+    }
+    if (NULL != rbuf1) {
+        recvbuf = rbuf2;
+        opal_cuda_memcpy_sync(recvbuf, rbuf1, bufsize);
+        free(rbuf1);
+    }
+#endif /* OPAL_CUDA_SUPPORT */
+
     OBJ_RELEASE(op);
     OMPI_ERRHANDLER_RETURN(err, comm, err, FUNC_NAME);
 }
Index: ompi/mpi/c/reduce.c
===================================================================
--- ompi/mpi/c/reduce.c	(revision 31741)
+++ ompi/mpi/c/reduce.c	(working copy)
@@ -13,6 +13,7 @@
  * Copyright (c) 2006      Cisco Systems, Inc.  All rights reserved.
  * Copyright (c) 2013      Los Alamos National Security, LLC.  All rights
  *                         reserved.
+ * Copyright (c) 2014      NVIDIA Corporation.  All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -22,6 +23,10 @@
 #include "ompi_config.h"
 #include <stdio.h>
 
+#if OPAL_CUDA_SUPPORT
+#include "opal/datatype/opal_convertor.h"
+#include "opal/datatype/opal_datatype_cuda.h"
+#endif /* OPAL_CUDA_SUPPORT */
 #include "ompi/mpi/c/bindings.h"
 #include "ompi/runtime/params.h"
 #include "ompi/communicator/communicator.h"
@@ -45,6 +50,12 @@
                MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) 
 {
     int err;
+#if OPAL_CUDA_SUPPORT
+    ptrdiff_t true_lb, true_extent, lb, extent;
+    char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2;
+    const char *sbuf2;
+    size_t bufsize;
+#endif /* OPAL_CUDA_SUPPORT */
 
     MEMCHECKER(
         memchecker_datatype(datatype);
@@ -132,10 +143,50 @@
     /* Invoke the coll component to perform the back-end operation */
 
     OBJ_RETAIN(op);
+
+#if OPAL_CUDA_SUPPORT
+    ompi_datatype_get_extent(datatype, &lb, &extent);
+    ompi_datatype_get_true_extent(datatype, &true_lb, &true_extent);
+    bufsize = true_extent + (ptrdiff_t)(count - 1) * extent;
+    if ((MPI_IN_PLACE != sendbuf) && (opal_cuda_check_bufs((char *)sendbuf, NULL))) {
+        sbuf1 = (char*)malloc(bufsize);
+        if (NULL == sbuf1) {
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        opal_cuda_memcpy_sync(sbuf1, sendbuf, bufsize);
+        sbuf2 = sendbuf; /* save away original buffer */
+        sendbuf = sbuf1 - lb;
+    }
+
+    if (opal_cuda_check_bufs(recvbuf, NULL)) {
+        rbuf1 = (char*)malloc(bufsize);
+        if (NULL == rbuf1) {
+            if (NULL != sbuf1) free(sbuf1);
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        opal_cuda_memcpy_sync(rbuf1, recvbuf, bufsize);
+        rbuf2 = recvbuf; /* save away original buffer */
+        recvbuf = rbuf1 - lb;
+    }
+#endif /* OPAL_CUDA_SUPPORT */
+
     /* XXX -- CONST -- do not cast away const -- update mca/coll */
     err = comm->c_coll.coll_reduce((void *) sendbuf, recvbuf, count,
                                    datatype, op, root, comm,
                                    comm->c_coll.coll_reduce_module);
+
+#if OPAL_CUDA_SUPPORT
+    if (NULL != sbuf1) {
+        free(sbuf1);
+        sendbuf = sbuf2;
+    }
+    if (NULL != rbuf1) {
+        recvbuf = rbuf2;
+        opal_cuda_memcpy_sync(recvbuf, rbuf1, bufsize);
+        free(rbuf1);
+    }
+#endif /* OPAL_CUDA_SUPPORT */
+
     OBJ_RELEASE(op);
     OMPI_ERRHANDLER_RETURN(err, comm, err, FUNC_NAME);
 }