oshmem: scoll: fixes strided alltoall
Signed-off-by: Alex Mikheev <alexm@mellanox.com>
Этот коммит содержится в:
Alex Mikheev
родитель
e7f91f8068
Коммит
cca67a69ea
@ -19,13 +19,19 @@
|
||||
#include "oshmem/mca/scoll/base/base.h"
|
||||
#include "scoll_basic.h"
|
||||
|
||||
static int _algorithm_simple(struct oshmem_group_t *group,
|
||||
void *target,
|
||||
const void *source,
|
||||
ptrdiff_t dst, ptrdiff_t sst,
|
||||
size_t nelems,
|
||||
size_t element_size,
|
||||
long *pSync);
|
||||
static int a2a_alg_simple(struct oshmem_group_t *group,
|
||||
void *target,
|
||||
const void *source,
|
||||
size_t nelems,
|
||||
size_t element_size);
|
||||
|
||||
static int a2as_alg_simple(struct oshmem_group_t *group,
|
||||
void *target,
|
||||
const void *source,
|
||||
ptrdiff_t dst, ptrdiff_t sst,
|
||||
size_t nelems,
|
||||
size_t element_size);
|
||||
|
||||
|
||||
int mca_scoll_basic_alltoall(struct oshmem_group_t *group,
|
||||
void *target,
|
||||
@ -36,88 +42,150 @@ int mca_scoll_basic_alltoall(struct oshmem_group_t *group,
|
||||
long *pSync,
|
||||
int alg)
|
||||
{
|
||||
int rc = OSHMEM_SUCCESS;
|
||||
int rc;
|
||||
int i;
|
||||
|
||||
/* Arguments validation */
|
||||
if (!group) {
|
||||
SCOLL_ERROR("Active set (group) of PE is not defined");
|
||||
rc = OSHMEM_ERR_BAD_PARAM;
|
||||
return OSHMEM_ERR_BAD_PARAM;
|
||||
}
|
||||
|
||||
/* Check if this PE is part of the group */
|
||||
if ((rc == OSHMEM_SUCCESS) && oshmem_proc_group_is_member(group)) {
|
||||
int i = 0;
|
||||
|
||||
if (pSync) {
|
||||
rc = _algorithm_simple(group,
|
||||
target,
|
||||
source,
|
||||
dst,
|
||||
sst,
|
||||
nelems,
|
||||
element_size,
|
||||
pSync);
|
||||
} else {
|
||||
SCOLL_ERROR("Incorrect argument pSync");
|
||||
rc = OSHMEM_ERR_BAD_PARAM;
|
||||
}
|
||||
|
||||
/* Restore initial values */
|
||||
SCOLL_VERBOSE(12,
|
||||
"PE#%d Restore special synchronization array",
|
||||
group->my_pe);
|
||||
for (i = 0; pSync && (i < _SHMEM_ALLTOALL_SYNC_SIZE); i++) {
|
||||
pSync[i] = _SHMEM_SYNC_VALUE;
|
||||
}
|
||||
if (!oshmem_proc_group_is_member(group)) {
|
||||
return OSHMEM_SUCCESS;
|
||||
}
|
||||
|
||||
return rc;
|
||||
}
|
||||
|
||||
static int _algorithm_simple(struct oshmem_group_t *group,
|
||||
void *target,
|
||||
const void *source,
|
||||
ptrdiff_t tst, ptrdiff_t sst,
|
||||
size_t nelems,
|
||||
size_t element_size,
|
||||
long *pSync)
|
||||
{
|
||||
int rc = OSHMEM_SUCCESS;
|
||||
int pe_cur;
|
||||
int i;
|
||||
int j;
|
||||
int k;
|
||||
|
||||
SCOLL_VERBOSE(14,
|
||||
"[#%d] send data to all PE in the group",
|
||||
group->my_pe);
|
||||
j = oshmem_proc_group_find_id(group, group->my_pe);
|
||||
for (i = 0; i < group->proc_count; i++) {
|
||||
/* index permutation for better distribution of traffic */
|
||||
k = (((j)+(i))%(group->proc_count));
|
||||
pe_cur = oshmem_proc_pe(group->proc_array[k]);
|
||||
rc = MCA_SPML_CALL(put(
|
||||
(void *)((char *)target + j * tst * nelems * element_size),
|
||||
nelems * element_size,
|
||||
(void *)((char *)source + i * sst * nelems * element_size),
|
||||
pe_cur));
|
||||
if (OSHMEM_SUCCESS != rc) {
|
||||
break;
|
||||
}
|
||||
if (!pSync) {
|
||||
SCOLL_ERROR("Incorrect argument pSync");
|
||||
return OSHMEM_ERR_BAD_PARAM;
|
||||
}
|
||||
|
||||
if ((sst == 1) && (dst == 1)) {
|
||||
rc = a2a_alg_simple(group, target, source, nelems, element_size);
|
||||
} else {
|
||||
rc = a2as_alg_simple(group, target, source, dst, sst, nelems,
|
||||
element_size);
|
||||
}
|
||||
|
||||
if (rc != OSHMEM_SUCCESS) {
|
||||
return rc;
|
||||
}
|
||||
|
||||
/* fence (which currently acts as quiet) is needed
|
||||
* because scoll level barrier does not guarantee put completion
|
||||
*/
|
||||
MCA_SPML_CALL(fence());
|
||||
|
||||
/* Wait for operation completion */
|
||||
if (rc == OSHMEM_SUCCESS) {
|
||||
SCOLL_VERBOSE(14, "[#%d] Wait for operation completion", group->my_pe);
|
||||
rc = BARRIER_FUNC(group,
|
||||
(pSync + 1),
|
||||
SCOLL_DEFAULT_ALG);
|
||||
SCOLL_VERBOSE(14, "[#%d] Wait for operation completion", group->my_pe);
|
||||
rc = BARRIER_FUNC(group, pSync + 1, SCOLL_DEFAULT_ALG);
|
||||
|
||||
/* Restore initial values */
|
||||
SCOLL_VERBOSE(12, "PE#%d Restore special synchronization array",
|
||||
group->my_pe);
|
||||
|
||||
for (i = 0; pSync && (i < _SHMEM_ALLTOALL_SYNC_SIZE); i++) {
|
||||
pSync[i] = _SHMEM_SYNC_VALUE;
|
||||
}
|
||||
|
||||
return rc;
|
||||
}
|
||||
|
||||
|
||||
static inline void *
|
||||
get_stride_elem(const void *base, ptrdiff_t sst, size_t nelems, size_t elem_size,
|
||||
int block_idx, int elem_idx)
|
||||
{
|
||||
/*
|
||||
* j th block starts at: nelems * element_size * sst * j
|
||||
* offset of the l th element in the block is: element_size * sst * l
|
||||
*/
|
||||
return (char *)base + elem_size * sst * (nelems * block_idx + elem_idx);
|
||||
}
|
||||
|
||||
static inline int
|
||||
get_dst_pe(struct oshmem_group_t *group, int src_blk_idx, int dst_blk_idx)
|
||||
{
|
||||
int dst_grp_pe;
|
||||
|
||||
/* index permutation for better distribution of traffic */
|
||||
dst_grp_pe = (dst_blk_idx + src_blk_idx) % group->proc_count;
|
||||
|
||||
/* convert to the global pe */
|
||||
return oshmem_proc_pe(group->proc_array[dst_grp_pe]);
|
||||
}
|
||||
|
||||
static int a2as_alg_simple(struct oshmem_group_t *group,
|
||||
void *target,
|
||||
const void *source,
|
||||
ptrdiff_t tst, ptrdiff_t sst,
|
||||
size_t nelems,
|
||||
size_t element_size)
|
||||
{
|
||||
int rc;
|
||||
int dst_pe;
|
||||
int src_blk_idx;
|
||||
int dst_blk_idx;
|
||||
size_t elem_idx;
|
||||
|
||||
SCOLL_VERBOSE(14,
|
||||
"[#%d] send data to all PE in the group",
|
||||
group->my_pe);
|
||||
|
||||
dst_blk_idx = oshmem_proc_group_find_id(group, group->my_pe);
|
||||
|
||||
for (src_blk_idx = 0; src_blk_idx < group->proc_count; src_blk_idx++) {
|
||||
|
||||
dst_pe = get_dst_pe(group, src_blk_idx, dst_blk_idx);
|
||||
for (elem_idx = 0; elem_idx < nelems; elem_idx++) {
|
||||
rc = MCA_SPML_CALL(put(
|
||||
get_stride_elem(target, tst, nelems, element_size,
|
||||
dst_blk_idx, elem_idx),
|
||||
element_size,
|
||||
get_stride_elem(source, sst, nelems, element_size,
|
||||
src_blk_idx, elem_idx),
|
||||
dst_pe));
|
||||
if (OSHMEM_SUCCESS != rc) {
|
||||
return rc;
|
||||
}
|
||||
}
|
||||
}
|
||||
return OSHMEM_SUCCESS;
|
||||
}
|
||||
|
||||
static int a2a_alg_simple(struct oshmem_group_t *group,
|
||||
void *target,
|
||||
const void *source,
|
||||
size_t nelems,
|
||||
size_t element_size)
|
||||
{
|
||||
int rc;
|
||||
int dst_pe;
|
||||
int src_blk_idx;
|
||||
int dst_blk_idx;
|
||||
void *dst_blk;
|
||||
|
||||
SCOLL_VERBOSE(14,
|
||||
"[#%d] send data to all PE in the group",
|
||||
group->my_pe);
|
||||
|
||||
dst_blk_idx = oshmem_proc_group_find_id(group, group->my_pe);
|
||||
|
||||
/* block start at stride 1 first elem */
|
||||
dst_blk = get_stride_elem(target, 1, nelems, element_size, dst_blk_idx, 0);
|
||||
|
||||
for (src_blk_idx = 0; src_blk_idx < group->proc_count; src_blk_idx++) {
|
||||
|
||||
dst_pe = get_dst_pe(group, src_blk_idx, dst_blk_idx);
|
||||
rc = MCA_SPML_CALL(put(dst_blk,
|
||||
nelems * element_size,
|
||||
get_stride_elem(source, 1, nelems,
|
||||
element_size, src_blk_idx, 0),
|
||||
dst_pe));
|
||||
if (OSHMEM_SUCCESS != rc) {
|
||||
return rc;
|
||||
}
|
||||
}
|
||||
return OSHMEM_SUCCESS;
|
||||
}
|
||||
|
||||
Загрузка…
Ссылка в новой задаче
Block a user