root/ompi/mca/coll/cuda/coll_cuda_reduce.c

DEFINITIONS

1. mca_coll_cuda_reduce

   1 /*
   2  * Copyright (c) 2004-2015 The University of Tennessee and The University
   3  *                         of Tennessee Research Foundation.  All rights
   4  *                         reserved.
   5  * Copyright (c) 2014-2015 NVIDIA Corporation.  All rights reserved.
   6  * $COPYRIGHT$
   7  *
   8  * Additional copyrights may follow
   9  *
  10  * $HEADER$
  11  */
  12 
  13 #include "ompi_config.h"
  14 #include "coll_cuda.h"
  15 
  16 #include <stdio.h>
  17 
  18 #include "ompi/op/op.h"
  19 #include "opal/datatype/opal_convertor.h"
  20 #include "opal/datatype/opal_datatype_cuda.h"
  21 
  22 /*
  23  *      reduce_log_inter
  24  *
  25  *      Function:       - reduction using O(N) algorithm
  26  *      Accepts:        - same as MPI_Reduce()
  27  *      Returns:        - MPI_SUCCESS or error code
  28  */
  29 int
  30 mca_coll_cuda_reduce(const void *sbuf, void *rbuf, int count,
  31                      struct ompi_datatype_t *dtype,
  32                      struct ompi_op_t *op,
  33                      int root, struct ompi_communicator_t *comm,
  34                      mca_coll_base_module_t *module)
  35 {
  36     mca_coll_cuda_module_t *s = (mca_coll_cuda_module_t*) module;
  37     ptrdiff_t gap;
  38     char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2 = NULL;
  39     const char *sbuf2;
  40     size_t bufsize;
  41     int rc;
  42 
  43     bufsize = opal_datatype_span(&dtype->super, count, &gap);
  44 
  45 
  46     if ((MPI_IN_PLACE != sbuf) && (opal_cuda_check_bufs((char *)sbuf, NULL))) {
  47         sbuf1 = (char*)malloc(bufsize);
  48         if (NULL == sbuf1) {
  49             return OMPI_ERR_OUT_OF_RESOURCE;
  50         }
  51         opal_cuda_memcpy_sync(sbuf1, sbuf, bufsize);
  52         sbuf2 = sbuf; /* save away original buffer */
  53         sbuf = sbuf1 - gap;
  54     }
  55 
  56     if (opal_cuda_check_bufs(rbuf, NULL)) {
  57         rbuf1 = (char*)malloc(bufsize);
  58         if (NULL == rbuf1) {
  59             if (NULL != sbuf1) free(sbuf1);
  60             return OMPI_ERR_OUT_OF_RESOURCE;
  61         }
  62         opal_cuda_memcpy_sync(rbuf1, rbuf, bufsize);
  63         rbuf2 = rbuf; /* save away original buffer */
  64         rbuf = rbuf1 - gap;
  65     }
  66     rc = s->c_coll.coll_reduce((void *) sbuf, rbuf, count,
  67                                dtype, op, root, comm,
  68                                s->c_coll.coll_reduce_module);
  69 
  70     if (NULL != sbuf1) {
  71         free(sbuf1);
  72     }
  73     if (NULL != rbuf1) {
  74         rbuf = rbuf2;
  75         opal_cuda_memcpy_sync(rbuf, rbuf1, bufsize);
  76         free(rbuf1);
  77     }
  78     return rc;
  79 }