|
| 1 | +/* |
| 2 | + * Copyright (c) 2013-2016 Mellanox Technologies, Inc. |
| 3 | + * All rights reserved. |
| 4 | + * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. |
| 5 | + * $COPYRIGHT$ |
| 6 | + * |
| 7 | + * Additional copyrights may follow |
| 8 | + * |
| 9 | + * $HEADER$ |
| 10 | + */ |
| 11 | + |
| 12 | +#include "oshmem_config.h" |
| 13 | +#include "oshmem/shmem/fortran/bindings.h" |
| 14 | +#include "oshmem/include/shmem.h" |
| 15 | +#include "oshmem/constants.h" |
| 16 | +#include "oshmem/mca/scoll/scoll.h" |
| 17 | +#include "oshmem/proc/proc.h" |
| 18 | +#include "oshmem/proc/proc_group_cache.h" |
| 19 | +#include "oshmem/op/op.h" |
| 20 | + |
| 21 | +#if OSHMEM_PROFILING |
| 22 | +#include "oshmem/shmem/fortran/profile/pbindings.h" |
| 23 | +SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_ALLTOALL32, shmem_alltoall32) |
| 24 | +SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_ALLTOALL64, shmem_alltoall64) |
| 25 | +SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_ALLTOALLS32, shmem_alltoalls32) |
| 26 | +SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_ALLTOALLS64, shmem_alltoalls64) |
| 27 | +#include "oshmem/shmem/fortran/profile/defines.h" |
| 28 | +#endif |
| 29 | + |
| 30 | +SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void, |
| 31 | + SHMEM_ALLTOALL32, |
| 32 | + shmem_alltoall32_, |
| 33 | + shmem_alltoall32__, |
| 34 | + shmem_alltoall32_f, |
| 35 | + (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync), |
| 36 | + (target, source, nlong, PE_start, logPE_stride, PE_size, pSync)) |
| 37 | + |
| 38 | +SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void, |
| 39 | + SHMEM_ALLTOALL64, |
| 40 | + shmem_alltoall64_, |
| 41 | + shmem_alltoall64__, |
| 42 | + shmem_alltoall64_f, |
| 43 | + (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync), |
| 44 | + (target, source, nlong, PE_start, logPE_stride, PE_size, pSync)) |
| 45 | + |
| 46 | +SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void, |
| 47 | + SHMEM_ALLTOALLS32, |
| 48 | + shmem_alltoalls32_, |
| 49 | + shmem_alltoalls32__, |
| 50 | + shmem_alltoalls32_f, |
| 51 | + (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *dst, MPI_Fint *sst, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync), |
| 52 | + (target, source, dst, sst, nlong, PE_start, logPE_stride, PE_size, pSync)) |
| 53 | + |
| 54 | +SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void, |
| 55 | + SHMEM_ALLTOALLS64, |
| 56 | + shmem_alltoalls64_, |
| 57 | + shmem_alltoalls64__, |
| 58 | + shmem_alltoalls64_f, |
| 59 | + (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *dst, MPI_Fint *sst, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync), |
| 60 | + (target, source, dst, sst, nlong, PE_start, logPE_stride, PE_size, pSync)) |
| 61 | + |
| 62 | +#define SHMEM_ALLTOALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ |
| 63 | + FORTRAN_POINTER_T source, \ |
| 64 | + MPI_Fint *nlong,\ |
| 65 | + MPI_Fint *PE_start, \ |
| 66 | + MPI_Fint *logPE_stride, \ |
| 67 | + MPI_Fint *PE_size, \ |
| 68 | + FORTRAN_POINTER_T pSync)\ |
| 69 | +{\ |
| 70 | + int rc = OSHMEM_SUCCESS;\ |
| 71 | + oshmem_group_t* group = NULL;\ |
| 72 | + {\ |
| 73 | + /* Create group basing PE_start, logPE_stride and PE_size */\ |
| 74 | + if (OSHMEM_GROUP_CACHE_ENABLED == 0)\ |
| 75 | + {\ |
| 76 | + group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \ |
| 77 | + (1 << OMPI_FINT_2_INT(*logPE_stride)), \ |
| 78 | + OMPI_FINT_2_INT(*PE_size));\ |
| 79 | + if (!group)\ |
| 80 | + rc = OSHMEM_ERROR;\ |
| 81 | + }\ |
| 82 | + else\ |
| 83 | + {\ |
| 84 | + group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ |
| 85 | + OMPI_FINT_2_INT(*logPE_stride),\ |
| 86 | + OMPI_FINT_2_INT(*PE_size));\ |
| 87 | + if (!group)\ |
| 88 | + {\ |
| 89 | + group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \ |
| 90 | + (1 << OMPI_FINT_2_INT(*logPE_stride)), \ |
| 91 | + OMPI_FINT_2_INT(*PE_size));\ |
| 92 | + if (!group)\ |
| 93 | + rc = OSHMEM_ERROR;\ |
| 94 | + cache_group(group,OMPI_FINT_2_INT(*PE_start),\ |
| 95 | + OMPI_FINT_2_INT(*logPE_stride),\ |
| 96 | + OMPI_FINT_2_INT(*PE_size));\ |
| 97 | + }\ |
| 98 | + } /* OSHMEM_GROUP_CACHE_ENABLED */\ |
| 99 | + /* Collective operation call */\ |
| 100 | + if ( rc == OSHMEM_SUCCESS )\ |
| 101 | + {\ |
| 102 | + oshmem_op_t* op = T_NAME;\ |
| 103 | +\ |
| 104 | + /* Call collective broadcast operation */\ |
| 105 | + rc = group->g_scoll.scoll_alltoall( group, \ |
| 106 | + FPTR_2_VOID_PTR(target), \ |
| 107 | + FPTR_2_VOID_PTR(source), \ |
| 108 | + 1, \ |
| 109 | + 1, \ |
| 110 | + OMPI_FINT_2_INT(*nlong), \ |
| 111 | + op->dt_size, \ |
| 112 | + FPTR_2_VOID_PTR(pSync), SCOLL_DEFAULT_ALG );\ |
| 113 | + }\ |
| 114 | + if (OSHMEM_GROUP_CACHE_ENABLED == 0) \ |
| 115 | + {\ |
| 116 | + if ( group )\ |
| 117 | + {\ |
| 118 | + oshmem_proc_group_destroy(group);\ |
| 119 | + }\ |
| 120 | + } /* OSHMEM_GROUP_CACHE_ENABLED */\ |
| 121 | + }\ |
| 122 | +} |
| 123 | + |
| 124 | +#define SHMEM_ALLTOALLS(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ |
| 125 | + FORTRAN_POINTER_T source, \ |
| 126 | + MPI_Fint *dst,\ |
| 127 | + MPI_Fint *sst,\ |
| 128 | + MPI_Fint *nlong,\ |
| 129 | + MPI_Fint *PE_start, \ |
| 130 | + MPI_Fint *logPE_stride, \ |
| 131 | + MPI_Fint *PE_size, \ |
| 132 | + FORTRAN_POINTER_T pSync)\ |
| 133 | +{\ |
| 134 | + int rc = OSHMEM_SUCCESS;\ |
| 135 | + oshmem_group_t* group = NULL;\ |
| 136 | + {\ |
| 137 | + /* Create group basing PE_start, logPE_stride and PE_size */\ |
| 138 | + if (OSHMEM_GROUP_CACHE_ENABLED == 0)\ |
| 139 | + {\ |
| 140 | + group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \ |
| 141 | + (1 << OMPI_FINT_2_INT(*logPE_stride)), \ |
| 142 | + OMPI_FINT_2_INT(*PE_size));\ |
| 143 | + if (!group)\ |
| 144 | + rc = OSHMEM_ERROR;\ |
| 145 | + }\ |
| 146 | + else\ |
| 147 | + {\ |
| 148 | + group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ |
| 149 | + OMPI_FINT_2_INT(*logPE_stride),\ |
| 150 | + OMPI_FINT_2_INT(*PE_size));\ |
| 151 | + if (!group)\ |
| 152 | + {\ |
| 153 | + group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \ |
| 154 | + (1 << OMPI_FINT_2_INT(*logPE_stride)), \ |
| 155 | + OMPI_FINT_2_INT(*PE_size));\ |
| 156 | + if (!group)\ |
| 157 | + rc = OSHMEM_ERROR;\ |
| 158 | + cache_group(group,OMPI_FINT_2_INT(*PE_start),\ |
| 159 | + OMPI_FINT_2_INT(*logPE_stride),\ |
| 160 | + OMPI_FINT_2_INT(*PE_size));\ |
| 161 | + }\ |
| 162 | + } /* OSHMEM_GROUP_CACHE_ENABLED */\ |
| 163 | + /* Collective operation call */\ |
| 164 | + if ( rc == OSHMEM_SUCCESS )\ |
| 165 | + {\ |
| 166 | + oshmem_op_t* op = T_NAME;\ |
| 167 | +\ |
| 168 | + /* Call collective broadcast operation */\ |
| 169 | + rc = group->g_scoll.scoll_alltoall( group, \ |
| 170 | + FPTR_2_VOID_PTR(target), \ |
| 171 | + FPTR_2_VOID_PTR(source), \ |
| 172 | + OMPI_FINT_2_INT(*dst), \ |
| 173 | + OMPI_FINT_2_INT(*sst), \ |
| 174 | + OMPI_FINT_2_INT(*nlong), \ |
| 175 | + op->dt_size, \ |
| 176 | + FPTR_2_VOID_PTR(pSync), SCOLL_DEFAULT_ALG );\ |
| 177 | + }\ |
| 178 | + if (OSHMEM_GROUP_CACHE_ENABLED == 0) \ |
| 179 | + {\ |
| 180 | + if ( group )\ |
| 181 | + {\ |
| 182 | + oshmem_proc_group_destroy(group);\ |
| 183 | + }\ |
| 184 | + } /* OSHMEM_GROUP_CACHE_ENABLED */\ |
| 185 | + }\ |
| 186 | +} |
| 187 | + |
| 188 | +SHMEM_ALLTOALL(shmem_alltoall32_f, oshmem_op_prod_fint4, OSHMEM_GROUP_CACHE_ENABLED) |
| 189 | +SHMEM_ALLTOALL(shmem_alltoall64_f, oshmem_op_prod_fint8, OSHMEM_GROUP_CACHE_ENABLED) |
| 190 | +SHMEM_ALLTOALLS(shmem_alltoalls32_f, oshmem_op_prod_fint4, OSHMEM_GROUP_CACHE_ENABLED) |
| 191 | +SHMEM_ALLTOALLS(shmem_alltoalls64_f, oshmem_op_prod_fint8, OSHMEM_GROUP_CACHE_ENABLED) |
0 commit comments