root/oshmem/shmem/c/shmem_reduce.c

/* [<][>][^][v][top][bottom][index][help] */
   1 /*
   2  * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
   3  *                         All rights reserved.
   4  * $COPYRIGHT$
   5  *
   6  * Additional copyrights may follow
   7  *
   8  * $HEADER$
   9  */
  10 #include "oshmem_config.h"
  11 
  12 #include "oshmem/constants.h"
  13 #include "oshmem/include/shmem.h"
  14 #include "oshmem/include/shmemx.h"
  15 
  16 #include "oshmem/runtime/runtime.h"
  17 
  18 #include "oshmem/mca/scoll/scoll.h"
  19 #include "oshmem/proc/proc.h"
  20 #include "oshmem/op/op.h"
  21 
  22 /*
  23  * The shared memory (SHMEM) reduction routines perform an associative binary operation
  24  * across symmetric arrays on multiple virtual PEs.
  25  * This routine returns the result of performing a operation on the source data
  26  * object of every PE in the active set. The active set of PEs is defined by the triple PE_start,
  27  * logPE_stride and PE_size.
  28  */
  29 #define SHMEM_TYPE_REDUCE_OP(name, type_name, type, prefix)                                 \
  30     void prefix##type_name##_##name##_to_all( type *target,                                 \
  31                                         const type *source,                                 \
  32                                         int nreduce,                                        \
  33                                         int PE_start,                                       \
  34                                         int logPE_stride,                                   \
  35                                         int PE_size,                                        \
  36                                         type *pWrk,                                         \
  37                                         long *pSync )                                       \
  38 {                                                                                           \
  39     int rc = OSHMEM_SUCCESS;                                                                \
  40     oshmem_group_t*  group = NULL;                                                          \
  41                                                                                             \
  42     RUNTIME_CHECK_INIT();                                                                   \
  43     RUNTIME_CHECK_ADDR_SIZE(target, nreduce);                                               \
  44     RUNTIME_CHECK_ADDR_SIZE(source, nreduce);                                               \
  45                                                                                             \
  46     {                                                                                       \
  47         group = oshmem_proc_group_create_nofail(PE_start, 1<<logPE_stride, PE_size);        \
  48         oshmem_op_t* op = oshmem_op_##name##type_name;                                      \
  49         size_t size = nreduce * op->dt_size;                                                \
  50                                                                                             \
  51         /* Call collective reduce operation */                                              \
  52         rc = group->g_scoll.scoll_reduce(                                                   \
  53                 group,                                                                      \
  54                 op,                                                                         \
  55                 (void*)target,                                                              \
  56                 (const void*)source,                                                        \
  57                 size,                                                                       \
  58                 pSync,                                                                      \
  59                 (void*)pWrk,                                                                \
  60                 SCOLL_DEFAULT_ALG );                                                        \
  61                                                                                             \
  62         oshmem_proc_group_destroy(group);                                                   \
  63     }                                                                                       \
  64     RUNTIME_CHECK_RC(rc);                                                                   \
  65 }
  66 
  67 #if OSHMEM_PROFILING
  68 #include "oshmem/include/pshmem.h"
  69 #pragma weak shmem_short_and_to_all     =  pshmem_short_and_to_all
  70 #pragma weak shmem_int_and_to_all       =  pshmem_int_and_to_all
  71 #pragma weak shmem_long_and_to_all      =  pshmem_long_and_to_all
  72 #pragma weak shmem_longlong_and_to_all  =  pshmem_longlong_and_to_all
  73 #pragma weak shmemx_int16_and_to_all    =  pshmemx_int16_and_to_all
  74 #pragma weak shmemx_int32_and_to_all    =  pshmemx_int32_and_to_all
  75 #pragma weak shmemx_int64_and_to_all    =  pshmemx_int64_and_to_all
  76 
  77 #pragma weak shmem_short_or_to_all      =  pshmem_short_or_to_all
  78 #pragma weak shmem_int_or_to_all        =  pshmem_int_or_to_all
  79 #pragma weak shmem_long_or_to_all       =  pshmem_long_or_to_all
  80 #pragma weak shmem_longlong_or_to_all   =  pshmem_longlong_or_to_all
  81 #pragma weak shmemx_int16_or_to_all     =  pshmemx_int16_or_to_all
  82 #pragma weak shmemx_int32_or_to_all     =  pshmemx_int32_or_to_all
  83 #pragma weak shmemx_int64_or_to_all     =  pshmemx_int64_or_to_all
  84 
  85 #pragma weak shmem_short_xor_to_all     =  pshmem_short_xor_to_all
  86 #pragma weak shmem_int_xor_to_all       =  pshmem_int_xor_to_all
  87 #pragma weak shmem_long_xor_to_all      =  pshmem_long_xor_to_all
  88 #pragma weak shmem_longlong_xor_to_all  =  pshmem_longlong_xor_to_all
  89 #pragma weak shmemx_int16_xor_to_all    =  pshmemx_int16_xor_to_all
  90 #pragma weak shmemx_int32_xor_to_all    =  pshmemx_int32_xor_to_all
  91 #pragma weak shmemx_int64_xor_to_all    =  pshmemx_int64_xor_to_all
  92 
  93 #pragma weak shmem_short_max_to_all     =  pshmem_short_max_to_all
  94 #pragma weak shmem_int_max_to_all       =  pshmem_int_max_to_all
  95 #pragma weak shmem_long_max_to_all      =  pshmem_long_max_to_all
  96 #pragma weak shmem_longlong_max_to_all  =  pshmem_longlong_max_to_all
  97 #pragma weak shmem_float_max_to_all     =  pshmem_float_max_to_all
  98 #pragma weak shmem_double_max_to_all    =  pshmem_double_max_to_all
  99 #pragma weak shmem_longdouble_max_to_all=  pshmem_longdouble_max_to_all
 100 #pragma weak shmemx_int16_max_to_all    =  pshmemx_int16_max_to_all
 101 #pragma weak shmemx_int32_max_to_all    =  pshmemx_int32_max_to_all
 102 #pragma weak shmemx_int64_max_to_all    =  pshmemx_int64_max_to_all
 103 
 104 #pragma weak shmem_short_min_to_all     =  pshmem_short_min_to_all
 105 #pragma weak shmem_int_min_to_all       =  pshmem_int_min_to_all
 106 #pragma weak shmem_long_min_to_all      =  pshmem_long_min_to_all
 107 #pragma weak shmem_longlong_min_to_all  =  pshmem_longlong_min_to_all
 108 #pragma weak shmem_float_min_to_all     =  pshmem_float_min_to_all
 109 #pragma weak shmem_double_min_to_all    =  pshmem_double_min_to_all
 110 #pragma weak shmem_longdouble_min_to_all=  pshmem_longdouble_min_to_all
 111 #pragma weak shmemx_int16_min_to_all    =  pshmemx_int16_min_to_all
 112 #pragma weak shmemx_int32_min_to_all    =  pshmemx_int32_min_to_all
 113 #pragma weak shmemx_int64_min_to_all    =  pshmemx_int64_min_to_all
 114 
 115 #pragma weak shmem_short_sum_to_all     =  pshmem_short_sum_to_all
 116 #pragma weak shmem_int_sum_to_all       =  pshmem_int_sum_to_all
 117 #pragma weak shmem_long_sum_to_all      =  pshmem_long_sum_to_all
 118 #pragma weak shmem_longlong_sum_to_all  =  pshmem_longlong_sum_to_all
 119 #pragma weak shmem_float_sum_to_all     =  pshmem_float_sum_to_all
 120 #pragma weak shmem_double_sum_to_all    =  pshmem_double_sum_to_all
 121 #pragma weak shmem_longdouble_sum_to_all=  pshmem_longdouble_sum_to_all
 122 #pragma weak shmem_complexf_sum_to_all  =  pshmem_complexf_sum_to_all
 123 #pragma weak shmem_complexd_sum_to_all  =  pshmem_complexd_sum_to_all
 124 #pragma weak shmemx_int16_sum_to_all    =  pshmemx_int16_sum_to_all
 125 #pragma weak shmemx_int32_sum_to_all    =  pshmemx_int32_sum_to_all
 126 #pragma weak shmemx_int64_sum_to_all    =  pshmemx_int64_sum_to_all
 127 
 128 #pragma weak shmem_short_prod_to_all    =  pshmem_short_prod_to_all
 129 #pragma weak shmem_int_prod_to_all      =  pshmem_int_prod_to_all
 130 #pragma weak shmem_long_prod_to_all     =  pshmem_long_prod_to_all
 131 #pragma weak shmem_longlong_prod_to_all =  pshmem_longlong_prod_to_all
 132 #pragma weak shmem_float_prod_to_all    =  pshmem_float_prod_to_all
 133 #pragma weak shmem_double_prod_to_all   =  pshmem_double_prod_to_all
 134 #pragma weak shmem_longdouble_prod_to_all = pshmem_longdouble_prod_to_all
 135 #pragma weak shmem_complexf_prod_to_all =  pshmem_complexf_prod_to_all
 136 #pragma weak shmem_complexd_prod_to_all =  pshmem_complexd_prod_to_all
 137 #pragma weak shmemx_int16_prod_to_all   =  pshmemx_int16_prod_to_all
 138 #pragma weak shmemx_int32_prod_to_all   =  pshmemx_int32_prod_to_all
 139 #pragma weak shmemx_int64_prod_to_all   =  pshmemx_int64_prod_to_all
 140 #include "oshmem/shmem/c/profile/defines.h"
 141 #endif /* OSHMEM_PROFILING */
 142 
 143 SHMEM_TYPE_REDUCE_OP(and, _short, short, shmem)
 144 SHMEM_TYPE_REDUCE_OP(and, _int, int, shmem)
 145 SHMEM_TYPE_REDUCE_OP(and, _long, long, shmem)
 146 SHMEM_TYPE_REDUCE_OP(and, _longlong, long long, shmem)
 147 SHMEM_TYPE_REDUCE_OP(and, _int16, int16_t, shmemx)
 148 SHMEM_TYPE_REDUCE_OP(and, _int32, int32_t, shmemx)
 149 SHMEM_TYPE_REDUCE_OP(and, _int64, int64_t, shmemx)
 150 
 151 SHMEM_TYPE_REDUCE_OP(or, _short, short, shmem)
 152 SHMEM_TYPE_REDUCE_OP(or, _int, int, shmem)
 153 SHMEM_TYPE_REDUCE_OP(or, _long, long, shmem)
 154 SHMEM_TYPE_REDUCE_OP(or, _longlong, long long, shmem)
 155 SHMEM_TYPE_REDUCE_OP(or, _int16, int16_t, shmemx)
 156 SHMEM_TYPE_REDUCE_OP(or, _int32, int32_t, shmemx)
 157 SHMEM_TYPE_REDUCE_OP(or, _int64, int64_t, shmemx)
 158 
 159 SHMEM_TYPE_REDUCE_OP(xor, _short, short, shmem)
 160 SHMEM_TYPE_REDUCE_OP(xor, _int, int, shmem)
 161 SHMEM_TYPE_REDUCE_OP(xor, _long, long, shmem)
 162 SHMEM_TYPE_REDUCE_OP(xor, _longlong, long long, shmem)
 163 SHMEM_TYPE_REDUCE_OP(xor, _int16, int16_t, shmemx)
 164 SHMEM_TYPE_REDUCE_OP(xor, _int32, int32_t, shmemx)
 165 SHMEM_TYPE_REDUCE_OP(xor, _int64, int64_t, shmemx)
 166 
 167 SHMEM_TYPE_REDUCE_OP(max, _short, short, shmem)
 168 SHMEM_TYPE_REDUCE_OP(max, _int, int, shmem)
 169 SHMEM_TYPE_REDUCE_OP(max, _long, long, shmem)
 170 SHMEM_TYPE_REDUCE_OP(max, _longlong, long long, shmem)
 171 SHMEM_TYPE_REDUCE_OP(max, _float, float, shmem)
 172 SHMEM_TYPE_REDUCE_OP(max, _double, double, shmem)
 173 SHMEM_TYPE_REDUCE_OP(max, _longdouble, long double, shmem)
 174 SHMEM_TYPE_REDUCE_OP(max, _int16, int16_t, shmemx)
 175 SHMEM_TYPE_REDUCE_OP(max, _int32, int32_t, shmemx)
 176 SHMEM_TYPE_REDUCE_OP(max, _int64, int64_t, shmemx)
 177 
 178 SHMEM_TYPE_REDUCE_OP(min, _short, short, shmem)
 179 SHMEM_TYPE_REDUCE_OP(min, _int, int, shmem)
 180 SHMEM_TYPE_REDUCE_OP(min, _long, long, shmem)
 181 SHMEM_TYPE_REDUCE_OP(min, _longlong, long long, shmem)
 182 SHMEM_TYPE_REDUCE_OP(min, _float, float, shmem)
 183 SHMEM_TYPE_REDUCE_OP(min, _double, double, shmem)
 184 SHMEM_TYPE_REDUCE_OP(min, _longdouble, long double, shmem)
 185 SHMEM_TYPE_REDUCE_OP(min, _int16, int16_t, shmemx)
 186 SHMEM_TYPE_REDUCE_OP(min, _int32, int32_t, shmemx)
 187 SHMEM_TYPE_REDUCE_OP(min, _int64, int64_t, shmemx)
 188 
 189 SHMEM_TYPE_REDUCE_OP(sum, _short, short, shmem)
 190 SHMEM_TYPE_REDUCE_OP(sum, _int, int, shmem)
 191 SHMEM_TYPE_REDUCE_OP(sum, _long, long, shmem)
 192 SHMEM_TYPE_REDUCE_OP(sum, _longlong, long long, shmem)
 193 SHMEM_TYPE_REDUCE_OP(sum, _float, float, shmem)
 194 SHMEM_TYPE_REDUCE_OP(sum, _double, double, shmem)
 195 SHMEM_TYPE_REDUCE_OP(sum, _longdouble, long double, shmem)
 196 SHMEM_TYPE_REDUCE_OP(sum, _complexf, float complex, shmem)
 197 SHMEM_TYPE_REDUCE_OP(sum, _complexd, double complex, shmem)
 198 SHMEM_TYPE_REDUCE_OP(sum, _int16, int16_t, shmemx)
 199 SHMEM_TYPE_REDUCE_OP(sum, _int32, int32_t, shmemx)
 200 SHMEM_TYPE_REDUCE_OP(sum, _int64, int64_t, shmemx)
 201 
 202 SHMEM_TYPE_REDUCE_OP(prod, _short, short, shmem)
 203 SHMEM_TYPE_REDUCE_OP(prod, _int, int, shmem)
 204 SHMEM_TYPE_REDUCE_OP(prod, _long, long, shmem)
 205 SHMEM_TYPE_REDUCE_OP(prod, _longlong, long long, shmem)
 206 SHMEM_TYPE_REDUCE_OP(prod, _float, float, shmem)
 207 SHMEM_TYPE_REDUCE_OP(prod, _double, double, shmem)
 208 SHMEM_TYPE_REDUCE_OP(prod, _longdouble, long double, shmem)
 209 SHMEM_TYPE_REDUCE_OP(prod, _complexf, float complex, shmem)
 210 SHMEM_TYPE_REDUCE_OP(prod, _complexd, double complex, shmem)
 211 SHMEM_TYPE_REDUCE_OP(prod, _int16, int16_t, shmemx)
 212 SHMEM_TYPE_REDUCE_OP(prod, _int32, int32_t, shmemx)
 213 SHMEM_TYPE_REDUCE_OP(prod, _int64, int64_t, shmemx)

/* [<][>][^][v][top][bottom][index][help] */