|
| 1 | +/* |
| 2 | + * Copyright (c) 2016 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 | + |
| 15 | +#include "orte/mca/grpcomm/grpcomm.h" |
| 16 | + |
| 17 | +#include "oshmem/runtime/runtime.h" |
| 18 | + |
| 19 | +#include "oshmem/mca/scoll/scoll.h" |
| 20 | + |
| 21 | +#include "oshmem/proc/proc.h" |
| 22 | +#include "oshmem/proc/proc_group_cache.h" |
| 23 | + |
| 24 | +static void _shmem_alltoall(void *target, |
| 25 | + const void *source, |
| 26 | + ptrdiff_t dst, ptrdiff_t sst, |
| 27 | + size_t nelems, |
| 28 | + size_t element_size, |
| 29 | + int PE_start, |
| 30 | + int logPE_stride, |
| 31 | + int PE_size, |
| 32 | + long *pSync); |
| 33 | + |
| 34 | +#define SHMEM_TYPE_ALLTOALL(name, element_size) \ |
| 35 | + void shmem##name(void *target, \ |
| 36 | + const void *source, \ |
| 37 | + size_t nelems, \ |
| 38 | + int PE_start, \ |
| 39 | + int logPE_stride, \ |
| 40 | + int PE_size, \ |
| 41 | + long *pSync) \ |
| 42 | +{ \ |
| 43 | + RUNTIME_CHECK_INIT(); \ |
| 44 | + RUNTIME_CHECK_ADDR(target); \ |
| 45 | + RUNTIME_CHECK_ADDR(source); \ |
| 46 | + \ |
| 47 | + _shmem_alltoall(target, source, 1, 1, nelems * element_size, \ |
| 48 | + PE_start, logPE_stride, PE_size, \ |
| 49 | + pSync); \ |
| 50 | +} |
| 51 | + |
| 52 | +#define SHMEM_TYPE_ALLTOALLS(name, element_size) \ |
| 53 | + void shmem##name(void *target, \ |
| 54 | + const void *source, \ |
| 55 | + ptrdiff_t dst, ptrdiff_t sst, \ |
| 56 | + size_t nelems, \ |
| 57 | + int PE_start, \ |
| 58 | + int logPE_stride, \ |
| 59 | + int PE_size, \ |
| 60 | + long *pSync) \ |
| 61 | +{ \ |
| 62 | + RUNTIME_CHECK_INIT(); \ |
| 63 | + RUNTIME_CHECK_ADDR(target); \ |
| 64 | + RUNTIME_CHECK_ADDR(source); \ |
| 65 | + \ |
| 66 | + _shmem_alltoall(target, source, dst, sst, nelems * element_size, \ |
| 67 | + PE_start, logPE_stride, PE_size, \ |
| 68 | + pSync); \ |
| 69 | +} |
| 70 | + |
| 71 | +static void _shmem_alltoall(void *target, |
| 72 | + const void *source, |
| 73 | + ptrdiff_t dst, ptrdiff_t sst, |
| 74 | + size_t nelems, |
| 75 | + size_t element_size, |
| 76 | + int PE_start, |
| 77 | + int logPE_stride, |
| 78 | + int PE_size, |
| 79 | + long *pSync) |
| 80 | +{ |
| 81 | + int rc = OSHMEM_SUCCESS; |
| 82 | + oshmem_group_t* group = NULL; |
| 83 | + |
| 84 | + if ((0 <= PE_start) && (0 <= logPE_stride)) { |
| 85 | + /* Create group basing PE_start, logPE_stride and PE_size */ |
| 86 | +#if OSHMEM_GROUP_CACHE_ENABLED == 0 |
| 87 | + group = oshmem_proc_group_create(PE_start, (1 << logPE_stride), PE_size); |
| 88 | + if (!group) |
| 89 | + rc = OSHMEM_ERROR; |
| 90 | +#else |
| 91 | + group = find_group_in_cache(PE_start, logPE_stride, PE_size); |
| 92 | + if (!group) { |
| 93 | + group = oshmem_proc_group_create(PE_start, |
| 94 | + (1 << logPE_stride), |
| 95 | + PE_size); |
| 96 | + if (!group) { |
| 97 | + rc = OSHMEM_ERROR; |
| 98 | + } else { |
| 99 | + cache_group(group, PE_start, logPE_stride, PE_size); |
| 100 | + } |
| 101 | + } |
| 102 | +#endif /* OSHMEM_GROUP_CACHE_ENABLED */ |
| 103 | + |
| 104 | + /* Collective operation call */ |
| 105 | + if (rc == OSHMEM_SUCCESS) { |
| 106 | + /* Call collective alltoall operation */ |
| 107 | + rc = group->g_scoll.scoll_alltoall(group, |
| 108 | + target, |
| 109 | + source, |
| 110 | + dst, |
| 111 | + sst, |
| 112 | + nelems, |
| 113 | + element_size, |
| 114 | + pSync, |
| 115 | + SCOLL_DEFAULT_ALG); |
| 116 | + } |
| 117 | +#if OSHMEM_GROUP_CACHE_ENABLED == 0 |
| 118 | + if ( rc == OSHMEM_SUCCESS ) { |
| 119 | + oshmem_proc_group_destroy(group); |
| 120 | + } |
| 121 | +#endif /* OSHMEM_GROUP_CACHE_ENABLED */ |
| 122 | + } |
| 123 | +} |
| 124 | + |
| 125 | +#if OSHMEM_PROFILING |
| 126 | +#include "oshmem/include/pshmem.h" |
| 127 | +#pragma weak shmem_alltoall32 = pshmem_alltoall32 |
| 128 | +#pragma weak shmem_alltoall64 = pshmem_alltoall64 |
| 129 | +#pragma weak shmem_alltoalls32 = pshmem_alltoalls32 |
| 130 | +#pragma weak shmem_alltoalls64 = pshmem_alltoalls64 |
| 131 | +#include "oshmem/shmem/c/profile/defines.h" |
| 132 | +#endif |
| 133 | + |
| 134 | +SHMEM_TYPE_ALLTOALL(_alltoall32, sizeof(uint32_t)) |
| 135 | +SHMEM_TYPE_ALLTOALL(_alltoall64, sizeof(uint64_t)) |
| 136 | +SHMEM_TYPE_ALLTOALLS(_alltoalls32, sizeof(uint32_t)) |
| 137 | +SHMEM_TYPE_ALLTOALLS(_alltoalls64, sizeof(uint64_t)) |
0 commit comments