Merge pull request #6104 from hoopoepg/topic/oshmem-zero-len-coll
OSHMEM: added processing of zero-length collectives
Этот коммит содержится в:
Коммит
39daf7a436
@ -61,6 +61,11 @@ int mca_scoll_basic_alltoall(struct oshmem_group_t *group,
|
||||
return OSHMEM_ERR_BAD_PARAM;
|
||||
}
|
||||
|
||||
/* Do nothing on zero-length request */
|
||||
if (OPAL_UNLIKELY(!nelems)) {
|
||||
return OPAL_SUCCESS;
|
||||
}
|
||||
|
||||
if ((sst == 1) && (dst == 1)) {
|
||||
rc = a2a_alg_simple(group, target, source, nelems, element_size);
|
||||
} else {
|
||||
|
@ -55,6 +55,11 @@ int mca_scoll_basic_broadcast(struct oshmem_group_t *group,
|
||||
if ((rc == OSHMEM_SUCCESS) && oshmem_proc_group_is_member(group)) {
|
||||
int i = 0;
|
||||
|
||||
/* Do nothing on zero-length request */
|
||||
if (OPAL_UNLIKELY(!nlong)) {
|
||||
return OSHMEM_SUCCESS;
|
||||
}
|
||||
|
||||
if (pSync) {
|
||||
alg = (alg == SCOLL_DEFAULT_ALG ?
|
||||
mca_scoll_basic_param_broadcast_algorithm : alg);
|
||||
|
@ -66,6 +66,11 @@ int mca_scoll_basic_collect(struct oshmem_group_t *group,
|
||||
if ((rc == OSHMEM_SUCCESS) && oshmem_proc_group_is_member(group)) {
|
||||
int i = 0;
|
||||
|
||||
/* Do nothing on zero-length request */
|
||||
if (OPAL_UNLIKELY(!nlong)) {
|
||||
return OPAL_SUCCESS;
|
||||
}
|
||||
|
||||
if (nlong_type) {
|
||||
alg = (alg == SCOLL_DEFAULT_ALG ?
|
||||
mca_scoll_basic_param_collect_algorithm : alg);
|
||||
|
@ -78,6 +78,11 @@ int mca_scoll_basic_reduce(struct oshmem_group_t *group,
|
||||
if ((rc == OSHMEM_SUCCESS) && oshmem_proc_group_is_member(group)) {
|
||||
int i = 0;
|
||||
|
||||
/* Do nothing on zero-length request */
|
||||
if (OPAL_UNLIKELY(!nlong)) {
|
||||
return OSHMEM_SUCCESS;
|
||||
}
|
||||
|
||||
if (pSync) {
|
||||
alg = (alg == SCOLL_DEFAULT_ALG ?
|
||||
mca_scoll_basic_param_reduce_algorithm : alg);
|
||||
|
@ -54,6 +54,12 @@ int mca_scoll_mpi_broadcast(struct oshmem_group_t *group,
|
||||
}
|
||||
dtype = &ompi_mpi_char.dt;
|
||||
root = oshmem_proc_group_find_id(group, PE_root);
|
||||
|
||||
/* Do nothing on zero-length request */
|
||||
if (OPAL_UNLIKELY(!nlong)) {
|
||||
return OSHMEM_SUCCESS;
|
||||
}
|
||||
|
||||
/* Open SHMEM specification has the following constrains (page 85):
|
||||
* "If using C/C++, nelems must be of type integer. If you are using Fortran, it must be a
|
||||
* default integer value". And also fortran signature says "INTEGER".
|
||||
@ -104,6 +110,12 @@ int mca_scoll_mpi_collect(struct oshmem_group_t *group,
|
||||
void *sbuf, *rbuf;
|
||||
MPI_COLL_VERBOSE(20,"RUNNING MPI ALLGATHER");
|
||||
mpi_module = (mca_scoll_mpi_module_t *) group->g_scoll.scoll_collect_module;
|
||||
|
||||
/* Do nothing on zero-length request */
|
||||
if (OPAL_UNLIKELY(!nlong)) {
|
||||
return OSHMEM_SUCCESS;
|
||||
}
|
||||
|
||||
if (nlong_type == true) {
|
||||
sbuf = (void *) source;
|
||||
rbuf = target;
|
||||
@ -177,6 +189,12 @@ int mca_scoll_mpi_reduce(struct oshmem_group_t *group,
|
||||
dtype = shmem_dtype_to_ompi_dtype(op);
|
||||
h_op = shmem_op_to_ompi_op(op->op);
|
||||
count = nlong/op->dt_size;
|
||||
|
||||
/* Do nothing on zero-length request */
|
||||
if (OPAL_UNLIKELY(!nlong)) {
|
||||
return OSHMEM_SUCCESS;
|
||||
}
|
||||
|
||||
/* Open SHMEM specification has the following constrains (page 85):
|
||||
* "If using C/C++, nelems must be of type integer. If you are using Fortran, it must be a
|
||||
* default integer value". And also fortran signature says "INTEGER".
|
||||
|
@ -200,6 +200,13 @@ OSHMEM_DECLSPEC int oshmem_shmem_register_params(void);
|
||||
RUNTIME_CHECK_ERROR("Required address %p is not in symmetric space\n", ((void*)x)); \
|
||||
oshmem_shmem_abort(-1); \
|
||||
}
|
||||
/* Check if address is in symmetric space or size is zero */
|
||||
#define RUNTIME_CHECK_ADDR_SIZE(x,s) \
|
||||
if (OPAL_UNLIKELY((s) && !MCA_MEMHEAP_CALL(is_symmetric_addr((x))))) \
|
||||
{ \
|
||||
RUNTIME_CHECK_ERROR("Required address %p is not in symmetric space\n", ((void*)x)); \
|
||||
oshmem_shmem_abort(-1); \
|
||||
}
|
||||
#define RUNTIME_CHECK_WITH_MEMHEAP_SIZE(x) \
|
||||
if (OPAL_UNLIKELY((long)(x) > MCA_MEMHEAP_CALL(size))) \
|
||||
{ \
|
||||
@ -212,6 +219,7 @@ OSHMEM_DECLSPEC int oshmem_shmem_register_params(void);
|
||||
#define RUNTIME_CHECK_INIT()
|
||||
#define RUNTIME_CHECK_PE(x)
|
||||
#define RUNTIME_CHECK_ADDR(x)
|
||||
#define RUNTIME_CHECK_ADDR_SIZE(x,s)
|
||||
#define RUNTIME_CHECK_WITH_MEMHEAP_SIZE(x)
|
||||
|
||||
#endif /* OSHMEM_PARAM_CHECK */
|
||||
|
@ -30,7 +30,7 @@ static void _shmem_alltoall(void *target,
|
||||
int PE_size,
|
||||
long *pSync);
|
||||
|
||||
#define SHMEM_TYPE_ALLTOALL(name, element_size) \
|
||||
#define SHMEM_TYPE_ALLTOALL(name, element_size) \
|
||||
void shmem##name(void *target, \
|
||||
const void *source, \
|
||||
size_t nelems, \
|
||||
@ -40,15 +40,15 @@ static void _shmem_alltoall(void *target,
|
||||
long *pSync) \
|
||||
{ \
|
||||
RUNTIME_CHECK_INIT(); \
|
||||
RUNTIME_CHECK_ADDR(target); \
|
||||
RUNTIME_CHECK_ADDR(source); \
|
||||
RUNTIME_CHECK_ADDR_SIZE(target, nelems); \
|
||||
RUNTIME_CHECK_ADDR_SIZE(source, nelems); \
|
||||
\
|
||||
_shmem_alltoall(target, source, 1, 1, nelems, element_size, \
|
||||
PE_start, logPE_stride, PE_size, \
|
||||
pSync); \
|
||||
}
|
||||
|
||||
#define SHMEM_TYPE_ALLTOALLS(name, element_size) \
|
||||
#define SHMEM_TYPE_ALLTOALLS(name, element_size) \
|
||||
void shmem##name(void *target, \
|
||||
const void *source, \
|
||||
ptrdiff_t dst, ptrdiff_t sst, \
|
||||
@ -59,8 +59,8 @@ static void _shmem_alltoall(void *target,
|
||||
long *pSync) \
|
||||
{ \
|
||||
RUNTIME_CHECK_INIT(); \
|
||||
RUNTIME_CHECK_ADDR(target); \
|
||||
RUNTIME_CHECK_ADDR(source); \
|
||||
RUNTIME_CHECK_ADDR_SIZE(target, nelems); \
|
||||
RUNTIME_CHECK_ADDR_SIZE(source, nelems); \
|
||||
\
|
||||
_shmem_alltoall(target, source, dst, sst, nelems, element_size, \
|
||||
PE_start, logPE_stride, PE_size, \
|
||||
|
@ -29,7 +29,7 @@ static void _shmem_broadcast(void *target,
|
||||
int PE_size,
|
||||
long *pSync);
|
||||
|
||||
#define SHMEM_TYPE_BROADCAST(name, element_size) \
|
||||
#define SHMEM_TYPE_BROADCAST(name, element_size) \
|
||||
void shmem##name( void *target, \
|
||||
const void *source, \
|
||||
size_t nelems, \
|
||||
@ -40,10 +40,10 @@ static void _shmem_broadcast(void *target,
|
||||
long *pSync) \
|
||||
{ \
|
||||
RUNTIME_CHECK_INIT(); \
|
||||
RUNTIME_CHECK_ADDR(target); \
|
||||
RUNTIME_CHECK_ADDR(source); \
|
||||
RUNTIME_CHECK_ADDR_SIZE(target, nelems); \
|
||||
RUNTIME_CHECK_ADDR_SIZE(source, nelems); \
|
||||
\
|
||||
_shmem_broadcast( target, source, nelems * element_size, \
|
||||
_shmem_broadcast( target, source, nelems * element_size, \
|
||||
PE_root, PE_start, logPE_stride, PE_size, \
|
||||
pSync); \
|
||||
}
|
||||
|
@ -39,10 +39,10 @@ static void _shmem_collect(void *target,
|
||||
long *pSync) \
|
||||
{ \
|
||||
RUNTIME_CHECK_INIT(); \
|
||||
RUNTIME_CHECK_ADDR(target); \
|
||||
RUNTIME_CHECK_ADDR(source); \
|
||||
RUNTIME_CHECK_ADDR_SIZE(target, nelems); \
|
||||
RUNTIME_CHECK_ADDR_SIZE(source, nelems); \
|
||||
\
|
||||
_shmem_collect( target, source, nelems * element_size, \
|
||||
_shmem_collect( target, source, nelems * element_size, \
|
||||
PE_start, logPE_stride, PE_size, \
|
||||
pSync, \
|
||||
nelems_type); \
|
||||
|
@ -26,8 +26,8 @@
|
||||
* object of every PE in the active set. The active set of PEs is defined by the triple PE_start,
|
||||
* logPE_stride and PE_size.
|
||||
*/
|
||||
#define SHMEM_TYPE_REDUCE_OP(name, type_name, type, prefix) \
|
||||
void prefix##type_name##_##name##_to_all( type *target, \
|
||||
#define SHMEM_TYPE_REDUCE_OP(name, type_name, type, prefix) \
|
||||
void prefix##type_name##_##name##_to_all( type *target, \
|
||||
const type *source, \
|
||||
int nreduce, \
|
||||
int PE_start, \
|
||||
@ -40,8 +40,8 @@
|
||||
oshmem_group_t* group = NULL; \
|
||||
\
|
||||
RUNTIME_CHECK_INIT(); \
|
||||
RUNTIME_CHECK_ADDR(target); \
|
||||
RUNTIME_CHECK_ADDR(source); \
|
||||
RUNTIME_CHECK_ADDR_SIZE(target, nreduce); \
|
||||
RUNTIME_CHECK_ADDR_SIZE(source, nreduce); \
|
||||
\
|
||||
{ \
|
||||
group = oshmem_proc_group_create_nofail(PE_start, 1<<logPE_stride, PE_size); \
|
||||
|
Загрузка…
x
Ссылка в новой задаче
Block a user