ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

oshmem: refactor group cache

Просмотр исходного кода 
- Use opal hash table instead of list for group lookup.
- Code cleanup/refactoring. Group cache is now a part
  of the proc_group.

Signed-off-by: Alex Mikheev <alexm@mellanox.com>
Этот коммит содержится в:
Alex Mikheev 2018-02-21 18:26:32 +02:00
родитель 7d0e02345b
Коммит 292d185c30
Не найден ключ, соответствующий данной подписи
Идентификатор ключа GPG: 0F964A1EF0579522
21 изменённых файлов: 631 добавлений и 1093 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

269
oshmem/proc/proc.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2014-2016 Research Organization for Information Science * Copyright (c) 2014-2016 Research Organization for Information Science
* and Technology (RIST). All rights reserved. * and Technology (RIST). All rights reserved.
@ -17,6 +17,7 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/runtime/runtime.h" #include "oshmem/runtime/runtime.h"
#include "oshmem/mca/scoll/base/base.h" #include "oshmem/mca/scoll/base/base.h"
#include "oshmem/proc/proc_group_cache.h"
#ifdef HAVE_STRINGS_H #ifdef HAVE_STRINGS_H
#include <strings.h> #include <strings.h>
@ -65,40 +66,67 @@ oshmem_group_t* oshmem_group_null = NULL;
OBJ_CLASS_INSTANCE(oshmem_group_t, opal_object_t, NULL, NULL); OBJ_CLASS_INSTANCE(oshmem_group_t, opal_object_t, NULL, NULL);
static void oshmem_proc_group_destroy_internal(oshmem_group_t* group,
int scoll_unselect);
int oshmem_proc_group_init(void) int oshmem_proc_group_init(void)
{ {
int rc;
rc = oshmem_group_cache_init();
if (OSHMEM_SUCCESS != rc) {
return rc;
}
/* Setup communicator array */ /* Setup communicator array */
OBJ_CONSTRUCT(&oshmem_group_array, opal_pointer_array_t); OBJ_CONSTRUCT(&oshmem_group_array, opal_pointer_array_t);
if (OPAL_SUCCESS
!= opal_pointer_array_init(&oshmem_group_array, rc = opal_pointer_array_init(&oshmem_group_array, 0,
0, ORTE_GLOBAL_ARRAY_MAX_SIZE, 1);
ORTE_GLOBAL_ARRAY_MAX_SIZE, if (OPAL_SUCCESS != rc) {
1)) { goto err1;
return OSHMEM_ERROR;
} }
/* Setup SHMEM_GROUP_ALL */ /* Setup SHMEM_GROUP_ALL */
if (NULL oshmem_group_all = oshmem_proc_group_create(0, 1, ompi_comm_size(oshmem_comm_world));
== (oshmem_group_all = if (NULL == oshmem_group_all) {
oshmem_proc_group_create(0, goto err2;
1,
ompi_comm_size(oshmem_comm_world)))) {
return OSHMEM_ERROR;
} }
/* Setup SHMEM_GROUP_SELF */ /* Setup SHMEM_GROUP_SELF */
if (NULL oshmem_group_self = oshmem_proc_group_create(oshmem_proc_pe(oshmem_proc_local()), 0, 1);
== (oshmem_group_self = oshmem_proc_group_create(oshmem_proc_pe(oshmem_proc_local()), if (NULL == oshmem_group_self) {
0, goto err3;
1))) {
oshmem_proc_group_destroy(oshmem_group_self);
return OSHMEM_ERROR;
} }
/* Setup SHMEM_GROUP_NULL */ /* Setup SHMEM_GROUP_NULL */
oshmem_group_null = NULL; oshmem_group_null = NULL;
return OSHMEM_SUCCESS; return OSHMEM_SUCCESS;
err3:
oshmem_proc_group_destroy_internal(oshmem_group_all, 1);
err2:
OBJ_DESTRUCT(&oshmem_group_array);
err1:
oshmem_group_cache_destroy();
return OSHMEM_ERROR;
}
void oshmem_proc_group_finalize_scoll(void)
{
int max, i;
oshmem_group_t *group;
/* Check whether we have some left */
max = opal_pointer_array_get_size(&oshmem_group_array);
for (i = 0; i < max; i++) {
group = (oshmem_group_t *) opal_pointer_array_get_item(&oshmem_group_array,
i);
if (NULL != group) {
mca_scoll_base_group_unselect(group);
}
}
} }
int oshmem_proc_group_finalize(void) int oshmem_proc_group_finalize(void)
@ -114,18 +142,17 @@ int oshmem_proc_group_finalize(void)
i); i);
if (NULL != group) { if (NULL != group) {
/* Group has not been freed before finalize */ /* Group has not been freed before finalize */
oshmem_proc_group_destroy(group); oshmem_proc_group_destroy_internal(group, 0);
} }
} }
OBJ_DESTRUCT(&oshmem_group_array); OBJ_DESTRUCT(&oshmem_group_array);
oshmem_group_cache_destroy();
return OSHMEM_SUCCESS; return OSHMEM_SUCCESS;
} }
oshmem_group_t* oshmem_proc_group_create(int pe_start, oshmem_group_t* oshmem_proc_group_create(int pe_start, int pe_stride, int pe_size)
int pe_stride,
size_t pe_size)
{ {
int cur_pe, count_pe; int cur_pe, count_pe;
int i; int i;
@ -135,107 +162,133 @@ oshmem_group_t* oshmem_proc_group_create(int pe_start,
assert(oshmem_proc_local()); assert(oshmem_proc_local());
group = oshmem_group_cache_find(pe_start, pe_stride, pe_size);
if (NULL != group) {
return group;
}
group = OBJ_NEW(oshmem_group_t); group = OBJ_NEW(oshmem_group_t);
if (NULL == group) {
return NULL;
}
if (group) { cur_pe = 0;
cur_pe = 0; count_pe = 0;
count_pe = 0;
OPAL_THREAD_LOCK(&oshmem_proc_lock); OPAL_THREAD_LOCK(&oshmem_proc_lock);
/* allocate an array */ /* allocate an array */
proc_array = (ompi_proc_t**) malloc(pe_size * sizeof(ompi_proc_t*)); proc_array = (ompi_proc_t**) malloc(pe_size * sizeof(ompi_proc_t*));
if (NULL == proc_array) { if (NULL == proc_array) {
OBJ_RELEASE(group); OBJ_RELEASE(group);
OPAL_THREAD_UNLOCK(&oshmem_proc_lock); OPAL_THREAD_UNLOCK(&oshmem_proc_lock);
return NULL ; return NULL ;
} }
group->my_pe = oshmem_proc_pe(oshmem_proc_local()); group->my_pe = oshmem_proc_pe(oshmem_proc_local());
group->is_member = 0; group->is_member = 0;
for (i = 0 ; i < ompi_comm_size(oshmem_comm_world) ; i++) { for (i = 0 ; i < ompi_comm_size(oshmem_comm_world) ; i++) {
proc = oshmem_proc_find(i); proc = oshmem_proc_find(i);
if (NULL == proc) { if (NULL == proc) {
opal_output(0,
"Error: Can not find proc object for pe = %d", i);
free(proc_array);
OBJ_RELEASE(group);
OPAL_THREAD_UNLOCK(&oshmem_proc_lock);
return NULL;
}
if (count_pe >= (int) pe_size) {
break;
} else if ((cur_pe >= pe_start)
&& ((pe_stride == 0)
|| (((cur_pe - pe_start) % pe_stride) == 0))) {
proc_array[count_pe++] = proc;
if (oshmem_proc_pe(proc) == group->my_pe)
group->is_member = 1;
}
cur_pe++;
}
group->proc_array = proc_array;
group->proc_count = (int) count_pe;
group->ompi_comm = NULL;
/* Prepare peers list */
OBJ_CONSTRUCT(&(group->peer_list), opal_list_t);
{
orte_namelist_t *peer = NULL;
for (i = 0; i < group->proc_count; i++) {
peer = OBJ_NEW(orte_namelist_t);
peer->name.jobid = OSHMEM_PROC_JOBID(group->proc_array[i]);
peer->name.vpid = OSHMEM_PROC_VPID(group->proc_array[i]);
opal_list_append(&(group->peer_list), &peer->super);
}
}
group->id = opal_pointer_array_add(&oshmem_group_array, group);
memset(&group->g_scoll, 0, sizeof(mca_scoll_base_group_scoll_t));
if (OSHMEM_SUCCESS != mca_scoll_base_select(group)) {
opal_output(0, opal_output(0,
"Error: No collective modules are available: group is not created, returning NULL"); "Error: Can not find proc object for pe = %d", i);
oshmem_proc_group_destroy(group); free(proc_array);
OBJ_RELEASE(group);
OPAL_THREAD_UNLOCK(&oshmem_proc_lock); OPAL_THREAD_UNLOCK(&oshmem_proc_lock);
return NULL; return NULL;
} }
if (count_pe >= (int) pe_size) {
break;
} else if ((cur_pe >= pe_start)
&& ((pe_stride == 0)
|| (((cur_pe - pe_start) % pe_stride) == 0))) {
proc_array[count_pe++] = proc;
if (oshmem_proc_pe(proc) == group->my_pe)
group->is_member = 1;
}
cur_pe++;
}
group->proc_array = proc_array;
group->proc_count = (int) count_pe;
group->ompi_comm = NULL;
/* Prepare peers list */
OBJ_CONSTRUCT(&(group->peer_list), opal_list_t);
{
orte_namelist_t *peer = NULL;
for (i = 0; i < group->proc_count; i++) {
peer = OBJ_NEW(orte_namelist_t);
peer->name.jobid = OSHMEM_PROC_JOBID(group->proc_array[i]);
peer->name.vpid = OSHMEM_PROC_VPID(group->proc_array[i]);
opal_list_append(&(group->peer_list), &peer->super);
}
}
group->id = opal_pointer_array_add(&oshmem_group_array, group);
memset(&group->g_scoll, 0, sizeof(mca_scoll_base_group_scoll_t));
if (OSHMEM_SUCCESS != mca_scoll_base_select(group)) {
opal_output(0,
"Error: No collective modules are available: group is not created, returning NULL");
oshmem_proc_group_destroy_internal(group, 0);
OPAL_THREAD_UNLOCK(&oshmem_proc_lock); OPAL_THREAD_UNLOCK(&oshmem_proc_lock);
return NULL;
} }
if (OSHMEM_SUCCESS != oshmem_group_cache_insert(group, pe_start,
pe_stride, pe_size)) {
oshmem_proc_group_destroy_internal(group, 1);
OPAL_THREAD_UNLOCK(&oshmem_proc_lock);
return NULL;
}
OPAL_THREAD_UNLOCK(&oshmem_proc_lock);
return group; return group;
} }
static void
oshmem_proc_group_destroy_internal(oshmem_group_t* group, int scoll_unselect)
{
if (NULL == group) {
return;
}
if (scoll_unselect) {
mca_scoll_base_group_unselect(group);
}
/* Destroy proc array */
if (group->proc_array) {
free(group->proc_array);
}
/* Destroy peer list */
{
opal_list_item_t *item;
while (NULL != (item = opal_list_remove_first(&(group->peer_list)))) {
/* destruct the item (we constructed it), then free the memory chunk */
OBJ_RELEASE(item);
}
OBJ_DESTRUCT(&(group->peer_list));
}
/* reset the oshmem_group_array entry - make sure that the
* entry is in the table */
if (NULL
!= opal_pointer_array_get_item(&oshmem_group_array,
group->id)) {
opal_pointer_array_set_item(&oshmem_group_array, group->id, NULL );
}
OBJ_RELEASE(group);
}
void oshmem_proc_group_destroy(oshmem_group_t* group) void oshmem_proc_group_destroy(oshmem_group_t* group)
{ {
if (group) { if (oshmem_group_cache_enabled()) {
mca_scoll_base_group_unselect(group); return;
/* Destroy proc array */
if (group->proc_array) {
free(group->proc_array);
}
/* Destroy peer list */
{
opal_list_item_t *item;
while (NULL != (item = opal_list_remove_first(&(group->peer_list)))) {
/* destruct the item (we constructed it), then free the memory chunk */
OBJ_RELEASE(item);
}
OBJ_DESTRUCT(&(group->peer_list));
}
/* reset the oshmem_group_array entry - make sure that the
* entry is in the table */
if (NULL
!= opal_pointer_array_get_item(&oshmem_group_array,
group->id)) {
opal_pointer_array_set_item(&oshmem_group_array, group->id, NULL );
}
OBJ_RELEASE(group);
} }
oshmem_proc_group_destroy_internal(group, 1);
} }

46
oshmem/proc/proc.h
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2016 Research Organization for Information Science * Copyright (c) 2016 Research Organization for Information Science
* and Technology (RIST). All rights reserved. * and Technology (RIST). All rights reserved.
@ -17,8 +17,6 @@
#include "oshmem/types.h" #include "oshmem/types.h"
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h"
#include "opal/class/opal_list.h" #include "opal/class/opal_list.h"
#include "opal/util/proc.h" #include "opal/util/proc.h"
#include "opal/dss/dss_types.h" #include "opal/dss/dss_types.h"
@ -30,6 +28,10 @@
#include "ompi/proc/proc.h" #include "ompi/proc/proc.h"
#include "ompi/communicator/communicator.h" #include "ompi/communicator/communicator.h"
#include "oshmem/mca/scoll/scoll.h"
#include "oshmem/runtime/runtime.h"
#include "oshmem/shmem/shmem_api_logger.h"
BEGIN_C_DECLS BEGIN_C_DECLS
/* ******************************************************************** */ /* ******************************************************************** */
@ -177,19 +179,17 @@ OSHMEM_DECLSPEC int oshmem_proc_group_init(void);
/** /**
* Finalize the OSHMEM process predefined groups * Finalize the OSHMEM process predefined groups
* *
* Initialize the Open SHMEM process predefined groups. This function will
* query the run-time environment and build a list of the proc
* instances in the current pe set. The local information not
* easily determined by the run-time ahead of time (architecture and
* hostname) will be published during this call.
*
* @note This is primarily used once during SHMEM setup.
*
* @retval OSHMEM_SUCESS System successfully initialized * @retval OSHMEM_SUCESS System successfully initialized
* @retval OSHMEM_ERROR Initialization failed due to unspecified error * @retval OSHMEM_ERROR Initialization failed due to unspecified error
*/ */
OSHMEM_DECLSPEC int oshmem_proc_group_finalize(void); OSHMEM_DECLSPEC int oshmem_proc_group_finalize(void);
/**
* Release collectives used by the groups. The function
* must be called prior to the oshmem_proc_group_finalize()
*/
OSHMEM_DECLSPEC void oshmem_proc_group_finalize_scoll(void);
/** /**
* Create processes group. * Create processes group.
* *
@ -205,7 +205,29 @@ OSHMEM_DECLSPEC int oshmem_proc_group_finalize(void);
*/ */
OSHMEM_DECLSPEC oshmem_group_t *oshmem_proc_group_create(int pe_start, OSHMEM_DECLSPEC oshmem_group_t *oshmem_proc_group_create(int pe_start,
int pe_stride, int pe_stride,
size_t pe_size); int pe_size);
/**
* same as above but abort on failure
*/
static inline oshmem_group_t *
oshmem_proc_group_create_nofail(int pe_start, int pe_stride, int pe_size)
{
oshmem_group_t *group;
group = oshmem_proc_group_create(pe_start, pe_stride, pe_size);
if (NULL == group) {
goto fatal;
}
return group;
fatal:
SHMEM_API_ERROR("Failed to create group (%d,%d,%d)",
pe_start, pe_stride, pe_size);
oshmem_shmem_abort(-1);
return NULL;
}
/** /**
* Destroy processes group. * Destroy processes group.

145
oshmem/proc/proc_group_cache.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2015 Research Organization for Information Science * Copyright (c) 2015 Research Organization for Information Science
* and Technology (RIST). All rights reserved. * and Technology (RIST). All rights reserved.
@ -13,104 +13,73 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/runtime/runtime.h" #include "oshmem/runtime/runtime.h"
OBJ_CLASS_INSTANCE(oshmem_group_cache_t, opal_object_t, NULL, NULL); #define OSHMEM_GROUP_CACHE_SIZE 1024
opal_list_t oshmem_group_cache_list = {{0}};
unsigned int oshmem_group_cache_size = 0;
oshmem_group_t* find_group_in_cache(int PE_start, int logPE_stride, int PE_size)
{
int cache_look_up_id[3] = { PE_start, logPE_stride, PE_size };
opal_list_item_t *item;
if (opal_list_is_empty(&oshmem_group_cache_list)) {
return NULL ;
}
for (item = opal_list_get_first(&oshmem_group_cache_list); static opal_hash_table_t group_cache;
item && (item != opal_list_get_end(&oshmem_group_cache_list));
item = opal_list_get_next(item)) {
if (!memcmp(((oshmem_group_cache_t *) item)->cache_id,
cache_look_up_id,
3 * sizeof(int))) {
return ((oshmem_group_cache_t *) item)->group;
}
}
return NULL ;
}
int cache_group(oshmem_group_t *group, typedef struct {
int PE_start, int pe_start;
int logPE_stride, int pe_size;
int PE_size) int pe_stride;
} oshmem_group_key_t;
static int group_cache_n_hits;
static int group_cache_n_lookups;
int oshmem_group_cache_init(void)
{ {
oshmem_group_cache_t *cached_group = NULL; OBJ_CONSTRUCT(&group_cache, opal_hash_table_t);
cached_group = OBJ_NEW(oshmem_group_cache_t); if (OPAL_SUCCESS != opal_hash_table_init(&group_cache, OSHMEM_GROUP_CACHE_SIZE)) {
#if OPAL_ENABLE_DEBUG return OSHMEM_ERROR;
cached_group->item.opal_list_item_belong_to = NULL;
cached_group->item.opal_list_item_refcount = 0;
#endif
cached_group->group = group;
cached_group->cache_id[0] = PE_start;
cached_group->cache_id[1] = logPE_stride;
cached_group->cache_id[2] = PE_size;
if (opal_list_get_size(&oshmem_group_cache_list)
< oshmem_group_cache_size) {
opal_list_append(&oshmem_group_cache_list,
(opal_list_item_t *)cached_group);
} else {
#if ABORT_ON_CACHE_OVERFLOW
opal_output(0,
"error: group cache overflow on rank %i: cache_size = %u: try increasing oshmem_proc_group_cache_size mca parameter",
group->my_pe,
oshmem_group_cache_size);
oshmem_shmem_abort(-1);
#else
/*This part of code makes FIFO group cache management. Define ABORT_ON_CACHE_OVERFLOW as 0 to enable this.*/
oshmem_group_cache_t *cached_group_to_remove = (oshmem_group_cache_t *)opal_list_remove_first(&oshmem_group_cache_list);
oshmem_proc_group_destroy(cached_group_to_remove->group);
OBJ_RELEASE(cached_group_to_remove);
opal_list_append(&oshmem_group_cache_list,(opal_list_item_t *)cached_group);
#endif
} }
return OSHMEM_SUCCESS; return OSHMEM_SUCCESS;
} }
int oshmem_group_cache_list_init(void) void oshmem_group_cache_destroy(void)
{ {
int mca_value; OBJ_DESTRUCT(&group_cache);
int cache_size_default = 1000;
OBJ_CONSTRUCT(&oshmem_group_cache_list, opal_list_t);
mca_value = cache_size_default;
(void) mca_base_var_register("oshmem",
"proc",
NULL,
"group_cache_size",
"The depth of the oshmem_group cache list used to speed up collective operations",
MCA_BASE_VAR_TYPE_INT,
NULL,
0,
MCA_BASE_VAR_FLAG_SETTABLE,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&mca_value);
if (mca_value < 0) {
opal_output(0,
"error: oshmem_proc_group_cache_size mca parameter was set to %i while it has to be positive value. Default value %i will be used.",
mca_value,
cache_size_default);
mca_value = cache_size_default;
}
oshmem_group_cache_size = (unsigned int) mca_value;
return OSHMEM_SUCCESS;
} }
int oshmem_group_cache_list_free(void) oshmem_group_t *oshmem_group_cache_find(int pe_start, int pe_stride, int pe_size)
{ {
oshmem_group_cache_t *cached_group = NULL; oshmem_group_key_t key;
opal_list_item_t *item; oshmem_group_t *group;
while (NULL != (item = opal_list_remove_first(&oshmem_group_cache_list))) {
cached_group = (oshmem_group_cache_t *) item; if (!oshmem_group_cache_enabled()) {
oshmem_proc_group_destroy(cached_group->group); return NULL;
OBJ_RELEASE(cached_group); }
key.pe_start = pe_start;
key.pe_size = pe_size;
key.pe_stride = pe_stride;
group_cache_n_lookups++;
if (OPAL_SUCCESS != opal_hash_table_get_value_ptr(&group_cache, &key,
sizeof(key), (void **)&group)) {
return NULL;
}
group_cache_n_hits++;
return group;
}
int oshmem_group_cache_insert(oshmem_group_t *group, int pe_start,
int pe_stride, int pe_size)
{
oshmem_group_key_t key;
if (!oshmem_group_cache_enabled()) {
return OSHMEM_SUCCESS;
}
key.pe_start = pe_start;
key.pe_size = pe_size;
key.pe_stride = pe_stride;
if (OPAL_SUCCESS != opal_hash_table_set_value_ptr(&group_cache, &key,
sizeof(key), group)) {
return OSHMEM_ERROR;
} }
return OSHMEM_SUCCESS; return OSHMEM_SUCCESS;
} }

42
oshmem/proc/proc_group_cache.h
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
* *
@ -13,29 +13,33 @@
#include "oshmem_config.h" #include "oshmem_config.h"
#include "proc.h" #include "proc.h"
#define OSHMEM_GROUP_CACHE_ENABLED 1 #define OSHMEM_GROUP_CACHE_ENABLED 1
#define ABORT_ON_CACHE_OVERFLOW 1
BEGIN_C_DECLS BEGIN_C_DECLS
struct oshmem_group_cache_t {
opal_list_item_t item;
oshmem_group_t *group;
int cache_id[3];
};
typedef struct oshmem_group_cache_t oshmem_group_cache_t; /**
OSHMEM_DECLSPEC OBJ_CLASS_DECLARATION(oshmem_group_cache_t); * A group cache.
OSHMEM_DECLSPEC extern opal_list_t oshmem_group_cache_list; *
* Deletion of a group is not implemented because it
* requires a synchronization between PEs
*
* If cache enabled every group is kept until the
* shmem_finalize() is called
*/
oshmem_group_t* find_group_in_cache(int PE_start, int logPE_stride, int PE_size); int oshmem_group_cache_init(void);
void oshmem_group_cache_destroy(void);
int cache_group(oshmem_group_t *group, oshmem_group_t* oshmem_group_cache_find(int pe_start, int pe_stride, int pe_size);
int PE_start,
int logPE_stride, int oshmem_group_cache_insert(oshmem_group_t *group, int pe_start,
int PE_size); int pe_stride, int pe_size);
int oshmem_group_cache_list_init(void);
int oshmem_group_cache_list_free(void); static inline int oshmem_group_cache_enabled(void)
{
return OSHMEM_GROUP_CACHE_ENABLED;
}
extern unsigned int oshmem_group_cache_size;
END_C_DECLS END_C_DECLS
#endif #endif

13
oshmem/runtime/oshmem_shmem_finalize.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2015 Research Organization for Information Science * Copyright (c) 2015 Research Organization for Information Science
* and Technology (RIST). All rights reserved. * and Technology (RIST). All rights reserved.
@ -102,17 +102,10 @@ static int _shmem_finalize(void)
if (OSHMEM_SUCCESS != (ret = oshmem_request_finalize())) { if (OSHMEM_SUCCESS != (ret = oshmem_request_finalize())) {
return ret; return ret;
} }
/* must free cached groups before we kill collectives */
if (OSHMEM_SUCCESS != (ret = oshmem_group_cache_list_free())) { oshmem_proc_group_finalize_scoll();
return ret;
}
/* We need to call mca_scoll_base_group_unselect explicitly for each group
* that are not freed by oshmem_group_cache_list_free. We can only release its collectives at this point */
mca_scoll_base_group_unselect(oshmem_group_all);
mca_scoll_base_group_unselect(oshmem_group_self);
/* Close down MCA modules */ /* Close down MCA modules */
if (OSHMEM_SUCCESS != (ret = mca_base_framework_close(&oshmem_atomic_base_framework) ) ) { if (OSHMEM_SUCCESS != (ret = mca_base_framework_close(&oshmem_atomic_base_framework) ) ) {
return ret; return ret;
} }

7
oshmem/runtime/oshmem_shmem_init.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2015-2016 Research Organization for Information Science * Copyright (c) 2015-2016 Research Organization for Information Science
* and Technology (RIST). All rights reserved. * and Technology (RIST). All rights reserved.
@ -259,11 +259,6 @@ static int _shmem_init(int argc, char **argv, int requested, int *provided)
goto error; goto error;
} }
if (OSHMEM_SUCCESS != (ret = oshmem_group_cache_list_init())) {
error = "oshmem_group_cache_list_init() failed";
goto error;
}
if (OSHMEM_SUCCESS != (ret = oshmem_op_init())) { if (OSHMEM_SUCCESS != (ret = oshmem_op_init())) {
error = "oshmem_op_init() failed"; error = "oshmem_op_init() failed";
goto error; goto error;

60
oshmem/shmem/c/shmem_alltoall.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2016 Mellanox Technologies, Inc. * Copyright (c) 2016-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
* *
@ -19,7 +19,6 @@
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
static void _shmem_alltoall(void *target, static void _shmem_alltoall(void *target,
const void *source, const void *source,
@ -78,48 +77,23 @@ static void _shmem_alltoall(void *target,
int PE_size, int PE_size,
long *pSync) long *pSync)
{ {
int rc = OSHMEM_SUCCESS; int rc;
oshmem_group_t* group = NULL; oshmem_group_t* group;
if ((0 <= PE_start) && (0 <= logPE_stride)) { /* Create group basing PE_start, logPE_stride and PE_size */
/* Create group basing PE_start, logPE_stride and PE_size */ group = oshmem_proc_group_create_nofail(PE_start, 1<<logPE_stride, PE_size);
#if OSHMEM_GROUP_CACHE_ENABLED == 0 /* Call collective alltoall operation */
group = oshmem_proc_group_create(PE_start, (1 << logPE_stride), PE_size); rc = group->g_scoll.scoll_alltoall(group,
if (!group) target,
rc = OSHMEM_ERROR; source,
#else dst,
group = find_group_in_cache(PE_start, logPE_stride, PE_size); sst,
if (!group) { nelems,
group = oshmem_proc_group_create(PE_start, element_size,
(1 << logPE_stride), pSync,
PE_size); SCOLL_DEFAULT_ALG);
if (!group) { oshmem_proc_group_destroy(group);
rc = OSHMEM_ERROR; RUNTIME_CHECK_RC(rc);
} else {
cache_group(group, PE_start, logPE_stride, PE_size);
}
}
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
/* Collective operation call */
if (rc == OSHMEM_SUCCESS) {
/* Call collective alltoall operation */
rc = group->g_scoll.scoll_alltoall(group,
target,
source,
dst,
sst,
nelems,
element_size,
pSync,
SCOLL_DEFAULT_ALG);
}
#if OSHMEM_GROUP_CACHE_ENABLED == 0
if ( rc == OSHMEM_SUCCESS ) {
oshmem_proc_group_destroy(group);
}
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
}
} }
#if OSHMEM_PROFILING #if OSHMEM_PROFILING

43
oshmem/shmem/c/shmem_barrier.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
* *
@ -18,7 +18,6 @@
#include "oshmem/mca/scoll/base/base.h" #include "oshmem/mca/scoll/base/base.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -30,8 +29,8 @@
void shmem_barrier(int PE_start, int logPE_stride, int PE_size, long *pSync) void shmem_barrier(int PE_start, int logPE_stride, int PE_size, long *pSync)
{ {
int rc = OSHMEM_SUCCESS; int rc;
oshmem_group_t* group = NULL; oshmem_group_t* group;
RUNTIME_CHECK_INIT(); RUNTIME_CHECK_INIT();
@ -40,38 +39,12 @@ void shmem_barrier(int PE_start, int logPE_stride, int PE_size, long *pSync)
shmem_fence(); shmem_fence();
#endif #endif
if ((0 <= PE_start) && (0 <= logPE_stride)) { /* Create group basing PE_start, logPE_stride and PE_size */
/* Create group basing PE_start, logPE_stride and PE_size */ group = oshmem_proc_group_create_nofail(PE_start, 1<<logPE_stride, PE_size);
#if OSHMEM_GROUP_CACHE_ENABLED == 0 /* Call barrier operation */
group = oshmem_proc_group_create(PE_start, (1 << logPE_stride), PE_size); rc = group->g_scoll.scoll_barrier(group, pSync, SCOLL_DEFAULT_ALG);
if (!group)
rc = OSHMEM_ERROR;
#else
group = find_group_in_cache(PE_start, logPE_stride, PE_size);
if (!group) {
group = oshmem_proc_group_create(PE_start,
(1 << logPE_stride),
PE_size);
if (!group) {
rc = OSHMEM_ERROR;
} else {
cache_group(group, PE_start, logPE_stride, PE_size);
}
}
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
/* Collective operation call */
if (rc == OSHMEM_SUCCESS) {
/* Call barrier operation */
rc = group->g_scoll.scoll_barrier(group, pSync, SCOLL_DEFAULT_ALG);
}
#if OSHMEM_GROUP_CACHE_ENABLED == 0 oshmem_proc_group_destroy(group);
if ( rc == OSHMEM_SUCCESS )
{
oshmem_proc_group_destroy(group);
}
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
}
RUNTIME_CHECK_RC(rc); RUNTIME_CHECK_RC(rc);
} }

59
oshmem/shmem/c/shmem_broadcast.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013-2017 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
* *
@ -19,7 +19,6 @@
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
static void _shmem_broadcast(void *target, static void _shmem_broadcast(void *target,
const void *source, const void *source,
@ -58,51 +57,31 @@ static void _shmem_broadcast(void *target,
int PE_size, int PE_size,
long *pSync) long *pSync)
{ {
int rc = OSHMEM_SUCCESS; int rc;
oshmem_group_t* group = NULL; oshmem_group_t *group;
if ((0 <= PE_root) && (PE_root < PE_size)) { if ((0 <= PE_root) && (PE_root < PE_size)) {
/* Create group basing PE_start, logPE_stride and PE_size */ /* Create group basing PE_start, logPE_stride and PE_size */
#if OSHMEM_GROUP_CACHE_ENABLED == 0 group = oshmem_proc_group_create_nofail(PE_start, 1 << logPE_stride, PE_size);
group = oshmem_proc_group_create(PE_start, (1 << logPE_stride), PE_size); if (PE_root >= group->proc_count) {
if (!group || (PE_root >= group->proc_count))
{
rc = OSHMEM_ERROR; rc = OSHMEM_ERROR;
goto out;
} }
#else
group = find_group_in_cache(PE_start, logPE_stride, PE_size);
if (!group) {
group = oshmem_proc_group_create(PE_start,
(1 << logPE_stride),
PE_size);
if (!group || (PE_root >= group->proc_count)) {
rc = OSHMEM_ERROR;
} else {
cache_group(group, PE_start, logPE_stride, PE_size);
}
}
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
/* Collective operation call */ /* Define actual PE using relative in active set */
if (rc == OSHMEM_SUCCESS) { PE_root = oshmem_proc_pe(group->proc_array[PE_root]);
/* Define actual PE using relative in active set */
PE_root = oshmem_proc_pe(group->proc_array[PE_root]);
/* Call collective broadcast operation */ /* Call collective broadcast operation */
rc = group->g_scoll.scoll_broadcast(group, rc = group->g_scoll.scoll_broadcast(group,
PE_root, PE_root,
target, target,
source, source,
nbytes, nbytes,
pSync, pSync,
SCOLL_DEFAULT_ALG); SCOLL_DEFAULT_ALG);
} out:
#if OSHMEM_GROUP_CACHE_ENABLED == 0 oshmem_proc_group_destroy(group);
if ( rc == OSHMEM_SUCCESS ) RUNTIME_CHECK_RC(rc);
{
oshmem_proc_group_destroy(group);
}
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
} }
} }

57
oshmem/shmem/c/shmem_collect.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
* *
@ -19,7 +19,6 @@
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
static void _shmem_collect(void *target, static void _shmem_collect(void *target,
const void *source, const void *source,
@ -58,47 +57,21 @@ static void _shmem_collect(void *target,
long *pSync, long *pSync,
bool array_type) bool array_type)
{ {
int rc = OSHMEM_SUCCESS; int rc;
oshmem_group_t* group = NULL; oshmem_group_t *group;
{ /* Create group basing PE_start, logPE_stride and PE_size */
/* Create group basing PE_start, logPE_stride and PE_size */ group = oshmem_proc_group_create_nofail(PE_start, 1<<logPE_stride, PE_size);
#if OSHMEM_GROUP_CACHE_ENABLED == 0 /* Call collective broadcast operation */
group = oshmem_proc_group_create(PE_start, (1 << logPE_stride), PE_size); rc = group->g_scoll.scoll_collect(group,
if (!group) target,
rc = OSHMEM_ERROR; source,
#else nbytes,
group = find_group_in_cache(PE_start, logPE_stride, PE_size); pSync,
if (!group) { array_type,
group = oshmem_proc_group_create(PE_start, SCOLL_DEFAULT_ALG);
(1 << logPE_stride), oshmem_proc_group_destroy(group);
PE_size); RUNTIME_CHECK_RC(rc);
if (!group) {
rc = OSHMEM_ERROR;
} else {
cache_group(group, PE_start, logPE_stride, PE_size);
}
}
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
/* Collective operation call */
if (rc == OSHMEM_SUCCESS) {
/* Call collective broadcast operation */
rc = group->g_scoll.scoll_collect(group,
target,
source,
nbytes,
pSync,
array_type,
SCOLL_DEFAULT_ALG);
}
#if OSHMEM_GROUP_CACHE_ENABLED == 0
if ( rc == OSHMEM_SUCCESS )
{
oshmem_proc_group_destroy(group);
}
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
}
} }
#if OSHMEM_PROFILING #if OSHMEM_PROFILING

60
oshmem/shmem/c/shmem_reduce.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
* *
@ -16,15 +16,8 @@
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_GROUP_CACHE_ENABLED == 0
static bool __group_cache_enabled = false;
#else
static bool __group_cache_enabled = true;
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
/* /*
* The shared memory (SHMEM) reduction routines perform an associative binary operation * The shared memory (SHMEM) reduction routines perform an associative binary operation
* across symmetric arrays on multiple virtual PEs. * across symmetric arrays on multiple virtual PEs.
@ -50,47 +43,22 @@ static bool __group_cache_enabled = true;
RUNTIME_CHECK_ADDR(source); \ RUNTIME_CHECK_ADDR(source); \
\ \
{ \ { \
/* Create group basing PE_start, logPE_stride and PE_size */ \ group = oshmem_proc_group_create_nofail(PE_start, 1<<logPE_stride, PE_size); \
if (!__group_cache_enabled) \
{ \
group = oshmem_proc_group_create(PE_start, (1 << logPE_stride), PE_size); \
if (!group) \
rc = OSHMEM_ERROR; \
} \
else \
{ \
group = find_group_in_cache(PE_start,logPE_stride,PE_size); \
if (!group) \
{ \
group = oshmem_proc_group_create(PE_start, (1 << logPE_stride), PE_size); \
if (!group) \
rc = OSHMEM_ERROR; \
cache_group(group,PE_start,logPE_stride,PE_size); \
} \
} \
\
/* Collective operation call */ \
if ( rc == OSHMEM_SUCCESS ) \
{ \
oshmem_op_t* op = oshmem_op_##name##type_name; \ oshmem_op_t* op = oshmem_op_##name##type_name; \
size_t size = nreduce * op->dt_size; \ size_t size = nreduce * op->dt_size; \
\ \
/* Call collective reduce operation */ \ /* Call collective reduce operation */ \
rc = group->g_scoll.scoll_reduce( \ rc = group->g_scoll.scoll_reduce( \
group, \ group, \
op, \ op, \
(void*)target, \ (void*)target, \
(const void*)source, \ (const void*)source, \
size, \ size, \
pSync, \ pSync, \
(void*)pWrk, \ (void*)pWrk, \
SCOLL_DEFAULT_ALG ); \ SCOLL_DEFAULT_ALG ); \
} \
\ \
if ( !__group_cache_enabled && (rc == OSHMEM_SUCCESS ) ) \ oshmem_proc_group_destroy(group); \
{ \
oshmem_proc_group_destroy(group); \
} \
} \ } \
RUNTIME_CHECK_RC(rc); \ RUNTIME_CHECK_RC(rc); \
} }

157
oshmem/shmem/fortran/shmem_alltoall_f.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013-2016 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
@ -15,7 +15,6 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -59,7 +58,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
(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), (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),
(target, source, dst, sst, nlong, PE_start, logPE_stride, PE_size, pSync)) (target, source, dst, sst, nlong, PE_start, logPE_stride, PE_size, pSync))
#define SHMEM_ALLTOALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_ALLTOALL(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *nlong,\ MPI_Fint *nlong,\
MPI_Fint *PE_start, \ MPI_Fint *PE_start, \
@ -67,61 +66,28 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
MPI_Fint *PE_size, \ MPI_Fint *PE_size, \
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
{\ /* Create group basing PE_start, logPE_stride and PE_size */\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start), \
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\ (1 << OMPI_FINT_2_INT(*logPE_stride)), \
{\ OMPI_FINT_2_INT(*PE_size));\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \ oshmem_op_t* op = T_NAME;\
(1 << OMPI_FINT_2_INT(*logPE_stride)), \ \
OMPI_FINT_2_INT(*PE_size));\ /* Call collective broadcast operation */\
if (!group)\ rc = group->g_scoll.scoll_alltoall( group, \
rc = OSHMEM_ERROR;\ FPTR_2_VOID_PTR(target), \
}\ FPTR_2_VOID_PTR(source), \
else\ 1, \
{\ 1, \
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ OMPI_FINT_2_INT(*nlong), \
OMPI_FINT_2_INT(*logPE_stride),\ op->dt_size, \
OMPI_FINT_2_INT(*PE_size));\ FPTR_2_VOID_PTR(pSync), SCOLL_DEFAULT_ALG );\
if (!group)\ oshmem_proc_group_destroy(group);\
{\ RUNTIME_CHECK_RC(rc);\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \
(1 << OMPI_FINT_2_INT(*logPE_stride)), \
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
rc = OSHMEM_ERROR;\
cache_group(group,OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
} /* OSHMEM_GROUP_CACHE_ENABLED */\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_op_t* op = T_NAME;\
\
/* Call collective broadcast operation */\
rc = group->g_scoll.scoll_alltoall( group, \
FPTR_2_VOID_PTR(target), \
FPTR_2_VOID_PTR(source), \
1, \
1, \
OMPI_FINT_2_INT(*nlong), \
op->dt_size, \
FPTR_2_VOID_PTR(pSync), SCOLL_DEFAULT_ALG );\
}\
if (OSHMEM_GROUP_CACHE_ENABLED == 0) \
{\
if ( group )\
{\
oshmem_proc_group_destroy(group);\
}\
} /* OSHMEM_GROUP_CACHE_ENABLED */\
}\
} }
#define SHMEM_ALLTOALLS(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_ALLTOALLS(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *dst,\ MPI_Fint *dst,\
MPI_Fint *sst,\ MPI_Fint *sst,\
@ -131,61 +97,28 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
MPI_Fint *PE_size, \ MPI_Fint *PE_size, \
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
{\ /* Create group basing PE_start, logPE_stride and PE_size */\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start), \
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\ (1 << OMPI_FINT_2_INT(*logPE_stride)), \
{\ OMPI_FINT_2_INT(*PE_size));\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \ oshmem_op_t* op = T_NAME;\
(1 << OMPI_FINT_2_INT(*logPE_stride)), \ \
OMPI_FINT_2_INT(*PE_size));\ /* Call collective broadcast operation */\
if (!group)\ rc = group->g_scoll.scoll_alltoall( group, \
rc = OSHMEM_ERROR;\ FPTR_2_VOID_PTR(target), \
}\ FPTR_2_VOID_PTR(source), \
else\ OMPI_FINT_2_INT(*dst), \
{\ OMPI_FINT_2_INT(*sst), \
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ OMPI_FINT_2_INT(*nlong), \
OMPI_FINT_2_INT(*logPE_stride),\ op->dt_size, \
OMPI_FINT_2_INT(*PE_size));\ FPTR_2_VOID_PTR(pSync), SCOLL_DEFAULT_ALG );\
if (!group)\ oshmem_proc_group_destroy(group);\
{\ RUNTIME_CHECK_RC(rc);\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \
(1 << OMPI_FINT_2_INT(*logPE_stride)), \
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
rc = OSHMEM_ERROR;\
cache_group(group,OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
} /* OSHMEM_GROUP_CACHE_ENABLED */\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_op_t* op = T_NAME;\
\
/* Call collective broadcast operation */\
rc = group->g_scoll.scoll_alltoall( group, \
FPTR_2_VOID_PTR(target), \
FPTR_2_VOID_PTR(source), \
OMPI_FINT_2_INT(*dst), \
OMPI_FINT_2_INT(*sst), \
OMPI_FINT_2_INT(*nlong), \
op->dt_size, \
FPTR_2_VOID_PTR(pSync), SCOLL_DEFAULT_ALG );\
}\
if (OSHMEM_GROUP_CACHE_ENABLED == 0) \
{\
if ( group )\
{\
oshmem_proc_group_destroy(group);\
}\
} /* OSHMEM_GROUP_CACHE_ENABLED */\
}\
} }
SHMEM_ALLTOALL(shmem_alltoall32_f, oshmem_op_prod_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_ALLTOALL(shmem_alltoall32_f, oshmem_op_prod_fint4)
SHMEM_ALLTOALL(shmem_alltoall64_f, oshmem_op_prod_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_ALLTOALL(shmem_alltoall64_f, oshmem_op_prod_fint8)
SHMEM_ALLTOALLS(shmem_alltoalls32_f, oshmem_op_prod_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_ALLTOALLS(shmem_alltoalls32_f, oshmem_op_prod_fint4)
SHMEM_ALLTOALLS(shmem_alltoalls64_f, oshmem_op_prod_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_ALLTOALLS(shmem_alltoalls64_f, oshmem_op_prod_fint8)

83
oshmem/shmem/fortran/shmem_and_to_all_f.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
@ -15,7 +15,6 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -50,7 +49,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync),
(target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) ) (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
#define SHMEM_AND_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_AND_TO_ALL(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *nreduce, \ MPI_Fint *nreduce, \
MPI_Fint *PE_start, \ MPI_Fint *PE_start, \
@ -59,61 +58,27 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
FORTRAN_POINTER_T *pWrk, \ FORTRAN_POINTER_T *pWrk, \
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
{\ /* Create group basing PE_start, logPE_stride and PE_size */\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start), \
if (0 == OSHMEM_GROUP_CACHE_ENABLED)\ (1 << OMPI_FINT_2_INT(*logPE_stride)), \
{\ OMPI_FINT_2_INT(*PE_size));\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \ oshmem_op_t* op = T_NAME;\
(1 << OMPI_FINT_2_INT(*logPE_stride)), \ size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
OMPI_FINT_2_INT(*PE_size));\ \
if (!group) \ /* Call collective reduce operation */\
rc = OSHMEM_ERROR;\ rc = group->g_scoll.scoll_reduce(group, \
}\ op, \
else \ FPTR_2_VOID_PTR(target), \
{\ FPTR_2_VOID_PTR(source), \
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ size, \
OMPI_FINT_2_INT(*logPE_stride),\ FPTR_2_VOID_PTR(pSync), \
OMPI_FINT_2_INT(*PE_size));\ FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
if (!group)\ oshmem_proc_group_destroy(group);\
{\ RUNTIME_CHECK_RC(rc); \
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \
(1 << OMPI_FINT_2_INT(*logPE_stride)), \
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
rc = OSHMEM_ERROR;\
cache_group(group,\
OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
}\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_op_t* op = T_NAME;\
size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
\
/* Call collective reduce operation */\
rc = group->g_scoll.scoll_reduce( group, \
op, \
FPTR_2_VOID_PTR(target), \
FPTR_2_VOID_PTR(source), \
size, \
FPTR_2_VOID_PTR(pSync), \
FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
}\
if (0 == OSHMEM_GROUP_CACHE_ENABLED)\
{\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_proc_group_destroy(group);\
}\
}\
}\
} }
SHMEM_AND_TO_ALL(shmem_int2_and_to_all_f, oshmem_op_and_fint2, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_AND_TO_ALL(shmem_int2_and_to_all_f, oshmem_op_and_fint2)
SHMEM_AND_TO_ALL(shmem_int4_and_to_all_f, oshmem_op_and_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_AND_TO_ALL(shmem_int4_and_to_all_f, oshmem_op_and_fint4)
SHMEM_AND_TO_ALL(shmem_int8_and_to_all_f, oshmem_op_and_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_AND_TO_ALL(shmem_int8_and_to_all_f, oshmem_op_and_fint8)

83
oshmem/shmem/fortran/shmem_broadcast_f.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
@ -15,7 +15,6 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -59,7 +58,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_root, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync), (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_root, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync),
(target, source, nlong, PE_root, PE_start, logPE_stride, PE_size, pSync)) (target, source, nlong, PE_root, PE_start, logPE_stride, PE_size, pSync))
#define SHMEM_BROADCAST(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_BROADCAST(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *nlong,\ MPI_Fint *nlong,\
MPI_Fint *PE_root, \ MPI_Fint *PE_root, \
@ -68,70 +67,40 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
MPI_Fint *PE_size, \ MPI_Fint *PE_size, \
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
int rel_PE_root = 0;\
oshmem_op_t* op = T_NAME;\
\ \
if ((0 <= OMPI_FINT_2_INT(*PE_root)) && \ if ((0 <= OMPI_FINT_2_INT(*PE_root)) && \
(OMPI_FINT_2_INT(*PE_root) < OMPI_FINT_2_INT(*PE_size)))\ (OMPI_FINT_2_INT(*PE_root) < OMPI_FINT_2_INT(*PE_size)))\
{\ {\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start), \
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\
{\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \
(1 << OMPI_FINT_2_INT(*logPE_stride)), \ (1 << OMPI_FINT_2_INT(*logPE_stride)), \
OMPI_FINT_2_INT(*PE_size));\ OMPI_FINT_2_INT(*PE_size));\
if (!group || (OMPI_FINT_2_INT(*PE_root) >= group->proc_count))\ if (OMPI_FINT_2_INT(*PE_root) >= group->proc_count)\
{\ {\
rc = OSHMEM_ERROR;\ rc = OSHMEM_ERROR;\
}\ goto out;\
}\ }\
else\ \
{\ /* Define actual PE using relative in active set */\
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ rel_PE_root = oshmem_proc_pe(group->proc_array[OMPI_FINT_2_INT(*PE_root)]);\
OMPI_FINT_2_INT(*logPE_stride),\ \
OMPI_FINT_2_INT(*PE_size));\ /* Call collective broadcast operation */\
if (!group)\ rc = group->g_scoll.scoll_broadcast( group, \
{\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\
OMPI_FINT_2_INT(*PE_size));\
if (!group || (OMPI_FINT_2_INT(*PE_root) >= group->proc_count))\
{\
rc = OSHMEM_ERROR;\
}\
cache_group(group,OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
} /* OSHMEM_GROUP_CACHE_ENABLED */\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
int rel_PE_root = 0;\
oshmem_op_t* op = T_NAME;\
\
/* Define actual PE using relative in active set */\
rel_PE_root = oshmem_proc_pe(group->proc_array[OMPI_FINT_2_INT(*PE_root)]);\
\
/* Call collective broadcast operation */\
rc = group->g_scoll.scoll_broadcast( group, \
rel_PE_root, \ rel_PE_root, \
FPTR_2_VOID_PTR(target), \ FPTR_2_VOID_PTR(target), \
FPTR_2_VOID_PTR(source), \ FPTR_2_VOID_PTR(source), \
OMPI_FINT_2_INT(*nlong) * op->dt_size, \ OMPI_FINT_2_INT(*nlong) * op->dt_size, \
FPTR_2_VOID_PTR(pSync), SCOLL_DEFAULT_ALG );\ FPTR_2_VOID_PTR(pSync), SCOLL_DEFAULT_ALG );\
}\ out: \
if (OSHMEM_GROUP_CACHE_ENABLED == 0) \ oshmem_proc_group_destroy(group);\
{\ RUNTIME_CHECK_RC(rc); \
if ( group )\ }\
{\
oshmem_proc_group_destroy(group);\
}\
} /* OSHMEM_GROUP_CACHE_ENABLED */\
}\
} }
SHMEM_BROADCAST(shmem_broadcast4_f, oshmem_op_prod_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_BROADCAST(shmem_broadcast4_f, oshmem_op_prod_fint4)
SHMEM_BROADCAST(shmem_broadcast8_f, oshmem_op_prod_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_BROADCAST(shmem_broadcast8_f, oshmem_op_prod_fint8)
SHMEM_BROADCAST(shmem_broadcast32_f, oshmem_op_prod_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_BROADCAST(shmem_broadcast32_f, oshmem_op_prod_fint4)
SHMEM_BROADCAST(shmem_broadcast64_f, oshmem_op_prod_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_BROADCAST(shmem_broadcast64_f, oshmem_op_prod_fint8)

86
oshmem/shmem/fortran/shmem_collect_f.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
@ -15,7 +15,6 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -95,7 +94,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
(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), (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),
(target,source,nlong,PE_start,logPE_stride,PE_size,pSync) ) (target,source,nlong,PE_start,logPE_stride,PE_size,pSync) )
#define SHMEM_COLLECT(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_COLLECT(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *nlong, \ MPI_Fint *nlong, \
MPI_Fint *PE_start, \ MPI_Fint *PE_start, \
@ -103,62 +102,29 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
MPI_Fint *PE_size, \ MPI_Fint *PE_size, \
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
{\ /* Create group basing PE_start, logPE_stride and PE_size */\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start), \
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\ (1 << OMPI_FINT_2_INT(*logPE_stride)), \
{\ OMPI_FINT_2_INT(*PE_size));\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \ oshmem_op_t* op = T_NAME;\
(1 << OMPI_FINT_2_INT(*logPE_stride)), \ /* Call collective broadcast operation */\
OMPI_FINT_2_INT(*PE_size));\ rc = group->g_scoll.scoll_collect( group, \
if (!group)\ FPTR_2_VOID_PTR(target), \
rc = OSHMEM_ERROR;\ FPTR_2_VOID_PTR(source), \
}\ OMPI_FINT_2_INT(*nlong) * op->dt_size, \
else\ FPTR_2_VOID_PTR(pSync), \
{\ false, SCOLL_DEFAULT_ALG);\
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ oshmem_proc_group_destroy(group);\
OMPI_FINT_2_INT(*logPE_stride),\ RUNTIME_CHECK_RC(rc);\
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
{\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \
(1 << OMPI_FINT_2_INT(*logPE_stride)), \
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
rc = OSHMEM_ERROR;\
cache_group(group,OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
} /* OSHMEM_GROUP_CACHE_ENABLED */\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_op_t* op = T_NAME;\
/* Call collective broadcast operation */\
rc = group->g_scoll.scoll_collect( group, \
FPTR_2_VOID_PTR(target), \
FPTR_2_VOID_PTR(source), \
OMPI_FINT_2_INT(*nlong) * op->dt_size, \
FPTR_2_VOID_PTR(pSync), \
false, SCOLL_DEFAULT_ALG);\
}\
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\
{\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_proc_group_destroy(group);\
}\
}/* OSHMEM_GROUP_CACHE_ENABLED */\
}\
} }
SHMEM_COLLECT(shmem_collect4_f, oshmem_op_prod_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_COLLECT(shmem_collect4_f, oshmem_op_prod_fint4)
SHMEM_COLLECT(shmem_collect8_f, oshmem_op_prod_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_COLLECT(shmem_collect8_f, oshmem_op_prod_fint8)
SHMEM_COLLECT(shmem_collect32_f, oshmem_op_prod_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_COLLECT(shmem_collect32_f, oshmem_op_prod_fint4)
SHMEM_COLLECT(shmem_collect64_f, oshmem_op_prod_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_COLLECT(shmem_collect64_f, oshmem_op_prod_fint8)
SHMEM_COLLECT(shmem_fcollect4_f, oshmem_op_prod_freal4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_COLLECT(shmem_fcollect4_f, oshmem_op_prod_freal4)
SHMEM_COLLECT(shmem_fcollect8_f, oshmem_op_prod_freal8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_COLLECT(shmem_fcollect8_f, oshmem_op_prod_freal8)
SHMEM_COLLECT(shmem_fcollect32_f, oshmem_op_prod_freal4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_COLLECT(shmem_fcollect32_f, oshmem_op_prod_freal4)
SHMEM_COLLECT(shmem_fcollect64_f, oshmem_op_prod_freal8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_COLLECT(shmem_fcollect64_f, oshmem_op_prod_freal8)

85
oshmem/shmem/fortran/shmem_max_to_all_f.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
@ -15,7 +15,6 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -77,7 +76,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync),
(target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) ) (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
#define SHMEM_MAX_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_MAX_TO_ALL(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *nreduce,\ MPI_Fint *nreduce,\
MPI_Fint *PE_start,\ MPI_Fint *PE_start,\
@ -86,61 +85,29 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
FORTRAN_POINTER_T *pWrk,\ FORTRAN_POINTER_T *pWrk,\
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
{\ /* Create group basing PE_start, logPE_stride and PE_size */\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start),\
if (OSHMEM_GROUP_CACHE_ENABLED == 0) {\ (1 << OMPI_FINT_2_INT(*logPE_stride)),\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\ OMPI_FINT_2_INT(*PE_size));\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\ oshmem_op_t* op = T_NAME;\
OMPI_FINT_2_INT(*PE_size));\ size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
if (!group)\ /* Call collective reduce operation */\
rc = OSHMEM_ERROR;\ rc = group->g_scoll.scoll_reduce( group,\
}\ op,\
else\ FPTR_2_VOID_PTR(target),\
{\ FPTR_2_VOID_PTR(source),\
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ size,\
OMPI_FINT_2_INT(*logPE_stride),\ FPTR_2_VOID_PTR(pSync),\
OMPI_FINT_2_INT(*PE_size));\ FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
if (!group)\ oshmem_proc_group_destroy(group);\
{\ RUNTIME_CHECK_RC(rc);\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
rc = OSHMEM_ERROR;\
cache_group(group,OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
}\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_op_t* op = T_NAME;\
size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
/* Call collective reduce operation */\
rc = group->g_scoll.scoll_reduce( group,\
op,\
FPTR_2_VOID_PTR(target),\
FPTR_2_VOID_PTR(source),\
size,\
FPTR_2_VOID_PTR(pSync),\
FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
}\
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\
{\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_proc_group_destroy(group);\
}\
}\
}\
} }
SHMEM_MAX_TO_ALL(shmem_int2_max_to_all_f, oshmem_op_max_fint2, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MAX_TO_ALL(shmem_int2_max_to_all_f, oshmem_op_max_fint2)
SHMEM_MAX_TO_ALL(shmem_int4_max_to_all_f, oshmem_op_max_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MAX_TO_ALL(shmem_int4_max_to_all_f, oshmem_op_max_fint4)
SHMEM_MAX_TO_ALL(shmem_int8_max_to_all_f, oshmem_op_max_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MAX_TO_ALL(shmem_int8_max_to_all_f, oshmem_op_max_fint8)
SHMEM_MAX_TO_ALL(shmem_real4_max_to_all_f, oshmem_op_max_freal4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MAX_TO_ALL(shmem_real4_max_to_all_f, oshmem_op_max_freal4)
SHMEM_MAX_TO_ALL(shmem_real8_max_to_all_f, oshmem_op_max_freal8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MAX_TO_ALL(shmem_real8_max_to_all_f, oshmem_op_max_freal8)
SHMEM_MAX_TO_ALL(shmem_real16_max_to_all_f, oshmem_op_max_freal16, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MAX_TO_ALL(shmem_real16_max_to_all_f, oshmem_op_max_freal16)

86
oshmem/shmem/fortran/shmem_min_to_all_f.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
@ -15,7 +15,6 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -78,7 +77,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
(target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) ) (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
#define SHMEM_MIN_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_MIN_TO_ALL(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *nreduce,\ MPI_Fint *nreduce,\
MPI_Fint *PE_start,\ MPI_Fint *PE_start,\
@ -87,61 +86,30 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
FORTRAN_POINTER_T *pWrk,\ FORTRAN_POINTER_T *pWrk,\
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
{\ /* Create group basing PE_start, logPE_stride and PE_size */\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start),\
if (OSHMEM_GROUP_CACHE_ENABLED == 0) {\ (1 << OMPI_FINT_2_INT(*logPE_stride)),\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\ OMPI_FINT_2_INT(*PE_size));\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\ /* Collective operation call */\
OMPI_FINT_2_INT(*PE_size));\ oshmem_op_t* op = T_NAME;\
if (!group)\ size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
rc = OSHMEM_ERROR;\ /* Call collective reduce operation */\
}\ rc = group->g_scoll.scoll_reduce( group,\
else\ op,\
{\ FPTR_2_VOID_PTR(target),\
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ FPTR_2_VOID_PTR(source),\
OMPI_FINT_2_INT(*logPE_stride),\ size,\
OMPI_FINT_2_INT(*PE_size));\ FPTR_2_VOID_PTR(pSync),\
if (!group)\ FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
{\ oshmem_proc_group_destroy(group);\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\ RUNTIME_CHECK_RC(rc);\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
rc = OSHMEM_ERROR;\
cache_group(group,OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
}\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_op_t* op = T_NAME;\
size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
/* Call collective reduce operation */\
rc = group->g_scoll.scoll_reduce( group,\
op,\
FPTR_2_VOID_PTR(target),\
FPTR_2_VOID_PTR(source),\
size,\
FPTR_2_VOID_PTR(pSync),\
FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
}\
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\
{\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_proc_group_destroy(group);\
}\
}\
}\
} }
SHMEM_MIN_TO_ALL(shmem_int2_min_to_all_f, oshmem_op_min_fint2, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MIN_TO_ALL(shmem_int2_min_to_all_f, oshmem_op_min_fint2)
SHMEM_MIN_TO_ALL(shmem_int4_min_to_all_f, oshmem_op_min_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MIN_TO_ALL(shmem_int4_min_to_all_f, oshmem_op_min_fint4)
SHMEM_MIN_TO_ALL(shmem_int8_min_to_all_f, oshmem_op_min_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MIN_TO_ALL(shmem_int8_min_to_all_f, oshmem_op_min_fint8)
SHMEM_MIN_TO_ALL(shmem_real4_min_to_all_f, oshmem_op_min_freal4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MIN_TO_ALL(shmem_real4_min_to_all_f, oshmem_op_min_freal4)
SHMEM_MIN_TO_ALL(shmem_real8_min_to_all_f, oshmem_op_min_freal8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MIN_TO_ALL(shmem_real8_min_to_all_f, oshmem_op_min_freal8)
SHMEM_MIN_TO_ALL(shmem_real16_min_to_all_f, oshmem_op_min_freal16, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_MIN_TO_ALL(shmem_real16_min_to_all_f, oshmem_op_min_freal16)

79
oshmem/shmem/fortran/shmem_or_to_all_f.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
@ -15,7 +15,6 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -50,7 +49,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync),
(target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) ) (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
#define SHMEM_OR_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_OR_TO_ALL(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *nreduce,\ MPI_Fint *nreduce,\
MPI_Fint *PE_start,\ MPI_Fint *PE_start,\
@ -59,58 +58,26 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
FORTRAN_POINTER_T *pWrk,\ FORTRAN_POINTER_T *pWrk,\
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
{\ /* Create group basing PE_start, logPE_stride and PE_size */\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start),\
if (OSHMEM_GROUP_CACHE_ENABLED == 0) {\ (1 << OMPI_FINT_2_INT(*logPE_stride)),\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\ OMPI_FINT_2_INT(*PE_size));\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\ oshmem_op_t* op = T_NAME;\
OMPI_FINT_2_INT(*PE_size));\ size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
if (!group)\ /* Call collective reduce operation */\
rc = OSHMEM_ERROR;\ rc = group->g_scoll.scoll_reduce( group,\
}\ op,\
else\ FPTR_2_VOID_PTR(target),\
{\ FPTR_2_VOID_PTR(source),\
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ size,\
OMPI_FINT_2_INT(*logPE_stride),\ FPTR_2_VOID_PTR(pSync),\
OMPI_FINT_2_INT(*PE_size));\ FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
if (!group)\ oshmem_proc_group_destroy(group);\
{\ RUNTIME_CHECK_RC(rc); \
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
rc = OSHMEM_ERROR;\
cache_group(group,OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
}\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_op_t* op = T_NAME;\
size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
/* Call collective reduce operation */\
rc = group->g_scoll.scoll_reduce( group,\
op,\
FPTR_2_VOID_PTR(target),\
FPTR_2_VOID_PTR(source),\
size,\
FPTR_2_VOID_PTR(pSync),\
FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
}\
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\
{\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_proc_group_destroy(group);\
}\
}\
}\
} }
SHMEM_OR_TO_ALL(shmem_int2_or_to_all_f, oshmem_op_or_fint2, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_OR_TO_ALL(shmem_int2_or_to_all_f, oshmem_op_or_fint2)
SHMEM_OR_TO_ALL(shmem_int4_or_to_all_f, oshmem_op_or_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_OR_TO_ALL(shmem_int4_or_to_all_f, oshmem_op_or_fint4)
SHMEM_OR_TO_ALL(shmem_int8_or_to_all_f, oshmem_op_or_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_OR_TO_ALL(shmem_int8_or_to_all_f, oshmem_op_or_fint8)

89
oshmem/shmem/fortran/shmem_prod_to_all_f.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
@ -15,7 +15,6 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -96,7 +95,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync),
(target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) ) (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
#define SHMEM_PROD_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_PROD_TO_ALL(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *nreduce,\ MPI_Fint *nreduce,\
MPI_Fint *PE_start,\ MPI_Fint *PE_start,\
@ -105,63 +104,31 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
FORTRAN_POINTER_T *pWrk,\ FORTRAN_POINTER_T *pWrk,\
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
{\ /* Create group basing PE_start, logPE_stride and PE_size */\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start),\
if (OSHMEM_GROUP_CACHE_ENABLED == 0) {\ (1 << OMPI_FINT_2_INT(*logPE_stride)),\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\ OMPI_FINT_2_INT(*PE_size));\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\ oshmem_op_t* op = T_NAME;\
OMPI_FINT_2_INT(*PE_size));\ size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
if (!group)\ /* Call collective reduce operation */\
rc = OSHMEM_ERROR;\ rc = group->g_scoll.scoll_reduce( group,\
}\ op,\
else\ FPTR_2_VOID_PTR(target),\
{\ FPTR_2_VOID_PTR(source),\
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ size,\
OMPI_FINT_2_INT(*logPE_stride),\ FPTR_2_VOID_PTR(pSync),\
OMPI_FINT_2_INT(*PE_size));\ FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
if (!group)\ oshmem_proc_group_destroy(group);\
{\ RUNTIME_CHECK_RC(rc);\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
rc = OSHMEM_ERROR;\
cache_group(group,OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
}\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_op_t* op = T_NAME;\
size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
/* Call collective reduce operation */\
rc = group->g_scoll.scoll_reduce( group,\
op,\
FPTR_2_VOID_PTR(target),\
FPTR_2_VOID_PTR(source),\
size,\
FPTR_2_VOID_PTR(pSync),\
FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
}\
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\
{\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_proc_group_destroy(group);\
}\
}\
}\
} }
SHMEM_PROD_TO_ALL(shmem_int2_prod_to_all_f, oshmem_op_prod_fint2, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_PROD_TO_ALL(shmem_int2_prod_to_all_f, oshmem_op_prod_fint2)
SHMEM_PROD_TO_ALL(shmem_int4_prod_to_all_f, oshmem_op_prod_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_PROD_TO_ALL(shmem_int4_prod_to_all_f, oshmem_op_prod_fint4)
SHMEM_PROD_TO_ALL(shmem_int8_prod_to_all_f, oshmem_op_prod_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_PROD_TO_ALL(shmem_int8_prod_to_all_f, oshmem_op_prod_fint8)
SHMEM_PROD_TO_ALL(shmem_comp4_prod_to_all_f, oshmem_op_prod_complexf, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_PROD_TO_ALL(shmem_comp4_prod_to_all_f, oshmem_op_prod_complexf)
SHMEM_PROD_TO_ALL(shmem_comp8_prod_to_all_f, oshmem_op_prod_complexd, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_PROD_TO_ALL(shmem_comp8_prod_to_all_f, oshmem_op_prod_complexd)
SHMEM_PROD_TO_ALL(shmem_real4_prod_to_all_f, oshmem_op_prod_freal4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_PROD_TO_ALL(shmem_real4_prod_to_all_f, oshmem_op_prod_freal4)
SHMEM_PROD_TO_ALL(shmem_real8_prod_to_all_f, oshmem_op_prod_freal8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_PROD_TO_ALL(shmem_real8_prod_to_all_f, oshmem_op_prod_freal8)
SHMEM_PROD_TO_ALL(shmem_real16_prod_to_all_f, oshmem_op_prod_freal16, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_PROD_TO_ALL(shmem_real16_prod_to_all_f, oshmem_op_prod_freal16)

89
oshmem/shmem/fortran/shmem_sum_to_all_f.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
@ -15,7 +15,6 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -95,7 +94,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync),
(target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) ) (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
#define SHMEM_SUM_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_SUM_TO_ALL(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *nreduce,\ MPI_Fint *nreduce,\
MPI_Fint *PE_start,\ MPI_Fint *PE_start,\
@ -104,63 +103,31 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
FORTRAN_POINTER_T *pWrk,\ FORTRAN_POINTER_T *pWrk,\
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
{\ /* Create group basing PE_start, logPE_stride and PE_size */\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start),\
if (OSHMEM_GROUP_CACHE_ENABLED == 0) {\ (1 << OMPI_FINT_2_INT(*logPE_stride)),\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\ OMPI_FINT_2_INT(*PE_size));\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\ oshmem_op_t* op = T_NAME;\
OMPI_FINT_2_INT(*PE_size));\ size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
if (!group)\ /* Call collective reduce operation */\
rc = OSHMEM_ERROR;\ rc = group->g_scoll.scoll_reduce( group,\
}\ op,\
else\ FPTR_2_VOID_PTR(target),\
{\ FPTR_2_VOID_PTR(source),\
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ size,\
OMPI_FINT_2_INT(*logPE_stride),\ FPTR_2_VOID_PTR(pSync),\
OMPI_FINT_2_INT(*PE_size));\ FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
if (!group)\ oshmem_proc_group_destroy(group);\
{\ RUNTIME_CHECK_RC(rc);\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start),\
(1 << OMPI_FINT_2_INT(*logPE_stride)),\
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
rc = OSHMEM_ERROR;\
cache_group(group,OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
}\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_op_t* op = T_NAME;\
size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
/* Call collective reduce operation */\
rc = group->g_scoll.scoll_reduce( group,\
op,\
FPTR_2_VOID_PTR(target),\
FPTR_2_VOID_PTR(source),\
size,\
FPTR_2_VOID_PTR(pSync),\
FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
}\
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\
{\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_proc_group_destroy(group);\
}\
}\
}\
} }
SHMEM_SUM_TO_ALL(shmem_int2_sum_to_all_f, oshmem_op_sum_fint2, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_SUM_TO_ALL(shmem_int2_sum_to_all_f, oshmem_op_sum_fint2)
SHMEM_SUM_TO_ALL(shmem_int4_sum_to_all_f, oshmem_op_sum_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_SUM_TO_ALL(shmem_int4_sum_to_all_f, oshmem_op_sum_fint4)
SHMEM_SUM_TO_ALL(shmem_int8_sum_to_all_f, oshmem_op_sum_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_SUM_TO_ALL(shmem_int8_sum_to_all_f, oshmem_op_sum_fint8)
SHMEM_SUM_TO_ALL(shmem_comp4_sum_to_all_f, oshmem_op_sum_complexf, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_SUM_TO_ALL(shmem_comp4_sum_to_all_f, oshmem_op_sum_complexf)
SHMEM_SUM_TO_ALL(shmem_comp8_sum_to_all_f, oshmem_op_sum_complexd, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_SUM_TO_ALL(shmem_comp8_sum_to_all_f, oshmem_op_sum_complexd)
SHMEM_SUM_TO_ALL(shmem_real4_sum_to_all_f, oshmem_op_sum_freal4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_SUM_TO_ALL(shmem_real4_sum_to_all_f, oshmem_op_sum_freal4)
SHMEM_SUM_TO_ALL(shmem_real8_sum_to_all_f, oshmem_op_sum_freal8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_SUM_TO_ALL(shmem_real8_sum_to_all_f, oshmem_op_sum_freal8)
SHMEM_SUM_TO_ALL(shmem_real16_sum_to_all_f, oshmem_op_sum_freal16, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_SUM_TO_ALL(shmem_real16_sum_to_all_f, oshmem_op_sum_freal16)

86
oshmem/shmem/fortran/shmem_xor_to_all_f.c
Просмотреть файл

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2013 Mellanox Technologies, Inc. * Copyright (c) 2013-2018 Mellanox Technologies, Inc.
* All rights reserved. * All rights reserved.
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$ * $COPYRIGHT$
@ -15,7 +15,6 @@
#include "oshmem/constants.h" #include "oshmem/constants.h"
#include "oshmem/mca/scoll/scoll.h" #include "oshmem/mca/scoll/scoll.h"
#include "oshmem/proc/proc.h" #include "oshmem/proc/proc.h"
#include "oshmem/proc/proc_group_cache.h"
#include "oshmem/op/op.h" #include "oshmem/op/op.h"
#if OSHMEM_PROFILING #if OSHMEM_PROFILING
@ -68,7 +67,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync),
(target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) ) (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
#define SHMEM_XOR_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \ #define SHMEM_XOR_TO_ALL(F_NAME, T_NAME) void F_NAME(FORTRAN_POINTER_T target, \
FORTRAN_POINTER_T source, \ FORTRAN_POINTER_T source, \
MPI_Fint *nreduce,\ MPI_Fint *nreduce,\
MPI_Fint *PE_start,\ MPI_Fint *PE_start,\
@ -77,62 +76,29 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
FORTRAN_POINTER_T *pWrk,\ FORTRAN_POINTER_T *pWrk,\
FORTRAN_POINTER_T pSync)\ FORTRAN_POINTER_T pSync)\
{\ {\
int rc = OSHMEM_SUCCESS;\ int rc;\
oshmem_group_t* group = NULL;\ oshmem_group_t *group;\
{\ /* Create group basing PE_start, logPE_stride and PE_size */\
/* Create group basing PE_start, logPE_stride and PE_size */\ group = oshmem_proc_group_create_nofail(OMPI_FINT_2_INT(*PE_start), \
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\ (1 << OMPI_FINT_2_INT(*logPE_stride)), \
{\ OMPI_FINT_2_INT(*PE_size));\
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \ oshmem_op_t* op = T_NAME;\
(1 << OMPI_FINT_2_INT(*logPE_stride)), \ size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
OMPI_FINT_2_INT(*PE_size));\ \
if (!group)\ /* Call collective reduce operation */\
rc = OSHMEM_ERROR;\ rc = group->g_scoll.scoll_reduce( group,\
}\ op,\
else\ FPTR_2_VOID_PTR(target),\
{\ FPTR_2_VOID_PTR(source),\
group = find_group_in_cache(OMPI_FINT_2_INT(*PE_start),\ size,\
OMPI_FINT_2_INT(*logPE_stride),\ FPTR_2_VOID_PTR(pSync),\
OMPI_FINT_2_INT(*PE_size));\ FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
if (!group)\ oshmem_proc_group_destroy(group);\
{\ RUNTIME_CHECK_RC(rc); \
group = oshmem_proc_group_create(OMPI_FINT_2_INT(*PE_start), \
(1 << OMPI_FINT_2_INT(*logPE_stride)), \
OMPI_FINT_2_INT(*PE_size));\
if (!group)\
rc = OSHMEM_ERROR;\
cache_group(group,OMPI_FINT_2_INT(*PE_start),\
OMPI_FINT_2_INT(*logPE_stride),\
OMPI_FINT_2_INT(*PE_size));\
}\
}\
/* Collective operation call */\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_op_t* op = T_NAME;\
size_t size = OMPI_FINT_2_INT(*nreduce) * op->dt_size;\
\
/* Call collective reduce operation */\
rc = group->g_scoll.scoll_reduce( group,\
op,\
FPTR_2_VOID_PTR(target),\
FPTR_2_VOID_PTR(source),\
size,\
FPTR_2_VOID_PTR(pSync),\
FPTR_2_VOID_PTR(*pWrk), SCOLL_DEFAULT_ALG);\
}\
if (OSHMEM_GROUP_CACHE_ENABLED == 0)\
{\
if ( rc == OSHMEM_SUCCESS )\
{\
oshmem_proc_group_destroy(group);\
}\
}\
}\
} }
SHMEM_XOR_TO_ALL(shmem_int2_xor_to_all_f, oshmem_op_xor_fint2, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_XOR_TO_ALL(shmem_int2_xor_to_all_f, oshmem_op_xor_fint2)
SHMEM_XOR_TO_ALL(shmem_int4_xor_to_all_f, oshmem_op_xor_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_XOR_TO_ALL(shmem_int4_xor_to_all_f, oshmem_op_xor_fint4)
SHMEM_XOR_TO_ALL(shmem_int8_xor_to_all_f, oshmem_op_xor_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_XOR_TO_ALL(shmem_int8_xor_to_all_f, oshmem_op_xor_fint8)
SHMEM_XOR_TO_ALL(shmem_comp4_xor_to_all_f, oshmem_op_xor_fint4, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_XOR_TO_ALL(shmem_comp4_xor_to_all_f, oshmem_op_xor_fint4)
SHMEM_XOR_TO_ALL(shmem_comp8_xor_to_all_f, oshmem_op_xor_fint8, OSHMEM_GROUP_CACHE_ENABLED) SHMEM_XOR_TO_ALL(shmem_comp8_xor_to_all_f, oshmem_op_xor_fint8)