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

Smallest overhead for contiguous predefined datatypes. There is now a special path

Просмотр исходного кода 
in the convertor code for predefined contiguous types.

This commit was SVN r9821.
Этот коммит содержится в:
George Bosilca 2006-05-04 22:21:09 +00:00
родитель 71d328c086
Коммит c53914ce31
2 изменённых файлов: 115 добавлений и 37 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

106
ompi/datatype/convertor.c
Просмотреть файл

@ -188,6 +188,45 @@ int32_t ompi_convertor_pack( ompi_convertor_t* pConv,
{
OMPI_CONVERTOR_SET_STATUS_BEFORE_PACK_UNPACK( pConv, iov, out_size, max_data );
if( (pConv->flags & (DT_FLAG_PREDEFINED | CONVERTOR_HOMOGENEOUS)) ==
(DT_FLAG_PREDEFINED | CONVERTOR_HOMOGENEOUS) ) {
/* We are doing conversion on a predefined datatype. The convertor contain
* minimal informations, we only use the bConverted to manage the conversion.
*/
uint32_t i;
char* base_pointer;
*max_data = pConv->bConverted;
for( i = 0; i < *out_size; i++ ) {
base_pointer = pConv->pBaseBuf + pConv->bConverted;
if( NULL == iov[i].iov_base ) {
iov[i].iov_base = base_pointer;
pConv->bConverted += iov[i].iov_len;
if( pConv->bConverted > pConv->local_size ) {
iov[i].iov_len -= (pConv->bConverted - pConv->local_size);
goto predefined_data_pack;
}
} else {
pConv->bConverted += iov[i].iov_len;
if( pConv->bConverted > pConv->local_size ) {
iov[i].iov_len -= (pConv->bConverted - pConv->local_size);
MEMCPY( iov[i].iov_base, base_pointer, iov[i].iov_len );
goto predefined_data_pack;
}
MEMCPY( iov[i].iov_base, base_pointer,
iov[i].iov_len );
}
}
*max_data = pConv->bConverted - (*max_data);
return 0;
predefined_data_pack:
*out_size = i;
*max_data = pConv->bConverted - (*max_data);
pConv->bConverted = pConv->local_size;
pConv->flags |= CONVERTOR_COMPLETED;
return 1;
}
/* There is no specific memory allocation. If the convertor notice that some memory
* is required in order to perform the operation (depend on the way the convertor
* was configured) it will call the attached function to get some memory. Any failure
@ -202,6 +241,36 @@ inline int32_t ompi_convertor_unpack( ompi_convertor_t* pConv,
{
OMPI_CONVERTOR_SET_STATUS_BEFORE_PACK_UNPACK( pConv, iov, out_size, max_data );
if( (pConv->flags & (DT_FLAG_PREDEFINED | CONVERTOR_HOMOGENEOUS)) ==
(DT_FLAG_PREDEFINED | CONVERTOR_HOMOGENEOUS) ) {
/* We are doing conversion on a predefined datatype. The convertor contain
* minimal informations, we only use the bConverted to manage the conversion.
*/
uint32_t i;
char* base_pointer;
*max_data = pConv->bConverted;
for( i = 0; i < *out_size; i++ ) {
base_pointer = pConv->pBaseBuf + pConv->bConverted;
pConv->bConverted += iov[i].iov_len;
if( pConv->bConverted > pConv->local_size ) {
pConv->bConverted = pConv->local_size;
iov[i].iov_len -= (pConv->bConverted - pConv->local_size);
MEMCPY( base_pointer, iov[i].iov_base, iov[i].iov_len );
goto predefined_data_unpack;
}
MEMCPY( base_pointer, iov[i].iov_base, iov[i].iov_len );
}
*max_data = pConv->bConverted - (*max_data);
return 0;
predefined_data_unpack:
*out_size = i;
*max_data = pConv->bConverted - (*max_data);
pConv->bConverted = pConv->local_size;
pConv->flags |= CONVERTOR_COMPLETED;
return 1;
}
return pConv->fAdvance( pConv, iov, out_size, max_data, freeAfter );
}
@ -291,20 +360,6 @@ int32_t ompi_convertor_set_position_nocheck( ompi_convertor_t* convertor,
{
int32_t rc;
/*
* Do not allow the convertor to go outside the data boundaries. This test include
* the check for datatype with size zero as well as for convertors with a count of zero.
*/
if( convertor->local_size <= *position) {
convertor->flags |= CONVERTOR_COMPLETED;
convertor->bConverted = convertor->local_size;
*position = convertor->bConverted;
return OMPI_SUCCESS;
}
/* Remove the completed flag if it's already set */
convertor->flags &= ~CONVERTOR_COMPLETED;
/*
* If we plan to rollback the convertor then first we have to set it
* at the beginning.
@ -323,17 +378,6 @@ int32_t ompi_convertor_set_position_nocheck( ompi_convertor_t* convertor,
return rc;
}
int32_t
ompi_convertor_personalize( ompi_convertor_t* convertor, uint32_t flags,
size_t* position, memalloc_fct_t allocfn, void* userdata )
{
convertor->flags |= flags;
convertor->memAlloc_fn = allocfn;
convertor->memAlloc_userdata = userdata;
return ompi_convertor_set_position( convertor, position );
}
/* This function will initialize a convertor based on a previously created convertor. The idea
* is the move outside these function the heavy selection of architecture features for the convertors.
*
@ -365,16 +409,22 @@ int ompi_convertor_prepare( ompi_convertor_t* convertor,
convertor->local_size = convertor->count * datatype->size;
/* If the data is empty we don't have to anything except mark the convertor as
* completed. With this flag set the pack and unpack functions will not do
* anything.
* anything. In order to decrease the data dependencies (and to speed-up this code)
* we will not test the convertor->local_size but we can test the 2 components.
*/
if( 0 == convertor->local_size ) {
if( (0 == convertor->count) || (0 == datatype->size) ) {
convertor->flags |= CONVERTOR_COMPLETED;
convertor->remote_size = 0;
convertor->local_size = convertor->remote_size = 0;
return OMPI_SUCCESS;
}
if( convertor->remoteArch == ompi_mpi_local_arch ) {
convertor->remote_size = convertor->local_size;
/* For predefined datatypes (contiguous) do nothing more */
if( (convertor->flags & DT_FLAG_PREDEFINED) == DT_FLAG_PREDEFINED ) {
convertor->bConverted = 0;
return OMPI_SUCCESS;
}
} else {
int i;
uint64_t bdt_mask = datatype->bdt_used >> DT_CHAR;

46
ompi/datatype/convertor.h
Просмотреть файл

@ -204,15 +204,6 @@ OMPI_DECLSPEC ompi_convertor_t* ompi_convertor_create( int32_t remote_arch, int3
*/
OMPI_DECLSPEC int ompi_convertor_cleanup( ompi_convertor_t* convertor );
/*
*
*/
OMPI_DECLSPEC int32_t
ompi_convertor_personalize( ompi_convertor_t* pConv, uint32_t flags,
size_t* starting_point,
memalloc_fct_t allocfn,
void* userdata );
/*
*
*/
@ -312,9 +303,46 @@ ompi_convertor_set_position( ompi_convertor_t* convertor,
* If the convertor is already at the correct position we are happy.
*/
if( (*position) == convertor->bConverted ) return OMPI_SUCCESS;
/*
* Do not allow the convertor to go outside the data boundaries. This test include
* the check for datatype with size zero as well as for convertors with a count of zero.
*/
if( convertor->local_size <= *position) {
convertor->flags |= CONVERTOR_COMPLETED;
convertor->bConverted = convertor->local_size;
*position = convertor->bConverted;
return OMPI_SUCCESS;
}
/* Remove the completed flag if it's already set */
convertor->flags &= ~CONVERTOR_COMPLETED;
if( (convertor->flags & (DT_FLAG_PREDEFINED | CONVERTOR_HOMOGENEOUS)) ==
(DT_FLAG_PREDEFINED | CONVERTOR_HOMOGENEOUS) ) {
/* basic predefined datatype (contiguous) */
convertor->bConverted = *position;
return OMPI_SUCCESS;
}
return ompi_convertor_set_position_nocheck( convertor, position );
}
/*
*
*/
static inline int32_t
ompi_convertor_personalize( ompi_convertor_t* convertor, uint32_t flags,
size_t* position,
memalloc_fct_t allocfn, void* userdata )
{
convertor->flags |= flags;
convertor->memAlloc_fn = allocfn;
convertor->memAlloc_userdata = userdata;
return ompi_convertor_set_position( convertor, position );
}
/*
*
*/